]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net> | |
7 | * | |
8 | * This is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License version 2.1, as published by the Free Software | |
11 | * Foundation. See file COPYING. | |
12 | * | |
13 | */ | |
14 | ||
15 | #include <errno.h> | |
f67539c2 | 16 | #include <ostream> |
7c673cae | 17 | #include <string> |
11fdf7f2 | 18 | #include <string_view> |
7c673cae FG |
19 | #include <map> |
20 | ||
21 | #include "MDCache.h" | |
22 | #include "MDSRank.h" | |
23 | #include "Server.h" | |
24 | #include "Locker.h" | |
25 | #include "MDLog.h" | |
26 | #include "MDBalancer.h" | |
27 | #include "Migrator.h" | |
28 | #include "ScrubStack.h" | |
29 | ||
30 | #include "SnapClient.h" | |
31 | ||
32 | #include "MDSMap.h" | |
33 | ||
34 | #include "CInode.h" | |
35 | #include "CDir.h" | |
36 | ||
37 | #include "Mutation.h" | |
38 | ||
39 | #include "include/ceph_fs.h" | |
40 | #include "include/filepath.h" | |
181888fb | 41 | #include "include/util.h" |
7c673cae | 42 | |
11fdf7f2 TL |
43 | #include "messages/MClientCaps.h" |
44 | ||
7c673cae FG |
45 | #include "msg/Message.h" |
46 | #include "msg/Messenger.h" | |
47 | ||
181888fb | 48 | #include "common/MemoryModel.h" |
7c673cae | 49 | #include "common/errno.h" |
7c673cae | 50 | #include "common/perf_counters.h" |
181888fb FG |
51 | #include "common/safe_io.h" |
52 | ||
7c673cae FG |
53 | #include "osdc/Journaler.h" |
54 | #include "osdc/Filer.h" | |
55 | ||
56 | #include "events/ESubtreeMap.h" | |
57 | #include "events/EUpdate.h" | |
f67539c2 | 58 | #include "events/EPeerUpdate.h" |
7c673cae FG |
59 | #include "events/EImportFinish.h" |
60 | #include "events/EFragment.h" | |
61 | #include "events/ECommitted.h" | |
9f95a23c | 62 | #include "events/EPurged.h" |
7c673cae FG |
63 | #include "events/ESessions.h" |
64 | ||
7c673cae FG |
65 | #include "InoTable.h" |
66 | ||
67 | #include "common/Timer.h" | |
68 | ||
69 | #include "perfglue/heap_profiler.h" | |
70 | ||
7c673cae FG |
71 | |
72 | #include "common/config.h" | |
11fdf7f2 | 73 | #include "include/ceph_assert.h" |
7c673cae FG |
74 | |
75 | #define dout_context g_ceph_context | |
76 | #define dout_subsys ceph_subsys_mds | |
77 | #undef dout_prefix | |
78 | #define dout_prefix _prefix(_dout, mds) | |
20effc67 TL |
79 | |
80 | using namespace std; | |
81 | ||
7c673cae FG |
82 | static ostream& _prefix(std::ostream *_dout, MDSRank *mds) { |
83 | return *_dout << "mds." << mds->get_nodeid() << ".cache "; | |
84 | } | |
85 | ||
86 | set<int> SimpleLock::empty_gather_set; | |
87 | ||
88 | ||
89 | /** | |
90 | * All non-I/O contexts that require a reference | |
91 | * to an MDCache instance descend from this. | |
92 | */ | |
11fdf7f2 | 93 | class MDCacheContext : public virtual MDSContext { |
7c673cae FG |
94 | protected: |
95 | MDCache *mdcache; | |
96 | MDSRank *get_mds() override | |
97 | { | |
11fdf7f2 | 98 | ceph_assert(mdcache != NULL); |
7c673cae FG |
99 | return mdcache->mds; |
100 | } | |
101 | public: | |
102 | explicit MDCacheContext(MDCache *mdc_) : mdcache(mdc_) {} | |
103 | }; | |
104 | ||
105 | ||
106 | /** | |
107 | * Only for contexts called back from an I/O completion | |
108 | * | |
109 | * Note: duplication of members wrt MDCacheContext, because | |
110 | * it'ls the lesser of two evils compared with introducing | |
111 | * yet another piece of (multiple) inheritance. | |
112 | */ | |
113 | class MDCacheIOContext : public virtual MDSIOContextBase { | |
114 | protected: | |
115 | MDCache *mdcache; | |
116 | MDSRank *get_mds() override | |
117 | { | |
11fdf7f2 | 118 | ceph_assert(mdcache != NULL); |
7c673cae FG |
119 | return mdcache->mds; |
120 | } | |
121 | public: | |
91327a77 AA |
122 | explicit MDCacheIOContext(MDCache *mdc_, bool track=true) : |
123 | MDSIOContextBase(track), mdcache(mdc_) {} | |
7c673cae FG |
124 | }; |
125 | ||
126 | class MDCacheLogContext : public virtual MDSLogContextBase { | |
127 | protected: | |
128 | MDCache *mdcache; | |
129 | MDSRank *get_mds() override | |
130 | { | |
11fdf7f2 | 131 | ceph_assert(mdcache != NULL); |
7c673cae FG |
132 | return mdcache->mds; |
133 | } | |
134 | public: | |
135 | explicit MDCacheLogContext(MDCache *mdc_) : mdcache(mdc_) {} | |
136 | }; | |
137 | ||
138 | MDCache::MDCache(MDSRank *m, PurgeQueue &purge_queue_) : | |
139 | mds(m), | |
9f95a23c | 140 | open_file_table(m), |
7c673cae | 141 | filer(m->objecter, m->finisher), |
a8e16298 | 142 | stray_manager(m, purge_queue_), |
9f95a23c TL |
143 | recovery_queue(m), |
144 | trim_counter(g_conf().get_val<double>("mds_cache_trim_decay_rate")) | |
7c673cae FG |
145 | { |
146 | migrator.reset(new Migrator(mds, this)); | |
7c673cae | 147 | |
11fdf7f2 TL |
148 | max_dir_commit_size = g_conf()->mds_dir_max_commit_size ? |
149 | (g_conf()->mds_dir_max_commit_size << 20) : | |
150 | (0.9 *(g_conf()->osd_max_write_size << 20)); | |
7c673cae | 151 | |
11fdf7f2 TL |
152 | cache_memory_limit = g_conf().get_val<Option::size_t>("mds_cache_memory_limit"); |
153 | cache_reservation = g_conf().get_val<double>("mds_cache_reservation"); | |
154 | cache_health_threshold = g_conf().get_val<double>("mds_health_cache_threshold"); | |
91327a77 | 155 | |
f6b5b4d7 TL |
156 | export_ephemeral_distributed_config = g_conf().get_val<bool>("mds_export_ephemeral_distributed"); |
157 | export_ephemeral_random_config = g_conf().get_val<bool>("mds_export_ephemeral_random"); | |
158 | export_ephemeral_random_max = g_conf().get_val<double>("mds_export_ephemeral_random_max"); | |
159 | ||
20effc67 TL |
160 | symlink_recovery = g_conf().get_val<bool>("mds_symlink_recovery"); |
161 | ||
11fdf7f2 | 162 | lru.lru_set_midpoint(g_conf().get_val<double>("mds_cache_mid")); |
7c673cae | 163 | |
31f18b77 FG |
164 | bottom_lru.lru_set_midpoint(0); |
165 | ||
11fdf7f2 | 166 | decayrate.set_halflife(g_conf()->mds_decay_halflife); |
7c673cae | 167 | |
b3b6e05e | 168 | upkeeper = std::thread(&MDCache::upkeep_main, this); |
7c673cae FG |
169 | } |
170 | ||
171 | MDCache::~MDCache() | |
172 | { | |
173 | if (logger) { | |
174 | g_ceph_context->get_perfcounters_collection()->remove(logger.get()); | |
175 | } | |
eafe8130 TL |
176 | if (upkeeper.joinable()) |
177 | upkeeper.join(); | |
7c673cae FG |
178 | } |
179 | ||
92f5a8d4 | 180 | void MDCache::handle_conf_change(const std::set<std::string>& changed, const MDSMap& mdsmap) |
91327a77 | 181 | { |
f6b5b4d7 | 182 | dout(20) << "config changes: " << changed << dendl; |
91327a77 | 183 | if (changed.count("mds_cache_memory_limit")) |
11fdf7f2 | 184 | cache_memory_limit = g_conf().get_val<Option::size_t>("mds_cache_memory_limit"); |
91327a77 | 185 | if (changed.count("mds_cache_reservation")) |
11fdf7f2 | 186 | cache_reservation = g_conf().get_val<double>("mds_cache_reservation"); |
f67539c2 TL |
187 | |
188 | bool ephemeral_pin_config_changed = false; | |
f6b5b4d7 TL |
189 | if (changed.count("mds_export_ephemeral_distributed")) { |
190 | export_ephemeral_distributed_config = g_conf().get_val<bool>("mds_export_ephemeral_distributed"); | |
191 | dout(10) << "Migrating any ephemeral distributed pinned inodes" << dendl; | |
192 | /* copy to vector to avoid removals during iteration */ | |
f67539c2 | 193 | ephemeral_pin_config_changed = true; |
f6b5b4d7 TL |
194 | } |
195 | if (changed.count("mds_export_ephemeral_random")) { | |
196 | export_ephemeral_random_config = g_conf().get_val<bool>("mds_export_ephemeral_random"); | |
197 | dout(10) << "Migrating any ephemeral random pinned inodes" << dendl; | |
198 | /* copy to vector to avoid removals during iteration */ | |
f67539c2 TL |
199 | ephemeral_pin_config_changed = true; |
200 | } | |
201 | if (ephemeral_pin_config_changed) { | |
f6b5b4d7 | 202 | std::vector<CInode*> migrate; |
f67539c2 | 203 | migrate.assign(export_ephemeral_pins.begin(), export_ephemeral_pins.end()); |
f6b5b4d7 | 204 | for (auto& in : migrate) { |
f67539c2 | 205 | in->maybe_export_pin(true); |
f6b5b4d7 | 206 | } |
f6b5b4d7 TL |
207 | } |
208 | if (changed.count("mds_export_ephemeral_random_max")) { | |
209 | export_ephemeral_random_max = g_conf().get_val<double>("mds_export_ephemeral_random_max"); | |
210 | } | |
91327a77 | 211 | if (changed.count("mds_health_cache_threshold")) |
11fdf7f2 | 212 | cache_health_threshold = g_conf().get_val<double>("mds_health_cache_threshold"); |
91327a77 | 213 | if (changed.count("mds_cache_mid")) |
11fdf7f2 | 214 | lru.lru_set_midpoint(g_conf().get_val<double>("mds_cache_mid")); |
a8e16298 | 215 | if (changed.count("mds_cache_trim_decay_rate")) { |
11fdf7f2 | 216 | trim_counter = DecayCounter(g_conf().get_val<double>("mds_cache_trim_decay_rate")); |
a8e16298 | 217 | } |
20effc67 TL |
218 | if (changed.count("mds_symlink_recovery")) { |
219 | symlink_recovery = g_conf().get_val<bool>("mds_symlink_recovery"); | |
220 | dout(10) << "Storing symlink targets on file object's head " << symlink_recovery << dendl; | |
221 | } | |
7c673cae | 222 | |
92f5a8d4 TL |
223 | migrator->handle_conf_change(changed, mdsmap); |
224 | mds->balancer->handle_conf_change(changed, mdsmap); | |
91327a77 | 225 | } |
7c673cae FG |
226 | |
227 | void MDCache::log_stat() | |
228 | { | |
7c673cae FG |
229 | mds->logger->set(l_mds_inodes, lru.lru_get_size()); |
230 | mds->logger->set(l_mds_inodes_pinned, lru.lru_get_num_pinned()); | |
231 | mds->logger->set(l_mds_inodes_top, lru.lru_get_top()); | |
232 | mds->logger->set(l_mds_inodes_bottom, lru.lru_get_bot()); | |
233 | mds->logger->set(l_mds_inodes_pin_tail, lru.lru_get_pintail()); | |
234 | mds->logger->set(l_mds_inodes_with_caps, num_inodes_with_caps); | |
235 | mds->logger->set(l_mds_caps, Capability::count()); | |
eafe8130 | 236 | if (root) { |
f67539c2 TL |
237 | mds->logger->set(l_mds_root_rfiles, root->get_inode()->rstat.rfiles); |
238 | mds->logger->set(l_mds_root_rbytes, root->get_inode()->rstat.rbytes); | |
239 | mds->logger->set(l_mds_root_rsnaps, root->get_inode()->rstat.rsnaps); | |
eafe8130 | 240 | } |
7c673cae FG |
241 | } |
242 | ||
243 | ||
244 | // | |
245 | ||
246 | bool MDCache::shutdown() | |
247 | { | |
eafe8130 TL |
248 | { |
249 | std::scoped_lock lock(upkeep_mutex); | |
250 | upkeep_trim_shutdown = true; | |
251 | upkeep_cvar.notify_one(); | |
252 | } | |
7c673cae FG |
253 | if (lru.lru_get_size() > 0) { |
254 | dout(7) << "WARNING: mdcache shutdown with non-empty cache" << dendl; | |
255 | //show_cache(); | |
256 | show_subtrees(); | |
257 | //dump(); | |
258 | } | |
259 | return true; | |
260 | } | |
261 | ||
262 | ||
263 | // ==================================================================== | |
264 | // some inode functions | |
265 | ||
a4b75251 | 266 | void MDCache::add_inode(CInode *in) |
7c673cae | 267 | { |
a4b75251 | 268 | // add to inode map |
b32b8144 FG |
269 | if (in->last == CEPH_NOSNAP) { |
270 | auto &p = inode_map[in->ino()]; | |
11fdf7f2 | 271 | ceph_assert(!p); // should be no dup inos! |
b32b8144 FG |
272 | p = in; |
273 | } else { | |
274 | auto &p = snap_inode_map[in->vino()]; | |
11fdf7f2 | 275 | ceph_assert(!p); // should be no dup inos! |
b32b8144 FG |
276 | p = in; |
277 | } | |
7c673cae FG |
278 | |
279 | if (in->ino() < MDS_INO_SYSTEM_BASE) { | |
b3b6e05e | 280 | if (in->ino() == CEPH_INO_ROOT) |
7c673cae FG |
281 | root = in; |
282 | else if (in->ino() == MDS_INO_MDSDIR(mds->get_nodeid())) | |
283 | myin = in; | |
284 | else if (in->is_stray()) { | |
285 | if (MDS_INO_STRAY_OWNER(in->ino()) == mds->get_nodeid()) { | |
286 | strays[MDS_INO_STRAY_INDEX(in->ino())] = in; | |
287 | } | |
288 | } | |
289 | if (in->is_base()) | |
290 | base_inodes.insert(in); | |
291 | } | |
7c673cae FG |
292 | } |
293 | ||
294 | void MDCache::remove_inode(CInode *o) | |
295 | { | |
296 | dout(14) << "remove_inode " << *o << dendl; | |
297 | ||
298 | if (o->get_parent_dn()) { | |
299 | // FIXME: multiple parents? | |
300 | CDentry *dn = o->get_parent_dn(); | |
11fdf7f2 | 301 | ceph_assert(!dn->is_dirty()); |
7c673cae FG |
302 | dn->dir->unlink_inode(dn); // leave dentry ... FIXME? |
303 | } | |
304 | ||
305 | if (o->is_dirty()) | |
306 | o->mark_clean(); | |
307 | if (o->is_dirty_parent()) | |
308 | o->clear_dirty_parent(); | |
309 | ||
310 | o->clear_scatter_dirty(); | |
311 | ||
f91f0fd5 TL |
312 | o->clear_clientwriteable(); |
313 | ||
7c673cae FG |
314 | o->item_open_file.remove_myself(); |
315 | ||
31f18b77 FG |
316 | if (o->state_test(CInode::STATE_QUEUEDEXPORTPIN)) |
317 | export_pin_queue.erase(o); | |
7c673cae | 318 | |
eafe8130 TL |
319 | if (o->state_test(CInode::STATE_DELAYEDEXPORTPIN)) |
320 | export_pin_delayed_queue.erase(o); | |
321 | ||
f67539c2 | 322 | o->clear_ephemeral_pin(true, true); |
f6b5b4d7 | 323 | |
7c673cae | 324 | // remove from inode map |
11fdf7f2 | 325 | if (o->last == CEPH_NOSNAP) { |
b32b8144 | 326 | inode_map.erase(o->ino()); |
11fdf7f2 TL |
327 | } else { |
328 | o->item_caps.remove_myself(); | |
b32b8144 | 329 | snap_inode_map.erase(o->vino()); |
11fdf7f2 | 330 | } |
7c673cae FG |
331 | |
332 | if (o->ino() < MDS_INO_SYSTEM_BASE) { | |
333 | if (o == root) root = 0; | |
334 | if (o == myin) myin = 0; | |
335 | if (o->is_stray()) { | |
336 | if (MDS_INO_STRAY_OWNER(o->ino()) == mds->get_nodeid()) { | |
337 | strays[MDS_INO_STRAY_INDEX(o->ino())] = 0; | |
338 | } | |
339 | } | |
340 | if (o->is_base()) | |
341 | base_inodes.erase(o); | |
11fdf7f2 | 342 | } |
7c673cae FG |
343 | |
344 | // delete it | |
11fdf7f2 | 345 | ceph_assert(o->get_num_ref() == 0); |
7c673cae FG |
346 | delete o; |
347 | } | |
348 | ||
349 | file_layout_t MDCache::gen_default_file_layout(const MDSMap &mdsmap) | |
350 | { | |
351 | file_layout_t result = file_layout_t::get_default(); | |
352 | result.pool_id = mdsmap.get_first_data_pool(); | |
353 | return result; | |
354 | } | |
355 | ||
356 | file_layout_t MDCache::gen_default_log_layout(const MDSMap &mdsmap) | |
357 | { | |
358 | file_layout_t result = file_layout_t::get_default(); | |
359 | result.pool_id = mdsmap.get_metadata_pool(); | |
11fdf7f2 TL |
360 | if (g_conf()->mds_log_segment_size > 0) { |
361 | result.object_size = g_conf()->mds_log_segment_size; | |
362 | result.stripe_unit = g_conf()->mds_log_segment_size; | |
7c673cae FG |
363 | } |
364 | return result; | |
365 | } | |
366 | ||
367 | void MDCache::init_layouts() | |
368 | { | |
369 | default_file_layout = gen_default_file_layout(*(mds->mdsmap)); | |
370 | default_log_layout = gen_default_log_layout(*(mds->mdsmap)); | |
371 | } | |
372 | ||
f67539c2 | 373 | void MDCache::create_unlinked_system_inode(CInode *in, inodeno_t ino, int mode) const |
7c673cae | 374 | { |
f67539c2 TL |
375 | auto _inode = in->_get_inode(); |
376 | _inode->ino = ino; | |
377 | _inode->version = 1; | |
378 | _inode->xattr_version = 1; | |
379 | _inode->mode = 0500 | mode; | |
380 | _inode->size = 0; | |
381 | _inode->ctime = _inode->mtime = _inode->btime = ceph_clock_now(); | |
382 | _inode->nlink = 1; | |
383 | _inode->truncate_size = -1ull; | |
384 | _inode->change_attr = 0; | |
385 | _inode->export_pin = MDS_RANK_NONE; | |
7c673cae | 386 | |
92f5a8d4 | 387 | // FIPS zeroization audit 20191117: this memset is not security related. |
f67539c2 TL |
388 | memset(&_inode->dir_layout, 0, sizeof(_inode->dir_layout)); |
389 | if (_inode->is_dir()) { | |
390 | _inode->dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; | |
391 | _inode->rstat.rsubdirs = 1; /* itself */ | |
392 | _inode->rstat.rctime = in->get_inode()->ctime; | |
7c673cae | 393 | } else { |
f67539c2 TL |
394 | _inode->layout = default_file_layout; |
395 | ++_inode->rstat.rfiles; | |
7c673cae | 396 | } |
f67539c2 | 397 | _inode->accounted_rstat = _inode->rstat; |
7c673cae FG |
398 | |
399 | if (in->is_base()) { | |
400 | if (in->is_root()) | |
401 | in->inode_auth = mds_authority_t(mds->get_nodeid(), CDIR_AUTH_UNKNOWN); | |
402 | else | |
403 | in->inode_auth = mds_authority_t(mds_rank_t(in->ino() - MDS_INO_MDSDIR_OFFSET), CDIR_AUTH_UNKNOWN); | |
404 | in->open_snaprealm(); // empty snaprealm | |
11fdf7f2 | 405 | ceph_assert(!in->snaprealm->parent); // created its own |
7c673cae FG |
406 | in->snaprealm->srnode.seq = 1; |
407 | } | |
408 | } | |
409 | ||
410 | CInode *MDCache::create_system_inode(inodeno_t ino, int mode) | |
411 | { | |
412 | dout(0) << "creating system inode with ino:" << ino << dendl; | |
413 | CInode *in = new CInode(this); | |
414 | create_unlinked_system_inode(in, ino, mode); | |
415 | add_inode(in); | |
416 | return in; | |
417 | } | |
418 | ||
419 | CInode *MDCache::create_root_inode() | |
420 | { | |
b3b6e05e | 421 | CInode *in = create_system_inode(CEPH_INO_ROOT, S_IFDIR|0755); |
f67539c2 TL |
422 | auto _inode = in->_get_inode(); |
423 | _inode->uid = g_conf()->mds_root_ino_uid; | |
424 | _inode->gid = g_conf()->mds_root_ino_gid; | |
425 | _inode->layout = default_file_layout; | |
426 | _inode->layout.pool_id = mds->mdsmap->get_first_data_pool(); | |
427 | return in; | |
7c673cae FG |
428 | } |
429 | ||
430 | void MDCache::create_empty_hierarchy(MDSGather *gather) | |
431 | { | |
432 | // create root dir | |
433 | CInode *root = create_root_inode(); | |
434 | ||
435 | // force empty root dir | |
436 | CDir *rootdir = root->get_or_open_dirfrag(this, frag_t()); | |
437 | adjust_subtree_auth(rootdir, mds->get_nodeid()); | |
438 | rootdir->dir_rep = CDir::REP_ALL; //NONE; | |
439 | ||
f67539c2 TL |
440 | ceph_assert(rootdir->get_fnode()->accounted_fragstat == rootdir->get_fnode()->fragstat); |
441 | ceph_assert(rootdir->get_fnode()->fragstat == root->get_inode()->dirstat); | |
442 | ceph_assert(rootdir->get_fnode()->accounted_rstat == rootdir->get_fnode()->rstat); | |
f64942e4 AA |
443 | /* Do no update rootdir rstat information of the fragment, rstat upkeep magic |
444 | * assume version 0 is stale/invalid. | |
445 | */ | |
7c673cae FG |
446 | |
447 | rootdir->mark_complete(); | |
f67539c2 TL |
448 | rootdir->_get_fnode()->version = rootdir->pre_dirty(); |
449 | rootdir->mark_dirty(mds->mdlog->get_current_segment()); | |
7c673cae FG |
450 | rootdir->commit(0, gather->new_sub()); |
451 | ||
f67539c2 | 452 | root->store(gather->new_sub()); |
b3b6e05e TL |
453 | root->mark_dirty_parent(mds->mdlog->get_current_segment(), true); |
454 | root->store_backtrace(gather->new_sub()); | |
7c673cae FG |
455 | } |
456 | ||
457 | void MDCache::create_mydir_hierarchy(MDSGather *gather) | |
458 | { | |
459 | // create mds dir | |
460 | CInode *my = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR); | |
461 | ||
462 | CDir *mydir = my->get_or_open_dirfrag(this, frag_t()); | |
f67539c2 TL |
463 | auto mydir_fnode = mydir->_get_fnode(); |
464 | ||
7c673cae FG |
465 | adjust_subtree_auth(mydir, mds->get_nodeid()); |
466 | ||
467 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
468 | ||
469 | // stray dir | |
470 | for (int i = 0; i < NUM_STRAY; ++i) { | |
471 | CInode *stray = create_system_inode(MDS_INO_STRAY(mds->get_nodeid(), i), S_IFDIR); | |
472 | CDir *straydir = stray->get_or_open_dirfrag(this, frag_t()); | |
f67539c2 TL |
473 | CachedStackStringStream css; |
474 | *css << "stray" << i; | |
475 | CDentry *sdn = mydir->add_primary_dentry(css->str(), stray, ""); | |
7c673cae FG |
476 | sdn->_mark_dirty(mds->mdlog->get_current_segment()); |
477 | ||
f67539c2 | 478 | stray->_get_inode()->dirstat = straydir->get_fnode()->fragstat; |
7c673cae | 479 | |
f67539c2 TL |
480 | mydir_fnode->rstat.add(stray->get_inode()->rstat); |
481 | mydir_fnode->fragstat.nsubdirs++; | |
7c673cae FG |
482 | // save them |
483 | straydir->mark_complete(); | |
f67539c2 TL |
484 | straydir->_get_fnode()->version = straydir->pre_dirty(); |
485 | straydir->mark_dirty(ls); | |
7c673cae | 486 | straydir->commit(0, gather->new_sub()); |
28e407b8 | 487 | stray->mark_dirty_parent(ls, true); |
7c673cae FG |
488 | stray->store_backtrace(gather->new_sub()); |
489 | } | |
490 | ||
f67539c2 TL |
491 | mydir_fnode->accounted_fragstat = mydir->get_fnode()->fragstat; |
492 | mydir_fnode->accounted_rstat = mydir->get_fnode()->rstat; | |
7c673cae | 493 | |
f67539c2 TL |
494 | auto inode = myin->_get_inode(); |
495 | inode->dirstat = mydir->get_fnode()->fragstat; | |
496 | inode->rstat = mydir->get_fnode()->rstat; | |
497 | ++inode->rstat.rsubdirs; | |
498 | inode->accounted_rstat = inode->rstat; | |
7c673cae FG |
499 | |
500 | mydir->mark_complete(); | |
f67539c2 TL |
501 | mydir_fnode->version = mydir->pre_dirty(); |
502 | mydir->mark_dirty(ls); | |
7c673cae FG |
503 | mydir->commit(0, gather->new_sub()); |
504 | ||
505 | myin->store(gather->new_sub()); | |
506 | } | |
507 | ||
508 | struct C_MDC_CreateSystemFile : public MDCacheLogContext { | |
509 | MutationRef mut; | |
510 | CDentry *dn; | |
511 | version_t dpv; | |
11fdf7f2 TL |
512 | MDSContext *fin; |
513 | C_MDC_CreateSystemFile(MDCache *c, MutationRef& mu, CDentry *d, version_t v, MDSContext *f) : | |
7c673cae FG |
514 | MDCacheLogContext(c), mut(mu), dn(d), dpv(v), fin(f) {} |
515 | void finish(int r) override { | |
516 | mdcache->_create_system_file_finish(mut, dn, dpv, fin); | |
517 | } | |
518 | }; | |
519 | ||
11fdf7f2 | 520 | void MDCache::_create_system_file(CDir *dir, std::string_view name, CInode *in, MDSContext *fin) |
7c673cae FG |
521 | { |
522 | dout(10) << "_create_system_file " << name << " in " << *dir << dendl; | |
523 | CDentry *dn = dir->add_null_dentry(name); | |
524 | ||
525 | dn->push_projected_linkage(in); | |
526 | version_t dpv = dn->pre_dirty(); | |
527 | ||
528 | CDir *mdir = 0; | |
f67539c2 TL |
529 | auto inode = in->_get_inode(); |
530 | if (in->is_dir()) { | |
531 | inode->rstat.rsubdirs = 1; | |
7c673cae FG |
532 | |
533 | mdir = in->get_or_open_dirfrag(this, frag_t()); | |
534 | mdir->mark_complete(); | |
f67539c2 TL |
535 | mdir->_get_fnode()->version = mdir->pre_dirty(); |
536 | } else { | |
537 | inode->rstat.rfiles = 1; | |
538 | } | |
539 | ||
540 | inode->version = dn->pre_dirty(); | |
7c673cae FG |
541 | |
542 | SnapRealm *realm = dir->get_inode()->find_snaprealm(); | |
543 | dn->first = in->first = realm->get_newest_seq() + 1; | |
544 | ||
545 | MutationRef mut(new MutationImpl()); | |
546 | ||
547 | // force some locks. hacky. | |
548 | mds->locker->wrlock_force(&dir->inode->filelock, mut); | |
549 | mds->locker->wrlock_force(&dir->inode->nestlock, mut); | |
550 | ||
551 | mut->ls = mds->mdlog->get_current_segment(); | |
552 | EUpdate *le = new EUpdate(mds->mdlog, "create system file"); | |
553 | mds->mdlog->start_entry(le); | |
554 | ||
555 | if (!in->is_mdsdir()) { | |
556 | predirty_journal_parents(mut, &le->metablob, in, dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
557 | le->metablob.add_primary_dentry(dn, in, true); | |
558 | } else { | |
559 | predirty_journal_parents(mut, &le->metablob, in, dir, PREDIRTY_DIR, 1); | |
560 | journal_dirty_inode(mut.get(), &le->metablob, in); | |
561 | dn->push_projected_linkage(in->ino(), in->d_type()); | |
562 | le->metablob.add_remote_dentry(dn, true, in->ino(), in->d_type()); | |
563 | le->metablob.add_root(true, in); | |
564 | } | |
565 | if (mdir) | |
566 | le->metablob.add_new_dir(mdir); // dirty AND complete AND new | |
567 | ||
568 | mds->mdlog->submit_entry(le, new C_MDC_CreateSystemFile(this, mut, dn, dpv, fin)); | |
569 | mds->mdlog->flush(); | |
570 | } | |
571 | ||
11fdf7f2 | 572 | void MDCache::_create_system_file_finish(MutationRef& mut, CDentry *dn, version_t dpv, MDSContext *fin) |
7c673cae FG |
573 | { |
574 | dout(10) << "_create_system_file_finish " << *dn << dendl; | |
575 | ||
576 | dn->pop_projected_linkage(); | |
577 | dn->mark_dirty(dpv, mut->ls); | |
578 | ||
579 | CInode *in = dn->get_linkage()->get_inode(); | |
f67539c2 | 580 | in->mark_dirty(mut->ls); |
7c673cae | 581 | |
f67539c2 | 582 | if (in->is_dir()) { |
7c673cae | 583 | CDir *dir = in->get_dirfrag(frag_t()); |
11fdf7f2 | 584 | ceph_assert(dir); |
f67539c2 | 585 | dir->mark_dirty(mut->ls); |
7c673cae FG |
586 | dir->mark_new(mut->ls); |
587 | } | |
588 | ||
589 | mut->apply(); | |
590 | mds->locker->drop_locks(mut.get()); | |
591 | mut->cleanup(); | |
592 | ||
593 | fin->complete(0); | |
594 | ||
595 | //if (dir && MDS_INO_IS_MDSDIR(in->ino())) | |
596 | //migrator->export_dir(dir, (int)in->ino() - MDS_INO_MDSDIR_OFFSET); | |
597 | } | |
598 | ||
599 | ||
600 | ||
601 | struct C_MDS_RetryOpenRoot : public MDSInternalContext { | |
602 | MDCache *cache; | |
603 | explicit C_MDS_RetryOpenRoot(MDCache *c) : MDSInternalContext(c->mds), cache(c) {} | |
604 | void finish(int r) override { | |
605 | if (r < 0) { | |
606 | // If we can't open root, something disastrous has happened: mark | |
607 | // this rank damaged for operator intervention. Note that | |
608 | // it is not okay to call suicide() here because we are in | |
609 | // a Finisher callback. | |
610 | cache->mds->damaged(); | |
611 | ceph_abort(); // damaged should never return | |
612 | } else { | |
613 | cache->open_root(); | |
614 | } | |
615 | } | |
616 | }; | |
617 | ||
11fdf7f2 | 618 | void MDCache::open_root_inode(MDSContext *c) |
7c673cae FG |
619 | { |
620 | if (mds->get_nodeid() == mds->mdsmap->get_root()) { | |
621 | CInode *in; | |
b3b6e05e | 622 | in = create_system_inode(CEPH_INO_ROOT, S_IFDIR|0755); // initially inaccurate! |
7c673cae FG |
623 | in->fetch(c); |
624 | } else { | |
b3b6e05e | 625 | discover_base_ino(CEPH_INO_ROOT, c, mds->mdsmap->get_root()); |
7c673cae FG |
626 | } |
627 | } | |
628 | ||
11fdf7f2 | 629 | void MDCache::open_mydir_inode(MDSContext *c) |
7c673cae | 630 | { |
7c673cae | 631 | CInode *in = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR|0755); // initially inaccurate! |
11fdf7f2 | 632 | in->fetch(c); |
7c673cae FG |
633 | } |
634 | ||
11fdf7f2 | 635 | void MDCache::open_mydir_frag(MDSContext *c) |
28e407b8 AA |
636 | { |
637 | open_mydir_inode( | |
638 | new MDSInternalContextWrapper(mds, | |
9f95a23c | 639 | new LambdaContext([this, c](int r) { |
28e407b8 AA |
640 | if (r < 0) { |
641 | c->complete(r); | |
642 | return; | |
643 | } | |
644 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); | |
11fdf7f2 | 645 | ceph_assert(mydir); |
28e407b8 AA |
646 | adjust_subtree_auth(mydir, mds->get_nodeid()); |
647 | mydir->fetch(c); | |
648 | }) | |
649 | ) | |
650 | ); | |
651 | } | |
652 | ||
7c673cae FG |
653 | void MDCache::open_root() |
654 | { | |
655 | dout(10) << "open_root" << dendl; | |
656 | ||
657 | if (!root) { | |
658 | open_root_inode(new C_MDS_RetryOpenRoot(this)); | |
659 | return; | |
660 | } | |
661 | if (mds->get_nodeid() == mds->mdsmap->get_root()) { | |
11fdf7f2 | 662 | ceph_assert(root->is_auth()); |
7c673cae | 663 | CDir *rootdir = root->get_or_open_dirfrag(this, frag_t()); |
11fdf7f2 | 664 | ceph_assert(rootdir); |
7c673cae FG |
665 | if (!rootdir->is_subtree_root()) |
666 | adjust_subtree_auth(rootdir, mds->get_nodeid()); | |
667 | if (!rootdir->is_complete()) { | |
668 | rootdir->fetch(new C_MDS_RetryOpenRoot(this)); | |
669 | return; | |
670 | } | |
671 | } else { | |
11fdf7f2 | 672 | ceph_assert(!root->is_auth()); |
7c673cae FG |
673 | CDir *rootdir = root->get_dirfrag(frag_t()); |
674 | if (!rootdir) { | |
224ce89b | 675 | open_remote_dirfrag(root, frag_t(), new C_MDS_RetryOpenRoot(this)); |
7c673cae FG |
676 | return; |
677 | } | |
678 | } | |
679 | ||
680 | if (!myin) { | |
681 | CInode *in = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR|0755); // initially inaccurate! | |
682 | in->fetch(new C_MDS_RetryOpenRoot(this)); | |
683 | return; | |
684 | } | |
685 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); | |
11fdf7f2 | 686 | ceph_assert(mydir); |
7c673cae FG |
687 | adjust_subtree_auth(mydir, mds->get_nodeid()); |
688 | ||
689 | populate_mydir(); | |
690 | } | |
691 | ||
f67539c2 TL |
692 | void MDCache::advance_stray() { |
693 | // check whether the directory has been fragmented | |
694 | if (stray_fragmenting_index >= 0) { | |
695 | auto&& dfs = strays[stray_fragmenting_index]->get_dirfrags(); | |
696 | bool any_fragmenting = false; | |
697 | for (const auto& dir : dfs) { | |
698 | if (dir->state_test(CDir::STATE_FRAGMENTING) || | |
699 | mds->balancer->is_fragment_pending(dir->dirfrag())) { | |
700 | any_fragmenting = true; | |
701 | break; | |
702 | } | |
703 | } | |
704 | if (!any_fragmenting) | |
705 | stray_fragmenting_index = -1; | |
706 | } | |
707 | ||
708 | for (int i = 1; i < NUM_STRAY; i++){ | |
709 | stray_index = (stray_index + i) % NUM_STRAY; | |
710 | if (stray_index != stray_fragmenting_index) | |
711 | break; | |
712 | } | |
713 | ||
714 | if (stray_fragmenting_index == -1 && is_open()) { | |
715 | // Fragment later stray dir in advance. We don't choose past | |
716 | // stray dir because in-flight requests may still use it. | |
717 | stray_fragmenting_index = (stray_index + 3) % NUM_STRAY; | |
718 | auto&& dfs = strays[stray_fragmenting_index]->get_dirfrags(); | |
719 | bool any_fragmenting = false; | |
720 | for (const auto& dir : dfs) { | |
721 | if (dir->should_split()) { | |
722 | mds->balancer->queue_split(dir, true); | |
723 | any_fragmenting = true; | |
724 | } else if (dir->should_merge()) { | |
725 | mds->balancer->queue_merge(dir); | |
726 | any_fragmenting = true; | |
727 | } | |
728 | } | |
729 | if (!any_fragmenting) | |
730 | stray_fragmenting_index = -1; | |
731 | } | |
732 | ||
733 | dout(10) << "advance_stray to index " << stray_index | |
734 | << " fragmenting index " << stray_fragmenting_index << dendl; | |
735 | } | |
736 | ||
7c673cae FG |
737 | void MDCache::populate_mydir() |
738 | { | |
11fdf7f2 | 739 | ceph_assert(myin); |
7c673cae | 740 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); |
11fdf7f2 | 741 | ceph_assert(mydir); |
7c673cae FG |
742 | |
743 | dout(10) << "populate_mydir " << *mydir << dendl; | |
744 | ||
745 | if (!mydir->is_complete()) { | |
746 | mydir->fetch(new C_MDS_RetryOpenRoot(this)); | |
747 | return; | |
748 | } | |
749 | ||
750 | if (mydir->get_version() == 0 && mydir->state_test(CDir::STATE_BADFRAG)) { | |
751 | // A missing dirfrag, we will recreate it. Before that, we must dirty | |
752 | // it before dirtying any of the strays we create within it. | |
753 | mds->clog->warn() << "fragment " << mydir->dirfrag() << " was unreadable, " | |
754 | "recreating it now"; | |
755 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
756 | mydir->state_clear(CDir::STATE_BADFRAG); | |
757 | mydir->mark_complete(); | |
f67539c2 TL |
758 | mydir->_get_fnode()->version = mydir->pre_dirty(); |
759 | mydir->mark_dirty(ls); | |
7c673cae FG |
760 | } |
761 | ||
762 | // open or create stray | |
763 | uint64_t num_strays = 0; | |
764 | for (int i = 0; i < NUM_STRAY; ++i) { | |
f67539c2 TL |
765 | CachedStackStringStream css; |
766 | *css << "stray" << i; | |
767 | CDentry *straydn = mydir->lookup(css->str()); | |
7c673cae FG |
768 | |
769 | // allow for older fs's with stray instead of stray0 | |
770 | if (straydn == NULL && i == 0) | |
771 | straydn = mydir->lookup("stray"); | |
772 | ||
773 | if (!straydn || !straydn->get_linkage()->get_inode()) { | |
f67539c2 | 774 | _create_system_file(mydir, css->strv(), create_system_inode(MDS_INO_STRAY(mds->get_nodeid(), i), S_IFDIR), |
7c673cae FG |
775 | new C_MDS_RetryOpenRoot(this)); |
776 | return; | |
777 | } | |
11fdf7f2 TL |
778 | ceph_assert(straydn); |
779 | ceph_assert(strays[i]); | |
7c673cae FG |
780 | // we make multiple passes through this method; make sure we only pin each stray once. |
781 | if (!strays[i]->state_test(CInode::STATE_STRAYPINNED)) { | |
782 | strays[i]->get(CInode::PIN_STRAY); | |
783 | strays[i]->state_set(CInode::STATE_STRAYPINNED); | |
784 | strays[i]->get_stickydirs(); | |
785 | } | |
786 | dout(20) << " stray num " << i << " is " << *strays[i] << dendl; | |
787 | ||
788 | // open all frags | |
11fdf7f2 TL |
789 | frag_vec_t leaves; |
790 | strays[i]->dirfragtree.get_leaves(leaves); | |
791 | for (const auto& leaf : leaves) { | |
792 | CDir *dir = strays[i]->get_dirfrag(leaf); | |
7c673cae | 793 | if (!dir) { |
11fdf7f2 | 794 | dir = strays[i]->get_or_open_dirfrag(this, leaf); |
7c673cae FG |
795 | } |
796 | ||
797 | // DamageTable applies special handling to strays: it will | |
798 | // have damaged() us out if one is damaged. | |
11fdf7f2 | 799 | ceph_assert(!dir->state_test(CDir::STATE_BADFRAG)); |
7c673cae FG |
800 | |
801 | if (dir->get_version() == 0) { | |
802 | dir->fetch(new C_MDS_RetryOpenRoot(this)); | |
803 | return; | |
804 | } | |
805 | ||
806 | if (dir->get_frag_size() > 0) | |
807 | num_strays += dir->get_frag_size(); | |
808 | } | |
809 | } | |
810 | ||
7c673cae FG |
811 | // okay! |
812 | dout(10) << "populate_mydir done" << dendl; | |
11fdf7f2 | 813 | ceph_assert(!open); |
7c673cae FG |
814 | open = true; |
815 | mds->queue_waiters(waiting_for_open); | |
816 | ||
11fdf7f2 TL |
817 | stray_manager.set_num_strays(num_strays); |
818 | stray_manager.activate(); | |
819 | ||
7c673cae FG |
820 | scan_stray_dir(); |
821 | } | |
822 | ||
11fdf7f2 | 823 | void MDCache::open_foreign_mdsdir(inodeno_t ino, MDSContext *fin) |
7c673cae FG |
824 | { |
825 | discover_base_ino(ino, fin, mds_rank_t(ino & (MAX_MDS-1))); | |
826 | } | |
827 | ||
828 | CDir *MDCache::get_stray_dir(CInode *in) | |
829 | { | |
830 | string straydname; | |
831 | in->name_stray_dentry(straydname); | |
832 | ||
833 | CInode *strayi = get_stray(); | |
11fdf7f2 | 834 | ceph_assert(strayi); |
7c673cae FG |
835 | frag_t fg = strayi->pick_dirfrag(straydname); |
836 | CDir *straydir = strayi->get_dirfrag(fg); | |
11fdf7f2 | 837 | ceph_assert(straydir); |
7c673cae FG |
838 | return straydir; |
839 | } | |
840 | ||
11fdf7f2 | 841 | MDSCacheObject *MDCache::get_object(const MDSCacheObjectInfo &info) |
7c673cae FG |
842 | { |
843 | // inode? | |
844 | if (info.ino) | |
845 | return get_inode(info.ino, info.snapid); | |
846 | ||
847 | // dir or dentry. | |
848 | CDir *dir = get_dirfrag(info.dirfrag); | |
849 | if (!dir) return 0; | |
850 | ||
851 | if (info.dname.length()) | |
852 | return dir->lookup(info.dname, info.snapid); | |
853 | else | |
854 | return dir; | |
855 | } | |
856 | ||
857 | ||
f6b5b4d7 TL |
858 | // ==================================================================== |
859 | // consistent hash ring | |
860 | ||
861 | /* | |
862 | * hashing implementation based on Lamping and Veach's Jump Consistent Hash: https://arxiv.org/pdf/1406.2294.pdf | |
863 | */ | |
f67539c2 | 864 | mds_rank_t MDCache::hash_into_rank_bucket(inodeno_t ino, frag_t fg) |
f6b5b4d7 TL |
865 | { |
866 | const mds_rank_t max_mds = mds->mdsmap->get_max_mds(); | |
867 | uint64_t hash = rjhash64(ino); | |
f67539c2 TL |
868 | if (fg) |
869 | hash = rjhash64(hash + rjhash64(fg.value())); | |
870 | ||
f6b5b4d7 TL |
871 | int64_t b = -1, j = 0; |
872 | while (j < max_mds) { | |
873 | b = j; | |
874 | hash = hash*2862933555777941757ULL + 1; | |
875 | j = (b + 1) * (double(1LL << 31) / double((hash >> 33) + 1)); | |
876 | } | |
877 | // verify bounds before returning | |
878 | auto result = mds_rank_t(b); | |
879 | ceph_assert(result >= 0 && result < max_mds); | |
880 | return result; | |
881 | } | |
7c673cae FG |
882 | |
883 | ||
884 | // ==================================================================== | |
885 | // subtree management | |
886 | ||
7c673cae FG |
887 | /* |
888 | * adjust the dir_auth of a subtree. | |
889 | * merge with parent and/or child subtrees, if is it appropriate. | |
890 | * merge can ONLY happen if both parent and child have unambiguous auth. | |
891 | */ | |
28e407b8 | 892 | void MDCache::adjust_subtree_auth(CDir *dir, mds_authority_t auth, bool adjust_pop) |
7c673cae FG |
893 | { |
894 | dout(7) << "adjust_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
895 | << " on " << *dir << dendl; | |
896 | ||
7c673cae FG |
897 | show_subtrees(); |
898 | ||
899 | CDir *root; | |
900 | if (dir->inode->is_base()) { | |
901 | root = dir; // bootstrap hack. | |
902 | if (subtrees.count(root) == 0) { | |
903 | subtrees[root]; | |
904 | root->get(CDir::PIN_SUBTREE); | |
905 | } | |
906 | } else { | |
907 | root = get_subtree_root(dir); // subtree root | |
908 | } | |
11fdf7f2 TL |
909 | ceph_assert(root); |
910 | ceph_assert(subtrees.count(root)); | |
7c673cae FG |
911 | dout(7) << " current root is " << *root << dendl; |
912 | ||
913 | if (root == dir) { | |
914 | // i am already a subtree. | |
915 | dir->set_dir_auth(auth); | |
916 | } else { | |
917 | // i am a new subtree. | |
918 | dout(10) << " new subtree at " << *dir << dendl; | |
11fdf7f2 | 919 | ceph_assert(subtrees.count(dir) == 0); |
7c673cae FG |
920 | subtrees[dir]; // create empty subtree bounds list for me. |
921 | dir->get(CDir::PIN_SUBTREE); | |
922 | ||
923 | // set dir_auth | |
924 | dir->set_dir_auth(auth); | |
925 | ||
926 | // move items nested beneath me, under me. | |
927 | set<CDir*>::iterator p = subtrees[root].begin(); | |
928 | while (p != subtrees[root].end()) { | |
929 | set<CDir*>::iterator next = p; | |
930 | ++next; | |
931 | if (get_subtree_root((*p)->get_parent_dir()) == dir) { | |
932 | // move under me | |
933 | dout(10) << " claiming child bound " << **p << dendl; | |
934 | subtrees[dir].insert(*p); | |
935 | subtrees[root].erase(p); | |
936 | } | |
937 | p = next; | |
938 | } | |
939 | ||
940 | // i am a bound of the parent subtree. | |
941 | subtrees[root].insert(dir); | |
942 | ||
943 | // i am now the subtree root. | |
944 | root = dir; | |
945 | ||
946 | // adjust recursive pop counters | |
28e407b8 | 947 | if (adjust_pop && dir->is_auth()) { |
7c673cae FG |
948 | CDir *p = dir->get_parent_dir(); |
949 | while (p) { | |
11fdf7f2 | 950 | p->pop_auth_subtree.sub(dir->pop_auth_subtree); |
7c673cae FG |
951 | if (p->is_subtree_root()) break; |
952 | p = p->inode->get_parent_dir(); | |
953 | } | |
954 | } | |
7c673cae FG |
955 | } |
956 | ||
957 | show_subtrees(); | |
958 | } | |
959 | ||
960 | ||
961 | void MDCache::try_subtree_merge(CDir *dir) | |
962 | { | |
963 | dout(7) << "try_subtree_merge " << *dir << dendl; | |
b32b8144 FG |
964 | // record my old bounds |
965 | auto oldbounds = subtrees.at(dir); | |
7c673cae | 966 | |
224ce89b | 967 | set<CInode*> to_eval; |
7c673cae | 968 | // try merge at my root |
224ce89b | 969 | try_subtree_merge_at(dir, &to_eval); |
7c673cae FG |
970 | |
971 | // try merge at my old bounds | |
224ce89b WB |
972 | for (auto bound : oldbounds) |
973 | try_subtree_merge_at(bound, &to_eval); | |
974 | ||
975 | if (!(mds->is_any_replay() || mds->is_resolve())) { | |
976 | for(auto in : to_eval) | |
977 | eval_subtree_root(in); | |
978 | } | |
7c673cae FG |
979 | } |
980 | ||
28e407b8 | 981 | void MDCache::try_subtree_merge_at(CDir *dir, set<CInode*> *to_eval, bool adjust_pop) |
7c673cae FG |
982 | { |
983 | dout(10) << "try_subtree_merge_at " << *dir << dendl; | |
b32b8144 FG |
984 | |
985 | if (dir->dir_auth.second != CDIR_AUTH_UNKNOWN || | |
986 | dir->state_test(CDir::STATE_EXPORTBOUND) || | |
987 | dir->state_test(CDir::STATE_AUXSUBTREE)) | |
988 | return; | |
989 | ||
990 | auto it = subtrees.find(dir); | |
11fdf7f2 | 991 | ceph_assert(it != subtrees.end()); |
7c673cae | 992 | |
7c673cae FG |
993 | // merge with parent? |
994 | CDir *parent = dir; | |
995 | if (!dir->inode->is_base()) | |
996 | parent = get_subtree_root(dir->get_parent_dir()); | |
997 | ||
b32b8144 FG |
998 | if (parent != dir && // we have a parent, |
999 | parent->dir_auth == dir->dir_auth) { // auth matches, | |
7c673cae FG |
1000 | // merge with parent. |
1001 | dout(10) << " subtree merge at " << *dir << dendl; | |
1002 | dir->set_dir_auth(CDIR_AUTH_DEFAULT); | |
1003 | ||
1004 | // move our bounds under the parent | |
b32b8144 | 1005 | subtrees[parent].insert(it->second.begin(), it->second.end()); |
7c673cae FG |
1006 | |
1007 | // we are no longer a subtree or bound | |
1008 | dir->put(CDir::PIN_SUBTREE); | |
b32b8144 | 1009 | subtrees.erase(it); |
7c673cae FG |
1010 | subtrees[parent].erase(dir); |
1011 | ||
1012 | // adjust popularity? | |
28e407b8 | 1013 | if (adjust_pop && dir->is_auth()) { |
28e407b8 | 1014 | CDir *cur = dir; |
7c673cae FG |
1015 | CDir *p = dir->get_parent_dir(); |
1016 | while (p) { | |
11fdf7f2 | 1017 | p->pop_auth_subtree.add(dir->pop_auth_subtree); |
28e407b8 | 1018 | p->pop_lru_subdirs.push_front(&cur->get_inode()->item_pop_lru); |
7c673cae | 1019 | if (p->is_subtree_root()) break; |
28e407b8 | 1020 | cur = p; |
7c673cae FG |
1021 | p = p->inode->get_parent_dir(); |
1022 | } | |
1023 | } | |
1024 | ||
224ce89b WB |
1025 | if (to_eval && dir->get_inode()->is_auth()) |
1026 | to_eval->insert(dir->get_inode()); | |
7c673cae | 1027 | |
181888fb FG |
1028 | show_subtrees(15); |
1029 | } | |
7c673cae FG |
1030 | } |
1031 | ||
7c673cae FG |
1032 | void MDCache::eval_subtree_root(CInode *diri) |
1033 | { | |
1034 | // evaluate subtree inode filelock? | |
1035 | // (we should scatter the filelock on subtree bounds) | |
11fdf7f2 | 1036 | ceph_assert(diri->is_auth()); |
224ce89b | 1037 | mds->locker->try_eval(diri, CEPH_LOCK_IFILE | CEPH_LOCK_INEST); |
7c673cae FG |
1038 | } |
1039 | ||
1040 | ||
11fdf7f2 | 1041 | void MDCache::adjust_bounded_subtree_auth(CDir *dir, const set<CDir*>& bounds, mds_authority_t auth) |
7c673cae FG |
1042 | { |
1043 | dout(7) << "adjust_bounded_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
1044 | << " on " << *dir | |
1045 | << " bounds " << bounds | |
1046 | << dendl; | |
1047 | ||
1048 | show_subtrees(); | |
1049 | ||
1050 | CDir *root; | |
b3b6e05e | 1051 | if (dir->ino() == CEPH_INO_ROOT) { |
7c673cae FG |
1052 | root = dir; // bootstrap hack. |
1053 | if (subtrees.count(root) == 0) { | |
1054 | subtrees[root]; | |
1055 | root->get(CDir::PIN_SUBTREE); | |
1056 | } | |
1057 | } else { | |
1058 | root = get_subtree_root(dir); // subtree root | |
1059 | } | |
11fdf7f2 TL |
1060 | ceph_assert(root); |
1061 | ceph_assert(subtrees.count(root)); | |
7c673cae FG |
1062 | dout(7) << " current root is " << *root << dendl; |
1063 | ||
1064 | mds_authority_t oldauth = dir->authority(); | |
1065 | ||
1066 | if (root == dir) { | |
1067 | // i am already a subtree. | |
1068 | dir->set_dir_auth(auth); | |
1069 | } else { | |
1070 | // i am a new subtree. | |
1071 | dout(10) << " new subtree at " << *dir << dendl; | |
11fdf7f2 | 1072 | ceph_assert(subtrees.count(dir) == 0); |
7c673cae FG |
1073 | subtrees[dir]; // create empty subtree bounds list for me. |
1074 | dir->get(CDir::PIN_SUBTREE); | |
1075 | ||
1076 | // set dir_auth | |
1077 | dir->set_dir_auth(auth); | |
1078 | ||
1079 | // move items nested beneath me, under me. | |
1080 | set<CDir*>::iterator p = subtrees[root].begin(); | |
1081 | while (p != subtrees[root].end()) { | |
1082 | set<CDir*>::iterator next = p; | |
1083 | ++next; | |
1084 | if (get_subtree_root((*p)->get_parent_dir()) == dir) { | |
1085 | // move under me | |
1086 | dout(10) << " claiming child bound " << **p << dendl; | |
1087 | subtrees[dir].insert(*p); | |
1088 | subtrees[root].erase(p); | |
1089 | } | |
1090 | p = next; | |
1091 | } | |
1092 | ||
1093 | // i am a bound of the parent subtree. | |
1094 | subtrees[root].insert(dir); | |
1095 | ||
1096 | // i am now the subtree root. | |
1097 | root = dir; | |
1098 | } | |
1099 | ||
224ce89b WB |
1100 | set<CInode*> to_eval; |
1101 | ||
7c673cae FG |
1102 | // verify/adjust bounds. |
1103 | // - these may be new, or | |
1104 | // - beneath existing ambiguous bounds (which will be collapsed), | |
1105 | // - but NOT beneath unambiguous bounds. | |
11fdf7f2 | 1106 | for (const auto& bound : bounds) { |
7c673cae FG |
1107 | // new bound? |
1108 | if (subtrees[dir].count(bound) == 0) { | |
1109 | if (get_subtree_root(bound) == dir) { | |
1110 | dout(10) << " new bound " << *bound << ", adjusting auth back to old " << oldauth << dendl; | |
1111 | adjust_subtree_auth(bound, oldauth); // otherwise, adjust at bound. | |
1112 | } | |
1113 | else { | |
1114 | dout(10) << " want bound " << *bound << dendl; | |
1115 | CDir *t = get_subtree_root(bound->get_parent_dir()); | |
1116 | if (subtrees[t].count(bound) == 0) { | |
11fdf7f2 | 1117 | ceph_assert(t != dir); |
7c673cae FG |
1118 | dout(10) << " new bound " << *bound << dendl; |
1119 | adjust_subtree_auth(bound, t->authority()); | |
1120 | } | |
1121 | // make sure it's nested beneath ambiguous subtree(s) | |
1122 | while (1) { | |
1123 | while (subtrees[dir].count(t) == 0) | |
1124 | t = get_subtree_root(t->get_parent_dir()); | |
1125 | dout(10) << " swallowing intervening subtree at " << *t << dendl; | |
1126 | adjust_subtree_auth(t, auth); | |
224ce89b | 1127 | try_subtree_merge_at(t, &to_eval); |
7c673cae FG |
1128 | t = get_subtree_root(bound->get_parent_dir()); |
1129 | if (t == dir) break; | |
1130 | } | |
1131 | } | |
1132 | } | |
1133 | else { | |
1134 | dout(10) << " already have bound " << *bound << dendl; | |
1135 | } | |
1136 | } | |
1137 | // merge stray bounds? | |
1138 | while (!subtrees[dir].empty()) { | |
1139 | set<CDir*> copy = subtrees[dir]; | |
1140 | for (set<CDir*>::iterator p = copy.begin(); p != copy.end(); ++p) { | |
1141 | if (bounds.count(*p) == 0) { | |
1142 | CDir *stray = *p; | |
1143 | dout(10) << " swallowing extra subtree at " << *stray << dendl; | |
1144 | adjust_subtree_auth(stray, auth); | |
224ce89b | 1145 | try_subtree_merge_at(stray, &to_eval); |
7c673cae FG |
1146 | } |
1147 | } | |
1148 | // swallowing subtree may add new subtree bounds | |
1149 | if (copy == subtrees[dir]) | |
1150 | break; | |
1151 | } | |
1152 | ||
1153 | // bound should now match. | |
1154 | verify_subtree_bounds(dir, bounds); | |
1155 | ||
1156 | show_subtrees(); | |
224ce89b WB |
1157 | |
1158 | if (!(mds->is_any_replay() || mds->is_resolve())) { | |
1159 | for(auto in : to_eval) | |
1160 | eval_subtree_root(in); | |
1161 | } | |
7c673cae FG |
1162 | } |
1163 | ||
1164 | ||
1165 | /* | |
1166 | * return a set of CDir*'s that correspond to the given bound set. Only adjust | |
1167 | * fragmentation as necessary to get an equivalent bounding set. That is, only | |
1168 | * split if one of our frags spans the provided bounding set. Never merge. | |
1169 | */ | |
11fdf7f2 | 1170 | void MDCache::get_force_dirfrag_bound_set(const vector<dirfrag_t>& dfs, set<CDir*>& bounds) |
7c673cae FG |
1171 | { |
1172 | dout(10) << "get_force_dirfrag_bound_set " << dfs << dendl; | |
1173 | ||
1174 | // sort by ino | |
1175 | map<inodeno_t, fragset_t> byino; | |
11fdf7f2 | 1176 | for (auto& frag : dfs) { |
9f95a23c | 1177 | byino[frag.ino].insert_raw(frag.frag); |
11fdf7f2 | 1178 | } |
7c673cae FG |
1179 | dout(10) << " by ino: " << byino << dendl; |
1180 | ||
1181 | for (map<inodeno_t,fragset_t>::iterator p = byino.begin(); p != byino.end(); ++p) { | |
9f95a23c | 1182 | p->second.simplify(); |
7c673cae FG |
1183 | CInode *diri = get_inode(p->first); |
1184 | if (!diri) | |
1185 | continue; | |
1186 | dout(10) << " checking fragset " << p->second.get() << " on " << *diri << dendl; | |
1187 | ||
1188 | fragtree_t tmpdft; | |
1189 | for (set<frag_t>::iterator q = p->second.begin(); q != p->second.end(); ++q) | |
1190 | tmpdft.force_to_leaf(g_ceph_context, *q); | |
1191 | ||
11fdf7f2 TL |
1192 | for (const auto& fg : p->second) { |
1193 | frag_vec_t leaves; | |
1194 | diri->dirfragtree.get_leaves_under(fg, leaves); | |
1195 | if (leaves.empty()) { | |
7c673cae | 1196 | frag_t approx_fg = diri->dirfragtree[fg.value()]; |
11fdf7f2 TL |
1197 | frag_vec_t approx_leaves; |
1198 | tmpdft.get_leaves_under(approx_fg, approx_leaves); | |
1199 | for (const auto& leaf : approx_leaves) { | |
1200 | if (p->second.get().count(leaf) == 0) { | |
7c673cae | 1201 | // not bound, so the resolve message is from auth MDS of the dirfrag |
11fdf7f2 | 1202 | force_dir_fragment(diri, leaf); |
7c673cae FG |
1203 | } |
1204 | } | |
7c673cae | 1205 | } |
f67539c2 TL |
1206 | |
1207 | auto&& [complete, sibs] = diri->get_dirfrags_under(fg); | |
1208 | for (const auto& sib : sibs) | |
1209 | bounds.insert(sib); | |
7c673cae FG |
1210 | } |
1211 | } | |
1212 | } | |
1213 | ||
11fdf7f2 | 1214 | void MDCache::adjust_bounded_subtree_auth(CDir *dir, const vector<dirfrag_t>& bound_dfs, const mds_authority_t &auth) |
7c673cae FG |
1215 | { |
1216 | dout(7) << "adjust_bounded_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
1217 | << " on " << *dir << " bound_dfs " << bound_dfs << dendl; | |
1218 | ||
1219 | set<CDir*> bounds; | |
1220 | get_force_dirfrag_bound_set(bound_dfs, bounds); | |
1221 | adjust_bounded_subtree_auth(dir, bounds, auth); | |
1222 | } | |
1223 | ||
11fdf7f2 | 1224 | void MDCache::map_dirfrag_set(const list<dirfrag_t>& dfs, set<CDir*>& result) |
7c673cae FG |
1225 | { |
1226 | dout(10) << "map_dirfrag_set " << dfs << dendl; | |
1227 | ||
1228 | // group by inode | |
1229 | map<inodeno_t, fragset_t> ino_fragset; | |
11fdf7f2 | 1230 | for (const auto &df : dfs) { |
9f95a23c | 1231 | ino_fragset[df.ino].insert_raw(df.frag); |
11fdf7f2 | 1232 | } |
7c673cae FG |
1233 | // get frags |
1234 | for (map<inodeno_t, fragset_t>::iterator p = ino_fragset.begin(); | |
1235 | p != ino_fragset.end(); | |
1236 | ++p) { | |
9f95a23c | 1237 | p->second.simplify(); |
7c673cae FG |
1238 | CInode *in = get_inode(p->first); |
1239 | if (!in) | |
1240 | continue; | |
1241 | ||
11fdf7f2 TL |
1242 | frag_vec_t fgs; |
1243 | for (const auto& fg : p->second) { | |
1244 | in->dirfragtree.get_leaves_under(fg, fgs); | |
1245 | } | |
7c673cae | 1246 | |
11fdf7f2 | 1247 | dout(15) << "map_dirfrag_set " << p->second << " -> " << fgs |
7c673cae FG |
1248 | << " on " << *in << dendl; |
1249 | ||
11fdf7f2 TL |
1250 | for (const auto& fg : fgs) { |
1251 | CDir *dir = in->get_dirfrag(fg); | |
7c673cae FG |
1252 | if (dir) |
1253 | result.insert(dir); | |
1254 | } | |
1255 | } | |
1256 | } | |
1257 | ||
1258 | ||
1259 | ||
1260 | CDir *MDCache::get_subtree_root(CDir *dir) | |
1261 | { | |
1262 | // find the underlying dir that delegates (or is about to delegate) auth | |
1263 | while (true) { | |
1264 | if (dir->is_subtree_root()) | |
1265 | return dir; | |
1266 | dir = dir->get_inode()->get_parent_dir(); | |
1267 | if (!dir) | |
1268 | return 0; // none | |
1269 | } | |
1270 | } | |
1271 | ||
1272 | CDir *MDCache::get_projected_subtree_root(CDir *dir) | |
1273 | { | |
1274 | // find the underlying dir that delegates (or is about to delegate) auth | |
1275 | while (true) { | |
1276 | if (dir->is_subtree_root()) | |
1277 | return dir; | |
1278 | dir = dir->get_inode()->get_projected_parent_dir(); | |
1279 | if (!dir) | |
1280 | return 0; // none | |
1281 | } | |
1282 | } | |
1283 | ||
1284 | void MDCache::remove_subtree(CDir *dir) | |
1285 | { | |
1286 | dout(10) << "remove_subtree " << *dir << dendl; | |
f6b5b4d7 TL |
1287 | auto it = subtrees.find(dir); |
1288 | ceph_assert(it != subtrees.end()); | |
1289 | subtrees.erase(it); | |
7c673cae FG |
1290 | dir->put(CDir::PIN_SUBTREE); |
1291 | if (dir->get_parent_dir()) { | |
1292 | CDir *p = get_subtree_root(dir->get_parent_dir()); | |
f6b5b4d7 TL |
1293 | auto it = subtrees.find(p); |
1294 | ceph_assert(it != subtrees.end()); | |
1295 | auto count = it->second.erase(dir); | |
1296 | ceph_assert(count == 1); | |
7c673cae FG |
1297 | } |
1298 | } | |
1299 | ||
1300 | void MDCache::get_subtree_bounds(CDir *dir, set<CDir*>& bounds) | |
1301 | { | |
11fdf7f2 | 1302 | ceph_assert(subtrees.count(dir)); |
7c673cae FG |
1303 | bounds = subtrees[dir]; |
1304 | } | |
1305 | ||
1306 | void MDCache::get_wouldbe_subtree_bounds(CDir *dir, set<CDir*>& bounds) | |
1307 | { | |
1308 | if (subtrees.count(dir)) { | |
1309 | // just copy them, dir is a subtree. | |
1310 | get_subtree_bounds(dir, bounds); | |
1311 | } else { | |
1312 | // find them | |
1313 | CDir *root = get_subtree_root(dir); | |
1314 | for (set<CDir*>::iterator p = subtrees[root].begin(); | |
1315 | p != subtrees[root].end(); | |
1316 | ++p) { | |
1317 | CDir *t = *p; | |
1318 | while (t != root) { | |
1319 | t = t->get_parent_dir(); | |
11fdf7f2 | 1320 | ceph_assert(t); |
7c673cae FG |
1321 | if (t == dir) { |
1322 | bounds.insert(*p); | |
1323 | continue; | |
1324 | } | |
1325 | } | |
1326 | } | |
1327 | } | |
1328 | } | |
1329 | ||
1330 | void MDCache::verify_subtree_bounds(CDir *dir, const set<CDir*>& bounds) | |
1331 | { | |
1332 | // for debugging only. | |
11fdf7f2 | 1333 | ceph_assert(subtrees.count(dir)); |
7c673cae FG |
1334 | if (bounds != subtrees[dir]) { |
1335 | dout(0) << "verify_subtree_bounds failed" << dendl; | |
1336 | set<CDir*> b = bounds; | |
1337 | for (auto &cd : subtrees[dir]) { | |
1338 | if (bounds.count(cd)) { | |
1339 | b.erase(cd); | |
1340 | continue; | |
1341 | } | |
1342 | dout(0) << " missing bound " << *cd << dendl; | |
1343 | } | |
1344 | for (const auto &cd : b) | |
1345 | dout(0) << " extra bound " << *cd << dendl; | |
1346 | } | |
11fdf7f2 | 1347 | ceph_assert(bounds == subtrees[dir]); |
7c673cae FG |
1348 | } |
1349 | ||
1350 | void MDCache::verify_subtree_bounds(CDir *dir, const list<dirfrag_t>& bounds) | |
1351 | { | |
1352 | // for debugging only. | |
11fdf7f2 | 1353 | ceph_assert(subtrees.count(dir)); |
7c673cae FG |
1354 | |
1355 | // make sure that any bounds i do have are properly noted as such. | |
1356 | int failed = 0; | |
1357 | for (const auto &fg : bounds) { | |
1358 | CDir *bd = get_dirfrag(fg); | |
1359 | if (!bd) continue; | |
1360 | if (subtrees[dir].count(bd) == 0) { | |
1361 | dout(0) << "verify_subtree_bounds failed: extra bound " << *bd << dendl; | |
1362 | failed++; | |
1363 | } | |
1364 | } | |
11fdf7f2 | 1365 | ceph_assert(failed == 0); |
7c673cae FG |
1366 | } |
1367 | ||
1368 | void MDCache::project_subtree_rename(CInode *diri, CDir *olddir, CDir *newdir) | |
1369 | { | |
1370 | dout(10) << "project_subtree_rename " << *diri << " from " << *olddir | |
1371 | << " to " << *newdir << dendl; | |
1372 | projected_subtree_renames[diri].push_back(pair<CDir*,CDir*>(olddir, newdir)); | |
1373 | } | |
1374 | ||
224ce89b | 1375 | void MDCache::adjust_subtree_after_rename(CInode *diri, CDir *olddir, bool pop) |
7c673cae FG |
1376 | { |
1377 | dout(10) << "adjust_subtree_after_rename " << *diri << " from " << *olddir << dendl; | |
1378 | ||
7c673cae FG |
1379 | CDir *newdir = diri->get_parent_dir(); |
1380 | ||
1381 | if (pop) { | |
1382 | map<CInode*,list<pair<CDir*,CDir*> > >::iterator p = projected_subtree_renames.find(diri); | |
11fdf7f2 TL |
1383 | ceph_assert(p != projected_subtree_renames.end()); |
1384 | ceph_assert(!p->second.empty()); | |
1385 | ceph_assert(p->second.front().first == olddir); | |
1386 | ceph_assert(p->second.front().second == newdir); | |
7c673cae FG |
1387 | p->second.pop_front(); |
1388 | if (p->second.empty()) | |
1389 | projected_subtree_renames.erase(p); | |
1390 | } | |
1391 | ||
11fdf7f2 TL |
1392 | // adjust total auth pin of freezing subtree |
1393 | if (olddir != newdir) { | |
9f95a23c TL |
1394 | auto&& dfls = diri->get_nested_dirfrags(); |
1395 | for (const auto& dir : dfls) | |
11fdf7f2 | 1396 | olddir->adjust_freeze_after_rename(dir); |
11fdf7f2 TL |
1397 | } |
1398 | ||
7c673cae | 1399 | // adjust subtree |
9f95a23c TL |
1400 | // N.B. make sure subtree dirfrags are at the front of the list |
1401 | auto dfls = diri->get_subtree_dirfrags(); | |
7c673cae | 1402 | diri->get_nested_dirfrags(dfls); |
9f95a23c | 1403 | for (const auto& dir : dfls) { |
7c673cae FG |
1404 | dout(10) << "dirfrag " << *dir << dendl; |
1405 | CDir *oldparent = get_subtree_root(olddir); | |
1406 | dout(10) << " old parent " << *oldparent << dendl; | |
1407 | CDir *newparent = get_subtree_root(newdir); | |
1408 | dout(10) << " new parent " << *newparent << dendl; | |
1409 | ||
9f95a23c TL |
1410 | auto& oldbounds = subtrees[oldparent]; |
1411 | auto& newbounds = subtrees[newparent]; | |
1412 | ||
28e407b8 | 1413 | if (olddir != newdir) |
11fdf7f2 | 1414 | mds->balancer->adjust_pop_for_rename(olddir, dir, false); |
28e407b8 | 1415 | |
7c673cae FG |
1416 | if (oldparent == newparent) { |
1417 | dout(10) << "parent unchanged for " << *dir << " at " << *oldparent << dendl; | |
28e407b8 | 1418 | } else if (dir->is_subtree_root()) { |
7c673cae FG |
1419 | // children are fine. change parent. |
1420 | dout(10) << "moving " << *dir << " from " << *oldparent << " to " << *newparent << dendl; | |
9f95a23c TL |
1421 | { |
1422 | auto n = oldbounds.erase(dir); | |
1423 | ceph_assert(n == 1); | |
1424 | } | |
1425 | newbounds.insert(dir); | |
224ce89b | 1426 | // caller is responsible for 'eval diri' |
28e407b8 | 1427 | try_subtree_merge_at(dir, NULL, false); |
7c673cae FG |
1428 | } else { |
1429 | // mid-subtree. | |
1430 | ||
1431 | // see if any old bounds move to the new parent. | |
9f95a23c TL |
1432 | std::vector<CDir*> tomove; |
1433 | for (const auto& bound : oldbounds) { | |
7c673cae FG |
1434 | CDir *broot = get_subtree_root(bound->get_parent_dir()); |
1435 | if (broot != oldparent) { | |
11fdf7f2 | 1436 | ceph_assert(broot == newparent); |
7c673cae FG |
1437 | tomove.push_back(bound); |
1438 | } | |
1439 | } | |
9f95a23c | 1440 | for (const auto& bound : tomove) { |
7c673cae | 1441 | dout(10) << "moving bound " << *bound << " from " << *oldparent << " to " << *newparent << dendl; |
9f95a23c TL |
1442 | oldbounds.erase(bound); |
1443 | newbounds.insert(bound); | |
7c673cae FG |
1444 | } |
1445 | ||
1446 | // did auth change? | |
1447 | if (oldparent->authority() != newparent->authority()) { | |
28e407b8 | 1448 | adjust_subtree_auth(dir, oldparent->authority(), false); |
224ce89b | 1449 | // caller is responsible for 'eval diri' |
28e407b8 | 1450 | try_subtree_merge_at(dir, NULL, false); |
7c673cae FG |
1451 | } |
1452 | } | |
28e407b8 AA |
1453 | |
1454 | if (olddir != newdir) | |
11fdf7f2 | 1455 | mds->balancer->adjust_pop_for_rename(newdir, dir, true); |
7c673cae FG |
1456 | } |
1457 | ||
1458 | show_subtrees(); | |
1459 | } | |
1460 | ||
7c673cae FG |
1461 | // =================================== |
1462 | // journal and snap/cow helpers | |
1463 | ||
1464 | ||
1465 | /* | |
1466 | * find first inode in cache that follows given snapid. otherwise, return current. | |
1467 | */ | |
1468 | CInode *MDCache::pick_inode_snap(CInode *in, snapid_t follows) | |
1469 | { | |
1470 | dout(10) << "pick_inode_snap follows " << follows << " on " << *in << dendl; | |
11fdf7f2 | 1471 | ceph_assert(in->last == CEPH_NOSNAP); |
7c673cae | 1472 | |
b32b8144 FG |
1473 | auto p = snap_inode_map.upper_bound(vinodeno_t(in->ino(), follows)); |
1474 | if (p != snap_inode_map.end() && p->second->ino() == in->ino()) { | |
1475 | dout(10) << "pick_inode_snap found " << *p->second << dendl; | |
1476 | in = p->second; | |
7c673cae | 1477 | } |
b32b8144 | 1478 | |
7c673cae FG |
1479 | return in; |
1480 | } | |
1481 | ||
1482 | ||
1483 | /* | |
1484 | * note: i'm currently cheating wrt dirty and inode.version on cow | |
1485 | * items. instead of doing a full dir predirty, i just take the | |
1486 | * original item's version, and set the dirty flag (via | |
1487 | * mutation::add_cow_{inode,dentry}() and mutation::apply(). that | |
1488 | * means a special case in the dir commit clean sweep assertions. | |
1489 | * bah. | |
1490 | */ | |
1491 | CInode *MDCache::cow_inode(CInode *in, snapid_t last) | |
1492 | { | |
11fdf7f2 | 1493 | ceph_assert(last >= in->first); |
7c673cae | 1494 | |
b32b8144 | 1495 | CInode *oldin = new CInode(this, true, in->first, last); |
f67539c2 TL |
1496 | auto _inode = CInode::allocate_inode(*in->get_previous_projected_inode()); |
1497 | _inode->trim_client_ranges(last); | |
1498 | oldin->reset_inode(std::move(_inode)); | |
1499 | auto _xattrs = in->get_previous_projected_xattrs(); | |
1500 | oldin->reset_xattrs(std::move(_xattrs)); | |
1501 | ||
11fdf7f2 | 1502 | oldin->symlink = in->symlink; |
7c673cae FG |
1503 | |
1504 | if (in->first < in->oldest_snap) | |
1505 | in->oldest_snap = in->first; | |
1506 | ||
1507 | in->first = last+1; | |
1508 | ||
1509 | dout(10) << "cow_inode " << *in << " to " << *oldin << dendl; | |
1510 | add_inode(oldin); | |
1511 | ||
1512 | if (in->last != CEPH_NOSNAP) { | |
1513 | CInode *head_in = get_inode(in->ino()); | |
11fdf7f2 | 1514 | ceph_assert(head_in); |
494da23a TL |
1515 | auto ret = head_in->split_need_snapflush(oldin, in); |
1516 | if (ret.first) { | |
7c673cae | 1517 | oldin->client_snap_caps = in->client_snap_caps; |
eafe8130 TL |
1518 | if (!oldin->client_snap_caps.empty()) { |
1519 | for (int i = 0; i < num_cinode_locks; i++) { | |
1520 | SimpleLock *lock = oldin->get_lock(cinode_lock_info[i].lock); | |
1521 | ceph_assert(lock); | |
494da23a TL |
1522 | if (lock->get_state() != LOCK_SNAP_SYNC) { |
1523 | ceph_assert(lock->is_stable()); | |
1524 | lock->set_state(LOCK_SNAP_SYNC); // gathering | |
1525 | oldin->auth_pin(lock); | |
1526 | } | |
7c673cae FG |
1527 | lock->get_wrlock(true); |
1528 | } | |
1529 | } | |
1530 | } | |
494da23a TL |
1531 | if (!ret.second) { |
1532 | auto client_snap_caps = std::move(in->client_snap_caps); | |
1533 | in->client_snap_caps.clear(); | |
1534 | in->item_open_file.remove_myself(); | |
1535 | in->item_caps.remove_myself(); | |
eafe8130 TL |
1536 | |
1537 | if (!client_snap_caps.empty()) { | |
1538 | MDSContext::vec finished; | |
1539 | for (int i = 0; i < num_cinode_locks; i++) { | |
1540 | SimpleLock *lock = in->get_lock(cinode_lock_info[i].lock); | |
1541 | ceph_assert(lock); | |
1542 | ceph_assert(lock->get_state() == LOCK_SNAP_SYNC); // gathering | |
494da23a | 1543 | lock->put_wrlock(); |
eafe8130 TL |
1544 | if (!lock->get_num_wrlocks()) { |
1545 | lock->set_state(LOCK_SYNC); | |
1546 | lock->take_waiting(SimpleLock::WAIT_STABLE|SimpleLock::WAIT_RD, finished); | |
1547 | in->auth_unpin(lock); | |
1548 | } | |
494da23a | 1549 | } |
eafe8130 | 1550 | mds->queue_waiters(finished); |
494da23a TL |
1551 | } |
1552 | } | |
7c673cae FG |
1553 | return oldin; |
1554 | } | |
1555 | ||
b32b8144 FG |
1556 | if (!in->client_caps.empty()) { |
1557 | const set<snapid_t>& snaps = in->find_snaprealm()->get_snaps(); | |
1558 | // clone caps? | |
94b18763 | 1559 | for (auto &p : in->client_caps) { |
b32b8144 | 1560 | client_t client = p.first; |
11fdf7f2 TL |
1561 | Capability *cap = &p.second; |
1562 | int issued = cap->need_snapflush() ? CEPH_CAP_ANY_WR : cap->issued(); | |
b32b8144 FG |
1563 | if ((issued & CEPH_CAP_ANY_WR) && |
1564 | cap->client_follows < last) { | |
eafe8130 TL |
1565 | dout(10) << " client." << client << " cap " << ccap_string(issued) << dendl; |
1566 | oldin->client_snap_caps.insert(client); | |
b32b8144 FG |
1567 | cap->client_follows = last; |
1568 | ||
1569 | // we need snapflushes for any intervening snaps | |
1570 | dout(10) << " snaps " << snaps << dendl; | |
1571 | for (auto q = snaps.lower_bound(oldin->first); | |
1572 | q != snaps.end() && *q <= last; | |
1573 | ++q) { | |
1574 | in->add_need_snapflush(oldin, *q, client); | |
1575 | } | |
1576 | } else { | |
1577 | dout(10) << " ignoring client." << client << " cap follows " << cap->client_follows << dendl; | |
7c673cae | 1578 | } |
7c673cae | 1579 | } |
eafe8130 TL |
1580 | |
1581 | if (!oldin->client_snap_caps.empty()) { | |
1582 | for (int i = 0; i < num_cinode_locks; i++) { | |
1583 | SimpleLock *lock = oldin->get_lock(cinode_lock_info[i].lock); | |
1584 | ceph_assert(lock); | |
1585 | if (lock->get_state() != LOCK_SNAP_SYNC) { | |
1586 | ceph_assert(lock->is_stable()); | |
1587 | lock->set_state(LOCK_SNAP_SYNC); // gathering | |
1588 | oldin->auth_pin(lock); | |
1589 | } | |
1590 | lock->get_wrlock(true); | |
1591 | } | |
1592 | } | |
7c673cae | 1593 | } |
7c673cae FG |
1594 | return oldin; |
1595 | } | |
1596 | ||
1597 | void MDCache::journal_cow_dentry(MutationImpl *mut, EMetaBlob *metablob, | |
1598 | CDentry *dn, snapid_t follows, | |
1599 | CInode **pcow_inode, CDentry::linkage_t *dnl) | |
1600 | { | |
1601 | if (!dn) { | |
1602 | dout(10) << "journal_cow_dentry got null CDentry, returning" << dendl; | |
1603 | return; | |
1604 | } | |
1605 | dout(10) << "journal_cow_dentry follows " << follows << " on " << *dn << dendl; | |
11fdf7f2 | 1606 | ceph_assert(dn->is_auth()); |
7c673cae FG |
1607 | |
1608 | // nothing to cow on a null dentry, fix caller | |
1609 | if (!dnl) | |
1610 | dnl = dn->get_projected_linkage(); | |
11fdf7f2 | 1611 | ceph_assert(!dnl->is_null()); |
7c673cae | 1612 | |
11fdf7f2 TL |
1613 | CInode *in = dnl->is_primary() ? dnl->get_inode() : NULL; |
1614 | bool cow_head = false; | |
1615 | if (in && in->state_test(CInode::STATE_AMBIGUOUSAUTH)) { | |
1616 | ceph_assert(in->is_frozen_inode()); | |
1617 | cow_head = true; | |
1618 | } | |
1619 | if (in && (in->is_multiversion() || cow_head)) { | |
7c673cae | 1620 | // multiversion inode. |
7c673cae FG |
1621 | SnapRealm *realm = NULL; |
1622 | ||
1623 | if (in->get_projected_parent_dn() != dn) { | |
11fdf7f2 | 1624 | ceph_assert(follows == CEPH_NOSNAP); |
7c673cae | 1625 | realm = dn->dir->inode->find_snaprealm(); |
11fdf7f2 TL |
1626 | snapid_t dir_follows = get_global_snaprealm()->get_newest_seq(); |
1627 | ceph_assert(dir_follows >= realm->get_newest_seq()); | |
7c673cae FG |
1628 | |
1629 | if (dir_follows+1 > dn->first) { | |
1630 | snapid_t oldfirst = dn->first; | |
1631 | dn->first = dir_follows+1; | |
1632 | if (realm->has_snaps_in_range(oldfirst, dir_follows)) { | |
f67539c2 TL |
1633 | CDir *dir = dn->dir; |
1634 | CDentry *olddn = dir->add_remote_dentry(dn->get_name(), in->ino(), in->d_type(), dn->alternate_name, oldfirst, dir_follows); | |
7c673cae | 1635 | dout(10) << " olddn " << *olddn << dendl; |
f67539c2 TL |
1636 | ceph_assert(dir->is_projected()); |
1637 | olddn->set_projected_version(dir->get_projected_version()); | |
7c673cae FG |
1638 | metablob->add_remote_dentry(olddn, true); |
1639 | mut->add_cow_dentry(olddn); | |
1640 | // FIXME: adjust link count here? hmm. | |
1641 | ||
1642 | if (dir_follows+1 > in->first) | |
11fdf7f2 | 1643 | in->cow_old_inode(dir_follows, cow_head); |
7c673cae FG |
1644 | } |
1645 | } | |
1646 | ||
11fdf7f2 | 1647 | follows = dir_follows; |
7c673cae FG |
1648 | if (in->snaprealm) { |
1649 | realm = in->snaprealm; | |
11fdf7f2 TL |
1650 | ceph_assert(follows >= realm->get_newest_seq()); |
1651 | } | |
7c673cae FG |
1652 | } else { |
1653 | realm = in->find_snaprealm(); | |
11fdf7f2 TL |
1654 | if (follows == CEPH_NOSNAP) { |
1655 | follows = get_global_snaprealm()->get_newest_seq(); | |
1656 | ceph_assert(follows >= realm->get_newest_seq()); | |
1657 | } | |
7c673cae FG |
1658 | } |
1659 | ||
1660 | // already cloned? | |
1661 | if (follows < in->first) { | |
1662 | dout(10) << "journal_cow_dentry follows " << follows << " < first on " << *in << dendl; | |
1663 | return; | |
1664 | } | |
1665 | ||
1666 | if (!realm->has_snaps_in_range(in->first, follows)) { | |
1667 | dout(10) << "journal_cow_dentry no snapshot follows " << follows << " on " << *in << dendl; | |
1668 | in->first = follows + 1; | |
1669 | return; | |
1670 | } | |
1671 | ||
11fdf7f2 | 1672 | in->cow_old_inode(follows, cow_head); |
7c673cae FG |
1673 | |
1674 | } else { | |
1675 | SnapRealm *realm = dn->dir->inode->find_snaprealm(); | |
11fdf7f2 TL |
1676 | if (follows == CEPH_NOSNAP) { |
1677 | follows = get_global_snaprealm()->get_newest_seq(); | |
1678 | ceph_assert(follows >= realm->get_newest_seq()); | |
1679 | } | |
7c673cae FG |
1680 | |
1681 | // already cloned? | |
1682 | if (follows < dn->first) { | |
1683 | dout(10) << "journal_cow_dentry follows " << follows << " < first on " << *dn << dendl; | |
1684 | return; | |
1685 | } | |
1686 | ||
1687 | // update dn.first before adding old dentry to cdir's map | |
1688 | snapid_t oldfirst = dn->first; | |
1689 | dn->first = follows+1; | |
1690 | ||
7c673cae FG |
1691 | if (!realm->has_snaps_in_range(oldfirst, follows)) { |
1692 | dout(10) << "journal_cow_dentry no snapshot follows " << follows << " on " << *dn << dendl; | |
1693 | if (in) | |
1694 | in->first = follows+1; | |
1695 | return; | |
1696 | } | |
1697 | ||
1698 | dout(10) << " dn " << *dn << dendl; | |
f67539c2 TL |
1699 | CDir *dir = dn->get_dir(); |
1700 | ceph_assert(dir->is_projected()); | |
1701 | ||
7c673cae FG |
1702 | if (in) { |
1703 | CInode *oldin = cow_inode(in, follows); | |
f67539c2 | 1704 | ceph_assert(in->is_projected()); |
7c673cae FG |
1705 | mut->add_cow_inode(oldin); |
1706 | if (pcow_inode) | |
1707 | *pcow_inode = oldin; | |
f67539c2 | 1708 | CDentry *olddn = dir->add_primary_dentry(dn->get_name(), oldin, dn->alternate_name, oldfirst, follows); |
7c673cae FG |
1709 | dout(10) << " olddn " << *olddn << dendl; |
1710 | bool need_snapflush = !oldin->client_snap_caps.empty(); | |
11fdf7f2 | 1711 | if (need_snapflush) { |
7c673cae | 1712 | mut->ls->open_files.push_back(&oldin->item_open_file); |
11fdf7f2 TL |
1713 | mds->locker->mark_need_snapflush_inode(oldin); |
1714 | } | |
f67539c2 | 1715 | olddn->set_projected_version(dir->get_projected_version()); |
7c673cae FG |
1716 | metablob->add_primary_dentry(olddn, 0, true, false, false, need_snapflush); |
1717 | mut->add_cow_dentry(olddn); | |
1718 | } else { | |
11fdf7f2 | 1719 | ceph_assert(dnl->is_remote()); |
f67539c2 | 1720 | CDentry *olddn = dir->add_remote_dentry(dn->get_name(), dnl->get_remote_ino(), dnl->get_remote_d_type(), dn->alternate_name, oldfirst, follows); |
7c673cae | 1721 | dout(10) << " olddn " << *olddn << dendl; |
f67539c2 TL |
1722 | |
1723 | olddn->set_projected_version(dir->get_projected_version()); | |
7c673cae FG |
1724 | metablob->add_remote_dentry(olddn, true); |
1725 | mut->add_cow_dentry(olddn); | |
1726 | } | |
1727 | } | |
1728 | } | |
1729 | ||
7c673cae FG |
1730 | void MDCache::journal_dirty_inode(MutationImpl *mut, EMetaBlob *metablob, CInode *in, snapid_t follows) |
1731 | { | |
1732 | if (in->is_base()) { | |
11fdf7f2 | 1733 | metablob->add_root(true, in); |
7c673cae FG |
1734 | } else { |
1735 | if (follows == CEPH_NOSNAP && in->last != CEPH_NOSNAP) | |
1736 | follows = in->first - 1; | |
1737 | CDentry *dn = in->get_projected_parent_dn(); | |
1738 | if (!dn->get_projected_linkage()->is_null()) // no need to cow a null dentry | |
1739 | journal_cow_dentry(mut, metablob, dn, follows); | |
1740 | if (in->get_projected_inode()->is_backtrace_updated()) { | |
1741 | bool dirty_pool = in->get_projected_inode()->layout.pool_id != | |
1742 | in->get_previous_projected_inode()->layout.pool_id; | |
1743 | metablob->add_primary_dentry(dn, in, true, true, dirty_pool); | |
1744 | } else { | |
1745 | metablob->add_primary_dentry(dn, in, true); | |
1746 | } | |
1747 | } | |
1748 | } | |
1749 | ||
1750 | ||
1751 | ||
1752 | // nested --------------------------------------------------------------- | |
1753 | ||
f67539c2 TL |
1754 | void MDCache::project_rstat_inode_to_frag(const MutationRef& mut, |
1755 | CInode *cur, CDir *parent, snapid_t first, | |
7c673cae FG |
1756 | int linkunlink, SnapRealm *prealm) |
1757 | { | |
1758 | CDentry *parentdn = cur->get_projected_parent_dn(); | |
7c673cae FG |
1759 | |
1760 | if (cur->first > first) | |
1761 | first = cur->first; | |
1762 | ||
1763 | dout(10) << "projected_rstat_inode_to_frag first " << first << " linkunlink " << linkunlink | |
1764 | << " " << *cur << dendl; | |
1765 | dout(20) << " frag head is [" << parent->first << ",head] " << dendl; | |
1766 | dout(20) << " inode update is [" << first << "," << cur->last << "]" << dendl; | |
1767 | ||
1768 | /* | |
1769 | * FIXME. this incompletely propagates rstats to _old_ parents | |
1770 | * (i.e. shortly after a directory rename). but we need full | |
1771 | * blown hard link backpointers to make this work properly... | |
1772 | */ | |
1773 | snapid_t floor = parentdn->first; | |
1774 | dout(20) << " floor of " << floor << " from parent dn " << *parentdn << dendl; | |
1775 | ||
1776 | if (!prealm) | |
1777 | prealm = parent->inode->find_snaprealm(); | |
1778 | const set<snapid_t> snaps = prealm->get_snaps(); | |
1779 | ||
1780 | if (cur->last != CEPH_NOSNAP) { | |
11fdf7f2 TL |
1781 | ceph_assert(cur->dirty_old_rstats.empty()); |
1782 | set<snapid_t>::const_iterator q = snaps.lower_bound(std::max(first, floor)); | |
7c673cae FG |
1783 | if (q == snaps.end() || *q > cur->last) |
1784 | return; | |
1785 | } | |
1786 | ||
1787 | if (cur->last >= floor) { | |
1788 | bool update = true; | |
1789 | if (cur->state_test(CInode::STATE_AMBIGUOUSAUTH) && cur->is_auth()) { | |
f67539c2 | 1790 | // rename src inode is not projected in the peer rename prep case. so we should |
7c673cae | 1791 | // avoid updateing the inode. |
11fdf7f2 TL |
1792 | ceph_assert(linkunlink < 0); |
1793 | ceph_assert(cur->is_frozen_inode()); | |
7c673cae FG |
1794 | update = false; |
1795 | } | |
f67539c2 TL |
1796 | // hacky |
1797 | const CInode::mempool_inode *pi; | |
1798 | if (update && mut->is_projected(cur)) { | |
1799 | pi = cur->_get_projected_inode(); | |
1800 | } else { | |
1801 | pi = cur->get_projected_inode().get(); | |
1802 | if (update) { | |
1803 | // new inode | |
1804 | ceph_assert(pi->rstat == pi->accounted_rstat); | |
1805 | update = false; | |
1806 | } | |
1807 | } | |
1808 | _project_rstat_inode_to_frag(pi, std::max(first, floor), cur->last, parent, | |
7c673cae FG |
1809 | linkunlink, update); |
1810 | } | |
1811 | ||
11fdf7f2 | 1812 | if (g_conf()->mds_snap_rstat) { |
94b18763 | 1813 | for (const auto &p : cur->dirty_old_rstats) { |
f67539c2 | 1814 | const auto &old = cur->get_old_inodes()->at(p); |
94b18763 FG |
1815 | snapid_t ofirst = std::max(old.first, floor); |
1816 | auto it = snaps.lower_bound(ofirst); | |
1817 | if (it == snaps.end() || *it > p) | |
7c673cae | 1818 | continue; |
94b18763 | 1819 | if (p >= floor) |
f67539c2 | 1820 | _project_rstat_inode_to_frag(&old.inode, ofirst, p, parent, 0, false); |
7c673cae FG |
1821 | } |
1822 | } | |
1823 | cur->dirty_old_rstats.clear(); | |
1824 | } | |
1825 | ||
1826 | ||
f67539c2 | 1827 | void MDCache::_project_rstat_inode_to_frag(const CInode::mempool_inode* inode, snapid_t ofirst, snapid_t last, |
7c673cae FG |
1828 | CDir *parent, int linkunlink, bool update_inode) |
1829 | { | |
1830 | dout(10) << "_project_rstat_inode_to_frag [" << ofirst << "," << last << "]" << dendl; | |
f67539c2 TL |
1831 | dout(20) << " inode rstat " << inode->rstat << dendl; |
1832 | dout(20) << " inode accounted_rstat " << inode->accounted_rstat << dendl; | |
7c673cae FG |
1833 | nest_info_t delta; |
1834 | if (linkunlink == 0) { | |
f67539c2 TL |
1835 | delta.add(inode->rstat); |
1836 | delta.sub(inode->accounted_rstat); | |
7c673cae | 1837 | } else if (linkunlink < 0) { |
f67539c2 | 1838 | delta.sub(inode->accounted_rstat); |
7c673cae | 1839 | } else { |
f67539c2 | 1840 | delta.add(inode->rstat); |
7c673cae FG |
1841 | } |
1842 | dout(20) << " delta " << delta << dendl; | |
1843 | ||
7c673cae FG |
1844 | |
1845 | while (last >= ofirst) { | |
1846 | /* | |
1847 | * pick fnode version to update. at each iteration, we want to | |
1848 | * pick a segment ending in 'last' to update. split as necessary | |
1849 | * to make that work. then, adjust first up so that we only | |
1850 | * update one segment at a time. then loop to cover the whole | |
1851 | * [ofirst,last] interval. | |
1852 | */ | |
1853 | nest_info_t *prstat; | |
1854 | snapid_t first; | |
f67539c2 | 1855 | auto pf = parent->_get_projected_fnode(); |
7c673cae | 1856 | if (last == CEPH_NOSNAP) { |
11fdf7f2 TL |
1857 | if (g_conf()->mds_snap_rstat) |
1858 | first = std::max(ofirst, parent->first); | |
7c673cae FG |
1859 | else |
1860 | first = parent->first; | |
1861 | prstat = &pf->rstat; | |
1862 | dout(20) << " projecting to head [" << first << "," << last << "] " << *prstat << dendl; | |
1863 | ||
1864 | if (first > parent->first && | |
1865 | !(pf->rstat == pf->accounted_rstat)) { | |
1866 | dout(10) << " target snapped and not fully accounted, cow to dirty_old_rstat [" | |
1867 | << parent->first << "," << (first-1) << "] " | |
1868 | << " " << *prstat << "/" << pf->accounted_rstat | |
1869 | << dendl; | |
1870 | parent->dirty_old_rstat[first-1].first = parent->first; | |
1871 | parent->dirty_old_rstat[first-1].rstat = pf->rstat; | |
1872 | parent->dirty_old_rstat[first-1].accounted_rstat = pf->accounted_rstat; | |
1873 | } | |
1874 | parent->first = first; | |
11fdf7f2 | 1875 | } else if (!g_conf()->mds_snap_rstat) { |
7c673cae FG |
1876 | // drop snapshots' rstats |
1877 | break; | |
1878 | } else if (last >= parent->first) { | |
1879 | first = parent->first; | |
1880 | parent->dirty_old_rstat[last].first = first; | |
1881 | parent->dirty_old_rstat[last].rstat = pf->rstat; | |
1882 | parent->dirty_old_rstat[last].accounted_rstat = pf->accounted_rstat; | |
1883 | prstat = &parent->dirty_old_rstat[last].rstat; | |
1884 | dout(10) << " projecting to newly split dirty_old_fnode [" << first << "," << last << "] " | |
1885 | << " " << *prstat << "/" << pf->accounted_rstat << dendl; | |
1886 | } else { | |
1887 | // be careful, dirty_old_rstat is a _sparse_ map. | |
1888 | // sorry, this is ugly. | |
1889 | first = ofirst; | |
1890 | ||
1891 | // find any intersection with last | |
94b18763 FG |
1892 | auto it = parent->dirty_old_rstat.lower_bound(last); |
1893 | if (it == parent->dirty_old_rstat.end()) { | |
7c673cae FG |
1894 | dout(20) << " no dirty_old_rstat with last >= last " << last << dendl; |
1895 | if (!parent->dirty_old_rstat.empty() && parent->dirty_old_rstat.rbegin()->first >= first) { | |
1896 | dout(20) << " last dirty_old_rstat ends at " << parent->dirty_old_rstat.rbegin()->first << dendl; | |
1897 | first = parent->dirty_old_rstat.rbegin()->first+1; | |
1898 | } | |
1899 | } else { | |
94b18763 FG |
1900 | // *it last is >= last |
1901 | if (it->second.first <= last) { | |
1902 | // *it intersects [first,last] | |
1903 | if (it->second.first < first) { | |
1904 | dout(10) << " splitting off left bit [" << it->second.first << "," << first-1 << "]" << dendl; | |
1905 | parent->dirty_old_rstat[first-1] = it->second; | |
1906 | it->second.first = first; | |
7c673cae | 1907 | } |
94b18763 FG |
1908 | if (it->second.first > first) |
1909 | first = it->second.first; | |
1910 | if (last < it->first) { | |
1911 | dout(10) << " splitting off right bit [" << last+1 << "," << it->first << "]" << dendl; | |
1912 | parent->dirty_old_rstat[last] = it->second; | |
1913 | it->second.first = last+1; | |
7c673cae FG |
1914 | } |
1915 | } else { | |
94b18763 FG |
1916 | // *it is to the _right_ of [first,last] |
1917 | it = parent->dirty_old_rstat.lower_bound(first); | |
1918 | // new *it last is >= first | |
1919 | if (it->second.first <= last && // new *it isn't also to the right, and | |
1920 | it->first >= first) { // it intersects our first bit, | |
1921 | dout(10) << " staying to the right of [" << it->second.first << "," << it->first << "]..." << dendl; | |
1922 | first = it->first+1; | |
7c673cae FG |
1923 | } |
1924 | dout(10) << " projecting to new dirty_old_rstat [" << first << "," << last << "]" << dendl; | |
1925 | } | |
1926 | } | |
1927 | dout(20) << " projecting to dirty_old_rstat [" << first << "," << last << "]" << dendl; | |
1928 | parent->dirty_old_rstat[last].first = first; | |
1929 | prstat = &parent->dirty_old_rstat[last].rstat; | |
1930 | } | |
1931 | ||
1932 | // apply | |
1933 | dout(20) << " project to [" << first << "," << last << "] " << *prstat << dendl; | |
11fdf7f2 | 1934 | ceph_assert(last >= first); |
7c673cae | 1935 | prstat->add(delta); |
7c673cae FG |
1936 | dout(20) << " result [" << first << "," << last << "] " << *prstat << " " << *parent << dendl; |
1937 | ||
1938 | last = first-1; | |
1939 | } | |
f67539c2 TL |
1940 | |
1941 | if (update_inode) { | |
1942 | auto _inode = const_cast<CInode::mempool_inode*>(inode); | |
1943 | _inode->accounted_rstat = _inode->rstat; | |
1944 | } | |
7c673cae FG |
1945 | } |
1946 | ||
f67539c2 TL |
1947 | void MDCache::project_rstat_frag_to_inode(const nest_info_t& rstat, |
1948 | const nest_info_t& accounted_rstat, | |
7c673cae FG |
1949 | snapid_t ofirst, snapid_t last, |
1950 | CInode *pin, bool cow_head) | |
1951 | { | |
1952 | dout(10) << "project_rstat_frag_to_inode [" << ofirst << "," << last << "]" << dendl; | |
1953 | dout(20) << " frag rstat " << rstat << dendl; | |
1954 | dout(20) << " frag accounted_rstat " << accounted_rstat << dendl; | |
1955 | nest_info_t delta = rstat; | |
1956 | delta.sub(accounted_rstat); | |
1957 | dout(20) << " delta " << delta << dendl; | |
1958 | ||
f67539c2 | 1959 | CInode::old_inode_map_ptr _old_inodes; |
7c673cae | 1960 | while (last >= ofirst) { |
94b18763 | 1961 | CInode::mempool_inode *pi; |
7c673cae FG |
1962 | snapid_t first; |
1963 | if (last == pin->last) { | |
f67539c2 | 1964 | pi = pin->_get_projected_inode(); |
11fdf7f2 | 1965 | first = std::max(ofirst, pin->first); |
7c673cae | 1966 | if (first > pin->first) { |
f67539c2 | 1967 | auto& old = pin->cow_old_inode(first-1, cow_head); |
7c673cae FG |
1968 | dout(20) << " cloned old_inode rstat is " << old.inode.rstat << dendl; |
1969 | } | |
1970 | } else { | |
f67539c2 TL |
1971 | if (!_old_inodes) { |
1972 | _old_inodes = CInode::allocate_old_inode_map(); | |
1973 | if (pin->is_any_old_inodes()) | |
1974 | *_old_inodes = *pin->get_old_inodes(); | |
1975 | } | |
7c673cae FG |
1976 | if (last >= pin->first) { |
1977 | first = pin->first; | |
1978 | pin->cow_old_inode(last, cow_head); | |
1979 | } else { | |
1980 | // our life is easier here because old_inodes is not sparse | |
1981 | // (although it may not begin at snapid 1) | |
f67539c2 TL |
1982 | auto it = _old_inodes->lower_bound(last); |
1983 | if (it == _old_inodes->end()) { | |
7c673cae FG |
1984 | dout(10) << " no old_inode <= " << last << ", done." << dendl; |
1985 | break; | |
1986 | } | |
94b18763 | 1987 | first = it->second.first; |
7c673cae | 1988 | if (first > last) { |
94b18763 | 1989 | dout(10) << " oldest old_inode is [" << first << "," << it->first << "], done." << dendl; |
7c673cae FG |
1990 | //assert(p == pin->old_inodes.begin()); |
1991 | break; | |
1992 | } | |
94b18763 FG |
1993 | if (it->first > last) { |
1994 | dout(10) << " splitting right old_inode [" << first << "," << it->first << "] to [" | |
1995 | << (last+1) << "," << it->first << "]" << dendl; | |
f67539c2 | 1996 | (*_old_inodes)[last] = it->second; |
94b18763 FG |
1997 | it->second.first = last+1; |
1998 | pin->dirty_old_rstats.insert(it->first); | |
7c673cae FG |
1999 | } |
2000 | } | |
2001 | if (first < ofirst) { | |
2002 | dout(10) << " splitting left old_inode [" << first << "," << last << "] to [" | |
2003 | << first << "," << ofirst-1 << "]" << dendl; | |
f67539c2 | 2004 | (*_old_inodes)[ofirst-1] = (*_old_inodes)[last]; |
7c673cae | 2005 | pin->dirty_old_rstats.insert(ofirst-1); |
f67539c2 | 2006 | (*_old_inodes)[last].first = first = ofirst; |
7c673cae | 2007 | } |
f67539c2 | 2008 | pi = &(*_old_inodes)[last].inode; |
7c673cae FG |
2009 | pin->dirty_old_rstats.insert(last); |
2010 | } | |
2011 | dout(20) << " projecting to [" << first << "," << last << "] " << pi->rstat << dendl; | |
2012 | pi->rstat.add(delta); | |
2013 | dout(20) << " result [" << first << "," << last << "] " << pi->rstat << dendl; | |
2014 | ||
2015 | last = first-1; | |
2016 | } | |
f67539c2 TL |
2017 | if (_old_inodes) |
2018 | pin->reset_old_inodes(std::move(_old_inodes)); | |
7c673cae FG |
2019 | } |
2020 | ||
a8e16298 | 2021 | void MDCache::broadcast_quota_to_client(CInode *in, client_t exclude_ct, bool quota_change) |
7c673cae | 2022 | { |
11fdf7f2 TL |
2023 | if (!(mds->is_active() || mds->is_stopping())) |
2024 | return; | |
2025 | ||
7c673cae FG |
2026 | if (!in->is_auth() || in->is_frozen()) |
2027 | return; | |
2028 | ||
f67539c2 TL |
2029 | const auto& pi = in->get_projected_inode(); |
2030 | if (!pi->quota.is_enable() && !quota_change) | |
7c673cae FG |
2031 | return; |
2032 | ||
11fdf7f2 TL |
2033 | // creaete snaprealm for quota inode (quota was set before mimic) |
2034 | if (!in->get_projected_srnode()) | |
2035 | mds->server->create_quota_realm(in); | |
7c673cae | 2036 | |
11fdf7f2 TL |
2037 | for (auto &p : in->client_caps) { |
2038 | Capability *cap = &p.second; | |
2039 | if (cap->is_noquota()) | |
2040 | continue; | |
28e407b8 | 2041 | |
11fdf7f2 | 2042 | if (exclude_ct >= 0 && exclude_ct != p.first) |
28e407b8 AA |
2043 | goto update; |
2044 | ||
f67539c2 TL |
2045 | if (cap->last_rbytes == pi->rstat.rbytes && |
2046 | cap->last_rsize == pi->rstat.rsize()) | |
7c673cae FG |
2047 | continue; |
2048 | ||
f67539c2 TL |
2049 | if (pi->quota.max_files > 0) { |
2050 | if (pi->rstat.rsize() >= pi->quota.max_files) | |
7c673cae FG |
2051 | goto update; |
2052 | ||
f67539c2 TL |
2053 | if ((abs(cap->last_rsize - pi->quota.max_files) >> 4) < |
2054 | abs(cap->last_rsize - pi->rstat.rsize())) | |
7c673cae FG |
2055 | goto update; |
2056 | } | |
2057 | ||
f67539c2 TL |
2058 | if (pi->quota.max_bytes > 0) { |
2059 | if (pi->rstat.rbytes > pi->quota.max_bytes - (pi->quota.max_bytes >> 3)) | |
7c673cae FG |
2060 | goto update; |
2061 | ||
f67539c2 TL |
2062 | if ((abs(cap->last_rbytes - pi->quota.max_bytes) >> 4) < |
2063 | abs(cap->last_rbytes - pi->rstat.rbytes)) | |
7c673cae FG |
2064 | goto update; |
2065 | } | |
2066 | ||
2067 | continue; | |
2068 | ||
2069 | update: | |
f67539c2 TL |
2070 | cap->last_rsize = pi->rstat.rsize(); |
2071 | cap->last_rbytes = pi->rstat.rbytes; | |
7c673cae | 2072 | |
9f95a23c | 2073 | auto msg = make_message<MClientQuota>(); |
7c673cae | 2074 | msg->ino = in->ino(); |
f67539c2 TL |
2075 | msg->rstat = pi->rstat; |
2076 | msg->quota = pi->quota; | |
11fdf7f2 | 2077 | mds->send_message_client_counted(msg, cap->get_session()); |
7c673cae | 2078 | } |
181888fb | 2079 | for (const auto &it : in->get_replicas()) { |
9f95a23c | 2080 | auto msg = make_message<MGatherCaps>(); |
7c673cae | 2081 | msg->ino = in->ino(); |
181888fb | 2082 | mds->send_message_mds(msg, it.first); |
7c673cae FG |
2083 | } |
2084 | } | |
2085 | ||
2086 | /* | |
2087 | * NOTE: we _have_ to delay the scatter if we are called during a | |
2088 | * rejoin, because we can't twiddle locks between when the | |
2089 | * rejoin_(weak|strong) is received and when we send the rejoin_ack. | |
2090 | * normally, this isn't a problem: a recover mds doesn't twiddle locks | |
2091 | * (no requests), and a survivor acks immediately. _except_ that | |
2092 | * during rejoin_(weak|strong) processing, we may complete a lock | |
2093 | * gather, and do a scatter_writebehind.. and we _can't_ twiddle the | |
2094 | * scatterlock state in that case or the lock states will get out of | |
2095 | * sync between the auth and replica. | |
2096 | * | |
2097 | * the simple solution is to never do the scatter here. instead, put | |
2098 | * the scatterlock on a list if it isn't already wrlockable. this is | |
2099 | * probably the best plan anyway, since we avoid too many | |
2100 | * scatters/locks under normal usage. | |
2101 | */ | |
2102 | /* | |
2103 | * some notes on dirlock/nestlock scatterlock semantics: | |
2104 | * | |
2105 | * the fragstat (dirlock) will never be updated without | |
2106 | * dirlock+nestlock wrlock held by the caller. | |
2107 | * | |
2108 | * the rstat (nestlock) _may_ get updated without a wrlock when nested | |
2109 | * data is pushed up the tree. this could be changed with some | |
2110 | * restructuring here, but in its current form we ensure that the | |
2111 | * fragstat+rstat _always_ reflect an accurrate summation over the dir | |
2112 | * frag, which is nice. and, we only need to track frags that need to | |
2113 | * be nudged (and not inodes with pending rstat changes that need to | |
2114 | * be pushed into the frag). a consequence of this is that the | |
2115 | * accounted_rstat on scatterlock sync may not match our current | |
2116 | * rstat. this is normal and expected. | |
2117 | */ | |
2118 | void MDCache::predirty_journal_parents(MutationRef mut, EMetaBlob *blob, | |
2119 | CInode *in, CDir *parent, | |
2120 | int flags, int linkunlink, | |
2121 | snapid_t cfollows) | |
2122 | { | |
2123 | bool primary_dn = flags & PREDIRTY_PRIMARY; | |
2124 | bool do_parent_mtime = flags & PREDIRTY_DIR; | |
2125 | bool shallow = flags & PREDIRTY_SHALLOW; | |
2126 | ||
11fdf7f2 | 2127 | ceph_assert(mds->mdlog->entry_is_open()); |
7c673cae FG |
2128 | |
2129 | // make sure stamp is set | |
2130 | if (mut->get_mds_stamp() == utime_t()) | |
2131 | mut->set_mds_stamp(ceph_clock_now()); | |
2132 | ||
2133 | if (in->is_base()) | |
2134 | return; | |
2135 | ||
2136 | dout(10) << "predirty_journal_parents" | |
2137 | << (do_parent_mtime ? " do_parent_mtime":"") | |
2138 | << " linkunlink=" << linkunlink | |
2139 | << (primary_dn ? " primary_dn":" remote_dn") | |
2140 | << (shallow ? " SHALLOW":"") | |
2141 | << " follows " << cfollows | |
2142 | << " " << *in << dendl; | |
2143 | ||
2144 | if (!parent) { | |
11fdf7f2 | 2145 | ceph_assert(primary_dn); |
7c673cae FG |
2146 | parent = in->get_projected_parent_dn()->get_dir(); |
2147 | } | |
2148 | ||
2149 | if (flags == 0 && linkunlink == 0) { | |
2150 | dout(10) << " no flags/linkunlink, just adding dir context to blob(s)" << dendl; | |
2151 | blob->add_dir_context(parent); | |
2152 | return; | |
2153 | } | |
2154 | ||
2155 | // build list of inodes to wrlock, dirty, and update | |
2156 | list<CInode*> lsi; | |
2157 | CInode *cur = in; | |
2158 | CDentry *parentdn = NULL; | |
2159 | bool first = true; | |
2160 | while (parent) { | |
2161 | //assert(cur->is_auth() || !primary_dn); // this breaks the rename auth twiddle hack | |
11fdf7f2 | 2162 | ceph_assert(parent->is_auth()); |
7c673cae FG |
2163 | |
2164 | // opportunistically adjust parent dirfrag | |
2165 | CInode *pin = parent->get_inode(); | |
2166 | ||
2167 | // inode -> dirfrag | |
2168 | mut->auth_pin(parent); | |
7c673cae | 2169 | |
f67539c2 | 2170 | auto pf = parent->project_fnode(mut); |
7c673cae FG |
2171 | pf->version = parent->pre_dirty(); |
2172 | ||
2173 | if (do_parent_mtime || linkunlink) { | |
11fdf7f2 TL |
2174 | ceph_assert(mut->is_wrlocked(&pin->filelock)); |
2175 | ceph_assert(mut->is_wrlocked(&pin->nestlock)); | |
2176 | ceph_assert(cfollows == CEPH_NOSNAP); | |
7c673cae FG |
2177 | |
2178 | // update stale fragstat/rstat? | |
2179 | parent->resync_accounted_fragstat(); | |
2180 | parent->resync_accounted_rstat(); | |
2181 | ||
2182 | if (do_parent_mtime) { | |
2183 | pf->fragstat.mtime = mut->get_op_stamp(); | |
2184 | pf->fragstat.change_attr++; | |
2185 | dout(10) << "predirty_journal_parents bumping change_attr to " << pf->fragstat.change_attr << " on " << parent << dendl; | |
2186 | if (pf->fragstat.mtime > pf->rstat.rctime) { | |
2187 | dout(10) << "predirty_journal_parents updating mtime on " << *parent << dendl; | |
2188 | pf->rstat.rctime = pf->fragstat.mtime; | |
2189 | } else { | |
2190 | dout(10) << "predirty_journal_parents updating mtime UNDERWATER on " << *parent << dendl; | |
2191 | } | |
2192 | } | |
2193 | if (linkunlink) { | |
2194 | dout(10) << "predirty_journal_parents updating size on " << *parent << dendl; | |
2195 | if (in->is_dir()) { | |
2196 | pf->fragstat.nsubdirs += linkunlink; | |
2197 | //pf->rstat.rsubdirs += linkunlink; | |
2198 | } else { | |
2199 | pf->fragstat.nfiles += linkunlink; | |
2200 | //pf->rstat.rfiles += linkunlink; | |
2201 | } | |
2202 | } | |
2203 | } | |
2204 | ||
2205 | // rstat | |
2206 | if (!primary_dn) { | |
2207 | // don't update parent this pass | |
2208 | } else if (!linkunlink && !(pin->nestlock.can_wrlock(-1) && | |
2209 | pin->versionlock.can_wrlock())) { | |
2210 | dout(20) << " unwritable parent nestlock " << pin->nestlock | |
2211 | << ", marking dirty rstat on " << *cur << dendl; | |
2212 | cur->mark_dirty_rstat(); | |
2213 | } else { | |
2214 | // if we don't hold a wrlock reference on this nestlock, take one, | |
2215 | // because we are about to write into the dirfrag fnode and that needs | |
2216 | // to commit before the lock can cycle. | |
2217 | if (linkunlink) { | |
f67539c2 | 2218 | ceph_assert(pin->nestlock.get_num_wrlocks() || mut->is_peer()); |
7c673cae FG |
2219 | } |
2220 | ||
11fdf7f2 | 2221 | if (!mut->is_wrlocked(&pin->nestlock)) { |
7c673cae FG |
2222 | dout(10) << " taking wrlock on " << pin->nestlock << " on " << *pin << dendl; |
2223 | mds->locker->wrlock_force(&pin->nestlock, mut); | |
2224 | } | |
2225 | ||
2226 | // now we can project the inode rstat diff the dirfrag | |
2227 | SnapRealm *prealm = pin->find_snaprealm(); | |
2228 | ||
2229 | snapid_t follows = cfollows; | |
2230 | if (follows == CEPH_NOSNAP) | |
2231 | follows = prealm->get_newest_seq(); | |
2232 | ||
2233 | snapid_t first = follows+1; | |
2234 | ||
2235 | // first, if the frag is stale, bring it back in sync. | |
2236 | parent->resync_accounted_rstat(); | |
2237 | ||
2238 | // now push inode rstats into frag | |
f67539c2 | 2239 | project_rstat_inode_to_frag(mut, cur, parent, first, linkunlink, prealm); |
7c673cae FG |
2240 | cur->clear_dirty_rstat(); |
2241 | } | |
2242 | ||
2243 | bool stop = false; | |
2244 | if (!pin->is_auth() || (!mut->is_auth_pinned(pin) && !pin->can_auth_pin())) { | |
2245 | dout(10) << "predirty_journal_parents !auth or ambig or can't authpin on " << *pin << dendl; | |
2246 | stop = true; | |
2247 | } | |
2248 | ||
2249 | // delay propagating until later? | |
2250 | if (!stop && !first && | |
11fdf7f2 | 2251 | g_conf()->mds_dirstat_min_interval > 0) { |
7c673cae | 2252 | double since_last_prop = mut->get_mds_stamp() - pin->last_dirstat_prop; |
11fdf7f2 | 2253 | if (since_last_prop < g_conf()->mds_dirstat_min_interval) { |
7c673cae | 2254 | dout(10) << "predirty_journal_parents last prop " << since_last_prop |
11fdf7f2 | 2255 | << " < " << g_conf()->mds_dirstat_min_interval |
7c673cae FG |
2256 | << ", stopping" << dendl; |
2257 | stop = true; | |
2258 | } else { | |
2259 | dout(10) << "predirty_journal_parents last prop " << since_last_prop << " ago, continuing" << dendl; | |
2260 | } | |
2261 | } | |
2262 | ||
2263 | // can cast only because i'm passing nowait=true in the sole user | |
7c673cae | 2264 | if (!stop && |
11fdf7f2 | 2265 | !mut->is_wrlocked(&pin->nestlock) && |
7c673cae | 2266 | (!pin->versionlock.can_wrlock() || // make sure we can take versionlock, too |
9f95a23c | 2267 | !mds->locker->wrlock_try(&pin->nestlock, mut) |
7c673cae FG |
2268 | )) { // ** do not initiate.. see above comment ** |
2269 | dout(10) << "predirty_journal_parents can't wrlock one of " << pin->versionlock << " or " << pin->nestlock | |
2270 | << " on " << *pin << dendl; | |
2271 | stop = true; | |
2272 | } | |
2273 | if (stop) { | |
2274 | dout(10) << "predirty_journal_parents stop. marking nestlock on " << *pin << dendl; | |
2275 | mds->locker->mark_updated_scatterlock(&pin->nestlock); | |
2276 | mut->ls->dirty_dirfrag_nest.push_back(&pin->item_dirty_dirfrag_nest); | |
2277 | mut->add_updated_lock(&pin->nestlock); | |
2278 | if (do_parent_mtime || linkunlink) { | |
2279 | mds->locker->mark_updated_scatterlock(&pin->filelock); | |
2280 | mut->ls->dirty_dirfrag_dir.push_back(&pin->item_dirty_dirfrag_dir); | |
2281 | mut->add_updated_lock(&pin->filelock); | |
2282 | } | |
2283 | break; | |
2284 | } | |
11fdf7f2 | 2285 | if (!mut->is_wrlocked(&pin->versionlock)) |
7c673cae FG |
2286 | mds->locker->local_wrlock_grab(&pin->versionlock, mut); |
2287 | ||
f67539c2 | 2288 | ceph_assert(mut->is_wrlocked(&pin->nestlock) || mut->is_peer()); |
7c673cae FG |
2289 | |
2290 | pin->last_dirstat_prop = mut->get_mds_stamp(); | |
2291 | ||
2292 | // dirfrag -> diri | |
2293 | mut->auth_pin(pin); | |
7c673cae FG |
2294 | lsi.push_front(pin); |
2295 | ||
2296 | pin->pre_cow_old_inode(); // avoid cow mayhem! | |
2297 | ||
f67539c2 TL |
2298 | auto pi = pin->project_inode(mut); |
2299 | pi.inode->version = pin->pre_dirty(); | |
7c673cae FG |
2300 | |
2301 | // dirstat | |
2302 | if (do_parent_mtime || linkunlink) { | |
2303 | dout(20) << "predirty_journal_parents add_delta " << pf->fragstat << dendl; | |
2304 | dout(20) << "predirty_journal_parents - " << pf->accounted_fragstat << dendl; | |
2305 | bool touched_mtime = false, touched_chattr = false; | |
f67539c2 | 2306 | pi.inode->dirstat.add_delta(pf->fragstat, pf->accounted_fragstat, &touched_mtime, &touched_chattr); |
7c673cae FG |
2307 | pf->accounted_fragstat = pf->fragstat; |
2308 | if (touched_mtime) | |
f67539c2 | 2309 | pi.inode->mtime = pi.inode->ctime = pi.inode->dirstat.mtime; |
7c673cae | 2310 | if (touched_chattr) |
f67539c2 TL |
2311 | pi.inode->change_attr = pi.inode->dirstat.change_attr; |
2312 | dout(20) << "predirty_journal_parents gives " << pi.inode->dirstat << " on " << *pin << dendl; | |
7c673cae FG |
2313 | |
2314 | if (parent->get_frag() == frag_t()) { // i.e., we are the only frag | |
f67539c2 | 2315 | if (pi.inode->dirstat.size() < 0) |
11fdf7f2 | 2316 | ceph_assert(!"negative dirstat size" == g_conf()->mds_verify_scatter); |
f67539c2 | 2317 | if (pi.inode->dirstat.size() != pf->fragstat.size()) { |
7c673cae | 2318 | mds->clog->error() << "unmatched fragstat size on single dirfrag " |
f67539c2 | 2319 | << parent->dirfrag() << ", inode has " << pi.inode->dirstat |
7c673cae FG |
2320 | << ", dirfrag has " << pf->fragstat; |
2321 | ||
2322 | // trust the dirfrag for now | |
f67539c2 | 2323 | pi.inode->dirstat = pf->fragstat; |
7c673cae | 2324 | |
11fdf7f2 | 2325 | ceph_assert(!"unmatched fragstat size" == g_conf()->mds_verify_scatter); |
7c673cae FG |
2326 | } |
2327 | } | |
2328 | } | |
2329 | ||
7c673cae FG |
2330 | // rstat |
2331 | dout(10) << "predirty_journal_parents frag->inode on " << *parent << dendl; | |
2332 | ||
2333 | // first, if the frag is stale, bring it back in sync. | |
2334 | parent->resync_accounted_rstat(); | |
2335 | ||
11fdf7f2 | 2336 | if (g_conf()->mds_snap_rstat) { |
94b18763 FG |
2337 | for (auto &p : parent->dirty_old_rstat) { |
2338 | project_rstat_frag_to_inode(p.second.rstat, p.second.accounted_rstat, p.second.first, | |
2339 | p.first, pin, true); | |
2340 | } | |
7c673cae FG |
2341 | } |
2342 | parent->dirty_old_rstat.clear(); | |
2343 | project_rstat_frag_to_inode(pf->rstat, pf->accounted_rstat, parent->first, CEPH_NOSNAP, pin, true);//false); | |
2344 | ||
2345 | pf->accounted_rstat = pf->rstat; | |
2346 | ||
2347 | if (parent->get_frag() == frag_t()) { // i.e., we are the only frag | |
f67539c2 | 2348 | if (pi.inode->rstat.rbytes != pf->rstat.rbytes) { |
7c673cae | 2349 | mds->clog->error() << "unmatched rstat rbytes on single dirfrag " |
f67539c2 | 2350 | << parent->dirfrag() << ", inode has " << pi.inode->rstat |
7c673cae FG |
2351 | << ", dirfrag has " << pf->rstat; |
2352 | ||
2353 | // trust the dirfrag for now | |
f67539c2 | 2354 | pi.inode->rstat = pf->rstat; |
7c673cae | 2355 | |
11fdf7f2 | 2356 | ceph_assert(!"unmatched rstat rbytes" == g_conf()->mds_verify_scatter); |
7c673cae FG |
2357 | } |
2358 | } | |
2359 | ||
2360 | parent->check_rstats(); | |
2361 | broadcast_quota_to_client(pin); | |
f67539c2 TL |
2362 | if (pin->is_base()) |
2363 | break; | |
7c673cae FG |
2364 | // next parent! |
2365 | cur = pin; | |
f67539c2 TL |
2366 | parentdn = pin->get_projected_parent_dn(); |
2367 | ceph_assert(parentdn); | |
7c673cae FG |
2368 | parent = parentdn->get_dir(); |
2369 | linkunlink = 0; | |
2370 | do_parent_mtime = false; | |
2371 | primary_dn = true; | |
2372 | first = false; | |
2373 | } | |
2374 | ||
2375 | // now, stick it in the blob | |
11fdf7f2 TL |
2376 | ceph_assert(parent); |
2377 | ceph_assert(parent->is_auth()); | |
7c673cae FG |
2378 | blob->add_dir_context(parent); |
2379 | blob->add_dir(parent, true); | |
9f95a23c TL |
2380 | for (const auto& in : lsi) { |
2381 | journal_dirty_inode(mut.get(), blob, in); | |
7c673cae FG |
2382 | } |
2383 | ||
2384 | } | |
2385 | ||
2386 | ||
2387 | ||
2388 | ||
2389 | ||
2390 | // =================================== | |
f67539c2 | 2391 | // peer requests |
7c673cae FG |
2392 | |
2393 | ||
2394 | /* | |
f67539c2 TL |
2395 | * some handlers for leader requests with peers. we need to make |
2396 | * sure leader journal commits before we forget we leadered them and | |
2397 | * remove them from the uncommitted_leaders map (used during recovery | |
2398 | * to commit|abort peers). | |
7c673cae | 2399 | */ |
f67539c2 | 2400 | struct C_MDC_CommittedLeader : public MDCacheLogContext { |
7c673cae | 2401 | metareqid_t reqid; |
f67539c2 | 2402 | C_MDC_CommittedLeader(MDCache *s, metareqid_t r) : MDCacheLogContext(s), reqid(r) {} |
7c673cae | 2403 | void finish(int r) override { |
f67539c2 | 2404 | mdcache->_logged_leader_commit(reqid); |
7c673cae FG |
2405 | } |
2406 | }; | |
2407 | ||
f67539c2 | 2408 | void MDCache::log_leader_commit(metareqid_t reqid) |
7c673cae | 2409 | { |
f67539c2 TL |
2410 | dout(10) << "log_leader_commit " << reqid << dendl; |
2411 | uncommitted_leaders[reqid].committing = true; | |
7c673cae | 2412 | mds->mdlog->start_submit_entry(new ECommitted(reqid), |
f67539c2 | 2413 | new C_MDC_CommittedLeader(this, reqid)); |
7c673cae FG |
2414 | } |
2415 | ||
f67539c2 | 2416 | void MDCache::_logged_leader_commit(metareqid_t reqid) |
7c673cae | 2417 | { |
f67539c2 TL |
2418 | dout(10) << "_logged_leader_commit " << reqid << dendl; |
2419 | ceph_assert(uncommitted_leaders.count(reqid)); | |
2420 | uncommitted_leaders[reqid].ls->uncommitted_leaders.erase(reqid); | |
2421 | mds->queue_waiters(uncommitted_leaders[reqid].waiters); | |
2422 | uncommitted_leaders.erase(reqid); | |
7c673cae FG |
2423 | } |
2424 | ||
2425 | // while active... | |
2426 | ||
f67539c2 | 2427 | void MDCache::committed_leader_peer(metareqid_t r, mds_rank_t from) |
7c673cae | 2428 | { |
f67539c2 TL |
2429 | dout(10) << "committed_leader_peer mds." << from << " on " << r << dendl; |
2430 | ceph_assert(uncommitted_leaders.count(r)); | |
2431 | uncommitted_leaders[r].peers.erase(from); | |
2432 | if (!uncommitted_leaders[r].recovering && uncommitted_leaders[r].peers.empty()) | |
2433 | log_leader_commit(r); | |
7c673cae FG |
2434 | } |
2435 | ||
f67539c2 | 2436 | void MDCache::logged_leader_update(metareqid_t reqid) |
7c673cae | 2437 | { |
f67539c2 TL |
2438 | dout(10) << "logged_leader_update " << reqid << dendl; |
2439 | ceph_assert(uncommitted_leaders.count(reqid)); | |
2440 | uncommitted_leaders[reqid].safe = true; | |
2441 | auto p = pending_leaders.find(reqid); | |
2442 | if (p != pending_leaders.end()) { | |
2443 | pending_leaders.erase(p); | |
2444 | if (pending_leaders.empty()) | |
7c673cae FG |
2445 | process_delayed_resolve(); |
2446 | } | |
2447 | } | |
2448 | ||
2449 | /* | |
f67539c2 TL |
2450 | * Leader may crash after receiving all peers' commit acks, but before journalling |
2451 | * the final commit. Peers may crash after journalling the peer commit, but before | |
2452 | * sending commit ack to the leader. Commit leaders with no uncommitted peer when | |
7c673cae FG |
2453 | * resolve finishes. |
2454 | */ | |
f67539c2 | 2455 | void MDCache::finish_committed_leaders() |
7c673cae | 2456 | { |
f67539c2 TL |
2457 | for (map<metareqid_t, uleader>::iterator p = uncommitted_leaders.begin(); |
2458 | p != uncommitted_leaders.end(); | |
7c673cae FG |
2459 | ++p) { |
2460 | p->second.recovering = false; | |
f67539c2 TL |
2461 | if (!p->second.committing && p->second.peers.empty()) { |
2462 | dout(10) << "finish_committed_leaders " << p->first << dendl; | |
2463 | log_leader_commit(p->first); | |
7c673cae FG |
2464 | } |
2465 | } | |
2466 | } | |
2467 | ||
2468 | /* | |
f67539c2 | 2469 | * at end of resolve... we must journal a commit|abort for all peer |
7c673cae FG |
2470 | * updates, before moving on. |
2471 | * | |
f67539c2 TL |
2472 | * this is so that the leader can safely journal ECommitted on ops it |
2473 | * leaders when it reaches up:active (all other recovering nodes must | |
7c673cae FG |
2474 | * complete resolve before that happens). |
2475 | */ | |
f67539c2 | 2476 | struct C_MDC_PeerCommit : public MDCacheLogContext { |
7c673cae FG |
2477 | mds_rank_t from; |
2478 | metareqid_t reqid; | |
f67539c2 | 2479 | C_MDC_PeerCommit(MDCache *c, int f, metareqid_t r) : MDCacheLogContext(c), from(f), reqid(r) {} |
7c673cae | 2480 | void finish(int r) override { |
f67539c2 | 2481 | mdcache->_logged_peer_commit(from, reqid); |
7c673cae FG |
2482 | } |
2483 | }; | |
2484 | ||
f67539c2 | 2485 | void MDCache::_logged_peer_commit(mds_rank_t from, metareqid_t reqid) |
7c673cae | 2486 | { |
f67539c2 | 2487 | dout(10) << "_logged_peer_commit from mds." << from << " " << reqid << dendl; |
7c673cae FG |
2488 | |
2489 | // send a message | |
f67539c2 | 2490 | auto req = make_message<MMDSPeerRequest>(reqid, 0, MMDSPeerRequest::OP_COMMITTED); |
7c673cae FG |
2491 | mds->send_message_mds(req, from); |
2492 | } | |
2493 | ||
2494 | ||
2495 | ||
2496 | ||
2497 | ||
2498 | ||
2499 | // ==================================================================== | |
2500 | // import map, recovery | |
2501 | ||
2502 | void MDCache::_move_subtree_map_bound(dirfrag_t df, dirfrag_t oldparent, dirfrag_t newparent, | |
2503 | map<dirfrag_t,vector<dirfrag_t> >& subtrees) | |
2504 | { | |
2505 | if (subtrees.count(oldparent)) { | |
2506 | vector<dirfrag_t>& v = subtrees[oldparent]; | |
2507 | dout(10) << " removing " << df << " from " << oldparent << " bounds " << v << dendl; | |
2508 | for (vector<dirfrag_t>::iterator it = v.begin(); it != v.end(); ++it) | |
2509 | if (*it == df) { | |
2510 | v.erase(it); | |
2511 | break; | |
2512 | } | |
2513 | } | |
2514 | if (subtrees.count(newparent)) { | |
2515 | vector<dirfrag_t>& v = subtrees[newparent]; | |
2516 | dout(10) << " adding " << df << " to " << newparent << " bounds " << v << dendl; | |
2517 | v.push_back(df); | |
2518 | } | |
2519 | } | |
2520 | ||
2521 | ESubtreeMap *MDCache::create_subtree_map() | |
2522 | { | |
2523 | dout(10) << "create_subtree_map " << num_subtrees() << " subtrees, " | |
2524 | << num_subtrees_fullauth() << " fullauth" | |
2525 | << dendl; | |
2526 | ||
2527 | show_subtrees(); | |
2528 | ||
2529 | ESubtreeMap *le = new ESubtreeMap(); | |
2530 | mds->mdlog->_start_entry(le); | |
2531 | ||
2532 | map<dirfrag_t, CDir*> dirs_to_add; | |
2533 | ||
2534 | if (myin) { | |
2535 | CDir* mydir = myin->get_dirfrag(frag_t()); | |
2536 | dirs_to_add[mydir->dirfrag()] = mydir; | |
2537 | } | |
2538 | ||
2539 | // include all auth subtrees, and their bounds. | |
2540 | // and a spanning tree to tie it to the root. | |
f6b5b4d7 | 2541 | for (auto& [dir, bounds] : subtrees) { |
7c673cae FG |
2542 | // journal subtree as "ours" if we are |
2543 | // me, -2 | |
2544 | // me, me | |
2545 | // me, !me (may be importing and ambiguous!) | |
2546 | ||
2547 | // so not | |
2548 | // !me, * | |
2549 | if (dir->get_dir_auth().first != mds->get_nodeid()) | |
2550 | continue; | |
2551 | ||
2552 | if (migrator->is_ambiguous_import(dir->dirfrag()) || | |
2553 | my_ambiguous_imports.count(dir->dirfrag())) { | |
2554 | dout(15) << " ambig subtree " << *dir << dendl; | |
2555 | le->ambiguous_subtrees.insert(dir->dirfrag()); | |
2556 | } else { | |
f6b5b4d7 | 2557 | dout(15) << " auth subtree " << *dir << dendl; |
7c673cae FG |
2558 | } |
2559 | ||
2560 | dirs_to_add[dir->dirfrag()] = dir; | |
2561 | le->subtrees[dir->dirfrag()].clear(); | |
2562 | ||
7c673cae | 2563 | // bounds |
f6b5b4d7 TL |
2564 | size_t nbounds = bounds.size(); |
2565 | if (nbounds > 3) { | |
2566 | dout(15) << " subtree has " << nbounds << " bounds" << dendl; | |
2567 | } | |
2568 | for (auto& bound : bounds) { | |
2569 | if (nbounds <= 3) { | |
2570 | dout(15) << " subtree bound " << *bound << dendl; | |
2571 | } | |
7c673cae FG |
2572 | dirs_to_add[bound->dirfrag()] = bound; |
2573 | le->subtrees[dir->dirfrag()].push_back(bound->dirfrag()); | |
2574 | } | |
2575 | } | |
2576 | ||
2577 | // apply projected renames | |
9f95a23c TL |
2578 | for (const auto& [diri, renames] : projected_subtree_renames) { |
2579 | for (const auto& [olddir, newdir] : renames) { | |
f6b5b4d7 | 2580 | dout(15) << " adjusting for projected rename of " << *diri << " to " << *newdir << dendl; |
7c673cae | 2581 | |
9f95a23c TL |
2582 | auto&& dfls = diri->get_dirfrags(); |
2583 | for (const auto& dir : dfls) { | |
f6b5b4d7 | 2584 | dout(15) << "dirfrag " << dir->dirfrag() << " " << *dir << dendl; |
7c673cae | 2585 | CDir *oldparent = get_projected_subtree_root(olddir); |
f6b5b4d7 | 2586 | dout(15) << " old parent " << oldparent->dirfrag() << " " << *oldparent << dendl; |
7c673cae | 2587 | CDir *newparent = get_projected_subtree_root(newdir); |
f6b5b4d7 | 2588 | dout(15) << " new parent " << newparent->dirfrag() << " " << *newparent << dendl; |
7c673cae FG |
2589 | |
2590 | if (oldparent == newparent) { | |
f6b5b4d7 | 2591 | dout(15) << "parent unchanged for " << dir->dirfrag() << " at " |
7c673cae FG |
2592 | << oldparent->dirfrag() << dendl; |
2593 | continue; | |
2594 | } | |
2595 | ||
2596 | if (dir->is_subtree_root()) { | |
2597 | if (le->subtrees.count(newparent->dirfrag()) && | |
2598 | oldparent->get_dir_auth() != newparent->get_dir_auth()) | |
2599 | dirs_to_add[dir->dirfrag()] = dir; | |
2600 | // children are fine. change parent. | |
2601 | _move_subtree_map_bound(dir->dirfrag(), oldparent->dirfrag(), newparent->dirfrag(), | |
2602 | le->subtrees); | |
2603 | } else { | |
2604 | // mid-subtree. | |
2605 | ||
2606 | if (oldparent->get_dir_auth() != newparent->get_dir_auth()) { | |
2607 | dout(10) << " creating subtree for " << dir->dirfrag() << dendl; | |
2608 | // if oldparent is auth, subtree is mine; include it. | |
2609 | if (le->subtrees.count(oldparent->dirfrag())) { | |
2610 | dirs_to_add[dir->dirfrag()] = dir; | |
2611 | le->subtrees[dir->dirfrag()].clear(); | |
2612 | } | |
2613 | // if newparent is auth, subtree is a new bound | |
2614 | if (le->subtrees.count(newparent->dirfrag())) { | |
2615 | dirs_to_add[dir->dirfrag()] = dir; | |
2616 | le->subtrees[newparent->dirfrag()].push_back(dir->dirfrag()); // newparent is auth; new bound | |
2617 | } | |
2618 | newparent = dir; | |
2619 | } | |
2620 | ||
2621 | // see if any old bounds move to the new parent. | |
f6b5b4d7 | 2622 | for (auto& bound : subtrees.at(oldparent)) { |
7c673cae FG |
2623 | if (dir->contains(bound->get_parent_dir())) |
2624 | _move_subtree_map_bound(bound->dirfrag(), oldparent->dirfrag(), newparent->dirfrag(), | |
2625 | le->subtrees); | |
2626 | } | |
2627 | } | |
2628 | } | |
2629 | } | |
2630 | } | |
2631 | ||
2632 | // simplify the journaled map. our in memory map may have more | |
2633 | // subtrees than needed due to migrations that are just getting | |
2634 | // started or just completing. but on replay, the "live" map will | |
2635 | // be simple and we can do a straight comparison. | |
f6b5b4d7 TL |
2636 | for (auto& [frag, bfrags] : le->subtrees) { |
2637 | if (le->ambiguous_subtrees.count(frag)) | |
7c673cae FG |
2638 | continue; |
2639 | unsigned i = 0; | |
f6b5b4d7 TL |
2640 | while (i < bfrags.size()) { |
2641 | dirfrag_t b = bfrags[i]; | |
7c673cae FG |
2642 | if (le->subtrees.count(b) && |
2643 | le->ambiguous_subtrees.count(b) == 0) { | |
f6b5b4d7 TL |
2644 | auto& bb = le->subtrees.at(b); |
2645 | dout(10) << "simplify: " << frag << " swallowing " << b << " with bounds " << bb << dendl; | |
2646 | for (auto& r : bb) { | |
2647 | bfrags.push_back(r); | |
2648 | } | |
7c673cae FG |
2649 | dirs_to_add.erase(b); |
2650 | le->subtrees.erase(b); | |
f6b5b4d7 | 2651 | bfrags.erase(bfrags.begin() + i); |
7c673cae FG |
2652 | } else { |
2653 | ++i; | |
2654 | } | |
2655 | } | |
2656 | } | |
2657 | ||
94b18763 | 2658 | for (auto &p : dirs_to_add) { |
7c673cae FG |
2659 | CDir *dir = p.second; |
2660 | le->metablob.add_dir_context(dir, EMetaBlob::TO_ROOT); | |
2661 | le->metablob.add_dir(dir, false); | |
2662 | } | |
2663 | ||
2664 | dout(15) << " subtrees " << le->subtrees << dendl; | |
2665 | dout(15) << " ambiguous_subtrees " << le->ambiguous_subtrees << dendl; | |
2666 | ||
2667 | //le->metablob.print(cout); | |
2668 | le->expire_pos = mds->mdlog->journaler->get_expire_pos(); | |
2669 | return le; | |
2670 | } | |
2671 | ||
2672 | void MDCache::dump_resolve_status(Formatter *f) const | |
2673 | { | |
2674 | f->open_object_section("resolve_status"); | |
2675 | f->dump_stream("resolve_gather") << resolve_gather; | |
2676 | f->dump_stream("resolve_ack_gather") << resolve_gather; | |
2677 | f->close_section(); | |
2678 | } | |
2679 | ||
11fdf7f2 | 2680 | void MDCache::resolve_start(MDSContext *resolve_done_) |
7c673cae FG |
2681 | { |
2682 | dout(10) << "resolve_start" << dendl; | |
11fdf7f2 | 2683 | ceph_assert(!resolve_done); |
7c673cae FG |
2684 | resolve_done.reset(resolve_done_); |
2685 | ||
2686 | if (mds->mdsmap->get_root() != mds->get_nodeid()) { | |
2687 | // if we don't have the root dir, adjust it to UNKNOWN. during | |
2688 | // resolve we want mds0 to explicit claim the portion of it that | |
2689 | // it owns, so that anything beyond its bounds get left as | |
2690 | // unknown. | |
2691 | CDir *rootdir = root->get_dirfrag(frag_t()); | |
2692 | if (rootdir) | |
2693 | adjust_subtree_auth(rootdir, CDIR_AUTH_UNKNOWN); | |
2694 | } | |
2695 | resolve_gather = recovery_set; | |
11fdf7f2 TL |
2696 | |
2697 | resolve_snapclient_commits = mds->snapclient->get_journaled_tids(); | |
7c673cae FG |
2698 | } |
2699 | ||
2700 | void MDCache::send_resolves() | |
2701 | { | |
f67539c2 | 2702 | send_peer_resolves(); |
11fdf7f2 TL |
2703 | |
2704 | if (!resolve_done) { | |
2705 | // I'm survivor: refresh snap cache | |
2706 | mds->snapclient->sync( | |
2707 | new MDSInternalContextWrapper(mds, | |
9f95a23c | 2708 | new LambdaContext([this](int r) { |
f67539c2 | 2709 | maybe_finish_peer_resolve(); |
11fdf7f2 TL |
2710 | }) |
2711 | ) | |
2712 | ); | |
2713 | dout(10) << "send_resolves waiting for snapclient cache to sync" << dendl; | |
2714 | return; | |
2715 | } | |
7c673cae FG |
2716 | if (!resolve_ack_gather.empty()) { |
2717 | dout(10) << "send_resolves still waiting for resolve ack from (" | |
2718 | << resolve_ack_gather << ")" << dendl; | |
2719 | return; | |
2720 | } | |
11fdf7f2 | 2721 | if (!resolve_need_rollback.empty()) { |
7c673cae | 2722 | dout(10) << "send_resolves still waiting for rollback to commit on (" |
11fdf7f2 | 2723 | << resolve_need_rollback << ")" << dendl; |
7c673cae FG |
2724 | return; |
2725 | } | |
11fdf7f2 | 2726 | |
7c673cae FG |
2727 | send_subtree_resolves(); |
2728 | } | |
2729 | ||
f67539c2 | 2730 | void MDCache::send_peer_resolves() |
7c673cae | 2731 | { |
f67539c2 | 2732 | dout(10) << "send_peer_resolves" << dendl; |
7c673cae | 2733 | |
9f95a23c | 2734 | map<mds_rank_t, ref_t<MMDSResolve>> resolves; |
7c673cae FG |
2735 | |
2736 | if (mds->is_resolve()) { | |
f67539c2 TL |
2737 | for (map<metareqid_t, upeer>::iterator p = uncommitted_peers.begin(); |
2738 | p != uncommitted_peers.end(); | |
7c673cae | 2739 | ++p) { |
f67539c2 TL |
2740 | mds_rank_t leader = p->second.leader; |
2741 | auto &m = resolves[leader]; | |
e306af50 | 2742 | if (!m) m = make_message<MMDSResolve>(); |
f67539c2 | 2743 | m->add_peer_request(p->first, false); |
7c673cae FG |
2744 | } |
2745 | } else { | |
2746 | set<mds_rank_t> resolve_set; | |
2747 | mds->mdsmap->get_mds_set(resolve_set, MDSMap::STATE_RESOLVE); | |
2748 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
2749 | p != active_requests.end(); | |
2750 | ++p) { | |
2751 | MDRequestRef& mdr = p->second; | |
f67539c2 | 2752 | if (!mdr->is_peer()) |
7c673cae | 2753 | continue; |
f67539c2 | 2754 | if (!mdr->peer_did_prepare() && !mdr->committing) { |
7c673cae FG |
2755 | continue; |
2756 | } | |
f67539c2 TL |
2757 | mds_rank_t leader = mdr->peer_to_mds; |
2758 | if (resolve_set.count(leader) || is_ambiguous_peer_update(p->first, leader)) { | |
7c673cae | 2759 | dout(10) << " including uncommitted " << *mdr << dendl; |
f67539c2 TL |
2760 | if (!resolves.count(leader)) |
2761 | resolves[leader] = make_message<MMDSResolve>(); | |
7c673cae FG |
2762 | if (!mdr->committing && |
2763 | mdr->has_more() && mdr->more()->is_inode_exporter) { | |
2764 | // re-send cap exports | |
2765 | CInode *in = mdr->more()->rename_inode; | |
2766 | map<client_t, Capability::Export> cap_map; | |
2767 | in->export_client_caps(cap_map); | |
2768 | bufferlist bl; | |
f67539c2 | 2769 | MMDSResolve::peer_inode_cap inode_caps(in->ino(), cap_map); |
9f95a23c | 2770 | encode(inode_caps, bl); |
f67539c2 | 2771 | resolves[leader]->add_peer_request(p->first, bl); |
7c673cae | 2772 | } else { |
f67539c2 | 2773 | resolves[leader]->add_peer_request(p->first, mdr->committing); |
7c673cae FG |
2774 | } |
2775 | } | |
2776 | } | |
2777 | } | |
2778 | ||
11fdf7f2 | 2779 | for (auto &p : resolves) { |
f67539c2 | 2780 | dout(10) << "sending peer resolve to mds." << p.first << dendl; |
11fdf7f2 TL |
2781 | mds->send_message_mds(p.second, p.first); |
2782 | resolve_ack_gather.insert(p.first); | |
7c673cae FG |
2783 | } |
2784 | } | |
2785 | ||
2786 | void MDCache::send_subtree_resolves() | |
2787 | { | |
2788 | dout(10) << "send_subtree_resolves" << dendl; | |
2789 | ||
2790 | if (migrator->is_exporting() || migrator->is_importing()) { | |
2791 | dout(7) << "send_subtree_resolves waiting, imports/exports still in progress" << dendl; | |
2792 | migrator->show_importing(); | |
2793 | migrator->show_exporting(); | |
2794 | resolves_pending = true; | |
2795 | return; // not now | |
2796 | } | |
2797 | ||
9f95a23c | 2798 | map<mds_rank_t, ref_t<MMDSResolve>> resolves; |
7c673cae FG |
2799 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); |
2800 | p != recovery_set.end(); | |
2801 | ++p) { | |
2802 | if (*p == mds->get_nodeid()) | |
2803 | continue; | |
2804 | if (mds->is_resolve() || mds->mdsmap->is_resolve(*p)) | |
9f95a23c | 2805 | resolves[*p] = make_message<MMDSResolve>(); |
7c673cae FG |
2806 | } |
2807 | ||
2808 | map<dirfrag_t, vector<dirfrag_t> > my_subtrees; | |
2809 | map<dirfrag_t, vector<dirfrag_t> > my_ambig_imports; | |
2810 | ||
2811 | // known | |
2812 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
2813 | p != subtrees.end(); | |
2814 | ++p) { | |
2815 | CDir *dir = p->first; | |
2816 | ||
2817 | // only our subtrees | |
2818 | if (dir->authority().first != mds->get_nodeid()) | |
2819 | continue; | |
2820 | ||
2821 | if (mds->is_resolve() && my_ambiguous_imports.count(dir->dirfrag())) | |
2822 | continue; // we'll add it below | |
2823 | ||
2824 | if (migrator->is_ambiguous_import(dir->dirfrag())) { | |
2825 | // ambiguous (mid-import) | |
2826 | set<CDir*> bounds; | |
2827 | get_subtree_bounds(dir, bounds); | |
2828 | vector<dirfrag_t> dfls; | |
2829 | for (set<CDir*>::iterator q = bounds.begin(); q != bounds.end(); ++q) | |
2830 | dfls.push_back((*q)->dirfrag()); | |
2831 | ||
2832 | my_ambig_imports[dir->dirfrag()] = dfls; | |
2833 | dout(10) << " ambig " << dir->dirfrag() << " " << dfls << dendl; | |
2834 | } else { | |
2835 | // not ambiguous. | |
11fdf7f2 TL |
2836 | for (auto &q : resolves) { |
2837 | resolves[q.first]->add_subtree(dir->dirfrag()); | |
2838 | } | |
7c673cae FG |
2839 | // bounds too |
2840 | vector<dirfrag_t> dfls; | |
2841 | for (set<CDir*>::iterator q = subtrees[dir].begin(); | |
2842 | q != subtrees[dir].end(); | |
2843 | ++q) { | |
2844 | CDir *bound = *q; | |
2845 | dfls.push_back(bound->dirfrag()); | |
2846 | } | |
2847 | ||
2848 | my_subtrees[dir->dirfrag()] = dfls; | |
2849 | dout(10) << " claim " << dir->dirfrag() << " " << dfls << dendl; | |
2850 | } | |
2851 | } | |
2852 | ||
2853 | // ambiguous | |
2854 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator p = my_ambiguous_imports.begin(); | |
2855 | p != my_ambiguous_imports.end(); | |
2856 | ++p) { | |
2857 | my_ambig_imports[p->first] = p->second; | |
2858 | dout(10) << " ambig " << p->first << " " << p->second << dendl; | |
2859 | } | |
2860 | ||
2861 | // simplify the claimed subtree. | |
2862 | for (auto p = my_subtrees.begin(); p != my_subtrees.end(); ++p) { | |
2863 | unsigned i = 0; | |
2864 | while (i < p->second.size()) { | |
2865 | dirfrag_t b = p->second[i]; | |
2866 | if (my_subtrees.count(b)) { | |
2867 | vector<dirfrag_t>& bb = my_subtrees[b]; | |
2868 | dout(10) << " simplify: " << p->first << " swallowing " << b << " with bounds " << bb << dendl; | |
2869 | for (vector<dirfrag_t>::iterator r = bb.begin(); r != bb.end(); ++r) | |
2870 | p->second.push_back(*r); | |
2871 | my_subtrees.erase(b); | |
2872 | p->second.erase(p->second.begin() + i); | |
2873 | } else { | |
2874 | ++i; | |
2875 | } | |
2876 | } | |
2877 | } | |
2878 | ||
2879 | // send | |
11fdf7f2 | 2880 | for (auto &p : resolves) { |
9f95a23c | 2881 | const ref_t<MMDSResolve> &m = p.second; |
11fdf7f2 TL |
2882 | if (mds->is_resolve()) { |
2883 | m->add_table_commits(TABLE_SNAP, resolve_snapclient_commits); | |
2884 | } else { | |
2885 | m->add_table_commits(TABLE_SNAP, mds->snapclient->get_journaled_tids()); | |
2886 | } | |
7c673cae FG |
2887 | m->subtrees = my_subtrees; |
2888 | m->ambiguous_imports = my_ambig_imports; | |
11fdf7f2 TL |
2889 | dout(10) << "sending subtee resolve to mds." << p.first << dendl; |
2890 | mds->send_message_mds(m, p.first); | |
7c673cae FG |
2891 | } |
2892 | resolves_pending = false; | |
2893 | } | |
2894 | ||
f67539c2 | 2895 | void MDCache::maybe_finish_peer_resolve() { |
11fdf7f2 TL |
2896 | if (resolve_ack_gather.empty() && resolve_need_rollback.empty()) { |
2897 | // snap cache get synced or I'm in resolve state | |
2898 | if (mds->snapclient->is_synced() || resolve_done) | |
2899 | send_subtree_resolves(); | |
2900 | process_delayed_resolve(); | |
2901 | } | |
2902 | } | |
2903 | ||
7c673cae FG |
2904 | void MDCache::handle_mds_failure(mds_rank_t who) |
2905 | { | |
2906 | dout(7) << "handle_mds_failure mds." << who << dendl; | |
2907 | ||
2908 | dout(1) << "handle_mds_failure mds." << who << " : recovery peers are " << recovery_set << dendl; | |
2909 | ||
2910 | resolve_gather.insert(who); | |
2911 | discard_delayed_resolve(who); | |
f67539c2 | 2912 | ambiguous_peer_updates.erase(who); |
7c673cae FG |
2913 | |
2914 | rejoin_gather.insert(who); | |
2915 | rejoin_sent.erase(who); // i need to send another | |
31f18b77 | 2916 | rejoin_ack_sent.erase(who); // i need to send another |
7c673cae FG |
2917 | rejoin_ack_gather.erase(who); // i'll need/get another. |
2918 | ||
2919 | dout(10) << " resolve_gather " << resolve_gather << dendl; | |
2920 | dout(10) << " resolve_ack_gather " << resolve_ack_gather << dendl; | |
2921 | dout(10) << " rejoin_sent " << rejoin_sent << dendl; | |
2922 | dout(10) << " rejoin_gather " << rejoin_gather << dendl; | |
2923 | dout(10) << " rejoin_ack_gather " << rejoin_ack_gather << dendl; | |
2924 | ||
2925 | ||
2926 | // tell the migrator too. | |
2927 | migrator->handle_mds_failure_or_stop(who); | |
2928 | ||
224ce89b WB |
2929 | // tell the balancer too. |
2930 | mds->balancer->handle_mds_failure(who); | |
2931 | ||
f67539c2 | 2932 | // clean up any requests peer to/from this node |
7c673cae FG |
2933 | list<MDRequestRef> finish; |
2934 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
2935 | p != active_requests.end(); | |
2936 | ++p) { | |
2937 | MDRequestRef& mdr = p->second; | |
f67539c2 TL |
2938 | // peer to the failed node? |
2939 | if (mdr->peer_to_mds == who) { | |
2940 | if (mdr->peer_did_prepare()) { | |
2941 | dout(10) << " peer request " << *mdr << " uncommitted, will resolve shortly" << dendl; | |
2942 | if (is_ambiguous_peer_update(p->first, mdr->peer_to_mds)) | |
2943 | remove_ambiguous_peer_update(p->first, mdr->peer_to_mds); | |
2944 | ||
2945 | if (!mdr->more()->waiting_on_peer.empty()) { | |
11fdf7f2 | 2946 | ceph_assert(mdr->more()->srcdn_auth_mds == mds->get_nodeid()); |
7c673cae | 2947 | // will rollback, no need to wait |
f67539c2 TL |
2948 | mdr->reset_peer_request(); |
2949 | mdr->more()->waiting_on_peer.clear(); | |
7c673cae FG |
2950 | } |
2951 | } else if (!mdr->committing) { | |
f67539c2 TL |
2952 | dout(10) << " peer request " << *mdr << " has no prepare, finishing up" << dendl; |
2953 | if (mdr->peer_request || mdr->peer_rolling_back()) | |
7c673cae FG |
2954 | mdr->aborted = true; |
2955 | else | |
2956 | finish.push_back(mdr); | |
2957 | } | |
2958 | } | |
2959 | ||
f67539c2 TL |
2960 | if (mdr->is_peer() && mdr->peer_did_prepare()) { |
2961 | if (mdr->more()->waiting_on_peer.count(who)) { | |
11fdf7f2 | 2962 | ceph_assert(mdr->more()->srcdn_auth_mds == mds->get_nodeid()); |
f67539c2 | 2963 | dout(10) << " peer request " << *mdr << " no longer need rename notity ack from mds." |
7c673cae | 2964 | << who << dendl; |
f67539c2 TL |
2965 | mdr->more()->waiting_on_peer.erase(who); |
2966 | if (mdr->more()->waiting_on_peer.empty() && mdr->peer_request) | |
7c673cae FG |
2967 | mds->queue_waiter(new C_MDS_RetryRequest(this, mdr)); |
2968 | } | |
2969 | ||
2970 | if (mdr->more()->srcdn_auth_mds == who && | |
f67539c2 | 2971 | mds->mdsmap->is_clientreplay_or_active_or_stopping(mdr->peer_to_mds)) { |
7c673cae | 2972 | // rename srcdn's auth mds failed, resolve even I'm a survivor. |
f67539c2 TL |
2973 | dout(10) << " peer request " << *mdr << " uncommitted, will resolve shortly" << dendl; |
2974 | add_ambiguous_peer_update(p->first, mdr->peer_to_mds); | |
7c673cae | 2975 | } |
f67539c2 TL |
2976 | } else if (mdr->peer_request) { |
2977 | const cref_t<MMDSPeerRequest> &peer_req = mdr->peer_request; | |
2978 | // FIXME: Peer rename request can arrive after we notice mds failure. | |
31f18b77 | 2979 | // This can cause mds to crash (does not affect integrity of FS). |
f67539c2 TL |
2980 | if (peer_req->get_op() == MMDSPeerRequest::OP_RENAMEPREP && |
2981 | peer_req->srcdn_auth == who) | |
2982 | peer_req->mark_interrupted(); | |
7c673cae FG |
2983 | } |
2984 | ||
f67539c2 TL |
2985 | // failed node is peer? |
2986 | if (mdr->is_leader() && !mdr->committing) { | |
7c673cae | 2987 | if (mdr->more()->srcdn_auth_mds == who) { |
f67539c2 | 2988 | dout(10) << " leader request " << *mdr << " waiting for rename srcdn's auth mds." |
7c673cae | 2989 | << who << " to recover" << dendl; |
11fdf7f2 | 2990 | ceph_assert(mdr->more()->witnessed.count(who) == 0); |
7c673cae FG |
2991 | if (mdr->more()->is_ambiguous_auth) |
2992 | mdr->clear_ambiguous_auth(); | |
2993 | // rename srcdn's auth mds failed, all witnesses will rollback | |
2994 | mdr->more()->witnessed.clear(); | |
f67539c2 | 2995 | pending_leaders.erase(p->first); |
7c673cae FG |
2996 | } |
2997 | ||
2998 | if (mdr->more()->witnessed.count(who)) { | |
2999 | mds_rank_t srcdn_auth = mdr->more()->srcdn_auth_mds; | |
f67539c2 TL |
3000 | if (srcdn_auth >= 0 && mdr->more()->waiting_on_peer.count(srcdn_auth)) { |
3001 | dout(10) << " leader request " << *mdr << " waiting for rename srcdn's auth mds." | |
7c673cae | 3002 | << mdr->more()->srcdn_auth_mds << " to reply" << dendl; |
f67539c2 TL |
3003 | // waiting for the peer (rename srcdn's auth mds), delay sending resolve ack |
3004 | // until either the request is committing or the peer also fails. | |
3005 | ceph_assert(mdr->more()->waiting_on_peer.size() == 1); | |
3006 | pending_leaders.insert(p->first); | |
7c673cae | 3007 | } else { |
f67539c2 | 3008 | dout(10) << " leader request " << *mdr << " no longer witnessed by peer mds." |
7c673cae FG |
3009 | << who << " to recover" << dendl; |
3010 | if (srcdn_auth >= 0) | |
11fdf7f2 | 3011 | ceph_assert(mdr->more()->witnessed.count(srcdn_auth) == 0); |
7c673cae FG |
3012 | |
3013 | // discard this peer's prepare (if any) | |
3014 | mdr->more()->witnessed.erase(who); | |
3015 | } | |
3016 | } | |
3017 | ||
f67539c2 TL |
3018 | if (mdr->more()->waiting_on_peer.count(who)) { |
3019 | dout(10) << " leader request " << *mdr << " waiting for peer mds." << who | |
7c673cae FG |
3020 | << " to recover" << dendl; |
3021 | // retry request when peer recovers | |
f67539c2 TL |
3022 | mdr->more()->waiting_on_peer.erase(who); |
3023 | if (mdr->more()->waiting_on_peer.empty()) | |
7c673cae FG |
3024 | mds->wait_for_active_peer(who, new C_MDS_RetryRequest(this, mdr)); |
3025 | } | |
3026 | ||
3027 | if (mdr->locking && mdr->locking_target_mds == who) | |
3028 | mdr->finish_locking(mdr->locking); | |
3029 | } | |
3030 | } | |
3031 | ||
f67539c2 TL |
3032 | for (map<metareqid_t, uleader>::iterator p = uncommitted_leaders.begin(); |
3033 | p != uncommitted_leaders.end(); | |
7c673cae | 3034 | ++p) { |
f67539c2 TL |
3035 | // The failed MDS may have already committed the peer update |
3036 | if (p->second.peers.count(who)) { | |
7c673cae | 3037 | p->second.recovering = true; |
f67539c2 | 3038 | p->second.peers.erase(who); |
7c673cae FG |
3039 | } |
3040 | } | |
3041 | ||
3042 | while (!finish.empty()) { | |
f67539c2 | 3043 | dout(10) << "cleaning up peer request " << *finish.front() << dendl; |
7c673cae FG |
3044 | request_finish(finish.front()); |
3045 | finish.pop_front(); | |
3046 | } | |
3047 | ||
3048 | kick_find_ino_peers(who); | |
3049 | kick_open_ino_peers(who); | |
3050 | ||
3051 | for (map<dirfrag_t,fragment_info_t>::iterator p = fragments.begin(); | |
3052 | p != fragments.end(); ) { | |
3053 | dirfrag_t df = p->first; | |
3054 | fragment_info_t& info = p->second; | |
a8e16298 TL |
3055 | |
3056 | if (info.is_fragmenting()) { | |
3057 | if (info.notify_ack_waiting.erase(who) && | |
3058 | info.notify_ack_waiting.empty()) { | |
3059 | fragment_drop_locks(info); | |
3060 | fragment_maybe_finish(p++); | |
3061 | } else { | |
3062 | ++p; | |
3063 | } | |
7c673cae | 3064 | continue; |
a8e16298 TL |
3065 | } |
3066 | ||
3067 | ++p; | |
7c673cae | 3068 | dout(10) << "cancelling fragment " << df << " bit " << info.bits << dendl; |
9f95a23c | 3069 | std::vector<CDir*> dirs; |
7c673cae FG |
3070 | info.dirs.swap(dirs); |
3071 | fragments.erase(df); | |
3072 | fragment_unmark_unfreeze_dirs(dirs); | |
3073 | } | |
3074 | ||
3075 | // MDCache::shutdown_export_strays() always exports strays to mds.0 | |
3076 | if (who == mds_rank_t(0)) | |
f64942e4 | 3077 | shutdown_exporting_strays.clear(); |
7c673cae FG |
3078 | |
3079 | show_subtrees(); | |
3080 | } | |
3081 | ||
3082 | /* | |
3083 | * handle_mds_recovery - called on another node's transition | |
3084 | * from resolve -> active. | |
3085 | */ | |
3086 | void MDCache::handle_mds_recovery(mds_rank_t who) | |
3087 | { | |
3088 | dout(7) << "handle_mds_recovery mds." << who << dendl; | |
3089 | ||
3090 | // exclude all discover waiters. kick_discovers() will do the job | |
3091 | static const uint64_t i_mask = CInode::WAIT_ANY_MASK & ~CInode::WAIT_DIR; | |
3092 | static const uint64_t d_mask = CDir::WAIT_ANY_MASK & ~CDir::WAIT_DENTRY; | |
3093 | ||
11fdf7f2 | 3094 | MDSContext::vec waiters; |
7c673cae FG |
3095 | |
3096 | // wake up any waiters in their subtrees | |
3097 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3098 | p != subtrees.end(); | |
3099 | ++p) { | |
3100 | CDir *dir = p->first; | |
3101 | ||
3102 | if (dir->authority().first != who || | |
3103 | dir->authority().second == mds->get_nodeid()) | |
3104 | continue; | |
11fdf7f2 | 3105 | ceph_assert(!dir->is_auth()); |
7c673cae FG |
3106 | |
3107 | // wake any waiters | |
9f95a23c TL |
3108 | std::queue<CDir*> q; |
3109 | q.push(dir); | |
7c673cae FG |
3110 | |
3111 | while (!q.empty()) { | |
3112 | CDir *d = q.front(); | |
9f95a23c | 3113 | q.pop(); |
7c673cae FG |
3114 | d->take_waiting(d_mask, waiters); |
3115 | ||
3116 | // inode waiters too | |
94b18763 FG |
3117 | for (auto &p : d->items) { |
3118 | CDentry *dn = p.second; | |
7c673cae FG |
3119 | CDentry::linkage_t *dnl = dn->get_linkage(); |
3120 | if (dnl->is_primary()) { | |
3121 | dnl->get_inode()->take_waiting(i_mask, waiters); | |
3122 | ||
3123 | // recurse? | |
9f95a23c TL |
3124 | auto&& ls = dnl->get_inode()->get_dirfrags(); |
3125 | for (const auto& subdir : ls) { | |
7c673cae | 3126 | if (!subdir->is_subtree_root()) |
9f95a23c | 3127 | q.push(subdir); |
7c673cae FG |
3128 | } |
3129 | } | |
3130 | } | |
3131 | } | |
3132 | } | |
3133 | ||
3134 | kick_open_ino_peers(who); | |
3135 | kick_find_ino_peers(who); | |
3136 | ||
3137 | // queue them up. | |
3138 | mds->queue_waiters(waiters); | |
3139 | } | |
3140 | ||
3141 | void MDCache::set_recovery_set(set<mds_rank_t>& s) | |
3142 | { | |
3143 | dout(7) << "set_recovery_set " << s << dendl; | |
3144 | recovery_set = s; | |
3145 | } | |
3146 | ||
3147 | ||
3148 | /* | |
3149 | * during resolve state, we share resolves to determine who | |
3150 | * is authoritative for which trees. we expect to get an resolve | |
3151 | * from _everyone_ in the recovery_set (the mds cluster at the time of | |
3152 | * the first failure). | |
3153 | * | |
3154 | * This functions puts the passed message before returning | |
3155 | */ | |
9f95a23c | 3156 | void MDCache::handle_resolve(const cref_t<MMDSResolve> &m) |
7c673cae FG |
3157 | { |
3158 | dout(7) << "handle_resolve from " << m->get_source() << dendl; | |
3159 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
3160 | ||
3161 | if (mds->get_state() < MDSMap::STATE_RESOLVE) { | |
3162 | if (mds->get_want_state() == CEPH_MDS_STATE_RESOLVE) { | |
3163 | mds->wait_for_resolve(new C_MDS_RetryMessage(mds, m)); | |
3164 | return; | |
3165 | } | |
3166 | // wait until we reach the resolve stage! | |
7c673cae FG |
3167 | return; |
3168 | } | |
3169 | ||
3170 | discard_delayed_resolve(from); | |
3171 | ||
f67539c2 TL |
3172 | // ambiguous peer requests? |
3173 | if (!m->peer_requests.empty()) { | |
7c673cae | 3174 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { |
f67539c2 TL |
3175 | for (auto p = m->peer_requests.begin(); p != m->peer_requests.end(); ++p) { |
3176 | if (uncommitted_leaders.count(p->first) && !uncommitted_leaders[p->first].safe) { | |
11fdf7f2 | 3177 | ceph_assert(!p->second.committing); |
f67539c2 | 3178 | pending_leaders.insert(p->first); |
7c673cae FG |
3179 | } |
3180 | } | |
3181 | ||
f67539c2 TL |
3182 | if (!pending_leaders.empty()) { |
3183 | dout(10) << " still have pending updates, delay processing peer resolve" << dendl; | |
7c673cae FG |
3184 | delayed_resolve[from] = m; |
3185 | return; | |
3186 | } | |
3187 | } | |
3188 | ||
9f95a23c | 3189 | auto ack = make_message<MMDSResolveAck>(); |
f67539c2 TL |
3190 | for (const auto &p : m->peer_requests) { |
3191 | if (uncommitted_leaders.count(p.first)) { //mds->sessionmap.have_completed_request(p.first)) { | |
7c673cae | 3192 | // COMMIT |
11fdf7f2 | 3193 | if (p.second.committing) { |
f67539c2 TL |
3194 | // already committing, waiting for the OP_COMMITTED peer reply |
3195 | dout(10) << " already committing peer request " << p << " noop "<< dendl; | |
7c673cae | 3196 | } else { |
f67539c2 | 3197 | dout(10) << " ambiguous peer request " << p << " will COMMIT" << dendl; |
11fdf7f2 | 3198 | ack->add_commit(p.first); |
7c673cae | 3199 | } |
f67539c2 | 3200 | uncommitted_leaders[p.first].peers.insert(from); // wait for peer OP_COMMITTED before we log ECommitted |
7c673cae | 3201 | |
11fdf7f2 | 3202 | if (p.second.inode_caps.length() > 0) { |
f67539c2 | 3203 | // peer wants to export caps (rename) |
11fdf7f2 | 3204 | ceph_assert(mds->is_resolve()); |
f67539c2 | 3205 | MMDSResolve::peer_inode_cap inode_caps; |
11fdf7f2 | 3206 | auto q = p.second.inode_caps.cbegin(); |
9f95a23c TL |
3207 | decode(inode_caps, q); |
3208 | inodeno_t ino = inode_caps.ino; | |
3209 | map<client_t,Capability::Export> cap_exports = inode_caps.cap_exports; | |
11fdf7f2 | 3210 | ceph_assert(get_inode(ino)); |
7c673cae FG |
3211 | |
3212 | for (map<client_t,Capability::Export>::iterator q = cap_exports.begin(); | |
3213 | q != cap_exports.end(); | |
3214 | ++q) { | |
3215 | Capability::Import& im = rejoin_imported_caps[from][ino][q->first]; | |
3216 | im.cap_id = ++last_cap_id; // assign a new cap ID | |
3217 | im.issue_seq = 1; | |
3218 | im.mseq = q->second.mseq; | |
28e407b8 AA |
3219 | |
3220 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
3221 | if (session) | |
3222 | rejoin_client_map.emplace(q->first, session->info.inst); | |
7c673cae FG |
3223 | } |
3224 | ||
3225 | // will process these caps in rejoin stage | |
f67539c2 TL |
3226 | rejoin_peer_exports[ino].first = from; |
3227 | rejoin_peer_exports[ino].second.swap(cap_exports); | |
7c673cae | 3228 | |
f67539c2 | 3229 | // send information of imported caps back to peer |
11fdf7f2 | 3230 | encode(rejoin_imported_caps[from][ino], ack->commit[p.first]); |
7c673cae FG |
3231 | } |
3232 | } else { | |
3233 | // ABORT | |
f67539c2 | 3234 | dout(10) << " ambiguous peer request " << p << " will ABORT" << dendl; |
11fdf7f2 TL |
3235 | ceph_assert(!p.second.committing); |
3236 | ack->add_abort(p.first); | |
7c673cae FG |
3237 | } |
3238 | } | |
3239 | mds->send_message(ack, m->get_connection()); | |
7c673cae FG |
3240 | return; |
3241 | } | |
3242 | ||
11fdf7f2 | 3243 | if (!resolve_ack_gather.empty() || !resolve_need_rollback.empty()) { |
7c673cae FG |
3244 | dout(10) << "delay processing subtree resolve" << dendl; |
3245 | delayed_resolve[from] = m; | |
3246 | return; | |
3247 | } | |
3248 | ||
3249 | bool survivor = false; | |
3250 | // am i a surviving ambiguous importer? | |
3251 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
3252 | survivor = true; | |
3253 | // check for any import success/failure (from this node) | |
3254 | map<dirfrag_t, vector<dirfrag_t> >::iterator p = my_ambiguous_imports.begin(); | |
3255 | while (p != my_ambiguous_imports.end()) { | |
3256 | map<dirfrag_t, vector<dirfrag_t> >::iterator next = p; | |
3257 | ++next; | |
3258 | CDir *dir = get_dirfrag(p->first); | |
11fdf7f2 | 3259 | ceph_assert(dir); |
7c673cae FG |
3260 | dout(10) << "checking ambiguous import " << *dir << dendl; |
3261 | if (migrator->is_importing(dir->dirfrag()) && | |
3262 | migrator->get_import_peer(dir->dirfrag()) == from) { | |
11fdf7f2 | 3263 | ceph_assert(migrator->get_import_state(dir->dirfrag()) == Migrator::IMPORT_ACKING); |
7c673cae FG |
3264 | |
3265 | // check if sender claims the subtree | |
3266 | bool claimed_by_sender = false; | |
11fdf7f2 | 3267 | for (const auto &q : m->subtrees) { |
7c673cae | 3268 | // an ambiguous import won't race with a refragmentation; it's appropriate to force here. |
11fdf7f2 | 3269 | CDir *base = get_force_dirfrag(q.first, false); |
7c673cae FG |
3270 | if (!base || !base->contains(dir)) |
3271 | continue; // base not dir or an ancestor of dir, clearly doesn't claim dir. | |
3272 | ||
3273 | bool inside = true; | |
3274 | set<CDir*> bounds; | |
11fdf7f2 | 3275 | get_force_dirfrag_bound_set(q.second, bounds); |
7c673cae FG |
3276 | for (set<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) { |
3277 | CDir *bound = *p; | |
3278 | if (bound->contains(dir)) { | |
3279 | inside = false; // nope, bound is dir or parent of dir, not inside. | |
3280 | break; | |
3281 | } | |
3282 | } | |
3283 | if (inside) | |
3284 | claimed_by_sender = true; | |
3285 | } | |
3286 | ||
3287 | my_ambiguous_imports.erase(p); // no longer ambiguous. | |
3288 | if (claimed_by_sender) { | |
3289 | dout(7) << "ambiguous import failed on " << *dir << dendl; | |
3290 | migrator->import_reverse(dir); | |
3291 | } else { | |
3292 | dout(7) << "ambiguous import succeeded on " << *dir << dendl; | |
3293 | migrator->import_finish(dir, true); | |
3294 | } | |
3295 | } | |
3296 | p = next; | |
3297 | } | |
3298 | } | |
3299 | ||
3300 | // update my dir_auth values | |
3301 | // need to do this on recoverying nodes _and_ bystanders (to resolve ambiguous | |
3302 | // migrations between other nodes) | |
11fdf7f2 TL |
3303 | for (const auto& p : m->subtrees) { |
3304 | dout(10) << "peer claims " << p.first << " bounds " << p.second << dendl; | |
3305 | CDir *dir = get_force_dirfrag(p.first, !survivor); | |
7c673cae FG |
3306 | if (!dir) |
3307 | continue; | |
11fdf7f2 | 3308 | adjust_bounded_subtree_auth(dir, p.second, from); |
7c673cae FG |
3309 | try_subtree_merge(dir); |
3310 | } | |
3311 | ||
3312 | show_subtrees(); | |
3313 | ||
3314 | // note ambiguous imports too | |
11fdf7f2 TL |
3315 | for (const auto& p : m->ambiguous_imports) { |
3316 | dout(10) << "noting ambiguous import on " << p.first << " bounds " << p.second << dendl; | |
3317 | other_ambiguous_imports[from][p.first] = p.second; | |
3318 | } | |
3319 | ||
3320 | // learn other mds' pendina snaptable commits. later when resolve finishes, we will reload | |
3321 | // snaptable cache from snapserver. By this way, snaptable cache get synced among all mds | |
3322 | for (const auto& p : m->table_clients) { | |
3323 | dout(10) << " noting " << get_mdstable_name(p.type) | |
3324 | << " pending_commits " << p.pending_commits << dendl; | |
3325 | MDSTableClient *client = mds->get_table_client(p.type); | |
3326 | for (const auto& q : p.pending_commits) | |
3327 | client->notify_commit(q); | |
7c673cae FG |
3328 | } |
3329 | ||
3330 | // did i get them all? | |
3331 | resolve_gather.erase(from); | |
3332 | ||
3333 | maybe_resolve_finish(); | |
7c673cae FG |
3334 | } |
3335 | ||
3336 | void MDCache::process_delayed_resolve() | |
3337 | { | |
3338 | dout(10) << "process_delayed_resolve" << dendl; | |
9f95a23c | 3339 | map<mds_rank_t, cref_t<MMDSResolve>> tmp; |
7c673cae | 3340 | tmp.swap(delayed_resolve); |
11fdf7f2 TL |
3341 | for (auto &p : tmp) { |
3342 | handle_resolve(p.second); | |
3343 | } | |
7c673cae FG |
3344 | } |
3345 | ||
3346 | void MDCache::discard_delayed_resolve(mds_rank_t who) | |
3347 | { | |
11fdf7f2 | 3348 | delayed_resolve.erase(who); |
7c673cae FG |
3349 | } |
3350 | ||
3351 | void MDCache::maybe_resolve_finish() | |
3352 | { | |
11fdf7f2 TL |
3353 | ceph_assert(resolve_ack_gather.empty()); |
3354 | ceph_assert(resolve_need_rollback.empty()); | |
7c673cae FG |
3355 | |
3356 | if (!resolve_gather.empty()) { | |
3357 | dout(10) << "maybe_resolve_finish still waiting for resolves (" | |
3358 | << resolve_gather << ")" << dendl; | |
3359 | return; | |
3360 | } | |
3361 | ||
3362 | dout(10) << "maybe_resolve_finish got all resolves+resolve_acks, done." << dendl; | |
3363 | disambiguate_my_imports(); | |
f67539c2 | 3364 | finish_committed_leaders(); |
7c673cae FG |
3365 | |
3366 | if (resolve_done) { | |
11fdf7f2 | 3367 | ceph_assert(mds->is_resolve()); |
7c673cae FG |
3368 | trim_unlinked_inodes(); |
3369 | recalc_auth_bits(false); | |
3370 | resolve_done.release()->complete(0); | |
3371 | } else { | |
11fdf7f2 | 3372 | // I am survivor. |
7c673cae FG |
3373 | maybe_send_pending_rejoins(); |
3374 | } | |
3375 | } | |
3376 | ||
9f95a23c | 3377 | void MDCache::handle_resolve_ack(const cref_t<MMDSResolveAck> &ack) |
7c673cae FG |
3378 | { |
3379 | dout(10) << "handle_resolve_ack " << *ack << " from " << ack->get_source() << dendl; | |
3380 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
3381 | ||
3382 | if (!resolve_ack_gather.count(from) || | |
3383 | mds->mdsmap->get_state(from) < MDSMap::STATE_RESOLVE) { | |
7c673cae FG |
3384 | return; |
3385 | } | |
3386 | ||
f67539c2 | 3387 | if (ambiguous_peer_updates.count(from)) { |
11fdf7f2 TL |
3388 | ceph_assert(mds->mdsmap->is_clientreplay_or_active_or_stopping(from)); |
3389 | ceph_assert(mds->is_clientreplay() || mds->is_active() || mds->is_stopping()); | |
7c673cae FG |
3390 | } |
3391 | ||
11fdf7f2 | 3392 | for (const auto &p : ack->commit) { |
f67539c2 | 3393 | dout(10) << " commit on peer " << p.first << dendl; |
7c673cae | 3394 | |
f67539c2 TL |
3395 | if (ambiguous_peer_updates.count(from)) { |
3396 | remove_ambiguous_peer_update(p.first, from); | |
7c673cae FG |
3397 | continue; |
3398 | } | |
3399 | ||
3400 | if (mds->is_resolve()) { | |
3401 | // replay | |
f67539c2 | 3402 | MDPeerUpdate *su = get_uncommitted_peer(p.first, from); |
11fdf7f2 | 3403 | ceph_assert(su); |
7c673cae FG |
3404 | |
3405 | // log commit | |
f67539c2 TL |
3406 | mds->mdlog->start_submit_entry(new EPeerUpdate(mds->mdlog, "unknown", p.first, from, |
3407 | EPeerUpdate::OP_COMMIT, su->origop), | |
3408 | new C_MDC_PeerCommit(this, from, p.first)); | |
7c673cae FG |
3409 | mds->mdlog->flush(); |
3410 | ||
f67539c2 | 3411 | finish_uncommitted_peer(p.first); |
7c673cae | 3412 | } else { |
11fdf7f2 | 3413 | MDRequestRef mdr = request_get(p.first); |
f67539c2 | 3414 | // information about leader imported caps |
11fdf7f2 TL |
3415 | if (p.second.length() > 0) |
3416 | mdr->more()->inode_import.share(p.second); | |
7c673cae | 3417 | |
f67539c2 | 3418 | ceph_assert(mdr->peer_request == 0); // shouldn't be doing anything! |
7c673cae FG |
3419 | request_finish(mdr); |
3420 | } | |
3421 | } | |
3422 | ||
11fdf7f2 | 3423 | for (const auto &metareq : ack->abort) { |
f67539c2 | 3424 | dout(10) << " abort on peer " << metareq << dendl; |
7c673cae FG |
3425 | |
3426 | if (mds->is_resolve()) { | |
f67539c2 | 3427 | MDPeerUpdate *su = get_uncommitted_peer(metareq, from); |
11fdf7f2 | 3428 | ceph_assert(su); |
7c673cae FG |
3429 | |
3430 | // perform rollback (and journal a rollback entry) | |
3431 | // note: this will hold up the resolve a bit, until the rollback entries journal. | |
3432 | MDRequestRef null_ref; | |
3433 | switch (su->origop) { | |
f67539c2 | 3434 | case EPeerUpdate::LINK: |
7c673cae FG |
3435 | mds->server->do_link_rollback(su->rollback, from, null_ref); |
3436 | break; | |
f67539c2 | 3437 | case EPeerUpdate::RENAME: |
7c673cae FG |
3438 | mds->server->do_rename_rollback(su->rollback, from, null_ref); |
3439 | break; | |
f67539c2 | 3440 | case EPeerUpdate::RMDIR: |
7c673cae FG |
3441 | mds->server->do_rmdir_rollback(su->rollback, from, null_ref); |
3442 | break; | |
3443 | default: | |
3444 | ceph_abort(); | |
3445 | } | |
3446 | } else { | |
11fdf7f2 | 3447 | MDRequestRef mdr = request_get(metareq); |
7c673cae | 3448 | mdr->aborted = true; |
f67539c2 TL |
3449 | if (mdr->peer_request) { |
3450 | if (mdr->peer_did_prepare()) // journaling peer prepare ? | |
11fdf7f2 | 3451 | add_rollback(metareq, from); |
7c673cae FG |
3452 | } else { |
3453 | request_finish(mdr); | |
3454 | } | |
3455 | } | |
3456 | } | |
3457 | ||
f67539c2 | 3458 | if (!ambiguous_peer_updates.count(from)) { |
7c673cae | 3459 | resolve_ack_gather.erase(from); |
f67539c2 | 3460 | maybe_finish_peer_resolve(); |
7c673cae | 3461 | } |
7c673cae FG |
3462 | } |
3463 | ||
f67539c2 | 3464 | void MDCache::add_uncommitted_peer(metareqid_t reqid, LogSegment *ls, mds_rank_t leader, MDPeerUpdate *su) |
7c673cae | 3465 | { |
f67539c2 | 3466 | auto const &ret = uncommitted_peers.emplace(std::piecewise_construct, |
e306af50 TL |
3467 | std::forward_as_tuple(reqid), |
3468 | std::forward_as_tuple()); | |
3469 | ceph_assert(ret.second); | |
f67539c2 TL |
3470 | ls->uncommitted_peers.insert(reqid); |
3471 | upeer &u = ret.first->second; | |
3472 | u.leader = leader; | |
e306af50 TL |
3473 | u.ls = ls; |
3474 | u.su = su; | |
3475 | if (su == nullptr) { | |
3476 | return; | |
3477 | } | |
7c673cae | 3478 | for(set<CInode*>::iterator p = su->olddirs.begin(); p != su->olddirs.end(); ++p) |
f67539c2 | 3479 | uncommitted_peer_rename_olddir[*p]++; |
7c673cae | 3480 | for(set<CInode*>::iterator p = su->unlinked.begin(); p != su->unlinked.end(); ++p) |
f67539c2 | 3481 | uncommitted_peer_unlink[*p]++; |
7c673cae FG |
3482 | } |
3483 | ||
f67539c2 | 3484 | void MDCache::finish_uncommitted_peer(metareqid_t reqid, bool assert_exist) |
7c673cae | 3485 | { |
f67539c2 TL |
3486 | auto it = uncommitted_peers.find(reqid); |
3487 | if (it == uncommitted_peers.end()) { | |
e306af50 TL |
3488 | ceph_assert(!assert_exist); |
3489 | return; | |
3490 | } | |
f67539c2 TL |
3491 | upeer &u = it->second; |
3492 | MDPeerUpdate* su = u.su; | |
e306af50 TL |
3493 | |
3494 | if (!u.waiters.empty()) { | |
3495 | mds->queue_waiters(u.waiters); | |
3496 | } | |
f67539c2 TL |
3497 | u.ls->uncommitted_peers.erase(reqid); |
3498 | uncommitted_peers.erase(it); | |
7c673cae | 3499 | |
e306af50 TL |
3500 | if (su == nullptr) { |
3501 | return; | |
3502 | } | |
7c673cae FG |
3503 | // discard the non-auth subtree we renamed out of |
3504 | for(set<CInode*>::iterator p = su->olddirs.begin(); p != su->olddirs.end(); ++p) { | |
3505 | CInode *diri = *p; | |
f67539c2 TL |
3506 | map<CInode*, int>::iterator it = uncommitted_peer_rename_olddir.find(diri); |
3507 | ceph_assert(it != uncommitted_peer_rename_olddir.end()); | |
7c673cae FG |
3508 | it->second--; |
3509 | if (it->second == 0) { | |
f67539c2 | 3510 | uncommitted_peer_rename_olddir.erase(it); |
9f95a23c TL |
3511 | auto&& ls = diri->get_dirfrags(); |
3512 | for (const auto& dir : ls) { | |
3513 | CDir *root = get_subtree_root(dir); | |
7c673cae FG |
3514 | if (root->get_dir_auth() == CDIR_AUTH_UNDEF) { |
3515 | try_trim_non_auth_subtree(root); | |
9f95a23c | 3516 | if (dir != root) |
7c673cae FG |
3517 | break; |
3518 | } | |
3519 | } | |
3520 | } else | |
11fdf7f2 | 3521 | ceph_assert(it->second > 0); |
7c673cae | 3522 | } |
f67539c2 | 3523 | // removed the inodes that were unlinked by peer update |
7c673cae FG |
3524 | for(set<CInode*>::iterator p = su->unlinked.begin(); p != su->unlinked.end(); ++p) { |
3525 | CInode *in = *p; | |
f67539c2 TL |
3526 | map<CInode*, int>::iterator it = uncommitted_peer_unlink.find(in); |
3527 | ceph_assert(it != uncommitted_peer_unlink.end()); | |
7c673cae FG |
3528 | it->second--; |
3529 | if (it->second == 0) { | |
f67539c2 | 3530 | uncommitted_peer_unlink.erase(it); |
7c673cae FG |
3531 | if (!in->get_projected_parent_dn()) |
3532 | mds->mdcache->remove_inode_recursive(in); | |
3533 | } else | |
11fdf7f2 | 3534 | ceph_assert(it->second > 0); |
7c673cae FG |
3535 | } |
3536 | delete su; | |
3537 | } | |
3538 | ||
f67539c2 | 3539 | MDPeerUpdate* MDCache::get_uncommitted_peer(metareqid_t reqid, mds_rank_t leader) |
7c673cae FG |
3540 | { |
3541 | ||
f67539c2 TL |
3542 | MDPeerUpdate* su = nullptr; |
3543 | auto it = uncommitted_peers.find(reqid); | |
3544 | if (it != uncommitted_peers.end() && | |
3545 | it->second.leader == leader) { | |
e306af50 | 3546 | su = it->second.su; |
7c673cae FG |
3547 | } |
3548 | return su; | |
3549 | } | |
3550 | ||
e306af50 | 3551 | void MDCache::finish_rollback(metareqid_t reqid, MDRequestRef& mdr) { |
f91f0fd5 | 3552 | auto p = resolve_need_rollback.find(reqid); |
11fdf7f2 | 3553 | ceph_assert(p != resolve_need_rollback.end()); |
e306af50 | 3554 | if (mds->is_resolve()) { |
f67539c2 | 3555 | finish_uncommitted_peer(reqid, false); |
e306af50 | 3556 | } else if (mdr) { |
f67539c2 | 3557 | finish_uncommitted_peer(mdr->reqid, mdr->more()->peer_update_journaled); |
e306af50 | 3558 | } |
11fdf7f2 | 3559 | resolve_need_rollback.erase(p); |
f67539c2 | 3560 | maybe_finish_peer_resolve(); |
7c673cae FG |
3561 | } |
3562 | ||
3563 | void MDCache::disambiguate_other_imports() | |
3564 | { | |
3565 | dout(10) << "disambiguate_other_imports" << dendl; | |
3566 | ||
3567 | bool recovering = !(mds->is_clientreplay() || mds->is_active() || mds->is_stopping()); | |
3568 | // other nodes' ambiguous imports | |
3569 | for (map<mds_rank_t, map<dirfrag_t, vector<dirfrag_t> > >::iterator p = other_ambiguous_imports.begin(); | |
3570 | p != other_ambiguous_imports.end(); | |
3571 | ++p) { | |
3572 | mds_rank_t who = p->first; | |
3573 | dout(10) << "ambiguous imports for mds." << who << dendl; | |
3574 | ||
3575 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator q = p->second.begin(); | |
3576 | q != p->second.end(); | |
3577 | ++q) { | |
3578 | dout(10) << " ambiguous import " << q->first << " bounds " << q->second << dendl; | |
3579 | // an ambiguous import will not race with a refragmentation; it's appropriate to force here. | |
3580 | CDir *dir = get_force_dirfrag(q->first, recovering); | |
3581 | if (!dir) continue; | |
3582 | ||
3583 | if (dir->is_ambiguous_auth() || // works for me_ambig or if i am a surviving bystander | |
3584 | dir->authority() == CDIR_AUTH_UNDEF) { // resolving | |
3585 | dout(10) << " mds." << who << " did import " << *dir << dendl; | |
3586 | adjust_bounded_subtree_auth(dir, q->second, who); | |
3587 | try_subtree_merge(dir); | |
3588 | } else { | |
3589 | dout(10) << " mds." << who << " did not import " << *dir << dendl; | |
3590 | } | |
3591 | } | |
3592 | } | |
3593 | other_ambiguous_imports.clear(); | |
3594 | } | |
3595 | ||
3596 | void MDCache::disambiguate_my_imports() | |
3597 | { | |
3598 | dout(10) << "disambiguate_my_imports" << dendl; | |
3599 | ||
3600 | if (!mds->is_resolve()) { | |
11fdf7f2 | 3601 | ceph_assert(my_ambiguous_imports.empty()); |
7c673cae FG |
3602 | return; |
3603 | } | |
3604 | ||
3605 | disambiguate_other_imports(); | |
3606 | ||
3607 | // my ambiguous imports | |
3608 | mds_authority_t me_ambig(mds->get_nodeid(), mds->get_nodeid()); | |
3609 | while (!my_ambiguous_imports.empty()) { | |
3610 | map<dirfrag_t, vector<dirfrag_t> >::iterator q = my_ambiguous_imports.begin(); | |
3611 | ||
3612 | CDir *dir = get_dirfrag(q->first); | |
11fdf7f2 | 3613 | ceph_assert(dir); |
7c673cae FG |
3614 | |
3615 | if (dir->authority() != me_ambig) { | |
3616 | dout(10) << "ambiguous import auth known, must not be me " << *dir << dendl; | |
3617 | cancel_ambiguous_import(dir); | |
3618 | ||
3619 | mds->mdlog->start_submit_entry(new EImportFinish(dir, false)); | |
3620 | ||
3621 | // subtree may have been swallowed by another node claiming dir | |
3622 | // as their own. | |
3623 | CDir *root = get_subtree_root(dir); | |
3624 | if (root != dir) | |
3625 | dout(10) << " subtree root is " << *root << dendl; | |
11fdf7f2 | 3626 | ceph_assert(root->dir_auth.first != mds->get_nodeid()); // no us! |
7c673cae FG |
3627 | try_trim_non_auth_subtree(root); |
3628 | } else { | |
3629 | dout(10) << "ambiguous import auth unclaimed, must be me " << *dir << dendl; | |
3630 | finish_ambiguous_import(q->first); | |
3631 | mds->mdlog->start_submit_entry(new EImportFinish(dir, true)); | |
3632 | } | |
3633 | } | |
11fdf7f2 | 3634 | ceph_assert(my_ambiguous_imports.empty()); |
7c673cae FG |
3635 | mds->mdlog->flush(); |
3636 | ||
3637 | // verify all my subtrees are unambiguous! | |
3638 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3639 | p != subtrees.end(); | |
3640 | ++p) { | |
3641 | CDir *dir = p->first; | |
3642 | if (dir->is_ambiguous_dir_auth()) { | |
3643 | dout(0) << "disambiguate_imports uh oh, dir_auth is still ambiguous for " << *dir << dendl; | |
3644 | } | |
11fdf7f2 | 3645 | ceph_assert(!dir->is_ambiguous_dir_auth()); |
7c673cae FG |
3646 | } |
3647 | ||
3648 | show_subtrees(); | |
3649 | } | |
3650 | ||
3651 | ||
3652 | void MDCache::add_ambiguous_import(dirfrag_t base, const vector<dirfrag_t>& bounds) | |
3653 | { | |
11fdf7f2 | 3654 | ceph_assert(my_ambiguous_imports.count(base) == 0); |
7c673cae FG |
3655 | my_ambiguous_imports[base] = bounds; |
3656 | } | |
3657 | ||
3658 | ||
3659 | void MDCache::add_ambiguous_import(CDir *base, const set<CDir*>& bounds) | |
3660 | { | |
3661 | // make a list | |
3662 | vector<dirfrag_t> binos; | |
3663 | for (set<CDir*>::iterator p = bounds.begin(); | |
3664 | p != bounds.end(); | |
3665 | ++p) | |
3666 | binos.push_back((*p)->dirfrag()); | |
3667 | ||
3668 | // note: this can get called twice if the exporter fails during recovery | |
3669 | if (my_ambiguous_imports.count(base->dirfrag())) | |
3670 | my_ambiguous_imports.erase(base->dirfrag()); | |
3671 | ||
3672 | add_ambiguous_import(base->dirfrag(), binos); | |
3673 | } | |
3674 | ||
3675 | void MDCache::cancel_ambiguous_import(CDir *dir) | |
3676 | { | |
3677 | dirfrag_t df = dir->dirfrag(); | |
11fdf7f2 | 3678 | ceph_assert(my_ambiguous_imports.count(df)); |
7c673cae FG |
3679 | dout(10) << "cancel_ambiguous_import " << df |
3680 | << " bounds " << my_ambiguous_imports[df] | |
3681 | << " " << *dir | |
3682 | << dendl; | |
3683 | my_ambiguous_imports.erase(df); | |
3684 | } | |
3685 | ||
3686 | void MDCache::finish_ambiguous_import(dirfrag_t df) | |
3687 | { | |
11fdf7f2 | 3688 | ceph_assert(my_ambiguous_imports.count(df)); |
7c673cae FG |
3689 | vector<dirfrag_t> bounds; |
3690 | bounds.swap(my_ambiguous_imports[df]); | |
3691 | my_ambiguous_imports.erase(df); | |
3692 | ||
3693 | dout(10) << "finish_ambiguous_import " << df | |
3694 | << " bounds " << bounds | |
3695 | << dendl; | |
3696 | CDir *dir = get_dirfrag(df); | |
11fdf7f2 | 3697 | ceph_assert(dir); |
7c673cae FG |
3698 | |
3699 | // adjust dir_auth, import maps | |
3700 | adjust_bounded_subtree_auth(dir, bounds, mds->get_nodeid()); | |
3701 | try_subtree_merge(dir); | |
3702 | } | |
3703 | ||
3704 | void MDCache::remove_inode_recursive(CInode *in) | |
3705 | { | |
3706 | dout(10) << "remove_inode_recursive " << *in << dendl; | |
9f95a23c TL |
3707 | auto&& ls = in->get_dirfrags(); |
3708 | for (const auto& subdir : ls) { | |
3709 | dout(10) << " removing dirfrag " << *subdir << dendl; | |
94b18763 FG |
3710 | auto it = subdir->items.begin(); |
3711 | while (it != subdir->items.end()) { | |
3712 | CDentry *dn = it->second; | |
3713 | ++it; | |
7c673cae FG |
3714 | CDentry::linkage_t *dnl = dn->get_linkage(); |
3715 | if (dnl->is_primary()) { | |
3716 | CInode *tin = dnl->get_inode(); | |
31f18b77 | 3717 | subdir->unlink_inode(dn, false); |
7c673cae FG |
3718 | remove_inode_recursive(tin); |
3719 | } | |
3720 | subdir->remove_dentry(dn); | |
3721 | } | |
3722 | ||
3723 | if (subdir->is_subtree_root()) | |
3724 | remove_subtree(subdir); | |
3725 | in->close_dirfrag(subdir->dirfrag().frag); | |
3726 | } | |
3727 | remove_inode(in); | |
3728 | } | |
3729 | ||
11fdf7f2 | 3730 | bool MDCache::expire_recursive(CInode *in, expiremap &expiremap) |
7c673cae | 3731 | { |
11fdf7f2 | 3732 | ceph_assert(!in->is_auth()); |
7c673cae FG |
3733 | |
3734 | dout(10) << __func__ << ":" << *in << dendl; | |
3735 | ||
3736 | // Recurse into any dirfrags beneath this inode | |
9f95a23c TL |
3737 | auto&& ls = in->get_dirfrags(); |
3738 | for (const auto& subdir : ls) { | |
7c673cae FG |
3739 | if (!in->is_mdsdir() && subdir->is_subtree_root()) { |
3740 | dout(10) << __func__ << ": stray still has subtree " << *in << dendl; | |
3741 | return true; | |
3742 | } | |
3743 | ||
1d09f67e TL |
3744 | for (auto it = subdir->items.begin(); it != subdir->items.end();) { |
3745 | CDentry *dn = it->second; | |
3746 | it++; | |
7c673cae FG |
3747 | CDentry::linkage_t *dnl = dn->get_linkage(); |
3748 | if (dnl->is_primary()) { | |
3749 | CInode *tin = dnl->get_inode(); | |
3750 | ||
3751 | /* Remote strays with linkage (i.e. hardlinks) should not be | |
3752 | * expired, because they may be the target of | |
3753 | * a rename() as the owning MDS shuts down */ | |
f67539c2 | 3754 | if (!tin->is_stray() && tin->get_inode()->nlink) { |
7c673cae FG |
3755 | dout(10) << __func__ << ": stray still has linkage " << *tin << dendl; |
3756 | return true; | |
3757 | } | |
3758 | ||
3759 | const bool abort = expire_recursive(tin, expiremap); | |
3760 | if (abort) { | |
3761 | return true; | |
3762 | } | |
3763 | } | |
3764 | if (dn->lru_is_expireable()) { | |
3765 | trim_dentry(dn, expiremap); | |
3766 | } else { | |
3767 | dout(10) << __func__ << ": stray dn is not expireable " << *dn << dendl; | |
3768 | return true; | |
3769 | } | |
3770 | } | |
3771 | } | |
3772 | ||
3773 | return false; | |
3774 | } | |
3775 | ||
3776 | void MDCache::trim_unlinked_inodes() | |
3777 | { | |
3778 | dout(7) << "trim_unlinked_inodes" << dendl; | |
81eedcae TL |
3779 | int count = 0; |
3780 | vector<CInode*> q; | |
94b18763 | 3781 | for (auto &p : inode_map) { |
b32b8144 | 3782 | CInode *in = p.second; |
7c673cae FG |
3783 | if (in->get_parent_dn() == NULL && !in->is_base()) { |
3784 | dout(7) << " will trim from " << *in << dendl; | |
3785 | q.push_back(in); | |
3786 | } | |
81eedcae | 3787 | |
33c7a0ef | 3788 | if (!(++count % mds->heartbeat_reset_grace())) |
81eedcae TL |
3789 | mds->heartbeat_reset(); |
3790 | } | |
81eedcae TL |
3791 | for (auto& in : q) { |
3792 | remove_inode_recursive(in); | |
3793 | ||
33c7a0ef | 3794 | if (!(++count % mds->heartbeat_reset_grace())) |
81eedcae | 3795 | mds->heartbeat_reset(); |
7c673cae | 3796 | } |
7c673cae FG |
3797 | } |
3798 | ||
3799 | /** recalc_auth_bits() | |
3800 | * once subtree auth is disambiguated, we need to adjust all the | |
3801 | * auth and dirty bits in our cache before moving on. | |
3802 | */ | |
3803 | void MDCache::recalc_auth_bits(bool replay) | |
3804 | { | |
3805 | dout(7) << "recalc_auth_bits " << (replay ? "(replay)" : "") << dendl; | |
3806 | ||
3807 | if (root) { | |
3808 | root->inode_auth.first = mds->mdsmap->get_root(); | |
3809 | bool auth = mds->get_nodeid() == root->inode_auth.first; | |
3810 | if (auth) { | |
3811 | root->state_set(CInode::STATE_AUTH); | |
3812 | } else { | |
3813 | root->state_clear(CInode::STATE_AUTH); | |
3814 | if (!replay) | |
3815 | root->state_set(CInode::STATE_REJOINING); | |
3816 | } | |
3817 | } | |
3818 | ||
3819 | set<CInode*> subtree_inodes; | |
3820 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3821 | p != subtrees.end(); | |
3822 | ++p) { | |
3823 | if (p->first->dir_auth.first == mds->get_nodeid()) | |
3824 | subtree_inodes.insert(p->first->inode); | |
3825 | } | |
3826 | ||
3827 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3828 | p != subtrees.end(); | |
3829 | ++p) { | |
3830 | if (p->first->inode->is_mdsdir()) { | |
3831 | CInode *in = p->first->inode; | |
3832 | bool auth = in->ino() == MDS_INO_MDSDIR(mds->get_nodeid()); | |
3833 | if (auth) { | |
3834 | in->state_set(CInode::STATE_AUTH); | |
3835 | } else { | |
3836 | in->state_clear(CInode::STATE_AUTH); | |
3837 | if (!replay) | |
3838 | in->state_set(CInode::STATE_REJOINING); | |
3839 | } | |
3840 | } | |
3841 | ||
9f95a23c TL |
3842 | std::queue<CDir*> dfq; // dirfrag queue |
3843 | dfq.push(p->first); | |
7c673cae FG |
3844 | |
3845 | bool auth = p->first->authority().first == mds->get_nodeid(); | |
3846 | dout(10) << " subtree auth=" << auth << " for " << *p->first << dendl; | |
3847 | ||
3848 | while (!dfq.empty()) { | |
3849 | CDir *dir = dfq.front(); | |
9f95a23c | 3850 | dfq.pop(); |
7c673cae FG |
3851 | |
3852 | // dir | |
3853 | if (auth) { | |
3854 | dir->state_set(CDir::STATE_AUTH); | |
3855 | } else { | |
3856 | dir->state_clear(CDir::STATE_AUTH); | |
3857 | if (!replay) { | |
3858 | // close empty non-auth dirfrag | |
3859 | if (!dir->is_subtree_root() && dir->get_num_any() == 0) { | |
3860 | dir->inode->close_dirfrag(dir->get_frag()); | |
3861 | continue; | |
3862 | } | |
3863 | dir->state_set(CDir::STATE_REJOINING); | |
3864 | dir->state_clear(CDir::STATE_COMPLETE); | |
3865 | if (dir->is_dirty()) | |
3866 | dir->mark_clean(); | |
3867 | } | |
3868 | } | |
3869 | ||
3870 | // dentries in this dir | |
94b18763 | 3871 | for (auto &p : dir->items) { |
7c673cae | 3872 | // dn |
94b18763 | 3873 | CDentry *dn = p.second; |
7c673cae FG |
3874 | CDentry::linkage_t *dnl = dn->get_linkage(); |
3875 | if (auth) { | |
3876 | dn->state_set(CDentry::STATE_AUTH); | |
3877 | } else { | |
3878 | dn->state_clear(CDentry::STATE_AUTH); | |
3879 | if (!replay) { | |
3880 | dn->state_set(CDentry::STATE_REJOINING); | |
3881 | if (dn->is_dirty()) | |
3882 | dn->mark_clean(); | |
3883 | } | |
3884 | } | |
3885 | ||
3886 | if (dnl->is_primary()) { | |
3887 | // inode | |
3888 | CInode *in = dnl->get_inode(); | |
3889 | if (auth) { | |
3890 | in->state_set(CInode::STATE_AUTH); | |
3891 | } else { | |
3892 | in->state_clear(CInode::STATE_AUTH); | |
3893 | if (!replay) { | |
3894 | in->state_set(CInode::STATE_REJOINING); | |
3895 | if (in->is_dirty()) | |
3896 | in->mark_clean(); | |
3897 | if (in->is_dirty_parent()) | |
3898 | in->clear_dirty_parent(); | |
3899 | // avoid touching scatterlocks for our subtree roots! | |
3900 | if (subtree_inodes.count(in) == 0) | |
3901 | in->clear_scatter_dirty(); | |
3902 | } | |
3903 | } | |
3904 | // recurse? | |
9f95a23c TL |
3905 | if (in->is_dir()) { |
3906 | auto&& dfv = in->get_nested_dirfrags(); | |
3907 | for (const auto& dir : dfv) { | |
3908 | dfq.push(dir); | |
3909 | } | |
3910 | } | |
7c673cae FG |
3911 | } |
3912 | } | |
3913 | } | |
3914 | } | |
3915 | ||
3916 | show_subtrees(); | |
3917 | show_cache(); | |
3918 | } | |
3919 | ||
3920 | ||
3921 | ||
3922 | // =========================================================================== | |
3923 | // REJOIN | |
3924 | ||
3925 | /* | |
3926 | * notes on scatterlock recovery: | |
3927 | * | |
3928 | * - recovering inode replica sends scatterlock data for any subtree | |
3929 | * roots (the only ones that are possibly dirty). | |
3930 | * | |
3931 | * - surviving auth incorporates any provided scatterlock data. any | |
3932 | * pending gathers are then finished, as with the other lock types. | |
3933 | * | |
3934 | * that takes care of surviving auth + (recovering replica)*. | |
3935 | * | |
3936 | * - surviving replica sends strong_inode, which includes current | |
3937 | * scatterlock state, AND any dirty scatterlock data. this | |
3938 | * provides the recovering auth with everything it might need. | |
3939 | * | |
3940 | * - recovering auth must pick initial scatterlock state based on | |
3941 | * (weak|strong) rejoins. | |
3942 | * - always assimilate scatterlock data (it can't hurt) | |
3943 | * - any surviving replica in SCATTER state -> SCATTER. otherwise, SYNC. | |
3944 | * - include base inode in ack for all inodes that saw scatterlock content | |
3945 | * | |
3946 | * also, for scatter gather, | |
3947 | * | |
3948 | * - auth increments {frag,r}stat.version on completion of any gather. | |
3949 | * | |
3950 | * - auth incorporates changes in a gather _only_ if the version | |
3951 | * matches. | |
3952 | * | |
3953 | * - replica discards changes any time the scatterlock syncs, and | |
3954 | * after recovery. | |
3955 | */ | |
3956 | ||
3957 | void MDCache::dump_rejoin_status(Formatter *f) const | |
3958 | { | |
3959 | f->open_object_section("rejoin_status"); | |
3960 | f->dump_stream("rejoin_gather") << rejoin_gather; | |
3961 | f->dump_stream("rejoin_ack_gather") << rejoin_ack_gather; | |
3962 | f->dump_unsigned("num_opening_inodes", cap_imports_num_opening); | |
3963 | f->close_section(); | |
3964 | } | |
3965 | ||
11fdf7f2 | 3966 | void MDCache::rejoin_start(MDSContext *rejoin_done_) |
7c673cae FG |
3967 | { |
3968 | dout(10) << "rejoin_start" << dendl; | |
11fdf7f2 | 3969 | ceph_assert(!rejoin_done); |
7c673cae FG |
3970 | rejoin_done.reset(rejoin_done_); |
3971 | ||
3972 | rejoin_gather = recovery_set; | |
3973 | // need finish opening cap inodes before sending cache rejoins | |
3974 | rejoin_gather.insert(mds->get_nodeid()); | |
3975 | process_imported_caps(); | |
3976 | } | |
3977 | ||
3978 | /* | |
3979 | * rejoin phase! | |
3980 | * | |
11fdf7f2 | 3981 | * this initiates rejoin. it should be called before we get any |
7c673cae FG |
3982 | * rejoin or rejoin_ack messages (or else mdsmap distribution is broken). |
3983 | * | |
3984 | * we start out by sending rejoins to everyone in the recovery set. | |
3985 | * | |
3986 | * if we are rejoin, send for all regions in our cache. | |
11fdf7f2 | 3987 | * if we are active|stopping, send only to nodes that are rejoining. |
7c673cae FG |
3988 | */ |
3989 | void MDCache::rejoin_send_rejoins() | |
3990 | { | |
3991 | dout(10) << "rejoin_send_rejoins with recovery_set " << recovery_set << dendl; | |
3992 | ||
3993 | if (rejoin_gather.count(mds->get_nodeid())) { | |
3994 | dout(7) << "rejoin_send_rejoins still processing imported caps, delaying" << dendl; | |
3995 | rejoins_pending = true; | |
3996 | return; | |
3997 | } | |
3998 | if (!resolve_gather.empty()) { | |
3999 | dout(7) << "rejoin_send_rejoins still waiting for resolves (" | |
4000 | << resolve_gather << ")" << dendl; | |
4001 | rejoins_pending = true; | |
4002 | return; | |
4003 | } | |
4004 | ||
11fdf7f2 TL |
4005 | ceph_assert(!migrator->is_importing()); |
4006 | ceph_assert(!migrator->is_exporting()); | |
7c673cae FG |
4007 | |
4008 | if (!mds->is_rejoin()) { | |
4009 | disambiguate_other_imports(); | |
4010 | } | |
4011 | ||
9f95a23c | 4012 | map<mds_rank_t, ref_t<MMDSCacheRejoin>> rejoins; |
7c673cae FG |
4013 | |
4014 | ||
4015 | // if i am rejoining, send a rejoin to everyone. | |
4016 | // otherwise, just send to others who are rejoining. | |
9f95a23c TL |
4017 | for (const auto& rank : recovery_set) { |
4018 | if (rank == mds->get_nodeid()) continue; // nothing to myself! | |
4019 | if (rejoin_sent.count(rank)) continue; // already sent a rejoin to this node! | |
7c673cae | 4020 | if (mds->is_rejoin()) |
9f95a23c TL |
4021 | rejoins[rank] = make_message<MMDSCacheRejoin>(MMDSCacheRejoin::OP_WEAK); |
4022 | else if (mds->mdsmap->is_rejoin(rank)) | |
4023 | rejoins[rank] = make_message<MMDSCacheRejoin>(MMDSCacheRejoin::OP_STRONG); | |
7c673cae FG |
4024 | } |
4025 | ||
4026 | if (mds->is_rejoin()) { | |
11fdf7f2 TL |
4027 | map<client_t, pair<Session*, set<mds_rank_t> > > client_exports; |
4028 | for (auto& p : cap_exports) { | |
4029 | mds_rank_t target = p.second.first; | |
7c673cae FG |
4030 | if (rejoins.count(target) == 0) |
4031 | continue; | |
11fdf7f2 TL |
4032 | for (auto q = p.second.second.begin(); q != p.second.second.end(); ) { |
4033 | Session *session = nullptr; | |
4034 | auto it = client_exports.find(q->first); | |
4035 | if (it != client_exports.end()) { | |
4036 | session = it->second.first; | |
4037 | if (session) | |
4038 | it->second.second.insert(target); | |
4039 | } else { | |
4040 | session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
4041 | auto& r = client_exports[q->first]; | |
4042 | r.first = session; | |
4043 | if (session) | |
4044 | r.second.insert(target); | |
4045 | } | |
4046 | if (session) { | |
4047 | ++q; | |
4048 | } else { | |
4049 | // remove reconnect with no session | |
4050 | p.second.second.erase(q++); | |
4051 | } | |
4052 | } | |
4053 | rejoins[target]->cap_exports[p.first] = p.second.second; | |
7c673cae | 4054 | } |
11fdf7f2 TL |
4055 | for (auto& p : client_exports) { |
4056 | Session *session = p.second.first; | |
4057 | for (auto& q : p.second.second) { | |
4058 | auto rejoin = rejoins[q]; | |
4059 | rejoin->client_map[p.first] = session->info.inst; | |
4060 | rejoin->client_metadata_map[p.first] = session->info.client_metadata; | |
4061 | } | |
7c673cae FG |
4062 | } |
4063 | } | |
4064 | ||
4065 | ||
4066 | // check all subtrees | |
4067 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
4068 | p != subtrees.end(); | |
4069 | ++p) { | |
4070 | CDir *dir = p->first; | |
11fdf7f2 | 4071 | ceph_assert(dir->is_subtree_root()); |
7c673cae FG |
4072 | if (dir->is_ambiguous_dir_auth()) { |
4073 | // exporter is recovering, importer is survivor. | |
11fdf7f2 TL |
4074 | ceph_assert(rejoins.count(dir->authority().first)); |
4075 | ceph_assert(!rejoins.count(dir->authority().second)); | |
7c673cae FG |
4076 | continue; |
4077 | } | |
4078 | ||
4079 | // my subtree? | |
4080 | if (dir->is_auth()) | |
4081 | continue; // skip my own regions! | |
4082 | ||
4083 | mds_rank_t auth = dir->get_dir_auth().first; | |
11fdf7f2 | 4084 | ceph_assert(auth >= 0); |
7c673cae FG |
4085 | if (rejoins.count(auth) == 0) |
4086 | continue; // don't care about this node's subtrees | |
4087 | ||
4088 | rejoin_walk(dir, rejoins[auth]); | |
4089 | } | |
4090 | ||
4091 | // rejoin root inodes, too | |
11fdf7f2 | 4092 | for (auto &p : rejoins) { |
7c673cae FG |
4093 | if (mds->is_rejoin()) { |
4094 | // weak | |
11fdf7f2 TL |
4095 | if (p.first == 0 && root) { |
4096 | p.second->add_weak_inode(root->vino()); | |
7c673cae FG |
4097 | if (root->is_dirty_scattered()) { |
4098 | dout(10) << " sending scatterlock state on root " << *root << dendl; | |
11fdf7f2 | 4099 | p.second->add_scatterlock_state(root); |
7c673cae FG |
4100 | } |
4101 | } | |
11fdf7f2 | 4102 | if (CInode *in = get_inode(MDS_INO_MDSDIR(p.first))) { |
7c673cae | 4103 | if (in) |
11fdf7f2 | 4104 | p.second->add_weak_inode(in->vino()); |
7c673cae FG |
4105 | } |
4106 | } else { | |
4107 | // strong | |
11fdf7f2 TL |
4108 | if (p.first == 0 && root) { |
4109 | p.second->add_strong_inode(root->vino(), | |
7c673cae FG |
4110 | root->get_replica_nonce(), |
4111 | root->get_caps_wanted(), | |
4112 | root->filelock.get_state(), | |
4113 | root->nestlock.get_state(), | |
4114 | root->dirfragtreelock.get_state()); | |
4115 | root->state_set(CInode::STATE_REJOINING); | |
4116 | if (root->is_dirty_scattered()) { | |
4117 | dout(10) << " sending scatterlock state on root " << *root << dendl; | |
11fdf7f2 | 4118 | p.second->add_scatterlock_state(root); |
7c673cae FG |
4119 | } |
4120 | } | |
4121 | ||
11fdf7f2 TL |
4122 | if (CInode *in = get_inode(MDS_INO_MDSDIR(p.first))) { |
4123 | p.second->add_strong_inode(in->vino(), | |
7c673cae FG |
4124 | in->get_replica_nonce(), |
4125 | in->get_caps_wanted(), | |
4126 | in->filelock.get_state(), | |
4127 | in->nestlock.get_state(), | |
4128 | in->dirfragtreelock.get_state()); | |
4129 | in->state_set(CInode::STATE_REJOINING); | |
4130 | } | |
4131 | } | |
4132 | } | |
4133 | ||
4134 | if (!mds->is_rejoin()) { | |
4135 | // i am survivor. send strong rejoin. | |
4136 | // note request remote_auth_pins, xlocks | |
4137 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
4138 | p != active_requests.end(); | |
4139 | ++p) { | |
4140 | MDRequestRef& mdr = p->second; | |
f67539c2 | 4141 | if (mdr->is_peer()) |
7c673cae FG |
4142 | continue; |
4143 | // auth pins | |
9f95a23c TL |
4144 | for (const auto& q : mdr->object_states) { |
4145 | if (q.second.remote_auth_pinned == MDS_RANK_NONE) | |
4146 | continue; | |
11fdf7f2 | 4147 | if (!q.first->is_auth()) { |
9f95a23c TL |
4148 | mds_rank_t target = q.second.remote_auth_pinned; |
4149 | ceph_assert(target == q.first->authority().first); | |
4150 | if (rejoins.count(target) == 0) continue; | |
4151 | const auto& rejoin = rejoins[target]; | |
7c673cae | 4152 | |
11fdf7f2 | 4153 | dout(15) << " " << *mdr << " authpin on " << *q.first << dendl; |
7c673cae | 4154 | MDSCacheObjectInfo i; |
11fdf7f2 | 4155 | q.first->set_object_info(i); |
7c673cae FG |
4156 | if (i.ino) |
4157 | rejoin->add_inode_authpin(vinodeno_t(i.ino, i.snapid), mdr->reqid, mdr->attempt); | |
4158 | else | |
4159 | rejoin->add_dentry_authpin(i.dirfrag, i.dname, i.snapid, mdr->reqid, mdr->attempt); | |
4160 | ||
4161 | if (mdr->has_more() && mdr->more()->is_remote_frozen_authpin && | |
11fdf7f2 | 4162 | mdr->more()->rename_inode == q.first) |
7c673cae FG |
4163 | rejoin->add_inode_frozen_authpin(vinodeno_t(i.ino, i.snapid), |
4164 | mdr->reqid, mdr->attempt); | |
4165 | } | |
4166 | } | |
4167 | // xlocks | |
11fdf7f2 TL |
4168 | for (const auto& q : mdr->locks) { |
4169 | auto lock = q.lock; | |
4170 | auto obj = lock->get_parent(); | |
4171 | if (q.is_xlock() && !obj->is_auth()) { | |
4172 | mds_rank_t who = obj->authority().first; | |
7c673cae | 4173 | if (rejoins.count(who) == 0) continue; |
9f95a23c | 4174 | const auto& rejoin = rejoins[who]; |
7c673cae | 4175 | |
11fdf7f2 | 4176 | dout(15) << " " << *mdr << " xlock on " << *lock << " " << *obj << dendl; |
7c673cae | 4177 | MDSCacheObjectInfo i; |
11fdf7f2 | 4178 | obj->set_object_info(i); |
7c673cae | 4179 | if (i.ino) |
11fdf7f2 | 4180 | rejoin->add_inode_xlock(vinodeno_t(i.ino, i.snapid), lock->get_type(), |
7c673cae FG |
4181 | mdr->reqid, mdr->attempt); |
4182 | else | |
4183 | rejoin->add_dentry_xlock(i.dirfrag, i.dname, i.snapid, | |
4184 | mdr->reqid, mdr->attempt); | |
11fdf7f2 TL |
4185 | } else if (q.is_remote_wrlock()) { |
4186 | mds_rank_t who = q.wrlock_target; | |
4187 | if (rejoins.count(who) == 0) continue; | |
9f95a23c | 4188 | const auto& rejoin = rejoins[who]; |
7c673cae | 4189 | |
11fdf7f2 TL |
4190 | dout(15) << " " << *mdr << " wrlock on " << *lock << " " << *obj << dendl; |
4191 | MDSCacheObjectInfo i; | |
4192 | obj->set_object_info(i); | |
4193 | ceph_assert(i.ino); | |
4194 | rejoin->add_inode_wrlock(vinodeno_t(i.ino, i.snapid), lock->get_type(), | |
4195 | mdr->reqid, mdr->attempt); | |
4196 | } | |
7c673cae FG |
4197 | } |
4198 | } | |
4199 | } | |
4200 | ||
4201 | // send the messages | |
11fdf7f2 TL |
4202 | for (auto &p : rejoins) { |
4203 | ceph_assert(rejoin_sent.count(p.first) == 0); | |
4204 | ceph_assert(rejoin_ack_gather.count(p.first) == 0); | |
4205 | rejoin_sent.insert(p.first); | |
4206 | rejoin_ack_gather.insert(p.first); | |
4207 | mds->send_message_mds(p.second, p.first); | |
7c673cae FG |
4208 | } |
4209 | rejoin_ack_gather.insert(mds->get_nodeid()); // we need to complete rejoin_gather_finish, too | |
4210 | rejoins_pending = false; | |
4211 | ||
4212 | // nothing? | |
28e407b8 | 4213 | if (mds->is_rejoin() && rejoin_gather.empty()) { |
7c673cae FG |
4214 | dout(10) << "nothing to rejoin" << dendl; |
4215 | rejoin_gather_finish(); | |
4216 | } | |
4217 | } | |
4218 | ||
4219 | ||
4220 | /** | |
4221 | * rejoin_walk - build rejoin declarations for a subtree | |
4222 | * | |
4223 | * @param dir subtree root | |
4224 | * @param rejoin rejoin message | |
4225 | * | |
4226 | * from a rejoining node: | |
4227 | * weak dirfrag | |
4228 | * weak dentries (w/ connectivity) | |
4229 | * | |
4230 | * from a surviving node: | |
4231 | * strong dirfrag | |
4232 | * strong dentries (no connectivity!) | |
4233 | * strong inodes | |
4234 | */ | |
9f95a23c | 4235 | void MDCache::rejoin_walk(CDir *dir, const ref_t<MMDSCacheRejoin> &rejoin) |
7c673cae FG |
4236 | { |
4237 | dout(10) << "rejoin_walk " << *dir << dendl; | |
4238 | ||
9f95a23c | 4239 | std::vector<CDir*> nested; // finish this dir, then do nested items |
7c673cae FG |
4240 | |
4241 | if (mds->is_rejoin()) { | |
4242 | // WEAK | |
4243 | rejoin->add_weak_dirfrag(dir->dirfrag()); | |
94b18763 FG |
4244 | for (auto &p : dir->items) { |
4245 | CDentry *dn = p.second; | |
11fdf7f2 | 4246 | ceph_assert(dn->last == CEPH_NOSNAP); |
7c673cae FG |
4247 | CDentry::linkage_t *dnl = dn->get_linkage(); |
4248 | dout(15) << " add_weak_primary_dentry " << *dn << dendl; | |
11fdf7f2 | 4249 | ceph_assert(dnl->is_primary()); |
7c673cae | 4250 | CInode *in = dnl->get_inode(); |
11fdf7f2 | 4251 | ceph_assert(dnl->get_inode()->is_dir()); |
94b18763 | 4252 | rejoin->add_weak_primary_dentry(dir->ino(), dn->get_name(), dn->first, dn->last, in->ino()); |
9f95a23c TL |
4253 | { |
4254 | auto&& dirs = in->get_nested_dirfrags(); | |
4255 | nested.insert(std::end(nested), std::begin(dirs), std::end(dirs)); | |
4256 | } | |
7c673cae FG |
4257 | if (in->is_dirty_scattered()) { |
4258 | dout(10) << " sending scatterlock state on " << *in << dendl; | |
4259 | rejoin->add_scatterlock_state(in); | |
4260 | } | |
4261 | } | |
4262 | } else { | |
4263 | // STRONG | |
4264 | dout(15) << " add_strong_dirfrag " << *dir << dendl; | |
4265 | rejoin->add_strong_dirfrag(dir->dirfrag(), dir->get_replica_nonce(), dir->get_dir_rep()); | |
4266 | dir->state_set(CDir::STATE_REJOINING); | |
4267 | ||
11fdf7f2 | 4268 | for (auto it = dir->items.begin(); it != dir->items.end(); ) { |
94b18763 | 4269 | CDentry *dn = it->second; |
11fdf7f2 TL |
4270 | ++it; |
4271 | dn->state_set(CDentry::STATE_REJOINING); | |
7c673cae | 4272 | CDentry::linkage_t *dnl = dn->get_linkage(); |
11fdf7f2 TL |
4273 | CInode *in = dnl->is_primary() ? dnl->get_inode() : NULL; |
4274 | ||
4275 | // trim snap dentries. because they may have been pruned by | |
4276 | // their auth mds (snap deleted) | |
4277 | if (dn->last != CEPH_NOSNAP) { | |
4278 | if (in && !in->remote_parents.empty()) { | |
4279 | // unlink any stale remote snap dentry. | |
4280 | for (auto it2 = in->remote_parents.begin(); it2 != in->remote_parents.end(); ) { | |
4281 | CDentry *remote_dn = *it2; | |
4282 | ++it2; | |
4283 | ceph_assert(remote_dn->last != CEPH_NOSNAP); | |
4284 | remote_dn->unlink_remote(remote_dn->get_linkage()); | |
4285 | } | |
4286 | } | |
4287 | if (dn->lru_is_expireable()) { | |
4288 | if (!dnl->is_null()) | |
4289 | dir->unlink_inode(dn, false); | |
4290 | if (in) | |
4291 | remove_inode(in); | |
4292 | dir->remove_dentry(dn); | |
4293 | continue; | |
4294 | } else { | |
4295 | // Inventing null/remote dentry shouldn't cause problem | |
4296 | ceph_assert(!dnl->is_primary()); | |
4297 | } | |
4298 | } | |
4299 | ||
7c673cae | 4300 | dout(15) << " add_strong_dentry " << *dn << dendl; |
f67539c2 TL |
4301 | rejoin->add_strong_dentry(dir->dirfrag(), dn->get_name(), dn->get_alternate_name(), |
4302 | dn->first, dn->last, | |
7c673cae FG |
4303 | dnl->is_primary() ? dnl->get_inode()->ino():inodeno_t(0), |
4304 | dnl->is_remote() ? dnl->get_remote_ino():inodeno_t(0), | |
4305 | dnl->is_remote() ? dnl->get_remote_d_type():0, | |
4306 | dn->get_replica_nonce(), | |
4307 | dn->lock.get_state()); | |
4308 | dn->state_set(CDentry::STATE_REJOINING); | |
4309 | if (dnl->is_primary()) { | |
4310 | CInode *in = dnl->get_inode(); | |
4311 | dout(15) << " add_strong_inode " << *in << dendl; | |
4312 | rejoin->add_strong_inode(in->vino(), | |
4313 | in->get_replica_nonce(), | |
4314 | in->get_caps_wanted(), | |
4315 | in->filelock.get_state(), | |
4316 | in->nestlock.get_state(), | |
4317 | in->dirfragtreelock.get_state()); | |
4318 | in->state_set(CInode::STATE_REJOINING); | |
9f95a23c TL |
4319 | { |
4320 | auto&& dirs = in->get_nested_dirfrags(); | |
4321 | nested.insert(std::end(nested), std::begin(dirs), std::end(dirs)); | |
4322 | } | |
7c673cae FG |
4323 | if (in->is_dirty_scattered()) { |
4324 | dout(10) << " sending scatterlock state on " << *in << dendl; | |
4325 | rejoin->add_scatterlock_state(in); | |
4326 | } | |
4327 | } | |
4328 | } | |
4329 | } | |
4330 | ||
4331 | // recurse into nested dirs | |
9f95a23c TL |
4332 | for (const auto& dir : nested) { |
4333 | rejoin_walk(dir, rejoin); | |
4334 | } | |
7c673cae FG |
4335 | } |
4336 | ||
4337 | ||
4338 | /* | |
4339 | * i got a rejoin. | |
4340 | * - reply with the lockstate | |
4341 | * | |
4342 | * if i am active|stopping, | |
4343 | * - remove source from replica list for everything not referenced here. | |
7c673cae | 4344 | */ |
9f95a23c | 4345 | void MDCache::handle_cache_rejoin(const cref_t<MMDSCacheRejoin> &m) |
7c673cae FG |
4346 | { |
4347 | dout(7) << "handle_cache_rejoin " << *m << " from " << m->get_source() | |
4348 | << " (" << m->get_payload().length() << " bytes)" | |
4349 | << dendl; | |
4350 | ||
4351 | switch (m->op) { | |
4352 | case MMDSCacheRejoin::OP_WEAK: | |
4353 | handle_cache_rejoin_weak(m); | |
4354 | break; | |
4355 | case MMDSCacheRejoin::OP_STRONG: | |
4356 | handle_cache_rejoin_strong(m); | |
4357 | break; | |
4358 | case MMDSCacheRejoin::OP_ACK: | |
4359 | handle_cache_rejoin_ack(m); | |
4360 | break; | |
4361 | ||
4362 | default: | |
4363 | ceph_abort(); | |
4364 | } | |
7c673cae FG |
4365 | } |
4366 | ||
4367 | ||
4368 | /* | |
4369 | * handle_cache_rejoin_weak | |
4370 | * | |
4371 | * the sender | |
4372 | * - is recovering from their journal. | |
4373 | * - may have incorrect (out of date) inode contents | |
4374 | * - will include weak dirfrag if sender is dirfrag auth and parent inode auth is recipient | |
4375 | * | |
4376 | * if the sender didn't trim_non_auth(), they | |
4377 | * - may have incorrect (out of date) dentry/inode linkage | |
4378 | * - may have deleted/purged inodes | |
4379 | * and i may have to go to disk to get accurate inode contents. yuck. | |
7c673cae | 4380 | */ |
9f95a23c | 4381 | void MDCache::handle_cache_rejoin_weak(const cref_t<MMDSCacheRejoin> &weak) |
7c673cae FG |
4382 | { |
4383 | mds_rank_t from = mds_rank_t(weak->get_source().num()); | |
4384 | ||
4385 | // possible response(s) | |
9f95a23c | 4386 | ref_t<MMDSCacheRejoin> ack; // if survivor |
7c673cae FG |
4387 | set<vinodeno_t> acked_inodes; // if survivor |
4388 | set<SimpleLock *> gather_locks; // if survivor | |
4389 | bool survivor = false; // am i a survivor? | |
4390 | ||
4391 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
4392 | survivor = true; | |
4393 | dout(10) << "i am a surivivor, and will ack immediately" << dendl; | |
9f95a23c | 4394 | ack = make_message<MMDSCacheRejoin>(MMDSCacheRejoin::OP_ACK); |
7c673cae FG |
4395 | |
4396 | map<inodeno_t,map<client_t,Capability::Import> > imported_caps; | |
4397 | ||
4398 | // check cap exports | |
4399 | for (auto p = weak->cap_exports.begin(); p != weak->cap_exports.end(); ++p) { | |
4400 | CInode *in = get_inode(p->first); | |
11fdf7f2 | 4401 | ceph_assert(!in || in->is_auth()); |
7c673cae FG |
4402 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { |
4403 | dout(10) << " claiming cap import " << p->first << " client." << q->first << " on " << *in << dendl; | |
4404 | Capability *cap = rejoin_import_cap(in, q->first, q->second, from); | |
4405 | Capability::Import& im = imported_caps[p->first][q->first]; | |
4406 | if (cap) { | |
4407 | im.cap_id = cap->get_cap_id(); | |
4408 | im.issue_seq = cap->get_last_seq(); | |
4409 | im.mseq = cap->get_mseq(); | |
4410 | } else { | |
4411 | // all are zero | |
4412 | } | |
4413 | } | |
4414 | mds->locker->eval(in, CEPH_CAP_LOCKS, true); | |
4415 | } | |
4416 | ||
11fdf7f2 | 4417 | encode(imported_caps, ack->imported_caps); |
7c673cae | 4418 | } else { |
11fdf7f2 | 4419 | ceph_assert(mds->is_rejoin()); |
7c673cae FG |
4420 | |
4421 | // we may have already received a strong rejoin from the sender. | |
4422 | rejoin_scour_survivor_replicas(from, NULL, acked_inodes, gather_locks); | |
11fdf7f2 | 4423 | ceph_assert(gather_locks.empty()); |
7c673cae FG |
4424 | |
4425 | // check cap exports. | |
4426 | rejoin_client_map.insert(weak->client_map.begin(), weak->client_map.end()); | |
11fdf7f2 TL |
4427 | rejoin_client_metadata_map.insert(weak->client_metadata_map.begin(), |
4428 | weak->client_metadata_map.end()); | |
7c673cae FG |
4429 | |
4430 | for (auto p = weak->cap_exports.begin(); p != weak->cap_exports.end(); ++p) { | |
4431 | CInode *in = get_inode(p->first); | |
11fdf7f2 | 4432 | ceph_assert(!in || in->is_auth()); |
7c673cae FG |
4433 | // note |
4434 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
4435 | dout(10) << " claiming cap import " << p->first << " client." << q->first << dendl; | |
4436 | cap_imports[p->first][q->first][from] = q->second; | |
4437 | } | |
4438 | } | |
4439 | } | |
4440 | ||
4441 | // assimilate any potentially dirty scatterlock state | |
11fdf7f2 TL |
4442 | for (const auto &p : weak->inode_scatterlocks) { |
4443 | CInode *in = get_inode(p.first); | |
4444 | ceph_assert(in); | |
4445 | in->decode_lock_state(CEPH_LOCK_IFILE, p.second.file); | |
4446 | in->decode_lock_state(CEPH_LOCK_INEST, p.second.nest); | |
4447 | in->decode_lock_state(CEPH_LOCK_IDFT, p.second.dft); | |
7c673cae FG |
4448 | if (!survivor) |
4449 | rejoin_potential_updated_scatterlocks.insert(in); | |
4450 | } | |
4451 | ||
4452 | // recovering peer may send incorrect dirfrags here. we need to | |
4453 | // infer which dirfrag they meant. the ack will include a | |
4454 | // strong_dirfrag that will set them straight on the fragmentation. | |
4455 | ||
4456 | // walk weak map | |
4457 | set<CDir*> dirs_to_share; | |
11fdf7f2 TL |
4458 | for (const auto &p : weak->weak_dirfrags) { |
4459 | CInode *diri = get_inode(p.ino); | |
7c673cae | 4460 | if (!diri) |
11fdf7f2 TL |
4461 | dout(0) << " missing dir ino " << p.ino << dendl; |
4462 | ceph_assert(diri); | |
7c673cae | 4463 | |
11fdf7f2 TL |
4464 | frag_vec_t leaves; |
4465 | if (diri->dirfragtree.is_leaf(p.frag)) { | |
4466 | leaves.push_back(p.frag); | |
7c673cae | 4467 | } else { |
11fdf7f2 TL |
4468 | diri->dirfragtree.get_leaves_under(p.frag, leaves); |
4469 | if (leaves.empty()) | |
4470 | leaves.push_back(diri->dirfragtree[p.frag.value()]); | |
7c673cae | 4471 | } |
11fdf7f2 TL |
4472 | for (const auto& leaf : leaves) { |
4473 | CDir *dir = diri->get_dirfrag(leaf); | |
7c673cae | 4474 | if (!dir) { |
11fdf7f2 | 4475 | dout(0) << " missing dir for " << p.frag << " (which maps to " << leaf << ") on " << *diri << dendl; |
7c673cae FG |
4476 | continue; |
4477 | } | |
11fdf7f2 | 4478 | ceph_assert(dir); |
7c673cae | 4479 | if (dirs_to_share.count(dir)) { |
11fdf7f2 | 4480 | dout(10) << " already have " << p.frag << " -> " << leaf << " " << *dir << dendl; |
7c673cae FG |
4481 | } else { |
4482 | dirs_to_share.insert(dir); | |
4483 | unsigned nonce = dir->add_replica(from); | |
11fdf7f2 | 4484 | dout(10) << " have " << p.frag << " -> " << leaf << " " << *dir << dendl; |
7c673cae FG |
4485 | if (ack) { |
4486 | ack->add_strong_dirfrag(dir->dirfrag(), nonce, dir->dir_rep); | |
4487 | ack->add_dirfrag_base(dir); | |
4488 | } | |
4489 | } | |
4490 | } | |
4491 | } | |
4492 | ||
11fdf7f2 TL |
4493 | for (const auto &p : weak->weak) { |
4494 | CInode *diri = get_inode(p.first); | |
7c673cae | 4495 | if (!diri) |
11fdf7f2 TL |
4496 | dout(0) << " missing dir ino " << p.first << dendl; |
4497 | ceph_assert(diri); | |
7c673cae FG |
4498 | |
4499 | // weak dentries | |
4500 | CDir *dir = 0; | |
11fdf7f2 | 4501 | for (const auto &q : p.second) { |
7c673cae FG |
4502 | // locate proper dirfrag. |
4503 | // optimize for common case (one dirfrag) to avoid dirs_to_share set check | |
11fdf7f2 | 4504 | frag_t fg = diri->pick_dirfrag(q.first.name); |
7c673cae FG |
4505 | if (!dir || dir->get_frag() != fg) { |
4506 | dir = diri->get_dirfrag(fg); | |
4507 | if (!dir) | |
4508 | dout(0) << " missing dir frag " << fg << " on " << *diri << dendl; | |
11fdf7f2 TL |
4509 | ceph_assert(dir); |
4510 | ceph_assert(dirs_to_share.count(dir)); | |
7c673cae FG |
4511 | } |
4512 | ||
4513 | // and dentry | |
11fdf7f2 TL |
4514 | CDentry *dn = dir->lookup(q.first.name, q.first.snapid); |
4515 | ceph_assert(dn); | |
7c673cae | 4516 | CDentry::linkage_t *dnl = dn->get_linkage(); |
11fdf7f2 | 4517 | ceph_assert(dnl->is_primary()); |
7c673cae FG |
4518 | |
4519 | if (survivor && dn->is_replica(from)) | |
4520 | dentry_remove_replica(dn, from, gather_locks); | |
4521 | unsigned dnonce = dn->add_replica(from); | |
4522 | dout(10) << " have " << *dn << dendl; | |
4523 | if (ack) | |
f67539c2 TL |
4524 | ack->add_strong_dentry(dir->dirfrag(), dn->get_name(), dn->get_alternate_name(), |
4525 | dn->first, dn->last, | |
7c673cae FG |
4526 | dnl->get_inode()->ino(), inodeno_t(0), 0, |
4527 | dnonce, dn->lock.get_replica_state()); | |
4528 | ||
4529 | // inode | |
4530 | CInode *in = dnl->get_inode(); | |
11fdf7f2 | 4531 | ceph_assert(in); |
7c673cae FG |
4532 | |
4533 | if (survivor && in->is_replica(from)) | |
4534 | inode_remove_replica(in, from, true, gather_locks); | |
4535 | unsigned inonce = in->add_replica(from); | |
4536 | dout(10) << " have " << *in << dendl; | |
4537 | ||
4538 | // scatter the dirlock, just in case? | |
4539 | if (!survivor && in->is_dir() && in->has_subtree_root_dirfrag()) | |
4540 | in->filelock.set_state(LOCK_MIX); | |
4541 | ||
4542 | if (ack) { | |
4543 | acked_inodes.insert(in->vino()); | |
4544 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4545 | bufferlist bl; | |
4546 | in->_encode_locks_state_for_rejoin(bl, from); | |
4547 | ack->add_inode_locks(in, inonce, bl); | |
4548 | } | |
4549 | } | |
4550 | } | |
4551 | ||
4552 | // weak base inodes? (root, stray, etc.) | |
4553 | for (set<vinodeno_t>::iterator p = weak->weak_inodes.begin(); | |
4554 | p != weak->weak_inodes.end(); | |
4555 | ++p) { | |
4556 | CInode *in = get_inode(*p); | |
11fdf7f2 | 4557 | ceph_assert(in); // hmm fixme wrt stray? |
7c673cae FG |
4558 | if (survivor && in->is_replica(from)) |
4559 | inode_remove_replica(in, from, true, gather_locks); | |
4560 | unsigned inonce = in->add_replica(from); | |
4561 | dout(10) << " have base " << *in << dendl; | |
4562 | ||
4563 | if (ack) { | |
4564 | acked_inodes.insert(in->vino()); | |
4565 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4566 | bufferlist bl; | |
4567 | in->_encode_locks_state_for_rejoin(bl, from); | |
4568 | ack->add_inode_locks(in, inonce, bl); | |
4569 | } | |
4570 | } | |
4571 | ||
11fdf7f2 | 4572 | ceph_assert(rejoin_gather.count(from)); |
7c673cae FG |
4573 | rejoin_gather.erase(from); |
4574 | if (survivor) { | |
4575 | // survivor. do everything now. | |
11fdf7f2 TL |
4576 | for (const auto &p : weak->inode_scatterlocks) { |
4577 | CInode *in = get_inode(p.first); | |
4578 | ceph_assert(in); | |
7c673cae FG |
4579 | dout(10) << " including base inode (due to potential scatterlock update) " << *in << dendl; |
4580 | acked_inodes.insert(in->vino()); | |
4581 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4582 | } | |
4583 | ||
4584 | rejoin_scour_survivor_replicas(from, ack, acked_inodes, gather_locks); | |
4585 | mds->send_message(ack, weak->get_connection()); | |
4586 | ||
4587 | for (set<SimpleLock*>::iterator p = gather_locks.begin(); p != gather_locks.end(); ++p) { | |
4588 | if (!(*p)->is_stable()) | |
4589 | mds->locker->eval_gather(*p); | |
4590 | } | |
4591 | } else { | |
4592 | // done? | |
28e407b8 | 4593 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) { |
7c673cae FG |
4594 | rejoin_gather_finish(); |
4595 | } else { | |
4596 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" << dendl; | |
4597 | } | |
4598 | } | |
4599 | } | |
4600 | ||
7c673cae FG |
4601 | /* |
4602 | * rejoin_scour_survivor_replica - remove source from replica list on unmentioned objects | |
4603 | * | |
4604 | * all validated replicas are acked with a strong nonce, etc. if that isn't in the | |
4605 | * ack, the replica dne, and we can remove it from our replica maps. | |
4606 | */ | |
9f95a23c | 4607 | void MDCache::rejoin_scour_survivor_replicas(mds_rank_t from, const cref_t<MMDSCacheRejoin> &ack, |
7c673cae FG |
4608 | set<vinodeno_t>& acked_inodes, |
4609 | set<SimpleLock *>& gather_locks) | |
4610 | { | |
4611 | dout(10) << "rejoin_scour_survivor_replicas from mds." << from << dendl; | |
4612 | ||
b32b8144 | 4613 | auto scour_func = [this, from, ack, &acked_inodes, &gather_locks] (CInode *in) { |
7c673cae FG |
4614 | // inode? |
4615 | if (in->is_auth() && | |
4616 | in->is_replica(from) && | |
b32b8144 | 4617 | (ack == NULL || acked_inodes.count(in->vino()) == 0)) { |
7c673cae FG |
4618 | inode_remove_replica(in, from, false, gather_locks); |
4619 | dout(10) << " rem " << *in << dendl; | |
4620 | } | |
4621 | ||
b32b8144 FG |
4622 | if (!in->is_dir()) |
4623 | return; | |
7c673cae | 4624 | |
9f95a23c TL |
4625 | const auto&& dfs = in->get_dirfrags(); |
4626 | for (const auto& dir : dfs) { | |
181888fb FG |
4627 | if (!dir->is_auth()) |
4628 | continue; | |
7c673cae | 4629 | |
181888fb | 4630 | if (dir->is_replica(from) && |
7c673cae FG |
4631 | (ack == NULL || ack->strong_dirfrags.count(dir->dirfrag()) == 0)) { |
4632 | dir->remove_replica(from); | |
4633 | dout(10) << " rem " << *dir << dendl; | |
4634 | } | |
4635 | ||
4636 | // dentries | |
94b18763 FG |
4637 | for (auto &p : dir->items) { |
4638 | CDentry *dn = p.second; | |
7c673cae | 4639 | |
11fdf7f2 TL |
4640 | if (dn->is_replica(from)) { |
4641 | if (ack) { | |
4642 | const auto it = ack->strong_dentries.find(dir->dirfrag()); | |
4643 | if (it != ack->strong_dentries.end() && it->second.count(string_snap_t(dn->get_name(), dn->last)) > 0) { | |
4644 | continue; | |
4645 | } | |
4646 | } | |
7c673cae FG |
4647 | dentry_remove_replica(dn, from, gather_locks); |
4648 | dout(10) << " rem " << *dn << dendl; | |
4649 | } | |
4650 | } | |
4651 | } | |
b32b8144 FG |
4652 | }; |
4653 | ||
94b18763 | 4654 | for (auto &p : inode_map) |
b32b8144 | 4655 | scour_func(p.second); |
94b18763 | 4656 | for (auto &p : snap_inode_map) |
b32b8144 | 4657 | scour_func(p.second); |
7c673cae FG |
4658 | } |
4659 | ||
4660 | ||
4661 | CInode *MDCache::rejoin_invent_inode(inodeno_t ino, snapid_t last) | |
4662 | { | |
f67539c2 TL |
4663 | CInode *in = new CInode(this, true, 2, last); |
4664 | in->_get_inode()->ino = ino; | |
7c673cae FG |
4665 | in->state_set(CInode::STATE_REJOINUNDEF); |
4666 | add_inode(in); | |
4667 | rejoin_undef_inodes.insert(in); | |
4668 | dout(10) << " invented " << *in << dendl; | |
4669 | return in; | |
4670 | } | |
4671 | ||
4672 | CDir *MDCache::rejoin_invent_dirfrag(dirfrag_t df) | |
4673 | { | |
4674 | CInode *in = get_inode(df.ino); | |
4675 | if (!in) | |
4676 | in = rejoin_invent_inode(df.ino, CEPH_NOSNAP); | |
4677 | if (!in->is_dir()) { | |
11fdf7f2 | 4678 | ceph_assert(in->state_test(CInode::STATE_REJOINUNDEF)); |
f67539c2 TL |
4679 | in->_get_inode()->mode = S_IFDIR; |
4680 | in->_get_inode()->dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; | |
7c673cae FG |
4681 | } |
4682 | CDir *dir = in->get_or_open_dirfrag(this, df.frag); | |
4683 | dir->state_set(CDir::STATE_REJOINUNDEF); | |
4684 | rejoin_undef_dirfrags.insert(dir); | |
4685 | dout(10) << " invented " << *dir << dendl; | |
4686 | return dir; | |
4687 | } | |
4688 | ||
9f95a23c | 4689 | void MDCache::handle_cache_rejoin_strong(const cref_t<MMDSCacheRejoin> &strong) |
7c673cae FG |
4690 | { |
4691 | mds_rank_t from = mds_rank_t(strong->get_source().num()); | |
4692 | ||
4693 | // only a recovering node will get a strong rejoin. | |
a8e16298 TL |
4694 | if (!mds->is_rejoin()) { |
4695 | if (mds->get_want_state() == MDSMap::STATE_REJOIN) { | |
4696 | mds->wait_for_rejoin(new C_MDS_RetryMessage(mds, strong)); | |
4697 | return; | |
4698 | } | |
11fdf7f2 | 4699 | ceph_abort_msg("got unexpected rejoin message during recovery"); |
a8e16298 | 4700 | } |
7c673cae FG |
4701 | |
4702 | // assimilate any potentially dirty scatterlock state | |
11fdf7f2 TL |
4703 | for (const auto &p : strong->inode_scatterlocks) { |
4704 | CInode *in = get_inode(p.first); | |
4705 | ceph_assert(in); | |
4706 | in->decode_lock_state(CEPH_LOCK_IFILE, p.second.file); | |
4707 | in->decode_lock_state(CEPH_LOCK_INEST, p.second.nest); | |
4708 | in->decode_lock_state(CEPH_LOCK_IDFT, p.second.dft); | |
7c673cae FG |
4709 | rejoin_potential_updated_scatterlocks.insert(in); |
4710 | } | |
4711 | ||
4712 | rejoin_unlinked_inodes[from].clear(); | |
4713 | ||
4714 | // surviving peer may send incorrect dirfrag here (maybe they didn't | |
4715 | // get the fragment notify, or maybe we rolled back?). we need to | |
4716 | // infer the right frag and get them with the program. somehow. | |
4717 | // we don't normally send ACK.. so we'll need to bundle this with | |
4718 | // MISSING or something. | |
4719 | ||
4720 | // strong dirfrags/dentries. | |
4721 | // also process auth_pins, xlocks. | |
11fdf7f2 TL |
4722 | for (const auto &p : strong->strong_dirfrags) { |
4723 | auto& dirfrag = p.first; | |
4724 | CInode *diri = get_inode(dirfrag.ino); | |
7c673cae | 4725 | if (!diri) |
11fdf7f2 TL |
4726 | diri = rejoin_invent_inode(dirfrag.ino, CEPH_NOSNAP); |
4727 | CDir *dir = diri->get_dirfrag(dirfrag.frag); | |
7c673cae FG |
4728 | bool refragged = false; |
4729 | if (dir) { | |
4730 | dout(10) << " have " << *dir << dendl; | |
4731 | } else { | |
4732 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) | |
4733 | dir = rejoin_invent_dirfrag(dirfrag_t(diri->ino(), frag_t())); | |
11fdf7f2 TL |
4734 | else if (diri->dirfragtree.is_leaf(dirfrag.frag)) |
4735 | dir = rejoin_invent_dirfrag(dirfrag); | |
7c673cae FG |
4736 | } |
4737 | if (dir) { | |
11fdf7f2 TL |
4738 | dir->add_replica(from, p.second.nonce); |
4739 | dir->dir_rep = p.second.dir_rep; | |
7c673cae | 4740 | } else { |
11fdf7f2 TL |
4741 | dout(10) << " frag " << dirfrag << " doesn't match dirfragtree " << *diri << dendl; |
4742 | frag_vec_t leaves; | |
4743 | diri->dirfragtree.get_leaves_under(dirfrag.frag, leaves); | |
4744 | if (leaves.empty()) | |
4745 | leaves.push_back(diri->dirfragtree[dirfrag.frag.value()]); | |
4746 | dout(10) << " maps to frag(s) " << leaves << dendl; | |
4747 | for (const auto& leaf : leaves) { | |
4748 | CDir *dir = diri->get_dirfrag(leaf); | |
7c673cae | 4749 | if (!dir) |
11fdf7f2 | 4750 | dir = rejoin_invent_dirfrag(dirfrag_t(diri->ino(), leaf)); |
7c673cae FG |
4751 | else |
4752 | dout(10) << " have(approx) " << *dir << dendl; | |
11fdf7f2 TL |
4753 | dir->add_replica(from, p.second.nonce); |
4754 | dir->dir_rep = p.second.dir_rep; | |
7c673cae FG |
4755 | } |
4756 | refragged = true; | |
4757 | } | |
4758 | ||
11fdf7f2 TL |
4759 | const auto it = strong->strong_dentries.find(dirfrag); |
4760 | if (it != strong->strong_dentries.end()) { | |
9f95a23c | 4761 | const auto& dmap = it->second; |
11fdf7f2 TL |
4762 | for (const auto &q : dmap) { |
4763 | const string_snap_t& ss = q.first; | |
4764 | const MMDSCacheRejoin::dn_strong& d = q.second; | |
4765 | CDentry *dn; | |
4766 | if (!refragged) | |
4767 | dn = dir->lookup(ss.name, ss.snapid); | |
4768 | else { | |
4769 | frag_t fg = diri->pick_dirfrag(ss.name); | |
4770 | dir = diri->get_dirfrag(fg); | |
4771 | ceph_assert(dir); | |
4772 | dn = dir->lookup(ss.name, ss.snapid); | |
4773 | } | |
4774 | if (!dn) { | |
4775 | if (d.is_remote()) { | |
f67539c2 | 4776 | dn = dir->add_remote_dentry(ss.name, d.remote_ino, d.remote_d_type, mempool::mds_co::string(d.alternate_name), d.first, ss.snapid); |
11fdf7f2 TL |
4777 | } else if (d.is_null()) { |
4778 | dn = dir->add_null_dentry(ss.name, d.first, ss.snapid); | |
4779 | } else { | |
4780 | CInode *in = get_inode(d.ino, ss.snapid); | |
4781 | if (!in) in = rejoin_invent_inode(d.ino, ss.snapid); | |
f67539c2 | 4782 | dn = dir->add_primary_dentry(ss.name, in, mempool::mds_co::string(d.alternate_name), d.first, ss.snapid); |
11fdf7f2 TL |
4783 | } |
4784 | dout(10) << " invented " << *dn << dendl; | |
4785 | } | |
4786 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4787 | ||
4788 | // dn auth_pin? | |
4789 | const auto pinned_it = strong->authpinned_dentries.find(dirfrag); | |
4790 | if (pinned_it != strong->authpinned_dentries.end()) { | |
f67539c2 TL |
4791 | const auto peer_reqid_it = pinned_it->second.find(ss); |
4792 | if (peer_reqid_it != pinned_it->second.end()) { | |
4793 | for (const auto &r : peer_reqid_it->second) { | |
11fdf7f2 TL |
4794 | dout(10) << " dn authpin by " << r << " on " << *dn << dendl; |
4795 | ||
f67539c2 | 4796 | // get/create peer mdrequest |
11fdf7f2 TL |
4797 | MDRequestRef mdr; |
4798 | if (have_request(r.reqid)) | |
4799 | mdr = request_get(r.reqid); | |
4800 | else | |
f67539c2 | 4801 | mdr = request_start_peer(r.reqid, r.attempt, strong); |
11fdf7f2 TL |
4802 | mdr->auth_pin(dn); |
4803 | } | |
4804 | } | |
7c673cae | 4805 | } |
7c673cae | 4806 | |
11fdf7f2 TL |
4807 | // dn xlock? |
4808 | const auto xlocked_it = strong->xlocked_dentries.find(dirfrag); | |
4809 | if (xlocked_it != strong->xlocked_dentries.end()) { | |
4810 | const auto ss_req_it = xlocked_it->second.find(ss); | |
4811 | if (ss_req_it != xlocked_it->second.end()) { | |
f67539c2 | 4812 | const MMDSCacheRejoin::peer_reqid& r = ss_req_it->second; |
11fdf7f2 TL |
4813 | dout(10) << " dn xlock by " << r << " on " << *dn << dendl; |
4814 | MDRequestRef mdr = request_get(r.reqid); // should have this from auth_pin above. | |
4815 | ceph_assert(mdr->is_auth_pinned(dn)); | |
4816 | if (!mdr->is_xlocked(&dn->versionlock)) { | |
4817 | ceph_assert(dn->versionlock.can_xlock_local()); | |
4818 | dn->versionlock.get_xlock(mdr, mdr->get_client()); | |
9f95a23c | 4819 | mdr->emplace_lock(&dn->versionlock, MutationImpl::LockOp::XLOCK); |
11fdf7f2 TL |
4820 | } |
4821 | if (dn->lock.is_stable()) | |
4822 | dn->auth_pin(&dn->lock); | |
4823 | dn->lock.set_state(LOCK_XLOCK); | |
4824 | dn->lock.get_xlock(mdr, mdr->get_client()); | |
9f95a23c | 4825 | mdr->emplace_lock(&dn->lock, MutationImpl::LockOp::XLOCK); |
11fdf7f2 TL |
4826 | } |
4827 | } | |
7c673cae | 4828 | |
11fdf7f2 TL |
4829 | dn->add_replica(from, d.nonce); |
4830 | dout(10) << " have " << *dn << dendl; | |
4831 | ||
4832 | if (dnl->is_primary()) { | |
4833 | if (d.is_primary()) { | |
4834 | if (vinodeno_t(d.ino, ss.snapid) != dnl->get_inode()->vino()) { | |
4835 | // the survivor missed MDentryUnlink+MDentryLink messages ? | |
4836 | ceph_assert(strong->strong_inodes.count(dnl->get_inode()->vino()) == 0); | |
4837 | CInode *in = get_inode(d.ino, ss.snapid); | |
4838 | ceph_assert(in); | |
4839 | ceph_assert(in->get_parent_dn()); | |
4840 | rejoin_unlinked_inodes[from].insert(in); | |
4841 | dout(7) << " sender has primary dentry but wrong inode" << dendl; | |
4842 | } | |
4843 | } else { | |
4844 | // the survivor missed MDentryLink message ? | |
4845 | ceph_assert(strong->strong_inodes.count(dnl->get_inode()->vino()) == 0); | |
4846 | dout(7) << " sender doesn't have primay dentry" << dendl; | |
4847 | } | |
4848 | } else { | |
4849 | if (d.is_primary()) { | |
4850 | // the survivor missed MDentryUnlink message ? | |
4851 | CInode *in = get_inode(d.ino, ss.snapid); | |
4852 | ceph_assert(in); | |
4853 | ceph_assert(in->get_parent_dn()); | |
7c673cae | 4854 | rejoin_unlinked_inodes[from].insert(in); |
11fdf7f2 | 4855 | dout(7) << " sender has primary dentry but we don't" << dendl; |
7c673cae | 4856 | } |
11fdf7f2 | 4857 | } |
7c673cae FG |
4858 | } |
4859 | } | |
4860 | } | |
4861 | ||
11fdf7f2 TL |
4862 | for (const auto &p : strong->strong_inodes) { |
4863 | CInode *in = get_inode(p.first); | |
4864 | ceph_assert(in); | |
4865 | in->add_replica(from, p.second.nonce); | |
7c673cae FG |
4866 | dout(10) << " have " << *in << dendl; |
4867 | ||
11fdf7f2 | 4868 | const MMDSCacheRejoin::inode_strong& is = p.second; |
7c673cae FG |
4869 | |
4870 | // caps_wanted | |
4871 | if (is.caps_wanted) { | |
11fdf7f2 | 4872 | in->set_mds_caps_wanted(from, is.caps_wanted); |
7c673cae FG |
4873 | dout(15) << " inode caps_wanted " << ccap_string(is.caps_wanted) |
4874 | << " on " << *in << dendl; | |
4875 | } | |
4876 | ||
4877 | // scatterlocks? | |
4878 | // infer state from replica state: | |
4879 | // * go to MIX if they might have wrlocks | |
4880 | // * go to LOCK if they are LOCK (just bc identify_files_to_recover might start twiddling filelock) | |
4881 | in->filelock.infer_state_from_strong_rejoin(is.filelock, !in->is_dir()); // maybe also go to LOCK | |
4882 | in->nestlock.infer_state_from_strong_rejoin(is.nestlock, false); | |
4883 | in->dirfragtreelock.infer_state_from_strong_rejoin(is.dftlock, false); | |
4884 | ||
4885 | // auth pin? | |
11fdf7f2 TL |
4886 | const auto authpinned_inodes_it = strong->authpinned_inodes.find(in->vino()); |
4887 | if (authpinned_inodes_it != strong->authpinned_inodes.end()) { | |
4888 | for (const auto& r : authpinned_inodes_it->second) { | |
4889 | dout(10) << " inode authpin by " << r << " on " << *in << dendl; | |
7c673cae | 4890 | |
f67539c2 | 4891 | // get/create peer mdrequest |
7c673cae | 4892 | MDRequestRef mdr; |
11fdf7f2 TL |
4893 | if (have_request(r.reqid)) |
4894 | mdr = request_get(r.reqid); | |
7c673cae | 4895 | else |
f67539c2 | 4896 | mdr = request_start_peer(r.reqid, r.attempt, strong); |
7c673cae | 4897 | if (strong->frozen_authpin_inodes.count(in->vino())) { |
11fdf7f2 | 4898 | ceph_assert(!in->get_num_auth_pins()); |
7c673cae FG |
4899 | mdr->freeze_auth_pin(in); |
4900 | } else { | |
11fdf7f2 | 4901 | ceph_assert(!in->is_frozen_auth_pin()); |
7c673cae FG |
4902 | } |
4903 | mdr->auth_pin(in); | |
4904 | } | |
4905 | } | |
4906 | // xlock(s)? | |
11fdf7f2 TL |
4907 | const auto xlocked_inodes_it = strong->xlocked_inodes.find(in->vino()); |
4908 | if (xlocked_inodes_it != strong->xlocked_inodes.end()) { | |
4909 | for (const auto &q : xlocked_inodes_it->second) { | |
4910 | SimpleLock *lock = in->get_lock(q.first); | |
4911 | dout(10) << " inode xlock by " << q.second << " on " << *lock << " on " << *in << dendl; | |
4912 | MDRequestRef mdr = request_get(q.second.reqid); // should have this from auth_pin above. | |
4913 | ceph_assert(mdr->is_auth_pinned(in)); | |
4914 | if (!mdr->is_xlocked(&in->versionlock)) { | |
4915 | ceph_assert(in->versionlock.can_xlock_local()); | |
7c673cae | 4916 | in->versionlock.get_xlock(mdr, mdr->get_client()); |
9f95a23c | 4917 | mdr->emplace_lock(&in->versionlock, MutationImpl::LockOp::XLOCK); |
7c673cae FG |
4918 | } |
4919 | if (lock->is_stable()) | |
4920 | in->auth_pin(lock); | |
4921 | lock->set_state(LOCK_XLOCK); | |
4922 | if (lock == &in->filelock) | |
4923 | in->loner_cap = -1; | |
4924 | lock->get_xlock(mdr, mdr->get_client()); | |
9f95a23c | 4925 | mdr->emplace_lock(lock, MutationImpl::LockOp::XLOCK); |
7c673cae FG |
4926 | } |
4927 | } | |
4928 | } | |
4929 | // wrlock(s)? | |
11fdf7f2 TL |
4930 | for (const auto &p : strong->wrlocked_inodes) { |
4931 | CInode *in = get_inode(p.first); | |
4932 | for (const auto &q : p.second) { | |
4933 | SimpleLock *lock = in->get_lock(q.first); | |
4934 | for (const auto &r : q.second) { | |
4935 | dout(10) << " inode wrlock by " << r << " on " << *lock << " on " << *in << dendl; | |
4936 | MDRequestRef mdr = request_get(r.reqid); // should have this from auth_pin above. | |
7c673cae | 4937 | if (in->is_auth()) |
11fdf7f2 | 4938 | ceph_assert(mdr->is_auth_pinned(in)); |
7c673cae FG |
4939 | lock->set_state(LOCK_MIX); |
4940 | if (lock == &in->filelock) | |
4941 | in->loner_cap = -1; | |
4942 | lock->get_wrlock(true); | |
9f95a23c | 4943 | mdr->emplace_lock(lock, MutationImpl::LockOp::WRLOCK); |
7c673cae FG |
4944 | } |
4945 | } | |
4946 | } | |
4947 | ||
4948 | // done? | |
11fdf7f2 | 4949 | ceph_assert(rejoin_gather.count(from)); |
7c673cae | 4950 | rejoin_gather.erase(from); |
28e407b8 | 4951 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) { |
7c673cae FG |
4952 | rejoin_gather_finish(); |
4953 | } else { | |
4954 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" << dendl; | |
4955 | } | |
4956 | } | |
4957 | ||
9f95a23c | 4958 | void MDCache::handle_cache_rejoin_ack(const cref_t<MMDSCacheRejoin> &ack) |
7c673cae FG |
4959 | { |
4960 | dout(7) << "handle_cache_rejoin_ack from " << ack->get_source() << dendl; | |
4961 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
4962 | ||
11fdf7f2 | 4963 | ceph_assert(mds->get_state() >= MDSMap::STATE_REJOIN); |
b32b8144 FG |
4964 | bool survivor = !mds->is_rejoin(); |
4965 | ||
7c673cae FG |
4966 | // for sending cache expire message |
4967 | set<CInode*> isolated_inodes; | |
4968 | set<CInode*> refragged_inodes; | |
11fdf7f2 | 4969 | list<pair<CInode*,int> > updated_realms; |
7c673cae FG |
4970 | |
4971 | // dirs | |
11fdf7f2 | 4972 | for (const auto &p : ack->strong_dirfrags) { |
7c673cae FG |
4973 | // we may have had incorrect dir fragmentation; refragment based |
4974 | // on what they auth tells us. | |
11fdf7f2 | 4975 | CDir *dir = get_dirfrag(p.first); |
7c673cae | 4976 | if (!dir) { |
11fdf7f2 | 4977 | dir = get_force_dirfrag(p.first, false); |
7c673cae FG |
4978 | if (dir) |
4979 | refragged_inodes.insert(dir->get_inode()); | |
4980 | } | |
4981 | if (!dir) { | |
11fdf7f2 | 4982 | CInode *diri = get_inode(p.first.ino); |
7c673cae FG |
4983 | if (!diri) { |
4984 | // barebones inode; the full inode loop below will clean up. | |
4985 | diri = new CInode(this, false); | |
f67539c2 TL |
4986 | auto _inode = diri->_get_inode(); |
4987 | _inode->ino = p.first.ino; | |
4988 | _inode->mode = S_IFDIR; | |
4989 | _inode->dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; | |
4990 | ||
7c673cae | 4991 | add_inode(diri); |
11fdf7f2 | 4992 | if (MDS_INO_MDSDIR(from) == p.first.ino) { |
7c673cae FG |
4993 | diri->inode_auth = mds_authority_t(from, CDIR_AUTH_UNKNOWN); |
4994 | dout(10) << " add inode " << *diri << dendl; | |
4995 | } else { | |
4996 | diri->inode_auth = CDIR_AUTH_DEFAULT; | |
4997 | isolated_inodes.insert(diri); | |
11fdf7f2 | 4998 | dout(10) << " unconnected dirfrag " << p.first << dendl; |
7c673cae FG |
4999 | } |
5000 | } | |
5001 | // barebones dirfrag; the full dirfrag loop below will clean up. | |
11fdf7f2 TL |
5002 | dir = diri->add_dirfrag(new CDir(diri, p.first.frag, this, false)); |
5003 | if (MDS_INO_MDSDIR(from) == p.first.ino || | |
7c673cae FG |
5004 | (dir->authority() != CDIR_AUTH_UNDEF && |
5005 | dir->authority().first != from)) | |
5006 | adjust_subtree_auth(dir, from); | |
5007 | dout(10) << " add dirfrag " << *dir << dendl; | |
5008 | } | |
5009 | ||
11fdf7f2 | 5010 | dir->set_replica_nonce(p.second.nonce); |
7c673cae FG |
5011 | dir->state_clear(CDir::STATE_REJOINING); |
5012 | dout(10) << " got " << *dir << dendl; | |
5013 | ||
5014 | // dentries | |
11fdf7f2 TL |
5015 | auto it = ack->strong_dentries.find(p.first); |
5016 | if (it != ack->strong_dentries.end()) { | |
5017 | for (const auto &q : it->second) { | |
5018 | CDentry *dn = dir->lookup(q.first.name, q.first.snapid); | |
5019 | if(!dn) | |
5020 | dn = dir->add_null_dentry(q.first.name, q.second.first, q.first.snapid); | |
5021 | ||
5022 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
5023 | ||
5024 | ceph_assert(dn->last == q.first.snapid); | |
5025 | if (dn->first != q.second.first) { | |
5026 | dout(10) << " adjust dn.first " << dn->first << " -> " << q.second.first << " on " << *dn << dendl; | |
5027 | dn->first = q.second.first; | |
5028 | } | |
7c673cae | 5029 | |
11fdf7f2 TL |
5030 | // may have bad linkage if we missed dentry link/unlink messages |
5031 | if (dnl->is_primary()) { | |
5032 | CInode *in = dnl->get_inode(); | |
5033 | if (!q.second.is_primary() || | |
5034 | vinodeno_t(q.second.ino, q.first.snapid) != in->vino()) { | |
5035 | dout(10) << " had bad linkage for " << *dn << ", unlinking " << *in << dendl; | |
5036 | dir->unlink_inode(dn); | |
5037 | } | |
5038 | } else if (dnl->is_remote()) { | |
5039 | if (!q.second.is_remote() || | |
5040 | q.second.remote_ino != dnl->get_remote_ino() || | |
5041 | q.second.remote_d_type != dnl->get_remote_d_type()) { | |
5042 | dout(10) << " had bad linkage for " << *dn << dendl; | |
5043 | dir->unlink_inode(dn); | |
5044 | } | |
5045 | } else { | |
5046 | if (!q.second.is_null()) | |
5047 | dout(10) << " had bad linkage for " << *dn << dendl; | |
5048 | } | |
7c673cae | 5049 | |
f67539c2 TL |
5050 | // hmm, did we have the proper linkage here? |
5051 | if (dnl->is_null() && !q.second.is_null()) { | |
11fdf7f2 TL |
5052 | if (q.second.is_remote()) { |
5053 | dn->dir->link_remote_inode(dn, q.second.remote_ino, q.second.remote_d_type); | |
5054 | } else { | |
5055 | CInode *in = get_inode(q.second.ino, q.first.snapid); | |
5056 | if (!in) { | |
5057 | // barebones inode; assume it's dir, the full inode loop below will clean up. | |
5058 | in = new CInode(this, false, q.second.first, q.first.snapid); | |
f67539c2 TL |
5059 | auto _inode = in->_get_inode(); |
5060 | _inode->ino = q.second.ino; | |
5061 | _inode->mode = S_IFDIR; | |
5062 | _inode->dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; | |
11fdf7f2 TL |
5063 | add_inode(in); |
5064 | dout(10) << " add inode " << *in << dendl; | |
5065 | } else if (in->get_parent_dn()) { | |
5066 | dout(10) << " had bad linkage for " << *(in->get_parent_dn()) | |
5067 | << ", unlinking " << *in << dendl; | |
5068 | in->get_parent_dir()->unlink_inode(in->get_parent_dn()); | |
5069 | } | |
5070 | dn->dir->link_primary_inode(dn, in); | |
5071 | isolated_inodes.erase(in); | |
7c673cae | 5072 | } |
f67539c2 | 5073 | } |
7c673cae | 5074 | |
11fdf7f2 TL |
5075 | dn->set_replica_nonce(q.second.nonce); |
5076 | dn->lock.set_state_rejoin(q.second.lock, rejoin_waiters, survivor); | |
5077 | dn->state_clear(CDentry::STATE_REJOINING); | |
5078 | dout(10) << " got " << *dn << dendl; | |
5079 | } | |
7c673cae FG |
5080 | } |
5081 | } | |
5082 | ||
9f95a23c TL |
5083 | for (const auto& in : refragged_inodes) { |
5084 | auto&& ls = in->get_nested_dirfrags(); | |
5085 | for (const auto& dir : ls) { | |
5086 | if (dir->is_auth() || ack->strong_dirfrags.count(dir->dirfrag())) | |
7c673cae | 5087 | continue; |
9f95a23c TL |
5088 | ceph_assert(dir->get_num_any() == 0); |
5089 | in->close_dirfrag(dir->get_frag()); | |
7c673cae FG |
5090 | } |
5091 | } | |
5092 | ||
5093 | // full dirfrags | |
11fdf7f2 TL |
5094 | for (const auto &p : ack->dirfrag_bases) { |
5095 | CDir *dir = get_dirfrag(p.first); | |
5096 | ceph_assert(dir); | |
5097 | auto q = p.second.cbegin(); | |
7c673cae FG |
5098 | dir->_decode_base(q); |
5099 | dout(10) << " got dir replica " << *dir << dendl; | |
5100 | } | |
5101 | ||
5102 | // full inodes | |
11fdf7f2 | 5103 | auto p = ack->inode_base.cbegin(); |
7c673cae FG |
5104 | while (!p.end()) { |
5105 | inodeno_t ino; | |
5106 | snapid_t last; | |
5107 | bufferlist basebl; | |
11fdf7f2 TL |
5108 | decode(ino, p); |
5109 | decode(last, p); | |
5110 | decode(basebl, p); | |
7c673cae | 5111 | CInode *in = get_inode(ino, last); |
11fdf7f2 TL |
5112 | ceph_assert(in); |
5113 | auto q = basebl.cbegin(); | |
5114 | snapid_t sseq = 0; | |
5115 | if (in->snaprealm) | |
5116 | sseq = in->snaprealm->srnode.seq; | |
7c673cae | 5117 | in->_decode_base(q); |
11fdf7f2 TL |
5118 | if (in->snaprealm && in->snaprealm->srnode.seq != sseq) { |
5119 | int snap_op = sseq > 0 ? CEPH_SNAP_OP_UPDATE : CEPH_SNAP_OP_SPLIT; | |
5120 | updated_realms.push_back(pair<CInode*,int>(in, snap_op)); | |
5121 | } | |
7c673cae FG |
5122 | dout(10) << " got inode base " << *in << dendl; |
5123 | } | |
5124 | ||
5125 | // inodes | |
11fdf7f2 | 5126 | p = ack->inode_locks.cbegin(); |
7c673cae FG |
5127 | //dout(10) << "inode_locks len " << ack->inode_locks.length() << " is " << ack->inode_locks << dendl; |
5128 | while (!p.end()) { | |
5129 | inodeno_t ino; | |
5130 | snapid_t last; | |
5131 | __u32 nonce; | |
5132 | bufferlist lockbl; | |
11fdf7f2 TL |
5133 | decode(ino, p); |
5134 | decode(last, p); | |
5135 | decode(nonce, p); | |
5136 | decode(lockbl, p); | |
7c673cae FG |
5137 | |
5138 | CInode *in = get_inode(ino, last); | |
11fdf7f2 | 5139 | ceph_assert(in); |
7c673cae | 5140 | in->set_replica_nonce(nonce); |
11fdf7f2 | 5141 | auto q = lockbl.cbegin(); |
b32b8144 | 5142 | in->_decode_locks_rejoin(q, rejoin_waiters, rejoin_eval_locks, survivor); |
7c673cae FG |
5143 | in->state_clear(CInode::STATE_REJOINING); |
5144 | dout(10) << " got inode locks " << *in << dendl; | |
5145 | } | |
5146 | ||
5147 | // FIXME: This can happen if entire subtree, together with the inode subtree root | |
5148 | // belongs to, were trimmed between sending cache rejoin and receiving rejoin ack. | |
11fdf7f2 | 5149 | ceph_assert(isolated_inodes.empty()); |
7c673cae FG |
5150 | |
5151 | map<inodeno_t,map<client_t,Capability::Import> > peer_imported; | |
11fdf7f2 TL |
5152 | auto bp = ack->imported_caps.cbegin(); |
5153 | decode(peer_imported, bp); | |
7c673cae FG |
5154 | |
5155 | for (map<inodeno_t,map<client_t,Capability::Import> >::iterator p = peer_imported.begin(); | |
5156 | p != peer_imported.end(); | |
5157 | ++p) { | |
28e407b8 | 5158 | auto& ex = cap_exports.at(p->first); |
11fdf7f2 | 5159 | ceph_assert(ex.first == from); |
7c673cae FG |
5160 | for (map<client_t,Capability::Import>::iterator q = p->second.begin(); |
5161 | q != p->second.end(); | |
5162 | ++q) { | |
28e407b8 | 5163 | auto r = ex.second.find(q->first); |
11fdf7f2 | 5164 | ceph_assert(r != ex.second.end()); |
7c673cae FG |
5165 | |
5166 | dout(10) << " exporting caps for client." << q->first << " ino " << p->first << dendl; | |
5167 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
28e407b8 AA |
5168 | if (!session) { |
5169 | dout(10) << " no session for client." << p->first << dendl; | |
5170 | ex.second.erase(r); | |
5171 | continue; | |
5172 | } | |
7c673cae FG |
5173 | |
5174 | // mark client caps stale. | |
9f95a23c | 5175 | auto m = make_message<MClientCaps>(CEPH_CAP_OP_EXPORT, p->first, 0, |
28e407b8 | 5176 | r->second.capinfo.cap_id, 0, |
7c673cae FG |
5177 | mds->get_osd_epoch_barrier()); |
5178 | m->set_cap_peer(q->second.cap_id, q->second.issue_seq, q->second.mseq, | |
5179 | (q->second.cap_id > 0 ? from : -1), 0); | |
5180 | mds->send_message_client_counted(m, session); | |
5181 | ||
28e407b8 | 5182 | ex.second.erase(r); |
7c673cae | 5183 | } |
11fdf7f2 TL |
5184 | ceph_assert(ex.second.empty()); |
5185 | } | |
5186 | ||
5187 | for (auto p : updated_realms) { | |
5188 | CInode *in = p.first; | |
5189 | bool notify_clients; | |
5190 | if (mds->is_rejoin()) { | |
5191 | if (!rejoin_pending_snaprealms.count(in)) { | |
5192 | in->get(CInode::PIN_OPENINGSNAPPARENTS); | |
5193 | rejoin_pending_snaprealms.insert(in); | |
5194 | } | |
5195 | notify_clients = false; | |
5196 | } else { | |
5197 | // notify clients if I'm survivor | |
5198 | notify_clients = true; | |
5199 | } | |
5200 | do_realm_invalidate_and_update_notify(in, p.second, notify_clients); | |
7c673cae FG |
5201 | } |
5202 | ||
5203 | // done? | |
11fdf7f2 | 5204 | ceph_assert(rejoin_ack_gather.count(from)); |
7c673cae | 5205 | rejoin_ack_gather.erase(from); |
b32b8144 | 5206 | if (!survivor) { |
7c673cae FG |
5207 | if (rejoin_gather.empty()) { |
5208 | // eval unstable scatter locks after all wrlocks are rejoined. | |
5209 | while (!rejoin_eval_locks.empty()) { | |
5210 | SimpleLock *lock = rejoin_eval_locks.front(); | |
5211 | rejoin_eval_locks.pop_front(); | |
5212 | if (!lock->is_stable()) | |
5213 | mds->locker->eval_gather(lock); | |
5214 | } | |
5215 | } | |
5216 | ||
5217 | if (rejoin_gather.empty() && // make sure we've gotten our FULL inodes, too. | |
5218 | rejoin_ack_gather.empty()) { | |
5219 | // finally, kickstart past snap parent opens | |
11fdf7f2 | 5220 | open_snaprealms(); |
7c673cae FG |
5221 | } else { |
5222 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" | |
5223 | << ", rejoin_ack from (" << rejoin_ack_gather << ")" << dendl; | |
5224 | } | |
5225 | } else { | |
5226 | // survivor. | |
5227 | mds->queue_waiters(rejoin_waiters); | |
5228 | } | |
5229 | } | |
5230 | ||
5231 | /** | |
5232 | * rejoin_trim_undef_inodes() -- remove REJOINUNDEF flagged inodes | |
5233 | * | |
5234 | * FIXME: wait, can this actually happen? a survivor should generate cache trim | |
5235 | * messages that clean these guys up... | |
5236 | */ | |
5237 | void MDCache::rejoin_trim_undef_inodes() | |
5238 | { | |
5239 | dout(10) << "rejoin_trim_undef_inodes" << dendl; | |
5240 | ||
5241 | while (!rejoin_undef_inodes.empty()) { | |
5242 | set<CInode*>::iterator p = rejoin_undef_inodes.begin(); | |
5243 | CInode *in = *p; | |
5244 | rejoin_undef_inodes.erase(p); | |
5245 | ||
5246 | in->clear_replica_map(); | |
5247 | ||
5248 | // close out dirfrags | |
5249 | if (in->is_dir()) { | |
9f95a23c TL |
5250 | const auto&& dfls = in->get_dirfrags(); |
5251 | for (const auto& dir : dfls) { | |
7c673cae FG |
5252 | dir->clear_replica_map(); |
5253 | ||
94b18763 FG |
5254 | for (auto &p : dir->items) { |
5255 | CDentry *dn = p.second; | |
7c673cae FG |
5256 | dn->clear_replica_map(); |
5257 | ||
5258 | dout(10) << " trimming " << *dn << dendl; | |
5259 | dir->remove_dentry(dn); | |
5260 | } | |
5261 | ||
5262 | dout(10) << " trimming " << *dir << dendl; | |
5263 | in->close_dirfrag(dir->dirfrag().frag); | |
5264 | } | |
5265 | } | |
5266 | ||
5267 | CDentry *dn = in->get_parent_dn(); | |
5268 | if (dn) { | |
5269 | dn->clear_replica_map(); | |
5270 | dout(10) << " trimming " << *dn << dendl; | |
5271 | dn->dir->remove_dentry(dn); | |
5272 | } else { | |
5273 | dout(10) << " trimming " << *in << dendl; | |
5274 | remove_inode(in); | |
5275 | } | |
5276 | } | |
5277 | ||
11fdf7f2 | 5278 | ceph_assert(rejoin_undef_inodes.empty()); |
7c673cae FG |
5279 | } |
5280 | ||
5281 | void MDCache::rejoin_gather_finish() | |
5282 | { | |
5283 | dout(10) << "rejoin_gather_finish" << dendl; | |
11fdf7f2 TL |
5284 | ceph_assert(mds->is_rejoin()); |
5285 | ceph_assert(rejoin_ack_gather.count(mds->get_nodeid())); | |
7c673cae FG |
5286 | |
5287 | if (open_undef_inodes_dirfrags()) | |
5288 | return; | |
5289 | ||
5290 | if (process_imported_caps()) | |
5291 | return; | |
5292 | ||
5293 | choose_lock_states_and_reconnect_caps(); | |
5294 | ||
5295 | identify_files_to_recover(); | |
5296 | rejoin_send_acks(); | |
5297 | ||
5298 | // signal completion of fetches, rejoin_gather_finish, etc. | |
7c673cae FG |
5299 | rejoin_ack_gather.erase(mds->get_nodeid()); |
5300 | ||
5301 | // did we already get our acks too? | |
5302 | if (rejoin_ack_gather.empty()) { | |
11fdf7f2 TL |
5303 | // finally, open snaprealms |
5304 | open_snaprealms(); | |
7c673cae FG |
5305 | } |
5306 | } | |
5307 | ||
5308 | class C_MDC_RejoinOpenInoFinish: public MDCacheContext { | |
5309 | inodeno_t ino; | |
5310 | public: | |
5311 | C_MDC_RejoinOpenInoFinish(MDCache *c, inodeno_t i) : MDCacheContext(c), ino(i) {} | |
5312 | void finish(int r) override { | |
5313 | mdcache->rejoin_open_ino_finish(ino, r); | |
5314 | } | |
5315 | }; | |
5316 | ||
5317 | void MDCache::rejoin_open_ino_finish(inodeno_t ino, int ret) | |
5318 | { | |
5319 | dout(10) << "open_caps_inode_finish ino " << ino << " ret " << ret << dendl; | |
5320 | ||
5321 | if (ret < 0) { | |
5322 | cap_imports_missing.insert(ino); | |
5323 | } else if (ret == mds->get_nodeid()) { | |
11fdf7f2 | 5324 | ceph_assert(get_inode(ino)); |
7c673cae FG |
5325 | } else { |
5326 | auto p = cap_imports.find(ino); | |
11fdf7f2 | 5327 | ceph_assert(p != cap_imports.end()); |
7c673cae | 5328 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { |
11fdf7f2 TL |
5329 | ceph_assert(q->second.count(MDS_RANK_NONE)); |
5330 | ceph_assert(q->second.size() == 1); | |
7c673cae FG |
5331 | rejoin_export_caps(p->first, q->first, q->second[MDS_RANK_NONE], ret); |
5332 | } | |
5333 | cap_imports.erase(p); | |
5334 | } | |
5335 | ||
11fdf7f2 | 5336 | ceph_assert(cap_imports_num_opening > 0); |
7c673cae FG |
5337 | cap_imports_num_opening--; |
5338 | ||
5339 | if (cap_imports_num_opening == 0) { | |
522d829b | 5340 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) |
7c673cae FG |
5341 | rejoin_gather_finish(); |
5342 | else if (rejoin_gather.count(mds->get_nodeid())) | |
5343 | process_imported_caps(); | |
5344 | } | |
5345 | } | |
5346 | ||
5347 | class C_MDC_RejoinSessionsOpened : public MDCacheLogContext { | |
5348 | public: | |
28e407b8 AA |
5349 | map<client_t,pair<Session*,uint64_t> > session_map; |
5350 | C_MDC_RejoinSessionsOpened(MDCache *c) : MDCacheLogContext(c) {} | |
7c673cae | 5351 | void finish(int r) override { |
11fdf7f2 | 5352 | ceph_assert(r == 0); |
28e407b8 | 5353 | mdcache->rejoin_open_sessions_finish(session_map); |
7c673cae FG |
5354 | } |
5355 | }; | |
5356 | ||
28e407b8 | 5357 | void MDCache::rejoin_open_sessions_finish(map<client_t,pair<Session*,uint64_t> >& session_map) |
7c673cae FG |
5358 | { |
5359 | dout(10) << "rejoin_open_sessions_finish" << dendl; | |
28e407b8 AA |
5360 | mds->server->finish_force_open_sessions(session_map); |
5361 | rejoin_session_map.swap(session_map); | |
522d829b | 5362 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) |
7c673cae FG |
5363 | rejoin_gather_finish(); |
5364 | } | |
5365 | ||
11fdf7f2 TL |
5366 | void MDCache::rejoin_prefetch_ino_finish(inodeno_t ino, int ret) |
5367 | { | |
5368 | auto p = cap_imports.find(ino); | |
5369 | if (p != cap_imports.end()) { | |
5370 | dout(10) << __func__ << " ino " << ino << " ret " << ret << dendl; | |
5371 | if (ret < 0) { | |
5372 | cap_imports_missing.insert(ino); | |
5373 | } else if (ret != mds->get_nodeid()) { | |
5374 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
5375 | ceph_assert(q->second.count(MDS_RANK_NONE)); | |
5376 | ceph_assert(q->second.size() == 1); | |
5377 | rejoin_export_caps(p->first, q->first, q->second[MDS_RANK_NONE], ret); | |
5378 | } | |
5379 | cap_imports.erase(p); | |
5380 | } | |
5381 | } | |
5382 | } | |
5383 | ||
7c673cae FG |
5384 | bool MDCache::process_imported_caps() |
5385 | { | |
5386 | dout(10) << "process_imported_caps" << dendl; | |
5387 | ||
11fdf7f2 TL |
5388 | if (!open_file_table.is_prefetched() && |
5389 | open_file_table.prefetch_inodes()) { | |
5390 | open_file_table.wait_for_prefetch( | |
5391 | new MDSInternalContextWrapper(mds, | |
9f95a23c | 5392 | new LambdaContext([this](int r) { |
11fdf7f2 TL |
5393 | ceph_assert(rejoin_gather.count(mds->get_nodeid())); |
5394 | process_imported_caps(); | |
5395 | }) | |
5396 | ) | |
5397 | ); | |
5398 | return true; | |
5399 | } | |
5400 | ||
f91f0fd5 TL |
5401 | for (auto& p : cap_imports) { |
5402 | CInode *in = get_inode(p.first); | |
7c673cae | 5403 | if (in) { |
11fdf7f2 | 5404 | ceph_assert(in->is_auth()); |
f91f0fd5 | 5405 | cap_imports_missing.erase(p.first); |
7c673cae FG |
5406 | continue; |
5407 | } | |
f91f0fd5 | 5408 | if (cap_imports_missing.count(p.first) > 0) |
7c673cae FG |
5409 | continue; |
5410 | ||
f91f0fd5 TL |
5411 | uint64_t parent_ino = 0; |
5412 | std::string_view d_name; | |
5413 | for (auto& q : p.second) { | |
5414 | for (auto& r : q.second) { | |
5415 | auto &icr = r.second; | |
5416 | if (icr.capinfo.pathbase && | |
5417 | icr.path.length() > 0 && | |
5418 | icr.path.find('/') == string::npos) { | |
5419 | parent_ino = icr.capinfo.pathbase; | |
5420 | d_name = icr.path; | |
5421 | break; | |
5422 | } | |
5423 | } | |
5424 | if (parent_ino) | |
5425 | break; | |
5426 | } | |
5427 | ||
5428 | dout(10) << " opening missing ino " << p.first << dendl; | |
7c673cae | 5429 | cap_imports_num_opening++; |
f91f0fd5 TL |
5430 | auto fin = new C_MDC_RejoinOpenInoFinish(this, p.first); |
5431 | if (parent_ino) { | |
5432 | vector<inode_backpointer_t> ancestors; | |
5433 | ancestors.push_back(inode_backpointer_t(parent_ino, string{d_name}, 0)); | |
5434 | open_ino(p.first, (int64_t)-1, fin, false, false, &ancestors); | |
5435 | } else { | |
5436 | open_ino(p.first, (int64_t)-1, fin, false); | |
5437 | } | |
33c7a0ef | 5438 | if (!(cap_imports_num_opening % mds->heartbeat_reset_grace())) |
28e407b8 | 5439 | mds->heartbeat_reset(); |
7c673cae FG |
5440 | } |
5441 | ||
5442 | if (cap_imports_num_opening > 0) | |
5443 | return true; | |
5444 | ||
5445 | // called by rejoin_gather_finish() ? | |
5446 | if (rejoin_gather.count(mds->get_nodeid()) == 0) { | |
28e407b8 AA |
5447 | if (!rejoin_client_map.empty() && |
5448 | rejoin_session_map.empty()) { | |
5449 | C_MDC_RejoinSessionsOpened *finish = new C_MDC_RejoinSessionsOpened(this); | |
5450 | version_t pv = mds->server->prepare_force_open_sessions(rejoin_client_map, | |
11fdf7f2 | 5451 | rejoin_client_metadata_map, |
28e407b8 | 5452 | finish->session_map); |
11fdf7f2 TL |
5453 | ESessions *le = new ESessions(pv, std::move(rejoin_client_map), |
5454 | std::move(rejoin_client_metadata_map)); | |
5455 | mds->mdlog->start_submit_entry(le, finish); | |
28e407b8 AA |
5456 | mds->mdlog->flush(); |
5457 | rejoin_client_map.clear(); | |
11fdf7f2 | 5458 | rejoin_client_metadata_map.clear(); |
28e407b8 | 5459 | return true; |
7c673cae | 5460 | } |
7c673cae | 5461 | |
f67539c2 TL |
5462 | // process caps that were exported by peer rename |
5463 | for (map<inodeno_t,pair<mds_rank_t,map<client_t,Capability::Export> > >::iterator p = rejoin_peer_exports.begin(); | |
5464 | p != rejoin_peer_exports.end(); | |
7c673cae FG |
5465 | ++p) { |
5466 | CInode *in = get_inode(p->first); | |
11fdf7f2 | 5467 | ceph_assert(in); |
7c673cae FG |
5468 | for (map<client_t,Capability::Export>::iterator q = p->second.second.begin(); |
5469 | q != p->second.second.end(); | |
5470 | ++q) { | |
28e407b8 AA |
5471 | auto r = rejoin_session_map.find(q->first); |
5472 | if (r == rejoin_session_map.end()) | |
5473 | continue; | |
7c673cae | 5474 | |
28e407b8 | 5475 | Session *session = r->second.first; |
7c673cae | 5476 | Capability *cap = in->get_client_cap(q->first); |
11fdf7f2 | 5477 | if (!cap) { |
7c673cae | 5478 | cap = in->add_client_cap(q->first, session); |
11fdf7f2 TL |
5479 | // add empty item to reconnected_caps |
5480 | (void)reconnected_caps[p->first][q->first]; | |
5481 | } | |
7c673cae FG |
5482 | cap->merge(q->second, true); |
5483 | ||
5484 | Capability::Import& im = rejoin_imported_caps[p->second.first][p->first][q->first]; | |
11fdf7f2 TL |
5485 | ceph_assert(cap->get_last_seq() == im.issue_seq); |
5486 | ceph_assert(cap->get_mseq() == im.mseq); | |
7c673cae FG |
5487 | cap->set_cap_id(im.cap_id); |
5488 | // send cap import because we assigned a new cap ID | |
5489 | do_cap_import(session, in, cap, q->second.cap_id, q->second.seq, q->second.mseq - 1, | |
5490 | p->second.first, CEPH_CAP_FLAG_AUTH); | |
5491 | } | |
5492 | } | |
f67539c2 | 5493 | rejoin_peer_exports.clear(); |
7c673cae FG |
5494 | rejoin_imported_caps.clear(); |
5495 | ||
5496 | // process cap imports | |
5497 | // ino -> client -> frommds -> capex | |
5498 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ) { | |
5499 | CInode *in = get_inode(p->first); | |
5500 | if (!in) { | |
5501 | dout(10) << " still missing ino " << p->first | |
5502 | << ", will try again after replayed client requests" << dendl; | |
5503 | ++p; | |
5504 | continue; | |
5505 | } | |
11fdf7f2 | 5506 | ceph_assert(in->is_auth()); |
7c673cae | 5507 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { |
28e407b8 AA |
5508 | Session *session; |
5509 | { | |
5510 | auto r = rejoin_session_map.find(q->first); | |
5511 | session = (r != rejoin_session_map.end() ? r->second.first : nullptr); | |
5512 | } | |
5513 | ||
7c673cae | 5514 | for (auto r = q->second.begin(); r != q->second.end(); ++r) { |
28e407b8 AA |
5515 | if (!session) { |
5516 | if (r->first >= 0) | |
5517 | (void)rejoin_imported_caps[r->first][p->first][q->first]; // all are zero | |
5518 | continue; | |
5519 | } | |
5520 | ||
7c673cae FG |
5521 | Capability *cap = in->reconnect_cap(q->first, r->second, session); |
5522 | add_reconnected_cap(q->first, in->ino(), r->second); | |
5523 | if (r->first >= 0) { | |
5524 | if (cap->get_last_seq() == 0) // don't increase mseq if cap already exists | |
5525 | cap->inc_mseq(); | |
5526 | do_cap_import(session, in, cap, r->second.capinfo.cap_id, 0, 0, r->first, 0); | |
5527 | ||
5528 | Capability::Import& im = rejoin_imported_caps[r->first][p->first][q->first]; | |
5529 | im.cap_id = cap->get_cap_id(); | |
5530 | im.issue_seq = cap->get_last_seq(); | |
5531 | im.mseq = cap->get_mseq(); | |
5532 | } | |
5533 | } | |
5534 | } | |
5535 | cap_imports.erase(p++); // remove and move on | |
5536 | } | |
5537 | } else { | |
5538 | trim_non_auth(); | |
5539 | ||
11fdf7f2 | 5540 | ceph_assert(rejoin_gather.count(mds->get_nodeid())); |
7c673cae | 5541 | rejoin_gather.erase(mds->get_nodeid()); |
11fdf7f2 | 5542 | ceph_assert(!rejoin_ack_gather.count(mds->get_nodeid())); |
7c673cae | 5543 | maybe_send_pending_rejoins(); |
7c673cae FG |
5544 | } |
5545 | return false; | |
5546 | } | |
5547 | ||
7c673cae FG |
5548 | void MDCache::rebuild_need_snapflush(CInode *head_in, SnapRealm *realm, |
5549 | client_t client, snapid_t snap_follows) | |
5550 | { | |
5551 | dout(10) << "rebuild_need_snapflush " << snap_follows << " on " << *head_in << dendl; | |
5552 | ||
11fdf7f2 TL |
5553 | if (!realm->has_snaps_in_range(snap_follows + 1, head_in->first - 1)) |
5554 | return; | |
5555 | ||
7c673cae FG |
5556 | const set<snapid_t>& snaps = realm->get_snaps(); |
5557 | snapid_t follows = snap_follows; | |
5558 | ||
5559 | while (true) { | |
5560 | CInode *in = pick_inode_snap(head_in, follows); | |
5561 | if (in == head_in) | |
5562 | break; | |
11fdf7f2 TL |
5563 | |
5564 | bool need_snapflush = false; | |
5565 | for (auto p = snaps.lower_bound(std::max<snapid_t>(in->first, (follows + 1))); | |
5566 | p != snaps.end() && *p <= in->last; | |
5567 | ++p) { | |
5568 | head_in->add_need_snapflush(in, *p, client); | |
5569 | need_snapflush = true; | |
5570 | } | |
5571 | follows = in->last; | |
5572 | if (!need_snapflush) | |
5573 | continue; | |
5574 | ||
7c673cae FG |
5575 | dout(10) << " need snapflush from client." << client << " on " << *in << dendl; |
5576 | ||
eafe8130 TL |
5577 | if (in->client_snap_caps.empty()) { |
5578 | for (int i = 0; i < num_cinode_locks; i++) { | |
5579 | int lockid = cinode_lock_info[i].lock; | |
5580 | SimpleLock *lock = in->get_lock(lockid); | |
5581 | ceph_assert(lock); | |
5582 | in->auth_pin(lock); | |
5583 | lock->set_state(LOCK_SNAP_SYNC); | |
5584 | lock->get_wrlock(true); | |
5585 | } | |
7c673cae | 5586 | } |
eafe8130 | 5587 | in->client_snap_caps.insert(client); |
11fdf7f2 | 5588 | mds->locker->mark_need_snapflush_inode(in); |
7c673cae FG |
5589 | } |
5590 | } | |
5591 | ||
5592 | /* | |
5593 | * choose lock states based on reconnected caps | |
5594 | */ | |
5595 | void MDCache::choose_lock_states_and_reconnect_caps() | |
5596 | { | |
5597 | dout(10) << "choose_lock_states_and_reconnect_caps" << dendl; | |
5598 | ||
81eedcae | 5599 | int count = 0; |
11fdf7f2 TL |
5600 | for (auto p : inode_map) { |
5601 | CInode *in = p.second; | |
7c673cae FG |
5602 | if (in->last != CEPH_NOSNAP) |
5603 | continue; | |
5604 | ||
f67539c2 | 5605 | if (in->is_auth() && !in->is_base() && in->get_inode()->is_dirty_rstat()) |
7c673cae FG |
5606 | in->mark_dirty_rstat(); |
5607 | ||
7c673cae | 5608 | int dirty_caps = 0; |
11fdf7f2 TL |
5609 | auto q = reconnected_caps.find(in->ino()); |
5610 | if (q != reconnected_caps.end()) { | |
5611 | for (const auto &it : q->second) | |
7c673cae FG |
5612 | dirty_caps |= it.second.dirty_caps; |
5613 | } | |
5614 | in->choose_lock_states(dirty_caps); | |
5615 | dout(15) << " chose lock states on " << *in << dendl; | |
5616 | ||
11fdf7f2 TL |
5617 | if (in->snaprealm && !rejoin_pending_snaprealms.count(in)) { |
5618 | in->get(CInode::PIN_OPENINGSNAPPARENTS); | |
5619 | rejoin_pending_snaprealms.insert(in); | |
7c673cae | 5620 | } |
81eedcae | 5621 | |
33c7a0ef | 5622 | if (!(++count % mds->heartbeat_reset_grace())) |
81eedcae | 5623 | mds->heartbeat_reset(); |
11fdf7f2 | 5624 | } |
7c673cae FG |
5625 | } |
5626 | ||
5627 | void MDCache::prepare_realm_split(SnapRealm *realm, client_t client, inodeno_t ino, | |
9f95a23c | 5628 | map<client_t,ref_t<MClientSnap>>& splits) |
7c673cae | 5629 | { |
9f95a23c | 5630 | ref_t<MClientSnap> snap; |
11fdf7f2 TL |
5631 | auto it = splits.find(client); |
5632 | if (it != splits.end()) { | |
5633 | snap = it->second; | |
5634 | snap->head.op = CEPH_SNAP_OP_SPLIT; | |
5635 | } else { | |
9f95a23c | 5636 | snap = make_message<MClientSnap>(CEPH_SNAP_OP_SPLIT); |
11fdf7f2 | 5637 | splits.emplace(std::piecewise_construct, std::forward_as_tuple(client), std::forward_as_tuple(snap)); |
7c673cae | 5638 | snap->head.split = realm->inode->ino(); |
11fdf7f2 | 5639 | snap->bl = realm->get_snap_trace(); |
7c673cae | 5640 | |
11fdf7f2 TL |
5641 | for (const auto& child : realm->open_children) |
5642 | snap->split_realms.push_back(child->inode->ino()); | |
5643 | } | |
7c673cae FG |
5644 | snap->split_inos.push_back(ino); |
5645 | } | |
5646 | ||
11fdf7f2 | 5647 | void MDCache::prepare_realm_merge(SnapRealm *realm, SnapRealm *parent_realm, |
9f95a23c | 5648 | map<client_t,ref_t<MClientSnap>>& splits) |
11fdf7f2 TL |
5649 | { |
5650 | ceph_assert(parent_realm); | |
5651 | ||
5652 | vector<inodeno_t> split_inos; | |
5653 | vector<inodeno_t> split_realms; | |
5654 | ||
f67539c2 | 5655 | for (auto p = realm->inodes_with_caps.begin(); !p.end(); ++p) |
11fdf7f2 TL |
5656 | split_inos.push_back((*p)->ino()); |
5657 | for (set<SnapRealm*>::iterator p = realm->open_children.begin(); | |
5658 | p != realm->open_children.end(); | |
5659 | ++p) | |
5660 | split_realms.push_back((*p)->inode->ino()); | |
5661 | ||
5662 | for (const auto& p : realm->client_caps) { | |
5663 | ceph_assert(!p.second->empty()); | |
5664 | auto em = splits.emplace(std::piecewise_construct, std::forward_as_tuple(p.first), std::forward_as_tuple()); | |
5665 | if (em.second) { | |
9f95a23c | 5666 | auto update = make_message<MClientSnap>(CEPH_SNAP_OP_SPLIT); |
11fdf7f2 TL |
5667 | update->head.split = parent_realm->inode->ino(); |
5668 | update->split_inos = split_inos; | |
5669 | update->split_realms = split_realms; | |
5670 | update->bl = parent_realm->get_snap_trace(); | |
5671 | em.first->second = std::move(update); | |
5672 | } | |
5673 | } | |
5674 | } | |
5675 | ||
9f95a23c | 5676 | void MDCache::send_snaps(map<client_t,ref_t<MClientSnap>>& splits) |
7c673cae FG |
5677 | { |
5678 | dout(10) << "send_snaps" << dendl; | |
5679 | ||
11fdf7f2 TL |
5680 | for (auto &p : splits) { |
5681 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(p.first.v)); | |
7c673cae | 5682 | if (session) { |
11fdf7f2 TL |
5683 | dout(10) << " client." << p.first |
5684 | << " split " << p.second->head.split | |
5685 | << " inos " << p.second->split_inos | |
7c673cae | 5686 | << dendl; |
11fdf7f2 | 5687 | mds->send_message_client_counted(p.second, session); |
7c673cae | 5688 | } else { |
11fdf7f2 | 5689 | dout(10) << " no session for client." << p.first << dendl; |
7c673cae FG |
5690 | } |
5691 | } | |
5692 | splits.clear(); | |
5693 | } | |
5694 | ||
5695 | ||
5696 | /* | |
5697 | * remove any items from logsegment open_file lists that don't have | |
5698 | * any caps | |
5699 | */ | |
5700 | void MDCache::clean_open_file_lists() | |
5701 | { | |
5702 | dout(10) << "clean_open_file_lists" << dendl; | |
5703 | ||
5704 | for (map<uint64_t,LogSegment*>::iterator p = mds->mdlog->segments.begin(); | |
5705 | p != mds->mdlog->segments.end(); | |
5706 | ++p) { | |
5707 | LogSegment *ls = p->second; | |
5708 | ||
5709 | elist<CInode*>::iterator q = ls->open_files.begin(member_offset(CInode, item_open_file)); | |
5710 | while (!q.end()) { | |
5711 | CInode *in = *q; | |
5712 | ++q; | |
5713 | if (in->last == CEPH_NOSNAP) { | |
11fdf7f2 TL |
5714 | dout(10) << " unlisting unwanted/capless inode " << *in << dendl; |
5715 | in->item_open_file.remove_myself(); | |
5716 | } else { | |
7c673cae FG |
5717 | if (in->client_snap_caps.empty()) { |
5718 | dout(10) << " unlisting flushed snap inode " << *in << dendl; | |
5719 | in->item_open_file.remove_myself(); | |
5720 | } | |
5721 | } | |
5722 | } | |
5723 | } | |
5724 | } | |
5725 | ||
11fdf7f2 TL |
5726 | void MDCache::dump_openfiles(Formatter *f) |
5727 | { | |
5728 | f->open_array_section("openfiles"); | |
5729 | for (auto p = mds->mdlog->segments.begin(); | |
5730 | p != mds->mdlog->segments.end(); | |
5731 | ++p) { | |
5732 | LogSegment *ls = p->second; | |
5733 | ||
5734 | auto q = ls->open_files.begin(member_offset(CInode, item_open_file)); | |
5735 | while (!q.end()) { | |
5736 | CInode *in = *q; | |
5737 | ++q; | |
5738 | if ((in->last == CEPH_NOSNAP && !in->is_any_caps_wanted()) | |
5739 | || (in->last != CEPH_NOSNAP && in->client_snap_caps.empty())) | |
5740 | continue; | |
5741 | f->open_object_section("file"); | |
5742 | in->dump(f, CInode::DUMP_PATH | CInode::DUMP_INODE_STORE_BASE | CInode::DUMP_CAPS); | |
5743 | f->close_section(); | |
5744 | } | |
5745 | } | |
5746 | f->close_section(); | |
5747 | } | |
7c673cae FG |
5748 | |
5749 | Capability* MDCache::rejoin_import_cap(CInode *in, client_t client, const cap_reconnect_t& icr, mds_rank_t frommds) | |
5750 | { | |
5751 | dout(10) << "rejoin_import_cap for client." << client << " from mds." << frommds | |
5752 | << " on " << *in << dendl; | |
5753 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(client.v)); | |
5754 | if (!session) { | |
5755 | dout(10) << " no session for client." << client << dendl; | |
5756 | return NULL; | |
5757 | } | |
5758 | ||
5759 | Capability *cap = in->reconnect_cap(client, icr, session); | |
5760 | ||
5761 | if (frommds >= 0) { | |
5762 | if (cap->get_last_seq() == 0) // don't increase mseq if cap already exists | |
5763 | cap->inc_mseq(); | |
5764 | do_cap_import(session, in, cap, icr.capinfo.cap_id, 0, 0, frommds, 0); | |
5765 | } | |
5766 | ||
5767 | return cap; | |
5768 | } | |
5769 | ||
5770 | void MDCache::export_remaining_imported_caps() | |
5771 | { | |
5772 | dout(10) << "export_remaining_imported_caps" << dendl; | |
5773 | ||
f67539c2 | 5774 | CachedStackStringStream css; |
7c673cae | 5775 | |
81eedcae | 5776 | int count = 0; |
7c673cae | 5777 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ++p) { |
f67539c2 | 5778 | *css << " ino " << p->first << "\n"; |
7c673cae FG |
5779 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { |
5780 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
5781 | if (session) { | |
5782 | // mark client caps stale. | |
9f95a23c TL |
5783 | auto stale = make_message<MClientCaps>(CEPH_CAP_OP_EXPORT, p->first, |
5784 | 0, 0, 0, | |
5785 | mds->get_osd_epoch_barrier()); | |
7c673cae FG |
5786 | stale->set_cap_peer(0, 0, 0, -1, 0); |
5787 | mds->send_message_client_counted(stale, q->first); | |
5788 | } | |
5789 | } | |
5790 | ||
33c7a0ef | 5791 | if (!(++count % mds->heartbeat_reset_grace())) |
81eedcae | 5792 | mds->heartbeat_reset(); |
7c673cae FG |
5793 | } |
5794 | ||
11fdf7f2 | 5795 | for (map<inodeno_t, MDSContext::vec >::iterator p = cap_reconnect_waiters.begin(); |
7c673cae FG |
5796 | p != cap_reconnect_waiters.end(); |
5797 | ++p) | |
5798 | mds->queue_waiters(p->second); | |
5799 | ||
5800 | cap_imports.clear(); | |
5801 | cap_reconnect_waiters.clear(); | |
5802 | ||
f67539c2 TL |
5803 | if (css->strv().length()) { |
5804 | mds->clog->warn() << "failed to reconnect caps for missing inodes:" | |
5805 | << css->strv(); | |
7c673cae FG |
5806 | } |
5807 | } | |
5808 | ||
a8e16298 | 5809 | Capability* MDCache::try_reconnect_cap(CInode *in, Session *session) |
7c673cae FG |
5810 | { |
5811 | client_t client = session->info.get_client(); | |
a8e16298 | 5812 | Capability *cap = nullptr; |
7c673cae FG |
5813 | const cap_reconnect_t *rc = get_replay_cap_reconnect(in->ino(), client); |
5814 | if (rc) { | |
a8e16298 | 5815 | cap = in->reconnect_cap(client, *rc, session); |
7c673cae FG |
5816 | dout(10) << "try_reconnect_cap client." << client |
5817 | << " reconnect wanted " << ccap_string(rc->capinfo.wanted) | |
5818 | << " issue " << ccap_string(rc->capinfo.issued) | |
5819 | << " on " << *in << dendl; | |
5820 | remove_replay_cap_reconnect(in->ino(), client); | |
5821 | ||
5822 | if (in->is_replicated()) { | |
5823 | mds->locker->try_eval(in, CEPH_CAP_LOCKS); | |
5824 | } else { | |
5825 | int dirty_caps = 0; | |
5826 | auto p = reconnected_caps.find(in->ino()); | |
5827 | if (p != reconnected_caps.end()) { | |
5828 | auto q = p->second.find(client); | |
5829 | if (q != p->second.end()) | |
5830 | dirty_caps = q->second.dirty_caps; | |
5831 | } | |
5832 | in->choose_lock_states(dirty_caps); | |
5833 | dout(15) << " chose lock states on " << *in << dendl; | |
5834 | } | |
5835 | ||
11fdf7f2 | 5836 | map<inodeno_t, MDSContext::vec >::iterator it = |
7c673cae FG |
5837 | cap_reconnect_waiters.find(in->ino()); |
5838 | if (it != cap_reconnect_waiters.end()) { | |
5839 | mds->queue_waiters(it->second); | |
5840 | cap_reconnect_waiters.erase(it); | |
5841 | } | |
5842 | } | |
a8e16298 | 5843 | return cap; |
7c673cae FG |
5844 | } |
5845 | ||
5846 | ||
5847 | ||
5848 | // ------- | |
5849 | // cap imports and delayed snap parent opens | |
5850 | ||
5851 | void MDCache::do_cap_import(Session *session, CInode *in, Capability *cap, | |
5852 | uint64_t p_cap_id, ceph_seq_t p_seq, ceph_seq_t p_mseq, | |
5853 | int peer, int p_flags) | |
5854 | { | |
7c673cae | 5855 | SnapRealm *realm = in->find_snaprealm(); |
f67539c2 TL |
5856 | dout(10) << "do_cap_import " << session->info.inst.name << " mseq " << cap->get_mseq() << " on " << *in << dendl; |
5857 | if (cap->get_last_seq() == 0) // reconnected cap | |
5858 | cap->inc_last_seq(); | |
5859 | cap->set_last_issue(); | |
5860 | cap->set_last_issue_stamp(ceph_clock_now()); | |
5861 | cap->clear_new(); | |
5862 | auto reap = make_message<MClientCaps>(CEPH_CAP_OP_IMPORT, | |
5863 | in->ino(), realm->inode->ino(), cap->get_cap_id(), | |
5864 | cap->get_last_seq(), cap->pending(), cap->wanted(), | |
5865 | 0, cap->get_mseq(), mds->get_osd_epoch_barrier()); | |
5866 | in->encode_cap_message(reap, cap); | |
5867 | reap->snapbl = realm->get_snap_trace(); | |
5868 | reap->set_cap_peer(p_cap_id, p_seq, p_mseq, peer, p_flags); | |
5869 | mds->send_message_client_counted(reap, session); | |
7c673cae FG |
5870 | } |
5871 | ||
5872 | void MDCache::do_delayed_cap_imports() | |
5873 | { | |
5874 | dout(10) << "do_delayed_cap_imports" << dendl; | |
5875 | ||
11fdf7f2 | 5876 | ceph_assert(delayed_imported_caps.empty()); |
7c673cae FG |
5877 | } |
5878 | ||
11fdf7f2 TL |
5879 | struct C_MDC_OpenSnapRealms : public MDCacheContext { |
5880 | explicit C_MDC_OpenSnapRealms(MDCache *c) : MDCacheContext(c) {} | |
7c673cae | 5881 | void finish(int r) override { |
11fdf7f2 | 5882 | mdcache->open_snaprealms(); |
7c673cae FG |
5883 | } |
5884 | }; | |
5885 | ||
11fdf7f2 | 5886 | void MDCache::open_snaprealms() |
7c673cae | 5887 | { |
11fdf7f2 | 5888 | dout(10) << "open_snaprealms" << dendl; |
7c673cae | 5889 | |
11fdf7f2 TL |
5890 | auto it = rejoin_pending_snaprealms.begin(); |
5891 | while (it != rejoin_pending_snaprealms.end()) { | |
5892 | CInode *in = *it; | |
5893 | SnapRealm *realm = in->snaprealm; | |
5894 | ceph_assert(realm); | |
f67539c2 TL |
5895 | |
5896 | map<client_t,ref_t<MClientSnap>> splits; | |
5897 | // finish off client snaprealm reconnects? | |
5898 | auto q = reconnected_snaprealms.find(in->ino()); | |
5899 | if (q != reconnected_snaprealms.end()) { | |
5900 | for (const auto& r : q->second) | |
5901 | finish_snaprealm_reconnect(r.first, realm, r.second, splits); | |
5902 | reconnected_snaprealms.erase(q); | |
5903 | } | |
5904 | ||
5905 | for (auto p = realm->inodes_with_caps.begin(); !p.end(); ++p) { | |
5906 | CInode *child = *p; | |
5907 | auto q = reconnected_caps.find(child->ino()); | |
5908 | ceph_assert(q != reconnected_caps.end()); | |
5909 | for (auto r = q->second.begin(); r != q->second.end(); ++r) { | |
5910 | Capability *cap = child->get_client_cap(r->first); | |
5911 | if (!cap) | |
5912 | continue; | |
5913 | if (r->second.snap_follows > 0) { | |
5914 | if (r->second.snap_follows < child->first - 1) { | |
5915 | rebuild_need_snapflush(child, realm, r->first, r->second.snap_follows); | |
5916 | } else if (r->second.snapflush) { | |
5917 | // When processing a cap flush message that is re-sent, it's possble | |
5918 | // that the sender has already released all WR caps. So we should | |
5919 | // force MDCache::cow_inode() to setup CInode::client_need_snapflush. | |
5920 | cap->mark_needsnapflush(); | |
7c673cae FG |
5921 | } |
5922 | } | |
f67539c2 TL |
5923 | // make sure client's cap is in the correct snaprealm. |
5924 | if (r->second.realm_ino != in->ino()) { | |
5925 | prepare_realm_split(realm, r->first, child->ino(), splits); | |
5926 | } | |
7c673cae | 5927 | } |
7c673cae | 5928 | } |
7c673cae | 5929 | |
f67539c2 TL |
5930 | rejoin_pending_snaprealms.erase(it++); |
5931 | in->put(CInode::PIN_OPENINGSNAPPARENTS); | |
11fdf7f2 | 5932 | |
f67539c2 | 5933 | send_snaps(splits); |
11fdf7f2 TL |
5934 | } |
5935 | ||
5936 | notify_global_snaprealm_update(CEPH_SNAP_OP_UPDATE); | |
5937 | ||
5938 | if (!reconnected_snaprealms.empty()) { | |
5939 | dout(5) << "open_snaprealms has unconnected snaprealm:" << dendl; | |
5940 | for (auto& p : reconnected_snaprealms) { | |
f67539c2 TL |
5941 | CachedStackStringStream css; |
5942 | *css << " " << p.first << " {"; | |
11fdf7f2 TL |
5943 | bool first = true; |
5944 | for (auto& q : p.second) { | |
5945 | if (!first) | |
f67539c2 TL |
5946 | *css << ", "; |
5947 | *css << "client." << q.first << "/" << q.second; | |
7c673cae | 5948 | } |
f67539c2 TL |
5949 | *css << "}"; |
5950 | dout(5) << css->strv() << dendl; | |
7c673cae | 5951 | } |
7c673cae | 5952 | } |
11fdf7f2 TL |
5953 | ceph_assert(rejoin_waiters.empty()); |
5954 | ceph_assert(rejoin_pending_snaprealms.empty()); | |
5955 | dout(10) << "open_snaprealms - all open" << dendl; | |
5956 | do_delayed_cap_imports(); | |
5957 | ||
5958 | ceph_assert(rejoin_done); | |
5959 | rejoin_done.release()->complete(0); | |
5960 | reconnected_caps.clear(); | |
7c673cae FG |
5961 | } |
5962 | ||
5963 | bool MDCache::open_undef_inodes_dirfrags() | |
5964 | { | |
5965 | dout(10) << "open_undef_inodes_dirfrags " | |
5966 | << rejoin_undef_inodes.size() << " inodes " | |
5967 | << rejoin_undef_dirfrags.size() << " dirfrags" << dendl; | |
5968 | ||
5969 | set<CDir*> fetch_queue = rejoin_undef_dirfrags; | |
5970 | ||
5971 | for (set<CInode*>::iterator p = rejoin_undef_inodes.begin(); | |
5972 | p != rejoin_undef_inodes.end(); | |
5973 | ++p) { | |
5974 | CInode *in = *p; | |
11fdf7f2 | 5975 | ceph_assert(!in->is_base()); |
f67539c2 | 5976 | ceph_assert(in->get_parent_dir()); |
7c673cae FG |
5977 | fetch_queue.insert(in->get_parent_dir()); |
5978 | } | |
5979 | ||
5980 | if (fetch_queue.empty()) | |
5981 | return false; | |
5982 | ||
28e407b8 AA |
5983 | MDSGatherBuilder gather(g_ceph_context, |
5984 | new MDSInternalContextWrapper(mds, | |
9f95a23c | 5985 | new LambdaContext([this](int r) { |
522d829b | 5986 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) |
28e407b8 AA |
5987 | rejoin_gather_finish(); |
5988 | }) | |
5989 | ) | |
5990 | ); | |
5991 | ||
7c673cae FG |
5992 | for (set<CDir*>::iterator p = fetch_queue.begin(); |
5993 | p != fetch_queue.end(); | |
5994 | ++p) { | |
5995 | CDir *dir = *p; | |
5996 | CInode *diri = dir->get_inode(); | |
5997 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) | |
5998 | continue; | |
5999 | if (dir->state_test(CDir::STATE_REJOINUNDEF)) | |
11fdf7f2 | 6000 | ceph_assert(diri->dirfragtree.is_leaf(dir->get_frag())); |
7c673cae FG |
6001 | dir->fetch(gather.new_sub()); |
6002 | } | |
11fdf7f2 | 6003 | ceph_assert(gather.has_subs()); |
7c673cae FG |
6004 | gather.activate(); |
6005 | return true; | |
6006 | } | |
6007 | ||
6008 | void MDCache::opened_undef_inode(CInode *in) { | |
6009 | dout(10) << "opened_undef_inode " << *in << dendl; | |
6010 | rejoin_undef_inodes.erase(in); | |
6011 | if (in->is_dir()) { | |
6012 | // FIXME: re-hash dentries if necessary | |
f67539c2 | 6013 | ceph_assert(in->get_inode()->dir_layout.dl_dir_hash == g_conf()->mds_default_dir_hash); |
9f95a23c | 6014 | if (in->get_num_dirfrags() && !in->dirfragtree.is_leaf(frag_t())) { |
7c673cae | 6015 | CDir *dir = in->get_dirfrag(frag_t()); |
11fdf7f2 | 6016 | ceph_assert(dir); |
7c673cae FG |
6017 | rejoin_undef_dirfrags.erase(dir); |
6018 | in->force_dirfrags(); | |
9f95a23c TL |
6019 | auto&& ls = in->get_dirfrags(); |
6020 | for (const auto& dir : ls) { | |
6021 | rejoin_undef_dirfrags.insert(dir); | |
6022 | } | |
7c673cae FG |
6023 | } |
6024 | } | |
6025 | } | |
6026 | ||
11fdf7f2 | 6027 | void MDCache::finish_snaprealm_reconnect(client_t client, SnapRealm *realm, snapid_t seq, |
9f95a23c | 6028 | map<client_t,ref_t<MClientSnap>>& updates) |
7c673cae FG |
6029 | { |
6030 | if (seq < realm->get_newest_seq()) { | |
6031 | dout(10) << "finish_snaprealm_reconnect client." << client << " has old seq " << seq << " < " | |
11fdf7f2 | 6032 | << realm->get_newest_seq() << " on " << *realm << dendl; |
9f95a23c | 6033 | auto snap = make_message<MClientSnap>(CEPH_SNAP_OP_UPDATE); |
11fdf7f2 TL |
6034 | snap->bl = realm->get_snap_trace(); |
6035 | for (const auto& child : realm->open_children) | |
6036 | snap->split_realms.push_back(child->inode->ino()); | |
6037 | updates.emplace(std::piecewise_construct, std::forward_as_tuple(client), std::forward_as_tuple(snap)); | |
7c673cae FG |
6038 | } else { |
6039 | dout(10) << "finish_snaprealm_reconnect client." << client << " up to date" | |
6040 | << " on " << *realm << dendl; | |
6041 | } | |
6042 | } | |
6043 | ||
6044 | ||
6045 | ||
6046 | void MDCache::rejoin_send_acks() | |
6047 | { | |
6048 | dout(7) << "rejoin_send_acks" << dendl; | |
6049 | ||
6050 | // replicate stray | |
6051 | for (map<mds_rank_t, set<CInode*> >::iterator p = rejoin_unlinked_inodes.begin(); | |
6052 | p != rejoin_unlinked_inodes.end(); | |
6053 | ++p) { | |
6054 | for (set<CInode*>::iterator q = p->second.begin(); | |
6055 | q != p->second.end(); | |
6056 | ++q) { | |
6057 | CInode *in = *q; | |
6058 | dout(7) << " unlinked inode " << *in << dendl; | |
6059 | // inode expired | |
6060 | if (!in->is_replica(p->first)) | |
6061 | continue; | |
6062 | while (1) { | |
6063 | CDentry *dn = in->get_parent_dn(); | |
6064 | if (dn->is_replica(p->first)) | |
6065 | break; | |
6066 | dn->add_replica(p->first); | |
6067 | CDir *dir = dn->get_dir(); | |
6068 | if (dir->is_replica(p->first)) | |
6069 | break; | |
6070 | dir->add_replica(p->first); | |
6071 | in = dir->get_inode(); | |
6072 | if (in->is_replica(p->first)) | |
6073 | break; | |
224ce89b | 6074 | in->add_replica(p->first); |
7c673cae FG |
6075 | if (in->is_base()) |
6076 | break; | |
6077 | } | |
6078 | } | |
6079 | } | |
6080 | rejoin_unlinked_inodes.clear(); | |
6081 | ||
6082 | // send acks to everyone in the recovery set | |
9f95a23c | 6083 | map<mds_rank_t,ref_t<MMDSCacheRejoin>> acks; |
7c673cae FG |
6084 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); |
6085 | p != recovery_set.end(); | |
31f18b77 FG |
6086 | ++p) { |
6087 | if (rejoin_ack_sent.count(*p)) | |
6088 | continue; | |
9f95a23c | 6089 | acks[*p] = make_message<MMDSCacheRejoin>(MMDSCacheRejoin::OP_ACK); |
31f18b77 FG |
6090 | } |
6091 | ||
6092 | rejoin_ack_sent = recovery_set; | |
7c673cae FG |
6093 | |
6094 | // walk subtrees | |
6095 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
6096 | p != subtrees.end(); | |
6097 | ++p) { | |
6098 | CDir *dir = p->first; | |
6099 | if (!dir->is_auth()) | |
6100 | continue; | |
6101 | dout(10) << "subtree " << *dir << dendl; | |
6102 | ||
6103 | // auth items in this subtree | |
9f95a23c TL |
6104 | std::queue<CDir*> dq; |
6105 | dq.push(dir); | |
7c673cae FG |
6106 | |
6107 | while (!dq.empty()) { | |
6108 | CDir *dir = dq.front(); | |
9f95a23c | 6109 | dq.pop(); |
7c673cae FG |
6110 | |
6111 | // dir | |
181888fb FG |
6112 | for (auto &r : dir->get_replicas()) { |
6113 | auto it = acks.find(r.first); | |
31f18b77 FG |
6114 | if (it == acks.end()) |
6115 | continue; | |
181888fb | 6116 | it->second->add_strong_dirfrag(dir->dirfrag(), ++r.second, dir->dir_rep); |
31f18b77 | 6117 | it->second->add_dirfrag_base(dir); |
7c673cae FG |
6118 | } |
6119 | ||
94b18763 FG |
6120 | for (auto &p : dir->items) { |
6121 | CDentry *dn = p.second; | |
7c673cae FG |
6122 | CDentry::linkage_t *dnl = dn->get_linkage(); |
6123 | ||
6124 | // inode | |
6125 | CInode *in = NULL; | |
6126 | if (dnl->is_primary()) | |
6127 | in = dnl->get_inode(); | |
6128 | ||
6129 | // dentry | |
181888fb FG |
6130 | for (auto &r : dn->get_replicas()) { |
6131 | auto it = acks.find(r.first); | |
31f18b77 FG |
6132 | if (it == acks.end()) |
6133 | continue; | |
f67539c2 TL |
6134 | it->second->add_strong_dentry(dir->dirfrag(), dn->get_name(), dn->get_alternate_name(), |
6135 | dn->first, dn->last, | |
7c673cae FG |
6136 | dnl->is_primary() ? dnl->get_inode()->ino():inodeno_t(0), |
6137 | dnl->is_remote() ? dnl->get_remote_ino():inodeno_t(0), | |
6138 | dnl->is_remote() ? dnl->get_remote_d_type():0, | |
181888fb | 6139 | ++r.second, |
7c673cae FG |
6140 | dn->lock.get_replica_state()); |
6141 | // peer missed MDentrylink message ? | |
181888fb FG |
6142 | if (in && !in->is_replica(r.first)) |
6143 | in->add_replica(r.first); | |
7c673cae FG |
6144 | } |
6145 | ||
6146 | if (!in) | |
6147 | continue; | |
6148 | ||
181888fb FG |
6149 | for (auto &r : in->get_replicas()) { |
6150 | auto it = acks.find(r.first); | |
31f18b77 FG |
6151 | if (it == acks.end()) |
6152 | continue; | |
6153 | it->second->add_inode_base(in, mds->mdsmap->get_up_features()); | |
7c673cae | 6154 | bufferlist bl; |
181888fb FG |
6155 | in->_encode_locks_state_for_rejoin(bl, r.first); |
6156 | it->second->add_inode_locks(in, ++r.second, bl); | |
7c673cae FG |
6157 | } |
6158 | ||
6159 | // subdirs in this subtree? | |
9f95a23c TL |
6160 | { |
6161 | auto&& dirs = in->get_nested_dirfrags(); | |
6162 | for (const auto& dir : dirs) { | |
6163 | dq.push(dir); | |
6164 | } | |
6165 | } | |
7c673cae FG |
6166 | } |
6167 | } | |
6168 | } | |
6169 | ||
6170 | // base inodes too | |
6171 | if (root && root->is_auth()) | |
181888fb FG |
6172 | for (auto &r : root->get_replicas()) { |
6173 | auto it = acks.find(r.first); | |
31f18b77 FG |
6174 | if (it == acks.end()) |
6175 | continue; | |
6176 | it->second->add_inode_base(root, mds->mdsmap->get_up_features()); | |
7c673cae | 6177 | bufferlist bl; |
181888fb FG |
6178 | root->_encode_locks_state_for_rejoin(bl, r.first); |
6179 | it->second->add_inode_locks(root, ++r.second, bl); | |
7c673cae FG |
6180 | } |
6181 | if (myin) | |
181888fb FG |
6182 | for (auto &r : myin->get_replicas()) { |
6183 | auto it = acks.find(r.first); | |
31f18b77 FG |
6184 | if (it == acks.end()) |
6185 | continue; | |
6186 | it->second->add_inode_base(myin, mds->mdsmap->get_up_features()); | |
7c673cae | 6187 | bufferlist bl; |
181888fb FG |
6188 | myin->_encode_locks_state_for_rejoin(bl, r.first); |
6189 | it->second->add_inode_locks(myin, ++r.second, bl); | |
7c673cae FG |
6190 | } |
6191 | ||
6192 | // include inode base for any inodes whose scatterlocks may have updated | |
6193 | for (set<CInode*>::iterator p = rejoin_potential_updated_scatterlocks.begin(); | |
6194 | p != rejoin_potential_updated_scatterlocks.end(); | |
6195 | ++p) { | |
6196 | CInode *in = *p; | |
181888fb FG |
6197 | for (const auto &r : in->get_replicas()) { |
6198 | auto it = acks.find(r.first); | |
31f18b77 FG |
6199 | if (it == acks.end()) |
6200 | continue; | |
6201 | it->second->add_inode_base(in, mds->mdsmap->get_up_features()); | |
6202 | } | |
7c673cae FG |
6203 | } |
6204 | ||
6205 | // send acks | |
31f18b77 | 6206 | for (auto p = acks.begin(); p != acks.end(); ++p) { |
11fdf7f2 | 6207 | encode(rejoin_imported_caps[p->first], p->second->imported_caps); |
7c673cae FG |
6208 | mds->send_message_mds(p->second, p->first); |
6209 | } | |
6210 | ||
6211 | rejoin_imported_caps.clear(); | |
6212 | } | |
6213 | ||
c07f9fc5 FG |
6214 | class C_MDC_ReIssueCaps : public MDCacheContext { |
6215 | CInode *in; | |
6216 | public: | |
6217 | C_MDC_ReIssueCaps(MDCache *mdc, CInode *i) : | |
6218 | MDCacheContext(mdc), in(i) | |
6219 | { | |
6220 | in->get(CInode::PIN_PTRWAITER); | |
6221 | } | |
6222 | void finish(int r) override { | |
6223 | if (!mdcache->mds->locker->eval(in, CEPH_CAP_LOCKS)) | |
6224 | mdcache->mds->locker->issue_caps(in); | |
6225 | in->put(CInode::PIN_PTRWAITER); | |
6226 | } | |
6227 | }; | |
7c673cae FG |
6228 | |
6229 | void MDCache::reissue_all_caps() | |
6230 | { | |
6231 | dout(10) << "reissue_all_caps" << dendl; | |
6232 | ||
81eedcae | 6233 | int count = 0; |
94b18763 | 6234 | for (auto &p : inode_map) { |
81eedcae | 6235 | int n = 1; |
b32b8144 | 6236 | CInode *in = p.second; |
7c673cae | 6237 | if (in->is_head() && in->is_any_caps()) { |
c07f9fc5 FG |
6238 | // called by MDSRank::active_start(). There shouldn't be any frozen subtree. |
6239 | if (in->is_frozen_inode()) { | |
6240 | in->add_waiter(CInode::WAIT_UNFREEZE, new C_MDC_ReIssueCaps(this, in)); | |
6241 | continue; | |
6242 | } | |
7c673cae | 6243 | if (!mds->locker->eval(in, CEPH_CAP_LOCKS)) |
81eedcae | 6244 | n += mds->locker->issue_caps(in); |
7c673cae | 6245 | } |
81eedcae | 6246 | |
33c7a0ef | 6247 | if ((count % mds->heartbeat_reset_grace()) + n >= mds->heartbeat_reset_grace()) |
81eedcae TL |
6248 | mds->heartbeat_reset(); |
6249 | count += n; | |
7c673cae FG |
6250 | } |
6251 | } | |
6252 | ||
6253 | ||
6254 | // =============================================================================== | |
6255 | ||
6256 | struct C_MDC_QueuedCow : public MDCacheContext { | |
6257 | CInode *in; | |
6258 | MutationRef mut; | |
6259 | C_MDC_QueuedCow(MDCache *mdc, CInode *i, MutationRef& m) : | |
6260 | MDCacheContext(mdc), in(i), mut(m) {} | |
6261 | void finish(int r) override { | |
6262 | mdcache->_queued_file_recover_cow(in, mut); | |
6263 | } | |
6264 | }; | |
6265 | ||
6266 | ||
6267 | void MDCache::queue_file_recover(CInode *in) | |
6268 | { | |
6269 | dout(10) << "queue_file_recover " << *in << dendl; | |
11fdf7f2 | 6270 | ceph_assert(in->is_auth()); |
7c673cae FG |
6271 | |
6272 | // cow? | |
6273 | /* | |
6274 | SnapRealm *realm = in->find_snaprealm(); | |
6275 | set<snapid_t> s = realm->get_snaps(); | |
6276 | while (!s.empty() && *s.begin() < in->first) | |
6277 | s.erase(s.begin()); | |
6278 | while (!s.empty() && *s.rbegin() > in->last) | |
6279 | s.erase(*s.rbegin()); | |
6280 | dout(10) << " snaps in [" << in->first << "," << in->last << "] are " << s << dendl; | |
6281 | if (s.size() > 1) { | |
f67539c2 TL |
6282 | auto pi = in->project_inode(mut); |
6283 | pi.inode.version = in->pre_dirty(); | |
7c673cae FG |
6284 | |
6285 | auto mut(std::make_shared<MutationImpl>()); | |
6286 | mut->ls = mds->mdlog->get_current_segment(); | |
6287 | EUpdate *le = new EUpdate(mds->mdlog, "queue_file_recover cow"); | |
6288 | mds->mdlog->start_entry(le); | |
6289 | predirty_journal_parents(mut, &le->metablob, in, 0, PREDIRTY_PRIMARY); | |
6290 | ||
6291 | s.erase(*s.begin()); | |
6292 | while (!s.empty()) { | |
6293 | snapid_t snapid = *s.begin(); | |
6294 | CInode *cow_inode = 0; | |
6295 | journal_cow_inode(mut, &le->metablob, in, snapid-1, &cow_inode); | |
11fdf7f2 | 6296 | ceph_assert(cow_inode); |
7c673cae FG |
6297 | recovery_queue.enqueue(cow_inode); |
6298 | s.erase(*s.begin()); | |
6299 | } | |
6300 | ||
6301 | in->parent->first = in->first; | |
6302 | le->metablob.add_primary_dentry(in->parent, in, true); | |
6303 | mds->mdlog->submit_entry(le, new C_MDC_QueuedCow(this, in, mut)); | |
6304 | mds->mdlog->flush(); | |
6305 | } | |
6306 | */ | |
6307 | ||
6308 | recovery_queue.enqueue(in); | |
6309 | } | |
6310 | ||
6311 | void MDCache::_queued_file_recover_cow(CInode *in, MutationRef& mut) | |
6312 | { | |
7c673cae FG |
6313 | mut->apply(); |
6314 | mds->locker->drop_locks(mut.get()); | |
6315 | mut->cleanup(); | |
6316 | } | |
6317 | ||
6318 | ||
6319 | /* | |
6320 | * called after recovery to recover file sizes for previously opened (for write) | |
6321 | * files. that is, those where max_size > size. | |
6322 | */ | |
6323 | void MDCache::identify_files_to_recover() | |
6324 | { | |
6325 | dout(10) << "identify_files_to_recover" << dendl; | |
81eedcae | 6326 | int count = 0; |
94b18763 | 6327 | for (auto &p : inode_map) { |
b32b8144 | 6328 | CInode *in = p.second; |
7c673cae FG |
6329 | if (!in->is_auth()) |
6330 | continue; | |
6331 | ||
6332 | if (in->last != CEPH_NOSNAP) | |
6333 | continue; | |
6334 | ||
6335 | // Only normal files need file size recovery | |
6336 | if (!in->is_file()) { | |
6337 | continue; | |
6338 | } | |
6339 | ||
6340 | bool recover = false; | |
f91f0fd5 TL |
6341 | const auto& client_ranges = in->get_projected_inode()->client_ranges; |
6342 | if (!client_ranges.empty()) { | |
6343 | in->mark_clientwriteable(); | |
6344 | for (auto& p : client_ranges) { | |
6345 | Capability *cap = in->get_client_cap(p.first); | |
6346 | if (cap) { | |
6347 | cap->mark_clientwriteable(); | |
6348 | } else { | |
6349 | dout(10) << " client." << p.first << " has range " << p.second << " but no cap on " << *in << dendl; | |
6350 | recover = true; | |
6351 | break; | |
6352 | } | |
7c673cae FG |
6353 | } |
6354 | } | |
6355 | ||
6356 | if (recover) { | |
6357 | if (in->filelock.is_stable()) { | |
6358 | in->auth_pin(&in->filelock); | |
6359 | } else { | |
11fdf7f2 | 6360 | ceph_assert(in->filelock.get_state() == LOCK_XLOCKSNAP); |
7c673cae FG |
6361 | } |
6362 | in->filelock.set_state(LOCK_PRE_SCAN); | |
6363 | rejoin_recover_q.push_back(in); | |
6364 | } else { | |
6365 | rejoin_check_q.push_back(in); | |
6366 | } | |
81eedcae | 6367 | |
33c7a0ef | 6368 | if (!(++count % mds->heartbeat_reset_grace())) |
81eedcae | 6369 | mds->heartbeat_reset(); |
7c673cae FG |
6370 | } |
6371 | } | |
6372 | ||
6373 | void MDCache::start_files_to_recover() | |
6374 | { | |
b3b6e05e | 6375 | int count = 0; |
7c673cae FG |
6376 | for (CInode *in : rejoin_check_q) { |
6377 | if (in->filelock.get_state() == LOCK_XLOCKSNAP) | |
6378 | mds->locker->issue_caps(in); | |
6379 | mds->locker->check_inode_max_size(in); | |
33c7a0ef | 6380 | if (!(++count % mds->heartbeat_reset_grace())) |
b3b6e05e | 6381 | mds->heartbeat_reset(); |
7c673cae FG |
6382 | } |
6383 | rejoin_check_q.clear(); | |
6384 | for (CInode *in : rejoin_recover_q) { | |
6385 | mds->locker->file_recover(&in->filelock); | |
33c7a0ef | 6386 | if (!(++count % mds->heartbeat_reset_grace())) |
b3b6e05e | 6387 | mds->heartbeat_reset(); |
7c673cae FG |
6388 | } |
6389 | if (!rejoin_recover_q.empty()) { | |
6390 | rejoin_recover_q.clear(); | |
6391 | do_file_recover(); | |
6392 | } | |
6393 | } | |
6394 | ||
6395 | void MDCache::do_file_recover() | |
6396 | { | |
6397 | recovery_queue.advance(); | |
6398 | } | |
6399 | ||
6400 | // =============================================================================== | |
6401 | ||
6402 | ||
6403 | // ---------------------------- | |
6404 | // truncate | |
6405 | ||
6406 | class C_MDC_RetryTruncate : public MDCacheContext { | |
6407 | CInode *in; | |
6408 | LogSegment *ls; | |
6409 | public: | |
6410 | C_MDC_RetryTruncate(MDCache *c, CInode *i, LogSegment *l) : | |
6411 | MDCacheContext(c), in(i), ls(l) {} | |
6412 | void finish(int r) override { | |
6413 | mdcache->_truncate_inode(in, ls); | |
6414 | } | |
6415 | }; | |
6416 | ||
6417 | void MDCache::truncate_inode(CInode *in, LogSegment *ls) | |
6418 | { | |
f67539c2 | 6419 | const auto& pi = in->get_projected_inode(); |
7c673cae FG |
6420 | dout(10) << "truncate_inode " |
6421 | << pi->truncate_from << " -> " << pi->truncate_size | |
6422 | << " on " << *in | |
6423 | << dendl; | |
6424 | ||
6425 | ls->truncating_inodes.insert(in); | |
6426 | in->get(CInode::PIN_TRUNCATING); | |
6427 | in->auth_pin(this); | |
6428 | ||
6429 | if (!in->client_need_snapflush.empty() && | |
6430 | (in->get_caps_issued() & CEPH_CAP_FILE_BUFFER)) { | |
11fdf7f2 | 6431 | ceph_assert(in->filelock.is_xlocked()); |
7c673cae FG |
6432 | in->filelock.set_xlock_snap_sync(new C_MDC_RetryTruncate(this, in, ls)); |
6433 | mds->locker->issue_caps(in); | |
6434 | return; | |
6435 | } | |
6436 | ||
6437 | _truncate_inode(in, ls); | |
6438 | } | |
6439 | ||
6440 | struct C_IO_MDC_TruncateFinish : public MDCacheIOContext { | |
6441 | CInode *in; | |
6442 | LogSegment *ls; | |
6443 | C_IO_MDC_TruncateFinish(MDCache *c, CInode *i, LogSegment *l) : | |
91327a77 AA |
6444 | MDCacheIOContext(c, false), in(i), ls(l) { |
6445 | } | |
7c673cae | 6446 | void finish(int r) override { |
f67539c2 | 6447 | ceph_assert(r == 0 || r == -CEPHFS_ENOENT); |
7c673cae FG |
6448 | mdcache->truncate_inode_finish(in, ls); |
6449 | } | |
91327a77 AA |
6450 | void print(ostream& out) const override { |
6451 | out << "file_truncate(" << in->ino() << ")"; | |
6452 | } | |
7c673cae FG |
6453 | }; |
6454 | ||
6455 | void MDCache::_truncate_inode(CInode *in, LogSegment *ls) | |
6456 | { | |
f67539c2 | 6457 | const auto& pi = in->get_inode(); |
7c673cae FG |
6458 | dout(10) << "_truncate_inode " |
6459 | << pi->truncate_from << " -> " << pi->truncate_size | |
6460 | << " on " << *in << dendl; | |
6461 | ||
11fdf7f2 TL |
6462 | ceph_assert(pi->is_truncating()); |
6463 | ceph_assert(pi->truncate_size < (1ULL << 63)); | |
6464 | ceph_assert(pi->truncate_from < (1ULL << 63)); | |
6465 | ceph_assert(pi->truncate_size < pi->truncate_from); | |
7c673cae FG |
6466 | |
6467 | ||
6468 | SnapRealm *realm = in->find_snaprealm(); | |
6469 | SnapContext nullsnap; | |
6470 | const SnapContext *snapc; | |
6471 | if (realm) { | |
6472 | dout(10) << " realm " << *realm << dendl; | |
6473 | snapc = &realm->get_snap_context(); | |
6474 | } else { | |
6475 | dout(10) << " NO realm, using null context" << dendl; | |
6476 | snapc = &nullsnap; | |
11fdf7f2 | 6477 | ceph_assert(in->last == CEPH_NOSNAP); |
7c673cae FG |
6478 | } |
6479 | dout(10) << "_truncate_inode snapc " << snapc << " on " << *in << dendl; | |
f67539c2 TL |
6480 | auto layout = pi->layout; |
6481 | filer.truncate(in->ino(), &layout, *snapc, | |
7c673cae FG |
6482 | pi->truncate_size, pi->truncate_from-pi->truncate_size, |
6483 | pi->truncate_seq, ceph::real_time::min(), 0, | |
6484 | new C_OnFinisher(new C_IO_MDC_TruncateFinish(this, in, ls), | |
6485 | mds->finisher)); | |
6486 | } | |
6487 | ||
6488 | struct C_MDC_TruncateLogged : public MDCacheLogContext { | |
6489 | CInode *in; | |
6490 | MutationRef mut; | |
6491 | C_MDC_TruncateLogged(MDCache *m, CInode *i, MutationRef& mu) : | |
6492 | MDCacheLogContext(m), in(i), mut(mu) {} | |
6493 | void finish(int r) override { | |
6494 | mdcache->truncate_inode_logged(in, mut); | |
6495 | } | |
6496 | }; | |
6497 | ||
6498 | void MDCache::truncate_inode_finish(CInode *in, LogSegment *ls) | |
6499 | { | |
6500 | dout(10) << "truncate_inode_finish " << *in << dendl; | |
6501 | ||
6502 | set<CInode*>::iterator p = ls->truncating_inodes.find(in); | |
11fdf7f2 | 6503 | ceph_assert(p != ls->truncating_inodes.end()); |
7c673cae FG |
6504 | ls->truncating_inodes.erase(p); |
6505 | ||
7c673cae FG |
6506 | MutationRef mut(new MutationImpl()); |
6507 | mut->ls = mds->mdlog->get_current_segment(); | |
f67539c2 TL |
6508 | |
6509 | // update | |
6510 | auto pi = in->project_inode(mut); | |
6511 | pi.inode->version = in->pre_dirty(); | |
6512 | pi.inode->truncate_from = 0; | |
6513 | pi.inode->truncate_pending--; | |
7c673cae FG |
6514 | |
6515 | EUpdate *le = new EUpdate(mds->mdlog, "truncate finish"); | |
6516 | mds->mdlog->start_entry(le); | |
7c673cae | 6517 | |
f67539c2 | 6518 | predirty_journal_parents(mut, &le->metablob, in, 0, PREDIRTY_PRIMARY); |
7c673cae | 6519 | journal_dirty_inode(mut.get(), &le->metablob, in); |
f67539c2 | 6520 | le->metablob.add_truncate_finish(in->ino(), ls->seq); |
7c673cae FG |
6521 | mds->mdlog->submit_entry(le, new C_MDC_TruncateLogged(this, in, mut)); |
6522 | ||
6523 | // flush immediately if there are readers/writers waiting | |
6524 | if (in->is_waiter_for(CInode::WAIT_TRUNC) || | |
6525 | (in->get_caps_wanted() & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR))) | |
6526 | mds->mdlog->flush(); | |
6527 | } | |
6528 | ||
6529 | void MDCache::truncate_inode_logged(CInode *in, MutationRef& mut) | |
6530 | { | |
6531 | dout(10) << "truncate_inode_logged " << *in << dendl; | |
6532 | mut->apply(); | |
6533 | mds->locker->drop_locks(mut.get()); | |
6534 | mut->cleanup(); | |
6535 | ||
6536 | in->put(CInode::PIN_TRUNCATING); | |
6537 | in->auth_unpin(this); | |
6538 | ||
11fdf7f2 | 6539 | MDSContext::vec waiters; |
7c673cae FG |
6540 | in->take_waiting(CInode::WAIT_TRUNC, waiters); |
6541 | mds->queue_waiters(waiters); | |
6542 | } | |
6543 | ||
6544 | ||
6545 | void MDCache::add_recovered_truncate(CInode *in, LogSegment *ls) | |
6546 | { | |
6547 | dout(20) << "add_recovered_truncate " << *in << " in log segment " | |
6548 | << ls->seq << "/" << ls->offset << dendl; | |
6549 | ls->truncating_inodes.insert(in); | |
6550 | in->get(CInode::PIN_TRUNCATING); | |
6551 | } | |
6552 | ||
6553 | void MDCache::remove_recovered_truncate(CInode *in, LogSegment *ls) | |
6554 | { | |
6555 | dout(20) << "remove_recovered_truncate " << *in << " in log segment " | |
6556 | << ls->seq << "/" << ls->offset << dendl; | |
6557 | // if we have the logseg the truncate started in, it must be in our list. | |
6558 | set<CInode*>::iterator p = ls->truncating_inodes.find(in); | |
11fdf7f2 | 6559 | ceph_assert(p != ls->truncating_inodes.end()); |
7c673cae FG |
6560 | ls->truncating_inodes.erase(p); |
6561 | in->put(CInode::PIN_TRUNCATING); | |
6562 | } | |
6563 | ||
6564 | void MDCache::start_recovered_truncates() | |
6565 | { | |
6566 | dout(10) << "start_recovered_truncates" << dendl; | |
6567 | for (map<uint64_t,LogSegment*>::iterator p = mds->mdlog->segments.begin(); | |
6568 | p != mds->mdlog->segments.end(); | |
6569 | ++p) { | |
6570 | LogSegment *ls = p->second; | |
6571 | for (set<CInode*>::iterator q = ls->truncating_inodes.begin(); | |
6572 | q != ls->truncating_inodes.end(); | |
6573 | ++q) { | |
6574 | CInode *in = *q; | |
6575 | in->auth_pin(this); | |
6576 | ||
6577 | if (!in->client_need_snapflush.empty() && | |
6578 | (in->get_caps_issued() & CEPH_CAP_FILE_BUFFER)) { | |
11fdf7f2 | 6579 | ceph_assert(in->filelock.is_stable()); |
7c673cae FG |
6580 | in->filelock.set_state(LOCK_XLOCKDONE); |
6581 | in->auth_pin(&in->filelock); | |
6582 | in->filelock.set_xlock_snap_sync(new C_MDC_RetryTruncate(this, in, ls)); | |
6583 | // start_files_to_recover will revoke caps | |
6584 | continue; | |
6585 | } | |
6586 | _truncate_inode(in, ls); | |
6587 | } | |
6588 | } | |
6589 | } | |
6590 | ||
6591 | ||
9f95a23c TL |
6592 | class C_MDS_purge_completed_finish : public MDCacheLogContext { |
6593 | interval_set<inodeno_t> inos; | |
9f95a23c | 6594 | LogSegment *ls; |
f67539c2 | 6595 | version_t inotablev; |
9f95a23c | 6596 | public: |
f67539c2 TL |
6597 | C_MDS_purge_completed_finish(MDCache *m, const interval_set<inodeno_t>& _inos, |
6598 | LogSegment *_ls, version_t iv) | |
6599 | : MDCacheLogContext(m), inos(_inos), ls(_ls), inotablev(iv) {} | |
9f95a23c | 6600 | void finish(int r) override { |
20effc67 | 6601 | ceph_assert(r == 0); |
9f95a23c | 6602 | if (inotablev) { |
f67539c2 | 6603 | get_mds()->inotable->apply_release_ids(inos); |
20effc67 | 6604 | ceph_assert(get_mds()->inotable->get_version() == inotablev); |
9f95a23c | 6605 | } |
f67539c2 | 6606 | ls->purge_inodes_finish(inos); |
9f95a23c TL |
6607 | } |
6608 | }; | |
7c673cae | 6609 | |
9f95a23c TL |
6610 | void MDCache::start_purge_inodes(){ |
6611 | dout(10) << "start_purge_inodes" << dendl; | |
6612 | for (auto& p : mds->mdlog->segments){ | |
6613 | LogSegment *ls = p.second; | |
f67539c2 TL |
6614 | if (ls->purging_inodes.size()){ |
6615 | purge_inodes(ls->purging_inodes, ls); | |
9f95a23c TL |
6616 | } |
6617 | } | |
6618 | } | |
7c673cae | 6619 | |
9f95a23c TL |
6620 | void MDCache::purge_inodes(const interval_set<inodeno_t>& inos, LogSegment *ls) |
6621 | { | |
f67539c2 TL |
6622 | dout(10) << __func__ << " purging inos " << inos << " logseg " << ls->seq << dendl; |
6623 | // FIXME: handle non-default data pool and namespace | |
6624 | ||
9f95a23c | 6625 | auto cb = new LambdaContext([this, inos, ls](int r){ |
20effc67 | 6626 | ceph_assert(r == 0 || r == -2); |
9f95a23c TL |
6627 | mds->inotable->project_release_ids(inos); |
6628 | version_t piv = mds->inotable->get_projected_version(); | |
20effc67 | 6629 | ceph_assert(piv != 0); |
f67539c2 TL |
6630 | mds->mdlog->start_submit_entry(new EPurged(inos, ls->seq, piv), |
6631 | new C_MDS_purge_completed_finish(this, inos, ls, piv)); | |
9f95a23c TL |
6632 | mds->mdlog->flush(); |
6633 | }); | |
6634 | ||
9f95a23c | 6635 | C_GatherBuilder gather(g_ceph_context, |
f67539c2 | 6636 | new C_OnFinisher(new MDSIOContextWrapper(mds, cb), mds->finisher)); |
9f95a23c | 6637 | SnapContext nullsnapc; |
f67539c2 TL |
6638 | for (const auto& [start, len] : inos) { |
6639 | for (auto i = start; i < start + len ; i += 1) { | |
6640 | filer.purge_range(i, &default_file_layout, nullsnapc, 0, 1, | |
6641 | ceph::real_clock::now(), 0, gather.new_sub()); | |
9f95a23c TL |
6642 | } |
6643 | } | |
6644 | gather.activate(); | |
6645 | } | |
7c673cae FG |
6646 | |
6647 | // ================================================================================ | |
6648 | // cache trimming | |
6649 | ||
11fdf7f2 | 6650 | std::pair<bool, uint64_t> MDCache::trim_lru(uint64_t count, expiremap& expiremap) |
181888fb | 6651 | { |
7c673cae | 6652 | bool is_standby_replay = mds->is_standby_replay(); |
181888fb FG |
6653 | std::vector<CDentry *> unexpirables; |
6654 | uint64_t trimmed = 0; | |
6655 | ||
11fdf7f2 | 6656 | auto trim_threshold = g_conf().get_val<Option::size_t>("mds_cache_trim_threshold"); |
a8e16298 | 6657 | |
181888fb FG |
6658 | dout(7) << "trim_lru trimming " << count |
6659 | << " items from LRU" | |
6660 | << " size=" << lru.lru_get_size() | |
6661 | << " mid=" << lru.lru_get_top() | |
6662 | << " pintail=" << lru.lru_get_pintail() | |
6663 | << " pinned=" << lru.lru_get_num_pinned() | |
6664 | << dendl; | |
7c673cae | 6665 | |
11fdf7f2 | 6666 | const uint64_t trim_counter_start = trim_counter.get(); |
a8e16298 TL |
6667 | bool throttled = false; |
6668 | while (1) { | |
6669 | throttled |= trim_counter_start+trimmed >= trim_threshold; | |
6670 | if (throttled) break; | |
31f18b77 FG |
6671 | CDentry *dn = static_cast<CDentry*>(bottom_lru.lru_expire()); |
6672 | if (!dn) | |
6673 | break; | |
6674 | if (trim_dentry(dn, expiremap)) { | |
6675 | unexpirables.push_back(dn); | |
181888fb FG |
6676 | } else { |
6677 | trimmed++; | |
31f18b77 FG |
6678 | } |
6679 | } | |
6680 | ||
181888fb | 6681 | for (auto &dn : unexpirables) { |
31f18b77 | 6682 | bottom_lru.lru_insert_mid(dn); |
181888fb | 6683 | } |
31f18b77 FG |
6684 | unexpirables.clear(); |
6685 | ||
181888fb | 6686 | // trim dentries from the LRU until count is reached |
b3b6e05e | 6687 | // if mds is in standby_replay and skip trimming the inodes |
494da23a | 6688 | while (!throttled && (cache_toofull() || count > 0 || is_standby_replay)) { |
a8e16298 TL |
6689 | throttled |= trim_counter_start+trimmed >= trim_threshold; |
6690 | if (throttled) break; | |
7c673cae FG |
6691 | CDentry *dn = static_cast<CDentry*>(lru.lru_expire()); |
6692 | if (!dn) { | |
6693 | break; | |
6694 | } | |
b3b6e05e | 6695 | if (is_standby_replay && dn->get_linkage()->inode) { |
494da23a TL |
6696 | // we move the inodes that need to be trimmed to the end of the lru queue. |
6697 | // refer to MDCache::standby_trim_segment | |
6698 | lru.lru_insert_bot(dn); | |
6699 | break; | |
181888fb FG |
6700 | } else if (trim_dentry(dn, expiremap)) { |
6701 | unexpirables.push_back(dn); | |
6702 | } else { | |
6703 | trimmed++; | |
3efd9988 | 6704 | if (count > 0) count--; |
7c673cae FG |
6705 | } |
6706 | } | |
11fdf7f2 | 6707 | trim_counter.hit(trimmed); |
181888fb FG |
6708 | |
6709 | for (auto &dn : unexpirables) { | |
31f18b77 | 6710 | lru.lru_insert_mid(dn); |
181888fb | 6711 | } |
31f18b77 | 6712 | unexpirables.clear(); |
7c673cae | 6713 | |
181888fb | 6714 | dout(7) << "trim_lru trimmed " << trimmed << " items" << dendl; |
a8e16298 | 6715 | return std::pair<bool, uint64_t>(throttled, trimmed); |
181888fb FG |
6716 | } |
6717 | ||
6718 | /* | |
6719 | * note: only called while MDS is active or stopping... NOT during recovery. | |
6720 | * however, we may expire a replica whose authority is recovering. | |
6721 | * | |
6722 | * @param count is number of dentries to try to expire | |
6723 | */ | |
a8e16298 | 6724 | std::pair<bool, uint64_t> MDCache::trim(uint64_t count) |
181888fb FG |
6725 | { |
6726 | uint64_t used = cache_size(); | |
91327a77 | 6727 | uint64_t limit = cache_memory_limit; |
11fdf7f2 | 6728 | expiremap expiremap; |
181888fb FG |
6729 | |
6730 | dout(7) << "trim bytes_used=" << bytes2str(used) | |
6731 | << " limit=" << bytes2str(limit) | |
91327a77 | 6732 | << " reservation=" << cache_reservation |
181888fb FG |
6733 | << "% count=" << count << dendl; |
6734 | ||
6735 | // process delayed eval_stray() | |
6736 | stray_manager.advance_delayed(); | |
6737 | ||
a8e16298 TL |
6738 | auto result = trim_lru(count, expiremap); |
6739 | auto& trimmed = result.second; | |
181888fb | 6740 | |
7c673cae | 6741 | // trim non-auth, non-bound subtrees |
181888fb | 6742 | for (auto p = subtrees.begin(); p != subtrees.end();) { |
7c673cae FG |
6743 | CDir *dir = p->first; |
6744 | ++p; | |
31f18b77 FG |
6745 | CInode *diri = dir->get_inode(); |
6746 | if (dir->is_auth()) { | |
f6b5b4d7 TL |
6747 | if (diri->is_auth() && !diri->is_base()) { |
6748 | /* this situation should correspond to an export pin */ | |
6749 | if (dir->get_num_head_items() == 0 && dir->get_num_ref() == 1) { | |
6750 | /* pinned empty subtree, try to drop */ | |
6751 | if (dir->state_test(CDir::STATE_AUXSUBTREE)) { | |
6752 | dout(20) << "trimming empty pinned subtree " << *dir << dendl; | |
6753 | dir->state_clear(CDir::STATE_AUXSUBTREE); | |
6754 | remove_subtree(dir); | |
6755 | diri->close_dirfrag(dir->dirfrag().frag); | |
6756 | } | |
6757 | } | |
6758 | } else if (!diri->is_auth() && !diri->is_base() && dir->get_num_head_items() == 0) { | |
6759 | if (dir->state_test(CDir::STATE_EXPORTING) || | |
6760 | !(mds->is_active() || mds->is_stopping()) || | |
6761 | dir->is_freezing() || dir->is_frozen()) | |
6762 | continue; | |
31f18b77 | 6763 | |
f6b5b4d7 | 6764 | migrator->export_empty_import(dir); |
a8e16298 | 6765 | ++trimmed; |
31f18b77 | 6766 | } |
f6b5b4d7 TL |
6767 | } else if (!diri->is_auth() && dir->get_num_ref() <= 1) { |
6768 | // only subtree pin | |
f91f0fd5 | 6769 | if (diri->get_num_ref() > diri->get_num_subtree_roots()) { |
f6b5b4d7 | 6770 | continue; |
f91f0fd5 | 6771 | } |
31f18b77 | 6772 | |
f6b5b4d7 TL |
6773 | // don't trim subtree root if its auth MDS is recovering. |
6774 | // This simplify the cache rejoin code. | |
6775 | if (dir->is_subtree_root() && rejoin_ack_gather.count(dir->get_dir_auth().first)) | |
6776 | continue; | |
6777 | trim_dirfrag(dir, 0, expiremap); | |
6778 | ++trimmed; | |
7c673cae FG |
6779 | } |
6780 | } | |
6781 | ||
6782 | // trim root? | |
181888fb | 6783 | if (mds->is_stopping() && root) { |
9f95a23c TL |
6784 | auto&& ls = root->get_dirfrags(); |
6785 | for (const auto& dir : ls) { | |
a8e16298 | 6786 | if (dir->get_num_ref() == 1) { // subtree pin |
7c673cae | 6787 | trim_dirfrag(dir, 0, expiremap); |
a8e16298 TL |
6788 | ++trimmed; |
6789 | } | |
7c673cae | 6790 | } |
a8e16298 | 6791 | if (root->get_num_ref() == 0) { |
7c673cae | 6792 | trim_inode(0, root, 0, expiremap); |
a8e16298 TL |
6793 | ++trimmed; |
6794 | } | |
7c673cae FG |
6795 | } |
6796 | ||
6797 | std::set<mds_rank_t> stopping; | |
6798 | mds->mdsmap->get_mds_set(stopping, MDSMap::STATE_STOPPING); | |
6799 | stopping.erase(mds->get_nodeid()); | |
6800 | for (auto rank : stopping) { | |
6801 | CInode* mdsdir_in = get_inode(MDS_INO_MDSDIR(rank)); | |
6802 | if (!mdsdir_in) | |
6803 | continue; | |
6804 | ||
11fdf7f2 TL |
6805 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(rank), std::forward_as_tuple()); |
6806 | if (em.second) { | |
9f95a23c | 6807 | em.first->second = make_message<MCacheExpire>(mds->get_nodeid()); |
7c673cae FG |
6808 | } |
6809 | ||
20effc67 | 6810 | dout(20) << __func__ << ": try expiring " << *mdsdir_in << " for stopping mds." << mds->get_nodeid() << dendl; |
7c673cae FG |
6811 | |
6812 | const bool aborted = expire_recursive(mdsdir_in, expiremap); | |
6813 | if (!aborted) { | |
6814 | dout(20) << __func__ << ": successfully expired mdsdir" << dendl; | |
9f95a23c | 6815 | auto&& ls = mdsdir_in->get_dirfrags(); |
7c673cae | 6816 | for (auto dir : ls) { |
a8e16298 | 6817 | if (dir->get_num_ref() == 1) { // subtree pin |
7c673cae | 6818 | trim_dirfrag(dir, dir, expiremap); |
a8e16298 TL |
6819 | ++trimmed; |
6820 | } | |
7c673cae | 6821 | } |
a8e16298 | 6822 | if (mdsdir_in->get_num_ref() == 0) { |
7c673cae | 6823 | trim_inode(NULL, mdsdir_in, NULL, expiremap); |
a8e16298 TL |
6824 | ++trimmed; |
6825 | } | |
7c673cae FG |
6826 | } else { |
6827 | dout(20) << __func__ << ": some unexpirable contents in mdsdir" << dendl; | |
6828 | } | |
6829 | } | |
6830 | ||
6831 | // Other rank's base inodes (when I'm stopping) | |
181888fb | 6832 | if (mds->is_stopping()) { |
7c673cae | 6833 | for (set<CInode*>::iterator p = base_inodes.begin(); |
11fdf7f2 TL |
6834 | p != base_inodes.end();) { |
6835 | CInode *base_in = *p; | |
6836 | ++p; | |
6837 | if (MDS_INO_IS_MDSDIR(base_in->ino()) && | |
6838 | MDS_INO_MDSDIR_OWNER(base_in->ino()) != mds->get_nodeid()) { | |
6839 | dout(20) << __func__ << ": maybe trimming base: " << *base_in << dendl; | |
6840 | if (base_in->get_num_ref() == 0) { | |
6841 | trim_inode(NULL, base_in, NULL, expiremap); | |
a8e16298 | 6842 | ++trimmed; |
7c673cae FG |
6843 | } |
6844 | } | |
6845 | } | |
6846 | } | |
6847 | ||
6848 | // send any expire messages | |
6849 | send_expire_messages(expiremap); | |
6850 | ||
a8e16298 | 6851 | return result; |
7c673cae FG |
6852 | } |
6853 | ||
11fdf7f2 | 6854 | void MDCache::send_expire_messages(expiremap& expiremap) |
7c673cae FG |
6855 | { |
6856 | // send expires | |
11fdf7f2 | 6857 | for (const auto &p : expiremap) { |
7c673cae | 6858 | if (mds->is_cluster_degraded() && |
11fdf7f2 TL |
6859 | (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || |
6860 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
6861 | rejoin_sent.count(p.first) == 0))) { | |
7c673cae FG |
6862 | continue; |
6863 | } | |
11fdf7f2 TL |
6864 | dout(7) << "sending cache_expire to " << p.first << dendl; |
6865 | mds->send_message_mds(p.second, p.first); | |
7c673cae | 6866 | } |
11fdf7f2 | 6867 | expiremap.clear(); |
7c673cae FG |
6868 | } |
6869 | ||
6870 | ||
11fdf7f2 | 6871 | bool MDCache::trim_dentry(CDentry *dn, expiremap& expiremap) |
7c673cae FG |
6872 | { |
6873 | dout(12) << "trim_dentry " << *dn << dendl; | |
6874 | ||
6875 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
6876 | ||
6877 | CDir *dir = dn->get_dir(); | |
11fdf7f2 | 6878 | ceph_assert(dir); |
7c673cae FG |
6879 | |
6880 | CDir *con = get_subtree_root(dir); | |
6881 | if (con) | |
6882 | dout(12) << " in container " << *con << dendl; | |
6883 | else { | |
6884 | dout(12) << " no container; under a not-yet-linked dir" << dendl; | |
11fdf7f2 | 6885 | ceph_assert(dn->is_auth()); |
7c673cae FG |
6886 | } |
6887 | ||
6888 | // If replica dentry is not readable, it's likely we will receive | |
6889 | // MDentryLink/MDentryUnlink message soon (It's possible we first | |
6890 | // receive a MDentryUnlink message, then MDentryLink message) | |
6891 | // MDentryLink message only replicates an inode, so we should | |
6892 | // avoid trimming the inode's parent dentry. This is because that | |
6893 | // unconnected replicas are problematic for subtree migration. | |
6894 | if (!dn->is_auth() && !dn->lock.can_read(-1) && | |
6895 | !dn->get_dir()->get_inode()->is_stray()) | |
6896 | return true; | |
6897 | ||
6898 | // adjust the dir state | |
6899 | // NOTE: we can safely remove a clean, null dentry without effecting | |
6900 | // directory completeness. | |
6901 | // (check this _before_ we unlink the inode, below!) | |
6902 | bool clear_complete = false; | |
6903 | if (!(dnl->is_null() && dn->is_clean())) | |
6904 | clear_complete = true; | |
6905 | ||
6906 | // unlink the dentry | |
6907 | if (dnl->is_remote()) { | |
6908 | // just unlink. | |
31f18b77 | 6909 | dir->unlink_inode(dn, false); |
7c673cae FG |
6910 | } else if (dnl->is_primary()) { |
6911 | // expire the inode, too. | |
6912 | CInode *in = dnl->get_inode(); | |
11fdf7f2 | 6913 | ceph_assert(in); |
7c673cae FG |
6914 | if (trim_inode(dn, in, con, expiremap)) |
6915 | return true; // purging stray instead of trimming | |
6916 | } else { | |
11fdf7f2 | 6917 | ceph_assert(dnl->is_null()); |
7c673cae FG |
6918 | } |
6919 | ||
6920 | if (!dn->is_auth()) { | |
6921 | // notify dentry authority. | |
6922 | mds_authority_t auth = dn->authority(); | |
6923 | ||
6924 | for (int p=0; p<2; p++) { | |
6925 | mds_rank_t a = auth.first; | |
6926 | if (p) a = auth.second; | |
6927 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6928 | if (mds->get_nodeid() == auth.second && | |
6929 | con->is_importing()) break; // don't send any expire while importing. | |
6930 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6931 | ||
6932 | dout(12) << " sending expire to mds." << a << " on " << *dn << dendl; | |
11fdf7f2 TL |
6933 | ceph_assert(a != mds->get_nodeid()); |
6934 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(a), std::forward_as_tuple()); | |
6935 | if (em.second) | |
9f95a23c | 6936 | em.first->second = make_message<MCacheExpire>(mds->get_nodeid()); |
11fdf7f2 | 6937 | em.first->second->add_dentry(con->dirfrag(), dir->dirfrag(), dn->get_name(), dn->last, dn->get_replica_nonce()); |
7c673cae FG |
6938 | } |
6939 | } | |
6940 | ||
6941 | // remove dentry | |
6942 | if (dn->last == CEPH_NOSNAP && dir->is_auth()) | |
6943 | dir->add_to_bloom(dn); | |
6944 | dir->remove_dentry(dn); | |
6945 | ||
6946 | if (clear_complete) | |
6947 | dir->state_clear(CDir::STATE_COMPLETE); | |
6948 | ||
7c673cae FG |
6949 | if (mds->logger) mds->logger->inc(l_mds_inodes_expired); |
6950 | return false; | |
6951 | } | |
6952 | ||
6953 | ||
11fdf7f2 | 6954 | void MDCache::trim_dirfrag(CDir *dir, CDir *con, expiremap& expiremap) |
7c673cae FG |
6955 | { |
6956 | dout(15) << "trim_dirfrag " << *dir << dendl; | |
6957 | ||
6958 | if (dir->is_subtree_root()) { | |
11fdf7f2 | 6959 | ceph_assert(!dir->is_auth() || |
7c673cae FG |
6960 | (!dir->is_replicated() && dir->inode->is_base())); |
6961 | remove_subtree(dir); // remove from subtree map | |
6962 | } | |
11fdf7f2 | 6963 | ceph_assert(dir->get_num_ref() == 0); |
7c673cae FG |
6964 | |
6965 | CInode *in = dir->get_inode(); | |
6966 | ||
6967 | if (!dir->is_auth()) { | |
6968 | mds_authority_t auth = dir->authority(); | |
6969 | ||
6970 | // was this an auth delegation? (if so, slightly modified container) | |
6971 | dirfrag_t condf; | |
6972 | if (dir->is_subtree_root()) { | |
6973 | dout(12) << " subtree root, container is " << *dir << dendl; | |
6974 | con = dir; | |
6975 | condf = dir->dirfrag(); | |
6976 | } else { | |
6977 | condf = con->dirfrag(); | |
6978 | } | |
6979 | ||
6980 | for (int p=0; p<2; p++) { | |
6981 | mds_rank_t a = auth.first; | |
6982 | if (p) a = auth.second; | |
6983 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6984 | if (mds->get_nodeid() == auth.second && | |
6985 | con->is_importing()) break; // don't send any expire while importing. | |
6986 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6987 | ||
6988 | dout(12) << " sending expire to mds." << a << " on " << *dir << dendl; | |
11fdf7f2 TL |
6989 | ceph_assert(a != mds->get_nodeid()); |
6990 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(a), std::forward_as_tuple()); | |
6991 | if (em.second) | |
9f95a23c | 6992 | em.first->second = make_message<MCacheExpire>(mds->get_nodeid()); /* new */ |
11fdf7f2 | 6993 | em.first->second->add_dir(condf, dir->dirfrag(), dir->replica_nonce); |
7c673cae FG |
6994 | } |
6995 | } | |
6996 | ||
6997 | in->close_dirfrag(dir->dirfrag().frag); | |
6998 | } | |
6999 | ||
7000 | /** | |
7001 | * Try trimming an inode from the cache | |
7002 | * | |
7003 | * @return true if the inode is still in cache, else false if it was trimmed | |
7004 | */ | |
11fdf7f2 | 7005 | bool MDCache::trim_inode(CDentry *dn, CInode *in, CDir *con, expiremap& expiremap) |
7c673cae FG |
7006 | { |
7007 | dout(15) << "trim_inode " << *in << dendl; | |
11fdf7f2 | 7008 | ceph_assert(in->get_num_ref() == 0); |
7c673cae FG |
7009 | |
7010 | if (in->is_dir()) { | |
7011 | // If replica inode's dirfragtreelock is not readable, it's likely | |
7012 | // some dirfrags of the inode are being fragmented and we will receive | |
7013 | // MMDSFragmentNotify soon. MMDSFragmentNotify only replicates the new | |
7014 | // dirfrags, so we should avoid trimming these dirfrags' parent inode. | |
7015 | // This is because that unconnected replicas are problematic for | |
7016 | // subtree migration. | |
7017 | // | |
9f95a23c | 7018 | if (!in->is_auth() && !mds->locker->rdlock_try(&in->dirfragtreelock, -1)) { |
7c673cae | 7019 | return true; |
28e407b8 | 7020 | } |
7c673cae FG |
7021 | |
7022 | // DIR | |
9f95a23c TL |
7023 | auto&& dfls = in->get_dirfrags(); |
7024 | for (const auto& dir : dfls) { | |
11fdf7f2 | 7025 | ceph_assert(!dir->is_subtree_root()); |
7c673cae FG |
7026 | trim_dirfrag(dir, con ? con:dir, expiremap); // if no container (e.g. root dirfrag), use *p |
7027 | } | |
7028 | } | |
7029 | ||
7030 | // INODE | |
7031 | if (in->is_auth()) { | |
7032 | // eval stray after closing dirfrags | |
7033 | if (dn && !dn->state_test(CDentry::STATE_PURGING)) { | |
7034 | maybe_eval_stray(in); | |
7035 | if (dn->state_test(CDentry::STATE_PURGING) || dn->get_num_ref() > 0) | |
7036 | return true; | |
7037 | } | |
7038 | } else { | |
7039 | mds_authority_t auth = in->authority(); | |
7040 | ||
7041 | dirfrag_t df; | |
7042 | if (con) | |
7043 | df = con->dirfrag(); | |
7044 | else | |
7045 | df = dirfrag_t(0,frag_t()); // must be a root or stray inode. | |
7046 | ||
7047 | for (int p=0; p<2; p++) { | |
7048 | mds_rank_t a = auth.first; | |
7049 | if (p) a = auth.second; | |
7050 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
7051 | if (con && mds->get_nodeid() == auth.second && | |
7052 | con->is_importing()) break; // don't send any expire while importing. | |
7053 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
7054 | ||
7055 | dout(12) << " sending expire to mds." << a << " on " << *in << dendl; | |
11fdf7f2 TL |
7056 | ceph_assert(a != mds->get_nodeid()); |
7057 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(a), std::forward_as_tuple()); | |
7058 | if (em.second) | |
9f95a23c | 7059 | em.first->second = make_message<MCacheExpire>(mds->get_nodeid()); /* new */ |
11fdf7f2 | 7060 | em.first->second->add_inode(df, in->vino(), in->get_replica_nonce()); |
7c673cae FG |
7061 | } |
7062 | } | |
7063 | ||
7064 | /* | |
7065 | if (in->is_auth()) { | |
7066 | if (in->hack_accessed) | |
7067 | mds->logger->inc("outt"); | |
7068 | else { | |
7069 | mds->logger->inc("outut"); | |
7070 | mds->logger->fset("oututl", ceph_clock_now() - in->hack_load_stamp); | |
7071 | } | |
7072 | } | |
7073 | */ | |
7074 | ||
7075 | // unlink | |
7076 | if (dn) | |
31f18b77 | 7077 | dn->get_dir()->unlink_inode(dn, false); |
7c673cae FG |
7078 | remove_inode(in); |
7079 | return false; | |
7080 | } | |
7081 | ||
7082 | ||
7083 | /** | |
7084 | * trim_non_auth - remove any non-auth items from our cache | |
7085 | * | |
7086 | * this reduces the amount of non-auth metadata in our cache, reducing the | |
7087 | * load incurred by the rejoin phase. | |
7088 | * | |
7089 | * the only non-auth items that remain are those that are needed to | |
7090 | * attach our own subtrees to the root. | |
7091 | * | |
7092 | * when we are done, all dentries will be in the top bit of the lru. | |
7093 | * | |
7094 | * why we have to do this: | |
7095 | * we may not have accurate linkage for non-auth items. which means we will | |
7096 | * know which subtree it falls into, and can not be sure to declare it to the | |
7097 | * correct authority. | |
7098 | */ | |
7099 | void MDCache::trim_non_auth() | |
7100 | { | |
7101 | dout(7) << "trim_non_auth" << dendl; | |
7102 | ||
7103 | // temporarily pin all subtree roots | |
7104 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
7105 | p != subtrees.end(); | |
7106 | ++p) | |
7107 | p->first->get(CDir::PIN_SUBTREETEMP); | |
7108 | ||
31f18b77 | 7109 | list<CDentry*> auth_list; |
7c673cae FG |
7110 | |
7111 | // trim non-auth items from the lru | |
31f18b77 FG |
7112 | for (;;) { |
7113 | CDentry *dn = NULL; | |
7114 | if (bottom_lru.lru_get_size() > 0) | |
7115 | dn = static_cast<CDentry*>(bottom_lru.lru_expire()); | |
7116 | if (!dn && lru.lru_get_size() > 0) | |
7117 | dn = static_cast<CDentry*>(lru.lru_expire()); | |
7118 | if (!dn) | |
7119 | break; | |
7120 | ||
7c673cae FG |
7121 | CDentry::linkage_t *dnl = dn->get_linkage(); |
7122 | ||
7123 | if (dn->is_auth()) { | |
7124 | // add back into lru (at the top) | |
31f18b77 | 7125 | auth_list.push_back(dn); |
7c673cae FG |
7126 | |
7127 | if (dnl->is_remote() && dnl->get_inode() && !dnl->get_inode()->is_auth()) | |
7128 | dn->unlink_remote(dnl); | |
7c673cae FG |
7129 | } else { |
7130 | // non-auth. expire. | |
7131 | CDir *dir = dn->get_dir(); | |
11fdf7f2 | 7132 | ceph_assert(dir); |
7c673cae FG |
7133 | |
7134 | // unlink the dentry | |
7135 | dout(10) << " removing " << *dn << dendl; | |
7136 | if (dnl->is_remote()) { | |
31f18b77 | 7137 | dir->unlink_inode(dn, false); |
7c673cae FG |
7138 | } |
7139 | else if (dnl->is_primary()) { | |
7140 | CInode *in = dnl->get_inode(); | |
7141 | dout(10) << " removing " << *in << dendl; | |
9f95a23c TL |
7142 | auto&& ls = in->get_dirfrags(); |
7143 | for (const auto& subdir : ls) { | |
11fdf7f2 | 7144 | ceph_assert(!subdir->is_subtree_root()); |
7c673cae FG |
7145 | in->close_dirfrag(subdir->dirfrag().frag); |
7146 | } | |
31f18b77 | 7147 | dir->unlink_inode(dn, false); |
7c673cae FG |
7148 | remove_inode(in); |
7149 | } | |
7150 | else { | |
11fdf7f2 | 7151 | ceph_assert(dnl->is_null()); |
7c673cae FG |
7152 | } |
7153 | ||
11fdf7f2 | 7154 | ceph_assert(!dir->has_bloom()); |
7c673cae FG |
7155 | dir->remove_dentry(dn); |
7156 | // adjust the dir state | |
7157 | dir->state_clear(CDir::STATE_COMPLETE); // dir incomplete! | |
7158 | // close empty non-auth dirfrag | |
7159 | if (!dir->is_subtree_root() && dir->get_num_any() == 0) | |
7160 | dir->inode->close_dirfrag(dir->get_frag()); | |
7161 | } | |
7162 | } | |
7163 | ||
9f95a23c | 7164 | for (const auto& dn : auth_list) { |
31f18b77 FG |
7165 | if (dn->state_test(CDentry::STATE_BOTTOMLRU)) |
7166 | bottom_lru.lru_insert_mid(dn); | |
7167 | else | |
7168 | lru.lru_insert_top(dn); | |
7169 | } | |
7170 | ||
7c673cae FG |
7171 | // move everything in the pintail to the top bit of the lru. |
7172 | lru.lru_touch_entire_pintail(); | |
7173 | ||
7174 | // unpin all subtrees | |
7175 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
7176 | p != subtrees.end(); | |
7177 | ++p) | |
7178 | p->first->put(CDir::PIN_SUBTREETEMP); | |
7179 | ||
31f18b77 FG |
7180 | if (lru.lru_get_size() == 0 && |
7181 | bottom_lru.lru_get_size() == 0) { | |
7c673cae | 7182 | // root, stray, etc.? |
b32b8144 | 7183 | auto p = inode_map.begin(); |
7c673cae | 7184 | while (p != inode_map.end()) { |
7c673cae | 7185 | CInode *in = p->second; |
b32b8144 | 7186 | ++p; |
7c673cae | 7187 | if (!in->is_auth()) { |
9f95a23c TL |
7188 | auto&& ls = in->get_dirfrags(); |
7189 | for (const auto& dir : ls) { | |
7190 | dout(10) << " removing " << *dir << dendl; | |
7191 | ceph_assert(dir->get_num_ref() == 1); // SUBTREE | |
7192 | remove_subtree(dir); | |
7193 | in->close_dirfrag(dir->dirfrag().frag); | |
7c673cae FG |
7194 | } |
7195 | dout(10) << " removing " << *in << dendl; | |
11fdf7f2 TL |
7196 | ceph_assert(!in->get_parent_dn()); |
7197 | ceph_assert(in->get_num_ref() == 0); | |
7c673cae FG |
7198 | remove_inode(in); |
7199 | } | |
7c673cae FG |
7200 | } |
7201 | } | |
7202 | ||
7203 | show_subtrees(); | |
7204 | } | |
7205 | ||
7206 | /** | |
7207 | * Recursively trim the subtree rooted at directory to remove all | |
7208 | * CInodes/CDentrys/CDirs that aren't links to remote MDSes, or ancestors | |
7209 | * of those links. This is used to clear invalid data out of the cache. | |
7210 | * Note that it doesn't clear the passed-in directory, since that's not | |
7211 | * always safe. | |
7212 | */ | |
7213 | bool MDCache::trim_non_auth_subtree(CDir *dir) | |
7214 | { | |
7215 | dout(10) << "trim_non_auth_subtree(" << dir << ") " << *dir << dendl; | |
7216 | ||
7217 | bool keep_dir = !can_trim_non_auth_dirfrag(dir); | |
7218 | ||
94b18763 FG |
7219 | auto j = dir->begin(); |
7220 | auto i = j; | |
7c673cae FG |
7221 | while (j != dir->end()) { |
7222 | i = j++; | |
7223 | CDentry *dn = i->second; | |
7224 | dout(10) << "trim_non_auth_subtree(" << dir << ") Checking dentry " << dn << dendl; | |
7225 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
7226 | if (dnl->is_primary()) { // check for subdirectories, etc | |
7227 | CInode *in = dnl->get_inode(); | |
7228 | bool keep_inode = false; | |
7229 | if (in->is_dir()) { | |
9f95a23c TL |
7230 | auto&& subdirs = in->get_dirfrags(); |
7231 | for (const auto& subdir : subdirs) { | |
7232 | if (subdir->is_subtree_root()) { | |
7c673cae | 7233 | keep_inode = true; |
9f95a23c | 7234 | dout(10) << "trim_non_auth_subtree(" << dir << ") keeping " << *subdir << dendl; |
7c673cae | 7235 | } else { |
9f95a23c | 7236 | if (trim_non_auth_subtree(subdir)) |
7c673cae FG |
7237 | keep_inode = true; |
7238 | else { | |
9f95a23c | 7239 | in->close_dirfrag(subdir->get_frag()); |
7c673cae FG |
7240 | dir->state_clear(CDir::STATE_COMPLETE); // now incomplete! |
7241 | } | |
7242 | } | |
7243 | } | |
7244 | ||
7245 | } | |
7246 | if (!keep_inode) { // remove it! | |
7247 | dout(20) << "trim_non_auth_subtree(" << dir << ") removing inode " << in << " with dentry" << dn << dendl; | |
31f18b77 | 7248 | dir->unlink_inode(dn, false); |
7c673cae | 7249 | remove_inode(in); |
11fdf7f2 | 7250 | ceph_assert(!dir->has_bloom()); |
7c673cae FG |
7251 | dir->remove_dentry(dn); |
7252 | } else { | |
7253 | dout(20) << "trim_non_auth_subtree(" << dir << ") keeping inode " << in << " with dentry " << dn <<dendl; | |
7254 | dn->state_clear(CDentry::STATE_AUTH); | |
7255 | in->state_clear(CInode::STATE_AUTH); | |
7256 | } | |
f67539c2 | 7257 | } else if (keep_dir && dnl->is_null()) { // keep null dentry for peer rollback |
7c673cae FG |
7258 | dout(20) << "trim_non_auth_subtree(" << dir << ") keeping dentry " << dn <<dendl; |
7259 | } else { // just remove it | |
7260 | dout(20) << "trim_non_auth_subtree(" << dir << ") removing dentry " << dn << dendl; | |
7261 | if (dnl->is_remote()) | |
31f18b77 | 7262 | dir->unlink_inode(dn, false); |
7c673cae FG |
7263 | dir->remove_dentry(dn); |
7264 | } | |
7265 | } | |
7266 | dir->state_clear(CDir::STATE_AUTH); | |
7267 | /** | |
7268 | * We've now checked all our children and deleted those that need it. | |
7269 | * Now return to caller, and tell them if *we're* a keeper. | |
7270 | */ | |
7271 | return keep_dir || dir->get_num_any(); | |
7272 | } | |
7273 | ||
7274 | /* | |
7275 | * during replay, when we determine a subtree is no longer ours, we | |
7276 | * try to trim it from our cache. because subtrees must be connected | |
7277 | * to the root, the fact that we can trim this tree may mean that our | |
7278 | * children or parents can also be trimmed. | |
7279 | */ | |
7280 | void MDCache::try_trim_non_auth_subtree(CDir *dir) | |
7281 | { | |
7282 | dout(10) << "try_trim_nonauth_subtree " << *dir << dendl; | |
7283 | ||
7284 | // can we now trim child subtrees? | |
7285 | set<CDir*> bounds; | |
7286 | get_subtree_bounds(dir, bounds); | |
7287 | for (set<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) { | |
7288 | CDir *bd = *p; | |
7289 | if (bd->get_dir_auth().first != mds->get_nodeid() && // we are not auth | |
7290 | bd->get_num_any() == 0 && // and empty | |
7291 | can_trim_non_auth_dirfrag(bd)) { | |
7292 | CInode *bi = bd->get_inode(); | |
7293 | dout(10) << " closing empty non-auth child subtree " << *bd << dendl; | |
7294 | remove_subtree(bd); | |
7295 | bd->mark_clean(); | |
7296 | bi->close_dirfrag(bd->get_frag()); | |
7297 | } | |
7298 | } | |
7299 | ||
7300 | if (trim_non_auth_subtree(dir)) { | |
7301 | // keep | |
7302 | try_subtree_merge(dir); | |
7303 | } else { | |
7304 | // can we trim this subtree (and possibly our ancestors) too? | |
7305 | while (true) { | |
7306 | CInode *diri = dir->get_inode(); | |
7307 | if (diri->is_base()) { | |
7308 | if (!diri->is_root() && diri->authority().first != mds->get_nodeid()) { | |
7309 | dout(10) << " closing empty non-auth subtree " << *dir << dendl; | |
7310 | remove_subtree(dir); | |
7311 | dir->mark_clean(); | |
7312 | diri->close_dirfrag(dir->get_frag()); | |
7313 | ||
7314 | dout(10) << " removing " << *diri << dendl; | |
11fdf7f2 TL |
7315 | ceph_assert(!diri->get_parent_dn()); |
7316 | ceph_assert(diri->get_num_ref() == 0); | |
7c673cae FG |
7317 | remove_inode(diri); |
7318 | } | |
7319 | break; | |
7320 | } | |
7321 | ||
7322 | CDir *psub = get_subtree_root(diri->get_parent_dir()); | |
7323 | dout(10) << " parent subtree is " << *psub << dendl; | |
7324 | if (psub->get_dir_auth().first == mds->get_nodeid()) | |
7325 | break; // we are auth, keep. | |
7326 | ||
7327 | dout(10) << " closing empty non-auth subtree " << *dir << dendl; | |
7328 | remove_subtree(dir); | |
7329 | dir->mark_clean(); | |
7330 | diri->close_dirfrag(dir->get_frag()); | |
7331 | ||
7332 | dout(10) << " parent subtree also non-auth: " << *psub << dendl; | |
7333 | if (trim_non_auth_subtree(psub)) | |
7334 | break; | |
7335 | dir = psub; | |
7336 | } | |
7337 | } | |
7338 | ||
7339 | show_subtrees(); | |
7340 | } | |
7341 | ||
7342 | void MDCache::standby_trim_segment(LogSegment *ls) | |
7343 | { | |
494da23a TL |
7344 | auto try_trim_inode = [this](CInode *in) { |
7345 | if (in->get_num_ref() == 0 && | |
7346 | !in->item_open_file.is_on_list() && | |
7347 | in->parent != NULL && | |
7348 | in->parent->get_num_ref() == 0){ | |
7349 | touch_dentry_bottom(in->parent); | |
7350 | } | |
7351 | }; | |
7352 | ||
7353 | auto try_trim_dentry = [this](CDentry *dn) { | |
7354 | if (dn->get_num_ref() > 0) | |
7355 | return; | |
7356 | auto in = dn->get_linkage()->inode; | |
7357 | if(in && in->item_open_file.is_on_list()) | |
7358 | return; | |
7359 | touch_dentry_bottom(dn); | |
7360 | }; | |
7361 | ||
7c673cae FG |
7362 | ls->new_dirfrags.clear_list(); |
7363 | ls->open_files.clear_list(); | |
7364 | ||
7365 | while (!ls->dirty_dirfrags.empty()) { | |
7366 | CDir *dir = ls->dirty_dirfrags.front(); | |
7367 | dir->mark_clean(); | |
494da23a TL |
7368 | if (dir->inode) |
7369 | try_trim_inode(dir->inode); | |
7c673cae FG |
7370 | } |
7371 | while (!ls->dirty_inodes.empty()) { | |
7372 | CInode *in = ls->dirty_inodes.front(); | |
7373 | in->mark_clean(); | |
494da23a | 7374 | try_trim_inode(in); |
7c673cae FG |
7375 | } |
7376 | while (!ls->dirty_dentries.empty()) { | |
7377 | CDentry *dn = ls->dirty_dentries.front(); | |
7378 | dn->mark_clean(); | |
494da23a | 7379 | try_trim_dentry(dn); |
7c673cae FG |
7380 | } |
7381 | while (!ls->dirty_parent_inodes.empty()) { | |
7382 | CInode *in = ls->dirty_parent_inodes.front(); | |
7383 | in->clear_dirty_parent(); | |
494da23a | 7384 | try_trim_inode(in); |
7c673cae FG |
7385 | } |
7386 | while (!ls->dirty_dirfrag_dir.empty()) { | |
7387 | CInode *in = ls->dirty_dirfrag_dir.front(); | |
7388 | in->filelock.remove_dirty(); | |
494da23a | 7389 | try_trim_inode(in); |
7c673cae FG |
7390 | } |
7391 | while (!ls->dirty_dirfrag_nest.empty()) { | |
7392 | CInode *in = ls->dirty_dirfrag_nest.front(); | |
7393 | in->nestlock.remove_dirty(); | |
494da23a | 7394 | try_trim_inode(in); |
7c673cae FG |
7395 | } |
7396 | while (!ls->dirty_dirfrag_dirfragtree.empty()) { | |
7397 | CInode *in = ls->dirty_dirfrag_dirfragtree.front(); | |
7398 | in->dirfragtreelock.remove_dirty(); | |
494da23a | 7399 | try_trim_inode(in); |
7c673cae | 7400 | } |
eafe8130 TL |
7401 | while (!ls->truncating_inodes.empty()) { |
7402 | auto it = ls->truncating_inodes.begin(); | |
7403 | CInode *in = *it; | |
7404 | ls->truncating_inodes.erase(it); | |
7405 | in->put(CInode::PIN_TRUNCATING); | |
7406 | try_trim_inode(in); | |
7407 | } | |
7c673cae FG |
7408 | } |
7409 | ||
9f95a23c | 7410 | void MDCache::handle_cache_expire(const cref_t<MCacheExpire> &m) |
7c673cae FG |
7411 | { |
7412 | mds_rank_t from = mds_rank_t(m->get_from()); | |
7413 | ||
7414 | dout(7) << "cache_expire from mds." << from << dendl; | |
7415 | ||
7416 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
7c673cae FG |
7417 | return; |
7418 | } | |
7419 | ||
7420 | set<SimpleLock *> gather_locks; | |
7421 | // loop over realms | |
11fdf7f2 | 7422 | for (const auto &p : m->realms) { |
7c673cae | 7423 | // check container? |
11fdf7f2 TL |
7424 | if (p.first.ino > 0) { |
7425 | CInode *expired_inode = get_inode(p.first.ino); | |
7426 | ceph_assert(expired_inode); // we had better have this. | |
7427 | CDir *parent_dir = expired_inode->get_approx_dirfrag(p.first.frag); | |
7428 | ceph_assert(parent_dir); | |
7c673cae FG |
7429 | |
7430 | int export_state = -1; | |
7431 | if (parent_dir->is_auth() && parent_dir->is_exporting()) { | |
7432 | export_state = migrator->get_export_state(parent_dir); | |
11fdf7f2 | 7433 | ceph_assert(export_state >= 0); |
7c673cae FG |
7434 | } |
7435 | ||
7436 | if (!parent_dir->is_auth() || | |
7437 | (export_state != -1 && | |
7438 | ((export_state == Migrator::EXPORT_WARNING && | |
7439 | migrator->export_has_warned(parent_dir,from)) || | |
7440 | export_state == Migrator::EXPORT_EXPORTING || | |
7441 | export_state == Migrator::EXPORT_LOGGINGFINISH || | |
7442 | (export_state == Migrator::EXPORT_NOTIFYING && | |
7443 | !migrator->export_has_notified(parent_dir,from))))) { | |
7444 | ||
7445 | // not auth. | |
7446 | dout(7) << "delaying nonauth|warned expires for " << *parent_dir << dendl; | |
11fdf7f2 | 7447 | ceph_assert(parent_dir->is_frozen_tree_root()); |
7c673cae FG |
7448 | |
7449 | // make a message container | |
11fdf7f2 TL |
7450 | |
7451 | auto em = delayed_expire[parent_dir].emplace(std::piecewise_construct, std::forward_as_tuple(from), std::forward_as_tuple()); | |
7452 | if (em.second) | |
9f95a23c | 7453 | em.first->second = make_message<MCacheExpire>(from); /* new */ |
11fdf7f2 | 7454 | |
7c673cae | 7455 | // merge these expires into it |
11fdf7f2 | 7456 | em.first->second->add_realm(p.first, p.second); |
7c673cae FG |
7457 | continue; |
7458 | } | |
11fdf7f2 | 7459 | ceph_assert(export_state <= Migrator::EXPORT_PREPPING || |
7c673cae FG |
7460 | (export_state == Migrator::EXPORT_WARNING && |
7461 | !migrator->export_has_warned(parent_dir, from))); | |
7462 | ||
7463 | dout(7) << "expires for " << *parent_dir << dendl; | |
7464 | } else { | |
7465 | dout(7) << "containerless expires (root, stray inodes)" << dendl; | |
7466 | } | |
7467 | ||
7468 | // INODES | |
11fdf7f2 TL |
7469 | for (const auto &q : p.second.inodes) { |
7470 | CInode *in = get_inode(q.first); | |
7471 | unsigned nonce = q.second; | |
7c673cae FG |
7472 | |
7473 | if (!in) { | |
11fdf7f2 | 7474 | dout(0) << " inode expire on " << q.first << " from " << from |
7c673cae | 7475 | << ", don't have it" << dendl; |
11fdf7f2 | 7476 | ceph_assert(in); |
7c673cae | 7477 | } |
11fdf7f2 | 7478 | ceph_assert(in->is_auth()); |
7c673cae FG |
7479 | dout(20) << __func__ << ": expiring inode " << *in << dendl; |
7480 | ||
7481 | // check nonce | |
7482 | if (nonce == in->get_replica_nonce(from)) { | |
7483 | // remove from our cached_by | |
7484 | dout(7) << " inode expire on " << *in << " from mds." << from | |
7485 | << " cached_by was " << in->get_replicas() << dendl; | |
7486 | inode_remove_replica(in, from, false, gather_locks); | |
7487 | } | |
7488 | else { | |
7489 | // this is an old nonce, ignore expire. | |
7490 | dout(7) << " inode expire on " << *in << " from mds." << from | |
7491 | << " with old nonce " << nonce | |
7492 | << " (current " << in->get_replica_nonce(from) << "), dropping" | |
7493 | << dendl; | |
7494 | } | |
7495 | } | |
7496 | ||
7497 | // DIRS | |
11fdf7f2 TL |
7498 | for (const auto &q : p.second.dirs) { |
7499 | CDir *dir = get_dirfrag(q.first); | |
7500 | unsigned nonce = q.second; | |
7c673cae FG |
7501 | |
7502 | if (!dir) { | |
11fdf7f2 | 7503 | CInode *diri = get_inode(q.first.ino); |
7c673cae FG |
7504 | if (diri) { |
7505 | if (mds->is_rejoin() && | |
7506 | rejoin_ack_gather.count(mds->get_nodeid()) && // haven't sent rejoin ack yet | |
7507 | !diri->is_replica(from)) { | |
9f95a23c | 7508 | auto&& ls = diri->get_nested_dirfrags(); |
11fdf7f2 | 7509 | dout(7) << " dir expire on dirfrag " << q.first << " from mds." << from |
7c673cae | 7510 | << " while rejoining, inode isn't replicated" << dendl; |
9f95a23c TL |
7511 | for (const auto& d : ls) { |
7512 | dir = d; | |
7c673cae FG |
7513 | if (dir->is_replica(from)) { |
7514 | dout(7) << " dir expire on " << *dir << " from mds." << from << dendl; | |
7515 | dir->remove_replica(from); | |
7516 | } | |
7517 | } | |
7518 | continue; | |
7519 | } | |
11fdf7f2 | 7520 | CDir *other = diri->get_approx_dirfrag(q.first.frag); |
7c673cae | 7521 | if (other) { |
11fdf7f2 | 7522 | dout(7) << " dir expire on dirfrag " << q.first << " from mds." << from |
7c673cae FG |
7523 | << " have " << *other << ", mismatched frags, dropping" << dendl; |
7524 | continue; | |
7525 | } | |
7526 | } | |
11fdf7f2 | 7527 | dout(0) << " dir expire on " << q.first << " from " << from |
7c673cae | 7528 | << ", don't have it" << dendl; |
11fdf7f2 | 7529 | ceph_assert(dir); |
7c673cae FG |
7530 | } |
7531 | dout(20) << __func__ << ": expiring dirfrag " << *dir << dendl; | |
7532 | ||
11fdf7f2 | 7533 | ceph_assert(dir->is_auth()); |
7c673cae FG |
7534 | |
7535 | // check nonce | |
7536 | if (nonce == dir->get_replica_nonce(from)) { | |
7537 | // remove from our cached_by | |
7538 | dout(7) << " dir expire on " << *dir << " from mds." << from | |
181888fb | 7539 | << " replicas was " << dir->get_replicas() << dendl; |
7c673cae FG |
7540 | dir->remove_replica(from); |
7541 | } | |
7542 | else { | |
7543 | // this is an old nonce, ignore expire. | |
7544 | dout(7) << " dir expire on " << *dir << " from mds." << from | |
7545 | << " with old nonce " << nonce << " (current " << dir->get_replica_nonce(from) | |
7546 | << "), dropping" << dendl; | |
7547 | } | |
7548 | } | |
7549 | ||
7550 | // DENTRIES | |
11fdf7f2 TL |
7551 | for (const auto &pd : p.second.dentries) { |
7552 | dout(10) << " dn expires in dir " << pd.first << dendl; | |
7553 | CInode *diri = get_inode(pd.first.ino); | |
7554 | ceph_assert(diri); | |
7555 | CDir *dir = diri->get_dirfrag(pd.first.frag); | |
7c673cae FG |
7556 | |
7557 | if (!dir) { | |
11fdf7f2 | 7558 | dout(0) << " dn expires on " << pd.first << " from " << from |
7c673cae FG |
7559 | << ", must have refragmented" << dendl; |
7560 | } else { | |
11fdf7f2 | 7561 | ceph_assert(dir->is_auth()); |
7c673cae FG |
7562 | } |
7563 | ||
11fdf7f2 TL |
7564 | for (const auto &p : pd.second) { |
7565 | unsigned nonce = p.second; | |
7c673cae FG |
7566 | CDentry *dn; |
7567 | ||
7568 | if (dir) { | |
11fdf7f2 | 7569 | dn = dir->lookup(p.first.first, p.first.second); |
7c673cae FG |
7570 | } else { |
7571 | // which dirfrag for this dentry? | |
11fdf7f2 TL |
7572 | CDir *dir = diri->get_dirfrag(diri->pick_dirfrag(p.first.first)); |
7573 | ceph_assert(dir); | |
7574 | ceph_assert(dir->is_auth()); | |
7575 | dn = dir->lookup(p.first.first, p.first.second); | |
7c673cae FG |
7576 | } |
7577 | ||
7578 | if (!dn) { | |
7579 | if (dir) | |
11fdf7f2 | 7580 | dout(0) << " missing dentry for " << p.first.first << " snap " << p.first.second << " in " << *dir << dendl; |
7c673cae | 7581 | else |
11fdf7f2 | 7582 | dout(0) << " missing dentry for " << p.first.first << " snap " << p.first.second << dendl; |
7c673cae | 7583 | } |
11fdf7f2 | 7584 | ceph_assert(dn); |
7c673cae FG |
7585 | |
7586 | if (nonce == dn->get_replica_nonce(from)) { | |
7587 | dout(7) << " dentry_expire on " << *dn << " from mds." << from << dendl; | |
7588 | dentry_remove_replica(dn, from, gather_locks); | |
7589 | } | |
7590 | else { | |
7591 | dout(7) << " dentry_expire on " << *dn << " from mds." << from | |
7592 | << " with old nonce " << nonce << " (current " << dn->get_replica_nonce(from) | |
7593 | << "), dropping" << dendl; | |
7594 | } | |
7595 | } | |
7596 | } | |
7597 | } | |
7598 | ||
7c673cae FG |
7599 | for (set<SimpleLock*>::iterator p = gather_locks.begin(); p != gather_locks.end(); ++p) { |
7600 | if (!(*p)->is_stable()) | |
7601 | mds->locker->eval_gather(*p); | |
7602 | } | |
7603 | } | |
7604 | ||
7605 | void MDCache::process_delayed_expire(CDir *dir) | |
7606 | { | |
7607 | dout(7) << "process_delayed_expire on " << *dir << dendl; | |
11fdf7f2 TL |
7608 | for (const auto &p : delayed_expire[dir]) { |
7609 | handle_cache_expire(p.second); | |
7610 | } | |
7c673cae FG |
7611 | delayed_expire.erase(dir); |
7612 | } | |
7613 | ||
7614 | void MDCache::discard_delayed_expire(CDir *dir) | |
7615 | { | |
7616 | dout(7) << "discard_delayed_expire on " << *dir << dendl; | |
7c673cae FG |
7617 | delayed_expire.erase(dir); |
7618 | } | |
7619 | ||
7620 | void MDCache::inode_remove_replica(CInode *in, mds_rank_t from, bool rejoin, | |
7621 | set<SimpleLock *>& gather_locks) | |
7622 | { | |
7623 | in->remove_replica(from); | |
11fdf7f2 | 7624 | in->set_mds_caps_wanted(from, 0); |
7c673cae FG |
7625 | |
7626 | // note: this code calls _eval more often than it needs to! | |
7627 | // fix lock | |
7628 | if (in->authlock.remove_replica(from)) gather_locks.insert(&in->authlock); | |
7629 | if (in->linklock.remove_replica(from)) gather_locks.insert(&in->linklock); | |
7630 | if (in->snaplock.remove_replica(from)) gather_locks.insert(&in->snaplock); | |
7631 | if (in->xattrlock.remove_replica(from)) gather_locks.insert(&in->xattrlock); | |
7632 | if (in->flocklock.remove_replica(from)) gather_locks.insert(&in->flocklock); | |
7633 | if (in->policylock.remove_replica(from)) gather_locks.insert(&in->policylock); | |
7634 | ||
7635 | // If 'rejoin' is true and the scatter lock is in LOCK_MIX_* state. | |
7636 | // Don't remove the recovering mds from lock's gathering list because | |
7637 | // it may hold rejoined wrlocks. | |
7638 | if (in->dirfragtreelock.remove_replica(from, rejoin)) gather_locks.insert(&in->dirfragtreelock); | |
7639 | if (in->filelock.remove_replica(from, rejoin)) gather_locks.insert(&in->filelock); | |
7640 | if (in->nestlock.remove_replica(from, rejoin)) gather_locks.insert(&in->nestlock); | |
7641 | } | |
7642 | ||
7643 | void MDCache::dentry_remove_replica(CDentry *dn, mds_rank_t from, set<SimpleLock *>& gather_locks) | |
7644 | { | |
7645 | dn->remove_replica(from); | |
7646 | ||
7647 | // fix lock | |
7648 | if (dn->lock.remove_replica(from)) | |
7649 | gather_locks.insert(&dn->lock); | |
7650 | ||
7651 | // Replicated strays might now be elegible for purge | |
11fdf7f2 | 7652 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); |
7c673cae FG |
7653 | if (dnl->is_primary()) { |
7654 | maybe_eval_stray(dnl->get_inode()); | |
7655 | } | |
7656 | } | |
7657 | ||
7658 | void MDCache::trim_client_leases() | |
7659 | { | |
7660 | utime_t now = ceph_clock_now(); | |
7661 | ||
7662 | dout(10) << "trim_client_leases" << dendl; | |
7663 | ||
eafe8130 TL |
7664 | std::size_t pool = 0; |
7665 | for (const auto& list : client_leases) { | |
7666 | pool += 1; | |
7667 | if (list.empty()) | |
7c673cae FG |
7668 | continue; |
7669 | ||
eafe8130 TL |
7670 | auto before = list.size(); |
7671 | while (!list.empty()) { | |
7672 | ClientLease *r = list.front(); | |
7c673cae FG |
7673 | if (r->ttl > now) break; |
7674 | CDentry *dn = static_cast<CDentry*>(r->parent); | |
7675 | dout(10) << " expiring client." << r->client << " lease of " << *dn << dendl; | |
7676 | dn->remove_client_lease(r, mds->locker); | |
7677 | } | |
eafe8130 | 7678 | auto after = list.size(); |
7c673cae FG |
7679 | dout(10) << "trim_client_leases pool " << pool << " trimmed " |
7680 | << (before-after) << " leases, " << after << " left" << dendl; | |
7681 | } | |
7682 | } | |
7683 | ||
7c673cae FG |
7684 | void MDCache::check_memory_usage() |
7685 | { | |
7686 | static MemoryModel mm(g_ceph_context); | |
7687 | static MemoryModel::snap last; | |
7688 | mm.sample(&last); | |
7689 | static MemoryModel::snap baseline = last; | |
7690 | ||
7691 | // check client caps | |
11fdf7f2 | 7692 | ceph_assert(CInode::count() == inode_map.size() + snap_inode_map.size() + num_shadow_inodes); |
181888fb | 7693 | double caps_per_inode = 0.0; |
7c673cae | 7694 | if (CInode::count()) |
181888fb | 7695 | caps_per_inode = (double)Capability::count() / (double)CInode::count(); |
7c673cae | 7696 | |
a8e16298 | 7697 | dout(2) << "Memory usage: " |
7c673cae FG |
7698 | << " total " << last.get_total() |
7699 | << ", rss " << last.get_rss() | |
7700 | << ", heap " << last.get_heap() | |
7701 | << ", baseline " << baseline.get_heap() | |
7c673cae FG |
7702 | << ", " << num_inodes_with_caps << " / " << CInode::count() << " inodes have caps" |
7703 | << ", " << Capability::count() << " caps, " << caps_per_inode << " caps per inode" | |
7704 | << dendl; | |
7705 | ||
c07f9fc5 | 7706 | mds->update_mlogger(); |
7c673cae FG |
7707 | mds->mlogger->set(l_mdm_rss, last.get_rss()); |
7708 | mds->mlogger->set(l_mdm_heap, last.get_heap()); | |
7c673cae FG |
7709 | } |
7710 | ||
7711 | ||
7712 | ||
7713 | // ========================================================================================= | |
7714 | // shutdown | |
7715 | ||
7716 | class C_MDC_ShutdownCheck : public MDCacheContext { | |
7717 | public: | |
7718 | explicit C_MDC_ShutdownCheck(MDCache *m) : MDCacheContext(m) {} | |
7719 | void finish(int) override { | |
7720 | mdcache->shutdown_check(); | |
7721 | } | |
7722 | }; | |
7723 | ||
7724 | void MDCache::shutdown_check() | |
7725 | { | |
7726 | dout(0) << "shutdown_check at " << ceph_clock_now() << dendl; | |
7727 | ||
7728 | // cache | |
7729 | char old_val[32] = { 0 }; | |
7730 | char *o = old_val; | |
11fdf7f2 TL |
7731 | g_conf().get_val("debug_mds", &o, sizeof(old_val)); |
7732 | g_conf().set_val("debug_mds", "10"); | |
7733 | g_conf().apply_changes(nullptr); | |
7c673cae | 7734 | show_cache(); |
11fdf7f2 TL |
7735 | g_conf().set_val("debug_mds", old_val); |
7736 | g_conf().apply_changes(nullptr); | |
7737 | mds->timer.add_event_after(g_conf()->mds_shutdown_check, new C_MDC_ShutdownCheck(this)); | |
7c673cae FG |
7738 | |
7739 | // this | |
31f18b77 | 7740 | dout(0) << "lru size now " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; |
7c673cae FG |
7741 | dout(0) << "log len " << mds->mdlog->get_num_events() << dendl; |
7742 | ||
7743 | ||
7744 | if (mds->objecter->is_active()) { | |
7745 | dout(0) << "objecter still active" << dendl; | |
7746 | mds->objecter->dump_active(); | |
7747 | } | |
7748 | } | |
7749 | ||
7750 | ||
7751 | void MDCache::shutdown_start() | |
7752 | { | |
a8e16298 | 7753 | dout(5) << "shutdown_start" << dendl; |
7c673cae | 7754 | |
11fdf7f2 TL |
7755 | if (g_conf()->mds_shutdown_check) |
7756 | mds->timer.add_event_after(g_conf()->mds_shutdown_check, new C_MDC_ShutdownCheck(this)); | |
7c673cae | 7757 | |
11fdf7f2 | 7758 | // g_conf()->debug_mds = 10; |
7c673cae FG |
7759 | } |
7760 | ||
7761 | ||
7762 | ||
7763 | bool MDCache::shutdown_pass() | |
7764 | { | |
7765 | dout(7) << "shutdown_pass" << dendl; | |
7766 | ||
7767 | if (mds->is_stopped()) { | |
7768 | dout(7) << " already shut down" << dendl; | |
7769 | show_cache(); | |
7770 | show_subtrees(); | |
7771 | return true; | |
7772 | } | |
7773 | ||
7774 | // empty stray dir | |
28e407b8 | 7775 | bool strays_all_exported = shutdown_export_strays(); |
7c673cae FG |
7776 | |
7777 | // trim cache | |
181888fb | 7778 | trim(UINT64_MAX); |
31f18b77 | 7779 | dout(5) << "lru size now " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; |
7c673cae | 7780 | |
28e407b8 | 7781 | // Export all subtrees to another active (usually rank 0) if not rank 0 |
7c673cae | 7782 | int num_auth_subtree = 0; |
f6b5b4d7 TL |
7783 | if (!subtrees.empty() && mds->get_nodeid() != 0) { |
7784 | dout(7) << "looking for subtrees to export" << dendl; | |
9f95a23c | 7785 | std::vector<CDir*> ls; |
f6b5b4d7 TL |
7786 | for (auto& [dir, bounds] : subtrees) { |
7787 | dout(10) << " examining " << *dir << " bounds " << bounds << dendl; | |
7788 | if (dir->get_inode()->is_mdsdir() || !dir->is_auth()) | |
7c673cae | 7789 | continue; |
f6b5b4d7 TL |
7790 | num_auth_subtree++; |
7791 | if (dir->is_frozen() || | |
7792 | dir->is_freezing() || | |
7793 | dir->is_ambiguous_dir_auth() || | |
7794 | dir->state_test(CDir::STATE_EXPORTING) || | |
7795 | dir->get_inode()->is_ephemerally_pinned()) { | |
7796 | continue; | |
7c673cae | 7797 | } |
f6b5b4d7 | 7798 | ls.push_back(dir); |
7c673cae | 7799 | } |
28e407b8 AA |
7800 | |
7801 | migrator->clear_export_queue(); | |
f67539c2 TL |
7802 | // stopping mds does not call MDBalancer::tick() |
7803 | mds->balancer->handle_export_pins(); | |
9f95a23c | 7804 | for (const auto& dir : ls) { |
7c673cae FG |
7805 | mds_rank_t dest = dir->get_inode()->authority().first; |
7806 | if (dest > 0 && !mds->mdsmap->is_active(dest)) | |
7807 | dest = 0; | |
7808 | dout(7) << "sending " << *dir << " back to mds." << dest << dendl; | |
7809 | migrator->export_dir_nicely(dir, dest); | |
7810 | } | |
7811 | } | |
7812 | ||
28e407b8 AA |
7813 | if (!strays_all_exported) { |
7814 | dout(7) << "waiting for strays to migrate" << dendl; | |
7815 | return false; | |
7816 | } | |
7817 | ||
7c673cae | 7818 | if (num_auth_subtree > 0) { |
11fdf7f2 | 7819 | ceph_assert(mds->get_nodeid() > 0); |
7c673cae FG |
7820 | dout(7) << "still have " << num_auth_subtree << " auth subtrees" << dendl; |
7821 | show_subtrees(); | |
7822 | return false; | |
7823 | } | |
7824 | ||
7825 | // close out any sessions (and open files!) before we try to trim the log, etc. | |
7826 | if (mds->sessionmap.have_unclosed_sessions()) { | |
7827 | if (!mds->server->terminating_sessions) | |
7828 | mds->server->terminate_sessions(); | |
7829 | return false; | |
7830 | } | |
7831 | ||
28e407b8 AA |
7832 | // Fully trim the log so that all objects in cache are clean and may be |
7833 | // trimmed by a future MDCache::trim. Note that MDSRank::tick does not | |
7834 | // trim the log such that the cache eventually becomes clean. | |
f64942e4 AA |
7835 | if (mds->mdlog->get_num_segments() > 0) { |
7836 | auto ls = mds->mdlog->get_current_segment(); | |
7837 | if (ls->num_events > 1 || !ls->dirty_dirfrags.empty()) { | |
7838 | // Current segment contains events other than subtreemap or | |
7839 | // there are dirty dirfrags (see CDir::log_mark_dirty()) | |
7840 | mds->mdlog->start_new_segment(); | |
7841 | mds->mdlog->flush(); | |
7842 | } | |
7843 | } | |
7844 | mds->mdlog->trim_all(); | |
28e407b8 AA |
7845 | if (mds->mdlog->get_num_segments() > 1) { |
7846 | dout(7) << "still >1 segments, waiting for log to trim" << dendl; | |
7847 | return false; | |
7848 | } | |
7849 | ||
7850 | // drop our reference to our stray dir inode | |
7851 | for (int i = 0; i < NUM_STRAY; ++i) { | |
7852 | if (strays[i] && | |
7853 | strays[i]->state_test(CInode::STATE_STRAYPINNED)) { | |
7854 | strays[i]->state_clear(CInode::STATE_STRAYPINNED); | |
7855 | strays[i]->put(CInode::PIN_STRAY); | |
7856 | strays[i]->put_stickydirs(); | |
7857 | } | |
7858 | } | |
7859 | ||
7c673cae FG |
7860 | CDir *mydir = myin ? myin->get_dirfrag(frag_t()) : NULL; |
7861 | if (mydir && !mydir->is_subtree_root()) | |
7862 | mydir = NULL; | |
7863 | ||
7864 | // subtrees map not empty yet? | |
7865 | if (subtrees.size() > (mydir ? 1 : 0)) { | |
7866 | dout(7) << "still have " << num_subtrees() << " subtrees" << dendl; | |
7867 | show_subtrees(); | |
7868 | migrator->show_importing(); | |
7869 | migrator->show_exporting(); | |
7870 | if (!migrator->is_importing() && !migrator->is_exporting()) | |
7871 | show_cache(); | |
7872 | return false; | |
7873 | } | |
11fdf7f2 TL |
7874 | ceph_assert(!migrator->is_exporting()); |
7875 | ceph_assert(!migrator->is_importing()); | |
7c673cae | 7876 | |
f64942e4 AA |
7877 | // replicas may dirty scatter locks |
7878 | if (myin && myin->is_replicated()) { | |
7879 | dout(7) << "still have replicated objects" << dendl; | |
7880 | return false; | |
7881 | } | |
7882 | ||
11fdf7f2 TL |
7883 | if ((myin && myin->get_num_auth_pins()) || |
7884 | (mydir && (mydir->get_auth_pins() || mydir->get_dir_auth_pins()))) { | |
181888fb FG |
7885 | dout(7) << "still have auth pinned objects" << dendl; |
7886 | return false; | |
7887 | } | |
7888 | ||
7c673cae FG |
7889 | // (only do this once!) |
7890 | if (!mds->mdlog->is_capped()) { | |
7891 | dout(7) << "capping the log" << dendl; | |
7892 | mds->mdlog->cap(); | |
7c673cae FG |
7893 | } |
7894 | ||
f64942e4 AA |
7895 | if (!mds->mdlog->empty()) |
7896 | mds->mdlog->trim(0); | |
7897 | ||
7c673cae FG |
7898 | if (!mds->mdlog->empty()) { |
7899 | dout(7) << "waiting for log to flush.. " << mds->mdlog->get_num_events() | |
7900 | << " in " << mds->mdlog->get_num_segments() << " segments" << dendl; | |
7901 | return false; | |
7902 | } | |
7903 | ||
7904 | if (!did_shutdown_log_cap) { | |
7905 | // flush journal header | |
7906 | dout(7) << "writing header for (now-empty) journal" << dendl; | |
11fdf7f2 | 7907 | ceph_assert(mds->mdlog->empty()); |
7c673cae FG |
7908 | mds->mdlog->write_head(0); |
7909 | // NOTE: filer active checker below will block us until this completes. | |
7910 | did_shutdown_log_cap = true; | |
7911 | return false; | |
7912 | } | |
7913 | ||
7914 | // filer active? | |
7915 | if (mds->objecter->is_active()) { | |
7916 | dout(7) << "objecter still active" << dendl; | |
7917 | mds->objecter->dump_active(); | |
7918 | return false; | |
7919 | } | |
7920 | ||
7921 | // trim what we can from the cache | |
31f18b77 FG |
7922 | if (lru.lru_get_size() > 0 || bottom_lru.lru_get_size() > 0) { |
7923 | dout(7) << "there's still stuff in the cache: " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; | |
7c673cae FG |
7924 | show_cache(); |
7925 | //dump(); | |
7926 | return false; | |
7927 | } | |
31f18b77 FG |
7928 | |
7929 | // make mydir subtree go away | |
7930 | if (mydir) { | |
7931 | if (mydir->get_num_ref() > 1) { // subtree pin | |
7932 | dout(7) << "there's still reference to mydir " << *mydir << dendl; | |
7933 | show_cache(); | |
7934 | return false; | |
7935 | } | |
7936 | ||
7937 | remove_subtree(mydir); | |
7938 | myin->close_dirfrag(mydir->get_frag()); | |
7939 | } | |
11fdf7f2 | 7940 | ceph_assert(subtrees.empty()); |
31f18b77 | 7941 | |
1adf2230 | 7942 | if (myin) { |
31f18b77 | 7943 | remove_inode(myin); |
11fdf7f2 | 7944 | ceph_assert(!myin); |
1adf2230 AA |
7945 | } |
7946 | ||
11fdf7f2 TL |
7947 | if (global_snaprealm) { |
7948 | remove_inode(global_snaprealm->inode); | |
7949 | global_snaprealm = nullptr; | |
7950 | } | |
7951 | ||
7c673cae | 7952 | // done! |
a8e16298 | 7953 | dout(5) << "shutdown done." << dendl; |
7c673cae FG |
7954 | return true; |
7955 | } | |
7956 | ||
7957 | bool MDCache::shutdown_export_strays() | |
7958 | { | |
f64942e4 AA |
7959 | static const unsigned MAX_EXPORTING = 100; |
7960 | ||
7c673cae FG |
7961 | if (mds->get_nodeid() == 0) |
7962 | return true; | |
f64942e4 AA |
7963 | |
7964 | if (shutdown_exporting_strays.size() * 3 >= MAX_EXPORTING * 2) | |
7965 | return false; | |
7966 | ||
7967 | dout(10) << "shutdown_export_strays " << shutdown_export_next.first | |
7968 | << " '" << shutdown_export_next.second << "'" << dendl; | |
7c673cae FG |
7969 | |
7970 | bool mds0_active = mds->mdsmap->is_active(mds_rank_t(0)); | |
f64942e4 | 7971 | bool all_exported = false; |
7c673cae | 7972 | |
f64942e4 AA |
7973 | again: |
7974 | auto next = shutdown_export_next; | |
7c673cae | 7975 | |
7c673cae | 7976 | for (int i = 0; i < NUM_STRAY; ++i) { |
f64942e4 AA |
7977 | CInode *strayi = strays[i]; |
7978 | if (!strayi || | |
7979 | !strayi->state_test(CInode::STATE_STRAYPINNED)) | |
7980 | continue; | |
7981 | if (strayi->ino() < next.first.ino) | |
7c673cae | 7982 | continue; |
7c673cae | 7983 | |
f64942e4 AA |
7984 | deque<CDir*> dfls; |
7985 | strayi->get_dirfrags(dfls); | |
7c673cae | 7986 | |
f64942e4 AA |
7987 | while (!dfls.empty()) { |
7988 | CDir *dir = dfls.front(); | |
7989 | dfls.pop_front(); | |
7990 | ||
7991 | if (dir->dirfrag() < next.first) | |
7c673cae | 7992 | continue; |
f64942e4 AA |
7993 | if (next.first < dir->dirfrag()) { |
7994 | next.first = dir->dirfrag(); | |
7995 | next.second.clear(); | |
7996 | } | |
7997 | ||
7998 | if (!dir->is_complete()) { | |
11fdf7f2 | 7999 | MDSContext *fin = nullptr; |
f64942e4 AA |
8000 | if (shutdown_exporting_strays.empty()) { |
8001 | fin = new MDSInternalContextWrapper(mds, | |
9f95a23c | 8002 | new LambdaContext([this](int r) { |
f64942e4 AA |
8003 | shutdown_export_strays(); |
8004 | }) | |
8005 | ); | |
8006 | } | |
8007 | dir->fetch(fin); | |
8008 | goto done; | |
7c673cae FG |
8009 | } |
8010 | ||
f64942e4 AA |
8011 | CDir::dentry_key_map::iterator it; |
8012 | if (next.second.empty()) { | |
8013 | it = dir->begin(); | |
7c673cae | 8014 | } else { |
f64942e4 AA |
8015 | auto hash = ceph_frag_value(strayi->hash_dentry_name(next.second)); |
8016 | it = dir->lower_bound(dentry_key_t(0, next.second, hash)); | |
7c673cae | 8017 | } |
f64942e4 AA |
8018 | |
8019 | for (; it != dir->end(); ++it) { | |
8020 | CDentry *dn = it->second; | |
8021 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
8022 | if (dnl->is_null()) | |
8023 | continue; | |
8024 | ||
8025 | if (!mds0_active && !dn->state_test(CDentry::STATE_PURGING)) { | |
11fdf7f2 | 8026 | next.second = it->first.name; |
f64942e4 AA |
8027 | goto done; |
8028 | } | |
8029 | ||
8030 | auto ret = shutdown_exporting_strays.insert(dnl->get_inode()->ino()); | |
8031 | if (!ret.second) { | |
8032 | dout(10) << "already exporting/purging " << *dn << dendl; | |
8033 | continue; | |
8034 | } | |
8035 | ||
8036 | // Don't try to migrate anything that is actually | |
8037 | // being purged right now | |
8038 | if (!dn->state_test(CDentry::STATE_PURGING)) | |
8039 | stray_manager.migrate_stray(dn, mds_rank_t(0)); // send to root! | |
8040 | ||
8041 | if (shutdown_exporting_strays.size() >= MAX_EXPORTING) { | |
8042 | ++it; | |
8043 | if (it != dir->end()) { | |
11fdf7f2 | 8044 | next.second = it->first.name; |
f64942e4 AA |
8045 | } else { |
8046 | if (dfls.empty()) | |
8047 | next.first.ino.val++; | |
8048 | else | |
8049 | next.first = dfls.front()->dirfrag(); | |
8050 | next.second.clear(); | |
8051 | } | |
8052 | goto done; | |
8053 | } | |
8054 | } | |
8055 | } | |
8056 | } | |
8057 | ||
8058 | if (shutdown_exporting_strays.empty()) { | |
8059 | dirfrag_t first_df(MDS_INO_STRAY(mds->get_nodeid(), 0), 0); | |
8060 | if (first_df < shutdown_export_next.first || | |
8061 | !shutdown_export_next.second.empty()) { | |
8062 | shutdown_export_next.first = first_df; | |
8063 | shutdown_export_next.second.clear(); | |
8064 | goto again; | |
7c673cae | 8065 | } |
f64942e4 | 8066 | all_exported = true; |
7c673cae FG |
8067 | } |
8068 | ||
f64942e4 AA |
8069 | done: |
8070 | shutdown_export_next = next; | |
8071 | return all_exported; | |
7c673cae FG |
8072 | } |
8073 | ||
8074 | // ========= messaging ============== | |
8075 | ||
9f95a23c | 8076 | void MDCache::dispatch(const cref_t<Message> &m) |
7c673cae FG |
8077 | { |
8078 | switch (m->get_type()) { | |
8079 | ||
8080 | // RESOLVE | |
8081 | case MSG_MDS_RESOLVE: | |
9f95a23c | 8082 | handle_resolve(ref_cast<MMDSResolve>(m)); |
7c673cae FG |
8083 | break; |
8084 | case MSG_MDS_RESOLVEACK: | |
9f95a23c | 8085 | handle_resolve_ack(ref_cast<MMDSResolveAck>(m)); |
7c673cae FG |
8086 | break; |
8087 | ||
8088 | // REJOIN | |
8089 | case MSG_MDS_CACHEREJOIN: | |
9f95a23c | 8090 | handle_cache_rejoin(ref_cast<MMDSCacheRejoin>(m)); |
7c673cae FG |
8091 | break; |
8092 | ||
8093 | case MSG_MDS_DISCOVER: | |
9f95a23c | 8094 | handle_discover(ref_cast<MDiscover>(m)); |
7c673cae FG |
8095 | break; |
8096 | case MSG_MDS_DISCOVERREPLY: | |
9f95a23c | 8097 | handle_discover_reply(ref_cast<MDiscoverReply>(m)); |
7c673cae FG |
8098 | break; |
8099 | ||
8100 | case MSG_MDS_DIRUPDATE: | |
9f95a23c | 8101 | handle_dir_update(ref_cast<MDirUpdate>(m)); |
7c673cae FG |
8102 | break; |
8103 | ||
8104 | case MSG_MDS_CACHEEXPIRE: | |
9f95a23c | 8105 | handle_cache_expire(ref_cast<MCacheExpire>(m)); |
7c673cae FG |
8106 | break; |
8107 | ||
8108 | case MSG_MDS_DENTRYLINK: | |
9f95a23c | 8109 | handle_dentry_link(ref_cast<MDentryLink>(m)); |
7c673cae FG |
8110 | break; |
8111 | case MSG_MDS_DENTRYUNLINK: | |
9f95a23c | 8112 | handle_dentry_unlink(ref_cast<MDentryUnlink>(m)); |
7c673cae FG |
8113 | break; |
8114 | ||
8115 | case MSG_MDS_FRAGMENTNOTIFY: | |
9f95a23c | 8116 | handle_fragment_notify(ref_cast<MMDSFragmentNotify>(m)); |
7c673cae | 8117 | break; |
a8e16298 | 8118 | case MSG_MDS_FRAGMENTNOTIFYACK: |
9f95a23c | 8119 | handle_fragment_notify_ack(ref_cast<MMDSFragmentNotifyAck>(m)); |
a8e16298 | 8120 | break; |
7c673cae FG |
8121 | |
8122 | case MSG_MDS_FINDINO: | |
9f95a23c | 8123 | handle_find_ino(ref_cast<MMDSFindIno>(m)); |
7c673cae FG |
8124 | break; |
8125 | case MSG_MDS_FINDINOREPLY: | |
9f95a23c | 8126 | handle_find_ino_reply(ref_cast<MMDSFindInoReply>(m)); |
7c673cae FG |
8127 | break; |
8128 | ||
8129 | case MSG_MDS_OPENINO: | |
9f95a23c | 8130 | handle_open_ino(ref_cast<MMDSOpenIno>(m)); |
7c673cae FG |
8131 | break; |
8132 | case MSG_MDS_OPENINOREPLY: | |
9f95a23c | 8133 | handle_open_ino_reply(ref_cast<MMDSOpenInoReply>(m)); |
11fdf7f2 TL |
8134 | break; |
8135 | ||
8136 | case MSG_MDS_SNAPUPDATE: | |
9f95a23c | 8137 | handle_snap_update(ref_cast<MMDSSnapUpdate>(m)); |
7c673cae FG |
8138 | break; |
8139 | ||
8140 | default: | |
8141 | derr << "cache unknown message " << m->get_type() << dendl; | |
11fdf7f2 | 8142 | ceph_abort_msg("cache unknown message"); |
7c673cae FG |
8143 | } |
8144 | } | |
8145 | ||
9f95a23c TL |
8146 | int MDCache::path_traverse(MDRequestRef& mdr, MDSContextFactory& cf, |
8147 | const filepath& path, int flags, | |
8148 | vector<CDentry*> *pdnvec, CInode **pin) | |
7c673cae | 8149 | { |
9f95a23c TL |
8150 | bool discover = (flags & MDS_TRAVERSE_DISCOVER); |
8151 | bool forward = !discover; | |
8152 | bool path_locked = (flags & MDS_TRAVERSE_PATH_LOCKED); | |
8153 | bool want_dentry = (flags & MDS_TRAVERSE_WANT_DENTRY); | |
8154 | bool want_auth = (flags & MDS_TRAVERSE_WANT_AUTH); | |
8155 | bool rdlock_snap = (flags & (MDS_TRAVERSE_RDLOCK_SNAP | MDS_TRAVERSE_RDLOCK_SNAP2)); | |
8156 | bool rdlock_path = (flags & MDS_TRAVERSE_RDLOCK_PATH); | |
8157 | bool xlock_dentry = (flags & MDS_TRAVERSE_XLOCK_DENTRY); | |
8158 | bool rdlock_authlock = (flags & MDS_TRAVERSE_RDLOCK_AUTHLOCK); | |
7c673cae | 8159 | |
9f95a23c TL |
8160 | if (forward) |
8161 | ceph_assert(mdr); // forward requires a request | |
7c673cae FG |
8162 | |
8163 | snapid_t snapid = CEPH_NOSNAP; | |
8164 | if (mdr) | |
8165 | mdr->snapid = snapid; | |
8166 | ||
b3b6e05e | 8167 | client_t client = mdr ? mdr->get_client() : -1; |
7c673cae FG |
8168 | |
8169 | if (mds->logger) mds->logger->inc(l_mds_traverse); | |
8170 | ||
8171 | dout(7) << "traverse: opening base ino " << path.get_ino() << " snap " << snapid << dendl; | |
8172 | CInode *cur = get_inode(path.get_ino()); | |
9f95a23c TL |
8173 | if (!cur) { |
8174 | if (MDS_INO_IS_MDSDIR(path.get_ino())) { | |
11fdf7f2 | 8175 | open_foreign_mdsdir(path.get_ino(), cf.build()); |
9f95a23c | 8176 | return 1; |
7c673cae | 8177 | } |
9f95a23c TL |
8178 | if (MDS_INO_IS_STRAY(path.get_ino())) { |
8179 | mds_rank_t rank = MDS_INO_STRAY_OWNER(path.get_ino()); | |
8180 | unsigned idx = MDS_INO_STRAY_INDEX(path.get_ino()); | |
8181 | filepath path(strays[idx]->get_parent_dn()->get_name(), | |
8182 | MDS_INO_MDSDIR(rank)); | |
8183 | MDRequestRef null_ref; | |
8184 | return path_traverse(null_ref, cf, path, MDS_TRAVERSE_DISCOVER, nullptr); | |
8185 | } | |
f67539c2 | 8186 | return -CEPHFS_ESTALE; |
7c673cae FG |
8187 | } |
8188 | if (cur->state_test(CInode::STATE_PURGING)) | |
f67539c2 | 8189 | return -CEPHFS_ESTALE; |
7c673cae | 8190 | |
9f95a23c TL |
8191 | if (flags & MDS_TRAVERSE_CHECK_LOCKCACHE) |
8192 | mds->locker->find_and_attach_lock_cache(mdr, cur); | |
8193 | ||
8194 | if (mdr && mdr->lock_cache) { | |
8195 | if (flags & MDS_TRAVERSE_WANT_DIRLAYOUT) | |
8196 | mdr->dir_layout = mdr->lock_cache->get_dir_layout(); | |
8197 | } else if (rdlock_snap) { | |
8198 | int n = (flags & MDS_TRAVERSE_RDLOCK_SNAP2) ? 1 : 0; | |
8199 | if ((n == 0 && !(mdr->locking_state & MutationImpl::SNAP_LOCKED)) || | |
8200 | (n == 1 && !(mdr->locking_state & MutationImpl::SNAP2_LOCKED))) { | |
8201 | bool want_layout = (flags & MDS_TRAVERSE_WANT_DIRLAYOUT); | |
8202 | if (!mds->locker->try_rdlock_snap_layout(cur, mdr, n, want_layout)) | |
8203 | return 1; | |
8204 | } | |
8205 | } | |
8206 | ||
7c673cae FG |
8207 | // start trace |
8208 | if (pdnvec) | |
8209 | pdnvec->clear(); | |
8210 | if (pin) | |
8211 | *pin = cur; | |
8212 | ||
9f95a23c TL |
8213 | MutationImpl::LockOpVec lov; |
8214 | ||
8215 | for (unsigned depth = 0; depth < path.depth(); ) { | |
7c673cae FG |
8216 | dout(12) << "traverse: path seg depth " << depth << " '" << path[depth] |
8217 | << "' snapid " << snapid << dendl; | |
8218 | ||
8219 | if (!cur->is_dir()) { | |
8220 | dout(7) << "traverse: " << *cur << " not a dir " << dendl; | |
f67539c2 | 8221 | return -CEPHFS_ENOTDIR; |
7c673cae FG |
8222 | } |
8223 | ||
8224 | // walk into snapdir? | |
8225 | if (path[depth].length() == 0) { | |
8226 | dout(10) << "traverse: snapdir" << dendl; | |
9f95a23c | 8227 | if (!mdr || depth > 0) // snapdir must be the first component |
f67539c2 | 8228 | return -CEPHFS_EINVAL; |
7c673cae FG |
8229 | snapid = CEPH_SNAPDIR; |
8230 | mdr->snapid = snapid; | |
8231 | depth++; | |
8232 | continue; | |
8233 | } | |
8234 | // walk thru snapdir? | |
8235 | if (snapid == CEPH_SNAPDIR) { | |
8236 | if (!mdr) | |
f67539c2 | 8237 | return -CEPHFS_EINVAL; |
7c673cae FG |
8238 | SnapRealm *realm = cur->find_snaprealm(); |
8239 | snapid = realm->resolve_snapname(path[depth], cur->ino()); | |
8240 | dout(10) << "traverse: snap " << path[depth] << " -> " << snapid << dendl; | |
11fdf7f2 | 8241 | if (!snapid) { |
9f95a23c TL |
8242 | if (pdnvec) |
8243 | pdnvec->clear(); // do not confuse likes of rdlock_path_pin_ref(); | |
f67539c2 | 8244 | return -CEPHFS_ENOENT; |
11fdf7f2 | 8245 | } |
7c673cae FG |
8246 | mdr->snapid = snapid; |
8247 | depth++; | |
8248 | continue; | |
8249 | } | |
8250 | ||
8251 | // open dir | |
8252 | frag_t fg = cur->pick_dirfrag(path[depth]); | |
8253 | CDir *curdir = cur->get_dirfrag(fg); | |
8254 | if (!curdir) { | |
8255 | if (cur->is_auth()) { | |
8256 | // parent dir frozen_dir? | |
8257 | if (cur->is_frozen()) { | |
8258 | dout(7) << "traverse: " << *cur << " is frozen, waiting" << dendl; | |
11fdf7f2 | 8259 | cur->add_waiter(CDir::WAIT_UNFREEZE, cf.build()); |
7c673cae FG |
8260 | return 1; |
8261 | } | |
8262 | curdir = cur->get_or_open_dirfrag(this, fg); | |
8263 | } else { | |
8264 | // discover? | |
8265 | dout(10) << "traverse: need dirfrag " << fg << ", doing discover from " << *cur << dendl; | |
11fdf7f2 | 8266 | discover_path(cur, snapid, path.postfixpath(depth), cf.build(), |
9f95a23c | 8267 | path_locked); |
7c673cae FG |
8268 | if (mds->logger) mds->logger->inc(l_mds_traverse_discover); |
8269 | return 1; | |
8270 | } | |
8271 | } | |
11fdf7f2 | 8272 | ceph_assert(curdir); |
7c673cae FG |
8273 | |
8274 | #ifdef MDS_VERIFY_FRAGSTAT | |
8275 | if (curdir->is_complete()) | |
8276 | curdir->verify_fragstat(); | |
8277 | #endif | |
8278 | ||
8279 | // frozen? | |
8280 | /* | |
8281 | if (curdir->is_frozen()) { | |
8282 | // doh! | |
8283 | // FIXME: traverse is allowed? | |
8284 | dout(7) << "traverse: " << *curdir << " is frozen, waiting" << dendl; | |
8285 | curdir->add_waiter(CDir::WAIT_UNFREEZE, _get_waiter(mdr, req, fin)); | |
8286 | if (onfinish) delete onfinish; | |
8287 | return 1; | |
8288 | } | |
8289 | */ | |
8290 | ||
9f95a23c TL |
8291 | if (want_auth && want_dentry && depth == path.depth() - 1) { |
8292 | if (curdir->is_ambiguous_auth()) { | |
8293 | dout(10) << "waiting for single auth on " << *curdir << dendl; | |
8294 | curdir->add_waiter(CInode::WAIT_SINGLEAUTH, cf.build()); | |
8295 | return 1; | |
8296 | } | |
8297 | if (!curdir->is_auth()) { | |
8298 | dout(10) << "fw to auth for " << *curdir << dendl; | |
8299 | request_forward(mdr, curdir->authority().first); | |
8300 | return 2; | |
8301 | } | |
8302 | } | |
8303 | ||
7c673cae FG |
8304 | // Before doing dirfrag->dn lookup, compare with DamageTable's |
8305 | // record of which dentries were unreadable | |
8306 | if (mds->damage_table.is_dentry_damaged(curdir, path[depth], snapid)) { | |
8307 | dout(4) << "traverse: stopped lookup at damaged dentry " | |
8308 | << *curdir << "/" << path[depth] << " snap=" << snapid << dendl; | |
f67539c2 | 8309 | return -CEPHFS_EIO; |
7c673cae FG |
8310 | } |
8311 | ||
8312 | // dentry | |
8313 | CDentry *dn = curdir->lookup(path[depth], snapid); | |
9f95a23c TL |
8314 | if (dn) { |
8315 | if (dn->state_test(CDentry::STATE_PURGING)) | |
f67539c2 | 8316 | return -CEPHFS_ENOENT; |
9f95a23c TL |
8317 | |
8318 | if (rdlock_path) { | |
8319 | lov.clear(); | |
8320 | if (xlock_dentry && depth == path.depth() - 1) { | |
8321 | if (depth > 0 || !mdr->lock_cache) { | |
8322 | lov.add_wrlock(&cur->filelock); | |
8323 | lov.add_wrlock(&cur->nestlock); | |
8324 | if (rdlock_authlock) | |
8325 | lov.add_rdlock(&cur->authlock); | |
8326 | } | |
8327 | lov.add_xlock(&dn->lock); | |
8328 | } else { | |
8329 | // force client to flush async dir operation if necessary | |
8330 | if (cur->filelock.is_cached()) | |
8331 | lov.add_wrlock(&cur->filelock); | |
8332 | lov.add_rdlock(&dn->lock); | |
8333 | } | |
8334 | if (!mds->locker->acquire_locks(mdr, lov)) { | |
8335 | dout(10) << "traverse: failed to rdlock " << dn->lock << " " << *dn << dendl; | |
8336 | return 1; | |
8337 | } | |
8338 | } else if (!path_locked && | |
8339 | !dn->lock.can_read(client) && | |
8340 | !(dn->lock.is_xlocked() && dn->lock.get_xlock_by() == mdr)) { | |
8341 | dout(10) << "traverse: non-readable dentry at " << *dn << dendl; | |
8342 | dn->lock.add_waiter(SimpleLock::WAIT_RD, cf.build()); | |
8343 | if (mds->logger) | |
8344 | mds->logger->inc(l_mds_traverse_lock); | |
8345 | if (dn->is_auth() && dn->lock.is_unstable_and_locked()) | |
8346 | mds->mdlog->flush(); | |
8347 | return 1; | |
8348 | } | |
7c673cae | 8349 | |
7c673cae FG |
8350 | if (pdnvec) |
8351 | pdnvec->push_back(dn); | |
7c673cae | 8352 | |
9f95a23c | 8353 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); |
f67539c2 | 8354 | // can we conclude CEPHFS_ENOENT? |
9f95a23c TL |
8355 | if (dnl->is_null()) { |
8356 | dout(10) << "traverse: null+readable dentry at " << *dn << dendl; | |
8357 | if (depth == path.depth() - 1) { | |
8358 | if (want_dentry) | |
8359 | break; | |
8360 | } else { | |
8361 | if (pdnvec) | |
7c673cae FG |
8362 | pdnvec->clear(); // do not confuse likes of rdlock_path_pin_ref(); |
8363 | } | |
f67539c2 | 8364 | return -CEPHFS_ENOENT; |
7c673cae | 8365 | } |
7c673cae | 8366 | |
7c673cae | 8367 | // do we have inode? |
9f95a23c | 8368 | CInode *in = dnl->get_inode(); |
7c673cae | 8369 | if (!in) { |
11fdf7f2 | 8370 | ceph_assert(dnl->is_remote()); |
7c673cae FG |
8371 | // do i have it? |
8372 | in = get_inode(dnl->get_remote_ino()); | |
8373 | if (in) { | |
8374 | dout(7) << "linking in remote in " << *in << dendl; | |
8375 | dn->link_remote(dnl, in); | |
8376 | } else { | |
8377 | dout(7) << "remote link to " << dnl->get_remote_ino() << ", which i don't have" << dendl; | |
11fdf7f2 | 8378 | ceph_assert(mdr); // we shouldn't hit non-primary dentries doing a non-mdr traversal! |
7c673cae FG |
8379 | if (mds->damage_table.is_remote_damaged(dnl->get_remote_ino())) { |
8380 | dout(4) << "traverse: remote dentry points to damaged ino " | |
8381 | << *dn << dendl; | |
f67539c2 | 8382 | return -CEPHFS_EIO; |
7c673cae | 8383 | } |
11fdf7f2 | 8384 | open_remote_dentry(dn, true, cf.build(), |
9f95a23c | 8385 | (path_locked && depth == path.depth() - 1)); |
7c673cae FG |
8386 | if (mds->logger) mds->logger->inc(l_mds_traverse_remote_ino); |
8387 | return 1; | |
9f95a23c | 8388 | } |
7c673cae FG |
8389 | } |
8390 | ||
8391 | cur = in; | |
7c673cae | 8392 | |
9f95a23c TL |
8393 | if (rdlock_snap && !(want_dentry && depth == path.depth() - 1)) { |
8394 | lov.clear(); | |
8395 | lov.add_rdlock(&cur->snaplock); | |
8396 | if (!mds->locker->acquire_locks(mdr, lov)) { | |
8397 | dout(10) << "traverse: failed to rdlock " << cur->snaplock << " " << *cur << dendl; | |
8398 | return 1; | |
8399 | } | |
8400 | } | |
8401 | ||
7c673cae FG |
8402 | // add to trace, continue. |
8403 | touch_inode(cur); | |
7c673cae FG |
8404 | if (pin) |
8405 | *pin = cur; | |
8406 | depth++; | |
8407 | continue; | |
8408 | } | |
9f95a23c TL |
8409 | |
8410 | ceph_assert(!dn); | |
7c673cae FG |
8411 | |
8412 | // MISS. dentry doesn't exist. | |
8413 | dout(12) << "traverse: miss on dentry " << path[depth] << " in " << *curdir << dendl; | |
8414 | ||
8415 | if (curdir->is_auth()) { | |
8416 | // dentry is mine. | |
8417 | if (curdir->is_complete() || | |
8418 | (snapid == CEPH_NOSNAP && | |
8419 | curdir->has_bloom() && | |
11fdf7f2 | 8420 | !curdir->is_in_bloom(path[depth]))) { |
7c673cae FG |
8421 | // file not found |
8422 | if (pdnvec) { | |
8423 | // instantiate a null dn? | |
9f95a23c | 8424 | if (depth < path.depth() - 1) { |
7c673cae | 8425 | dout(20) << " didn't traverse full path; not returning pdnvec" << dendl; |
7c673cae FG |
8426 | } else if (snapid < CEPH_MAXSNAP) { |
8427 | dout(20) << " not adding null for snapid " << snapid << dendl; | |
9f95a23c TL |
8428 | } else if (curdir->is_frozen()) { |
8429 | dout(7) << "traverse: " << *curdir << " is frozen, waiting" << dendl; | |
8430 | curdir->add_waiter(CDir::WAIT_UNFREEZE, cf.build()); | |
8431 | return 1; | |
7c673cae FG |
8432 | } else { |
8433 | // create a null dentry | |
8434 | dn = curdir->add_null_dentry(path[depth]); | |
8435 | dout(20) << " added null " << *dn << dendl; | |
9f95a23c TL |
8436 | |
8437 | if (rdlock_path) { | |
8438 | lov.clear(); | |
8439 | if (xlock_dentry) { | |
8440 | if (depth > 0 || !mdr->lock_cache) { | |
8441 | lov.add_wrlock(&cur->filelock); | |
8442 | lov.add_wrlock(&cur->nestlock); | |
8443 | if (rdlock_authlock) | |
8444 | lov.add_rdlock(&cur->authlock); | |
8445 | } | |
8446 | lov.add_xlock(&dn->lock); | |
8447 | } else { | |
8448 | // force client to flush async dir operation if necessary | |
8449 | if (cur->filelock.is_cached()) | |
8450 | lov.add_wrlock(&cur->filelock); | |
8451 | lov.add_rdlock(&dn->lock); | |
8452 | } | |
8453 | if (!mds->locker->acquire_locks(mdr, lov)) { | |
8454 | dout(10) << "traverse: failed to rdlock " << dn->lock << " " << *dn << dendl; | |
8455 | return 1; | |
8456 | } | |
8457 | } | |
7c673cae | 8458 | } |
9f95a23c | 8459 | if (dn) { |
7c673cae | 8460 | pdnvec->push_back(dn); |
9f95a23c TL |
8461 | if (want_dentry) |
8462 | break; | |
8463 | } else { | |
7c673cae | 8464 | pdnvec->clear(); // do not confuse likes of rdlock_path_pin_ref(); |
9f95a23c | 8465 | } |
7c673cae | 8466 | } |
f67539c2 | 8467 | return -CEPHFS_ENOENT; |
7c673cae FG |
8468 | } else { |
8469 | ||
8470 | // Check DamageTable for missing fragments before trying to fetch | |
8471 | // this | |
8472 | if (mds->damage_table.is_dirfrag_damaged(curdir)) { | |
8473 | dout(4) << "traverse: damaged dirfrag " << *curdir | |
8474 | << ", blocking fetch" << dendl; | |
f67539c2 | 8475 | return -CEPHFS_EIO; |
7c673cae FG |
8476 | } |
8477 | ||
8478 | // directory isn't complete; reload | |
8479 | dout(7) << "traverse: incomplete dir contents for " << *cur << ", fetching" << dendl; | |
8480 | touch_inode(cur); | |
11fdf7f2 | 8481 | curdir->fetch(cf.build(), path[depth]); |
7c673cae FG |
8482 | if (mds->logger) mds->logger->inc(l_mds_traverse_dir_fetch); |
8483 | return 1; | |
8484 | } | |
8485 | } else { | |
8486 | // dirfrag/dentry is not mine. | |
8487 | mds_authority_t dauth = curdir->authority(); | |
8488 | ||
f91f0fd5 | 8489 | if (forward && |
11fdf7f2 | 8490 | mdr && mdr->client_request && |
9f95a23c | 8491 | (int)depth < mdr->client_request->get_num_fwd()){ |
7c673cae FG |
8492 | dout(7) << "traverse: snap " << snapid << " and depth " << depth |
8493 | << " < fwd " << mdr->client_request->get_num_fwd() | |
8494 | << ", discovering instead of forwarding" << dendl; | |
8495 | discover = true; | |
8496 | } | |
8497 | ||
9f95a23c | 8498 | if ((discover)) { |
7c673cae | 8499 | dout(7) << "traverse: discover from " << path[depth] << " from " << *curdir << dendl; |
11fdf7f2 | 8500 | discover_path(curdir, snapid, path.postfixpath(depth), cf.build(), |
9f95a23c | 8501 | path_locked); |
7c673cae FG |
8502 | if (mds->logger) mds->logger->inc(l_mds_traverse_discover); |
8503 | return 1; | |
8504 | } | |
8505 | if (forward) { | |
8506 | // forward | |
8507 | dout(7) << "traverse: not auth for " << path << " in " << *curdir << dendl; | |
8508 | ||
8509 | if (curdir->is_ambiguous_auth()) { | |
8510 | // wait | |
8511 | dout(7) << "traverse: waiting for single auth in " << *curdir << dendl; | |
11fdf7f2 | 8512 | curdir->add_waiter(CDir::WAIT_SINGLEAUTH, cf.build()); |
7c673cae FG |
8513 | return 1; |
8514 | } | |
8515 | ||
8516 | dout(7) << "traverse: forwarding, not auth for " << *curdir << dendl; | |
11fdf7f2 TL |
8517 | |
8518 | request_forward(mdr, dauth.first); | |
8519 | ||
7c673cae | 8520 | if (mds->logger) mds->logger->inc(l_mds_traverse_forward); |
7c673cae | 8521 | return 2; |
11fdf7f2 | 8522 | } |
7c673cae | 8523 | } |
11fdf7f2 | 8524 | |
7c673cae FG |
8525 | ceph_abort(); // i shouldn't get here |
8526 | } | |
9f95a23c TL |
8527 | |
8528 | if (want_auth && !want_dentry) { | |
8529 | if (cur->is_ambiguous_auth()) { | |
8530 | dout(10) << "waiting for single auth on " << *cur << dendl; | |
8531 | cur->add_waiter(CInode::WAIT_SINGLEAUTH, cf.build()); | |
8532 | return 1; | |
8533 | } | |
8534 | if (!cur->is_auth()) { | |
8535 | dout(10) << "fw to auth for " << *cur << dendl; | |
8536 | request_forward(mdr, cur->authority().first); | |
8537 | return 2; | |
8538 | } | |
8539 | } | |
7c673cae FG |
8540 | |
8541 | // success. | |
8542 | if (mds->logger) mds->logger->inc(l_mds_traverse_hit); | |
8543 | dout(10) << "path_traverse finish on snapid " << snapid << dendl; | |
8544 | if (mdr) | |
11fdf7f2 | 8545 | ceph_assert(mdr->snapid == snapid); |
9f95a23c TL |
8546 | |
8547 | if (flags & MDS_TRAVERSE_RDLOCK_SNAP) | |
8548 | mdr->locking_state |= MutationImpl::SNAP_LOCKED; | |
8549 | else if (flags & MDS_TRAVERSE_RDLOCK_SNAP2) | |
8550 | mdr->locking_state |= MutationImpl::SNAP2_LOCKED; | |
8551 | ||
8552 | if (rdlock_path) | |
8553 | mdr->locking_state |= MutationImpl::PATH_LOCKED; | |
8554 | ||
7c673cae FG |
8555 | return 0; |
8556 | } | |
8557 | ||
8558 | CInode *MDCache::cache_traverse(const filepath& fp) | |
8559 | { | |
8560 | dout(10) << "cache_traverse " << fp << dendl; | |
8561 | ||
8562 | CInode *in; | |
f67539c2 | 8563 | unsigned depth = 0; |
20effc67 TL |
8564 | char mdsdir_name[16]; |
8565 | sprintf(mdsdir_name, "~mds%d", mds->get_nodeid()); | |
f67539c2 TL |
8566 | |
8567 | if (fp.get_ino()) { | |
7c673cae | 8568 | in = get_inode(fp.get_ino()); |
20effc67 | 8569 | } else if (fp.depth() > 0 && (fp[0] == "~mdsdir" || fp[0] == mdsdir_name)) { |
f67539c2 TL |
8570 | in = myin; |
8571 | depth = 1; | |
8572 | } else { | |
7c673cae | 8573 | in = root; |
f67539c2 | 8574 | } |
7c673cae FG |
8575 | if (!in) |
8576 | return NULL; | |
8577 | ||
f67539c2 TL |
8578 | for (; depth < fp.depth(); depth++) { |
8579 | std::string_view dname = fp[depth]; | |
7c673cae | 8580 | frag_t fg = in->pick_dirfrag(dname); |
f67539c2 | 8581 | dout(20) << " " << depth << " " << dname << " frag " << fg << " from " << *in << dendl; |
7c673cae FG |
8582 | CDir *curdir = in->get_dirfrag(fg); |
8583 | if (!curdir) | |
8584 | return NULL; | |
8585 | CDentry *dn = curdir->lookup(dname, CEPH_NOSNAP); | |
8586 | if (!dn) | |
8587 | return NULL; | |
8588 | in = dn->get_linkage()->get_inode(); | |
8589 | if (!in) | |
8590 | return NULL; | |
8591 | } | |
8592 | dout(10) << " got " << *in << dendl; | |
8593 | return in; | |
8594 | } | |
8595 | ||
8596 | ||
8597 | /** | |
8598 | * open_remote_dir -- open up a remote dirfrag | |
8599 | * | |
8600 | * @param diri base inode | |
8601 | * @param approxfg approximate fragment. | |
8602 | * @param fin completion callback | |
8603 | */ | |
11fdf7f2 | 8604 | void MDCache::open_remote_dirfrag(CInode *diri, frag_t approxfg, MDSContext *fin) |
7c673cae FG |
8605 | { |
8606 | dout(10) << "open_remote_dir on " << *diri << dendl; | |
11fdf7f2 TL |
8607 | ceph_assert(diri->is_dir()); |
8608 | ceph_assert(!diri->is_auth()); | |
8609 | ceph_assert(diri->get_dirfrag(approxfg) == 0); | |
7c673cae | 8610 | |
224ce89b | 8611 | discover_dir_frag(diri, approxfg, fin); |
7c673cae FG |
8612 | } |
8613 | ||
8614 | ||
8615 | /** | |
8616 | * get_dentry_inode - get or open inode | |
8617 | * | |
8618 | * @param dn the dentry | |
8619 | * @param mdr current request | |
8620 | * | |
8621 | * will return inode for primary, or link up/open up remote link's inode as necessary. | |
8622 | * If it's not available right now, puts mdr on wait list and returns null. | |
8623 | */ | |
8624 | CInode *MDCache::get_dentry_inode(CDentry *dn, MDRequestRef& mdr, bool projected) | |
8625 | { | |
8626 | CDentry::linkage_t *dnl; | |
8627 | if (projected) | |
8628 | dnl = dn->get_projected_linkage(); | |
8629 | else | |
8630 | dnl = dn->get_linkage(); | |
8631 | ||
11fdf7f2 | 8632 | ceph_assert(!dnl->is_null()); |
7c673cae FG |
8633 | |
8634 | if (dnl->is_primary()) | |
8635 | return dnl->inode; | |
8636 | ||
11fdf7f2 | 8637 | ceph_assert(dnl->is_remote()); |
7c673cae FG |
8638 | CInode *in = get_inode(dnl->get_remote_ino()); |
8639 | if (in) { | |
8640 | dout(7) << "get_dentry_inode linking in remote in " << *in << dendl; | |
8641 | dn->link_remote(dnl, in); | |
8642 | return in; | |
8643 | } else { | |
8644 | dout(10) << "get_dentry_inode on remote dn, opening inode for " << *dn << dendl; | |
8645 | open_remote_dentry(dn, projected, new C_MDS_RetryRequest(this, mdr)); | |
8646 | return 0; | |
8647 | } | |
8648 | } | |
8649 | ||
8650 | struct C_MDC_OpenRemoteDentry : public MDCacheContext { | |
8651 | CDentry *dn; | |
8652 | inodeno_t ino; | |
11fdf7f2 | 8653 | MDSContext *onfinish; |
7c673cae | 8654 | bool want_xlocked; |
11fdf7f2 | 8655 | C_MDC_OpenRemoteDentry(MDCache *m, CDentry *d, inodeno_t i, MDSContext *f, bool wx) : |
31f18b77 FG |
8656 | MDCacheContext(m), dn(d), ino(i), onfinish(f), want_xlocked(wx) { |
8657 | dn->get(MDSCacheObject::PIN_PTRWAITER); | |
8658 | } | |
7c673cae FG |
8659 | void finish(int r) override { |
8660 | mdcache->_open_remote_dentry_finish(dn, ino, onfinish, want_xlocked, r); | |
31f18b77 | 8661 | dn->put(MDSCacheObject::PIN_PTRWAITER); |
7c673cae FG |
8662 | } |
8663 | }; | |
8664 | ||
11fdf7f2 | 8665 | void MDCache::open_remote_dentry(CDentry *dn, bool projected, MDSContext *fin, bool want_xlocked) |
7c673cae FG |
8666 | { |
8667 | dout(10) << "open_remote_dentry " << *dn << dendl; | |
8668 | CDentry::linkage_t *dnl = projected ? dn->get_projected_linkage() : dn->get_linkage(); | |
8669 | inodeno_t ino = dnl->get_remote_ino(); | |
b3b6e05e | 8670 | int64_t pool = dnl->get_remote_d_type() == DT_DIR ? mds->get_metadata_pool() : -1; |
7c673cae FG |
8671 | open_ino(ino, pool, |
8672 | new C_MDC_OpenRemoteDentry(this, dn, ino, fin, want_xlocked), true, want_xlocked); // backtrace | |
8673 | } | |
8674 | ||
11fdf7f2 | 8675 | void MDCache::_open_remote_dentry_finish(CDentry *dn, inodeno_t ino, MDSContext *fin, |
7c673cae FG |
8676 | bool want_xlocked, int r) |
8677 | { | |
8678 | if (r < 0) { | |
31f18b77 FG |
8679 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); |
8680 | if (dnl->is_remote() && dnl->get_remote_ino() == ino) { | |
7c673cae FG |
8681 | dout(0) << "open_remote_dentry_finish bad remote dentry " << *dn << dendl; |
8682 | dn->state_set(CDentry::STATE_BADREMOTEINO); | |
8683 | ||
8684 | std::string path; | |
8685 | CDir *dir = dn->get_dir(); | |
8686 | if (dir) { | |
31f18b77 | 8687 | dir->get_inode()->make_path_string(path); |
94b18763 | 8688 | path += "/"; |
11fdf7f2 | 8689 | path += dn->get_name(); |
7c673cae FG |
8690 | } |
8691 | ||
31f18b77 | 8692 | bool fatal = mds->damage_table.notify_remote_damaged(ino, path); |
7c673cae | 8693 | if (fatal) { |
31f18b77 FG |
8694 | mds->damaged(); |
8695 | ceph_abort(); // unreachable, damaged() respawns us | |
7c673cae | 8696 | } |
31f18b77 FG |
8697 | } else { |
8698 | r = 0; | |
8699 | } | |
7c673cae FG |
8700 | } |
8701 | fin->complete(r < 0 ? r : 0); | |
8702 | } | |
8703 | ||
8704 | ||
8705 | void MDCache::make_trace(vector<CDentry*>& trace, CInode *in) | |
8706 | { | |
8707 | // empty trace if we're a base inode | |
8708 | if (in->is_base()) | |
8709 | return; | |
8710 | ||
8711 | CInode *parent = in->get_parent_inode(); | |
11fdf7f2 | 8712 | ceph_assert(parent); |
7c673cae FG |
8713 | make_trace(trace, parent); |
8714 | ||
8715 | CDentry *dn = in->get_parent_dn(); | |
8716 | dout(15) << "make_trace adding " << *dn << dendl; | |
8717 | trace.push_back(dn); | |
8718 | } | |
8719 | ||
8720 | ||
8721 | // ------------------------------------------------------------------------------- | |
8722 | // Open inode by inode number | |
8723 | ||
8724 | class C_IO_MDC_OpenInoBacktraceFetched : public MDCacheIOContext { | |
8725 | inodeno_t ino; | |
8726 | public: | |
8727 | bufferlist bl; | |
8728 | C_IO_MDC_OpenInoBacktraceFetched(MDCache *c, inodeno_t i) : | |
8729 | MDCacheIOContext(c), ino(i) {} | |
8730 | void finish(int r) override { | |
8731 | mdcache->_open_ino_backtrace_fetched(ino, bl, r); | |
8732 | } | |
91327a77 AA |
8733 | void print(ostream& out) const override { |
8734 | out << "openino_backtrace_fetch" << ino << ")"; | |
8735 | } | |
7c673cae FG |
8736 | }; |
8737 | ||
8738 | struct C_MDC_OpenInoTraverseDir : public MDCacheContext { | |
8739 | inodeno_t ino; | |
9f95a23c | 8740 | cref_t<MMDSOpenIno> msg; |
7c673cae FG |
8741 | bool parent; |
8742 | public: | |
9f95a23c | 8743 | C_MDC_OpenInoTraverseDir(MDCache *c, inodeno_t i, const cref_t<MMDSOpenIno> &m, bool p) : |
7c673cae FG |
8744 | MDCacheContext(c), ino(i), msg(m), parent(p) {} |
8745 | void finish(int r) override { | |
8746 | if (r < 0 && !parent) | |
f67539c2 | 8747 | r = -CEPHFS_EAGAIN; |
7c673cae FG |
8748 | if (msg) { |
8749 | mdcache->handle_open_ino(msg, r); | |
8750 | return; | |
8751 | } | |
11fdf7f2 TL |
8752 | auto& info = mdcache->opening_inodes.at(ino); |
8753 | mdcache->_open_ino_traverse_dir(ino, info, r); | |
7c673cae FG |
8754 | } |
8755 | }; | |
8756 | ||
8757 | struct C_MDC_OpenInoParentOpened : public MDCacheContext { | |
8758 | inodeno_t ino; | |
8759 | public: | |
8760 | C_MDC_OpenInoParentOpened(MDCache *c, inodeno_t i) : MDCacheContext(c), ino(i) {} | |
8761 | void finish(int r) override { | |
8762 | mdcache->_open_ino_parent_opened(ino, r); | |
8763 | } | |
8764 | }; | |
8765 | ||
8766 | void MDCache::_open_ino_backtrace_fetched(inodeno_t ino, bufferlist& bl, int err) | |
8767 | { | |
8768 | dout(10) << "_open_ino_backtrace_fetched ino " << ino << " errno " << err << dendl; | |
8769 | ||
11fdf7f2 | 8770 | open_ino_info_t& info = opening_inodes.at(ino); |
7c673cae FG |
8771 | |
8772 | CInode *in = get_inode(ino); | |
8773 | if (in) { | |
8774 | dout(10) << " found cached " << *in << dendl; | |
8775 | open_ino_finish(ino, info, in->authority().first); | |
8776 | return; | |
8777 | } | |
8778 | ||
8779 | inode_backtrace_t backtrace; | |
8780 | if (err == 0) { | |
8781 | try { | |
11fdf7f2 | 8782 | decode(backtrace, bl); |
7c673cae FG |
8783 | } catch (const buffer::error &decode_exc) { |
8784 | derr << "corrupt backtrace on ino x0" << std::hex << ino | |
f67539c2 TL |
8785 | << std::dec << ": " << decode_exc.what() << dendl; |
8786 | open_ino_finish(ino, info, -CEPHFS_EIO); | |
7c673cae FG |
8787 | return; |
8788 | } | |
8789 | if (backtrace.pool != info.pool && backtrace.pool != -1) { | |
8790 | dout(10) << " old object in pool " << info.pool | |
8791 | << ", retrying pool " << backtrace.pool << dendl; | |
8792 | info.pool = backtrace.pool; | |
8793 | C_IO_MDC_OpenInoBacktraceFetched *fin = | |
8794 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8795 | fetch_backtrace(ino, info.pool, fin->bl, | |
8796 | new C_OnFinisher(fin, mds->finisher)); | |
8797 | return; | |
8798 | } | |
f67539c2 | 8799 | } else if (err == -CEPHFS_ENOENT) { |
b3b6e05e | 8800 | int64_t meta_pool = mds->get_metadata_pool(); |
7c673cae FG |
8801 | if (info.pool != meta_pool) { |
8802 | dout(10) << " no object in pool " << info.pool | |
8803 | << ", retrying pool " << meta_pool << dendl; | |
8804 | info.pool = meta_pool; | |
8805 | C_IO_MDC_OpenInoBacktraceFetched *fin = | |
8806 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8807 | fetch_backtrace(ino, info.pool, fin->bl, | |
8808 | new C_OnFinisher(fin, mds->finisher)); | |
8809 | return; | |
8810 | } | |
8811 | err = 0; // backtrace.ancestors.empty() is checked below | |
8812 | } | |
8813 | ||
8814 | if (err == 0) { | |
8815 | if (backtrace.ancestors.empty()) { | |
8816 | dout(10) << " got empty backtrace " << dendl; | |
f67539c2 | 8817 | err = -CEPHFS_ESTALE; |
7c673cae FG |
8818 | } else if (!info.ancestors.empty()) { |
8819 | if (info.ancestors[0] == backtrace.ancestors[0]) { | |
8820 | dout(10) << " got same parents " << info.ancestors[0] << " 2 times" << dendl; | |
f67539c2 | 8821 | err = -CEPHFS_EINVAL; |
7c673cae FG |
8822 | } else { |
8823 | info.last_err = 0; | |
8824 | } | |
8825 | } | |
8826 | } | |
8827 | if (err) { | |
8828 | dout(0) << " failed to open ino " << ino << " err " << err << "/" << info.last_err << dendl; | |
8829 | if (info.last_err) | |
8830 | err = info.last_err; | |
8831 | open_ino_finish(ino, info, err); | |
8832 | return; | |
8833 | } | |
8834 | ||
8835 | dout(10) << " got backtrace " << backtrace << dendl; | |
8836 | info.ancestors = backtrace.ancestors; | |
8837 | ||
8838 | _open_ino_traverse_dir(ino, info, 0); | |
8839 | } | |
8840 | ||
8841 | void MDCache::_open_ino_parent_opened(inodeno_t ino, int ret) | |
8842 | { | |
8843 | dout(10) << "_open_ino_parent_opened ino " << ino << " ret " << ret << dendl; | |
8844 | ||
11fdf7f2 | 8845 | open_ino_info_t& info = opening_inodes.at(ino); |
7c673cae FG |
8846 | |
8847 | CInode *in = get_inode(ino); | |
8848 | if (in) { | |
8849 | dout(10) << " found cached " << *in << dendl; | |
8850 | open_ino_finish(ino, info, in->authority().first); | |
8851 | return; | |
8852 | } | |
8853 | ||
8854 | if (ret == mds->get_nodeid()) { | |
8855 | _open_ino_traverse_dir(ino, info, 0); | |
8856 | } else { | |
8857 | if (ret >= 0) { | |
8858 | mds_rank_t checked_rank = mds_rank_t(ret); | |
8859 | info.check_peers = true; | |
8860 | info.auth_hint = checked_rank; | |
8861 | info.checked.erase(checked_rank); | |
8862 | } | |
8863 | do_open_ino(ino, info, ret); | |
8864 | } | |
8865 | } | |
8866 | ||
8867 | void MDCache::_open_ino_traverse_dir(inodeno_t ino, open_ino_info_t& info, int ret) | |
8868 | { | |
8869 | dout(10) << __func__ << ": ino " << ino << " ret " << ret << dendl; | |
8870 | ||
8871 | CInode *in = get_inode(ino); | |
8872 | if (in) { | |
8873 | dout(10) << " found cached " << *in << dendl; | |
8874 | open_ino_finish(ino, info, in->authority().first); | |
8875 | return; | |
8876 | } | |
8877 | ||
8878 | if (ret) { | |
8879 | do_open_ino(ino, info, ret); | |
8880 | return; | |
8881 | } | |
8882 | ||
8883 | mds_rank_t hint = info.auth_hint; | |
8884 | ret = open_ino_traverse_dir(ino, NULL, info.ancestors, | |
8885 | info.discover, info.want_xlocked, &hint); | |
8886 | if (ret > 0) | |
8887 | return; | |
8888 | if (hint != mds->get_nodeid()) | |
8889 | info.auth_hint = hint; | |
8890 | do_open_ino(ino, info, ret); | |
8891 | } | |
8892 | ||
9f95a23c | 8893 | void MDCache::_open_ino_fetch_dir(inodeno_t ino, const cref_t<MMDSOpenIno> &m, CDir *dir, bool parent) |
7c673cae FG |
8894 | { |
8895 | if (dir->state_test(CDir::STATE_REJOINUNDEF)) | |
11fdf7f2 | 8896 | ceph_assert(dir->get_inode()->dirfragtree.is_leaf(dir->get_frag())); |
7c673cae | 8897 | dir->fetch(new C_MDC_OpenInoTraverseDir(this, ino, m, parent)); |
11fdf7f2 TL |
8898 | if (mds->logger) |
8899 | mds->logger->inc(l_mds_openino_dir_fetch); | |
7c673cae FG |
8900 | } |
8901 | ||
9f95a23c | 8902 | int MDCache::open_ino_traverse_dir(inodeno_t ino, const cref_t<MMDSOpenIno> &m, |
11fdf7f2 | 8903 | const vector<inode_backpointer_t>& ancestors, |
7c673cae FG |
8904 | bool discover, bool want_xlocked, mds_rank_t *hint) |
8905 | { | |
8906 | dout(10) << "open_ino_traverse_dir ino " << ino << " " << ancestors << dendl; | |
8907 | int err = 0; | |
8908 | for (unsigned i = 0; i < ancestors.size(); i++) { | |
11fdf7f2 TL |
8909 | const auto& ancestor = ancestors.at(i); |
8910 | CInode *diri = get_inode(ancestor.dirino); | |
7c673cae FG |
8911 | |
8912 | if (!diri) { | |
11fdf7f2 TL |
8913 | if (discover && MDS_INO_IS_MDSDIR(ancestor.dirino)) { |
8914 | open_foreign_mdsdir(ancestor.dirino, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
7c673cae FG |
8915 | return 1; |
8916 | } | |
8917 | continue; | |
8918 | } | |
8919 | ||
8920 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) { | |
8921 | CDir *dir = diri->get_parent_dir(); | |
8922 | while (dir->state_test(CDir::STATE_REJOINUNDEF) && | |
8923 | dir->get_inode()->state_test(CInode::STATE_REJOINUNDEF)) | |
8924 | dir = dir->get_inode()->get_parent_dir(); | |
8925 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8926 | return 1; | |
8927 | } | |
8928 | ||
8929 | if (!diri->is_dir()) { | |
8930 | dout(10) << " " << *diri << " is not dir" << dendl; | |
8931 | if (i == 0) | |
f67539c2 | 8932 | err = -CEPHFS_ENOTDIR; |
7c673cae FG |
8933 | break; |
8934 | } | |
8935 | ||
11fdf7f2 | 8936 | const string& name = ancestor.dname; |
7c673cae FG |
8937 | frag_t fg = diri->pick_dirfrag(name); |
8938 | CDir *dir = diri->get_dirfrag(fg); | |
8939 | if (!dir) { | |
8940 | if (diri->is_auth()) { | |
8941 | if (diri->is_frozen()) { | |
8942 | dout(10) << " " << *diri << " is frozen, waiting " << dendl; | |
8943 | diri->add_waiter(CDir::WAIT_UNFREEZE, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8944 | return 1; | |
8945 | } | |
8946 | dir = diri->get_or_open_dirfrag(this, fg); | |
8947 | } else if (discover) { | |
8948 | open_remote_dirfrag(diri, fg, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8949 | return 1; | |
8950 | } | |
8951 | } | |
8952 | if (dir) { | |
11fdf7f2 | 8953 | inodeno_t next_ino = i > 0 ? ancestors.at(i-1).dirino : ino; |
7c673cae FG |
8954 | CDentry *dn = dir->lookup(name); |
8955 | CDentry::linkage_t *dnl = dn ? dn->get_linkage() : NULL; | |
8956 | if (dir->is_auth()) { | |
8957 | if (dnl && dnl->is_primary() && | |
8958 | dnl->get_inode()->state_test(CInode::STATE_REJOINUNDEF)) { | |
8959 | dout(10) << " fetching undef " << *dnl->get_inode() << dendl; | |
8960 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8961 | return 1; | |
8962 | } | |
8963 | ||
8964 | if (!dnl && !dir->is_complete() && | |
8965 | (!dir->has_bloom() || dir->is_in_bloom(name))) { | |
8966 | dout(10) << " fetching incomplete " << *dir << dendl; | |
8967 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8968 | return 1; | |
8969 | } | |
8970 | ||
8971 | dout(10) << " no ino " << next_ino << " in " << *dir << dendl; | |
8972 | if (i == 0) | |
f67539c2 | 8973 | err = -CEPHFS_ENOENT; |
7c673cae FG |
8974 | } else if (discover) { |
8975 | if (!dnl) { | |
8976 | filepath path(name, 0); | |
8977 | discover_path(dir, CEPH_NOSNAP, path, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0), | |
8978 | (i == 0 && want_xlocked)); | |
8979 | return 1; | |
8980 | } | |
8981 | if (dnl->is_null() && !dn->lock.can_read(-1)) { | |
8982 | dout(10) << " null " << *dn << " is not readable, waiting" << dendl; | |
8983 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8984 | return 1; | |
8985 | } | |
8986 | dout(10) << " no ino " << next_ino << " in " << *dir << dendl; | |
8987 | if (i == 0) | |
f67539c2 | 8988 | err = -CEPHFS_ENOENT; |
7c673cae FG |
8989 | } |
8990 | } | |
8991 | if (hint && i == 0) | |
8992 | *hint = dir ? dir->authority().first : diri->authority().first; | |
8993 | break; | |
8994 | } | |
8995 | return err; | |
8996 | } | |
8997 | ||
8998 | void MDCache::open_ino_finish(inodeno_t ino, open_ino_info_t& info, int ret) | |
8999 | { | |
9000 | dout(10) << "open_ino_finish ino " << ino << " ret " << ret << dendl; | |
9001 | ||
11fdf7f2 | 9002 | MDSContext::vec waiters; |
7c673cae FG |
9003 | waiters.swap(info.waiters); |
9004 | opening_inodes.erase(ino); | |
9005 | finish_contexts(g_ceph_context, waiters, ret); | |
9006 | } | |
9007 | ||
9008 | void MDCache::do_open_ino(inodeno_t ino, open_ino_info_t& info, int err) | |
9009 | { | |
f67539c2 | 9010 | if (err < 0 && err != -CEPHFS_EAGAIN) { |
7c673cae | 9011 | info.checked.clear(); |
7c673cae FG |
9012 | info.checking = MDS_RANK_NONE; |
9013 | info.check_peers = true; | |
9014 | info.fetch_backtrace = true; | |
9015 | if (info.discover) { | |
9016 | info.discover = false; | |
9017 | info.ancestors.clear(); | |
9018 | } | |
f67539c2 | 9019 | if (err != -CEPHFS_ENOENT && err != -CEPHFS_ENOTDIR) |
7c673cae FG |
9020 | info.last_err = err; |
9021 | } | |
9022 | ||
d2e6a577 FG |
9023 | if (info.check_peers || info.discover) { |
9024 | if (info.discover) { | |
9025 | // got backtrace from peer, but failed to find inode. re-check peers | |
9026 | info.discover = false; | |
9027 | info.ancestors.clear(); | |
9028 | info.checked.clear(); | |
9029 | } | |
7c673cae FG |
9030 | info.check_peers = false; |
9031 | info.checking = MDS_RANK_NONE; | |
9032 | do_open_ino_peer(ino, info); | |
9033 | } else if (info.fetch_backtrace) { | |
9034 | info.check_peers = true; | |
9035 | info.fetch_backtrace = false; | |
9036 | info.checking = mds->get_nodeid(); | |
9037 | info.checked.clear(); | |
7c673cae FG |
9038 | C_IO_MDC_OpenInoBacktraceFetched *fin = |
9039 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
9040 | fetch_backtrace(ino, info.pool, fin->bl, | |
9041 | new C_OnFinisher(fin, mds->finisher)); | |
9042 | } else { | |
11fdf7f2 | 9043 | ceph_assert(!info.ancestors.empty()); |
7c673cae | 9044 | info.checking = mds->get_nodeid(); |
b3b6e05e | 9045 | open_ino(info.ancestors[0].dirino, mds->get_metadata_pool(), |
7c673cae FG |
9046 | new C_MDC_OpenInoParentOpened(this, ino), info.want_replica); |
9047 | } | |
9048 | } | |
9049 | ||
9050 | void MDCache::do_open_ino_peer(inodeno_t ino, open_ino_info_t& info) | |
9051 | { | |
9052 | set<mds_rank_t> all, active; | |
9053 | mds->mdsmap->get_mds_set(all); | |
7c673cae | 9054 | if (mds->get_state() == MDSMap::STATE_REJOIN) |
1adf2230 AA |
9055 | mds->mdsmap->get_mds_set_lower_bound(active, MDSMap::STATE_REJOIN); |
9056 | else | |
9057 | mds->mdsmap->get_mds_set_lower_bound(active, MDSMap::STATE_CLIENTREPLAY); | |
7c673cae FG |
9058 | |
9059 | dout(10) << "do_open_ino_peer " << ino << " active " << active | |
9060 | << " all " << all << " checked " << info.checked << dendl; | |
9061 | ||
11fdf7f2 | 9062 | mds_rank_t whoami = mds->get_nodeid(); |
7c673cae | 9063 | mds_rank_t peer = MDS_RANK_NONE; |
11fdf7f2 | 9064 | if (info.auth_hint >= 0 && info.auth_hint != whoami) { |
7c673cae FG |
9065 | if (active.count(info.auth_hint)) { |
9066 | peer = info.auth_hint; | |
9067 | info.auth_hint = MDS_RANK_NONE; | |
9068 | } | |
9069 | } else { | |
9070 | for (set<mds_rank_t>::iterator p = active.begin(); p != active.end(); ++p) | |
11fdf7f2 | 9071 | if (*p != whoami && info.checked.count(*p) == 0) { |
7c673cae FG |
9072 | peer = *p; |
9073 | break; | |
9074 | } | |
9075 | } | |
9076 | if (peer < 0) { | |
11fdf7f2 | 9077 | all.erase(whoami); |
d2e6a577 | 9078 | if (all != info.checked) { |
7c673cae FG |
9079 | dout(10) << " waiting for more peers to be active" << dendl; |
9080 | } else { | |
9081 | dout(10) << " all MDS peers have been checked " << dendl; | |
9082 | do_open_ino(ino, info, 0); | |
9083 | } | |
9084 | } else { | |
9085 | info.checking = peer; | |
9086 | vector<inode_backpointer_t> *pa = NULL; | |
9087 | // got backtrace from peer or backtrace just fetched | |
9088 | if (info.discover || !info.fetch_backtrace) | |
9089 | pa = &info.ancestors; | |
9f95a23c | 9090 | mds->send_message_mds(make_message<MMDSOpenIno>(info.tid, ino, pa), peer); |
11fdf7f2 TL |
9091 | if (mds->logger) |
9092 | mds->logger->inc(l_mds_openino_peer_discover); | |
7c673cae FG |
9093 | } |
9094 | } | |
9095 | ||
9f95a23c | 9096 | void MDCache::handle_open_ino(const cref_t<MMDSOpenIno> &m, int err) |
7c673cae FG |
9097 | { |
9098 | if (mds->get_state() < MDSMap::STATE_REJOIN && | |
9099 | mds->get_want_state() != CEPH_MDS_STATE_REJOIN) { | |
7c673cae FG |
9100 | return; |
9101 | } | |
9102 | ||
9103 | dout(10) << "handle_open_ino " << *m << " err " << err << dendl; | |
9104 | ||
11fdf7f2 | 9105 | auto from = mds_rank_t(m->get_source().num()); |
7c673cae | 9106 | inodeno_t ino = m->ino; |
9f95a23c | 9107 | ref_t<MMDSOpenInoReply> reply; |
7c673cae FG |
9108 | CInode *in = get_inode(ino); |
9109 | if (in) { | |
9110 | dout(10) << " have " << *in << dendl; | |
9f95a23c | 9111 | reply = make_message<MMDSOpenInoReply>(m->get_tid(), ino, mds_rank_t(0)); |
7c673cae FG |
9112 | if (in->is_auth()) { |
9113 | touch_inode(in); | |
9114 | while (1) { | |
9115 | CDentry *pdn = in->get_parent_dn(); | |
9116 | if (!pdn) | |
9117 | break; | |
9118 | CInode *diri = pdn->get_dir()->get_inode(); | |
94b18763 | 9119 | reply->ancestors.push_back(inode_backpointer_t(diri->ino(), pdn->get_name(), |
f67539c2 | 9120 | in->get_version())); |
7c673cae FG |
9121 | in = diri; |
9122 | } | |
9123 | } else { | |
9124 | reply->hint = in->authority().first; | |
9125 | } | |
9126 | } else if (err < 0) { | |
9f95a23c | 9127 | reply = make_message<MMDSOpenInoReply>(m->get_tid(), ino, MDS_RANK_NONE, err); |
7c673cae FG |
9128 | } else { |
9129 | mds_rank_t hint = MDS_RANK_NONE; | |
9130 | int ret = open_ino_traverse_dir(ino, m, m->ancestors, false, false, &hint); | |
9131 | if (ret > 0) | |
9132 | return; | |
9f95a23c | 9133 | reply = make_message<MMDSOpenInoReply>(m->get_tid(), ino, hint, ret); |
7c673cae | 9134 | } |
11fdf7f2 | 9135 | mds->send_message_mds(reply, from); |
7c673cae FG |
9136 | } |
9137 | ||
9f95a23c | 9138 | void MDCache::handle_open_ino_reply(const cref_t<MMDSOpenInoReply> &m) |
7c673cae FG |
9139 | { |
9140 | dout(10) << "handle_open_ino_reply " << *m << dendl; | |
9141 | ||
9142 | inodeno_t ino = m->ino; | |
9143 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
9144 | auto it = opening_inodes.find(ino); | |
9145 | if (it != opening_inodes.end() && it->second.checking == from) { | |
9146 | open_ino_info_t& info = it->second; | |
9147 | info.checking = MDS_RANK_NONE; | |
9148 | info.checked.insert(from); | |
9149 | ||
9150 | CInode *in = get_inode(ino); | |
9151 | if (in) { | |
9152 | dout(10) << " found cached " << *in << dendl; | |
9153 | open_ino_finish(ino, info, in->authority().first); | |
9154 | } else if (!m->ancestors.empty()) { | |
9155 | dout(10) << " found ino " << ino << " on mds." << from << dendl; | |
9156 | if (!info.want_replica) { | |
9157 | open_ino_finish(ino, info, from); | |
7c673cae FG |
9158 | return; |
9159 | } | |
9160 | ||
9161 | info.ancestors = m->ancestors; | |
9162 | info.auth_hint = from; | |
9163 | info.checking = mds->get_nodeid(); | |
9164 | info.discover = true; | |
9165 | _open_ino_traverse_dir(ino, info, 0); | |
9166 | } else if (m->error) { | |
9167 | dout(10) << " error " << m->error << " from mds." << from << dendl; | |
9168 | do_open_ino(ino, info, m->error); | |
9169 | } else { | |
9170 | if (m->hint >= 0 && m->hint != mds->get_nodeid()) { | |
9171 | info.auth_hint = m->hint; | |
9172 | info.checked.erase(m->hint); | |
9173 | } | |
9174 | do_open_ino_peer(ino, info); | |
9175 | } | |
9176 | } | |
7c673cae FG |
9177 | } |
9178 | ||
9179 | void MDCache::kick_open_ino_peers(mds_rank_t who) | |
9180 | { | |
9181 | dout(10) << "kick_open_ino_peers mds." << who << dendl; | |
9182 | ||
9183 | for (map<inodeno_t, open_ino_info_t>::iterator p = opening_inodes.begin(); | |
9184 | p != opening_inodes.end(); | |
9185 | ++p) { | |
9186 | open_ino_info_t& info = p->second; | |
9187 | if (info.checking == who) { | |
9188 | dout(10) << " kicking ino " << p->first << " who was checking mds." << who << dendl; | |
9189 | info.checking = MDS_RANK_NONE; | |
9190 | do_open_ino_peer(p->first, info); | |
9191 | } else if (info.checking == MDS_RANK_NONE) { | |
9192 | dout(10) << " kicking ino " << p->first << " who was waiting" << dendl; | |
9193 | do_open_ino_peer(p->first, info); | |
9194 | } | |
9195 | } | |
9196 | } | |
9197 | ||
11fdf7f2 | 9198 | void MDCache::open_ino(inodeno_t ino, int64_t pool, MDSContext* fin, |
f91f0fd5 TL |
9199 | bool want_replica, bool want_xlocked, |
9200 | vector<inode_backpointer_t> *ancestors_hint, | |
9201 | mds_rank_t auth_hint) | |
7c673cae FG |
9202 | { |
9203 | dout(10) << "open_ino " << ino << " pool " << pool << " want_replica " | |
9204 | << want_replica << dendl; | |
9205 | ||
11fdf7f2 TL |
9206 | auto it = opening_inodes.find(ino); |
9207 | if (it != opening_inodes.end()) { | |
9208 | open_ino_info_t& info = it->second; | |
7c673cae FG |
9209 | if (want_replica) { |
9210 | info.want_replica = true; | |
9211 | if (want_xlocked && !info.want_xlocked) { | |
9212 | if (!info.ancestors.empty()) { | |
9213 | CInode *diri = get_inode(info.ancestors[0].dirino); | |
9214 | if (diri) { | |
9215 | frag_t fg = diri->pick_dirfrag(info.ancestors[0].dname); | |
9216 | CDir *dir = diri->get_dirfrag(fg); | |
9217 | if (dir && !dir->is_auth()) { | |
9218 | filepath path(info.ancestors[0].dname, 0); | |
9219 | discover_path(dir, CEPH_NOSNAP, path, NULL, true); | |
9220 | } | |
9221 | } | |
9222 | } | |
9223 | info.want_xlocked = true; | |
9224 | } | |
9225 | } | |
9226 | info.waiters.push_back(fin); | |
9227 | } else { | |
9228 | open_ino_info_t& info = opening_inodes[ino]; | |
7c673cae FG |
9229 | info.want_replica = want_replica; |
9230 | info.want_xlocked = want_xlocked; | |
9231 | info.tid = ++open_ino_last_tid; | |
9232 | info.pool = pool >= 0 ? pool : default_file_layout.pool_id; | |
9233 | info.waiters.push_back(fin); | |
f91f0fd5 TL |
9234 | if (auth_hint != MDS_RANK_NONE) |
9235 | info.auth_hint = auth_hint; | |
9236 | if (ancestors_hint) { | |
9237 | info.ancestors = std::move(*ancestors_hint); | |
11fdf7f2 TL |
9238 | info.fetch_backtrace = false; |
9239 | info.checking = mds->get_nodeid(); | |
9240 | _open_ino_traverse_dir(ino, info, 0); | |
9241 | } else { | |
9242 | do_open_ino(ino, info, 0); | |
9243 | } | |
7c673cae FG |
9244 | } |
9245 | } | |
9246 | ||
9247 | /* ---------------------------- */ | |
9248 | ||
9249 | /* | |
9250 | * search for a given inode on MDS peers. optionally start with the given node. | |
9251 | ||
9252 | ||
9253 | TODO | |
9254 | - recover from mds node failure, recovery | |
9255 | - traverse path | |
9256 | ||
9257 | */ | |
9f95a23c TL |
9258 | void MDCache::find_ino_peers(inodeno_t ino, MDSContext *c, |
9259 | mds_rank_t hint, bool path_locked) | |
7c673cae FG |
9260 | { |
9261 | dout(5) << "find_ino_peers " << ino << " hint " << hint << dendl; | |
b32b8144 FG |
9262 | CInode *in = get_inode(ino); |
9263 | if (in && in->state_test(CInode::STATE_PURGING)) { | |
f67539c2 | 9264 | c->complete(-CEPHFS_ESTALE); |
b32b8144 FG |
9265 | return; |
9266 | } | |
11fdf7f2 | 9267 | ceph_assert(!in); |
7c673cae FG |
9268 | |
9269 | ceph_tid_t tid = ++find_ino_peer_last_tid; | |
9270 | find_ino_peer_info_t& fip = find_ino_peer[tid]; | |
9271 | fip.ino = ino; | |
9272 | fip.tid = tid; | |
9273 | fip.fin = c; | |
9f95a23c | 9274 | fip.path_locked = path_locked; |
7c673cae | 9275 | fip.hint = hint; |
7c673cae FG |
9276 | _do_find_ino_peer(fip); |
9277 | } | |
9278 | ||
9279 | void MDCache::_do_find_ino_peer(find_ino_peer_info_t& fip) | |
9280 | { | |
9281 | set<mds_rank_t> all, active; | |
9282 | mds->mdsmap->get_mds_set(all); | |
1adf2230 | 9283 | mds->mdsmap->get_mds_set_lower_bound(active, MDSMap::STATE_CLIENTREPLAY); |
7c673cae FG |
9284 | |
9285 | dout(10) << "_do_find_ino_peer " << fip.tid << " " << fip.ino | |
9286 | << " active " << active << " all " << all | |
9287 | << " checked " << fip.checked | |
9288 | << dendl; | |
9289 | ||
9290 | mds_rank_t m = MDS_RANK_NONE; | |
9291 | if (fip.hint >= 0) { | |
9292 | m = fip.hint; | |
9293 | fip.hint = MDS_RANK_NONE; | |
9294 | } else { | |
9295 | for (set<mds_rank_t>::iterator p = active.begin(); p != active.end(); ++p) | |
9296 | if (*p != mds->get_nodeid() && | |
9297 | fip.checked.count(*p) == 0) { | |
9298 | m = *p; | |
9299 | break; | |
9300 | } | |
9301 | } | |
9302 | if (m == MDS_RANK_NONE) { | |
d2e6a577 FG |
9303 | all.erase(mds->get_nodeid()); |
9304 | if (all != fip.checked) { | |
7c673cae FG |
9305 | dout(10) << "_do_find_ino_peer waiting for more peers to be active" << dendl; |
9306 | } else { | |
9307 | dout(10) << "_do_find_ino_peer failed on " << fip.ino << dendl; | |
f67539c2 | 9308 | fip.fin->complete(-CEPHFS_ESTALE); |
7c673cae FG |
9309 | find_ino_peer.erase(fip.tid); |
9310 | } | |
9311 | } else { | |
9312 | fip.checking = m; | |
9f95a23c | 9313 | mds->send_message_mds(make_message<MMDSFindIno>(fip.tid, fip.ino), m); |
7c673cae FG |
9314 | } |
9315 | } | |
9316 | ||
9f95a23c | 9317 | void MDCache::handle_find_ino(const cref_t<MMDSFindIno> &m) |
7c673cae FG |
9318 | { |
9319 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
7c673cae FG |
9320 | return; |
9321 | } | |
9322 | ||
9323 | dout(10) << "handle_find_ino " << *m << dendl; | |
9f95a23c | 9324 | auto r = make_message<MMDSFindInoReply>(m->tid); |
7c673cae FG |
9325 | CInode *in = get_inode(m->ino); |
9326 | if (in) { | |
9327 | in->make_path(r->path); | |
9328 | dout(10) << " have " << r->path << " " << *in << dendl; | |
33c7a0ef TL |
9329 | |
9330 | /* | |
9331 | * If the the CInode was just created by using openc in current | |
9332 | * auth MDS, but the client just sends a getattr request to another | |
9333 | * replica MDS. Then here it will make a path of '#INODE-NUMBER' | |
9334 | * only because the CInode hasn't been linked yet, and the replica | |
9335 | * MDS will keep retrying until the auth MDS flushes the mdlog and | |
9336 | * the C_MDS_openc_finish and link_primary_inode are called at most | |
9337 | * 5 seconds later. | |
9338 | */ | |
9339 | if (!in->get_parent_dn() && in->is_auth()) { | |
9340 | mds->mdlog->flush(); | |
9341 | } | |
7c673cae | 9342 | } |
11fdf7f2 | 9343 | mds->send_message_mds(r, mds_rank_t(m->get_source().num())); |
7c673cae FG |
9344 | } |
9345 | ||
9346 | ||
9f95a23c | 9347 | void MDCache::handle_find_ino_reply(const cref_t<MMDSFindInoReply> &m) |
7c673cae | 9348 | { |
9f95a23c | 9349 | auto p = find_ino_peer.find(m->tid); |
7c673cae FG |
9350 | if (p != find_ino_peer.end()) { |
9351 | dout(10) << "handle_find_ino_reply " << *m << dendl; | |
9352 | find_ino_peer_info_t& fip = p->second; | |
9353 | ||
9354 | // success? | |
9355 | if (get_inode(fip.ino)) { | |
9356 | dout(10) << "handle_find_ino_reply successfully found " << fip.ino << dendl; | |
9357 | mds->queue_waiter(fip.fin); | |
9358 | find_ino_peer.erase(p); | |
7c673cae FG |
9359 | return; |
9360 | } | |
9361 | ||
9362 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
9363 | if (fip.checking == from) | |
9364 | fip.checking = MDS_RANK_NONE; | |
9365 | fip.checked.insert(from); | |
9366 | ||
9367 | if (!m->path.empty()) { | |
9368 | // we got a path! | |
9369 | vector<CDentry*> trace; | |
11fdf7f2 | 9370 | CF_MDS_RetryMessageFactory cf(mds, m); |
7c673cae | 9371 | MDRequestRef null_ref; |
9f95a23c TL |
9372 | int flags = MDS_TRAVERSE_DISCOVER; |
9373 | if (fip.path_locked) | |
9374 | flags |= MDS_TRAVERSE_PATH_LOCKED; | |
9375 | int r = path_traverse(null_ref, cf, m->path, flags, &trace); | |
7c673cae FG |
9376 | if (r > 0) |
9377 | return; | |
9378 | dout(0) << "handle_find_ino_reply failed with " << r << " on " << m->path | |
9379 | << ", retrying" << dendl; | |
9380 | fip.checked.clear(); | |
9381 | _do_find_ino_peer(fip); | |
9382 | } else { | |
9383 | // nope, continue. | |
9384 | _do_find_ino_peer(fip); | |
9385 | } | |
9386 | } else { | |
9387 | dout(10) << "handle_find_ino_reply tid " << m->tid << " dne" << dendl; | |
9388 | } | |
7c673cae FG |
9389 | } |
9390 | ||
9391 | void MDCache::kick_find_ino_peers(mds_rank_t who) | |
9392 | { | |
9393 | // find_ino_peers requests we should move on from | |
9394 | for (map<ceph_tid_t,find_ino_peer_info_t>::iterator p = find_ino_peer.begin(); | |
9395 | p != find_ino_peer.end(); | |
9396 | ++p) { | |
9397 | find_ino_peer_info_t& fip = p->second; | |
9398 | if (fip.checking == who) { | |
9399 | dout(10) << "kicking find_ino_peer " << fip.tid << " who was checking mds." << who << dendl; | |
9400 | fip.checking = MDS_RANK_NONE; | |
9401 | _do_find_ino_peer(fip); | |
9402 | } else if (fip.checking == MDS_RANK_NONE) { | |
9403 | dout(10) << "kicking find_ino_peer " << fip.tid << " who was waiting" << dendl; | |
9404 | _do_find_ino_peer(fip); | |
9405 | } | |
9406 | } | |
9407 | } | |
9408 | ||
9409 | /* ---------------------------- */ | |
9410 | ||
9411 | int MDCache::get_num_client_requests() | |
9412 | { | |
9413 | int count = 0; | |
9414 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
9415 | p != active_requests.end(); | |
9416 | ++p) { | |
9417 | MDRequestRef& mdr = p->second; | |
f67539c2 | 9418 | if (mdr->reqid.name.is_client() && !mdr->is_peer()) |
7c673cae FG |
9419 | count++; |
9420 | } | |
9421 | return count; | |
9422 | } | |
9423 | ||
9f95a23c | 9424 | MDRequestRef MDCache::request_start(const cref_t<MClientRequest>& req) |
7c673cae | 9425 | { |
f67539c2 | 9426 | // did we win a forward race against a peer? |
7c673cae FG |
9427 | if (active_requests.count(req->get_reqid())) { |
9428 | MDRequestRef& mdr = active_requests[req->get_reqid()]; | |
11fdf7f2 | 9429 | ceph_assert(mdr); |
f67539c2 | 9430 | if (mdr->is_peer()) { |
7c673cae FG |
9431 | dout(10) << "request_start already had " << *mdr << ", waiting for finish" << dendl; |
9432 | mdr->more()->waiting_for_finish.push_back(new C_MDS_RetryMessage(mds, req)); | |
9433 | } else { | |
9434 | dout(10) << "request_start already processing " << *mdr << ", dropping new msg" << dendl; | |
7c673cae FG |
9435 | } |
9436 | return MDRequestRef(); | |
9437 | } | |
9438 | ||
9439 | // register new client request | |
9440 | MDRequestImpl::Params params; | |
9441 | params.reqid = req->get_reqid(); | |
9442 | params.attempt = req->get_num_fwd(); | |
9443 | params.client_req = req; | |
9444 | params.initiated = req->get_recv_stamp(); | |
9445 | params.throttled = req->get_throttle_stamp(); | |
9446 | params.all_read = req->get_recv_complete_stamp(); | |
9447 | params.dispatched = req->get_dispatch_stamp(); | |
9448 | ||
9449 | MDRequestRef mdr = | |
11fdf7f2 | 9450 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params*>(¶ms); |
7c673cae FG |
9451 | active_requests[params.reqid] = mdr; |
9452 | mdr->set_op_stamp(req->get_stamp()); | |
9453 | dout(7) << "request_start " << *mdr << dendl; | |
9454 | return mdr; | |
9455 | } | |
9456 | ||
f67539c2 | 9457 | MDRequestRef MDCache::request_start_peer(metareqid_t ri, __u32 attempt, const cref_t<Message> &m) |
7c673cae FG |
9458 | { |
9459 | int by = m->get_source().num(); | |
9460 | MDRequestImpl::Params params; | |
9461 | params.reqid = ri; | |
9462 | params.attempt = attempt; | |
f67539c2 TL |
9463 | params.triggering_peer_req = m; |
9464 | params.peer_to = by; | |
7c673cae FG |
9465 | params.initiated = m->get_recv_stamp(); |
9466 | params.throttled = m->get_throttle_stamp(); | |
9467 | params.all_read = m->get_recv_complete_stamp(); | |
9468 | params.dispatched = m->get_dispatch_stamp(); | |
9469 | MDRequestRef mdr = | |
11fdf7f2 TL |
9470 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params*>(¶ms); |
9471 | ceph_assert(active_requests.count(mdr->reqid) == 0); | |
7c673cae | 9472 | active_requests[mdr->reqid] = mdr; |
f67539c2 | 9473 | dout(7) << "request_start_peer " << *mdr << " by mds." << by << dendl; |
7c673cae FG |
9474 | return mdr; |
9475 | } | |
9476 | ||
9477 | MDRequestRef MDCache::request_start_internal(int op) | |
9478 | { | |
91327a77 | 9479 | utime_t now = ceph_clock_now(); |
7c673cae FG |
9480 | MDRequestImpl::Params params; |
9481 | params.reqid.name = entity_name_t::MDS(mds->get_nodeid()); | |
9482 | params.reqid.tid = mds->issue_tid(); | |
91327a77 AA |
9483 | params.initiated = now; |
9484 | params.throttled = now; | |
9485 | params.all_read = now; | |
9486 | params.dispatched = now; | |
7c673cae FG |
9487 | params.internal_op = op; |
9488 | MDRequestRef mdr = | |
11fdf7f2 | 9489 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params*>(¶ms); |
7c673cae | 9490 | |
11fdf7f2 | 9491 | ceph_assert(active_requests.count(mdr->reqid) == 0); |
7c673cae FG |
9492 | active_requests[mdr->reqid] = mdr; |
9493 | dout(7) << "request_start_internal " << *mdr << " op " << op << dendl; | |
9494 | return mdr; | |
9495 | } | |
9496 | ||
9497 | MDRequestRef MDCache::request_get(metareqid_t rid) | |
9498 | { | |
9499 | ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.find(rid); | |
11fdf7f2 | 9500 | ceph_assert(p != active_requests.end()); |
7c673cae FG |
9501 | dout(7) << "request_get " << rid << " " << *p->second << dendl; |
9502 | return p->second; | |
9503 | } | |
9504 | ||
9505 | void MDCache::request_finish(MDRequestRef& mdr) | |
9506 | { | |
9507 | dout(7) << "request_finish " << *mdr << dendl; | |
9508 | mdr->mark_event("finishing request"); | |
9509 | ||
f67539c2 TL |
9510 | // peer finisher? |
9511 | if (mdr->has_more() && mdr->more()->peer_commit) { | |
9512 | Context *fin = mdr->more()->peer_commit; | |
9513 | mdr->more()->peer_commit = 0; | |
7c673cae FG |
9514 | int ret; |
9515 | if (mdr->aborted) { | |
9516 | mdr->aborted = false; | |
9517 | ret = -1; | |
f67539c2 | 9518 | mdr->more()->peer_rolling_back = true; |
7c673cae FG |
9519 | } else { |
9520 | ret = 0; | |
9521 | mdr->committing = true; | |
9522 | } | |
9523 | fin->complete(ret); // this must re-call request_finish. | |
9524 | return; | |
9525 | } | |
9526 | ||
d2e6a577 FG |
9527 | switch(mdr->internal_op) { |
9528 | case CEPH_MDS_OP_FRAGMENTDIR: | |
9529 | logger->inc(l_mdss_ireq_fragmentdir); | |
9530 | break; | |
9531 | case CEPH_MDS_OP_EXPORTDIR: | |
9532 | logger->inc(l_mdss_ireq_exportdir); | |
9533 | break; | |
9534 | case CEPH_MDS_OP_ENQUEUE_SCRUB: | |
9535 | logger->inc(l_mdss_ireq_enqueue_scrub); | |
9536 | break; | |
9537 | case CEPH_MDS_OP_FLUSH: | |
9538 | logger->inc(l_mdss_ireq_flush); | |
9539 | break; | |
9540 | case CEPH_MDS_OP_REPAIR_FRAGSTATS: | |
9541 | logger->inc(l_mdss_ireq_fragstats); | |
9542 | break; | |
9543 | case CEPH_MDS_OP_REPAIR_INODESTATS: | |
9544 | logger->inc(l_mdss_ireq_inodestats); | |
9545 | break; | |
9546 | } | |
9547 | ||
7c673cae FG |
9548 | request_cleanup(mdr); |
9549 | } | |
9550 | ||
9551 | ||
9552 | void MDCache::request_forward(MDRequestRef& mdr, mds_rank_t who, int port) | |
9553 | { | |
f91f0fd5 TL |
9554 | CachedStackStringStream css; |
9555 | *css << "forwarding request to mds." << who; | |
9556 | mdr->mark_event(css->strv()); | |
7c673cae FG |
9557 | if (mdr->client_request && mdr->client_request->get_source().is_client()) { |
9558 | dout(7) << "request_forward " << *mdr << " to mds." << who << " req " | |
9559 | << *mdr->client_request << dendl; | |
f91f0fd5 TL |
9560 | if (mdr->is_batch_head()) { |
9561 | mdr->release_batch_op()->forward(who); | |
9f95a23c TL |
9562 | } else { |
9563 | mds->forward_message_mds(mdr->release_client_request(), who); | |
9564 | } | |
7c673cae FG |
9565 | if (mds->logger) mds->logger->inc(l_mds_forward); |
9566 | } else if (mdr->internal_op >= 0) { | |
9567 | dout(10) << "request_forward on internal op; cancelling" << dendl; | |
f67539c2 | 9568 | mdr->internal_op_finish->complete(-CEPHFS_EXDEV); |
7c673cae FG |
9569 | } else { |
9570 | dout(7) << "request_forward drop " << *mdr << " req " << *mdr->client_request | |
9571 | << " was from mds" << dendl; | |
9572 | } | |
9573 | request_cleanup(mdr); | |
9574 | } | |
9575 | ||
9576 | ||
9577 | void MDCache::dispatch_request(MDRequestRef& mdr) | |
9578 | { | |
9579 | if (mdr->client_request) { | |
9580 | mds->server->dispatch_client_request(mdr); | |
f67539c2 TL |
9581 | } else if (mdr->peer_request) { |
9582 | mds->server->dispatch_peer_request(mdr); | |
7c673cae FG |
9583 | } else { |
9584 | switch (mdr->internal_op) { | |
9585 | case CEPH_MDS_OP_FRAGMENTDIR: | |
9586 | dispatch_fragment_dir(mdr); | |
9587 | break; | |
9588 | case CEPH_MDS_OP_EXPORTDIR: | |
9589 | migrator->dispatch_export_dir(mdr, 0); | |
9590 | break; | |
9591 | case CEPH_MDS_OP_ENQUEUE_SCRUB: | |
9592 | enqueue_scrub_work(mdr); | |
9593 | break; | |
9594 | case CEPH_MDS_OP_FLUSH: | |
9595 | flush_dentry_work(mdr); | |
9596 | break; | |
9597 | case CEPH_MDS_OP_REPAIR_FRAGSTATS: | |
9598 | repair_dirfrag_stats_work(mdr); | |
9599 | break; | |
9600 | case CEPH_MDS_OP_REPAIR_INODESTATS: | |
9601 | repair_inode_stats_work(mdr); | |
9602 | break; | |
f67539c2 TL |
9603 | case CEPH_MDS_OP_RDLOCK_FRAGSSTATS: |
9604 | rdlock_dirfrags_stats_work(mdr); | |
11fdf7f2 | 9605 | break; |
7c673cae FG |
9606 | default: |
9607 | ceph_abort(); | |
9608 | } | |
9609 | } | |
9610 | } | |
9611 | ||
9612 | ||
9613 | void MDCache::request_drop_foreign_locks(MDRequestRef& mdr) | |
9614 | { | |
9615 | if (!mdr->has_more()) | |
9616 | return; | |
9617 | ||
f67539c2 | 9618 | // clean up peers |
7c673cae | 9619 | // (will implicitly drop remote dn pins) |
f67539c2 TL |
9620 | for (set<mds_rank_t>::iterator p = mdr->more()->peers.begin(); |
9621 | p != mdr->more()->peers.end(); | |
7c673cae | 9622 | ++p) { |
f67539c2 TL |
9623 | auto r = make_message<MMDSPeerRequest>(mdr->reqid, mdr->attempt, |
9624 | MMDSPeerRequest::OP_FINISH); | |
7c673cae FG |
9625 | |
9626 | if (mdr->killed && !mdr->committing) { | |
9627 | r->mark_abort(); | |
9628 | } else if (mdr->more()->srcdn_auth_mds == *p && | |
9629 | mdr->more()->inode_import.length() > 0) { | |
9630 | // information about rename imported caps | |
f67539c2 | 9631 | r->inode_export = std::move(mdr->more()->inode_import); |
7c673cae FG |
9632 | } |
9633 | ||
9634 | mds->send_message_mds(r, *p); | |
9635 | } | |
9636 | ||
9637 | /* strip foreign xlocks out of lock lists, since the OP_FINISH drops them | |
9638 | * implicitly. Note that we don't call the finishers -- there shouldn't | |
9639 | * be any on a remote lock and the request finish wakes up all | |
9640 | * the waiters anyway! */ | |
7c673cae | 9641 | |
11fdf7f2 TL |
9642 | for (auto it = mdr->locks.begin(); it != mdr->locks.end(); ) { |
9643 | SimpleLock *lock = it->lock; | |
9644 | if (it->is_xlock() && !lock->get_parent()->is_auth()) { | |
9645 | dout(10) << "request_drop_foreign_locks forgetting lock " << *lock | |
9646 | << " on " << lock->get_parent() << dendl; | |
9647 | lock->put_xlock(); | |
9648 | mdr->locks.erase(it++); | |
9649 | } else if (it->is_remote_wrlock()) { | |
9650 | dout(10) << "request_drop_foreign_locks forgetting remote_wrlock " << *lock | |
9651 | << " on mds." << it->wrlock_target << " on " << *lock->get_parent() << dendl; | |
9652 | if (it->is_wrlock()) { | |
9653 | it->clear_remote_wrlock(); | |
9654 | ++it; | |
9655 | } else { | |
9656 | mdr->locks.erase(it++); | |
9657 | } | |
9658 | } else { | |
9659 | ++it; | |
9660 | } | |
7c673cae FG |
9661 | } |
9662 | ||
f67539c2 | 9663 | mdr->more()->peers.clear(); /* we no longer have requests out to them, and |
7c673cae FG |
9664 | * leaving them in can cause double-notifies as |
9665 | * this function can get called more than once */ | |
9666 | } | |
9667 | ||
9668 | void MDCache::request_drop_non_rdlocks(MDRequestRef& mdr) | |
9669 | { | |
9670 | request_drop_foreign_locks(mdr); | |
9671 | mds->locker->drop_non_rdlocks(mdr.get()); | |
9672 | } | |
9673 | ||
9674 | void MDCache::request_drop_locks(MDRequestRef& mdr) | |
9675 | { | |
9676 | request_drop_foreign_locks(mdr); | |
9677 | mds->locker->drop_locks(mdr.get()); | |
9678 | } | |
9679 | ||
9680 | void MDCache::request_cleanup(MDRequestRef& mdr) | |
9681 | { | |
9682 | dout(15) << "request_cleanup " << *mdr << dendl; | |
9683 | ||
9684 | if (mdr->has_more()) { | |
9685 | if (mdr->more()->is_ambiguous_auth) | |
9686 | mdr->clear_ambiguous_auth(); | |
9687 | if (!mdr->more()->waiting_for_finish.empty()) | |
9688 | mds->queue_waiters(mdr->more()->waiting_for_finish); | |
9689 | } | |
9690 | ||
9691 | request_drop_locks(mdr); | |
9692 | ||
9693 | // drop (local) auth pins | |
9694 | mdr->drop_local_auth_pins(); | |
9695 | ||
9696 | // drop stickydirs | |
11fdf7f2 | 9697 | mdr->put_stickydirs(); |
7c673cae FG |
9698 | |
9699 | mds->locker->kick_cap_releases(mdr); | |
9700 | ||
9701 | // drop cache pins | |
9702 | mdr->drop_pins(); | |
9703 | ||
9704 | // remove from session | |
9705 | mdr->item_session_request.remove_myself(); | |
9706 | ||
9707 | // remove from map | |
9708 | active_requests.erase(mdr->reqid); | |
9709 | ||
9710 | if (mds->logger) | |
9711 | log_stat(); | |
9712 | ||
9713 | mdr->mark_event("cleaned up request"); | |
9714 | } | |
9715 | ||
9716 | void MDCache::request_kill(MDRequestRef& mdr) | |
9717 | { | |
f67539c2 | 9718 | // rollback peer requests is tricky. just let the request proceed. |
94b18763 | 9719 | if (mdr->has_more() && |
f67539c2 | 9720 | (!mdr->more()->witnessed.empty() || !mdr->more()->waiting_on_peer.empty())) { |
9f95a23c | 9721 | if (!(mdr->locking_state & MutationImpl::ALL_LOCKED)) { |
11fdf7f2 | 9722 | ceph_assert(mdr->more()->witnessed.empty()); |
94b18763 | 9723 | mdr->aborted = true; |
f67539c2 | 9724 | dout(10) << "request_kill " << *mdr << " -- waiting for peer reply, delaying" << dendl; |
94b18763 | 9725 | } else { |
f67539c2 | 9726 | dout(10) << "request_kill " << *mdr << " -- already started peer prep, no-op" << dendl; |
94b18763 | 9727 | } |
7c673cae | 9728 | |
11fdf7f2 TL |
9729 | ceph_assert(mdr->used_prealloc_ino == 0); |
9730 | ceph_assert(mdr->prealloc_inos.empty()); | |
7c673cae FG |
9731 | |
9732 | mdr->session = NULL; | |
9733 | mdr->item_session_request.remove_myself(); | |
9734 | return; | |
9735 | } | |
9736 | ||
9737 | mdr->killed = true; | |
9738 | mdr->mark_event("killing request"); | |
9739 | ||
9740 | if (mdr->committing) { | |
f67539c2 TL |
9741 | dout(10) << "request_kill " << *mdr << " -- already committing, remove it from sesssion requests" << dendl; |
9742 | mdr->item_session_request.remove_myself(); | |
7c673cae FG |
9743 | } else { |
9744 | dout(10) << "request_kill " << *mdr << dendl; | |
9745 | request_cleanup(mdr); | |
9746 | } | |
9747 | } | |
9748 | ||
9749 | // ------------------------------------------------------------------------------- | |
9750 | // SNAPREALMS | |
9751 | ||
11fdf7f2 | 9752 | void MDCache::create_global_snaprealm() |
7c673cae | 9753 | { |
11fdf7f2 | 9754 | CInode *in = new CInode(this); // dummy inode |
b3b6e05e | 9755 | create_unlinked_system_inode(in, CEPH_INO_GLOBAL_SNAPREALM, S_IFDIR|0755); |
11fdf7f2 TL |
9756 | add_inode(in); |
9757 | global_snaprealm = in->snaprealm; | |
7c673cae FG |
9758 | } |
9759 | ||
11fdf7f2 | 9760 | void MDCache::do_realm_invalidate_and_update_notify(CInode *in, int snapop, bool notify_clients) |
7c673cae FG |
9761 | { |
9762 | dout(10) << "do_realm_invalidate_and_update_notify " << *in->snaprealm << " " << *in << dendl; | |
9763 | ||
9764 | vector<inodeno_t> split_inos; | |
9765 | vector<inodeno_t> split_realms; | |
9766 | ||
11fdf7f2 | 9767 | if (notify_clients) { |
11fdf7f2 TL |
9768 | if (snapop == CEPH_SNAP_OP_SPLIT) { |
9769 | // notify clients of update|split | |
f67539c2 | 9770 | for (auto p = in->snaprealm->inodes_with_caps.begin(); !p.end(); ++p) |
11fdf7f2 | 9771 | split_inos.push_back((*p)->ino()); |
7c673cae | 9772 | |
f67539c2 TL |
9773 | for (auto& r : in->snaprealm->open_children) |
9774 | split_realms.push_back(r->inode->ino()); | |
11fdf7f2 TL |
9775 | } |
9776 | } | |
7c673cae | 9777 | |
9f95a23c | 9778 | map<client_t, ref_t<MClientSnap>> updates; |
7c673cae FG |
9779 | list<SnapRealm*> q; |
9780 | q.push_back(in->snaprealm); | |
9781 | while (!q.empty()) { | |
9782 | SnapRealm *realm = q.front(); | |
9783 | q.pop_front(); | |
9784 | ||
9785 | dout(10) << " realm " << *realm << " on " << *realm->inode << dendl; | |
9786 | realm->invalidate_cached_snaps(); | |
9787 | ||
11fdf7f2 TL |
9788 | if (notify_clients) { |
9789 | for (const auto& p : realm->client_caps) { | |
9790 | const auto& client = p.first; | |
9791 | const auto& caps = p.second; | |
9792 | ceph_assert(!caps->empty()); | |
9793 | ||
9794 | auto em = updates.emplace(std::piecewise_construct, std::forward_as_tuple(client), std::forward_as_tuple()); | |
9795 | if (em.second) { | |
9f95a23c | 9796 | auto update = make_message<MClientSnap>(CEPH_SNAP_OP_SPLIT); |
11fdf7f2 TL |
9797 | update->head.split = in->ino(); |
9798 | update->split_inos = split_inos; | |
9799 | update->split_realms = split_realms; | |
9800 | update->bl = in->snaprealm->get_snap_trace(); | |
9801 | em.first->second = std::move(update); | |
9802 | } | |
7c673cae FG |
9803 | } |
9804 | } | |
9805 | ||
7c673cae FG |
9806 | // notify for active children, too. |
9807 | dout(10) << " " << realm << " open_children are " << realm->open_children << dendl; | |
f67539c2 TL |
9808 | for (auto& r : realm->open_children) |
9809 | q.push_back(r); | |
7c673cae FG |
9810 | } |
9811 | ||
11fdf7f2 | 9812 | if (notify_clients) |
7c673cae | 9813 | send_snaps(updates); |
7c673cae FG |
9814 | } |
9815 | ||
11fdf7f2 | 9816 | void MDCache::send_snap_update(CInode *in, version_t stid, int snap_op) |
7c673cae | 9817 | { |
11fdf7f2 TL |
9818 | dout(10) << __func__ << " " << *in << " stid " << stid << dendl; |
9819 | ceph_assert(in->is_auth()); | |
7c673cae | 9820 | |
11fdf7f2 TL |
9821 | set<mds_rank_t> mds_set; |
9822 | if (stid > 0) { | |
9823 | mds->mdsmap->get_mds_set_lower_bound(mds_set, MDSMap::STATE_RESOLVE); | |
9824 | mds_set.erase(mds->get_nodeid()); | |
9825 | } else { | |
9826 | in->list_replicas(mds_set); | |
9827 | } | |
7c673cae | 9828 | |
11fdf7f2 TL |
9829 | if (!mds_set.empty()) { |
9830 | bufferlist snap_blob; | |
9831 | in->encode_snap(snap_blob); | |
7c673cae | 9832 | |
11fdf7f2 | 9833 | for (auto p : mds_set) { |
9f95a23c | 9834 | auto m = make_message<MMDSSnapUpdate>(in->ino(), stid, snap_op); |
11fdf7f2 TL |
9835 | m->snap_blob = snap_blob; |
9836 | mds->send_message_mds(m, p); | |
9837 | } | |
9838 | } | |
7c673cae | 9839 | |
11fdf7f2 TL |
9840 | if (stid > 0) |
9841 | notify_global_snaprealm_update(snap_op); | |
9842 | } | |
7c673cae | 9843 | |
9f95a23c | 9844 | void MDCache::handle_snap_update(const cref_t<MMDSSnapUpdate> &m) |
11fdf7f2 TL |
9845 | { |
9846 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
9847 | dout(10) << __func__ << " " << *m << " from mds." << from << dendl; | |
7c673cae | 9848 | |
11fdf7f2 TL |
9849 | if (mds->get_state() < MDSMap::STATE_RESOLVE && |
9850 | mds->get_want_state() != CEPH_MDS_STATE_RESOLVE) { | |
9851 | return; | |
9852 | } | |
7c673cae | 9853 | |
11fdf7f2 TL |
9854 | // null rejoin_done means open_snaprealms() has already been called |
9855 | bool notify_clients = mds->get_state() > MDSMap::STATE_REJOIN || | |
9856 | (mds->is_rejoin() && !rejoin_done); | |
9857 | ||
9858 | if (m->get_tid() > 0) { | |
9859 | mds->snapclient->notify_commit(m->get_tid()); | |
9860 | if (notify_clients) | |
9861 | notify_global_snaprealm_update(m->get_snap_op()); | |
9862 | } | |
9863 | ||
9864 | CInode *in = get_inode(m->get_ino()); | |
9865 | if (in) { | |
9866 | ceph_assert(!in->is_auth()); | |
9867 | if (mds->get_state() > MDSMap::STATE_REJOIN || | |
9868 | (mds->is_rejoin() && !in->is_rejoining())) { | |
9869 | auto p = m->snap_blob.cbegin(); | |
9870 | in->decode_snap(p); | |
9871 | ||
9872 | if (!notify_clients) { | |
9873 | if (!rejoin_pending_snaprealms.count(in)) { | |
9874 | in->get(CInode::PIN_OPENINGSNAPPARENTS); | |
9875 | rejoin_pending_snaprealms.insert(in); | |
9876 | } | |
9877 | } | |
9878 | do_realm_invalidate_and_update_notify(in, m->get_snap_op(), notify_clients); | |
9879 | } | |
9880 | } | |
7c673cae FG |
9881 | } |
9882 | ||
11fdf7f2 TL |
9883 | void MDCache::notify_global_snaprealm_update(int snap_op) |
9884 | { | |
9885 | if (snap_op != CEPH_SNAP_OP_DESTROY) | |
9886 | snap_op = CEPH_SNAP_OP_UPDATE; | |
9887 | set<Session*> sessions; | |
9888 | mds->sessionmap.get_client_session_set(sessions); | |
9889 | for (auto &session : sessions) { | |
9890 | if (!session->is_open() && !session->is_stale()) | |
9891 | continue; | |
9f95a23c | 9892 | auto update = make_message<MClientSnap>(snap_op); |
11fdf7f2 TL |
9893 | update->head.split = global_snaprealm->inode->ino(); |
9894 | update->bl = global_snaprealm->get_snap_trace(); | |
9895 | mds->send_message_client_counted(update, session); | |
9896 | } | |
9897 | } | |
7c673cae FG |
9898 | |
9899 | // ------------------------------------------------------------------------------- | |
9900 | // STRAYS | |
9901 | ||
9902 | struct C_MDC_RetryScanStray : public MDCacheContext { | |
9903 | dirfrag_t next; | |
9904 | C_MDC_RetryScanStray(MDCache *c, dirfrag_t n) : MDCacheContext(c), next(n) { } | |
9905 | void finish(int r) override { | |
9906 | mdcache->scan_stray_dir(next); | |
9907 | } | |
9908 | }; | |
9909 | ||
9910 | void MDCache::scan_stray_dir(dirfrag_t next) | |
9911 | { | |
9912 | dout(10) << "scan_stray_dir " << next << dendl; | |
9913 | ||
f67539c2 TL |
9914 | if (next.ino) |
9915 | next.frag = strays[MDS_INO_STRAY_INDEX(next.ino)]->dirfragtree[next.frag.value()]; | |
9916 | ||
7c673cae FG |
9917 | for (int i = 0; i < NUM_STRAY; ++i) { |
9918 | if (strays[i]->ino() < next.ino) | |
9919 | continue; | |
f67539c2 TL |
9920 | |
9921 | std::vector<CDir*> ls; | |
7c673cae | 9922 | strays[i]->get_dirfrags(ls); |
7c673cae | 9923 | |
f67539c2 TL |
9924 | for (const auto& dir : ls) { |
9925 | if (dir->get_frag() < next.frag) | |
9926 | continue; | |
9927 | ||
9928 | if (!dir->can_auth_pin()) { | |
9929 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDC_RetryScanStray(this, dir->dirfrag())); | |
9930 | return; | |
9931 | } | |
9932 | ||
9933 | if (!dir->is_complete()) { | |
9934 | dir->fetch(new C_MDC_RetryScanStray(this, dir->dirfrag())); | |
9935 | return; | |
9936 | } | |
9937 | ||
9938 | for (auto &p : dir->items) { | |
9939 | CDentry *dn = p.second; | |
9940 | dn->state_set(CDentry::STATE_STRAY); | |
9941 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
9942 | if (dnl->is_primary()) { | |
9943 | CInode *in = dnl->get_inode(); | |
9944 | if (in->get_inode()->nlink == 0) | |
9945 | in->state_set(CInode::STATE_ORPHAN); | |
9946 | maybe_eval_stray(in); | |
9947 | } | |
7c673cae FG |
9948 | } |
9949 | } | |
9950 | } | |
9951 | } | |
9952 | ||
7c673cae FG |
9953 | void MDCache::fetch_backtrace(inodeno_t ino, int64_t pool, bufferlist& bl, Context *fin) |
9954 | { | |
9955 | object_t oid = CInode::get_object_name(ino, frag_t(), ""); | |
9956 | mds->objecter->getxattr(oid, object_locator_t(pool), "parent", CEPH_NOSNAP, &bl, 0, fin); | |
11fdf7f2 TL |
9957 | if (mds->logger) |
9958 | mds->logger->inc(l_mds_openino_backtrace_fetch); | |
7c673cae FG |
9959 | } |
9960 | ||
9961 | ||
9962 | ||
9963 | ||
9964 | ||
9965 | // ======================================================================================== | |
9966 | // DISCOVER | |
9967 | /* | |
9968 | ||
9969 | - for all discovers (except base_inos, e.g. root, stray), waiters are attached | |
9970 | to the parent metadata object in the cache (pinning it). | |
9971 | ||
9972 | - all discovers are tracked by tid, so that we can ignore potentially dup replies. | |
9973 | ||
9974 | */ | |
9975 | ||
9976 | void MDCache::_send_discover(discover_info_t& d) | |
9977 | { | |
9f95a23c TL |
9978 | auto dis = make_message<MDiscover>(d.ino, d.frag, d.snap, d.want_path, |
9979 | d.want_base_dir, d.path_locked); | |
7c673cae FG |
9980 | dis->set_tid(d.tid); |
9981 | mds->send_message_mds(dis, d.mds); | |
9982 | } | |
9983 | ||
9984 | void MDCache::discover_base_ino(inodeno_t want_ino, | |
11fdf7f2 | 9985 | MDSContext *onfinish, |
7c673cae FG |
9986 | mds_rank_t from) |
9987 | { | |
9988 | dout(7) << "discover_base_ino " << want_ino << " from mds." << from << dendl; | |
9989 | if (waiting_for_base_ino[from].count(want_ino) == 0) { | |
9990 | discover_info_t& d = _create_discover(from); | |
9991 | d.ino = want_ino; | |
9992 | _send_discover(d); | |
9993 | } | |
9994 | waiting_for_base_ino[from][want_ino].push_back(onfinish); | |
9995 | } | |
9996 | ||
9997 | ||
9998 | void MDCache::discover_dir_frag(CInode *base, | |
9999 | frag_t approx_fg, | |
11fdf7f2 | 10000 | MDSContext *onfinish, |
7c673cae FG |
10001 | mds_rank_t from) |
10002 | { | |
10003 | if (from < 0) | |
10004 | from = base->authority().first; | |
10005 | ||
10006 | dirfrag_t df(base->ino(), approx_fg); | |
10007 | dout(7) << "discover_dir_frag " << df | |
10008 | << " from mds." << from << dendl; | |
10009 | ||
10010 | if (!base->is_waiting_for_dir(approx_fg) || !onfinish) { | |
10011 | discover_info_t& d = _create_discover(from); | |
10012 | d.pin_base(base); | |
10013 | d.ino = base->ino(); | |
10014 | d.frag = approx_fg; | |
10015 | d.want_base_dir = true; | |
10016 | _send_discover(d); | |
10017 | } | |
10018 | ||
10019 | if (onfinish) | |
10020 | base->add_dir_waiter(approx_fg, onfinish); | |
10021 | } | |
10022 | ||
10023 | struct C_MDC_RetryDiscoverPath : public MDCacheContext { | |
10024 | CInode *base; | |
10025 | snapid_t snapid; | |
10026 | filepath path; | |
10027 | mds_rank_t from; | |
10028 | C_MDC_RetryDiscoverPath(MDCache *c, CInode *b, snapid_t s, filepath &p, mds_rank_t f) : | |
10029 | MDCacheContext(c), base(b), snapid(s), path(p), from(f) {} | |
10030 | void finish(int r) override { | |
10031 | mdcache->discover_path(base, snapid, path, 0, from); | |
10032 | } | |
10033 | }; | |
10034 | ||
10035 | void MDCache::discover_path(CInode *base, | |
10036 | snapid_t snap, | |
10037 | filepath want_path, | |
11fdf7f2 | 10038 | MDSContext *onfinish, |
9f95a23c | 10039 | bool path_locked, |
7c673cae FG |
10040 | mds_rank_t from) |
10041 | { | |
10042 | if (from < 0) | |
10043 | from = base->authority().first; | |
10044 | ||
10045 | dout(7) << "discover_path " << base->ino() << " " << want_path << " snap " << snap << " from mds." << from | |
9f95a23c | 10046 | << (path_locked ? " path_locked":"") |
7c673cae FG |
10047 | << dendl; |
10048 | ||
10049 | if (base->is_ambiguous_auth()) { | |
10050 | dout(10) << " waiting for single auth on " << *base << dendl; | |
10051 | if (!onfinish) | |
10052 | onfinish = new C_MDC_RetryDiscoverPath(this, base, snap, want_path, from); | |
10053 | base->add_waiter(CInode::WAIT_SINGLEAUTH, onfinish); | |
10054 | return; | |
10055 | } else if (from == mds->get_nodeid()) { | |
11fdf7f2 | 10056 | MDSContext::vec finished; |
7c673cae FG |
10057 | base->take_waiting(CInode::WAIT_DIR, finished); |
10058 | mds->queue_waiters(finished); | |
10059 | return; | |
10060 | } | |
10061 | ||
10062 | frag_t fg = base->pick_dirfrag(want_path[0]); | |
9f95a23c | 10063 | if ((path_locked && want_path.depth() == 1) || |
7c673cae FG |
10064 | !base->is_waiting_for_dir(fg) || !onfinish) { |
10065 | discover_info_t& d = _create_discover(from); | |
10066 | d.ino = base->ino(); | |
10067 | d.pin_base(base); | |
10068 | d.frag = fg; | |
10069 | d.snap = snap; | |
10070 | d.want_path = want_path; | |
10071 | d.want_base_dir = true; | |
9f95a23c | 10072 | d.path_locked = path_locked; |
7c673cae FG |
10073 | _send_discover(d); |
10074 | } | |
10075 | ||
10076 | // register + wait | |
10077 | if (onfinish) | |
10078 | base->add_dir_waiter(fg, onfinish); | |
10079 | } | |
10080 | ||
10081 | struct C_MDC_RetryDiscoverPath2 : public MDCacheContext { | |
10082 | CDir *base; | |
10083 | snapid_t snapid; | |
10084 | filepath path; | |
10085 | C_MDC_RetryDiscoverPath2(MDCache *c, CDir *b, snapid_t s, filepath &p) : | |
10086 | MDCacheContext(c), base(b), snapid(s), path(p) {} | |
10087 | void finish(int r) override { | |
10088 | mdcache->discover_path(base, snapid, path, 0); | |
10089 | } | |
10090 | }; | |
10091 | ||
10092 | void MDCache::discover_path(CDir *base, | |
10093 | snapid_t snap, | |
10094 | filepath want_path, | |
11fdf7f2 | 10095 | MDSContext *onfinish, |
9f95a23c | 10096 | bool path_locked) |
7c673cae FG |
10097 | { |
10098 | mds_rank_t from = base->authority().first; | |
10099 | ||
10100 | dout(7) << "discover_path " << base->dirfrag() << " " << want_path << " snap " << snap << " from mds." << from | |
9f95a23c | 10101 | << (path_locked ? " path_locked":"") |
7c673cae FG |
10102 | << dendl; |
10103 | ||
10104 | if (base->is_ambiguous_auth()) { | |
10105 | dout(7) << " waiting for single auth on " << *base << dendl; | |
10106 | if (!onfinish) | |
10107 | onfinish = new C_MDC_RetryDiscoverPath2(this, base, snap, want_path); | |
10108 | base->add_waiter(CDir::WAIT_SINGLEAUTH, onfinish); | |
10109 | return; | |
10110 | } else if (from == mds->get_nodeid()) { | |
11fdf7f2 | 10111 | MDSContext::vec finished; |
7c673cae FG |
10112 | base->take_sub_waiting(finished); |
10113 | mds->queue_waiters(finished); | |
10114 | return; | |
10115 | } | |
10116 | ||
9f95a23c | 10117 | if ((path_locked && want_path.depth() == 1) || |
7c673cae FG |
10118 | !base->is_waiting_for_dentry(want_path[0].c_str(), snap) || !onfinish) { |
10119 | discover_info_t& d = _create_discover(from); | |
10120 | d.ino = base->ino(); | |
31f18b77 | 10121 | d.pin_base(base->inode); |
7c673cae FG |
10122 | d.frag = base->get_frag(); |
10123 | d.snap = snap; | |
10124 | d.want_path = want_path; | |
10125 | d.want_base_dir = false; | |
9f95a23c | 10126 | d.path_locked = path_locked; |
7c673cae FG |
10127 | _send_discover(d); |
10128 | } | |
10129 | ||
10130 | // register + wait | |
10131 | if (onfinish) | |
10132 | base->add_dentry_waiter(want_path[0], snap, onfinish); | |
10133 | } | |
10134 | ||
10135 | void MDCache::kick_discovers(mds_rank_t who) | |
10136 | { | |
10137 | for (map<ceph_tid_t,discover_info_t>::iterator p = discovers.begin(); | |
10138 | p != discovers.end(); | |
10139 | ++p) { | |
10140 | if (p->second.mds != who) | |
10141 | continue; | |
10142 | _send_discover(p->second); | |
10143 | } | |
10144 | } | |
10145 | ||
10146 | ||
9f95a23c | 10147 | void MDCache::handle_discover(const cref_t<MDiscover> &dis) |
7c673cae FG |
10148 | { |
10149 | mds_rank_t whoami = mds->get_nodeid(); | |
10150 | mds_rank_t from = mds_rank_t(dis->get_source().num()); | |
10151 | ||
11fdf7f2 | 10152 | ceph_assert(from != whoami); |
7c673cae FG |
10153 | |
10154 | if (mds->get_state() <= MDSMap::STATE_REJOIN) { | |
10155 | if (mds->get_state() < MDSMap::STATE_REJOIN && | |
d2e6a577 | 10156 | mds->get_want_state() < CEPH_MDS_STATE_REJOIN) { |
7c673cae FG |
10157 | return; |
10158 | } | |
10159 | ||
10160 | // proceed if requester is in the REJOIN stage, the request is from parallel_fetch(). | |
10161 | // delay processing request from survivor because we may not yet choose lock states. | |
10162 | if (!mds->mdsmap->is_rejoin(from)) { | |
10163 | dout(0) << "discover_reply not yet active(|still rejoining), delaying" << dendl; | |
10164 | mds->wait_for_replay(new C_MDS_RetryMessage(mds, dis)); | |
10165 | return; | |
10166 | } | |
10167 | } | |
10168 | ||
10169 | ||
10170 | CInode *cur = 0; | |
9f95a23c | 10171 | auto reply = make_message<MDiscoverReply>(*dis); |
7c673cae FG |
10172 | |
10173 | snapid_t snapid = dis->get_snapid(); | |
10174 | ||
10175 | // get started. | |
10176 | if (MDS_INO_IS_BASE(dis->get_base_ino()) && | |
10177 | !dis->wants_base_dir() && dis->get_want().depth() == 0) { | |
10178 | // wants root | |
10179 | dout(7) << "handle_discover from mds." << from | |
10180 | << " wants base + " << dis->get_want().get_path() | |
10181 | << " snap " << snapid | |
10182 | << dendl; | |
10183 | ||
10184 | cur = get_inode(dis->get_base_ino()); | |
11fdf7f2 | 10185 | ceph_assert(cur); |
7c673cae FG |
10186 | |
10187 | // add root | |
10188 | reply->starts_with = MDiscoverReply::INODE; | |
9f95a23c | 10189 | encode_replica_inode(cur, from, reply->trace, mds->mdsmap->get_up_features()); |
7c673cae FG |
10190 | dout(10) << "added base " << *cur << dendl; |
10191 | } | |
10192 | else { | |
10193 | // there's a base inode | |
10194 | cur = get_inode(dis->get_base_ino(), snapid); | |
10195 | if (!cur && snapid != CEPH_NOSNAP) { | |
10196 | cur = get_inode(dis->get_base_ino()); | |
10197 | if (cur && !cur->is_multiversion()) | |
10198 | cur = NULL; // nope! | |
10199 | } | |
10200 | ||
10201 | if (!cur) { | |
10202 | dout(7) << "handle_discover mds." << from | |
10203 | << " don't have base ino " << dis->get_base_ino() << "." << snapid | |
10204 | << dendl; | |
10205 | if (!dis->wants_base_dir() && dis->get_want().depth() > 0) | |
10206 | reply->set_error_dentry(dis->get_dentry(0)); | |
10207 | reply->set_flag_error_dir(); | |
10208 | } else if (dis->wants_base_dir()) { | |
10209 | dout(7) << "handle_discover mds." << from | |
10210 | << " wants basedir+" << dis->get_want().get_path() | |
10211 | << " has " << *cur | |
10212 | << dendl; | |
10213 | } else { | |
10214 | dout(7) << "handle_discover mds." << from | |
10215 | << " wants " << dis->get_want().get_path() | |
10216 | << " has " << *cur | |
10217 | << dendl; | |
10218 | } | |
10219 | } | |
10220 | ||
11fdf7f2 | 10221 | ceph_assert(reply); |
7c673cae FG |
10222 | |
10223 | // add content | |
10224 | // do some fidgeting to include a dir if they asked for the base dir, or just root. | |
10225 | for (unsigned i = 0; | |
10226 | cur && (i < dis->get_want().depth() || dis->get_want().depth() == 0); | |
10227 | i++) { | |
10228 | ||
10229 | // -- figure out the dir | |
10230 | ||
10231 | // is *cur even a dir at all? | |
10232 | if (!cur->is_dir()) { | |
10233 | dout(7) << *cur << " not a dir" << dendl; | |
10234 | reply->set_flag_error_dir(); | |
10235 | break; | |
10236 | } | |
10237 | ||
10238 | // pick frag | |
10239 | frag_t fg; | |
10240 | if (dis->get_want().depth()) { | |
10241 | // dentry specifies | |
10242 | fg = cur->pick_dirfrag(dis->get_dentry(i)); | |
10243 | } else { | |
10244 | // requester explicity specified the frag | |
11fdf7f2 | 10245 | ceph_assert(dis->wants_base_dir() || MDS_INO_IS_BASE(dis->get_base_ino())); |
7c673cae FG |
10246 | fg = dis->get_base_dir_frag(); |
10247 | if (!cur->dirfragtree.is_leaf(fg)) | |
10248 | fg = cur->dirfragtree[fg.value()]; | |
10249 | } | |
10250 | CDir *curdir = cur->get_dirfrag(fg); | |
10251 | ||
10252 | if ((!curdir && !cur->is_auth()) || | |
10253 | (curdir && !curdir->is_auth())) { | |
10254 | ||
10255 | /* before: | |
10256 | * ONLY set flag if empty!! | |
10257 | * otherwise requester will wake up waiter(s) _and_ continue with discover, | |
10258 | * resulting in duplicate discovers in flight, | |
10259 | * which can wreak havoc when discovering rename srcdn (which may move) | |
10260 | */ | |
10261 | ||
10262 | if (reply->is_empty()) { | |
10263 | // only hint if empty. | |
10264 | // someday this could be better, but right now the waiter logic isn't smart enough. | |
10265 | ||
10266 | // hint | |
10267 | if (curdir) { | |
10268 | dout(7) << " not dirfrag auth, setting dir_auth_hint for " << *curdir << dendl; | |
10269 | reply->set_dir_auth_hint(curdir->authority().first); | |
10270 | } else { | |
10271 | dout(7) << " dirfrag not open, not inode auth, setting dir_auth_hint for " | |
10272 | << *cur << dendl; | |
10273 | reply->set_dir_auth_hint(cur->authority().first); | |
10274 | } | |
10275 | ||
10276 | // note error dentry, if any | |
10277 | // NOTE: important, as it allows requester to issue an equivalent discover | |
10278 | // to whomever we hint at. | |
10279 | if (dis->get_want().depth() > i) | |
10280 | reply->set_error_dentry(dis->get_dentry(i)); | |
10281 | } | |
10282 | ||
10283 | break; | |
10284 | } | |
10285 | ||
10286 | if (!curdir) { // open dir? | |
10287 | if (cur->is_frozen()) { | |
10288 | if (!reply->is_empty()) { | |
10289 | dout(7) << *cur << " is frozen, non-empty reply, stopping" << dendl; | |
10290 | break; | |
10291 | } | |
10292 | dout(7) << *cur << " is frozen, empty reply, waiting" << dendl; | |
10293 | cur->add_waiter(CInode::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10294 | return; |
10295 | } | |
10296 | curdir = cur->get_or_open_dirfrag(this, fg); | |
10297 | } else if (curdir->is_frozen_tree() || | |
10298 | (curdir->is_frozen_dir() && fragment_are_all_frozen(curdir))) { | |
31f18b77 FG |
10299 | if (!reply->is_empty()) { |
10300 | dout(7) << *curdir << " is frozen, non-empty reply, stopping" << dendl; | |
10301 | break; | |
10302 | } | |
7c673cae FG |
10303 | if (dis->wants_base_dir() && dis->get_base_dir_frag() != curdir->get_frag()) { |
10304 | dout(7) << *curdir << " is frozen, dirfrag mismatch, stopping" << dendl; | |
10305 | reply->set_flag_error_dir(); | |
10306 | break; | |
10307 | } | |
7c673cae FG |
10308 | dout(7) << *curdir << " is frozen, empty reply, waiting" << dendl; |
10309 | curdir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10310 | return; |
10311 | } | |
10312 | ||
10313 | // add dir | |
10314 | if (curdir->get_version() == 0) { | |
10315 | // fetch newly opened dir | |
10316 | } else if (reply->is_empty() && !dis->wants_base_dir()) { | |
10317 | dout(7) << "handle_discover not adding unwanted base dir " << *curdir << dendl; | |
10318 | // make sure the base frag is correct, though, in there was a refragment since the | |
10319 | // original request was sent. | |
10320 | reply->set_base_dir_frag(curdir->get_frag()); | |
10321 | } else { | |
11fdf7f2 | 10322 | ceph_assert(!curdir->is_ambiguous_auth()); // would be frozen. |
7c673cae FG |
10323 | if (!reply->trace.length()) |
10324 | reply->starts_with = MDiscoverReply::DIR; | |
9f95a23c | 10325 | encode_replica_dir(curdir, from, reply->trace); |
7c673cae FG |
10326 | dout(7) << "handle_discover added dir " << *curdir << dendl; |
10327 | } | |
10328 | ||
10329 | // lookup | |
10330 | CDentry *dn = 0; | |
10331 | if (curdir->get_version() == 0) { | |
10332 | // fetch newly opened dir | |
11fdf7f2 | 10333 | ceph_assert(!curdir->has_bloom()); |
7c673cae FG |
10334 | } else if (dis->get_want().depth() > 0) { |
10335 | // lookup dentry | |
10336 | dn = curdir->lookup(dis->get_dentry(i), snapid); | |
10337 | } else | |
10338 | break; // done! | |
10339 | ||
10340 | // incomplete dir? | |
10341 | if (!dn) { | |
31f18b77 | 10342 | if (!curdir->is_complete() && |
11fdf7f2 TL |
10343 | !(snapid == CEPH_NOSNAP && |
10344 | curdir->has_bloom() && | |
10345 | !curdir->is_in_bloom(dis->get_dentry(i)))) { | |
7c673cae FG |
10346 | // readdir |
10347 | dout(7) << "incomplete dir contents for " << *curdir << ", fetching" << dendl; | |
10348 | if (reply->is_empty()) { | |
10349 | // fetch and wait | |
10350 | curdir->fetch(new C_MDS_RetryMessage(mds, dis), | |
10351 | dis->wants_base_dir() && curdir->get_version() == 0); | |
7c673cae FG |
10352 | return; |
10353 | } else { | |
10354 | // initiate fetch, but send what we have so far | |
10355 | curdir->fetch(0); | |
10356 | break; | |
10357 | } | |
10358 | } | |
10359 | ||
11fdf7f2 TL |
10360 | if (snapid != CEPH_NOSNAP && !reply->is_empty()) { |
10361 | dout(7) << "dentry " << dis->get_dentry(i) << " snap " << snapid | |
10362 | << " dne, non-empty reply, stopping" << dendl; | |
10363 | break; | |
10364 | } | |
10365 | ||
7c673cae FG |
10366 | // send null dentry |
10367 | dout(7) << "dentry " << dis->get_dentry(i) << " dne, returning null in " | |
10368 | << *curdir << dendl; | |
11fdf7f2 TL |
10369 | if (snapid == CEPH_NOSNAP) |
10370 | dn = curdir->add_null_dentry(dis->get_dentry(i)); | |
10371 | else | |
10372 | dn = curdir->add_null_dentry(dis->get_dentry(i), snapid, snapid); | |
7c673cae | 10373 | } |
11fdf7f2 | 10374 | ceph_assert(dn); |
7c673cae | 10375 | |
31f18b77 FG |
10376 | // don't add replica to purging dentry/inode |
10377 | if (dn->state_test(CDentry::STATE_PURGING)) { | |
10378 | if (reply->is_empty()) | |
10379 | reply->set_flag_error_dn(dis->get_dentry(i)); | |
10380 | break; | |
10381 | } | |
10382 | ||
7c673cae FG |
10383 | CDentry::linkage_t *dnl = dn->get_linkage(); |
10384 | ||
10385 | // xlocked dentry? | |
10386 | // ...always block on non-tail items (they are unrelated) | |
10387 | // ...allow xlocked tail disocvery _only_ if explicitly requested | |
7c673cae FG |
10388 | if (dn->lock.is_xlocked()) { |
10389 | // is this the last (tail) item in the discover traversal? | |
9f95a23c TL |
10390 | if (dis->is_path_locked()) { |
10391 | dout(7) << "handle_discover allowing discovery of xlocked " << *dn << dendl; | |
7c673cae FG |
10392 | } else if (reply->is_empty()) { |
10393 | dout(7) << "handle_discover blocking on xlocked " << *dn << dendl; | |
10394 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10395 | return; |
10396 | } else { | |
10397 | dout(7) << "handle_discover non-empty reply, xlocked tail " << *dn << dendl; | |
10398 | break; | |
10399 | } | |
10400 | } | |
10401 | ||
10402 | // frozen inode? | |
9f95a23c | 10403 | bool tailitem = (dis->get_want().depth() == 0) || (i == dis->get_want().depth() - 1); |
7c673cae | 10404 | if (dnl->is_primary() && dnl->get_inode()->is_frozen_inode()) { |
9f95a23c | 10405 | if (tailitem && dis->is_path_locked()) { |
7c673cae FG |
10406 | dout(7) << "handle_discover allowing discovery of frozen tail " << *dnl->get_inode() << dendl; |
10407 | } else if (reply->is_empty()) { | |
10408 | dout(7) << *dnl->get_inode() << " is frozen, empty reply, waiting" << dendl; | |
10409 | dnl->get_inode()->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10410 | return; |
10411 | } else { | |
10412 | dout(7) << *dnl->get_inode() << " is frozen, non-empty reply, stopping" << dendl; | |
10413 | break; | |
10414 | } | |
10415 | } | |
10416 | ||
10417 | // add dentry | |
10418 | if (!reply->trace.length()) | |
10419 | reply->starts_with = MDiscoverReply::DENTRY; | |
9f95a23c | 10420 | encode_replica_dentry(dn, from, reply->trace); |
7c673cae FG |
10421 | dout(7) << "handle_discover added dentry " << *dn << dendl; |
10422 | ||
10423 | if (!dnl->is_primary()) break; // stop on null or remote link. | |
10424 | ||
10425 | // add inode | |
10426 | CInode *next = dnl->get_inode(); | |
11fdf7f2 | 10427 | ceph_assert(next->is_auth()); |
7c673cae | 10428 | |
9f95a23c | 10429 | encode_replica_inode(next, from, reply->trace, mds->mdsmap->get_up_features()); |
7c673cae FG |
10430 | dout(7) << "handle_discover added inode " << *next << dendl; |
10431 | ||
10432 | // descend, keep going. | |
10433 | cur = next; | |
10434 | continue; | |
10435 | } | |
10436 | ||
10437 | // how did we do? | |
11fdf7f2 | 10438 | ceph_assert(!reply->is_empty()); |
7c673cae FG |
10439 | dout(7) << "handle_discover sending result back to asker mds." << from << dendl; |
10440 | mds->send_message(reply, dis->get_connection()); | |
7c673cae FG |
10441 | } |
10442 | ||
9f95a23c | 10443 | void MDCache::handle_discover_reply(const cref_t<MDiscoverReply> &m) |
7c673cae FG |
10444 | { |
10445 | /* | |
10446 | if (mds->get_state() < MDSMap::STATE_ACTIVE) { | |
10447 | dout(0) << "discover_reply NOT ACTIVE YET" << dendl; | |
7c673cae FG |
10448 | return; |
10449 | } | |
10450 | */ | |
10451 | dout(7) << "discover_reply " << *m << dendl; | |
10452 | if (m->is_flag_error_dir()) | |
10453 | dout(7) << " flag error, dir" << dendl; | |
10454 | if (m->is_flag_error_dn()) | |
10455 | dout(7) << " flag error, dentry = " << m->get_error_dentry() << dendl; | |
10456 | ||
11fdf7f2 | 10457 | MDSContext::vec finished, error; |
7c673cae FG |
10458 | mds_rank_t from = mds_rank_t(m->get_source().num()); |
10459 | ||
10460 | // starting point | |
10461 | CInode *cur = get_inode(m->get_base_ino()); | |
11fdf7f2 | 10462 | auto p = m->trace.cbegin(); |
7c673cae FG |
10463 | |
10464 | int next = m->starts_with; | |
10465 | ||
10466 | // decrement discover counters | |
10467 | if (m->get_tid()) { | |
10468 | map<ceph_tid_t,discover_info_t>::iterator p = discovers.find(m->get_tid()); | |
10469 | if (p != discovers.end()) { | |
10470 | dout(10) << " found tid " << m->get_tid() << dendl; | |
10471 | discovers.erase(p); | |
10472 | } else { | |
10473 | dout(10) << " tid " << m->get_tid() << " not found, must be dup reply" << dendl; | |
10474 | } | |
10475 | } | |
10476 | ||
10477 | // discover may start with an inode | |
10478 | if (!p.end() && next == MDiscoverReply::INODE) { | |
9f95a23c | 10479 | decode_replica_inode(cur, p, NULL, finished); |
7c673cae | 10480 | dout(7) << "discover_reply got base inode " << *cur << dendl; |
11fdf7f2 | 10481 | ceph_assert(cur->is_base()); |
7c673cae FG |
10482 | |
10483 | next = MDiscoverReply::DIR; | |
10484 | ||
10485 | // take waiters? | |
10486 | if (cur->is_base() && | |
10487 | waiting_for_base_ino[from].count(cur->ino())) { | |
10488 | finished.swap(waiting_for_base_ino[from][cur->ino()]); | |
10489 | waiting_for_base_ino[from].erase(cur->ino()); | |
10490 | } | |
10491 | } | |
11fdf7f2 | 10492 | ceph_assert(cur); |
7c673cae FG |
10493 | |
10494 | // loop over discover results. | |
10495 | // indexes follow each ([[dir] dentry] inode) | |
10496 | // can start, end with any type. | |
10497 | while (!p.end()) { | |
10498 | // dir | |
10499 | frag_t fg; | |
9f95a23c | 10500 | CDir *curdir = nullptr; |
7c673cae | 10501 | if (next == MDiscoverReply::DIR) { |
9f95a23c | 10502 | decode_replica_dir(curdir, p, cur, mds_rank_t(m->get_source().num()), finished); |
7c673cae | 10503 | if (cur->ino() == m->get_base_ino() && curdir->get_frag() != m->get_base_dir_frag()) { |
11fdf7f2 | 10504 | ceph_assert(m->get_wanted_base_dir()); |
7c673cae FG |
10505 | cur->take_dir_waiting(m->get_base_dir_frag(), finished); |
10506 | } | |
10507 | } else { | |
10508 | // note: this can only happen our first way around this loop. | |
10509 | if (p.end() && m->is_flag_error_dn()) { | |
10510 | fg = cur->pick_dirfrag(m->get_error_dentry()); | |
10511 | curdir = cur->get_dirfrag(fg); | |
10512 | } else | |
10513 | curdir = cur->get_dirfrag(m->get_base_dir_frag()); | |
10514 | } | |
10515 | ||
10516 | if (p.end()) | |
10517 | break; | |
10518 | ||
10519 | // dentry | |
9f95a23c TL |
10520 | CDentry *dn = nullptr; |
10521 | decode_replica_dentry(dn, p, curdir, finished); | |
7c673cae FG |
10522 | |
10523 | if (p.end()) | |
10524 | break; | |
10525 | ||
10526 | // inode | |
9f95a23c | 10527 | decode_replica_inode(cur, p, dn, finished); |
7c673cae FG |
10528 | |
10529 | next = MDiscoverReply::DIR; | |
10530 | } | |
10531 | ||
10532 | // dir error? | |
10533 | // or dir_auth hint? | |
10534 | if (m->is_flag_error_dir() && !cur->is_dir()) { | |
10535 | // not a dir. | |
10536 | cur->take_waiting(CInode::WAIT_DIR, error); | |
10537 | } else if (m->is_flag_error_dir() || m->get_dir_auth_hint() != CDIR_AUTH_UNKNOWN) { | |
10538 | mds_rank_t who = m->get_dir_auth_hint(); | |
10539 | if (who == mds->get_nodeid()) who = -1; | |
10540 | if (who >= 0) | |
10541 | dout(7) << " dir_auth_hint is " << m->get_dir_auth_hint() << dendl; | |
10542 | ||
7c673cae FG |
10543 | |
10544 | if (m->get_wanted_base_dir()) { | |
31f18b77 FG |
10545 | frag_t fg = m->get_base_dir_frag(); |
10546 | CDir *dir = cur->get_dirfrag(fg); | |
10547 | ||
7c673cae FG |
10548 | if (cur->is_waiting_for_dir(fg)) { |
10549 | if (cur->is_auth()) | |
10550 | cur->take_waiting(CInode::WAIT_DIR, finished); | |
10551 | else if (dir || !cur->dirfragtree.is_leaf(fg)) | |
10552 | cur->take_dir_waiting(fg, finished); | |
10553 | else | |
10554 | discover_dir_frag(cur, fg, 0, who); | |
10555 | } else | |
10556 | dout(7) << " doing nothing, nobody is waiting for dir" << dendl; | |
10557 | } | |
10558 | ||
10559 | // try again? | |
10560 | if (m->get_error_dentry().length()) { | |
31f18b77 FG |
10561 | frag_t fg = cur->pick_dirfrag(m->get_error_dentry()); |
10562 | CDir *dir = cur->get_dirfrag(fg); | |
7c673cae FG |
10563 | // wanted a dentry |
10564 | if (dir && dir->is_waiting_for_dentry(m->get_error_dentry(), m->get_wanted_snapid())) { | |
10565 | if (dir->is_auth() || dir->lookup(m->get_error_dentry())) { | |
10566 | dir->take_dentry_waiting(m->get_error_dentry(), m->get_wanted_snapid(), | |
10567 | m->get_wanted_snapid(), finished); | |
10568 | } else { | |
10569 | filepath relpath(m->get_error_dentry(), 0); | |
9f95a23c | 10570 | discover_path(dir, m->get_wanted_snapid(), relpath, 0, m->is_path_locked()); |
7c673cae FG |
10571 | } |
10572 | } else | |
10573 | dout(7) << " doing nothing, have dir but nobody is waiting on dentry " | |
10574 | << m->get_error_dentry() << dendl; | |
10575 | } | |
31f18b77 FG |
10576 | } else if (m->is_flag_error_dn()) { |
10577 | frag_t fg = cur->pick_dirfrag(m->get_error_dentry()); | |
10578 | CDir *dir = cur->get_dirfrag(fg); | |
10579 | if (dir) { | |
10580 | if (dir->is_auth()) { | |
10581 | dir->take_sub_waiting(finished); | |
10582 | } else { | |
10583 | dir->take_dentry_waiting(m->get_error_dentry(), m->get_wanted_snapid(), | |
10584 | m->get_wanted_snapid(), error); | |
10585 | } | |
10586 | } | |
7c673cae FG |
10587 | } |
10588 | ||
10589 | // waiters | |
f67539c2 | 10590 | finish_contexts(g_ceph_context, error, -CEPHFS_ENOENT); // finish errors directly |
7c673cae | 10591 | mds->queue_waiters(finished); |
7c673cae FG |
10592 | } |
10593 | ||
10594 | ||
10595 | ||
10596 | // ---------------------------- | |
10597 | // REPLICAS | |
10598 | ||
b32b8144 | 10599 | |
9f95a23c | 10600 | void MDCache::encode_replica_dir(CDir *dir, mds_rank_t to, bufferlist& bl) |
b32b8144 | 10601 | { |
9f95a23c | 10602 | ENCODE_START(1, 1, bl); |
b32b8144 | 10603 | dirfrag_t df = dir->dirfrag(); |
11fdf7f2 | 10604 | encode(df, bl); |
9f95a23c TL |
10605 | __u32 nonce = dir->add_replica(to); |
10606 | encode(nonce, bl); | |
10607 | dir->_encode_base(bl); | |
10608 | ENCODE_FINISH(bl); | |
b32b8144 FG |
10609 | } |
10610 | ||
9f95a23c | 10611 | void MDCache::encode_replica_dentry(CDentry *dn, mds_rank_t to, bufferlist& bl) |
b32b8144 | 10612 | { |
f67539c2 | 10613 | ENCODE_START(2, 1, bl); |
11fdf7f2 TL |
10614 | encode(dn->get_name(), bl); |
10615 | encode(dn->last, bl); | |
9f95a23c TL |
10616 | |
10617 | __u32 nonce = dn->add_replica(to); | |
10618 | encode(nonce, bl); | |
10619 | encode(dn->first, bl); | |
10620 | encode(dn->linkage.remote_ino, bl); | |
10621 | encode(dn->linkage.remote_d_type, bl); | |
10622 | dn->lock.encode_state_for_replica(bl); | |
10623 | bool need_recover = mds->get_state() < MDSMap::STATE_ACTIVE; | |
10624 | encode(need_recover, bl); | |
f67539c2 | 10625 | encode(dn->alternate_name, bl); |
9f95a23c | 10626 | ENCODE_FINISH(bl); |
b32b8144 FG |
10627 | } |
10628 | ||
9f95a23c | 10629 | void MDCache::encode_replica_inode(CInode *in, mds_rank_t to, bufferlist& bl, |
b32b8144 FG |
10630 | uint64_t features) |
10631 | { | |
9f95a23c | 10632 | ceph_assert(in->is_auth()); |
f67539c2 TL |
10633 | |
10634 | ENCODE_START(2, 1, bl); | |
10635 | encode(in->ino(), bl); // bleh, minor assymetry here | |
11fdf7f2 | 10636 | encode(in->last, bl); |
9f95a23c TL |
10637 | |
10638 | __u32 nonce = in->add_replica(to); | |
10639 | encode(nonce, bl); | |
10640 | ||
10641 | in->_encode_base(bl, features); | |
10642 | in->_encode_locks_state_for_replica(bl, mds->get_state() < MDSMap::STATE_ACTIVE); | |
f6b5b4d7 TL |
10643 | |
10644 | __u32 state = in->state; | |
10645 | encode(state, bl); | |
10646 | ||
9f95a23c | 10647 | ENCODE_FINISH(bl); |
b32b8144 FG |
10648 | } |
10649 | ||
9f95a23c | 10650 | void MDCache::decode_replica_dir(CDir *&dir, bufferlist::const_iterator& p, CInode *diri, mds_rank_t from, |
11fdf7f2 | 10651 | MDSContext::vec& finished) |
7c673cae | 10652 | { |
9f95a23c | 10653 | DECODE_START(1, p); |
7c673cae | 10654 | dirfrag_t df; |
11fdf7f2 | 10655 | decode(df, p); |
7c673cae | 10656 | |
11fdf7f2 | 10657 | ceph_assert(diri->ino() == df.ino); |
7c673cae FG |
10658 | |
10659 | // add it (_replica_) | |
9f95a23c | 10660 | dir = diri->get_dirfrag(df.frag); |
7c673cae FG |
10661 | |
10662 | if (dir) { | |
10663 | // had replica. update w/ new nonce. | |
9f95a23c TL |
10664 | __u32 nonce; |
10665 | decode(nonce, p); | |
10666 | dir->set_replica_nonce(nonce); | |
10667 | dir->_decode_base(p); | |
10668 | dout(7) << __func__ << " had " << *dir << " nonce " << dir->replica_nonce << dendl; | |
7c673cae FG |
10669 | } else { |
10670 | // force frag to leaf in the diri tree | |
10671 | if (!diri->dirfragtree.is_leaf(df.frag)) { | |
9f95a23c | 10672 | dout(7) << __func__ << " forcing frag " << df.frag << " to leaf in the fragtree " |
7c673cae FG |
10673 | << diri->dirfragtree << dendl; |
10674 | diri->dirfragtree.force_to_leaf(g_ceph_context, df.frag); | |
10675 | } | |
7c673cae FG |
10676 | // add replica. |
10677 | dir = diri->add_dirfrag( new CDir(diri, df.frag, this, false) ); | |
9f95a23c TL |
10678 | __u32 nonce; |
10679 | decode(nonce, p); | |
10680 | dir->set_replica_nonce(nonce); | |
10681 | dir->_decode_base(p); | |
7c673cae FG |
10682 | // is this a dir_auth delegation boundary? |
10683 | if (from != diri->authority().first || | |
10684 | diri->is_ambiguous_auth() || | |
10685 | diri->is_base()) | |
10686 | adjust_subtree_auth(dir, from); | |
10687 | ||
9f95a23c | 10688 | dout(7) << __func__ << " added " << *dir << " nonce " << dir->replica_nonce << dendl; |
7c673cae FG |
10689 | // get waiters |
10690 | diri->take_dir_waiting(df.frag, finished); | |
10691 | } | |
9f95a23c | 10692 | DECODE_FINISH(p); |
7c673cae FG |
10693 | } |
10694 | ||
9f95a23c | 10695 | void MDCache::decode_replica_dentry(CDentry *&dn, bufferlist::const_iterator& p, CDir *dir, MDSContext::vec& finished) |
7c673cae | 10696 | { |
9f95a23c | 10697 | DECODE_START(1, p); |
7c673cae FG |
10698 | string name; |
10699 | snapid_t last; | |
11fdf7f2 TL |
10700 | decode(name, p); |
10701 | decode(last, p); | |
7c673cae | 10702 | |
9f95a23c | 10703 | dn = dir->lookup(name, last); |
7c673cae FG |
10704 | |
10705 | // have it? | |
9f95a23c | 10706 | bool is_new = false; |
7c673cae | 10707 | if (dn) { |
9f95a23c TL |
10708 | is_new = false; |
10709 | dout(7) << __func__ << " had " << *dn << dendl; | |
7c673cae | 10710 | } else { |
9f95a23c | 10711 | is_new = true; |
7c673cae | 10712 | dn = dir->add_null_dentry(name, 1 /* this will get updated below */, last); |
9f95a23c | 10713 | dout(7) << __func__ << " added " << *dn << dendl; |
7c673cae | 10714 | } |
9f95a23c TL |
10715 | |
10716 | __u32 nonce; | |
10717 | decode(nonce, p); | |
10718 | dn->set_replica_nonce(nonce); | |
10719 | decode(dn->first, p); | |
7c673cae | 10720 | |
9f95a23c TL |
10721 | inodeno_t rino; |
10722 | unsigned char rdtype; | |
10723 | decode(rino, p); | |
10724 | decode(rdtype, p); | |
10725 | dn->lock.decode_state(p, is_new); | |
7c673cae | 10726 | |
9f95a23c TL |
10727 | bool need_recover; |
10728 | decode(need_recover, p); | |
10729 | ||
f67539c2 TL |
10730 | mempool::mds_co::string alternate_name; |
10731 | if (struct_v >= 2) { | |
10732 | decode(alternate_name, p); | |
10733 | } | |
10734 | ||
9f95a23c | 10735 | if (is_new) { |
f67539c2 | 10736 | dn->set_alternate_name(std::move(alternate_name)); |
9f95a23c TL |
10737 | if (rino) |
10738 | dir->link_remote_inode(dn, rino, rdtype); | |
10739 | if (need_recover) | |
10740 | dn->lock.mark_need_recover(); | |
f67539c2 TL |
10741 | } else { |
10742 | ceph_assert(dn->alternate_name == alternate_name); | |
9f95a23c TL |
10743 | } |
10744 | ||
10745 | dir->take_dentry_waiting(name, dn->first, dn->last, finished); | |
10746 | DECODE_FINISH(p); | |
7c673cae FG |
10747 | } |
10748 | ||
9f95a23c | 10749 | void MDCache::decode_replica_inode(CInode *&in, bufferlist::const_iterator& p, CDentry *dn, MDSContext::vec& finished) |
7c673cae | 10750 | { |
f6b5b4d7 | 10751 | DECODE_START(2, p); |
7c673cae FG |
10752 | inodeno_t ino; |
10753 | snapid_t last; | |
9f95a23c | 10754 | __u32 nonce; |
11fdf7f2 TL |
10755 | decode(ino, p); |
10756 | decode(last, p); | |
9f95a23c TL |
10757 | decode(nonce, p); |
10758 | in = get_inode(ino, last); | |
7c673cae | 10759 | if (!in) { |
f67539c2 | 10760 | in = new CInode(this, false, 2, last); |
9f95a23c TL |
10761 | in->set_replica_nonce(nonce); |
10762 | in->_decode_base(p); | |
10763 | in->_decode_locks_state_for_replica(p, true); | |
7c673cae | 10764 | add_inode(in); |
b3b6e05e | 10765 | if (in->ino() == CEPH_INO_ROOT) |
7c673cae FG |
10766 | in->inode_auth.first = 0; |
10767 | else if (in->is_mdsdir()) | |
10768 | in->inode_auth.first = in->ino() - MDS_INO_MDSDIR_OFFSET; | |
9f95a23c | 10769 | dout(10) << __func__ << " added " << *in << dendl; |
7c673cae | 10770 | if (dn) { |
11fdf7f2 | 10771 | ceph_assert(dn->get_linkage()->is_null()); |
7c673cae FG |
10772 | dn->dir->link_primary_inode(dn, in); |
10773 | } | |
10774 | } else { | |
9f95a23c TL |
10775 | in->set_replica_nonce(nonce); |
10776 | in->_decode_base(p); | |
10777 | in->_decode_locks_state_for_replica(p, false); | |
10778 | dout(10) << __func__ << " had " << *in << dendl; | |
7c673cae FG |
10779 | } |
10780 | ||
10781 | if (dn) { | |
10782 | if (!dn->get_linkage()->is_primary() || dn->get_linkage()->get_inode() != in) | |
9f95a23c | 10783 | dout(10) << __func__ << " different linkage in dentry " << *dn << dendl; |
7c673cae | 10784 | } |
f6b5b4d7 TL |
10785 | |
10786 | if (struct_v >= 2) { | |
10787 | __u32 s; | |
10788 | decode(s, p); | |
10789 | s &= CInode::MASK_STATE_REPLICATED; | |
10790 | if (s & CInode::STATE_RANDEPHEMERALPIN) { | |
10791 | dout(10) << "replica inode is random ephemeral pinned" << dendl; | |
f67539c2 | 10792 | in->set_ephemeral_pin(false, true); |
f6b5b4d7 TL |
10793 | } |
10794 | } | |
10795 | ||
9f95a23c | 10796 | DECODE_FINISH(p); |
7c673cae FG |
10797 | } |
10798 | ||
10799 | ||
9f95a23c | 10800 | void MDCache::encode_replica_stray(CDentry *straydn, mds_rank_t who, bufferlist& bl) |
7c673cae | 10801 | { |
f67539c2 | 10802 | ceph_assert(straydn->get_num_auth_pins()); |
33c7a0ef | 10803 | ENCODE_START(2, 1, bl); |
7c673cae | 10804 | uint64_t features = mds->mdsmap->get_up_features(); |
9f95a23c TL |
10805 | encode_replica_inode(get_myin(), who, bl, features); |
10806 | encode_replica_dir(straydn->get_dir()->inode->get_parent_dn()->get_dir(), who, bl); | |
10807 | encode_replica_dentry(straydn->get_dir()->inode->get_parent_dn(), who, bl); | |
10808 | encode_replica_inode(straydn->get_dir()->inode, who, bl, features); | |
10809 | encode_replica_dir(straydn->get_dir(), who, bl); | |
10810 | encode_replica_dentry(straydn, who, bl); | |
33c7a0ef TL |
10811 | if (!straydn->get_projected_linkage()->is_null()) { |
10812 | encode_replica_inode(straydn->get_projected_linkage()->get_inode(), who, bl, features); | |
10813 | } | |
9f95a23c | 10814 | ENCODE_FINISH(bl); |
7c673cae FG |
10815 | } |
10816 | ||
33c7a0ef | 10817 | void MDCache::decode_replica_stray(CDentry *&straydn, CInode **in, const bufferlist &bl, mds_rank_t from) |
7c673cae | 10818 | { |
11fdf7f2 TL |
10819 | MDSContext::vec finished; |
10820 | auto p = bl.cbegin(); | |
7c673cae | 10821 | |
33c7a0ef | 10822 | DECODE_START(2, p); |
9f95a23c TL |
10823 | CInode *mdsin = nullptr; |
10824 | decode_replica_inode(mdsin, p, NULL, finished); | |
10825 | CDir *mdsdir = nullptr; | |
10826 | decode_replica_dir(mdsdir, p, mdsin, from, finished); | |
10827 | CDentry *straydirdn = nullptr; | |
10828 | decode_replica_dentry(straydirdn, p, mdsdir, finished); | |
10829 | CInode *strayin = nullptr; | |
10830 | decode_replica_inode(strayin, p, straydirdn, finished); | |
10831 | CDir *straydir = nullptr; | |
10832 | decode_replica_dir(straydir, p, strayin, from, finished); | |
10833 | ||
10834 | decode_replica_dentry(straydn, p, straydir, finished); | |
33c7a0ef TL |
10835 | if (struct_v >= 2 && in) { |
10836 | decode_replica_inode(*in, p, straydn, finished); | |
10837 | } | |
7c673cae FG |
10838 | if (!finished.empty()) |
10839 | mds->queue_waiters(finished); | |
9f95a23c | 10840 | DECODE_FINISH(p); |
7c673cae FG |
10841 | } |
10842 | ||
10843 | ||
10844 | int MDCache::send_dir_updates(CDir *dir, bool bcast) | |
10845 | { | |
10846 | // this is an FYI, re: replication | |
10847 | ||
10848 | set<mds_rank_t> who; | |
10849 | if (bcast) { | |
f67539c2 TL |
10850 | set<mds_rank_t> mds_set; |
10851 | mds->get_mds_map()->get_active_mds_set(mds_set); | |
10852 | ||
10853 | set<mds_rank_t> replica_set; | |
10854 | for (const auto &p : dir->get_replicas()) { | |
10855 | replica_set.insert(p.first); | |
10856 | } | |
10857 | ||
10858 | std::set_difference(mds_set.begin(), mds_set.end(), | |
10859 | replica_set.begin(), replica_set.end(), | |
10860 | std::inserter(who, who.end())); | |
7c673cae | 10861 | } else { |
181888fb FG |
10862 | for (const auto &p : dir->get_replicas()) { |
10863 | who.insert(p.first); | |
10864 | } | |
7c673cae FG |
10865 | } |
10866 | ||
10867 | dout(7) << "sending dir_update on " << *dir << " bcast " << bcast << " to " << who << dendl; | |
10868 | ||
10869 | filepath path; | |
10870 | dir->inode->make_path(path); | |
10871 | ||
f67539c2 TL |
10872 | std::set<int32_t> dir_rep_set; |
10873 | for (const auto &r : dir->dir_rep_by) { | |
10874 | dir_rep_set.insert(r); | |
10875 | } | |
10876 | ||
7c673cae FG |
10877 | mds_rank_t whoami = mds->get_nodeid(); |
10878 | for (set<mds_rank_t>::iterator it = who.begin(); | |
10879 | it != who.end(); | |
10880 | ++it) { | |
10881 | if (*it == whoami) continue; | |
10882 | //if (*it == except) continue; | |
10883 | dout(7) << "sending dir_update on " << *dir << " to " << *it << dendl; | |
10884 | ||
f67539c2 | 10885 | mds->send_message_mds(make_message<MDirUpdate>(mds->get_nodeid(), dir->dirfrag(), dir->dir_rep, dir_rep_set, path, bcast), *it); |
7c673cae FG |
10886 | } |
10887 | ||
10888 | return 0; | |
10889 | } | |
10890 | ||
9f95a23c | 10891 | void MDCache::handle_dir_update(const cref_t<MDirUpdate> &m) |
7c673cae | 10892 | { |
224ce89b WB |
10893 | dirfrag_t df = m->get_dirfrag(); |
10894 | CDir *dir = get_dirfrag(df); | |
7c673cae | 10895 | if (!dir) { |
224ce89b | 10896 | dout(5) << "dir_update on " << df << ", don't have it" << dendl; |
7c673cae FG |
10897 | |
10898 | // discover it? | |
10899 | if (m->should_discover()) { | |
10900 | // only try once! | |
10901 | // this is key to avoid a fragtree update race, among other things. | |
224ce89b | 10902 | m->inc_tried_discover(); |
7c673cae FG |
10903 | vector<CDentry*> trace; |
10904 | CInode *in; | |
10905 | filepath path = m->get_path(); | |
10906 | dout(5) << "trying discover on dir_update for " << path << dendl; | |
11fdf7f2 | 10907 | CF_MDS_RetryMessageFactory cf(mds, m); |
7c673cae | 10908 | MDRequestRef null_ref; |
9f95a23c | 10909 | int r = path_traverse(null_ref, cf, path, MDS_TRAVERSE_DISCOVER, &trace, &in); |
7c673cae FG |
10910 | if (r > 0) |
10911 | return; | |
224ce89b WB |
10912 | if (r == 0 && |
10913 | in->ino() == df.ino && | |
10914 | in->get_approx_dirfrag(df.frag) == NULL) { | |
10915 | open_remote_dirfrag(in, df.frag, new C_MDS_RetryMessage(mds, m)); | |
10916 | return; | |
10917 | } | |
7c673cae FG |
10918 | } |
10919 | ||
7c673cae FG |
10920 | return; |
10921 | } | |
10922 | ||
224ce89b WB |
10923 | if (!m->has_tried_discover()) { |
10924 | // Update if it already exists. Othwerwise it got updated by discover reply. | |
10925 | dout(5) << "dir_update on " << *dir << dendl; | |
10926 | dir->dir_rep = m->get_dir_rep(); | |
94b18763 FG |
10927 | dir->dir_rep_by.clear(); |
10928 | for (const auto &e : m->get_dir_rep_by()) { | |
10929 | dir->dir_rep_by.insert(e); | |
10930 | } | |
224ce89b | 10931 | } |
7c673cae FG |
10932 | } |
10933 | ||
10934 | ||
10935 | ||
10936 | ||
10937 | ||
10938 | // LINK | |
10939 | ||
9f95a23c TL |
10940 | void MDCache::encode_remote_dentry_link(CDentry::linkage_t *dnl, bufferlist& bl) |
10941 | { | |
10942 | ENCODE_START(1, 1, bl); | |
10943 | inodeno_t ino = dnl->get_remote_ino(); | |
10944 | encode(ino, bl); | |
10945 | __u8 d_type = dnl->get_remote_d_type(); | |
10946 | encode(d_type, bl); | |
10947 | ENCODE_FINISH(bl); | |
10948 | } | |
10949 | ||
10950 | void MDCache::decode_remote_dentry_link(CDir *dir, CDentry *dn, bufferlist::const_iterator& p) | |
10951 | { | |
10952 | DECODE_START(1, p); | |
10953 | inodeno_t ino; | |
10954 | __u8 d_type; | |
10955 | decode(ino, p); | |
10956 | decode(d_type, p); | |
10957 | dout(10) << __func__ << " remote " << ino << " " << d_type << dendl; | |
10958 | dir->link_remote_inode(dn, ino, d_type); | |
10959 | DECODE_FINISH(p); | |
10960 | } | |
10961 | ||
7c673cae FG |
10962 | void MDCache::send_dentry_link(CDentry *dn, MDRequestRef& mdr) |
10963 | { | |
9f95a23c | 10964 | dout(7) << __func__ << " " << *dn << dendl; |
7c673cae FG |
10965 | |
10966 | CDir *subtree = get_subtree_root(dn->get_dir()); | |
181888fb | 10967 | for (const auto &p : dn->get_replicas()) { |
7c673cae | 10968 | // don't tell (rename) witnesses; they already know |
181888fb | 10969 | if (mdr.get() && mdr->more()->witnessed.count(p.first)) |
7c673cae | 10970 | continue; |
181888fb FG |
10971 | if (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || |
10972 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
10973 | rejoin_gather.count(p.first))) | |
7c673cae FG |
10974 | continue; |
10975 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
9f95a23c | 10976 | auto m = make_message<MDentryLink>(subtree->dirfrag(), dn->get_dir()->dirfrag(), dn->get_name(), dnl->is_primary()); |
7c673cae | 10977 | if (dnl->is_primary()) { |
9f95a23c TL |
10978 | dout(10) << __func__ << " primary " << *dnl->get_inode() << dendl; |
10979 | encode_replica_inode(dnl->get_inode(), p.first, m->bl, | |
7c673cae FG |
10980 | mds->mdsmap->get_up_features()); |
10981 | } else if (dnl->is_remote()) { | |
9f95a23c | 10982 | encode_remote_dentry_link(dnl, m->bl); |
7c673cae FG |
10983 | } else |
10984 | ceph_abort(); // aie, bad caller! | |
181888fb | 10985 | mds->send_message_mds(m, p.first); |
7c673cae FG |
10986 | } |
10987 | } | |
10988 | ||
9f95a23c | 10989 | void MDCache::handle_dentry_link(const cref_t<MDentryLink> &m) |
7c673cae | 10990 | { |
7c673cae FG |
10991 | CDentry *dn = NULL; |
10992 | CDir *dir = get_dirfrag(m->get_dirfrag()); | |
10993 | if (!dir) { | |
9f95a23c | 10994 | dout(7) << __func__ << " don't have dirfrag " << m->get_dirfrag() << dendl; |
7c673cae FG |
10995 | } else { |
10996 | dn = dir->lookup(m->get_dn()); | |
10997 | if (!dn) { | |
9f95a23c | 10998 | dout(7) << __func__ << " don't have dentry " << *dir << " dn " << m->get_dn() << dendl; |
7c673cae | 10999 | } else { |
9f95a23c | 11000 | dout(7) << __func__ << " on " << *dn << dendl; |
7c673cae FG |
11001 | CDentry::linkage_t *dnl = dn->get_linkage(); |
11002 | ||
11fdf7f2 TL |
11003 | ceph_assert(!dn->is_auth()); |
11004 | ceph_assert(dnl->is_null()); | |
7c673cae FG |
11005 | } |
11006 | } | |
11007 | ||
11fdf7f2 TL |
11008 | auto p = m->bl.cbegin(); |
11009 | MDSContext::vec finished; | |
7c673cae FG |
11010 | if (dn) { |
11011 | if (m->get_is_primary()) { | |
11012 | // primary link. | |
9f95a23c TL |
11013 | CInode *in = nullptr; |
11014 | decode_replica_inode(in, p, dn, finished); | |
7c673cae FG |
11015 | } else { |
11016 | // remote link, easy enough. | |
9f95a23c | 11017 | decode_remote_dentry_link(dir, dn, p); |
7c673cae FG |
11018 | } |
11019 | } else { | |
11020 | ceph_abort(); | |
11021 | } | |
11022 | ||
11023 | if (!finished.empty()) | |
11024 | mds->queue_waiters(finished); | |
11025 | ||
7c673cae FG |
11026 | return; |
11027 | } | |
11028 | ||
11029 | ||
11030 | // UNLINK | |
11031 | ||
11032 | void MDCache::send_dentry_unlink(CDentry *dn, CDentry *straydn, MDRequestRef& mdr) | |
11033 | { | |
9f95a23c | 11034 | dout(10) << __func__ << " " << *dn << dendl; |
7c673cae FG |
11035 | // share unlink news with replicas |
11036 | set<mds_rank_t> replicas; | |
11037 | dn->list_replicas(replicas); | |
11fdf7f2 TL |
11038 | bufferlist snapbl; |
11039 | if (straydn) { | |
7c673cae | 11040 | straydn->list_replicas(replicas); |
11fdf7f2 TL |
11041 | CInode *strayin = straydn->get_linkage()->get_inode(); |
11042 | strayin->encode_snap_blob(snapbl); | |
11043 | } | |
7c673cae FG |
11044 | for (set<mds_rank_t>::iterator it = replicas.begin(); |
11045 | it != replicas.end(); | |
11046 | ++it) { | |
11047 | // don't tell (rmdir) witnesses; they already know | |
11048 | if (mdr.get() && mdr->more()->witnessed.count(*it)) | |
11049 | continue; | |
11050 | ||
11051 | if (mds->mdsmap->get_state(*it) < MDSMap::STATE_REJOIN || | |
11052 | (mds->mdsmap->get_state(*it) == MDSMap::STATE_REJOIN && | |
11053 | rejoin_gather.count(*it))) | |
11054 | continue; | |
11055 | ||
9f95a23c | 11056 | auto unlink = make_message<MDentryUnlink>(dn->get_dir()->dirfrag(), dn->get_name()); |
11fdf7f2 | 11057 | if (straydn) { |
9f95a23c | 11058 | encode_replica_stray(straydn, *it, unlink->straybl); |
11fdf7f2 TL |
11059 | unlink->snapbl = snapbl; |
11060 | } | |
7c673cae FG |
11061 | mds->send_message_mds(unlink, *it); |
11062 | } | |
11063 | } | |
11064 | ||
9f95a23c | 11065 | void MDCache::handle_dentry_unlink(const cref_t<MDentryUnlink> &m) |
7c673cae FG |
11066 | { |
11067 | // straydn | |
9f95a23c | 11068 | CDentry *straydn = nullptr; |
33c7a0ef | 11069 | CInode *strayin = nullptr; |
7c673cae | 11070 | if (m->straybl.length()) |
33c7a0ef | 11071 | decode_replica_stray(straydn, &strayin, m->straybl, mds_rank_t(m->get_source().num())); |
7c673cae FG |
11072 | |
11073 | CDir *dir = get_dirfrag(m->get_dirfrag()); | |
11074 | if (!dir) { | |
9f95a23c | 11075 | dout(7) << __func__ << " don't have dirfrag " << m->get_dirfrag() << dendl; |
7c673cae FG |
11076 | } else { |
11077 | CDentry *dn = dir->lookup(m->get_dn()); | |
11078 | if (!dn) { | |
9f95a23c | 11079 | dout(7) << __func__ << " don't have dentry " << *dir << " dn " << m->get_dn() << dendl; |
7c673cae | 11080 | } else { |
9f95a23c | 11081 | dout(7) << __func__ << " on " << *dn << dendl; |
7c673cae FG |
11082 | CDentry::linkage_t *dnl = dn->get_linkage(); |
11083 | ||
11084 | // open inode? | |
11085 | if (dnl->is_primary()) { | |
11086 | CInode *in = dnl->get_inode(); | |
11087 | dn->dir->unlink_inode(dn); | |
11fdf7f2 | 11088 | ceph_assert(straydn); |
7c673cae FG |
11089 | straydn->dir->link_primary_inode(straydn, in); |
11090 | ||
11091 | // in->first is lazily updated on replica; drag it forward so | |
11092 | // that we always keep it in sync with the dnq | |
11fdf7f2 | 11093 | ceph_assert(straydn->first >= in->first); |
7c673cae FG |
11094 | in->first = straydn->first; |
11095 | ||
11096 | // update subtree map? | |
11097 | if (in->is_dir()) | |
11098 | adjust_subtree_after_rename(in, dir, false); | |
11099 | ||
11fdf7f2 TL |
11100 | if (m->snapbl.length()) { |
11101 | bool hadrealm = (in->snaprealm ? true : false); | |
11102 | in->decode_snap_blob(m->snapbl); | |
11103 | ceph_assert(in->snaprealm); | |
11fdf7f2 TL |
11104 | if (!hadrealm) |
11105 | do_realm_invalidate_and_update_notify(in, CEPH_SNAP_OP_SPLIT, false); | |
11106 | } | |
11107 | ||
7c673cae FG |
11108 | // send caps to auth (if we're not already) |
11109 | if (in->is_any_caps() && | |
11110 | !in->state_test(CInode::STATE_EXPORTINGCAPS)) | |
11111 | migrator->export_caps(in); | |
11112 | ||
7c673cae FG |
11113 | straydn = NULL; |
11114 | } else { | |
11fdf7f2 TL |
11115 | ceph_assert(!straydn); |
11116 | ceph_assert(dnl->is_remote()); | |
7c673cae FG |
11117 | dn->dir->unlink_inode(dn); |
11118 | } | |
11fdf7f2 | 11119 | ceph_assert(dnl->is_null()); |
7c673cae FG |
11120 | } |
11121 | } | |
11122 | ||
11123 | // race with trim_dentry() | |
11124 | if (straydn) { | |
11fdf7f2 TL |
11125 | ceph_assert(straydn->get_num_ref() == 0); |
11126 | ceph_assert(straydn->get_linkage()->is_null()); | |
11127 | expiremap ex; | |
11128 | trim_dentry(straydn, ex); | |
11129 | send_expire_messages(ex); | |
7c673cae | 11130 | } |
7c673cae FG |
11131 | } |
11132 | ||
11133 | ||
11134 | ||
11135 | ||
11136 | ||
11137 | ||
11138 | // =================================================================== | |
11139 | ||
11140 | ||
11141 | ||
11142 | // =================================================================== | |
11143 | // FRAGMENT | |
11144 | ||
11145 | ||
11146 | /** | |
11147 | * adjust_dir_fragments -- adjust fragmentation for a directory | |
11148 | * | |
11149 | * @param diri directory inode | |
11150 | * @param basefrag base fragment | |
11151 | * @param bits bit adjustment. positive for split, negative for merge. | |
11152 | */ | |
11153 | void MDCache::adjust_dir_fragments(CInode *diri, frag_t basefrag, int bits, | |
9f95a23c | 11154 | std::vector<CDir*>* resultfrags, |
11fdf7f2 | 11155 | MDSContext::vec& waiters, |
7c673cae FG |
11156 | bool replay) |
11157 | { | |
11158 | dout(10) << "adjust_dir_fragments " << basefrag << " " << bits | |
11159 | << " on " << *diri << dendl; | |
11160 | ||
9f95a23c | 11161 | auto&& p = diri->get_dirfrags_under(basefrag); |
7c673cae | 11162 | |
9f95a23c | 11163 | adjust_dir_fragments(diri, p.second, basefrag, bits, resultfrags, waiters, replay); |
7c673cae FG |
11164 | } |
11165 | ||
11166 | CDir *MDCache::force_dir_fragment(CInode *diri, frag_t fg, bool replay) | |
11167 | { | |
11168 | CDir *dir = diri->get_dirfrag(fg); | |
11169 | if (dir) | |
11170 | return dir; | |
11171 | ||
11172 | dout(10) << "force_dir_fragment " << fg << " on " << *diri << dendl; | |
11173 | ||
9f95a23c | 11174 | std::vector<CDir*> src, result; |
11fdf7f2 | 11175 | MDSContext::vec waiters; |
7c673cae FG |
11176 | |
11177 | // split a parent? | |
11178 | frag_t parent = diri->dirfragtree.get_branch_or_leaf(fg); | |
11179 | while (1) { | |
11180 | CDir *pdir = diri->get_dirfrag(parent); | |
11181 | if (pdir) { | |
11182 | int split = fg.bits() - parent.bits(); | |
11183 | dout(10) << " splitting parent by " << split << " " << *pdir << dendl; | |
11184 | src.push_back(pdir); | |
9f95a23c | 11185 | adjust_dir_fragments(diri, src, parent, split, &result, waiters, replay); |
7c673cae FG |
11186 | dir = diri->get_dirfrag(fg); |
11187 | if (dir) { | |
11188 | dout(10) << "force_dir_fragment result " << *dir << dendl; | |
11189 | break; | |
11190 | } | |
11191 | } | |
11192 | if (parent == frag_t()) | |
11193 | break; | |
11194 | frag_t last = parent; | |
11195 | parent = parent.parent(); | |
11196 | dout(10) << " " << last << " parent is " << parent << dendl; | |
11197 | } | |
11198 | ||
11199 | if (!dir) { | |
11200 | // hoover up things under fg? | |
9f95a23c TL |
11201 | { |
11202 | auto&& p = diri->get_dirfrags_under(fg); | |
11203 | src.insert(std::end(src), std::cbegin(p.second), std::cend(p.second)); | |
11204 | } | |
7c673cae FG |
11205 | if (src.empty()) { |
11206 | dout(10) << "force_dir_fragment no frags under " << fg << dendl; | |
11207 | } else { | |
11208 | dout(10) << " will combine frags under " << fg << ": " << src << dendl; | |
9f95a23c | 11209 | adjust_dir_fragments(diri, src, fg, 0, &result, waiters, replay); |
7c673cae FG |
11210 | dir = result.front(); |
11211 | dout(10) << "force_dir_fragment result " << *dir << dendl; | |
11212 | } | |
11213 | } | |
11214 | if (!replay) | |
11215 | mds->queue_waiters(waiters); | |
11216 | return dir; | |
11217 | } | |
11218 | ||
11219 | void MDCache::adjust_dir_fragments(CInode *diri, | |
9f95a23c | 11220 | const std::vector<CDir*>& srcfrags, |
7c673cae | 11221 | frag_t basefrag, int bits, |
9f95a23c | 11222 | std::vector<CDir*>* resultfrags, |
11fdf7f2 | 11223 | MDSContext::vec& waiters, |
7c673cae FG |
11224 | bool replay) |
11225 | { | |
11226 | dout(10) << "adjust_dir_fragments " << basefrag << " bits " << bits | |
11227 | << " srcfrags " << srcfrags | |
11228 | << " on " << *diri << dendl; | |
11229 | ||
11230 | // adjust fragtree | |
11231 | // yuck. we may have discovered the inode while it was being fragmented. | |
11232 | if (!diri->dirfragtree.is_leaf(basefrag)) | |
11233 | diri->dirfragtree.force_to_leaf(g_ceph_context, basefrag); | |
11234 | ||
11235 | if (bits > 0) | |
11236 | diri->dirfragtree.split(basefrag, bits); | |
11237 | dout(10) << " new fragtree is " << diri->dirfragtree << dendl; | |
11238 | ||
11239 | if (srcfrags.empty()) | |
11240 | return; | |
11241 | ||
11242 | // split | |
11243 | CDir *parent_dir = diri->get_parent_dir(); | |
11244 | CDir *parent_subtree = 0; | |
11245 | if (parent_dir) | |
11246 | parent_subtree = get_subtree_root(parent_dir); | |
11247 | ||
9f95a23c | 11248 | ceph_assert(srcfrags.size() >= 1); |
7c673cae FG |
11249 | if (bits > 0) { |
11250 | // SPLIT | |
11fdf7f2 | 11251 | ceph_assert(srcfrags.size() == 1); |
7c673cae FG |
11252 | CDir *dir = srcfrags.front(); |
11253 | ||
11254 | dir->split(bits, resultfrags, waiters, replay); | |
11255 | ||
11256 | // did i change the subtree map? | |
11257 | if (dir->is_subtree_root()) { | |
11258 | // new frags are now separate subtrees | |
9f95a23c TL |
11259 | for (const auto& dir : *resultfrags) { |
11260 | subtrees[dir].clear(); // new frag is now its own subtree | |
11261 | } | |
7c673cae FG |
11262 | |
11263 | // was i a bound? | |
11264 | if (parent_subtree) { | |
11fdf7f2 | 11265 | ceph_assert(subtrees[parent_subtree].count(dir)); |
7c673cae | 11266 | subtrees[parent_subtree].erase(dir); |
9f95a23c TL |
11267 | for (const auto& dir : *resultfrags) { |
11268 | ceph_assert(dir->is_subtree_root()); | |
11269 | subtrees[parent_subtree].insert(dir); | |
7c673cae FG |
11270 | } |
11271 | } | |
11272 | ||
11273 | // adjust my bounds. | |
11274 | set<CDir*> bounds; | |
11275 | bounds.swap(subtrees[dir]); | |
11276 | subtrees.erase(dir); | |
11277 | for (set<CDir*>::iterator p = bounds.begin(); | |
11278 | p != bounds.end(); | |
11279 | ++p) { | |
11280 | CDir *frag = get_subtree_root((*p)->get_parent_dir()); | |
11281 | subtrees[frag].insert(*p); | |
11282 | } | |
11283 | ||
11284 | show_subtrees(10); | |
7c673cae FG |
11285 | } |
11286 | ||
11287 | diri->close_dirfrag(dir->get_frag()); | |
11288 | ||
11289 | } else { | |
11290 | // MERGE | |
11291 | ||
11292 | // are my constituent bits subtrees? if so, i will be too. | |
11293 | // (it's all or none, actually.) | |
11fdf7f2 | 11294 | bool any_subtree = false, any_non_subtree = false; |
9f95a23c | 11295 | for (const auto& dir : srcfrags) { |
11fdf7f2 | 11296 | if (dir->is_subtree_root()) |
31f18b77 | 11297 | any_subtree = true; |
11fdf7f2 TL |
11298 | else |
11299 | any_non_subtree = true; | |
31f18b77 | 11300 | } |
11fdf7f2 TL |
11301 | ceph_assert(!any_subtree || !any_non_subtree); |
11302 | ||
31f18b77 FG |
11303 | set<CDir*> new_bounds; |
11304 | if (any_subtree) { | |
9f95a23c | 11305 | for (const auto& dir : srcfrags) { |
31f18b77 FG |
11306 | // this simplifies the code that find subtrees underneath the dirfrag |
11307 | if (!dir->is_subtree_root()) { | |
11308 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
11309 | adjust_subtree_auth(dir, mds->get_nodeid()); | |
11310 | } | |
11311 | } | |
11312 | ||
9f95a23c | 11313 | for (const auto& dir : srcfrags) { |
11fdf7f2 | 11314 | ceph_assert(dir->is_subtree_root()); |
7c673cae | 11315 | dout(10) << " taking srcfrag subtree bounds from " << *dir << dendl; |
7c673cae FG |
11316 | map<CDir*, set<CDir*> >::iterator q = subtrees.find(dir); |
11317 | set<CDir*>::iterator r = q->second.begin(); | |
11318 | while (r != subtrees[dir].end()) { | |
11319 | new_bounds.insert(*r); | |
11320 | subtrees[dir].erase(r++); | |
11321 | } | |
11322 | subtrees.erase(q); | |
31f18b77 | 11323 | |
7c673cae FG |
11324 | // remove myself as my parent's bound |
11325 | if (parent_subtree) | |
11326 | subtrees[parent_subtree].erase(dir); | |
11327 | } | |
11328 | } | |
11329 | ||
11330 | // merge | |
11331 | CDir *f = new CDir(diri, basefrag, this, srcfrags.front()->is_auth()); | |
11332 | f->merge(srcfrags, waiters, replay); | |
7c673cae | 11333 | |
31f18b77 | 11334 | if (any_subtree) { |
11fdf7f2 | 11335 | ceph_assert(f->is_subtree_root()); |
7c673cae FG |
11336 | subtrees[f].swap(new_bounds); |
11337 | if (parent_subtree) | |
11338 | subtrees[parent_subtree].insert(f); | |
11339 | ||
11340 | show_subtrees(10); | |
11341 | } | |
11342 | ||
9f95a23c | 11343 | resultfrags->push_back(f); |
7c673cae FG |
11344 | } |
11345 | } | |
11346 | ||
11347 | ||
11348 | class C_MDC_FragmentFrozen : public MDSInternalContext { | |
11349 | MDCache *mdcache; | |
11350 | MDRequestRef mdr; | |
11351 | public: | |
11352 | C_MDC_FragmentFrozen(MDCache *m, MDRequestRef& r) : | |
11353 | MDSInternalContext(m->mds), mdcache(m), mdr(r) {} | |
11354 | void finish(int r) override { | |
11355 | mdcache->fragment_frozen(mdr, r); | |
11356 | } | |
11357 | }; | |
11358 | ||
9f95a23c | 11359 | bool MDCache::can_fragment(CInode *diri, const std::vector<CDir*>& dirs) |
7c673cae FG |
11360 | { |
11361 | if (is_readonly()) { | |
11362 | dout(7) << "can_fragment: read-only FS, no fragmenting for now" << dendl; | |
11363 | return false; | |
11364 | } | |
11365 | if (mds->is_cluster_degraded()) { | |
11366 | dout(7) << "can_fragment: cluster degraded, no fragmenting for now" << dendl; | |
11367 | return false; | |
11368 | } | |
11369 | if (diri->get_parent_dir() && | |
11370 | diri->get_parent_dir()->get_inode()->is_stray()) { | |
11371 | dout(7) << "can_fragment: i won't merge|split anything in stray" << dendl; | |
11372 | return false; | |
11373 | } | |
b3b6e05e | 11374 | if (diri->is_mdsdir() || diri->ino() == CEPH_INO_CEPH) { |
f67539c2 | 11375 | dout(7) << "can_fragment: i won't fragment mdsdir or .ceph" << dendl; |
7c673cae FG |
11376 | return false; |
11377 | } | |
11378 | ||
9f95a23c | 11379 | for (const auto& dir : dirs) { |
f67539c2 TL |
11380 | if (dir->scrub_is_in_progress()) { |
11381 | dout(7) << "can_fragment: scrub in progress " << *dir << dendl; | |
11382 | return false; | |
11383 | } | |
11384 | ||
7c673cae FG |
11385 | if (dir->state_test(CDir::STATE_FRAGMENTING)) { |
11386 | dout(7) << "can_fragment: already fragmenting " << *dir << dendl; | |
11387 | return false; | |
11388 | } | |
11389 | if (!dir->is_auth()) { | |
11390 | dout(7) << "can_fragment: not auth on " << *dir << dendl; | |
11391 | return false; | |
11392 | } | |
11393 | if (dir->is_bad()) { | |
11394 | dout(7) << "can_fragment: bad dirfrag " << *dir << dendl; | |
11395 | return false; | |
11396 | } | |
11397 | if (dir->is_frozen() || | |
11398 | dir->is_freezing()) { | |
11399 | dout(7) << "can_fragment: can't merge, freezing|frozen. wait for other exports to finish first." << dendl; | |
11400 | return false; | |
11401 | } | |
11402 | } | |
11403 | ||
11404 | return true; | |
11405 | } | |
11406 | ||
11407 | void MDCache::split_dir(CDir *dir, int bits) | |
11408 | { | |
11409 | dout(7) << __func__ << " " << *dir << " bits " << bits << dendl; | |
11fdf7f2 | 11410 | ceph_assert(dir->is_auth()); |
7c673cae FG |
11411 | CInode *diri = dir->inode; |
11412 | ||
9f95a23c | 11413 | std::vector<CDir*> dirs; |
7c673cae FG |
11414 | dirs.push_back(dir); |
11415 | ||
11416 | if (!can_fragment(diri, dirs)) { | |
11417 | dout(7) << __func__ << " cannot fragment right now, dropping" << dendl; | |
11418 | return; | |
11419 | } | |
11420 | ||
31f18b77 FG |
11421 | if (dir->frag.bits() + bits > 24) { |
11422 | dout(7) << __func__ << " frag bits > 24, dropping" << dendl; | |
11423 | return; | |
11424 | } | |
11425 | ||
7c673cae FG |
11426 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FRAGMENTDIR); |
11427 | mdr->more()->fragment_base = dir->dirfrag(); | |
11428 | ||
11fdf7f2 | 11429 | ceph_assert(fragments.count(dir->dirfrag()) == 0); |
7c673cae FG |
11430 | fragment_info_t& info = fragments[dir->dirfrag()]; |
11431 | info.mdr = mdr; | |
11432 | info.dirs.push_back(dir); | |
11433 | info.bits = bits; | |
11434 | info.last_cum_auth_pins_change = ceph_clock_now(); | |
11435 | ||
11436 | fragment_freeze_dirs(dirs); | |
11437 | // initial mark+complete pass | |
11438 | fragment_mark_and_complete(mdr); | |
11439 | } | |
11440 | ||
11441 | void MDCache::merge_dir(CInode *diri, frag_t frag) | |
11442 | { | |
11443 | dout(7) << "merge_dir to " << frag << " on " << *diri << dendl; | |
11444 | ||
9f95a23c TL |
11445 | auto&& [all, dirs] = diri->get_dirfrags_under(frag); |
11446 | if (!all) { | |
7c673cae FG |
11447 | dout(7) << "don't have all frags under " << frag << " for " << *diri << dendl; |
11448 | return; | |
11449 | } | |
11450 | ||
11451 | if (diri->dirfragtree.is_leaf(frag)) { | |
11452 | dout(10) << " " << frag << " already a leaf for " << *diri << dendl; | |
11453 | return; | |
11454 | } | |
11455 | ||
11456 | if (!can_fragment(diri, dirs)) | |
11457 | return; | |
11458 | ||
11459 | CDir *first = dirs.front(); | |
11460 | int bits = first->get_frag().bits() - frag.bits(); | |
1911f103 | 11461 | dout(10) << " we are merging by " << bits << " bits" << dendl; |
7c673cae FG |
11462 | |
11463 | dirfrag_t basedirfrag(diri->ino(), frag); | |
11464 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FRAGMENTDIR); | |
11465 | mdr->more()->fragment_base = basedirfrag; | |
11466 | ||
11fdf7f2 | 11467 | ceph_assert(fragments.count(basedirfrag) == 0); |
7c673cae FG |
11468 | fragment_info_t& info = fragments[basedirfrag]; |
11469 | info.mdr = mdr; | |
11470 | info.dirs = dirs; | |
11471 | info.bits = -bits; | |
11472 | info.last_cum_auth_pins_change = ceph_clock_now(); | |
11473 | ||
11474 | fragment_freeze_dirs(dirs); | |
11475 | // initial mark+complete pass | |
11476 | fragment_mark_and_complete(mdr); | |
11477 | } | |
11478 | ||
9f95a23c | 11479 | void MDCache::fragment_freeze_dirs(const std::vector<CDir*>& dirs) |
7c673cae | 11480 | { |
11fdf7f2 | 11481 | bool any_subtree = false, any_non_subtree = false; |
9f95a23c | 11482 | for (const auto& dir : dirs) { |
7c673cae FG |
11483 | dir->auth_pin(dir); // until we mark and complete them |
11484 | dir->state_set(CDir::STATE_FRAGMENTING); | |
11485 | dir->freeze_dir(); | |
11fdf7f2 TL |
11486 | ceph_assert(dir->is_freezing_dir()); |
11487 | ||
11488 | if (dir->is_subtree_root()) | |
11489 | any_subtree = true; | |
11490 | else | |
11491 | any_non_subtree = true; | |
11492 | } | |
11493 | ||
11494 | if (any_subtree && any_non_subtree) { | |
11495 | // either all dirfrags are subtree roots or all are not. | |
9f95a23c | 11496 | for (const auto& dir : dirs) { |
11fdf7f2 TL |
11497 | if (dir->is_subtree_root()) { |
11498 | ceph_assert(dir->state_test(CDir::STATE_AUXSUBTREE)); | |
11499 | } else { | |
11500 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
11501 | adjust_subtree_auth(dir, mds->get_nodeid()); | |
11502 | } | |
11503 | } | |
7c673cae FG |
11504 | } |
11505 | } | |
11506 | ||
11507 | class C_MDC_FragmentMarking : public MDCacheContext { | |
11508 | MDRequestRef mdr; | |
11509 | public: | |
11510 | C_MDC_FragmentMarking(MDCache *m, MDRequestRef& r) : MDCacheContext(m), mdr(r) {} | |
11511 | void finish(int r) override { | |
11512 | mdcache->fragment_mark_and_complete(mdr); | |
11513 | } | |
11514 | }; | |
11515 | ||
11516 | void MDCache::fragment_mark_and_complete(MDRequestRef& mdr) | |
11517 | { | |
11518 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11519 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11520 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11521 | dout(7) << "fragment_mark_and_complete " << basedirfrag << " must have aborted" << dendl; | |
11522 | request_finish(mdr); | |
11523 | return; | |
11524 | } | |
11525 | ||
11526 | fragment_info_t& info = it->second; | |
11527 | CInode *diri = info.dirs.front()->get_inode(); | |
11528 | dout(10) << "fragment_mark_and_complete " << info.dirs << " on " << *diri << dendl; | |
11529 | ||
11530 | MDSGatherBuilder gather(g_ceph_context); | |
11531 | ||
9f95a23c | 11532 | for (const auto& dir : info.dirs) { |
7c673cae FG |
11533 | bool ready = true; |
11534 | if (!dir->is_complete()) { | |
11535 | dout(15) << " fetching incomplete " << *dir << dendl; | |
11536 | dir->fetch(gather.new_sub(), true); // ignore authpinnability | |
11537 | ready = false; | |
11538 | } else if (dir->get_frag() == frag_t()) { | |
11539 | // The COMPLETE flag gets lost if we fragment a new dirfrag, then rollback | |
11540 | // the operation. To avoid CDir::fetch() complaining about missing object, | |
11541 | // we commit new dirfrag first. | |
11542 | if (dir->state_test(CDir::STATE_CREATING)) { | |
11543 | dout(15) << " waiting until new dir gets journaled " << *dir << dendl; | |
11544 | dir->add_waiter(CDir::WAIT_CREATED, gather.new_sub()); | |
11545 | ready = false; | |
11546 | } else if (dir->is_new()) { | |
11547 | dout(15) << " committing new " << *dir << dendl; | |
11fdf7f2 | 11548 | ceph_assert(dir->is_dirty()); |
7c673cae FG |
11549 | dir->commit(0, gather.new_sub(), true); |
11550 | ready = false; | |
11551 | } | |
11552 | } | |
11553 | if (!ready) | |
11554 | continue; | |
11555 | ||
11556 | if (!dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
11557 | dout(15) << " marking " << *dir << dendl; | |
94b18763 FG |
11558 | for (auto &p : dir->items) { |
11559 | CDentry *dn = p.second; | |
7c673cae | 11560 | dn->get(CDentry::PIN_FRAGMENTING); |
11fdf7f2 | 11561 | ceph_assert(!dn->state_test(CDentry::STATE_FRAGMENTING)); |
7c673cae FG |
11562 | dn->state_set(CDentry::STATE_FRAGMENTING); |
11563 | } | |
11564 | dir->state_set(CDir::STATE_DNPINNEDFRAG); | |
11565 | dir->auth_unpin(dir); | |
11566 | } else { | |
11567 | dout(15) << " already marked " << *dir << dendl; | |
11568 | } | |
11569 | } | |
11570 | if (gather.has_subs()) { | |
11571 | gather.set_finisher(new C_MDC_FragmentMarking(this, mdr)); | |
11572 | gather.activate(); | |
11573 | return; | |
11574 | } | |
11575 | ||
9f95a23c | 11576 | for (const auto& dir : info.dirs) { |
7c673cae | 11577 | if (!dir->is_frozen_dir()) { |
11fdf7f2 | 11578 | ceph_assert(dir->is_freezing_dir()); |
7c673cae FG |
11579 | dir->add_waiter(CDir::WAIT_FROZEN, gather.new_sub()); |
11580 | } | |
11581 | } | |
11582 | if (gather.has_subs()) { | |
11583 | gather.set_finisher(new C_MDC_FragmentFrozen(this, mdr)); | |
11584 | gather.activate(); | |
11585 | // flush log so that request auth_pins are retired | |
11586 | mds->mdlog->flush(); | |
11587 | return; | |
11588 | } | |
11589 | ||
11590 | fragment_frozen(mdr, 0); | |
11591 | } | |
11592 | ||
9f95a23c | 11593 | void MDCache::fragment_unmark_unfreeze_dirs(const std::vector<CDir*>& dirs) |
7c673cae FG |
11594 | { |
11595 | dout(10) << "fragment_unmark_unfreeze_dirs " << dirs << dendl; | |
9f95a23c | 11596 | for (const auto& dir : dirs) { |
7c673cae FG |
11597 | dout(10) << " frag " << *dir << dendl; |
11598 | ||
11fdf7f2 | 11599 | ceph_assert(dir->state_test(CDir::STATE_FRAGMENTING)); |
7c673cae FG |
11600 | dir->state_clear(CDir::STATE_FRAGMENTING); |
11601 | ||
11602 | if (dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
11603 | dir->state_clear(CDir::STATE_DNPINNEDFRAG); | |
11604 | ||
94b18763 FG |
11605 | for (auto &p : dir->items) { |
11606 | CDentry *dn = p.second; | |
11fdf7f2 | 11607 | ceph_assert(dn->state_test(CDentry::STATE_FRAGMENTING)); |
7c673cae FG |
11608 | dn->state_clear(CDentry::STATE_FRAGMENTING); |
11609 | dn->put(CDentry::PIN_FRAGMENTING); | |
11610 | } | |
11611 | } else { | |
11612 | dir->auth_unpin(dir); | |
11613 | } | |
11614 | ||
11615 | dir->unfreeze_dir(); | |
11616 | } | |
11617 | } | |
11618 | ||
11619 | bool MDCache::fragment_are_all_frozen(CDir *dir) | |
11620 | { | |
11fdf7f2 | 11621 | ceph_assert(dir->is_frozen_dir()); |
7c673cae FG |
11622 | map<dirfrag_t,fragment_info_t>::iterator p; |
11623 | for (p = fragments.lower_bound(dirfrag_t(dir->ino(), 0)); | |
11624 | p != fragments.end() && p->first.ino == dir->ino(); | |
11625 | ++p) { | |
11626 | if (p->first.frag.contains(dir->get_frag())) | |
11627 | return p->second.all_frozen; | |
11628 | } | |
11629 | ceph_abort(); | |
11630 | return false; | |
11631 | } | |
11632 | ||
11633 | void MDCache::fragment_freeze_inc_num_waiters(CDir *dir) | |
11634 | { | |
11635 | map<dirfrag_t,fragment_info_t>::iterator p; | |
11636 | for (p = fragments.lower_bound(dirfrag_t(dir->ino(), 0)); | |
11637 | p != fragments.end() && p->first.ino == dir->ino(); | |
11638 | ++p) { | |
11639 | if (p->first.frag.contains(dir->get_frag())) { | |
11640 | p->second.num_remote_waiters++; | |
11641 | return; | |
11642 | } | |
11643 | } | |
11644 | ceph_abort(); | |
11645 | } | |
11646 | ||
11647 | void MDCache::find_stale_fragment_freeze() | |
11648 | { | |
11649 | dout(10) << "find_stale_fragment_freeze" << dendl; | |
11650 | // see comment in Migrator::find_stale_export_freeze() | |
11651 | utime_t now = ceph_clock_now(); | |
11652 | utime_t cutoff = now; | |
11fdf7f2 | 11653 | cutoff -= g_conf()->mds_freeze_tree_timeout; |
7c673cae FG |
11654 | |
11655 | for (map<dirfrag_t,fragment_info_t>::iterator p = fragments.begin(); | |
11656 | p != fragments.end(); ) { | |
11657 | dirfrag_t df = p->first; | |
11658 | fragment_info_t& info = p->second; | |
11659 | ++p; | |
11660 | if (info.all_frozen) | |
11661 | continue; | |
11662 | CDir *dir; | |
11663 | int total_auth_pins = 0; | |
9f95a23c TL |
11664 | for (const auto& d : info.dirs) { |
11665 | dir = d; | |
7c673cae FG |
11666 | if (!dir->state_test(CDir::STATE_DNPINNEDFRAG)) { |
11667 | total_auth_pins = -1; | |
11668 | break; | |
11669 | } | |
11670 | if (dir->is_frozen_dir()) | |
11671 | continue; | |
11672 | total_auth_pins += dir->get_auth_pins() + dir->get_dir_auth_pins(); | |
11673 | } | |
11674 | if (total_auth_pins < 0) | |
11675 | continue; | |
11676 | if (info.last_cum_auth_pins != total_auth_pins) { | |
11677 | info.last_cum_auth_pins = total_auth_pins; | |
11678 | info.last_cum_auth_pins_change = now; | |
11679 | continue; | |
11680 | } | |
11681 | if (info.last_cum_auth_pins_change >= cutoff) | |
11682 | continue; | |
11683 | dir = info.dirs.front(); | |
11684 | if (info.num_remote_waiters > 0 || | |
11685 | (!dir->inode->is_root() && dir->get_parent_dir()->is_freezing())) { | |
11686 | dout(10) << " cancel fragmenting " << df << " bit " << info.bits << dendl; | |
9f95a23c | 11687 | std::vector<CDir*> dirs; |
7c673cae FG |
11688 | info.dirs.swap(dirs); |
11689 | fragments.erase(df); | |
11690 | fragment_unmark_unfreeze_dirs(dirs); | |
11691 | } | |
11692 | } | |
11693 | } | |
11694 | ||
11695 | class C_MDC_FragmentPrep : public MDCacheLogContext { | |
11696 | MDRequestRef mdr; | |
11697 | public: | |
11698 | C_MDC_FragmentPrep(MDCache *m, MDRequestRef& r) : MDCacheLogContext(m), mdr(r) {} | |
11699 | void finish(int r) override { | |
11700 | mdcache->_fragment_logged(mdr); | |
11701 | } | |
11702 | }; | |
11703 | ||
11704 | class C_MDC_FragmentStore : public MDCacheContext { | |
11705 | MDRequestRef mdr; | |
11706 | public: | |
11707 | C_MDC_FragmentStore(MDCache *m, MDRequestRef& r) : MDCacheContext(m), mdr(r) {} | |
11708 | void finish(int r) override { | |
11709 | mdcache->_fragment_stored(mdr); | |
11710 | } | |
11711 | }; | |
11712 | ||
11713 | class C_MDC_FragmentCommit : public MDCacheLogContext { | |
11714 | dirfrag_t basedirfrag; | |
a8e16298 | 11715 | MDRequestRef mdr; |
7c673cae | 11716 | public: |
a8e16298 TL |
11717 | C_MDC_FragmentCommit(MDCache *m, dirfrag_t df, const MDRequestRef& r) : |
11718 | MDCacheLogContext(m), basedirfrag(df), mdr(r) {} | |
7c673cae | 11719 | void finish(int r) override { |
a8e16298 | 11720 | mdcache->_fragment_committed(basedirfrag, mdr); |
7c673cae FG |
11721 | } |
11722 | }; | |
11723 | ||
a8e16298 | 11724 | class C_IO_MDC_FragmentPurgeOld : public MDCacheIOContext { |
7c673cae | 11725 | dirfrag_t basedirfrag; |
a8e16298 TL |
11726 | int bits; |
11727 | MDRequestRef mdr; | |
7c673cae | 11728 | public: |
a8e16298 TL |
11729 | C_IO_MDC_FragmentPurgeOld(MDCache *m, dirfrag_t f, int b, |
11730 | const MDRequestRef& r) : | |
11731 | MDCacheIOContext(m), basedirfrag(f), bits(b), mdr(r) {} | |
7c673cae | 11732 | void finish(int r) override { |
f67539c2 | 11733 | ceph_assert(r == 0 || r == -CEPHFS_ENOENT); |
a8e16298 | 11734 | mdcache->_fragment_old_purged(basedirfrag, bits, mdr); |
7c673cae | 11735 | } |
91327a77 | 11736 | void print(ostream& out) const override { |
a8e16298 | 11737 | out << "fragment_purge_old(" << basedirfrag << ")"; |
91327a77 | 11738 | } |
7c673cae FG |
11739 | }; |
11740 | ||
11741 | void MDCache::fragment_frozen(MDRequestRef& mdr, int r) | |
11742 | { | |
11743 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11744 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11745 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11746 | dout(7) << "fragment_frozen " << basedirfrag << " must have aborted" << dendl; | |
11747 | request_finish(mdr); | |
11748 | return; | |
11749 | } | |
11750 | ||
11fdf7f2 | 11751 | ceph_assert(r == 0); |
7c673cae FG |
11752 | fragment_info_t& info = it->second; |
11753 | dout(10) << "fragment_frozen " << basedirfrag.frag << " by " << info.bits | |
11754 | << " on " << info.dirs.front()->get_inode() << dendl; | |
11755 | ||
11756 | info.all_frozen = true; | |
11757 | dispatch_fragment_dir(mdr); | |
11758 | } | |
11759 | ||
11760 | void MDCache::dispatch_fragment_dir(MDRequestRef& mdr) | |
11761 | { | |
11762 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11763 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11764 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11765 | dout(7) << "dispatch_fragment_dir " << basedirfrag << " must have aborted" << dendl; | |
11766 | request_finish(mdr); | |
11767 | return; | |
11768 | } | |
11769 | ||
11770 | fragment_info_t& info = it->second; | |
11771 | CInode *diri = info.dirs.front()->get_inode(); | |
11772 | ||
11773 | dout(10) << "dispatch_fragment_dir " << basedirfrag << " bits " << info.bits | |
11774 | << " on " << *diri << dendl; | |
9f95a23c | 11775 | |
f67539c2 | 11776 | if (mdr->more()->peer_error) |
9f95a23c TL |
11777 | mdr->aborted = true; |
11778 | ||
7c673cae | 11779 | if (!mdr->aborted) { |
11fdf7f2 TL |
11780 | MutationImpl::LockOpVec lov; |
11781 | lov.add_wrlock(&diri->dirfragtreelock); | |
7c673cae | 11782 | // prevent a racing gather on any other scatterlocks too |
9f95a23c TL |
11783 | lov.lock_scatter_gather(&diri->nestlock); |
11784 | lov.lock_scatter_gather(&diri->filelock); | |
11785 | if (!mds->locker->acquire_locks(mdr, lov, NULL, true)) { | |
7c673cae FG |
11786 | if (!mdr->aborted) |
11787 | return; | |
9f95a23c | 11788 | } |
7c673cae FG |
11789 | } |
11790 | ||
11791 | if (mdr->aborted) { | |
11792 | dout(10) << " can't auth_pin " << *diri << ", requeuing dir " | |
11793 | << info.dirs.front()->dirfrag() << dendl; | |
11794 | if (info.bits > 0) | |
11795 | mds->balancer->queue_split(info.dirs.front(), false); | |
11796 | else | |
11797 | mds->balancer->queue_merge(info.dirs.front()); | |
11798 | fragment_unmark_unfreeze_dirs(info.dirs); | |
11799 | fragments.erase(it); | |
11800 | request_finish(mdr); | |
11801 | return; | |
11802 | } | |
11803 | ||
11804 | mdr->ls = mds->mdlog->get_current_segment(); | |
11805 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_PREPARE, basedirfrag, info.bits); | |
11806 | mds->mdlog->start_entry(le); | |
11807 | ||
9f95a23c | 11808 | for (const auto& dir : info.dirs) { |
7c673cae FG |
11809 | dirfrag_rollback rollback; |
11810 | rollback.fnode = dir->fnode; | |
11811 | le->add_orig_frag(dir->get_frag(), &rollback); | |
11812 | } | |
11813 | ||
11814 | // refragment | |
11fdf7f2 | 11815 | MDSContext::vec waiters; |
7c673cae | 11816 | adjust_dir_fragments(diri, info.dirs, basedirfrag.frag, info.bits, |
9f95a23c | 11817 | &info.resultfrags, waiters, false); |
11fdf7f2 | 11818 | if (g_conf()->mds_debug_frag) |
7c673cae FG |
11819 | diri->verify_dirfrags(); |
11820 | mds->queue_waiters(waiters); | |
11821 | ||
11fdf7f2 TL |
11822 | for (const auto& fg : le->orig_frags) |
11823 | ceph_assert(!diri->dirfragtree.is_leaf(fg)); | |
7c673cae | 11824 | |
9f95a23c TL |
11825 | le->metablob.add_dir_context(info.resultfrags.front()); |
11826 | for (const auto& dir : info.resultfrags) { | |
7c673cae | 11827 | if (diri->is_auth()) { |
9f95a23c | 11828 | le->metablob.add_fragmented_dir(dir, false, false); |
7c673cae | 11829 | } else { |
9f95a23c TL |
11830 | dir->state_set(CDir::STATE_DIRTYDFT); |
11831 | le->metablob.add_fragmented_dir(dir, false, true); | |
7c673cae FG |
11832 | } |
11833 | } | |
11834 | ||
11835 | // dft lock | |
11836 | if (diri->is_auth()) { | |
11837 | // journal dirfragtree | |
f67539c2 TL |
11838 | auto pi = diri->project_inode(mdr); |
11839 | pi.inode->version = diri->pre_dirty(); | |
11840 | predirty_journal_parents(mdr, &le->metablob, diri, 0, PREDIRTY_PRIMARY); | |
7c673cae FG |
11841 | journal_dirty_inode(mdr.get(), &le->metablob, diri); |
11842 | } else { | |
11843 | mds->locker->mark_updated_scatterlock(&diri->dirfragtreelock); | |
11844 | mdr->ls->dirty_dirfrag_dirfragtree.push_back(&diri->item_dirty_dirfrag_dirfragtree); | |
11845 | mdr->add_updated_lock(&diri->dirfragtreelock); | |
11846 | } | |
11847 | ||
11848 | /* | |
11849 | // filelock | |
11850 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
11851 | mut->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
11852 | mut->add_updated_lock(&diri->filelock); | |
11853 | ||
11854 | // dirlock | |
11855 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
11856 | mut->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
11857 | mut->add_updated_lock(&diri->nestlock); | |
11858 | */ | |
11859 | ||
11860 | add_uncommitted_fragment(basedirfrag, info.bits, le->orig_frags, mdr->ls); | |
11861 | mds->server->submit_mdlog_entry(le, new C_MDC_FragmentPrep(this, mdr), | |
11862 | mdr, __func__); | |
11863 | mds->mdlog->flush(); | |
11864 | } | |
11865 | ||
11866 | void MDCache::_fragment_logged(MDRequestRef& mdr) | |
11867 | { | |
11868 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
a8e16298 | 11869 | auto& info = fragments.at(basedirfrag); |
7c673cae FG |
11870 | CInode *diri = info.resultfrags.front()->get_inode(); |
11871 | ||
11872 | dout(10) << "fragment_logged " << basedirfrag << " bits " << info.bits | |
11873 | << " on " << *diri << dendl; | |
a8e16298 | 11874 | mdr->mark_event("prepare logged"); |
7c673cae | 11875 | |
7c673cae FG |
11876 | mdr->apply(); // mark scatterlock |
11877 | ||
11878 | // store resulting frags | |
11879 | MDSGatherBuilder gather(g_ceph_context, new C_MDC_FragmentStore(this, mdr)); | |
11880 | ||
9f95a23c | 11881 | for (const auto& dir : info.resultfrags) { |
7c673cae FG |
11882 | dout(10) << " storing result frag " << *dir << dendl; |
11883 | ||
f67539c2 | 11884 | dir->mark_dirty(mdr->ls); |
f91f0fd5 TL |
11885 | dir->mark_new(mdr->ls); |
11886 | ||
7c673cae FG |
11887 | // freeze and store them too |
11888 | dir->auth_pin(this); | |
11889 | dir->state_set(CDir::STATE_FRAGMENTING); | |
11890 | dir->commit(0, gather.new_sub(), true); // ignore authpinnability | |
11891 | } | |
11892 | ||
11893 | gather.activate(); | |
11894 | } | |
11895 | ||
11896 | void MDCache::_fragment_stored(MDRequestRef& mdr) | |
11897 | { | |
11898 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
a8e16298 TL |
11899 | fragment_info_t &info = fragments.at(basedirfrag); |
11900 | CDir *first = info.resultfrags.front(); | |
11901 | CInode *diri = first->get_inode(); | |
7c673cae FG |
11902 | |
11903 | dout(10) << "fragment_stored " << basedirfrag << " bits " << info.bits | |
11904 | << " on " << *diri << dendl; | |
a8e16298 | 11905 | mdr->mark_event("new frags stored"); |
7c673cae FG |
11906 | |
11907 | // tell peers | |
a8e16298 TL |
11908 | mds_rank_t diri_auth = (first->is_subtree_root() && !diri->is_auth()) ? |
11909 | diri->authority().first : CDIR_AUTH_UNKNOWN; | |
181888fb FG |
11910 | for (const auto &p : first->get_replicas()) { |
11911 | if (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || | |
11912 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
11913 | rejoin_gather.count(p.first))) | |
7c673cae FG |
11914 | continue; |
11915 | ||
9f95a23c | 11916 | auto notify = make_message<MMDSFragmentNotify>(basedirfrag, info.bits, mdr->reqid.tid); |
a8e16298 TL |
11917 | if (diri_auth != CDIR_AUTH_UNKNOWN && // subtree root |
11918 | diri_auth != p.first) { // not auth mds of diri | |
11919 | /* | |
11920 | * In the nornal case, mds does not trim dir inode whose child dirfrags | |
11921 | * are likely being fragmented (see trim_inode()). But when fragmenting | |
11922 | * subtree roots, following race can happen: | |
11923 | * | |
11924 | * - mds.a (auth mds of dirfrag) sends fragment_notify message to | |
11925 | * mds.c and drops wrlock on dirfragtreelock. | |
11926 | * - mds.b (auth mds of dir inode) changes dirfragtreelock state to | |
11927 | * SYNC and send lock message mds.c | |
11928 | * - mds.c receives the lock message and changes dirfragtreelock state | |
11929 | * to SYNC | |
11930 | * - mds.c trim dirfrag and dir inode from its cache | |
11931 | * - mds.c receives the fragment_notify message | |
11932 | * | |
11933 | * So we need to ensure replicas have received the notify, then unlock | |
11934 | * the dirfragtreelock. | |
11935 | */ | |
11936 | notify->mark_ack_wanted(); | |
11937 | info.notify_ack_waiting.insert(p.first); | |
11938 | } | |
7c673cae FG |
11939 | |
11940 | // freshly replicate new dirs to peers | |
9f95a23c TL |
11941 | for (const auto& dir : info.resultfrags) { |
11942 | encode_replica_dir(dir, p.first, notify->basebl); | |
11943 | } | |
7c673cae | 11944 | |
181888fb | 11945 | mds->send_message_mds(notify, p.first); |
7c673cae FG |
11946 | } |
11947 | ||
11948 | // journal commit | |
11949 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_COMMIT, basedirfrag, info.bits); | |
a8e16298 | 11950 | mds->mdlog->start_submit_entry(le, new C_MDC_FragmentCommit(this, basedirfrag, mdr)); |
7c673cae | 11951 | |
7c673cae FG |
11952 | |
11953 | // unfreeze resulting frags | |
9f95a23c | 11954 | for (const auto& dir : info.resultfrags) { |
7c673cae FG |
11955 | dout(10) << " result frag " << *dir << dendl; |
11956 | ||
94b18763 FG |
11957 | for (auto &p : dir->items) { |
11958 | CDentry *dn = p.second; | |
11fdf7f2 | 11959 | ceph_assert(dn->state_test(CDentry::STATE_FRAGMENTING)); |
7c673cae FG |
11960 | dn->state_clear(CDentry::STATE_FRAGMENTING); |
11961 | dn->put(CDentry::PIN_FRAGMENTING); | |
11962 | } | |
11963 | ||
11964 | // unfreeze | |
11965 | dir->unfreeze_dir(); | |
11966 | } | |
11967 | ||
a8e16298 TL |
11968 | if (info.notify_ack_waiting.empty()) { |
11969 | fragment_drop_locks(info); | |
11970 | } else { | |
11971 | mds->locker->drop_locks_for_fragment_unfreeze(mdr.get()); | |
11972 | } | |
7c673cae FG |
11973 | } |
11974 | ||
a8e16298 | 11975 | void MDCache::_fragment_committed(dirfrag_t basedirfrag, const MDRequestRef& mdr) |
7c673cae FG |
11976 | { |
11977 | dout(10) << "fragment_committed " << basedirfrag << dendl; | |
a8e16298 TL |
11978 | if (mdr) |
11979 | mdr->mark_event("commit logged"); | |
11980 | ||
11981 | ufragment &uf = uncommitted_fragments.at(basedirfrag); | |
7c673cae FG |
11982 | |
11983 | // remove old frags | |
11984 | C_GatherBuilder gather( | |
11985 | g_ceph_context, | |
11986 | new C_OnFinisher( | |
a8e16298 | 11987 | new C_IO_MDC_FragmentPurgeOld(this, basedirfrag, uf.bits, mdr), |
7c673cae FG |
11988 | mds->finisher)); |
11989 | ||
11990 | SnapContext nullsnapc; | |
b3b6e05e | 11991 | object_locator_t oloc(mds->get_metadata_pool()); |
11fdf7f2 TL |
11992 | for (const auto& fg : uf.old_frags) { |
11993 | object_t oid = CInode::get_object_name(basedirfrag.ino, fg, ""); | |
7c673cae | 11994 | ObjectOperation op; |
11fdf7f2 | 11995 | if (fg == frag_t()) { |
7c673cae FG |
11996 | // backtrace object |
11997 | dout(10) << " truncate orphan dirfrag " << oid << dendl; | |
11998 | op.truncate(0); | |
11999 | op.omap_clear(); | |
12000 | } else { | |
12001 | dout(10) << " removing orphan dirfrag " << oid << dendl; | |
12002 | op.remove(); | |
12003 | } | |
12004 | mds->objecter->mutate(oid, oloc, op, nullsnapc, | |
12005 | ceph::real_clock::now(), | |
12006 | 0, gather.new_sub()); | |
12007 | } | |
12008 | ||
11fdf7f2 | 12009 | ceph_assert(gather.has_subs()); |
7c673cae FG |
12010 | gather.activate(); |
12011 | } | |
12012 | ||
a8e16298 | 12013 | void MDCache::_fragment_old_purged(dirfrag_t basedirfrag, int bits, const MDRequestRef& mdr) |
7c673cae | 12014 | { |
a8e16298 TL |
12015 | dout(10) << "fragment_old_purged " << basedirfrag << dendl; |
12016 | if (mdr) | |
12017 | mdr->mark_event("old frags purged"); | |
12018 | ||
12019 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_FINISH, basedirfrag, bits); | |
12020 | mds->mdlog->start_submit_entry(le); | |
12021 | ||
12022 | finish_uncommitted_fragment(basedirfrag, EFragment::OP_FINISH); | |
12023 | ||
12024 | if (mds->logger) { | |
12025 | if (bits > 0) { | |
12026 | mds->logger->inc(l_mds_dir_split); | |
12027 | } else { | |
12028 | mds->logger->inc(l_mds_dir_merge); | |
12029 | } | |
12030 | } | |
12031 | ||
12032 | if (mdr) { | |
12033 | auto it = fragments.find(basedirfrag); | |
12034 | ceph_assert(it != fragments.end()); | |
12035 | it->second.finishing = true; | |
12036 | if (it->second.notify_ack_waiting.empty()) | |
12037 | fragment_maybe_finish(it); | |
12038 | else | |
12039 | mdr->mark_event("wating for notify acks"); | |
12040 | } | |
12041 | } | |
12042 | ||
12043 | void MDCache::fragment_drop_locks(fragment_info_t& info) | |
12044 | { | |
12045 | mds->locker->drop_locks(info.mdr.get()); | |
12046 | request_finish(info.mdr); | |
12047 | //info.mdr.reset(); | |
12048 | } | |
12049 | ||
12050 | void MDCache::fragment_maybe_finish(const fragment_info_iterator& it) | |
12051 | { | |
12052 | if (!it->second.finishing) | |
12053 | return; | |
7c673cae FG |
12054 | |
12055 | // unmark & auth_unpin | |
a8e16298 | 12056 | for (const auto &dir : it->second.resultfrags) { |
7c673cae FG |
12057 | dir->state_clear(CDir::STATE_FRAGMENTING); |
12058 | dir->auth_unpin(this); | |
12059 | ||
12060 | // In case the resulting fragments are beyond the split size, | |
12061 | // we might need to split them again right away (they could | |
12062 | // have been taking inserts between unfreezing and getting | |
12063 | // here) | |
12064 | mds->balancer->maybe_fragment(dir, false); | |
12065 | } | |
12066 | ||
a8e16298 TL |
12067 | fragments.erase(it); |
12068 | } | |
12069 | ||
12070 | ||
9f95a23c | 12071 | void MDCache::handle_fragment_notify_ack(const cref_t<MMDSFragmentNotifyAck> &ack) |
a8e16298 TL |
12072 | { |
12073 | dout(10) << "handle_fragment_notify_ack " << *ack << " from " << ack->get_source() << dendl; | |
12074 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
12075 | ||
12076 | if (mds->get_state() < MDSMap::STATE_ACTIVE) { | |
a8e16298 | 12077 | return; |
7c673cae FG |
12078 | } |
12079 | ||
a8e16298 TL |
12080 | auto it = fragments.find(ack->get_base_dirfrag()); |
12081 | if (it == fragments.end() || | |
12082 | it->second.get_tid() != ack->get_tid()) { | |
12083 | dout(10) << "handle_fragment_notify_ack obsolete message, dropping" << dendl; | |
a8e16298 TL |
12084 | return; |
12085 | } | |
7c673cae | 12086 | |
a8e16298 TL |
12087 | if (it->second.notify_ack_waiting.erase(from) && |
12088 | it->second.notify_ack_waiting.empty()) { | |
12089 | fragment_drop_locks(it->second); | |
12090 | fragment_maybe_finish(it); | |
12091 | } | |
7c673cae FG |
12092 | } |
12093 | ||
9f95a23c | 12094 | void MDCache::handle_fragment_notify(const cref_t<MMDSFragmentNotify> ¬ify) |
7c673cae FG |
12095 | { |
12096 | dout(10) << "handle_fragment_notify " << *notify << " from " << notify->get_source() << dendl; | |
a8e16298 | 12097 | mds_rank_t from = mds_rank_t(notify->get_source().num()); |
7c673cae FG |
12098 | |
12099 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
7c673cae FG |
12100 | return; |
12101 | } | |
12102 | ||
12103 | CInode *diri = get_inode(notify->get_ino()); | |
12104 | if (diri) { | |
12105 | frag_t base = notify->get_basefrag(); | |
12106 | int bits = notify->get_bits(); | |
12107 | ||
12108 | /* | |
12109 | if ((bits < 0 && diri->dirfragtree.is_leaf(base)) || | |
12110 | (bits > 0 && !diri->dirfragtree.is_leaf(base))) { | |
12111 | dout(10) << " dft " << diri->dirfragtree << " state doesn't match " << base << " by " << bits | |
12112 | << ", must have found out during resolve/rejoin? ignoring. " << *diri << dendl; | |
7c673cae FG |
12113 | return; |
12114 | } | |
12115 | */ | |
12116 | ||
12117 | // refragment | |
11fdf7f2 | 12118 | MDSContext::vec waiters; |
9f95a23c TL |
12119 | std::vector<CDir*> resultfrags; |
12120 | adjust_dir_fragments(diri, base, bits, &resultfrags, waiters, false); | |
11fdf7f2 | 12121 | if (g_conf()->mds_debug_frag) |
7c673cae FG |
12122 | diri->verify_dirfrags(); |
12123 | ||
9f95a23c TL |
12124 | for (const auto& dir : resultfrags) { |
12125 | diri->take_dir_waiting(dir->get_frag(), waiters); | |
12126 | } | |
7c673cae FG |
12127 | |
12128 | // add new replica dirs values | |
11fdf7f2 | 12129 | auto p = notify->basebl.cbegin(); |
9f95a23c TL |
12130 | while (!p.end()) { |
12131 | CDir *tmp_dir = nullptr; | |
12132 | decode_replica_dir(tmp_dir, p, diri, from, waiters); | |
12133 | } | |
7c673cae FG |
12134 | |
12135 | mds->queue_waiters(waiters); | |
12136 | } else { | |
12137 | ceph_abort(); | |
12138 | } | |
12139 | ||
a8e16298 | 12140 | if (notify->is_ack_wanted()) { |
9f95a23c | 12141 | auto ack = make_message<MMDSFragmentNotifyAck>(notify->get_base_dirfrag(), |
11fdf7f2 | 12142 | notify->get_bits(), notify->get_tid()); |
a8e16298 TL |
12143 | mds->send_message_mds(ack, from); |
12144 | } | |
7c673cae FG |
12145 | } |
12146 | ||
11fdf7f2 | 12147 | void MDCache::add_uncommitted_fragment(dirfrag_t basedirfrag, int bits, const frag_vec_t& old_frags, |
7c673cae FG |
12148 | LogSegment *ls, bufferlist *rollback) |
12149 | { | |
12150 | dout(10) << "add_uncommitted_fragment: base dirfrag " << basedirfrag << " bits " << bits << dendl; | |
11fdf7f2 | 12151 | ceph_assert(!uncommitted_fragments.count(basedirfrag)); |
7c673cae FG |
12152 | ufragment& uf = uncommitted_fragments[basedirfrag]; |
12153 | uf.old_frags = old_frags; | |
12154 | uf.bits = bits; | |
12155 | uf.ls = ls; | |
12156 | ls->uncommitted_fragments.insert(basedirfrag); | |
12157 | if (rollback) | |
12158 | uf.rollback.swap(*rollback); | |
12159 | } | |
12160 | ||
12161 | void MDCache::finish_uncommitted_fragment(dirfrag_t basedirfrag, int op) | |
12162 | { | |
12163 | dout(10) << "finish_uncommitted_fragments: base dirfrag " << basedirfrag | |
12164 | << " op " << EFragment::op_name(op) << dendl; | |
12165 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
12166 | if (it != uncommitted_fragments.end()) { | |
12167 | ufragment& uf = it->second; | |
12168 | if (op != EFragment::OP_FINISH && !uf.old_frags.empty()) { | |
12169 | uf.committed = true; | |
12170 | } else { | |
12171 | uf.ls->uncommitted_fragments.erase(basedirfrag); | |
12172 | mds->queue_waiters(uf.waiters); | |
12173 | uncommitted_fragments.erase(it); | |
12174 | } | |
12175 | } | |
12176 | } | |
12177 | ||
11fdf7f2 | 12178 | void MDCache::rollback_uncommitted_fragment(dirfrag_t basedirfrag, frag_vec_t&& old_frags) |
7c673cae FG |
12179 | { |
12180 | dout(10) << "rollback_uncommitted_fragment: base dirfrag " << basedirfrag | |
12181 | << " old_frags (" << old_frags << ")" << dendl; | |
12182 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
12183 | if (it != uncommitted_fragments.end()) { | |
12184 | ufragment& uf = it->second; | |
12185 | if (!uf.old_frags.empty()) { | |
11fdf7f2 | 12186 | uf.old_frags = std::move(old_frags); |
7c673cae FG |
12187 | uf.committed = true; |
12188 | } else { | |
12189 | uf.ls->uncommitted_fragments.erase(basedirfrag); | |
12190 | uncommitted_fragments.erase(it); | |
12191 | } | |
12192 | } | |
12193 | } | |
12194 | ||
f91f0fd5 | 12195 | void MDCache::wait_for_uncommitted_fragments(MDSContext* finisher) |
e306af50 | 12196 | { |
f91f0fd5 TL |
12197 | MDSGatherBuilder gather(g_ceph_context, finisher); |
12198 | for (auto& p : uncommitted_fragments) { | |
12199 | p.second.waiters.push_back(gather.new_sub()); | |
12200 | } | |
12201 | gather.activate(); | |
e306af50 TL |
12202 | } |
12203 | ||
f67539c2 TL |
12204 | struct C_MDC_FragmentRollback : public MDCacheLogContext { |
12205 | MutationRef mut; | |
12206 | C_MDC_FragmentRollback(MDCache *c, MutationRef& m) : | |
12207 | MDCacheLogContext(c), mut(m) {} | |
12208 | void finish(int r) override { | |
12209 | mut->apply(); | |
12210 | get_mds()->locker->drop_locks(mut.get()); | |
12211 | mut->cleanup(); | |
12212 | } | |
12213 | }; | |
12214 | ||
7c673cae FG |
12215 | void MDCache::rollback_uncommitted_fragments() |
12216 | { | |
12217 | dout(10) << "rollback_uncommitted_fragments: " << uncommitted_fragments.size() << " pending" << dendl; | |
12218 | for (map<dirfrag_t, ufragment>::iterator p = uncommitted_fragments.begin(); | |
12219 | p != uncommitted_fragments.end(); | |
12220 | ++p) { | |
12221 | ufragment &uf = p->second; | |
12222 | CInode *diri = get_inode(p->first.ino); | |
11fdf7f2 | 12223 | ceph_assert(diri); |
7c673cae FG |
12224 | |
12225 | if (uf.committed) { | |
a8e16298 | 12226 | _fragment_committed(p->first, MDRequestRef()); |
7c673cae FG |
12227 | continue; |
12228 | } | |
12229 | ||
12230 | dout(10) << " rolling back " << p->first << " refragment by " << uf.bits << " bits" << dendl; | |
12231 | ||
f67539c2 TL |
12232 | MutationRef mut(new MutationImpl()); |
12233 | mut->ls = mds->mdlog->get_current_segment(); | |
7c673cae FG |
12234 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_ROLLBACK, p->first, uf.bits); |
12235 | mds->mdlog->start_entry(le); | |
12236 | bool diri_auth = (diri->authority() != CDIR_AUTH_UNDEF); | |
12237 | ||
11fdf7f2 | 12238 | frag_vec_t old_frags; |
7c673cae FG |
12239 | diri->dirfragtree.get_leaves_under(p->first.frag, old_frags); |
12240 | ||
9f95a23c | 12241 | std::vector<CDir*> resultfrags; |
7c673cae FG |
12242 | if (uf.old_frags.empty()) { |
12243 | // created by old format EFragment | |
11fdf7f2 | 12244 | MDSContext::vec waiters; |
9f95a23c | 12245 | adjust_dir_fragments(diri, p->first.frag, -uf.bits, &resultfrags, waiters, true); |
7c673cae | 12246 | } else { |
11fdf7f2 TL |
12247 | auto bp = uf.rollback.cbegin(); |
12248 | for (const auto& fg : uf.old_frags) { | |
12249 | CDir *dir = force_dir_fragment(diri, fg); | |
7c673cae FG |
12250 | resultfrags.push_back(dir); |
12251 | ||
12252 | dirfrag_rollback rollback; | |
11fdf7f2 | 12253 | decode(rollback, bp); |
7c673cae | 12254 | |
7c673cae FG |
12255 | dir->fnode = rollback.fnode; |
12256 | ||
f67539c2 | 12257 | dir->mark_dirty(mut->ls); |
7c673cae | 12258 | |
f67539c2 | 12259 | if (!(dir->get_fnode()->rstat == dir->get_fnode()->accounted_rstat)) { |
7c673cae | 12260 | dout(10) << " dirty nestinfo on " << *dir << dendl; |
f67539c2 TL |
12261 | mds->locker->mark_updated_scatterlock(&diri->nestlock); |
12262 | mut->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
12263 | mut->add_updated_lock(&diri->nestlock); | |
7c673cae | 12264 | } |
f67539c2 | 12265 | if (!(dir->get_fnode()->fragstat == dir->get_fnode()->accounted_fragstat)) { |
7c673cae | 12266 | dout(10) << " dirty fragstat on " << *dir << dendl; |
f67539c2 TL |
12267 | mds->locker->mark_updated_scatterlock(&diri->filelock); |
12268 | mut->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
12269 | mut->add_updated_lock(&diri->filelock); | |
7c673cae FG |
12270 | } |
12271 | ||
12272 | le->add_orig_frag(dir->get_frag()); | |
12273 | le->metablob.add_dir_context(dir); | |
12274 | if (diri_auth) { | |
12275 | le->metablob.add_fragmented_dir(dir, true, false); | |
12276 | } else { | |
12277 | dout(10) << " dirty dirfragtree on " << *dir << dendl; | |
12278 | dir->state_set(CDir::STATE_DIRTYDFT); | |
12279 | le->metablob.add_fragmented_dir(dir, true, true); | |
12280 | } | |
12281 | } | |
12282 | } | |
12283 | ||
12284 | if (diri_auth) { | |
f67539c2 TL |
12285 | auto pi = diri->project_inode(mut); |
12286 | pi.inode->version = diri->pre_dirty(); | |
12287 | predirty_journal_parents(mut, &le->metablob, diri, 0, PREDIRTY_PRIMARY); | |
7c673cae FG |
12288 | le->metablob.add_primary_dentry(diri->get_projected_parent_dn(), diri, true); |
12289 | } else { | |
12290 | mds->locker->mark_updated_scatterlock(&diri->dirfragtreelock); | |
f67539c2 TL |
12291 | mut->ls->dirty_dirfrag_dirfragtree.push_back(&diri->item_dirty_dirfrag_dirfragtree); |
12292 | mut->add_updated_lock(&diri->dirfragtreelock); | |
7c673cae FG |
12293 | } |
12294 | ||
11fdf7f2 | 12295 | if (g_conf()->mds_debug_frag) |
7c673cae FG |
12296 | diri->verify_dirfrags(); |
12297 | ||
11fdf7f2 TL |
12298 | for (const auto& leaf : old_frags) { |
12299 | ceph_assert(!diri->dirfragtree.is_leaf(leaf)); | |
12300 | } | |
7c673cae | 12301 | |
f67539c2 | 12302 | mds->mdlog->submit_entry(le, new C_MDC_FragmentRollback(this, mut)); |
7c673cae FG |
12303 | |
12304 | uf.old_frags.swap(old_frags); | |
a8e16298 | 12305 | _fragment_committed(p->first, MDRequestRef()); |
7c673cae FG |
12306 | } |
12307 | } | |
12308 | ||
12309 | void MDCache::force_readonly() | |
12310 | { | |
12311 | if (is_readonly()) | |
12312 | return; | |
12313 | ||
12314 | dout(1) << "force file system read-only" << dendl; | |
12315 | mds->clog->warn() << "force file system read-only"; | |
12316 | ||
12317 | set_readonly(); | |
12318 | ||
12319 | mds->server->force_clients_readonly(); | |
12320 | ||
12321 | // revoke write caps | |
81eedcae | 12322 | int count = 0; |
94b18763 | 12323 | for (auto &p : inode_map) { |
b32b8144 | 12324 | CInode *in = p.second; |
7c673cae FG |
12325 | if (in->is_head()) |
12326 | mds->locker->eval(in, CEPH_CAP_LOCKS); | |
33c7a0ef | 12327 | if (!(++count % mds->heartbeat_reset_grace())) |
81eedcae | 12328 | mds->heartbeat_reset(); |
7c673cae FG |
12329 | } |
12330 | ||
12331 | mds->mdlog->flush(); | |
12332 | } | |
12333 | ||
12334 | ||
12335 | // ============================================================== | |
12336 | // debug crap | |
12337 | ||
81eedcae | 12338 | void MDCache::show_subtrees(int dbl, bool force_print) |
7c673cae | 12339 | { |
11fdf7f2 | 12340 | if (g_conf()->mds_thrash_exports) |
7c673cae FG |
12341 | dbl += 15; |
12342 | ||
12343 | //dout(10) << "show_subtrees" << dendl; | |
12344 | ||
11fdf7f2 | 12345 | if (!g_conf()->subsys.should_gather(ceph_subsys_mds, dbl)) |
7c673cae FG |
12346 | return; // i won't print anything. |
12347 | ||
12348 | if (subtrees.empty()) { | |
11fdf7f2 TL |
12349 | dout(ceph::dout::need_dynamic(dbl)) << "show_subtrees - no subtrees" |
12350 | << dendl; | |
7c673cae FG |
12351 | return; |
12352 | } | |
12353 | ||
81eedcae TL |
12354 | if (!force_print && subtrees.size() > SUBTREES_COUNT_THRESHOLD && |
12355 | !g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) { | |
12356 | dout(ceph::dout::need_dynamic(dbl)) << "number of subtrees = " << subtrees.size() << "; not " | |
12357 | "printing subtrees" << dendl; | |
12358 | return; | |
12359 | } | |
12360 | ||
7c673cae | 12361 | // root frags |
9f95a23c | 12362 | std::vector<CDir*> basefrags; |
7c673cae FG |
12363 | for (set<CInode*>::iterator p = base_inodes.begin(); |
12364 | p != base_inodes.end(); | |
12365 | ++p) | |
12366 | (*p)->get_dirfrags(basefrags); | |
12367 | //dout(15) << "show_subtrees, base dirfrags " << basefrags << dendl; | |
12368 | dout(15) << "show_subtrees" << dendl; | |
12369 | ||
12370 | // queue stuff | |
12371 | list<pair<CDir*,int> > q; | |
12372 | string indent; | |
12373 | set<CDir*> seen; | |
12374 | ||
12375 | // calc max depth | |
9f95a23c TL |
12376 | for (const auto& dir : basefrags) { |
12377 | q.emplace_back(dir, 0); | |
12378 | } | |
7c673cae FG |
12379 | |
12380 | set<CDir*> subtrees_seen; | |
12381 | ||
81eedcae | 12382 | unsigned int depth = 0; |
7c673cae FG |
12383 | while (!q.empty()) { |
12384 | CDir *dir = q.front().first; | |
81eedcae | 12385 | unsigned int d = q.front().second; |
7c673cae FG |
12386 | q.pop_front(); |
12387 | ||
12388 | if (subtrees.count(dir) == 0) continue; | |
12389 | ||
12390 | subtrees_seen.insert(dir); | |
12391 | ||
12392 | if (d > depth) depth = d; | |
12393 | ||
12394 | // sanity check | |
12395 | //dout(25) << "saw depth " << d << " " << *dir << dendl; | |
12396 | if (seen.count(dir)) dout(0) << "aah, already seen " << *dir << dendl; | |
11fdf7f2 | 12397 | ceph_assert(seen.count(dir) == 0); |
7c673cae FG |
12398 | seen.insert(dir); |
12399 | ||
12400 | // nested items? | |
12401 | if (!subtrees[dir].empty()) { | |
12402 | for (set<CDir*>::iterator p = subtrees[dir].begin(); | |
12403 | p != subtrees[dir].end(); | |
12404 | ++p) { | |
12405 | //dout(25) << " saw sub " << **p << dendl; | |
12406 | q.push_front(pair<CDir*,int>(*p, d+1)); | |
12407 | } | |
12408 | } | |
12409 | } | |
12410 | ||
81eedcae TL |
12411 | if (!force_print && depth > SUBTREES_DEPTH_THRESHOLD && |
12412 | !g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) { | |
12413 | dout(ceph::dout::need_dynamic(dbl)) << "max depth among subtrees = " << depth << "; not printing " | |
12414 | "subtrees" << dendl; | |
12415 | return; | |
12416 | } | |
7c673cae FG |
12417 | |
12418 | // print tree | |
9f95a23c TL |
12419 | for (const auto& dir : basefrags) { |
12420 | q.emplace_back(dir, 0); | |
12421 | } | |
7c673cae FG |
12422 | |
12423 | while (!q.empty()) { | |
12424 | CDir *dir = q.front().first; | |
12425 | int d = q.front().second; | |
12426 | q.pop_front(); | |
12427 | ||
12428 | if (subtrees.count(dir) == 0) continue; | |
12429 | ||
12430 | // adjust indenter | |
12431 | while ((unsigned)d < indent.size()) | |
12432 | indent.resize(d); | |
12433 | ||
12434 | // pad | |
12435 | string pad = "______________________________________"; | |
12436 | pad.resize(depth*2+1-indent.size()); | |
12437 | if (!subtrees[dir].empty()) | |
12438 | pad[0] = '.'; // parent | |
12439 | ||
12440 | ||
12441 | string auth; | |
12442 | if (dir->is_auth()) | |
12443 | auth = "auth "; | |
12444 | else | |
12445 | auth = " rep "; | |
12446 | ||
12447 | char s[10]; | |
12448 | if (dir->get_dir_auth().second == CDIR_AUTH_UNKNOWN) | |
12449 | snprintf(s, sizeof(s), "%2d ", int(dir->get_dir_auth().first)); | |
12450 | else | |
12451 | snprintf(s, sizeof(s), "%2d,%2d", int(dir->get_dir_auth().first), int(dir->get_dir_auth().second)); | |
12452 | ||
12453 | ||
11fdf7f2 TL |
12454 | dout(ceph::dout::need_dynamic(dbl)) << indent << "|_" << pad << s |
12455 | << " " << auth << *dir << dendl; | |
7c673cae | 12456 | |
b3b6e05e | 12457 | if (dir->ino() == CEPH_INO_ROOT) |
11fdf7f2 | 12458 | ceph_assert(dir->inode == root); |
7c673cae | 12459 | if (dir->ino() == MDS_INO_MDSDIR(mds->get_nodeid())) |
11fdf7f2 | 12460 | ceph_assert(dir->inode == myin); |
7c673cae | 12461 | if (dir->inode->is_stray() && (MDS_INO_STRAY_OWNER(dir->ino()) == mds->get_nodeid())) |
11fdf7f2 | 12462 | ceph_assert(strays[MDS_INO_STRAY_INDEX(dir->ino())] == dir->inode); |
7c673cae FG |
12463 | |
12464 | // nested items? | |
12465 | if (!subtrees[dir].empty()) { | |
12466 | // more at my level? | |
12467 | if (!q.empty() && q.front().second == d) | |
12468 | indent += "| "; | |
12469 | else | |
12470 | indent += " "; | |
12471 | ||
12472 | for (set<CDir*>::iterator p = subtrees[dir].begin(); | |
12473 | p != subtrees[dir].end(); | |
12474 | ++p) | |
12475 | q.push_front(pair<CDir*,int>(*p, d+2)); | |
12476 | } | |
12477 | } | |
12478 | ||
12479 | // verify there isn't stray crap in subtree map | |
12480 | int lost = 0; | |
12481 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
12482 | p != subtrees.end(); | |
12483 | ++p) { | |
12484 | if (subtrees_seen.count(p->first)) continue; | |
12485 | dout(10) << "*** stray/lost entry in subtree map: " << *p->first << dendl; | |
12486 | lost++; | |
12487 | } | |
11fdf7f2 | 12488 | ceph_assert(lost == 0); |
7c673cae FG |
12489 | } |
12490 | ||
7c673cae FG |
12491 | void MDCache::show_cache() |
12492 | { | |
f67539c2 TL |
12493 | if (!g_conf()->subsys.should_gather<ceph_subsys_mds, 7>()) |
12494 | return; | |
7c673cae | 12495 | dout(7) << "show_cache" << dendl; |
b32b8144 FG |
12496 | |
12497 | auto show_func = [this](CInode *in) { | |
7c673cae | 12498 | // unlinked? |
b32b8144 FG |
12499 | if (!in->parent) |
12500 | dout(7) << " unlinked " << *in << dendl; | |
12501 | ||
7c673cae | 12502 | // dirfrags? |
9f95a23c TL |
12503 | auto&& dfs = in->get_dirfrags(); |
12504 | for (const auto& dir : dfs) { | |
7c673cae | 12505 | dout(7) << " dirfrag " << *dir << dendl; |
b32b8144 | 12506 | |
94b18763 FG |
12507 | for (auto &p : dir->items) { |
12508 | CDentry *dn = p.second; | |
7c673cae FG |
12509 | dout(7) << " dentry " << *dn << dendl; |
12510 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
12511 | if (dnl->is_primary() && dnl->get_inode()) | |
12512 | dout(7) << " inode " << *dnl->get_inode() << dendl; | |
12513 | } | |
12514 | } | |
b32b8144 FG |
12515 | }; |
12516 | ||
94b18763 | 12517 | for (auto &p : inode_map) |
b32b8144 | 12518 | show_func(p.second); |
94b18763 | 12519 | for (auto &p : snap_inode_map) |
b32b8144 | 12520 | show_func(p.second); |
7c673cae FG |
12521 | } |
12522 | ||
f64942e4 | 12523 | void MDCache::cache_status(Formatter *f) |
181888fb FG |
12524 | { |
12525 | f->open_object_section("cache"); | |
12526 | ||
12527 | f->open_object_section("pool"); | |
12528 | mempool::get_pool(mempool::mds_co::id).dump(f); | |
12529 | f->close_section(); | |
12530 | ||
12531 | f->close_section(); | |
181888fb FG |
12532 | } |
12533 | ||
11fdf7f2 | 12534 | void MDCache::dump_tree(CInode *in, const int cur_depth, const int max_depth, Formatter *f) |
7c673cae | 12535 | { |
11fdf7f2 TL |
12536 | ceph_assert(in); |
12537 | if ((max_depth >= 0) && (cur_depth > max_depth)) { | |
12538 | return; | |
12539 | } | |
9f95a23c | 12540 | auto&& ls = in->get_dirfrags(); |
11fdf7f2 TL |
12541 | for (const auto &subdir : ls) { |
12542 | for (const auto &p : subdir->items) { | |
12543 | CDentry *dn = p.second; | |
12544 | CInode *in = dn->get_linkage()->get_inode(); | |
12545 | if (in) { | |
12546 | dump_tree(in, cur_depth + 1, max_depth, f); | |
12547 | } | |
12548 | } | |
12549 | } | |
12550 | f->open_object_section("inode"); | |
12551 | in->dump(f, CInode::DUMP_DEFAULT | CInode::DUMP_DIRFRAGS); | |
12552 | f->close_section(); | |
7c673cae FG |
12553 | } |
12554 | ||
20effc67 | 12555 | int MDCache::dump_cache(std::string_view file_name, double timeout) |
7c673cae | 12556 | { |
20effc67 | 12557 | return dump_cache(file_name, NULL, timeout); |
7c673cae FG |
12558 | } |
12559 | ||
20effc67 | 12560 | int MDCache::dump_cache(Formatter *f, double timeout) |
7c673cae | 12561 | { |
20effc67 | 12562 | return dump_cache(std::string_view(""), f, timeout); |
7c673cae FG |
12563 | } |
12564 | ||
12565 | /** | |
12566 | * Dump the metadata cache, either to a Formatter, if | |
12567 | * provided, else to a plain text file. | |
12568 | */ | |
20effc67 | 12569 | int MDCache::dump_cache(std::string_view fn, Formatter *f, double timeout) |
7c673cae FG |
12570 | { |
12571 | int r = 0; | |
f64942e4 AA |
12572 | |
12573 | // dumping large caches may cause mds to hang or worse get killed. | |
12574 | // so, disallow the dump if the cache size exceeds the configured | |
12575 | // threshold, which is 1G for formatter and unlimited for file (note | |
12576 | // that this can be jacked up by the admin... and is nothing but foot | |
12577 | // shooting, but the option itself is for devs and hence dangerous to | |
12578 | // tune). TODO: remove this when fixed. | |
12579 | uint64_t threshold = f ? | |
11fdf7f2 TL |
12580 | g_conf().get_val<Option::size_t>("mds_dump_cache_threshold_formatter") : |
12581 | g_conf().get_val<Option::size_t>("mds_dump_cache_threshold_file"); | |
f64942e4 AA |
12582 | |
12583 | if (threshold && cache_size() > threshold) { | |
12584 | if (f) { | |
f67539c2 TL |
12585 | CachedStackStringStream css; |
12586 | *css << "cache usage exceeds dump threshold"; | |
f64942e4 | 12587 | f->open_object_section("result"); |
f67539c2 | 12588 | f->dump_string("error", css->strv()); |
f64942e4 AA |
12589 | f->close_section(); |
12590 | } else { | |
12591 | derr << "cache usage exceeds dump threshold" << dendl; | |
f67539c2 | 12592 | r = -CEPHFS_EINVAL; |
f64942e4 AA |
12593 | } |
12594 | return r; | |
12595 | } | |
12596 | ||
12597 | r = 0; | |
7c673cae FG |
12598 | int fd = -1; |
12599 | ||
12600 | if (f) { | |
12601 | f->open_array_section("inodes"); | |
12602 | } else { | |
94b18763 FG |
12603 | char path[PATH_MAX] = ""; |
12604 | if (fn.length()) { | |
12605 | snprintf(path, sizeof path, "%s", fn.data()); | |
12606 | } else { | |
12607 | snprintf(path, sizeof path, "cachedump.%d.mds%d", (int)mds->mdsmap->get_epoch(), int(mds->get_nodeid())); | |
7c673cae FG |
12608 | } |
12609 | ||
94b18763 | 12610 | dout(1) << "dump_cache to " << path << dendl; |
7c673cae | 12611 | |
91327a77 | 12612 | fd = ::open(path, O_WRONLY|O_CREAT|O_EXCL|O_CLOEXEC, 0600); |
7c673cae | 12613 | if (fd < 0) { |
94b18763 | 12614 | derr << "failed to open " << path << ": " << cpp_strerror(errno) << dendl; |
31f18b77 | 12615 | return errno; |
7c673cae FG |
12616 | } |
12617 | } | |
12618 | ||
11fdf7f2 | 12619 | auto dump_func = [fd, f](CInode *in) { |
b32b8144 | 12620 | int r; |
7c673cae FG |
12621 | if (f) { |
12622 | f->open_object_section("inode"); | |
11fdf7f2 TL |
12623 | in->dump(f, CInode::DUMP_DEFAULT | CInode::DUMP_DIRFRAGS); |
12624 | f->close_section(); | |
12625 | return 1; | |
12626 | } | |
f67539c2 TL |
12627 | CachedStackStringStream css; |
12628 | *css << *in << std::endl; | |
12629 | auto sv = css->strv(); | |
12630 | r = safe_write(fd, sv.data(), sv.size()); | |
11fdf7f2 TL |
12631 | if (r < 0) |
12632 | return r; | |
9f95a23c | 12633 | auto&& dfs = in->get_dirfrags(); |
11fdf7f2 | 12634 | for (auto &dir : dfs) { |
f67539c2 TL |
12635 | CachedStackStringStream css2; |
12636 | *css2 << " " << *dir << std::endl; | |
12637 | auto sv = css2->strv(); | |
12638 | r = safe_write(fd, sv.data(), sv.size()); | |
11fdf7f2 TL |
12639 | if (r < 0) |
12640 | return r; | |
94b18763 FG |
12641 | for (auto &p : dir->items) { |
12642 | CDentry *dn = p.second; | |
f67539c2 TL |
12643 | CachedStackStringStream css3; |
12644 | *css3 << " " << *dn << std::endl; | |
12645 | auto sv = css3->strv(); | |
12646 | r = safe_write(fd, sv.data(), sv.size()); | |
11fdf7f2 TL |
12647 | if (r < 0) |
12648 | return r; | |
7c673cae FG |
12649 | } |
12650 | dir->check_rstats(); | |
7c673cae | 12651 | } |
b32b8144 FG |
12652 | return 1; |
12653 | }; | |
12654 | ||
20effc67 TL |
12655 | auto start = mono_clock::now(); |
12656 | int64_t count = 0; | |
94b18763 | 12657 | for (auto &p : inode_map) { |
b32b8144 FG |
12658 | r = dump_func(p.second); |
12659 | if (r < 0) | |
12660 | goto out; | |
20effc67 TL |
12661 | if (!(++count % 1000) && |
12662 | timeout > 0 && | |
12663 | std::chrono::duration<double>(mono_clock::now() - start).count() > timeout) { | |
12664 | r = -ETIMEDOUT; | |
12665 | goto out; | |
12666 | } | |
b32b8144 | 12667 | } |
94b18763 | 12668 | for (auto &p : snap_inode_map) { |
b32b8144 FG |
12669 | r = dump_func(p.second); |
12670 | if (r < 0) | |
12671 | goto out; | |
20effc67 TL |
12672 | if (!(++count % 1000) && |
12673 | timeout > 0 && | |
12674 | std::chrono::duration<double>(mono_clock::now() - start).count() > timeout) { | |
12675 | r = -ETIMEDOUT; | |
12676 | goto out; | |
12677 | } | |
12678 | ||
7c673cae | 12679 | } |
b32b8144 | 12680 | r = 0; |
7c673cae FG |
12681 | |
12682 | out: | |
12683 | if (f) { | |
20effc67 TL |
12684 | if (r == -ETIMEDOUT) |
12685 | { | |
12686 | f->close_section(); | |
12687 | f->open_object_section("result"); | |
12688 | f->dump_string("error", "the operation timeout"); | |
12689 | } | |
7c673cae FG |
12690 | f->close_section(); // inodes |
12691 | } else { | |
20effc67 TL |
12692 | if (r == -ETIMEDOUT) |
12693 | { | |
12694 | CachedStackStringStream css; | |
12695 | *css << "error : the operation timeout" << std::endl; | |
12696 | auto sv = css->strv(); | |
12697 | r = safe_write(fd, sv.data(), sv.size()); | |
12698 | } | |
7c673cae FG |
12699 | ::close(fd); |
12700 | } | |
31f18b77 | 12701 | return r; |
7c673cae FG |
12702 | } |
12703 | ||
7c673cae FG |
12704 | void C_MDS_RetryRequest::finish(int r) |
12705 | { | |
12706 | mdr->retry++; | |
12707 | cache->dispatch_request(mdr); | |
12708 | } | |
12709 | ||
f67539c2 TL |
12710 | MDSContext *CF_MDS_RetryRequestFactory::build() |
12711 | { | |
12712 | if (drop_locks) { | |
12713 | mdcache->mds->locker->drop_locks(mdr.get(), nullptr); | |
12714 | mdr->drop_local_auth_pins(); | |
12715 | } | |
12716 | return new C_MDS_RetryRequest(mdcache, mdr); | |
12717 | } | |
7c673cae FG |
12718 | |
12719 | class C_MDS_EnqueueScrub : public Context | |
12720 | { | |
11fdf7f2 | 12721 | std::string tag; |
7c673cae FG |
12722 | Formatter *formatter; |
12723 | Context *on_finish; | |
12724 | public: | |
12725 | ScrubHeaderRef header; | |
11fdf7f2 TL |
12726 | C_MDS_EnqueueScrub(std::string_view tag, Formatter *f, Context *fin) : |
12727 | tag(tag), formatter(f), on_finish(fin), header(nullptr) {} | |
7c673cae | 12728 | |
7c673cae | 12729 | void finish(int r) override { |
f67539c2 TL |
12730 | formatter->open_object_section("results"); |
12731 | formatter->dump_int("return_code", r); | |
11fdf7f2 | 12732 | if (r == 0) { |
f67539c2 TL |
12733 | formatter->dump_string("scrub_tag", tag); |
12734 | formatter->dump_string("mode", "asynchronous"); | |
7c673cae | 12735 | } |
f67539c2 TL |
12736 | formatter->close_section(); |
12737 | ||
12738 | r = 0; | |
7c673cae FG |
12739 | if (on_finish) |
12740 | on_finish->complete(r); | |
12741 | } | |
12742 | }; | |
12743 | ||
12744 | void MDCache::enqueue_scrub( | |
11fdf7f2 TL |
12745 | std::string_view path, |
12746 | std::string_view tag, | |
7c673cae FG |
12747 | bool force, bool recursive, bool repair, |
12748 | Formatter *f, Context *fin) | |
12749 | { | |
11fdf7f2 | 12750 | dout(10) << __func__ << " " << path << dendl; |
f67539c2 TL |
12751 | |
12752 | filepath fp; | |
12753 | if (path.compare(0, 4, "~mds") == 0) { | |
12754 | mds_rank_t rank; | |
12755 | if (path == "~mdsdir") { | |
12756 | rank = mds->get_nodeid(); | |
12757 | } else { | |
12758 | std::string err; | |
12759 | rank = strict_strtoll(path.substr(4), 10, &err); | |
12760 | if (!err.empty()) | |
12761 | rank = MDS_RANK_NONE; | |
12762 | } | |
12763 | if (rank >= 0 && rank < MAX_MDS) | |
12764 | fp.set_path("", MDS_INO_MDSDIR(rank)); | |
11fdf7f2 | 12765 | } |
f67539c2 TL |
12766 | if (fp.get_ino() == inodeno_t(0)) |
12767 | fp.set_path(path); | |
12768 | ||
12769 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_ENQUEUE_SCRUB); | |
12770 | mdr->set_filepath(fp); | |
11fdf7f2 TL |
12771 | |
12772 | bool is_internal = false; | |
12773 | std::string tag_str(tag); | |
12774 | if (tag_str.empty()) { | |
12775 | uuid_d uuid_gen; | |
12776 | uuid_gen.generate_random(); | |
12777 | tag_str = uuid_gen.to_string(); | |
12778 | is_internal = true; | |
12779 | } | |
7c673cae | 12780 | |
11fdf7f2 | 12781 | C_MDS_EnqueueScrub *cs = new C_MDS_EnqueueScrub(tag_str, f, fin); |
f67539c2 | 12782 | cs->header = std::make_shared<ScrubHeader>(tag_str, is_internal, force, recursive, repair); |
7c673cae FG |
12783 | |
12784 | mdr->internal_op_finish = cs; | |
12785 | enqueue_scrub_work(mdr); | |
12786 | } | |
12787 | ||
12788 | void MDCache::enqueue_scrub_work(MDRequestRef& mdr) | |
12789 | { | |
f67539c2 TL |
12790 | CInode *in; |
12791 | CF_MDS_RetryRequestFactory cf(this, mdr, true); | |
12792 | int r = path_traverse(mdr, cf, mdr->get_filepath(), | |
12793 | MDS_TRAVERSE_DISCOVER | MDS_TRAVERSE_RDLOCK_PATH, | |
12794 | nullptr, &in); | |
12795 | if (r > 0) | |
7c673cae | 12796 | return; |
f67539c2 TL |
12797 | if (r < 0) { |
12798 | mds->server->respond_to_request(mdr, r); | |
12799 | return; | |
12800 | } | |
7c673cae FG |
12801 | |
12802 | // Cannot scrub same dentry twice at same time | |
11fdf7f2 | 12803 | if (in->scrub_is_in_progress()) { |
f67539c2 | 12804 | mds->server->respond_to_request(mdr, -CEPHFS_EBUSY); |
7c673cae FG |
12805 | return; |
12806 | } else { | |
12807 | in->scrub_info(); | |
12808 | } | |
12809 | ||
f67539c2 TL |
12810 | C_MDS_EnqueueScrub *cs = static_cast<C_MDS_EnqueueScrub*>(mdr->internal_op_finish); |
12811 | ScrubHeaderRef& header = cs->header; | |
7c673cae | 12812 | |
f67539c2 | 12813 | r = mds->scrubstack->enqueue(in, header, !header->get_recursive()); |
7c673cae | 12814 | |
f67539c2 | 12815 | mds->server->respond_to_request(mdr, r); |
11fdf7f2 TL |
12816 | } |
12817 | ||
12818 | struct C_MDC_RespondInternalRequest : public MDCacheLogContext { | |
7c673cae | 12819 | MDRequestRef mdr; |
11fdf7f2 | 12820 | C_MDC_RespondInternalRequest(MDCache *c, MDRequestRef& m) : |
7c673cae FG |
12821 | MDCacheLogContext(c), mdr(m) {} |
12822 | void finish(int r) override { | |
12823 | mdr->apply(); | |
12824 | get_mds()->server->respond_to_request(mdr, r); | |
12825 | } | |
12826 | }; | |
12827 | ||
f67539c2 TL |
12828 | struct C_MDC_ScrubRepaired : public MDCacheContext { |
12829 | ScrubHeaderRef header; | |
12830 | public: | |
12831 | C_MDC_ScrubRepaired(MDCache *m, const ScrubHeaderRef& h) | |
12832 | : MDCacheContext(m), header(h) { | |
12833 | header->inc_num_pending(); | |
12834 | } | |
12835 | void finish(int r) override { | |
12836 | header->dec_num_pending(); | |
12837 | } | |
12838 | }; | |
12839 | ||
7c673cae FG |
12840 | void MDCache::repair_dirfrag_stats(CDir *dir) |
12841 | { | |
12842 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_REPAIR_FRAGSTATS); | |
12843 | mdr->pin(dir); | |
12844 | mdr->internal_op_private = dir; | |
f67539c2 TL |
12845 | if (dir->scrub_is_in_progress()) |
12846 | mdr->internal_op_finish = new C_MDC_ScrubRepaired(this, dir->get_scrub_header()); | |
12847 | else | |
12848 | mdr->internal_op_finish = new C_MDSInternalNoop; | |
7c673cae FG |
12849 | repair_dirfrag_stats_work(mdr); |
12850 | } | |
12851 | ||
12852 | void MDCache::repair_dirfrag_stats_work(MDRequestRef& mdr) | |
12853 | { | |
12854 | CDir *dir = static_cast<CDir*>(mdr->internal_op_private); | |
12855 | dout(10) << __func__ << " " << *dir << dendl; | |
12856 | ||
12857 | if (!dir->is_auth()) { | |
f67539c2 | 12858 | mds->server->respond_to_request(mdr, -CEPHFS_ESTALE); |
7c673cae FG |
12859 | return; |
12860 | } | |
12861 | ||
12862 | if (!mdr->is_auth_pinned(dir) && !dir->can_auth_pin()) { | |
224ce89b WB |
12863 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryRequest(this, mdr)); |
12864 | ||
7c673cae FG |
12865 | mds->locker->drop_locks(mdr.get()); |
12866 | mdr->drop_local_auth_pins(); | |
9f95a23c | 12867 | if (mdr->is_any_remote_auth_pin()) |
224ce89b | 12868 | mds->locker->notify_freeze_waiter(dir); |
7c673cae FG |
12869 | return; |
12870 | } | |
12871 | ||
12872 | mdr->auth_pin(dir); | |
12873 | ||
11fdf7f2 | 12874 | MutationImpl::LockOpVec lov; |
7c673cae | 12875 | CInode *diri = dir->inode; |
11fdf7f2 TL |
12876 | lov.add_rdlock(&diri->dirfragtreelock); |
12877 | lov.add_wrlock(&diri->nestlock); | |
12878 | lov.add_wrlock(&diri->filelock); | |
12879 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
12880 | return; |
12881 | ||
12882 | if (!dir->is_complete()) { | |
12883 | dir->fetch(new C_MDS_RetryRequest(this, mdr)); | |
12884 | return; | |
12885 | } | |
12886 | ||
12887 | frag_info_t frag_info; | |
12888 | nest_info_t nest_info; | |
94b18763 | 12889 | for (auto it = dir->begin(); it != dir->end(); ++it) { |
7c673cae FG |
12890 | CDentry *dn = it->second; |
12891 | if (dn->last != CEPH_NOSNAP) | |
12892 | continue; | |
12893 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
12894 | if (dnl->is_primary()) { | |
12895 | CInode *in = dnl->get_inode(); | |
12896 | nest_info.add(in->get_projected_inode()->accounted_rstat); | |
12897 | if (in->is_dir()) | |
12898 | frag_info.nsubdirs++; | |
12899 | else | |
12900 | frag_info.nfiles++; | |
12901 | } else if (dnl->is_remote()) | |
12902 | frag_info.nfiles++; | |
12903 | } | |
12904 | ||
f67539c2 | 12905 | auto pf = dir->get_projected_fnode(); |
7c673cae FG |
12906 | bool good_fragstat = frag_info.same_sums(pf->fragstat); |
12907 | bool good_rstat = nest_info.same_sums(pf->rstat); | |
12908 | if (good_fragstat && good_rstat) { | |
12909 | dout(10) << __func__ << " no corruption found" << dendl; | |
12910 | mds->server->respond_to_request(mdr, 0); | |
12911 | return; | |
12912 | } | |
12913 | ||
f67539c2 TL |
12914 | auto _pf = dir->project_fnode(mdr); |
12915 | _pf->version = dir->pre_dirty(); | |
12916 | pf = _pf; | |
7c673cae FG |
12917 | |
12918 | mdr->ls = mds->mdlog->get_current_segment(); | |
12919 | EUpdate *le = new EUpdate(mds->mdlog, "repair_dirfrag"); | |
12920 | mds->mdlog->start_entry(le); | |
12921 | ||
12922 | if (!good_fragstat) { | |
12923 | if (pf->fragstat.mtime > frag_info.mtime) | |
12924 | frag_info.mtime = pf->fragstat.mtime; | |
12925 | if (pf->fragstat.change_attr > frag_info.change_attr) | |
12926 | frag_info.change_attr = pf->fragstat.change_attr; | |
f67539c2 | 12927 | _pf->fragstat = frag_info; |
7c673cae FG |
12928 | mds->locker->mark_updated_scatterlock(&diri->filelock); |
12929 | mdr->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
12930 | mdr->add_updated_lock(&diri->filelock); | |
12931 | } | |
12932 | ||
12933 | if (!good_rstat) { | |
12934 | if (pf->rstat.rctime > nest_info.rctime) | |
12935 | nest_info.rctime = pf->rstat.rctime; | |
f67539c2 | 12936 | _pf->rstat = nest_info; |
7c673cae FG |
12937 | mds->locker->mark_updated_scatterlock(&diri->nestlock); |
12938 | mdr->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
12939 | mdr->add_updated_lock(&diri->nestlock); | |
12940 | } | |
12941 | ||
12942 | le->metablob.add_dir_context(dir); | |
12943 | le->metablob.add_dir(dir, true); | |
12944 | ||
11fdf7f2 | 12945 | mds->mdlog->submit_entry(le, new C_MDC_RespondInternalRequest(this, mdr)); |
7c673cae FG |
12946 | } |
12947 | ||
12948 | void MDCache::repair_inode_stats(CInode *diri) | |
12949 | { | |
12950 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_REPAIR_INODESTATS); | |
f67539c2 | 12951 | mdr->auth_pin(diri); // already auth pinned by CInode::validate_disk_state() |
7c673cae | 12952 | mdr->internal_op_private = diri; |
f67539c2 TL |
12953 | if (diri->scrub_is_in_progress()) |
12954 | mdr->internal_op_finish = new C_MDC_ScrubRepaired(this, diri->get_scrub_header()); | |
12955 | else | |
12956 | mdr->internal_op_finish = new C_MDSInternalNoop; | |
7c673cae FG |
12957 | repair_inode_stats_work(mdr); |
12958 | } | |
12959 | ||
12960 | void MDCache::repair_inode_stats_work(MDRequestRef& mdr) | |
12961 | { | |
12962 | CInode *diri = static_cast<CInode*>(mdr->internal_op_private); | |
12963 | dout(10) << __func__ << " " << *diri << dendl; | |
12964 | ||
12965 | if (!diri->is_auth()) { | |
f67539c2 | 12966 | mds->server->respond_to_request(mdr, -CEPHFS_ESTALE); |
7c673cae FG |
12967 | return; |
12968 | } | |
12969 | if (!diri->is_dir()) { | |
f67539c2 | 12970 | mds->server->respond_to_request(mdr, -CEPHFS_ENOTDIR); |
7c673cae FG |
12971 | return; |
12972 | } | |
12973 | ||
11fdf7f2 | 12974 | MutationImpl::LockOpVec lov; |
7c673cae FG |
12975 | |
12976 | if (mdr->ls) // already marked filelock/nestlock dirty ? | |
12977 | goto do_rdlocks; | |
12978 | ||
11fdf7f2 TL |
12979 | lov.add_rdlock(&diri->dirfragtreelock); |
12980 | lov.add_wrlock(&diri->nestlock); | |
12981 | lov.add_wrlock(&diri->filelock); | |
12982 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
12983 | return; |
12984 | ||
12985 | // Fetch all dirfrags and mark filelock/nestlock dirty. This will tirgger | |
12986 | // the scatter-gather process, which will fix any fragstat/rstat errors. | |
11fdf7f2 TL |
12987 | { |
12988 | frag_vec_t leaves; | |
12989 | diri->dirfragtree.get_leaves(leaves); | |
12990 | for (const auto& leaf : leaves) { | |
12991 | CDir *dir = diri->get_dirfrag(leaf); | |
12992 | if (!dir) { | |
12993 | ceph_assert(mdr->is_auth_pinned(diri)); | |
12994 | dir = diri->get_or_open_dirfrag(this, leaf); | |
12995 | } | |
12996 | if (dir->get_version() == 0) { | |
12997 | ceph_assert(dir->is_auth()); | |
12998 | dir->fetch(new C_MDS_RetryRequest(this, mdr)); | |
12999 | return; | |
13000 | } | |
7c673cae FG |
13001 | } |
13002 | } | |
13003 | ||
13004 | diri->state_set(CInode::STATE_REPAIRSTATS); | |
13005 | mdr->ls = mds->mdlog->get_current_segment(); | |
13006 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
13007 | mdr->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
13008 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
13009 | mdr->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
13010 | ||
13011 | mds->locker->drop_locks(mdr.get()); | |
13012 | ||
13013 | do_rdlocks: | |
13014 | // force the scatter-gather process | |
11fdf7f2 TL |
13015 | lov.clear(); |
13016 | lov.add_rdlock(&diri->dirfragtreelock); | |
13017 | lov.add_rdlock(&diri->nestlock); | |
13018 | lov.add_rdlock(&diri->filelock); | |
13019 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
13020 | return; |
13021 | ||
13022 | diri->state_clear(CInode::STATE_REPAIRSTATS); | |
13023 | ||
13024 | frag_info_t dir_info; | |
13025 | nest_info_t nest_info; | |
11fdf7f2 TL |
13026 | nest_info.rsubdirs = 1; // it gets one to account for self |
13027 | if (const sr_t *srnode = diri->get_projected_srnode(); srnode) | |
13028 | nest_info.rsnaps = srnode->snaps.size(); | |
7c673cae | 13029 | |
11fdf7f2 TL |
13030 | { |
13031 | frag_vec_t leaves; | |
13032 | diri->dirfragtree.get_leaves(leaves); | |
13033 | for (const auto& leaf : leaves) { | |
13034 | CDir *dir = diri->get_dirfrag(leaf); | |
13035 | ceph_assert(dir); | |
13036 | ceph_assert(dir->get_version() > 0); | |
f67539c2 TL |
13037 | dir_info.add(dir->get_fnode()->accounted_fragstat); |
13038 | nest_info.add(dir->get_fnode()->accounted_rstat); | |
11fdf7f2 | 13039 | } |
7c673cae FG |
13040 | } |
13041 | ||
f67539c2 TL |
13042 | if (!dir_info.same_sums(diri->get_inode()->dirstat) || |
13043 | !nest_info.same_sums(diri->get_inode()->rstat)) { | |
7c673cae FG |
13044 | dout(10) << __func__ << " failed to fix fragstat/rstat on " |
13045 | << *diri << dendl; | |
13046 | } | |
13047 | ||
13048 | mds->server->respond_to_request(mdr, 0); | |
13049 | } | |
13050 | ||
f67539c2 | 13051 | void MDCache::rdlock_dirfrags_stats(CInode *diri, MDSInternalContext* fin) |
11fdf7f2 | 13052 | { |
f67539c2 TL |
13053 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_RDLOCK_FRAGSSTATS); |
13054 | mdr->auth_pin(diri); // already auth pinned by CInode::validate_disk_state() | |
13055 | mdr->internal_op_private = diri; | |
13056 | mdr->internal_op_finish = fin; | |
13057 | return rdlock_dirfrags_stats_work(mdr); | |
11fdf7f2 TL |
13058 | } |
13059 | ||
f67539c2 | 13060 | void MDCache::rdlock_dirfrags_stats_work(MDRequestRef& mdr) |
11fdf7f2 | 13061 | { |
f67539c2 TL |
13062 | CInode *diri = static_cast<CInode*>(mdr->internal_op_private); |
13063 | dout(10) << __func__ << " " << *diri << dendl; | |
13064 | if (!diri->is_auth()) { | |
13065 | mds->server->respond_to_request(mdr, -CEPHFS_ESTALE); | |
13066 | return; | |
13067 | } | |
13068 | if (!diri->is_dir()) { | |
13069 | mds->server->respond_to_request(mdr, -CEPHFS_ENOTDIR); | |
11fdf7f2 TL |
13070 | return; |
13071 | } | |
13072 | ||
13073 | MutationImpl::LockOpVec lov; | |
f67539c2 TL |
13074 | lov.add_rdlock(&diri->dirfragtreelock); |
13075 | lov.add_rdlock(&diri->nestlock); | |
13076 | lov.add_rdlock(&diri->filelock); | |
11fdf7f2 TL |
13077 | if (!mds->locker->acquire_locks(mdr, lov)) |
13078 | return; | |
f67539c2 | 13079 | dout(10) << __func__ << " start dirfrags : " << *diri << dendl; |
11fdf7f2 | 13080 | |
f67539c2 TL |
13081 | mds->server->respond_to_request(mdr, 0); |
13082 | return; | |
11fdf7f2 TL |
13083 | } |
13084 | ||
13085 | void MDCache::flush_dentry(std::string_view path, Context *fin) | |
7c673cae FG |
13086 | { |
13087 | if (is_readonly()) { | |
13088 | dout(10) << __func__ << ": read-only FS" << dendl; | |
f67539c2 | 13089 | fin->complete(-CEPHFS_EROFS); |
7c673cae FG |
13090 | return; |
13091 | } | |
13092 | dout(10) << "flush_dentry " << path << dendl; | |
13093 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FLUSH); | |
94b18763 | 13094 | filepath fp(path); |
7c673cae FG |
13095 | mdr->set_filepath(fp); |
13096 | mdr->internal_op_finish = fin; | |
13097 | flush_dentry_work(mdr); | |
13098 | } | |
13099 | ||
11fdf7f2 | 13100 | class C_FinishIOMDR : public MDSContext { |
7c673cae FG |
13101 | protected: |
13102 | MDSRank *mds; | |
13103 | MDRequestRef mdr; | |
13104 | MDSRank *get_mds() override { return mds; } | |
13105 | public: | |
13106 | C_FinishIOMDR(MDSRank *mds_, MDRequestRef& mdr_) : mds(mds_), mdr(mdr_) {} | |
13107 | void finish(int r) override { mds->server->respond_to_request(mdr, r); } | |
13108 | }; | |
13109 | ||
13110 | void MDCache::flush_dentry_work(MDRequestRef& mdr) | |
13111 | { | |
11fdf7f2 | 13112 | MutationImpl::LockOpVec lov; |
9f95a23c TL |
13113 | CInode *in = mds->server->rdlock_path_pin_ref(mdr, true); |
13114 | if (!in) | |
7c673cae FG |
13115 | return; |
13116 | ||
11fdf7f2 | 13117 | ceph_assert(in->is_auth()); |
7c673cae FG |
13118 | in->flush(new C_FinishIOMDR(mds, mdr)); |
13119 | } | |
13120 | ||
13121 | ||
13122 | /** | |
13123 | * Initialize performance counters with global perfcounter | |
13124 | * collection. | |
13125 | */ | |
13126 | void MDCache::register_perfcounters() | |
13127 | { | |
91327a77 AA |
13128 | PerfCountersBuilder pcb(g_ceph_context, "mds_cache", l_mdc_first, l_mdc_last); |
13129 | ||
13130 | // Stray/purge statistics | |
13131 | pcb.add_u64(l_mdc_num_strays, "num_strays", "Stray dentries", "stry", | |
13132 | PerfCountersBuilder::PRIO_INTERESTING); | |
13133 | pcb.add_u64(l_mdc_num_recovering_enqueued, | |
13134 | "num_recovering_enqueued", "Files waiting for recovery", "recy", | |
13135 | PerfCountersBuilder::PRIO_INTERESTING); | |
13136 | pcb.add_u64_counter(l_mdc_recovery_completed, | |
13137 | "recovery_completed", "File recoveries completed", "recd", | |
13138 | PerfCountersBuilder::PRIO_INTERESTING); | |
13139 | ||
13140 | // useful recovery queue statistics | |
13141 | pcb.set_prio_default(PerfCountersBuilder::PRIO_USEFUL); | |
13142 | pcb.add_u64(l_mdc_num_recovering_processing, "num_recovering_processing", | |
13143 | "Files currently being recovered"); | |
13144 | pcb.add_u64(l_mdc_num_recovering_prioritized, "num_recovering_prioritized", | |
13145 | "Files waiting for recovery with elevated priority"); | |
13146 | pcb.add_u64_counter(l_mdc_recovery_started, "recovery_started", | |
13147 | "File recoveries started"); | |
13148 | ||
13149 | // along with other stray dentries stats | |
13150 | pcb.add_u64(l_mdc_num_strays_delayed, "num_strays_delayed", | |
13151 | "Stray dentries delayed"); | |
13152 | pcb.add_u64(l_mdc_num_strays_enqueuing, "num_strays_enqueuing", | |
13153 | "Stray dentries enqueuing for purge"); | |
13154 | pcb.add_u64_counter(l_mdc_strays_created, "strays_created", | |
13155 | "Stray dentries created"); | |
7c673cae | 13156 | pcb.add_u64_counter(l_mdc_strays_enqueued, "strays_enqueued", |
91327a77 AA |
13157 | "Stray dentries enqueued for purge"); |
13158 | pcb.add_u64_counter(l_mdc_strays_reintegrated, "strays_reintegrated", | |
13159 | "Stray dentries reintegrated"); | |
13160 | pcb.add_u64_counter(l_mdc_strays_migrated, "strays_migrated", | |
13161 | "Stray dentries migrated"); | |
7c673cae | 13162 | |
91327a77 | 13163 | // low prio internal request stats |
d2e6a577 | 13164 | pcb.add_u64_counter(l_mdss_ireq_enqueue_scrub, "ireq_enqueue_scrub", |
91327a77 | 13165 | "Internal Request type enqueue scrub"); |
d2e6a577 | 13166 | pcb.add_u64_counter(l_mdss_ireq_exportdir, "ireq_exportdir", |
91327a77 | 13167 | "Internal Request type export dir"); |
d2e6a577 | 13168 | pcb.add_u64_counter(l_mdss_ireq_flush, "ireq_flush", |
91327a77 | 13169 | "Internal Request type flush"); |
d2e6a577 | 13170 | pcb.add_u64_counter(l_mdss_ireq_fragmentdir, "ireq_fragmentdir", |
91327a77 | 13171 | "Internal Request type fragmentdir"); |
d2e6a577 | 13172 | pcb.add_u64_counter(l_mdss_ireq_fragstats, "ireq_fragstats", |
91327a77 | 13173 | "Internal Request type frag stats"); |
d2e6a577 | 13174 | pcb.add_u64_counter(l_mdss_ireq_inodestats, "ireq_inodestats", |
91327a77 | 13175 | "Internal Request type inode stats"); |
d2e6a577 | 13176 | |
7c673cae FG |
13177 | logger.reset(pcb.create_perf_counters()); |
13178 | g_ceph_context->get_perfcounters_collection()->add(logger.get()); | |
13179 | recovery_queue.set_logger(logger.get()); | |
13180 | stray_manager.set_logger(logger.get()); | |
13181 | } | |
13182 | ||
7c673cae FG |
13183 | /** |
13184 | * Call this when putting references to an inode/dentry or | |
13185 | * when attempting to trim it. | |
13186 | * | |
13187 | * If this inode is no longer linked by anyone, and this MDS | |
13188 | * rank holds the primary dentry, and that dentry is in a stray | |
13189 | * directory, then give up the dentry to the StrayManager, never | |
13190 | * to be seen again by MDCache. | |
13191 | * | |
13192 | * @param delay if true, then purgeable inodes are stashed til | |
13193 | * the next trim(), rather than being purged right | |
13194 | * away. | |
13195 | */ | |
13196 | void MDCache::maybe_eval_stray(CInode *in, bool delay) { | |
f67539c2 | 13197 | if (in->get_inode()->nlink > 0 || in->is_base() || is_readonly() || |
224ce89b | 13198 | mds->get_state() <= MDSMap::STATE_REJOIN) |
7c673cae | 13199 | return; |
224ce89b | 13200 | |
7c673cae FG |
13201 | CDentry *dn = in->get_projected_parent_dn(); |
13202 | ||
13203 | if (dn->state_test(CDentry::STATE_PURGING)) { | |
13204 | /* We have already entered the purging process, no need | |
13205 | * to re-evaluate me ! */ | |
13206 | return; | |
13207 | } | |
13208 | ||
11fdf7f2 TL |
13209 | if (dn->get_dir()->get_inode()->is_stray()) { |
13210 | if (delay) | |
13211 | stray_manager.queue_delayed(dn); | |
13212 | else | |
13213 | stray_manager.eval_stray(dn); | |
7c673cae FG |
13214 | } |
13215 | } | |
13216 | ||
31f18b77 FG |
13217 | void MDCache::clear_dirty_bits_for_stray(CInode* diri) { |
13218 | dout(10) << __func__ << " " << *diri << dendl; | |
11fdf7f2 | 13219 | ceph_assert(diri->get_projected_parent_dir()->inode->is_stray()); |
9f95a23c | 13220 | auto&& ls = diri->get_dirfrags(); |
94b18763 | 13221 | for (auto &p : ls) { |
31f18b77 FG |
13222 | if (p->is_auth() && !(p->is_frozen() || p->is_freezing())) |
13223 | p->try_remove_dentries_for_stray(); | |
13224 | } | |
13225 | if (!diri->snaprealm) { | |
13226 | if (diri->is_auth()) | |
13227 | diri->clear_dirty_rstat(); | |
13228 | diri->clear_scatter_dirty(); | |
13229 | } | |
13230 | } | |
13231 | ||
11fdf7f2 TL |
13232 | bool MDCache::dump_inode(Formatter *f, uint64_t number) { |
13233 | CInode *in = get_inode(number); | |
13234 | if (!in) { | |
13235 | return false; | |
13236 | } | |
13237 | f->open_object_section("inode"); | |
13238 | in->dump(f, CInode::DUMP_DEFAULT | CInode::DUMP_PATH); | |
13239 | f->close_section(); | |
13240 | return true; | |
13241 | } | |
eafe8130 | 13242 | |
f6b5b4d7 | 13243 | void MDCache::handle_mdsmap(const MDSMap &mdsmap, const MDSMap &oldmap) { |
f67539c2 TL |
13244 | const mds_rank_t max_mds = mdsmap.get_max_mds(); |
13245 | ||
eafe8130 TL |
13246 | // process export_pin_delayed_queue whenever a new MDSMap received |
13247 | auto &q = export_pin_delayed_queue; | |
13248 | for (auto it = q.begin(); it != q.end(); ) { | |
13249 | auto *in = *it; | |
13250 | mds_rank_t export_pin = in->get_export_pin(false); | |
13251 | dout(10) << " delayed export_pin=" << export_pin << " on " << *in | |
f67539c2 | 13252 | << " max_mds=" << max_mds << dendl; |
eafe8130 TL |
13253 | if (export_pin >= mdsmap.get_max_mds()) { |
13254 | it++; | |
13255 | continue; | |
13256 | } | |
13257 | ||
13258 | in->state_clear(CInode::STATE_DELAYEDEXPORTPIN); | |
13259 | it = q.erase(it); | |
f6b5b4d7 | 13260 | in->queue_export_pin(export_pin); |
eafe8130 | 13261 | } |
eafe8130 | 13262 | |
f6b5b4d7 TL |
13263 | if (mdsmap.get_max_mds() != oldmap.get_max_mds()) { |
13264 | dout(10) << "Checking ephemerally pinned directories for redistribute due to max_mds change." << dendl; | |
13265 | /* copy to vector to avoid removals during iteration */ | |
13266 | std::vector<CInode*> migrate; | |
f67539c2 | 13267 | migrate.assign(export_ephemeral_pins.begin(), export_ephemeral_pins.end()); |
f6b5b4d7 | 13268 | for (auto& in : migrate) { |
f67539c2 | 13269 | in->maybe_export_pin(); |
f6b5b4d7 TL |
13270 | } |
13271 | } | |
f67539c2 TL |
13272 | |
13273 | if (max_mds <= 1) { | |
13274 | export_ephemeral_dist_frag_bits = 0; | |
13275 | } else { | |
13276 | double want = g_conf().get_val<double>("mds_export_ephemeral_distributed_factor"); | |
13277 | want *= max_mds; | |
13278 | unsigned n = 0; | |
13279 | while ((1U << n) < (unsigned)want) | |
13280 | ++n; | |
13281 | export_ephemeral_dist_frag_bits = n; | |
13282 | } | |
f6b5b4d7 | 13283 | } |
b3b6e05e TL |
13284 | |
13285 | void MDCache::upkeep_main(void) | |
13286 | { | |
13287 | std::unique_lock lock(upkeep_mutex); | |
13288 | while (!upkeep_trim_shutdown.load()) { | |
13289 | auto now = clock::now(); | |
13290 | auto since = now-upkeep_last_trim; | |
13291 | auto trim_interval = clock::duration(g_conf().get_val<std::chrono::seconds>("mds_cache_trim_interval")); | |
13292 | if (since >= trim_interval*.90) { | |
13293 | lock.unlock(); /* mds_lock -> upkeep_mutex */ | |
13294 | std::scoped_lock mds_lock(mds->mds_lock); | |
13295 | lock.lock(); | |
13296 | if (upkeep_trim_shutdown.load()) | |
13297 | return; | |
13298 | check_memory_usage(); | |
13299 | if (mds->is_cache_trimmable()) { | |
13300 | dout(20) << "upkeep thread trimming cache; last trim " << since << " ago" << dendl; | |
13301 | bool active_with_clients = mds->is_active() || mds->is_clientreplay() || mds->is_stopping(); | |
13302 | if (active_with_clients) { | |
13303 | trim_client_leases(); | |
13304 | } | |
a4b75251 TL |
13305 | if (is_open()) { |
13306 | trim(); | |
13307 | } | |
b3b6e05e TL |
13308 | if (active_with_clients) { |
13309 | auto recall_flags = Server::RecallFlags::ENFORCE_MAX|Server::RecallFlags::ENFORCE_LIVENESS; | |
13310 | if (cache_toofull()) { | |
13311 | recall_flags = recall_flags|Server::RecallFlags::TRIM; | |
13312 | } | |
13313 | mds->server->recall_client_state(nullptr, recall_flags); | |
13314 | } | |
13315 | upkeep_last_trim = now = clock::now(); | |
13316 | } else { | |
13317 | dout(10) << "cache not ready for trimming" << dendl; | |
13318 | } | |
13319 | } else { | |
13320 | trim_interval -= since; | |
13321 | } | |
13322 | since = now-upkeep_last_release; | |
13323 | auto release_interval = clock::duration(g_conf().get_val<std::chrono::seconds>("mds_cache_release_free_interval")); | |
13324 | if (since >= release_interval*.90) { | |
13325 | /* XXX not necessary once MDCache uses PriorityCache */ | |
13326 | dout(10) << "releasing free memory" << dendl; | |
13327 | ceph_heap_release_free_memory(); | |
13328 | upkeep_last_release = clock::now(); | |
13329 | } else { | |
13330 | release_interval -= since; | |
13331 | } | |
13332 | auto interval = std::min(release_interval, trim_interval); | |
13333 | dout(20) << "upkeep thread waiting interval " << interval << dendl; | |
13334 | upkeep_cvar.wait_for(lock, interval); | |
13335 | } | |
13336 | } |