]>
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> | |
16 | #include <fstream> | |
17 | #include <iostream> | |
18 | #include <sstream> | |
19 | #include <string> | |
11fdf7f2 | 20 | #include <string_view> |
7c673cae FG |
21 | #include <map> |
22 | ||
23 | #include "MDCache.h" | |
24 | #include "MDSRank.h" | |
25 | #include "Server.h" | |
26 | #include "Locker.h" | |
27 | #include "MDLog.h" | |
28 | #include "MDBalancer.h" | |
29 | #include "Migrator.h" | |
30 | #include "ScrubStack.h" | |
31 | ||
32 | #include "SnapClient.h" | |
33 | ||
34 | #include "MDSMap.h" | |
35 | ||
36 | #include "CInode.h" | |
37 | #include "CDir.h" | |
38 | ||
39 | #include "Mutation.h" | |
40 | ||
41 | #include "include/ceph_fs.h" | |
42 | #include "include/filepath.h" | |
181888fb | 43 | #include "include/util.h" |
7c673cae | 44 | |
11fdf7f2 TL |
45 | #include "messages/MClientCaps.h" |
46 | ||
7c673cae FG |
47 | #include "msg/Message.h" |
48 | #include "msg/Messenger.h" | |
49 | ||
181888fb | 50 | #include "common/MemoryModel.h" |
7c673cae | 51 | #include "common/errno.h" |
7c673cae | 52 | #include "common/perf_counters.h" |
181888fb FG |
53 | #include "common/safe_io.h" |
54 | ||
7c673cae FG |
55 | #include "osdc/Journaler.h" |
56 | #include "osdc/Filer.h" | |
57 | ||
58 | #include "events/ESubtreeMap.h" | |
59 | #include "events/EUpdate.h" | |
60 | #include "events/ESlaveUpdate.h" | |
61 | #include "events/EImportFinish.h" | |
62 | #include "events/EFragment.h" | |
63 | #include "events/ECommitted.h" | |
64 | #include "events/ESessions.h" | |
65 | ||
7c673cae FG |
66 | #include "InoTable.h" |
67 | ||
68 | #include "common/Timer.h" | |
69 | ||
70 | #include "perfglue/heap_profiler.h" | |
71 | ||
7c673cae FG |
72 | |
73 | #include "common/config.h" | |
11fdf7f2 | 74 | #include "include/ceph_assert.h" |
7c673cae FG |
75 | |
76 | #define dout_context g_ceph_context | |
77 | #define dout_subsys ceph_subsys_mds | |
78 | #undef dout_prefix | |
79 | #define dout_prefix _prefix(_dout, mds) | |
80 | static ostream& _prefix(std::ostream *_dout, MDSRank *mds) { | |
81 | return *_dout << "mds." << mds->get_nodeid() << ".cache "; | |
82 | } | |
83 | ||
84 | set<int> SimpleLock::empty_gather_set; | |
85 | ||
86 | ||
87 | /** | |
88 | * All non-I/O contexts that require a reference | |
89 | * to an MDCache instance descend from this. | |
90 | */ | |
11fdf7f2 | 91 | class MDCacheContext : public virtual MDSContext { |
7c673cae FG |
92 | protected: |
93 | MDCache *mdcache; | |
94 | MDSRank *get_mds() override | |
95 | { | |
11fdf7f2 | 96 | ceph_assert(mdcache != NULL); |
7c673cae FG |
97 | return mdcache->mds; |
98 | } | |
99 | public: | |
100 | explicit MDCacheContext(MDCache *mdc_) : mdcache(mdc_) {} | |
101 | }; | |
102 | ||
103 | ||
104 | /** | |
105 | * Only for contexts called back from an I/O completion | |
106 | * | |
107 | * Note: duplication of members wrt MDCacheContext, because | |
108 | * it'ls the lesser of two evils compared with introducing | |
109 | * yet another piece of (multiple) inheritance. | |
110 | */ | |
111 | class MDCacheIOContext : public virtual MDSIOContextBase { | |
112 | protected: | |
113 | MDCache *mdcache; | |
114 | MDSRank *get_mds() override | |
115 | { | |
11fdf7f2 | 116 | ceph_assert(mdcache != NULL); |
7c673cae FG |
117 | return mdcache->mds; |
118 | } | |
119 | public: | |
91327a77 AA |
120 | explicit MDCacheIOContext(MDCache *mdc_, bool track=true) : |
121 | MDSIOContextBase(track), mdcache(mdc_) {} | |
7c673cae FG |
122 | }; |
123 | ||
124 | class MDCacheLogContext : public virtual MDSLogContextBase { | |
125 | protected: | |
126 | MDCache *mdcache; | |
127 | MDSRank *get_mds() override | |
128 | { | |
11fdf7f2 | 129 | ceph_assert(mdcache != NULL); |
7c673cae FG |
130 | return mdcache->mds; |
131 | } | |
132 | public: | |
133 | explicit MDCacheLogContext(MDCache *mdc_) : mdcache(mdc_) {} | |
134 | }; | |
135 | ||
136 | MDCache::MDCache(MDSRank *m, PurgeQueue &purge_queue_) : | |
137 | mds(m), | |
138 | filer(m->objecter, m->finisher), | |
139 | exceeded_size_limit(false), | |
140 | recovery_queue(m), | |
a8e16298 | 141 | stray_manager(m, purge_queue_), |
11fdf7f2 TL |
142 | trim_counter(g_conf().get_val<double>("mds_cache_trim_decay_rate")), |
143 | open_file_table(m) | |
7c673cae FG |
144 | { |
145 | migrator.reset(new Migrator(mds, this)); | |
146 | root = NULL; | |
147 | myin = NULL; | |
148 | readonly = false; | |
149 | ||
150 | stray_index = 0; | |
151 | for (int i = 0; i < NUM_STRAY; ++i) { | |
152 | strays[i] = NULL; | |
153 | } | |
154 | ||
b32b8144 | 155 | num_shadow_inodes = 0; |
7c673cae FG |
156 | num_inodes_with_caps = 0; |
157 | ||
11fdf7f2 TL |
158 | max_dir_commit_size = g_conf()->mds_dir_max_commit_size ? |
159 | (g_conf()->mds_dir_max_commit_size << 20) : | |
160 | (0.9 *(g_conf()->osd_max_write_size << 20)); | |
7c673cae FG |
161 | |
162 | discover_last_tid = 0; | |
163 | open_ino_last_tid = 0; | |
164 | find_ino_peer_last_tid = 0; | |
165 | ||
166 | last_cap_id = 0; | |
167 | ||
168 | client_lease_durations[0] = 5.0; | |
169 | client_lease_durations[1] = 30.0; | |
170 | client_lease_durations[2] = 300.0; | |
171 | ||
172 | resolves_pending = false; | |
173 | rejoins_pending = false; | |
174 | cap_imports_num_opening = 0; | |
175 | ||
176 | opening_root = open = false; | |
91327a77 | 177 | |
11fdf7f2 TL |
178 | cache_inode_limit = g_conf().get_val<int64_t>("mds_cache_size"); |
179 | cache_memory_limit = g_conf().get_val<Option::size_t>("mds_cache_memory_limit"); | |
180 | cache_reservation = g_conf().get_val<double>("mds_cache_reservation"); | |
181 | cache_health_threshold = g_conf().get_val<double>("mds_health_cache_threshold"); | |
91327a77 | 182 | |
11fdf7f2 | 183 | lru.lru_set_midpoint(g_conf().get_val<double>("mds_cache_mid")); |
7c673cae | 184 | |
31f18b77 FG |
185 | bottom_lru.lru_set_midpoint(0); |
186 | ||
11fdf7f2 | 187 | decayrate.set_halflife(g_conf()->mds_decay_halflife); |
7c673cae FG |
188 | |
189 | did_shutdown_log_cap = false; | |
11fdf7f2 TL |
190 | |
191 | global_snaprealm = NULL; | |
7c673cae FG |
192 | } |
193 | ||
194 | MDCache::~MDCache() | |
195 | { | |
196 | if (logger) { | |
197 | g_ceph_context->get_perfcounters_collection()->remove(logger.get()); | |
198 | } | |
199 | } | |
200 | ||
11fdf7f2 | 201 | void MDCache::handle_conf_change(const ConfigProxy& conf, |
91327a77 AA |
202 | const std::set <std::string> &changed, |
203 | const MDSMap &mdsmap) | |
204 | { | |
205 | if (changed.count("mds_cache_size")) | |
11fdf7f2 | 206 | cache_inode_limit = g_conf().get_val<int64_t>("mds_cache_size"); |
91327a77 | 207 | if (changed.count("mds_cache_memory_limit")) |
11fdf7f2 | 208 | cache_memory_limit = g_conf().get_val<Option::size_t>("mds_cache_memory_limit"); |
91327a77 | 209 | if (changed.count("mds_cache_reservation")) |
11fdf7f2 | 210 | cache_reservation = g_conf().get_val<double>("mds_cache_reservation"); |
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 | } |
7c673cae | 218 | |
91327a77 AA |
219 | migrator->handle_conf_change(conf, changed, mdsmap); |
220 | mds->balancer->handle_conf_change(conf, changed, mdsmap); | |
221 | } | |
7c673cae FG |
222 | |
223 | void MDCache::log_stat() | |
224 | { | |
91327a77 | 225 | mds->logger->set(l_mds_inode_max, cache_inode_limit ? : INT_MAX); |
7c673cae FG |
226 | mds->logger->set(l_mds_inodes, lru.lru_get_size()); |
227 | mds->logger->set(l_mds_inodes_pinned, lru.lru_get_num_pinned()); | |
228 | mds->logger->set(l_mds_inodes_top, lru.lru_get_top()); | |
229 | mds->logger->set(l_mds_inodes_bottom, lru.lru_get_bot()); | |
230 | mds->logger->set(l_mds_inodes_pin_tail, lru.lru_get_pintail()); | |
231 | mds->logger->set(l_mds_inodes_with_caps, num_inodes_with_caps); | |
232 | mds->logger->set(l_mds_caps, Capability::count()); | |
233 | } | |
234 | ||
235 | ||
236 | // | |
237 | ||
238 | bool MDCache::shutdown() | |
239 | { | |
240 | if (lru.lru_get_size() > 0) { | |
241 | dout(7) << "WARNING: mdcache shutdown with non-empty cache" << dendl; | |
242 | //show_cache(); | |
243 | show_subtrees(); | |
244 | //dump(); | |
245 | } | |
246 | return true; | |
247 | } | |
248 | ||
249 | ||
250 | // ==================================================================== | |
251 | // some inode functions | |
252 | ||
253 | void MDCache::add_inode(CInode *in) | |
254 | { | |
255 | // add to lru, inode map | |
b32b8144 FG |
256 | if (in->last == CEPH_NOSNAP) { |
257 | auto &p = inode_map[in->ino()]; | |
11fdf7f2 | 258 | ceph_assert(!p); // should be no dup inos! |
b32b8144 FG |
259 | p = in; |
260 | } else { | |
261 | auto &p = snap_inode_map[in->vino()]; | |
11fdf7f2 | 262 | ceph_assert(!p); // should be no dup inos! |
b32b8144 FG |
263 | p = in; |
264 | } | |
7c673cae FG |
265 | |
266 | if (in->ino() < MDS_INO_SYSTEM_BASE) { | |
267 | if (in->ino() == MDS_INO_ROOT) | |
268 | root = in; | |
269 | else if (in->ino() == MDS_INO_MDSDIR(mds->get_nodeid())) | |
270 | myin = in; | |
271 | else if (in->is_stray()) { | |
272 | if (MDS_INO_STRAY_OWNER(in->ino()) == mds->get_nodeid()) { | |
273 | strays[MDS_INO_STRAY_INDEX(in->ino())] = in; | |
274 | } | |
275 | } | |
276 | if (in->is_base()) | |
277 | base_inodes.insert(in); | |
278 | } | |
279 | ||
181888fb | 280 | if (cache_toofull()) { |
7c673cae FG |
281 | exceeded_size_limit = true; |
282 | } | |
283 | } | |
284 | ||
285 | void MDCache::remove_inode(CInode *o) | |
286 | { | |
287 | dout(14) << "remove_inode " << *o << dendl; | |
288 | ||
289 | if (o->get_parent_dn()) { | |
290 | // FIXME: multiple parents? | |
291 | CDentry *dn = o->get_parent_dn(); | |
11fdf7f2 | 292 | ceph_assert(!dn->is_dirty()); |
7c673cae FG |
293 | dn->dir->unlink_inode(dn); // leave dentry ... FIXME? |
294 | } | |
295 | ||
296 | if (o->is_dirty()) | |
297 | o->mark_clean(); | |
298 | if (o->is_dirty_parent()) | |
299 | o->clear_dirty_parent(); | |
300 | ||
301 | o->clear_scatter_dirty(); | |
302 | ||
303 | o->item_open_file.remove_myself(); | |
304 | ||
31f18b77 FG |
305 | if (o->state_test(CInode::STATE_QUEUEDEXPORTPIN)) |
306 | export_pin_queue.erase(o); | |
7c673cae FG |
307 | |
308 | // remove from inode map | |
11fdf7f2 | 309 | if (o->last == CEPH_NOSNAP) { |
b32b8144 | 310 | inode_map.erase(o->ino()); |
11fdf7f2 TL |
311 | } else { |
312 | o->item_caps.remove_myself(); | |
b32b8144 | 313 | snap_inode_map.erase(o->vino()); |
11fdf7f2 | 314 | } |
7c673cae FG |
315 | |
316 | if (o->ino() < MDS_INO_SYSTEM_BASE) { | |
317 | if (o == root) root = 0; | |
318 | if (o == myin) myin = 0; | |
319 | if (o->is_stray()) { | |
320 | if (MDS_INO_STRAY_OWNER(o->ino()) == mds->get_nodeid()) { | |
321 | strays[MDS_INO_STRAY_INDEX(o->ino())] = 0; | |
322 | } | |
323 | } | |
324 | if (o->is_base()) | |
325 | base_inodes.erase(o); | |
11fdf7f2 | 326 | } |
7c673cae FG |
327 | |
328 | // delete it | |
11fdf7f2 | 329 | ceph_assert(o->get_num_ref() == 0); |
7c673cae FG |
330 | delete o; |
331 | } | |
332 | ||
333 | file_layout_t MDCache::gen_default_file_layout(const MDSMap &mdsmap) | |
334 | { | |
335 | file_layout_t result = file_layout_t::get_default(); | |
336 | result.pool_id = mdsmap.get_first_data_pool(); | |
337 | return result; | |
338 | } | |
339 | ||
340 | file_layout_t MDCache::gen_default_log_layout(const MDSMap &mdsmap) | |
341 | { | |
342 | file_layout_t result = file_layout_t::get_default(); | |
343 | result.pool_id = mdsmap.get_metadata_pool(); | |
11fdf7f2 TL |
344 | if (g_conf()->mds_log_segment_size > 0) { |
345 | result.object_size = g_conf()->mds_log_segment_size; | |
346 | result.stripe_unit = g_conf()->mds_log_segment_size; | |
7c673cae FG |
347 | } |
348 | return result; | |
349 | } | |
350 | ||
351 | void MDCache::init_layouts() | |
352 | { | |
353 | default_file_layout = gen_default_file_layout(*(mds->mdsmap)); | |
354 | default_log_layout = gen_default_log_layout(*(mds->mdsmap)); | |
355 | } | |
356 | ||
357 | void MDCache::create_unlinked_system_inode(CInode *in, inodeno_t ino, | |
358 | int mode) const | |
359 | { | |
360 | in->inode.ino = ino; | |
361 | in->inode.version = 1; | |
362 | in->inode.xattr_version = 1; | |
363 | in->inode.mode = 0500 | mode; | |
364 | in->inode.size = 0; | |
365 | in->inode.ctime = | |
366 | in->inode.mtime = | |
367 | in->inode.btime = ceph_clock_now(); | |
368 | in->inode.nlink = 1; | |
369 | in->inode.truncate_size = -1ull; | |
370 | in->inode.change_attr = 0; | |
371 | in->inode.export_pin = MDS_RANK_NONE; | |
372 | ||
373 | memset(&in->inode.dir_layout, 0, sizeof(in->inode.dir_layout)); | |
374 | if (in->inode.is_dir()) { | |
11fdf7f2 | 375 | in->inode.dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; |
f64942e4 AA |
376 | in->inode.rstat.rsubdirs = 1; /* itself */ |
377 | in->inode.rstat.rctime = in->inode.ctime; | |
7c673cae FG |
378 | } else { |
379 | in->inode.layout = default_file_layout; | |
380 | ++in->inode.rstat.rfiles; | |
381 | } | |
382 | in->inode.accounted_rstat = in->inode.rstat; | |
383 | ||
384 | if (in->is_base()) { | |
385 | if (in->is_root()) | |
386 | in->inode_auth = mds_authority_t(mds->get_nodeid(), CDIR_AUTH_UNKNOWN); | |
387 | else | |
388 | in->inode_auth = mds_authority_t(mds_rank_t(in->ino() - MDS_INO_MDSDIR_OFFSET), CDIR_AUTH_UNKNOWN); | |
389 | in->open_snaprealm(); // empty snaprealm | |
11fdf7f2 | 390 | ceph_assert(!in->snaprealm->parent); // created its own |
7c673cae FG |
391 | in->snaprealm->srnode.seq = 1; |
392 | } | |
393 | } | |
394 | ||
395 | CInode *MDCache::create_system_inode(inodeno_t ino, int mode) | |
396 | { | |
397 | dout(0) << "creating system inode with ino:" << ino << dendl; | |
398 | CInode *in = new CInode(this); | |
399 | create_unlinked_system_inode(in, ino, mode); | |
400 | add_inode(in); | |
401 | return in; | |
402 | } | |
403 | ||
404 | CInode *MDCache::create_root_inode() | |
405 | { | |
406 | CInode *i = create_system_inode(MDS_INO_ROOT, S_IFDIR|0755); | |
11fdf7f2 TL |
407 | i->inode.uid = g_conf()->mds_root_ino_uid; |
408 | i->inode.gid = g_conf()->mds_root_ino_gid; | |
7c673cae FG |
409 | i->inode.layout = default_file_layout; |
410 | i->inode.layout.pool_id = mds->mdsmap->get_first_data_pool(); | |
411 | return i; | |
412 | } | |
413 | ||
414 | void MDCache::create_empty_hierarchy(MDSGather *gather) | |
415 | { | |
416 | // create root dir | |
417 | CInode *root = create_root_inode(); | |
418 | ||
419 | // force empty root dir | |
420 | CDir *rootdir = root->get_or_open_dirfrag(this, frag_t()); | |
421 | adjust_subtree_auth(rootdir, mds->get_nodeid()); | |
422 | rootdir->dir_rep = CDir::REP_ALL; //NONE; | |
423 | ||
11fdf7f2 TL |
424 | ceph_assert(rootdir->fnode.accounted_fragstat == rootdir->fnode.fragstat); |
425 | ceph_assert(rootdir->fnode.fragstat == root->inode.dirstat); | |
426 | ceph_assert(rootdir->fnode.accounted_rstat == rootdir->fnode.rstat); | |
f64942e4 AA |
427 | /* Do no update rootdir rstat information of the fragment, rstat upkeep magic |
428 | * assume version 0 is stale/invalid. | |
429 | */ | |
7c673cae FG |
430 | |
431 | rootdir->mark_complete(); | |
432 | rootdir->mark_dirty(rootdir->pre_dirty(), mds->mdlog->get_current_segment()); | |
433 | rootdir->commit(0, gather->new_sub()); | |
434 | ||
28e407b8 AA |
435 | root->mark_clean(); |
436 | root->mark_dirty(root->pre_dirty(), mds->mdlog->get_current_segment()); | |
437 | root->mark_dirty_parent(mds->mdlog->get_current_segment(), true); | |
438 | root->flush(gather->new_sub()); | |
7c673cae FG |
439 | } |
440 | ||
441 | void MDCache::create_mydir_hierarchy(MDSGather *gather) | |
442 | { | |
443 | // create mds dir | |
444 | CInode *my = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR); | |
445 | ||
446 | CDir *mydir = my->get_or_open_dirfrag(this, frag_t()); | |
447 | adjust_subtree_auth(mydir, mds->get_nodeid()); | |
448 | ||
449 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
450 | ||
451 | // stray dir | |
452 | for (int i = 0; i < NUM_STRAY; ++i) { | |
453 | CInode *stray = create_system_inode(MDS_INO_STRAY(mds->get_nodeid(), i), S_IFDIR); | |
454 | CDir *straydir = stray->get_or_open_dirfrag(this, frag_t()); | |
455 | stringstream name; | |
456 | name << "stray" << i; | |
457 | CDentry *sdn = mydir->add_primary_dentry(name.str(), stray); | |
458 | sdn->_mark_dirty(mds->mdlog->get_current_segment()); | |
459 | ||
460 | stray->inode.dirstat = straydir->fnode.fragstat; | |
461 | ||
462 | mydir->fnode.rstat.add(stray->inode.rstat); | |
463 | mydir->fnode.fragstat.nsubdirs++; | |
464 | // save them | |
465 | straydir->mark_complete(); | |
466 | straydir->mark_dirty(straydir->pre_dirty(), ls); | |
467 | straydir->commit(0, gather->new_sub()); | |
28e407b8 | 468 | stray->mark_dirty_parent(ls, true); |
7c673cae FG |
469 | stray->store_backtrace(gather->new_sub()); |
470 | } | |
471 | ||
472 | mydir->fnode.accounted_fragstat = mydir->fnode.fragstat; | |
473 | mydir->fnode.accounted_rstat = mydir->fnode.rstat; | |
474 | ||
475 | myin->inode.dirstat = mydir->fnode.fragstat; | |
476 | myin->inode.rstat = mydir->fnode.rstat; | |
477 | ++myin->inode.rstat.rsubdirs; | |
478 | myin->inode.accounted_rstat = myin->inode.rstat; | |
479 | ||
480 | mydir->mark_complete(); | |
481 | mydir->mark_dirty(mydir->pre_dirty(), ls); | |
482 | mydir->commit(0, gather->new_sub()); | |
483 | ||
484 | myin->store(gather->new_sub()); | |
485 | } | |
486 | ||
487 | struct C_MDC_CreateSystemFile : public MDCacheLogContext { | |
488 | MutationRef mut; | |
489 | CDentry *dn; | |
490 | version_t dpv; | |
11fdf7f2 TL |
491 | MDSContext *fin; |
492 | C_MDC_CreateSystemFile(MDCache *c, MutationRef& mu, CDentry *d, version_t v, MDSContext *f) : | |
7c673cae FG |
493 | MDCacheLogContext(c), mut(mu), dn(d), dpv(v), fin(f) {} |
494 | void finish(int r) override { | |
495 | mdcache->_create_system_file_finish(mut, dn, dpv, fin); | |
496 | } | |
497 | }; | |
498 | ||
11fdf7f2 | 499 | void MDCache::_create_system_file(CDir *dir, std::string_view name, CInode *in, MDSContext *fin) |
7c673cae FG |
500 | { |
501 | dout(10) << "_create_system_file " << name << " in " << *dir << dendl; | |
502 | CDentry *dn = dir->add_null_dentry(name); | |
503 | ||
504 | dn->push_projected_linkage(in); | |
505 | version_t dpv = dn->pre_dirty(); | |
506 | ||
507 | CDir *mdir = 0; | |
508 | if (in->inode.is_dir()) { | |
509 | in->inode.rstat.rsubdirs = 1; | |
510 | ||
511 | mdir = in->get_or_open_dirfrag(this, frag_t()); | |
512 | mdir->mark_complete(); | |
513 | mdir->pre_dirty(); | |
514 | } else | |
515 | in->inode.rstat.rfiles = 1; | |
516 | in->inode.version = dn->pre_dirty(); | |
517 | ||
518 | SnapRealm *realm = dir->get_inode()->find_snaprealm(); | |
519 | dn->first = in->first = realm->get_newest_seq() + 1; | |
520 | ||
521 | MutationRef mut(new MutationImpl()); | |
522 | ||
523 | // force some locks. hacky. | |
524 | mds->locker->wrlock_force(&dir->inode->filelock, mut); | |
525 | mds->locker->wrlock_force(&dir->inode->nestlock, mut); | |
526 | ||
527 | mut->ls = mds->mdlog->get_current_segment(); | |
528 | EUpdate *le = new EUpdate(mds->mdlog, "create system file"); | |
529 | mds->mdlog->start_entry(le); | |
530 | ||
531 | if (!in->is_mdsdir()) { | |
532 | predirty_journal_parents(mut, &le->metablob, in, dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
533 | le->metablob.add_primary_dentry(dn, in, true); | |
534 | } else { | |
535 | predirty_journal_parents(mut, &le->metablob, in, dir, PREDIRTY_DIR, 1); | |
536 | journal_dirty_inode(mut.get(), &le->metablob, in); | |
537 | dn->push_projected_linkage(in->ino(), in->d_type()); | |
538 | le->metablob.add_remote_dentry(dn, true, in->ino(), in->d_type()); | |
539 | le->metablob.add_root(true, in); | |
540 | } | |
541 | if (mdir) | |
542 | le->metablob.add_new_dir(mdir); // dirty AND complete AND new | |
543 | ||
544 | mds->mdlog->submit_entry(le, new C_MDC_CreateSystemFile(this, mut, dn, dpv, fin)); | |
545 | mds->mdlog->flush(); | |
546 | } | |
547 | ||
11fdf7f2 | 548 | void MDCache::_create_system_file_finish(MutationRef& mut, CDentry *dn, version_t dpv, MDSContext *fin) |
7c673cae FG |
549 | { |
550 | dout(10) << "_create_system_file_finish " << *dn << dendl; | |
551 | ||
552 | dn->pop_projected_linkage(); | |
553 | dn->mark_dirty(dpv, mut->ls); | |
554 | ||
555 | CInode *in = dn->get_linkage()->get_inode(); | |
556 | in->inode.version--; | |
557 | in->mark_dirty(in->inode.version + 1, mut->ls); | |
558 | ||
559 | if (in->inode.is_dir()) { | |
560 | CDir *dir = in->get_dirfrag(frag_t()); | |
11fdf7f2 | 561 | ceph_assert(dir); |
7c673cae FG |
562 | dir->mark_dirty(1, mut->ls); |
563 | dir->mark_new(mut->ls); | |
564 | } | |
565 | ||
566 | mut->apply(); | |
567 | mds->locker->drop_locks(mut.get()); | |
568 | mut->cleanup(); | |
569 | ||
570 | fin->complete(0); | |
571 | ||
572 | //if (dir && MDS_INO_IS_MDSDIR(in->ino())) | |
573 | //migrator->export_dir(dir, (int)in->ino() - MDS_INO_MDSDIR_OFFSET); | |
574 | } | |
575 | ||
576 | ||
577 | ||
578 | struct C_MDS_RetryOpenRoot : public MDSInternalContext { | |
579 | MDCache *cache; | |
580 | explicit C_MDS_RetryOpenRoot(MDCache *c) : MDSInternalContext(c->mds), cache(c) {} | |
581 | void finish(int r) override { | |
582 | if (r < 0) { | |
583 | // If we can't open root, something disastrous has happened: mark | |
584 | // this rank damaged for operator intervention. Note that | |
585 | // it is not okay to call suicide() here because we are in | |
586 | // a Finisher callback. | |
587 | cache->mds->damaged(); | |
588 | ceph_abort(); // damaged should never return | |
589 | } else { | |
590 | cache->open_root(); | |
591 | } | |
592 | } | |
593 | }; | |
594 | ||
11fdf7f2 | 595 | void MDCache::open_root_inode(MDSContext *c) |
7c673cae FG |
596 | { |
597 | if (mds->get_nodeid() == mds->mdsmap->get_root()) { | |
598 | CInode *in; | |
599 | in = create_system_inode(MDS_INO_ROOT, S_IFDIR|0755); // initially inaccurate! | |
600 | in->fetch(c); | |
601 | } else { | |
602 | discover_base_ino(MDS_INO_ROOT, c, mds->mdsmap->get_root()); | |
603 | } | |
604 | } | |
605 | ||
11fdf7f2 | 606 | void MDCache::open_mydir_inode(MDSContext *c) |
7c673cae | 607 | { |
7c673cae | 608 | CInode *in = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR|0755); // initially inaccurate! |
11fdf7f2 | 609 | in->fetch(c); |
7c673cae FG |
610 | } |
611 | ||
11fdf7f2 | 612 | void MDCache::open_mydir_frag(MDSContext *c) |
28e407b8 AA |
613 | { |
614 | open_mydir_inode( | |
615 | new MDSInternalContextWrapper(mds, | |
616 | new FunctionContext([this, c](int r) { | |
617 | if (r < 0) { | |
618 | c->complete(r); | |
619 | return; | |
620 | } | |
621 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); | |
11fdf7f2 | 622 | ceph_assert(mydir); |
28e407b8 AA |
623 | adjust_subtree_auth(mydir, mds->get_nodeid()); |
624 | mydir->fetch(c); | |
625 | }) | |
626 | ) | |
627 | ); | |
628 | } | |
629 | ||
7c673cae FG |
630 | void MDCache::open_root() |
631 | { | |
632 | dout(10) << "open_root" << dendl; | |
633 | ||
634 | if (!root) { | |
635 | open_root_inode(new C_MDS_RetryOpenRoot(this)); | |
636 | return; | |
637 | } | |
638 | if (mds->get_nodeid() == mds->mdsmap->get_root()) { | |
11fdf7f2 | 639 | ceph_assert(root->is_auth()); |
7c673cae | 640 | CDir *rootdir = root->get_or_open_dirfrag(this, frag_t()); |
11fdf7f2 | 641 | ceph_assert(rootdir); |
7c673cae FG |
642 | if (!rootdir->is_subtree_root()) |
643 | adjust_subtree_auth(rootdir, mds->get_nodeid()); | |
644 | if (!rootdir->is_complete()) { | |
645 | rootdir->fetch(new C_MDS_RetryOpenRoot(this)); | |
646 | return; | |
647 | } | |
648 | } else { | |
11fdf7f2 | 649 | ceph_assert(!root->is_auth()); |
7c673cae FG |
650 | CDir *rootdir = root->get_dirfrag(frag_t()); |
651 | if (!rootdir) { | |
224ce89b | 652 | open_remote_dirfrag(root, frag_t(), new C_MDS_RetryOpenRoot(this)); |
7c673cae FG |
653 | return; |
654 | } | |
655 | } | |
656 | ||
657 | if (!myin) { | |
658 | CInode *in = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR|0755); // initially inaccurate! | |
659 | in->fetch(new C_MDS_RetryOpenRoot(this)); | |
660 | return; | |
661 | } | |
662 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); | |
11fdf7f2 | 663 | ceph_assert(mydir); |
7c673cae FG |
664 | adjust_subtree_auth(mydir, mds->get_nodeid()); |
665 | ||
666 | populate_mydir(); | |
667 | } | |
668 | ||
669 | void MDCache::populate_mydir() | |
670 | { | |
11fdf7f2 | 671 | ceph_assert(myin); |
7c673cae | 672 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); |
11fdf7f2 | 673 | ceph_assert(mydir); |
7c673cae FG |
674 | |
675 | dout(10) << "populate_mydir " << *mydir << dendl; | |
676 | ||
677 | if (!mydir->is_complete()) { | |
678 | mydir->fetch(new C_MDS_RetryOpenRoot(this)); | |
679 | return; | |
680 | } | |
681 | ||
682 | if (mydir->get_version() == 0 && mydir->state_test(CDir::STATE_BADFRAG)) { | |
683 | // A missing dirfrag, we will recreate it. Before that, we must dirty | |
684 | // it before dirtying any of the strays we create within it. | |
685 | mds->clog->warn() << "fragment " << mydir->dirfrag() << " was unreadable, " | |
686 | "recreating it now"; | |
687 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
688 | mydir->state_clear(CDir::STATE_BADFRAG); | |
689 | mydir->mark_complete(); | |
690 | mydir->mark_dirty(mydir->pre_dirty(), ls); | |
691 | } | |
692 | ||
693 | // open or create stray | |
694 | uint64_t num_strays = 0; | |
695 | for (int i = 0; i < NUM_STRAY; ++i) { | |
696 | stringstream name; | |
697 | name << "stray" << i; | |
698 | CDentry *straydn = mydir->lookup(name.str()); | |
699 | ||
700 | // allow for older fs's with stray instead of stray0 | |
701 | if (straydn == NULL && i == 0) | |
702 | straydn = mydir->lookup("stray"); | |
703 | ||
704 | if (!straydn || !straydn->get_linkage()->get_inode()) { | |
705 | _create_system_file(mydir, name.str().c_str(), create_system_inode(MDS_INO_STRAY(mds->get_nodeid(), i), S_IFDIR), | |
706 | new C_MDS_RetryOpenRoot(this)); | |
707 | return; | |
708 | } | |
11fdf7f2 TL |
709 | ceph_assert(straydn); |
710 | ceph_assert(strays[i]); | |
7c673cae FG |
711 | // we make multiple passes through this method; make sure we only pin each stray once. |
712 | if (!strays[i]->state_test(CInode::STATE_STRAYPINNED)) { | |
713 | strays[i]->get(CInode::PIN_STRAY); | |
714 | strays[i]->state_set(CInode::STATE_STRAYPINNED); | |
715 | strays[i]->get_stickydirs(); | |
716 | } | |
717 | dout(20) << " stray num " << i << " is " << *strays[i] << dendl; | |
718 | ||
719 | // open all frags | |
11fdf7f2 TL |
720 | frag_vec_t leaves; |
721 | strays[i]->dirfragtree.get_leaves(leaves); | |
722 | for (const auto& leaf : leaves) { | |
723 | CDir *dir = strays[i]->get_dirfrag(leaf); | |
7c673cae | 724 | if (!dir) { |
11fdf7f2 | 725 | dir = strays[i]->get_or_open_dirfrag(this, leaf); |
7c673cae FG |
726 | } |
727 | ||
728 | // DamageTable applies special handling to strays: it will | |
729 | // have damaged() us out if one is damaged. | |
11fdf7f2 | 730 | ceph_assert(!dir->state_test(CDir::STATE_BADFRAG)); |
7c673cae FG |
731 | |
732 | if (dir->get_version() == 0) { | |
733 | dir->fetch(new C_MDS_RetryOpenRoot(this)); | |
734 | return; | |
735 | } | |
736 | ||
737 | if (dir->get_frag_size() > 0) | |
738 | num_strays += dir->get_frag_size(); | |
739 | } | |
740 | } | |
741 | ||
7c673cae FG |
742 | // okay! |
743 | dout(10) << "populate_mydir done" << dendl; | |
11fdf7f2 | 744 | ceph_assert(!open); |
7c673cae FG |
745 | open = true; |
746 | mds->queue_waiters(waiting_for_open); | |
747 | ||
11fdf7f2 TL |
748 | stray_manager.set_num_strays(num_strays); |
749 | stray_manager.activate(); | |
750 | ||
7c673cae FG |
751 | scan_stray_dir(); |
752 | } | |
753 | ||
11fdf7f2 | 754 | void MDCache::open_foreign_mdsdir(inodeno_t ino, MDSContext *fin) |
7c673cae FG |
755 | { |
756 | discover_base_ino(ino, fin, mds_rank_t(ino & (MAX_MDS-1))); | |
757 | } | |
758 | ||
759 | CDir *MDCache::get_stray_dir(CInode *in) | |
760 | { | |
761 | string straydname; | |
762 | in->name_stray_dentry(straydname); | |
763 | ||
764 | CInode *strayi = get_stray(); | |
11fdf7f2 | 765 | ceph_assert(strayi); |
7c673cae FG |
766 | frag_t fg = strayi->pick_dirfrag(straydname); |
767 | CDir *straydir = strayi->get_dirfrag(fg); | |
11fdf7f2 | 768 | ceph_assert(straydir); |
7c673cae FG |
769 | return straydir; |
770 | } | |
771 | ||
772 | CDentry *MDCache::get_or_create_stray_dentry(CInode *in) | |
773 | { | |
774 | CDir *straydir = get_stray_dir(in); | |
775 | string straydname; | |
776 | in->name_stray_dentry(straydname); | |
777 | CDentry *straydn = straydir->lookup(straydname); | |
778 | if (!straydn) { | |
779 | straydn = straydir->add_null_dentry(straydname); | |
780 | straydn->mark_new(); | |
781 | } else { | |
11fdf7f2 | 782 | ceph_assert(straydn->get_projected_linkage()->is_null()); |
7c673cae FG |
783 | } |
784 | ||
785 | straydn->state_set(CDentry::STATE_STRAY); | |
786 | return straydn; | |
787 | } | |
788 | ||
789 | ||
790 | ||
11fdf7f2 | 791 | MDSCacheObject *MDCache::get_object(const MDSCacheObjectInfo &info) |
7c673cae FG |
792 | { |
793 | // inode? | |
794 | if (info.ino) | |
795 | return get_inode(info.ino, info.snapid); | |
796 | ||
797 | // dir or dentry. | |
798 | CDir *dir = get_dirfrag(info.dirfrag); | |
799 | if (!dir) return 0; | |
800 | ||
801 | if (info.dname.length()) | |
802 | return dir->lookup(info.dname, info.snapid); | |
803 | else | |
804 | return dir; | |
805 | } | |
806 | ||
807 | ||
808 | ||
809 | ||
810 | // ==================================================================== | |
811 | // subtree management | |
812 | ||
7c673cae FG |
813 | /* |
814 | * adjust the dir_auth of a subtree. | |
815 | * merge with parent and/or child subtrees, if is it appropriate. | |
816 | * merge can ONLY happen if both parent and child have unambiguous auth. | |
817 | */ | |
28e407b8 | 818 | void MDCache::adjust_subtree_auth(CDir *dir, mds_authority_t auth, bool adjust_pop) |
7c673cae FG |
819 | { |
820 | dout(7) << "adjust_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
821 | << " on " << *dir << dendl; | |
822 | ||
7c673cae FG |
823 | show_subtrees(); |
824 | ||
825 | CDir *root; | |
826 | if (dir->inode->is_base()) { | |
827 | root = dir; // bootstrap hack. | |
828 | if (subtrees.count(root) == 0) { | |
829 | subtrees[root]; | |
830 | root->get(CDir::PIN_SUBTREE); | |
831 | } | |
832 | } else { | |
833 | root = get_subtree_root(dir); // subtree root | |
834 | } | |
11fdf7f2 TL |
835 | ceph_assert(root); |
836 | ceph_assert(subtrees.count(root)); | |
7c673cae FG |
837 | dout(7) << " current root is " << *root << dendl; |
838 | ||
839 | if (root == dir) { | |
840 | // i am already a subtree. | |
841 | dir->set_dir_auth(auth); | |
842 | } else { | |
843 | // i am a new subtree. | |
844 | dout(10) << " new subtree at " << *dir << dendl; | |
11fdf7f2 | 845 | ceph_assert(subtrees.count(dir) == 0); |
7c673cae FG |
846 | subtrees[dir]; // create empty subtree bounds list for me. |
847 | dir->get(CDir::PIN_SUBTREE); | |
848 | ||
849 | // set dir_auth | |
850 | dir->set_dir_auth(auth); | |
851 | ||
852 | // move items nested beneath me, under me. | |
853 | set<CDir*>::iterator p = subtrees[root].begin(); | |
854 | while (p != subtrees[root].end()) { | |
855 | set<CDir*>::iterator next = p; | |
856 | ++next; | |
857 | if (get_subtree_root((*p)->get_parent_dir()) == dir) { | |
858 | // move under me | |
859 | dout(10) << " claiming child bound " << **p << dendl; | |
860 | subtrees[dir].insert(*p); | |
861 | subtrees[root].erase(p); | |
862 | } | |
863 | p = next; | |
864 | } | |
865 | ||
866 | // i am a bound of the parent subtree. | |
867 | subtrees[root].insert(dir); | |
868 | ||
869 | // i am now the subtree root. | |
870 | root = dir; | |
871 | ||
872 | // adjust recursive pop counters | |
28e407b8 | 873 | if (adjust_pop && dir->is_auth()) { |
7c673cae FG |
874 | CDir *p = dir->get_parent_dir(); |
875 | while (p) { | |
11fdf7f2 | 876 | p->pop_auth_subtree.sub(dir->pop_auth_subtree); |
7c673cae FG |
877 | if (p->is_subtree_root()) break; |
878 | p = p->inode->get_parent_dir(); | |
879 | } | |
880 | } | |
7c673cae FG |
881 | } |
882 | ||
883 | show_subtrees(); | |
884 | } | |
885 | ||
886 | ||
887 | void MDCache::try_subtree_merge(CDir *dir) | |
888 | { | |
889 | dout(7) << "try_subtree_merge " << *dir << dendl; | |
b32b8144 FG |
890 | // record my old bounds |
891 | auto oldbounds = subtrees.at(dir); | |
7c673cae | 892 | |
224ce89b | 893 | set<CInode*> to_eval; |
7c673cae | 894 | // try merge at my root |
224ce89b | 895 | try_subtree_merge_at(dir, &to_eval); |
7c673cae FG |
896 | |
897 | // try merge at my old bounds | |
224ce89b WB |
898 | for (auto bound : oldbounds) |
899 | try_subtree_merge_at(bound, &to_eval); | |
900 | ||
901 | if (!(mds->is_any_replay() || mds->is_resolve())) { | |
902 | for(auto in : to_eval) | |
903 | eval_subtree_root(in); | |
904 | } | |
7c673cae FG |
905 | } |
906 | ||
907 | class C_MDC_SubtreeMergeWB : public MDCacheLogContext { | |
908 | CInode *in; | |
909 | MutationRef mut; | |
910 | public: | |
911 | C_MDC_SubtreeMergeWB(MDCache *mdc, CInode *i, MutationRef& m) : MDCacheLogContext(mdc), in(i), mut(m) {} | |
912 | void finish(int r) override { | |
913 | mdcache->subtree_merge_writebehind_finish(in, mut); | |
914 | } | |
915 | }; | |
916 | ||
28e407b8 | 917 | void MDCache::try_subtree_merge_at(CDir *dir, set<CInode*> *to_eval, bool adjust_pop) |
7c673cae FG |
918 | { |
919 | dout(10) << "try_subtree_merge_at " << *dir << dendl; | |
b32b8144 FG |
920 | |
921 | if (dir->dir_auth.second != CDIR_AUTH_UNKNOWN || | |
922 | dir->state_test(CDir::STATE_EXPORTBOUND) || | |
923 | dir->state_test(CDir::STATE_AUXSUBTREE)) | |
924 | return; | |
925 | ||
926 | auto it = subtrees.find(dir); | |
11fdf7f2 | 927 | ceph_assert(it != subtrees.end()); |
7c673cae | 928 | |
7c673cae FG |
929 | // merge with parent? |
930 | CDir *parent = dir; | |
931 | if (!dir->inode->is_base()) | |
932 | parent = get_subtree_root(dir->get_parent_dir()); | |
933 | ||
b32b8144 FG |
934 | if (parent != dir && // we have a parent, |
935 | parent->dir_auth == dir->dir_auth) { // auth matches, | |
7c673cae FG |
936 | // merge with parent. |
937 | dout(10) << " subtree merge at " << *dir << dendl; | |
938 | dir->set_dir_auth(CDIR_AUTH_DEFAULT); | |
939 | ||
940 | // move our bounds under the parent | |
b32b8144 | 941 | subtrees[parent].insert(it->second.begin(), it->second.end()); |
7c673cae FG |
942 | |
943 | // we are no longer a subtree or bound | |
944 | dir->put(CDir::PIN_SUBTREE); | |
b32b8144 | 945 | subtrees.erase(it); |
7c673cae FG |
946 | subtrees[parent].erase(dir); |
947 | ||
948 | // adjust popularity? | |
28e407b8 | 949 | if (adjust_pop && dir->is_auth()) { |
28e407b8 | 950 | CDir *cur = dir; |
7c673cae FG |
951 | CDir *p = dir->get_parent_dir(); |
952 | while (p) { | |
11fdf7f2 | 953 | p->pop_auth_subtree.add(dir->pop_auth_subtree); |
28e407b8 | 954 | p->pop_lru_subdirs.push_front(&cur->get_inode()->item_pop_lru); |
7c673cae | 955 | if (p->is_subtree_root()) break; |
28e407b8 | 956 | cur = p; |
7c673cae FG |
957 | p = p->inode->get_parent_dir(); |
958 | } | |
959 | } | |
960 | ||
224ce89b WB |
961 | if (to_eval && dir->get_inode()->is_auth()) |
962 | to_eval->insert(dir->get_inode()); | |
7c673cae | 963 | |
181888fb FG |
964 | show_subtrees(15); |
965 | } | |
7c673cae FG |
966 | } |
967 | ||
968 | void MDCache::subtree_merge_writebehind_finish(CInode *in, MutationRef& mut) | |
969 | { | |
970 | dout(10) << "subtree_merge_writebehind_finish on " << in << dendl; | |
971 | in->pop_and_dirty_projected_inode(mut->ls); | |
972 | ||
973 | mut->apply(); | |
974 | mds->locker->drop_locks(mut.get()); | |
975 | mut->cleanup(); | |
976 | ||
977 | in->auth_unpin(this); | |
978 | } | |
979 | ||
980 | void MDCache::eval_subtree_root(CInode *diri) | |
981 | { | |
982 | // evaluate subtree inode filelock? | |
983 | // (we should scatter the filelock on subtree bounds) | |
11fdf7f2 | 984 | ceph_assert(diri->is_auth()); |
224ce89b | 985 | mds->locker->try_eval(diri, CEPH_LOCK_IFILE | CEPH_LOCK_INEST); |
7c673cae FG |
986 | } |
987 | ||
988 | ||
11fdf7f2 | 989 | void MDCache::adjust_bounded_subtree_auth(CDir *dir, const set<CDir*>& bounds, mds_authority_t auth) |
7c673cae FG |
990 | { |
991 | dout(7) << "adjust_bounded_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
992 | << " on " << *dir | |
993 | << " bounds " << bounds | |
994 | << dendl; | |
995 | ||
996 | show_subtrees(); | |
997 | ||
998 | CDir *root; | |
999 | if (dir->ino() == MDS_INO_ROOT) { | |
1000 | root = dir; // bootstrap hack. | |
1001 | if (subtrees.count(root) == 0) { | |
1002 | subtrees[root]; | |
1003 | root->get(CDir::PIN_SUBTREE); | |
1004 | } | |
1005 | } else { | |
1006 | root = get_subtree_root(dir); // subtree root | |
1007 | } | |
11fdf7f2 TL |
1008 | ceph_assert(root); |
1009 | ceph_assert(subtrees.count(root)); | |
7c673cae FG |
1010 | dout(7) << " current root is " << *root << dendl; |
1011 | ||
1012 | mds_authority_t oldauth = dir->authority(); | |
1013 | ||
1014 | if (root == dir) { | |
1015 | // i am already a subtree. | |
1016 | dir->set_dir_auth(auth); | |
1017 | } else { | |
1018 | // i am a new subtree. | |
1019 | dout(10) << " new subtree at " << *dir << dendl; | |
11fdf7f2 | 1020 | ceph_assert(subtrees.count(dir) == 0); |
7c673cae FG |
1021 | subtrees[dir]; // create empty subtree bounds list for me. |
1022 | dir->get(CDir::PIN_SUBTREE); | |
1023 | ||
1024 | // set dir_auth | |
1025 | dir->set_dir_auth(auth); | |
1026 | ||
1027 | // move items nested beneath me, under me. | |
1028 | set<CDir*>::iterator p = subtrees[root].begin(); | |
1029 | while (p != subtrees[root].end()) { | |
1030 | set<CDir*>::iterator next = p; | |
1031 | ++next; | |
1032 | if (get_subtree_root((*p)->get_parent_dir()) == dir) { | |
1033 | // move under me | |
1034 | dout(10) << " claiming child bound " << **p << dendl; | |
1035 | subtrees[dir].insert(*p); | |
1036 | subtrees[root].erase(p); | |
1037 | } | |
1038 | p = next; | |
1039 | } | |
1040 | ||
1041 | // i am a bound of the parent subtree. | |
1042 | subtrees[root].insert(dir); | |
1043 | ||
1044 | // i am now the subtree root. | |
1045 | root = dir; | |
1046 | } | |
1047 | ||
224ce89b WB |
1048 | set<CInode*> to_eval; |
1049 | ||
7c673cae FG |
1050 | // verify/adjust bounds. |
1051 | // - these may be new, or | |
1052 | // - beneath existing ambiguous bounds (which will be collapsed), | |
1053 | // - but NOT beneath unambiguous bounds. | |
11fdf7f2 | 1054 | for (const auto& bound : bounds) { |
7c673cae FG |
1055 | // new bound? |
1056 | if (subtrees[dir].count(bound) == 0) { | |
1057 | if (get_subtree_root(bound) == dir) { | |
1058 | dout(10) << " new bound " << *bound << ", adjusting auth back to old " << oldauth << dendl; | |
1059 | adjust_subtree_auth(bound, oldauth); // otherwise, adjust at bound. | |
1060 | } | |
1061 | else { | |
1062 | dout(10) << " want bound " << *bound << dendl; | |
1063 | CDir *t = get_subtree_root(bound->get_parent_dir()); | |
1064 | if (subtrees[t].count(bound) == 0) { | |
11fdf7f2 | 1065 | ceph_assert(t != dir); |
7c673cae FG |
1066 | dout(10) << " new bound " << *bound << dendl; |
1067 | adjust_subtree_auth(bound, t->authority()); | |
1068 | } | |
1069 | // make sure it's nested beneath ambiguous subtree(s) | |
1070 | while (1) { | |
1071 | while (subtrees[dir].count(t) == 0) | |
1072 | t = get_subtree_root(t->get_parent_dir()); | |
1073 | dout(10) << " swallowing intervening subtree at " << *t << dendl; | |
1074 | adjust_subtree_auth(t, auth); | |
224ce89b | 1075 | try_subtree_merge_at(t, &to_eval); |
7c673cae FG |
1076 | t = get_subtree_root(bound->get_parent_dir()); |
1077 | if (t == dir) break; | |
1078 | } | |
1079 | } | |
1080 | } | |
1081 | else { | |
1082 | dout(10) << " already have bound " << *bound << dendl; | |
1083 | } | |
1084 | } | |
1085 | // merge stray bounds? | |
1086 | while (!subtrees[dir].empty()) { | |
1087 | set<CDir*> copy = subtrees[dir]; | |
1088 | for (set<CDir*>::iterator p = copy.begin(); p != copy.end(); ++p) { | |
1089 | if (bounds.count(*p) == 0) { | |
1090 | CDir *stray = *p; | |
1091 | dout(10) << " swallowing extra subtree at " << *stray << dendl; | |
1092 | adjust_subtree_auth(stray, auth); | |
224ce89b | 1093 | try_subtree_merge_at(stray, &to_eval); |
7c673cae FG |
1094 | } |
1095 | } | |
1096 | // swallowing subtree may add new subtree bounds | |
1097 | if (copy == subtrees[dir]) | |
1098 | break; | |
1099 | } | |
1100 | ||
1101 | // bound should now match. | |
1102 | verify_subtree_bounds(dir, bounds); | |
1103 | ||
1104 | show_subtrees(); | |
224ce89b WB |
1105 | |
1106 | if (!(mds->is_any_replay() || mds->is_resolve())) { | |
1107 | for(auto in : to_eval) | |
1108 | eval_subtree_root(in); | |
1109 | } | |
7c673cae FG |
1110 | } |
1111 | ||
1112 | ||
1113 | /* | |
1114 | * return a set of CDir*'s that correspond to the given bound set. Only adjust | |
1115 | * fragmentation as necessary to get an equivalent bounding set. That is, only | |
1116 | * split if one of our frags spans the provided bounding set. Never merge. | |
1117 | */ | |
11fdf7f2 | 1118 | void MDCache::get_force_dirfrag_bound_set(const vector<dirfrag_t>& dfs, set<CDir*>& bounds) |
7c673cae FG |
1119 | { |
1120 | dout(10) << "get_force_dirfrag_bound_set " << dfs << dendl; | |
1121 | ||
1122 | // sort by ino | |
1123 | map<inodeno_t, fragset_t> byino; | |
11fdf7f2 TL |
1124 | for (auto& frag : dfs) { |
1125 | byino[frag.ino].insert(frag.frag); | |
1126 | } | |
7c673cae FG |
1127 | dout(10) << " by ino: " << byino << dendl; |
1128 | ||
1129 | for (map<inodeno_t,fragset_t>::iterator p = byino.begin(); p != byino.end(); ++p) { | |
1130 | CInode *diri = get_inode(p->first); | |
1131 | if (!diri) | |
1132 | continue; | |
1133 | dout(10) << " checking fragset " << p->second.get() << " on " << *diri << dendl; | |
1134 | ||
1135 | fragtree_t tmpdft; | |
1136 | for (set<frag_t>::iterator q = p->second.begin(); q != p->second.end(); ++q) | |
1137 | tmpdft.force_to_leaf(g_ceph_context, *q); | |
1138 | ||
11fdf7f2 TL |
1139 | for (const auto& fg : p->second) { |
1140 | frag_vec_t leaves; | |
1141 | diri->dirfragtree.get_leaves_under(fg, leaves); | |
1142 | if (leaves.empty()) { | |
7c673cae FG |
1143 | bool all = true; |
1144 | frag_t approx_fg = diri->dirfragtree[fg.value()]; | |
11fdf7f2 TL |
1145 | frag_vec_t approx_leaves; |
1146 | tmpdft.get_leaves_under(approx_fg, approx_leaves); | |
1147 | for (const auto& leaf : approx_leaves) { | |
1148 | if (p->second.get().count(leaf) == 0) { | |
7c673cae | 1149 | // not bound, so the resolve message is from auth MDS of the dirfrag |
11fdf7f2 | 1150 | force_dir_fragment(diri, leaf); |
7c673cae FG |
1151 | all = false; |
1152 | } | |
1153 | } | |
1154 | if (all) | |
11fdf7f2 | 1155 | leaves.push_back(approx_fg); |
7c673cae | 1156 | else |
11fdf7f2 | 1157 | diri->dirfragtree.get_leaves_under(fg, leaves); |
7c673cae | 1158 | } |
11fdf7f2 TL |
1159 | dout(10) << " frag " << fg << " contains " << leaves << dendl; |
1160 | for (const auto& leaf : leaves) { | |
1161 | CDir *dir = diri->get_dirfrag(leaf); | |
7c673cae FG |
1162 | if (dir) |
1163 | bounds.insert(dir); | |
1164 | } | |
1165 | } | |
1166 | } | |
1167 | } | |
1168 | ||
11fdf7f2 | 1169 | void MDCache::adjust_bounded_subtree_auth(CDir *dir, const vector<dirfrag_t>& bound_dfs, const mds_authority_t &auth) |
7c673cae FG |
1170 | { |
1171 | dout(7) << "adjust_bounded_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
1172 | << " on " << *dir << " bound_dfs " << bound_dfs << dendl; | |
1173 | ||
1174 | set<CDir*> bounds; | |
1175 | get_force_dirfrag_bound_set(bound_dfs, bounds); | |
1176 | adjust_bounded_subtree_auth(dir, bounds, auth); | |
1177 | } | |
1178 | ||
11fdf7f2 | 1179 | void MDCache::map_dirfrag_set(const list<dirfrag_t>& dfs, set<CDir*>& result) |
7c673cae FG |
1180 | { |
1181 | dout(10) << "map_dirfrag_set " << dfs << dendl; | |
1182 | ||
1183 | // group by inode | |
1184 | map<inodeno_t, fragset_t> ino_fragset; | |
11fdf7f2 TL |
1185 | for (const auto &df : dfs) { |
1186 | ino_fragset[df.ino].insert(df.frag); | |
1187 | } | |
7c673cae FG |
1188 | |
1189 | // get frags | |
1190 | for (map<inodeno_t, fragset_t>::iterator p = ino_fragset.begin(); | |
1191 | p != ino_fragset.end(); | |
1192 | ++p) { | |
1193 | CInode *in = get_inode(p->first); | |
1194 | if (!in) | |
1195 | continue; | |
1196 | ||
11fdf7f2 TL |
1197 | frag_vec_t fgs; |
1198 | for (const auto& fg : p->second) { | |
1199 | in->dirfragtree.get_leaves_under(fg, fgs); | |
1200 | } | |
7c673cae | 1201 | |
11fdf7f2 | 1202 | dout(15) << "map_dirfrag_set " << p->second << " -> " << fgs |
7c673cae FG |
1203 | << " on " << *in << dendl; |
1204 | ||
11fdf7f2 TL |
1205 | for (const auto& fg : fgs) { |
1206 | CDir *dir = in->get_dirfrag(fg); | |
7c673cae FG |
1207 | if (dir) |
1208 | result.insert(dir); | |
1209 | } | |
1210 | } | |
1211 | } | |
1212 | ||
1213 | ||
1214 | ||
1215 | CDir *MDCache::get_subtree_root(CDir *dir) | |
1216 | { | |
1217 | // find the underlying dir that delegates (or is about to delegate) auth | |
1218 | while (true) { | |
1219 | if (dir->is_subtree_root()) | |
1220 | return dir; | |
1221 | dir = dir->get_inode()->get_parent_dir(); | |
1222 | if (!dir) | |
1223 | return 0; // none | |
1224 | } | |
1225 | } | |
1226 | ||
1227 | CDir *MDCache::get_projected_subtree_root(CDir *dir) | |
1228 | { | |
1229 | // find the underlying dir that delegates (or is about to delegate) auth | |
1230 | while (true) { | |
1231 | if (dir->is_subtree_root()) | |
1232 | return dir; | |
1233 | dir = dir->get_inode()->get_projected_parent_dir(); | |
1234 | if (!dir) | |
1235 | return 0; // none | |
1236 | } | |
1237 | } | |
1238 | ||
1239 | void MDCache::remove_subtree(CDir *dir) | |
1240 | { | |
1241 | dout(10) << "remove_subtree " << *dir << dendl; | |
11fdf7f2 TL |
1242 | ceph_assert(subtrees.count(dir)); |
1243 | ceph_assert(subtrees[dir].empty()); | |
7c673cae FG |
1244 | subtrees.erase(dir); |
1245 | dir->put(CDir::PIN_SUBTREE); | |
1246 | if (dir->get_parent_dir()) { | |
1247 | CDir *p = get_subtree_root(dir->get_parent_dir()); | |
11fdf7f2 | 1248 | ceph_assert(subtrees[p].count(dir)); |
7c673cae FG |
1249 | subtrees[p].erase(dir); |
1250 | } | |
1251 | } | |
1252 | ||
1253 | void MDCache::get_subtree_bounds(CDir *dir, set<CDir*>& bounds) | |
1254 | { | |
11fdf7f2 | 1255 | ceph_assert(subtrees.count(dir)); |
7c673cae FG |
1256 | bounds = subtrees[dir]; |
1257 | } | |
1258 | ||
1259 | void MDCache::get_wouldbe_subtree_bounds(CDir *dir, set<CDir*>& bounds) | |
1260 | { | |
1261 | if (subtrees.count(dir)) { | |
1262 | // just copy them, dir is a subtree. | |
1263 | get_subtree_bounds(dir, bounds); | |
1264 | } else { | |
1265 | // find them | |
1266 | CDir *root = get_subtree_root(dir); | |
1267 | for (set<CDir*>::iterator p = subtrees[root].begin(); | |
1268 | p != subtrees[root].end(); | |
1269 | ++p) { | |
1270 | CDir *t = *p; | |
1271 | while (t != root) { | |
1272 | t = t->get_parent_dir(); | |
11fdf7f2 | 1273 | ceph_assert(t); |
7c673cae FG |
1274 | if (t == dir) { |
1275 | bounds.insert(*p); | |
1276 | continue; | |
1277 | } | |
1278 | } | |
1279 | } | |
1280 | } | |
1281 | } | |
1282 | ||
1283 | void MDCache::verify_subtree_bounds(CDir *dir, const set<CDir*>& bounds) | |
1284 | { | |
1285 | // for debugging only. | |
11fdf7f2 | 1286 | ceph_assert(subtrees.count(dir)); |
7c673cae FG |
1287 | if (bounds != subtrees[dir]) { |
1288 | dout(0) << "verify_subtree_bounds failed" << dendl; | |
1289 | set<CDir*> b = bounds; | |
1290 | for (auto &cd : subtrees[dir]) { | |
1291 | if (bounds.count(cd)) { | |
1292 | b.erase(cd); | |
1293 | continue; | |
1294 | } | |
1295 | dout(0) << " missing bound " << *cd << dendl; | |
1296 | } | |
1297 | for (const auto &cd : b) | |
1298 | dout(0) << " extra bound " << *cd << dendl; | |
1299 | } | |
11fdf7f2 | 1300 | ceph_assert(bounds == subtrees[dir]); |
7c673cae FG |
1301 | } |
1302 | ||
1303 | void MDCache::verify_subtree_bounds(CDir *dir, const list<dirfrag_t>& bounds) | |
1304 | { | |
1305 | // for debugging only. | |
11fdf7f2 | 1306 | ceph_assert(subtrees.count(dir)); |
7c673cae FG |
1307 | |
1308 | // make sure that any bounds i do have are properly noted as such. | |
1309 | int failed = 0; | |
1310 | for (const auto &fg : bounds) { | |
1311 | CDir *bd = get_dirfrag(fg); | |
1312 | if (!bd) continue; | |
1313 | if (subtrees[dir].count(bd) == 0) { | |
1314 | dout(0) << "verify_subtree_bounds failed: extra bound " << *bd << dendl; | |
1315 | failed++; | |
1316 | } | |
1317 | } | |
11fdf7f2 | 1318 | ceph_assert(failed == 0); |
7c673cae FG |
1319 | } |
1320 | ||
1321 | void MDCache::project_subtree_rename(CInode *diri, CDir *olddir, CDir *newdir) | |
1322 | { | |
1323 | dout(10) << "project_subtree_rename " << *diri << " from " << *olddir | |
1324 | << " to " << *newdir << dendl; | |
1325 | projected_subtree_renames[diri].push_back(pair<CDir*,CDir*>(olddir, newdir)); | |
1326 | } | |
1327 | ||
224ce89b | 1328 | void MDCache::adjust_subtree_after_rename(CInode *diri, CDir *olddir, bool pop) |
7c673cae FG |
1329 | { |
1330 | dout(10) << "adjust_subtree_after_rename " << *diri << " from " << *olddir << dendl; | |
1331 | ||
7c673cae FG |
1332 | CDir *newdir = diri->get_parent_dir(); |
1333 | ||
1334 | if (pop) { | |
1335 | map<CInode*,list<pair<CDir*,CDir*> > >::iterator p = projected_subtree_renames.find(diri); | |
11fdf7f2 TL |
1336 | ceph_assert(p != projected_subtree_renames.end()); |
1337 | ceph_assert(!p->second.empty()); | |
1338 | ceph_assert(p->second.front().first == olddir); | |
1339 | ceph_assert(p->second.front().second == newdir); | |
7c673cae FG |
1340 | p->second.pop_front(); |
1341 | if (p->second.empty()) | |
1342 | projected_subtree_renames.erase(p); | |
1343 | } | |
1344 | ||
11fdf7f2 TL |
1345 | vector<CDir*> dfls; |
1346 | ||
1347 | // adjust total auth pin of freezing subtree | |
1348 | if (olddir != newdir) { | |
1349 | diri->get_nested_dirfrags(dfls); | |
1350 | for (auto dir : dfls) | |
1351 | olddir->adjust_freeze_after_rename(dir); | |
1352 | dfls.clear(); | |
1353 | } | |
1354 | ||
7c673cae | 1355 | // adjust subtree |
7c673cae FG |
1356 | // make sure subtree dirfrags are at the front of the list |
1357 | diri->get_subtree_dirfrags(dfls); | |
1358 | diri->get_nested_dirfrags(dfls); | |
11fdf7f2 | 1359 | for (auto dir : dfls) { |
7c673cae FG |
1360 | dout(10) << "dirfrag " << *dir << dendl; |
1361 | CDir *oldparent = get_subtree_root(olddir); | |
1362 | dout(10) << " old parent " << *oldparent << dendl; | |
1363 | CDir *newparent = get_subtree_root(newdir); | |
1364 | dout(10) << " new parent " << *newparent << dendl; | |
1365 | ||
28e407b8 | 1366 | if (olddir != newdir) |
11fdf7f2 | 1367 | mds->balancer->adjust_pop_for_rename(olddir, dir, false); |
28e407b8 | 1368 | |
7c673cae FG |
1369 | if (oldparent == newparent) { |
1370 | dout(10) << "parent unchanged for " << *dir << " at " << *oldparent << dendl; | |
28e407b8 | 1371 | } else if (dir->is_subtree_root()) { |
7c673cae FG |
1372 | // children are fine. change parent. |
1373 | dout(10) << "moving " << *dir << " from " << *oldparent << " to " << *newparent << dendl; | |
11fdf7f2 | 1374 | ceph_assert(subtrees[oldparent].count(dir)); |
7c673cae | 1375 | subtrees[oldparent].erase(dir); |
11fdf7f2 | 1376 | ceph_assert(subtrees.count(newparent)); |
7c673cae | 1377 | subtrees[newparent].insert(dir); |
224ce89b | 1378 | // caller is responsible for 'eval diri' |
28e407b8 | 1379 | try_subtree_merge_at(dir, NULL, false); |
7c673cae FG |
1380 | } else { |
1381 | // mid-subtree. | |
1382 | ||
1383 | // see if any old bounds move to the new parent. | |
1384 | list<CDir*> tomove; | |
1385 | for (set<CDir*>::iterator p = subtrees[oldparent].begin(); | |
1386 | p != subtrees[oldparent].end(); | |
1387 | ++p) { | |
1388 | CDir *bound = *p; | |
1389 | CDir *broot = get_subtree_root(bound->get_parent_dir()); | |
1390 | if (broot != oldparent) { | |
11fdf7f2 | 1391 | ceph_assert(broot == newparent); |
7c673cae FG |
1392 | tomove.push_back(bound); |
1393 | } | |
1394 | } | |
1395 | for (list<CDir*>::iterator p = tomove.begin(); p != tomove.end(); ++p) { | |
1396 | CDir *bound = *p; | |
1397 | dout(10) << "moving bound " << *bound << " from " << *oldparent << " to " << *newparent << dendl; | |
1398 | subtrees[oldparent].erase(bound); | |
1399 | subtrees[newparent].insert(bound); | |
1400 | } | |
1401 | ||
1402 | // did auth change? | |
1403 | if (oldparent->authority() != newparent->authority()) { | |
28e407b8 | 1404 | adjust_subtree_auth(dir, oldparent->authority(), false); |
224ce89b | 1405 | // caller is responsible for 'eval diri' |
28e407b8 | 1406 | try_subtree_merge_at(dir, NULL, false); |
7c673cae FG |
1407 | } |
1408 | } | |
28e407b8 AA |
1409 | |
1410 | if (olddir != newdir) | |
11fdf7f2 | 1411 | mds->balancer->adjust_pop_for_rename(newdir, dir, true); |
7c673cae FG |
1412 | } |
1413 | ||
1414 | show_subtrees(); | |
1415 | } | |
1416 | ||
7c673cae FG |
1417 | // =================================== |
1418 | // journal and snap/cow helpers | |
1419 | ||
1420 | ||
1421 | /* | |
1422 | * find first inode in cache that follows given snapid. otherwise, return current. | |
1423 | */ | |
1424 | CInode *MDCache::pick_inode_snap(CInode *in, snapid_t follows) | |
1425 | { | |
1426 | dout(10) << "pick_inode_snap follows " << follows << " on " << *in << dendl; | |
11fdf7f2 | 1427 | ceph_assert(in->last == CEPH_NOSNAP); |
7c673cae | 1428 | |
b32b8144 FG |
1429 | auto p = snap_inode_map.upper_bound(vinodeno_t(in->ino(), follows)); |
1430 | if (p != snap_inode_map.end() && p->second->ino() == in->ino()) { | |
1431 | dout(10) << "pick_inode_snap found " << *p->second << dendl; | |
1432 | in = p->second; | |
7c673cae | 1433 | } |
b32b8144 | 1434 | |
7c673cae FG |
1435 | return in; |
1436 | } | |
1437 | ||
1438 | ||
1439 | /* | |
1440 | * note: i'm currently cheating wrt dirty and inode.version on cow | |
1441 | * items. instead of doing a full dir predirty, i just take the | |
1442 | * original item's version, and set the dirty flag (via | |
1443 | * mutation::add_cow_{inode,dentry}() and mutation::apply(). that | |
1444 | * means a special case in the dir commit clean sweep assertions. | |
1445 | * bah. | |
1446 | */ | |
1447 | CInode *MDCache::cow_inode(CInode *in, snapid_t last) | |
1448 | { | |
11fdf7f2 | 1449 | ceph_assert(last >= in->first); |
7c673cae | 1450 | |
b32b8144 | 1451 | CInode *oldin = new CInode(this, true, in->first, last); |
7c673cae | 1452 | oldin->inode = *in->get_previous_projected_inode(); |
7c673cae | 1453 | oldin->xattrs = *in->get_previous_projected_xattrs(); |
11fdf7f2 | 1454 | oldin->symlink = in->symlink; |
7c673cae FG |
1455 | oldin->inode.trim_client_ranges(last); |
1456 | ||
1457 | if (in->first < in->oldest_snap) | |
1458 | in->oldest_snap = in->first; | |
1459 | ||
1460 | in->first = last+1; | |
1461 | ||
1462 | dout(10) << "cow_inode " << *in << " to " << *oldin << dendl; | |
1463 | add_inode(oldin); | |
1464 | ||
1465 | if (in->last != CEPH_NOSNAP) { | |
1466 | CInode *head_in = get_inode(in->ino()); | |
11fdf7f2 | 1467 | ceph_assert(head_in); |
7c673cae FG |
1468 | if (head_in->split_need_snapflush(oldin, in)) { |
1469 | oldin->client_snap_caps = in->client_snap_caps; | |
94b18763 FG |
1470 | for (const auto &p : in->client_snap_caps) { |
1471 | SimpleLock *lock = oldin->get_lock(p.first); | |
11fdf7f2 | 1472 | ceph_assert(lock); |
94b18763 | 1473 | for (const auto &q : p.second) { |
7c673cae FG |
1474 | oldin->auth_pin(lock); |
1475 | lock->set_state(LOCK_SNAP_SYNC); // gathering | |
1476 | lock->get_wrlock(true); | |
94b18763 | 1477 | (void)q; /* unused */ |
7c673cae FG |
1478 | } |
1479 | } | |
1480 | } | |
1481 | return oldin; | |
1482 | } | |
1483 | ||
b32b8144 FG |
1484 | if (!in->client_caps.empty()) { |
1485 | const set<snapid_t>& snaps = in->find_snaprealm()->get_snaps(); | |
1486 | // clone caps? | |
94b18763 | 1487 | for (auto &p : in->client_caps) { |
b32b8144 | 1488 | client_t client = p.first; |
11fdf7f2 TL |
1489 | Capability *cap = &p.second; |
1490 | int issued = cap->need_snapflush() ? CEPH_CAP_ANY_WR : cap->issued(); | |
b32b8144 FG |
1491 | if ((issued & CEPH_CAP_ANY_WR) && |
1492 | cap->client_follows < last) { | |
1493 | // note in oldin | |
1494 | for (int i = 0; i < num_cinode_locks; i++) { | |
1495 | if (issued & cinode_lock_info[i].wr_caps) { | |
1496 | int lockid = cinode_lock_info[i].lock; | |
1497 | SimpleLock *lock = oldin->get_lock(lockid); | |
11fdf7f2 | 1498 | ceph_assert(lock); |
b32b8144 FG |
1499 | oldin->client_snap_caps[lockid].insert(client); |
1500 | oldin->auth_pin(lock); | |
1501 | lock->set_state(LOCK_SNAP_SYNC); // gathering | |
1502 | lock->get_wrlock(true); | |
1503 | dout(10) << " client." << client << " cap " << ccap_string(issued & cinode_lock_info[i].wr_caps) | |
1504 | << " wrlock lock " << *lock << " on " << *oldin << dendl; | |
1505 | } | |
7c673cae | 1506 | } |
b32b8144 FG |
1507 | cap->client_follows = last; |
1508 | ||
1509 | // we need snapflushes for any intervening snaps | |
1510 | dout(10) << " snaps " << snaps << dendl; | |
1511 | for (auto q = snaps.lower_bound(oldin->first); | |
1512 | q != snaps.end() && *q <= last; | |
1513 | ++q) { | |
1514 | in->add_need_snapflush(oldin, *q, client); | |
1515 | } | |
1516 | } else { | |
1517 | dout(10) << " ignoring client." << client << " cap follows " << cap->client_follows << dendl; | |
7c673cae | 1518 | } |
7c673cae FG |
1519 | } |
1520 | } | |
7c673cae FG |
1521 | return oldin; |
1522 | } | |
1523 | ||
1524 | void MDCache::journal_cow_dentry(MutationImpl *mut, EMetaBlob *metablob, | |
1525 | CDentry *dn, snapid_t follows, | |
1526 | CInode **pcow_inode, CDentry::linkage_t *dnl) | |
1527 | { | |
1528 | if (!dn) { | |
1529 | dout(10) << "journal_cow_dentry got null CDentry, returning" << dendl; | |
1530 | return; | |
1531 | } | |
1532 | dout(10) << "journal_cow_dentry follows " << follows << " on " << *dn << dendl; | |
11fdf7f2 | 1533 | ceph_assert(dn->is_auth()); |
7c673cae FG |
1534 | |
1535 | // nothing to cow on a null dentry, fix caller | |
1536 | if (!dnl) | |
1537 | dnl = dn->get_projected_linkage(); | |
11fdf7f2 | 1538 | ceph_assert(!dnl->is_null()); |
7c673cae | 1539 | |
11fdf7f2 TL |
1540 | CInode *in = dnl->is_primary() ? dnl->get_inode() : NULL; |
1541 | bool cow_head = false; | |
1542 | if (in && in->state_test(CInode::STATE_AMBIGUOUSAUTH)) { | |
1543 | ceph_assert(in->is_frozen_inode()); | |
1544 | cow_head = true; | |
1545 | } | |
1546 | if (in && (in->is_multiversion() || cow_head)) { | |
7c673cae | 1547 | // multiversion inode. |
7c673cae FG |
1548 | SnapRealm *realm = NULL; |
1549 | ||
1550 | if (in->get_projected_parent_dn() != dn) { | |
11fdf7f2 | 1551 | ceph_assert(follows == CEPH_NOSNAP); |
7c673cae | 1552 | realm = dn->dir->inode->find_snaprealm(); |
11fdf7f2 TL |
1553 | snapid_t dir_follows = get_global_snaprealm()->get_newest_seq(); |
1554 | ceph_assert(dir_follows >= realm->get_newest_seq()); | |
7c673cae FG |
1555 | |
1556 | if (dir_follows+1 > dn->first) { | |
1557 | snapid_t oldfirst = dn->first; | |
1558 | dn->first = dir_follows+1; | |
1559 | if (realm->has_snaps_in_range(oldfirst, dir_follows)) { | |
94b18763 | 1560 | CDentry *olddn = dn->dir->add_remote_dentry(dn->get_name(), in->ino(), in->d_type(), |
7c673cae FG |
1561 | oldfirst, dir_follows); |
1562 | olddn->pre_dirty(); | |
1563 | dout(10) << " olddn " << *olddn << dendl; | |
1564 | metablob->add_remote_dentry(olddn, true); | |
1565 | mut->add_cow_dentry(olddn); | |
1566 | // FIXME: adjust link count here? hmm. | |
1567 | ||
1568 | if (dir_follows+1 > in->first) | |
11fdf7f2 | 1569 | in->cow_old_inode(dir_follows, cow_head); |
7c673cae FG |
1570 | } |
1571 | } | |
1572 | ||
11fdf7f2 | 1573 | follows = dir_follows; |
7c673cae FG |
1574 | if (in->snaprealm) { |
1575 | realm = in->snaprealm; | |
11fdf7f2 TL |
1576 | ceph_assert(follows >= realm->get_newest_seq()); |
1577 | } | |
7c673cae FG |
1578 | } else { |
1579 | realm = in->find_snaprealm(); | |
11fdf7f2 TL |
1580 | if (follows == CEPH_NOSNAP) { |
1581 | follows = get_global_snaprealm()->get_newest_seq(); | |
1582 | ceph_assert(follows >= realm->get_newest_seq()); | |
1583 | } | |
7c673cae FG |
1584 | } |
1585 | ||
1586 | // already cloned? | |
1587 | if (follows < in->first) { | |
1588 | dout(10) << "journal_cow_dentry follows " << follows << " < first on " << *in << dendl; | |
1589 | return; | |
1590 | } | |
1591 | ||
1592 | if (!realm->has_snaps_in_range(in->first, follows)) { | |
1593 | dout(10) << "journal_cow_dentry no snapshot follows " << follows << " on " << *in << dendl; | |
1594 | in->first = follows + 1; | |
1595 | return; | |
1596 | } | |
1597 | ||
11fdf7f2 | 1598 | in->cow_old_inode(follows, cow_head); |
7c673cae FG |
1599 | |
1600 | } else { | |
1601 | SnapRealm *realm = dn->dir->inode->find_snaprealm(); | |
11fdf7f2 TL |
1602 | if (follows == CEPH_NOSNAP) { |
1603 | follows = get_global_snaprealm()->get_newest_seq(); | |
1604 | ceph_assert(follows >= realm->get_newest_seq()); | |
1605 | } | |
7c673cae FG |
1606 | |
1607 | // already cloned? | |
1608 | if (follows < dn->first) { | |
1609 | dout(10) << "journal_cow_dentry follows " << follows << " < first on " << *dn << dendl; | |
1610 | return; | |
1611 | } | |
1612 | ||
1613 | // update dn.first before adding old dentry to cdir's map | |
1614 | snapid_t oldfirst = dn->first; | |
1615 | dn->first = follows+1; | |
1616 | ||
7c673cae FG |
1617 | if (!realm->has_snaps_in_range(oldfirst, follows)) { |
1618 | dout(10) << "journal_cow_dentry no snapshot follows " << follows << " on " << *dn << dendl; | |
1619 | if (in) | |
1620 | in->first = follows+1; | |
1621 | return; | |
1622 | } | |
1623 | ||
1624 | dout(10) << " dn " << *dn << dendl; | |
1625 | if (in) { | |
1626 | CInode *oldin = cow_inode(in, follows); | |
1627 | mut->add_cow_inode(oldin); | |
1628 | if (pcow_inode) | |
1629 | *pcow_inode = oldin; | |
11fdf7f2 | 1630 | CDentry *olddn = dn->dir->add_primary_dentry(dn->get_name(), oldin, oldfirst, follows); |
7c673cae FG |
1631 | oldin->inode.version = olddn->pre_dirty(); |
1632 | dout(10) << " olddn " << *olddn << dendl; | |
1633 | bool need_snapflush = !oldin->client_snap_caps.empty(); | |
11fdf7f2 | 1634 | if (need_snapflush) { |
7c673cae | 1635 | mut->ls->open_files.push_back(&oldin->item_open_file); |
11fdf7f2 TL |
1636 | mds->locker->mark_need_snapflush_inode(oldin); |
1637 | } | |
7c673cae FG |
1638 | metablob->add_primary_dentry(olddn, 0, true, false, false, need_snapflush); |
1639 | mut->add_cow_dentry(olddn); | |
1640 | } else { | |
11fdf7f2 | 1641 | ceph_assert(dnl->is_remote()); |
94b18763 | 1642 | CDentry *olddn = dn->dir->add_remote_dentry(dn->get_name(), dnl->get_remote_ino(), dnl->get_remote_d_type(), |
7c673cae FG |
1643 | oldfirst, follows); |
1644 | olddn->pre_dirty(); | |
1645 | dout(10) << " olddn " << *olddn << dendl; | |
1646 | metablob->add_remote_dentry(olddn, true); | |
1647 | mut->add_cow_dentry(olddn); | |
1648 | } | |
1649 | } | |
1650 | } | |
1651 | ||
1652 | ||
1653 | void MDCache::journal_cow_inode(MutationRef& mut, EMetaBlob *metablob, | |
1654 | CInode *in, snapid_t follows, | |
1655 | CInode **pcow_inode) | |
1656 | { | |
1657 | dout(10) << "journal_cow_inode follows " << follows << " on " << *in << dendl; | |
1658 | CDentry *dn = in->get_projected_parent_dn(); | |
1659 | journal_cow_dentry(mut.get(), metablob, dn, follows, pcow_inode); | |
1660 | } | |
1661 | ||
1662 | void MDCache::journal_dirty_inode(MutationImpl *mut, EMetaBlob *metablob, CInode *in, snapid_t follows) | |
1663 | { | |
1664 | if (in->is_base()) { | |
11fdf7f2 | 1665 | metablob->add_root(true, in); |
7c673cae FG |
1666 | } else { |
1667 | if (follows == CEPH_NOSNAP && in->last != CEPH_NOSNAP) | |
1668 | follows = in->first - 1; | |
1669 | CDentry *dn = in->get_projected_parent_dn(); | |
1670 | if (!dn->get_projected_linkage()->is_null()) // no need to cow a null dentry | |
1671 | journal_cow_dentry(mut, metablob, dn, follows); | |
1672 | if (in->get_projected_inode()->is_backtrace_updated()) { | |
1673 | bool dirty_pool = in->get_projected_inode()->layout.pool_id != | |
1674 | in->get_previous_projected_inode()->layout.pool_id; | |
1675 | metablob->add_primary_dentry(dn, in, true, true, dirty_pool); | |
1676 | } else { | |
1677 | metablob->add_primary_dentry(dn, in, true); | |
1678 | } | |
1679 | } | |
1680 | } | |
1681 | ||
1682 | ||
1683 | ||
1684 | // nested --------------------------------------------------------------- | |
1685 | ||
1686 | void MDCache::project_rstat_inode_to_frag(CInode *cur, CDir *parent, snapid_t first, | |
1687 | int linkunlink, SnapRealm *prealm) | |
1688 | { | |
1689 | CDentry *parentdn = cur->get_projected_parent_dn(); | |
94b18763 | 1690 | CInode::mempool_inode *curi = cur->get_projected_inode(); |
7c673cae FG |
1691 | |
1692 | if (cur->first > first) | |
1693 | first = cur->first; | |
1694 | ||
1695 | dout(10) << "projected_rstat_inode_to_frag first " << first << " linkunlink " << linkunlink | |
1696 | << " " << *cur << dendl; | |
1697 | dout(20) << " frag head is [" << parent->first << ",head] " << dendl; | |
1698 | dout(20) << " inode update is [" << first << "," << cur->last << "]" << dendl; | |
1699 | ||
1700 | /* | |
1701 | * FIXME. this incompletely propagates rstats to _old_ parents | |
1702 | * (i.e. shortly after a directory rename). but we need full | |
1703 | * blown hard link backpointers to make this work properly... | |
1704 | */ | |
1705 | snapid_t floor = parentdn->first; | |
1706 | dout(20) << " floor of " << floor << " from parent dn " << *parentdn << dendl; | |
1707 | ||
1708 | if (!prealm) | |
1709 | prealm = parent->inode->find_snaprealm(); | |
1710 | const set<snapid_t> snaps = prealm->get_snaps(); | |
1711 | ||
1712 | if (cur->last != CEPH_NOSNAP) { | |
11fdf7f2 TL |
1713 | ceph_assert(cur->dirty_old_rstats.empty()); |
1714 | set<snapid_t>::const_iterator q = snaps.lower_bound(std::max(first, floor)); | |
7c673cae FG |
1715 | if (q == snaps.end() || *q > cur->last) |
1716 | return; | |
1717 | } | |
1718 | ||
1719 | if (cur->last >= floor) { | |
1720 | bool update = true; | |
1721 | if (cur->state_test(CInode::STATE_AMBIGUOUSAUTH) && cur->is_auth()) { | |
1722 | // rename src inode is not projected in the slave rename prep case. so we should | |
1723 | // avoid updateing the inode. | |
11fdf7f2 TL |
1724 | ceph_assert(linkunlink < 0); |
1725 | ceph_assert(cur->is_frozen_inode()); | |
7c673cae FG |
1726 | update = false; |
1727 | } | |
11fdf7f2 | 1728 | _project_rstat_inode_to_frag(*curi, std::max(first, floor), cur->last, parent, |
7c673cae FG |
1729 | linkunlink, update); |
1730 | } | |
1731 | ||
11fdf7f2 | 1732 | if (g_conf()->mds_snap_rstat) { |
94b18763 FG |
1733 | for (const auto &p : cur->dirty_old_rstats) { |
1734 | auto &old = cur->old_inodes[p]; | |
1735 | snapid_t ofirst = std::max(old.first, floor); | |
1736 | auto it = snaps.lower_bound(ofirst); | |
1737 | if (it == snaps.end() || *it > p) | |
7c673cae | 1738 | continue; |
94b18763 FG |
1739 | if (p >= floor) |
1740 | _project_rstat_inode_to_frag(old.inode, ofirst, p, parent, 0, false); | |
7c673cae FG |
1741 | } |
1742 | } | |
1743 | cur->dirty_old_rstats.clear(); | |
1744 | } | |
1745 | ||
1746 | ||
94b18763 | 1747 | void MDCache::_project_rstat_inode_to_frag(CInode::mempool_inode& inode, snapid_t ofirst, snapid_t last, |
7c673cae FG |
1748 | CDir *parent, int linkunlink, bool update_inode) |
1749 | { | |
1750 | dout(10) << "_project_rstat_inode_to_frag [" << ofirst << "," << last << "]" << dendl; | |
1751 | dout(20) << " inode rstat " << inode.rstat << dendl; | |
1752 | dout(20) << " inode accounted_rstat " << inode.accounted_rstat << dendl; | |
1753 | nest_info_t delta; | |
1754 | if (linkunlink == 0) { | |
1755 | delta.add(inode.rstat); | |
1756 | delta.sub(inode.accounted_rstat); | |
1757 | } else if (linkunlink < 0) { | |
1758 | delta.sub(inode.accounted_rstat); | |
1759 | } else { | |
1760 | delta.add(inode.rstat); | |
1761 | } | |
1762 | dout(20) << " delta " << delta << dendl; | |
1763 | ||
1764 | if (update_inode) | |
1765 | inode.accounted_rstat = inode.rstat; | |
1766 | ||
1767 | while (last >= ofirst) { | |
1768 | /* | |
1769 | * pick fnode version to update. at each iteration, we want to | |
1770 | * pick a segment ending in 'last' to update. split as necessary | |
1771 | * to make that work. then, adjust first up so that we only | |
1772 | * update one segment at a time. then loop to cover the whole | |
1773 | * [ofirst,last] interval. | |
1774 | */ | |
1775 | nest_info_t *prstat; | |
1776 | snapid_t first; | |
1777 | fnode_t *pf = parent->get_projected_fnode(); | |
1778 | if (last == CEPH_NOSNAP) { | |
11fdf7f2 TL |
1779 | if (g_conf()->mds_snap_rstat) |
1780 | first = std::max(ofirst, parent->first); | |
7c673cae FG |
1781 | else |
1782 | first = parent->first; | |
1783 | prstat = &pf->rstat; | |
1784 | dout(20) << " projecting to head [" << first << "," << last << "] " << *prstat << dendl; | |
1785 | ||
1786 | if (first > parent->first && | |
1787 | !(pf->rstat == pf->accounted_rstat)) { | |
1788 | dout(10) << " target snapped and not fully accounted, cow to dirty_old_rstat [" | |
1789 | << parent->first << "," << (first-1) << "] " | |
1790 | << " " << *prstat << "/" << pf->accounted_rstat | |
1791 | << dendl; | |
1792 | parent->dirty_old_rstat[first-1].first = parent->first; | |
1793 | parent->dirty_old_rstat[first-1].rstat = pf->rstat; | |
1794 | parent->dirty_old_rstat[first-1].accounted_rstat = pf->accounted_rstat; | |
1795 | } | |
1796 | parent->first = first; | |
11fdf7f2 | 1797 | } else if (!g_conf()->mds_snap_rstat) { |
7c673cae FG |
1798 | // drop snapshots' rstats |
1799 | break; | |
1800 | } else if (last >= parent->first) { | |
1801 | first = parent->first; | |
1802 | parent->dirty_old_rstat[last].first = first; | |
1803 | parent->dirty_old_rstat[last].rstat = pf->rstat; | |
1804 | parent->dirty_old_rstat[last].accounted_rstat = pf->accounted_rstat; | |
1805 | prstat = &parent->dirty_old_rstat[last].rstat; | |
1806 | dout(10) << " projecting to newly split dirty_old_fnode [" << first << "," << last << "] " | |
1807 | << " " << *prstat << "/" << pf->accounted_rstat << dendl; | |
1808 | } else { | |
1809 | // be careful, dirty_old_rstat is a _sparse_ map. | |
1810 | // sorry, this is ugly. | |
1811 | first = ofirst; | |
1812 | ||
1813 | // find any intersection with last | |
94b18763 FG |
1814 | auto it = parent->dirty_old_rstat.lower_bound(last); |
1815 | if (it == parent->dirty_old_rstat.end()) { | |
7c673cae FG |
1816 | dout(20) << " no dirty_old_rstat with last >= last " << last << dendl; |
1817 | if (!parent->dirty_old_rstat.empty() && parent->dirty_old_rstat.rbegin()->first >= first) { | |
1818 | dout(20) << " last dirty_old_rstat ends at " << parent->dirty_old_rstat.rbegin()->first << dendl; | |
1819 | first = parent->dirty_old_rstat.rbegin()->first+1; | |
1820 | } | |
1821 | } else { | |
94b18763 FG |
1822 | // *it last is >= last |
1823 | if (it->second.first <= last) { | |
1824 | // *it intersects [first,last] | |
1825 | if (it->second.first < first) { | |
1826 | dout(10) << " splitting off left bit [" << it->second.first << "," << first-1 << "]" << dendl; | |
1827 | parent->dirty_old_rstat[first-1] = it->second; | |
1828 | it->second.first = first; | |
7c673cae | 1829 | } |
94b18763 FG |
1830 | if (it->second.first > first) |
1831 | first = it->second.first; | |
1832 | if (last < it->first) { | |
1833 | dout(10) << " splitting off right bit [" << last+1 << "," << it->first << "]" << dendl; | |
1834 | parent->dirty_old_rstat[last] = it->second; | |
1835 | it->second.first = last+1; | |
7c673cae FG |
1836 | } |
1837 | } else { | |
94b18763 FG |
1838 | // *it is to the _right_ of [first,last] |
1839 | it = parent->dirty_old_rstat.lower_bound(first); | |
1840 | // new *it last is >= first | |
1841 | if (it->second.first <= last && // new *it isn't also to the right, and | |
1842 | it->first >= first) { // it intersects our first bit, | |
1843 | dout(10) << " staying to the right of [" << it->second.first << "," << it->first << "]..." << dendl; | |
1844 | first = it->first+1; | |
7c673cae FG |
1845 | } |
1846 | dout(10) << " projecting to new dirty_old_rstat [" << first << "," << last << "]" << dendl; | |
1847 | } | |
1848 | } | |
1849 | dout(20) << " projecting to dirty_old_rstat [" << first << "," << last << "]" << dendl; | |
1850 | parent->dirty_old_rstat[last].first = first; | |
1851 | prstat = &parent->dirty_old_rstat[last].rstat; | |
1852 | } | |
1853 | ||
1854 | // apply | |
1855 | dout(20) << " project to [" << first << "," << last << "] " << *prstat << dendl; | |
11fdf7f2 | 1856 | ceph_assert(last >= first); |
7c673cae FG |
1857 | prstat->add(delta); |
1858 | if (update_inode) | |
1859 | inode.accounted_rstat = inode.rstat; | |
1860 | dout(20) << " result [" << first << "," << last << "] " << *prstat << " " << *parent << dendl; | |
1861 | ||
1862 | last = first-1; | |
1863 | } | |
1864 | } | |
1865 | ||
1866 | void MDCache::project_rstat_frag_to_inode(nest_info_t& rstat, nest_info_t& accounted_rstat, | |
1867 | snapid_t ofirst, snapid_t last, | |
1868 | CInode *pin, bool cow_head) | |
1869 | { | |
1870 | dout(10) << "project_rstat_frag_to_inode [" << ofirst << "," << last << "]" << dendl; | |
1871 | dout(20) << " frag rstat " << rstat << dendl; | |
1872 | dout(20) << " frag accounted_rstat " << accounted_rstat << dendl; | |
1873 | nest_info_t delta = rstat; | |
1874 | delta.sub(accounted_rstat); | |
1875 | dout(20) << " delta " << delta << dendl; | |
1876 | ||
1877 | while (last >= ofirst) { | |
94b18763 | 1878 | CInode::mempool_inode *pi; |
7c673cae FG |
1879 | snapid_t first; |
1880 | if (last == pin->last) { | |
1881 | pi = pin->get_projected_inode(); | |
11fdf7f2 | 1882 | first = std::max(ofirst, pin->first); |
7c673cae | 1883 | if (first > pin->first) { |
94b18763 | 1884 | auto &old = pin->cow_old_inode(first-1, cow_head); |
7c673cae FG |
1885 | dout(20) << " cloned old_inode rstat is " << old.inode.rstat << dendl; |
1886 | } | |
1887 | } else { | |
1888 | if (last >= pin->first) { | |
1889 | first = pin->first; | |
1890 | pin->cow_old_inode(last, cow_head); | |
1891 | } else { | |
1892 | // our life is easier here because old_inodes is not sparse | |
1893 | // (although it may not begin at snapid 1) | |
94b18763 FG |
1894 | auto it = pin->old_inodes.lower_bound(last); |
1895 | if (it == pin->old_inodes.end()) { | |
7c673cae FG |
1896 | dout(10) << " no old_inode <= " << last << ", done." << dendl; |
1897 | break; | |
1898 | } | |
94b18763 | 1899 | first = it->second.first; |
7c673cae | 1900 | if (first > last) { |
94b18763 | 1901 | dout(10) << " oldest old_inode is [" << first << "," << it->first << "], done." << dendl; |
7c673cae FG |
1902 | //assert(p == pin->old_inodes.begin()); |
1903 | break; | |
1904 | } | |
94b18763 FG |
1905 | if (it->first > last) { |
1906 | dout(10) << " splitting right old_inode [" << first << "," << it->first << "] to [" | |
1907 | << (last+1) << "," << it->first << "]" << dendl; | |
1908 | pin->old_inodes[last] = it->second; | |
1909 | it->second.first = last+1; | |
1910 | pin->dirty_old_rstats.insert(it->first); | |
7c673cae FG |
1911 | } |
1912 | } | |
1913 | if (first < ofirst) { | |
1914 | dout(10) << " splitting left old_inode [" << first << "," << last << "] to [" | |
1915 | << first << "," << ofirst-1 << "]" << dendl; | |
1916 | pin->old_inodes[ofirst-1] = pin->old_inodes[last]; | |
1917 | pin->dirty_old_rstats.insert(ofirst-1); | |
1918 | pin->old_inodes[last].first = first = ofirst; | |
1919 | } | |
1920 | pi = &pin->old_inodes[last].inode; | |
1921 | pin->dirty_old_rstats.insert(last); | |
1922 | } | |
1923 | dout(20) << " projecting to [" << first << "," << last << "] " << pi->rstat << dendl; | |
1924 | pi->rstat.add(delta); | |
1925 | dout(20) << " result [" << first << "," << last << "] " << pi->rstat << dendl; | |
1926 | ||
1927 | last = first-1; | |
1928 | } | |
1929 | } | |
1930 | ||
a8e16298 | 1931 | void MDCache::broadcast_quota_to_client(CInode *in, client_t exclude_ct, bool quota_change) |
7c673cae | 1932 | { |
11fdf7f2 TL |
1933 | if (!(mds->is_active() || mds->is_stopping())) |
1934 | return; | |
1935 | ||
7c673cae FG |
1936 | if (!in->is_auth() || in->is_frozen()) |
1937 | return; | |
1938 | ||
94b18763 | 1939 | auto i = in->get_projected_inode(); |
a8e16298 TL |
1940 | |
1941 | if (!i->quota.is_enable() && | |
1942 | !quota_change) | |
7c673cae FG |
1943 | return; |
1944 | ||
11fdf7f2 TL |
1945 | // creaete snaprealm for quota inode (quota was set before mimic) |
1946 | if (!in->get_projected_srnode()) | |
1947 | mds->server->create_quota_realm(in); | |
7c673cae | 1948 | |
11fdf7f2 TL |
1949 | for (auto &p : in->client_caps) { |
1950 | Capability *cap = &p.second; | |
1951 | if (cap->is_noquota()) | |
1952 | continue; | |
28e407b8 | 1953 | |
11fdf7f2 | 1954 | if (exclude_ct >= 0 && exclude_ct != p.first) |
28e407b8 AA |
1955 | goto update; |
1956 | ||
7c673cae FG |
1957 | if (cap->last_rbytes == i->rstat.rbytes && |
1958 | cap->last_rsize == i->rstat.rsize()) | |
1959 | continue; | |
1960 | ||
1961 | if (i->quota.max_files > 0) { | |
1962 | if (i->rstat.rsize() >= i->quota.max_files) | |
1963 | goto update; | |
1964 | ||
1965 | if ((abs(cap->last_rsize - i->quota.max_files) >> 4) < | |
1966 | abs(cap->last_rsize - i->rstat.rsize())) | |
1967 | goto update; | |
1968 | } | |
1969 | ||
1970 | if (i->quota.max_bytes > 0) { | |
1971 | if (i->rstat.rbytes > i->quota.max_bytes - (i->quota.max_bytes >> 3)) | |
1972 | goto update; | |
1973 | ||
1974 | if ((abs(cap->last_rbytes - i->quota.max_bytes) >> 4) < | |
1975 | abs(cap->last_rbytes - i->rstat.rbytes)) | |
1976 | goto update; | |
1977 | } | |
1978 | ||
1979 | continue; | |
1980 | ||
1981 | update: | |
1982 | cap->last_rsize = i->rstat.rsize(); | |
1983 | cap->last_rbytes = i->rstat.rbytes; | |
1984 | ||
11fdf7f2 | 1985 | auto msg = MClientQuota::create(); |
7c673cae FG |
1986 | msg->ino = in->ino(); |
1987 | msg->rstat = i->rstat; | |
1988 | msg->quota = i->quota; | |
11fdf7f2 | 1989 | mds->send_message_client_counted(msg, cap->get_session()); |
7c673cae | 1990 | } |
181888fb | 1991 | for (const auto &it : in->get_replicas()) { |
11fdf7f2 | 1992 | auto msg = MGatherCaps::create(); |
7c673cae | 1993 | msg->ino = in->ino(); |
181888fb | 1994 | mds->send_message_mds(msg, it.first); |
7c673cae FG |
1995 | } |
1996 | } | |
1997 | ||
1998 | /* | |
1999 | * NOTE: we _have_ to delay the scatter if we are called during a | |
2000 | * rejoin, because we can't twiddle locks between when the | |
2001 | * rejoin_(weak|strong) is received and when we send the rejoin_ack. | |
2002 | * normally, this isn't a problem: a recover mds doesn't twiddle locks | |
2003 | * (no requests), and a survivor acks immediately. _except_ that | |
2004 | * during rejoin_(weak|strong) processing, we may complete a lock | |
2005 | * gather, and do a scatter_writebehind.. and we _can't_ twiddle the | |
2006 | * scatterlock state in that case or the lock states will get out of | |
2007 | * sync between the auth and replica. | |
2008 | * | |
2009 | * the simple solution is to never do the scatter here. instead, put | |
2010 | * the scatterlock on a list if it isn't already wrlockable. this is | |
2011 | * probably the best plan anyway, since we avoid too many | |
2012 | * scatters/locks under normal usage. | |
2013 | */ | |
2014 | /* | |
2015 | * some notes on dirlock/nestlock scatterlock semantics: | |
2016 | * | |
2017 | * the fragstat (dirlock) will never be updated without | |
2018 | * dirlock+nestlock wrlock held by the caller. | |
2019 | * | |
2020 | * the rstat (nestlock) _may_ get updated without a wrlock when nested | |
2021 | * data is pushed up the tree. this could be changed with some | |
2022 | * restructuring here, but in its current form we ensure that the | |
2023 | * fragstat+rstat _always_ reflect an accurrate summation over the dir | |
2024 | * frag, which is nice. and, we only need to track frags that need to | |
2025 | * be nudged (and not inodes with pending rstat changes that need to | |
2026 | * be pushed into the frag). a consequence of this is that the | |
2027 | * accounted_rstat on scatterlock sync may not match our current | |
2028 | * rstat. this is normal and expected. | |
2029 | */ | |
2030 | void MDCache::predirty_journal_parents(MutationRef mut, EMetaBlob *blob, | |
2031 | CInode *in, CDir *parent, | |
2032 | int flags, int linkunlink, | |
2033 | snapid_t cfollows) | |
2034 | { | |
2035 | bool primary_dn = flags & PREDIRTY_PRIMARY; | |
2036 | bool do_parent_mtime = flags & PREDIRTY_DIR; | |
2037 | bool shallow = flags & PREDIRTY_SHALLOW; | |
2038 | ||
11fdf7f2 | 2039 | ceph_assert(mds->mdlog->entry_is_open()); |
7c673cae FG |
2040 | |
2041 | // make sure stamp is set | |
2042 | if (mut->get_mds_stamp() == utime_t()) | |
2043 | mut->set_mds_stamp(ceph_clock_now()); | |
2044 | ||
2045 | if (in->is_base()) | |
2046 | return; | |
2047 | ||
2048 | dout(10) << "predirty_journal_parents" | |
2049 | << (do_parent_mtime ? " do_parent_mtime":"") | |
2050 | << " linkunlink=" << linkunlink | |
2051 | << (primary_dn ? " primary_dn":" remote_dn") | |
2052 | << (shallow ? " SHALLOW":"") | |
2053 | << " follows " << cfollows | |
2054 | << " " << *in << dendl; | |
2055 | ||
2056 | if (!parent) { | |
11fdf7f2 | 2057 | ceph_assert(primary_dn); |
7c673cae FG |
2058 | parent = in->get_projected_parent_dn()->get_dir(); |
2059 | } | |
2060 | ||
2061 | if (flags == 0 && linkunlink == 0) { | |
2062 | dout(10) << " no flags/linkunlink, just adding dir context to blob(s)" << dendl; | |
2063 | blob->add_dir_context(parent); | |
2064 | return; | |
2065 | } | |
2066 | ||
2067 | // build list of inodes to wrlock, dirty, and update | |
2068 | list<CInode*> lsi; | |
2069 | CInode *cur = in; | |
2070 | CDentry *parentdn = NULL; | |
2071 | bool first = true; | |
2072 | while (parent) { | |
2073 | //assert(cur->is_auth() || !primary_dn); // this breaks the rename auth twiddle hack | |
11fdf7f2 | 2074 | ceph_assert(parent->is_auth()); |
7c673cae FG |
2075 | |
2076 | // opportunistically adjust parent dirfrag | |
2077 | CInode *pin = parent->get_inode(); | |
2078 | ||
2079 | // inode -> dirfrag | |
2080 | mut->auth_pin(parent); | |
2081 | mut->add_projected_fnode(parent); | |
2082 | ||
2083 | fnode_t *pf = parent->project_fnode(); | |
2084 | pf->version = parent->pre_dirty(); | |
2085 | ||
2086 | if (do_parent_mtime || linkunlink) { | |
11fdf7f2 TL |
2087 | ceph_assert(mut->is_wrlocked(&pin->filelock)); |
2088 | ceph_assert(mut->is_wrlocked(&pin->nestlock)); | |
2089 | ceph_assert(cfollows == CEPH_NOSNAP); | |
7c673cae FG |
2090 | |
2091 | // update stale fragstat/rstat? | |
2092 | parent->resync_accounted_fragstat(); | |
2093 | parent->resync_accounted_rstat(); | |
2094 | ||
2095 | if (do_parent_mtime) { | |
2096 | pf->fragstat.mtime = mut->get_op_stamp(); | |
2097 | pf->fragstat.change_attr++; | |
2098 | dout(10) << "predirty_journal_parents bumping change_attr to " << pf->fragstat.change_attr << " on " << parent << dendl; | |
2099 | if (pf->fragstat.mtime > pf->rstat.rctime) { | |
2100 | dout(10) << "predirty_journal_parents updating mtime on " << *parent << dendl; | |
2101 | pf->rstat.rctime = pf->fragstat.mtime; | |
2102 | } else { | |
2103 | dout(10) << "predirty_journal_parents updating mtime UNDERWATER on " << *parent << dendl; | |
2104 | } | |
2105 | } | |
2106 | if (linkunlink) { | |
2107 | dout(10) << "predirty_journal_parents updating size on " << *parent << dendl; | |
2108 | if (in->is_dir()) { | |
2109 | pf->fragstat.nsubdirs += linkunlink; | |
2110 | //pf->rstat.rsubdirs += linkunlink; | |
2111 | } else { | |
2112 | pf->fragstat.nfiles += linkunlink; | |
2113 | //pf->rstat.rfiles += linkunlink; | |
2114 | } | |
2115 | } | |
2116 | } | |
2117 | ||
2118 | // rstat | |
2119 | if (!primary_dn) { | |
2120 | // don't update parent this pass | |
2121 | } else if (!linkunlink && !(pin->nestlock.can_wrlock(-1) && | |
2122 | pin->versionlock.can_wrlock())) { | |
2123 | dout(20) << " unwritable parent nestlock " << pin->nestlock | |
2124 | << ", marking dirty rstat on " << *cur << dendl; | |
2125 | cur->mark_dirty_rstat(); | |
2126 | } else { | |
2127 | // if we don't hold a wrlock reference on this nestlock, take one, | |
2128 | // because we are about to write into the dirfrag fnode and that needs | |
2129 | // to commit before the lock can cycle. | |
2130 | if (linkunlink) { | |
11fdf7f2 | 2131 | ceph_assert(pin->nestlock.get_num_wrlocks() || mut->is_slave()); |
7c673cae FG |
2132 | } |
2133 | ||
11fdf7f2 | 2134 | if (!mut->is_wrlocked(&pin->nestlock)) { |
7c673cae FG |
2135 | dout(10) << " taking wrlock on " << pin->nestlock << " on " << *pin << dendl; |
2136 | mds->locker->wrlock_force(&pin->nestlock, mut); | |
2137 | } | |
2138 | ||
2139 | // now we can project the inode rstat diff the dirfrag | |
2140 | SnapRealm *prealm = pin->find_snaprealm(); | |
2141 | ||
2142 | snapid_t follows = cfollows; | |
2143 | if (follows == CEPH_NOSNAP) | |
2144 | follows = prealm->get_newest_seq(); | |
2145 | ||
2146 | snapid_t first = follows+1; | |
2147 | ||
2148 | // first, if the frag is stale, bring it back in sync. | |
2149 | parent->resync_accounted_rstat(); | |
2150 | ||
2151 | // now push inode rstats into frag | |
2152 | project_rstat_inode_to_frag(cur, parent, first, linkunlink, prealm); | |
2153 | cur->clear_dirty_rstat(); | |
2154 | } | |
2155 | ||
2156 | bool stop = false; | |
2157 | if (!pin->is_auth() || (!mut->is_auth_pinned(pin) && !pin->can_auth_pin())) { | |
2158 | dout(10) << "predirty_journal_parents !auth or ambig or can't authpin on " << *pin << dendl; | |
2159 | stop = true; | |
2160 | } | |
2161 | ||
2162 | // delay propagating until later? | |
2163 | if (!stop && !first && | |
11fdf7f2 | 2164 | g_conf()->mds_dirstat_min_interval > 0) { |
7c673cae | 2165 | double since_last_prop = mut->get_mds_stamp() - pin->last_dirstat_prop; |
11fdf7f2 | 2166 | if (since_last_prop < g_conf()->mds_dirstat_min_interval) { |
7c673cae | 2167 | dout(10) << "predirty_journal_parents last prop " << since_last_prop |
11fdf7f2 | 2168 | << " < " << g_conf()->mds_dirstat_min_interval |
7c673cae FG |
2169 | << ", stopping" << dendl; |
2170 | stop = true; | |
2171 | } else { | |
2172 | dout(10) << "predirty_journal_parents last prop " << since_last_prop << " ago, continuing" << dendl; | |
2173 | } | |
2174 | } | |
2175 | ||
2176 | // can cast only because i'm passing nowait=true in the sole user | |
2177 | MDRequestRef mdmut = static_cast<MDRequestImpl*>(mut.get()); | |
2178 | if (!stop && | |
11fdf7f2 | 2179 | !mut->is_wrlocked(&pin->nestlock) && |
7c673cae FG |
2180 | (!pin->versionlock.can_wrlock() || // make sure we can take versionlock, too |
2181 | //true | |
2182 | !mds->locker->wrlock_start(&pin->nestlock, mdmut, true) | |
2183 | )) { // ** do not initiate.. see above comment ** | |
2184 | dout(10) << "predirty_journal_parents can't wrlock one of " << pin->versionlock << " or " << pin->nestlock | |
2185 | << " on " << *pin << dendl; | |
2186 | stop = true; | |
2187 | } | |
2188 | if (stop) { | |
2189 | dout(10) << "predirty_journal_parents stop. marking nestlock on " << *pin << dendl; | |
2190 | mds->locker->mark_updated_scatterlock(&pin->nestlock); | |
2191 | mut->ls->dirty_dirfrag_nest.push_back(&pin->item_dirty_dirfrag_nest); | |
2192 | mut->add_updated_lock(&pin->nestlock); | |
2193 | if (do_parent_mtime || linkunlink) { | |
2194 | mds->locker->mark_updated_scatterlock(&pin->filelock); | |
2195 | mut->ls->dirty_dirfrag_dir.push_back(&pin->item_dirty_dirfrag_dir); | |
2196 | mut->add_updated_lock(&pin->filelock); | |
2197 | } | |
2198 | break; | |
2199 | } | |
11fdf7f2 | 2200 | if (!mut->is_wrlocked(&pin->versionlock)) |
7c673cae FG |
2201 | mds->locker->local_wrlock_grab(&pin->versionlock, mut); |
2202 | ||
11fdf7f2 | 2203 | ceph_assert(mut->is_wrlocked(&pin->nestlock) || mut->is_slave()); |
7c673cae FG |
2204 | |
2205 | pin->last_dirstat_prop = mut->get_mds_stamp(); | |
2206 | ||
2207 | // dirfrag -> diri | |
2208 | mut->auth_pin(pin); | |
2209 | mut->add_projected_inode(pin); | |
2210 | lsi.push_front(pin); | |
2211 | ||
2212 | pin->pre_cow_old_inode(); // avoid cow mayhem! | |
2213 | ||
94b18763 FG |
2214 | auto &pi = pin->project_inode(); |
2215 | pi.inode.version = pin->pre_dirty(); | |
7c673cae FG |
2216 | |
2217 | // dirstat | |
2218 | if (do_parent_mtime || linkunlink) { | |
2219 | dout(20) << "predirty_journal_parents add_delta " << pf->fragstat << dendl; | |
2220 | dout(20) << "predirty_journal_parents - " << pf->accounted_fragstat << dendl; | |
2221 | bool touched_mtime = false, touched_chattr = false; | |
94b18763 | 2222 | pi.inode.dirstat.add_delta(pf->fragstat, pf->accounted_fragstat, &touched_mtime, &touched_chattr); |
7c673cae FG |
2223 | pf->accounted_fragstat = pf->fragstat; |
2224 | if (touched_mtime) | |
94b18763 | 2225 | pi.inode.mtime = pi.inode.ctime = pi.inode.dirstat.mtime; |
7c673cae | 2226 | if (touched_chattr) |
94b18763 FG |
2227 | pi.inode.change_attr = pi.inode.dirstat.change_attr; |
2228 | dout(20) << "predirty_journal_parents gives " << pi.inode.dirstat << " on " << *pin << dendl; | |
7c673cae FG |
2229 | |
2230 | if (parent->get_frag() == frag_t()) { // i.e., we are the only frag | |
94b18763 | 2231 | if (pi.inode.dirstat.size() < 0) |
11fdf7f2 | 2232 | ceph_assert(!"negative dirstat size" == g_conf()->mds_verify_scatter); |
94b18763 | 2233 | if (pi.inode.dirstat.size() != pf->fragstat.size()) { |
7c673cae | 2234 | mds->clog->error() << "unmatched fragstat size on single dirfrag " |
94b18763 | 2235 | << parent->dirfrag() << ", inode has " << pi.inode.dirstat |
7c673cae FG |
2236 | << ", dirfrag has " << pf->fragstat; |
2237 | ||
2238 | // trust the dirfrag for now | |
94b18763 | 2239 | pi.inode.dirstat = pf->fragstat; |
7c673cae | 2240 | |
11fdf7f2 | 2241 | ceph_assert(!"unmatched fragstat size" == g_conf()->mds_verify_scatter); |
7c673cae FG |
2242 | } |
2243 | } | |
2244 | } | |
2245 | ||
2246 | /* | |
2247 | * the rule here is to follow the _oldest_ parent with dirty rstat | |
2248 | * data. if we don't propagate all data, we add ourselves to the | |
2249 | * nudge list. that way all rstat data will (eventually) get | |
2250 | * pushed up the tree. | |
2251 | * | |
2252 | * actually, no. for now, silently drop rstats for old parents. we need | |
2253 | * hard link backpointers to do the above properly. | |
2254 | */ | |
2255 | ||
2256 | // stop? | |
2257 | if (pin->is_base()) | |
2258 | break; | |
2259 | parentdn = pin->get_projected_parent_dn(); | |
11fdf7f2 | 2260 | ceph_assert(parentdn); |
7c673cae FG |
2261 | |
2262 | // rstat | |
2263 | dout(10) << "predirty_journal_parents frag->inode on " << *parent << dendl; | |
2264 | ||
2265 | // first, if the frag is stale, bring it back in sync. | |
2266 | parent->resync_accounted_rstat(); | |
2267 | ||
11fdf7f2 | 2268 | if (g_conf()->mds_snap_rstat) { |
94b18763 FG |
2269 | for (auto &p : parent->dirty_old_rstat) { |
2270 | project_rstat_frag_to_inode(p.second.rstat, p.second.accounted_rstat, p.second.first, | |
2271 | p.first, pin, true); | |
2272 | } | |
7c673cae FG |
2273 | } |
2274 | parent->dirty_old_rstat.clear(); | |
2275 | project_rstat_frag_to_inode(pf->rstat, pf->accounted_rstat, parent->first, CEPH_NOSNAP, pin, true);//false); | |
2276 | ||
2277 | pf->accounted_rstat = pf->rstat; | |
2278 | ||
2279 | if (parent->get_frag() == frag_t()) { // i.e., we are the only frag | |
94b18763 | 2280 | if (pi.inode.rstat.rbytes != pf->rstat.rbytes) { |
7c673cae | 2281 | mds->clog->error() << "unmatched rstat rbytes on single dirfrag " |
94b18763 | 2282 | << parent->dirfrag() << ", inode has " << pi.inode.rstat |
7c673cae FG |
2283 | << ", dirfrag has " << pf->rstat; |
2284 | ||
2285 | // trust the dirfrag for now | |
94b18763 | 2286 | pi.inode.rstat = pf->rstat; |
7c673cae | 2287 | |
11fdf7f2 | 2288 | ceph_assert(!"unmatched rstat rbytes" == g_conf()->mds_verify_scatter); |
7c673cae FG |
2289 | } |
2290 | } | |
2291 | ||
2292 | parent->check_rstats(); | |
2293 | broadcast_quota_to_client(pin); | |
2294 | // next parent! | |
2295 | cur = pin; | |
2296 | parent = parentdn->get_dir(); | |
2297 | linkunlink = 0; | |
2298 | do_parent_mtime = false; | |
2299 | primary_dn = true; | |
2300 | first = false; | |
2301 | } | |
2302 | ||
2303 | // now, stick it in the blob | |
11fdf7f2 TL |
2304 | ceph_assert(parent); |
2305 | ceph_assert(parent->is_auth()); | |
7c673cae FG |
2306 | blob->add_dir_context(parent); |
2307 | blob->add_dir(parent, true); | |
2308 | for (list<CInode*>::iterator p = lsi.begin(); | |
2309 | p != lsi.end(); | |
2310 | ++p) { | |
2311 | CInode *cur = *p; | |
2312 | journal_dirty_inode(mut.get(), blob, cur); | |
2313 | } | |
2314 | ||
2315 | } | |
2316 | ||
2317 | ||
2318 | ||
2319 | ||
2320 | ||
2321 | // =================================== | |
2322 | // slave requests | |
2323 | ||
2324 | ||
2325 | /* | |
2326 | * some handlers for master requests with slaves. we need to make | |
2327 | * sure slaves journal commits before we forget we mastered them and | |
2328 | * remove them from the uncommitted_masters map (used during recovery | |
2329 | * to commit|abort slaves). | |
2330 | */ | |
2331 | struct C_MDC_CommittedMaster : public MDCacheLogContext { | |
2332 | metareqid_t reqid; | |
2333 | C_MDC_CommittedMaster(MDCache *s, metareqid_t r) : MDCacheLogContext(s), reqid(r) {} | |
2334 | void finish(int r) override { | |
2335 | mdcache->_logged_master_commit(reqid); | |
2336 | } | |
2337 | }; | |
2338 | ||
2339 | void MDCache::log_master_commit(metareqid_t reqid) | |
2340 | { | |
2341 | dout(10) << "log_master_commit " << reqid << dendl; | |
2342 | uncommitted_masters[reqid].committing = true; | |
2343 | mds->mdlog->start_submit_entry(new ECommitted(reqid), | |
2344 | new C_MDC_CommittedMaster(this, reqid)); | |
2345 | } | |
2346 | ||
2347 | void MDCache::_logged_master_commit(metareqid_t reqid) | |
2348 | { | |
2349 | dout(10) << "_logged_master_commit " << reqid << dendl; | |
11fdf7f2 | 2350 | ceph_assert(uncommitted_masters.count(reqid)); |
7c673cae FG |
2351 | uncommitted_masters[reqid].ls->uncommitted_masters.erase(reqid); |
2352 | mds->queue_waiters(uncommitted_masters[reqid].waiters); | |
2353 | uncommitted_masters.erase(reqid); | |
2354 | } | |
2355 | ||
2356 | // while active... | |
2357 | ||
2358 | void MDCache::committed_master_slave(metareqid_t r, mds_rank_t from) | |
2359 | { | |
2360 | dout(10) << "committed_master_slave mds." << from << " on " << r << dendl; | |
11fdf7f2 | 2361 | ceph_assert(uncommitted_masters.count(r)); |
7c673cae FG |
2362 | uncommitted_masters[r].slaves.erase(from); |
2363 | if (!uncommitted_masters[r].recovering && uncommitted_masters[r].slaves.empty()) | |
2364 | log_master_commit(r); | |
2365 | } | |
2366 | ||
2367 | void MDCache::logged_master_update(metareqid_t reqid) | |
2368 | { | |
2369 | dout(10) << "logged_master_update " << reqid << dendl; | |
11fdf7f2 | 2370 | ceph_assert(uncommitted_masters.count(reqid)); |
7c673cae | 2371 | uncommitted_masters[reqid].safe = true; |
11fdf7f2 TL |
2372 | auto p = pending_masters.find(reqid); |
2373 | if (p != pending_masters.end()) { | |
2374 | pending_masters.erase(p); | |
7c673cae FG |
2375 | if (pending_masters.empty()) |
2376 | process_delayed_resolve(); | |
2377 | } | |
2378 | } | |
2379 | ||
2380 | /* | |
2381 | * Master may crash after receiving all slaves' commit acks, but before journalling | |
2382 | * the final commit. Slaves may crash after journalling the slave commit, but before | |
2383 | * sending commit ack to the master. Commit masters with no uncommitted slave when | |
2384 | * resolve finishes. | |
2385 | */ | |
2386 | void MDCache::finish_committed_masters() | |
2387 | { | |
2388 | for (map<metareqid_t, umaster>::iterator p = uncommitted_masters.begin(); | |
2389 | p != uncommitted_masters.end(); | |
2390 | ++p) { | |
2391 | p->second.recovering = false; | |
2392 | if (!p->second.committing && p->second.slaves.empty()) { | |
2393 | dout(10) << "finish_committed_masters " << p->first << dendl; | |
2394 | log_master_commit(p->first); | |
2395 | } | |
2396 | } | |
2397 | } | |
2398 | ||
2399 | /* | |
2400 | * at end of resolve... we must journal a commit|abort for all slave | |
2401 | * updates, before moving on. | |
2402 | * | |
2403 | * this is so that the master can safely journal ECommitted on ops it | |
2404 | * masters when it reaches up:active (all other recovering nodes must | |
2405 | * complete resolve before that happens). | |
2406 | */ | |
2407 | struct C_MDC_SlaveCommit : public MDCacheLogContext { | |
2408 | mds_rank_t from; | |
2409 | metareqid_t reqid; | |
2410 | C_MDC_SlaveCommit(MDCache *c, int f, metareqid_t r) : MDCacheLogContext(c), from(f), reqid(r) {} | |
2411 | void finish(int r) override { | |
2412 | mdcache->_logged_slave_commit(from, reqid); | |
2413 | } | |
2414 | }; | |
2415 | ||
2416 | void MDCache::_logged_slave_commit(mds_rank_t from, metareqid_t reqid) | |
2417 | { | |
2418 | dout(10) << "_logged_slave_commit from mds." << from << " " << reqid << dendl; | |
2419 | ||
2420 | // send a message | |
11fdf7f2 | 2421 | auto req = MMDSSlaveRequest::create(reqid, 0, MMDSSlaveRequest::OP_COMMITTED); |
7c673cae FG |
2422 | mds->send_message_mds(req, from); |
2423 | } | |
2424 | ||
2425 | ||
2426 | ||
2427 | ||
2428 | ||
2429 | ||
2430 | // ==================================================================== | |
2431 | // import map, recovery | |
2432 | ||
2433 | void MDCache::_move_subtree_map_bound(dirfrag_t df, dirfrag_t oldparent, dirfrag_t newparent, | |
2434 | map<dirfrag_t,vector<dirfrag_t> >& subtrees) | |
2435 | { | |
2436 | if (subtrees.count(oldparent)) { | |
2437 | vector<dirfrag_t>& v = subtrees[oldparent]; | |
2438 | dout(10) << " removing " << df << " from " << oldparent << " bounds " << v << dendl; | |
2439 | for (vector<dirfrag_t>::iterator it = v.begin(); it != v.end(); ++it) | |
2440 | if (*it == df) { | |
2441 | v.erase(it); | |
2442 | break; | |
2443 | } | |
2444 | } | |
2445 | if (subtrees.count(newparent)) { | |
2446 | vector<dirfrag_t>& v = subtrees[newparent]; | |
2447 | dout(10) << " adding " << df << " to " << newparent << " bounds " << v << dendl; | |
2448 | v.push_back(df); | |
2449 | } | |
2450 | } | |
2451 | ||
2452 | ESubtreeMap *MDCache::create_subtree_map() | |
2453 | { | |
2454 | dout(10) << "create_subtree_map " << num_subtrees() << " subtrees, " | |
2455 | << num_subtrees_fullauth() << " fullauth" | |
2456 | << dendl; | |
2457 | ||
2458 | show_subtrees(); | |
2459 | ||
2460 | ESubtreeMap *le = new ESubtreeMap(); | |
2461 | mds->mdlog->_start_entry(le); | |
2462 | ||
2463 | map<dirfrag_t, CDir*> dirs_to_add; | |
2464 | ||
2465 | if (myin) { | |
2466 | CDir* mydir = myin->get_dirfrag(frag_t()); | |
2467 | dirs_to_add[mydir->dirfrag()] = mydir; | |
2468 | } | |
2469 | ||
2470 | // include all auth subtrees, and their bounds. | |
2471 | // and a spanning tree to tie it to the root. | |
2472 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
2473 | p != subtrees.end(); | |
2474 | ++p) { | |
2475 | CDir *dir = p->first; | |
2476 | ||
2477 | // journal subtree as "ours" if we are | |
2478 | // me, -2 | |
2479 | // me, me | |
2480 | // me, !me (may be importing and ambiguous!) | |
2481 | ||
2482 | // so not | |
2483 | // !me, * | |
2484 | if (dir->get_dir_auth().first != mds->get_nodeid()) | |
2485 | continue; | |
2486 | ||
2487 | if (migrator->is_ambiguous_import(dir->dirfrag()) || | |
2488 | my_ambiguous_imports.count(dir->dirfrag())) { | |
2489 | dout(15) << " ambig subtree " << *dir << dendl; | |
2490 | le->ambiguous_subtrees.insert(dir->dirfrag()); | |
2491 | } else { | |
2492 | dout(15) << " subtree " << *dir << dendl; | |
2493 | } | |
2494 | ||
2495 | dirs_to_add[dir->dirfrag()] = dir; | |
2496 | le->subtrees[dir->dirfrag()].clear(); | |
2497 | ||
2498 | ||
2499 | // bounds | |
2500 | for (set<CDir*>::iterator q = p->second.begin(); | |
2501 | q != p->second.end(); | |
2502 | ++q) { | |
2503 | CDir *bound = *q; | |
2504 | dout(15) << " subtree bound " << *bound << dendl; | |
2505 | dirs_to_add[bound->dirfrag()] = bound; | |
2506 | le->subtrees[dir->dirfrag()].push_back(bound->dirfrag()); | |
2507 | } | |
2508 | } | |
2509 | ||
2510 | // apply projected renames | |
2511 | for (map<CInode*,list<pair<CDir*,CDir*> > >::iterator p = projected_subtree_renames.begin(); | |
2512 | p != projected_subtree_renames.end(); | |
2513 | ++p) { | |
2514 | for (list<pair<CDir*,CDir*> >::iterator q = p->second.begin(); q != p->second.end(); ++q) { | |
2515 | CInode *diri = p->first; | |
2516 | CDir *olddir = q->first; | |
2517 | CDir *newdir = q->second; | |
2518 | dout(10) << " adjusting for projected rename of " << *diri << " to " << *newdir << dendl; | |
2519 | ||
2520 | list<CDir*> dfls; | |
2521 | diri->get_dirfrags(dfls); | |
2522 | for (list<CDir*>::iterator p = dfls.begin(); p != dfls.end(); ++p) { | |
2523 | CDir *dir = *p; | |
2524 | dout(10) << "dirfrag " << dir->dirfrag() << " " << *dir << dendl; | |
2525 | CDir *oldparent = get_projected_subtree_root(olddir); | |
2526 | dout(10) << " old parent " << oldparent->dirfrag() << " " << *oldparent << dendl; | |
2527 | CDir *newparent = get_projected_subtree_root(newdir); | |
2528 | dout(10) << " new parent " << newparent->dirfrag() << " " << *newparent << dendl; | |
2529 | ||
2530 | if (oldparent == newparent) { | |
2531 | dout(10) << "parent unchanged for " << dir->dirfrag() << " at " | |
2532 | << oldparent->dirfrag() << dendl; | |
2533 | continue; | |
2534 | } | |
2535 | ||
2536 | if (dir->is_subtree_root()) { | |
2537 | if (le->subtrees.count(newparent->dirfrag()) && | |
2538 | oldparent->get_dir_auth() != newparent->get_dir_auth()) | |
2539 | dirs_to_add[dir->dirfrag()] = dir; | |
2540 | // children are fine. change parent. | |
2541 | _move_subtree_map_bound(dir->dirfrag(), oldparent->dirfrag(), newparent->dirfrag(), | |
2542 | le->subtrees); | |
2543 | } else { | |
2544 | // mid-subtree. | |
2545 | ||
2546 | if (oldparent->get_dir_auth() != newparent->get_dir_auth()) { | |
2547 | dout(10) << " creating subtree for " << dir->dirfrag() << dendl; | |
2548 | // if oldparent is auth, subtree is mine; include it. | |
2549 | if (le->subtrees.count(oldparent->dirfrag())) { | |
2550 | dirs_to_add[dir->dirfrag()] = dir; | |
2551 | le->subtrees[dir->dirfrag()].clear(); | |
2552 | } | |
2553 | // if newparent is auth, subtree is a new bound | |
2554 | if (le->subtrees.count(newparent->dirfrag())) { | |
2555 | dirs_to_add[dir->dirfrag()] = dir; | |
2556 | le->subtrees[newparent->dirfrag()].push_back(dir->dirfrag()); // newparent is auth; new bound | |
2557 | } | |
2558 | newparent = dir; | |
2559 | } | |
2560 | ||
2561 | // see if any old bounds move to the new parent. | |
2562 | for (set<CDir*>::iterator p = subtrees[oldparent].begin(); | |
2563 | p != subtrees[oldparent].end(); | |
2564 | ++p) { | |
2565 | CDir *bound = *p; | |
2566 | if (dir->contains(bound->get_parent_dir())) | |
2567 | _move_subtree_map_bound(bound->dirfrag(), oldparent->dirfrag(), newparent->dirfrag(), | |
2568 | le->subtrees); | |
2569 | } | |
2570 | } | |
2571 | } | |
2572 | } | |
2573 | } | |
2574 | ||
2575 | // simplify the journaled map. our in memory map may have more | |
2576 | // subtrees than needed due to migrations that are just getting | |
2577 | // started or just completing. but on replay, the "live" map will | |
2578 | // be simple and we can do a straight comparison. | |
2579 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator p = le->subtrees.begin(); p != le->subtrees.end(); ++p) { | |
2580 | if (le->ambiguous_subtrees.count(p->first)) | |
2581 | continue; | |
2582 | unsigned i = 0; | |
2583 | while (i < p->second.size()) { | |
2584 | dirfrag_t b = p->second[i]; | |
2585 | if (le->subtrees.count(b) && | |
2586 | le->ambiguous_subtrees.count(b) == 0) { | |
2587 | vector<dirfrag_t>& bb = le->subtrees[b]; | |
2588 | dout(10) << "simplify: " << p->first << " swallowing " << b << " with bounds " << bb << dendl; | |
2589 | for (vector<dirfrag_t>::iterator r = bb.begin(); r != bb.end(); ++r) | |
2590 | p->second.push_back(*r); | |
2591 | dirs_to_add.erase(b); | |
2592 | le->subtrees.erase(b); | |
2593 | p->second.erase(p->second.begin() + i); | |
2594 | } else { | |
2595 | ++i; | |
2596 | } | |
2597 | } | |
2598 | } | |
2599 | ||
94b18763 | 2600 | for (auto &p : dirs_to_add) { |
7c673cae FG |
2601 | CDir *dir = p.second; |
2602 | le->metablob.add_dir_context(dir, EMetaBlob::TO_ROOT); | |
2603 | le->metablob.add_dir(dir, false); | |
2604 | } | |
2605 | ||
2606 | dout(15) << " subtrees " << le->subtrees << dendl; | |
2607 | dout(15) << " ambiguous_subtrees " << le->ambiguous_subtrees << dendl; | |
2608 | ||
2609 | //le->metablob.print(cout); | |
2610 | le->expire_pos = mds->mdlog->journaler->get_expire_pos(); | |
2611 | return le; | |
2612 | } | |
2613 | ||
2614 | void MDCache::dump_resolve_status(Formatter *f) const | |
2615 | { | |
2616 | f->open_object_section("resolve_status"); | |
2617 | f->dump_stream("resolve_gather") << resolve_gather; | |
2618 | f->dump_stream("resolve_ack_gather") << resolve_gather; | |
2619 | f->close_section(); | |
2620 | } | |
2621 | ||
11fdf7f2 | 2622 | void MDCache::resolve_start(MDSContext *resolve_done_) |
7c673cae FG |
2623 | { |
2624 | dout(10) << "resolve_start" << dendl; | |
11fdf7f2 | 2625 | ceph_assert(!resolve_done); |
7c673cae FG |
2626 | resolve_done.reset(resolve_done_); |
2627 | ||
2628 | if (mds->mdsmap->get_root() != mds->get_nodeid()) { | |
2629 | // if we don't have the root dir, adjust it to UNKNOWN. during | |
2630 | // resolve we want mds0 to explicit claim the portion of it that | |
2631 | // it owns, so that anything beyond its bounds get left as | |
2632 | // unknown. | |
2633 | CDir *rootdir = root->get_dirfrag(frag_t()); | |
2634 | if (rootdir) | |
2635 | adjust_subtree_auth(rootdir, CDIR_AUTH_UNKNOWN); | |
2636 | } | |
2637 | resolve_gather = recovery_set; | |
11fdf7f2 TL |
2638 | |
2639 | resolve_snapclient_commits = mds->snapclient->get_journaled_tids(); | |
7c673cae FG |
2640 | } |
2641 | ||
2642 | void MDCache::send_resolves() | |
2643 | { | |
2644 | send_slave_resolves(); | |
11fdf7f2 TL |
2645 | |
2646 | if (!resolve_done) { | |
2647 | // I'm survivor: refresh snap cache | |
2648 | mds->snapclient->sync( | |
2649 | new MDSInternalContextWrapper(mds, | |
2650 | new FunctionContext([this](int r) { | |
2651 | maybe_finish_slave_resolve(); | |
2652 | }) | |
2653 | ) | |
2654 | ); | |
2655 | dout(10) << "send_resolves waiting for snapclient cache to sync" << dendl; | |
2656 | return; | |
2657 | } | |
7c673cae FG |
2658 | if (!resolve_ack_gather.empty()) { |
2659 | dout(10) << "send_resolves still waiting for resolve ack from (" | |
2660 | << resolve_ack_gather << ")" << dendl; | |
2661 | return; | |
2662 | } | |
11fdf7f2 | 2663 | if (!resolve_need_rollback.empty()) { |
7c673cae | 2664 | dout(10) << "send_resolves still waiting for rollback to commit on (" |
11fdf7f2 | 2665 | << resolve_need_rollback << ")" << dendl; |
7c673cae FG |
2666 | return; |
2667 | } | |
11fdf7f2 | 2668 | |
7c673cae FG |
2669 | send_subtree_resolves(); |
2670 | } | |
2671 | ||
2672 | void MDCache::send_slave_resolves() | |
2673 | { | |
2674 | dout(10) << "send_slave_resolves" << dendl; | |
2675 | ||
11fdf7f2 | 2676 | map<mds_rank_t, MMDSResolve::ref> resolves; |
7c673cae FG |
2677 | |
2678 | if (mds->is_resolve()) { | |
2679 | for (map<mds_rank_t, map<metareqid_t, MDSlaveUpdate*> >::iterator p = uncommitted_slave_updates.begin(); | |
2680 | p != uncommitted_slave_updates.end(); | |
2681 | ++p) { | |
11fdf7f2 | 2682 | resolves[p->first] = MMDSResolve::create(); |
7c673cae FG |
2683 | for (map<metareqid_t, MDSlaveUpdate*>::iterator q = p->second.begin(); |
2684 | q != p->second.end(); | |
2685 | ++q) { | |
2686 | dout(10) << " including uncommitted " << q->first << dendl; | |
2687 | resolves[p->first]->add_slave_request(q->first, false); | |
2688 | } | |
2689 | } | |
2690 | } else { | |
2691 | set<mds_rank_t> resolve_set; | |
2692 | mds->mdsmap->get_mds_set(resolve_set, MDSMap::STATE_RESOLVE); | |
2693 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
2694 | p != active_requests.end(); | |
2695 | ++p) { | |
2696 | MDRequestRef& mdr = p->second; | |
2697 | if (!mdr->is_slave()) | |
2698 | continue; | |
2699 | if (!mdr->slave_did_prepare() && !mdr->committing) { | |
2700 | continue; | |
2701 | } | |
2702 | mds_rank_t master = mdr->slave_to_mds; | |
2703 | if (resolve_set.count(master) || is_ambiguous_slave_update(p->first, master)) { | |
2704 | dout(10) << " including uncommitted " << *mdr << dendl; | |
2705 | if (!resolves.count(master)) | |
11fdf7f2 | 2706 | resolves[master] = MMDSResolve::create(); |
7c673cae FG |
2707 | if (!mdr->committing && |
2708 | mdr->has_more() && mdr->more()->is_inode_exporter) { | |
2709 | // re-send cap exports | |
2710 | CInode *in = mdr->more()->rename_inode; | |
2711 | map<client_t, Capability::Export> cap_map; | |
2712 | in->export_client_caps(cap_map); | |
2713 | bufferlist bl; | |
11fdf7f2 TL |
2714 | encode(in->ino(), bl); |
2715 | encode(cap_map, bl); | |
7c673cae FG |
2716 | resolves[master]->add_slave_request(p->first, bl); |
2717 | } else { | |
2718 | resolves[master]->add_slave_request(p->first, mdr->committing); | |
2719 | } | |
2720 | } | |
2721 | } | |
2722 | } | |
2723 | ||
11fdf7f2 TL |
2724 | for (auto &p : resolves) { |
2725 | dout(10) << "sending slave resolve to mds." << p.first << dendl; | |
2726 | mds->send_message_mds(p.second, p.first); | |
2727 | resolve_ack_gather.insert(p.first); | |
7c673cae FG |
2728 | } |
2729 | } | |
2730 | ||
2731 | void MDCache::send_subtree_resolves() | |
2732 | { | |
2733 | dout(10) << "send_subtree_resolves" << dendl; | |
2734 | ||
2735 | if (migrator->is_exporting() || migrator->is_importing()) { | |
2736 | dout(7) << "send_subtree_resolves waiting, imports/exports still in progress" << dendl; | |
2737 | migrator->show_importing(); | |
2738 | migrator->show_exporting(); | |
2739 | resolves_pending = true; | |
2740 | return; // not now | |
2741 | } | |
2742 | ||
11fdf7f2 | 2743 | map<mds_rank_t, MMDSResolve::ref> resolves; |
7c673cae FG |
2744 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); |
2745 | p != recovery_set.end(); | |
2746 | ++p) { | |
2747 | if (*p == mds->get_nodeid()) | |
2748 | continue; | |
2749 | if (mds->is_resolve() || mds->mdsmap->is_resolve(*p)) | |
11fdf7f2 | 2750 | resolves[*p] = MMDSResolve::create(); |
7c673cae FG |
2751 | } |
2752 | ||
2753 | map<dirfrag_t, vector<dirfrag_t> > my_subtrees; | |
2754 | map<dirfrag_t, vector<dirfrag_t> > my_ambig_imports; | |
2755 | ||
2756 | // known | |
2757 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
2758 | p != subtrees.end(); | |
2759 | ++p) { | |
2760 | CDir *dir = p->first; | |
2761 | ||
2762 | // only our subtrees | |
2763 | if (dir->authority().first != mds->get_nodeid()) | |
2764 | continue; | |
2765 | ||
2766 | if (mds->is_resolve() && my_ambiguous_imports.count(dir->dirfrag())) | |
2767 | continue; // we'll add it below | |
2768 | ||
2769 | if (migrator->is_ambiguous_import(dir->dirfrag())) { | |
2770 | // ambiguous (mid-import) | |
2771 | set<CDir*> bounds; | |
2772 | get_subtree_bounds(dir, bounds); | |
2773 | vector<dirfrag_t> dfls; | |
2774 | for (set<CDir*>::iterator q = bounds.begin(); q != bounds.end(); ++q) | |
2775 | dfls.push_back((*q)->dirfrag()); | |
2776 | ||
2777 | my_ambig_imports[dir->dirfrag()] = dfls; | |
2778 | dout(10) << " ambig " << dir->dirfrag() << " " << dfls << dendl; | |
2779 | } else { | |
2780 | // not ambiguous. | |
11fdf7f2 TL |
2781 | for (auto &q : resolves) { |
2782 | resolves[q.first]->add_subtree(dir->dirfrag()); | |
2783 | } | |
7c673cae FG |
2784 | // bounds too |
2785 | vector<dirfrag_t> dfls; | |
2786 | for (set<CDir*>::iterator q = subtrees[dir].begin(); | |
2787 | q != subtrees[dir].end(); | |
2788 | ++q) { | |
2789 | CDir *bound = *q; | |
2790 | dfls.push_back(bound->dirfrag()); | |
2791 | } | |
2792 | ||
2793 | my_subtrees[dir->dirfrag()] = dfls; | |
2794 | dout(10) << " claim " << dir->dirfrag() << " " << dfls << dendl; | |
2795 | } | |
2796 | } | |
2797 | ||
2798 | // ambiguous | |
2799 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator p = my_ambiguous_imports.begin(); | |
2800 | p != my_ambiguous_imports.end(); | |
2801 | ++p) { | |
2802 | my_ambig_imports[p->first] = p->second; | |
2803 | dout(10) << " ambig " << p->first << " " << p->second << dendl; | |
2804 | } | |
2805 | ||
2806 | // simplify the claimed subtree. | |
2807 | for (auto p = my_subtrees.begin(); p != my_subtrees.end(); ++p) { | |
2808 | unsigned i = 0; | |
2809 | while (i < p->second.size()) { | |
2810 | dirfrag_t b = p->second[i]; | |
2811 | if (my_subtrees.count(b)) { | |
2812 | vector<dirfrag_t>& bb = my_subtrees[b]; | |
2813 | dout(10) << " simplify: " << p->first << " swallowing " << b << " with bounds " << bb << dendl; | |
2814 | for (vector<dirfrag_t>::iterator r = bb.begin(); r != bb.end(); ++r) | |
2815 | p->second.push_back(*r); | |
2816 | my_subtrees.erase(b); | |
2817 | p->second.erase(p->second.begin() + i); | |
2818 | } else { | |
2819 | ++i; | |
2820 | } | |
2821 | } | |
2822 | } | |
2823 | ||
2824 | // send | |
11fdf7f2 TL |
2825 | for (auto &p : resolves) { |
2826 | const MMDSResolve::ref &m = p.second; | |
2827 | if (mds->is_resolve()) { | |
2828 | m->add_table_commits(TABLE_SNAP, resolve_snapclient_commits); | |
2829 | } else { | |
2830 | m->add_table_commits(TABLE_SNAP, mds->snapclient->get_journaled_tids()); | |
2831 | } | |
7c673cae FG |
2832 | m->subtrees = my_subtrees; |
2833 | m->ambiguous_imports = my_ambig_imports; | |
11fdf7f2 TL |
2834 | dout(10) << "sending subtee resolve to mds." << p.first << dendl; |
2835 | mds->send_message_mds(m, p.first); | |
7c673cae FG |
2836 | } |
2837 | resolves_pending = false; | |
2838 | } | |
2839 | ||
11fdf7f2 TL |
2840 | void MDCache::maybe_finish_slave_resolve() { |
2841 | if (resolve_ack_gather.empty() && resolve_need_rollback.empty()) { | |
2842 | // snap cache get synced or I'm in resolve state | |
2843 | if (mds->snapclient->is_synced() || resolve_done) | |
2844 | send_subtree_resolves(); | |
2845 | process_delayed_resolve(); | |
2846 | } | |
2847 | } | |
2848 | ||
7c673cae FG |
2849 | void MDCache::handle_mds_failure(mds_rank_t who) |
2850 | { | |
2851 | dout(7) << "handle_mds_failure mds." << who << dendl; | |
2852 | ||
2853 | dout(1) << "handle_mds_failure mds." << who << " : recovery peers are " << recovery_set << dendl; | |
2854 | ||
2855 | resolve_gather.insert(who); | |
2856 | discard_delayed_resolve(who); | |
2857 | ambiguous_slave_updates.erase(who); | |
2858 | ||
2859 | rejoin_gather.insert(who); | |
2860 | rejoin_sent.erase(who); // i need to send another | |
31f18b77 | 2861 | rejoin_ack_sent.erase(who); // i need to send another |
7c673cae FG |
2862 | rejoin_ack_gather.erase(who); // i'll need/get another. |
2863 | ||
2864 | dout(10) << " resolve_gather " << resolve_gather << dendl; | |
2865 | dout(10) << " resolve_ack_gather " << resolve_ack_gather << dendl; | |
2866 | dout(10) << " rejoin_sent " << rejoin_sent << dendl; | |
2867 | dout(10) << " rejoin_gather " << rejoin_gather << dendl; | |
2868 | dout(10) << " rejoin_ack_gather " << rejoin_ack_gather << dendl; | |
2869 | ||
2870 | ||
2871 | // tell the migrator too. | |
2872 | migrator->handle_mds_failure_or_stop(who); | |
2873 | ||
224ce89b WB |
2874 | // tell the balancer too. |
2875 | mds->balancer->handle_mds_failure(who); | |
2876 | ||
7c673cae FG |
2877 | // clean up any requests slave to/from this node |
2878 | list<MDRequestRef> finish; | |
2879 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
2880 | p != active_requests.end(); | |
2881 | ++p) { | |
2882 | MDRequestRef& mdr = p->second; | |
2883 | // slave to the failed node? | |
2884 | if (mdr->slave_to_mds == who) { | |
2885 | if (mdr->slave_did_prepare()) { | |
2886 | dout(10) << " slave request " << *mdr << " uncommitted, will resolve shortly" << dendl; | |
2887 | if (is_ambiguous_slave_update(p->first, mdr->slave_to_mds)) | |
2888 | remove_ambiguous_slave_update(p->first, mdr->slave_to_mds); | |
2889 | ||
2890 | if (!mdr->more()->waiting_on_slave.empty()) { | |
11fdf7f2 | 2891 | ceph_assert(mdr->more()->srcdn_auth_mds == mds->get_nodeid()); |
7c673cae | 2892 | // will rollback, no need to wait |
91327a77 | 2893 | mdr->reset_slave_request(); |
7c673cae FG |
2894 | mdr->more()->waiting_on_slave.clear(); |
2895 | } | |
2896 | } else if (!mdr->committing) { | |
2897 | dout(10) << " slave request " << *mdr << " has no prepare, finishing up" << dendl; | |
2898 | if (mdr->slave_request || mdr->slave_rolling_back()) | |
2899 | mdr->aborted = true; | |
2900 | else | |
2901 | finish.push_back(mdr); | |
2902 | } | |
2903 | } | |
2904 | ||
2905 | if (mdr->is_slave() && mdr->slave_did_prepare()) { | |
2906 | if (mdr->more()->waiting_on_slave.count(who)) { | |
11fdf7f2 | 2907 | ceph_assert(mdr->more()->srcdn_auth_mds == mds->get_nodeid()); |
7c673cae FG |
2908 | dout(10) << " slave request " << *mdr << " no longer need rename notity ack from mds." |
2909 | << who << dendl; | |
2910 | mdr->more()->waiting_on_slave.erase(who); | |
2911 | if (mdr->more()->waiting_on_slave.empty() && mdr->slave_request) | |
2912 | mds->queue_waiter(new C_MDS_RetryRequest(this, mdr)); | |
2913 | } | |
2914 | ||
2915 | if (mdr->more()->srcdn_auth_mds == who && | |
2916 | mds->mdsmap->is_clientreplay_or_active_or_stopping(mdr->slave_to_mds)) { | |
2917 | // rename srcdn's auth mds failed, resolve even I'm a survivor. | |
2918 | dout(10) << " slave request " << *mdr << " uncommitted, will resolve shortly" << dendl; | |
2919 | add_ambiguous_slave_update(p->first, mdr->slave_to_mds); | |
2920 | } | |
31f18b77 | 2921 | } else if (mdr->slave_request) { |
11fdf7f2 | 2922 | const MMDSSlaveRequest::const_ref &slave_req = mdr->slave_request; |
31f18b77 FG |
2923 | // FIXME: Slave rename request can arrive after we notice mds failure. |
2924 | // This can cause mds to crash (does not affect integrity of FS). | |
2925 | if (slave_req->get_op() == MMDSSlaveRequest::OP_RENAMEPREP && | |
2926 | slave_req->srcdn_auth == who) | |
2927 | slave_req->mark_interrupted(); | |
7c673cae FG |
2928 | } |
2929 | ||
2930 | // failed node is slave? | |
2931 | if (mdr->is_master() && !mdr->committing) { | |
2932 | if (mdr->more()->srcdn_auth_mds == who) { | |
2933 | dout(10) << " master request " << *mdr << " waiting for rename srcdn's auth mds." | |
2934 | << who << " to recover" << dendl; | |
11fdf7f2 | 2935 | ceph_assert(mdr->more()->witnessed.count(who) == 0); |
7c673cae FG |
2936 | if (mdr->more()->is_ambiguous_auth) |
2937 | mdr->clear_ambiguous_auth(); | |
2938 | // rename srcdn's auth mds failed, all witnesses will rollback | |
2939 | mdr->more()->witnessed.clear(); | |
2940 | pending_masters.erase(p->first); | |
2941 | } | |
2942 | ||
2943 | if (mdr->more()->witnessed.count(who)) { | |
2944 | mds_rank_t srcdn_auth = mdr->more()->srcdn_auth_mds; | |
2945 | if (srcdn_auth >= 0 && mdr->more()->waiting_on_slave.count(srcdn_auth)) { | |
2946 | dout(10) << " master request " << *mdr << " waiting for rename srcdn's auth mds." | |
2947 | << mdr->more()->srcdn_auth_mds << " to reply" << dendl; | |
2948 | // waiting for the slave (rename srcdn's auth mds), delay sending resolve ack | |
2949 | // until either the request is committing or the slave also fails. | |
11fdf7f2 | 2950 | ceph_assert(mdr->more()->waiting_on_slave.size() == 1); |
7c673cae FG |
2951 | pending_masters.insert(p->first); |
2952 | } else { | |
2953 | dout(10) << " master request " << *mdr << " no longer witnessed by slave mds." | |
2954 | << who << " to recover" << dendl; | |
2955 | if (srcdn_auth >= 0) | |
11fdf7f2 | 2956 | ceph_assert(mdr->more()->witnessed.count(srcdn_auth) == 0); |
7c673cae FG |
2957 | |
2958 | // discard this peer's prepare (if any) | |
2959 | mdr->more()->witnessed.erase(who); | |
2960 | } | |
2961 | } | |
2962 | ||
2963 | if (mdr->more()->waiting_on_slave.count(who)) { | |
2964 | dout(10) << " master request " << *mdr << " waiting for slave mds." << who | |
2965 | << " to recover" << dendl; | |
2966 | // retry request when peer recovers | |
2967 | mdr->more()->waiting_on_slave.erase(who); | |
2968 | if (mdr->more()->waiting_on_slave.empty()) | |
2969 | mds->wait_for_active_peer(who, new C_MDS_RetryRequest(this, mdr)); | |
2970 | } | |
2971 | ||
2972 | if (mdr->locking && mdr->locking_target_mds == who) | |
2973 | mdr->finish_locking(mdr->locking); | |
2974 | } | |
2975 | } | |
2976 | ||
2977 | for (map<metareqid_t, umaster>::iterator p = uncommitted_masters.begin(); | |
2978 | p != uncommitted_masters.end(); | |
2979 | ++p) { | |
2980 | // The failed MDS may have already committed the slave update | |
2981 | if (p->second.slaves.count(who)) { | |
2982 | p->second.recovering = true; | |
2983 | p->second.slaves.erase(who); | |
2984 | } | |
2985 | } | |
2986 | ||
2987 | while (!finish.empty()) { | |
2988 | dout(10) << "cleaning up slave request " << *finish.front() << dendl; | |
2989 | request_finish(finish.front()); | |
2990 | finish.pop_front(); | |
2991 | } | |
2992 | ||
2993 | kick_find_ino_peers(who); | |
2994 | kick_open_ino_peers(who); | |
2995 | ||
2996 | for (map<dirfrag_t,fragment_info_t>::iterator p = fragments.begin(); | |
2997 | p != fragments.end(); ) { | |
2998 | dirfrag_t df = p->first; | |
2999 | fragment_info_t& info = p->second; | |
a8e16298 TL |
3000 | |
3001 | if (info.is_fragmenting()) { | |
3002 | if (info.notify_ack_waiting.erase(who) && | |
3003 | info.notify_ack_waiting.empty()) { | |
3004 | fragment_drop_locks(info); | |
3005 | fragment_maybe_finish(p++); | |
3006 | } else { | |
3007 | ++p; | |
3008 | } | |
7c673cae | 3009 | continue; |
a8e16298 TL |
3010 | } |
3011 | ||
3012 | ++p; | |
7c673cae FG |
3013 | dout(10) << "cancelling fragment " << df << " bit " << info.bits << dendl; |
3014 | list<CDir*> dirs; | |
3015 | info.dirs.swap(dirs); | |
3016 | fragments.erase(df); | |
3017 | fragment_unmark_unfreeze_dirs(dirs); | |
3018 | } | |
3019 | ||
3020 | // MDCache::shutdown_export_strays() always exports strays to mds.0 | |
3021 | if (who == mds_rank_t(0)) | |
f64942e4 | 3022 | shutdown_exporting_strays.clear(); |
7c673cae FG |
3023 | |
3024 | show_subtrees(); | |
3025 | } | |
3026 | ||
3027 | /* | |
3028 | * handle_mds_recovery - called on another node's transition | |
3029 | * from resolve -> active. | |
3030 | */ | |
3031 | void MDCache::handle_mds_recovery(mds_rank_t who) | |
3032 | { | |
3033 | dout(7) << "handle_mds_recovery mds." << who << dendl; | |
3034 | ||
3035 | // exclude all discover waiters. kick_discovers() will do the job | |
3036 | static const uint64_t i_mask = CInode::WAIT_ANY_MASK & ~CInode::WAIT_DIR; | |
3037 | static const uint64_t d_mask = CDir::WAIT_ANY_MASK & ~CDir::WAIT_DENTRY; | |
3038 | ||
11fdf7f2 | 3039 | MDSContext::vec waiters; |
7c673cae FG |
3040 | |
3041 | // wake up any waiters in their subtrees | |
3042 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3043 | p != subtrees.end(); | |
3044 | ++p) { | |
3045 | CDir *dir = p->first; | |
3046 | ||
3047 | if (dir->authority().first != who || | |
3048 | dir->authority().second == mds->get_nodeid()) | |
3049 | continue; | |
11fdf7f2 | 3050 | ceph_assert(!dir->is_auth()); |
7c673cae FG |
3051 | |
3052 | // wake any waiters | |
3053 | list<CDir*> q; | |
3054 | q.push_back(dir); | |
3055 | ||
3056 | while (!q.empty()) { | |
3057 | CDir *d = q.front(); | |
3058 | q.pop_front(); | |
3059 | d->take_waiting(d_mask, waiters); | |
3060 | ||
3061 | // inode waiters too | |
94b18763 FG |
3062 | for (auto &p : d->items) { |
3063 | CDentry *dn = p.second; | |
7c673cae FG |
3064 | CDentry::linkage_t *dnl = dn->get_linkage(); |
3065 | if (dnl->is_primary()) { | |
3066 | dnl->get_inode()->take_waiting(i_mask, waiters); | |
3067 | ||
3068 | // recurse? | |
3069 | list<CDir*> ls; | |
3070 | dnl->get_inode()->get_dirfrags(ls); | |
3071 | for (list<CDir*>::iterator p = ls.begin(); | |
3072 | p != ls.end(); | |
3073 | ++p) { | |
3074 | CDir *subdir = *p; | |
3075 | if (!subdir->is_subtree_root()) | |
3076 | q.push_back(subdir); | |
3077 | } | |
3078 | } | |
3079 | } | |
3080 | } | |
3081 | } | |
3082 | ||
3083 | kick_open_ino_peers(who); | |
3084 | kick_find_ino_peers(who); | |
3085 | ||
3086 | // queue them up. | |
3087 | mds->queue_waiters(waiters); | |
3088 | } | |
3089 | ||
3090 | void MDCache::set_recovery_set(set<mds_rank_t>& s) | |
3091 | { | |
3092 | dout(7) << "set_recovery_set " << s << dendl; | |
3093 | recovery_set = s; | |
3094 | } | |
3095 | ||
3096 | ||
3097 | /* | |
3098 | * during resolve state, we share resolves to determine who | |
3099 | * is authoritative for which trees. we expect to get an resolve | |
3100 | * from _everyone_ in the recovery_set (the mds cluster at the time of | |
3101 | * the first failure). | |
3102 | * | |
3103 | * This functions puts the passed message before returning | |
3104 | */ | |
11fdf7f2 | 3105 | void MDCache::handle_resolve(const MMDSResolve::const_ref &m) |
7c673cae FG |
3106 | { |
3107 | dout(7) << "handle_resolve from " << m->get_source() << dendl; | |
3108 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
3109 | ||
3110 | if (mds->get_state() < MDSMap::STATE_RESOLVE) { | |
3111 | if (mds->get_want_state() == CEPH_MDS_STATE_RESOLVE) { | |
3112 | mds->wait_for_resolve(new C_MDS_RetryMessage(mds, m)); | |
3113 | return; | |
3114 | } | |
3115 | // wait until we reach the resolve stage! | |
7c673cae FG |
3116 | return; |
3117 | } | |
3118 | ||
3119 | discard_delayed_resolve(from); | |
3120 | ||
3121 | // ambiguous slave requests? | |
3122 | if (!m->slave_requests.empty()) { | |
3123 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
3124 | for (auto p = m->slave_requests.begin(); p != m->slave_requests.end(); ++p) { | |
3125 | if (uncommitted_masters.count(p->first) && !uncommitted_masters[p->first].safe) { | |
11fdf7f2 | 3126 | ceph_assert(!p->second.committing); |
7c673cae FG |
3127 | pending_masters.insert(p->first); |
3128 | } | |
3129 | } | |
3130 | ||
3131 | if (!pending_masters.empty()) { | |
3132 | dout(10) << " still have pending updates, delay processing slave resolve" << dendl; | |
3133 | delayed_resolve[from] = m; | |
3134 | return; | |
3135 | } | |
3136 | } | |
3137 | ||
11fdf7f2 TL |
3138 | auto ack = MMDSResolveAck::create(); |
3139 | for (const auto &p : m->slave_requests) { | |
3140 | if (uncommitted_masters.count(p.first)) { //mds->sessionmap.have_completed_request(p.first)) { | |
7c673cae | 3141 | // COMMIT |
11fdf7f2 | 3142 | if (p.second.committing) { |
7c673cae | 3143 | // already committing, waiting for the OP_COMMITTED slave reply |
11fdf7f2 | 3144 | dout(10) << " already committing slave request " << p << " noop "<< dendl; |
7c673cae | 3145 | } else { |
11fdf7f2 TL |
3146 | dout(10) << " ambiguous slave request " << p << " will COMMIT" << dendl; |
3147 | ack->add_commit(p.first); | |
7c673cae | 3148 | } |
11fdf7f2 | 3149 | uncommitted_masters[p.first].slaves.insert(from); // wait for slave OP_COMMITTED before we log ECommitted |
7c673cae | 3150 | |
11fdf7f2 | 3151 | if (p.second.inode_caps.length() > 0) { |
7c673cae | 3152 | // slave wants to export caps (rename) |
11fdf7f2 | 3153 | ceph_assert(mds->is_resolve()); |
7c673cae FG |
3154 | |
3155 | inodeno_t ino; | |
3156 | map<client_t,Capability::Export> cap_exports; | |
11fdf7f2 TL |
3157 | auto q = p.second.inode_caps.cbegin(); |
3158 | decode(ino, q); | |
3159 | decode(cap_exports, q); | |
7c673cae | 3160 | |
11fdf7f2 | 3161 | ceph_assert(get_inode(ino)); |
7c673cae FG |
3162 | |
3163 | for (map<client_t,Capability::Export>::iterator q = cap_exports.begin(); | |
3164 | q != cap_exports.end(); | |
3165 | ++q) { | |
3166 | Capability::Import& im = rejoin_imported_caps[from][ino][q->first]; | |
3167 | im.cap_id = ++last_cap_id; // assign a new cap ID | |
3168 | im.issue_seq = 1; | |
3169 | im.mseq = q->second.mseq; | |
28e407b8 AA |
3170 | |
3171 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
3172 | if (session) | |
3173 | rejoin_client_map.emplace(q->first, session->info.inst); | |
7c673cae FG |
3174 | } |
3175 | ||
3176 | // will process these caps in rejoin stage | |
3177 | rejoin_slave_exports[ino].first = from; | |
3178 | rejoin_slave_exports[ino].second.swap(cap_exports); | |
3179 | ||
3180 | // send information of imported caps back to slave | |
11fdf7f2 | 3181 | encode(rejoin_imported_caps[from][ino], ack->commit[p.first]); |
7c673cae FG |
3182 | } |
3183 | } else { | |
3184 | // ABORT | |
11fdf7f2 TL |
3185 | dout(10) << " ambiguous slave request " << p << " will ABORT" << dendl; |
3186 | ceph_assert(!p.second.committing); | |
3187 | ack->add_abort(p.first); | |
7c673cae FG |
3188 | } |
3189 | } | |
3190 | mds->send_message(ack, m->get_connection()); | |
7c673cae FG |
3191 | return; |
3192 | } | |
3193 | ||
11fdf7f2 | 3194 | if (!resolve_ack_gather.empty() || !resolve_need_rollback.empty()) { |
7c673cae FG |
3195 | dout(10) << "delay processing subtree resolve" << dendl; |
3196 | delayed_resolve[from] = m; | |
3197 | return; | |
3198 | } | |
3199 | ||
3200 | bool survivor = false; | |
3201 | // am i a surviving ambiguous importer? | |
3202 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
3203 | survivor = true; | |
3204 | // check for any import success/failure (from this node) | |
3205 | map<dirfrag_t, vector<dirfrag_t> >::iterator p = my_ambiguous_imports.begin(); | |
3206 | while (p != my_ambiguous_imports.end()) { | |
3207 | map<dirfrag_t, vector<dirfrag_t> >::iterator next = p; | |
3208 | ++next; | |
3209 | CDir *dir = get_dirfrag(p->first); | |
11fdf7f2 | 3210 | ceph_assert(dir); |
7c673cae FG |
3211 | dout(10) << "checking ambiguous import " << *dir << dendl; |
3212 | if (migrator->is_importing(dir->dirfrag()) && | |
3213 | migrator->get_import_peer(dir->dirfrag()) == from) { | |
11fdf7f2 | 3214 | ceph_assert(migrator->get_import_state(dir->dirfrag()) == Migrator::IMPORT_ACKING); |
7c673cae FG |
3215 | |
3216 | // check if sender claims the subtree | |
3217 | bool claimed_by_sender = false; | |
11fdf7f2 | 3218 | for (const auto &q : m->subtrees) { |
7c673cae | 3219 | // an ambiguous import won't race with a refragmentation; it's appropriate to force here. |
11fdf7f2 | 3220 | CDir *base = get_force_dirfrag(q.first, false); |
7c673cae FG |
3221 | if (!base || !base->contains(dir)) |
3222 | continue; // base not dir or an ancestor of dir, clearly doesn't claim dir. | |
3223 | ||
3224 | bool inside = true; | |
3225 | set<CDir*> bounds; | |
11fdf7f2 | 3226 | get_force_dirfrag_bound_set(q.second, bounds); |
7c673cae FG |
3227 | for (set<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) { |
3228 | CDir *bound = *p; | |
3229 | if (bound->contains(dir)) { | |
3230 | inside = false; // nope, bound is dir or parent of dir, not inside. | |
3231 | break; | |
3232 | } | |
3233 | } | |
3234 | if (inside) | |
3235 | claimed_by_sender = true; | |
3236 | } | |
3237 | ||
3238 | my_ambiguous_imports.erase(p); // no longer ambiguous. | |
3239 | if (claimed_by_sender) { | |
3240 | dout(7) << "ambiguous import failed on " << *dir << dendl; | |
3241 | migrator->import_reverse(dir); | |
3242 | } else { | |
3243 | dout(7) << "ambiguous import succeeded on " << *dir << dendl; | |
3244 | migrator->import_finish(dir, true); | |
3245 | } | |
3246 | } | |
3247 | p = next; | |
3248 | } | |
3249 | } | |
3250 | ||
3251 | // update my dir_auth values | |
3252 | // need to do this on recoverying nodes _and_ bystanders (to resolve ambiguous | |
3253 | // migrations between other nodes) | |
11fdf7f2 TL |
3254 | for (const auto& p : m->subtrees) { |
3255 | dout(10) << "peer claims " << p.first << " bounds " << p.second << dendl; | |
3256 | CDir *dir = get_force_dirfrag(p.first, !survivor); | |
7c673cae FG |
3257 | if (!dir) |
3258 | continue; | |
11fdf7f2 | 3259 | adjust_bounded_subtree_auth(dir, p.second, from); |
7c673cae FG |
3260 | try_subtree_merge(dir); |
3261 | } | |
3262 | ||
3263 | show_subtrees(); | |
3264 | ||
3265 | // note ambiguous imports too | |
11fdf7f2 TL |
3266 | for (const auto& p : m->ambiguous_imports) { |
3267 | dout(10) << "noting ambiguous import on " << p.first << " bounds " << p.second << dendl; | |
3268 | other_ambiguous_imports[from][p.first] = p.second; | |
3269 | } | |
3270 | ||
3271 | // learn other mds' pendina snaptable commits. later when resolve finishes, we will reload | |
3272 | // snaptable cache from snapserver. By this way, snaptable cache get synced among all mds | |
3273 | for (const auto& p : m->table_clients) { | |
3274 | dout(10) << " noting " << get_mdstable_name(p.type) | |
3275 | << " pending_commits " << p.pending_commits << dendl; | |
3276 | MDSTableClient *client = mds->get_table_client(p.type); | |
3277 | for (const auto& q : p.pending_commits) | |
3278 | client->notify_commit(q); | |
7c673cae FG |
3279 | } |
3280 | ||
3281 | // did i get them all? | |
3282 | resolve_gather.erase(from); | |
3283 | ||
3284 | maybe_resolve_finish(); | |
7c673cae FG |
3285 | } |
3286 | ||
3287 | void MDCache::process_delayed_resolve() | |
3288 | { | |
3289 | dout(10) << "process_delayed_resolve" << dendl; | |
11fdf7f2 | 3290 | map<mds_rank_t, MMDSResolve::const_ref> tmp; |
7c673cae | 3291 | tmp.swap(delayed_resolve); |
11fdf7f2 TL |
3292 | for (auto &p : tmp) { |
3293 | handle_resolve(p.second); | |
3294 | } | |
7c673cae FG |
3295 | } |
3296 | ||
3297 | void MDCache::discard_delayed_resolve(mds_rank_t who) | |
3298 | { | |
11fdf7f2 | 3299 | delayed_resolve.erase(who); |
7c673cae FG |
3300 | } |
3301 | ||
3302 | void MDCache::maybe_resolve_finish() | |
3303 | { | |
11fdf7f2 TL |
3304 | ceph_assert(resolve_ack_gather.empty()); |
3305 | ceph_assert(resolve_need_rollback.empty()); | |
7c673cae FG |
3306 | |
3307 | if (!resolve_gather.empty()) { | |
3308 | dout(10) << "maybe_resolve_finish still waiting for resolves (" | |
3309 | << resolve_gather << ")" << dendl; | |
3310 | return; | |
3311 | } | |
3312 | ||
3313 | dout(10) << "maybe_resolve_finish got all resolves+resolve_acks, done." << dendl; | |
3314 | disambiguate_my_imports(); | |
3315 | finish_committed_masters(); | |
3316 | ||
3317 | if (resolve_done) { | |
11fdf7f2 | 3318 | ceph_assert(mds->is_resolve()); |
7c673cae FG |
3319 | trim_unlinked_inodes(); |
3320 | recalc_auth_bits(false); | |
3321 | resolve_done.release()->complete(0); | |
3322 | } else { | |
11fdf7f2 | 3323 | // I am survivor. |
7c673cae FG |
3324 | maybe_send_pending_rejoins(); |
3325 | } | |
3326 | } | |
3327 | ||
11fdf7f2 | 3328 | void MDCache::handle_resolve_ack(const MMDSResolveAck::const_ref &ack) |
7c673cae FG |
3329 | { |
3330 | dout(10) << "handle_resolve_ack " << *ack << " from " << ack->get_source() << dendl; | |
3331 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
3332 | ||
3333 | if (!resolve_ack_gather.count(from) || | |
3334 | mds->mdsmap->get_state(from) < MDSMap::STATE_RESOLVE) { | |
7c673cae FG |
3335 | return; |
3336 | } | |
3337 | ||
3338 | if (ambiguous_slave_updates.count(from)) { | |
11fdf7f2 TL |
3339 | ceph_assert(mds->mdsmap->is_clientreplay_or_active_or_stopping(from)); |
3340 | ceph_assert(mds->is_clientreplay() || mds->is_active() || mds->is_stopping()); | |
7c673cae FG |
3341 | } |
3342 | ||
11fdf7f2 TL |
3343 | for (const auto &p : ack->commit) { |
3344 | dout(10) << " commit on slave " << p.first << dendl; | |
7c673cae FG |
3345 | |
3346 | if (ambiguous_slave_updates.count(from)) { | |
11fdf7f2 | 3347 | remove_ambiguous_slave_update(p.first, from); |
7c673cae FG |
3348 | continue; |
3349 | } | |
3350 | ||
3351 | if (mds->is_resolve()) { | |
3352 | // replay | |
11fdf7f2 TL |
3353 | MDSlaveUpdate *su = get_uncommitted_slave_update(p.first, from); |
3354 | ceph_assert(su); | |
7c673cae FG |
3355 | |
3356 | // log commit | |
11fdf7f2 | 3357 | mds->mdlog->start_submit_entry(new ESlaveUpdate(mds->mdlog, "unknown", p.first, from, |
7c673cae | 3358 | ESlaveUpdate::OP_COMMIT, su->origop), |
11fdf7f2 | 3359 | new C_MDC_SlaveCommit(this, from, p.first)); |
7c673cae FG |
3360 | mds->mdlog->flush(); |
3361 | ||
11fdf7f2 | 3362 | finish_uncommitted_slave_update(p.first, from); |
7c673cae | 3363 | } else { |
11fdf7f2 | 3364 | MDRequestRef mdr = request_get(p.first); |
7c673cae | 3365 | // information about master imported caps |
11fdf7f2 TL |
3366 | if (p.second.length() > 0) |
3367 | mdr->more()->inode_import.share(p.second); | |
7c673cae | 3368 | |
11fdf7f2 | 3369 | ceph_assert(mdr->slave_request == 0); // shouldn't be doing anything! |
7c673cae FG |
3370 | request_finish(mdr); |
3371 | } | |
3372 | } | |
3373 | ||
11fdf7f2 TL |
3374 | for (const auto &metareq : ack->abort) { |
3375 | dout(10) << " abort on slave " << metareq << dendl; | |
7c673cae FG |
3376 | |
3377 | if (mds->is_resolve()) { | |
11fdf7f2 TL |
3378 | MDSlaveUpdate *su = get_uncommitted_slave_update(metareq, from); |
3379 | ceph_assert(su); | |
7c673cae FG |
3380 | |
3381 | // perform rollback (and journal a rollback entry) | |
3382 | // note: this will hold up the resolve a bit, until the rollback entries journal. | |
3383 | MDRequestRef null_ref; | |
3384 | switch (su->origop) { | |
3385 | case ESlaveUpdate::LINK: | |
3386 | mds->server->do_link_rollback(su->rollback, from, null_ref); | |
3387 | break; | |
3388 | case ESlaveUpdate::RENAME: | |
3389 | mds->server->do_rename_rollback(su->rollback, from, null_ref); | |
3390 | break; | |
3391 | case ESlaveUpdate::RMDIR: | |
3392 | mds->server->do_rmdir_rollback(su->rollback, from, null_ref); | |
3393 | break; | |
3394 | default: | |
3395 | ceph_abort(); | |
3396 | } | |
3397 | } else { | |
11fdf7f2 | 3398 | MDRequestRef mdr = request_get(metareq); |
7c673cae FG |
3399 | mdr->aborted = true; |
3400 | if (mdr->slave_request) { | |
3401 | if (mdr->slave_did_prepare()) // journaling slave prepare ? | |
11fdf7f2 | 3402 | add_rollback(metareq, from); |
7c673cae FG |
3403 | } else { |
3404 | request_finish(mdr); | |
3405 | } | |
3406 | } | |
3407 | } | |
3408 | ||
11fdf7f2 | 3409 | if (!ambiguous_slave_updates.count(from)) { |
7c673cae | 3410 | resolve_ack_gather.erase(from); |
11fdf7f2 | 3411 | maybe_finish_slave_resolve(); |
7c673cae | 3412 | } |
7c673cae FG |
3413 | } |
3414 | ||
3415 | void MDCache::add_uncommitted_slave_update(metareqid_t reqid, mds_rank_t master, MDSlaveUpdate *su) | |
3416 | { | |
11fdf7f2 | 3417 | ceph_assert(uncommitted_slave_updates[master].count(reqid) == 0); |
7c673cae FG |
3418 | uncommitted_slave_updates[master][reqid] = su; |
3419 | for(set<CInode*>::iterator p = su->olddirs.begin(); p != su->olddirs.end(); ++p) | |
3420 | uncommitted_slave_rename_olddir[*p]++; | |
3421 | for(set<CInode*>::iterator p = su->unlinked.begin(); p != su->unlinked.end(); ++p) | |
3422 | uncommitted_slave_unlink[*p]++; | |
3423 | } | |
3424 | ||
3425 | void MDCache::finish_uncommitted_slave_update(metareqid_t reqid, mds_rank_t master) | |
3426 | { | |
11fdf7f2 | 3427 | ceph_assert(uncommitted_slave_updates[master].count(reqid)); |
7c673cae FG |
3428 | MDSlaveUpdate* su = uncommitted_slave_updates[master][reqid]; |
3429 | ||
3430 | uncommitted_slave_updates[master].erase(reqid); | |
3431 | if (uncommitted_slave_updates[master].empty()) | |
3432 | uncommitted_slave_updates.erase(master); | |
3433 | // discard the non-auth subtree we renamed out of | |
3434 | for(set<CInode*>::iterator p = su->olddirs.begin(); p != su->olddirs.end(); ++p) { | |
3435 | CInode *diri = *p; | |
3436 | map<CInode*, int>::iterator it = uncommitted_slave_rename_olddir.find(diri); | |
11fdf7f2 | 3437 | ceph_assert(it != uncommitted_slave_rename_olddir.end()); |
7c673cae FG |
3438 | it->second--; |
3439 | if (it->second == 0) { | |
3440 | uncommitted_slave_rename_olddir.erase(it); | |
3441 | list<CDir*> ls; | |
3442 | diri->get_dirfrags(ls); | |
3443 | for (list<CDir*>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
3444 | CDir *root = get_subtree_root(*q); | |
3445 | if (root->get_dir_auth() == CDIR_AUTH_UNDEF) { | |
3446 | try_trim_non_auth_subtree(root); | |
3447 | if (*q != root) | |
3448 | break; | |
3449 | } | |
3450 | } | |
3451 | } else | |
11fdf7f2 | 3452 | ceph_assert(it->second > 0); |
7c673cae FG |
3453 | } |
3454 | // removed the inodes that were unlinked by slave update | |
3455 | for(set<CInode*>::iterator p = su->unlinked.begin(); p != su->unlinked.end(); ++p) { | |
3456 | CInode *in = *p; | |
3457 | map<CInode*, int>::iterator it = uncommitted_slave_unlink.find(in); | |
11fdf7f2 | 3458 | ceph_assert(it != uncommitted_slave_unlink.end()); |
7c673cae FG |
3459 | it->second--; |
3460 | if (it->second == 0) { | |
3461 | uncommitted_slave_unlink.erase(it); | |
3462 | if (!in->get_projected_parent_dn()) | |
3463 | mds->mdcache->remove_inode_recursive(in); | |
3464 | } else | |
11fdf7f2 | 3465 | ceph_assert(it->second > 0); |
7c673cae FG |
3466 | } |
3467 | delete su; | |
3468 | } | |
3469 | ||
3470 | MDSlaveUpdate* MDCache::get_uncommitted_slave_update(metareqid_t reqid, mds_rank_t master) | |
3471 | { | |
3472 | ||
3473 | MDSlaveUpdate* su = NULL; | |
3474 | if (uncommitted_slave_updates.count(master) && | |
3475 | uncommitted_slave_updates[master].count(reqid)) { | |
3476 | su = uncommitted_slave_updates[master][reqid]; | |
11fdf7f2 | 3477 | ceph_assert(su); |
7c673cae FG |
3478 | } |
3479 | return su; | |
3480 | } | |
3481 | ||
3482 | void MDCache::finish_rollback(metareqid_t reqid) { | |
11fdf7f2 TL |
3483 | auto p = resolve_need_rollback.find(reqid); |
3484 | ceph_assert(p != resolve_need_rollback.end()); | |
7c673cae | 3485 | if (mds->is_resolve()) |
11fdf7f2 TL |
3486 | finish_uncommitted_slave_update(reqid, p->second); |
3487 | resolve_need_rollback.erase(p); | |
3488 | maybe_finish_slave_resolve(); | |
7c673cae FG |
3489 | } |
3490 | ||
3491 | void MDCache::disambiguate_other_imports() | |
3492 | { | |
3493 | dout(10) << "disambiguate_other_imports" << dendl; | |
3494 | ||
3495 | bool recovering = !(mds->is_clientreplay() || mds->is_active() || mds->is_stopping()); | |
3496 | // other nodes' ambiguous imports | |
3497 | for (map<mds_rank_t, map<dirfrag_t, vector<dirfrag_t> > >::iterator p = other_ambiguous_imports.begin(); | |
3498 | p != other_ambiguous_imports.end(); | |
3499 | ++p) { | |
3500 | mds_rank_t who = p->first; | |
3501 | dout(10) << "ambiguous imports for mds." << who << dendl; | |
3502 | ||
3503 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator q = p->second.begin(); | |
3504 | q != p->second.end(); | |
3505 | ++q) { | |
3506 | dout(10) << " ambiguous import " << q->first << " bounds " << q->second << dendl; | |
3507 | // an ambiguous import will not race with a refragmentation; it's appropriate to force here. | |
3508 | CDir *dir = get_force_dirfrag(q->first, recovering); | |
3509 | if (!dir) continue; | |
3510 | ||
3511 | if (dir->is_ambiguous_auth() || // works for me_ambig or if i am a surviving bystander | |
3512 | dir->authority() == CDIR_AUTH_UNDEF) { // resolving | |
3513 | dout(10) << " mds." << who << " did import " << *dir << dendl; | |
3514 | adjust_bounded_subtree_auth(dir, q->second, who); | |
3515 | try_subtree_merge(dir); | |
3516 | } else { | |
3517 | dout(10) << " mds." << who << " did not import " << *dir << dendl; | |
3518 | } | |
3519 | } | |
3520 | } | |
3521 | other_ambiguous_imports.clear(); | |
3522 | } | |
3523 | ||
3524 | void MDCache::disambiguate_my_imports() | |
3525 | { | |
3526 | dout(10) << "disambiguate_my_imports" << dendl; | |
3527 | ||
3528 | if (!mds->is_resolve()) { | |
11fdf7f2 | 3529 | ceph_assert(my_ambiguous_imports.empty()); |
7c673cae FG |
3530 | return; |
3531 | } | |
3532 | ||
3533 | disambiguate_other_imports(); | |
3534 | ||
3535 | // my ambiguous imports | |
3536 | mds_authority_t me_ambig(mds->get_nodeid(), mds->get_nodeid()); | |
3537 | while (!my_ambiguous_imports.empty()) { | |
3538 | map<dirfrag_t, vector<dirfrag_t> >::iterator q = my_ambiguous_imports.begin(); | |
3539 | ||
3540 | CDir *dir = get_dirfrag(q->first); | |
11fdf7f2 | 3541 | ceph_assert(dir); |
7c673cae FG |
3542 | |
3543 | if (dir->authority() != me_ambig) { | |
3544 | dout(10) << "ambiguous import auth known, must not be me " << *dir << dendl; | |
3545 | cancel_ambiguous_import(dir); | |
3546 | ||
3547 | mds->mdlog->start_submit_entry(new EImportFinish(dir, false)); | |
3548 | ||
3549 | // subtree may have been swallowed by another node claiming dir | |
3550 | // as their own. | |
3551 | CDir *root = get_subtree_root(dir); | |
3552 | if (root != dir) | |
3553 | dout(10) << " subtree root is " << *root << dendl; | |
11fdf7f2 | 3554 | ceph_assert(root->dir_auth.first != mds->get_nodeid()); // no us! |
7c673cae FG |
3555 | try_trim_non_auth_subtree(root); |
3556 | } else { | |
3557 | dout(10) << "ambiguous import auth unclaimed, must be me " << *dir << dendl; | |
3558 | finish_ambiguous_import(q->first); | |
3559 | mds->mdlog->start_submit_entry(new EImportFinish(dir, true)); | |
3560 | } | |
3561 | } | |
11fdf7f2 | 3562 | ceph_assert(my_ambiguous_imports.empty()); |
7c673cae FG |
3563 | mds->mdlog->flush(); |
3564 | ||
3565 | // verify all my subtrees are unambiguous! | |
3566 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3567 | p != subtrees.end(); | |
3568 | ++p) { | |
3569 | CDir *dir = p->first; | |
3570 | if (dir->is_ambiguous_dir_auth()) { | |
3571 | dout(0) << "disambiguate_imports uh oh, dir_auth is still ambiguous for " << *dir << dendl; | |
3572 | } | |
11fdf7f2 | 3573 | ceph_assert(!dir->is_ambiguous_dir_auth()); |
7c673cae FG |
3574 | } |
3575 | ||
3576 | show_subtrees(); | |
3577 | } | |
3578 | ||
3579 | ||
3580 | void MDCache::add_ambiguous_import(dirfrag_t base, const vector<dirfrag_t>& bounds) | |
3581 | { | |
11fdf7f2 | 3582 | ceph_assert(my_ambiguous_imports.count(base) == 0); |
7c673cae FG |
3583 | my_ambiguous_imports[base] = bounds; |
3584 | } | |
3585 | ||
3586 | ||
3587 | void MDCache::add_ambiguous_import(CDir *base, const set<CDir*>& bounds) | |
3588 | { | |
3589 | // make a list | |
3590 | vector<dirfrag_t> binos; | |
3591 | for (set<CDir*>::iterator p = bounds.begin(); | |
3592 | p != bounds.end(); | |
3593 | ++p) | |
3594 | binos.push_back((*p)->dirfrag()); | |
3595 | ||
3596 | // note: this can get called twice if the exporter fails during recovery | |
3597 | if (my_ambiguous_imports.count(base->dirfrag())) | |
3598 | my_ambiguous_imports.erase(base->dirfrag()); | |
3599 | ||
3600 | add_ambiguous_import(base->dirfrag(), binos); | |
3601 | } | |
3602 | ||
3603 | void MDCache::cancel_ambiguous_import(CDir *dir) | |
3604 | { | |
3605 | dirfrag_t df = dir->dirfrag(); | |
11fdf7f2 | 3606 | ceph_assert(my_ambiguous_imports.count(df)); |
7c673cae FG |
3607 | dout(10) << "cancel_ambiguous_import " << df |
3608 | << " bounds " << my_ambiguous_imports[df] | |
3609 | << " " << *dir | |
3610 | << dendl; | |
3611 | my_ambiguous_imports.erase(df); | |
3612 | } | |
3613 | ||
3614 | void MDCache::finish_ambiguous_import(dirfrag_t df) | |
3615 | { | |
11fdf7f2 | 3616 | ceph_assert(my_ambiguous_imports.count(df)); |
7c673cae FG |
3617 | vector<dirfrag_t> bounds; |
3618 | bounds.swap(my_ambiguous_imports[df]); | |
3619 | my_ambiguous_imports.erase(df); | |
3620 | ||
3621 | dout(10) << "finish_ambiguous_import " << df | |
3622 | << " bounds " << bounds | |
3623 | << dendl; | |
3624 | CDir *dir = get_dirfrag(df); | |
11fdf7f2 | 3625 | ceph_assert(dir); |
7c673cae FG |
3626 | |
3627 | // adjust dir_auth, import maps | |
3628 | adjust_bounded_subtree_auth(dir, bounds, mds->get_nodeid()); | |
3629 | try_subtree_merge(dir); | |
3630 | } | |
3631 | ||
3632 | void MDCache::remove_inode_recursive(CInode *in) | |
3633 | { | |
3634 | dout(10) << "remove_inode_recursive " << *in << dendl; | |
3635 | list<CDir*> ls; | |
3636 | in->get_dirfrags(ls); | |
3637 | list<CDir*>::iterator p = ls.begin(); | |
3638 | while (p != ls.end()) { | |
3639 | CDir *subdir = *p++; | |
3640 | ||
3641 | dout(10) << " removing dirfrag " << subdir << dendl; | |
94b18763 FG |
3642 | auto it = subdir->items.begin(); |
3643 | while (it != subdir->items.end()) { | |
3644 | CDentry *dn = it->second; | |
3645 | ++it; | |
7c673cae FG |
3646 | CDentry::linkage_t *dnl = dn->get_linkage(); |
3647 | if (dnl->is_primary()) { | |
3648 | CInode *tin = dnl->get_inode(); | |
31f18b77 | 3649 | subdir->unlink_inode(dn, false); |
7c673cae FG |
3650 | remove_inode_recursive(tin); |
3651 | } | |
3652 | subdir->remove_dentry(dn); | |
3653 | } | |
3654 | ||
3655 | if (subdir->is_subtree_root()) | |
3656 | remove_subtree(subdir); | |
3657 | in->close_dirfrag(subdir->dirfrag().frag); | |
3658 | } | |
3659 | remove_inode(in); | |
3660 | } | |
3661 | ||
11fdf7f2 | 3662 | bool MDCache::expire_recursive(CInode *in, expiremap &expiremap) |
7c673cae | 3663 | { |
11fdf7f2 | 3664 | ceph_assert(!in->is_auth()); |
7c673cae FG |
3665 | |
3666 | dout(10) << __func__ << ":" << *in << dendl; | |
3667 | ||
3668 | // Recurse into any dirfrags beneath this inode | |
3669 | list<CDir*> ls; | |
3670 | in->get_dirfrags(ls); | |
3671 | for (auto subdir : ls) { | |
3672 | if (!in->is_mdsdir() && subdir->is_subtree_root()) { | |
3673 | dout(10) << __func__ << ": stray still has subtree " << *in << dendl; | |
3674 | return true; | |
3675 | } | |
3676 | ||
3677 | for (auto &it : subdir->items) { | |
3678 | CDentry *dn = it.second; | |
3679 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
3680 | if (dnl->is_primary()) { | |
3681 | CInode *tin = dnl->get_inode(); | |
3682 | ||
3683 | /* Remote strays with linkage (i.e. hardlinks) should not be | |
3684 | * expired, because they may be the target of | |
3685 | * a rename() as the owning MDS shuts down */ | |
3686 | if (!tin->is_stray() && tin->inode.nlink) { | |
3687 | dout(10) << __func__ << ": stray still has linkage " << *tin << dendl; | |
3688 | return true; | |
3689 | } | |
3690 | ||
3691 | const bool abort = expire_recursive(tin, expiremap); | |
3692 | if (abort) { | |
3693 | return true; | |
3694 | } | |
3695 | } | |
3696 | if (dn->lru_is_expireable()) { | |
3697 | trim_dentry(dn, expiremap); | |
3698 | } else { | |
3699 | dout(10) << __func__ << ": stray dn is not expireable " << *dn << dendl; | |
3700 | return true; | |
3701 | } | |
3702 | } | |
3703 | } | |
3704 | ||
3705 | return false; | |
3706 | } | |
3707 | ||
3708 | void MDCache::trim_unlinked_inodes() | |
3709 | { | |
3710 | dout(7) << "trim_unlinked_inodes" << dendl; | |
81eedcae TL |
3711 | int count = 0; |
3712 | vector<CInode*> q; | |
94b18763 | 3713 | for (auto &p : inode_map) { |
b32b8144 | 3714 | CInode *in = p.second; |
7c673cae FG |
3715 | if (in->get_parent_dn() == NULL && !in->is_base()) { |
3716 | dout(7) << " will trim from " << *in << dendl; | |
3717 | q.push_back(in); | |
3718 | } | |
81eedcae TL |
3719 | |
3720 | if (!(++count % 1000)) | |
3721 | mds->heartbeat_reset(); | |
3722 | } | |
3723 | ||
3724 | for (auto& in : q) { | |
3725 | remove_inode_recursive(in); | |
3726 | ||
3727 | if (!(++count % 1000)) | |
3728 | mds->heartbeat_reset(); | |
7c673cae | 3729 | } |
7c673cae FG |
3730 | } |
3731 | ||
3732 | /** recalc_auth_bits() | |
3733 | * once subtree auth is disambiguated, we need to adjust all the | |
3734 | * auth and dirty bits in our cache before moving on. | |
3735 | */ | |
3736 | void MDCache::recalc_auth_bits(bool replay) | |
3737 | { | |
3738 | dout(7) << "recalc_auth_bits " << (replay ? "(replay)" : "") << dendl; | |
3739 | ||
3740 | if (root) { | |
3741 | root->inode_auth.first = mds->mdsmap->get_root(); | |
3742 | bool auth = mds->get_nodeid() == root->inode_auth.first; | |
3743 | if (auth) { | |
3744 | root->state_set(CInode::STATE_AUTH); | |
3745 | } else { | |
3746 | root->state_clear(CInode::STATE_AUTH); | |
3747 | if (!replay) | |
3748 | root->state_set(CInode::STATE_REJOINING); | |
3749 | } | |
3750 | } | |
3751 | ||
3752 | set<CInode*> subtree_inodes; | |
3753 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3754 | p != subtrees.end(); | |
3755 | ++p) { | |
3756 | if (p->first->dir_auth.first == mds->get_nodeid()) | |
3757 | subtree_inodes.insert(p->first->inode); | |
3758 | } | |
3759 | ||
3760 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3761 | p != subtrees.end(); | |
3762 | ++p) { | |
3763 | if (p->first->inode->is_mdsdir()) { | |
3764 | CInode *in = p->first->inode; | |
3765 | bool auth = in->ino() == MDS_INO_MDSDIR(mds->get_nodeid()); | |
3766 | if (auth) { | |
3767 | in->state_set(CInode::STATE_AUTH); | |
3768 | } else { | |
3769 | in->state_clear(CInode::STATE_AUTH); | |
3770 | if (!replay) | |
3771 | in->state_set(CInode::STATE_REJOINING); | |
3772 | } | |
3773 | } | |
3774 | ||
3775 | list<CDir*> dfq; // dirfrag queue | |
3776 | dfq.push_back(p->first); | |
3777 | ||
3778 | bool auth = p->first->authority().first == mds->get_nodeid(); | |
3779 | dout(10) << " subtree auth=" << auth << " for " << *p->first << dendl; | |
3780 | ||
3781 | while (!dfq.empty()) { | |
3782 | CDir *dir = dfq.front(); | |
3783 | dfq.pop_front(); | |
3784 | ||
3785 | // dir | |
3786 | if (auth) { | |
3787 | dir->state_set(CDir::STATE_AUTH); | |
3788 | } else { | |
3789 | dir->state_clear(CDir::STATE_AUTH); | |
3790 | if (!replay) { | |
3791 | // close empty non-auth dirfrag | |
3792 | if (!dir->is_subtree_root() && dir->get_num_any() == 0) { | |
3793 | dir->inode->close_dirfrag(dir->get_frag()); | |
3794 | continue; | |
3795 | } | |
3796 | dir->state_set(CDir::STATE_REJOINING); | |
3797 | dir->state_clear(CDir::STATE_COMPLETE); | |
3798 | if (dir->is_dirty()) | |
3799 | dir->mark_clean(); | |
3800 | } | |
3801 | } | |
3802 | ||
3803 | // dentries in this dir | |
94b18763 | 3804 | for (auto &p : dir->items) { |
7c673cae | 3805 | // dn |
94b18763 | 3806 | CDentry *dn = p.second; |
7c673cae FG |
3807 | CDentry::linkage_t *dnl = dn->get_linkage(); |
3808 | if (auth) { | |
3809 | dn->state_set(CDentry::STATE_AUTH); | |
3810 | } else { | |
3811 | dn->state_clear(CDentry::STATE_AUTH); | |
3812 | if (!replay) { | |
3813 | dn->state_set(CDentry::STATE_REJOINING); | |
3814 | if (dn->is_dirty()) | |
3815 | dn->mark_clean(); | |
3816 | } | |
3817 | } | |
3818 | ||
3819 | if (dnl->is_primary()) { | |
3820 | // inode | |
3821 | CInode *in = dnl->get_inode(); | |
3822 | if (auth) { | |
3823 | in->state_set(CInode::STATE_AUTH); | |
3824 | } else { | |
3825 | in->state_clear(CInode::STATE_AUTH); | |
3826 | if (!replay) { | |
3827 | in->state_set(CInode::STATE_REJOINING); | |
3828 | if (in->is_dirty()) | |
3829 | in->mark_clean(); | |
3830 | if (in->is_dirty_parent()) | |
3831 | in->clear_dirty_parent(); | |
3832 | // avoid touching scatterlocks for our subtree roots! | |
3833 | if (subtree_inodes.count(in) == 0) | |
3834 | in->clear_scatter_dirty(); | |
3835 | } | |
3836 | } | |
3837 | // recurse? | |
3838 | if (in->is_dir()) | |
3839 | in->get_nested_dirfrags(dfq); | |
3840 | } | |
3841 | } | |
3842 | } | |
3843 | } | |
3844 | ||
3845 | show_subtrees(); | |
3846 | show_cache(); | |
3847 | } | |
3848 | ||
3849 | ||
3850 | ||
3851 | // =========================================================================== | |
3852 | // REJOIN | |
3853 | ||
3854 | /* | |
3855 | * notes on scatterlock recovery: | |
3856 | * | |
3857 | * - recovering inode replica sends scatterlock data for any subtree | |
3858 | * roots (the only ones that are possibly dirty). | |
3859 | * | |
3860 | * - surviving auth incorporates any provided scatterlock data. any | |
3861 | * pending gathers are then finished, as with the other lock types. | |
3862 | * | |
3863 | * that takes care of surviving auth + (recovering replica)*. | |
3864 | * | |
3865 | * - surviving replica sends strong_inode, which includes current | |
3866 | * scatterlock state, AND any dirty scatterlock data. this | |
3867 | * provides the recovering auth with everything it might need. | |
3868 | * | |
3869 | * - recovering auth must pick initial scatterlock state based on | |
3870 | * (weak|strong) rejoins. | |
3871 | * - always assimilate scatterlock data (it can't hurt) | |
3872 | * - any surviving replica in SCATTER state -> SCATTER. otherwise, SYNC. | |
3873 | * - include base inode in ack for all inodes that saw scatterlock content | |
3874 | * | |
3875 | * also, for scatter gather, | |
3876 | * | |
3877 | * - auth increments {frag,r}stat.version on completion of any gather. | |
3878 | * | |
3879 | * - auth incorporates changes in a gather _only_ if the version | |
3880 | * matches. | |
3881 | * | |
3882 | * - replica discards changes any time the scatterlock syncs, and | |
3883 | * after recovery. | |
3884 | */ | |
3885 | ||
3886 | void MDCache::dump_rejoin_status(Formatter *f) const | |
3887 | { | |
3888 | f->open_object_section("rejoin_status"); | |
3889 | f->dump_stream("rejoin_gather") << rejoin_gather; | |
3890 | f->dump_stream("rejoin_ack_gather") << rejoin_ack_gather; | |
3891 | f->dump_unsigned("num_opening_inodes", cap_imports_num_opening); | |
3892 | f->close_section(); | |
3893 | } | |
3894 | ||
11fdf7f2 | 3895 | void MDCache::rejoin_start(MDSContext *rejoin_done_) |
7c673cae FG |
3896 | { |
3897 | dout(10) << "rejoin_start" << dendl; | |
11fdf7f2 | 3898 | ceph_assert(!rejoin_done); |
7c673cae FG |
3899 | rejoin_done.reset(rejoin_done_); |
3900 | ||
3901 | rejoin_gather = recovery_set; | |
3902 | // need finish opening cap inodes before sending cache rejoins | |
3903 | rejoin_gather.insert(mds->get_nodeid()); | |
3904 | process_imported_caps(); | |
3905 | } | |
3906 | ||
3907 | /* | |
3908 | * rejoin phase! | |
3909 | * | |
11fdf7f2 | 3910 | * this initiates rejoin. it should be called before we get any |
7c673cae FG |
3911 | * rejoin or rejoin_ack messages (or else mdsmap distribution is broken). |
3912 | * | |
3913 | * we start out by sending rejoins to everyone in the recovery set. | |
3914 | * | |
3915 | * if we are rejoin, send for all regions in our cache. | |
11fdf7f2 | 3916 | * if we are active|stopping, send only to nodes that are rejoining. |
7c673cae FG |
3917 | */ |
3918 | void MDCache::rejoin_send_rejoins() | |
3919 | { | |
3920 | dout(10) << "rejoin_send_rejoins with recovery_set " << recovery_set << dendl; | |
3921 | ||
3922 | if (rejoin_gather.count(mds->get_nodeid())) { | |
3923 | dout(7) << "rejoin_send_rejoins still processing imported caps, delaying" << dendl; | |
3924 | rejoins_pending = true; | |
3925 | return; | |
3926 | } | |
3927 | if (!resolve_gather.empty()) { | |
3928 | dout(7) << "rejoin_send_rejoins still waiting for resolves (" | |
3929 | << resolve_gather << ")" << dendl; | |
3930 | rejoins_pending = true; | |
3931 | return; | |
3932 | } | |
3933 | ||
11fdf7f2 TL |
3934 | ceph_assert(!migrator->is_importing()); |
3935 | ceph_assert(!migrator->is_exporting()); | |
7c673cae FG |
3936 | |
3937 | if (!mds->is_rejoin()) { | |
3938 | disambiguate_other_imports(); | |
3939 | } | |
3940 | ||
11fdf7f2 | 3941 | map<mds_rank_t, MMDSCacheRejoin::ref> rejoins; |
7c673cae FG |
3942 | |
3943 | ||
3944 | // if i am rejoining, send a rejoin to everyone. | |
3945 | // otherwise, just send to others who are rejoining. | |
3946 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); | |
3947 | p != recovery_set.end(); | |
3948 | ++p) { | |
3949 | if (*p == mds->get_nodeid()) continue; // nothing to myself! | |
3950 | if (rejoin_sent.count(*p)) continue; // already sent a rejoin to this node! | |
3951 | if (mds->is_rejoin()) | |
11fdf7f2 | 3952 | rejoins[*p] = MMDSCacheRejoin::create(MMDSCacheRejoin::OP_WEAK); |
7c673cae | 3953 | else if (mds->mdsmap->is_rejoin(*p)) |
11fdf7f2 | 3954 | rejoins[*p] = MMDSCacheRejoin::create(MMDSCacheRejoin::OP_STRONG); |
7c673cae FG |
3955 | } |
3956 | ||
3957 | if (mds->is_rejoin()) { | |
11fdf7f2 TL |
3958 | map<client_t, pair<Session*, set<mds_rank_t> > > client_exports; |
3959 | for (auto& p : cap_exports) { | |
3960 | mds_rank_t target = p.second.first; | |
7c673cae FG |
3961 | if (rejoins.count(target) == 0) |
3962 | continue; | |
11fdf7f2 TL |
3963 | for (auto q = p.second.second.begin(); q != p.second.second.end(); ) { |
3964 | Session *session = nullptr; | |
3965 | auto it = client_exports.find(q->first); | |
3966 | if (it != client_exports.end()) { | |
3967 | session = it->second.first; | |
3968 | if (session) | |
3969 | it->second.second.insert(target); | |
3970 | } else { | |
3971 | session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
3972 | auto& r = client_exports[q->first]; | |
3973 | r.first = session; | |
3974 | if (session) | |
3975 | r.second.insert(target); | |
3976 | } | |
3977 | if (session) { | |
3978 | ++q; | |
3979 | } else { | |
3980 | // remove reconnect with no session | |
3981 | p.second.second.erase(q++); | |
3982 | } | |
3983 | } | |
3984 | rejoins[target]->cap_exports[p.first] = p.second.second; | |
7c673cae | 3985 | } |
11fdf7f2 TL |
3986 | for (auto& p : client_exports) { |
3987 | Session *session = p.second.first; | |
3988 | for (auto& q : p.second.second) { | |
3989 | auto rejoin = rejoins[q]; | |
3990 | rejoin->client_map[p.first] = session->info.inst; | |
3991 | rejoin->client_metadata_map[p.first] = session->info.client_metadata; | |
3992 | } | |
7c673cae FG |
3993 | } |
3994 | } | |
3995 | ||
3996 | ||
3997 | // check all subtrees | |
3998 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
3999 | p != subtrees.end(); | |
4000 | ++p) { | |
4001 | CDir *dir = p->first; | |
11fdf7f2 | 4002 | ceph_assert(dir->is_subtree_root()); |
7c673cae FG |
4003 | if (dir->is_ambiguous_dir_auth()) { |
4004 | // exporter is recovering, importer is survivor. | |
11fdf7f2 TL |
4005 | ceph_assert(rejoins.count(dir->authority().first)); |
4006 | ceph_assert(!rejoins.count(dir->authority().second)); | |
7c673cae FG |
4007 | continue; |
4008 | } | |
4009 | ||
4010 | // my subtree? | |
4011 | if (dir->is_auth()) | |
4012 | continue; // skip my own regions! | |
4013 | ||
4014 | mds_rank_t auth = dir->get_dir_auth().first; | |
11fdf7f2 | 4015 | ceph_assert(auth >= 0); |
7c673cae FG |
4016 | if (rejoins.count(auth) == 0) |
4017 | continue; // don't care about this node's subtrees | |
4018 | ||
4019 | rejoin_walk(dir, rejoins[auth]); | |
4020 | } | |
4021 | ||
4022 | // rejoin root inodes, too | |
11fdf7f2 | 4023 | for (auto &p : rejoins) { |
7c673cae FG |
4024 | if (mds->is_rejoin()) { |
4025 | // weak | |
11fdf7f2 TL |
4026 | if (p.first == 0 && root) { |
4027 | p.second->add_weak_inode(root->vino()); | |
7c673cae FG |
4028 | if (root->is_dirty_scattered()) { |
4029 | dout(10) << " sending scatterlock state on root " << *root << dendl; | |
11fdf7f2 | 4030 | p.second->add_scatterlock_state(root); |
7c673cae FG |
4031 | } |
4032 | } | |
11fdf7f2 | 4033 | if (CInode *in = get_inode(MDS_INO_MDSDIR(p.first))) { |
7c673cae | 4034 | if (in) |
11fdf7f2 | 4035 | p.second->add_weak_inode(in->vino()); |
7c673cae FG |
4036 | } |
4037 | } else { | |
4038 | // strong | |
11fdf7f2 TL |
4039 | if (p.first == 0 && root) { |
4040 | p.second->add_strong_inode(root->vino(), | |
7c673cae FG |
4041 | root->get_replica_nonce(), |
4042 | root->get_caps_wanted(), | |
4043 | root->filelock.get_state(), | |
4044 | root->nestlock.get_state(), | |
4045 | root->dirfragtreelock.get_state()); | |
4046 | root->state_set(CInode::STATE_REJOINING); | |
4047 | if (root->is_dirty_scattered()) { | |
4048 | dout(10) << " sending scatterlock state on root " << *root << dendl; | |
11fdf7f2 | 4049 | p.second->add_scatterlock_state(root); |
7c673cae FG |
4050 | } |
4051 | } | |
4052 | ||
11fdf7f2 TL |
4053 | if (CInode *in = get_inode(MDS_INO_MDSDIR(p.first))) { |
4054 | p.second->add_strong_inode(in->vino(), | |
7c673cae FG |
4055 | in->get_replica_nonce(), |
4056 | in->get_caps_wanted(), | |
4057 | in->filelock.get_state(), | |
4058 | in->nestlock.get_state(), | |
4059 | in->dirfragtreelock.get_state()); | |
4060 | in->state_set(CInode::STATE_REJOINING); | |
4061 | } | |
4062 | } | |
4063 | } | |
4064 | ||
4065 | if (!mds->is_rejoin()) { | |
4066 | // i am survivor. send strong rejoin. | |
4067 | // note request remote_auth_pins, xlocks | |
4068 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
4069 | p != active_requests.end(); | |
4070 | ++p) { | |
4071 | MDRequestRef& mdr = p->second; | |
4072 | if (mdr->is_slave()) | |
4073 | continue; | |
4074 | // auth pins | |
11fdf7f2 TL |
4075 | for (const auto& q : mdr->remote_auth_pins) { |
4076 | if (!q.first->is_auth()) { | |
4077 | ceph_assert(q.second == q.first->authority().first); | |
4078 | if (rejoins.count(q.second) == 0) continue; | |
4079 | const MMDSCacheRejoin::ref &rejoin = rejoins[q.second]; | |
7c673cae | 4080 | |
11fdf7f2 | 4081 | dout(15) << " " << *mdr << " authpin on " << *q.first << dendl; |
7c673cae | 4082 | MDSCacheObjectInfo i; |
11fdf7f2 | 4083 | q.first->set_object_info(i); |
7c673cae FG |
4084 | if (i.ino) |
4085 | rejoin->add_inode_authpin(vinodeno_t(i.ino, i.snapid), mdr->reqid, mdr->attempt); | |
4086 | else | |
4087 | rejoin->add_dentry_authpin(i.dirfrag, i.dname, i.snapid, mdr->reqid, mdr->attempt); | |
4088 | ||
4089 | if (mdr->has_more() && mdr->more()->is_remote_frozen_authpin && | |
11fdf7f2 | 4090 | mdr->more()->rename_inode == q.first) |
7c673cae FG |
4091 | rejoin->add_inode_frozen_authpin(vinodeno_t(i.ino, i.snapid), |
4092 | mdr->reqid, mdr->attempt); | |
4093 | } | |
4094 | } | |
4095 | // xlocks | |
11fdf7f2 TL |
4096 | for (const auto& q : mdr->locks) { |
4097 | auto lock = q.lock; | |
4098 | auto obj = lock->get_parent(); | |
4099 | if (q.is_xlock() && !obj->is_auth()) { | |
4100 | mds_rank_t who = obj->authority().first; | |
7c673cae | 4101 | if (rejoins.count(who) == 0) continue; |
11fdf7f2 | 4102 | const MMDSCacheRejoin::ref &rejoin = rejoins[who]; |
7c673cae | 4103 | |
11fdf7f2 | 4104 | dout(15) << " " << *mdr << " xlock on " << *lock << " " << *obj << dendl; |
7c673cae | 4105 | MDSCacheObjectInfo i; |
11fdf7f2 | 4106 | obj->set_object_info(i); |
7c673cae | 4107 | if (i.ino) |
11fdf7f2 | 4108 | rejoin->add_inode_xlock(vinodeno_t(i.ino, i.snapid), lock->get_type(), |
7c673cae FG |
4109 | mdr->reqid, mdr->attempt); |
4110 | else | |
4111 | rejoin->add_dentry_xlock(i.dirfrag, i.dname, i.snapid, | |
4112 | mdr->reqid, mdr->attempt); | |
11fdf7f2 TL |
4113 | } else if (q.is_remote_wrlock()) { |
4114 | mds_rank_t who = q.wrlock_target; | |
4115 | if (rejoins.count(who) == 0) continue; | |
4116 | const MMDSCacheRejoin::ref &rejoin = rejoins[who]; | |
7c673cae | 4117 | |
11fdf7f2 TL |
4118 | dout(15) << " " << *mdr << " wrlock on " << *lock << " " << *obj << dendl; |
4119 | MDSCacheObjectInfo i; | |
4120 | obj->set_object_info(i); | |
4121 | ceph_assert(i.ino); | |
4122 | rejoin->add_inode_wrlock(vinodeno_t(i.ino, i.snapid), lock->get_type(), | |
4123 | mdr->reqid, mdr->attempt); | |
4124 | } | |
7c673cae FG |
4125 | } |
4126 | } | |
4127 | } | |
4128 | ||
4129 | // send the messages | |
11fdf7f2 TL |
4130 | for (auto &p : rejoins) { |
4131 | ceph_assert(rejoin_sent.count(p.first) == 0); | |
4132 | ceph_assert(rejoin_ack_gather.count(p.first) == 0); | |
4133 | rejoin_sent.insert(p.first); | |
4134 | rejoin_ack_gather.insert(p.first); | |
4135 | mds->send_message_mds(p.second, p.first); | |
7c673cae FG |
4136 | } |
4137 | rejoin_ack_gather.insert(mds->get_nodeid()); // we need to complete rejoin_gather_finish, too | |
4138 | rejoins_pending = false; | |
4139 | ||
4140 | // nothing? | |
28e407b8 | 4141 | if (mds->is_rejoin() && rejoin_gather.empty()) { |
7c673cae FG |
4142 | dout(10) << "nothing to rejoin" << dendl; |
4143 | rejoin_gather_finish(); | |
4144 | } | |
4145 | } | |
4146 | ||
4147 | ||
4148 | /** | |
4149 | * rejoin_walk - build rejoin declarations for a subtree | |
4150 | * | |
4151 | * @param dir subtree root | |
4152 | * @param rejoin rejoin message | |
4153 | * | |
4154 | * from a rejoining node: | |
4155 | * weak dirfrag | |
4156 | * weak dentries (w/ connectivity) | |
4157 | * | |
4158 | * from a surviving node: | |
4159 | * strong dirfrag | |
4160 | * strong dentries (no connectivity!) | |
4161 | * strong inodes | |
4162 | */ | |
11fdf7f2 | 4163 | void MDCache::rejoin_walk(CDir *dir, const MMDSCacheRejoin::ref &rejoin) |
7c673cae FG |
4164 | { |
4165 | dout(10) << "rejoin_walk " << *dir << dendl; | |
4166 | ||
4167 | list<CDir*> nested; // finish this dir, then do nested items | |
4168 | ||
4169 | if (mds->is_rejoin()) { | |
4170 | // WEAK | |
4171 | rejoin->add_weak_dirfrag(dir->dirfrag()); | |
94b18763 FG |
4172 | for (auto &p : dir->items) { |
4173 | CDentry *dn = p.second; | |
11fdf7f2 | 4174 | ceph_assert(dn->last == CEPH_NOSNAP); |
7c673cae FG |
4175 | CDentry::linkage_t *dnl = dn->get_linkage(); |
4176 | dout(15) << " add_weak_primary_dentry " << *dn << dendl; | |
11fdf7f2 | 4177 | ceph_assert(dnl->is_primary()); |
7c673cae | 4178 | CInode *in = dnl->get_inode(); |
11fdf7f2 | 4179 | ceph_assert(dnl->get_inode()->is_dir()); |
94b18763 | 4180 | rejoin->add_weak_primary_dentry(dir->ino(), dn->get_name(), dn->first, dn->last, in->ino()); |
7c673cae FG |
4181 | in->get_nested_dirfrags(nested); |
4182 | if (in->is_dirty_scattered()) { | |
4183 | dout(10) << " sending scatterlock state on " << *in << dendl; | |
4184 | rejoin->add_scatterlock_state(in); | |
4185 | } | |
4186 | } | |
4187 | } else { | |
4188 | // STRONG | |
4189 | dout(15) << " add_strong_dirfrag " << *dir << dendl; | |
4190 | rejoin->add_strong_dirfrag(dir->dirfrag(), dir->get_replica_nonce(), dir->get_dir_rep()); | |
4191 | dir->state_set(CDir::STATE_REJOINING); | |
4192 | ||
11fdf7f2 | 4193 | for (auto it = dir->items.begin(); it != dir->items.end(); ) { |
94b18763 | 4194 | CDentry *dn = it->second; |
11fdf7f2 TL |
4195 | ++it; |
4196 | dn->state_set(CDentry::STATE_REJOINING); | |
7c673cae | 4197 | CDentry::linkage_t *dnl = dn->get_linkage(); |
11fdf7f2 TL |
4198 | CInode *in = dnl->is_primary() ? dnl->get_inode() : NULL; |
4199 | ||
4200 | // trim snap dentries. because they may have been pruned by | |
4201 | // their auth mds (snap deleted) | |
4202 | if (dn->last != CEPH_NOSNAP) { | |
4203 | if (in && !in->remote_parents.empty()) { | |
4204 | // unlink any stale remote snap dentry. | |
4205 | for (auto it2 = in->remote_parents.begin(); it2 != in->remote_parents.end(); ) { | |
4206 | CDentry *remote_dn = *it2; | |
4207 | ++it2; | |
4208 | ceph_assert(remote_dn->last != CEPH_NOSNAP); | |
4209 | remote_dn->unlink_remote(remote_dn->get_linkage()); | |
4210 | } | |
4211 | } | |
4212 | if (dn->lru_is_expireable()) { | |
4213 | if (!dnl->is_null()) | |
4214 | dir->unlink_inode(dn, false); | |
4215 | if (in) | |
4216 | remove_inode(in); | |
4217 | dir->remove_dentry(dn); | |
4218 | continue; | |
4219 | } else { | |
4220 | // Inventing null/remote dentry shouldn't cause problem | |
4221 | ceph_assert(!dnl->is_primary()); | |
4222 | } | |
4223 | } | |
4224 | ||
7c673cae | 4225 | dout(15) << " add_strong_dentry " << *dn << dendl; |
94b18763 | 4226 | rejoin->add_strong_dentry(dir->dirfrag(), dn->get_name(), dn->first, dn->last, |
7c673cae FG |
4227 | dnl->is_primary() ? dnl->get_inode()->ino():inodeno_t(0), |
4228 | dnl->is_remote() ? dnl->get_remote_ino():inodeno_t(0), | |
4229 | dnl->is_remote() ? dnl->get_remote_d_type():0, | |
4230 | dn->get_replica_nonce(), | |
4231 | dn->lock.get_state()); | |
4232 | dn->state_set(CDentry::STATE_REJOINING); | |
4233 | if (dnl->is_primary()) { | |
4234 | CInode *in = dnl->get_inode(); | |
4235 | dout(15) << " add_strong_inode " << *in << dendl; | |
4236 | rejoin->add_strong_inode(in->vino(), | |
4237 | in->get_replica_nonce(), | |
4238 | in->get_caps_wanted(), | |
4239 | in->filelock.get_state(), | |
4240 | in->nestlock.get_state(), | |
4241 | in->dirfragtreelock.get_state()); | |
4242 | in->state_set(CInode::STATE_REJOINING); | |
4243 | in->get_nested_dirfrags(nested); | |
4244 | if (in->is_dirty_scattered()) { | |
4245 | dout(10) << " sending scatterlock state on " << *in << dendl; | |
4246 | rejoin->add_scatterlock_state(in); | |
4247 | } | |
4248 | } | |
4249 | } | |
4250 | } | |
4251 | ||
4252 | // recurse into nested dirs | |
4253 | for (list<CDir*>::iterator p = nested.begin(); | |
4254 | p != nested.end(); | |
4255 | ++p) | |
4256 | rejoin_walk(*p, rejoin); | |
4257 | } | |
4258 | ||
4259 | ||
4260 | /* | |
4261 | * i got a rejoin. | |
4262 | * - reply with the lockstate | |
4263 | * | |
4264 | * if i am active|stopping, | |
4265 | * - remove source from replica list for everything not referenced here. | |
7c673cae | 4266 | */ |
11fdf7f2 | 4267 | void MDCache::handle_cache_rejoin(const MMDSCacheRejoin::const_ref &m) |
7c673cae FG |
4268 | { |
4269 | dout(7) << "handle_cache_rejoin " << *m << " from " << m->get_source() | |
4270 | << " (" << m->get_payload().length() << " bytes)" | |
4271 | << dendl; | |
4272 | ||
4273 | switch (m->op) { | |
4274 | case MMDSCacheRejoin::OP_WEAK: | |
4275 | handle_cache_rejoin_weak(m); | |
4276 | break; | |
4277 | case MMDSCacheRejoin::OP_STRONG: | |
4278 | handle_cache_rejoin_strong(m); | |
4279 | break; | |
4280 | case MMDSCacheRejoin::OP_ACK: | |
4281 | handle_cache_rejoin_ack(m); | |
4282 | break; | |
4283 | ||
4284 | default: | |
4285 | ceph_abort(); | |
4286 | } | |
7c673cae FG |
4287 | } |
4288 | ||
4289 | ||
4290 | /* | |
4291 | * handle_cache_rejoin_weak | |
4292 | * | |
4293 | * the sender | |
4294 | * - is recovering from their journal. | |
4295 | * - may have incorrect (out of date) inode contents | |
4296 | * - will include weak dirfrag if sender is dirfrag auth and parent inode auth is recipient | |
4297 | * | |
4298 | * if the sender didn't trim_non_auth(), they | |
4299 | * - may have incorrect (out of date) dentry/inode linkage | |
4300 | * - may have deleted/purged inodes | |
4301 | * and i may have to go to disk to get accurate inode contents. yuck. | |
7c673cae | 4302 | */ |
11fdf7f2 | 4303 | void MDCache::handle_cache_rejoin_weak(const MMDSCacheRejoin::const_ref &weak) |
7c673cae FG |
4304 | { |
4305 | mds_rank_t from = mds_rank_t(weak->get_source().num()); | |
4306 | ||
4307 | // possible response(s) | |
11fdf7f2 | 4308 | MMDSCacheRejoin::ref ack; // if survivor |
7c673cae FG |
4309 | set<vinodeno_t> acked_inodes; // if survivor |
4310 | set<SimpleLock *> gather_locks; // if survivor | |
4311 | bool survivor = false; // am i a survivor? | |
4312 | ||
4313 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
4314 | survivor = true; | |
4315 | dout(10) << "i am a surivivor, and will ack immediately" << dendl; | |
11fdf7f2 | 4316 | ack = MMDSCacheRejoin::create(MMDSCacheRejoin::OP_ACK); |
7c673cae FG |
4317 | |
4318 | map<inodeno_t,map<client_t,Capability::Import> > imported_caps; | |
4319 | ||
4320 | // check cap exports | |
4321 | for (auto p = weak->cap_exports.begin(); p != weak->cap_exports.end(); ++p) { | |
4322 | CInode *in = get_inode(p->first); | |
11fdf7f2 | 4323 | ceph_assert(!in || in->is_auth()); |
7c673cae FG |
4324 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { |
4325 | dout(10) << " claiming cap import " << p->first << " client." << q->first << " on " << *in << dendl; | |
4326 | Capability *cap = rejoin_import_cap(in, q->first, q->second, from); | |
4327 | Capability::Import& im = imported_caps[p->first][q->first]; | |
4328 | if (cap) { | |
4329 | im.cap_id = cap->get_cap_id(); | |
4330 | im.issue_seq = cap->get_last_seq(); | |
4331 | im.mseq = cap->get_mseq(); | |
4332 | } else { | |
4333 | // all are zero | |
4334 | } | |
4335 | } | |
4336 | mds->locker->eval(in, CEPH_CAP_LOCKS, true); | |
4337 | } | |
4338 | ||
11fdf7f2 | 4339 | encode(imported_caps, ack->imported_caps); |
7c673cae | 4340 | } else { |
11fdf7f2 | 4341 | ceph_assert(mds->is_rejoin()); |
7c673cae FG |
4342 | |
4343 | // we may have already received a strong rejoin from the sender. | |
4344 | rejoin_scour_survivor_replicas(from, NULL, acked_inodes, gather_locks); | |
11fdf7f2 | 4345 | ceph_assert(gather_locks.empty()); |
7c673cae FG |
4346 | |
4347 | // check cap exports. | |
4348 | rejoin_client_map.insert(weak->client_map.begin(), weak->client_map.end()); | |
11fdf7f2 TL |
4349 | rejoin_client_metadata_map.insert(weak->client_metadata_map.begin(), |
4350 | weak->client_metadata_map.end()); | |
7c673cae FG |
4351 | |
4352 | for (auto p = weak->cap_exports.begin(); p != weak->cap_exports.end(); ++p) { | |
4353 | CInode *in = get_inode(p->first); | |
11fdf7f2 | 4354 | ceph_assert(!in || in->is_auth()); |
7c673cae FG |
4355 | // note |
4356 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
4357 | dout(10) << " claiming cap import " << p->first << " client." << q->first << dendl; | |
4358 | cap_imports[p->first][q->first][from] = q->second; | |
4359 | } | |
4360 | } | |
4361 | } | |
4362 | ||
4363 | // assimilate any potentially dirty scatterlock state | |
11fdf7f2 TL |
4364 | for (const auto &p : weak->inode_scatterlocks) { |
4365 | CInode *in = get_inode(p.first); | |
4366 | ceph_assert(in); | |
4367 | in->decode_lock_state(CEPH_LOCK_IFILE, p.second.file); | |
4368 | in->decode_lock_state(CEPH_LOCK_INEST, p.second.nest); | |
4369 | in->decode_lock_state(CEPH_LOCK_IDFT, p.second.dft); | |
7c673cae FG |
4370 | if (!survivor) |
4371 | rejoin_potential_updated_scatterlocks.insert(in); | |
4372 | } | |
4373 | ||
4374 | // recovering peer may send incorrect dirfrags here. we need to | |
4375 | // infer which dirfrag they meant. the ack will include a | |
4376 | // strong_dirfrag that will set them straight on the fragmentation. | |
4377 | ||
4378 | // walk weak map | |
4379 | set<CDir*> dirs_to_share; | |
11fdf7f2 TL |
4380 | for (const auto &p : weak->weak_dirfrags) { |
4381 | CInode *diri = get_inode(p.ino); | |
7c673cae | 4382 | if (!diri) |
11fdf7f2 TL |
4383 | dout(0) << " missing dir ino " << p.ino << dendl; |
4384 | ceph_assert(diri); | |
7c673cae | 4385 | |
11fdf7f2 TL |
4386 | frag_vec_t leaves; |
4387 | if (diri->dirfragtree.is_leaf(p.frag)) { | |
4388 | leaves.push_back(p.frag); | |
7c673cae | 4389 | } else { |
11fdf7f2 TL |
4390 | diri->dirfragtree.get_leaves_under(p.frag, leaves); |
4391 | if (leaves.empty()) | |
4392 | leaves.push_back(diri->dirfragtree[p.frag.value()]); | |
7c673cae | 4393 | } |
11fdf7f2 TL |
4394 | for (const auto& leaf : leaves) { |
4395 | CDir *dir = diri->get_dirfrag(leaf); | |
7c673cae | 4396 | if (!dir) { |
11fdf7f2 | 4397 | dout(0) << " missing dir for " << p.frag << " (which maps to " << leaf << ") on " << *diri << dendl; |
7c673cae FG |
4398 | continue; |
4399 | } | |
11fdf7f2 | 4400 | ceph_assert(dir); |
7c673cae | 4401 | if (dirs_to_share.count(dir)) { |
11fdf7f2 | 4402 | dout(10) << " already have " << p.frag << " -> " << leaf << " " << *dir << dendl; |
7c673cae FG |
4403 | } else { |
4404 | dirs_to_share.insert(dir); | |
4405 | unsigned nonce = dir->add_replica(from); | |
11fdf7f2 | 4406 | dout(10) << " have " << p.frag << " -> " << leaf << " " << *dir << dendl; |
7c673cae FG |
4407 | if (ack) { |
4408 | ack->add_strong_dirfrag(dir->dirfrag(), nonce, dir->dir_rep); | |
4409 | ack->add_dirfrag_base(dir); | |
4410 | } | |
4411 | } | |
4412 | } | |
4413 | } | |
4414 | ||
11fdf7f2 TL |
4415 | for (const auto &p : weak->weak) { |
4416 | CInode *diri = get_inode(p.first); | |
7c673cae | 4417 | if (!diri) |
11fdf7f2 TL |
4418 | dout(0) << " missing dir ino " << p.first << dendl; |
4419 | ceph_assert(diri); | |
7c673cae FG |
4420 | |
4421 | // weak dentries | |
4422 | CDir *dir = 0; | |
11fdf7f2 | 4423 | for (const auto &q : p.second) { |
7c673cae FG |
4424 | // locate proper dirfrag. |
4425 | // optimize for common case (one dirfrag) to avoid dirs_to_share set check | |
11fdf7f2 | 4426 | frag_t fg = diri->pick_dirfrag(q.first.name); |
7c673cae FG |
4427 | if (!dir || dir->get_frag() != fg) { |
4428 | dir = diri->get_dirfrag(fg); | |
4429 | if (!dir) | |
4430 | dout(0) << " missing dir frag " << fg << " on " << *diri << dendl; | |
11fdf7f2 TL |
4431 | ceph_assert(dir); |
4432 | ceph_assert(dirs_to_share.count(dir)); | |
7c673cae FG |
4433 | } |
4434 | ||
4435 | // and dentry | |
11fdf7f2 TL |
4436 | CDentry *dn = dir->lookup(q.first.name, q.first.snapid); |
4437 | ceph_assert(dn); | |
7c673cae | 4438 | CDentry::linkage_t *dnl = dn->get_linkage(); |
11fdf7f2 | 4439 | ceph_assert(dnl->is_primary()); |
7c673cae FG |
4440 | |
4441 | if (survivor && dn->is_replica(from)) | |
4442 | dentry_remove_replica(dn, from, gather_locks); | |
4443 | unsigned dnonce = dn->add_replica(from); | |
4444 | dout(10) << " have " << *dn << dendl; | |
4445 | if (ack) | |
94b18763 | 4446 | ack->add_strong_dentry(dir->dirfrag(), dn->get_name(), dn->first, dn->last, |
7c673cae FG |
4447 | dnl->get_inode()->ino(), inodeno_t(0), 0, |
4448 | dnonce, dn->lock.get_replica_state()); | |
4449 | ||
4450 | // inode | |
4451 | CInode *in = dnl->get_inode(); | |
11fdf7f2 | 4452 | ceph_assert(in); |
7c673cae FG |
4453 | |
4454 | if (survivor && in->is_replica(from)) | |
4455 | inode_remove_replica(in, from, true, gather_locks); | |
4456 | unsigned inonce = in->add_replica(from); | |
4457 | dout(10) << " have " << *in << dendl; | |
4458 | ||
4459 | // scatter the dirlock, just in case? | |
4460 | if (!survivor && in->is_dir() && in->has_subtree_root_dirfrag()) | |
4461 | in->filelock.set_state(LOCK_MIX); | |
4462 | ||
4463 | if (ack) { | |
4464 | acked_inodes.insert(in->vino()); | |
4465 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4466 | bufferlist bl; | |
4467 | in->_encode_locks_state_for_rejoin(bl, from); | |
4468 | ack->add_inode_locks(in, inonce, bl); | |
4469 | } | |
4470 | } | |
4471 | } | |
4472 | ||
4473 | // weak base inodes? (root, stray, etc.) | |
4474 | for (set<vinodeno_t>::iterator p = weak->weak_inodes.begin(); | |
4475 | p != weak->weak_inodes.end(); | |
4476 | ++p) { | |
4477 | CInode *in = get_inode(*p); | |
11fdf7f2 | 4478 | ceph_assert(in); // hmm fixme wrt stray? |
7c673cae FG |
4479 | if (survivor && in->is_replica(from)) |
4480 | inode_remove_replica(in, from, true, gather_locks); | |
4481 | unsigned inonce = in->add_replica(from); | |
4482 | dout(10) << " have base " << *in << dendl; | |
4483 | ||
4484 | if (ack) { | |
4485 | acked_inodes.insert(in->vino()); | |
4486 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4487 | bufferlist bl; | |
4488 | in->_encode_locks_state_for_rejoin(bl, from); | |
4489 | ack->add_inode_locks(in, inonce, bl); | |
4490 | } | |
4491 | } | |
4492 | ||
11fdf7f2 | 4493 | ceph_assert(rejoin_gather.count(from)); |
7c673cae FG |
4494 | rejoin_gather.erase(from); |
4495 | if (survivor) { | |
4496 | // survivor. do everything now. | |
11fdf7f2 TL |
4497 | for (const auto &p : weak->inode_scatterlocks) { |
4498 | CInode *in = get_inode(p.first); | |
4499 | ceph_assert(in); | |
7c673cae FG |
4500 | dout(10) << " including base inode (due to potential scatterlock update) " << *in << dendl; |
4501 | acked_inodes.insert(in->vino()); | |
4502 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4503 | } | |
4504 | ||
4505 | rejoin_scour_survivor_replicas(from, ack, acked_inodes, gather_locks); | |
4506 | mds->send_message(ack, weak->get_connection()); | |
4507 | ||
4508 | for (set<SimpleLock*>::iterator p = gather_locks.begin(); p != gather_locks.end(); ++p) { | |
4509 | if (!(*p)->is_stable()) | |
4510 | mds->locker->eval_gather(*p); | |
4511 | } | |
4512 | } else { | |
4513 | // done? | |
28e407b8 | 4514 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) { |
7c673cae FG |
4515 | rejoin_gather_finish(); |
4516 | } else { | |
4517 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" << dendl; | |
4518 | } | |
4519 | } | |
4520 | } | |
4521 | ||
7c673cae FG |
4522 | /* |
4523 | * rejoin_scour_survivor_replica - remove source from replica list on unmentioned objects | |
4524 | * | |
4525 | * all validated replicas are acked with a strong nonce, etc. if that isn't in the | |
4526 | * ack, the replica dne, and we can remove it from our replica maps. | |
4527 | */ | |
11fdf7f2 | 4528 | void MDCache::rejoin_scour_survivor_replicas(mds_rank_t from, const MMDSCacheRejoin::const_ref &ack, |
7c673cae FG |
4529 | set<vinodeno_t>& acked_inodes, |
4530 | set<SimpleLock *>& gather_locks) | |
4531 | { | |
4532 | dout(10) << "rejoin_scour_survivor_replicas from mds." << from << dendl; | |
4533 | ||
b32b8144 | 4534 | auto scour_func = [this, from, ack, &acked_inodes, &gather_locks] (CInode *in) { |
7c673cae FG |
4535 | // inode? |
4536 | if (in->is_auth() && | |
4537 | in->is_replica(from) && | |
b32b8144 | 4538 | (ack == NULL || acked_inodes.count(in->vino()) == 0)) { |
7c673cae FG |
4539 | inode_remove_replica(in, from, false, gather_locks); |
4540 | dout(10) << " rem " << *in << dendl; | |
4541 | } | |
4542 | ||
b32b8144 FG |
4543 | if (!in->is_dir()) |
4544 | return; | |
7c673cae FG |
4545 | |
4546 | list<CDir*> dfs; | |
4547 | in->get_dirfrags(dfs); | |
4548 | for (list<CDir*>::iterator p = dfs.begin(); | |
4549 | p != dfs.end(); | |
4550 | ++p) { | |
4551 | CDir *dir = *p; | |
181888fb FG |
4552 | if (!dir->is_auth()) |
4553 | continue; | |
7c673cae | 4554 | |
181888fb | 4555 | if (dir->is_replica(from) && |
7c673cae FG |
4556 | (ack == NULL || ack->strong_dirfrags.count(dir->dirfrag()) == 0)) { |
4557 | dir->remove_replica(from); | |
4558 | dout(10) << " rem " << *dir << dendl; | |
4559 | } | |
4560 | ||
4561 | // dentries | |
94b18763 FG |
4562 | for (auto &p : dir->items) { |
4563 | CDentry *dn = p.second; | |
7c673cae | 4564 | |
11fdf7f2 TL |
4565 | if (dn->is_replica(from)) { |
4566 | if (ack) { | |
4567 | const auto it = ack->strong_dentries.find(dir->dirfrag()); | |
4568 | if (it != ack->strong_dentries.end() && it->second.count(string_snap_t(dn->get_name(), dn->last)) > 0) { | |
4569 | continue; | |
4570 | } | |
4571 | } | |
7c673cae FG |
4572 | dentry_remove_replica(dn, from, gather_locks); |
4573 | dout(10) << " rem " << *dn << dendl; | |
4574 | } | |
4575 | } | |
4576 | } | |
b32b8144 FG |
4577 | }; |
4578 | ||
94b18763 | 4579 | for (auto &p : inode_map) |
b32b8144 | 4580 | scour_func(p.second); |
94b18763 | 4581 | for (auto &p : snap_inode_map) |
b32b8144 | 4582 | scour_func(p.second); |
7c673cae FG |
4583 | } |
4584 | ||
4585 | ||
4586 | CInode *MDCache::rejoin_invent_inode(inodeno_t ino, snapid_t last) | |
4587 | { | |
4588 | CInode *in = new CInode(this, true, 1, last); | |
4589 | in->inode.ino = ino; | |
4590 | in->state_set(CInode::STATE_REJOINUNDEF); | |
4591 | add_inode(in); | |
4592 | rejoin_undef_inodes.insert(in); | |
4593 | dout(10) << " invented " << *in << dendl; | |
4594 | return in; | |
4595 | } | |
4596 | ||
4597 | CDir *MDCache::rejoin_invent_dirfrag(dirfrag_t df) | |
4598 | { | |
4599 | CInode *in = get_inode(df.ino); | |
4600 | if (!in) | |
4601 | in = rejoin_invent_inode(df.ino, CEPH_NOSNAP); | |
4602 | if (!in->is_dir()) { | |
11fdf7f2 | 4603 | ceph_assert(in->state_test(CInode::STATE_REJOINUNDEF)); |
7c673cae | 4604 | in->inode.mode = S_IFDIR; |
11fdf7f2 | 4605 | in->inode.dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; |
7c673cae FG |
4606 | } |
4607 | CDir *dir = in->get_or_open_dirfrag(this, df.frag); | |
4608 | dir->state_set(CDir::STATE_REJOINUNDEF); | |
4609 | rejoin_undef_dirfrags.insert(dir); | |
4610 | dout(10) << " invented " << *dir << dendl; | |
4611 | return dir; | |
4612 | } | |
4613 | ||
11fdf7f2 | 4614 | void MDCache::handle_cache_rejoin_strong(const MMDSCacheRejoin::const_ref &strong) |
7c673cae FG |
4615 | { |
4616 | mds_rank_t from = mds_rank_t(strong->get_source().num()); | |
4617 | ||
4618 | // only a recovering node will get a strong rejoin. | |
a8e16298 TL |
4619 | if (!mds->is_rejoin()) { |
4620 | if (mds->get_want_state() == MDSMap::STATE_REJOIN) { | |
4621 | mds->wait_for_rejoin(new C_MDS_RetryMessage(mds, strong)); | |
4622 | return; | |
4623 | } | |
11fdf7f2 | 4624 | ceph_abort_msg("got unexpected rejoin message during recovery"); |
a8e16298 | 4625 | } |
7c673cae FG |
4626 | |
4627 | // assimilate any potentially dirty scatterlock state | |
11fdf7f2 TL |
4628 | for (const auto &p : strong->inode_scatterlocks) { |
4629 | CInode *in = get_inode(p.first); | |
4630 | ceph_assert(in); | |
4631 | in->decode_lock_state(CEPH_LOCK_IFILE, p.second.file); | |
4632 | in->decode_lock_state(CEPH_LOCK_INEST, p.second.nest); | |
4633 | in->decode_lock_state(CEPH_LOCK_IDFT, p.second.dft); | |
7c673cae FG |
4634 | rejoin_potential_updated_scatterlocks.insert(in); |
4635 | } | |
4636 | ||
4637 | rejoin_unlinked_inodes[from].clear(); | |
4638 | ||
4639 | // surviving peer may send incorrect dirfrag here (maybe they didn't | |
4640 | // get the fragment notify, or maybe we rolled back?). we need to | |
4641 | // infer the right frag and get them with the program. somehow. | |
4642 | // we don't normally send ACK.. so we'll need to bundle this with | |
4643 | // MISSING or something. | |
4644 | ||
4645 | // strong dirfrags/dentries. | |
4646 | // also process auth_pins, xlocks. | |
11fdf7f2 TL |
4647 | for (const auto &p : strong->strong_dirfrags) { |
4648 | auto& dirfrag = p.first; | |
4649 | CInode *diri = get_inode(dirfrag.ino); | |
7c673cae | 4650 | if (!diri) |
11fdf7f2 TL |
4651 | diri = rejoin_invent_inode(dirfrag.ino, CEPH_NOSNAP); |
4652 | CDir *dir = diri->get_dirfrag(dirfrag.frag); | |
7c673cae FG |
4653 | bool refragged = false; |
4654 | if (dir) { | |
4655 | dout(10) << " have " << *dir << dendl; | |
4656 | } else { | |
4657 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) | |
4658 | dir = rejoin_invent_dirfrag(dirfrag_t(diri->ino(), frag_t())); | |
11fdf7f2 TL |
4659 | else if (diri->dirfragtree.is_leaf(dirfrag.frag)) |
4660 | dir = rejoin_invent_dirfrag(dirfrag); | |
7c673cae FG |
4661 | } |
4662 | if (dir) { | |
11fdf7f2 TL |
4663 | dir->add_replica(from, p.second.nonce); |
4664 | dir->dir_rep = p.second.dir_rep; | |
7c673cae | 4665 | } else { |
11fdf7f2 TL |
4666 | dout(10) << " frag " << dirfrag << " doesn't match dirfragtree " << *diri << dendl; |
4667 | frag_vec_t leaves; | |
4668 | diri->dirfragtree.get_leaves_under(dirfrag.frag, leaves); | |
4669 | if (leaves.empty()) | |
4670 | leaves.push_back(diri->dirfragtree[dirfrag.frag.value()]); | |
4671 | dout(10) << " maps to frag(s) " << leaves << dendl; | |
4672 | for (const auto& leaf : leaves) { | |
4673 | CDir *dir = diri->get_dirfrag(leaf); | |
7c673cae | 4674 | if (!dir) |
11fdf7f2 | 4675 | dir = rejoin_invent_dirfrag(dirfrag_t(diri->ino(), leaf)); |
7c673cae FG |
4676 | else |
4677 | dout(10) << " have(approx) " << *dir << dendl; | |
11fdf7f2 TL |
4678 | dir->add_replica(from, p.second.nonce); |
4679 | dir->dir_rep = p.second.dir_rep; | |
7c673cae FG |
4680 | } |
4681 | refragged = true; | |
4682 | } | |
4683 | ||
11fdf7f2 TL |
4684 | const auto it = strong->strong_dentries.find(dirfrag); |
4685 | if (it != strong->strong_dentries.end()) { | |
4686 | const map<string_snap_t,MMDSCacheRejoin::dn_strong>& dmap = it->second; | |
4687 | for (const auto &q : dmap) { | |
4688 | const string_snap_t& ss = q.first; | |
4689 | const MMDSCacheRejoin::dn_strong& d = q.second; | |
4690 | CDentry *dn; | |
4691 | if (!refragged) | |
4692 | dn = dir->lookup(ss.name, ss.snapid); | |
4693 | else { | |
4694 | frag_t fg = diri->pick_dirfrag(ss.name); | |
4695 | dir = diri->get_dirfrag(fg); | |
4696 | ceph_assert(dir); | |
4697 | dn = dir->lookup(ss.name, ss.snapid); | |
4698 | } | |
4699 | if (!dn) { | |
4700 | if (d.is_remote()) { | |
4701 | dn = dir->add_remote_dentry(ss.name, d.remote_ino, d.remote_d_type, d.first, ss.snapid); | |
4702 | } else if (d.is_null()) { | |
4703 | dn = dir->add_null_dentry(ss.name, d.first, ss.snapid); | |
4704 | } else { | |
4705 | CInode *in = get_inode(d.ino, ss.snapid); | |
4706 | if (!in) in = rejoin_invent_inode(d.ino, ss.snapid); | |
4707 | dn = dir->add_primary_dentry(ss.name, in, d.first, ss.snapid); | |
4708 | } | |
4709 | dout(10) << " invented " << *dn << dendl; | |
4710 | } | |
4711 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4712 | ||
4713 | // dn auth_pin? | |
4714 | const auto pinned_it = strong->authpinned_dentries.find(dirfrag); | |
4715 | if (pinned_it != strong->authpinned_dentries.end()) { | |
4716 | const auto slave_reqid_it = pinned_it->second.find(ss); | |
4717 | if (slave_reqid_it != pinned_it->second.end()) { | |
4718 | for (const auto &r : slave_reqid_it->second) { | |
4719 | dout(10) << " dn authpin by " << r << " on " << *dn << dendl; | |
4720 | ||
4721 | // get/create slave mdrequest | |
4722 | MDRequestRef mdr; | |
4723 | if (have_request(r.reqid)) | |
4724 | mdr = request_get(r.reqid); | |
4725 | else | |
4726 | mdr = request_start_slave(r.reqid, r.attempt, strong); | |
4727 | mdr->auth_pin(dn); | |
4728 | } | |
4729 | } | |
7c673cae | 4730 | } |
7c673cae | 4731 | |
11fdf7f2 TL |
4732 | // dn xlock? |
4733 | const auto xlocked_it = strong->xlocked_dentries.find(dirfrag); | |
4734 | if (xlocked_it != strong->xlocked_dentries.end()) { | |
4735 | const auto ss_req_it = xlocked_it->second.find(ss); | |
4736 | if (ss_req_it != xlocked_it->second.end()) { | |
4737 | const MMDSCacheRejoin::slave_reqid& r = ss_req_it->second; | |
4738 | dout(10) << " dn xlock by " << r << " on " << *dn << dendl; | |
4739 | MDRequestRef mdr = request_get(r.reqid); // should have this from auth_pin above. | |
4740 | ceph_assert(mdr->is_auth_pinned(dn)); | |
4741 | if (!mdr->is_xlocked(&dn->versionlock)) { | |
4742 | ceph_assert(dn->versionlock.can_xlock_local()); | |
4743 | dn->versionlock.get_xlock(mdr, mdr->get_client()); | |
4744 | mdr->locks.emplace(&dn->versionlock, MutationImpl::LockOp::XLOCK); | |
4745 | } | |
4746 | if (dn->lock.is_stable()) | |
4747 | dn->auth_pin(&dn->lock); | |
4748 | dn->lock.set_state(LOCK_XLOCK); | |
4749 | dn->lock.get_xlock(mdr, mdr->get_client()); | |
4750 | mdr->locks.emplace(&dn->lock, MutationImpl::LockOp::XLOCK); | |
4751 | } | |
4752 | } | |
7c673cae | 4753 | |
11fdf7f2 TL |
4754 | dn->add_replica(from, d.nonce); |
4755 | dout(10) << " have " << *dn << dendl; | |
4756 | ||
4757 | if (dnl->is_primary()) { | |
4758 | if (d.is_primary()) { | |
4759 | if (vinodeno_t(d.ino, ss.snapid) != dnl->get_inode()->vino()) { | |
4760 | // the survivor missed MDentryUnlink+MDentryLink messages ? | |
4761 | ceph_assert(strong->strong_inodes.count(dnl->get_inode()->vino()) == 0); | |
4762 | CInode *in = get_inode(d.ino, ss.snapid); | |
4763 | ceph_assert(in); | |
4764 | ceph_assert(in->get_parent_dn()); | |
4765 | rejoin_unlinked_inodes[from].insert(in); | |
4766 | dout(7) << " sender has primary dentry but wrong inode" << dendl; | |
4767 | } | |
4768 | } else { | |
4769 | // the survivor missed MDentryLink message ? | |
4770 | ceph_assert(strong->strong_inodes.count(dnl->get_inode()->vino()) == 0); | |
4771 | dout(7) << " sender doesn't have primay dentry" << dendl; | |
4772 | } | |
4773 | } else { | |
4774 | if (d.is_primary()) { | |
4775 | // the survivor missed MDentryUnlink message ? | |
4776 | CInode *in = get_inode(d.ino, ss.snapid); | |
4777 | ceph_assert(in); | |
4778 | ceph_assert(in->get_parent_dn()); | |
7c673cae | 4779 | rejoin_unlinked_inodes[from].insert(in); |
11fdf7f2 | 4780 | dout(7) << " sender has primary dentry but we don't" << dendl; |
7c673cae | 4781 | } |
11fdf7f2 | 4782 | } |
7c673cae FG |
4783 | } |
4784 | } | |
4785 | } | |
4786 | ||
11fdf7f2 TL |
4787 | for (const auto &p : strong->strong_inodes) { |
4788 | CInode *in = get_inode(p.first); | |
4789 | ceph_assert(in); | |
4790 | in->add_replica(from, p.second.nonce); | |
7c673cae FG |
4791 | dout(10) << " have " << *in << dendl; |
4792 | ||
11fdf7f2 | 4793 | const MMDSCacheRejoin::inode_strong& is = p.second; |
7c673cae FG |
4794 | |
4795 | // caps_wanted | |
4796 | if (is.caps_wanted) { | |
11fdf7f2 | 4797 | in->set_mds_caps_wanted(from, is.caps_wanted); |
7c673cae FG |
4798 | dout(15) << " inode caps_wanted " << ccap_string(is.caps_wanted) |
4799 | << " on " << *in << dendl; | |
4800 | } | |
4801 | ||
4802 | // scatterlocks? | |
4803 | // infer state from replica state: | |
4804 | // * go to MIX if they might have wrlocks | |
4805 | // * go to LOCK if they are LOCK (just bc identify_files_to_recover might start twiddling filelock) | |
4806 | in->filelock.infer_state_from_strong_rejoin(is.filelock, !in->is_dir()); // maybe also go to LOCK | |
4807 | in->nestlock.infer_state_from_strong_rejoin(is.nestlock, false); | |
4808 | in->dirfragtreelock.infer_state_from_strong_rejoin(is.dftlock, false); | |
4809 | ||
4810 | // auth pin? | |
11fdf7f2 TL |
4811 | const auto authpinned_inodes_it = strong->authpinned_inodes.find(in->vino()); |
4812 | if (authpinned_inodes_it != strong->authpinned_inodes.end()) { | |
4813 | for (const auto& r : authpinned_inodes_it->second) { | |
4814 | dout(10) << " inode authpin by " << r << " on " << *in << dendl; | |
7c673cae FG |
4815 | |
4816 | // get/create slave mdrequest | |
4817 | MDRequestRef mdr; | |
11fdf7f2 TL |
4818 | if (have_request(r.reqid)) |
4819 | mdr = request_get(r.reqid); | |
7c673cae | 4820 | else |
11fdf7f2 | 4821 | mdr = request_start_slave(r.reqid, r.attempt, strong); |
7c673cae | 4822 | if (strong->frozen_authpin_inodes.count(in->vino())) { |
11fdf7f2 | 4823 | ceph_assert(!in->get_num_auth_pins()); |
7c673cae FG |
4824 | mdr->freeze_auth_pin(in); |
4825 | } else { | |
11fdf7f2 | 4826 | ceph_assert(!in->is_frozen_auth_pin()); |
7c673cae FG |
4827 | } |
4828 | mdr->auth_pin(in); | |
4829 | } | |
4830 | } | |
4831 | // xlock(s)? | |
11fdf7f2 TL |
4832 | const auto xlocked_inodes_it = strong->xlocked_inodes.find(in->vino()); |
4833 | if (xlocked_inodes_it != strong->xlocked_inodes.end()) { | |
4834 | for (const auto &q : xlocked_inodes_it->second) { | |
4835 | SimpleLock *lock = in->get_lock(q.first); | |
4836 | dout(10) << " inode xlock by " << q.second << " on " << *lock << " on " << *in << dendl; | |
4837 | MDRequestRef mdr = request_get(q.second.reqid); // should have this from auth_pin above. | |
4838 | ceph_assert(mdr->is_auth_pinned(in)); | |
4839 | if (!mdr->is_xlocked(&in->versionlock)) { | |
4840 | ceph_assert(in->versionlock.can_xlock_local()); | |
7c673cae | 4841 | in->versionlock.get_xlock(mdr, mdr->get_client()); |
11fdf7f2 | 4842 | mdr->locks.emplace(&in->versionlock, MutationImpl::LockOp::XLOCK); |
7c673cae FG |
4843 | } |
4844 | if (lock->is_stable()) | |
4845 | in->auth_pin(lock); | |
4846 | lock->set_state(LOCK_XLOCK); | |
4847 | if (lock == &in->filelock) | |
4848 | in->loner_cap = -1; | |
4849 | lock->get_xlock(mdr, mdr->get_client()); | |
11fdf7f2 | 4850 | mdr->locks.emplace(lock, MutationImpl::LockOp::XLOCK); |
7c673cae FG |
4851 | } |
4852 | } | |
4853 | } | |
4854 | // wrlock(s)? | |
11fdf7f2 TL |
4855 | for (const auto &p : strong->wrlocked_inodes) { |
4856 | CInode *in = get_inode(p.first); | |
4857 | for (const auto &q : p.second) { | |
4858 | SimpleLock *lock = in->get_lock(q.first); | |
4859 | for (const auto &r : q.second) { | |
4860 | dout(10) << " inode wrlock by " << r << " on " << *lock << " on " << *in << dendl; | |
4861 | MDRequestRef mdr = request_get(r.reqid); // should have this from auth_pin above. | |
7c673cae | 4862 | if (in->is_auth()) |
11fdf7f2 | 4863 | ceph_assert(mdr->is_auth_pinned(in)); |
7c673cae FG |
4864 | lock->set_state(LOCK_MIX); |
4865 | if (lock == &in->filelock) | |
4866 | in->loner_cap = -1; | |
4867 | lock->get_wrlock(true); | |
11fdf7f2 | 4868 | mdr->locks.emplace(lock, MutationImpl::LockOp::WRLOCK); |
7c673cae FG |
4869 | } |
4870 | } | |
4871 | } | |
4872 | ||
4873 | // done? | |
11fdf7f2 | 4874 | ceph_assert(rejoin_gather.count(from)); |
7c673cae | 4875 | rejoin_gather.erase(from); |
28e407b8 | 4876 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) { |
7c673cae FG |
4877 | rejoin_gather_finish(); |
4878 | } else { | |
4879 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" << dendl; | |
4880 | } | |
4881 | } | |
4882 | ||
11fdf7f2 | 4883 | void MDCache::handle_cache_rejoin_ack(const MMDSCacheRejoin::const_ref &ack) |
7c673cae FG |
4884 | { |
4885 | dout(7) << "handle_cache_rejoin_ack from " << ack->get_source() << dendl; | |
4886 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
4887 | ||
11fdf7f2 | 4888 | ceph_assert(mds->get_state() >= MDSMap::STATE_REJOIN); |
b32b8144 FG |
4889 | bool survivor = !mds->is_rejoin(); |
4890 | ||
7c673cae FG |
4891 | // for sending cache expire message |
4892 | set<CInode*> isolated_inodes; | |
4893 | set<CInode*> refragged_inodes; | |
11fdf7f2 | 4894 | list<pair<CInode*,int> > updated_realms; |
7c673cae FG |
4895 | |
4896 | // dirs | |
11fdf7f2 | 4897 | for (const auto &p : ack->strong_dirfrags) { |
7c673cae FG |
4898 | // we may have had incorrect dir fragmentation; refragment based |
4899 | // on what they auth tells us. | |
11fdf7f2 | 4900 | CDir *dir = get_dirfrag(p.first); |
7c673cae | 4901 | if (!dir) { |
11fdf7f2 | 4902 | dir = get_force_dirfrag(p.first, false); |
7c673cae FG |
4903 | if (dir) |
4904 | refragged_inodes.insert(dir->get_inode()); | |
4905 | } | |
4906 | if (!dir) { | |
11fdf7f2 | 4907 | CInode *diri = get_inode(p.first.ino); |
7c673cae FG |
4908 | if (!diri) { |
4909 | // barebones inode; the full inode loop below will clean up. | |
4910 | diri = new CInode(this, false); | |
11fdf7f2 | 4911 | diri->inode.ino = p.first.ino; |
7c673cae | 4912 | diri->inode.mode = S_IFDIR; |
11fdf7f2 | 4913 | diri->inode.dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; |
7c673cae | 4914 | add_inode(diri); |
11fdf7f2 | 4915 | if (MDS_INO_MDSDIR(from) == p.first.ino) { |
7c673cae FG |
4916 | diri->inode_auth = mds_authority_t(from, CDIR_AUTH_UNKNOWN); |
4917 | dout(10) << " add inode " << *diri << dendl; | |
4918 | } else { | |
4919 | diri->inode_auth = CDIR_AUTH_DEFAULT; | |
4920 | isolated_inodes.insert(diri); | |
11fdf7f2 | 4921 | dout(10) << " unconnected dirfrag " << p.first << dendl; |
7c673cae FG |
4922 | } |
4923 | } | |
4924 | // barebones dirfrag; the full dirfrag loop below will clean up. | |
11fdf7f2 TL |
4925 | dir = diri->add_dirfrag(new CDir(diri, p.first.frag, this, false)); |
4926 | if (MDS_INO_MDSDIR(from) == p.first.ino || | |
7c673cae FG |
4927 | (dir->authority() != CDIR_AUTH_UNDEF && |
4928 | dir->authority().first != from)) | |
4929 | adjust_subtree_auth(dir, from); | |
4930 | dout(10) << " add dirfrag " << *dir << dendl; | |
4931 | } | |
4932 | ||
11fdf7f2 | 4933 | dir->set_replica_nonce(p.second.nonce); |
7c673cae FG |
4934 | dir->state_clear(CDir::STATE_REJOINING); |
4935 | dout(10) << " got " << *dir << dendl; | |
4936 | ||
4937 | // dentries | |
11fdf7f2 TL |
4938 | auto it = ack->strong_dentries.find(p.first); |
4939 | if (it != ack->strong_dentries.end()) { | |
4940 | for (const auto &q : it->second) { | |
4941 | CDentry *dn = dir->lookup(q.first.name, q.first.snapid); | |
4942 | if(!dn) | |
4943 | dn = dir->add_null_dentry(q.first.name, q.second.first, q.first.snapid); | |
4944 | ||
4945 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4946 | ||
4947 | ceph_assert(dn->last == q.first.snapid); | |
4948 | if (dn->first != q.second.first) { | |
4949 | dout(10) << " adjust dn.first " << dn->first << " -> " << q.second.first << " on " << *dn << dendl; | |
4950 | dn->first = q.second.first; | |
4951 | } | |
7c673cae | 4952 | |
11fdf7f2 TL |
4953 | // may have bad linkage if we missed dentry link/unlink messages |
4954 | if (dnl->is_primary()) { | |
4955 | CInode *in = dnl->get_inode(); | |
4956 | if (!q.second.is_primary() || | |
4957 | vinodeno_t(q.second.ino, q.first.snapid) != in->vino()) { | |
4958 | dout(10) << " had bad linkage for " << *dn << ", unlinking " << *in << dendl; | |
4959 | dir->unlink_inode(dn); | |
4960 | } | |
4961 | } else if (dnl->is_remote()) { | |
4962 | if (!q.second.is_remote() || | |
4963 | q.second.remote_ino != dnl->get_remote_ino() || | |
4964 | q.second.remote_d_type != dnl->get_remote_d_type()) { | |
4965 | dout(10) << " had bad linkage for " << *dn << dendl; | |
4966 | dir->unlink_inode(dn); | |
4967 | } | |
4968 | } else { | |
4969 | if (!q.second.is_null()) | |
4970 | dout(10) << " had bad linkage for " << *dn << dendl; | |
4971 | } | |
7c673cae | 4972 | |
11fdf7f2 TL |
4973 | // hmm, did we have the proper linkage here? |
4974 | if (dnl->is_null() && !q.second.is_null()) { | |
4975 | if (q.second.is_remote()) { | |
4976 | dn->dir->link_remote_inode(dn, q.second.remote_ino, q.second.remote_d_type); | |
4977 | } else { | |
4978 | CInode *in = get_inode(q.second.ino, q.first.snapid); | |
4979 | if (!in) { | |
4980 | // barebones inode; assume it's dir, the full inode loop below will clean up. | |
4981 | in = new CInode(this, false, q.second.first, q.first.snapid); | |
4982 | in->inode.ino = q.second.ino; | |
4983 | in->inode.mode = S_IFDIR; | |
4984 | in->inode.dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; | |
4985 | add_inode(in); | |
4986 | dout(10) << " add inode " << *in << dendl; | |
4987 | } else if (in->get_parent_dn()) { | |
4988 | dout(10) << " had bad linkage for " << *(in->get_parent_dn()) | |
4989 | << ", unlinking " << *in << dendl; | |
4990 | in->get_parent_dir()->unlink_inode(in->get_parent_dn()); | |
4991 | } | |
4992 | dn->dir->link_primary_inode(dn, in); | |
4993 | isolated_inodes.erase(in); | |
7c673cae | 4994 | } |
11fdf7f2 | 4995 | } |
7c673cae | 4996 | |
11fdf7f2 TL |
4997 | dn->set_replica_nonce(q.second.nonce); |
4998 | dn->lock.set_state_rejoin(q.second.lock, rejoin_waiters, survivor); | |
4999 | dn->state_clear(CDentry::STATE_REJOINING); | |
5000 | dout(10) << " got " << *dn << dendl; | |
5001 | } | |
7c673cae FG |
5002 | } |
5003 | } | |
5004 | ||
5005 | for (set<CInode*>::iterator p = refragged_inodes.begin(); | |
5006 | p != refragged_inodes.end(); | |
5007 | ++p) { | |
5008 | list<CDir*> ls; | |
5009 | (*p)->get_nested_dirfrags(ls); | |
5010 | for (list<CDir*>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
5011 | if ((*q)->is_auth() || ack->strong_dirfrags.count((*q)->dirfrag())) | |
5012 | continue; | |
11fdf7f2 | 5013 | ceph_assert((*q)->get_num_any() == 0); |
7c673cae FG |
5014 | (*p)->close_dirfrag((*q)->get_frag()); |
5015 | } | |
5016 | } | |
5017 | ||
5018 | // full dirfrags | |
11fdf7f2 TL |
5019 | for (const auto &p : ack->dirfrag_bases) { |
5020 | CDir *dir = get_dirfrag(p.first); | |
5021 | ceph_assert(dir); | |
5022 | auto q = p.second.cbegin(); | |
7c673cae FG |
5023 | dir->_decode_base(q); |
5024 | dout(10) << " got dir replica " << *dir << dendl; | |
5025 | } | |
5026 | ||
5027 | // full inodes | |
11fdf7f2 | 5028 | auto p = ack->inode_base.cbegin(); |
7c673cae FG |
5029 | while (!p.end()) { |
5030 | inodeno_t ino; | |
5031 | snapid_t last; | |
5032 | bufferlist basebl; | |
11fdf7f2 TL |
5033 | decode(ino, p); |
5034 | decode(last, p); | |
5035 | decode(basebl, p); | |
7c673cae | 5036 | CInode *in = get_inode(ino, last); |
11fdf7f2 TL |
5037 | ceph_assert(in); |
5038 | auto q = basebl.cbegin(); | |
5039 | snapid_t sseq = 0; | |
5040 | if (in->snaprealm) | |
5041 | sseq = in->snaprealm->srnode.seq; | |
7c673cae | 5042 | in->_decode_base(q); |
11fdf7f2 TL |
5043 | if (in->snaprealm && in->snaprealm->srnode.seq != sseq) { |
5044 | int snap_op = sseq > 0 ? CEPH_SNAP_OP_UPDATE : CEPH_SNAP_OP_SPLIT; | |
5045 | updated_realms.push_back(pair<CInode*,int>(in, snap_op)); | |
5046 | } | |
7c673cae FG |
5047 | dout(10) << " got inode base " << *in << dendl; |
5048 | } | |
5049 | ||
5050 | // inodes | |
11fdf7f2 | 5051 | p = ack->inode_locks.cbegin(); |
7c673cae FG |
5052 | //dout(10) << "inode_locks len " << ack->inode_locks.length() << " is " << ack->inode_locks << dendl; |
5053 | while (!p.end()) { | |
5054 | inodeno_t ino; | |
5055 | snapid_t last; | |
5056 | __u32 nonce; | |
5057 | bufferlist lockbl; | |
11fdf7f2 TL |
5058 | decode(ino, p); |
5059 | decode(last, p); | |
5060 | decode(nonce, p); | |
5061 | decode(lockbl, p); | |
7c673cae FG |
5062 | |
5063 | CInode *in = get_inode(ino, last); | |
11fdf7f2 | 5064 | ceph_assert(in); |
7c673cae | 5065 | in->set_replica_nonce(nonce); |
11fdf7f2 | 5066 | auto q = lockbl.cbegin(); |
b32b8144 | 5067 | in->_decode_locks_rejoin(q, rejoin_waiters, rejoin_eval_locks, survivor); |
7c673cae FG |
5068 | in->state_clear(CInode::STATE_REJOINING); |
5069 | dout(10) << " got inode locks " << *in << dendl; | |
5070 | } | |
5071 | ||
5072 | // FIXME: This can happen if entire subtree, together with the inode subtree root | |
5073 | // belongs to, were trimmed between sending cache rejoin and receiving rejoin ack. | |
11fdf7f2 | 5074 | ceph_assert(isolated_inodes.empty()); |
7c673cae FG |
5075 | |
5076 | map<inodeno_t,map<client_t,Capability::Import> > peer_imported; | |
11fdf7f2 TL |
5077 | auto bp = ack->imported_caps.cbegin(); |
5078 | decode(peer_imported, bp); | |
7c673cae FG |
5079 | |
5080 | for (map<inodeno_t,map<client_t,Capability::Import> >::iterator p = peer_imported.begin(); | |
5081 | p != peer_imported.end(); | |
5082 | ++p) { | |
28e407b8 | 5083 | auto& ex = cap_exports.at(p->first); |
11fdf7f2 | 5084 | ceph_assert(ex.first == from); |
7c673cae FG |
5085 | for (map<client_t,Capability::Import>::iterator q = p->second.begin(); |
5086 | q != p->second.end(); | |
5087 | ++q) { | |
28e407b8 | 5088 | auto r = ex.second.find(q->first); |
11fdf7f2 | 5089 | ceph_assert(r != ex.second.end()); |
7c673cae FG |
5090 | |
5091 | dout(10) << " exporting caps for client." << q->first << " ino " << p->first << dendl; | |
5092 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
28e407b8 AA |
5093 | if (!session) { |
5094 | dout(10) << " no session for client." << p->first << dendl; | |
5095 | ex.second.erase(r); | |
5096 | continue; | |
5097 | } | |
7c673cae FG |
5098 | |
5099 | // mark client caps stale. | |
11fdf7f2 | 5100 | auto m = MClientCaps::create(CEPH_CAP_OP_EXPORT, p->first, 0, |
28e407b8 | 5101 | r->second.capinfo.cap_id, 0, |
7c673cae FG |
5102 | mds->get_osd_epoch_barrier()); |
5103 | m->set_cap_peer(q->second.cap_id, q->second.issue_seq, q->second.mseq, | |
5104 | (q->second.cap_id > 0 ? from : -1), 0); | |
5105 | mds->send_message_client_counted(m, session); | |
5106 | ||
28e407b8 | 5107 | ex.second.erase(r); |
7c673cae | 5108 | } |
11fdf7f2 TL |
5109 | ceph_assert(ex.second.empty()); |
5110 | } | |
5111 | ||
5112 | for (auto p : updated_realms) { | |
5113 | CInode *in = p.first; | |
5114 | bool notify_clients; | |
5115 | if (mds->is_rejoin()) { | |
5116 | if (!rejoin_pending_snaprealms.count(in)) { | |
5117 | in->get(CInode::PIN_OPENINGSNAPPARENTS); | |
5118 | rejoin_pending_snaprealms.insert(in); | |
5119 | } | |
5120 | notify_clients = false; | |
5121 | } else { | |
5122 | // notify clients if I'm survivor | |
5123 | notify_clients = true; | |
5124 | } | |
5125 | do_realm_invalidate_and_update_notify(in, p.second, notify_clients); | |
7c673cae FG |
5126 | } |
5127 | ||
5128 | // done? | |
11fdf7f2 | 5129 | ceph_assert(rejoin_ack_gather.count(from)); |
7c673cae | 5130 | rejoin_ack_gather.erase(from); |
b32b8144 | 5131 | if (!survivor) { |
7c673cae FG |
5132 | if (rejoin_gather.empty()) { |
5133 | // eval unstable scatter locks after all wrlocks are rejoined. | |
5134 | while (!rejoin_eval_locks.empty()) { | |
5135 | SimpleLock *lock = rejoin_eval_locks.front(); | |
5136 | rejoin_eval_locks.pop_front(); | |
5137 | if (!lock->is_stable()) | |
5138 | mds->locker->eval_gather(lock); | |
5139 | } | |
5140 | } | |
5141 | ||
5142 | if (rejoin_gather.empty() && // make sure we've gotten our FULL inodes, too. | |
5143 | rejoin_ack_gather.empty()) { | |
5144 | // finally, kickstart past snap parent opens | |
11fdf7f2 | 5145 | open_snaprealms(); |
7c673cae FG |
5146 | } else { |
5147 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" | |
5148 | << ", rejoin_ack from (" << rejoin_ack_gather << ")" << dendl; | |
5149 | } | |
5150 | } else { | |
5151 | // survivor. | |
5152 | mds->queue_waiters(rejoin_waiters); | |
5153 | } | |
5154 | } | |
5155 | ||
5156 | /** | |
5157 | * rejoin_trim_undef_inodes() -- remove REJOINUNDEF flagged inodes | |
5158 | * | |
5159 | * FIXME: wait, can this actually happen? a survivor should generate cache trim | |
5160 | * messages that clean these guys up... | |
5161 | */ | |
5162 | void MDCache::rejoin_trim_undef_inodes() | |
5163 | { | |
5164 | dout(10) << "rejoin_trim_undef_inodes" << dendl; | |
5165 | ||
5166 | while (!rejoin_undef_inodes.empty()) { | |
5167 | set<CInode*>::iterator p = rejoin_undef_inodes.begin(); | |
5168 | CInode *in = *p; | |
5169 | rejoin_undef_inodes.erase(p); | |
5170 | ||
5171 | in->clear_replica_map(); | |
5172 | ||
5173 | // close out dirfrags | |
5174 | if (in->is_dir()) { | |
5175 | list<CDir*> dfls; | |
5176 | in->get_dirfrags(dfls); | |
5177 | for (list<CDir*>::iterator p = dfls.begin(); | |
5178 | p != dfls.end(); | |
5179 | ++p) { | |
5180 | CDir *dir = *p; | |
5181 | dir->clear_replica_map(); | |
5182 | ||
94b18763 FG |
5183 | for (auto &p : dir->items) { |
5184 | CDentry *dn = p.second; | |
7c673cae FG |
5185 | dn->clear_replica_map(); |
5186 | ||
5187 | dout(10) << " trimming " << *dn << dendl; | |
5188 | dir->remove_dentry(dn); | |
5189 | } | |
5190 | ||
5191 | dout(10) << " trimming " << *dir << dendl; | |
5192 | in->close_dirfrag(dir->dirfrag().frag); | |
5193 | } | |
5194 | } | |
5195 | ||
5196 | CDentry *dn = in->get_parent_dn(); | |
5197 | if (dn) { | |
5198 | dn->clear_replica_map(); | |
5199 | dout(10) << " trimming " << *dn << dendl; | |
5200 | dn->dir->remove_dentry(dn); | |
5201 | } else { | |
5202 | dout(10) << " trimming " << *in << dendl; | |
5203 | remove_inode(in); | |
5204 | } | |
5205 | } | |
5206 | ||
11fdf7f2 | 5207 | ceph_assert(rejoin_undef_inodes.empty()); |
7c673cae FG |
5208 | } |
5209 | ||
5210 | void MDCache::rejoin_gather_finish() | |
5211 | { | |
5212 | dout(10) << "rejoin_gather_finish" << dendl; | |
11fdf7f2 TL |
5213 | ceph_assert(mds->is_rejoin()); |
5214 | ceph_assert(rejoin_ack_gather.count(mds->get_nodeid())); | |
7c673cae FG |
5215 | |
5216 | if (open_undef_inodes_dirfrags()) | |
5217 | return; | |
5218 | ||
5219 | if (process_imported_caps()) | |
5220 | return; | |
5221 | ||
5222 | choose_lock_states_and_reconnect_caps(); | |
5223 | ||
5224 | identify_files_to_recover(); | |
5225 | rejoin_send_acks(); | |
5226 | ||
5227 | // signal completion of fetches, rejoin_gather_finish, etc. | |
7c673cae FG |
5228 | rejoin_ack_gather.erase(mds->get_nodeid()); |
5229 | ||
5230 | // did we already get our acks too? | |
5231 | if (rejoin_ack_gather.empty()) { | |
11fdf7f2 TL |
5232 | // finally, open snaprealms |
5233 | open_snaprealms(); | |
7c673cae FG |
5234 | } |
5235 | } | |
5236 | ||
5237 | class C_MDC_RejoinOpenInoFinish: public MDCacheContext { | |
5238 | inodeno_t ino; | |
5239 | public: | |
5240 | C_MDC_RejoinOpenInoFinish(MDCache *c, inodeno_t i) : MDCacheContext(c), ino(i) {} | |
5241 | void finish(int r) override { | |
5242 | mdcache->rejoin_open_ino_finish(ino, r); | |
5243 | } | |
5244 | }; | |
5245 | ||
5246 | void MDCache::rejoin_open_ino_finish(inodeno_t ino, int ret) | |
5247 | { | |
5248 | dout(10) << "open_caps_inode_finish ino " << ino << " ret " << ret << dendl; | |
5249 | ||
5250 | if (ret < 0) { | |
5251 | cap_imports_missing.insert(ino); | |
5252 | } else if (ret == mds->get_nodeid()) { | |
11fdf7f2 | 5253 | ceph_assert(get_inode(ino)); |
7c673cae FG |
5254 | } else { |
5255 | auto p = cap_imports.find(ino); | |
11fdf7f2 | 5256 | ceph_assert(p != cap_imports.end()); |
7c673cae | 5257 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { |
11fdf7f2 TL |
5258 | ceph_assert(q->second.count(MDS_RANK_NONE)); |
5259 | ceph_assert(q->second.size() == 1); | |
7c673cae FG |
5260 | rejoin_export_caps(p->first, q->first, q->second[MDS_RANK_NONE], ret); |
5261 | } | |
5262 | cap_imports.erase(p); | |
5263 | } | |
5264 | ||
11fdf7f2 | 5265 | ceph_assert(cap_imports_num_opening > 0); |
7c673cae FG |
5266 | cap_imports_num_opening--; |
5267 | ||
5268 | if (cap_imports_num_opening == 0) { | |
5269 | if (rejoin_gather.empty()) | |
5270 | rejoin_gather_finish(); | |
5271 | else if (rejoin_gather.count(mds->get_nodeid())) | |
5272 | process_imported_caps(); | |
5273 | } | |
5274 | } | |
5275 | ||
5276 | class C_MDC_RejoinSessionsOpened : public MDCacheLogContext { | |
5277 | public: | |
28e407b8 AA |
5278 | map<client_t,pair<Session*,uint64_t> > session_map; |
5279 | C_MDC_RejoinSessionsOpened(MDCache *c) : MDCacheLogContext(c) {} | |
7c673cae | 5280 | void finish(int r) override { |
11fdf7f2 | 5281 | ceph_assert(r == 0); |
28e407b8 | 5282 | mdcache->rejoin_open_sessions_finish(session_map); |
7c673cae FG |
5283 | } |
5284 | }; | |
5285 | ||
28e407b8 | 5286 | void MDCache::rejoin_open_sessions_finish(map<client_t,pair<Session*,uint64_t> >& session_map) |
7c673cae FG |
5287 | { |
5288 | dout(10) << "rejoin_open_sessions_finish" << dendl; | |
28e407b8 AA |
5289 | mds->server->finish_force_open_sessions(session_map); |
5290 | rejoin_session_map.swap(session_map); | |
7c673cae FG |
5291 | if (rejoin_gather.empty()) |
5292 | rejoin_gather_finish(); | |
5293 | } | |
5294 | ||
11fdf7f2 TL |
5295 | void MDCache::rejoin_prefetch_ino_finish(inodeno_t ino, int ret) |
5296 | { | |
5297 | auto p = cap_imports.find(ino); | |
5298 | if (p != cap_imports.end()) { | |
5299 | dout(10) << __func__ << " ino " << ino << " ret " << ret << dendl; | |
5300 | if (ret < 0) { | |
5301 | cap_imports_missing.insert(ino); | |
5302 | } else if (ret != mds->get_nodeid()) { | |
5303 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
5304 | ceph_assert(q->second.count(MDS_RANK_NONE)); | |
5305 | ceph_assert(q->second.size() == 1); | |
5306 | rejoin_export_caps(p->first, q->first, q->second[MDS_RANK_NONE], ret); | |
5307 | } | |
5308 | cap_imports.erase(p); | |
5309 | } | |
5310 | } | |
5311 | } | |
5312 | ||
7c673cae FG |
5313 | bool MDCache::process_imported_caps() |
5314 | { | |
5315 | dout(10) << "process_imported_caps" << dendl; | |
5316 | ||
11fdf7f2 TL |
5317 | if (!open_file_table.is_prefetched() && |
5318 | open_file_table.prefetch_inodes()) { | |
5319 | open_file_table.wait_for_prefetch( | |
5320 | new MDSInternalContextWrapper(mds, | |
5321 | new FunctionContext([this](int r) { | |
5322 | ceph_assert(rejoin_gather.count(mds->get_nodeid())); | |
5323 | process_imported_caps(); | |
5324 | }) | |
5325 | ) | |
5326 | ); | |
5327 | return true; | |
5328 | } | |
5329 | ||
7c673cae FG |
5330 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ++p) { |
5331 | CInode *in = get_inode(p->first); | |
5332 | if (in) { | |
11fdf7f2 | 5333 | ceph_assert(in->is_auth()); |
7c673cae FG |
5334 | cap_imports_missing.erase(p->first); |
5335 | continue; | |
5336 | } | |
5337 | if (cap_imports_missing.count(p->first) > 0) | |
5338 | continue; | |
5339 | ||
5340 | cap_imports_num_opening++; | |
5341 | dout(10) << " opening missing ino " << p->first << dendl; | |
5342 | open_ino(p->first, (int64_t)-1, new C_MDC_RejoinOpenInoFinish(this, p->first), false); | |
28e407b8 AA |
5343 | if (!(cap_imports_num_opening % 1000)) |
5344 | mds->heartbeat_reset(); | |
7c673cae FG |
5345 | } |
5346 | ||
5347 | if (cap_imports_num_opening > 0) | |
5348 | return true; | |
5349 | ||
5350 | // called by rejoin_gather_finish() ? | |
5351 | if (rejoin_gather.count(mds->get_nodeid()) == 0) { | |
28e407b8 AA |
5352 | if (!rejoin_client_map.empty() && |
5353 | rejoin_session_map.empty()) { | |
5354 | C_MDC_RejoinSessionsOpened *finish = new C_MDC_RejoinSessionsOpened(this); | |
5355 | version_t pv = mds->server->prepare_force_open_sessions(rejoin_client_map, | |
11fdf7f2 | 5356 | rejoin_client_metadata_map, |
28e407b8 | 5357 | finish->session_map); |
11fdf7f2 TL |
5358 | ESessions *le = new ESessions(pv, std::move(rejoin_client_map), |
5359 | std::move(rejoin_client_metadata_map)); | |
5360 | mds->mdlog->start_submit_entry(le, finish); | |
28e407b8 AA |
5361 | mds->mdlog->flush(); |
5362 | rejoin_client_map.clear(); | |
11fdf7f2 | 5363 | rejoin_client_metadata_map.clear(); |
28e407b8 | 5364 | return true; |
7c673cae | 5365 | } |
7c673cae FG |
5366 | |
5367 | // process caps that were exported by slave rename | |
5368 | for (map<inodeno_t,pair<mds_rank_t,map<client_t,Capability::Export> > >::iterator p = rejoin_slave_exports.begin(); | |
5369 | p != rejoin_slave_exports.end(); | |
5370 | ++p) { | |
5371 | CInode *in = get_inode(p->first); | |
11fdf7f2 | 5372 | ceph_assert(in); |
7c673cae FG |
5373 | for (map<client_t,Capability::Export>::iterator q = p->second.second.begin(); |
5374 | q != p->second.second.end(); | |
5375 | ++q) { | |
28e407b8 AA |
5376 | auto r = rejoin_session_map.find(q->first); |
5377 | if (r == rejoin_session_map.end()) | |
5378 | continue; | |
7c673cae | 5379 | |
28e407b8 | 5380 | Session *session = r->second.first; |
7c673cae | 5381 | Capability *cap = in->get_client_cap(q->first); |
11fdf7f2 | 5382 | if (!cap) { |
7c673cae | 5383 | cap = in->add_client_cap(q->first, session); |
11fdf7f2 TL |
5384 | // add empty item to reconnected_caps |
5385 | (void)reconnected_caps[p->first][q->first]; | |
5386 | } | |
7c673cae FG |
5387 | cap->merge(q->second, true); |
5388 | ||
5389 | Capability::Import& im = rejoin_imported_caps[p->second.first][p->first][q->first]; | |
11fdf7f2 TL |
5390 | ceph_assert(cap->get_last_seq() == im.issue_seq); |
5391 | ceph_assert(cap->get_mseq() == im.mseq); | |
7c673cae FG |
5392 | cap->set_cap_id(im.cap_id); |
5393 | // send cap import because we assigned a new cap ID | |
5394 | do_cap_import(session, in, cap, q->second.cap_id, q->second.seq, q->second.mseq - 1, | |
5395 | p->second.first, CEPH_CAP_FLAG_AUTH); | |
5396 | } | |
5397 | } | |
5398 | rejoin_slave_exports.clear(); | |
5399 | rejoin_imported_caps.clear(); | |
5400 | ||
5401 | // process cap imports | |
5402 | // ino -> client -> frommds -> capex | |
5403 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ) { | |
5404 | CInode *in = get_inode(p->first); | |
5405 | if (!in) { | |
5406 | dout(10) << " still missing ino " << p->first | |
5407 | << ", will try again after replayed client requests" << dendl; | |
5408 | ++p; | |
5409 | continue; | |
5410 | } | |
11fdf7f2 | 5411 | ceph_assert(in->is_auth()); |
7c673cae | 5412 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { |
28e407b8 AA |
5413 | Session *session; |
5414 | { | |
5415 | auto r = rejoin_session_map.find(q->first); | |
5416 | session = (r != rejoin_session_map.end() ? r->second.first : nullptr); | |
5417 | } | |
5418 | ||
7c673cae | 5419 | for (auto r = q->second.begin(); r != q->second.end(); ++r) { |
28e407b8 AA |
5420 | if (!session) { |
5421 | if (r->first >= 0) | |
5422 | (void)rejoin_imported_caps[r->first][p->first][q->first]; // all are zero | |
5423 | continue; | |
5424 | } | |
5425 | ||
7c673cae FG |
5426 | Capability *cap = in->reconnect_cap(q->first, r->second, session); |
5427 | add_reconnected_cap(q->first, in->ino(), r->second); | |
5428 | if (r->first >= 0) { | |
5429 | if (cap->get_last_seq() == 0) // don't increase mseq if cap already exists | |
5430 | cap->inc_mseq(); | |
5431 | do_cap_import(session, in, cap, r->second.capinfo.cap_id, 0, 0, r->first, 0); | |
5432 | ||
5433 | Capability::Import& im = rejoin_imported_caps[r->first][p->first][q->first]; | |
5434 | im.cap_id = cap->get_cap_id(); | |
5435 | im.issue_seq = cap->get_last_seq(); | |
5436 | im.mseq = cap->get_mseq(); | |
5437 | } | |
5438 | } | |
5439 | } | |
5440 | cap_imports.erase(p++); // remove and move on | |
5441 | } | |
5442 | } else { | |
5443 | trim_non_auth(); | |
5444 | ||
11fdf7f2 | 5445 | ceph_assert(rejoin_gather.count(mds->get_nodeid())); |
7c673cae | 5446 | rejoin_gather.erase(mds->get_nodeid()); |
11fdf7f2 | 5447 | ceph_assert(!rejoin_ack_gather.count(mds->get_nodeid())); |
7c673cae | 5448 | maybe_send_pending_rejoins(); |
7c673cae FG |
5449 | } |
5450 | return false; | |
5451 | } | |
5452 | ||
7c673cae FG |
5453 | void MDCache::rebuild_need_snapflush(CInode *head_in, SnapRealm *realm, |
5454 | client_t client, snapid_t snap_follows) | |
5455 | { | |
5456 | dout(10) << "rebuild_need_snapflush " << snap_follows << " on " << *head_in << dendl; | |
5457 | ||
11fdf7f2 TL |
5458 | if (!realm->has_snaps_in_range(snap_follows + 1, head_in->first - 1)) |
5459 | return; | |
5460 | ||
7c673cae FG |
5461 | const set<snapid_t>& snaps = realm->get_snaps(); |
5462 | snapid_t follows = snap_follows; | |
5463 | ||
5464 | while (true) { | |
5465 | CInode *in = pick_inode_snap(head_in, follows); | |
5466 | if (in == head_in) | |
5467 | break; | |
11fdf7f2 TL |
5468 | |
5469 | bool need_snapflush = false; | |
5470 | for (auto p = snaps.lower_bound(std::max<snapid_t>(in->first, (follows + 1))); | |
5471 | p != snaps.end() && *p <= in->last; | |
5472 | ++p) { | |
5473 | head_in->add_need_snapflush(in, *p, client); | |
5474 | need_snapflush = true; | |
5475 | } | |
5476 | follows = in->last; | |
5477 | if (!need_snapflush) | |
5478 | continue; | |
5479 | ||
7c673cae FG |
5480 | dout(10) << " need snapflush from client." << client << " on " << *in << dendl; |
5481 | ||
11fdf7f2 | 5482 | /* TODO: we can check the reconnected/flushing caps to find |
7c673cae FG |
5483 | * which locks need gathering */ |
5484 | for (int i = 0; i < num_cinode_locks; i++) { | |
5485 | int lockid = cinode_lock_info[i].lock; | |
5486 | SimpleLock *lock = in->get_lock(lockid); | |
11fdf7f2 | 5487 | ceph_assert(lock); |
7c673cae FG |
5488 | in->client_snap_caps[lockid].insert(client); |
5489 | in->auth_pin(lock); | |
5490 | lock->set_state(LOCK_SNAP_SYNC); | |
5491 | lock->get_wrlock(true); | |
5492 | } | |
11fdf7f2 | 5493 | mds->locker->mark_need_snapflush_inode(in); |
7c673cae FG |
5494 | } |
5495 | } | |
5496 | ||
5497 | /* | |
5498 | * choose lock states based on reconnected caps | |
5499 | */ | |
5500 | void MDCache::choose_lock_states_and_reconnect_caps() | |
5501 | { | |
5502 | dout(10) << "choose_lock_states_and_reconnect_caps" << dendl; | |
5503 | ||
81eedcae | 5504 | int count = 0; |
11fdf7f2 TL |
5505 | for (auto p : inode_map) { |
5506 | CInode *in = p.second; | |
7c673cae FG |
5507 | if (in->last != CEPH_NOSNAP) |
5508 | continue; | |
5509 | ||
5510 | if (in->is_auth() && !in->is_base() && in->inode.is_dirty_rstat()) | |
5511 | in->mark_dirty_rstat(); | |
5512 | ||
7c673cae | 5513 | int dirty_caps = 0; |
11fdf7f2 TL |
5514 | auto q = reconnected_caps.find(in->ino()); |
5515 | if (q != reconnected_caps.end()) { | |
5516 | for (const auto &it : q->second) | |
7c673cae FG |
5517 | dirty_caps |= it.second.dirty_caps; |
5518 | } | |
5519 | in->choose_lock_states(dirty_caps); | |
5520 | dout(15) << " chose lock states on " << *in << dendl; | |
5521 | ||
11fdf7f2 TL |
5522 | if (in->snaprealm && !rejoin_pending_snaprealms.count(in)) { |
5523 | in->get(CInode::PIN_OPENINGSNAPPARENTS); | |
5524 | rejoin_pending_snaprealms.insert(in); | |
7c673cae | 5525 | } |
81eedcae TL |
5526 | |
5527 | if (!(++count % 1000)) | |
5528 | mds->heartbeat_reset(); | |
11fdf7f2 | 5529 | } |
7c673cae FG |
5530 | } |
5531 | ||
5532 | void MDCache::prepare_realm_split(SnapRealm *realm, client_t client, inodeno_t ino, | |
11fdf7f2 | 5533 | map<client_t,MClientSnap::ref>& splits) |
7c673cae | 5534 | { |
11fdf7f2 TL |
5535 | MClientSnap::ref snap; |
5536 | auto it = splits.find(client); | |
5537 | if (it != splits.end()) { | |
5538 | snap = it->second; | |
5539 | snap->head.op = CEPH_SNAP_OP_SPLIT; | |
5540 | } else { | |
5541 | snap = MClientSnap::create(CEPH_SNAP_OP_SPLIT); | |
5542 | splits.emplace(std::piecewise_construct, std::forward_as_tuple(client), std::forward_as_tuple(snap)); | |
7c673cae | 5543 | snap->head.split = realm->inode->ino(); |
11fdf7f2 | 5544 | snap->bl = realm->get_snap_trace(); |
7c673cae | 5545 | |
11fdf7f2 TL |
5546 | for (const auto& child : realm->open_children) |
5547 | snap->split_realms.push_back(child->inode->ino()); | |
5548 | } | |
7c673cae FG |
5549 | snap->split_inos.push_back(ino); |
5550 | } | |
5551 | ||
11fdf7f2 TL |
5552 | void MDCache::prepare_realm_merge(SnapRealm *realm, SnapRealm *parent_realm, |
5553 | map<client_t,MClientSnap::ref>& splits) | |
5554 | { | |
5555 | ceph_assert(parent_realm); | |
5556 | ||
5557 | vector<inodeno_t> split_inos; | |
5558 | vector<inodeno_t> split_realms; | |
5559 | ||
5560 | for (elist<CInode*>::iterator p = realm->inodes_with_caps.begin(member_offset(CInode, item_caps)); | |
5561 | !p.end(); | |
5562 | ++p) | |
5563 | split_inos.push_back((*p)->ino()); | |
5564 | for (set<SnapRealm*>::iterator p = realm->open_children.begin(); | |
5565 | p != realm->open_children.end(); | |
5566 | ++p) | |
5567 | split_realms.push_back((*p)->inode->ino()); | |
5568 | ||
5569 | for (const auto& p : realm->client_caps) { | |
5570 | ceph_assert(!p.second->empty()); | |
5571 | auto em = splits.emplace(std::piecewise_construct, std::forward_as_tuple(p.first), std::forward_as_tuple()); | |
5572 | if (em.second) { | |
5573 | auto update = MClientSnap::create(CEPH_SNAP_OP_SPLIT); | |
5574 | update->head.split = parent_realm->inode->ino(); | |
5575 | update->split_inos = split_inos; | |
5576 | update->split_realms = split_realms; | |
5577 | update->bl = parent_realm->get_snap_trace(); | |
5578 | em.first->second = std::move(update); | |
5579 | } | |
5580 | } | |
5581 | } | |
5582 | ||
5583 | void MDCache::send_snaps(map<client_t,MClientSnap::ref>& splits) | |
7c673cae FG |
5584 | { |
5585 | dout(10) << "send_snaps" << dendl; | |
5586 | ||
11fdf7f2 TL |
5587 | for (auto &p : splits) { |
5588 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(p.first.v)); | |
7c673cae | 5589 | if (session) { |
11fdf7f2 TL |
5590 | dout(10) << " client." << p.first |
5591 | << " split " << p.second->head.split | |
5592 | << " inos " << p.second->split_inos | |
7c673cae | 5593 | << dendl; |
11fdf7f2 | 5594 | mds->send_message_client_counted(p.second, session); |
7c673cae | 5595 | } else { |
11fdf7f2 | 5596 | dout(10) << " no session for client." << p.first << dendl; |
7c673cae FG |
5597 | } |
5598 | } | |
5599 | splits.clear(); | |
5600 | } | |
5601 | ||
5602 | ||
5603 | /* | |
5604 | * remove any items from logsegment open_file lists that don't have | |
5605 | * any caps | |
5606 | */ | |
5607 | void MDCache::clean_open_file_lists() | |
5608 | { | |
5609 | dout(10) << "clean_open_file_lists" << dendl; | |
5610 | ||
5611 | for (map<uint64_t,LogSegment*>::iterator p = mds->mdlog->segments.begin(); | |
5612 | p != mds->mdlog->segments.end(); | |
5613 | ++p) { | |
5614 | LogSegment *ls = p->second; | |
5615 | ||
5616 | elist<CInode*>::iterator q = ls->open_files.begin(member_offset(CInode, item_open_file)); | |
5617 | while (!q.end()) { | |
5618 | CInode *in = *q; | |
5619 | ++q; | |
5620 | if (in->last == CEPH_NOSNAP) { | |
11fdf7f2 TL |
5621 | dout(10) << " unlisting unwanted/capless inode " << *in << dendl; |
5622 | in->item_open_file.remove_myself(); | |
5623 | } else { | |
7c673cae FG |
5624 | if (in->client_snap_caps.empty()) { |
5625 | dout(10) << " unlisting flushed snap inode " << *in << dendl; | |
5626 | in->item_open_file.remove_myself(); | |
5627 | } | |
5628 | } | |
5629 | } | |
5630 | } | |
5631 | } | |
5632 | ||
11fdf7f2 TL |
5633 | void MDCache::dump_openfiles(Formatter *f) |
5634 | { | |
5635 | f->open_array_section("openfiles"); | |
5636 | for (auto p = mds->mdlog->segments.begin(); | |
5637 | p != mds->mdlog->segments.end(); | |
5638 | ++p) { | |
5639 | LogSegment *ls = p->second; | |
5640 | ||
5641 | auto q = ls->open_files.begin(member_offset(CInode, item_open_file)); | |
5642 | while (!q.end()) { | |
5643 | CInode *in = *q; | |
5644 | ++q; | |
5645 | if ((in->last == CEPH_NOSNAP && !in->is_any_caps_wanted()) | |
5646 | || (in->last != CEPH_NOSNAP && in->client_snap_caps.empty())) | |
5647 | continue; | |
5648 | f->open_object_section("file"); | |
5649 | in->dump(f, CInode::DUMP_PATH | CInode::DUMP_INODE_STORE_BASE | CInode::DUMP_CAPS); | |
5650 | f->close_section(); | |
5651 | } | |
5652 | } | |
5653 | f->close_section(); | |
5654 | } | |
7c673cae FG |
5655 | |
5656 | Capability* MDCache::rejoin_import_cap(CInode *in, client_t client, const cap_reconnect_t& icr, mds_rank_t frommds) | |
5657 | { | |
5658 | dout(10) << "rejoin_import_cap for client." << client << " from mds." << frommds | |
5659 | << " on " << *in << dendl; | |
5660 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(client.v)); | |
5661 | if (!session) { | |
5662 | dout(10) << " no session for client." << client << dendl; | |
5663 | return NULL; | |
5664 | } | |
5665 | ||
5666 | Capability *cap = in->reconnect_cap(client, icr, session); | |
5667 | ||
5668 | if (frommds >= 0) { | |
5669 | if (cap->get_last_seq() == 0) // don't increase mseq if cap already exists | |
5670 | cap->inc_mseq(); | |
5671 | do_cap_import(session, in, cap, icr.capinfo.cap_id, 0, 0, frommds, 0); | |
5672 | } | |
5673 | ||
5674 | return cap; | |
5675 | } | |
5676 | ||
5677 | void MDCache::export_remaining_imported_caps() | |
5678 | { | |
5679 | dout(10) << "export_remaining_imported_caps" << dendl; | |
5680 | ||
5681 | stringstream warn_str; | |
5682 | ||
81eedcae | 5683 | int count = 0; |
7c673cae FG |
5684 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ++p) { |
5685 | warn_str << " ino " << p->first << "\n"; | |
5686 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
5687 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
5688 | if (session) { | |
5689 | // mark client caps stale. | |
11fdf7f2 | 5690 | auto stale = MClientCaps::create(CEPH_CAP_OP_EXPORT, p->first, 0, 0, 0, mds->get_osd_epoch_barrier()); |
7c673cae FG |
5691 | stale->set_cap_peer(0, 0, 0, -1, 0); |
5692 | mds->send_message_client_counted(stale, q->first); | |
5693 | } | |
5694 | } | |
5695 | ||
81eedcae TL |
5696 | if (!(++count % 1000)) |
5697 | mds->heartbeat_reset(); | |
7c673cae FG |
5698 | } |
5699 | ||
11fdf7f2 | 5700 | for (map<inodeno_t, MDSContext::vec >::iterator p = cap_reconnect_waiters.begin(); |
7c673cae FG |
5701 | p != cap_reconnect_waiters.end(); |
5702 | ++p) | |
5703 | mds->queue_waiters(p->second); | |
5704 | ||
5705 | cap_imports.clear(); | |
5706 | cap_reconnect_waiters.clear(); | |
5707 | ||
5708 | if (warn_str.peek() != EOF) { | |
5709 | mds->clog->warn() << "failed to reconnect caps for missing inodes:"; | |
5710 | mds->clog->warn(warn_str); | |
5711 | } | |
5712 | } | |
5713 | ||
a8e16298 | 5714 | Capability* MDCache::try_reconnect_cap(CInode *in, Session *session) |
7c673cae FG |
5715 | { |
5716 | client_t client = session->info.get_client(); | |
a8e16298 | 5717 | Capability *cap = nullptr; |
7c673cae FG |
5718 | const cap_reconnect_t *rc = get_replay_cap_reconnect(in->ino(), client); |
5719 | if (rc) { | |
a8e16298 | 5720 | cap = in->reconnect_cap(client, *rc, session); |
7c673cae FG |
5721 | dout(10) << "try_reconnect_cap client." << client |
5722 | << " reconnect wanted " << ccap_string(rc->capinfo.wanted) | |
5723 | << " issue " << ccap_string(rc->capinfo.issued) | |
5724 | << " on " << *in << dendl; | |
5725 | remove_replay_cap_reconnect(in->ino(), client); | |
5726 | ||
5727 | if (in->is_replicated()) { | |
5728 | mds->locker->try_eval(in, CEPH_CAP_LOCKS); | |
5729 | } else { | |
5730 | int dirty_caps = 0; | |
5731 | auto p = reconnected_caps.find(in->ino()); | |
5732 | if (p != reconnected_caps.end()) { | |
5733 | auto q = p->second.find(client); | |
5734 | if (q != p->second.end()) | |
5735 | dirty_caps = q->second.dirty_caps; | |
5736 | } | |
5737 | in->choose_lock_states(dirty_caps); | |
5738 | dout(15) << " chose lock states on " << *in << dendl; | |
5739 | } | |
5740 | ||
11fdf7f2 | 5741 | map<inodeno_t, MDSContext::vec >::iterator it = |
7c673cae FG |
5742 | cap_reconnect_waiters.find(in->ino()); |
5743 | if (it != cap_reconnect_waiters.end()) { | |
5744 | mds->queue_waiters(it->second); | |
5745 | cap_reconnect_waiters.erase(it); | |
5746 | } | |
5747 | } | |
a8e16298 | 5748 | return cap; |
7c673cae FG |
5749 | } |
5750 | ||
5751 | ||
5752 | ||
5753 | // ------- | |
5754 | // cap imports and delayed snap parent opens | |
5755 | ||
5756 | void MDCache::do_cap_import(Session *session, CInode *in, Capability *cap, | |
5757 | uint64_t p_cap_id, ceph_seq_t p_seq, ceph_seq_t p_mseq, | |
5758 | int peer, int p_flags) | |
5759 | { | |
7c673cae FG |
5760 | SnapRealm *realm = in->find_snaprealm(); |
5761 | if (realm->have_past_parents_open()) { | |
5762 | dout(10) << "do_cap_import " << session->info.inst.name << " mseq " << cap->get_mseq() << " on " << *in << dendl; | |
5763 | if (cap->get_last_seq() == 0) // reconnected cap | |
5764 | cap->inc_last_seq(); | |
5765 | cap->set_last_issue(); | |
5766 | cap->set_last_issue_stamp(ceph_clock_now()); | |
5767 | cap->clear_new(); | |
11fdf7f2 | 5768 | auto reap = MClientCaps::create(CEPH_CAP_OP_IMPORT, in->ino(), realm->inode->ino(), cap->get_cap_id(), cap->get_last_seq(), cap->pending(), cap->wanted(), 0, cap->get_mseq(), mds->get_osd_epoch_barrier()); |
7c673cae | 5769 | in->encode_cap_message(reap, cap); |
11fdf7f2 | 5770 | reap->snapbl = realm->get_snap_trace(); |
7c673cae FG |
5771 | reap->set_cap_peer(p_cap_id, p_seq, p_mseq, peer, p_flags); |
5772 | mds->send_message_client_counted(reap, session); | |
5773 | } else { | |
11fdf7f2 | 5774 | ceph_abort(); |
7c673cae FG |
5775 | } |
5776 | } | |
5777 | ||
5778 | void MDCache::do_delayed_cap_imports() | |
5779 | { | |
5780 | dout(10) << "do_delayed_cap_imports" << dendl; | |
5781 | ||
11fdf7f2 | 5782 | ceph_assert(delayed_imported_caps.empty()); |
7c673cae FG |
5783 | } |
5784 | ||
11fdf7f2 TL |
5785 | struct C_MDC_OpenSnapRealms : public MDCacheContext { |
5786 | explicit C_MDC_OpenSnapRealms(MDCache *c) : MDCacheContext(c) {} | |
7c673cae | 5787 | void finish(int r) override { |
11fdf7f2 | 5788 | mdcache->open_snaprealms(); |
7c673cae FG |
5789 | } |
5790 | }; | |
5791 | ||
11fdf7f2 | 5792 | void MDCache::open_snaprealms() |
7c673cae | 5793 | { |
11fdf7f2 | 5794 | dout(10) << "open_snaprealms" << dendl; |
7c673cae | 5795 | |
7c673cae FG |
5796 | MDSGatherBuilder gather(g_ceph_context); |
5797 | ||
11fdf7f2 TL |
5798 | auto it = rejoin_pending_snaprealms.begin(); |
5799 | while (it != rejoin_pending_snaprealms.end()) { | |
5800 | CInode *in = *it; | |
5801 | SnapRealm *realm = in->snaprealm; | |
5802 | ceph_assert(realm); | |
5803 | if (realm->have_past_parents_open() || | |
5804 | realm->open_parents(gather.new_sub())) { | |
7c673cae FG |
5805 | dout(10) << " past parents now open on " << *in << dendl; |
5806 | ||
11fdf7f2 TL |
5807 | map<client_t,MClientSnap::ref> splits; |
5808 | // finish off client snaprealm reconnects? | |
5809 | map<inodeno_t,map<client_t,snapid_t> >::iterator q = reconnected_snaprealms.find(in->ino()); | |
5810 | if (q != reconnected_snaprealms.end()) { | |
5811 | for (const auto& r : q->second) | |
5812 | finish_snaprealm_reconnect(r.first, realm, r.second, splits); | |
5813 | reconnected_snaprealms.erase(q); | |
5814 | } | |
5815 | ||
5816 | for (elist<CInode*>::iterator p = realm->inodes_with_caps.begin(member_offset(CInode, item_caps)); | |
5817 | !p.end(); ++p) { | |
5818 | CInode *child = *p; | |
7c673cae | 5819 | auto q = reconnected_caps.find(child->ino()); |
11fdf7f2 | 5820 | ceph_assert(q != reconnected_caps.end()); |
7c673cae | 5821 | for (auto r = q->second.begin(); r != q->second.end(); ++r) { |
11fdf7f2 TL |
5822 | Capability *cap = child->get_client_cap(r->first); |
5823 | if (!cap) | |
5824 | continue; | |
5825 | if (r->second.snap_follows > 0) { | |
5826 | if (r->second.snap_follows < child->first - 1) { | |
5827 | rebuild_need_snapflush(child, realm, r->first, r->second.snap_follows); | |
5828 | } else if (r->second.snapflush) { | |
5829 | // When processing a cap flush message that is re-sent, it's possble | |
5830 | // that the sender has already released all WR caps. So we should | |
5831 | // force MDCache::cow_inode() to setup CInode::client_need_snapflush. | |
5832 | cap->mark_needsnapflush(); | |
5833 | } | |
7c673cae FG |
5834 | } |
5835 | // make sure client's cap is in the correct snaprealm. | |
5836 | if (r->second.realm_ino != in->ino()) { | |
11fdf7f2 | 5837 | prepare_realm_split(realm, r->first, child->ino(), splits); |
7c673cae FG |
5838 | } |
5839 | } | |
5840 | } | |
5841 | ||
11fdf7f2 | 5842 | rejoin_pending_snaprealms.erase(it++); |
7c673cae FG |
5843 | in->put(CInode::PIN_OPENINGSNAPPARENTS); |
5844 | ||
11fdf7f2 | 5845 | send_snaps(splits); |
7c673cae FG |
5846 | } else { |
5847 | dout(10) << " opening past parents on " << *in << dendl; | |
11fdf7f2 | 5848 | ++it; |
7c673cae FG |
5849 | } |
5850 | } | |
5851 | ||
7c673cae | 5852 | if (gather.has_subs()) { |
11fdf7f2 TL |
5853 | if (gather.num_subs_remaining() == 0) { |
5854 | // cleanup gather | |
5855 | gather.set_finisher(new C_MDSInternalNoop); | |
5856 | gather.activate(); | |
5857 | } else { | |
5858 | // for multimds, must succeed the first time | |
5859 | ceph_assert(recovery_set.empty()); | |
5860 | ||
5861 | dout(10) << "open_snaprealms - waiting for " | |
5862 | << gather.num_subs_remaining() << dendl; | |
5863 | gather.set_finisher(new C_MDC_OpenSnapRealms(this)); | |
5864 | gather.activate(); | |
5865 | return; | |
5866 | } | |
5867 | } | |
5868 | ||
5869 | notify_global_snaprealm_update(CEPH_SNAP_OP_UPDATE); | |
5870 | ||
5871 | if (!reconnected_snaprealms.empty()) { | |
5872 | dout(5) << "open_snaprealms has unconnected snaprealm:" << dendl; | |
5873 | for (auto& p : reconnected_snaprealms) { | |
7c673cae | 5874 | stringstream warn_str; |
11fdf7f2 TL |
5875 | warn_str << " " << p.first << " {"; |
5876 | bool first = true; | |
5877 | for (auto& q : p.second) { | |
5878 | if (!first) | |
5879 | warn_str << ", "; | |
5880 | warn_str << "client." << q.first << "/" << q.second; | |
7c673cae | 5881 | } |
11fdf7f2 TL |
5882 | warn_str << "}"; |
5883 | dout(5) << warn_str.str() << dendl; | |
7c673cae | 5884 | } |
7c673cae | 5885 | } |
11fdf7f2 TL |
5886 | ceph_assert(rejoin_waiters.empty()); |
5887 | ceph_assert(rejoin_pending_snaprealms.empty()); | |
5888 | dout(10) << "open_snaprealms - all open" << dendl; | |
5889 | do_delayed_cap_imports(); | |
5890 | ||
5891 | ceph_assert(rejoin_done); | |
5892 | rejoin_done.release()->complete(0); | |
5893 | reconnected_caps.clear(); | |
7c673cae FG |
5894 | } |
5895 | ||
5896 | bool MDCache::open_undef_inodes_dirfrags() | |
5897 | { | |
5898 | dout(10) << "open_undef_inodes_dirfrags " | |
5899 | << rejoin_undef_inodes.size() << " inodes " | |
5900 | << rejoin_undef_dirfrags.size() << " dirfrags" << dendl; | |
5901 | ||
5902 | set<CDir*> fetch_queue = rejoin_undef_dirfrags; | |
5903 | ||
5904 | for (set<CInode*>::iterator p = rejoin_undef_inodes.begin(); | |
5905 | p != rejoin_undef_inodes.end(); | |
5906 | ++p) { | |
5907 | CInode *in = *p; | |
11fdf7f2 | 5908 | ceph_assert(!in->is_base()); |
7c673cae FG |
5909 | fetch_queue.insert(in->get_parent_dir()); |
5910 | } | |
5911 | ||
5912 | if (fetch_queue.empty()) | |
5913 | return false; | |
5914 | ||
28e407b8 AA |
5915 | MDSGatherBuilder gather(g_ceph_context, |
5916 | new MDSInternalContextWrapper(mds, | |
5917 | new FunctionContext([this](int r) { | |
5918 | if (rejoin_gather.empty()) | |
5919 | rejoin_gather_finish(); | |
5920 | }) | |
5921 | ) | |
5922 | ); | |
5923 | ||
7c673cae FG |
5924 | for (set<CDir*>::iterator p = fetch_queue.begin(); |
5925 | p != fetch_queue.end(); | |
5926 | ++p) { | |
5927 | CDir *dir = *p; | |
5928 | CInode *diri = dir->get_inode(); | |
5929 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) | |
5930 | continue; | |
5931 | if (dir->state_test(CDir::STATE_REJOINUNDEF)) | |
11fdf7f2 | 5932 | ceph_assert(diri->dirfragtree.is_leaf(dir->get_frag())); |
7c673cae FG |
5933 | dir->fetch(gather.new_sub()); |
5934 | } | |
11fdf7f2 | 5935 | ceph_assert(gather.has_subs()); |
7c673cae FG |
5936 | gather.activate(); |
5937 | return true; | |
5938 | } | |
5939 | ||
5940 | void MDCache::opened_undef_inode(CInode *in) { | |
5941 | dout(10) << "opened_undef_inode " << *in << dendl; | |
5942 | rejoin_undef_inodes.erase(in); | |
5943 | if (in->is_dir()) { | |
5944 | // FIXME: re-hash dentries if necessary | |
11fdf7f2 | 5945 | ceph_assert(in->inode.dir_layout.dl_dir_hash == g_conf()->mds_default_dir_hash); |
7c673cae FG |
5946 | if (in->has_dirfrags() && !in->dirfragtree.is_leaf(frag_t())) { |
5947 | CDir *dir = in->get_dirfrag(frag_t()); | |
11fdf7f2 | 5948 | ceph_assert(dir); |
7c673cae FG |
5949 | rejoin_undef_dirfrags.erase(dir); |
5950 | in->force_dirfrags(); | |
5951 | list<CDir*> ls; | |
5952 | in->get_dirfrags(ls); | |
5953 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) | |
5954 | rejoin_undef_dirfrags.insert(*p); | |
5955 | } | |
5956 | } | |
5957 | } | |
5958 | ||
11fdf7f2 TL |
5959 | void MDCache::finish_snaprealm_reconnect(client_t client, SnapRealm *realm, snapid_t seq, |
5960 | map<client_t,MClientSnap::ref>& updates) | |
7c673cae FG |
5961 | { |
5962 | if (seq < realm->get_newest_seq()) { | |
5963 | dout(10) << "finish_snaprealm_reconnect client." << client << " has old seq " << seq << " < " | |
11fdf7f2 TL |
5964 | << realm->get_newest_seq() << " on " << *realm << dendl; |
5965 | auto snap = MClientSnap::create(CEPH_SNAP_OP_UPDATE); | |
5966 | snap->bl = realm->get_snap_trace(); | |
5967 | for (const auto& child : realm->open_children) | |
5968 | snap->split_realms.push_back(child->inode->ino()); | |
5969 | updates.emplace(std::piecewise_construct, std::forward_as_tuple(client), std::forward_as_tuple(snap)); | |
7c673cae FG |
5970 | } else { |
5971 | dout(10) << "finish_snaprealm_reconnect client." << client << " up to date" | |
5972 | << " on " << *realm << dendl; | |
5973 | } | |
5974 | } | |
5975 | ||
5976 | ||
5977 | ||
5978 | void MDCache::rejoin_send_acks() | |
5979 | { | |
5980 | dout(7) << "rejoin_send_acks" << dendl; | |
5981 | ||
5982 | // replicate stray | |
5983 | for (map<mds_rank_t, set<CInode*> >::iterator p = rejoin_unlinked_inodes.begin(); | |
5984 | p != rejoin_unlinked_inodes.end(); | |
5985 | ++p) { | |
5986 | for (set<CInode*>::iterator q = p->second.begin(); | |
5987 | q != p->second.end(); | |
5988 | ++q) { | |
5989 | CInode *in = *q; | |
5990 | dout(7) << " unlinked inode " << *in << dendl; | |
5991 | // inode expired | |
5992 | if (!in->is_replica(p->first)) | |
5993 | continue; | |
5994 | while (1) { | |
5995 | CDentry *dn = in->get_parent_dn(); | |
5996 | if (dn->is_replica(p->first)) | |
5997 | break; | |
5998 | dn->add_replica(p->first); | |
5999 | CDir *dir = dn->get_dir(); | |
6000 | if (dir->is_replica(p->first)) | |
6001 | break; | |
6002 | dir->add_replica(p->first); | |
6003 | in = dir->get_inode(); | |
6004 | if (in->is_replica(p->first)) | |
6005 | break; | |
224ce89b | 6006 | in->add_replica(p->first); |
7c673cae FG |
6007 | if (in->is_base()) |
6008 | break; | |
6009 | } | |
6010 | } | |
6011 | } | |
6012 | rejoin_unlinked_inodes.clear(); | |
6013 | ||
6014 | // send acks to everyone in the recovery set | |
11fdf7f2 | 6015 | map<mds_rank_t,MMDSCacheRejoin::ref> acks; |
7c673cae FG |
6016 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); |
6017 | p != recovery_set.end(); | |
31f18b77 FG |
6018 | ++p) { |
6019 | if (rejoin_ack_sent.count(*p)) | |
6020 | continue; | |
11fdf7f2 | 6021 | acks[*p] = MMDSCacheRejoin::create(MMDSCacheRejoin::OP_ACK); |
31f18b77 FG |
6022 | } |
6023 | ||
6024 | rejoin_ack_sent = recovery_set; | |
7c673cae FG |
6025 | |
6026 | // walk subtrees | |
6027 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
6028 | p != subtrees.end(); | |
6029 | ++p) { | |
6030 | CDir *dir = p->first; | |
6031 | if (!dir->is_auth()) | |
6032 | continue; | |
6033 | dout(10) << "subtree " << *dir << dendl; | |
6034 | ||
6035 | // auth items in this subtree | |
6036 | list<CDir*> dq; | |
6037 | dq.push_back(dir); | |
6038 | ||
6039 | while (!dq.empty()) { | |
6040 | CDir *dir = dq.front(); | |
6041 | dq.pop_front(); | |
6042 | ||
6043 | // dir | |
181888fb FG |
6044 | for (auto &r : dir->get_replicas()) { |
6045 | auto it = acks.find(r.first); | |
31f18b77 FG |
6046 | if (it == acks.end()) |
6047 | continue; | |
181888fb | 6048 | it->second->add_strong_dirfrag(dir->dirfrag(), ++r.second, dir->dir_rep); |
31f18b77 | 6049 | it->second->add_dirfrag_base(dir); |
7c673cae FG |
6050 | } |
6051 | ||
94b18763 FG |
6052 | for (auto &p : dir->items) { |
6053 | CDentry *dn = p.second; | |
7c673cae FG |
6054 | CDentry::linkage_t *dnl = dn->get_linkage(); |
6055 | ||
6056 | // inode | |
6057 | CInode *in = NULL; | |
6058 | if (dnl->is_primary()) | |
6059 | in = dnl->get_inode(); | |
6060 | ||
6061 | // dentry | |
181888fb FG |
6062 | for (auto &r : dn->get_replicas()) { |
6063 | auto it = acks.find(r.first); | |
31f18b77 FG |
6064 | if (it == acks.end()) |
6065 | continue; | |
94b18763 | 6066 | it->second->add_strong_dentry(dir->dirfrag(), dn->get_name(), dn->first, dn->last, |
7c673cae FG |
6067 | dnl->is_primary() ? dnl->get_inode()->ino():inodeno_t(0), |
6068 | dnl->is_remote() ? dnl->get_remote_ino():inodeno_t(0), | |
6069 | dnl->is_remote() ? dnl->get_remote_d_type():0, | |
181888fb | 6070 | ++r.second, |
7c673cae FG |
6071 | dn->lock.get_replica_state()); |
6072 | // peer missed MDentrylink message ? | |
181888fb FG |
6073 | if (in && !in->is_replica(r.first)) |
6074 | in->add_replica(r.first); | |
7c673cae FG |
6075 | } |
6076 | ||
6077 | if (!in) | |
6078 | continue; | |
6079 | ||
181888fb FG |
6080 | for (auto &r : in->get_replicas()) { |
6081 | auto it = acks.find(r.first); | |
31f18b77 FG |
6082 | if (it == acks.end()) |
6083 | continue; | |
6084 | it->second->add_inode_base(in, mds->mdsmap->get_up_features()); | |
7c673cae | 6085 | bufferlist bl; |
181888fb FG |
6086 | in->_encode_locks_state_for_rejoin(bl, r.first); |
6087 | it->second->add_inode_locks(in, ++r.second, bl); | |
7c673cae FG |
6088 | } |
6089 | ||
6090 | // subdirs in this subtree? | |
6091 | in->get_nested_dirfrags(dq); | |
6092 | } | |
6093 | } | |
6094 | } | |
6095 | ||
6096 | // base inodes too | |
6097 | if (root && root->is_auth()) | |
181888fb FG |
6098 | for (auto &r : root->get_replicas()) { |
6099 | auto it = acks.find(r.first); | |
31f18b77 FG |
6100 | if (it == acks.end()) |
6101 | continue; | |
6102 | it->second->add_inode_base(root, mds->mdsmap->get_up_features()); | |
7c673cae | 6103 | bufferlist bl; |
181888fb FG |
6104 | root->_encode_locks_state_for_rejoin(bl, r.first); |
6105 | it->second->add_inode_locks(root, ++r.second, bl); | |
7c673cae FG |
6106 | } |
6107 | if (myin) | |
181888fb FG |
6108 | for (auto &r : myin->get_replicas()) { |
6109 | auto it = acks.find(r.first); | |
31f18b77 FG |
6110 | if (it == acks.end()) |
6111 | continue; | |
6112 | it->second->add_inode_base(myin, mds->mdsmap->get_up_features()); | |
7c673cae | 6113 | bufferlist bl; |
181888fb FG |
6114 | myin->_encode_locks_state_for_rejoin(bl, r.first); |
6115 | it->second->add_inode_locks(myin, ++r.second, bl); | |
7c673cae FG |
6116 | } |
6117 | ||
6118 | // include inode base for any inodes whose scatterlocks may have updated | |
6119 | for (set<CInode*>::iterator p = rejoin_potential_updated_scatterlocks.begin(); | |
6120 | p != rejoin_potential_updated_scatterlocks.end(); | |
6121 | ++p) { | |
6122 | CInode *in = *p; | |
181888fb FG |
6123 | for (const auto &r : in->get_replicas()) { |
6124 | auto it = acks.find(r.first); | |
31f18b77 FG |
6125 | if (it == acks.end()) |
6126 | continue; | |
6127 | it->second->add_inode_base(in, mds->mdsmap->get_up_features()); | |
6128 | } | |
7c673cae FG |
6129 | } |
6130 | ||
6131 | // send acks | |
31f18b77 | 6132 | for (auto p = acks.begin(); p != acks.end(); ++p) { |
11fdf7f2 | 6133 | encode(rejoin_imported_caps[p->first], p->second->imported_caps); |
7c673cae FG |
6134 | mds->send_message_mds(p->second, p->first); |
6135 | } | |
6136 | ||
6137 | rejoin_imported_caps.clear(); | |
6138 | } | |
6139 | ||
c07f9fc5 FG |
6140 | class C_MDC_ReIssueCaps : public MDCacheContext { |
6141 | CInode *in; | |
6142 | public: | |
6143 | C_MDC_ReIssueCaps(MDCache *mdc, CInode *i) : | |
6144 | MDCacheContext(mdc), in(i) | |
6145 | { | |
6146 | in->get(CInode::PIN_PTRWAITER); | |
6147 | } | |
6148 | void finish(int r) override { | |
6149 | if (!mdcache->mds->locker->eval(in, CEPH_CAP_LOCKS)) | |
6150 | mdcache->mds->locker->issue_caps(in); | |
6151 | in->put(CInode::PIN_PTRWAITER); | |
6152 | } | |
6153 | }; | |
7c673cae FG |
6154 | |
6155 | void MDCache::reissue_all_caps() | |
6156 | { | |
6157 | dout(10) << "reissue_all_caps" << dendl; | |
6158 | ||
81eedcae | 6159 | int count = 0; |
94b18763 | 6160 | for (auto &p : inode_map) { |
81eedcae | 6161 | int n = 1; |
b32b8144 | 6162 | CInode *in = p.second; |
7c673cae | 6163 | if (in->is_head() && in->is_any_caps()) { |
c07f9fc5 FG |
6164 | // called by MDSRank::active_start(). There shouldn't be any frozen subtree. |
6165 | if (in->is_frozen_inode()) { | |
6166 | in->add_waiter(CInode::WAIT_UNFREEZE, new C_MDC_ReIssueCaps(this, in)); | |
6167 | continue; | |
6168 | } | |
7c673cae | 6169 | if (!mds->locker->eval(in, CEPH_CAP_LOCKS)) |
81eedcae | 6170 | n += mds->locker->issue_caps(in); |
7c673cae | 6171 | } |
81eedcae TL |
6172 | |
6173 | if ((count % 1000) + n >= 1000) | |
6174 | mds->heartbeat_reset(); | |
6175 | count += n; | |
7c673cae FG |
6176 | } |
6177 | } | |
6178 | ||
6179 | ||
6180 | // =============================================================================== | |
6181 | ||
6182 | struct C_MDC_QueuedCow : public MDCacheContext { | |
6183 | CInode *in; | |
6184 | MutationRef mut; | |
6185 | C_MDC_QueuedCow(MDCache *mdc, CInode *i, MutationRef& m) : | |
6186 | MDCacheContext(mdc), in(i), mut(m) {} | |
6187 | void finish(int r) override { | |
6188 | mdcache->_queued_file_recover_cow(in, mut); | |
6189 | } | |
6190 | }; | |
6191 | ||
6192 | ||
6193 | void MDCache::queue_file_recover(CInode *in) | |
6194 | { | |
6195 | dout(10) << "queue_file_recover " << *in << dendl; | |
11fdf7f2 | 6196 | ceph_assert(in->is_auth()); |
7c673cae FG |
6197 | |
6198 | // cow? | |
6199 | /* | |
6200 | SnapRealm *realm = in->find_snaprealm(); | |
6201 | set<snapid_t> s = realm->get_snaps(); | |
6202 | while (!s.empty() && *s.begin() < in->first) | |
6203 | s.erase(s.begin()); | |
6204 | while (!s.empty() && *s.rbegin() > in->last) | |
6205 | s.erase(*s.rbegin()); | |
6206 | dout(10) << " snaps in [" << in->first << "," << in->last << "] are " << s << dendl; | |
6207 | if (s.size() > 1) { | |
94b18763 | 6208 | CInode::mempool_inode pi = in->project_inode(); |
7c673cae FG |
6209 | pi->version = in->pre_dirty(); |
6210 | ||
6211 | auto mut(std::make_shared<MutationImpl>()); | |
6212 | mut->ls = mds->mdlog->get_current_segment(); | |
6213 | EUpdate *le = new EUpdate(mds->mdlog, "queue_file_recover cow"); | |
6214 | mds->mdlog->start_entry(le); | |
6215 | predirty_journal_parents(mut, &le->metablob, in, 0, PREDIRTY_PRIMARY); | |
6216 | ||
6217 | s.erase(*s.begin()); | |
6218 | while (!s.empty()) { | |
6219 | snapid_t snapid = *s.begin(); | |
6220 | CInode *cow_inode = 0; | |
6221 | journal_cow_inode(mut, &le->metablob, in, snapid-1, &cow_inode); | |
11fdf7f2 | 6222 | ceph_assert(cow_inode); |
7c673cae FG |
6223 | recovery_queue.enqueue(cow_inode); |
6224 | s.erase(*s.begin()); | |
6225 | } | |
6226 | ||
6227 | in->parent->first = in->first; | |
6228 | le->metablob.add_primary_dentry(in->parent, in, true); | |
6229 | mds->mdlog->submit_entry(le, new C_MDC_QueuedCow(this, in, mut)); | |
6230 | mds->mdlog->flush(); | |
6231 | } | |
6232 | */ | |
6233 | ||
6234 | recovery_queue.enqueue(in); | |
6235 | } | |
6236 | ||
6237 | void MDCache::_queued_file_recover_cow(CInode *in, MutationRef& mut) | |
6238 | { | |
6239 | in->pop_and_dirty_projected_inode(mut->ls); | |
6240 | mut->apply(); | |
6241 | mds->locker->drop_locks(mut.get()); | |
6242 | mut->cleanup(); | |
6243 | } | |
6244 | ||
6245 | ||
6246 | /* | |
6247 | * called after recovery to recover file sizes for previously opened (for write) | |
6248 | * files. that is, those where max_size > size. | |
6249 | */ | |
6250 | void MDCache::identify_files_to_recover() | |
6251 | { | |
6252 | dout(10) << "identify_files_to_recover" << dendl; | |
81eedcae | 6253 | int count = 0; |
94b18763 | 6254 | for (auto &p : inode_map) { |
b32b8144 | 6255 | CInode *in = p.second; |
7c673cae FG |
6256 | if (!in->is_auth()) |
6257 | continue; | |
6258 | ||
6259 | if (in->last != CEPH_NOSNAP) | |
6260 | continue; | |
6261 | ||
6262 | // Only normal files need file size recovery | |
6263 | if (!in->is_file()) { | |
6264 | continue; | |
6265 | } | |
6266 | ||
6267 | bool recover = false; | |
6268 | for (map<client_t,client_writeable_range_t>::iterator p = in->inode.client_ranges.begin(); | |
6269 | p != in->inode.client_ranges.end(); | |
6270 | ++p) { | |
6271 | Capability *cap = in->get_client_cap(p->first); | |
a8e16298 TL |
6272 | if (cap) { |
6273 | cap->mark_clientwriteable(); | |
6274 | } else { | |
7c673cae FG |
6275 | dout(10) << " client." << p->first << " has range " << p->second << " but no cap on " << *in << dendl; |
6276 | recover = true; | |
6277 | break; | |
6278 | } | |
6279 | } | |
6280 | ||
6281 | if (recover) { | |
6282 | if (in->filelock.is_stable()) { | |
6283 | in->auth_pin(&in->filelock); | |
6284 | } else { | |
11fdf7f2 | 6285 | ceph_assert(in->filelock.get_state() == LOCK_XLOCKSNAP); |
7c673cae FG |
6286 | } |
6287 | in->filelock.set_state(LOCK_PRE_SCAN); | |
6288 | rejoin_recover_q.push_back(in); | |
6289 | } else { | |
6290 | rejoin_check_q.push_back(in); | |
6291 | } | |
81eedcae TL |
6292 | |
6293 | if (!(++count % 1000)) | |
6294 | mds->heartbeat_reset(); | |
7c673cae FG |
6295 | } |
6296 | } | |
6297 | ||
6298 | void MDCache::start_files_to_recover() | |
6299 | { | |
6300 | for (CInode *in : rejoin_check_q) { | |
6301 | if (in->filelock.get_state() == LOCK_XLOCKSNAP) | |
6302 | mds->locker->issue_caps(in); | |
6303 | mds->locker->check_inode_max_size(in); | |
6304 | } | |
6305 | rejoin_check_q.clear(); | |
6306 | for (CInode *in : rejoin_recover_q) { | |
6307 | mds->locker->file_recover(&in->filelock); | |
6308 | } | |
6309 | if (!rejoin_recover_q.empty()) { | |
6310 | rejoin_recover_q.clear(); | |
6311 | do_file_recover(); | |
6312 | } | |
6313 | } | |
6314 | ||
6315 | void MDCache::do_file_recover() | |
6316 | { | |
6317 | recovery_queue.advance(); | |
6318 | } | |
6319 | ||
6320 | // =============================================================================== | |
6321 | ||
6322 | ||
6323 | // ---------------------------- | |
6324 | // truncate | |
6325 | ||
6326 | class C_MDC_RetryTruncate : public MDCacheContext { | |
6327 | CInode *in; | |
6328 | LogSegment *ls; | |
6329 | public: | |
6330 | C_MDC_RetryTruncate(MDCache *c, CInode *i, LogSegment *l) : | |
6331 | MDCacheContext(c), in(i), ls(l) {} | |
6332 | void finish(int r) override { | |
6333 | mdcache->_truncate_inode(in, ls); | |
6334 | } | |
6335 | }; | |
6336 | ||
6337 | void MDCache::truncate_inode(CInode *in, LogSegment *ls) | |
6338 | { | |
94b18763 | 6339 | auto pi = in->get_projected_inode(); |
7c673cae FG |
6340 | dout(10) << "truncate_inode " |
6341 | << pi->truncate_from << " -> " << pi->truncate_size | |
6342 | << " on " << *in | |
6343 | << dendl; | |
6344 | ||
6345 | ls->truncating_inodes.insert(in); | |
6346 | in->get(CInode::PIN_TRUNCATING); | |
6347 | in->auth_pin(this); | |
6348 | ||
6349 | if (!in->client_need_snapflush.empty() && | |
6350 | (in->get_caps_issued() & CEPH_CAP_FILE_BUFFER)) { | |
11fdf7f2 | 6351 | ceph_assert(in->filelock.is_xlocked()); |
7c673cae FG |
6352 | in->filelock.set_xlock_snap_sync(new C_MDC_RetryTruncate(this, in, ls)); |
6353 | mds->locker->issue_caps(in); | |
6354 | return; | |
6355 | } | |
6356 | ||
6357 | _truncate_inode(in, ls); | |
6358 | } | |
6359 | ||
6360 | struct C_IO_MDC_TruncateFinish : public MDCacheIOContext { | |
6361 | CInode *in; | |
6362 | LogSegment *ls; | |
6363 | C_IO_MDC_TruncateFinish(MDCache *c, CInode *i, LogSegment *l) : | |
91327a77 AA |
6364 | MDCacheIOContext(c, false), in(i), ls(l) { |
6365 | } | |
7c673cae | 6366 | void finish(int r) override { |
11fdf7f2 | 6367 | ceph_assert(r == 0 || r == -ENOENT); |
7c673cae FG |
6368 | mdcache->truncate_inode_finish(in, ls); |
6369 | } | |
91327a77 AA |
6370 | void print(ostream& out) const override { |
6371 | out << "file_truncate(" << in->ino() << ")"; | |
6372 | } | |
7c673cae FG |
6373 | }; |
6374 | ||
6375 | void MDCache::_truncate_inode(CInode *in, LogSegment *ls) | |
6376 | { | |
94b18763 | 6377 | auto pi = &in->inode; |
7c673cae FG |
6378 | dout(10) << "_truncate_inode " |
6379 | << pi->truncate_from << " -> " << pi->truncate_size | |
6380 | << " on " << *in << dendl; | |
6381 | ||
11fdf7f2 TL |
6382 | ceph_assert(pi->is_truncating()); |
6383 | ceph_assert(pi->truncate_size < (1ULL << 63)); | |
6384 | ceph_assert(pi->truncate_from < (1ULL << 63)); | |
6385 | ceph_assert(pi->truncate_size < pi->truncate_from); | |
7c673cae FG |
6386 | |
6387 | ||
6388 | SnapRealm *realm = in->find_snaprealm(); | |
6389 | SnapContext nullsnap; | |
6390 | const SnapContext *snapc; | |
6391 | if (realm) { | |
6392 | dout(10) << " realm " << *realm << dendl; | |
6393 | snapc = &realm->get_snap_context(); | |
6394 | } else { | |
6395 | dout(10) << " NO realm, using null context" << dendl; | |
6396 | snapc = &nullsnap; | |
11fdf7f2 | 6397 | ceph_assert(in->last == CEPH_NOSNAP); |
7c673cae FG |
6398 | } |
6399 | dout(10) << "_truncate_inode snapc " << snapc << " on " << *in << dendl; | |
6400 | filer.truncate(in->inode.ino, &in->inode.layout, *snapc, | |
6401 | pi->truncate_size, pi->truncate_from-pi->truncate_size, | |
6402 | pi->truncate_seq, ceph::real_time::min(), 0, | |
6403 | new C_OnFinisher(new C_IO_MDC_TruncateFinish(this, in, ls), | |
6404 | mds->finisher)); | |
6405 | } | |
6406 | ||
6407 | struct C_MDC_TruncateLogged : public MDCacheLogContext { | |
6408 | CInode *in; | |
6409 | MutationRef mut; | |
6410 | C_MDC_TruncateLogged(MDCache *m, CInode *i, MutationRef& mu) : | |
6411 | MDCacheLogContext(m), in(i), mut(mu) {} | |
6412 | void finish(int r) override { | |
6413 | mdcache->truncate_inode_logged(in, mut); | |
6414 | } | |
6415 | }; | |
6416 | ||
6417 | void MDCache::truncate_inode_finish(CInode *in, LogSegment *ls) | |
6418 | { | |
6419 | dout(10) << "truncate_inode_finish " << *in << dendl; | |
6420 | ||
6421 | set<CInode*>::iterator p = ls->truncating_inodes.find(in); | |
11fdf7f2 | 6422 | ceph_assert(p != ls->truncating_inodes.end()); |
7c673cae FG |
6423 | ls->truncating_inodes.erase(p); |
6424 | ||
6425 | // update | |
94b18763 FG |
6426 | auto &pi = in->project_inode(); |
6427 | pi.inode.version = in->pre_dirty(); | |
6428 | pi.inode.truncate_from = 0; | |
6429 | pi.inode.truncate_pending--; | |
7c673cae FG |
6430 | |
6431 | MutationRef mut(new MutationImpl()); | |
6432 | mut->ls = mds->mdlog->get_current_segment(); | |
6433 | mut->add_projected_inode(in); | |
6434 | ||
6435 | EUpdate *le = new EUpdate(mds->mdlog, "truncate finish"); | |
6436 | mds->mdlog->start_entry(le); | |
6437 | CDentry *dn = in->get_projected_parent_dn(); | |
6438 | le->metablob.add_dir_context(dn->get_dir()); | |
6439 | le->metablob.add_primary_dentry(dn, in, true); | |
6440 | le->metablob.add_truncate_finish(in->ino(), ls->seq); | |
6441 | ||
6442 | journal_dirty_inode(mut.get(), &le->metablob, in); | |
6443 | mds->mdlog->submit_entry(le, new C_MDC_TruncateLogged(this, in, mut)); | |
6444 | ||
6445 | // flush immediately if there are readers/writers waiting | |
6446 | if (in->is_waiter_for(CInode::WAIT_TRUNC) || | |
6447 | (in->get_caps_wanted() & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR))) | |
6448 | mds->mdlog->flush(); | |
6449 | } | |
6450 | ||
6451 | void MDCache::truncate_inode_logged(CInode *in, MutationRef& mut) | |
6452 | { | |
6453 | dout(10) << "truncate_inode_logged " << *in << dendl; | |
6454 | mut->apply(); | |
6455 | mds->locker->drop_locks(mut.get()); | |
6456 | mut->cleanup(); | |
6457 | ||
6458 | in->put(CInode::PIN_TRUNCATING); | |
6459 | in->auth_unpin(this); | |
6460 | ||
11fdf7f2 | 6461 | MDSContext::vec waiters; |
7c673cae FG |
6462 | in->take_waiting(CInode::WAIT_TRUNC, waiters); |
6463 | mds->queue_waiters(waiters); | |
6464 | } | |
6465 | ||
6466 | ||
6467 | void MDCache::add_recovered_truncate(CInode *in, LogSegment *ls) | |
6468 | { | |
6469 | dout(20) << "add_recovered_truncate " << *in << " in log segment " | |
6470 | << ls->seq << "/" << ls->offset << dendl; | |
6471 | ls->truncating_inodes.insert(in); | |
6472 | in->get(CInode::PIN_TRUNCATING); | |
6473 | } | |
6474 | ||
6475 | void MDCache::remove_recovered_truncate(CInode *in, LogSegment *ls) | |
6476 | { | |
6477 | dout(20) << "remove_recovered_truncate " << *in << " in log segment " | |
6478 | << ls->seq << "/" << ls->offset << dendl; | |
6479 | // if we have the logseg the truncate started in, it must be in our list. | |
6480 | set<CInode*>::iterator p = ls->truncating_inodes.find(in); | |
11fdf7f2 | 6481 | ceph_assert(p != ls->truncating_inodes.end()); |
7c673cae FG |
6482 | ls->truncating_inodes.erase(p); |
6483 | in->put(CInode::PIN_TRUNCATING); | |
6484 | } | |
6485 | ||
6486 | void MDCache::start_recovered_truncates() | |
6487 | { | |
6488 | dout(10) << "start_recovered_truncates" << dendl; | |
6489 | for (map<uint64_t,LogSegment*>::iterator p = mds->mdlog->segments.begin(); | |
6490 | p != mds->mdlog->segments.end(); | |
6491 | ++p) { | |
6492 | LogSegment *ls = p->second; | |
6493 | for (set<CInode*>::iterator q = ls->truncating_inodes.begin(); | |
6494 | q != ls->truncating_inodes.end(); | |
6495 | ++q) { | |
6496 | CInode *in = *q; | |
6497 | in->auth_pin(this); | |
6498 | ||
6499 | if (!in->client_need_snapflush.empty() && | |
6500 | (in->get_caps_issued() & CEPH_CAP_FILE_BUFFER)) { | |
11fdf7f2 | 6501 | ceph_assert(in->filelock.is_stable()); |
7c673cae FG |
6502 | in->filelock.set_state(LOCK_XLOCKDONE); |
6503 | in->auth_pin(&in->filelock); | |
6504 | in->filelock.set_xlock_snap_sync(new C_MDC_RetryTruncate(this, in, ls)); | |
6505 | // start_files_to_recover will revoke caps | |
6506 | continue; | |
6507 | } | |
6508 | _truncate_inode(in, ls); | |
6509 | } | |
6510 | } | |
6511 | } | |
6512 | ||
6513 | ||
6514 | ||
6515 | ||
6516 | ||
6517 | ||
6518 | // ================================================================================ | |
6519 | // cache trimming | |
6520 | ||
11fdf7f2 | 6521 | std::pair<bool, uint64_t> MDCache::trim_lru(uint64_t count, expiremap& expiremap) |
181888fb | 6522 | { |
7c673cae | 6523 | bool is_standby_replay = mds->is_standby_replay(); |
181888fb FG |
6524 | std::vector<CDentry *> unexpirables; |
6525 | uint64_t trimmed = 0; | |
6526 | ||
11fdf7f2 | 6527 | auto trim_threshold = g_conf().get_val<Option::size_t>("mds_cache_trim_threshold"); |
a8e16298 | 6528 | |
181888fb FG |
6529 | dout(7) << "trim_lru trimming " << count |
6530 | << " items from LRU" | |
6531 | << " size=" << lru.lru_get_size() | |
6532 | << " mid=" << lru.lru_get_top() | |
6533 | << " pintail=" << lru.lru_get_pintail() | |
6534 | << " pinned=" << lru.lru_get_num_pinned() | |
6535 | << dendl; | |
7c673cae | 6536 | |
11fdf7f2 | 6537 | const uint64_t trim_counter_start = trim_counter.get(); |
a8e16298 TL |
6538 | bool throttled = false; |
6539 | while (1) { | |
6540 | throttled |= trim_counter_start+trimmed >= trim_threshold; | |
6541 | if (throttled) break; | |
31f18b77 FG |
6542 | CDentry *dn = static_cast<CDentry*>(bottom_lru.lru_expire()); |
6543 | if (!dn) | |
6544 | break; | |
6545 | if (trim_dentry(dn, expiremap)) { | |
6546 | unexpirables.push_back(dn); | |
181888fb FG |
6547 | } else { |
6548 | trimmed++; | |
31f18b77 FG |
6549 | } |
6550 | } | |
6551 | ||
181888fb | 6552 | for (auto &dn : unexpirables) { |
31f18b77 | 6553 | bottom_lru.lru_insert_mid(dn); |
181888fb | 6554 | } |
31f18b77 FG |
6555 | unexpirables.clear(); |
6556 | ||
181888fb | 6557 | // trim dentries from the LRU until count is reached |
a8e16298 TL |
6558 | while (!throttled && (cache_toofull() || count > 0)) { |
6559 | throttled |= trim_counter_start+trimmed >= trim_threshold; | |
6560 | if (throttled) break; | |
7c673cae FG |
6561 | CDentry *dn = static_cast<CDentry*>(lru.lru_expire()); |
6562 | if (!dn) { | |
6563 | break; | |
6564 | } | |
7c673cae | 6565 | if ((is_standby_replay && dn->get_linkage()->inode && |
181888fb | 6566 | dn->get_linkage()->inode->item_open_file.is_on_list())) { |
7c673cae | 6567 | unexpirables.push_back(dn); |
181888fb FG |
6568 | } else if (trim_dentry(dn, expiremap)) { |
6569 | unexpirables.push_back(dn); | |
6570 | } else { | |
6571 | trimmed++; | |
3efd9988 | 6572 | if (count > 0) count--; |
7c673cae FG |
6573 | } |
6574 | } | |
11fdf7f2 | 6575 | trim_counter.hit(trimmed); |
181888fb FG |
6576 | |
6577 | for (auto &dn : unexpirables) { | |
31f18b77 | 6578 | lru.lru_insert_mid(dn); |
181888fb | 6579 | } |
31f18b77 | 6580 | unexpirables.clear(); |
7c673cae | 6581 | |
181888fb | 6582 | dout(7) << "trim_lru trimmed " << trimmed << " items" << dendl; |
a8e16298 | 6583 | return std::pair<bool, uint64_t>(throttled, trimmed); |
181888fb FG |
6584 | } |
6585 | ||
6586 | /* | |
6587 | * note: only called while MDS is active or stopping... NOT during recovery. | |
6588 | * however, we may expire a replica whose authority is recovering. | |
6589 | * | |
6590 | * @param count is number of dentries to try to expire | |
6591 | */ | |
a8e16298 | 6592 | std::pair<bool, uint64_t> MDCache::trim(uint64_t count) |
181888fb FG |
6593 | { |
6594 | uint64_t used = cache_size(); | |
91327a77 | 6595 | uint64_t limit = cache_memory_limit; |
11fdf7f2 | 6596 | expiremap expiremap; |
181888fb FG |
6597 | |
6598 | dout(7) << "trim bytes_used=" << bytes2str(used) | |
6599 | << " limit=" << bytes2str(limit) | |
91327a77 | 6600 | << " reservation=" << cache_reservation |
181888fb FG |
6601 | << "% count=" << count << dendl; |
6602 | ||
6603 | // process delayed eval_stray() | |
6604 | stray_manager.advance_delayed(); | |
6605 | ||
a8e16298 TL |
6606 | auto result = trim_lru(count, expiremap); |
6607 | auto& trimmed = result.second; | |
181888fb | 6608 | |
7c673cae | 6609 | // trim non-auth, non-bound subtrees |
181888fb | 6610 | for (auto p = subtrees.begin(); p != subtrees.end();) { |
7c673cae FG |
6611 | CDir *dir = p->first; |
6612 | ++p; | |
31f18b77 FG |
6613 | CInode *diri = dir->get_inode(); |
6614 | if (dir->is_auth()) { | |
6615 | if (!diri->is_auth() && !diri->is_base() && | |
6616 | dir->get_num_head_items() == 0) { | |
6617 | if (dir->state_test(CDir::STATE_EXPORTING) || | |
181888fb | 6618 | !(mds->is_active() || mds->is_stopping()) || |
31f18b77 FG |
6619 | dir->is_freezing() || dir->is_frozen()) |
6620 | continue; | |
6621 | ||
6622 | migrator->export_empty_import(dir); | |
a8e16298 | 6623 | ++trimmed; |
31f18b77 FG |
6624 | } |
6625 | } else { | |
6626 | if (!diri->is_auth()) { | |
6627 | if (dir->get_num_ref() > 1) // only subtree pin | |
6628 | continue; | |
6629 | list<CDir*> ls; | |
6630 | diri->get_subtree_dirfrags(ls); | |
6631 | if (diri->get_num_ref() > (int)ls.size()) // only pinned by subtrees | |
6632 | continue; | |
6633 | ||
6634 | // don't trim subtree root if its auth MDS is recovering. | |
6635 | // This simplify the cache rejoin code. | |
6636 | if (dir->is_subtree_root() && | |
6637 | rejoin_ack_gather.count(dir->get_dir_auth().first)) | |
6638 | continue; | |
7c673cae | 6639 | trim_dirfrag(dir, 0, expiremap); |
a8e16298 | 6640 | ++trimmed; |
31f18b77 | 6641 | } |
7c673cae FG |
6642 | } |
6643 | } | |
6644 | ||
6645 | // trim root? | |
181888fb | 6646 | if (mds->is_stopping() && root) { |
7c673cae FG |
6647 | list<CDir*> ls; |
6648 | root->get_dirfrags(ls); | |
6649 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
6650 | CDir *dir = *p; | |
a8e16298 | 6651 | if (dir->get_num_ref() == 1) { // subtree pin |
7c673cae | 6652 | trim_dirfrag(dir, 0, expiremap); |
a8e16298 TL |
6653 | ++trimmed; |
6654 | } | |
7c673cae | 6655 | } |
a8e16298 | 6656 | if (root->get_num_ref() == 0) { |
7c673cae | 6657 | trim_inode(0, root, 0, expiremap); |
a8e16298 TL |
6658 | ++trimmed; |
6659 | } | |
7c673cae FG |
6660 | } |
6661 | ||
6662 | std::set<mds_rank_t> stopping; | |
6663 | mds->mdsmap->get_mds_set(stopping, MDSMap::STATE_STOPPING); | |
6664 | stopping.erase(mds->get_nodeid()); | |
6665 | for (auto rank : stopping) { | |
6666 | CInode* mdsdir_in = get_inode(MDS_INO_MDSDIR(rank)); | |
6667 | if (!mdsdir_in) | |
6668 | continue; | |
6669 | ||
11fdf7f2 TL |
6670 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(rank), std::forward_as_tuple()); |
6671 | if (em.second) { | |
6672 | em.first->second = MCacheExpire::create(mds->get_nodeid()); | |
7c673cae FG |
6673 | } |
6674 | ||
6675 | dout(20) << __func__ << ": try expiring " << *mdsdir_in << " for stopping mds." << mds << dendl; | |
6676 | ||
6677 | const bool aborted = expire_recursive(mdsdir_in, expiremap); | |
6678 | if (!aborted) { | |
6679 | dout(20) << __func__ << ": successfully expired mdsdir" << dendl; | |
6680 | list<CDir*> ls; | |
6681 | mdsdir_in->get_dirfrags(ls); | |
6682 | for (auto dir : ls) { | |
a8e16298 | 6683 | if (dir->get_num_ref() == 1) { // subtree pin |
7c673cae | 6684 | trim_dirfrag(dir, dir, expiremap); |
a8e16298 TL |
6685 | ++trimmed; |
6686 | } | |
7c673cae | 6687 | } |
a8e16298 | 6688 | if (mdsdir_in->get_num_ref() == 0) { |
7c673cae | 6689 | trim_inode(NULL, mdsdir_in, NULL, expiremap); |
a8e16298 TL |
6690 | ++trimmed; |
6691 | } | |
7c673cae FG |
6692 | } else { |
6693 | dout(20) << __func__ << ": some unexpirable contents in mdsdir" << dendl; | |
6694 | } | |
6695 | } | |
6696 | ||
6697 | // Other rank's base inodes (when I'm stopping) | |
181888fb | 6698 | if (mds->is_stopping()) { |
7c673cae | 6699 | for (set<CInode*>::iterator p = base_inodes.begin(); |
11fdf7f2 TL |
6700 | p != base_inodes.end();) { |
6701 | CInode *base_in = *p; | |
6702 | ++p; | |
6703 | if (MDS_INO_IS_MDSDIR(base_in->ino()) && | |
6704 | MDS_INO_MDSDIR_OWNER(base_in->ino()) != mds->get_nodeid()) { | |
6705 | dout(20) << __func__ << ": maybe trimming base: " << *base_in << dendl; | |
6706 | if (base_in->get_num_ref() == 0) { | |
6707 | trim_inode(NULL, base_in, NULL, expiremap); | |
a8e16298 | 6708 | ++trimmed; |
7c673cae FG |
6709 | } |
6710 | } | |
6711 | } | |
6712 | } | |
6713 | ||
6714 | // send any expire messages | |
6715 | send_expire_messages(expiremap); | |
6716 | ||
a8e16298 | 6717 | return result; |
7c673cae FG |
6718 | } |
6719 | ||
11fdf7f2 | 6720 | void MDCache::send_expire_messages(expiremap& expiremap) |
7c673cae FG |
6721 | { |
6722 | // send expires | |
11fdf7f2 | 6723 | for (const auto &p : expiremap) { |
7c673cae | 6724 | if (mds->is_cluster_degraded() && |
11fdf7f2 TL |
6725 | (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || |
6726 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
6727 | rejoin_sent.count(p.first) == 0))) { | |
7c673cae FG |
6728 | continue; |
6729 | } | |
11fdf7f2 TL |
6730 | dout(7) << "sending cache_expire to " << p.first << dendl; |
6731 | mds->send_message_mds(p.second, p.first); | |
7c673cae | 6732 | } |
11fdf7f2 | 6733 | expiremap.clear(); |
7c673cae FG |
6734 | } |
6735 | ||
6736 | ||
11fdf7f2 | 6737 | bool MDCache::trim_dentry(CDentry *dn, expiremap& expiremap) |
7c673cae FG |
6738 | { |
6739 | dout(12) << "trim_dentry " << *dn << dendl; | |
6740 | ||
6741 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
6742 | ||
6743 | CDir *dir = dn->get_dir(); | |
11fdf7f2 | 6744 | ceph_assert(dir); |
7c673cae FG |
6745 | |
6746 | CDir *con = get_subtree_root(dir); | |
6747 | if (con) | |
6748 | dout(12) << " in container " << *con << dendl; | |
6749 | else { | |
6750 | dout(12) << " no container; under a not-yet-linked dir" << dendl; | |
11fdf7f2 | 6751 | ceph_assert(dn->is_auth()); |
7c673cae FG |
6752 | } |
6753 | ||
6754 | // If replica dentry is not readable, it's likely we will receive | |
6755 | // MDentryLink/MDentryUnlink message soon (It's possible we first | |
6756 | // receive a MDentryUnlink message, then MDentryLink message) | |
6757 | // MDentryLink message only replicates an inode, so we should | |
6758 | // avoid trimming the inode's parent dentry. This is because that | |
6759 | // unconnected replicas are problematic for subtree migration. | |
6760 | if (!dn->is_auth() && !dn->lock.can_read(-1) && | |
6761 | !dn->get_dir()->get_inode()->is_stray()) | |
6762 | return true; | |
6763 | ||
6764 | // adjust the dir state | |
6765 | // NOTE: we can safely remove a clean, null dentry without effecting | |
6766 | // directory completeness. | |
6767 | // (check this _before_ we unlink the inode, below!) | |
6768 | bool clear_complete = false; | |
6769 | if (!(dnl->is_null() && dn->is_clean())) | |
6770 | clear_complete = true; | |
6771 | ||
6772 | // unlink the dentry | |
6773 | if (dnl->is_remote()) { | |
6774 | // just unlink. | |
31f18b77 | 6775 | dir->unlink_inode(dn, false); |
7c673cae FG |
6776 | } else if (dnl->is_primary()) { |
6777 | // expire the inode, too. | |
6778 | CInode *in = dnl->get_inode(); | |
11fdf7f2 | 6779 | ceph_assert(in); |
7c673cae FG |
6780 | if (trim_inode(dn, in, con, expiremap)) |
6781 | return true; // purging stray instead of trimming | |
6782 | } else { | |
11fdf7f2 | 6783 | ceph_assert(dnl->is_null()); |
7c673cae FG |
6784 | } |
6785 | ||
6786 | if (!dn->is_auth()) { | |
6787 | // notify dentry authority. | |
6788 | mds_authority_t auth = dn->authority(); | |
6789 | ||
6790 | for (int p=0; p<2; p++) { | |
6791 | mds_rank_t a = auth.first; | |
6792 | if (p) a = auth.second; | |
6793 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6794 | if (mds->get_nodeid() == auth.second && | |
6795 | con->is_importing()) break; // don't send any expire while importing. | |
6796 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6797 | ||
6798 | dout(12) << " sending expire to mds." << a << " on " << *dn << dendl; | |
11fdf7f2 TL |
6799 | ceph_assert(a != mds->get_nodeid()); |
6800 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(a), std::forward_as_tuple()); | |
6801 | if (em.second) | |
6802 | em.first->second = MCacheExpire::create(mds->get_nodeid()); | |
6803 | em.first->second->add_dentry(con->dirfrag(), dir->dirfrag(), dn->get_name(), dn->last, dn->get_replica_nonce()); | |
7c673cae FG |
6804 | } |
6805 | } | |
6806 | ||
6807 | // remove dentry | |
6808 | if (dn->last == CEPH_NOSNAP && dir->is_auth()) | |
6809 | dir->add_to_bloom(dn); | |
6810 | dir->remove_dentry(dn); | |
6811 | ||
6812 | if (clear_complete) | |
6813 | dir->state_clear(CDir::STATE_COMPLETE); | |
6814 | ||
7c673cae FG |
6815 | if (mds->logger) mds->logger->inc(l_mds_inodes_expired); |
6816 | return false; | |
6817 | } | |
6818 | ||
6819 | ||
11fdf7f2 | 6820 | void MDCache::trim_dirfrag(CDir *dir, CDir *con, expiremap& expiremap) |
7c673cae FG |
6821 | { |
6822 | dout(15) << "trim_dirfrag " << *dir << dendl; | |
6823 | ||
6824 | if (dir->is_subtree_root()) { | |
11fdf7f2 | 6825 | ceph_assert(!dir->is_auth() || |
7c673cae FG |
6826 | (!dir->is_replicated() && dir->inode->is_base())); |
6827 | remove_subtree(dir); // remove from subtree map | |
6828 | } | |
11fdf7f2 | 6829 | ceph_assert(dir->get_num_ref() == 0); |
7c673cae FG |
6830 | |
6831 | CInode *in = dir->get_inode(); | |
6832 | ||
6833 | if (!dir->is_auth()) { | |
6834 | mds_authority_t auth = dir->authority(); | |
6835 | ||
6836 | // was this an auth delegation? (if so, slightly modified container) | |
6837 | dirfrag_t condf; | |
6838 | if (dir->is_subtree_root()) { | |
6839 | dout(12) << " subtree root, container is " << *dir << dendl; | |
6840 | con = dir; | |
6841 | condf = dir->dirfrag(); | |
6842 | } else { | |
6843 | condf = con->dirfrag(); | |
6844 | } | |
6845 | ||
6846 | for (int p=0; p<2; p++) { | |
6847 | mds_rank_t a = auth.first; | |
6848 | if (p) a = auth.second; | |
6849 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6850 | if (mds->get_nodeid() == auth.second && | |
6851 | con->is_importing()) break; // don't send any expire while importing. | |
6852 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6853 | ||
6854 | dout(12) << " sending expire to mds." << a << " on " << *dir << dendl; | |
11fdf7f2 TL |
6855 | ceph_assert(a != mds->get_nodeid()); |
6856 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(a), std::forward_as_tuple()); | |
6857 | if (em.second) | |
6858 | em.first->second = MCacheExpire::create(mds->get_nodeid()); /* new */ | |
6859 | em.first->second->add_dir(condf, dir->dirfrag(), dir->replica_nonce); | |
7c673cae FG |
6860 | } |
6861 | } | |
6862 | ||
6863 | in->close_dirfrag(dir->dirfrag().frag); | |
6864 | } | |
6865 | ||
6866 | /** | |
6867 | * Try trimming an inode from the cache | |
6868 | * | |
6869 | * @return true if the inode is still in cache, else false if it was trimmed | |
6870 | */ | |
11fdf7f2 | 6871 | bool MDCache::trim_inode(CDentry *dn, CInode *in, CDir *con, expiremap& expiremap) |
7c673cae FG |
6872 | { |
6873 | dout(15) << "trim_inode " << *in << dendl; | |
11fdf7f2 | 6874 | ceph_assert(in->get_num_ref() == 0); |
7c673cae FG |
6875 | |
6876 | if (in->is_dir()) { | |
6877 | // If replica inode's dirfragtreelock is not readable, it's likely | |
6878 | // some dirfrags of the inode are being fragmented and we will receive | |
6879 | // MMDSFragmentNotify soon. MMDSFragmentNotify only replicates the new | |
6880 | // dirfrags, so we should avoid trimming these dirfrags' parent inode. | |
6881 | // This is because that unconnected replicas are problematic for | |
6882 | // subtree migration. | |
6883 | // | |
28e407b8 | 6884 | if (!in->is_auth() && !mds->locker->rdlock_try(&in->dirfragtreelock, -1, nullptr)) { |
7c673cae | 6885 | return true; |
28e407b8 | 6886 | } |
7c673cae FG |
6887 | |
6888 | // DIR | |
6889 | list<CDir*> dfls; | |
6890 | in->get_dirfrags(dfls); | |
6891 | for (list<CDir*>::iterator p = dfls.begin(); p != dfls.end(); ++p) { | |
6892 | CDir *dir = *p; | |
11fdf7f2 | 6893 | ceph_assert(!dir->is_subtree_root()); |
7c673cae FG |
6894 | trim_dirfrag(dir, con ? con:dir, expiremap); // if no container (e.g. root dirfrag), use *p |
6895 | } | |
6896 | } | |
6897 | ||
6898 | // INODE | |
6899 | if (in->is_auth()) { | |
6900 | // eval stray after closing dirfrags | |
6901 | if (dn && !dn->state_test(CDentry::STATE_PURGING)) { | |
6902 | maybe_eval_stray(in); | |
6903 | if (dn->state_test(CDentry::STATE_PURGING) || dn->get_num_ref() > 0) | |
6904 | return true; | |
6905 | } | |
6906 | } else { | |
6907 | mds_authority_t auth = in->authority(); | |
6908 | ||
6909 | dirfrag_t df; | |
6910 | if (con) | |
6911 | df = con->dirfrag(); | |
6912 | else | |
6913 | df = dirfrag_t(0,frag_t()); // must be a root or stray inode. | |
6914 | ||
6915 | for (int p=0; p<2; p++) { | |
6916 | mds_rank_t a = auth.first; | |
6917 | if (p) a = auth.second; | |
6918 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6919 | if (con && mds->get_nodeid() == auth.second && | |
6920 | con->is_importing()) break; // don't send any expire while importing. | |
6921 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6922 | ||
6923 | dout(12) << " sending expire to mds." << a << " on " << *in << dendl; | |
11fdf7f2 TL |
6924 | ceph_assert(a != mds->get_nodeid()); |
6925 | auto em = expiremap.emplace(std::piecewise_construct, std::forward_as_tuple(a), std::forward_as_tuple()); | |
6926 | if (em.second) | |
6927 | em.first->second = MCacheExpire::create(mds->get_nodeid()); /* new */ | |
6928 | em.first->second->add_inode(df, in->vino(), in->get_replica_nonce()); | |
7c673cae FG |
6929 | } |
6930 | } | |
6931 | ||
6932 | /* | |
6933 | if (in->is_auth()) { | |
6934 | if (in->hack_accessed) | |
6935 | mds->logger->inc("outt"); | |
6936 | else { | |
6937 | mds->logger->inc("outut"); | |
6938 | mds->logger->fset("oututl", ceph_clock_now() - in->hack_load_stamp); | |
6939 | } | |
6940 | } | |
6941 | */ | |
6942 | ||
6943 | // unlink | |
6944 | if (dn) | |
31f18b77 | 6945 | dn->get_dir()->unlink_inode(dn, false); |
7c673cae FG |
6946 | remove_inode(in); |
6947 | return false; | |
6948 | } | |
6949 | ||
6950 | ||
6951 | /** | |
6952 | * trim_non_auth - remove any non-auth items from our cache | |
6953 | * | |
6954 | * this reduces the amount of non-auth metadata in our cache, reducing the | |
6955 | * load incurred by the rejoin phase. | |
6956 | * | |
6957 | * the only non-auth items that remain are those that are needed to | |
6958 | * attach our own subtrees to the root. | |
6959 | * | |
6960 | * when we are done, all dentries will be in the top bit of the lru. | |
6961 | * | |
6962 | * why we have to do this: | |
6963 | * we may not have accurate linkage for non-auth items. which means we will | |
6964 | * know which subtree it falls into, and can not be sure to declare it to the | |
6965 | * correct authority. | |
6966 | */ | |
6967 | void MDCache::trim_non_auth() | |
6968 | { | |
6969 | dout(7) << "trim_non_auth" << dendl; | |
6970 | ||
6971 | // temporarily pin all subtree roots | |
6972 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
6973 | p != subtrees.end(); | |
6974 | ++p) | |
6975 | p->first->get(CDir::PIN_SUBTREETEMP); | |
6976 | ||
31f18b77 | 6977 | list<CDentry*> auth_list; |
7c673cae FG |
6978 | |
6979 | // trim non-auth items from the lru | |
31f18b77 FG |
6980 | for (;;) { |
6981 | CDentry *dn = NULL; | |
6982 | if (bottom_lru.lru_get_size() > 0) | |
6983 | dn = static_cast<CDentry*>(bottom_lru.lru_expire()); | |
6984 | if (!dn && lru.lru_get_size() > 0) | |
6985 | dn = static_cast<CDentry*>(lru.lru_expire()); | |
6986 | if (!dn) | |
6987 | break; | |
6988 | ||
7c673cae FG |
6989 | CDentry::linkage_t *dnl = dn->get_linkage(); |
6990 | ||
6991 | if (dn->is_auth()) { | |
6992 | // add back into lru (at the top) | |
31f18b77 | 6993 | auth_list.push_back(dn); |
7c673cae FG |
6994 | |
6995 | if (dnl->is_remote() && dnl->get_inode() && !dnl->get_inode()->is_auth()) | |
6996 | dn->unlink_remote(dnl); | |
7c673cae FG |
6997 | } else { |
6998 | // non-auth. expire. | |
6999 | CDir *dir = dn->get_dir(); | |
11fdf7f2 | 7000 | ceph_assert(dir); |
7c673cae FG |
7001 | |
7002 | // unlink the dentry | |
7003 | dout(10) << " removing " << *dn << dendl; | |
7004 | if (dnl->is_remote()) { | |
31f18b77 | 7005 | dir->unlink_inode(dn, false); |
7c673cae FG |
7006 | } |
7007 | else if (dnl->is_primary()) { | |
7008 | CInode *in = dnl->get_inode(); | |
7009 | dout(10) << " removing " << *in << dendl; | |
7010 | list<CDir*> ls; | |
7011 | in->get_dirfrags(ls); | |
7012 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
7013 | CDir *subdir = *p; | |
11fdf7f2 | 7014 | ceph_assert(!subdir->is_subtree_root()); |
7c673cae FG |
7015 | in->close_dirfrag(subdir->dirfrag().frag); |
7016 | } | |
31f18b77 | 7017 | dir->unlink_inode(dn, false); |
7c673cae FG |
7018 | remove_inode(in); |
7019 | } | |
7020 | else { | |
11fdf7f2 | 7021 | ceph_assert(dnl->is_null()); |
7c673cae FG |
7022 | } |
7023 | ||
11fdf7f2 | 7024 | ceph_assert(!dir->has_bloom()); |
7c673cae FG |
7025 | dir->remove_dentry(dn); |
7026 | // adjust the dir state | |
7027 | dir->state_clear(CDir::STATE_COMPLETE); // dir incomplete! | |
7028 | // close empty non-auth dirfrag | |
7029 | if (!dir->is_subtree_root() && dir->get_num_any() == 0) | |
7030 | dir->inode->close_dirfrag(dir->get_frag()); | |
7031 | } | |
7032 | } | |
7033 | ||
31f18b77 FG |
7034 | for (auto dn : auth_list) { |
7035 | if (dn->state_test(CDentry::STATE_BOTTOMLRU)) | |
7036 | bottom_lru.lru_insert_mid(dn); | |
7037 | else | |
7038 | lru.lru_insert_top(dn); | |
7039 | } | |
7040 | ||
7c673cae FG |
7041 | // move everything in the pintail to the top bit of the lru. |
7042 | lru.lru_touch_entire_pintail(); | |
7043 | ||
7044 | // unpin all subtrees | |
7045 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
7046 | p != subtrees.end(); | |
7047 | ++p) | |
7048 | p->first->put(CDir::PIN_SUBTREETEMP); | |
7049 | ||
31f18b77 FG |
7050 | if (lru.lru_get_size() == 0 && |
7051 | bottom_lru.lru_get_size() == 0) { | |
7c673cae | 7052 | // root, stray, etc.? |
b32b8144 | 7053 | auto p = inode_map.begin(); |
7c673cae | 7054 | while (p != inode_map.end()) { |
7c673cae | 7055 | CInode *in = p->second; |
b32b8144 | 7056 | ++p; |
7c673cae FG |
7057 | if (!in->is_auth()) { |
7058 | list<CDir*> ls; | |
7059 | in->get_dirfrags(ls); | |
7060 | for (list<CDir*>::iterator p = ls.begin(); | |
7061 | p != ls.end(); | |
7062 | ++p) { | |
7063 | dout(10) << " removing " << **p << dendl; | |
11fdf7f2 | 7064 | ceph_assert((*p)->get_num_ref() == 1); // SUBTREE |
7c673cae FG |
7065 | remove_subtree((*p)); |
7066 | in->close_dirfrag((*p)->dirfrag().frag); | |
7067 | } | |
7068 | dout(10) << " removing " << *in << dendl; | |
11fdf7f2 TL |
7069 | ceph_assert(!in->get_parent_dn()); |
7070 | ceph_assert(in->get_num_ref() == 0); | |
7c673cae FG |
7071 | remove_inode(in); |
7072 | } | |
7c673cae FG |
7073 | } |
7074 | } | |
7075 | ||
7076 | show_subtrees(); | |
7077 | } | |
7078 | ||
7079 | /** | |
7080 | * Recursively trim the subtree rooted at directory to remove all | |
7081 | * CInodes/CDentrys/CDirs that aren't links to remote MDSes, or ancestors | |
7082 | * of those links. This is used to clear invalid data out of the cache. | |
7083 | * Note that it doesn't clear the passed-in directory, since that's not | |
7084 | * always safe. | |
7085 | */ | |
7086 | bool MDCache::trim_non_auth_subtree(CDir *dir) | |
7087 | { | |
7088 | dout(10) << "trim_non_auth_subtree(" << dir << ") " << *dir << dendl; | |
7089 | ||
7090 | bool keep_dir = !can_trim_non_auth_dirfrag(dir); | |
7091 | ||
94b18763 FG |
7092 | auto j = dir->begin(); |
7093 | auto i = j; | |
7c673cae FG |
7094 | while (j != dir->end()) { |
7095 | i = j++; | |
7096 | CDentry *dn = i->second; | |
7097 | dout(10) << "trim_non_auth_subtree(" << dir << ") Checking dentry " << dn << dendl; | |
7098 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
7099 | if (dnl->is_primary()) { // check for subdirectories, etc | |
7100 | CInode *in = dnl->get_inode(); | |
7101 | bool keep_inode = false; | |
7102 | if (in->is_dir()) { | |
7103 | list<CDir*> subdirs; | |
7104 | in->get_dirfrags(subdirs); | |
7105 | for (list<CDir*>::iterator subdir = subdirs.begin(); | |
7106 | subdir != subdirs.end(); | |
7107 | ++subdir) { | |
7108 | if ((*subdir)->is_subtree_root()) { | |
7109 | keep_inode = true; | |
7110 | dout(10) << "trim_non_auth_subtree(" << dir << ") keeping " << **subdir << dendl; | |
7111 | } else { | |
7112 | if (trim_non_auth_subtree(*subdir)) | |
7113 | keep_inode = true; | |
7114 | else { | |
7115 | in->close_dirfrag((*subdir)->get_frag()); | |
7116 | dir->state_clear(CDir::STATE_COMPLETE); // now incomplete! | |
7117 | } | |
7118 | } | |
7119 | } | |
7120 | ||
7121 | } | |
7122 | if (!keep_inode) { // remove it! | |
7123 | dout(20) << "trim_non_auth_subtree(" << dir << ") removing inode " << in << " with dentry" << dn << dendl; | |
31f18b77 | 7124 | dir->unlink_inode(dn, false); |
7c673cae | 7125 | remove_inode(in); |
11fdf7f2 | 7126 | ceph_assert(!dir->has_bloom()); |
7c673cae FG |
7127 | dir->remove_dentry(dn); |
7128 | } else { | |
7129 | dout(20) << "trim_non_auth_subtree(" << dir << ") keeping inode " << in << " with dentry " << dn <<dendl; | |
7130 | dn->state_clear(CDentry::STATE_AUTH); | |
7131 | in->state_clear(CInode::STATE_AUTH); | |
7132 | } | |
7133 | } else if (keep_dir && dnl->is_null()) { // keep null dentry for slave rollback | |
7134 | dout(20) << "trim_non_auth_subtree(" << dir << ") keeping dentry " << dn <<dendl; | |
7135 | } else { // just remove it | |
7136 | dout(20) << "trim_non_auth_subtree(" << dir << ") removing dentry " << dn << dendl; | |
7137 | if (dnl->is_remote()) | |
31f18b77 | 7138 | dir->unlink_inode(dn, false); |
7c673cae FG |
7139 | dir->remove_dentry(dn); |
7140 | } | |
7141 | } | |
7142 | dir->state_clear(CDir::STATE_AUTH); | |
7143 | /** | |
7144 | * We've now checked all our children and deleted those that need it. | |
7145 | * Now return to caller, and tell them if *we're* a keeper. | |
7146 | */ | |
7147 | return keep_dir || dir->get_num_any(); | |
7148 | } | |
7149 | ||
7150 | /* | |
7151 | * during replay, when we determine a subtree is no longer ours, we | |
7152 | * try to trim it from our cache. because subtrees must be connected | |
7153 | * to the root, the fact that we can trim this tree may mean that our | |
7154 | * children or parents can also be trimmed. | |
7155 | */ | |
7156 | void MDCache::try_trim_non_auth_subtree(CDir *dir) | |
7157 | { | |
7158 | dout(10) << "try_trim_nonauth_subtree " << *dir << dendl; | |
7159 | ||
7160 | // can we now trim child subtrees? | |
7161 | set<CDir*> bounds; | |
7162 | get_subtree_bounds(dir, bounds); | |
7163 | for (set<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) { | |
7164 | CDir *bd = *p; | |
7165 | if (bd->get_dir_auth().first != mds->get_nodeid() && // we are not auth | |
7166 | bd->get_num_any() == 0 && // and empty | |
7167 | can_trim_non_auth_dirfrag(bd)) { | |
7168 | CInode *bi = bd->get_inode(); | |
7169 | dout(10) << " closing empty non-auth child subtree " << *bd << dendl; | |
7170 | remove_subtree(bd); | |
7171 | bd->mark_clean(); | |
7172 | bi->close_dirfrag(bd->get_frag()); | |
7173 | } | |
7174 | } | |
7175 | ||
7176 | if (trim_non_auth_subtree(dir)) { | |
7177 | // keep | |
7178 | try_subtree_merge(dir); | |
7179 | } else { | |
7180 | // can we trim this subtree (and possibly our ancestors) too? | |
7181 | while (true) { | |
7182 | CInode *diri = dir->get_inode(); | |
7183 | if (diri->is_base()) { | |
7184 | if (!diri->is_root() && diri->authority().first != mds->get_nodeid()) { | |
7185 | dout(10) << " closing empty non-auth subtree " << *dir << dendl; | |
7186 | remove_subtree(dir); | |
7187 | dir->mark_clean(); | |
7188 | diri->close_dirfrag(dir->get_frag()); | |
7189 | ||
7190 | dout(10) << " removing " << *diri << dendl; | |
11fdf7f2 TL |
7191 | ceph_assert(!diri->get_parent_dn()); |
7192 | ceph_assert(diri->get_num_ref() == 0); | |
7c673cae FG |
7193 | remove_inode(diri); |
7194 | } | |
7195 | break; | |
7196 | } | |
7197 | ||
7198 | CDir *psub = get_subtree_root(diri->get_parent_dir()); | |
7199 | dout(10) << " parent subtree is " << *psub << dendl; | |
7200 | if (psub->get_dir_auth().first == mds->get_nodeid()) | |
7201 | break; // we are auth, keep. | |
7202 | ||
7203 | dout(10) << " closing empty non-auth subtree " << *dir << dendl; | |
7204 | remove_subtree(dir); | |
7205 | dir->mark_clean(); | |
7206 | diri->close_dirfrag(dir->get_frag()); | |
7207 | ||
7208 | dout(10) << " parent subtree also non-auth: " << *psub << dendl; | |
7209 | if (trim_non_auth_subtree(psub)) | |
7210 | break; | |
7211 | dir = psub; | |
7212 | } | |
7213 | } | |
7214 | ||
7215 | show_subtrees(); | |
7216 | } | |
7217 | ||
7218 | void MDCache::standby_trim_segment(LogSegment *ls) | |
7219 | { | |
7220 | ls->new_dirfrags.clear_list(); | |
7221 | ls->open_files.clear_list(); | |
7222 | ||
7223 | while (!ls->dirty_dirfrags.empty()) { | |
7224 | CDir *dir = ls->dirty_dirfrags.front(); | |
7225 | dir->mark_clean(); | |
7226 | } | |
7227 | while (!ls->dirty_inodes.empty()) { | |
7228 | CInode *in = ls->dirty_inodes.front(); | |
7229 | in->mark_clean(); | |
7230 | } | |
7231 | while (!ls->dirty_dentries.empty()) { | |
7232 | CDentry *dn = ls->dirty_dentries.front(); | |
7233 | dn->mark_clean(); | |
7234 | } | |
7235 | while (!ls->dirty_parent_inodes.empty()) { | |
7236 | CInode *in = ls->dirty_parent_inodes.front(); | |
7237 | in->clear_dirty_parent(); | |
7238 | } | |
7239 | while (!ls->dirty_dirfrag_dir.empty()) { | |
7240 | CInode *in = ls->dirty_dirfrag_dir.front(); | |
7241 | in->filelock.remove_dirty(); | |
7242 | } | |
7243 | while (!ls->dirty_dirfrag_nest.empty()) { | |
7244 | CInode *in = ls->dirty_dirfrag_nest.front(); | |
7245 | in->nestlock.remove_dirty(); | |
7246 | } | |
7247 | while (!ls->dirty_dirfrag_dirfragtree.empty()) { | |
7248 | CInode *in = ls->dirty_dirfrag_dirfragtree.front(); | |
7249 | in->dirfragtreelock.remove_dirty(); | |
7250 | } | |
7251 | } | |
7252 | ||
11fdf7f2 | 7253 | void MDCache::handle_cache_expire(const MCacheExpire::const_ref &m) |
7c673cae FG |
7254 | { |
7255 | mds_rank_t from = mds_rank_t(m->get_from()); | |
7256 | ||
7257 | dout(7) << "cache_expire from mds." << from << dendl; | |
7258 | ||
7259 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
7c673cae FG |
7260 | return; |
7261 | } | |
7262 | ||
7263 | set<SimpleLock *> gather_locks; | |
7264 | // loop over realms | |
11fdf7f2 | 7265 | for (const auto &p : m->realms) { |
7c673cae | 7266 | // check container? |
11fdf7f2 TL |
7267 | if (p.first.ino > 0) { |
7268 | CInode *expired_inode = get_inode(p.first.ino); | |
7269 | ceph_assert(expired_inode); // we had better have this. | |
7270 | CDir *parent_dir = expired_inode->get_approx_dirfrag(p.first.frag); | |
7271 | ceph_assert(parent_dir); | |
7c673cae FG |
7272 | |
7273 | int export_state = -1; | |
7274 | if (parent_dir->is_auth() && parent_dir->is_exporting()) { | |
7275 | export_state = migrator->get_export_state(parent_dir); | |
11fdf7f2 | 7276 | ceph_assert(export_state >= 0); |
7c673cae FG |
7277 | } |
7278 | ||
7279 | if (!parent_dir->is_auth() || | |
7280 | (export_state != -1 && | |
7281 | ((export_state == Migrator::EXPORT_WARNING && | |
7282 | migrator->export_has_warned(parent_dir,from)) || | |
7283 | export_state == Migrator::EXPORT_EXPORTING || | |
7284 | export_state == Migrator::EXPORT_LOGGINGFINISH || | |
7285 | (export_state == Migrator::EXPORT_NOTIFYING && | |
7286 | !migrator->export_has_notified(parent_dir,from))))) { | |
7287 | ||
7288 | // not auth. | |
7289 | dout(7) << "delaying nonauth|warned expires for " << *parent_dir << dendl; | |
11fdf7f2 | 7290 | ceph_assert(parent_dir->is_frozen_tree_root()); |
7c673cae FG |
7291 | |
7292 | // make a message container | |
11fdf7f2 TL |
7293 | |
7294 | auto em = delayed_expire[parent_dir].emplace(std::piecewise_construct, std::forward_as_tuple(from), std::forward_as_tuple()); | |
7295 | if (em.second) | |
7296 | em.first->second = MCacheExpire::create(from); /* new */ | |
7297 | ||
7c673cae | 7298 | // merge these expires into it |
11fdf7f2 | 7299 | em.first->second->add_realm(p.first, p.second); |
7c673cae FG |
7300 | continue; |
7301 | } | |
11fdf7f2 | 7302 | ceph_assert(export_state <= Migrator::EXPORT_PREPPING || |
7c673cae FG |
7303 | (export_state == Migrator::EXPORT_WARNING && |
7304 | !migrator->export_has_warned(parent_dir, from))); | |
7305 | ||
7306 | dout(7) << "expires for " << *parent_dir << dendl; | |
7307 | } else { | |
7308 | dout(7) << "containerless expires (root, stray inodes)" << dendl; | |
7309 | } | |
7310 | ||
7311 | // INODES | |
11fdf7f2 TL |
7312 | for (const auto &q : p.second.inodes) { |
7313 | CInode *in = get_inode(q.first); | |
7314 | unsigned nonce = q.second; | |
7c673cae FG |
7315 | |
7316 | if (!in) { | |
11fdf7f2 | 7317 | dout(0) << " inode expire on " << q.first << " from " << from |
7c673cae | 7318 | << ", don't have it" << dendl; |
11fdf7f2 | 7319 | ceph_assert(in); |
7c673cae | 7320 | } |
11fdf7f2 | 7321 | ceph_assert(in->is_auth()); |
7c673cae FG |
7322 | dout(20) << __func__ << ": expiring inode " << *in << dendl; |
7323 | ||
7324 | // check nonce | |
7325 | if (nonce == in->get_replica_nonce(from)) { | |
7326 | // remove from our cached_by | |
7327 | dout(7) << " inode expire on " << *in << " from mds." << from | |
7328 | << " cached_by was " << in->get_replicas() << dendl; | |
7329 | inode_remove_replica(in, from, false, gather_locks); | |
7330 | } | |
7331 | else { | |
7332 | // this is an old nonce, ignore expire. | |
7333 | dout(7) << " inode expire on " << *in << " from mds." << from | |
7334 | << " with old nonce " << nonce | |
7335 | << " (current " << in->get_replica_nonce(from) << "), dropping" | |
7336 | << dendl; | |
7337 | } | |
7338 | } | |
7339 | ||
7340 | // DIRS | |
11fdf7f2 TL |
7341 | for (const auto &q : p.second.dirs) { |
7342 | CDir *dir = get_dirfrag(q.first); | |
7343 | unsigned nonce = q.second; | |
7c673cae FG |
7344 | |
7345 | if (!dir) { | |
11fdf7f2 | 7346 | CInode *diri = get_inode(q.first.ino); |
7c673cae FG |
7347 | if (diri) { |
7348 | if (mds->is_rejoin() && | |
7349 | rejoin_ack_gather.count(mds->get_nodeid()) && // haven't sent rejoin ack yet | |
7350 | !diri->is_replica(from)) { | |
7351 | list<CDir*> ls; | |
7352 | diri->get_nested_dirfrags(ls); | |
11fdf7f2 | 7353 | dout(7) << " dir expire on dirfrag " << q.first << " from mds." << from |
7c673cae FG |
7354 | << " while rejoining, inode isn't replicated" << dendl; |
7355 | for (list<CDir*>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
7356 | dir = *q; | |
7357 | if (dir->is_replica(from)) { | |
7358 | dout(7) << " dir expire on " << *dir << " from mds." << from << dendl; | |
7359 | dir->remove_replica(from); | |
7360 | } | |
7361 | } | |
7362 | continue; | |
7363 | } | |
11fdf7f2 | 7364 | CDir *other = diri->get_approx_dirfrag(q.first.frag); |
7c673cae | 7365 | if (other) { |
11fdf7f2 | 7366 | dout(7) << " dir expire on dirfrag " << q.first << " from mds." << from |
7c673cae FG |
7367 | << " have " << *other << ", mismatched frags, dropping" << dendl; |
7368 | continue; | |
7369 | } | |
7370 | } | |
11fdf7f2 | 7371 | dout(0) << " dir expire on " << q.first << " from " << from |
7c673cae | 7372 | << ", don't have it" << dendl; |
11fdf7f2 | 7373 | ceph_assert(dir); |
7c673cae FG |
7374 | } |
7375 | dout(20) << __func__ << ": expiring dirfrag " << *dir << dendl; | |
7376 | ||
11fdf7f2 | 7377 | ceph_assert(dir->is_auth()); |
7c673cae FG |
7378 | |
7379 | // check nonce | |
7380 | if (nonce == dir->get_replica_nonce(from)) { | |
7381 | // remove from our cached_by | |
7382 | dout(7) << " dir expire on " << *dir << " from mds." << from | |
181888fb | 7383 | << " replicas was " << dir->get_replicas() << dendl; |
7c673cae FG |
7384 | dir->remove_replica(from); |
7385 | } | |
7386 | else { | |
7387 | // this is an old nonce, ignore expire. | |
7388 | dout(7) << " dir expire on " << *dir << " from mds." << from | |
7389 | << " with old nonce " << nonce << " (current " << dir->get_replica_nonce(from) | |
7390 | << "), dropping" << dendl; | |
7391 | } | |
7392 | } | |
7393 | ||
7394 | // DENTRIES | |
11fdf7f2 TL |
7395 | for (const auto &pd : p.second.dentries) { |
7396 | dout(10) << " dn expires in dir " << pd.first << dendl; | |
7397 | CInode *diri = get_inode(pd.first.ino); | |
7398 | ceph_assert(diri); | |
7399 | CDir *dir = diri->get_dirfrag(pd.first.frag); | |
7c673cae FG |
7400 | |
7401 | if (!dir) { | |
11fdf7f2 | 7402 | dout(0) << " dn expires on " << pd.first << " from " << from |
7c673cae FG |
7403 | << ", must have refragmented" << dendl; |
7404 | } else { | |
11fdf7f2 | 7405 | ceph_assert(dir->is_auth()); |
7c673cae FG |
7406 | } |
7407 | ||
11fdf7f2 TL |
7408 | for (const auto &p : pd.second) { |
7409 | unsigned nonce = p.second; | |
7c673cae FG |
7410 | CDentry *dn; |
7411 | ||
7412 | if (dir) { | |
11fdf7f2 | 7413 | dn = dir->lookup(p.first.first, p.first.second); |
7c673cae FG |
7414 | } else { |
7415 | // which dirfrag for this dentry? | |
11fdf7f2 TL |
7416 | CDir *dir = diri->get_dirfrag(diri->pick_dirfrag(p.first.first)); |
7417 | ceph_assert(dir); | |
7418 | ceph_assert(dir->is_auth()); | |
7419 | dn = dir->lookup(p.first.first, p.first.second); | |
7c673cae FG |
7420 | } |
7421 | ||
7422 | if (!dn) { | |
7423 | if (dir) | |
11fdf7f2 | 7424 | dout(0) << " missing dentry for " << p.first.first << " snap " << p.first.second << " in " << *dir << dendl; |
7c673cae | 7425 | else |
11fdf7f2 | 7426 | dout(0) << " missing dentry for " << p.first.first << " snap " << p.first.second << dendl; |
7c673cae | 7427 | } |
11fdf7f2 | 7428 | ceph_assert(dn); |
7c673cae FG |
7429 | |
7430 | if (nonce == dn->get_replica_nonce(from)) { | |
7431 | dout(7) << " dentry_expire on " << *dn << " from mds." << from << dendl; | |
7432 | dentry_remove_replica(dn, from, gather_locks); | |
7433 | } | |
7434 | else { | |
7435 | dout(7) << " dentry_expire on " << *dn << " from mds." << from | |
7436 | << " with old nonce " << nonce << " (current " << dn->get_replica_nonce(from) | |
7437 | << "), dropping" << dendl; | |
7438 | } | |
7439 | } | |
7440 | } | |
7441 | } | |
7442 | ||
7c673cae FG |
7443 | for (set<SimpleLock*>::iterator p = gather_locks.begin(); p != gather_locks.end(); ++p) { |
7444 | if (!(*p)->is_stable()) | |
7445 | mds->locker->eval_gather(*p); | |
7446 | } | |
7447 | } | |
7448 | ||
7449 | void MDCache::process_delayed_expire(CDir *dir) | |
7450 | { | |
7451 | dout(7) << "process_delayed_expire on " << *dir << dendl; | |
11fdf7f2 TL |
7452 | for (const auto &p : delayed_expire[dir]) { |
7453 | handle_cache_expire(p.second); | |
7454 | } | |
7c673cae FG |
7455 | delayed_expire.erase(dir); |
7456 | } | |
7457 | ||
7458 | void MDCache::discard_delayed_expire(CDir *dir) | |
7459 | { | |
7460 | dout(7) << "discard_delayed_expire on " << *dir << dendl; | |
7c673cae FG |
7461 | delayed_expire.erase(dir); |
7462 | } | |
7463 | ||
7464 | void MDCache::inode_remove_replica(CInode *in, mds_rank_t from, bool rejoin, | |
7465 | set<SimpleLock *>& gather_locks) | |
7466 | { | |
7467 | in->remove_replica(from); | |
11fdf7f2 | 7468 | in->set_mds_caps_wanted(from, 0); |
7c673cae FG |
7469 | |
7470 | // note: this code calls _eval more often than it needs to! | |
7471 | // fix lock | |
7472 | if (in->authlock.remove_replica(from)) gather_locks.insert(&in->authlock); | |
7473 | if (in->linklock.remove_replica(from)) gather_locks.insert(&in->linklock); | |
7474 | if (in->snaplock.remove_replica(from)) gather_locks.insert(&in->snaplock); | |
7475 | if (in->xattrlock.remove_replica(from)) gather_locks.insert(&in->xattrlock); | |
7476 | if (in->flocklock.remove_replica(from)) gather_locks.insert(&in->flocklock); | |
7477 | if (in->policylock.remove_replica(from)) gather_locks.insert(&in->policylock); | |
7478 | ||
7479 | // If 'rejoin' is true and the scatter lock is in LOCK_MIX_* state. | |
7480 | // Don't remove the recovering mds from lock's gathering list because | |
7481 | // it may hold rejoined wrlocks. | |
7482 | if (in->dirfragtreelock.remove_replica(from, rejoin)) gather_locks.insert(&in->dirfragtreelock); | |
7483 | if (in->filelock.remove_replica(from, rejoin)) gather_locks.insert(&in->filelock); | |
7484 | if (in->nestlock.remove_replica(from, rejoin)) gather_locks.insert(&in->nestlock); | |
7485 | } | |
7486 | ||
7487 | void MDCache::dentry_remove_replica(CDentry *dn, mds_rank_t from, set<SimpleLock *>& gather_locks) | |
7488 | { | |
7489 | dn->remove_replica(from); | |
7490 | ||
7491 | // fix lock | |
7492 | if (dn->lock.remove_replica(from)) | |
7493 | gather_locks.insert(&dn->lock); | |
7494 | ||
7495 | // Replicated strays might now be elegible for purge | |
11fdf7f2 | 7496 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); |
7c673cae FG |
7497 | if (dnl->is_primary()) { |
7498 | maybe_eval_stray(dnl->get_inode()); | |
7499 | } | |
7500 | } | |
7501 | ||
7502 | void MDCache::trim_client_leases() | |
7503 | { | |
7504 | utime_t now = ceph_clock_now(); | |
7505 | ||
7506 | dout(10) << "trim_client_leases" << dendl; | |
7507 | ||
7508 | for (int pool=0; pool<client_lease_pools; pool++) { | |
7509 | int before = client_leases[pool].size(); | |
7510 | if (client_leases[pool].empty()) | |
7511 | continue; | |
7512 | ||
7513 | while (!client_leases[pool].empty()) { | |
7514 | ClientLease *r = client_leases[pool].front(); | |
7515 | if (r->ttl > now) break; | |
7516 | CDentry *dn = static_cast<CDentry*>(r->parent); | |
7517 | dout(10) << " expiring client." << r->client << " lease of " << *dn << dendl; | |
7518 | dn->remove_client_lease(r, mds->locker); | |
7519 | } | |
7520 | int after = client_leases[pool].size(); | |
7521 | dout(10) << "trim_client_leases pool " << pool << " trimmed " | |
7522 | << (before-after) << " leases, " << after << " left" << dendl; | |
7523 | } | |
7524 | } | |
7525 | ||
7526 | ||
7527 | void MDCache::check_memory_usage() | |
7528 | { | |
7529 | static MemoryModel mm(g_ceph_context); | |
7530 | static MemoryModel::snap last; | |
7531 | mm.sample(&last); | |
7532 | static MemoryModel::snap baseline = last; | |
7533 | ||
7534 | // check client caps | |
11fdf7f2 | 7535 | ceph_assert(CInode::count() == inode_map.size() + snap_inode_map.size() + num_shadow_inodes); |
181888fb | 7536 | double caps_per_inode = 0.0; |
7c673cae | 7537 | if (CInode::count()) |
181888fb | 7538 | caps_per_inode = (double)Capability::count() / (double)CInode::count(); |
7c673cae | 7539 | |
a8e16298 | 7540 | dout(2) << "Memory usage: " |
7c673cae FG |
7541 | << " total " << last.get_total() |
7542 | << ", rss " << last.get_rss() | |
7543 | << ", heap " << last.get_heap() | |
7544 | << ", baseline " << baseline.get_heap() | |
7c673cae FG |
7545 | << ", " << num_inodes_with_caps << " / " << CInode::count() << " inodes have caps" |
7546 | << ", " << Capability::count() << " caps, " << caps_per_inode << " caps per inode" | |
7547 | << dendl; | |
7548 | ||
c07f9fc5 | 7549 | mds->update_mlogger(); |
7c673cae FG |
7550 | mds->mlogger->set(l_mdm_rss, last.get_rss()); |
7551 | mds->mlogger->set(l_mdm_heap, last.get_heap()); | |
7552 | ||
181888fb | 7553 | if (cache_toofull()) { |
a8e16298 | 7554 | mds->server->recall_client_state(nullptr); |
7c673cae FG |
7555 | } |
7556 | ||
7557 | // If the cache size had exceeded its limit, but we're back in bounds | |
7558 | // now, free any unused pool memory so that our memory usage isn't | |
7559 | // permanently bloated. | |
181888fb | 7560 | if (exceeded_size_limit && !cache_toofull()) { |
7c673cae FG |
7561 | // Only do this once we are back in bounds: otherwise the releases would |
7562 | // slow down whatever process caused us to exceed bounds to begin with | |
7563 | if (ceph_using_tcmalloc()) { | |
a8e16298 | 7564 | dout(5) << "check_memory_usage: releasing unused space from tcmalloc" |
7c673cae FG |
7565 | << dendl; |
7566 | ceph_heap_release_free_memory(); | |
7567 | } | |
7568 | exceeded_size_limit = false; | |
7569 | } | |
7570 | } | |
7571 | ||
7572 | ||
7573 | ||
7574 | // ========================================================================================= | |
7575 | // shutdown | |
7576 | ||
7577 | class C_MDC_ShutdownCheck : public MDCacheContext { | |
7578 | public: | |
7579 | explicit C_MDC_ShutdownCheck(MDCache *m) : MDCacheContext(m) {} | |
7580 | void finish(int) override { | |
7581 | mdcache->shutdown_check(); | |
7582 | } | |
7583 | }; | |
7584 | ||
7585 | void MDCache::shutdown_check() | |
7586 | { | |
7587 | dout(0) << "shutdown_check at " << ceph_clock_now() << dendl; | |
7588 | ||
7589 | // cache | |
7590 | char old_val[32] = { 0 }; | |
7591 | char *o = old_val; | |
11fdf7f2 TL |
7592 | g_conf().get_val("debug_mds", &o, sizeof(old_val)); |
7593 | g_conf().set_val("debug_mds", "10"); | |
7594 | g_conf().apply_changes(nullptr); | |
7c673cae | 7595 | show_cache(); |
11fdf7f2 TL |
7596 | g_conf().set_val("debug_mds", old_val); |
7597 | g_conf().apply_changes(nullptr); | |
7598 | mds->timer.add_event_after(g_conf()->mds_shutdown_check, new C_MDC_ShutdownCheck(this)); | |
7c673cae FG |
7599 | |
7600 | // this | |
31f18b77 | 7601 | dout(0) << "lru size now " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; |
7c673cae FG |
7602 | dout(0) << "log len " << mds->mdlog->get_num_events() << dendl; |
7603 | ||
7604 | ||
7605 | if (mds->objecter->is_active()) { | |
7606 | dout(0) << "objecter still active" << dendl; | |
7607 | mds->objecter->dump_active(); | |
7608 | } | |
7609 | } | |
7610 | ||
7611 | ||
7612 | void MDCache::shutdown_start() | |
7613 | { | |
a8e16298 | 7614 | dout(5) << "shutdown_start" << dendl; |
7c673cae | 7615 | |
11fdf7f2 TL |
7616 | if (g_conf()->mds_shutdown_check) |
7617 | mds->timer.add_event_after(g_conf()->mds_shutdown_check, new C_MDC_ShutdownCheck(this)); | |
7c673cae | 7618 | |
11fdf7f2 | 7619 | // g_conf()->debug_mds = 10; |
7c673cae FG |
7620 | } |
7621 | ||
7622 | ||
7623 | ||
7624 | bool MDCache::shutdown_pass() | |
7625 | { | |
7626 | dout(7) << "shutdown_pass" << dendl; | |
7627 | ||
7628 | if (mds->is_stopped()) { | |
7629 | dout(7) << " already shut down" << dendl; | |
7630 | show_cache(); | |
7631 | show_subtrees(); | |
7632 | return true; | |
7633 | } | |
7634 | ||
7635 | // empty stray dir | |
28e407b8 | 7636 | bool strays_all_exported = shutdown_export_strays(); |
7c673cae FG |
7637 | |
7638 | // trim cache | |
181888fb | 7639 | trim(UINT64_MAX); |
31f18b77 | 7640 | dout(5) << "lru size now " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; |
7c673cae | 7641 | |
28e407b8 | 7642 | // Export all subtrees to another active (usually rank 0) if not rank 0 |
7c673cae FG |
7643 | int num_auth_subtree = 0; |
7644 | if (!subtrees.empty() && | |
28e407b8 | 7645 | mds->get_nodeid() != 0) { |
7c673cae FG |
7646 | dout(7) << "looking for subtrees to export to mds0" << dendl; |
7647 | list<CDir*> ls; | |
7648 | for (map<CDir*, set<CDir*> >::iterator it = subtrees.begin(); | |
7649 | it != subtrees.end(); | |
7650 | ++it) { | |
7651 | CDir *dir = it->first; | |
7652 | if (dir->get_inode()->is_mdsdir()) | |
7653 | continue; | |
7654 | if (dir->is_auth()) { | |
7655 | num_auth_subtree++; | |
7656 | if (dir->is_frozen() || | |
7657 | dir->is_freezing() || | |
7658 | dir->is_ambiguous_dir_auth() || | |
7659 | dir->state_test(CDir::STATE_EXPORTING)) | |
7660 | continue; | |
7661 | ls.push_back(dir); | |
7662 | } | |
7663 | } | |
28e407b8 AA |
7664 | |
7665 | migrator->clear_export_queue(); | |
7c673cae FG |
7666 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { |
7667 | CDir *dir = *p; | |
7668 | mds_rank_t dest = dir->get_inode()->authority().first; | |
7669 | if (dest > 0 && !mds->mdsmap->is_active(dest)) | |
7670 | dest = 0; | |
7671 | dout(7) << "sending " << *dir << " back to mds." << dest << dendl; | |
7672 | migrator->export_dir_nicely(dir, dest); | |
7673 | } | |
7674 | } | |
7675 | ||
28e407b8 AA |
7676 | if (!strays_all_exported) { |
7677 | dout(7) << "waiting for strays to migrate" << dendl; | |
7678 | return false; | |
7679 | } | |
7680 | ||
7c673cae | 7681 | if (num_auth_subtree > 0) { |
11fdf7f2 | 7682 | ceph_assert(mds->get_nodeid() > 0); |
7c673cae FG |
7683 | dout(7) << "still have " << num_auth_subtree << " auth subtrees" << dendl; |
7684 | show_subtrees(); | |
7685 | return false; | |
7686 | } | |
7687 | ||
7688 | // close out any sessions (and open files!) before we try to trim the log, etc. | |
7689 | if (mds->sessionmap.have_unclosed_sessions()) { | |
7690 | if (!mds->server->terminating_sessions) | |
7691 | mds->server->terminate_sessions(); | |
7692 | return false; | |
7693 | } | |
7694 | ||
28e407b8 AA |
7695 | // Fully trim the log so that all objects in cache are clean and may be |
7696 | // trimmed by a future MDCache::trim. Note that MDSRank::tick does not | |
7697 | // trim the log such that the cache eventually becomes clean. | |
f64942e4 AA |
7698 | if (mds->mdlog->get_num_segments() > 0) { |
7699 | auto ls = mds->mdlog->get_current_segment(); | |
7700 | if (ls->num_events > 1 || !ls->dirty_dirfrags.empty()) { | |
7701 | // Current segment contains events other than subtreemap or | |
7702 | // there are dirty dirfrags (see CDir::log_mark_dirty()) | |
7703 | mds->mdlog->start_new_segment(); | |
7704 | mds->mdlog->flush(); | |
7705 | } | |
7706 | } | |
7707 | mds->mdlog->trim_all(); | |
28e407b8 AA |
7708 | if (mds->mdlog->get_num_segments() > 1) { |
7709 | dout(7) << "still >1 segments, waiting for log to trim" << dendl; | |
7710 | return false; | |
7711 | } | |
7712 | ||
7713 | // drop our reference to our stray dir inode | |
7714 | for (int i = 0; i < NUM_STRAY; ++i) { | |
7715 | if (strays[i] && | |
7716 | strays[i]->state_test(CInode::STATE_STRAYPINNED)) { | |
7717 | strays[i]->state_clear(CInode::STATE_STRAYPINNED); | |
7718 | strays[i]->put(CInode::PIN_STRAY); | |
7719 | strays[i]->put_stickydirs(); | |
7720 | } | |
7721 | } | |
7722 | ||
7c673cae FG |
7723 | CDir *mydir = myin ? myin->get_dirfrag(frag_t()) : NULL; |
7724 | if (mydir && !mydir->is_subtree_root()) | |
7725 | mydir = NULL; | |
7726 | ||
7727 | // subtrees map not empty yet? | |
7728 | if (subtrees.size() > (mydir ? 1 : 0)) { | |
7729 | dout(7) << "still have " << num_subtrees() << " subtrees" << dendl; | |
7730 | show_subtrees(); | |
7731 | migrator->show_importing(); | |
7732 | migrator->show_exporting(); | |
7733 | if (!migrator->is_importing() && !migrator->is_exporting()) | |
7734 | show_cache(); | |
7735 | return false; | |
7736 | } | |
11fdf7f2 TL |
7737 | ceph_assert(!migrator->is_exporting()); |
7738 | ceph_assert(!migrator->is_importing()); | |
7c673cae | 7739 | |
f64942e4 AA |
7740 | // replicas may dirty scatter locks |
7741 | if (myin && myin->is_replicated()) { | |
7742 | dout(7) << "still have replicated objects" << dendl; | |
7743 | return false; | |
7744 | } | |
7745 | ||
11fdf7f2 TL |
7746 | if ((myin && myin->get_num_auth_pins()) || |
7747 | (mydir && (mydir->get_auth_pins() || mydir->get_dir_auth_pins()))) { | |
181888fb FG |
7748 | dout(7) << "still have auth pinned objects" << dendl; |
7749 | return false; | |
7750 | } | |
7751 | ||
7c673cae FG |
7752 | // (only do this once!) |
7753 | if (!mds->mdlog->is_capped()) { | |
7754 | dout(7) << "capping the log" << dendl; | |
7755 | mds->mdlog->cap(); | |
7c673cae FG |
7756 | } |
7757 | ||
f64942e4 AA |
7758 | if (!mds->mdlog->empty()) |
7759 | mds->mdlog->trim(0); | |
7760 | ||
7c673cae FG |
7761 | if (!mds->mdlog->empty()) { |
7762 | dout(7) << "waiting for log to flush.. " << mds->mdlog->get_num_events() | |
7763 | << " in " << mds->mdlog->get_num_segments() << " segments" << dendl; | |
7764 | return false; | |
7765 | } | |
7766 | ||
7767 | if (!did_shutdown_log_cap) { | |
7768 | // flush journal header | |
7769 | dout(7) << "writing header for (now-empty) journal" << dendl; | |
11fdf7f2 | 7770 | ceph_assert(mds->mdlog->empty()); |
7c673cae FG |
7771 | mds->mdlog->write_head(0); |
7772 | // NOTE: filer active checker below will block us until this completes. | |
7773 | did_shutdown_log_cap = true; | |
7774 | return false; | |
7775 | } | |
7776 | ||
7777 | // filer active? | |
7778 | if (mds->objecter->is_active()) { | |
7779 | dout(7) << "objecter still active" << dendl; | |
7780 | mds->objecter->dump_active(); | |
7781 | return false; | |
7782 | } | |
7783 | ||
7784 | // trim what we can from the cache | |
31f18b77 FG |
7785 | if (lru.lru_get_size() > 0 || bottom_lru.lru_get_size() > 0) { |
7786 | dout(7) << "there's still stuff in the cache: " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; | |
7c673cae FG |
7787 | show_cache(); |
7788 | //dump(); | |
7789 | return false; | |
7790 | } | |
31f18b77 FG |
7791 | |
7792 | // make mydir subtree go away | |
7793 | if (mydir) { | |
7794 | if (mydir->get_num_ref() > 1) { // subtree pin | |
7795 | dout(7) << "there's still reference to mydir " << *mydir << dendl; | |
7796 | show_cache(); | |
7797 | return false; | |
7798 | } | |
7799 | ||
7800 | remove_subtree(mydir); | |
7801 | myin->close_dirfrag(mydir->get_frag()); | |
7802 | } | |
11fdf7f2 | 7803 | ceph_assert(subtrees.empty()); |
31f18b77 | 7804 | |
1adf2230 | 7805 | if (myin) { |
31f18b77 | 7806 | remove_inode(myin); |
11fdf7f2 | 7807 | ceph_assert(!myin); |
1adf2230 AA |
7808 | } |
7809 | ||
11fdf7f2 TL |
7810 | if (global_snaprealm) { |
7811 | remove_inode(global_snaprealm->inode); | |
7812 | global_snaprealm = nullptr; | |
7813 | } | |
7814 | ||
7c673cae | 7815 | // done! |
a8e16298 | 7816 | dout(5) << "shutdown done." << dendl; |
7c673cae FG |
7817 | return true; |
7818 | } | |
7819 | ||
7820 | bool MDCache::shutdown_export_strays() | |
7821 | { | |
f64942e4 AA |
7822 | static const unsigned MAX_EXPORTING = 100; |
7823 | ||
7c673cae FG |
7824 | if (mds->get_nodeid() == 0) |
7825 | return true; | |
f64942e4 AA |
7826 | |
7827 | if (shutdown_exporting_strays.size() * 3 >= MAX_EXPORTING * 2) | |
7828 | return false; | |
7829 | ||
7830 | dout(10) << "shutdown_export_strays " << shutdown_export_next.first | |
7831 | << " '" << shutdown_export_next.second << "'" << dendl; | |
7c673cae FG |
7832 | |
7833 | bool mds0_active = mds->mdsmap->is_active(mds_rank_t(0)); | |
f64942e4 | 7834 | bool all_exported = false; |
7c673cae | 7835 | |
f64942e4 AA |
7836 | again: |
7837 | auto next = shutdown_export_next; | |
7c673cae | 7838 | |
7c673cae | 7839 | for (int i = 0; i < NUM_STRAY; ++i) { |
f64942e4 AA |
7840 | CInode *strayi = strays[i]; |
7841 | if (!strayi || | |
7842 | !strayi->state_test(CInode::STATE_STRAYPINNED)) | |
7843 | continue; | |
7844 | if (strayi->ino() < next.first.ino) | |
7c673cae | 7845 | continue; |
7c673cae | 7846 | |
f64942e4 AA |
7847 | deque<CDir*> dfls; |
7848 | strayi->get_dirfrags(dfls); | |
7c673cae | 7849 | |
f64942e4 AA |
7850 | while (!dfls.empty()) { |
7851 | CDir *dir = dfls.front(); | |
7852 | dfls.pop_front(); | |
7853 | ||
7854 | if (dir->dirfrag() < next.first) | |
7c673cae | 7855 | continue; |
f64942e4 AA |
7856 | if (next.first < dir->dirfrag()) { |
7857 | next.first = dir->dirfrag(); | |
7858 | next.second.clear(); | |
7859 | } | |
7860 | ||
7861 | if (!dir->is_complete()) { | |
11fdf7f2 | 7862 | MDSContext *fin = nullptr; |
f64942e4 AA |
7863 | if (shutdown_exporting_strays.empty()) { |
7864 | fin = new MDSInternalContextWrapper(mds, | |
7865 | new FunctionContext([this](int r) { | |
7866 | shutdown_export_strays(); | |
7867 | }) | |
7868 | ); | |
7869 | } | |
7870 | dir->fetch(fin); | |
7871 | goto done; | |
7c673cae FG |
7872 | } |
7873 | ||
f64942e4 AA |
7874 | CDir::dentry_key_map::iterator it; |
7875 | if (next.second.empty()) { | |
7876 | it = dir->begin(); | |
7c673cae | 7877 | } else { |
f64942e4 AA |
7878 | auto hash = ceph_frag_value(strayi->hash_dentry_name(next.second)); |
7879 | it = dir->lower_bound(dentry_key_t(0, next.second, hash)); | |
7c673cae | 7880 | } |
f64942e4 AA |
7881 | |
7882 | for (; it != dir->end(); ++it) { | |
7883 | CDentry *dn = it->second; | |
7884 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
7885 | if (dnl->is_null()) | |
7886 | continue; | |
7887 | ||
7888 | if (!mds0_active && !dn->state_test(CDentry::STATE_PURGING)) { | |
11fdf7f2 | 7889 | next.second = it->first.name; |
f64942e4 AA |
7890 | goto done; |
7891 | } | |
7892 | ||
7893 | auto ret = shutdown_exporting_strays.insert(dnl->get_inode()->ino()); | |
7894 | if (!ret.second) { | |
7895 | dout(10) << "already exporting/purging " << *dn << dendl; | |
7896 | continue; | |
7897 | } | |
7898 | ||
7899 | // Don't try to migrate anything that is actually | |
7900 | // being purged right now | |
7901 | if (!dn->state_test(CDentry::STATE_PURGING)) | |
7902 | stray_manager.migrate_stray(dn, mds_rank_t(0)); // send to root! | |
7903 | ||
7904 | if (shutdown_exporting_strays.size() >= MAX_EXPORTING) { | |
7905 | ++it; | |
7906 | if (it != dir->end()) { | |
11fdf7f2 | 7907 | next.second = it->first.name; |
f64942e4 AA |
7908 | } else { |
7909 | if (dfls.empty()) | |
7910 | next.first.ino.val++; | |
7911 | else | |
7912 | next.first = dfls.front()->dirfrag(); | |
7913 | next.second.clear(); | |
7914 | } | |
7915 | goto done; | |
7916 | } | |
7917 | } | |
7918 | } | |
7919 | } | |
7920 | ||
7921 | if (shutdown_exporting_strays.empty()) { | |
7922 | dirfrag_t first_df(MDS_INO_STRAY(mds->get_nodeid(), 0), 0); | |
7923 | if (first_df < shutdown_export_next.first || | |
7924 | !shutdown_export_next.second.empty()) { | |
7925 | shutdown_export_next.first = first_df; | |
7926 | shutdown_export_next.second.clear(); | |
7927 | goto again; | |
7c673cae | 7928 | } |
f64942e4 | 7929 | all_exported = true; |
7c673cae FG |
7930 | } |
7931 | ||
f64942e4 AA |
7932 | done: |
7933 | shutdown_export_next = next; | |
7934 | return all_exported; | |
7c673cae FG |
7935 | } |
7936 | ||
7937 | // ========= messaging ============== | |
7938 | ||
11fdf7f2 | 7939 | void MDCache::dispatch(const Message::const_ref &m) |
7c673cae FG |
7940 | { |
7941 | switch (m->get_type()) { | |
7942 | ||
7943 | // RESOLVE | |
7944 | case MSG_MDS_RESOLVE: | |
11fdf7f2 | 7945 | handle_resolve(MMDSResolve::msgref_cast(m)); |
7c673cae FG |
7946 | break; |
7947 | case MSG_MDS_RESOLVEACK: | |
11fdf7f2 | 7948 | handle_resolve_ack(MMDSResolveAck::msgref_cast(m)); |
7c673cae FG |
7949 | break; |
7950 | ||
7951 | // REJOIN | |
7952 | case MSG_MDS_CACHEREJOIN: | |
11fdf7f2 | 7953 | handle_cache_rejoin(MMDSCacheRejoin::msgref_cast(m)); |
7c673cae FG |
7954 | break; |
7955 | ||
7956 | case MSG_MDS_DISCOVER: | |
11fdf7f2 | 7957 | handle_discover(MDiscover::msgref_cast(m)); |
7c673cae FG |
7958 | break; |
7959 | case MSG_MDS_DISCOVERREPLY: | |
11fdf7f2 | 7960 | handle_discover_reply(MDiscoverReply::msgref_cast(m)); |
7c673cae FG |
7961 | break; |
7962 | ||
7963 | case MSG_MDS_DIRUPDATE: | |
11fdf7f2 | 7964 | handle_dir_update(MDirUpdate::msgref_cast(m)); |
7c673cae FG |
7965 | break; |
7966 | ||
7967 | case MSG_MDS_CACHEEXPIRE: | |
11fdf7f2 | 7968 | handle_cache_expire(MCacheExpire::msgref_cast(m)); |
7c673cae FG |
7969 | break; |
7970 | ||
7971 | case MSG_MDS_DENTRYLINK: | |
11fdf7f2 | 7972 | handle_dentry_link(MDentryLink::msgref_cast(m)); |
7c673cae FG |
7973 | break; |
7974 | case MSG_MDS_DENTRYUNLINK: | |
11fdf7f2 | 7975 | handle_dentry_unlink(MDentryUnlink::msgref_cast(m)); |
7c673cae FG |
7976 | break; |
7977 | ||
7978 | case MSG_MDS_FRAGMENTNOTIFY: | |
11fdf7f2 | 7979 | handle_fragment_notify(MMDSFragmentNotify::msgref_cast(m)); |
7c673cae | 7980 | break; |
a8e16298 | 7981 | case MSG_MDS_FRAGMENTNOTIFYACK: |
11fdf7f2 | 7982 | handle_fragment_notify_ack(MMDSFragmentNotifyAck::msgref_cast(m)); |
a8e16298 | 7983 | break; |
7c673cae FG |
7984 | |
7985 | case MSG_MDS_FINDINO: | |
11fdf7f2 | 7986 | handle_find_ino(MMDSFindIno::msgref_cast(m)); |
7c673cae FG |
7987 | break; |
7988 | case MSG_MDS_FINDINOREPLY: | |
11fdf7f2 | 7989 | handle_find_ino_reply(MMDSFindInoReply::msgref_cast(m)); |
7c673cae FG |
7990 | break; |
7991 | ||
7992 | case MSG_MDS_OPENINO: | |
11fdf7f2 | 7993 | handle_open_ino(MMDSOpenIno::msgref_cast(m)); |
7c673cae FG |
7994 | break; |
7995 | case MSG_MDS_OPENINOREPLY: | |
11fdf7f2 TL |
7996 | handle_open_ino_reply(MMDSOpenInoReply::msgref_cast(m)); |
7997 | break; | |
7998 | ||
7999 | case MSG_MDS_SNAPUPDATE: | |
8000 | handle_snap_update(MMDSSnapUpdate::msgref_cast(m)); | |
7c673cae FG |
8001 | break; |
8002 | ||
8003 | default: | |
8004 | derr << "cache unknown message " << m->get_type() << dendl; | |
11fdf7f2 | 8005 | ceph_abort_msg("cache unknown message"); |
7c673cae FG |
8006 | } |
8007 | } | |
8008 | ||
11fdf7f2 TL |
8009 | int MDCache::path_traverse(MDRequestRef& mdr, MDSContextFactory& cf, // who |
8010 | const filepath& path, // what | |
7c673cae | 8011 | vector<CDentry*> *pdnvec, // result |
11fdf7f2 | 8012 | CInode **pin, |
7c673cae FG |
8013 | int onfail) |
8014 | { | |
8015 | bool discover = (onfail == MDS_TRAVERSE_DISCOVER); | |
8016 | bool null_okay = (onfail == MDS_TRAVERSE_DISCOVERXLOCK); | |
8017 | bool forward = (onfail == MDS_TRAVERSE_FORWARD); | |
8018 | ||
11fdf7f2 | 8019 | ceph_assert(!forward || mdr); // forward requires a request |
7c673cae FG |
8020 | |
8021 | snapid_t snapid = CEPH_NOSNAP; | |
8022 | if (mdr) | |
8023 | mdr->snapid = snapid; | |
8024 | ||
8025 | client_t client = (mdr && mdr->reqid.name.is_client()) ? mdr->reqid.name.num() : -1; | |
8026 | ||
8027 | if (mds->logger) mds->logger->inc(l_mds_traverse); | |
8028 | ||
8029 | dout(7) << "traverse: opening base ino " << path.get_ino() << " snap " << snapid << dendl; | |
8030 | CInode *cur = get_inode(path.get_ino()); | |
8031 | if (cur == NULL) { | |
8032 | if (MDS_INO_IS_MDSDIR(path.get_ino())) | |
11fdf7f2 | 8033 | open_foreign_mdsdir(path.get_ino(), cf.build()); |
7c673cae FG |
8034 | else { |
8035 | //ceph_abort(); // hrm.. broken | |
8036 | return -ESTALE; | |
8037 | } | |
8038 | return 1; | |
8039 | } | |
8040 | if (cur->state_test(CInode::STATE_PURGING)) | |
8041 | return -ESTALE; | |
8042 | ||
8043 | // make sure snaprealm are open... | |
11fdf7f2 TL |
8044 | if (mdr && cur->snaprealm && !cur->snaprealm->have_past_parents_open() && |
8045 | !cur->snaprealm->open_parents(cf.build())) { | |
7c673cae FG |
8046 | return 1; |
8047 | } | |
8048 | ||
8049 | // start trace | |
8050 | if (pdnvec) | |
8051 | pdnvec->clear(); | |
8052 | if (pin) | |
8053 | *pin = cur; | |
8054 | ||
8055 | unsigned depth = 0; | |
8056 | while (depth < path.depth()) { | |
8057 | dout(12) << "traverse: path seg depth " << depth << " '" << path[depth] | |
8058 | << "' snapid " << snapid << dendl; | |
8059 | ||
8060 | if (!cur->is_dir()) { | |
8061 | dout(7) << "traverse: " << *cur << " not a dir " << dendl; | |
8062 | return -ENOTDIR; | |
8063 | } | |
8064 | ||
8065 | // walk into snapdir? | |
8066 | if (path[depth].length() == 0) { | |
8067 | dout(10) << "traverse: snapdir" << dendl; | |
8068 | if (!mdr) | |
8069 | return -EINVAL; | |
8070 | snapid = CEPH_SNAPDIR; | |
8071 | mdr->snapid = snapid; | |
8072 | depth++; | |
8073 | continue; | |
8074 | } | |
8075 | // walk thru snapdir? | |
8076 | if (snapid == CEPH_SNAPDIR) { | |
8077 | if (!mdr) | |
8078 | return -EINVAL; | |
8079 | SnapRealm *realm = cur->find_snaprealm(); | |
8080 | snapid = realm->resolve_snapname(path[depth], cur->ino()); | |
8081 | dout(10) << "traverse: snap " << path[depth] << " -> " << snapid << dendl; | |
11fdf7f2 TL |
8082 | if (!snapid) { |
8083 | CInode *t = cur; | |
8084 | while (t) { | |
8085 | // if snaplock isn't readable, it's possible that other mds is creating | |
8086 | // snapshot, but snap update message hasn't been received. | |
8087 | if (!t->snaplock.can_read(client)) { | |
8088 | dout(10) << " non-readable snaplock on " << *t << dendl; | |
8089 | t->snaplock.add_waiter(SimpleLock::WAIT_RD, cf.build()); | |
8090 | return 1; | |
8091 | } | |
8092 | CDentry *pdn = t->get_projected_parent_dn(); | |
8093 | t = pdn ? pdn->get_dir()->get_inode() : NULL; | |
8094 | } | |
7c673cae | 8095 | return -ENOENT; |
11fdf7f2 | 8096 | } |
7c673cae FG |
8097 | mdr->snapid = snapid; |
8098 | depth++; | |
8099 | continue; | |
8100 | } | |
8101 | ||
8102 | // open dir | |
8103 | frag_t fg = cur->pick_dirfrag(path[depth]); | |
8104 | CDir *curdir = cur->get_dirfrag(fg); | |
8105 | if (!curdir) { | |
8106 | if (cur->is_auth()) { | |
8107 | // parent dir frozen_dir? | |
8108 | if (cur->is_frozen()) { | |
8109 | dout(7) << "traverse: " << *cur << " is frozen, waiting" << dendl; | |
11fdf7f2 | 8110 | cur->add_waiter(CDir::WAIT_UNFREEZE, cf.build()); |
7c673cae FG |
8111 | return 1; |
8112 | } | |
8113 | curdir = cur->get_or_open_dirfrag(this, fg); | |
8114 | } else { | |
8115 | // discover? | |
8116 | dout(10) << "traverse: need dirfrag " << fg << ", doing discover from " << *cur << dendl; | |
11fdf7f2 | 8117 | discover_path(cur, snapid, path.postfixpath(depth), cf.build(), |
7c673cae FG |
8118 | null_okay); |
8119 | if (mds->logger) mds->logger->inc(l_mds_traverse_discover); | |
8120 | return 1; | |
8121 | } | |
8122 | } | |
11fdf7f2 | 8123 | ceph_assert(curdir); |
7c673cae FG |
8124 | |
8125 | #ifdef MDS_VERIFY_FRAGSTAT | |
8126 | if (curdir->is_complete()) | |
8127 | curdir->verify_fragstat(); | |
8128 | #endif | |
8129 | ||
8130 | // frozen? | |
8131 | /* | |
8132 | if (curdir->is_frozen()) { | |
8133 | // doh! | |
8134 | // FIXME: traverse is allowed? | |
8135 | dout(7) << "traverse: " << *curdir << " is frozen, waiting" << dendl; | |
8136 | curdir->add_waiter(CDir::WAIT_UNFREEZE, _get_waiter(mdr, req, fin)); | |
8137 | if (onfinish) delete onfinish; | |
8138 | return 1; | |
8139 | } | |
8140 | */ | |
8141 | ||
8142 | // Before doing dirfrag->dn lookup, compare with DamageTable's | |
8143 | // record of which dentries were unreadable | |
8144 | if (mds->damage_table.is_dentry_damaged(curdir, path[depth], snapid)) { | |
8145 | dout(4) << "traverse: stopped lookup at damaged dentry " | |
8146 | << *curdir << "/" << path[depth] << " snap=" << snapid << dendl; | |
8147 | return -EIO; | |
8148 | } | |
8149 | ||
8150 | // dentry | |
8151 | CDentry *dn = curdir->lookup(path[depth], snapid); | |
8152 | CDentry::linkage_t *dnl = dn ? dn->get_projected_linkage() : 0; | |
8153 | ||
8154 | // null and last_bit and xlocked by me? | |
8155 | if (dnl && dnl->is_null() && null_okay) { | |
8156 | dout(10) << "traverse: hit null dentry at tail of traverse, succeeding" << dendl; | |
8157 | if (pdnvec) | |
8158 | pdnvec->push_back(dn); | |
8159 | if (pin) | |
8160 | *pin = 0; | |
8161 | break; // done! | |
8162 | } | |
8163 | ||
8164 | if (dnl && | |
8165 | dn->lock.is_xlocked() && | |
8166 | dn->lock.get_xlock_by() != mdr && | |
8167 | !dn->lock.can_read(client) && | |
8168 | (dnl->is_null() || forward)) { | |
8169 | dout(10) << "traverse: xlocked dentry at " << *dn << dendl; | |
11fdf7f2 | 8170 | dn->lock.add_waiter(SimpleLock::WAIT_RD, cf.build()); |
7c673cae FG |
8171 | if (mds->logger) mds->logger->inc(l_mds_traverse_lock); |
8172 | mds->mdlog->flush(); | |
8173 | return 1; | |
8174 | } | |
8175 | ||
8176 | // can we conclude ENOENT? | |
8177 | if (dnl && dnl->is_null()) { | |
8178 | if (dn->lock.can_read(client) || | |
8179 | (dn->lock.is_xlocked() && dn->lock.get_xlock_by() == mdr)) { | |
8180 | dout(10) << "traverse: miss on null+readable dentry " << path[depth] << " " << *dn << dendl; | |
8181 | if (pdnvec) { | |
8182 | if (depth == path.depth() - 1) | |
8183 | pdnvec->push_back(dn); | |
8184 | else | |
8185 | pdnvec->clear(); // do not confuse likes of rdlock_path_pin_ref(); | |
8186 | } | |
8187 | return -ENOENT; | |
8188 | } else { | |
8189 | dout(10) << "miss on dentry " << *dn << ", can't read due to lock" << dendl; | |
11fdf7f2 | 8190 | dn->lock.add_waiter(SimpleLock::WAIT_RD, cf.build()); |
7c673cae FG |
8191 | return 1; |
8192 | } | |
8193 | } | |
8194 | ||
8195 | if (dnl && !dnl->is_null()) { | |
8196 | CInode *in = dnl->get_inode(); | |
8197 | ||
8198 | // do we have inode? | |
8199 | if (!in) { | |
11fdf7f2 | 8200 | ceph_assert(dnl->is_remote()); |
7c673cae FG |
8201 | // do i have it? |
8202 | in = get_inode(dnl->get_remote_ino()); | |
8203 | if (in) { | |
8204 | dout(7) << "linking in remote in " << *in << dendl; | |
8205 | dn->link_remote(dnl, in); | |
8206 | } else { | |
8207 | dout(7) << "remote link to " << dnl->get_remote_ino() << ", which i don't have" << dendl; | |
11fdf7f2 | 8208 | ceph_assert(mdr); // we shouldn't hit non-primary dentries doing a non-mdr traversal! |
7c673cae FG |
8209 | if (mds->damage_table.is_remote_damaged(dnl->get_remote_ino())) { |
8210 | dout(4) << "traverse: remote dentry points to damaged ino " | |
8211 | << *dn << dendl; | |
8212 | return -EIO; | |
8213 | } | |
11fdf7f2 | 8214 | open_remote_dentry(dn, true, cf.build(), |
7c673cae FG |
8215 | (null_okay && depth == path.depth() - 1)); |
8216 | if (mds->logger) mds->logger->inc(l_mds_traverse_remote_ino); | |
8217 | return 1; | |
8218 | } | |
8219 | } | |
8220 | ||
8221 | cur = in; | |
8222 | // make sure snaprealm are open... | |
11fdf7f2 TL |
8223 | if (mdr && cur->snaprealm && !cur->snaprealm->have_past_parents_open() && |
8224 | !cur->snaprealm->open_parents(cf.build())) { | |
7c673cae FG |
8225 | return 1; |
8226 | } | |
8227 | ||
8228 | // add to trace, continue. | |
8229 | touch_inode(cur); | |
8230 | if (pdnvec) | |
8231 | pdnvec->push_back(dn); | |
8232 | if (pin) | |
8233 | *pin = cur; | |
8234 | depth++; | |
8235 | continue; | |
8236 | } | |
8237 | ||
8238 | ||
8239 | // MISS. dentry doesn't exist. | |
8240 | dout(12) << "traverse: miss on dentry " << path[depth] << " in " << *curdir << dendl; | |
8241 | ||
8242 | if (curdir->is_auth()) { | |
8243 | // dentry is mine. | |
8244 | if (curdir->is_complete() || | |
8245 | (snapid == CEPH_NOSNAP && | |
8246 | curdir->has_bloom() && | |
11fdf7f2 | 8247 | !curdir->is_in_bloom(path[depth]))) { |
7c673cae FG |
8248 | // file not found |
8249 | if (pdnvec) { | |
8250 | // instantiate a null dn? | |
8251 | if (depth < path.depth()-1){ | |
8252 | dout(20) << " didn't traverse full path; not returning pdnvec" << dendl; | |
8253 | dn = NULL; | |
8254 | } else if (dn) { | |
8255 | ceph_abort(); // should have fallen out in ->is_null() check above | |
8256 | } else if (curdir->is_frozen()) { | |
8257 | dout(20) << " not adding null to frozen dir " << dendl; | |
8258 | } else if (snapid < CEPH_MAXSNAP) { | |
8259 | dout(20) << " not adding null for snapid " << snapid << dendl; | |
8260 | } else { | |
8261 | // create a null dentry | |
8262 | dn = curdir->add_null_dentry(path[depth]); | |
8263 | dout(20) << " added null " << *dn << dendl; | |
8264 | } | |
8265 | if (dn) | |
8266 | pdnvec->push_back(dn); | |
8267 | else | |
8268 | pdnvec->clear(); // do not confuse likes of rdlock_path_pin_ref(); | |
8269 | } | |
8270 | return -ENOENT; | |
8271 | } else { | |
8272 | ||
8273 | // Check DamageTable for missing fragments before trying to fetch | |
8274 | // this | |
8275 | if (mds->damage_table.is_dirfrag_damaged(curdir)) { | |
8276 | dout(4) << "traverse: damaged dirfrag " << *curdir | |
8277 | << ", blocking fetch" << dendl; | |
8278 | return -EIO; | |
8279 | } | |
8280 | ||
8281 | // directory isn't complete; reload | |
8282 | dout(7) << "traverse: incomplete dir contents for " << *cur << ", fetching" << dendl; | |
8283 | touch_inode(cur); | |
11fdf7f2 | 8284 | curdir->fetch(cf.build(), path[depth]); |
7c673cae FG |
8285 | if (mds->logger) mds->logger->inc(l_mds_traverse_dir_fetch); |
8286 | return 1; | |
8287 | } | |
8288 | } else { | |
8289 | // dirfrag/dentry is not mine. | |
8290 | mds_authority_t dauth = curdir->authority(); | |
8291 | ||
8292 | if (forward && | |
11fdf7f2 | 8293 | mdr && mdr->client_request && |
7c673cae FG |
8294 | (int)depth < mdr->client_request->get_num_fwd()) { |
8295 | dout(7) << "traverse: snap " << snapid << " and depth " << depth | |
8296 | << " < fwd " << mdr->client_request->get_num_fwd() | |
8297 | << ", discovering instead of forwarding" << dendl; | |
8298 | discover = true; | |
8299 | } | |
8300 | ||
8301 | if ((discover || null_okay)) { | |
8302 | dout(7) << "traverse: discover from " << path[depth] << " from " << *curdir << dendl; | |
11fdf7f2 | 8303 | discover_path(curdir, snapid, path.postfixpath(depth), cf.build(), |
7c673cae FG |
8304 | null_okay); |
8305 | if (mds->logger) mds->logger->inc(l_mds_traverse_discover); | |
8306 | return 1; | |
8307 | } | |
8308 | if (forward) { | |
8309 | // forward | |
8310 | dout(7) << "traverse: not auth for " << path << " in " << *curdir << dendl; | |
8311 | ||
8312 | if (curdir->is_ambiguous_auth()) { | |
8313 | // wait | |
8314 | dout(7) << "traverse: waiting for single auth in " << *curdir << dendl; | |
11fdf7f2 | 8315 | curdir->add_waiter(CDir::WAIT_SINGLEAUTH, cf.build()); |
7c673cae FG |
8316 | return 1; |
8317 | } | |
8318 | ||
8319 | dout(7) << "traverse: forwarding, not auth for " << *curdir << dendl; | |
11fdf7f2 TL |
8320 | |
8321 | request_forward(mdr, dauth.first); | |
8322 | ||
7c673cae | 8323 | if (mds->logger) mds->logger->inc(l_mds_traverse_forward); |
7c673cae | 8324 | return 2; |
11fdf7f2 | 8325 | } |
7c673cae | 8326 | } |
11fdf7f2 | 8327 | |
7c673cae FG |
8328 | ceph_abort(); // i shouldn't get here |
8329 | } | |
8330 | ||
8331 | // success. | |
8332 | if (mds->logger) mds->logger->inc(l_mds_traverse_hit); | |
8333 | dout(10) << "path_traverse finish on snapid " << snapid << dendl; | |
8334 | if (mdr) | |
11fdf7f2 | 8335 | ceph_assert(mdr->snapid == snapid); |
7c673cae FG |
8336 | return 0; |
8337 | } | |
8338 | ||
8339 | CInode *MDCache::cache_traverse(const filepath& fp) | |
8340 | { | |
8341 | dout(10) << "cache_traverse " << fp << dendl; | |
8342 | ||
8343 | CInode *in; | |
8344 | if (fp.get_ino()) | |
8345 | in = get_inode(fp.get_ino()); | |
8346 | else | |
8347 | in = root; | |
8348 | if (!in) | |
8349 | return NULL; | |
8350 | ||
8351 | for (unsigned i = 0; i < fp.depth(); i++) { | |
11fdf7f2 | 8352 | std::string_view dname = fp[i]; |
7c673cae FG |
8353 | frag_t fg = in->pick_dirfrag(dname); |
8354 | dout(20) << " " << i << " " << dname << " frag " << fg << " from " << *in << dendl; | |
8355 | CDir *curdir = in->get_dirfrag(fg); | |
8356 | if (!curdir) | |
8357 | return NULL; | |
8358 | CDentry *dn = curdir->lookup(dname, CEPH_NOSNAP); | |
8359 | if (!dn) | |
8360 | return NULL; | |
8361 | in = dn->get_linkage()->get_inode(); | |
8362 | if (!in) | |
8363 | return NULL; | |
8364 | } | |
8365 | dout(10) << " got " << *in << dendl; | |
8366 | return in; | |
8367 | } | |
8368 | ||
8369 | ||
8370 | /** | |
8371 | * open_remote_dir -- open up a remote dirfrag | |
8372 | * | |
8373 | * @param diri base inode | |
8374 | * @param approxfg approximate fragment. | |
8375 | * @param fin completion callback | |
8376 | */ | |
11fdf7f2 | 8377 | void MDCache::open_remote_dirfrag(CInode *diri, frag_t approxfg, MDSContext *fin) |
7c673cae FG |
8378 | { |
8379 | dout(10) << "open_remote_dir on " << *diri << dendl; | |
11fdf7f2 TL |
8380 | ceph_assert(diri->is_dir()); |
8381 | ceph_assert(!diri->is_auth()); | |
8382 | ceph_assert(diri->get_dirfrag(approxfg) == 0); | |
7c673cae | 8383 | |
224ce89b | 8384 | discover_dir_frag(diri, approxfg, fin); |
7c673cae FG |
8385 | } |
8386 | ||
8387 | ||
8388 | /** | |
8389 | * get_dentry_inode - get or open inode | |
8390 | * | |
8391 | * @param dn the dentry | |
8392 | * @param mdr current request | |
8393 | * | |
8394 | * will return inode for primary, or link up/open up remote link's inode as necessary. | |
8395 | * If it's not available right now, puts mdr on wait list and returns null. | |
8396 | */ | |
8397 | CInode *MDCache::get_dentry_inode(CDentry *dn, MDRequestRef& mdr, bool projected) | |
8398 | { | |
8399 | CDentry::linkage_t *dnl; | |
8400 | if (projected) | |
8401 | dnl = dn->get_projected_linkage(); | |
8402 | else | |
8403 | dnl = dn->get_linkage(); | |
8404 | ||
11fdf7f2 | 8405 | ceph_assert(!dnl->is_null()); |
7c673cae FG |
8406 | |
8407 | if (dnl->is_primary()) | |
8408 | return dnl->inode; | |
8409 | ||
11fdf7f2 | 8410 | ceph_assert(dnl->is_remote()); |
7c673cae FG |
8411 | CInode *in = get_inode(dnl->get_remote_ino()); |
8412 | if (in) { | |
8413 | dout(7) << "get_dentry_inode linking in remote in " << *in << dendl; | |
8414 | dn->link_remote(dnl, in); | |
8415 | return in; | |
8416 | } else { | |
8417 | dout(10) << "get_dentry_inode on remote dn, opening inode for " << *dn << dendl; | |
8418 | open_remote_dentry(dn, projected, new C_MDS_RetryRequest(this, mdr)); | |
8419 | return 0; | |
8420 | } | |
8421 | } | |
8422 | ||
8423 | struct C_MDC_OpenRemoteDentry : public MDCacheContext { | |
8424 | CDentry *dn; | |
8425 | inodeno_t ino; | |
11fdf7f2 | 8426 | MDSContext *onfinish; |
7c673cae | 8427 | bool want_xlocked; |
11fdf7f2 | 8428 | C_MDC_OpenRemoteDentry(MDCache *m, CDentry *d, inodeno_t i, MDSContext *f, bool wx) : |
31f18b77 FG |
8429 | MDCacheContext(m), dn(d), ino(i), onfinish(f), want_xlocked(wx) { |
8430 | dn->get(MDSCacheObject::PIN_PTRWAITER); | |
8431 | } | |
7c673cae FG |
8432 | void finish(int r) override { |
8433 | mdcache->_open_remote_dentry_finish(dn, ino, onfinish, want_xlocked, r); | |
31f18b77 | 8434 | dn->put(MDSCacheObject::PIN_PTRWAITER); |
7c673cae FG |
8435 | } |
8436 | }; | |
8437 | ||
11fdf7f2 | 8438 | void MDCache::open_remote_dentry(CDentry *dn, bool projected, MDSContext *fin, bool want_xlocked) |
7c673cae FG |
8439 | { |
8440 | dout(10) << "open_remote_dentry " << *dn << dendl; | |
8441 | CDentry::linkage_t *dnl = projected ? dn->get_projected_linkage() : dn->get_linkage(); | |
8442 | inodeno_t ino = dnl->get_remote_ino(); | |
8443 | int64_t pool = dnl->get_remote_d_type() == DT_DIR ? mds->mdsmap->get_metadata_pool() : -1; | |
8444 | open_ino(ino, pool, | |
8445 | new C_MDC_OpenRemoteDentry(this, dn, ino, fin, want_xlocked), true, want_xlocked); // backtrace | |
8446 | } | |
8447 | ||
11fdf7f2 | 8448 | void MDCache::_open_remote_dentry_finish(CDentry *dn, inodeno_t ino, MDSContext *fin, |
7c673cae FG |
8449 | bool want_xlocked, int r) |
8450 | { | |
8451 | if (r < 0) { | |
31f18b77 FG |
8452 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); |
8453 | if (dnl->is_remote() && dnl->get_remote_ino() == ino) { | |
7c673cae FG |
8454 | dout(0) << "open_remote_dentry_finish bad remote dentry " << *dn << dendl; |
8455 | dn->state_set(CDentry::STATE_BADREMOTEINO); | |
8456 | ||
8457 | std::string path; | |
8458 | CDir *dir = dn->get_dir(); | |
8459 | if (dir) { | |
31f18b77 | 8460 | dir->get_inode()->make_path_string(path); |
94b18763 | 8461 | path += "/"; |
11fdf7f2 | 8462 | path += dn->get_name(); |
7c673cae FG |
8463 | } |
8464 | ||
31f18b77 | 8465 | bool fatal = mds->damage_table.notify_remote_damaged(ino, path); |
7c673cae | 8466 | if (fatal) { |
31f18b77 FG |
8467 | mds->damaged(); |
8468 | ceph_abort(); // unreachable, damaged() respawns us | |
7c673cae | 8469 | } |
31f18b77 FG |
8470 | } else { |
8471 | r = 0; | |
8472 | } | |
7c673cae FG |
8473 | } |
8474 | fin->complete(r < 0 ? r : 0); | |
8475 | } | |
8476 | ||
8477 | ||
8478 | void MDCache::make_trace(vector<CDentry*>& trace, CInode *in) | |
8479 | { | |
8480 | // empty trace if we're a base inode | |
8481 | if (in->is_base()) | |
8482 | return; | |
8483 | ||
8484 | CInode *parent = in->get_parent_inode(); | |
11fdf7f2 | 8485 | ceph_assert(parent); |
7c673cae FG |
8486 | make_trace(trace, parent); |
8487 | ||
8488 | CDentry *dn = in->get_parent_dn(); | |
8489 | dout(15) << "make_trace adding " << *dn << dendl; | |
8490 | trace.push_back(dn); | |
8491 | } | |
8492 | ||
8493 | ||
8494 | // ------------------------------------------------------------------------------- | |
8495 | // Open inode by inode number | |
8496 | ||
8497 | class C_IO_MDC_OpenInoBacktraceFetched : public MDCacheIOContext { | |
8498 | inodeno_t ino; | |
8499 | public: | |
8500 | bufferlist bl; | |
8501 | C_IO_MDC_OpenInoBacktraceFetched(MDCache *c, inodeno_t i) : | |
8502 | MDCacheIOContext(c), ino(i) {} | |
8503 | void finish(int r) override { | |
8504 | mdcache->_open_ino_backtrace_fetched(ino, bl, r); | |
8505 | } | |
91327a77 AA |
8506 | void print(ostream& out) const override { |
8507 | out << "openino_backtrace_fetch" << ino << ")"; | |
8508 | } | |
7c673cae FG |
8509 | }; |
8510 | ||
8511 | struct C_MDC_OpenInoTraverseDir : public MDCacheContext { | |
8512 | inodeno_t ino; | |
11fdf7f2 | 8513 | MMDSOpenIno::const_ref msg; |
7c673cae FG |
8514 | bool parent; |
8515 | public: | |
11fdf7f2 | 8516 | C_MDC_OpenInoTraverseDir(MDCache *c, inodeno_t i, const MMDSOpenIno::const_ref &m, bool p) : |
7c673cae FG |
8517 | MDCacheContext(c), ino(i), msg(m), parent(p) {} |
8518 | void finish(int r) override { | |
8519 | if (r < 0 && !parent) | |
8520 | r = -EAGAIN; | |
8521 | if (msg) { | |
8522 | mdcache->handle_open_ino(msg, r); | |
8523 | return; | |
8524 | } | |
11fdf7f2 TL |
8525 | auto& info = mdcache->opening_inodes.at(ino); |
8526 | mdcache->_open_ino_traverse_dir(ino, info, r); | |
7c673cae FG |
8527 | } |
8528 | }; | |
8529 | ||
8530 | struct C_MDC_OpenInoParentOpened : public MDCacheContext { | |
8531 | inodeno_t ino; | |
8532 | public: | |
8533 | C_MDC_OpenInoParentOpened(MDCache *c, inodeno_t i) : MDCacheContext(c), ino(i) {} | |
8534 | void finish(int r) override { | |
8535 | mdcache->_open_ino_parent_opened(ino, r); | |
8536 | } | |
8537 | }; | |
8538 | ||
8539 | void MDCache::_open_ino_backtrace_fetched(inodeno_t ino, bufferlist& bl, int err) | |
8540 | { | |
8541 | dout(10) << "_open_ino_backtrace_fetched ino " << ino << " errno " << err << dendl; | |
8542 | ||
11fdf7f2 | 8543 | open_ino_info_t& info = opening_inodes.at(ino); |
7c673cae FG |
8544 | |
8545 | CInode *in = get_inode(ino); | |
8546 | if (in) { | |
8547 | dout(10) << " found cached " << *in << dendl; | |
8548 | open_ino_finish(ino, info, in->authority().first); | |
8549 | return; | |
8550 | } | |
8551 | ||
8552 | inode_backtrace_t backtrace; | |
8553 | if (err == 0) { | |
8554 | try { | |
11fdf7f2 | 8555 | decode(backtrace, bl); |
7c673cae FG |
8556 | } catch (const buffer::error &decode_exc) { |
8557 | derr << "corrupt backtrace on ino x0" << std::hex << ino | |
8558 | << std::dec << ": " << decode_exc << dendl; | |
8559 | open_ino_finish(ino, info, -EIO); | |
8560 | return; | |
8561 | } | |
8562 | if (backtrace.pool != info.pool && backtrace.pool != -1) { | |
8563 | dout(10) << " old object in pool " << info.pool | |
8564 | << ", retrying pool " << backtrace.pool << dendl; | |
8565 | info.pool = backtrace.pool; | |
8566 | C_IO_MDC_OpenInoBacktraceFetched *fin = | |
8567 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8568 | fetch_backtrace(ino, info.pool, fin->bl, | |
8569 | new C_OnFinisher(fin, mds->finisher)); | |
8570 | return; | |
8571 | } | |
8572 | } else if (err == -ENOENT) { | |
8573 | int64_t meta_pool = mds->mdsmap->get_metadata_pool(); | |
8574 | if (info.pool != meta_pool) { | |
8575 | dout(10) << " no object in pool " << info.pool | |
8576 | << ", retrying pool " << meta_pool << dendl; | |
8577 | info.pool = meta_pool; | |
8578 | C_IO_MDC_OpenInoBacktraceFetched *fin = | |
8579 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8580 | fetch_backtrace(ino, info.pool, fin->bl, | |
8581 | new C_OnFinisher(fin, mds->finisher)); | |
8582 | return; | |
8583 | } | |
8584 | err = 0; // backtrace.ancestors.empty() is checked below | |
8585 | } | |
8586 | ||
8587 | if (err == 0) { | |
8588 | if (backtrace.ancestors.empty()) { | |
8589 | dout(10) << " got empty backtrace " << dendl; | |
8590 | err = -EIO; | |
8591 | } else if (!info.ancestors.empty()) { | |
8592 | if (info.ancestors[0] == backtrace.ancestors[0]) { | |
8593 | dout(10) << " got same parents " << info.ancestors[0] << " 2 times" << dendl; | |
8594 | err = -EINVAL; | |
8595 | } else { | |
8596 | info.last_err = 0; | |
8597 | } | |
8598 | } | |
8599 | } | |
8600 | if (err) { | |
8601 | dout(0) << " failed to open ino " << ino << " err " << err << "/" << info.last_err << dendl; | |
8602 | if (info.last_err) | |
8603 | err = info.last_err; | |
8604 | open_ino_finish(ino, info, err); | |
8605 | return; | |
8606 | } | |
8607 | ||
8608 | dout(10) << " got backtrace " << backtrace << dendl; | |
8609 | info.ancestors = backtrace.ancestors; | |
8610 | ||
8611 | _open_ino_traverse_dir(ino, info, 0); | |
8612 | } | |
8613 | ||
8614 | void MDCache::_open_ino_parent_opened(inodeno_t ino, int ret) | |
8615 | { | |
8616 | dout(10) << "_open_ino_parent_opened ino " << ino << " ret " << ret << dendl; | |
8617 | ||
11fdf7f2 | 8618 | open_ino_info_t& info = opening_inodes.at(ino); |
7c673cae FG |
8619 | |
8620 | CInode *in = get_inode(ino); | |
8621 | if (in) { | |
8622 | dout(10) << " found cached " << *in << dendl; | |
8623 | open_ino_finish(ino, info, in->authority().first); | |
8624 | return; | |
8625 | } | |
8626 | ||
8627 | if (ret == mds->get_nodeid()) { | |
8628 | _open_ino_traverse_dir(ino, info, 0); | |
8629 | } else { | |
8630 | if (ret >= 0) { | |
8631 | mds_rank_t checked_rank = mds_rank_t(ret); | |
8632 | info.check_peers = true; | |
8633 | info.auth_hint = checked_rank; | |
8634 | info.checked.erase(checked_rank); | |
8635 | } | |
8636 | do_open_ino(ino, info, ret); | |
8637 | } | |
8638 | } | |
8639 | ||
8640 | void MDCache::_open_ino_traverse_dir(inodeno_t ino, open_ino_info_t& info, int ret) | |
8641 | { | |
8642 | dout(10) << __func__ << ": ino " << ino << " ret " << ret << dendl; | |
8643 | ||
8644 | CInode *in = get_inode(ino); | |
8645 | if (in) { | |
8646 | dout(10) << " found cached " << *in << dendl; | |
8647 | open_ino_finish(ino, info, in->authority().first); | |
8648 | return; | |
8649 | } | |
8650 | ||
8651 | if (ret) { | |
8652 | do_open_ino(ino, info, ret); | |
8653 | return; | |
8654 | } | |
8655 | ||
8656 | mds_rank_t hint = info.auth_hint; | |
8657 | ret = open_ino_traverse_dir(ino, NULL, info.ancestors, | |
8658 | info.discover, info.want_xlocked, &hint); | |
8659 | if (ret > 0) | |
8660 | return; | |
8661 | if (hint != mds->get_nodeid()) | |
8662 | info.auth_hint = hint; | |
8663 | do_open_ino(ino, info, ret); | |
8664 | } | |
8665 | ||
11fdf7f2 | 8666 | void MDCache::_open_ino_fetch_dir(inodeno_t ino, const MMDSOpenIno::const_ref &m, CDir *dir, bool parent) |
7c673cae FG |
8667 | { |
8668 | if (dir->state_test(CDir::STATE_REJOINUNDEF)) | |
11fdf7f2 | 8669 | ceph_assert(dir->get_inode()->dirfragtree.is_leaf(dir->get_frag())); |
7c673cae | 8670 | dir->fetch(new C_MDC_OpenInoTraverseDir(this, ino, m, parent)); |
11fdf7f2 TL |
8671 | if (mds->logger) |
8672 | mds->logger->inc(l_mds_openino_dir_fetch); | |
7c673cae FG |
8673 | } |
8674 | ||
11fdf7f2 TL |
8675 | int MDCache::open_ino_traverse_dir(inodeno_t ino, const MMDSOpenIno::const_ref &m, |
8676 | const vector<inode_backpointer_t>& ancestors, | |
7c673cae FG |
8677 | bool discover, bool want_xlocked, mds_rank_t *hint) |
8678 | { | |
8679 | dout(10) << "open_ino_traverse_dir ino " << ino << " " << ancestors << dendl; | |
8680 | int err = 0; | |
8681 | for (unsigned i = 0; i < ancestors.size(); i++) { | |
11fdf7f2 TL |
8682 | const auto& ancestor = ancestors.at(i); |
8683 | CInode *diri = get_inode(ancestor.dirino); | |
7c673cae FG |
8684 | |
8685 | if (!diri) { | |
11fdf7f2 TL |
8686 | if (discover && MDS_INO_IS_MDSDIR(ancestor.dirino)) { |
8687 | open_foreign_mdsdir(ancestor.dirino, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
7c673cae FG |
8688 | return 1; |
8689 | } | |
8690 | continue; | |
8691 | } | |
8692 | ||
8693 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) { | |
8694 | CDir *dir = diri->get_parent_dir(); | |
8695 | while (dir->state_test(CDir::STATE_REJOINUNDEF) && | |
8696 | dir->get_inode()->state_test(CInode::STATE_REJOINUNDEF)) | |
8697 | dir = dir->get_inode()->get_parent_dir(); | |
8698 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8699 | return 1; | |
8700 | } | |
8701 | ||
8702 | if (!diri->is_dir()) { | |
8703 | dout(10) << " " << *diri << " is not dir" << dendl; | |
8704 | if (i == 0) | |
8705 | err = -ENOTDIR; | |
8706 | break; | |
8707 | } | |
8708 | ||
11fdf7f2 | 8709 | const string& name = ancestor.dname; |
7c673cae FG |
8710 | frag_t fg = diri->pick_dirfrag(name); |
8711 | CDir *dir = diri->get_dirfrag(fg); | |
8712 | if (!dir) { | |
8713 | if (diri->is_auth()) { | |
8714 | if (diri->is_frozen()) { | |
8715 | dout(10) << " " << *diri << " is frozen, waiting " << dendl; | |
8716 | diri->add_waiter(CDir::WAIT_UNFREEZE, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8717 | return 1; | |
8718 | } | |
8719 | dir = diri->get_or_open_dirfrag(this, fg); | |
8720 | } else if (discover) { | |
8721 | open_remote_dirfrag(diri, fg, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8722 | return 1; | |
8723 | } | |
8724 | } | |
8725 | if (dir) { | |
11fdf7f2 | 8726 | inodeno_t next_ino = i > 0 ? ancestors.at(i-1).dirino : ino; |
7c673cae FG |
8727 | CDentry *dn = dir->lookup(name); |
8728 | CDentry::linkage_t *dnl = dn ? dn->get_linkage() : NULL; | |
8729 | if (dir->is_auth()) { | |
8730 | if (dnl && dnl->is_primary() && | |
8731 | dnl->get_inode()->state_test(CInode::STATE_REJOINUNDEF)) { | |
8732 | dout(10) << " fetching undef " << *dnl->get_inode() << dendl; | |
8733 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8734 | return 1; | |
8735 | } | |
8736 | ||
8737 | if (!dnl && !dir->is_complete() && | |
8738 | (!dir->has_bloom() || dir->is_in_bloom(name))) { | |
8739 | dout(10) << " fetching incomplete " << *dir << dendl; | |
8740 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8741 | return 1; | |
8742 | } | |
8743 | ||
8744 | dout(10) << " no ino " << next_ino << " in " << *dir << dendl; | |
8745 | if (i == 0) | |
8746 | err = -ENOENT; | |
8747 | } else if (discover) { | |
8748 | if (!dnl) { | |
8749 | filepath path(name, 0); | |
8750 | discover_path(dir, CEPH_NOSNAP, path, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0), | |
8751 | (i == 0 && want_xlocked)); | |
8752 | return 1; | |
8753 | } | |
8754 | if (dnl->is_null() && !dn->lock.can_read(-1)) { | |
8755 | dout(10) << " null " << *dn << " is not readable, waiting" << dendl; | |
8756 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8757 | return 1; | |
8758 | } | |
8759 | dout(10) << " no ino " << next_ino << " in " << *dir << dendl; | |
8760 | if (i == 0) | |
8761 | err = -ENOENT; | |
8762 | } | |
8763 | } | |
8764 | if (hint && i == 0) | |
8765 | *hint = dir ? dir->authority().first : diri->authority().first; | |
8766 | break; | |
8767 | } | |
8768 | return err; | |
8769 | } | |
8770 | ||
8771 | void MDCache::open_ino_finish(inodeno_t ino, open_ino_info_t& info, int ret) | |
8772 | { | |
8773 | dout(10) << "open_ino_finish ino " << ino << " ret " << ret << dendl; | |
8774 | ||
11fdf7f2 | 8775 | MDSContext::vec waiters; |
7c673cae FG |
8776 | waiters.swap(info.waiters); |
8777 | opening_inodes.erase(ino); | |
8778 | finish_contexts(g_ceph_context, waiters, ret); | |
8779 | } | |
8780 | ||
8781 | void MDCache::do_open_ino(inodeno_t ino, open_ino_info_t& info, int err) | |
8782 | { | |
8783 | if (err < 0 && err != -EAGAIN) { | |
8784 | info.checked.clear(); | |
7c673cae FG |
8785 | info.checking = MDS_RANK_NONE; |
8786 | info.check_peers = true; | |
8787 | info.fetch_backtrace = true; | |
8788 | if (info.discover) { | |
8789 | info.discover = false; | |
8790 | info.ancestors.clear(); | |
8791 | } | |
8792 | if (err != -ENOENT && err != -ENOTDIR) | |
8793 | info.last_err = err; | |
8794 | } | |
8795 | ||
d2e6a577 FG |
8796 | if (info.check_peers || info.discover) { |
8797 | if (info.discover) { | |
8798 | // got backtrace from peer, but failed to find inode. re-check peers | |
8799 | info.discover = false; | |
8800 | info.ancestors.clear(); | |
8801 | info.checked.clear(); | |
8802 | } | |
7c673cae FG |
8803 | info.check_peers = false; |
8804 | info.checking = MDS_RANK_NONE; | |
8805 | do_open_ino_peer(ino, info); | |
8806 | } else if (info.fetch_backtrace) { | |
8807 | info.check_peers = true; | |
8808 | info.fetch_backtrace = false; | |
8809 | info.checking = mds->get_nodeid(); | |
8810 | info.checked.clear(); | |
7c673cae FG |
8811 | C_IO_MDC_OpenInoBacktraceFetched *fin = |
8812 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8813 | fetch_backtrace(ino, info.pool, fin->bl, | |
8814 | new C_OnFinisher(fin, mds->finisher)); | |
8815 | } else { | |
11fdf7f2 | 8816 | ceph_assert(!info.ancestors.empty()); |
7c673cae FG |
8817 | info.checking = mds->get_nodeid(); |
8818 | open_ino(info.ancestors[0].dirino, mds->mdsmap->get_metadata_pool(), | |
8819 | new C_MDC_OpenInoParentOpened(this, ino), info.want_replica); | |
8820 | } | |
8821 | } | |
8822 | ||
8823 | void MDCache::do_open_ino_peer(inodeno_t ino, open_ino_info_t& info) | |
8824 | { | |
8825 | set<mds_rank_t> all, active; | |
8826 | mds->mdsmap->get_mds_set(all); | |
7c673cae | 8827 | if (mds->get_state() == MDSMap::STATE_REJOIN) |
1adf2230 AA |
8828 | mds->mdsmap->get_mds_set_lower_bound(active, MDSMap::STATE_REJOIN); |
8829 | else | |
8830 | mds->mdsmap->get_mds_set_lower_bound(active, MDSMap::STATE_CLIENTREPLAY); | |
7c673cae FG |
8831 | |
8832 | dout(10) << "do_open_ino_peer " << ino << " active " << active | |
8833 | << " all " << all << " checked " << info.checked << dendl; | |
8834 | ||
11fdf7f2 | 8835 | mds_rank_t whoami = mds->get_nodeid(); |
7c673cae | 8836 | mds_rank_t peer = MDS_RANK_NONE; |
11fdf7f2 | 8837 | if (info.auth_hint >= 0 && info.auth_hint != whoami) { |
7c673cae FG |
8838 | if (active.count(info.auth_hint)) { |
8839 | peer = info.auth_hint; | |
8840 | info.auth_hint = MDS_RANK_NONE; | |
8841 | } | |
8842 | } else { | |
8843 | for (set<mds_rank_t>::iterator p = active.begin(); p != active.end(); ++p) | |
11fdf7f2 | 8844 | if (*p != whoami && info.checked.count(*p) == 0) { |
7c673cae FG |
8845 | peer = *p; |
8846 | break; | |
8847 | } | |
8848 | } | |
8849 | if (peer < 0) { | |
11fdf7f2 | 8850 | all.erase(whoami); |
d2e6a577 | 8851 | if (all != info.checked) { |
7c673cae FG |
8852 | dout(10) << " waiting for more peers to be active" << dendl; |
8853 | } else { | |
8854 | dout(10) << " all MDS peers have been checked " << dendl; | |
8855 | do_open_ino(ino, info, 0); | |
8856 | } | |
8857 | } else { | |
8858 | info.checking = peer; | |
8859 | vector<inode_backpointer_t> *pa = NULL; | |
8860 | // got backtrace from peer or backtrace just fetched | |
8861 | if (info.discover || !info.fetch_backtrace) | |
8862 | pa = &info.ancestors; | |
11fdf7f2 TL |
8863 | mds->send_message_mds(MMDSOpenIno::create(info.tid, ino, pa), peer); |
8864 | if (mds->logger) | |
8865 | mds->logger->inc(l_mds_openino_peer_discover); | |
7c673cae FG |
8866 | } |
8867 | } | |
8868 | ||
11fdf7f2 | 8869 | void MDCache::handle_open_ino(const MMDSOpenIno::const_ref &m, int err) |
7c673cae FG |
8870 | { |
8871 | if (mds->get_state() < MDSMap::STATE_REJOIN && | |
8872 | mds->get_want_state() != CEPH_MDS_STATE_REJOIN) { | |
7c673cae FG |
8873 | return; |
8874 | } | |
8875 | ||
8876 | dout(10) << "handle_open_ino " << *m << " err " << err << dendl; | |
8877 | ||
11fdf7f2 | 8878 | auto from = mds_rank_t(m->get_source().num()); |
7c673cae | 8879 | inodeno_t ino = m->ino; |
11fdf7f2 | 8880 | MMDSOpenInoReply::ref reply; |
7c673cae FG |
8881 | CInode *in = get_inode(ino); |
8882 | if (in) { | |
8883 | dout(10) << " have " << *in << dendl; | |
11fdf7f2 | 8884 | reply = MMDSOpenInoReply::create(m->get_tid(), ino, mds_rank_t(0)); |
7c673cae FG |
8885 | if (in->is_auth()) { |
8886 | touch_inode(in); | |
8887 | while (1) { | |
8888 | CDentry *pdn = in->get_parent_dn(); | |
8889 | if (!pdn) | |
8890 | break; | |
8891 | CInode *diri = pdn->get_dir()->get_inode(); | |
94b18763 | 8892 | reply->ancestors.push_back(inode_backpointer_t(diri->ino(), pdn->get_name(), |
7c673cae FG |
8893 | in->inode.version)); |
8894 | in = diri; | |
8895 | } | |
8896 | } else { | |
8897 | reply->hint = in->authority().first; | |
8898 | } | |
8899 | } else if (err < 0) { | |
11fdf7f2 | 8900 | reply = MMDSOpenInoReply::create(m->get_tid(), ino, MDS_RANK_NONE, err); |
7c673cae FG |
8901 | } else { |
8902 | mds_rank_t hint = MDS_RANK_NONE; | |
8903 | int ret = open_ino_traverse_dir(ino, m, m->ancestors, false, false, &hint); | |
8904 | if (ret > 0) | |
8905 | return; | |
11fdf7f2 | 8906 | reply = MMDSOpenInoReply::create(m->get_tid(), ino, hint, ret); |
7c673cae | 8907 | } |
11fdf7f2 | 8908 | mds->send_message_mds(reply, from); |
7c673cae FG |
8909 | } |
8910 | ||
11fdf7f2 | 8911 | void MDCache::handle_open_ino_reply(const MMDSOpenInoReply::const_ref &m) |
7c673cae FG |
8912 | { |
8913 | dout(10) << "handle_open_ino_reply " << *m << dendl; | |
8914 | ||
8915 | inodeno_t ino = m->ino; | |
8916 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
8917 | auto it = opening_inodes.find(ino); | |
8918 | if (it != opening_inodes.end() && it->second.checking == from) { | |
8919 | open_ino_info_t& info = it->second; | |
8920 | info.checking = MDS_RANK_NONE; | |
8921 | info.checked.insert(from); | |
8922 | ||
8923 | CInode *in = get_inode(ino); | |
8924 | if (in) { | |
8925 | dout(10) << " found cached " << *in << dendl; | |
8926 | open_ino_finish(ino, info, in->authority().first); | |
8927 | } else if (!m->ancestors.empty()) { | |
8928 | dout(10) << " found ino " << ino << " on mds." << from << dendl; | |
8929 | if (!info.want_replica) { | |
8930 | open_ino_finish(ino, info, from); | |
7c673cae FG |
8931 | return; |
8932 | } | |
8933 | ||
8934 | info.ancestors = m->ancestors; | |
8935 | info.auth_hint = from; | |
8936 | info.checking = mds->get_nodeid(); | |
8937 | info.discover = true; | |
8938 | _open_ino_traverse_dir(ino, info, 0); | |
8939 | } else if (m->error) { | |
8940 | dout(10) << " error " << m->error << " from mds." << from << dendl; | |
8941 | do_open_ino(ino, info, m->error); | |
8942 | } else { | |
8943 | if (m->hint >= 0 && m->hint != mds->get_nodeid()) { | |
8944 | info.auth_hint = m->hint; | |
8945 | info.checked.erase(m->hint); | |
8946 | } | |
8947 | do_open_ino_peer(ino, info); | |
8948 | } | |
8949 | } | |
7c673cae FG |
8950 | } |
8951 | ||
8952 | void MDCache::kick_open_ino_peers(mds_rank_t who) | |
8953 | { | |
8954 | dout(10) << "kick_open_ino_peers mds." << who << dendl; | |
8955 | ||
8956 | for (map<inodeno_t, open_ino_info_t>::iterator p = opening_inodes.begin(); | |
8957 | p != opening_inodes.end(); | |
8958 | ++p) { | |
8959 | open_ino_info_t& info = p->second; | |
8960 | if (info.checking == who) { | |
8961 | dout(10) << " kicking ino " << p->first << " who was checking mds." << who << dendl; | |
8962 | info.checking = MDS_RANK_NONE; | |
8963 | do_open_ino_peer(p->first, info); | |
8964 | } else if (info.checking == MDS_RANK_NONE) { | |
8965 | dout(10) << " kicking ino " << p->first << " who was waiting" << dendl; | |
8966 | do_open_ino_peer(p->first, info); | |
8967 | } | |
8968 | } | |
8969 | } | |
8970 | ||
11fdf7f2 | 8971 | void MDCache::open_ino(inodeno_t ino, int64_t pool, MDSContext* fin, |
7c673cae FG |
8972 | bool want_replica, bool want_xlocked) |
8973 | { | |
8974 | dout(10) << "open_ino " << ino << " pool " << pool << " want_replica " | |
8975 | << want_replica << dendl; | |
8976 | ||
11fdf7f2 TL |
8977 | auto it = opening_inodes.find(ino); |
8978 | if (it != opening_inodes.end()) { | |
8979 | open_ino_info_t& info = it->second; | |
7c673cae FG |
8980 | if (want_replica) { |
8981 | info.want_replica = true; | |
8982 | if (want_xlocked && !info.want_xlocked) { | |
8983 | if (!info.ancestors.empty()) { | |
8984 | CInode *diri = get_inode(info.ancestors[0].dirino); | |
8985 | if (diri) { | |
8986 | frag_t fg = diri->pick_dirfrag(info.ancestors[0].dname); | |
8987 | CDir *dir = diri->get_dirfrag(fg); | |
8988 | if (dir && !dir->is_auth()) { | |
8989 | filepath path(info.ancestors[0].dname, 0); | |
8990 | discover_path(dir, CEPH_NOSNAP, path, NULL, true); | |
8991 | } | |
8992 | } | |
8993 | } | |
8994 | info.want_xlocked = true; | |
8995 | } | |
8996 | } | |
8997 | info.waiters.push_back(fin); | |
8998 | } else { | |
8999 | open_ino_info_t& info = opening_inodes[ino]; | |
7c673cae FG |
9000 | info.want_replica = want_replica; |
9001 | info.want_xlocked = want_xlocked; | |
9002 | info.tid = ++open_ino_last_tid; | |
9003 | info.pool = pool >= 0 ? pool : default_file_layout.pool_id; | |
9004 | info.waiters.push_back(fin); | |
11fdf7f2 TL |
9005 | if (mds->is_rejoin() && |
9006 | open_file_table.get_ancestors(ino, info.ancestors, info.auth_hint)) { | |
9007 | info.fetch_backtrace = false; | |
9008 | info.checking = mds->get_nodeid(); | |
9009 | _open_ino_traverse_dir(ino, info, 0); | |
9010 | } else { | |
9011 | do_open_ino(ino, info, 0); | |
9012 | } | |
7c673cae FG |
9013 | } |
9014 | } | |
9015 | ||
9016 | /* ---------------------------- */ | |
9017 | ||
9018 | /* | |
9019 | * search for a given inode on MDS peers. optionally start with the given node. | |
9020 | ||
9021 | ||
9022 | TODO | |
9023 | - recover from mds node failure, recovery | |
9024 | - traverse path | |
9025 | ||
9026 | */ | |
11fdf7f2 | 9027 | void MDCache::find_ino_peers(inodeno_t ino, MDSContext *c, mds_rank_t hint) |
7c673cae FG |
9028 | { |
9029 | dout(5) << "find_ino_peers " << ino << " hint " << hint << dendl; | |
b32b8144 FG |
9030 | CInode *in = get_inode(ino); |
9031 | if (in && in->state_test(CInode::STATE_PURGING)) { | |
9032 | c->complete(-ESTALE); | |
9033 | return; | |
9034 | } | |
11fdf7f2 | 9035 | ceph_assert(!in); |
7c673cae FG |
9036 | |
9037 | ceph_tid_t tid = ++find_ino_peer_last_tid; | |
9038 | find_ino_peer_info_t& fip = find_ino_peer[tid]; | |
9039 | fip.ino = ino; | |
9040 | fip.tid = tid; | |
9041 | fip.fin = c; | |
9042 | fip.hint = hint; | |
7c673cae FG |
9043 | _do_find_ino_peer(fip); |
9044 | } | |
9045 | ||
9046 | void MDCache::_do_find_ino_peer(find_ino_peer_info_t& fip) | |
9047 | { | |
9048 | set<mds_rank_t> all, active; | |
9049 | mds->mdsmap->get_mds_set(all); | |
1adf2230 | 9050 | mds->mdsmap->get_mds_set_lower_bound(active, MDSMap::STATE_CLIENTREPLAY); |
7c673cae FG |
9051 | |
9052 | dout(10) << "_do_find_ino_peer " << fip.tid << " " << fip.ino | |
9053 | << " active " << active << " all " << all | |
9054 | << " checked " << fip.checked | |
9055 | << dendl; | |
9056 | ||
9057 | mds_rank_t m = MDS_RANK_NONE; | |
9058 | if (fip.hint >= 0) { | |
9059 | m = fip.hint; | |
9060 | fip.hint = MDS_RANK_NONE; | |
9061 | } else { | |
9062 | for (set<mds_rank_t>::iterator p = active.begin(); p != active.end(); ++p) | |
9063 | if (*p != mds->get_nodeid() && | |
9064 | fip.checked.count(*p) == 0) { | |
9065 | m = *p; | |
9066 | break; | |
9067 | } | |
9068 | } | |
9069 | if (m == MDS_RANK_NONE) { | |
d2e6a577 FG |
9070 | all.erase(mds->get_nodeid()); |
9071 | if (all != fip.checked) { | |
7c673cae FG |
9072 | dout(10) << "_do_find_ino_peer waiting for more peers to be active" << dendl; |
9073 | } else { | |
9074 | dout(10) << "_do_find_ino_peer failed on " << fip.ino << dendl; | |
9075 | fip.fin->complete(-ESTALE); | |
9076 | find_ino_peer.erase(fip.tid); | |
9077 | } | |
9078 | } else { | |
9079 | fip.checking = m; | |
11fdf7f2 | 9080 | mds->send_message_mds(MMDSFindIno::create(fip.tid, fip.ino), m); |
7c673cae FG |
9081 | } |
9082 | } | |
9083 | ||
11fdf7f2 | 9084 | void MDCache::handle_find_ino(const MMDSFindIno::const_ref &m) |
7c673cae FG |
9085 | { |
9086 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
7c673cae FG |
9087 | return; |
9088 | } | |
9089 | ||
9090 | dout(10) << "handle_find_ino " << *m << dendl; | |
11fdf7f2 | 9091 | auto r = MMDSFindInoReply::create(m->tid); |
7c673cae FG |
9092 | CInode *in = get_inode(m->ino); |
9093 | if (in) { | |
9094 | in->make_path(r->path); | |
9095 | dout(10) << " have " << r->path << " " << *in << dendl; | |
9096 | } | |
11fdf7f2 | 9097 | mds->send_message_mds(r, mds_rank_t(m->get_source().num())); |
7c673cae FG |
9098 | } |
9099 | ||
9100 | ||
11fdf7f2 | 9101 | void MDCache::handle_find_ino_reply(const MMDSFindInoReply::const_ref &m) |
7c673cae FG |
9102 | { |
9103 | map<ceph_tid_t, find_ino_peer_info_t>::iterator p = find_ino_peer.find(m->tid); | |
9104 | if (p != find_ino_peer.end()) { | |
9105 | dout(10) << "handle_find_ino_reply " << *m << dendl; | |
9106 | find_ino_peer_info_t& fip = p->second; | |
9107 | ||
9108 | // success? | |
9109 | if (get_inode(fip.ino)) { | |
9110 | dout(10) << "handle_find_ino_reply successfully found " << fip.ino << dendl; | |
9111 | mds->queue_waiter(fip.fin); | |
9112 | find_ino_peer.erase(p); | |
7c673cae FG |
9113 | return; |
9114 | } | |
9115 | ||
9116 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
9117 | if (fip.checking == from) | |
9118 | fip.checking = MDS_RANK_NONE; | |
9119 | fip.checked.insert(from); | |
9120 | ||
9121 | if (!m->path.empty()) { | |
9122 | // we got a path! | |
9123 | vector<CDentry*> trace; | |
11fdf7f2 | 9124 | CF_MDS_RetryMessageFactory cf(mds, m); |
7c673cae | 9125 | MDRequestRef null_ref; |
11fdf7f2 | 9126 | int r = path_traverse(null_ref, cf, m->path, &trace, NULL, MDS_TRAVERSE_DISCOVER); |
7c673cae FG |
9127 | if (r > 0) |
9128 | return; | |
9129 | dout(0) << "handle_find_ino_reply failed with " << r << " on " << m->path | |
9130 | << ", retrying" << dendl; | |
9131 | fip.checked.clear(); | |
9132 | _do_find_ino_peer(fip); | |
9133 | } else { | |
9134 | // nope, continue. | |
9135 | _do_find_ino_peer(fip); | |
9136 | } | |
9137 | } else { | |
9138 | dout(10) << "handle_find_ino_reply tid " << m->tid << " dne" << dendl; | |
9139 | } | |
7c673cae FG |
9140 | } |
9141 | ||
9142 | void MDCache::kick_find_ino_peers(mds_rank_t who) | |
9143 | { | |
9144 | // find_ino_peers requests we should move on from | |
9145 | for (map<ceph_tid_t,find_ino_peer_info_t>::iterator p = find_ino_peer.begin(); | |
9146 | p != find_ino_peer.end(); | |
9147 | ++p) { | |
9148 | find_ino_peer_info_t& fip = p->second; | |
9149 | if (fip.checking == who) { | |
9150 | dout(10) << "kicking find_ino_peer " << fip.tid << " who was checking mds." << who << dendl; | |
9151 | fip.checking = MDS_RANK_NONE; | |
9152 | _do_find_ino_peer(fip); | |
9153 | } else if (fip.checking == MDS_RANK_NONE) { | |
9154 | dout(10) << "kicking find_ino_peer " << fip.tid << " who was waiting" << dendl; | |
9155 | _do_find_ino_peer(fip); | |
9156 | } | |
9157 | } | |
9158 | } | |
9159 | ||
9160 | /* ---------------------------- */ | |
9161 | ||
9162 | int MDCache::get_num_client_requests() | |
9163 | { | |
9164 | int count = 0; | |
9165 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
9166 | p != active_requests.end(); | |
9167 | ++p) { | |
9168 | MDRequestRef& mdr = p->second; | |
9169 | if (mdr->reqid.name.is_client() && !mdr->is_slave()) | |
9170 | count++; | |
9171 | } | |
9172 | return count; | |
9173 | } | |
9174 | ||
11fdf7f2 | 9175 | MDRequestRef MDCache::request_start(const MClientRequest::const_ref& req) |
7c673cae FG |
9176 | { |
9177 | // did we win a forward race against a slave? | |
9178 | if (active_requests.count(req->get_reqid())) { | |
9179 | MDRequestRef& mdr = active_requests[req->get_reqid()]; | |
11fdf7f2 | 9180 | ceph_assert(mdr); |
7c673cae FG |
9181 | if (mdr->is_slave()) { |
9182 | dout(10) << "request_start already had " << *mdr << ", waiting for finish" << dendl; | |
9183 | mdr->more()->waiting_for_finish.push_back(new C_MDS_RetryMessage(mds, req)); | |
9184 | } else { | |
9185 | dout(10) << "request_start already processing " << *mdr << ", dropping new msg" << dendl; | |
7c673cae FG |
9186 | } |
9187 | return MDRequestRef(); | |
9188 | } | |
9189 | ||
9190 | // register new client request | |
9191 | MDRequestImpl::Params params; | |
9192 | params.reqid = req->get_reqid(); | |
9193 | params.attempt = req->get_num_fwd(); | |
9194 | params.client_req = req; | |
9195 | params.initiated = req->get_recv_stamp(); | |
9196 | params.throttled = req->get_throttle_stamp(); | |
9197 | params.all_read = req->get_recv_complete_stamp(); | |
9198 | params.dispatched = req->get_dispatch_stamp(); | |
9199 | ||
9200 | MDRequestRef mdr = | |
11fdf7f2 | 9201 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params*>(¶ms); |
7c673cae FG |
9202 | active_requests[params.reqid] = mdr; |
9203 | mdr->set_op_stamp(req->get_stamp()); | |
9204 | dout(7) << "request_start " << *mdr << dendl; | |
9205 | return mdr; | |
9206 | } | |
9207 | ||
11fdf7f2 | 9208 | MDRequestRef MDCache::request_start_slave(metareqid_t ri, __u32 attempt, const Message::const_ref &m) |
7c673cae FG |
9209 | { |
9210 | int by = m->get_source().num(); | |
9211 | MDRequestImpl::Params params; | |
9212 | params.reqid = ri; | |
9213 | params.attempt = attempt; | |
9214 | params.triggering_slave_req = m; | |
9215 | params.slave_to = by; | |
9216 | params.initiated = m->get_recv_stamp(); | |
9217 | params.throttled = m->get_throttle_stamp(); | |
9218 | params.all_read = m->get_recv_complete_stamp(); | |
9219 | params.dispatched = m->get_dispatch_stamp(); | |
9220 | MDRequestRef mdr = | |
11fdf7f2 TL |
9221 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params*>(¶ms); |
9222 | ceph_assert(active_requests.count(mdr->reqid) == 0); | |
7c673cae FG |
9223 | active_requests[mdr->reqid] = mdr; |
9224 | dout(7) << "request_start_slave " << *mdr << " by mds." << by << dendl; | |
9225 | return mdr; | |
9226 | } | |
9227 | ||
9228 | MDRequestRef MDCache::request_start_internal(int op) | |
9229 | { | |
91327a77 | 9230 | utime_t now = ceph_clock_now(); |
7c673cae FG |
9231 | MDRequestImpl::Params params; |
9232 | params.reqid.name = entity_name_t::MDS(mds->get_nodeid()); | |
9233 | params.reqid.tid = mds->issue_tid(); | |
91327a77 AA |
9234 | params.initiated = now; |
9235 | params.throttled = now; | |
9236 | params.all_read = now; | |
9237 | params.dispatched = now; | |
7c673cae FG |
9238 | params.internal_op = op; |
9239 | MDRequestRef mdr = | |
11fdf7f2 | 9240 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params*>(¶ms); |
7c673cae | 9241 | |
11fdf7f2 | 9242 | ceph_assert(active_requests.count(mdr->reqid) == 0); |
7c673cae FG |
9243 | active_requests[mdr->reqid] = mdr; |
9244 | dout(7) << "request_start_internal " << *mdr << " op " << op << dendl; | |
9245 | return mdr; | |
9246 | } | |
9247 | ||
9248 | MDRequestRef MDCache::request_get(metareqid_t rid) | |
9249 | { | |
9250 | ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.find(rid); | |
11fdf7f2 | 9251 | ceph_assert(p != active_requests.end()); |
7c673cae FG |
9252 | dout(7) << "request_get " << rid << " " << *p->second << dendl; |
9253 | return p->second; | |
9254 | } | |
9255 | ||
9256 | void MDCache::request_finish(MDRequestRef& mdr) | |
9257 | { | |
9258 | dout(7) << "request_finish " << *mdr << dendl; | |
9259 | mdr->mark_event("finishing request"); | |
9260 | ||
9261 | // slave finisher? | |
9262 | if (mdr->has_more() && mdr->more()->slave_commit) { | |
9263 | Context *fin = mdr->more()->slave_commit; | |
9264 | mdr->more()->slave_commit = 0; | |
9265 | int ret; | |
9266 | if (mdr->aborted) { | |
9267 | mdr->aborted = false; | |
9268 | ret = -1; | |
9269 | mdr->more()->slave_rolling_back = true; | |
9270 | } else { | |
9271 | ret = 0; | |
9272 | mdr->committing = true; | |
9273 | } | |
9274 | fin->complete(ret); // this must re-call request_finish. | |
9275 | return; | |
9276 | } | |
9277 | ||
d2e6a577 FG |
9278 | switch(mdr->internal_op) { |
9279 | case CEPH_MDS_OP_FRAGMENTDIR: | |
9280 | logger->inc(l_mdss_ireq_fragmentdir); | |
9281 | break; | |
9282 | case CEPH_MDS_OP_EXPORTDIR: | |
9283 | logger->inc(l_mdss_ireq_exportdir); | |
9284 | break; | |
9285 | case CEPH_MDS_OP_ENQUEUE_SCRUB: | |
9286 | logger->inc(l_mdss_ireq_enqueue_scrub); | |
9287 | break; | |
9288 | case CEPH_MDS_OP_FLUSH: | |
9289 | logger->inc(l_mdss_ireq_flush); | |
9290 | break; | |
9291 | case CEPH_MDS_OP_REPAIR_FRAGSTATS: | |
9292 | logger->inc(l_mdss_ireq_fragstats); | |
9293 | break; | |
9294 | case CEPH_MDS_OP_REPAIR_INODESTATS: | |
9295 | logger->inc(l_mdss_ireq_inodestats); | |
9296 | break; | |
9297 | } | |
9298 | ||
7c673cae FG |
9299 | request_cleanup(mdr); |
9300 | } | |
9301 | ||
9302 | ||
9303 | void MDCache::request_forward(MDRequestRef& mdr, mds_rank_t who, int port) | |
9304 | { | |
9305 | mdr->mark_event("forwarding request"); | |
9306 | if (mdr->client_request && mdr->client_request->get_source().is_client()) { | |
9307 | dout(7) << "request_forward " << *mdr << " to mds." << who << " req " | |
9308 | << *mdr->client_request << dendl; | |
91327a77 | 9309 | mds->forward_message_mds(mdr->release_client_request(), who); |
7c673cae FG |
9310 | if (mds->logger) mds->logger->inc(l_mds_forward); |
9311 | } else if (mdr->internal_op >= 0) { | |
9312 | dout(10) << "request_forward on internal op; cancelling" << dendl; | |
9313 | mdr->internal_op_finish->complete(-EXDEV); | |
9314 | } else { | |
9315 | dout(7) << "request_forward drop " << *mdr << " req " << *mdr->client_request | |
9316 | << " was from mds" << dendl; | |
9317 | } | |
9318 | request_cleanup(mdr); | |
9319 | } | |
9320 | ||
9321 | ||
9322 | void MDCache::dispatch_request(MDRequestRef& mdr) | |
9323 | { | |
9324 | if (mdr->client_request) { | |
9325 | mds->server->dispatch_client_request(mdr); | |
9326 | } else if (mdr->slave_request) { | |
9327 | mds->server->dispatch_slave_request(mdr); | |
9328 | } else { | |
9329 | switch (mdr->internal_op) { | |
9330 | case CEPH_MDS_OP_FRAGMENTDIR: | |
9331 | dispatch_fragment_dir(mdr); | |
9332 | break; | |
9333 | case CEPH_MDS_OP_EXPORTDIR: | |
9334 | migrator->dispatch_export_dir(mdr, 0); | |
9335 | break; | |
9336 | case CEPH_MDS_OP_ENQUEUE_SCRUB: | |
9337 | enqueue_scrub_work(mdr); | |
9338 | break; | |
9339 | case CEPH_MDS_OP_FLUSH: | |
9340 | flush_dentry_work(mdr); | |
9341 | break; | |
9342 | case CEPH_MDS_OP_REPAIR_FRAGSTATS: | |
9343 | repair_dirfrag_stats_work(mdr); | |
9344 | break; | |
9345 | case CEPH_MDS_OP_REPAIR_INODESTATS: | |
9346 | repair_inode_stats_work(mdr); | |
9347 | break; | |
11fdf7f2 TL |
9348 | case CEPH_MDS_OP_UPGRADE_SNAPREALM: |
9349 | upgrade_inode_snaprealm_work(mdr); | |
9350 | break; | |
7c673cae FG |
9351 | default: |
9352 | ceph_abort(); | |
9353 | } | |
9354 | } | |
9355 | } | |
9356 | ||
9357 | ||
9358 | void MDCache::request_drop_foreign_locks(MDRequestRef& mdr) | |
9359 | { | |
9360 | if (!mdr->has_more()) | |
9361 | return; | |
9362 | ||
9363 | // clean up slaves | |
9364 | // (will implicitly drop remote dn pins) | |
9365 | for (set<mds_rank_t>::iterator p = mdr->more()->slaves.begin(); | |
9366 | p != mdr->more()->slaves.end(); | |
9367 | ++p) { | |
11fdf7f2 | 9368 | auto r = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, |
7c673cae FG |
9369 | MMDSSlaveRequest::OP_FINISH); |
9370 | ||
9371 | if (mdr->killed && !mdr->committing) { | |
9372 | r->mark_abort(); | |
9373 | } else if (mdr->more()->srcdn_auth_mds == *p && | |
9374 | mdr->more()->inode_import.length() > 0) { | |
9375 | // information about rename imported caps | |
9376 | r->inode_export.claim(mdr->more()->inode_import); | |
9377 | } | |
9378 | ||
9379 | mds->send_message_mds(r, *p); | |
9380 | } | |
9381 | ||
9382 | /* strip foreign xlocks out of lock lists, since the OP_FINISH drops them | |
9383 | * implicitly. Note that we don't call the finishers -- there shouldn't | |
9384 | * be any on a remote lock and the request finish wakes up all | |
9385 | * the waiters anyway! */ | |
7c673cae | 9386 | |
11fdf7f2 TL |
9387 | for (auto it = mdr->locks.begin(); it != mdr->locks.end(); ) { |
9388 | SimpleLock *lock = it->lock; | |
9389 | if (it->is_xlock() && !lock->get_parent()->is_auth()) { | |
9390 | dout(10) << "request_drop_foreign_locks forgetting lock " << *lock | |
9391 | << " on " << lock->get_parent() << dendl; | |
9392 | lock->put_xlock(); | |
9393 | mdr->locks.erase(it++); | |
9394 | } else if (it->is_remote_wrlock()) { | |
9395 | dout(10) << "request_drop_foreign_locks forgetting remote_wrlock " << *lock | |
9396 | << " on mds." << it->wrlock_target << " on " << *lock->get_parent() << dendl; | |
9397 | if (it->is_wrlock()) { | |
9398 | it->clear_remote_wrlock(); | |
9399 | ++it; | |
9400 | } else { | |
9401 | mdr->locks.erase(it++); | |
9402 | } | |
9403 | } else { | |
9404 | ++it; | |
9405 | } | |
7c673cae FG |
9406 | } |
9407 | ||
9408 | mdr->more()->slaves.clear(); /* we no longer have requests out to them, and | |
9409 | * leaving them in can cause double-notifies as | |
9410 | * this function can get called more than once */ | |
9411 | } | |
9412 | ||
9413 | void MDCache::request_drop_non_rdlocks(MDRequestRef& mdr) | |
9414 | { | |
9415 | request_drop_foreign_locks(mdr); | |
9416 | mds->locker->drop_non_rdlocks(mdr.get()); | |
9417 | } | |
9418 | ||
9419 | void MDCache::request_drop_locks(MDRequestRef& mdr) | |
9420 | { | |
9421 | request_drop_foreign_locks(mdr); | |
9422 | mds->locker->drop_locks(mdr.get()); | |
9423 | } | |
9424 | ||
9425 | void MDCache::request_cleanup(MDRequestRef& mdr) | |
9426 | { | |
9427 | dout(15) << "request_cleanup " << *mdr << dendl; | |
9428 | ||
9429 | if (mdr->has_more()) { | |
9430 | if (mdr->more()->is_ambiguous_auth) | |
9431 | mdr->clear_ambiguous_auth(); | |
9432 | if (!mdr->more()->waiting_for_finish.empty()) | |
9433 | mds->queue_waiters(mdr->more()->waiting_for_finish); | |
9434 | } | |
9435 | ||
9436 | request_drop_locks(mdr); | |
9437 | ||
9438 | // drop (local) auth pins | |
9439 | mdr->drop_local_auth_pins(); | |
9440 | ||
9441 | // drop stickydirs | |
11fdf7f2 | 9442 | mdr->put_stickydirs(); |
7c673cae FG |
9443 | |
9444 | mds->locker->kick_cap_releases(mdr); | |
9445 | ||
9446 | // drop cache pins | |
9447 | mdr->drop_pins(); | |
9448 | ||
9449 | // remove from session | |
9450 | mdr->item_session_request.remove_myself(); | |
9451 | ||
9452 | // remove from map | |
9453 | active_requests.erase(mdr->reqid); | |
9454 | ||
9455 | if (mds->logger) | |
9456 | log_stat(); | |
9457 | ||
9458 | mdr->mark_event("cleaned up request"); | |
9459 | } | |
9460 | ||
9461 | void MDCache::request_kill(MDRequestRef& mdr) | |
9462 | { | |
9463 | // rollback slave requests is tricky. just let the request proceed. | |
94b18763 | 9464 | if (mdr->has_more() && |
7c673cae | 9465 | (!mdr->more()->witnessed.empty() || !mdr->more()->waiting_on_slave.empty())) { |
94b18763 | 9466 | if (!mdr->done_locking) { |
11fdf7f2 | 9467 | ceph_assert(mdr->more()->witnessed.empty()); |
94b18763 FG |
9468 | mdr->aborted = true; |
9469 | dout(10) << "request_kill " << *mdr << " -- waiting for slave reply, delaying" << dendl; | |
9470 | } else { | |
9471 | dout(10) << "request_kill " << *mdr << " -- already started slave prep, no-op" << dendl; | |
9472 | } | |
7c673cae | 9473 | |
11fdf7f2 TL |
9474 | ceph_assert(mdr->used_prealloc_ino == 0); |
9475 | ceph_assert(mdr->prealloc_inos.empty()); | |
7c673cae FG |
9476 | |
9477 | mdr->session = NULL; | |
9478 | mdr->item_session_request.remove_myself(); | |
9479 | return; | |
9480 | } | |
9481 | ||
9482 | mdr->killed = true; | |
9483 | mdr->mark_event("killing request"); | |
9484 | ||
9485 | if (mdr->committing) { | |
9486 | dout(10) << "request_kill " << *mdr << " -- already committing, no-op" << dendl; | |
9487 | } else { | |
9488 | dout(10) << "request_kill " << *mdr << dendl; | |
9489 | request_cleanup(mdr); | |
9490 | } | |
9491 | } | |
9492 | ||
9493 | // ------------------------------------------------------------------------------- | |
9494 | // SNAPREALMS | |
9495 | ||
11fdf7f2 | 9496 | void MDCache::create_global_snaprealm() |
7c673cae | 9497 | { |
11fdf7f2 TL |
9498 | CInode *in = new CInode(this); // dummy inode |
9499 | create_unlinked_system_inode(in, MDS_INO_GLOBAL_SNAPREALM, S_IFDIR|0755); | |
9500 | add_inode(in); | |
9501 | global_snaprealm = in->snaprealm; | |
7c673cae FG |
9502 | } |
9503 | ||
11fdf7f2 | 9504 | void MDCache::do_realm_invalidate_and_update_notify(CInode *in, int snapop, bool notify_clients) |
7c673cae FG |
9505 | { |
9506 | dout(10) << "do_realm_invalidate_and_update_notify " << *in->snaprealm << " " << *in << dendl; | |
9507 | ||
9508 | vector<inodeno_t> split_inos; | |
9509 | vector<inodeno_t> split_realms; | |
9510 | ||
11fdf7f2 TL |
9511 | if (notify_clients) { |
9512 | ceph_assert(in->snaprealm->have_past_parents_open()); | |
9513 | if (snapop == CEPH_SNAP_OP_SPLIT) { | |
9514 | // notify clients of update|split | |
9515 | for (elist<CInode*>::iterator p = in->snaprealm->inodes_with_caps.begin(member_offset(CInode, item_caps)); | |
9516 | !p.end(); ++p) | |
9517 | split_inos.push_back((*p)->ino()); | |
7c673cae | 9518 | |
11fdf7f2 TL |
9519 | for (set<SnapRealm*>::iterator p = in->snaprealm->open_children.begin(); |
9520 | p != in->snaprealm->open_children.end(); | |
9521 | ++p) | |
9522 | split_realms.push_back((*p)->inode->ino()); | |
9523 | } | |
9524 | } | |
7c673cae FG |
9525 | |
9526 | set<SnapRealm*> past_children; | |
11fdf7f2 | 9527 | map<client_t, MClientSnap::ref> updates; |
7c673cae FG |
9528 | list<SnapRealm*> q; |
9529 | q.push_back(in->snaprealm); | |
9530 | while (!q.empty()) { | |
9531 | SnapRealm *realm = q.front(); | |
9532 | q.pop_front(); | |
9533 | ||
9534 | dout(10) << " realm " << *realm << " on " << *realm->inode << dendl; | |
9535 | realm->invalidate_cached_snaps(); | |
9536 | ||
11fdf7f2 TL |
9537 | if (notify_clients) { |
9538 | for (const auto& p : realm->client_caps) { | |
9539 | const auto& client = p.first; | |
9540 | const auto& caps = p.second; | |
9541 | ceph_assert(!caps->empty()); | |
9542 | ||
9543 | auto em = updates.emplace(std::piecewise_construct, std::forward_as_tuple(client), std::forward_as_tuple()); | |
9544 | if (em.second) { | |
9545 | auto update = MClientSnap::create(CEPH_SNAP_OP_SPLIT); | |
9546 | update->head.split = in->ino(); | |
9547 | update->split_inos = split_inos; | |
9548 | update->split_realms = split_realms; | |
9549 | update->bl = in->snaprealm->get_snap_trace(); | |
9550 | em.first->second = std::move(update); | |
9551 | } | |
7c673cae FG |
9552 | } |
9553 | } | |
9554 | ||
9555 | if (snapop == CEPH_SNAP_OP_UPDATE || snapop == CEPH_SNAP_OP_DESTROY) { | |
9556 | for (set<SnapRealm*>::iterator p = realm->open_past_children.begin(); | |
9557 | p != realm->open_past_children.end(); | |
9558 | ++p) | |
9559 | past_children.insert(*p); | |
9560 | } | |
9561 | ||
9562 | // notify for active children, too. | |
9563 | dout(10) << " " << realm << " open_children are " << realm->open_children << dendl; | |
9564 | for (set<SnapRealm*>::iterator p = realm->open_children.begin(); | |
9565 | p != realm->open_children.end(); | |
9566 | ++p) | |
9567 | q.push_back(*p); | |
9568 | } | |
9569 | ||
11fdf7f2 | 9570 | if (notify_clients) |
7c673cae FG |
9571 | send_snaps(updates); |
9572 | ||
9573 | // notify past children and their descendants if we update/delete old snapshots | |
9574 | for (set<SnapRealm*>::iterator p = past_children.begin(); | |
9575 | p != past_children.end(); | |
9576 | ++p) | |
9577 | q.push_back(*p); | |
9578 | ||
9579 | while (!q.empty()) { | |
9580 | SnapRealm *realm = q.front(); | |
9581 | q.pop_front(); | |
9582 | ||
9583 | realm->invalidate_cached_snaps(); | |
9584 | ||
9585 | for (set<SnapRealm*>::iterator p = realm->open_children.begin(); | |
9586 | p != realm->open_children.end(); | |
9587 | ++p) { | |
9588 | if (past_children.count(*p) == 0) | |
9589 | q.push_back(*p); | |
9590 | } | |
9591 | ||
9592 | for (set<SnapRealm*>::iterator p = realm->open_past_children.begin(); | |
9593 | p != realm->open_past_children.end(); | |
9594 | ++p) { | |
9595 | if (past_children.count(*p) == 0) { | |
9596 | q.push_back(*p); | |
9597 | past_children.insert(*p); | |
9598 | } | |
9599 | } | |
9600 | } | |
9601 | ||
9602 | if (snapop == CEPH_SNAP_OP_DESTROY) { | |
9603 | // eval stray inodes if we delete snapshot from their past ancestor snaprealm | |
9604 | for (set<SnapRealm*>::iterator p = past_children.begin(); | |
9605 | p != past_children.end(); | |
9606 | ++p) | |
9607 | maybe_eval_stray((*p)->inode, true); | |
9608 | } | |
9609 | } | |
9610 | ||
11fdf7f2 | 9611 | void MDCache::send_snap_update(CInode *in, version_t stid, int snap_op) |
7c673cae | 9612 | { |
11fdf7f2 TL |
9613 | dout(10) << __func__ << " " << *in << " stid " << stid << dendl; |
9614 | ceph_assert(in->is_auth()); | |
7c673cae | 9615 | |
11fdf7f2 TL |
9616 | set<mds_rank_t> mds_set; |
9617 | if (stid > 0) { | |
9618 | mds->mdsmap->get_mds_set_lower_bound(mds_set, MDSMap::STATE_RESOLVE); | |
9619 | mds_set.erase(mds->get_nodeid()); | |
9620 | } else { | |
9621 | in->list_replicas(mds_set); | |
9622 | } | |
7c673cae | 9623 | |
11fdf7f2 TL |
9624 | if (!mds_set.empty()) { |
9625 | bufferlist snap_blob; | |
9626 | in->encode_snap(snap_blob); | |
7c673cae | 9627 | |
11fdf7f2 TL |
9628 | for (auto p : mds_set) { |
9629 | auto m = MMDSSnapUpdate::create(in->ino(), stid, snap_op); | |
9630 | m->snap_blob = snap_blob; | |
9631 | mds->send_message_mds(m, p); | |
9632 | } | |
9633 | } | |
7c673cae | 9634 | |
11fdf7f2 TL |
9635 | if (stid > 0) |
9636 | notify_global_snaprealm_update(snap_op); | |
9637 | } | |
7c673cae | 9638 | |
11fdf7f2 TL |
9639 | void MDCache::handle_snap_update(const MMDSSnapUpdate::const_ref &m) |
9640 | { | |
9641 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
9642 | dout(10) << __func__ << " " << *m << " from mds." << from << dendl; | |
7c673cae | 9643 | |
11fdf7f2 TL |
9644 | if (mds->get_state() < MDSMap::STATE_RESOLVE && |
9645 | mds->get_want_state() != CEPH_MDS_STATE_RESOLVE) { | |
9646 | return; | |
9647 | } | |
7c673cae | 9648 | |
11fdf7f2 TL |
9649 | // null rejoin_done means open_snaprealms() has already been called |
9650 | bool notify_clients = mds->get_state() > MDSMap::STATE_REJOIN || | |
9651 | (mds->is_rejoin() && !rejoin_done); | |
9652 | ||
9653 | if (m->get_tid() > 0) { | |
9654 | mds->snapclient->notify_commit(m->get_tid()); | |
9655 | if (notify_clients) | |
9656 | notify_global_snaprealm_update(m->get_snap_op()); | |
9657 | } | |
9658 | ||
9659 | CInode *in = get_inode(m->get_ino()); | |
9660 | if (in) { | |
9661 | ceph_assert(!in->is_auth()); | |
9662 | if (mds->get_state() > MDSMap::STATE_REJOIN || | |
9663 | (mds->is_rejoin() && !in->is_rejoining())) { | |
9664 | auto p = m->snap_blob.cbegin(); | |
9665 | in->decode_snap(p); | |
9666 | ||
9667 | if (!notify_clients) { | |
9668 | if (!rejoin_pending_snaprealms.count(in)) { | |
9669 | in->get(CInode::PIN_OPENINGSNAPPARENTS); | |
9670 | rejoin_pending_snaprealms.insert(in); | |
9671 | } | |
9672 | } | |
9673 | do_realm_invalidate_and_update_notify(in, m->get_snap_op(), notify_clients); | |
9674 | } | |
9675 | } | |
7c673cae FG |
9676 | } |
9677 | ||
11fdf7f2 TL |
9678 | void MDCache::notify_global_snaprealm_update(int snap_op) |
9679 | { | |
9680 | if (snap_op != CEPH_SNAP_OP_DESTROY) | |
9681 | snap_op = CEPH_SNAP_OP_UPDATE; | |
9682 | set<Session*> sessions; | |
9683 | mds->sessionmap.get_client_session_set(sessions); | |
9684 | for (auto &session : sessions) { | |
9685 | if (!session->is_open() && !session->is_stale()) | |
9686 | continue; | |
9687 | auto update = MClientSnap::create(snap_op); | |
9688 | update->head.split = global_snaprealm->inode->ino(); | |
9689 | update->bl = global_snaprealm->get_snap_trace(); | |
9690 | mds->send_message_client_counted(update, session); | |
9691 | } | |
9692 | } | |
7c673cae FG |
9693 | |
9694 | // ------------------------------------------------------------------------------- | |
9695 | // STRAYS | |
9696 | ||
9697 | struct C_MDC_RetryScanStray : public MDCacheContext { | |
9698 | dirfrag_t next; | |
9699 | C_MDC_RetryScanStray(MDCache *c, dirfrag_t n) : MDCacheContext(c), next(n) { } | |
9700 | void finish(int r) override { | |
9701 | mdcache->scan_stray_dir(next); | |
9702 | } | |
9703 | }; | |
9704 | ||
9705 | void MDCache::scan_stray_dir(dirfrag_t next) | |
9706 | { | |
9707 | dout(10) << "scan_stray_dir " << next << dendl; | |
9708 | ||
9709 | list<CDir*> ls; | |
9710 | for (int i = 0; i < NUM_STRAY; ++i) { | |
9711 | if (strays[i]->ino() < next.ino) | |
9712 | continue; | |
9713 | strays[i]->get_dirfrags(ls); | |
9714 | } | |
9715 | ||
9716 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
9717 | CDir *dir = *p; | |
9718 | if (dir->dirfrag() < next) | |
9719 | continue; | |
9720 | if (!dir->is_complete()) { | |
9721 | dir->fetch(new C_MDC_RetryScanStray(this, dir->dirfrag())); | |
9722 | return; | |
9723 | } | |
94b18763 FG |
9724 | for (auto &p : dir->items) { |
9725 | CDentry *dn = p.second; | |
7c673cae FG |
9726 | dn->state_set(CDentry::STATE_STRAY); |
9727 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
9728 | if (dnl->is_primary()) { | |
9729 | CInode *in = dnl->get_inode(); | |
9730 | if (in->inode.nlink == 0) | |
9731 | in->state_set(CInode::STATE_ORPHAN); | |
9732 | maybe_eval_stray(in); | |
9733 | } | |
9734 | } | |
9735 | } | |
9736 | } | |
9737 | ||
7c673cae FG |
9738 | void MDCache::fetch_backtrace(inodeno_t ino, int64_t pool, bufferlist& bl, Context *fin) |
9739 | { | |
9740 | object_t oid = CInode::get_object_name(ino, frag_t(), ""); | |
9741 | mds->objecter->getxattr(oid, object_locator_t(pool), "parent", CEPH_NOSNAP, &bl, 0, fin); | |
11fdf7f2 TL |
9742 | if (mds->logger) |
9743 | mds->logger->inc(l_mds_openino_backtrace_fetch); | |
7c673cae FG |
9744 | } |
9745 | ||
9746 | ||
9747 | ||
9748 | ||
9749 | ||
9750 | // ======================================================================================== | |
9751 | // DISCOVER | |
9752 | /* | |
9753 | ||
9754 | - for all discovers (except base_inos, e.g. root, stray), waiters are attached | |
9755 | to the parent metadata object in the cache (pinning it). | |
9756 | ||
9757 | - all discovers are tracked by tid, so that we can ignore potentially dup replies. | |
9758 | ||
9759 | */ | |
9760 | ||
9761 | void MDCache::_send_discover(discover_info_t& d) | |
9762 | { | |
11fdf7f2 | 9763 | auto dis = MDiscover::create(d.ino, d.frag, d.snap, d.want_path, d.want_base_dir, d.want_xlocked); |
7c673cae FG |
9764 | dis->set_tid(d.tid); |
9765 | mds->send_message_mds(dis, d.mds); | |
9766 | } | |
9767 | ||
9768 | void MDCache::discover_base_ino(inodeno_t want_ino, | |
11fdf7f2 | 9769 | MDSContext *onfinish, |
7c673cae FG |
9770 | mds_rank_t from) |
9771 | { | |
9772 | dout(7) << "discover_base_ino " << want_ino << " from mds." << from << dendl; | |
9773 | if (waiting_for_base_ino[from].count(want_ino) == 0) { | |
9774 | discover_info_t& d = _create_discover(from); | |
9775 | d.ino = want_ino; | |
9776 | _send_discover(d); | |
9777 | } | |
9778 | waiting_for_base_ino[from][want_ino].push_back(onfinish); | |
9779 | } | |
9780 | ||
9781 | ||
9782 | void MDCache::discover_dir_frag(CInode *base, | |
9783 | frag_t approx_fg, | |
11fdf7f2 | 9784 | MDSContext *onfinish, |
7c673cae FG |
9785 | mds_rank_t from) |
9786 | { | |
9787 | if (from < 0) | |
9788 | from = base->authority().first; | |
9789 | ||
9790 | dirfrag_t df(base->ino(), approx_fg); | |
9791 | dout(7) << "discover_dir_frag " << df | |
9792 | << " from mds." << from << dendl; | |
9793 | ||
9794 | if (!base->is_waiting_for_dir(approx_fg) || !onfinish) { | |
9795 | discover_info_t& d = _create_discover(from); | |
9796 | d.pin_base(base); | |
9797 | d.ino = base->ino(); | |
9798 | d.frag = approx_fg; | |
9799 | d.want_base_dir = true; | |
9800 | _send_discover(d); | |
9801 | } | |
9802 | ||
9803 | if (onfinish) | |
9804 | base->add_dir_waiter(approx_fg, onfinish); | |
9805 | } | |
9806 | ||
9807 | struct C_MDC_RetryDiscoverPath : public MDCacheContext { | |
9808 | CInode *base; | |
9809 | snapid_t snapid; | |
9810 | filepath path; | |
9811 | mds_rank_t from; | |
9812 | C_MDC_RetryDiscoverPath(MDCache *c, CInode *b, snapid_t s, filepath &p, mds_rank_t f) : | |
9813 | MDCacheContext(c), base(b), snapid(s), path(p), from(f) {} | |
9814 | void finish(int r) override { | |
9815 | mdcache->discover_path(base, snapid, path, 0, from); | |
9816 | } | |
9817 | }; | |
9818 | ||
9819 | void MDCache::discover_path(CInode *base, | |
9820 | snapid_t snap, | |
9821 | filepath want_path, | |
11fdf7f2 | 9822 | MDSContext *onfinish, |
7c673cae FG |
9823 | bool want_xlocked, |
9824 | mds_rank_t from) | |
9825 | { | |
9826 | if (from < 0) | |
9827 | from = base->authority().first; | |
9828 | ||
9829 | dout(7) << "discover_path " << base->ino() << " " << want_path << " snap " << snap << " from mds." << from | |
9830 | << (want_xlocked ? " want_xlocked":"") | |
9831 | << dendl; | |
9832 | ||
9833 | if (base->is_ambiguous_auth()) { | |
9834 | dout(10) << " waiting for single auth on " << *base << dendl; | |
9835 | if (!onfinish) | |
9836 | onfinish = new C_MDC_RetryDiscoverPath(this, base, snap, want_path, from); | |
9837 | base->add_waiter(CInode::WAIT_SINGLEAUTH, onfinish); | |
9838 | return; | |
9839 | } else if (from == mds->get_nodeid()) { | |
11fdf7f2 | 9840 | MDSContext::vec finished; |
7c673cae FG |
9841 | base->take_waiting(CInode::WAIT_DIR, finished); |
9842 | mds->queue_waiters(finished); | |
9843 | return; | |
9844 | } | |
9845 | ||
9846 | frag_t fg = base->pick_dirfrag(want_path[0]); | |
9847 | if ((want_xlocked && want_path.depth() == 1) || | |
9848 | !base->is_waiting_for_dir(fg) || !onfinish) { | |
9849 | discover_info_t& d = _create_discover(from); | |
9850 | d.ino = base->ino(); | |
9851 | d.pin_base(base); | |
9852 | d.frag = fg; | |
9853 | d.snap = snap; | |
9854 | d.want_path = want_path; | |
9855 | d.want_base_dir = true; | |
9856 | d.want_xlocked = want_xlocked; | |
9857 | _send_discover(d); | |
9858 | } | |
9859 | ||
9860 | // register + wait | |
9861 | if (onfinish) | |
9862 | base->add_dir_waiter(fg, onfinish); | |
9863 | } | |
9864 | ||
9865 | struct C_MDC_RetryDiscoverPath2 : public MDCacheContext { | |
9866 | CDir *base; | |
9867 | snapid_t snapid; | |
9868 | filepath path; | |
9869 | C_MDC_RetryDiscoverPath2(MDCache *c, CDir *b, snapid_t s, filepath &p) : | |
9870 | MDCacheContext(c), base(b), snapid(s), path(p) {} | |
9871 | void finish(int r) override { | |
9872 | mdcache->discover_path(base, snapid, path, 0); | |
9873 | } | |
9874 | }; | |
9875 | ||
9876 | void MDCache::discover_path(CDir *base, | |
9877 | snapid_t snap, | |
9878 | filepath want_path, | |
11fdf7f2 | 9879 | MDSContext *onfinish, |
7c673cae FG |
9880 | bool want_xlocked) |
9881 | { | |
9882 | mds_rank_t from = base->authority().first; | |
9883 | ||
9884 | dout(7) << "discover_path " << base->dirfrag() << " " << want_path << " snap " << snap << " from mds." << from | |
9885 | << (want_xlocked ? " want_xlocked":"") | |
9886 | << dendl; | |
9887 | ||
9888 | if (base->is_ambiguous_auth()) { | |
9889 | dout(7) << " waiting for single auth on " << *base << dendl; | |
9890 | if (!onfinish) | |
9891 | onfinish = new C_MDC_RetryDiscoverPath2(this, base, snap, want_path); | |
9892 | base->add_waiter(CDir::WAIT_SINGLEAUTH, onfinish); | |
9893 | return; | |
9894 | } else if (from == mds->get_nodeid()) { | |
11fdf7f2 | 9895 | MDSContext::vec finished; |
7c673cae FG |
9896 | base->take_sub_waiting(finished); |
9897 | mds->queue_waiters(finished); | |
9898 | return; | |
9899 | } | |
9900 | ||
9901 | if ((want_xlocked && want_path.depth() == 1) || | |
9902 | !base->is_waiting_for_dentry(want_path[0].c_str(), snap) || !onfinish) { | |
9903 | discover_info_t& d = _create_discover(from); | |
9904 | d.ino = base->ino(); | |
31f18b77 | 9905 | d.pin_base(base->inode); |
7c673cae FG |
9906 | d.frag = base->get_frag(); |
9907 | d.snap = snap; | |
9908 | d.want_path = want_path; | |
9909 | d.want_base_dir = false; | |
9910 | d.want_xlocked = want_xlocked; | |
9911 | _send_discover(d); | |
9912 | } | |
9913 | ||
9914 | // register + wait | |
9915 | if (onfinish) | |
9916 | base->add_dentry_waiter(want_path[0], snap, onfinish); | |
9917 | } | |
9918 | ||
9919 | void MDCache::kick_discovers(mds_rank_t who) | |
9920 | { | |
9921 | for (map<ceph_tid_t,discover_info_t>::iterator p = discovers.begin(); | |
9922 | p != discovers.end(); | |
9923 | ++p) { | |
9924 | if (p->second.mds != who) | |
9925 | continue; | |
9926 | _send_discover(p->second); | |
9927 | } | |
9928 | } | |
9929 | ||
9930 | ||
11fdf7f2 | 9931 | void MDCache::handle_discover(const MDiscover::const_ref &dis) |
7c673cae FG |
9932 | { |
9933 | mds_rank_t whoami = mds->get_nodeid(); | |
9934 | mds_rank_t from = mds_rank_t(dis->get_source().num()); | |
9935 | ||
11fdf7f2 | 9936 | ceph_assert(from != whoami); |
7c673cae FG |
9937 | |
9938 | if (mds->get_state() <= MDSMap::STATE_REJOIN) { | |
9939 | if (mds->get_state() < MDSMap::STATE_REJOIN && | |
d2e6a577 | 9940 | mds->get_want_state() < CEPH_MDS_STATE_REJOIN) { |
7c673cae FG |
9941 | return; |
9942 | } | |
9943 | ||
9944 | // proceed if requester is in the REJOIN stage, the request is from parallel_fetch(). | |
9945 | // delay processing request from survivor because we may not yet choose lock states. | |
9946 | if (!mds->mdsmap->is_rejoin(from)) { | |
9947 | dout(0) << "discover_reply not yet active(|still rejoining), delaying" << dendl; | |
9948 | mds->wait_for_replay(new C_MDS_RetryMessage(mds, dis)); | |
9949 | return; | |
9950 | } | |
9951 | } | |
9952 | ||
9953 | ||
9954 | CInode *cur = 0; | |
11fdf7f2 | 9955 | auto reply = MDiscoverReply::create(*dis); |
7c673cae FG |
9956 | |
9957 | snapid_t snapid = dis->get_snapid(); | |
9958 | ||
9959 | // get started. | |
9960 | if (MDS_INO_IS_BASE(dis->get_base_ino()) && | |
9961 | !dis->wants_base_dir() && dis->get_want().depth() == 0) { | |
9962 | // wants root | |
9963 | dout(7) << "handle_discover from mds." << from | |
9964 | << " wants base + " << dis->get_want().get_path() | |
9965 | << " snap " << snapid | |
9966 | << dendl; | |
9967 | ||
9968 | cur = get_inode(dis->get_base_ino()); | |
11fdf7f2 | 9969 | ceph_assert(cur); |
7c673cae FG |
9970 | |
9971 | // add root | |
9972 | reply->starts_with = MDiscoverReply::INODE; | |
9973 | replicate_inode(cur, from, reply->trace, mds->mdsmap->get_up_features()); | |
9974 | dout(10) << "added base " << *cur << dendl; | |
9975 | } | |
9976 | else { | |
9977 | // there's a base inode | |
9978 | cur = get_inode(dis->get_base_ino(), snapid); | |
9979 | if (!cur && snapid != CEPH_NOSNAP) { | |
9980 | cur = get_inode(dis->get_base_ino()); | |
9981 | if (cur && !cur->is_multiversion()) | |
9982 | cur = NULL; // nope! | |
9983 | } | |
9984 | ||
9985 | if (!cur) { | |
9986 | dout(7) << "handle_discover mds." << from | |
9987 | << " don't have base ino " << dis->get_base_ino() << "." << snapid | |
9988 | << dendl; | |
9989 | if (!dis->wants_base_dir() && dis->get_want().depth() > 0) | |
9990 | reply->set_error_dentry(dis->get_dentry(0)); | |
9991 | reply->set_flag_error_dir(); | |
9992 | } else if (dis->wants_base_dir()) { | |
9993 | dout(7) << "handle_discover mds." << from | |
9994 | << " wants basedir+" << dis->get_want().get_path() | |
9995 | << " has " << *cur | |
9996 | << dendl; | |
9997 | } else { | |
9998 | dout(7) << "handle_discover mds." << from | |
9999 | << " wants " << dis->get_want().get_path() | |
10000 | << " has " << *cur | |
10001 | << dendl; | |
10002 | } | |
10003 | } | |
10004 | ||
11fdf7f2 | 10005 | ceph_assert(reply); |
7c673cae FG |
10006 | |
10007 | // add content | |
10008 | // do some fidgeting to include a dir if they asked for the base dir, or just root. | |
10009 | for (unsigned i = 0; | |
10010 | cur && (i < dis->get_want().depth() || dis->get_want().depth() == 0); | |
10011 | i++) { | |
10012 | ||
10013 | // -- figure out the dir | |
10014 | ||
10015 | // is *cur even a dir at all? | |
10016 | if (!cur->is_dir()) { | |
10017 | dout(7) << *cur << " not a dir" << dendl; | |
10018 | reply->set_flag_error_dir(); | |
10019 | break; | |
10020 | } | |
10021 | ||
10022 | // pick frag | |
10023 | frag_t fg; | |
10024 | if (dis->get_want().depth()) { | |
10025 | // dentry specifies | |
10026 | fg = cur->pick_dirfrag(dis->get_dentry(i)); | |
10027 | } else { | |
10028 | // requester explicity specified the frag | |
11fdf7f2 | 10029 | ceph_assert(dis->wants_base_dir() || MDS_INO_IS_BASE(dis->get_base_ino())); |
7c673cae FG |
10030 | fg = dis->get_base_dir_frag(); |
10031 | if (!cur->dirfragtree.is_leaf(fg)) | |
10032 | fg = cur->dirfragtree[fg.value()]; | |
10033 | } | |
10034 | CDir *curdir = cur->get_dirfrag(fg); | |
10035 | ||
10036 | if ((!curdir && !cur->is_auth()) || | |
10037 | (curdir && !curdir->is_auth())) { | |
10038 | ||
10039 | /* before: | |
10040 | * ONLY set flag if empty!! | |
10041 | * otherwise requester will wake up waiter(s) _and_ continue with discover, | |
10042 | * resulting in duplicate discovers in flight, | |
10043 | * which can wreak havoc when discovering rename srcdn (which may move) | |
10044 | */ | |
10045 | ||
10046 | if (reply->is_empty()) { | |
10047 | // only hint if empty. | |
10048 | // someday this could be better, but right now the waiter logic isn't smart enough. | |
10049 | ||
10050 | // hint | |
10051 | if (curdir) { | |
10052 | dout(7) << " not dirfrag auth, setting dir_auth_hint for " << *curdir << dendl; | |
10053 | reply->set_dir_auth_hint(curdir->authority().first); | |
10054 | } else { | |
10055 | dout(7) << " dirfrag not open, not inode auth, setting dir_auth_hint for " | |
10056 | << *cur << dendl; | |
10057 | reply->set_dir_auth_hint(cur->authority().first); | |
10058 | } | |
10059 | ||
10060 | // note error dentry, if any | |
10061 | // NOTE: important, as it allows requester to issue an equivalent discover | |
10062 | // to whomever we hint at. | |
10063 | if (dis->get_want().depth() > i) | |
10064 | reply->set_error_dentry(dis->get_dentry(i)); | |
10065 | } | |
10066 | ||
10067 | break; | |
10068 | } | |
10069 | ||
10070 | if (!curdir) { // open dir? | |
10071 | if (cur->is_frozen()) { | |
10072 | if (!reply->is_empty()) { | |
10073 | dout(7) << *cur << " is frozen, non-empty reply, stopping" << dendl; | |
10074 | break; | |
10075 | } | |
10076 | dout(7) << *cur << " is frozen, empty reply, waiting" << dendl; | |
10077 | cur->add_waiter(CInode::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10078 | return; |
10079 | } | |
10080 | curdir = cur->get_or_open_dirfrag(this, fg); | |
10081 | } else if (curdir->is_frozen_tree() || | |
10082 | (curdir->is_frozen_dir() && fragment_are_all_frozen(curdir))) { | |
31f18b77 FG |
10083 | if (!reply->is_empty()) { |
10084 | dout(7) << *curdir << " is frozen, non-empty reply, stopping" << dendl; | |
10085 | break; | |
10086 | } | |
7c673cae FG |
10087 | if (dis->wants_base_dir() && dis->get_base_dir_frag() != curdir->get_frag()) { |
10088 | dout(7) << *curdir << " is frozen, dirfrag mismatch, stopping" << dendl; | |
10089 | reply->set_flag_error_dir(); | |
10090 | break; | |
10091 | } | |
7c673cae FG |
10092 | dout(7) << *curdir << " is frozen, empty reply, waiting" << dendl; |
10093 | curdir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10094 | return; |
10095 | } | |
10096 | ||
10097 | // add dir | |
10098 | if (curdir->get_version() == 0) { | |
10099 | // fetch newly opened dir | |
10100 | } else if (reply->is_empty() && !dis->wants_base_dir()) { | |
10101 | dout(7) << "handle_discover not adding unwanted base dir " << *curdir << dendl; | |
10102 | // make sure the base frag is correct, though, in there was a refragment since the | |
10103 | // original request was sent. | |
10104 | reply->set_base_dir_frag(curdir->get_frag()); | |
10105 | } else { | |
11fdf7f2 | 10106 | ceph_assert(!curdir->is_ambiguous_auth()); // would be frozen. |
7c673cae FG |
10107 | if (!reply->trace.length()) |
10108 | reply->starts_with = MDiscoverReply::DIR; | |
10109 | replicate_dir(curdir, from, reply->trace); | |
10110 | dout(7) << "handle_discover added dir " << *curdir << dendl; | |
10111 | } | |
10112 | ||
10113 | // lookup | |
10114 | CDentry *dn = 0; | |
10115 | if (curdir->get_version() == 0) { | |
10116 | // fetch newly opened dir | |
11fdf7f2 | 10117 | ceph_assert(!curdir->has_bloom()); |
7c673cae FG |
10118 | } else if (dis->get_want().depth() > 0) { |
10119 | // lookup dentry | |
10120 | dn = curdir->lookup(dis->get_dentry(i), snapid); | |
10121 | } else | |
10122 | break; // done! | |
10123 | ||
10124 | // incomplete dir? | |
10125 | if (!dn) { | |
31f18b77 | 10126 | if (!curdir->is_complete() && |
11fdf7f2 TL |
10127 | !(snapid == CEPH_NOSNAP && |
10128 | curdir->has_bloom() && | |
10129 | !curdir->is_in_bloom(dis->get_dentry(i)))) { | |
7c673cae FG |
10130 | // readdir |
10131 | dout(7) << "incomplete dir contents for " << *curdir << ", fetching" << dendl; | |
10132 | if (reply->is_empty()) { | |
10133 | // fetch and wait | |
10134 | curdir->fetch(new C_MDS_RetryMessage(mds, dis), | |
10135 | dis->wants_base_dir() && curdir->get_version() == 0); | |
7c673cae FG |
10136 | return; |
10137 | } else { | |
10138 | // initiate fetch, but send what we have so far | |
10139 | curdir->fetch(0); | |
10140 | break; | |
10141 | } | |
10142 | } | |
10143 | ||
11fdf7f2 TL |
10144 | if (snapid != CEPH_NOSNAP && !reply->is_empty()) { |
10145 | dout(7) << "dentry " << dis->get_dentry(i) << " snap " << snapid | |
10146 | << " dne, non-empty reply, stopping" << dendl; | |
10147 | break; | |
10148 | } | |
10149 | ||
7c673cae FG |
10150 | // send null dentry |
10151 | dout(7) << "dentry " << dis->get_dentry(i) << " dne, returning null in " | |
10152 | << *curdir << dendl; | |
11fdf7f2 TL |
10153 | if (snapid == CEPH_NOSNAP) |
10154 | dn = curdir->add_null_dentry(dis->get_dentry(i)); | |
10155 | else | |
10156 | dn = curdir->add_null_dentry(dis->get_dentry(i), snapid, snapid); | |
7c673cae | 10157 | } |
11fdf7f2 | 10158 | ceph_assert(dn); |
7c673cae | 10159 | |
31f18b77 FG |
10160 | // don't add replica to purging dentry/inode |
10161 | if (dn->state_test(CDentry::STATE_PURGING)) { | |
10162 | if (reply->is_empty()) | |
10163 | reply->set_flag_error_dn(dis->get_dentry(i)); | |
10164 | break; | |
10165 | } | |
10166 | ||
7c673cae FG |
10167 | CDentry::linkage_t *dnl = dn->get_linkage(); |
10168 | ||
10169 | // xlocked dentry? | |
10170 | // ...always block on non-tail items (they are unrelated) | |
10171 | // ...allow xlocked tail disocvery _only_ if explicitly requested | |
10172 | bool tailitem = (dis->get_want().depth() == 0) || (i == dis->get_want().depth() - 1); | |
10173 | if (dn->lock.is_xlocked()) { | |
10174 | // is this the last (tail) item in the discover traversal? | |
10175 | if (tailitem && dis->wants_xlocked()) { | |
10176 | dout(7) << "handle_discover allowing discovery of xlocked tail " << *dn << dendl; | |
10177 | } else if (reply->is_empty()) { | |
10178 | dout(7) << "handle_discover blocking on xlocked " << *dn << dendl; | |
10179 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10180 | return; |
10181 | } else { | |
10182 | dout(7) << "handle_discover non-empty reply, xlocked tail " << *dn << dendl; | |
10183 | break; | |
10184 | } | |
10185 | } | |
10186 | ||
10187 | // frozen inode? | |
10188 | if (dnl->is_primary() && dnl->get_inode()->is_frozen_inode()) { | |
10189 | if (tailitem && dis->wants_xlocked()) { | |
10190 | dout(7) << "handle_discover allowing discovery of frozen tail " << *dnl->get_inode() << dendl; | |
10191 | } else if (reply->is_empty()) { | |
10192 | dout(7) << *dnl->get_inode() << " is frozen, empty reply, waiting" << dendl; | |
10193 | dnl->get_inode()->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
7c673cae FG |
10194 | return; |
10195 | } else { | |
10196 | dout(7) << *dnl->get_inode() << " is frozen, non-empty reply, stopping" << dendl; | |
10197 | break; | |
10198 | } | |
10199 | } | |
10200 | ||
10201 | // add dentry | |
10202 | if (!reply->trace.length()) | |
10203 | reply->starts_with = MDiscoverReply::DENTRY; | |
10204 | replicate_dentry(dn, from, reply->trace); | |
10205 | dout(7) << "handle_discover added dentry " << *dn << dendl; | |
10206 | ||
10207 | if (!dnl->is_primary()) break; // stop on null or remote link. | |
10208 | ||
10209 | // add inode | |
10210 | CInode *next = dnl->get_inode(); | |
11fdf7f2 | 10211 | ceph_assert(next->is_auth()); |
7c673cae FG |
10212 | |
10213 | replicate_inode(next, from, reply->trace, mds->mdsmap->get_up_features()); | |
10214 | dout(7) << "handle_discover added inode " << *next << dendl; | |
10215 | ||
10216 | // descend, keep going. | |
10217 | cur = next; | |
10218 | continue; | |
10219 | } | |
10220 | ||
10221 | // how did we do? | |
11fdf7f2 | 10222 | ceph_assert(!reply->is_empty()); |
7c673cae FG |
10223 | dout(7) << "handle_discover sending result back to asker mds." << from << dendl; |
10224 | mds->send_message(reply, dis->get_connection()); | |
7c673cae FG |
10225 | } |
10226 | ||
11fdf7f2 | 10227 | void MDCache::handle_discover_reply(const MDiscoverReply::const_ref &m) |
7c673cae FG |
10228 | { |
10229 | /* | |
10230 | if (mds->get_state() < MDSMap::STATE_ACTIVE) { | |
10231 | dout(0) << "discover_reply NOT ACTIVE YET" << dendl; | |
7c673cae FG |
10232 | return; |
10233 | } | |
10234 | */ | |
10235 | dout(7) << "discover_reply " << *m << dendl; | |
10236 | if (m->is_flag_error_dir()) | |
10237 | dout(7) << " flag error, dir" << dendl; | |
10238 | if (m->is_flag_error_dn()) | |
10239 | dout(7) << " flag error, dentry = " << m->get_error_dentry() << dendl; | |
10240 | ||
11fdf7f2 | 10241 | MDSContext::vec finished, error; |
7c673cae FG |
10242 | mds_rank_t from = mds_rank_t(m->get_source().num()); |
10243 | ||
10244 | // starting point | |
10245 | CInode *cur = get_inode(m->get_base_ino()); | |
11fdf7f2 | 10246 | auto p = m->trace.cbegin(); |
7c673cae FG |
10247 | |
10248 | int next = m->starts_with; | |
10249 | ||
10250 | // decrement discover counters | |
10251 | if (m->get_tid()) { | |
10252 | map<ceph_tid_t,discover_info_t>::iterator p = discovers.find(m->get_tid()); | |
10253 | if (p != discovers.end()) { | |
10254 | dout(10) << " found tid " << m->get_tid() << dendl; | |
10255 | discovers.erase(p); | |
10256 | } else { | |
10257 | dout(10) << " tid " << m->get_tid() << " not found, must be dup reply" << dendl; | |
10258 | } | |
10259 | } | |
10260 | ||
10261 | // discover may start with an inode | |
10262 | if (!p.end() && next == MDiscoverReply::INODE) { | |
10263 | cur = add_replica_inode(p, NULL, finished); | |
10264 | dout(7) << "discover_reply got base inode " << *cur << dendl; | |
11fdf7f2 | 10265 | ceph_assert(cur->is_base()); |
7c673cae FG |
10266 | |
10267 | next = MDiscoverReply::DIR; | |
10268 | ||
10269 | // take waiters? | |
10270 | if (cur->is_base() && | |
10271 | waiting_for_base_ino[from].count(cur->ino())) { | |
10272 | finished.swap(waiting_for_base_ino[from][cur->ino()]); | |
10273 | waiting_for_base_ino[from].erase(cur->ino()); | |
10274 | } | |
10275 | } | |
11fdf7f2 | 10276 | ceph_assert(cur); |
7c673cae FG |
10277 | |
10278 | // loop over discover results. | |
10279 | // indexes follow each ([[dir] dentry] inode) | |
10280 | // can start, end with any type. | |
10281 | while (!p.end()) { | |
10282 | // dir | |
10283 | frag_t fg; | |
10284 | CDir *curdir = 0; | |
10285 | if (next == MDiscoverReply::DIR) { | |
10286 | curdir = add_replica_dir(p, cur, mds_rank_t(m->get_source().num()), finished); | |
10287 | if (cur->ino() == m->get_base_ino() && curdir->get_frag() != m->get_base_dir_frag()) { | |
11fdf7f2 | 10288 | ceph_assert(m->get_wanted_base_dir()); |
7c673cae FG |
10289 | cur->take_dir_waiting(m->get_base_dir_frag(), finished); |
10290 | } | |
10291 | } else { | |
10292 | // note: this can only happen our first way around this loop. | |
10293 | if (p.end() && m->is_flag_error_dn()) { | |
10294 | fg = cur->pick_dirfrag(m->get_error_dentry()); | |
10295 | curdir = cur->get_dirfrag(fg); | |
10296 | } else | |
10297 | curdir = cur->get_dirfrag(m->get_base_dir_frag()); | |
10298 | } | |
10299 | ||
10300 | if (p.end()) | |
10301 | break; | |
10302 | ||
10303 | // dentry | |
10304 | CDentry *dn = add_replica_dentry(p, curdir, finished); | |
10305 | ||
10306 | if (p.end()) | |
10307 | break; | |
10308 | ||
10309 | // inode | |
10310 | cur = add_replica_inode(p, dn, finished); | |
10311 | ||
10312 | next = MDiscoverReply::DIR; | |
10313 | } | |
10314 | ||
10315 | // dir error? | |
10316 | // or dir_auth hint? | |
10317 | if (m->is_flag_error_dir() && !cur->is_dir()) { | |
10318 | // not a dir. | |
10319 | cur->take_waiting(CInode::WAIT_DIR, error); | |
10320 | } else if (m->is_flag_error_dir() || m->get_dir_auth_hint() != CDIR_AUTH_UNKNOWN) { | |
10321 | mds_rank_t who = m->get_dir_auth_hint(); | |
10322 | if (who == mds->get_nodeid()) who = -1; | |
10323 | if (who >= 0) | |
10324 | dout(7) << " dir_auth_hint is " << m->get_dir_auth_hint() << dendl; | |
10325 | ||
7c673cae FG |
10326 | |
10327 | if (m->get_wanted_base_dir()) { | |
31f18b77 FG |
10328 | frag_t fg = m->get_base_dir_frag(); |
10329 | CDir *dir = cur->get_dirfrag(fg); | |
10330 | ||
7c673cae FG |
10331 | if (cur->is_waiting_for_dir(fg)) { |
10332 | if (cur->is_auth()) | |
10333 | cur->take_waiting(CInode::WAIT_DIR, finished); | |
10334 | else if (dir || !cur->dirfragtree.is_leaf(fg)) | |
10335 | cur->take_dir_waiting(fg, finished); | |
10336 | else | |
10337 | discover_dir_frag(cur, fg, 0, who); | |
10338 | } else | |
10339 | dout(7) << " doing nothing, nobody is waiting for dir" << dendl; | |
10340 | } | |
10341 | ||
10342 | // try again? | |
10343 | if (m->get_error_dentry().length()) { | |
31f18b77 FG |
10344 | frag_t fg = cur->pick_dirfrag(m->get_error_dentry()); |
10345 | CDir *dir = cur->get_dirfrag(fg); | |
7c673cae FG |
10346 | // wanted a dentry |
10347 | if (dir && dir->is_waiting_for_dentry(m->get_error_dentry(), m->get_wanted_snapid())) { | |
10348 | if (dir->is_auth() || dir->lookup(m->get_error_dentry())) { | |
10349 | dir->take_dentry_waiting(m->get_error_dentry(), m->get_wanted_snapid(), | |
10350 | m->get_wanted_snapid(), finished); | |
10351 | } else { | |
10352 | filepath relpath(m->get_error_dentry(), 0); | |
10353 | discover_path(dir, m->get_wanted_snapid(), relpath, 0, m->get_wanted_xlocked()); | |
10354 | } | |
10355 | } else | |
10356 | dout(7) << " doing nothing, have dir but nobody is waiting on dentry " | |
10357 | << m->get_error_dentry() << dendl; | |
10358 | } | |
31f18b77 FG |
10359 | } else if (m->is_flag_error_dn()) { |
10360 | frag_t fg = cur->pick_dirfrag(m->get_error_dentry()); | |
10361 | CDir *dir = cur->get_dirfrag(fg); | |
10362 | if (dir) { | |
10363 | if (dir->is_auth()) { | |
10364 | dir->take_sub_waiting(finished); | |
10365 | } else { | |
10366 | dir->take_dentry_waiting(m->get_error_dentry(), m->get_wanted_snapid(), | |
10367 | m->get_wanted_snapid(), error); | |
10368 | } | |
10369 | } | |
7c673cae FG |
10370 | } |
10371 | ||
10372 | // waiters | |
10373 | finish_contexts(g_ceph_context, error, -ENOENT); // finish errors directly | |
10374 | mds->queue_waiters(finished); | |
7c673cae FG |
10375 | } |
10376 | ||
10377 | ||
10378 | ||
10379 | // ---------------------------- | |
10380 | // REPLICAS | |
10381 | ||
b32b8144 FG |
10382 | |
10383 | void MDCache::replicate_dir(CDir *dir, mds_rank_t to, bufferlist& bl) | |
10384 | { | |
10385 | dirfrag_t df = dir->dirfrag(); | |
11fdf7f2 | 10386 | encode(df, bl); |
b32b8144 FG |
10387 | dir->encode_replica(to, bl); |
10388 | } | |
10389 | ||
10390 | void MDCache::replicate_dentry(CDentry *dn, mds_rank_t to, bufferlist& bl) | |
10391 | { | |
11fdf7f2 TL |
10392 | encode(dn->get_name(), bl); |
10393 | encode(dn->last, bl); | |
b32b8144 FG |
10394 | dn->encode_replica(to, bl, mds->get_state() < MDSMap::STATE_ACTIVE); |
10395 | } | |
10396 | ||
10397 | void MDCache::replicate_inode(CInode *in, mds_rank_t to, bufferlist& bl, | |
10398 | uint64_t features) | |
10399 | { | |
11fdf7f2 TL |
10400 | encode(in->inode.ino, bl); // bleh, minor assymetry here |
10401 | encode(in->last, bl); | |
b32b8144 FG |
10402 | in->encode_replica(to, bl, features, mds->get_state() < MDSMap::STATE_ACTIVE); |
10403 | } | |
10404 | ||
11fdf7f2 TL |
10405 | CDir *MDCache::add_replica_dir(bufferlist::const_iterator& p, CInode *diri, mds_rank_t from, |
10406 | MDSContext::vec& finished) | |
7c673cae FG |
10407 | { |
10408 | dirfrag_t df; | |
11fdf7f2 | 10409 | decode(df, p); |
7c673cae | 10410 | |
11fdf7f2 | 10411 | ceph_assert(diri->ino() == df.ino); |
7c673cae FG |
10412 | |
10413 | // add it (_replica_) | |
10414 | CDir *dir = diri->get_dirfrag(df.frag); | |
10415 | ||
10416 | if (dir) { | |
10417 | // had replica. update w/ new nonce. | |
10418 | dir->decode_replica(p); | |
10419 | dout(7) << "add_replica_dir had " << *dir << " nonce " << dir->replica_nonce << dendl; | |
10420 | } else { | |
10421 | // force frag to leaf in the diri tree | |
10422 | if (!diri->dirfragtree.is_leaf(df.frag)) { | |
10423 | dout(7) << "add_replica_dir forcing frag " << df.frag << " to leaf in the fragtree " | |
10424 | << diri->dirfragtree << dendl; | |
10425 | diri->dirfragtree.force_to_leaf(g_ceph_context, df.frag); | |
10426 | } | |
10427 | ||
10428 | // add replica. | |
10429 | dir = diri->add_dirfrag( new CDir(diri, df.frag, this, false) ); | |
10430 | dir->decode_replica(p); | |
10431 | ||
10432 | // is this a dir_auth delegation boundary? | |
10433 | if (from != diri->authority().first || | |
10434 | diri->is_ambiguous_auth() || | |
10435 | diri->is_base()) | |
10436 | adjust_subtree_auth(dir, from); | |
10437 | ||
10438 | dout(7) << "add_replica_dir added " << *dir << " nonce " << dir->replica_nonce << dendl; | |
10439 | ||
10440 | // get waiters | |
10441 | diri->take_dir_waiting(df.frag, finished); | |
10442 | } | |
10443 | ||
10444 | return dir; | |
10445 | } | |
10446 | ||
11fdf7f2 | 10447 | CDentry *MDCache::add_replica_dentry(bufferlist::const_iterator& p, CDir *dir, MDSContext::vec& finished) |
7c673cae FG |
10448 | { |
10449 | string name; | |
10450 | snapid_t last; | |
11fdf7f2 TL |
10451 | decode(name, p); |
10452 | decode(last, p); | |
7c673cae FG |
10453 | |
10454 | CDentry *dn = dir->lookup(name, last); | |
10455 | ||
10456 | // have it? | |
10457 | if (dn) { | |
10458 | dn->decode_replica(p, false); | |
10459 | dout(7) << "add_replica_dentry had " << *dn << dendl; | |
10460 | } else { | |
10461 | dn = dir->add_null_dentry(name, 1 /* this will get updated below */, last); | |
10462 | dn->decode_replica(p, true); | |
10463 | dout(7) << "add_replica_dentry added " << *dn << dendl; | |
10464 | } | |
10465 | ||
10466 | dir->take_dentry_waiting(name, dn->first, dn->last, finished); | |
10467 | ||
10468 | return dn; | |
10469 | } | |
10470 | ||
11fdf7f2 | 10471 | CInode *MDCache::add_replica_inode(bufferlist::const_iterator& p, CDentry *dn, MDSContext::vec& finished) |
7c673cae FG |
10472 | { |
10473 | inodeno_t ino; | |
10474 | snapid_t last; | |
11fdf7f2 TL |
10475 | decode(ino, p); |
10476 | decode(last, p); | |
7c673cae FG |
10477 | CInode *in = get_inode(ino, last); |
10478 | if (!in) { | |
10479 | in = new CInode(this, false, 1, last); | |
10480 | in->decode_replica(p, true); | |
10481 | add_inode(in); | |
10482 | if (in->ino() == MDS_INO_ROOT) | |
10483 | in->inode_auth.first = 0; | |
10484 | else if (in->is_mdsdir()) | |
10485 | in->inode_auth.first = in->ino() - MDS_INO_MDSDIR_OFFSET; | |
10486 | dout(10) << "add_replica_inode added " << *in << dendl; | |
10487 | if (dn) { | |
11fdf7f2 | 10488 | ceph_assert(dn->get_linkage()->is_null()); |
7c673cae FG |
10489 | dn->dir->link_primary_inode(dn, in); |
10490 | } | |
10491 | } else { | |
10492 | in->decode_replica(p, false); | |
10493 | dout(10) << "add_replica_inode had " << *in << dendl; | |
10494 | } | |
10495 | ||
10496 | if (dn) { | |
10497 | if (!dn->get_linkage()->is_primary() || dn->get_linkage()->get_inode() != in) | |
10498 | dout(10) << "add_replica_inode different linkage in dentry " << *dn << dendl; | |
10499 | } | |
10500 | ||
10501 | return in; | |
10502 | } | |
10503 | ||
10504 | ||
10505 | void MDCache::replicate_stray(CDentry *straydn, mds_rank_t who, bufferlist& bl) | |
10506 | { | |
10507 | uint64_t features = mds->mdsmap->get_up_features(); | |
10508 | replicate_inode(get_myin(), who, bl, features); | |
10509 | replicate_dir(straydn->get_dir()->inode->get_parent_dn()->get_dir(), who, bl); | |
10510 | replicate_dentry(straydn->get_dir()->inode->get_parent_dn(), who, bl); | |
10511 | replicate_inode(straydn->get_dir()->inode, who, bl, features); | |
10512 | replicate_dir(straydn->get_dir(), who, bl); | |
10513 | replicate_dentry(straydn, who, bl); | |
10514 | } | |
10515 | ||
11fdf7f2 | 10516 | CDentry *MDCache::add_replica_stray(const bufferlist &bl, mds_rank_t from) |
7c673cae | 10517 | { |
11fdf7f2 TL |
10518 | MDSContext::vec finished; |
10519 | auto p = bl.cbegin(); | |
7c673cae FG |
10520 | |
10521 | CInode *mdsin = add_replica_inode(p, NULL, finished); | |
10522 | CDir *mdsdir = add_replica_dir(p, mdsin, from, finished); | |
10523 | CDentry *straydirdn = add_replica_dentry(p, mdsdir, finished); | |
10524 | CInode *strayin = add_replica_inode(p, straydirdn, finished); | |
10525 | CDir *straydir = add_replica_dir(p, strayin, from, finished); | |
10526 | CDentry *straydn = add_replica_dentry(p, straydir, finished); | |
10527 | if (!finished.empty()) | |
10528 | mds->queue_waiters(finished); | |
10529 | ||
10530 | return straydn; | |
10531 | } | |
10532 | ||
10533 | ||
10534 | int MDCache::send_dir_updates(CDir *dir, bool bcast) | |
10535 | { | |
10536 | // this is an FYI, re: replication | |
10537 | ||
10538 | set<mds_rank_t> who; | |
10539 | if (bcast) { | |
10540 | mds->get_mds_map()->get_active_mds_set(who); | |
10541 | } else { | |
181888fb FG |
10542 | for (const auto &p : dir->get_replicas()) { |
10543 | who.insert(p.first); | |
10544 | } | |
7c673cae FG |
10545 | } |
10546 | ||
10547 | dout(7) << "sending dir_update on " << *dir << " bcast " << bcast << " to " << who << dendl; | |
10548 | ||
10549 | filepath path; | |
10550 | dir->inode->make_path(path); | |
10551 | ||
10552 | mds_rank_t whoami = mds->get_nodeid(); | |
10553 | for (set<mds_rank_t>::iterator it = who.begin(); | |
10554 | it != who.end(); | |
10555 | ++it) { | |
10556 | if (*it == whoami) continue; | |
10557 | //if (*it == except) continue; | |
10558 | dout(7) << "sending dir_update on " << *dir << " to " << *it << dendl; | |
10559 | ||
94b18763 FG |
10560 | std::set<int32_t> s; |
10561 | for (const auto &r : dir->dir_rep_by) { | |
10562 | s.insert(r); | |
10563 | } | |
11fdf7f2 | 10564 | mds->send_message_mds(MDirUpdate::create(mds->get_nodeid(), dir->dirfrag(), dir->dir_rep, s, path, bcast), *it); |
7c673cae FG |
10565 | } |
10566 | ||
10567 | return 0; | |
10568 | } | |
10569 | ||
11fdf7f2 | 10570 | void MDCache::handle_dir_update(const MDirUpdate::const_ref &m) |
7c673cae | 10571 | { |
224ce89b WB |
10572 | dirfrag_t df = m->get_dirfrag(); |
10573 | CDir *dir = get_dirfrag(df); | |
7c673cae | 10574 | if (!dir) { |
224ce89b | 10575 | dout(5) << "dir_update on " << df << ", don't have it" << dendl; |
7c673cae FG |
10576 | |
10577 | // discover it? | |
10578 | if (m->should_discover()) { | |
10579 | // only try once! | |
10580 | // this is key to avoid a fragtree update race, among other things. | |
224ce89b | 10581 | m->inc_tried_discover(); |
7c673cae FG |
10582 | vector<CDentry*> trace; |
10583 | CInode *in; | |
10584 | filepath path = m->get_path(); | |
10585 | dout(5) << "trying discover on dir_update for " << path << dendl; | |
11fdf7f2 | 10586 | CF_MDS_RetryMessageFactory cf(mds, m); |
7c673cae | 10587 | MDRequestRef null_ref; |
11fdf7f2 | 10588 | int r = path_traverse(null_ref, cf, path, &trace, &in, MDS_TRAVERSE_DISCOVER); |
7c673cae FG |
10589 | if (r > 0) |
10590 | return; | |
224ce89b WB |
10591 | if (r == 0 && |
10592 | in->ino() == df.ino && | |
10593 | in->get_approx_dirfrag(df.frag) == NULL) { | |
10594 | open_remote_dirfrag(in, df.frag, new C_MDS_RetryMessage(mds, m)); | |
10595 | return; | |
10596 | } | |
7c673cae FG |
10597 | } |
10598 | ||
7c673cae FG |
10599 | return; |
10600 | } | |
10601 | ||
224ce89b WB |
10602 | if (!m->has_tried_discover()) { |
10603 | // Update if it already exists. Othwerwise it got updated by discover reply. | |
10604 | dout(5) << "dir_update on " << *dir << dendl; | |
10605 | dir->dir_rep = m->get_dir_rep(); | |
94b18763 FG |
10606 | dir->dir_rep_by.clear(); |
10607 | for (const auto &e : m->get_dir_rep_by()) { | |
10608 | dir->dir_rep_by.insert(e); | |
10609 | } | |
224ce89b | 10610 | } |
7c673cae FG |
10611 | } |
10612 | ||
10613 | ||
10614 | ||
10615 | ||
10616 | ||
10617 | // LINK | |
10618 | ||
10619 | void MDCache::send_dentry_link(CDentry *dn, MDRequestRef& mdr) | |
10620 | { | |
10621 | dout(7) << "send_dentry_link " << *dn << dendl; | |
10622 | ||
10623 | CDir *subtree = get_subtree_root(dn->get_dir()); | |
181888fb | 10624 | for (const auto &p : dn->get_replicas()) { |
7c673cae | 10625 | // don't tell (rename) witnesses; they already know |
181888fb | 10626 | if (mdr.get() && mdr->more()->witnessed.count(p.first)) |
7c673cae | 10627 | continue; |
181888fb FG |
10628 | if (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || |
10629 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
10630 | rejoin_gather.count(p.first))) | |
7c673cae FG |
10631 | continue; |
10632 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
11fdf7f2 | 10633 | auto m = MDentryLink::create(subtree->dirfrag(), dn->get_dir()->dirfrag(), dn->get_name(), dnl->is_primary()); |
7c673cae FG |
10634 | if (dnl->is_primary()) { |
10635 | dout(10) << " primary " << *dnl->get_inode() << dendl; | |
181888fb | 10636 | replicate_inode(dnl->get_inode(), p.first, m->bl, |
7c673cae FG |
10637 | mds->mdsmap->get_up_features()); |
10638 | } else if (dnl->is_remote()) { | |
10639 | inodeno_t ino = dnl->get_remote_ino(); | |
10640 | __u8 d_type = dnl->get_remote_d_type(); | |
10641 | dout(10) << " remote " << ino << " " << d_type << dendl; | |
11fdf7f2 TL |
10642 | encode(ino, m->bl); |
10643 | encode(d_type, m->bl); | |
7c673cae FG |
10644 | } else |
10645 | ceph_abort(); // aie, bad caller! | |
181888fb | 10646 | mds->send_message_mds(m, p.first); |
7c673cae FG |
10647 | } |
10648 | } | |
10649 | ||
11fdf7f2 | 10650 | void MDCache::handle_dentry_link(const MDentryLink::const_ref &m) |
7c673cae | 10651 | { |
7c673cae FG |
10652 | CDentry *dn = NULL; |
10653 | CDir *dir = get_dirfrag(m->get_dirfrag()); | |
10654 | if (!dir) { | |
10655 | dout(7) << "handle_dentry_link don't have dirfrag " << m->get_dirfrag() << dendl; | |
10656 | } else { | |
10657 | dn = dir->lookup(m->get_dn()); | |
10658 | if (!dn) { | |
10659 | dout(7) << "handle_dentry_link don't have dentry " << *dir << " dn " << m->get_dn() << dendl; | |
10660 | } else { | |
10661 | dout(7) << "handle_dentry_link on " << *dn << dendl; | |
10662 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
10663 | ||
11fdf7f2 TL |
10664 | ceph_assert(!dn->is_auth()); |
10665 | ceph_assert(dnl->is_null()); | |
7c673cae FG |
10666 | } |
10667 | } | |
10668 | ||
11fdf7f2 TL |
10669 | auto p = m->bl.cbegin(); |
10670 | MDSContext::vec finished; | |
7c673cae FG |
10671 | if (dn) { |
10672 | if (m->get_is_primary()) { | |
10673 | // primary link. | |
10674 | add_replica_inode(p, dn, finished); | |
10675 | } else { | |
10676 | // remote link, easy enough. | |
10677 | inodeno_t ino; | |
10678 | __u8 d_type; | |
11fdf7f2 TL |
10679 | decode(ino, p); |
10680 | decode(d_type, p); | |
7c673cae FG |
10681 | dir->link_remote_inode(dn, ino, d_type); |
10682 | } | |
10683 | } else { | |
10684 | ceph_abort(); | |
10685 | } | |
10686 | ||
10687 | if (!finished.empty()) | |
10688 | mds->queue_waiters(finished); | |
10689 | ||
7c673cae FG |
10690 | return; |
10691 | } | |
10692 | ||
10693 | ||
10694 | // UNLINK | |
10695 | ||
10696 | void MDCache::send_dentry_unlink(CDentry *dn, CDentry *straydn, MDRequestRef& mdr) | |
10697 | { | |
10698 | dout(10) << "send_dentry_unlink " << *dn << dendl; | |
10699 | // share unlink news with replicas | |
10700 | set<mds_rank_t> replicas; | |
10701 | dn->list_replicas(replicas); | |
11fdf7f2 TL |
10702 | bufferlist snapbl; |
10703 | if (straydn) { | |
7c673cae | 10704 | straydn->list_replicas(replicas); |
11fdf7f2 TL |
10705 | CInode *strayin = straydn->get_linkage()->get_inode(); |
10706 | strayin->encode_snap_blob(snapbl); | |
10707 | } | |
7c673cae FG |
10708 | for (set<mds_rank_t>::iterator it = replicas.begin(); |
10709 | it != replicas.end(); | |
10710 | ++it) { | |
10711 | // don't tell (rmdir) witnesses; they already know | |
10712 | if (mdr.get() && mdr->more()->witnessed.count(*it)) | |
10713 | continue; | |
10714 | ||
10715 | if (mds->mdsmap->get_state(*it) < MDSMap::STATE_REJOIN || | |
10716 | (mds->mdsmap->get_state(*it) == MDSMap::STATE_REJOIN && | |
10717 | rejoin_gather.count(*it))) | |
10718 | continue; | |
10719 | ||
11fdf7f2 TL |
10720 | auto unlink = MDentryUnlink::create(dn->get_dir()->dirfrag(), dn->get_name()); |
10721 | if (straydn) { | |
7c673cae | 10722 | replicate_stray(straydn, *it, unlink->straybl); |
11fdf7f2 TL |
10723 | unlink->snapbl = snapbl; |
10724 | } | |
7c673cae FG |
10725 | mds->send_message_mds(unlink, *it); |
10726 | } | |
10727 | } | |
10728 | ||
11fdf7f2 | 10729 | void MDCache::handle_dentry_unlink(const MDentryUnlink::const_ref &m) |
7c673cae FG |
10730 | { |
10731 | // straydn | |
10732 | CDentry *straydn = NULL; | |
10733 | if (m->straybl.length()) | |
10734 | straydn = add_replica_stray(m->straybl, mds_rank_t(m->get_source().num())); | |
10735 | ||
10736 | CDir *dir = get_dirfrag(m->get_dirfrag()); | |
10737 | if (!dir) { | |
10738 | dout(7) << "handle_dentry_unlink don't have dirfrag " << m->get_dirfrag() << dendl; | |
10739 | } else { | |
10740 | CDentry *dn = dir->lookup(m->get_dn()); | |
10741 | if (!dn) { | |
10742 | dout(7) << "handle_dentry_unlink don't have dentry " << *dir << " dn " << m->get_dn() << dendl; | |
10743 | } else { | |
10744 | dout(7) << "handle_dentry_unlink on " << *dn << dendl; | |
10745 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
10746 | ||
10747 | // open inode? | |
10748 | if (dnl->is_primary()) { | |
10749 | CInode *in = dnl->get_inode(); | |
10750 | dn->dir->unlink_inode(dn); | |
11fdf7f2 | 10751 | ceph_assert(straydn); |
7c673cae FG |
10752 | straydn->dir->link_primary_inode(straydn, in); |
10753 | ||
10754 | // in->first is lazily updated on replica; drag it forward so | |
10755 | // that we always keep it in sync with the dnq | |
11fdf7f2 | 10756 | ceph_assert(straydn->first >= in->first); |
7c673cae FG |
10757 | in->first = straydn->first; |
10758 | ||
10759 | // update subtree map? | |
10760 | if (in->is_dir()) | |
10761 | adjust_subtree_after_rename(in, dir, false); | |
10762 | ||
11fdf7f2 TL |
10763 | if (m->snapbl.length()) { |
10764 | bool hadrealm = (in->snaprealm ? true : false); | |
10765 | in->decode_snap_blob(m->snapbl); | |
10766 | ceph_assert(in->snaprealm); | |
10767 | ceph_assert(in->snaprealm->have_past_parents_open()); | |
10768 | if (!hadrealm) | |
10769 | do_realm_invalidate_and_update_notify(in, CEPH_SNAP_OP_SPLIT, false); | |
10770 | } | |
10771 | ||
7c673cae FG |
10772 | // send caps to auth (if we're not already) |
10773 | if (in->is_any_caps() && | |
10774 | !in->state_test(CInode::STATE_EXPORTINGCAPS)) | |
10775 | migrator->export_caps(in); | |
10776 | ||
7c673cae FG |
10777 | straydn = NULL; |
10778 | } else { | |
11fdf7f2 TL |
10779 | ceph_assert(!straydn); |
10780 | ceph_assert(dnl->is_remote()); | |
7c673cae FG |
10781 | dn->dir->unlink_inode(dn); |
10782 | } | |
11fdf7f2 | 10783 | ceph_assert(dnl->is_null()); |
7c673cae FG |
10784 | } |
10785 | } | |
10786 | ||
10787 | // race with trim_dentry() | |
10788 | if (straydn) { | |
11fdf7f2 TL |
10789 | ceph_assert(straydn->get_num_ref() == 0); |
10790 | ceph_assert(straydn->get_linkage()->is_null()); | |
10791 | expiremap ex; | |
10792 | trim_dentry(straydn, ex); | |
10793 | send_expire_messages(ex); | |
7c673cae | 10794 | } |
7c673cae FG |
10795 | } |
10796 | ||
10797 | ||
10798 | ||
10799 | ||
10800 | ||
10801 | ||
10802 | // =================================================================== | |
10803 | ||
10804 | ||
10805 | ||
10806 | // =================================================================== | |
10807 | // FRAGMENT | |
10808 | ||
10809 | ||
10810 | /** | |
10811 | * adjust_dir_fragments -- adjust fragmentation for a directory | |
10812 | * | |
10813 | * @param diri directory inode | |
10814 | * @param basefrag base fragment | |
10815 | * @param bits bit adjustment. positive for split, negative for merge. | |
10816 | */ | |
10817 | void MDCache::adjust_dir_fragments(CInode *diri, frag_t basefrag, int bits, | |
10818 | list<CDir*>& resultfrags, | |
11fdf7f2 | 10819 | MDSContext::vec& waiters, |
7c673cae FG |
10820 | bool replay) |
10821 | { | |
10822 | dout(10) << "adjust_dir_fragments " << basefrag << " " << bits | |
10823 | << " on " << *diri << dendl; | |
10824 | ||
10825 | list<CDir*> srcfrags; | |
10826 | diri->get_dirfrags_under(basefrag, srcfrags); | |
10827 | ||
10828 | adjust_dir_fragments(diri, srcfrags, basefrag, bits, resultfrags, waiters, replay); | |
10829 | } | |
10830 | ||
10831 | CDir *MDCache::force_dir_fragment(CInode *diri, frag_t fg, bool replay) | |
10832 | { | |
10833 | CDir *dir = diri->get_dirfrag(fg); | |
10834 | if (dir) | |
10835 | return dir; | |
10836 | ||
10837 | dout(10) << "force_dir_fragment " << fg << " on " << *diri << dendl; | |
10838 | ||
10839 | list<CDir*> src, result; | |
11fdf7f2 | 10840 | MDSContext::vec waiters; |
7c673cae FG |
10841 | |
10842 | // split a parent? | |
10843 | frag_t parent = diri->dirfragtree.get_branch_or_leaf(fg); | |
10844 | while (1) { | |
10845 | CDir *pdir = diri->get_dirfrag(parent); | |
10846 | if (pdir) { | |
10847 | int split = fg.bits() - parent.bits(); | |
10848 | dout(10) << " splitting parent by " << split << " " << *pdir << dendl; | |
10849 | src.push_back(pdir); | |
10850 | adjust_dir_fragments(diri, src, parent, split, result, waiters, replay); | |
10851 | dir = diri->get_dirfrag(fg); | |
10852 | if (dir) { | |
10853 | dout(10) << "force_dir_fragment result " << *dir << dendl; | |
10854 | break; | |
10855 | } | |
10856 | } | |
10857 | if (parent == frag_t()) | |
10858 | break; | |
10859 | frag_t last = parent; | |
10860 | parent = parent.parent(); | |
10861 | dout(10) << " " << last << " parent is " << parent << dendl; | |
10862 | } | |
10863 | ||
10864 | if (!dir) { | |
10865 | // hoover up things under fg? | |
10866 | diri->get_dirfrags_under(fg, src); | |
10867 | if (src.empty()) { | |
10868 | dout(10) << "force_dir_fragment no frags under " << fg << dendl; | |
10869 | } else { | |
10870 | dout(10) << " will combine frags under " << fg << ": " << src << dendl; | |
10871 | adjust_dir_fragments(diri, src, fg, 0, result, waiters, replay); | |
10872 | dir = result.front(); | |
10873 | dout(10) << "force_dir_fragment result " << *dir << dendl; | |
10874 | } | |
10875 | } | |
10876 | if (!replay) | |
10877 | mds->queue_waiters(waiters); | |
10878 | return dir; | |
10879 | } | |
10880 | ||
10881 | void MDCache::adjust_dir_fragments(CInode *diri, | |
10882 | list<CDir*>& srcfrags, | |
10883 | frag_t basefrag, int bits, | |
10884 | list<CDir*>& resultfrags, | |
11fdf7f2 | 10885 | MDSContext::vec& waiters, |
7c673cae FG |
10886 | bool replay) |
10887 | { | |
10888 | dout(10) << "adjust_dir_fragments " << basefrag << " bits " << bits | |
10889 | << " srcfrags " << srcfrags | |
10890 | << " on " << *diri << dendl; | |
10891 | ||
10892 | // adjust fragtree | |
10893 | // yuck. we may have discovered the inode while it was being fragmented. | |
10894 | if (!diri->dirfragtree.is_leaf(basefrag)) | |
10895 | diri->dirfragtree.force_to_leaf(g_ceph_context, basefrag); | |
10896 | ||
10897 | if (bits > 0) | |
10898 | diri->dirfragtree.split(basefrag, bits); | |
10899 | dout(10) << " new fragtree is " << diri->dirfragtree << dendl; | |
10900 | ||
10901 | if (srcfrags.empty()) | |
10902 | return; | |
10903 | ||
10904 | // split | |
10905 | CDir *parent_dir = diri->get_parent_dir(); | |
10906 | CDir *parent_subtree = 0; | |
10907 | if (parent_dir) | |
10908 | parent_subtree = get_subtree_root(parent_dir); | |
10909 | ||
10910 | if (bits > 0) { | |
10911 | // SPLIT | |
11fdf7f2 | 10912 | ceph_assert(srcfrags.size() == 1); |
7c673cae FG |
10913 | CDir *dir = srcfrags.front(); |
10914 | ||
10915 | dir->split(bits, resultfrags, waiters, replay); | |
10916 | ||
10917 | // did i change the subtree map? | |
10918 | if (dir->is_subtree_root()) { | |
10919 | // new frags are now separate subtrees | |
10920 | for (list<CDir*>::iterator p = resultfrags.begin(); | |
10921 | p != resultfrags.end(); | |
10922 | ++p) | |
10923 | subtrees[*p].clear(); // new frag is now its own subtree | |
10924 | ||
10925 | // was i a bound? | |
10926 | if (parent_subtree) { | |
11fdf7f2 | 10927 | ceph_assert(subtrees[parent_subtree].count(dir)); |
7c673cae FG |
10928 | subtrees[parent_subtree].erase(dir); |
10929 | for (list<CDir*>::iterator p = resultfrags.begin(); | |
10930 | p != resultfrags.end(); | |
10931 | ++p) { | |
11fdf7f2 | 10932 | ceph_assert((*p)->is_subtree_root()); |
7c673cae FG |
10933 | subtrees[parent_subtree].insert(*p); |
10934 | } | |
10935 | } | |
10936 | ||
10937 | // adjust my bounds. | |
10938 | set<CDir*> bounds; | |
10939 | bounds.swap(subtrees[dir]); | |
10940 | subtrees.erase(dir); | |
10941 | for (set<CDir*>::iterator p = bounds.begin(); | |
10942 | p != bounds.end(); | |
10943 | ++p) { | |
10944 | CDir *frag = get_subtree_root((*p)->get_parent_dir()); | |
10945 | subtrees[frag].insert(*p); | |
10946 | } | |
10947 | ||
10948 | show_subtrees(10); | |
7c673cae FG |
10949 | } |
10950 | ||
10951 | diri->close_dirfrag(dir->get_frag()); | |
10952 | ||
10953 | } else { | |
10954 | // MERGE | |
10955 | ||
10956 | // are my constituent bits subtrees? if so, i will be too. | |
10957 | // (it's all or none, actually.) | |
11fdf7f2 | 10958 | bool any_subtree = false, any_non_subtree = false; |
31f18b77 | 10959 | for (CDir *dir : srcfrags) { |
11fdf7f2 | 10960 | if (dir->is_subtree_root()) |
31f18b77 | 10961 | any_subtree = true; |
11fdf7f2 TL |
10962 | else |
10963 | any_non_subtree = true; | |
31f18b77 | 10964 | } |
11fdf7f2 TL |
10965 | ceph_assert(!any_subtree || !any_non_subtree); |
10966 | ||
31f18b77 FG |
10967 | set<CDir*> new_bounds; |
10968 | if (any_subtree) { | |
10969 | for (CDir *dir : srcfrags) { | |
10970 | // this simplifies the code that find subtrees underneath the dirfrag | |
10971 | if (!dir->is_subtree_root()) { | |
10972 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
10973 | adjust_subtree_auth(dir, mds->get_nodeid()); | |
10974 | } | |
10975 | } | |
10976 | ||
10977 | for (CDir *dir : srcfrags) { | |
11fdf7f2 | 10978 | ceph_assert(dir->is_subtree_root()); |
7c673cae | 10979 | dout(10) << " taking srcfrag subtree bounds from " << *dir << dendl; |
7c673cae FG |
10980 | map<CDir*, set<CDir*> >::iterator q = subtrees.find(dir); |
10981 | set<CDir*>::iterator r = q->second.begin(); | |
10982 | while (r != subtrees[dir].end()) { | |
10983 | new_bounds.insert(*r); | |
10984 | subtrees[dir].erase(r++); | |
10985 | } | |
10986 | subtrees.erase(q); | |
31f18b77 | 10987 | |
7c673cae FG |
10988 | // remove myself as my parent's bound |
10989 | if (parent_subtree) | |
10990 | subtrees[parent_subtree].erase(dir); | |
10991 | } | |
10992 | } | |
10993 | ||
10994 | // merge | |
10995 | CDir *f = new CDir(diri, basefrag, this, srcfrags.front()->is_auth()); | |
10996 | f->merge(srcfrags, waiters, replay); | |
7c673cae | 10997 | |
31f18b77 | 10998 | if (any_subtree) { |
11fdf7f2 | 10999 | ceph_assert(f->is_subtree_root()); |
7c673cae FG |
11000 | subtrees[f].swap(new_bounds); |
11001 | if (parent_subtree) | |
11002 | subtrees[parent_subtree].insert(f); | |
11003 | ||
11004 | show_subtrees(10); | |
11005 | } | |
11006 | ||
11007 | resultfrags.push_back(f); | |
11008 | } | |
11009 | } | |
11010 | ||
11011 | ||
11012 | class C_MDC_FragmentFrozen : public MDSInternalContext { | |
11013 | MDCache *mdcache; | |
11014 | MDRequestRef mdr; | |
11015 | public: | |
11016 | C_MDC_FragmentFrozen(MDCache *m, MDRequestRef& r) : | |
11017 | MDSInternalContext(m->mds), mdcache(m), mdr(r) {} | |
11018 | void finish(int r) override { | |
11019 | mdcache->fragment_frozen(mdr, r); | |
11020 | } | |
11021 | }; | |
11022 | ||
11023 | bool MDCache::can_fragment(CInode *diri, list<CDir*>& dirs) | |
11024 | { | |
11025 | if (is_readonly()) { | |
11026 | dout(7) << "can_fragment: read-only FS, no fragmenting for now" << dendl; | |
11027 | return false; | |
11028 | } | |
11029 | if (mds->is_cluster_degraded()) { | |
11030 | dout(7) << "can_fragment: cluster degraded, no fragmenting for now" << dendl; | |
11031 | return false; | |
11032 | } | |
11033 | if (diri->get_parent_dir() && | |
11034 | diri->get_parent_dir()->get_inode()->is_stray()) { | |
11035 | dout(7) << "can_fragment: i won't merge|split anything in stray" << dendl; | |
11036 | return false; | |
11037 | } | |
11038 | if (diri->is_mdsdir() || diri->is_stray() || diri->ino() == MDS_INO_CEPH) { | |
11039 | dout(7) << "can_fragment: i won't fragment the mdsdir or straydir or .ceph" << dendl; | |
11040 | return false; | |
11041 | } | |
11042 | ||
11043 | if (diri->scrub_is_in_progress()) { | |
11044 | dout(7) << "can_fragment: scrub in progress" << dendl; | |
11045 | return false; | |
11046 | } | |
11047 | ||
11048 | for (list<CDir*>::iterator p = dirs.begin(); p != dirs.end(); ++p) { | |
11049 | CDir *dir = *p; | |
11050 | if (dir->state_test(CDir::STATE_FRAGMENTING)) { | |
11051 | dout(7) << "can_fragment: already fragmenting " << *dir << dendl; | |
11052 | return false; | |
11053 | } | |
11054 | if (!dir->is_auth()) { | |
11055 | dout(7) << "can_fragment: not auth on " << *dir << dendl; | |
11056 | return false; | |
11057 | } | |
11058 | if (dir->is_bad()) { | |
11059 | dout(7) << "can_fragment: bad dirfrag " << *dir << dendl; | |
11060 | return false; | |
11061 | } | |
11062 | if (dir->is_frozen() || | |
11063 | dir->is_freezing()) { | |
11064 | dout(7) << "can_fragment: can't merge, freezing|frozen. wait for other exports to finish first." << dendl; | |
11065 | return false; | |
11066 | } | |
11067 | } | |
11068 | ||
11069 | return true; | |
11070 | } | |
11071 | ||
11072 | void MDCache::split_dir(CDir *dir, int bits) | |
11073 | { | |
11074 | dout(7) << __func__ << " " << *dir << " bits " << bits << dendl; | |
11fdf7f2 | 11075 | ceph_assert(dir->is_auth()); |
7c673cae FG |
11076 | CInode *diri = dir->inode; |
11077 | ||
11078 | list<CDir*> dirs; | |
11079 | dirs.push_back(dir); | |
11080 | ||
11081 | if (!can_fragment(diri, dirs)) { | |
11082 | dout(7) << __func__ << " cannot fragment right now, dropping" << dendl; | |
11083 | return; | |
11084 | } | |
11085 | ||
31f18b77 FG |
11086 | if (dir->frag.bits() + bits > 24) { |
11087 | dout(7) << __func__ << " frag bits > 24, dropping" << dendl; | |
11088 | return; | |
11089 | } | |
11090 | ||
7c673cae FG |
11091 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FRAGMENTDIR); |
11092 | mdr->more()->fragment_base = dir->dirfrag(); | |
11093 | ||
11fdf7f2 | 11094 | ceph_assert(fragments.count(dir->dirfrag()) == 0); |
7c673cae FG |
11095 | fragment_info_t& info = fragments[dir->dirfrag()]; |
11096 | info.mdr = mdr; | |
11097 | info.dirs.push_back(dir); | |
11098 | info.bits = bits; | |
11099 | info.last_cum_auth_pins_change = ceph_clock_now(); | |
11100 | ||
11101 | fragment_freeze_dirs(dirs); | |
11102 | // initial mark+complete pass | |
11103 | fragment_mark_and_complete(mdr); | |
11104 | } | |
11105 | ||
11106 | void MDCache::merge_dir(CInode *diri, frag_t frag) | |
11107 | { | |
11108 | dout(7) << "merge_dir to " << frag << " on " << *diri << dendl; | |
11109 | ||
11110 | list<CDir*> dirs; | |
11111 | if (!diri->get_dirfrags_under(frag, dirs)) { | |
11112 | dout(7) << "don't have all frags under " << frag << " for " << *diri << dendl; | |
11113 | return; | |
11114 | } | |
11115 | ||
11116 | if (diri->dirfragtree.is_leaf(frag)) { | |
11117 | dout(10) << " " << frag << " already a leaf for " << *diri << dendl; | |
11118 | return; | |
11119 | } | |
11120 | ||
11121 | if (!can_fragment(diri, dirs)) | |
11122 | return; | |
11123 | ||
11124 | CDir *first = dirs.front(); | |
11125 | int bits = first->get_frag().bits() - frag.bits(); | |
11126 | dout(10) << " we are merginb by " << bits << " bits" << dendl; | |
11127 | ||
11128 | dirfrag_t basedirfrag(diri->ino(), frag); | |
11129 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FRAGMENTDIR); | |
11130 | mdr->more()->fragment_base = basedirfrag; | |
11131 | ||
11fdf7f2 | 11132 | ceph_assert(fragments.count(basedirfrag) == 0); |
7c673cae FG |
11133 | fragment_info_t& info = fragments[basedirfrag]; |
11134 | info.mdr = mdr; | |
11135 | info.dirs = dirs; | |
11136 | info.bits = -bits; | |
11137 | info.last_cum_auth_pins_change = ceph_clock_now(); | |
11138 | ||
11139 | fragment_freeze_dirs(dirs); | |
11140 | // initial mark+complete pass | |
11141 | fragment_mark_and_complete(mdr); | |
11142 | } | |
11143 | ||
11144 | void MDCache::fragment_freeze_dirs(list<CDir*>& dirs) | |
11145 | { | |
11fdf7f2 TL |
11146 | bool any_subtree = false, any_non_subtree = false; |
11147 | for (CDir* dir : dirs) { | |
7c673cae FG |
11148 | dir->auth_pin(dir); // until we mark and complete them |
11149 | dir->state_set(CDir::STATE_FRAGMENTING); | |
11150 | dir->freeze_dir(); | |
11fdf7f2 TL |
11151 | ceph_assert(dir->is_freezing_dir()); |
11152 | ||
11153 | if (dir->is_subtree_root()) | |
11154 | any_subtree = true; | |
11155 | else | |
11156 | any_non_subtree = true; | |
11157 | } | |
11158 | ||
11159 | if (any_subtree && any_non_subtree) { | |
11160 | // either all dirfrags are subtree roots or all are not. | |
11161 | for (CDir *dir : dirs) { | |
11162 | if (dir->is_subtree_root()) { | |
11163 | ceph_assert(dir->state_test(CDir::STATE_AUXSUBTREE)); | |
11164 | } else { | |
11165 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
11166 | adjust_subtree_auth(dir, mds->get_nodeid()); | |
11167 | } | |
11168 | } | |
7c673cae FG |
11169 | } |
11170 | } | |
11171 | ||
11172 | class C_MDC_FragmentMarking : public MDCacheContext { | |
11173 | MDRequestRef mdr; | |
11174 | public: | |
11175 | C_MDC_FragmentMarking(MDCache *m, MDRequestRef& r) : MDCacheContext(m), mdr(r) {} | |
11176 | void finish(int r) override { | |
11177 | mdcache->fragment_mark_and_complete(mdr); | |
11178 | } | |
11179 | }; | |
11180 | ||
11181 | void MDCache::fragment_mark_and_complete(MDRequestRef& mdr) | |
11182 | { | |
11183 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11184 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11185 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11186 | dout(7) << "fragment_mark_and_complete " << basedirfrag << " must have aborted" << dendl; | |
11187 | request_finish(mdr); | |
11188 | return; | |
11189 | } | |
11190 | ||
11191 | fragment_info_t& info = it->second; | |
11192 | CInode *diri = info.dirs.front()->get_inode(); | |
11193 | dout(10) << "fragment_mark_and_complete " << info.dirs << " on " << *diri << dendl; | |
11194 | ||
11195 | MDSGatherBuilder gather(g_ceph_context); | |
11196 | ||
11197 | for (list<CDir*>::iterator p = info.dirs.begin(); | |
11198 | p != info.dirs.end(); | |
11199 | ++p) { | |
11200 | CDir *dir = *p; | |
11201 | ||
11202 | bool ready = true; | |
11203 | if (!dir->is_complete()) { | |
11204 | dout(15) << " fetching incomplete " << *dir << dendl; | |
11205 | dir->fetch(gather.new_sub(), true); // ignore authpinnability | |
11206 | ready = false; | |
11207 | } else if (dir->get_frag() == frag_t()) { | |
11208 | // The COMPLETE flag gets lost if we fragment a new dirfrag, then rollback | |
11209 | // the operation. To avoid CDir::fetch() complaining about missing object, | |
11210 | // we commit new dirfrag first. | |
11211 | if (dir->state_test(CDir::STATE_CREATING)) { | |
11212 | dout(15) << " waiting until new dir gets journaled " << *dir << dendl; | |
11213 | dir->add_waiter(CDir::WAIT_CREATED, gather.new_sub()); | |
11214 | ready = false; | |
11215 | } else if (dir->is_new()) { | |
11216 | dout(15) << " committing new " << *dir << dendl; | |
11fdf7f2 | 11217 | ceph_assert(dir->is_dirty()); |
7c673cae FG |
11218 | dir->commit(0, gather.new_sub(), true); |
11219 | ready = false; | |
11220 | } | |
11221 | } | |
11222 | if (!ready) | |
11223 | continue; | |
11224 | ||
11225 | if (!dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
11226 | dout(15) << " marking " << *dir << dendl; | |
94b18763 FG |
11227 | for (auto &p : dir->items) { |
11228 | CDentry *dn = p.second; | |
7c673cae | 11229 | dn->get(CDentry::PIN_FRAGMENTING); |
11fdf7f2 | 11230 | ceph_assert(!dn->state_test(CDentry::STATE_FRAGMENTING)); |
7c673cae FG |
11231 | dn->state_set(CDentry::STATE_FRAGMENTING); |
11232 | } | |
11233 | dir->state_set(CDir::STATE_DNPINNEDFRAG); | |
11234 | dir->auth_unpin(dir); | |
11235 | } else { | |
11236 | dout(15) << " already marked " << *dir << dendl; | |
11237 | } | |
11238 | } | |
11239 | if (gather.has_subs()) { | |
11240 | gather.set_finisher(new C_MDC_FragmentMarking(this, mdr)); | |
11241 | gather.activate(); | |
11242 | return; | |
11243 | } | |
11244 | ||
11245 | for (list<CDir*>::iterator p = info.dirs.begin(); | |
11246 | p != info.dirs.end(); | |
11247 | ++p) { | |
11248 | CDir *dir = *p; | |
11249 | if (!dir->is_frozen_dir()) { | |
11fdf7f2 | 11250 | ceph_assert(dir->is_freezing_dir()); |
7c673cae FG |
11251 | dir->add_waiter(CDir::WAIT_FROZEN, gather.new_sub()); |
11252 | } | |
11253 | } | |
11254 | if (gather.has_subs()) { | |
11255 | gather.set_finisher(new C_MDC_FragmentFrozen(this, mdr)); | |
11256 | gather.activate(); | |
11257 | // flush log so that request auth_pins are retired | |
11258 | mds->mdlog->flush(); | |
11259 | return; | |
11260 | } | |
11261 | ||
11262 | fragment_frozen(mdr, 0); | |
11263 | } | |
11264 | ||
11265 | void MDCache::fragment_unmark_unfreeze_dirs(list<CDir*>& dirs) | |
11266 | { | |
11267 | dout(10) << "fragment_unmark_unfreeze_dirs " << dirs << dendl; | |
11268 | for (list<CDir*>::iterator p = dirs.begin(); p != dirs.end(); ++p) { | |
11269 | CDir *dir = *p; | |
11270 | dout(10) << " frag " << *dir << dendl; | |
11271 | ||
11fdf7f2 | 11272 | ceph_assert(dir->state_test(CDir::STATE_FRAGMENTING)); |
7c673cae FG |
11273 | dir->state_clear(CDir::STATE_FRAGMENTING); |
11274 | ||
11275 | if (dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
11276 | dir->state_clear(CDir::STATE_DNPINNEDFRAG); | |
11277 | ||
94b18763 FG |
11278 | for (auto &p : dir->items) { |
11279 | CDentry *dn = p.second; | |
11fdf7f2 | 11280 | ceph_assert(dn->state_test(CDentry::STATE_FRAGMENTING)); |
7c673cae FG |
11281 | dn->state_clear(CDentry::STATE_FRAGMENTING); |
11282 | dn->put(CDentry::PIN_FRAGMENTING); | |
11283 | } | |
11284 | } else { | |
11285 | dir->auth_unpin(dir); | |
11286 | } | |
11287 | ||
11288 | dir->unfreeze_dir(); | |
11289 | } | |
11290 | } | |
11291 | ||
11292 | bool MDCache::fragment_are_all_frozen(CDir *dir) | |
11293 | { | |
11fdf7f2 | 11294 | ceph_assert(dir->is_frozen_dir()); |
7c673cae FG |
11295 | map<dirfrag_t,fragment_info_t>::iterator p; |
11296 | for (p = fragments.lower_bound(dirfrag_t(dir->ino(), 0)); | |
11297 | p != fragments.end() && p->first.ino == dir->ino(); | |
11298 | ++p) { | |
11299 | if (p->first.frag.contains(dir->get_frag())) | |
11300 | return p->second.all_frozen; | |
11301 | } | |
11302 | ceph_abort(); | |
11303 | return false; | |
11304 | } | |
11305 | ||
11306 | void MDCache::fragment_freeze_inc_num_waiters(CDir *dir) | |
11307 | { | |
11308 | map<dirfrag_t,fragment_info_t>::iterator p; | |
11309 | for (p = fragments.lower_bound(dirfrag_t(dir->ino(), 0)); | |
11310 | p != fragments.end() && p->first.ino == dir->ino(); | |
11311 | ++p) { | |
11312 | if (p->first.frag.contains(dir->get_frag())) { | |
11313 | p->second.num_remote_waiters++; | |
11314 | return; | |
11315 | } | |
11316 | } | |
11317 | ceph_abort(); | |
11318 | } | |
11319 | ||
11320 | void MDCache::find_stale_fragment_freeze() | |
11321 | { | |
11322 | dout(10) << "find_stale_fragment_freeze" << dendl; | |
11323 | // see comment in Migrator::find_stale_export_freeze() | |
11324 | utime_t now = ceph_clock_now(); | |
11325 | utime_t cutoff = now; | |
11fdf7f2 | 11326 | cutoff -= g_conf()->mds_freeze_tree_timeout; |
7c673cae FG |
11327 | |
11328 | for (map<dirfrag_t,fragment_info_t>::iterator p = fragments.begin(); | |
11329 | p != fragments.end(); ) { | |
11330 | dirfrag_t df = p->first; | |
11331 | fragment_info_t& info = p->second; | |
11332 | ++p; | |
11333 | if (info.all_frozen) | |
11334 | continue; | |
11335 | CDir *dir; | |
11336 | int total_auth_pins = 0; | |
11337 | for (list<CDir*>::iterator q = info.dirs.begin(); | |
11338 | q != info.dirs.end(); | |
11339 | ++q) { | |
11340 | dir = *q; | |
11341 | if (!dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
11342 | total_auth_pins = -1; | |
11343 | break; | |
11344 | } | |
11345 | if (dir->is_frozen_dir()) | |
11346 | continue; | |
11347 | total_auth_pins += dir->get_auth_pins() + dir->get_dir_auth_pins(); | |
11348 | } | |
11349 | if (total_auth_pins < 0) | |
11350 | continue; | |
11351 | if (info.last_cum_auth_pins != total_auth_pins) { | |
11352 | info.last_cum_auth_pins = total_auth_pins; | |
11353 | info.last_cum_auth_pins_change = now; | |
11354 | continue; | |
11355 | } | |
11356 | if (info.last_cum_auth_pins_change >= cutoff) | |
11357 | continue; | |
11358 | dir = info.dirs.front(); | |
11359 | if (info.num_remote_waiters > 0 || | |
11360 | (!dir->inode->is_root() && dir->get_parent_dir()->is_freezing())) { | |
11361 | dout(10) << " cancel fragmenting " << df << " bit " << info.bits << dendl; | |
11362 | list<CDir*> dirs; | |
11363 | info.dirs.swap(dirs); | |
11364 | fragments.erase(df); | |
11365 | fragment_unmark_unfreeze_dirs(dirs); | |
11366 | } | |
11367 | } | |
11368 | } | |
11369 | ||
11370 | class C_MDC_FragmentPrep : public MDCacheLogContext { | |
11371 | MDRequestRef mdr; | |
11372 | public: | |
11373 | C_MDC_FragmentPrep(MDCache *m, MDRequestRef& r) : MDCacheLogContext(m), mdr(r) {} | |
11374 | void finish(int r) override { | |
11375 | mdcache->_fragment_logged(mdr); | |
11376 | } | |
11377 | }; | |
11378 | ||
11379 | class C_MDC_FragmentStore : public MDCacheContext { | |
11380 | MDRequestRef mdr; | |
11381 | public: | |
11382 | C_MDC_FragmentStore(MDCache *m, MDRequestRef& r) : MDCacheContext(m), mdr(r) {} | |
11383 | void finish(int r) override { | |
11384 | mdcache->_fragment_stored(mdr); | |
11385 | } | |
11386 | }; | |
11387 | ||
11388 | class C_MDC_FragmentCommit : public MDCacheLogContext { | |
11389 | dirfrag_t basedirfrag; | |
a8e16298 | 11390 | MDRequestRef mdr; |
7c673cae | 11391 | public: |
a8e16298 TL |
11392 | C_MDC_FragmentCommit(MDCache *m, dirfrag_t df, const MDRequestRef& r) : |
11393 | MDCacheLogContext(m), basedirfrag(df), mdr(r) {} | |
7c673cae | 11394 | void finish(int r) override { |
a8e16298 | 11395 | mdcache->_fragment_committed(basedirfrag, mdr); |
7c673cae FG |
11396 | } |
11397 | }; | |
11398 | ||
a8e16298 | 11399 | class C_IO_MDC_FragmentPurgeOld : public MDCacheIOContext { |
7c673cae | 11400 | dirfrag_t basedirfrag; |
a8e16298 TL |
11401 | int bits; |
11402 | MDRequestRef mdr; | |
7c673cae | 11403 | public: |
a8e16298 TL |
11404 | C_IO_MDC_FragmentPurgeOld(MDCache *m, dirfrag_t f, int b, |
11405 | const MDRequestRef& r) : | |
11406 | MDCacheIOContext(m), basedirfrag(f), bits(b), mdr(r) {} | |
7c673cae | 11407 | void finish(int r) override { |
11fdf7f2 | 11408 | ceph_assert(r == 0 || r == -ENOENT); |
a8e16298 | 11409 | mdcache->_fragment_old_purged(basedirfrag, bits, mdr); |
7c673cae | 11410 | } |
91327a77 | 11411 | void print(ostream& out) const override { |
a8e16298 | 11412 | out << "fragment_purge_old(" << basedirfrag << ")"; |
91327a77 | 11413 | } |
7c673cae FG |
11414 | }; |
11415 | ||
11416 | void MDCache::fragment_frozen(MDRequestRef& mdr, int r) | |
11417 | { | |
11418 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11419 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11420 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11421 | dout(7) << "fragment_frozen " << basedirfrag << " must have aborted" << dendl; | |
11422 | request_finish(mdr); | |
11423 | return; | |
11424 | } | |
11425 | ||
11fdf7f2 | 11426 | ceph_assert(r == 0); |
7c673cae FG |
11427 | fragment_info_t& info = it->second; |
11428 | dout(10) << "fragment_frozen " << basedirfrag.frag << " by " << info.bits | |
11429 | << " on " << info.dirs.front()->get_inode() << dendl; | |
11430 | ||
11431 | info.all_frozen = true; | |
11432 | dispatch_fragment_dir(mdr); | |
11433 | } | |
11434 | ||
11435 | void MDCache::dispatch_fragment_dir(MDRequestRef& mdr) | |
11436 | { | |
11437 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11438 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11439 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11440 | dout(7) << "dispatch_fragment_dir " << basedirfrag << " must have aborted" << dendl; | |
11441 | request_finish(mdr); | |
11442 | return; | |
11443 | } | |
11444 | ||
11445 | fragment_info_t& info = it->second; | |
11446 | CInode *diri = info.dirs.front()->get_inode(); | |
11447 | ||
11448 | dout(10) << "dispatch_fragment_dir " << basedirfrag << " bits " << info.bits | |
11449 | << " on " << *diri << dendl; | |
11450 | if (!mdr->aborted) { | |
11fdf7f2 TL |
11451 | MutationImpl::LockOpVec lov; |
11452 | lov.add_wrlock(&diri->dirfragtreelock); | |
7c673cae | 11453 | // prevent a racing gather on any other scatterlocks too |
11fdf7f2 TL |
11454 | lov.add_wrlock(&diri->nestlock); |
11455 | lov.add_wrlock(&diri->filelock); | |
11456 | if (!mds->locker->acquire_locks(mdr, lov, NULL, true)) | |
7c673cae FG |
11457 | if (!mdr->aborted) |
11458 | return; | |
11459 | } | |
11460 | ||
11461 | if (mdr->aborted) { | |
11462 | dout(10) << " can't auth_pin " << *diri << ", requeuing dir " | |
11463 | << info.dirs.front()->dirfrag() << dendl; | |
11464 | if (info.bits > 0) | |
11465 | mds->balancer->queue_split(info.dirs.front(), false); | |
11466 | else | |
11467 | mds->balancer->queue_merge(info.dirs.front()); | |
11468 | fragment_unmark_unfreeze_dirs(info.dirs); | |
11469 | fragments.erase(it); | |
11470 | request_finish(mdr); | |
11471 | return; | |
11472 | } | |
11473 | ||
11474 | mdr->ls = mds->mdlog->get_current_segment(); | |
11475 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_PREPARE, basedirfrag, info.bits); | |
11476 | mds->mdlog->start_entry(le); | |
11477 | ||
11478 | for (list<CDir*>::iterator p = info.dirs.begin(); p != info.dirs.end(); ++p) { | |
11479 | CDir *dir = *p; | |
11480 | dirfrag_rollback rollback; | |
11481 | rollback.fnode = dir->fnode; | |
11482 | le->add_orig_frag(dir->get_frag(), &rollback); | |
11483 | } | |
11484 | ||
11485 | // refragment | |
11fdf7f2 | 11486 | MDSContext::vec waiters; |
7c673cae FG |
11487 | adjust_dir_fragments(diri, info.dirs, basedirfrag.frag, info.bits, |
11488 | info.resultfrags, waiters, false); | |
11fdf7f2 | 11489 | if (g_conf()->mds_debug_frag) |
7c673cae FG |
11490 | diri->verify_dirfrags(); |
11491 | mds->queue_waiters(waiters); | |
11492 | ||
11fdf7f2 TL |
11493 | for (const auto& fg : le->orig_frags) |
11494 | ceph_assert(!diri->dirfragtree.is_leaf(fg)); | |
7c673cae FG |
11495 | |
11496 | le->metablob.add_dir_context(*info.resultfrags.begin()); | |
11497 | for (list<CDir*>::iterator p = info.resultfrags.begin(); | |
11498 | p != info.resultfrags.end(); | |
11499 | ++p) { | |
11500 | if (diri->is_auth()) { | |
11501 | le->metablob.add_fragmented_dir(*p, false, false); | |
11502 | } else { | |
11503 | (*p)->state_set(CDir::STATE_DIRTYDFT); | |
11504 | le->metablob.add_fragmented_dir(*p, false, true); | |
11505 | } | |
11506 | } | |
11507 | ||
11508 | // dft lock | |
11509 | if (diri->is_auth()) { | |
11510 | // journal dirfragtree | |
94b18763 FG |
11511 | auto &pi = diri->project_inode(); |
11512 | pi.inode.version = diri->pre_dirty(); | |
7c673cae FG |
11513 | journal_dirty_inode(mdr.get(), &le->metablob, diri); |
11514 | } else { | |
11515 | mds->locker->mark_updated_scatterlock(&diri->dirfragtreelock); | |
11516 | mdr->ls->dirty_dirfrag_dirfragtree.push_back(&diri->item_dirty_dirfrag_dirfragtree); | |
11517 | mdr->add_updated_lock(&diri->dirfragtreelock); | |
11518 | } | |
11519 | ||
11520 | /* | |
11521 | // filelock | |
11522 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
11523 | mut->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
11524 | mut->add_updated_lock(&diri->filelock); | |
11525 | ||
11526 | // dirlock | |
11527 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
11528 | mut->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
11529 | mut->add_updated_lock(&diri->nestlock); | |
11530 | */ | |
11531 | ||
11532 | add_uncommitted_fragment(basedirfrag, info.bits, le->orig_frags, mdr->ls); | |
11533 | mds->server->submit_mdlog_entry(le, new C_MDC_FragmentPrep(this, mdr), | |
11534 | mdr, __func__); | |
11535 | mds->mdlog->flush(); | |
11536 | } | |
11537 | ||
11538 | void MDCache::_fragment_logged(MDRequestRef& mdr) | |
11539 | { | |
11540 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
a8e16298 | 11541 | auto& info = fragments.at(basedirfrag); |
7c673cae FG |
11542 | CInode *diri = info.resultfrags.front()->get_inode(); |
11543 | ||
11544 | dout(10) << "fragment_logged " << basedirfrag << " bits " << info.bits | |
11545 | << " on " << *diri << dendl; | |
a8e16298 | 11546 | mdr->mark_event("prepare logged"); |
7c673cae FG |
11547 | |
11548 | if (diri->is_auth()) | |
11549 | diri->pop_and_dirty_projected_inode(mdr->ls); | |
11550 | ||
11551 | mdr->apply(); // mark scatterlock | |
11552 | ||
11553 | // store resulting frags | |
11554 | MDSGatherBuilder gather(g_ceph_context, new C_MDC_FragmentStore(this, mdr)); | |
11555 | ||
11556 | for (list<CDir*>::iterator p = info.resultfrags.begin(); | |
11557 | p != info.resultfrags.end(); | |
11558 | ++p) { | |
11559 | CDir *dir = *p; | |
11560 | dout(10) << " storing result frag " << *dir << dendl; | |
11561 | ||
11562 | // freeze and store them too | |
11563 | dir->auth_pin(this); | |
11564 | dir->state_set(CDir::STATE_FRAGMENTING); | |
11565 | dir->commit(0, gather.new_sub(), true); // ignore authpinnability | |
11566 | } | |
11567 | ||
11568 | gather.activate(); | |
11569 | } | |
11570 | ||
11571 | void MDCache::_fragment_stored(MDRequestRef& mdr) | |
11572 | { | |
11573 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
a8e16298 TL |
11574 | fragment_info_t &info = fragments.at(basedirfrag); |
11575 | CDir *first = info.resultfrags.front(); | |
11576 | CInode *diri = first->get_inode(); | |
7c673cae FG |
11577 | |
11578 | dout(10) << "fragment_stored " << basedirfrag << " bits " << info.bits | |
11579 | << " on " << *diri << dendl; | |
a8e16298 | 11580 | mdr->mark_event("new frags stored"); |
7c673cae FG |
11581 | |
11582 | // tell peers | |
a8e16298 TL |
11583 | mds_rank_t diri_auth = (first->is_subtree_root() && !diri->is_auth()) ? |
11584 | diri->authority().first : CDIR_AUTH_UNKNOWN; | |
181888fb FG |
11585 | for (const auto &p : first->get_replicas()) { |
11586 | if (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || | |
11587 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
11588 | rejoin_gather.count(p.first))) | |
7c673cae FG |
11589 | continue; |
11590 | ||
11fdf7f2 | 11591 | auto notify = MMDSFragmentNotify::create(basedirfrag, info.bits, mdr->reqid.tid); |
a8e16298 TL |
11592 | if (diri_auth != CDIR_AUTH_UNKNOWN && // subtree root |
11593 | diri_auth != p.first) { // not auth mds of diri | |
11594 | /* | |
11595 | * In the nornal case, mds does not trim dir inode whose child dirfrags | |
11596 | * are likely being fragmented (see trim_inode()). But when fragmenting | |
11597 | * subtree roots, following race can happen: | |
11598 | * | |
11599 | * - mds.a (auth mds of dirfrag) sends fragment_notify message to | |
11600 | * mds.c and drops wrlock on dirfragtreelock. | |
11601 | * - mds.b (auth mds of dir inode) changes dirfragtreelock state to | |
11602 | * SYNC and send lock message mds.c | |
11603 | * - mds.c receives the lock message and changes dirfragtreelock state | |
11604 | * to SYNC | |
11605 | * - mds.c trim dirfrag and dir inode from its cache | |
11606 | * - mds.c receives the fragment_notify message | |
11607 | * | |
11608 | * So we need to ensure replicas have received the notify, then unlock | |
11609 | * the dirfragtreelock. | |
11610 | */ | |
11611 | notify->mark_ack_wanted(); | |
11612 | info.notify_ack_waiting.insert(p.first); | |
11613 | } | |
7c673cae FG |
11614 | |
11615 | // freshly replicate new dirs to peers | |
11616 | for (list<CDir*>::iterator q = info.resultfrags.begin(); | |
11617 | q != info.resultfrags.end(); | |
11618 | ++q) | |
181888fb | 11619 | replicate_dir(*q, p.first, notify->basebl); |
7c673cae | 11620 | |
181888fb | 11621 | mds->send_message_mds(notify, p.first); |
7c673cae FG |
11622 | } |
11623 | ||
11624 | // journal commit | |
11625 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_COMMIT, basedirfrag, info.bits); | |
a8e16298 | 11626 | mds->mdlog->start_submit_entry(le, new C_MDC_FragmentCommit(this, basedirfrag, mdr)); |
7c673cae | 11627 | |
7c673cae FG |
11628 | |
11629 | // unfreeze resulting frags | |
11630 | for (list<CDir*>::iterator p = info.resultfrags.begin(); | |
11631 | p != info.resultfrags.end(); | |
11632 | ++p) { | |
11633 | CDir *dir = *p; | |
11634 | dout(10) << " result frag " << *dir << dendl; | |
11635 | ||
94b18763 FG |
11636 | for (auto &p : dir->items) { |
11637 | CDentry *dn = p.second; | |
11fdf7f2 | 11638 | ceph_assert(dn->state_test(CDentry::STATE_FRAGMENTING)); |
7c673cae FG |
11639 | dn->state_clear(CDentry::STATE_FRAGMENTING); |
11640 | dn->put(CDentry::PIN_FRAGMENTING); | |
11641 | } | |
11642 | ||
11643 | // unfreeze | |
11644 | dir->unfreeze_dir(); | |
11645 | } | |
11646 | ||
a8e16298 TL |
11647 | if (info.notify_ack_waiting.empty()) { |
11648 | fragment_drop_locks(info); | |
11649 | } else { | |
11650 | mds->locker->drop_locks_for_fragment_unfreeze(mdr.get()); | |
11651 | } | |
7c673cae FG |
11652 | } |
11653 | ||
a8e16298 | 11654 | void MDCache::_fragment_committed(dirfrag_t basedirfrag, const MDRequestRef& mdr) |
7c673cae FG |
11655 | { |
11656 | dout(10) << "fragment_committed " << basedirfrag << dendl; | |
a8e16298 TL |
11657 | if (mdr) |
11658 | mdr->mark_event("commit logged"); | |
11659 | ||
11660 | ufragment &uf = uncommitted_fragments.at(basedirfrag); | |
7c673cae FG |
11661 | |
11662 | // remove old frags | |
11663 | C_GatherBuilder gather( | |
11664 | g_ceph_context, | |
11665 | new C_OnFinisher( | |
a8e16298 | 11666 | new C_IO_MDC_FragmentPurgeOld(this, basedirfrag, uf.bits, mdr), |
7c673cae FG |
11667 | mds->finisher)); |
11668 | ||
11669 | SnapContext nullsnapc; | |
11670 | object_locator_t oloc(mds->mdsmap->get_metadata_pool()); | |
11fdf7f2 TL |
11671 | for (const auto& fg : uf.old_frags) { |
11672 | object_t oid = CInode::get_object_name(basedirfrag.ino, fg, ""); | |
7c673cae | 11673 | ObjectOperation op; |
11fdf7f2 | 11674 | if (fg == frag_t()) { |
7c673cae FG |
11675 | // backtrace object |
11676 | dout(10) << " truncate orphan dirfrag " << oid << dendl; | |
11677 | op.truncate(0); | |
11678 | op.omap_clear(); | |
11679 | } else { | |
11680 | dout(10) << " removing orphan dirfrag " << oid << dendl; | |
11681 | op.remove(); | |
11682 | } | |
11683 | mds->objecter->mutate(oid, oloc, op, nullsnapc, | |
11684 | ceph::real_clock::now(), | |
11685 | 0, gather.new_sub()); | |
11686 | } | |
11687 | ||
11fdf7f2 | 11688 | ceph_assert(gather.has_subs()); |
7c673cae FG |
11689 | gather.activate(); |
11690 | } | |
11691 | ||
a8e16298 | 11692 | void MDCache::_fragment_old_purged(dirfrag_t basedirfrag, int bits, const MDRequestRef& mdr) |
7c673cae | 11693 | { |
a8e16298 TL |
11694 | dout(10) << "fragment_old_purged " << basedirfrag << dendl; |
11695 | if (mdr) | |
11696 | mdr->mark_event("old frags purged"); | |
11697 | ||
11698 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_FINISH, basedirfrag, bits); | |
11699 | mds->mdlog->start_submit_entry(le); | |
11700 | ||
11701 | finish_uncommitted_fragment(basedirfrag, EFragment::OP_FINISH); | |
11702 | ||
11703 | if (mds->logger) { | |
11704 | if (bits > 0) { | |
11705 | mds->logger->inc(l_mds_dir_split); | |
11706 | } else { | |
11707 | mds->logger->inc(l_mds_dir_merge); | |
11708 | } | |
11709 | } | |
11710 | ||
11711 | if (mdr) { | |
11712 | auto it = fragments.find(basedirfrag); | |
11713 | ceph_assert(it != fragments.end()); | |
11714 | it->second.finishing = true; | |
11715 | if (it->second.notify_ack_waiting.empty()) | |
11716 | fragment_maybe_finish(it); | |
11717 | else | |
11718 | mdr->mark_event("wating for notify acks"); | |
11719 | } | |
11720 | } | |
11721 | ||
11722 | void MDCache::fragment_drop_locks(fragment_info_t& info) | |
11723 | { | |
11724 | mds->locker->drop_locks(info.mdr.get()); | |
11725 | request_finish(info.mdr); | |
11726 | //info.mdr.reset(); | |
11727 | } | |
11728 | ||
11729 | void MDCache::fragment_maybe_finish(const fragment_info_iterator& it) | |
11730 | { | |
11731 | if (!it->second.finishing) | |
11732 | return; | |
7c673cae FG |
11733 | |
11734 | // unmark & auth_unpin | |
a8e16298 | 11735 | for (const auto &dir : it->second.resultfrags) { |
7c673cae FG |
11736 | dir->state_clear(CDir::STATE_FRAGMENTING); |
11737 | dir->auth_unpin(this); | |
11738 | ||
11739 | // In case the resulting fragments are beyond the split size, | |
11740 | // we might need to split them again right away (they could | |
11741 | // have been taking inserts between unfreezing and getting | |
11742 | // here) | |
11743 | mds->balancer->maybe_fragment(dir, false); | |
11744 | } | |
11745 | ||
a8e16298 TL |
11746 | fragments.erase(it); |
11747 | } | |
11748 | ||
11749 | ||
11fdf7f2 | 11750 | void MDCache::handle_fragment_notify_ack(const MMDSFragmentNotifyAck::const_ref &ack) |
a8e16298 TL |
11751 | { |
11752 | dout(10) << "handle_fragment_notify_ack " << *ack << " from " << ack->get_source() << dendl; | |
11753 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
11754 | ||
11755 | if (mds->get_state() < MDSMap::STATE_ACTIVE) { | |
a8e16298 | 11756 | return; |
7c673cae FG |
11757 | } |
11758 | ||
a8e16298 TL |
11759 | auto it = fragments.find(ack->get_base_dirfrag()); |
11760 | if (it == fragments.end() || | |
11761 | it->second.get_tid() != ack->get_tid()) { | |
11762 | dout(10) << "handle_fragment_notify_ack obsolete message, dropping" << dendl; | |
a8e16298 TL |
11763 | return; |
11764 | } | |
7c673cae | 11765 | |
a8e16298 TL |
11766 | if (it->second.notify_ack_waiting.erase(from) && |
11767 | it->second.notify_ack_waiting.empty()) { | |
11768 | fragment_drop_locks(it->second); | |
11769 | fragment_maybe_finish(it); | |
11770 | } | |
7c673cae FG |
11771 | } |
11772 | ||
11fdf7f2 | 11773 | void MDCache::handle_fragment_notify(const MMDSFragmentNotify::const_ref ¬ify) |
7c673cae FG |
11774 | { |
11775 | dout(10) << "handle_fragment_notify " << *notify << " from " << notify->get_source() << dendl; | |
a8e16298 | 11776 | mds_rank_t from = mds_rank_t(notify->get_source().num()); |
7c673cae FG |
11777 | |
11778 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
7c673cae FG |
11779 | return; |
11780 | } | |
11781 | ||
11782 | CInode *diri = get_inode(notify->get_ino()); | |
11783 | if (diri) { | |
11784 | frag_t base = notify->get_basefrag(); | |
11785 | int bits = notify->get_bits(); | |
11786 | ||
11787 | /* | |
11788 | if ((bits < 0 && diri->dirfragtree.is_leaf(base)) || | |
11789 | (bits > 0 && !diri->dirfragtree.is_leaf(base))) { | |
11790 | dout(10) << " dft " << diri->dirfragtree << " state doesn't match " << base << " by " << bits | |
11791 | << ", must have found out during resolve/rejoin? ignoring. " << *diri << dendl; | |
7c673cae FG |
11792 | return; |
11793 | } | |
11794 | */ | |
11795 | ||
11796 | // refragment | |
11fdf7f2 | 11797 | MDSContext::vec waiters; |
7c673cae FG |
11798 | list<CDir*> resultfrags; |
11799 | adjust_dir_fragments(diri, base, bits, resultfrags, waiters, false); | |
11fdf7f2 | 11800 | if (g_conf()->mds_debug_frag) |
7c673cae FG |
11801 | diri->verify_dirfrags(); |
11802 | ||
11803 | for (list<CDir*>::iterator p = resultfrags.begin(); p != resultfrags.end(); ++p) | |
11804 | diri->take_dir_waiting((*p)->get_frag(), waiters); | |
11805 | ||
11806 | // add new replica dirs values | |
11fdf7f2 | 11807 | auto p = notify->basebl.cbegin(); |
7c673cae | 11808 | while (!p.end()) |
a8e16298 | 11809 | add_replica_dir(p, diri, from, waiters); |
7c673cae FG |
11810 | |
11811 | mds->queue_waiters(waiters); | |
11812 | } else { | |
11813 | ceph_abort(); | |
11814 | } | |
11815 | ||
a8e16298 | 11816 | if (notify->is_ack_wanted()) { |
11fdf7f2 TL |
11817 | auto ack = MMDSFragmentNotifyAck::create(notify->get_base_dirfrag(), |
11818 | notify->get_bits(), notify->get_tid()); | |
a8e16298 TL |
11819 | mds->send_message_mds(ack, from); |
11820 | } | |
7c673cae FG |
11821 | } |
11822 | ||
11fdf7f2 | 11823 | void MDCache::add_uncommitted_fragment(dirfrag_t basedirfrag, int bits, const frag_vec_t& old_frags, |
7c673cae FG |
11824 | LogSegment *ls, bufferlist *rollback) |
11825 | { | |
11826 | dout(10) << "add_uncommitted_fragment: base dirfrag " << basedirfrag << " bits " << bits << dendl; | |
11fdf7f2 | 11827 | ceph_assert(!uncommitted_fragments.count(basedirfrag)); |
7c673cae FG |
11828 | ufragment& uf = uncommitted_fragments[basedirfrag]; |
11829 | uf.old_frags = old_frags; | |
11830 | uf.bits = bits; | |
11831 | uf.ls = ls; | |
11832 | ls->uncommitted_fragments.insert(basedirfrag); | |
11833 | if (rollback) | |
11834 | uf.rollback.swap(*rollback); | |
11835 | } | |
11836 | ||
11837 | void MDCache::finish_uncommitted_fragment(dirfrag_t basedirfrag, int op) | |
11838 | { | |
11839 | dout(10) << "finish_uncommitted_fragments: base dirfrag " << basedirfrag | |
11840 | << " op " << EFragment::op_name(op) << dendl; | |
11841 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
11842 | if (it != uncommitted_fragments.end()) { | |
11843 | ufragment& uf = it->second; | |
11844 | if (op != EFragment::OP_FINISH && !uf.old_frags.empty()) { | |
11845 | uf.committed = true; | |
11846 | } else { | |
11847 | uf.ls->uncommitted_fragments.erase(basedirfrag); | |
11848 | mds->queue_waiters(uf.waiters); | |
11849 | uncommitted_fragments.erase(it); | |
11850 | } | |
11851 | } | |
11852 | } | |
11853 | ||
11fdf7f2 | 11854 | void MDCache::rollback_uncommitted_fragment(dirfrag_t basedirfrag, frag_vec_t&& old_frags) |
7c673cae FG |
11855 | { |
11856 | dout(10) << "rollback_uncommitted_fragment: base dirfrag " << basedirfrag | |
11857 | << " old_frags (" << old_frags << ")" << dendl; | |
11858 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
11859 | if (it != uncommitted_fragments.end()) { | |
11860 | ufragment& uf = it->second; | |
11861 | if (!uf.old_frags.empty()) { | |
11fdf7f2 | 11862 | uf.old_frags = std::move(old_frags); |
7c673cae FG |
11863 | uf.committed = true; |
11864 | } else { | |
11865 | uf.ls->uncommitted_fragments.erase(basedirfrag); | |
11866 | uncommitted_fragments.erase(it); | |
11867 | } | |
11868 | } | |
11869 | } | |
11870 | ||
11871 | void MDCache::rollback_uncommitted_fragments() | |
11872 | { | |
11873 | dout(10) << "rollback_uncommitted_fragments: " << uncommitted_fragments.size() << " pending" << dendl; | |
11874 | for (map<dirfrag_t, ufragment>::iterator p = uncommitted_fragments.begin(); | |
11875 | p != uncommitted_fragments.end(); | |
11876 | ++p) { | |
11877 | ufragment &uf = p->second; | |
11878 | CInode *diri = get_inode(p->first.ino); | |
11fdf7f2 | 11879 | ceph_assert(diri); |
7c673cae FG |
11880 | |
11881 | if (uf.committed) { | |
a8e16298 | 11882 | _fragment_committed(p->first, MDRequestRef()); |
7c673cae FG |
11883 | continue; |
11884 | } | |
11885 | ||
11886 | dout(10) << " rolling back " << p->first << " refragment by " << uf.bits << " bits" << dendl; | |
11887 | ||
11888 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
11889 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_ROLLBACK, p->first, uf.bits); | |
11890 | mds->mdlog->start_entry(le); | |
11891 | bool diri_auth = (diri->authority() != CDIR_AUTH_UNDEF); | |
11892 | ||
11fdf7f2 | 11893 | frag_vec_t old_frags; |
7c673cae FG |
11894 | diri->dirfragtree.get_leaves_under(p->first.frag, old_frags); |
11895 | ||
11896 | list<CDir*> resultfrags; | |
11897 | if (uf.old_frags.empty()) { | |
11898 | // created by old format EFragment | |
11fdf7f2 | 11899 | MDSContext::vec waiters; |
7c673cae FG |
11900 | adjust_dir_fragments(diri, p->first.frag, -uf.bits, resultfrags, waiters, true); |
11901 | } else { | |
11fdf7f2 TL |
11902 | auto bp = uf.rollback.cbegin(); |
11903 | for (const auto& fg : uf.old_frags) { | |
11904 | CDir *dir = force_dir_fragment(diri, fg); | |
7c673cae FG |
11905 | resultfrags.push_back(dir); |
11906 | ||
11907 | dirfrag_rollback rollback; | |
11fdf7f2 | 11908 | decode(rollback, bp); |
7c673cae FG |
11909 | |
11910 | dir->set_version(rollback.fnode.version); | |
11911 | dir->fnode = rollback.fnode; | |
11912 | ||
11913 | dir->_mark_dirty(ls); | |
11914 | ||
11915 | if (!(dir->fnode.rstat == dir->fnode.accounted_rstat)) { | |
11916 | dout(10) << " dirty nestinfo on " << *dir << dendl; | |
11917 | mds->locker->mark_updated_scatterlock(&dir->inode->nestlock); | |
11918 | ls->dirty_dirfrag_nest.push_back(&dir->inode->item_dirty_dirfrag_nest); | |
11919 | } | |
11920 | if (!(dir->fnode.fragstat == dir->fnode.accounted_fragstat)) { | |
11921 | dout(10) << " dirty fragstat on " << *dir << dendl; | |
11922 | mds->locker->mark_updated_scatterlock(&dir->inode->filelock); | |
11923 | ls->dirty_dirfrag_dir.push_back(&dir->inode->item_dirty_dirfrag_dir); | |
11924 | } | |
11925 | ||
11926 | le->add_orig_frag(dir->get_frag()); | |
11927 | le->metablob.add_dir_context(dir); | |
11928 | if (diri_auth) { | |
11929 | le->metablob.add_fragmented_dir(dir, true, false); | |
11930 | } else { | |
11931 | dout(10) << " dirty dirfragtree on " << *dir << dendl; | |
11932 | dir->state_set(CDir::STATE_DIRTYDFT); | |
11933 | le->metablob.add_fragmented_dir(dir, true, true); | |
11934 | } | |
11935 | } | |
11936 | } | |
11937 | ||
11938 | if (diri_auth) { | |
94b18763 FG |
11939 | auto &pi = diri->project_inode(); |
11940 | pi.inode.version = diri->pre_dirty(); | |
7c673cae FG |
11941 | diri->pop_and_dirty_projected_inode(ls); // hacky |
11942 | le->metablob.add_primary_dentry(diri->get_projected_parent_dn(), diri, true); | |
11943 | } else { | |
11944 | mds->locker->mark_updated_scatterlock(&diri->dirfragtreelock); | |
11945 | ls->dirty_dirfrag_dirfragtree.push_back(&diri->item_dirty_dirfrag_dirfragtree); | |
11946 | } | |
11947 | ||
11fdf7f2 | 11948 | if (g_conf()->mds_debug_frag) |
7c673cae FG |
11949 | diri->verify_dirfrags(); |
11950 | ||
11fdf7f2 TL |
11951 | for (const auto& leaf : old_frags) { |
11952 | ceph_assert(!diri->dirfragtree.is_leaf(leaf)); | |
11953 | } | |
7c673cae | 11954 | |
7c673cae FG |
11955 | mds->mdlog->submit_entry(le); |
11956 | ||
11957 | uf.old_frags.swap(old_frags); | |
a8e16298 | 11958 | _fragment_committed(p->first, MDRequestRef()); |
7c673cae FG |
11959 | } |
11960 | } | |
11961 | ||
11962 | void MDCache::force_readonly() | |
11963 | { | |
11964 | if (is_readonly()) | |
11965 | return; | |
11966 | ||
11967 | dout(1) << "force file system read-only" << dendl; | |
11968 | mds->clog->warn() << "force file system read-only"; | |
11969 | ||
11970 | set_readonly(); | |
11971 | ||
11972 | mds->server->force_clients_readonly(); | |
11973 | ||
11974 | // revoke write caps | |
81eedcae | 11975 | int count = 0; |
94b18763 | 11976 | for (auto &p : inode_map) { |
b32b8144 | 11977 | CInode *in = p.second; |
7c673cae FG |
11978 | if (in->is_head()) |
11979 | mds->locker->eval(in, CEPH_CAP_LOCKS); | |
81eedcae TL |
11980 | if (!(++count % 1000)) |
11981 | mds->heartbeat_reset(); | |
7c673cae FG |
11982 | } |
11983 | ||
11984 | mds->mdlog->flush(); | |
11985 | } | |
11986 | ||
11987 | ||
11988 | // ============================================================== | |
11989 | // debug crap | |
11990 | ||
81eedcae | 11991 | void MDCache::show_subtrees(int dbl, bool force_print) |
7c673cae | 11992 | { |
11fdf7f2 | 11993 | if (g_conf()->mds_thrash_exports) |
7c673cae FG |
11994 | dbl += 15; |
11995 | ||
11996 | //dout(10) << "show_subtrees" << dendl; | |
11997 | ||
11fdf7f2 | 11998 | if (!g_conf()->subsys.should_gather(ceph_subsys_mds, dbl)) |
7c673cae FG |
11999 | return; // i won't print anything. |
12000 | ||
12001 | if (subtrees.empty()) { | |
11fdf7f2 TL |
12002 | dout(ceph::dout::need_dynamic(dbl)) << "show_subtrees - no subtrees" |
12003 | << dendl; | |
7c673cae FG |
12004 | return; |
12005 | } | |
12006 | ||
81eedcae TL |
12007 | if (!force_print && subtrees.size() > SUBTREES_COUNT_THRESHOLD && |
12008 | !g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) { | |
12009 | dout(ceph::dout::need_dynamic(dbl)) << "number of subtrees = " << subtrees.size() << "; not " | |
12010 | "printing subtrees" << dendl; | |
12011 | return; | |
12012 | } | |
12013 | ||
7c673cae FG |
12014 | // root frags |
12015 | list<CDir*> basefrags; | |
12016 | for (set<CInode*>::iterator p = base_inodes.begin(); | |
12017 | p != base_inodes.end(); | |
12018 | ++p) | |
12019 | (*p)->get_dirfrags(basefrags); | |
12020 | //dout(15) << "show_subtrees, base dirfrags " << basefrags << dendl; | |
12021 | dout(15) << "show_subtrees" << dendl; | |
12022 | ||
12023 | // queue stuff | |
12024 | list<pair<CDir*,int> > q; | |
12025 | string indent; | |
12026 | set<CDir*> seen; | |
12027 | ||
12028 | // calc max depth | |
12029 | for (list<CDir*>::iterator p = basefrags.begin(); p != basefrags.end(); ++p) | |
12030 | q.push_back(pair<CDir*,int>(*p, 0)); | |
12031 | ||
12032 | set<CDir*> subtrees_seen; | |
12033 | ||
81eedcae | 12034 | unsigned int depth = 0; |
7c673cae FG |
12035 | while (!q.empty()) { |
12036 | CDir *dir = q.front().first; | |
81eedcae | 12037 | unsigned int d = q.front().second; |
7c673cae FG |
12038 | q.pop_front(); |
12039 | ||
12040 | if (subtrees.count(dir) == 0) continue; | |
12041 | ||
12042 | subtrees_seen.insert(dir); | |
12043 | ||
12044 | if (d > depth) depth = d; | |
12045 | ||
12046 | // sanity check | |
12047 | //dout(25) << "saw depth " << d << " " << *dir << dendl; | |
12048 | if (seen.count(dir)) dout(0) << "aah, already seen " << *dir << dendl; | |
11fdf7f2 | 12049 | ceph_assert(seen.count(dir) == 0); |
7c673cae FG |
12050 | seen.insert(dir); |
12051 | ||
12052 | // nested items? | |
12053 | if (!subtrees[dir].empty()) { | |
12054 | for (set<CDir*>::iterator p = subtrees[dir].begin(); | |
12055 | p != subtrees[dir].end(); | |
12056 | ++p) { | |
12057 | //dout(25) << " saw sub " << **p << dendl; | |
12058 | q.push_front(pair<CDir*,int>(*p, d+1)); | |
12059 | } | |
12060 | } | |
12061 | } | |
12062 | ||
81eedcae TL |
12063 | if (!force_print && depth > SUBTREES_DEPTH_THRESHOLD && |
12064 | !g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) { | |
12065 | dout(ceph::dout::need_dynamic(dbl)) << "max depth among subtrees = " << depth << "; not printing " | |
12066 | "subtrees" << dendl; | |
12067 | return; | |
12068 | } | |
7c673cae FG |
12069 | |
12070 | // print tree | |
12071 | for (list<CDir*>::iterator p = basefrags.begin(); p != basefrags.end(); ++p) | |
12072 | q.push_back(pair<CDir*,int>(*p, 0)); | |
12073 | ||
12074 | while (!q.empty()) { | |
12075 | CDir *dir = q.front().first; | |
12076 | int d = q.front().second; | |
12077 | q.pop_front(); | |
12078 | ||
12079 | if (subtrees.count(dir) == 0) continue; | |
12080 | ||
12081 | // adjust indenter | |
12082 | while ((unsigned)d < indent.size()) | |
12083 | indent.resize(d); | |
12084 | ||
12085 | // pad | |
12086 | string pad = "______________________________________"; | |
12087 | pad.resize(depth*2+1-indent.size()); | |
12088 | if (!subtrees[dir].empty()) | |
12089 | pad[0] = '.'; // parent | |
12090 | ||
12091 | ||
12092 | string auth; | |
12093 | if (dir->is_auth()) | |
12094 | auth = "auth "; | |
12095 | else | |
12096 | auth = " rep "; | |
12097 | ||
12098 | char s[10]; | |
12099 | if (dir->get_dir_auth().second == CDIR_AUTH_UNKNOWN) | |
12100 | snprintf(s, sizeof(s), "%2d ", int(dir->get_dir_auth().first)); | |
12101 | else | |
12102 | snprintf(s, sizeof(s), "%2d,%2d", int(dir->get_dir_auth().first), int(dir->get_dir_auth().second)); | |
12103 | ||
12104 | ||
11fdf7f2 TL |
12105 | dout(ceph::dout::need_dynamic(dbl)) << indent << "|_" << pad << s |
12106 | << " " << auth << *dir << dendl; | |
7c673cae FG |
12107 | |
12108 | if (dir->ino() == MDS_INO_ROOT) | |
11fdf7f2 | 12109 | ceph_assert(dir->inode == root); |
7c673cae | 12110 | if (dir->ino() == MDS_INO_MDSDIR(mds->get_nodeid())) |
11fdf7f2 | 12111 | ceph_assert(dir->inode == myin); |
7c673cae | 12112 | if (dir->inode->is_stray() && (MDS_INO_STRAY_OWNER(dir->ino()) == mds->get_nodeid())) |
11fdf7f2 | 12113 | ceph_assert(strays[MDS_INO_STRAY_INDEX(dir->ino())] == dir->inode); |
7c673cae FG |
12114 | |
12115 | // nested items? | |
12116 | if (!subtrees[dir].empty()) { | |
12117 | // more at my level? | |
12118 | if (!q.empty() && q.front().second == d) | |
12119 | indent += "| "; | |
12120 | else | |
12121 | indent += " "; | |
12122 | ||
12123 | for (set<CDir*>::iterator p = subtrees[dir].begin(); | |
12124 | p != subtrees[dir].end(); | |
12125 | ++p) | |
12126 | q.push_front(pair<CDir*,int>(*p, d+2)); | |
12127 | } | |
12128 | } | |
12129 | ||
12130 | // verify there isn't stray crap in subtree map | |
12131 | int lost = 0; | |
12132 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
12133 | p != subtrees.end(); | |
12134 | ++p) { | |
12135 | if (subtrees_seen.count(p->first)) continue; | |
12136 | dout(10) << "*** stray/lost entry in subtree map: " << *p->first << dendl; | |
12137 | lost++; | |
12138 | } | |
11fdf7f2 | 12139 | ceph_assert(lost == 0); |
7c673cae FG |
12140 | } |
12141 | ||
7c673cae FG |
12142 | void MDCache::show_cache() |
12143 | { | |
12144 | dout(7) << "show_cache" << dendl; | |
b32b8144 FG |
12145 | |
12146 | auto show_func = [this](CInode *in) { | |
7c673cae | 12147 | // unlinked? |
b32b8144 FG |
12148 | if (!in->parent) |
12149 | dout(7) << " unlinked " << *in << dendl; | |
12150 | ||
7c673cae FG |
12151 | // dirfrags? |
12152 | list<CDir*> dfs; | |
b32b8144 | 12153 | in->get_dirfrags(dfs); |
7c673cae FG |
12154 | for (list<CDir*>::iterator p = dfs.begin(); p != dfs.end(); ++p) { |
12155 | CDir *dir = *p; | |
12156 | dout(7) << " dirfrag " << *dir << dendl; | |
b32b8144 | 12157 | |
94b18763 FG |
12158 | for (auto &p : dir->items) { |
12159 | CDentry *dn = p.second; | |
7c673cae FG |
12160 | dout(7) << " dentry " << *dn << dendl; |
12161 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
12162 | if (dnl->is_primary() && dnl->get_inode()) | |
12163 | dout(7) << " inode " << *dnl->get_inode() << dendl; | |
12164 | } | |
12165 | } | |
b32b8144 FG |
12166 | }; |
12167 | ||
94b18763 | 12168 | for (auto &p : inode_map) |
b32b8144 | 12169 | show_func(p.second); |
94b18763 | 12170 | for (auto &p : snap_inode_map) |
b32b8144 | 12171 | show_func(p.second); |
7c673cae FG |
12172 | } |
12173 | ||
f64942e4 | 12174 | void MDCache::cache_status(Formatter *f) |
181888fb FG |
12175 | { |
12176 | f->open_object_section("cache"); | |
12177 | ||
12178 | f->open_object_section("pool"); | |
12179 | mempool::get_pool(mempool::mds_co::id).dump(f); | |
12180 | f->close_section(); | |
12181 | ||
12182 | f->close_section(); | |
181888fb FG |
12183 | } |
12184 | ||
11fdf7f2 | 12185 | void MDCache::dump_tree(CInode *in, const int cur_depth, const int max_depth, Formatter *f) |
7c673cae | 12186 | { |
11fdf7f2 TL |
12187 | ceph_assert(in); |
12188 | if ((max_depth >= 0) && (cur_depth > max_depth)) { | |
12189 | return; | |
12190 | } | |
12191 | list<CDir*> ls; | |
12192 | in->get_dirfrags(ls); | |
12193 | for (const auto &subdir : ls) { | |
12194 | for (const auto &p : subdir->items) { | |
12195 | CDentry *dn = p.second; | |
12196 | CInode *in = dn->get_linkage()->get_inode(); | |
12197 | if (in) { | |
12198 | dump_tree(in, cur_depth + 1, max_depth, f); | |
12199 | } | |
12200 | } | |
12201 | } | |
12202 | f->open_object_section("inode"); | |
12203 | in->dump(f, CInode::DUMP_DEFAULT | CInode::DUMP_DIRFRAGS); | |
12204 | f->close_section(); | |
7c673cae FG |
12205 | } |
12206 | ||
11fdf7f2 | 12207 | int MDCache::dump_cache(std::string_view file_name) |
7c673cae | 12208 | { |
11fdf7f2 | 12209 | return dump_cache(file_name, NULL); |
7c673cae FG |
12210 | } |
12211 | ||
11fdf7f2 | 12212 | int MDCache::dump_cache(Formatter *f) |
7c673cae | 12213 | { |
11fdf7f2 | 12214 | return dump_cache(std::string_view(""), f); |
7c673cae FG |
12215 | } |
12216 | ||
12217 | /** | |
12218 | * Dump the metadata cache, either to a Formatter, if | |
12219 | * provided, else to a plain text file. | |
12220 | */ | |
11fdf7f2 | 12221 | int MDCache::dump_cache(std::string_view fn, Formatter *f) |
7c673cae FG |
12222 | { |
12223 | int r = 0; | |
f64942e4 AA |
12224 | |
12225 | // dumping large caches may cause mds to hang or worse get killed. | |
12226 | // so, disallow the dump if the cache size exceeds the configured | |
12227 | // threshold, which is 1G for formatter and unlimited for file (note | |
12228 | // that this can be jacked up by the admin... and is nothing but foot | |
12229 | // shooting, but the option itself is for devs and hence dangerous to | |
12230 | // tune). TODO: remove this when fixed. | |
12231 | uint64_t threshold = f ? | |
11fdf7f2 TL |
12232 | g_conf().get_val<Option::size_t>("mds_dump_cache_threshold_formatter") : |
12233 | g_conf().get_val<Option::size_t>("mds_dump_cache_threshold_file"); | |
f64942e4 AA |
12234 | |
12235 | if (threshold && cache_size() > threshold) { | |
12236 | if (f) { | |
12237 | std::stringstream ss; | |
12238 | ss << "cache usage exceeds dump threshold"; | |
12239 | f->open_object_section("result"); | |
12240 | f->dump_string("error", ss.str()); | |
12241 | f->close_section(); | |
12242 | } else { | |
12243 | derr << "cache usage exceeds dump threshold" << dendl; | |
12244 | r = -EINVAL; | |
12245 | } | |
12246 | return r; | |
12247 | } | |
12248 | ||
12249 | r = 0; | |
7c673cae FG |
12250 | int fd = -1; |
12251 | ||
12252 | if (f) { | |
12253 | f->open_array_section("inodes"); | |
12254 | } else { | |
94b18763 FG |
12255 | char path[PATH_MAX] = ""; |
12256 | if (fn.length()) { | |
12257 | snprintf(path, sizeof path, "%s", fn.data()); | |
12258 | } else { | |
12259 | snprintf(path, sizeof path, "cachedump.%d.mds%d", (int)mds->mdsmap->get_epoch(), int(mds->get_nodeid())); | |
7c673cae FG |
12260 | } |
12261 | ||
94b18763 | 12262 | dout(1) << "dump_cache to " << path << dendl; |
7c673cae | 12263 | |
91327a77 | 12264 | fd = ::open(path, O_WRONLY|O_CREAT|O_EXCL|O_CLOEXEC, 0600); |
7c673cae | 12265 | if (fd < 0) { |
94b18763 | 12266 | derr << "failed to open " << path << ": " << cpp_strerror(errno) << dendl; |
31f18b77 | 12267 | return errno; |
7c673cae FG |
12268 | } |
12269 | } | |
12270 | ||
11fdf7f2 | 12271 | auto dump_func = [fd, f](CInode *in) { |
b32b8144 | 12272 | int r; |
7c673cae FG |
12273 | if (f) { |
12274 | f->open_object_section("inode"); | |
11fdf7f2 TL |
12275 | in->dump(f, CInode::DUMP_DEFAULT | CInode::DUMP_DIRFRAGS); |
12276 | f->close_section(); | |
12277 | return 1; | |
12278 | } | |
12279 | ostringstream ss; | |
12280 | ss << *in << std::endl; | |
12281 | std::string s = ss.str(); | |
12282 | r = safe_write(fd, s.c_str(), s.length()); | |
12283 | if (r < 0) | |
12284 | return r; | |
7c673cae FG |
12285 | list<CDir*> dfs; |
12286 | in->get_dirfrags(dfs); | |
11fdf7f2 TL |
12287 | for (auto &dir : dfs) { |
12288 | ostringstream tt; | |
12289 | tt << " " << *dir << std::endl; | |
12290 | std::string t = tt.str(); | |
12291 | r = safe_write(fd, t.c_str(), t.length()); | |
12292 | if (r < 0) | |
12293 | return r; | |
94b18763 FG |
12294 | for (auto &p : dir->items) { |
12295 | CDentry *dn = p.second; | |
11fdf7f2 TL |
12296 | ostringstream uu; |
12297 | uu << " " << *dn << std::endl; | |
12298 | std::string u = uu.str(); | |
12299 | r = safe_write(fd, u.c_str(), u.length()); | |
12300 | if (r < 0) | |
12301 | return r; | |
7c673cae FG |
12302 | } |
12303 | dir->check_rstats(); | |
7c673cae | 12304 | } |
b32b8144 FG |
12305 | return 1; |
12306 | }; | |
12307 | ||
94b18763 | 12308 | for (auto &p : inode_map) { |
b32b8144 FG |
12309 | r = dump_func(p.second); |
12310 | if (r < 0) | |
12311 | goto out; | |
12312 | } | |
94b18763 | 12313 | for (auto &p : snap_inode_map) { |
b32b8144 FG |
12314 | r = dump_func(p.second); |
12315 | if (r < 0) | |
12316 | goto out; | |
7c673cae | 12317 | } |
b32b8144 | 12318 | r = 0; |
7c673cae FG |
12319 | |
12320 | out: | |
12321 | if (f) { | |
12322 | f->close_section(); // inodes | |
12323 | } else { | |
12324 | ::close(fd); | |
12325 | } | |
31f18b77 | 12326 | return r; |
7c673cae FG |
12327 | } |
12328 | ||
12329 | ||
12330 | ||
12331 | C_MDS_RetryRequest::C_MDS_RetryRequest(MDCache *c, MDRequestRef& r) | |
12332 | : MDSInternalContext(c->mds), cache(c), mdr(r) | |
12333 | {} | |
12334 | ||
12335 | void C_MDS_RetryRequest::finish(int r) | |
12336 | { | |
12337 | mdr->retry++; | |
12338 | cache->dispatch_request(mdr); | |
12339 | } | |
12340 | ||
12341 | ||
12342 | class C_MDS_EnqueueScrub : public Context | |
12343 | { | |
11fdf7f2 | 12344 | std::string tag; |
7c673cae FG |
12345 | Formatter *formatter; |
12346 | Context *on_finish; | |
12347 | public: | |
12348 | ScrubHeaderRef header; | |
11fdf7f2 TL |
12349 | C_MDS_EnqueueScrub(std::string_view tag, Formatter *f, Context *fin) : |
12350 | tag(tag), formatter(f), on_finish(fin), header(nullptr) {} | |
7c673cae FG |
12351 | |
12352 | Context *take_finisher() { | |
12353 | Context *fin = on_finish; | |
12354 | on_finish = NULL; | |
12355 | return fin; | |
12356 | } | |
12357 | ||
12358 | void finish(int r) override { | |
11fdf7f2 TL |
12359 | if (r == 0) { |
12360 | // since recursive scrub is asynchronous, dump minimal output | |
12361 | // to not upset cli tools. | |
12362 | if (header && header->get_recursive()) { | |
12363 | formatter->open_object_section("results"); | |
12364 | formatter->dump_int("return_code", 0); | |
12365 | formatter->dump_string("scrub_tag", tag); | |
12366 | formatter->dump_string("mode", "asynchronous"); | |
12367 | formatter->close_section(); // results | |
12368 | } | |
12369 | } else { // we failed the lookup or something; dump ourselves | |
7c673cae FG |
12370 | formatter->open_object_section("results"); |
12371 | formatter->dump_int("return_code", r); | |
12372 | formatter->close_section(); // results | |
11fdf7f2 | 12373 | r = 0; // already dumped in formatter |
7c673cae FG |
12374 | } |
12375 | if (on_finish) | |
12376 | on_finish->complete(r); | |
12377 | } | |
12378 | }; | |
12379 | ||
12380 | void MDCache::enqueue_scrub( | |
11fdf7f2 TL |
12381 | std::string_view path, |
12382 | std::string_view tag, | |
7c673cae FG |
12383 | bool force, bool recursive, bool repair, |
12384 | Formatter *f, Context *fin) | |
12385 | { | |
11fdf7f2 | 12386 | dout(10) << __func__ << " " << path << dendl; |
7c673cae | 12387 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_ENQUEUE_SCRUB); |
11fdf7f2 TL |
12388 | if (path == "~mdsdir") { |
12389 | filepath fp(MDS_INO_MDSDIR(mds->get_nodeid())); | |
12390 | mdr->set_filepath(fp); | |
12391 | } else { | |
12392 | filepath fp(path); | |
12393 | mdr->set_filepath(path); | |
12394 | } | |
12395 | ||
12396 | bool is_internal = false; | |
12397 | std::string tag_str(tag); | |
12398 | if (tag_str.empty()) { | |
12399 | uuid_d uuid_gen; | |
12400 | uuid_gen.generate_random(); | |
12401 | tag_str = uuid_gen.to_string(); | |
12402 | is_internal = true; | |
12403 | } | |
7c673cae | 12404 | |
11fdf7f2 | 12405 | C_MDS_EnqueueScrub *cs = new C_MDS_EnqueueScrub(tag_str, f, fin); |
7c673cae | 12406 | cs->header = std::make_shared<ScrubHeader>( |
11fdf7f2 | 12407 | tag_str, is_internal, force, recursive, repair, f); |
7c673cae FG |
12408 | |
12409 | mdr->internal_op_finish = cs; | |
12410 | enqueue_scrub_work(mdr); | |
12411 | } | |
12412 | ||
12413 | void MDCache::enqueue_scrub_work(MDRequestRef& mdr) | |
12414 | { | |
11fdf7f2 TL |
12415 | MutationImpl::LockOpVec lov; |
12416 | CInode *in = mds->server->rdlock_path_pin_ref(mdr, 0, lov, true); | |
7c673cae FG |
12417 | if (NULL == in) |
12418 | return; | |
12419 | ||
12420 | // TODO: Remove this restriction | |
11fdf7f2 | 12421 | ceph_assert(in->is_auth()); |
7c673cae | 12422 | |
11fdf7f2 | 12423 | bool locked = mds->locker->acquire_locks(mdr, lov); |
7c673cae FG |
12424 | if (!locked) |
12425 | return; | |
12426 | ||
12427 | C_MDS_EnqueueScrub *cs = static_cast<C_MDS_EnqueueScrub*>(mdr->internal_op_finish); | |
11fdf7f2 | 12428 | ScrubHeaderRef header = cs->header; |
7c673cae FG |
12429 | |
12430 | // Cannot scrub same dentry twice at same time | |
11fdf7f2 | 12431 | if (in->scrub_is_in_progress()) { |
7c673cae FG |
12432 | mds->server->respond_to_request(mdr, -EBUSY); |
12433 | return; | |
12434 | } else { | |
12435 | in->scrub_info(); | |
12436 | } | |
12437 | ||
12438 | header->set_origin(in); | |
12439 | ||
11fdf7f2 TL |
12440 | Context *fin; |
12441 | if (header->get_recursive()) { | |
12442 | header->get_origin()->get(CInode::PIN_SCRUBQUEUE); | |
12443 | fin = new MDSInternalContextWrapper(mds, | |
12444 | new FunctionContext([this, header](int r) { | |
12445 | recursive_scrub_finish(header); | |
12446 | header->get_origin()->put(CInode::PIN_SCRUBQUEUE); | |
12447 | }) | |
12448 | ); | |
12449 | } else { | |
b32b8144 FG |
12450 | fin = cs->take_finisher(); |
12451 | } | |
12452 | ||
12453 | // If the scrub did some repair, then flush the journal at the end of | |
12454 | // the scrub. Otherwise in the case of e.g. rewriting a backtrace | |
12455 | // the on disk state will still look damaged. | |
28e407b8 AA |
12456 | auto scrub_finish = new FunctionContext([this, header, fin](int r){ |
12457 | if (!header->get_repaired()) { | |
12458 | if (fin) | |
12459 | fin->complete(r); | |
12460 | return; | |
12461 | } | |
12462 | ||
12463 | auto flush_finish = new FunctionContext([this, fin](int r){ | |
12464 | dout(4) << "Expiring log segments because scrub did some repairs" << dendl; | |
12465 | mds->mdlog->trim_all(); | |
12466 | ||
12467 | if (fin) { | |
12468 | MDSGatherBuilder gather(g_ceph_context); | |
12469 | auto& expiring_segments = mds->mdlog->get_expiring_segments(); | |
12470 | for (auto logseg : expiring_segments) | |
12471 | logseg->wait_for_expiry(gather.new_sub()); | |
11fdf7f2 | 12472 | ceph_assert(gather.has_subs()); |
28e407b8 AA |
12473 | gather.set_finisher(new MDSInternalContextWrapper(mds, fin)); |
12474 | gather.activate(); | |
b32b8144 | 12475 | } |
28e407b8 AA |
12476 | }); |
12477 | ||
12478 | dout(4) << "Flushing journal because scrub did some repairs" << dendl; | |
12479 | mds->mdlog->start_new_segment(); | |
12480 | mds->mdlog->flush(); | |
12481 | mds->mdlog->wait_for_safe(new MDSInternalContextWrapper(mds, flush_finish)); | |
b32b8144 FG |
12482 | }); |
12483 | ||
7c673cae | 12484 | if (!header->get_recursive()) { |
7c673cae | 12485 | mds->scrubstack->enqueue_inode_top(in, header, |
28e407b8 | 12486 | new MDSInternalContextWrapper(mds, scrub_finish)); |
b32b8144 FG |
12487 | } else { |
12488 | mds->scrubstack->enqueue_inode_bottom(in, header, | |
28e407b8 | 12489 | new MDSInternalContextWrapper(mds, scrub_finish)); |
b32b8144 | 12490 | } |
7c673cae FG |
12491 | |
12492 | mds->server->respond_to_request(mdr, 0); | |
12493 | return; | |
12494 | } | |
12495 | ||
11fdf7f2 TL |
12496 | void MDCache::recursive_scrub_finish(const ScrubHeaderRef& header) |
12497 | { | |
12498 | if (header->get_origin()->is_base() && | |
12499 | header->get_force() && header->get_repair()) { | |
12500 | // notify snapserver that base directory is recursively scrubbed. | |
12501 | // After both root and mdsdir are recursively scrubbed, snapserver | |
12502 | // knows that all old format snaprealms are converted to the new | |
12503 | // format. | |
12504 | if (mds->mdsmap->get_num_in_mds() == 1 && | |
12505 | mds->mdsmap->get_num_failed_mds() == 0 && | |
12506 | mds->mdsmap->get_tableserver() == mds->get_nodeid()) { | |
12507 | mds->mark_base_recursively_scrubbed(header->get_origin()->ino()); | |
12508 | } | |
12509 | } | |
12510 | } | |
12511 | ||
12512 | struct C_MDC_RespondInternalRequest : public MDCacheLogContext { | |
7c673cae | 12513 | MDRequestRef mdr; |
11fdf7f2 | 12514 | C_MDC_RespondInternalRequest(MDCache *c, MDRequestRef& m) : |
7c673cae FG |
12515 | MDCacheLogContext(c), mdr(m) {} |
12516 | void finish(int r) override { | |
12517 | mdr->apply(); | |
12518 | get_mds()->server->respond_to_request(mdr, r); | |
12519 | } | |
12520 | }; | |
12521 | ||
12522 | void MDCache::repair_dirfrag_stats(CDir *dir) | |
12523 | { | |
12524 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_REPAIR_FRAGSTATS); | |
12525 | mdr->pin(dir); | |
12526 | mdr->internal_op_private = dir; | |
12527 | mdr->internal_op_finish = new C_MDSInternalNoop; | |
12528 | repair_dirfrag_stats_work(mdr); | |
12529 | } | |
12530 | ||
12531 | void MDCache::repair_dirfrag_stats_work(MDRequestRef& mdr) | |
12532 | { | |
12533 | CDir *dir = static_cast<CDir*>(mdr->internal_op_private); | |
12534 | dout(10) << __func__ << " " << *dir << dendl; | |
12535 | ||
12536 | if (!dir->is_auth()) { | |
12537 | mds->server->respond_to_request(mdr, -ESTALE); | |
12538 | return; | |
12539 | } | |
12540 | ||
12541 | if (!mdr->is_auth_pinned(dir) && !dir->can_auth_pin()) { | |
224ce89b WB |
12542 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryRequest(this, mdr)); |
12543 | ||
7c673cae FG |
12544 | mds->locker->drop_locks(mdr.get()); |
12545 | mdr->drop_local_auth_pins(); | |
224ce89b WB |
12546 | if (!mdr->remote_auth_pins.empty()) |
12547 | mds->locker->notify_freeze_waiter(dir); | |
7c673cae FG |
12548 | return; |
12549 | } | |
12550 | ||
12551 | mdr->auth_pin(dir); | |
12552 | ||
11fdf7f2 | 12553 | MutationImpl::LockOpVec lov; |
7c673cae | 12554 | CInode *diri = dir->inode; |
11fdf7f2 TL |
12555 | lov.add_rdlock(&diri->dirfragtreelock); |
12556 | lov.add_wrlock(&diri->nestlock); | |
12557 | lov.add_wrlock(&diri->filelock); | |
12558 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
12559 | return; |
12560 | ||
12561 | if (!dir->is_complete()) { | |
12562 | dir->fetch(new C_MDS_RetryRequest(this, mdr)); | |
12563 | return; | |
12564 | } | |
12565 | ||
12566 | frag_info_t frag_info; | |
12567 | nest_info_t nest_info; | |
94b18763 | 12568 | for (auto it = dir->begin(); it != dir->end(); ++it) { |
7c673cae FG |
12569 | CDentry *dn = it->second; |
12570 | if (dn->last != CEPH_NOSNAP) | |
12571 | continue; | |
12572 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
12573 | if (dnl->is_primary()) { | |
12574 | CInode *in = dnl->get_inode(); | |
12575 | nest_info.add(in->get_projected_inode()->accounted_rstat); | |
12576 | if (in->is_dir()) | |
12577 | frag_info.nsubdirs++; | |
12578 | else | |
12579 | frag_info.nfiles++; | |
12580 | } else if (dnl->is_remote()) | |
12581 | frag_info.nfiles++; | |
12582 | } | |
12583 | ||
12584 | fnode_t *pf = dir->get_projected_fnode(); | |
12585 | bool good_fragstat = frag_info.same_sums(pf->fragstat); | |
12586 | bool good_rstat = nest_info.same_sums(pf->rstat); | |
12587 | if (good_fragstat && good_rstat) { | |
12588 | dout(10) << __func__ << " no corruption found" << dendl; | |
12589 | mds->server->respond_to_request(mdr, 0); | |
12590 | return; | |
12591 | } | |
12592 | ||
12593 | pf = dir->project_fnode(); | |
12594 | pf->version = dir->pre_dirty(); | |
12595 | mdr->add_projected_fnode(dir); | |
12596 | ||
12597 | mdr->ls = mds->mdlog->get_current_segment(); | |
12598 | EUpdate *le = new EUpdate(mds->mdlog, "repair_dirfrag"); | |
12599 | mds->mdlog->start_entry(le); | |
12600 | ||
12601 | if (!good_fragstat) { | |
12602 | if (pf->fragstat.mtime > frag_info.mtime) | |
12603 | frag_info.mtime = pf->fragstat.mtime; | |
12604 | if (pf->fragstat.change_attr > frag_info.change_attr) | |
12605 | frag_info.change_attr = pf->fragstat.change_attr; | |
12606 | pf->fragstat = frag_info; | |
12607 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
12608 | mdr->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
12609 | mdr->add_updated_lock(&diri->filelock); | |
12610 | } | |
12611 | ||
12612 | if (!good_rstat) { | |
12613 | if (pf->rstat.rctime > nest_info.rctime) | |
12614 | nest_info.rctime = pf->rstat.rctime; | |
12615 | pf->rstat = nest_info; | |
12616 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
12617 | mdr->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
12618 | mdr->add_updated_lock(&diri->nestlock); | |
12619 | } | |
12620 | ||
12621 | le->metablob.add_dir_context(dir); | |
12622 | le->metablob.add_dir(dir, true); | |
12623 | ||
11fdf7f2 | 12624 | mds->mdlog->submit_entry(le, new C_MDC_RespondInternalRequest(this, mdr)); |
7c673cae FG |
12625 | } |
12626 | ||
12627 | void MDCache::repair_inode_stats(CInode *diri) | |
12628 | { | |
12629 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_REPAIR_INODESTATS); | |
12630 | mdr->pin(diri); | |
12631 | mdr->internal_op_private = diri; | |
12632 | mdr->internal_op_finish = new C_MDSInternalNoop; | |
12633 | repair_inode_stats_work(mdr); | |
12634 | } | |
12635 | ||
12636 | void MDCache::repair_inode_stats_work(MDRequestRef& mdr) | |
12637 | { | |
12638 | CInode *diri = static_cast<CInode*>(mdr->internal_op_private); | |
12639 | dout(10) << __func__ << " " << *diri << dendl; | |
12640 | ||
12641 | if (!diri->is_auth()) { | |
12642 | mds->server->respond_to_request(mdr, -ESTALE); | |
12643 | return; | |
12644 | } | |
12645 | if (!diri->is_dir()) { | |
12646 | mds->server->respond_to_request(mdr, -ENOTDIR); | |
12647 | return; | |
12648 | } | |
12649 | ||
11fdf7f2 | 12650 | MutationImpl::LockOpVec lov; |
7c673cae FG |
12651 | |
12652 | if (mdr->ls) // already marked filelock/nestlock dirty ? | |
12653 | goto do_rdlocks; | |
12654 | ||
11fdf7f2 TL |
12655 | lov.add_rdlock(&diri->dirfragtreelock); |
12656 | lov.add_wrlock(&diri->nestlock); | |
12657 | lov.add_wrlock(&diri->filelock); | |
12658 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
12659 | return; |
12660 | ||
12661 | // Fetch all dirfrags and mark filelock/nestlock dirty. This will tirgger | |
12662 | // the scatter-gather process, which will fix any fragstat/rstat errors. | |
11fdf7f2 TL |
12663 | { |
12664 | frag_vec_t leaves; | |
12665 | diri->dirfragtree.get_leaves(leaves); | |
12666 | for (const auto& leaf : leaves) { | |
12667 | CDir *dir = diri->get_dirfrag(leaf); | |
12668 | if (!dir) { | |
12669 | ceph_assert(mdr->is_auth_pinned(diri)); | |
12670 | dir = diri->get_or_open_dirfrag(this, leaf); | |
12671 | } | |
12672 | if (dir->get_version() == 0) { | |
12673 | ceph_assert(dir->is_auth()); | |
12674 | dir->fetch(new C_MDS_RetryRequest(this, mdr)); | |
12675 | return; | |
12676 | } | |
7c673cae FG |
12677 | } |
12678 | } | |
12679 | ||
12680 | diri->state_set(CInode::STATE_REPAIRSTATS); | |
12681 | mdr->ls = mds->mdlog->get_current_segment(); | |
12682 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
12683 | mdr->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
12684 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
12685 | mdr->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
12686 | ||
12687 | mds->locker->drop_locks(mdr.get()); | |
12688 | ||
12689 | do_rdlocks: | |
12690 | // force the scatter-gather process | |
11fdf7f2 TL |
12691 | lov.clear(); |
12692 | lov.add_rdlock(&diri->dirfragtreelock); | |
12693 | lov.add_rdlock(&diri->nestlock); | |
12694 | lov.add_rdlock(&diri->filelock); | |
12695 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
12696 | return; |
12697 | ||
12698 | diri->state_clear(CInode::STATE_REPAIRSTATS); | |
12699 | ||
12700 | frag_info_t dir_info; | |
12701 | nest_info_t nest_info; | |
11fdf7f2 TL |
12702 | nest_info.rsubdirs = 1; // it gets one to account for self |
12703 | if (const sr_t *srnode = diri->get_projected_srnode(); srnode) | |
12704 | nest_info.rsnaps = srnode->snaps.size(); | |
7c673cae | 12705 | |
11fdf7f2 TL |
12706 | { |
12707 | frag_vec_t leaves; | |
12708 | diri->dirfragtree.get_leaves(leaves); | |
12709 | for (const auto& leaf : leaves) { | |
12710 | CDir *dir = diri->get_dirfrag(leaf); | |
12711 | ceph_assert(dir); | |
12712 | ceph_assert(dir->get_version() > 0); | |
12713 | dir_info.add(dir->fnode.accounted_fragstat); | |
12714 | nest_info.add(dir->fnode.accounted_rstat); | |
12715 | } | |
7c673cae FG |
12716 | } |
12717 | ||
12718 | if (!dir_info.same_sums(diri->inode.dirstat) || | |
12719 | !nest_info.same_sums(diri->inode.rstat)) { | |
12720 | dout(10) << __func__ << " failed to fix fragstat/rstat on " | |
12721 | << *diri << dendl; | |
12722 | } | |
12723 | ||
12724 | mds->server->respond_to_request(mdr, 0); | |
12725 | } | |
12726 | ||
11fdf7f2 TL |
12727 | void MDCache::upgrade_inode_snaprealm(CInode *in) |
12728 | { | |
12729 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_UPGRADE_SNAPREALM); | |
12730 | mdr->pin(in); | |
12731 | mdr->internal_op_private = in; | |
12732 | mdr->internal_op_finish = new C_MDSInternalNoop; | |
12733 | upgrade_inode_snaprealm_work(mdr); | |
12734 | } | |
12735 | ||
12736 | void MDCache::upgrade_inode_snaprealm_work(MDRequestRef& mdr) | |
12737 | { | |
12738 | CInode *in = static_cast<CInode*>(mdr->internal_op_private); | |
12739 | dout(10) << __func__ << " " << *in << dendl; | |
12740 | ||
12741 | if (!in->is_auth()) { | |
12742 | mds->server->respond_to_request(mdr, -ESTALE); | |
12743 | return; | |
12744 | } | |
12745 | ||
12746 | MutationImpl::LockOpVec lov; | |
12747 | mds->locker->include_snap_rdlocks(in, lov); | |
12748 | lov.erase_rdlock(&in->snaplock); | |
12749 | lov.add_xlock(&in->snaplock); | |
12750 | ||
12751 | if (!mds->locker->acquire_locks(mdr, lov)) | |
12752 | return; | |
12753 | ||
12754 | // project_snaprealm() upgrades snaprealm format | |
12755 | auto &pi = in->project_inode(false, true); | |
12756 | mdr->add_projected_inode(in); | |
12757 | pi.inode.version = in->pre_dirty(); | |
12758 | ||
12759 | mdr->ls = mds->mdlog->get_current_segment(); | |
12760 | EUpdate *le = new EUpdate(mds->mdlog, "upgrade_snaprealm"); | |
12761 | mds->mdlog->start_entry(le); | |
12762 | ||
12763 | if (in->is_base()) { | |
12764 | le->metablob.add_root(true, in); | |
12765 | } else { | |
12766 | CDentry *pdn = in->get_projected_parent_dn(); | |
12767 | le->metablob.add_dir_context(pdn->get_dir()); | |
12768 | le->metablob.add_primary_dentry(pdn, in, true); | |
12769 | } | |
12770 | ||
12771 | mds->mdlog->submit_entry(le, new C_MDC_RespondInternalRequest(this, mdr)); | |
12772 | } | |
12773 | ||
12774 | void MDCache::flush_dentry(std::string_view path, Context *fin) | |
7c673cae FG |
12775 | { |
12776 | if (is_readonly()) { | |
12777 | dout(10) << __func__ << ": read-only FS" << dendl; | |
12778 | fin->complete(-EROFS); | |
12779 | return; | |
12780 | } | |
12781 | dout(10) << "flush_dentry " << path << dendl; | |
12782 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FLUSH); | |
94b18763 | 12783 | filepath fp(path); |
7c673cae FG |
12784 | mdr->set_filepath(fp); |
12785 | mdr->internal_op_finish = fin; | |
12786 | flush_dentry_work(mdr); | |
12787 | } | |
12788 | ||
11fdf7f2 | 12789 | class C_FinishIOMDR : public MDSContext { |
7c673cae FG |
12790 | protected: |
12791 | MDSRank *mds; | |
12792 | MDRequestRef mdr; | |
12793 | MDSRank *get_mds() override { return mds; } | |
12794 | public: | |
12795 | C_FinishIOMDR(MDSRank *mds_, MDRequestRef& mdr_) : mds(mds_), mdr(mdr_) {} | |
12796 | void finish(int r) override { mds->server->respond_to_request(mdr, r); } | |
12797 | }; | |
12798 | ||
12799 | void MDCache::flush_dentry_work(MDRequestRef& mdr) | |
12800 | { | |
11fdf7f2 TL |
12801 | MutationImpl::LockOpVec lov; |
12802 | CInode *in = mds->server->rdlock_path_pin_ref(mdr, 0, lov, true); | |
7c673cae FG |
12803 | if (NULL == in) |
12804 | return; | |
12805 | ||
12806 | // TODO: Is this necessary? Fix it if so | |
11fdf7f2 TL |
12807 | ceph_assert(in->is_auth()); |
12808 | bool locked = mds->locker->acquire_locks(mdr, lov); | |
7c673cae FG |
12809 | if (!locked) |
12810 | return; | |
12811 | in->flush(new C_FinishIOMDR(mds, mdr)); | |
12812 | } | |
12813 | ||
12814 | ||
12815 | /** | |
12816 | * Initialize performance counters with global perfcounter | |
12817 | * collection. | |
12818 | */ | |
12819 | void MDCache::register_perfcounters() | |
12820 | { | |
91327a77 AA |
12821 | PerfCountersBuilder pcb(g_ceph_context, "mds_cache", l_mdc_first, l_mdc_last); |
12822 | ||
12823 | // Stray/purge statistics | |
12824 | pcb.add_u64(l_mdc_num_strays, "num_strays", "Stray dentries", "stry", | |
12825 | PerfCountersBuilder::PRIO_INTERESTING); | |
12826 | pcb.add_u64(l_mdc_num_recovering_enqueued, | |
12827 | "num_recovering_enqueued", "Files waiting for recovery", "recy", | |
12828 | PerfCountersBuilder::PRIO_INTERESTING); | |
12829 | pcb.add_u64_counter(l_mdc_recovery_completed, | |
12830 | "recovery_completed", "File recoveries completed", "recd", | |
12831 | PerfCountersBuilder::PRIO_INTERESTING); | |
12832 | ||
12833 | // useful recovery queue statistics | |
12834 | pcb.set_prio_default(PerfCountersBuilder::PRIO_USEFUL); | |
12835 | pcb.add_u64(l_mdc_num_recovering_processing, "num_recovering_processing", | |
12836 | "Files currently being recovered"); | |
12837 | pcb.add_u64(l_mdc_num_recovering_prioritized, "num_recovering_prioritized", | |
12838 | "Files waiting for recovery with elevated priority"); | |
12839 | pcb.add_u64_counter(l_mdc_recovery_started, "recovery_started", | |
12840 | "File recoveries started"); | |
12841 | ||
12842 | // along with other stray dentries stats | |
12843 | pcb.add_u64(l_mdc_num_strays_delayed, "num_strays_delayed", | |
12844 | "Stray dentries delayed"); | |
12845 | pcb.add_u64(l_mdc_num_strays_enqueuing, "num_strays_enqueuing", | |
12846 | "Stray dentries enqueuing for purge"); | |
12847 | pcb.add_u64_counter(l_mdc_strays_created, "strays_created", | |
12848 | "Stray dentries created"); | |
7c673cae | 12849 | pcb.add_u64_counter(l_mdc_strays_enqueued, "strays_enqueued", |
91327a77 AA |
12850 | "Stray dentries enqueued for purge"); |
12851 | pcb.add_u64_counter(l_mdc_strays_reintegrated, "strays_reintegrated", | |
12852 | "Stray dentries reintegrated"); | |
12853 | pcb.add_u64_counter(l_mdc_strays_migrated, "strays_migrated", | |
12854 | "Stray dentries migrated"); | |
7c673cae | 12855 | |
91327a77 | 12856 | // low prio internal request stats |
d2e6a577 | 12857 | pcb.add_u64_counter(l_mdss_ireq_enqueue_scrub, "ireq_enqueue_scrub", |
91327a77 | 12858 | "Internal Request type enqueue scrub"); |
d2e6a577 | 12859 | pcb.add_u64_counter(l_mdss_ireq_exportdir, "ireq_exportdir", |
91327a77 | 12860 | "Internal Request type export dir"); |
d2e6a577 | 12861 | pcb.add_u64_counter(l_mdss_ireq_flush, "ireq_flush", |
91327a77 | 12862 | "Internal Request type flush"); |
d2e6a577 | 12863 | pcb.add_u64_counter(l_mdss_ireq_fragmentdir, "ireq_fragmentdir", |
91327a77 | 12864 | "Internal Request type fragmentdir"); |
d2e6a577 | 12865 | pcb.add_u64_counter(l_mdss_ireq_fragstats, "ireq_fragstats", |
91327a77 | 12866 | "Internal Request type frag stats"); |
d2e6a577 | 12867 | pcb.add_u64_counter(l_mdss_ireq_inodestats, "ireq_inodestats", |
91327a77 | 12868 | "Internal Request type inode stats"); |
d2e6a577 | 12869 | |
7c673cae FG |
12870 | logger.reset(pcb.create_perf_counters()); |
12871 | g_ceph_context->get_perfcounters_collection()->add(logger.get()); | |
12872 | recovery_queue.set_logger(logger.get()); | |
12873 | stray_manager.set_logger(logger.get()); | |
12874 | } | |
12875 | ||
7c673cae FG |
12876 | /** |
12877 | * Call this when putting references to an inode/dentry or | |
12878 | * when attempting to trim it. | |
12879 | * | |
12880 | * If this inode is no longer linked by anyone, and this MDS | |
12881 | * rank holds the primary dentry, and that dentry is in a stray | |
12882 | * directory, then give up the dentry to the StrayManager, never | |
12883 | * to be seen again by MDCache. | |
12884 | * | |
12885 | * @param delay if true, then purgeable inodes are stashed til | |
12886 | * the next trim(), rather than being purged right | |
12887 | * away. | |
12888 | */ | |
12889 | void MDCache::maybe_eval_stray(CInode *in, bool delay) { | |
224ce89b WB |
12890 | if (in->inode.nlink > 0 || in->is_base() || is_readonly() || |
12891 | mds->get_state() <= MDSMap::STATE_REJOIN) | |
7c673cae | 12892 | return; |
224ce89b | 12893 | |
7c673cae FG |
12894 | CDentry *dn = in->get_projected_parent_dn(); |
12895 | ||
12896 | if (dn->state_test(CDentry::STATE_PURGING)) { | |
12897 | /* We have already entered the purging process, no need | |
12898 | * to re-evaluate me ! */ | |
12899 | return; | |
12900 | } | |
12901 | ||
11fdf7f2 TL |
12902 | if (dn->get_dir()->get_inode()->is_stray()) { |
12903 | if (delay) | |
12904 | stray_manager.queue_delayed(dn); | |
12905 | else | |
12906 | stray_manager.eval_stray(dn); | |
7c673cae FG |
12907 | } |
12908 | } | |
12909 | ||
31f18b77 FG |
12910 | void MDCache::clear_dirty_bits_for_stray(CInode* diri) { |
12911 | dout(10) << __func__ << " " << *diri << dendl; | |
11fdf7f2 | 12912 | ceph_assert(diri->get_projected_parent_dir()->inode->is_stray()); |
31f18b77 FG |
12913 | list<CDir*> ls; |
12914 | diri->get_dirfrags(ls); | |
94b18763 | 12915 | for (auto &p : ls) { |
31f18b77 FG |
12916 | if (p->is_auth() && !(p->is_frozen() || p->is_freezing())) |
12917 | p->try_remove_dentries_for_stray(); | |
12918 | } | |
12919 | if (!diri->snaprealm) { | |
12920 | if (diri->is_auth()) | |
12921 | diri->clear_dirty_rstat(); | |
12922 | diri->clear_scatter_dirty(); | |
12923 | } | |
12924 | } | |
12925 | ||
11fdf7f2 TL |
12926 | bool MDCache::dump_inode(Formatter *f, uint64_t number) { |
12927 | CInode *in = get_inode(number); | |
12928 | if (!in) { | |
12929 | return false; | |
12930 | } | |
12931 | f->open_object_section("inode"); | |
12932 | in->dump(f, CInode::DUMP_DEFAULT | CInode::DUMP_PATH); | |
12933 | f->close_section(); | |
12934 | return true; | |
12935 | } |