]>
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> | |
20 | #include <map> | |
21 | ||
22 | #include "MDCache.h" | |
23 | #include "MDSRank.h" | |
24 | #include "Server.h" | |
25 | #include "Locker.h" | |
26 | #include "MDLog.h" | |
27 | #include "MDBalancer.h" | |
28 | #include "Migrator.h" | |
29 | #include "ScrubStack.h" | |
30 | ||
31 | #include "SnapClient.h" | |
32 | ||
33 | #include "MDSMap.h" | |
34 | ||
35 | #include "CInode.h" | |
36 | #include "CDir.h" | |
37 | ||
38 | #include "Mutation.h" | |
39 | ||
40 | #include "include/ceph_fs.h" | |
41 | #include "include/filepath.h" | |
181888fb | 42 | #include "include/util.h" |
7c673cae FG |
43 | |
44 | #include "msg/Message.h" | |
45 | #include "msg/Messenger.h" | |
46 | ||
181888fb | 47 | #include "common/MemoryModel.h" |
7c673cae | 48 | #include "common/errno.h" |
7c673cae | 49 | #include "common/perf_counters.h" |
181888fb FG |
50 | #include "common/safe_io.h" |
51 | ||
7c673cae FG |
52 | #include "osdc/Journaler.h" |
53 | #include "osdc/Filer.h" | |
54 | ||
55 | #include "events/ESubtreeMap.h" | |
56 | #include "events/EUpdate.h" | |
57 | #include "events/ESlaveUpdate.h" | |
58 | #include "events/EImportFinish.h" | |
59 | #include "events/EFragment.h" | |
60 | #include "events/ECommitted.h" | |
61 | #include "events/ESessions.h" | |
62 | ||
63 | #include "messages/MGenericMessage.h" | |
64 | ||
65 | #include "messages/MMDSResolve.h" | |
66 | #include "messages/MMDSResolveAck.h" | |
67 | #include "messages/MMDSCacheRejoin.h" | |
68 | ||
69 | #include "messages/MDiscover.h" | |
70 | #include "messages/MDiscoverReply.h" | |
71 | ||
72 | //#include "messages/MInodeUpdate.h" | |
73 | #include "messages/MDirUpdate.h" | |
74 | #include "messages/MCacheExpire.h" | |
75 | ||
76 | #include "messages/MInodeFileCaps.h" | |
77 | ||
78 | #include "messages/MLock.h" | |
79 | #include "messages/MDentryLink.h" | |
80 | #include "messages/MDentryUnlink.h" | |
81 | ||
82 | #include "messages/MMDSFindIno.h" | |
83 | #include "messages/MMDSFindInoReply.h" | |
84 | ||
85 | #include "messages/MMDSOpenIno.h" | |
86 | #include "messages/MMDSOpenInoReply.h" | |
87 | ||
88 | #include "messages/MClientRequest.h" | |
89 | #include "messages/MClientCaps.h" | |
90 | #include "messages/MClientSnap.h" | |
91 | #include "messages/MClientQuota.h" | |
92 | ||
93 | #include "messages/MMDSSlaveRequest.h" | |
94 | ||
95 | #include "messages/MMDSFragmentNotify.h" | |
96 | ||
97 | #include "messages/MGatherCaps.h" | |
98 | ||
99 | #include "InoTable.h" | |
100 | ||
101 | #include "common/Timer.h" | |
102 | ||
103 | #include "perfglue/heap_profiler.h" | |
104 | ||
105 | using namespace std; | |
106 | ||
107 | #include "common/config.h" | |
108 | #include "include/assert.h" | |
109 | ||
110 | #define dout_context g_ceph_context | |
111 | #define dout_subsys ceph_subsys_mds | |
112 | #undef dout_prefix | |
113 | #define dout_prefix _prefix(_dout, mds) | |
114 | static ostream& _prefix(std::ostream *_dout, MDSRank *mds) { | |
115 | return *_dout << "mds." << mds->get_nodeid() << ".cache "; | |
116 | } | |
117 | ||
118 | set<int> SimpleLock::empty_gather_set; | |
119 | ||
120 | ||
121 | /** | |
122 | * All non-I/O contexts that require a reference | |
123 | * to an MDCache instance descend from this. | |
124 | */ | |
125 | class MDCacheContext : public virtual MDSInternalContextBase { | |
126 | protected: | |
127 | MDCache *mdcache; | |
128 | MDSRank *get_mds() override | |
129 | { | |
130 | assert(mdcache != NULL); | |
131 | return mdcache->mds; | |
132 | } | |
133 | public: | |
134 | explicit MDCacheContext(MDCache *mdc_) : mdcache(mdc_) {} | |
135 | }; | |
136 | ||
137 | ||
138 | /** | |
139 | * Only for contexts called back from an I/O completion | |
140 | * | |
141 | * Note: duplication of members wrt MDCacheContext, because | |
142 | * it'ls the lesser of two evils compared with introducing | |
143 | * yet another piece of (multiple) inheritance. | |
144 | */ | |
145 | class MDCacheIOContext : public virtual MDSIOContextBase { | |
146 | protected: | |
147 | MDCache *mdcache; | |
148 | MDSRank *get_mds() override | |
149 | { | |
150 | assert(mdcache != NULL); | |
151 | return mdcache->mds; | |
152 | } | |
153 | public: | |
154 | explicit MDCacheIOContext(MDCache *mdc_) : mdcache(mdc_) {} | |
155 | }; | |
156 | ||
157 | class MDCacheLogContext : public virtual MDSLogContextBase { | |
158 | protected: | |
159 | MDCache *mdcache; | |
160 | MDSRank *get_mds() override | |
161 | { | |
162 | assert(mdcache != NULL); | |
163 | return mdcache->mds; | |
164 | } | |
165 | public: | |
166 | explicit MDCacheLogContext(MDCache *mdc_) : mdcache(mdc_) {} | |
167 | }; | |
168 | ||
169 | MDCache::MDCache(MDSRank *m, PurgeQueue &purge_queue_) : | |
170 | mds(m), | |
171 | filer(m->objecter, m->finisher), | |
172 | exceeded_size_limit(false), | |
173 | recovery_queue(m), | |
174 | stray_manager(m, purge_queue_) | |
175 | { | |
176 | migrator.reset(new Migrator(mds, this)); | |
177 | root = NULL; | |
178 | myin = NULL; | |
179 | readonly = false; | |
180 | ||
181 | stray_index = 0; | |
182 | for (int i = 0; i < NUM_STRAY; ++i) { | |
183 | strays[i] = NULL; | |
184 | } | |
185 | ||
186 | num_inodes_with_caps = 0; | |
187 | ||
188 | max_dir_commit_size = g_conf->mds_dir_max_commit_size ? | |
189 | (g_conf->mds_dir_max_commit_size << 20) : | |
190 | (0.9 *(g_conf->osd_max_write_size << 20)); | |
191 | ||
192 | discover_last_tid = 0; | |
193 | open_ino_last_tid = 0; | |
194 | find_ino_peer_last_tid = 0; | |
195 | ||
196 | last_cap_id = 0; | |
197 | ||
198 | client_lease_durations[0] = 5.0; | |
199 | client_lease_durations[1] = 30.0; | |
200 | client_lease_durations[2] = 300.0; | |
201 | ||
202 | resolves_pending = false; | |
203 | rejoins_pending = false; | |
204 | cap_imports_num_opening = 0; | |
205 | ||
206 | opening_root = open = false; | |
181888fb | 207 | lru.lru_set_midpoint(cache_mid()); |
7c673cae | 208 | |
31f18b77 FG |
209 | bottom_lru.lru_set_midpoint(0); |
210 | ||
7c673cae FG |
211 | decayrate.set_halflife(g_conf->mds_decay_halflife); |
212 | ||
213 | did_shutdown_log_cap = false; | |
214 | } | |
215 | ||
216 | MDCache::~MDCache() | |
217 | { | |
218 | if (logger) { | |
219 | g_ceph_context->get_perfcounters_collection()->remove(logger.get()); | |
220 | } | |
221 | } | |
222 | ||
223 | ||
224 | ||
225 | void MDCache::log_stat() | |
226 | { | |
181888fb | 227 | mds->logger->set(l_mds_inode_max, cache_limit_inodes() == 0 ? INT_MAX : cache_limit_inodes()); |
7c673cae FG |
228 | mds->logger->set(l_mds_inodes, lru.lru_get_size()); |
229 | mds->logger->set(l_mds_inodes_pinned, lru.lru_get_num_pinned()); | |
230 | mds->logger->set(l_mds_inodes_top, lru.lru_get_top()); | |
231 | mds->logger->set(l_mds_inodes_bottom, lru.lru_get_bot()); | |
232 | mds->logger->set(l_mds_inodes_pin_tail, lru.lru_get_pintail()); | |
233 | mds->logger->set(l_mds_inodes_with_caps, num_inodes_with_caps); | |
234 | mds->logger->set(l_mds_caps, Capability::count()); | |
235 | } | |
236 | ||
237 | ||
238 | // | |
239 | ||
240 | bool MDCache::shutdown() | |
241 | { | |
242 | if (lru.lru_get_size() > 0) { | |
243 | dout(7) << "WARNING: mdcache shutdown with non-empty cache" << dendl; | |
244 | //show_cache(); | |
245 | show_subtrees(); | |
246 | //dump(); | |
247 | } | |
248 | return true; | |
249 | } | |
250 | ||
251 | ||
252 | // ==================================================================== | |
253 | // some inode functions | |
254 | ||
255 | void MDCache::add_inode(CInode *in) | |
256 | { | |
257 | // add to lru, inode map | |
258 | assert(inode_map.count(in->vino()) == 0); // should be no dup inos! | |
259 | inode_map[ in->vino() ] = in; | |
260 | ||
261 | if (in->ino() < MDS_INO_SYSTEM_BASE) { | |
262 | if (in->ino() == MDS_INO_ROOT) | |
263 | root = in; | |
264 | else if (in->ino() == MDS_INO_MDSDIR(mds->get_nodeid())) | |
265 | myin = in; | |
266 | else if (in->is_stray()) { | |
267 | if (MDS_INO_STRAY_OWNER(in->ino()) == mds->get_nodeid()) { | |
268 | strays[MDS_INO_STRAY_INDEX(in->ino())] = in; | |
269 | } | |
270 | } | |
271 | if (in->is_base()) | |
272 | base_inodes.insert(in); | |
273 | } | |
274 | ||
181888fb | 275 | if (cache_toofull()) { |
7c673cae FG |
276 | exceeded_size_limit = true; |
277 | } | |
278 | } | |
279 | ||
280 | void MDCache::remove_inode(CInode *o) | |
281 | { | |
282 | dout(14) << "remove_inode " << *o << dendl; | |
283 | ||
284 | if (o->get_parent_dn()) { | |
285 | // FIXME: multiple parents? | |
286 | CDentry *dn = o->get_parent_dn(); | |
287 | assert(!dn->is_dirty()); | |
288 | dn->dir->unlink_inode(dn); // leave dentry ... FIXME? | |
289 | } | |
290 | ||
291 | if (o->is_dirty()) | |
292 | o->mark_clean(); | |
293 | if (o->is_dirty_parent()) | |
294 | o->clear_dirty_parent(); | |
295 | ||
296 | o->clear_scatter_dirty(); | |
297 | ||
298 | o->item_open_file.remove_myself(); | |
299 | ||
31f18b77 FG |
300 | if (o->state_test(CInode::STATE_QUEUEDEXPORTPIN)) |
301 | export_pin_queue.erase(o); | |
7c673cae FG |
302 | |
303 | // remove from inode map | |
304 | inode_map.erase(o->vino()); | |
305 | ||
306 | if (o->ino() < MDS_INO_SYSTEM_BASE) { | |
307 | if (o == root) root = 0; | |
308 | if (o == myin) myin = 0; | |
309 | if (o->is_stray()) { | |
310 | if (MDS_INO_STRAY_OWNER(o->ino()) == mds->get_nodeid()) { | |
311 | strays[MDS_INO_STRAY_INDEX(o->ino())] = 0; | |
312 | } | |
313 | } | |
314 | if (o->is_base()) | |
315 | base_inodes.erase(o); | |
316 | } | |
317 | ||
318 | // delete it | |
319 | assert(o->get_num_ref() == 0); | |
320 | delete o; | |
321 | } | |
322 | ||
323 | file_layout_t MDCache::gen_default_file_layout(const MDSMap &mdsmap) | |
324 | { | |
325 | file_layout_t result = file_layout_t::get_default(); | |
326 | result.pool_id = mdsmap.get_first_data_pool(); | |
327 | return result; | |
328 | } | |
329 | ||
330 | file_layout_t MDCache::gen_default_log_layout(const MDSMap &mdsmap) | |
331 | { | |
332 | file_layout_t result = file_layout_t::get_default(); | |
333 | result.pool_id = mdsmap.get_metadata_pool(); | |
334 | if (g_conf->mds_log_segment_size > 0) { | |
335 | result.object_size = g_conf->mds_log_segment_size; | |
336 | result.stripe_unit = g_conf->mds_log_segment_size; | |
337 | } | |
338 | return result; | |
339 | } | |
340 | ||
341 | void MDCache::init_layouts() | |
342 | { | |
343 | default_file_layout = gen_default_file_layout(*(mds->mdsmap)); | |
344 | default_log_layout = gen_default_log_layout(*(mds->mdsmap)); | |
345 | } | |
346 | ||
347 | void MDCache::create_unlinked_system_inode(CInode *in, inodeno_t ino, | |
348 | int mode) const | |
349 | { | |
350 | in->inode.ino = ino; | |
351 | in->inode.version = 1; | |
352 | in->inode.xattr_version = 1; | |
353 | in->inode.mode = 0500 | mode; | |
354 | in->inode.size = 0; | |
355 | in->inode.ctime = | |
356 | in->inode.mtime = | |
357 | in->inode.btime = ceph_clock_now(); | |
358 | in->inode.nlink = 1; | |
359 | in->inode.truncate_size = -1ull; | |
360 | in->inode.change_attr = 0; | |
361 | in->inode.export_pin = MDS_RANK_NONE; | |
362 | ||
363 | memset(&in->inode.dir_layout, 0, sizeof(in->inode.dir_layout)); | |
364 | if (in->inode.is_dir()) { | |
365 | in->inode.dir_layout.dl_dir_hash = g_conf->mds_default_dir_hash; | |
366 | ++in->inode.rstat.rsubdirs; | |
367 | } else { | |
368 | in->inode.layout = default_file_layout; | |
369 | ++in->inode.rstat.rfiles; | |
370 | } | |
371 | in->inode.accounted_rstat = in->inode.rstat; | |
372 | ||
373 | if (in->is_base()) { | |
374 | if (in->is_root()) | |
375 | in->inode_auth = mds_authority_t(mds->get_nodeid(), CDIR_AUTH_UNKNOWN); | |
376 | else | |
377 | in->inode_auth = mds_authority_t(mds_rank_t(in->ino() - MDS_INO_MDSDIR_OFFSET), CDIR_AUTH_UNKNOWN); | |
378 | in->open_snaprealm(); // empty snaprealm | |
379 | assert(!in->snaprealm->parent); // created its own | |
380 | in->snaprealm->srnode.seq = 1; | |
381 | } | |
382 | } | |
383 | ||
384 | CInode *MDCache::create_system_inode(inodeno_t ino, int mode) | |
385 | { | |
386 | dout(0) << "creating system inode with ino:" << ino << dendl; | |
387 | CInode *in = new CInode(this); | |
388 | create_unlinked_system_inode(in, ino, mode); | |
389 | add_inode(in); | |
390 | return in; | |
391 | } | |
392 | ||
393 | CInode *MDCache::create_root_inode() | |
394 | { | |
395 | CInode *i = create_system_inode(MDS_INO_ROOT, S_IFDIR|0755); | |
396 | i->inode.uid = g_conf->mds_root_ino_uid; | |
397 | i->inode.gid = g_conf->mds_root_ino_gid; | |
398 | i->inode.layout = default_file_layout; | |
399 | i->inode.layout.pool_id = mds->mdsmap->get_first_data_pool(); | |
400 | return i; | |
401 | } | |
402 | ||
403 | void MDCache::create_empty_hierarchy(MDSGather *gather) | |
404 | { | |
405 | // create root dir | |
406 | CInode *root = create_root_inode(); | |
407 | ||
408 | // force empty root dir | |
409 | CDir *rootdir = root->get_or_open_dirfrag(this, frag_t()); | |
410 | adjust_subtree_auth(rootdir, mds->get_nodeid()); | |
411 | rootdir->dir_rep = CDir::REP_ALL; //NONE; | |
412 | ||
413 | rootdir->fnode.accounted_fragstat = rootdir->fnode.fragstat; | |
414 | rootdir->fnode.accounted_rstat = rootdir->fnode.rstat; | |
415 | ||
416 | root->inode.dirstat = rootdir->fnode.fragstat; | |
417 | root->inode.rstat = rootdir->fnode.rstat; | |
418 | ++root->inode.rstat.rsubdirs; | |
419 | root->inode.accounted_rstat = root->inode.rstat; | |
420 | ||
421 | rootdir->mark_complete(); | |
422 | rootdir->mark_dirty(rootdir->pre_dirty(), mds->mdlog->get_current_segment()); | |
423 | rootdir->commit(0, gather->new_sub()); | |
424 | ||
425 | root->store(gather->new_sub()); | |
426 | } | |
427 | ||
428 | void MDCache::create_mydir_hierarchy(MDSGather *gather) | |
429 | { | |
430 | // create mds dir | |
431 | CInode *my = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR); | |
432 | ||
433 | CDir *mydir = my->get_or_open_dirfrag(this, frag_t()); | |
434 | adjust_subtree_auth(mydir, mds->get_nodeid()); | |
435 | ||
436 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
437 | ||
438 | // stray dir | |
439 | for (int i = 0; i < NUM_STRAY; ++i) { | |
440 | CInode *stray = create_system_inode(MDS_INO_STRAY(mds->get_nodeid(), i), S_IFDIR); | |
441 | CDir *straydir = stray->get_or_open_dirfrag(this, frag_t()); | |
442 | stringstream name; | |
443 | name << "stray" << i; | |
444 | CDentry *sdn = mydir->add_primary_dentry(name.str(), stray); | |
445 | sdn->_mark_dirty(mds->mdlog->get_current_segment()); | |
446 | ||
447 | stray->inode.dirstat = straydir->fnode.fragstat; | |
448 | ||
449 | mydir->fnode.rstat.add(stray->inode.rstat); | |
450 | mydir->fnode.fragstat.nsubdirs++; | |
451 | // save them | |
452 | straydir->mark_complete(); | |
453 | straydir->mark_dirty(straydir->pre_dirty(), ls); | |
454 | straydir->commit(0, gather->new_sub()); | |
455 | stray->_mark_dirty_parent(ls, true); | |
456 | stray->store_backtrace(gather->new_sub()); | |
457 | } | |
458 | ||
459 | mydir->fnode.accounted_fragstat = mydir->fnode.fragstat; | |
460 | mydir->fnode.accounted_rstat = mydir->fnode.rstat; | |
461 | ||
462 | myin->inode.dirstat = mydir->fnode.fragstat; | |
463 | myin->inode.rstat = mydir->fnode.rstat; | |
464 | ++myin->inode.rstat.rsubdirs; | |
465 | myin->inode.accounted_rstat = myin->inode.rstat; | |
466 | ||
467 | mydir->mark_complete(); | |
468 | mydir->mark_dirty(mydir->pre_dirty(), ls); | |
469 | mydir->commit(0, gather->new_sub()); | |
470 | ||
471 | myin->store(gather->new_sub()); | |
472 | } | |
473 | ||
474 | struct C_MDC_CreateSystemFile : public MDCacheLogContext { | |
475 | MutationRef mut; | |
476 | CDentry *dn; | |
477 | version_t dpv; | |
478 | MDSInternalContextBase *fin; | |
479 | C_MDC_CreateSystemFile(MDCache *c, MutationRef& mu, CDentry *d, version_t v, MDSInternalContextBase *f) : | |
480 | MDCacheLogContext(c), mut(mu), dn(d), dpv(v), fin(f) {} | |
481 | void finish(int r) override { | |
482 | mdcache->_create_system_file_finish(mut, dn, dpv, fin); | |
483 | } | |
484 | }; | |
485 | ||
486 | void MDCache::_create_system_file(CDir *dir, const char *name, CInode *in, MDSInternalContextBase *fin) | |
487 | { | |
488 | dout(10) << "_create_system_file " << name << " in " << *dir << dendl; | |
489 | CDentry *dn = dir->add_null_dentry(name); | |
490 | ||
491 | dn->push_projected_linkage(in); | |
492 | version_t dpv = dn->pre_dirty(); | |
493 | ||
494 | CDir *mdir = 0; | |
495 | if (in->inode.is_dir()) { | |
496 | in->inode.rstat.rsubdirs = 1; | |
497 | ||
498 | mdir = in->get_or_open_dirfrag(this, frag_t()); | |
499 | mdir->mark_complete(); | |
500 | mdir->pre_dirty(); | |
501 | } else | |
502 | in->inode.rstat.rfiles = 1; | |
503 | in->inode.version = dn->pre_dirty(); | |
504 | ||
505 | SnapRealm *realm = dir->get_inode()->find_snaprealm(); | |
506 | dn->first = in->first = realm->get_newest_seq() + 1; | |
507 | ||
508 | MutationRef mut(new MutationImpl()); | |
509 | ||
510 | // force some locks. hacky. | |
511 | mds->locker->wrlock_force(&dir->inode->filelock, mut); | |
512 | mds->locker->wrlock_force(&dir->inode->nestlock, mut); | |
513 | ||
514 | mut->ls = mds->mdlog->get_current_segment(); | |
515 | EUpdate *le = new EUpdate(mds->mdlog, "create system file"); | |
516 | mds->mdlog->start_entry(le); | |
517 | ||
518 | if (!in->is_mdsdir()) { | |
519 | predirty_journal_parents(mut, &le->metablob, in, dir, PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
520 | le->metablob.add_primary_dentry(dn, in, true); | |
521 | } else { | |
522 | predirty_journal_parents(mut, &le->metablob, in, dir, PREDIRTY_DIR, 1); | |
523 | journal_dirty_inode(mut.get(), &le->metablob, in); | |
524 | dn->push_projected_linkage(in->ino(), in->d_type()); | |
525 | le->metablob.add_remote_dentry(dn, true, in->ino(), in->d_type()); | |
526 | le->metablob.add_root(true, in); | |
527 | } | |
528 | if (mdir) | |
529 | le->metablob.add_new_dir(mdir); // dirty AND complete AND new | |
530 | ||
531 | mds->mdlog->submit_entry(le, new C_MDC_CreateSystemFile(this, mut, dn, dpv, fin)); | |
532 | mds->mdlog->flush(); | |
533 | } | |
534 | ||
535 | void MDCache::_create_system_file_finish(MutationRef& mut, CDentry *dn, version_t dpv, MDSInternalContextBase *fin) | |
536 | { | |
537 | dout(10) << "_create_system_file_finish " << *dn << dendl; | |
538 | ||
539 | dn->pop_projected_linkage(); | |
540 | dn->mark_dirty(dpv, mut->ls); | |
541 | ||
542 | CInode *in = dn->get_linkage()->get_inode(); | |
543 | in->inode.version--; | |
544 | in->mark_dirty(in->inode.version + 1, mut->ls); | |
545 | ||
546 | if (in->inode.is_dir()) { | |
547 | CDir *dir = in->get_dirfrag(frag_t()); | |
548 | assert(dir); | |
549 | dir->mark_dirty(1, mut->ls); | |
550 | dir->mark_new(mut->ls); | |
551 | } | |
552 | ||
553 | mut->apply(); | |
554 | mds->locker->drop_locks(mut.get()); | |
555 | mut->cleanup(); | |
556 | ||
557 | fin->complete(0); | |
558 | ||
559 | //if (dir && MDS_INO_IS_MDSDIR(in->ino())) | |
560 | //migrator->export_dir(dir, (int)in->ino() - MDS_INO_MDSDIR_OFFSET); | |
561 | } | |
562 | ||
563 | ||
564 | ||
565 | struct C_MDS_RetryOpenRoot : public MDSInternalContext { | |
566 | MDCache *cache; | |
567 | explicit C_MDS_RetryOpenRoot(MDCache *c) : MDSInternalContext(c->mds), cache(c) {} | |
568 | void finish(int r) override { | |
569 | if (r < 0) { | |
570 | // If we can't open root, something disastrous has happened: mark | |
571 | // this rank damaged for operator intervention. Note that | |
572 | // it is not okay to call suicide() here because we are in | |
573 | // a Finisher callback. | |
574 | cache->mds->damaged(); | |
575 | ceph_abort(); // damaged should never return | |
576 | } else { | |
577 | cache->open_root(); | |
578 | } | |
579 | } | |
580 | }; | |
581 | ||
582 | void MDCache::open_root_inode(MDSInternalContextBase *c) | |
583 | { | |
584 | if (mds->get_nodeid() == mds->mdsmap->get_root()) { | |
585 | CInode *in; | |
586 | in = create_system_inode(MDS_INO_ROOT, S_IFDIR|0755); // initially inaccurate! | |
587 | in->fetch(c); | |
588 | } else { | |
589 | discover_base_ino(MDS_INO_ROOT, c, mds->mdsmap->get_root()); | |
590 | } | |
591 | } | |
592 | ||
593 | void MDCache::open_mydir_inode(MDSInternalContextBase *c) | |
594 | { | |
595 | MDSGatherBuilder gather(g_ceph_context); | |
596 | ||
597 | CInode *in = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR|0755); // initially inaccurate! | |
598 | in->fetch(gather.new_sub()); | |
599 | ||
600 | gather.set_finisher(c); | |
601 | gather.activate(); | |
602 | } | |
603 | ||
604 | void MDCache::open_root() | |
605 | { | |
606 | dout(10) << "open_root" << dendl; | |
607 | ||
608 | if (!root) { | |
609 | open_root_inode(new C_MDS_RetryOpenRoot(this)); | |
610 | return; | |
611 | } | |
612 | if (mds->get_nodeid() == mds->mdsmap->get_root()) { | |
613 | assert(root->is_auth()); | |
614 | CDir *rootdir = root->get_or_open_dirfrag(this, frag_t()); | |
615 | assert(rootdir); | |
616 | if (!rootdir->is_subtree_root()) | |
617 | adjust_subtree_auth(rootdir, mds->get_nodeid()); | |
618 | if (!rootdir->is_complete()) { | |
619 | rootdir->fetch(new C_MDS_RetryOpenRoot(this)); | |
620 | return; | |
621 | } | |
622 | } else { | |
623 | assert(!root->is_auth()); | |
624 | CDir *rootdir = root->get_dirfrag(frag_t()); | |
625 | if (!rootdir) { | |
224ce89b | 626 | open_remote_dirfrag(root, frag_t(), new C_MDS_RetryOpenRoot(this)); |
7c673cae FG |
627 | return; |
628 | } | |
629 | } | |
630 | ||
631 | if (!myin) { | |
632 | CInode *in = create_system_inode(MDS_INO_MDSDIR(mds->get_nodeid()), S_IFDIR|0755); // initially inaccurate! | |
633 | in->fetch(new C_MDS_RetryOpenRoot(this)); | |
634 | return; | |
635 | } | |
636 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); | |
637 | assert(mydir); | |
638 | adjust_subtree_auth(mydir, mds->get_nodeid()); | |
639 | ||
640 | populate_mydir(); | |
641 | } | |
642 | ||
643 | void MDCache::populate_mydir() | |
644 | { | |
645 | assert(myin); | |
646 | CDir *mydir = myin->get_or_open_dirfrag(this, frag_t()); | |
647 | assert(mydir); | |
648 | ||
649 | dout(10) << "populate_mydir " << *mydir << dendl; | |
650 | ||
651 | if (!mydir->is_complete()) { | |
652 | mydir->fetch(new C_MDS_RetryOpenRoot(this)); | |
653 | return; | |
654 | } | |
655 | ||
656 | if (mydir->get_version() == 0 && mydir->state_test(CDir::STATE_BADFRAG)) { | |
657 | // A missing dirfrag, we will recreate it. Before that, we must dirty | |
658 | // it before dirtying any of the strays we create within it. | |
659 | mds->clog->warn() << "fragment " << mydir->dirfrag() << " was unreadable, " | |
660 | "recreating it now"; | |
661 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
662 | mydir->state_clear(CDir::STATE_BADFRAG); | |
663 | mydir->mark_complete(); | |
664 | mydir->mark_dirty(mydir->pre_dirty(), ls); | |
665 | } | |
666 | ||
667 | // open or create stray | |
668 | uint64_t num_strays = 0; | |
669 | for (int i = 0; i < NUM_STRAY; ++i) { | |
670 | stringstream name; | |
671 | name << "stray" << i; | |
672 | CDentry *straydn = mydir->lookup(name.str()); | |
673 | ||
674 | // allow for older fs's with stray instead of stray0 | |
675 | if (straydn == NULL && i == 0) | |
676 | straydn = mydir->lookup("stray"); | |
677 | ||
678 | if (!straydn || !straydn->get_linkage()->get_inode()) { | |
679 | _create_system_file(mydir, name.str().c_str(), create_system_inode(MDS_INO_STRAY(mds->get_nodeid(), i), S_IFDIR), | |
680 | new C_MDS_RetryOpenRoot(this)); | |
681 | return; | |
682 | } | |
683 | assert(straydn); | |
684 | assert(strays[i]); | |
685 | // we make multiple passes through this method; make sure we only pin each stray once. | |
686 | if (!strays[i]->state_test(CInode::STATE_STRAYPINNED)) { | |
687 | strays[i]->get(CInode::PIN_STRAY); | |
688 | strays[i]->state_set(CInode::STATE_STRAYPINNED); | |
689 | strays[i]->get_stickydirs(); | |
690 | } | |
691 | dout(20) << " stray num " << i << " is " << *strays[i] << dendl; | |
692 | ||
693 | // open all frags | |
694 | list<frag_t> ls; | |
695 | strays[i]->dirfragtree.get_leaves(ls); | |
696 | for (list<frag_t>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
697 | frag_t fg = *p; | |
698 | CDir *dir = strays[i]->get_dirfrag(fg); | |
699 | if (!dir) { | |
700 | dir = strays[i]->get_or_open_dirfrag(this, fg); | |
701 | } | |
702 | ||
703 | // DamageTable applies special handling to strays: it will | |
704 | // have damaged() us out if one is damaged. | |
705 | assert(!dir->state_test(CDir::STATE_BADFRAG)); | |
706 | ||
707 | if (dir->get_version() == 0) { | |
708 | dir->fetch(new C_MDS_RetryOpenRoot(this)); | |
709 | return; | |
710 | } | |
711 | ||
712 | if (dir->get_frag_size() > 0) | |
713 | num_strays += dir->get_frag_size(); | |
714 | } | |
715 | } | |
716 | ||
717 | stray_manager.set_num_strays(num_strays); | |
718 | ||
719 | // okay! | |
720 | dout(10) << "populate_mydir done" << dendl; | |
721 | assert(!open); | |
722 | open = true; | |
723 | mds->queue_waiters(waiting_for_open); | |
724 | ||
725 | scan_stray_dir(); | |
726 | } | |
727 | ||
728 | void MDCache::open_foreign_mdsdir(inodeno_t ino, MDSInternalContextBase *fin) | |
729 | { | |
730 | discover_base_ino(ino, fin, mds_rank_t(ino & (MAX_MDS-1))); | |
731 | } | |
732 | ||
733 | CDir *MDCache::get_stray_dir(CInode *in) | |
734 | { | |
735 | string straydname; | |
736 | in->name_stray_dentry(straydname); | |
737 | ||
738 | CInode *strayi = get_stray(); | |
739 | assert(strayi); | |
740 | frag_t fg = strayi->pick_dirfrag(straydname); | |
741 | CDir *straydir = strayi->get_dirfrag(fg); | |
742 | assert(straydir); | |
743 | return straydir; | |
744 | } | |
745 | ||
746 | CDentry *MDCache::get_or_create_stray_dentry(CInode *in) | |
747 | { | |
748 | CDir *straydir = get_stray_dir(in); | |
749 | string straydname; | |
750 | in->name_stray_dentry(straydname); | |
751 | CDentry *straydn = straydir->lookup(straydname); | |
752 | if (!straydn) { | |
753 | straydn = straydir->add_null_dentry(straydname); | |
754 | straydn->mark_new(); | |
755 | } else { | |
756 | assert(straydn->get_projected_linkage()->is_null()); | |
757 | } | |
758 | ||
759 | straydn->state_set(CDentry::STATE_STRAY); | |
760 | return straydn; | |
761 | } | |
762 | ||
763 | ||
764 | ||
765 | MDSCacheObject *MDCache::get_object(MDSCacheObjectInfo &info) | |
766 | { | |
767 | // inode? | |
768 | if (info.ino) | |
769 | return get_inode(info.ino, info.snapid); | |
770 | ||
771 | // dir or dentry. | |
772 | CDir *dir = get_dirfrag(info.dirfrag); | |
773 | if (!dir) return 0; | |
774 | ||
775 | if (info.dname.length()) | |
776 | return dir->lookup(info.dname, info.snapid); | |
777 | else | |
778 | return dir; | |
779 | } | |
780 | ||
781 | ||
782 | ||
783 | ||
784 | // ==================================================================== | |
785 | // subtree management | |
786 | ||
787 | void MDCache::list_subtrees(list<CDir*>& ls) | |
788 | { | |
789 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
790 | p != subtrees.end(); | |
791 | ++p) | |
792 | ls.push_back(p->first); | |
793 | } | |
794 | ||
795 | /* | |
796 | * adjust the dir_auth of a subtree. | |
797 | * merge with parent and/or child subtrees, if is it appropriate. | |
798 | * merge can ONLY happen if both parent and child have unambiguous auth. | |
799 | */ | |
224ce89b | 800 | void MDCache::adjust_subtree_auth(CDir *dir, mds_authority_t auth) |
7c673cae FG |
801 | { |
802 | dout(7) << "adjust_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
803 | << " on " << *dir << dendl; | |
804 | ||
7c673cae FG |
805 | show_subtrees(); |
806 | ||
807 | CDir *root; | |
808 | if (dir->inode->is_base()) { | |
809 | root = dir; // bootstrap hack. | |
810 | if (subtrees.count(root) == 0) { | |
811 | subtrees[root]; | |
812 | root->get(CDir::PIN_SUBTREE); | |
813 | } | |
814 | } else { | |
815 | root = get_subtree_root(dir); // subtree root | |
816 | } | |
817 | assert(root); | |
818 | assert(subtrees.count(root)); | |
819 | dout(7) << " current root is " << *root << dendl; | |
820 | ||
821 | if (root == dir) { | |
822 | // i am already a subtree. | |
823 | dir->set_dir_auth(auth); | |
824 | } else { | |
825 | // i am a new subtree. | |
826 | dout(10) << " new subtree at " << *dir << dendl; | |
827 | assert(subtrees.count(dir) == 0); | |
828 | subtrees[dir]; // create empty subtree bounds list for me. | |
829 | dir->get(CDir::PIN_SUBTREE); | |
830 | ||
831 | // set dir_auth | |
832 | dir->set_dir_auth(auth); | |
833 | ||
834 | // move items nested beneath me, under me. | |
835 | set<CDir*>::iterator p = subtrees[root].begin(); | |
836 | while (p != subtrees[root].end()) { | |
837 | set<CDir*>::iterator next = p; | |
838 | ++next; | |
839 | if (get_subtree_root((*p)->get_parent_dir()) == dir) { | |
840 | // move under me | |
841 | dout(10) << " claiming child bound " << **p << dendl; | |
842 | subtrees[dir].insert(*p); | |
843 | subtrees[root].erase(p); | |
844 | } | |
845 | p = next; | |
846 | } | |
847 | ||
848 | // i am a bound of the parent subtree. | |
849 | subtrees[root].insert(dir); | |
850 | ||
851 | // i am now the subtree root. | |
852 | root = dir; | |
853 | ||
854 | // adjust recursive pop counters | |
855 | if (dir->is_auth()) { | |
856 | utime_t now = ceph_clock_now(); | |
857 | CDir *p = dir->get_parent_dir(); | |
858 | while (p) { | |
859 | p->pop_auth_subtree.sub(now, decayrate, dir->pop_auth_subtree); | |
860 | if (p->is_subtree_root()) break; | |
861 | p = p->inode->get_parent_dir(); | |
862 | } | |
863 | } | |
7c673cae FG |
864 | } |
865 | ||
866 | show_subtrees(); | |
867 | } | |
868 | ||
869 | ||
870 | void MDCache::try_subtree_merge(CDir *dir) | |
871 | { | |
872 | dout(7) << "try_subtree_merge " << *dir << dendl; | |
873 | assert(subtrees.count(dir)); | |
874 | set<CDir*> oldbounds = subtrees[dir]; | |
875 | ||
224ce89b | 876 | set<CInode*> to_eval; |
7c673cae | 877 | // try merge at my root |
224ce89b | 878 | try_subtree_merge_at(dir, &to_eval); |
7c673cae FG |
879 | |
880 | // try merge at my old bounds | |
224ce89b WB |
881 | for (auto bound : oldbounds) |
882 | try_subtree_merge_at(bound, &to_eval); | |
883 | ||
884 | if (!(mds->is_any_replay() || mds->is_resolve())) { | |
885 | for(auto in : to_eval) | |
886 | eval_subtree_root(in); | |
887 | } | |
7c673cae FG |
888 | } |
889 | ||
890 | class C_MDC_SubtreeMergeWB : public MDCacheLogContext { | |
891 | CInode *in; | |
892 | MutationRef mut; | |
893 | public: | |
894 | C_MDC_SubtreeMergeWB(MDCache *mdc, CInode *i, MutationRef& m) : MDCacheLogContext(mdc), in(i), mut(m) {} | |
895 | void finish(int r) override { | |
896 | mdcache->subtree_merge_writebehind_finish(in, mut); | |
897 | } | |
898 | }; | |
899 | ||
224ce89b | 900 | void MDCache::try_subtree_merge_at(CDir *dir, set<CInode*> *to_eval) |
7c673cae FG |
901 | { |
902 | dout(10) << "try_subtree_merge_at " << *dir << dendl; | |
903 | assert(subtrees.count(dir)); | |
904 | ||
7c673cae FG |
905 | // merge with parent? |
906 | CDir *parent = dir; | |
907 | if (!dir->inode->is_base()) | |
908 | parent = get_subtree_root(dir->get_parent_dir()); | |
909 | ||
910 | if (parent != dir && // we have a parent, | |
911 | parent->dir_auth == dir->dir_auth && // auth matches, | |
912 | dir->dir_auth.second == CDIR_AUTH_UNKNOWN && // auth is unambiguous, | |
913 | !dir->state_test(CDir::STATE_EXPORTBOUND) && // not an exportbound, | |
914 | !dir->state_test(CDir::STATE_AUXSUBTREE)) { // not aux subtree | |
915 | // merge with parent. | |
916 | dout(10) << " subtree merge at " << *dir << dendl; | |
917 | dir->set_dir_auth(CDIR_AUTH_DEFAULT); | |
918 | ||
919 | // move our bounds under the parent | |
920 | for (set<CDir*>::iterator p = subtrees[dir].begin(); | |
921 | p != subtrees[dir].end(); | |
922 | ++p) | |
923 | subtrees[parent].insert(*p); | |
924 | ||
925 | // we are no longer a subtree or bound | |
926 | dir->put(CDir::PIN_SUBTREE); | |
927 | subtrees.erase(dir); | |
928 | subtrees[parent].erase(dir); | |
929 | ||
930 | // adjust popularity? | |
931 | if (dir->is_auth()) { | |
932 | utime_t now = ceph_clock_now(); | |
933 | CDir *p = dir->get_parent_dir(); | |
934 | while (p) { | |
935 | p->pop_auth_subtree.add(now, decayrate, dir->pop_auth_subtree); | |
936 | if (p->is_subtree_root()) break; | |
937 | p = p->inode->get_parent_dir(); | |
938 | } | |
939 | } | |
940 | ||
224ce89b WB |
941 | if (to_eval && dir->get_inode()->is_auth()) |
942 | to_eval->insert(dir->get_inode()); | |
7c673cae | 943 | |
181888fb FG |
944 | show_subtrees(15); |
945 | } | |
7c673cae FG |
946 | } |
947 | ||
948 | void MDCache::subtree_merge_writebehind_finish(CInode *in, MutationRef& mut) | |
949 | { | |
950 | dout(10) << "subtree_merge_writebehind_finish on " << in << dendl; | |
951 | in->pop_and_dirty_projected_inode(mut->ls); | |
952 | ||
953 | mut->apply(); | |
954 | mds->locker->drop_locks(mut.get()); | |
955 | mut->cleanup(); | |
956 | ||
957 | in->auth_unpin(this); | |
958 | } | |
959 | ||
960 | void MDCache::eval_subtree_root(CInode *diri) | |
961 | { | |
962 | // evaluate subtree inode filelock? | |
963 | // (we should scatter the filelock on subtree bounds) | |
224ce89b WB |
964 | assert(diri->is_auth()); |
965 | mds->locker->try_eval(diri, CEPH_LOCK_IFILE | CEPH_LOCK_INEST); | |
7c673cae FG |
966 | } |
967 | ||
968 | ||
969 | void MDCache::adjust_bounded_subtree_auth(CDir *dir, set<CDir*>& bounds, mds_authority_t auth) | |
970 | { | |
971 | dout(7) << "adjust_bounded_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
972 | << " on " << *dir | |
973 | << " bounds " << bounds | |
974 | << dendl; | |
975 | ||
976 | show_subtrees(); | |
977 | ||
978 | CDir *root; | |
979 | if (dir->ino() == MDS_INO_ROOT) { | |
980 | root = dir; // bootstrap hack. | |
981 | if (subtrees.count(root) == 0) { | |
982 | subtrees[root]; | |
983 | root->get(CDir::PIN_SUBTREE); | |
984 | } | |
985 | } else { | |
986 | root = get_subtree_root(dir); // subtree root | |
987 | } | |
988 | assert(root); | |
989 | assert(subtrees.count(root)); | |
990 | dout(7) << " current root is " << *root << dendl; | |
991 | ||
992 | mds_authority_t oldauth = dir->authority(); | |
993 | ||
994 | if (root == dir) { | |
995 | // i am already a subtree. | |
996 | dir->set_dir_auth(auth); | |
997 | } else { | |
998 | // i am a new subtree. | |
999 | dout(10) << " new subtree at " << *dir << dendl; | |
1000 | assert(subtrees.count(dir) == 0); | |
1001 | subtrees[dir]; // create empty subtree bounds list for me. | |
1002 | dir->get(CDir::PIN_SUBTREE); | |
1003 | ||
1004 | // set dir_auth | |
1005 | dir->set_dir_auth(auth); | |
1006 | ||
1007 | // move items nested beneath me, under me. | |
1008 | set<CDir*>::iterator p = subtrees[root].begin(); | |
1009 | while (p != subtrees[root].end()) { | |
1010 | set<CDir*>::iterator next = p; | |
1011 | ++next; | |
1012 | if (get_subtree_root((*p)->get_parent_dir()) == dir) { | |
1013 | // move under me | |
1014 | dout(10) << " claiming child bound " << **p << dendl; | |
1015 | subtrees[dir].insert(*p); | |
1016 | subtrees[root].erase(p); | |
1017 | } | |
1018 | p = next; | |
1019 | } | |
1020 | ||
1021 | // i am a bound of the parent subtree. | |
1022 | subtrees[root].insert(dir); | |
1023 | ||
1024 | // i am now the subtree root. | |
1025 | root = dir; | |
1026 | } | |
1027 | ||
224ce89b WB |
1028 | set<CInode*> to_eval; |
1029 | ||
7c673cae FG |
1030 | // verify/adjust bounds. |
1031 | // - these may be new, or | |
1032 | // - beneath existing ambiguous bounds (which will be collapsed), | |
1033 | // - but NOT beneath unambiguous bounds. | |
1034 | for (set<CDir*>::iterator p = bounds.begin(); | |
1035 | p != bounds.end(); | |
1036 | ++p) { | |
1037 | CDir *bound = *p; | |
1038 | ||
1039 | // new bound? | |
1040 | if (subtrees[dir].count(bound) == 0) { | |
1041 | if (get_subtree_root(bound) == dir) { | |
1042 | dout(10) << " new bound " << *bound << ", adjusting auth back to old " << oldauth << dendl; | |
1043 | adjust_subtree_auth(bound, oldauth); // otherwise, adjust at bound. | |
1044 | } | |
1045 | else { | |
1046 | dout(10) << " want bound " << *bound << dendl; | |
1047 | CDir *t = get_subtree_root(bound->get_parent_dir()); | |
1048 | if (subtrees[t].count(bound) == 0) { | |
1049 | assert(t != dir); | |
1050 | dout(10) << " new bound " << *bound << dendl; | |
1051 | adjust_subtree_auth(bound, t->authority()); | |
1052 | } | |
1053 | // make sure it's nested beneath ambiguous subtree(s) | |
1054 | while (1) { | |
1055 | while (subtrees[dir].count(t) == 0) | |
1056 | t = get_subtree_root(t->get_parent_dir()); | |
1057 | dout(10) << " swallowing intervening subtree at " << *t << dendl; | |
1058 | adjust_subtree_auth(t, auth); | |
224ce89b | 1059 | try_subtree_merge_at(t, &to_eval); |
7c673cae FG |
1060 | t = get_subtree_root(bound->get_parent_dir()); |
1061 | if (t == dir) break; | |
1062 | } | |
1063 | } | |
1064 | } | |
1065 | else { | |
1066 | dout(10) << " already have bound " << *bound << dendl; | |
1067 | } | |
1068 | } | |
1069 | // merge stray bounds? | |
1070 | while (!subtrees[dir].empty()) { | |
1071 | set<CDir*> copy = subtrees[dir]; | |
1072 | for (set<CDir*>::iterator p = copy.begin(); p != copy.end(); ++p) { | |
1073 | if (bounds.count(*p) == 0) { | |
1074 | CDir *stray = *p; | |
1075 | dout(10) << " swallowing extra subtree at " << *stray << dendl; | |
1076 | adjust_subtree_auth(stray, auth); | |
224ce89b | 1077 | try_subtree_merge_at(stray, &to_eval); |
7c673cae FG |
1078 | } |
1079 | } | |
1080 | // swallowing subtree may add new subtree bounds | |
1081 | if (copy == subtrees[dir]) | |
1082 | break; | |
1083 | } | |
1084 | ||
1085 | // bound should now match. | |
1086 | verify_subtree_bounds(dir, bounds); | |
1087 | ||
1088 | show_subtrees(); | |
224ce89b WB |
1089 | |
1090 | if (!(mds->is_any_replay() || mds->is_resolve())) { | |
1091 | for(auto in : to_eval) | |
1092 | eval_subtree_root(in); | |
1093 | } | |
7c673cae FG |
1094 | } |
1095 | ||
1096 | ||
1097 | /* | |
1098 | * return a set of CDir*'s that correspond to the given bound set. Only adjust | |
1099 | * fragmentation as necessary to get an equivalent bounding set. That is, only | |
1100 | * split if one of our frags spans the provided bounding set. Never merge. | |
1101 | */ | |
1102 | void MDCache::get_force_dirfrag_bound_set(vector<dirfrag_t>& dfs, set<CDir*>& bounds) | |
1103 | { | |
1104 | dout(10) << "get_force_dirfrag_bound_set " << dfs << dendl; | |
1105 | ||
1106 | // sort by ino | |
1107 | map<inodeno_t, fragset_t> byino; | |
1108 | for (vector<dirfrag_t>::iterator p = dfs.begin(); p != dfs.end(); ++p) | |
1109 | byino[p->ino].insert(p->frag); | |
1110 | dout(10) << " by ino: " << byino << dendl; | |
1111 | ||
1112 | for (map<inodeno_t,fragset_t>::iterator p = byino.begin(); p != byino.end(); ++p) { | |
1113 | CInode *diri = get_inode(p->first); | |
1114 | if (!diri) | |
1115 | continue; | |
1116 | dout(10) << " checking fragset " << p->second.get() << " on " << *diri << dendl; | |
1117 | ||
1118 | fragtree_t tmpdft; | |
1119 | for (set<frag_t>::iterator q = p->second.begin(); q != p->second.end(); ++q) | |
1120 | tmpdft.force_to_leaf(g_ceph_context, *q); | |
1121 | ||
1122 | for (set<frag_t>::iterator q = p->second.begin(); q != p->second.end(); ++q) { | |
1123 | frag_t fg = *q; | |
1124 | list<frag_t> fgls; | |
1125 | diri->dirfragtree.get_leaves_under(fg, fgls); | |
1126 | if (fgls.empty()) { | |
1127 | bool all = true; | |
1128 | frag_t approx_fg = diri->dirfragtree[fg.value()]; | |
1129 | list<frag_t> ls; | |
1130 | tmpdft.get_leaves_under(approx_fg, ls); | |
1131 | for (list<frag_t>::iterator r = ls.begin(); r != ls.end(); ++r) { | |
1132 | if (p->second.get().count(*r) == 0) { | |
1133 | // not bound, so the resolve message is from auth MDS of the dirfrag | |
1134 | force_dir_fragment(diri, *r); | |
1135 | all = false; | |
1136 | } | |
1137 | } | |
1138 | if (all) | |
1139 | fgls.push_back(approx_fg); | |
1140 | else | |
1141 | diri->dirfragtree.get_leaves_under(fg, fgls); | |
1142 | } | |
1143 | dout(10) << " frag " << fg << " contains " << fgls << dendl; | |
1144 | for (list<frag_t>::iterator r = fgls.begin(); r != fgls.end(); ++r) { | |
1145 | CDir *dir = diri->get_dirfrag(*r); | |
1146 | if (dir) | |
1147 | bounds.insert(dir); | |
1148 | } | |
1149 | } | |
1150 | } | |
1151 | } | |
1152 | ||
1153 | void MDCache::adjust_bounded_subtree_auth(CDir *dir, vector<dirfrag_t>& bound_dfs, mds_authority_t auth) | |
1154 | { | |
1155 | dout(7) << "adjust_bounded_subtree_auth " << dir->get_dir_auth() << " -> " << auth | |
1156 | << " on " << *dir << " bound_dfs " << bound_dfs << dendl; | |
1157 | ||
1158 | set<CDir*> bounds; | |
1159 | get_force_dirfrag_bound_set(bound_dfs, bounds); | |
1160 | adjust_bounded_subtree_auth(dir, bounds, auth); | |
1161 | } | |
1162 | ||
1163 | void MDCache::map_dirfrag_set(list<dirfrag_t>& dfs, set<CDir*>& result) | |
1164 | { | |
1165 | dout(10) << "map_dirfrag_set " << dfs << dendl; | |
1166 | ||
1167 | // group by inode | |
1168 | map<inodeno_t, fragset_t> ino_fragset; | |
1169 | for (list<dirfrag_t>::iterator p = dfs.begin(); p != dfs.end(); ++p) | |
1170 | ino_fragset[p->ino].insert(p->frag); | |
1171 | ||
1172 | // get frags | |
1173 | for (map<inodeno_t, fragset_t>::iterator p = ino_fragset.begin(); | |
1174 | p != ino_fragset.end(); | |
1175 | ++p) { | |
1176 | CInode *in = get_inode(p->first); | |
1177 | if (!in) | |
1178 | continue; | |
1179 | ||
1180 | list<frag_t> fglist; | |
1181 | for (set<frag_t>::iterator q = p->second.begin(); q != p->second.end(); ++q) | |
1182 | in->dirfragtree.get_leaves_under(*q, fglist); | |
1183 | ||
1184 | dout(15) << "map_dirfrag_set " << p->second << " -> " << fglist | |
1185 | << " on " << *in << dendl; | |
1186 | ||
1187 | for (list<frag_t>::iterator q = fglist.begin(); q != fglist.end(); ++q) { | |
1188 | CDir *dir = in->get_dirfrag(*q); | |
1189 | if (dir) | |
1190 | result.insert(dir); | |
1191 | } | |
1192 | } | |
1193 | } | |
1194 | ||
1195 | ||
1196 | ||
1197 | CDir *MDCache::get_subtree_root(CDir *dir) | |
1198 | { | |
1199 | // find the underlying dir that delegates (or is about to delegate) auth | |
1200 | while (true) { | |
1201 | if (dir->is_subtree_root()) | |
1202 | return dir; | |
1203 | dir = dir->get_inode()->get_parent_dir(); | |
1204 | if (!dir) | |
1205 | return 0; // none | |
1206 | } | |
1207 | } | |
1208 | ||
1209 | CDir *MDCache::get_projected_subtree_root(CDir *dir) | |
1210 | { | |
1211 | // find the underlying dir that delegates (or is about to delegate) auth | |
1212 | while (true) { | |
1213 | if (dir->is_subtree_root()) | |
1214 | return dir; | |
1215 | dir = dir->get_inode()->get_projected_parent_dir(); | |
1216 | if (!dir) | |
1217 | return 0; // none | |
1218 | } | |
1219 | } | |
1220 | ||
1221 | void MDCache::remove_subtree(CDir *dir) | |
1222 | { | |
1223 | dout(10) << "remove_subtree " << *dir << dendl; | |
1224 | assert(subtrees.count(dir)); | |
1225 | assert(subtrees[dir].empty()); | |
1226 | subtrees.erase(dir); | |
1227 | dir->put(CDir::PIN_SUBTREE); | |
1228 | if (dir->get_parent_dir()) { | |
1229 | CDir *p = get_subtree_root(dir->get_parent_dir()); | |
1230 | assert(subtrees[p].count(dir)); | |
1231 | subtrees[p].erase(dir); | |
1232 | } | |
1233 | } | |
1234 | ||
1235 | void MDCache::get_subtree_bounds(CDir *dir, set<CDir*>& bounds) | |
1236 | { | |
1237 | assert(subtrees.count(dir)); | |
1238 | bounds = subtrees[dir]; | |
1239 | } | |
1240 | ||
1241 | void MDCache::get_wouldbe_subtree_bounds(CDir *dir, set<CDir*>& bounds) | |
1242 | { | |
1243 | if (subtrees.count(dir)) { | |
1244 | // just copy them, dir is a subtree. | |
1245 | get_subtree_bounds(dir, bounds); | |
1246 | } else { | |
1247 | // find them | |
1248 | CDir *root = get_subtree_root(dir); | |
1249 | for (set<CDir*>::iterator p = subtrees[root].begin(); | |
1250 | p != subtrees[root].end(); | |
1251 | ++p) { | |
1252 | CDir *t = *p; | |
1253 | while (t != root) { | |
1254 | t = t->get_parent_dir(); | |
1255 | assert(t); | |
1256 | if (t == dir) { | |
1257 | bounds.insert(*p); | |
1258 | continue; | |
1259 | } | |
1260 | } | |
1261 | } | |
1262 | } | |
1263 | } | |
1264 | ||
1265 | void MDCache::verify_subtree_bounds(CDir *dir, const set<CDir*>& bounds) | |
1266 | { | |
1267 | // for debugging only. | |
1268 | assert(subtrees.count(dir)); | |
1269 | if (bounds != subtrees[dir]) { | |
1270 | dout(0) << "verify_subtree_bounds failed" << dendl; | |
1271 | set<CDir*> b = bounds; | |
1272 | for (auto &cd : subtrees[dir]) { | |
1273 | if (bounds.count(cd)) { | |
1274 | b.erase(cd); | |
1275 | continue; | |
1276 | } | |
1277 | dout(0) << " missing bound " << *cd << dendl; | |
1278 | } | |
1279 | for (const auto &cd : b) | |
1280 | dout(0) << " extra bound " << *cd << dendl; | |
1281 | } | |
1282 | assert(bounds == subtrees[dir]); | |
1283 | } | |
1284 | ||
1285 | void MDCache::verify_subtree_bounds(CDir *dir, const list<dirfrag_t>& bounds) | |
1286 | { | |
1287 | // for debugging only. | |
1288 | assert(subtrees.count(dir)); | |
1289 | ||
1290 | // make sure that any bounds i do have are properly noted as such. | |
1291 | int failed = 0; | |
1292 | for (const auto &fg : bounds) { | |
1293 | CDir *bd = get_dirfrag(fg); | |
1294 | if (!bd) continue; | |
1295 | if (subtrees[dir].count(bd) == 0) { | |
1296 | dout(0) << "verify_subtree_bounds failed: extra bound " << *bd << dendl; | |
1297 | failed++; | |
1298 | } | |
1299 | } | |
1300 | assert(failed == 0); | |
1301 | } | |
1302 | ||
1303 | void MDCache::project_subtree_rename(CInode *diri, CDir *olddir, CDir *newdir) | |
1304 | { | |
1305 | dout(10) << "project_subtree_rename " << *diri << " from " << *olddir | |
1306 | << " to " << *newdir << dendl; | |
1307 | projected_subtree_renames[diri].push_back(pair<CDir*,CDir*>(olddir, newdir)); | |
1308 | } | |
1309 | ||
224ce89b | 1310 | void MDCache::adjust_subtree_after_rename(CInode *diri, CDir *olddir, bool pop) |
7c673cae FG |
1311 | { |
1312 | dout(10) << "adjust_subtree_after_rename " << *diri << " from " << *olddir << dendl; | |
1313 | ||
1314 | //show_subtrees(); | |
1315 | ||
1316 | CDir *newdir = diri->get_parent_dir(); | |
1317 | ||
1318 | if (pop) { | |
1319 | map<CInode*,list<pair<CDir*,CDir*> > >::iterator p = projected_subtree_renames.find(diri); | |
1320 | assert(p != projected_subtree_renames.end()); | |
1321 | assert(!p->second.empty()); | |
1322 | assert(p->second.front().first == olddir); | |
1323 | assert(p->second.front().second == newdir); | |
1324 | p->second.pop_front(); | |
1325 | if (p->second.empty()) | |
1326 | projected_subtree_renames.erase(p); | |
1327 | } | |
1328 | ||
1329 | // adjust subtree | |
1330 | list<CDir*> dfls; | |
1331 | // make sure subtree dirfrags are at the front of the list | |
1332 | diri->get_subtree_dirfrags(dfls); | |
1333 | diri->get_nested_dirfrags(dfls); | |
1334 | for (list<CDir*>::iterator p = dfls.begin(); p != dfls.end(); ++p) { | |
1335 | CDir *dir = *p; | |
1336 | ||
1337 | dout(10) << "dirfrag " << *dir << dendl; | |
1338 | CDir *oldparent = get_subtree_root(olddir); | |
1339 | dout(10) << " old parent " << *oldparent << dendl; | |
1340 | CDir *newparent = get_subtree_root(newdir); | |
1341 | dout(10) << " new parent " << *newparent << dendl; | |
1342 | ||
1343 | if (oldparent == newparent) { | |
1344 | dout(10) << "parent unchanged for " << *dir << " at " << *oldparent << dendl; | |
1345 | continue; | |
1346 | } | |
1347 | ||
1348 | if (dir->is_subtree_root()) { | |
1349 | // children are fine. change parent. | |
1350 | dout(10) << "moving " << *dir << " from " << *oldparent << " to " << *newparent << dendl; | |
1351 | assert(subtrees[oldparent].count(dir)); | |
1352 | subtrees[oldparent].erase(dir); | |
1353 | assert(subtrees.count(newparent)); | |
1354 | subtrees[newparent].insert(dir); | |
224ce89b WB |
1355 | // caller is responsible for 'eval diri' |
1356 | try_subtree_merge_at(dir, NULL); | |
7c673cae FG |
1357 | } else { |
1358 | // mid-subtree. | |
1359 | ||
1360 | // see if any old bounds move to the new parent. | |
1361 | list<CDir*> tomove; | |
1362 | for (set<CDir*>::iterator p = subtrees[oldparent].begin(); | |
1363 | p != subtrees[oldparent].end(); | |
1364 | ++p) { | |
1365 | CDir *bound = *p; | |
1366 | CDir *broot = get_subtree_root(bound->get_parent_dir()); | |
1367 | if (broot != oldparent) { | |
1368 | assert(broot == newparent); | |
1369 | tomove.push_back(bound); | |
1370 | } | |
1371 | } | |
1372 | for (list<CDir*>::iterator p = tomove.begin(); p != tomove.end(); ++p) { | |
1373 | CDir *bound = *p; | |
1374 | dout(10) << "moving bound " << *bound << " from " << *oldparent << " to " << *newparent << dendl; | |
1375 | subtrees[oldparent].erase(bound); | |
1376 | subtrees[newparent].insert(bound); | |
1377 | } | |
1378 | ||
1379 | // did auth change? | |
1380 | if (oldparent->authority() != newparent->authority()) { | |
224ce89b WB |
1381 | adjust_subtree_auth(dir, oldparent->authority()); |
1382 | // caller is responsible for 'eval diri' | |
1383 | try_subtree_merge_at(dir, NULL); | |
7c673cae FG |
1384 | } |
1385 | } | |
1386 | } | |
1387 | ||
1388 | show_subtrees(); | |
1389 | } | |
1390 | ||
1391 | ||
1392 | void MDCache::get_fullauth_subtrees(set<CDir*>& s) | |
1393 | { | |
1394 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
1395 | p != subtrees.end(); | |
1396 | ++p) { | |
1397 | CDir *root = p->first; | |
1398 | if (root->is_full_dir_auth()) | |
1399 | s.insert(root); | |
1400 | } | |
1401 | } | |
1402 | void MDCache::get_auth_subtrees(set<CDir*>& s) | |
1403 | { | |
1404 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
1405 | p != subtrees.end(); | |
1406 | ++p) { | |
1407 | CDir *root = p->first; | |
1408 | if (root->is_auth()) | |
1409 | s.insert(root); | |
1410 | } | |
1411 | } | |
1412 | ||
1413 | ||
1414 | // count. | |
1415 | ||
1416 | int MDCache::num_subtrees() | |
1417 | { | |
1418 | return subtrees.size(); | |
1419 | } | |
1420 | ||
1421 | int MDCache::num_subtrees_fullauth() | |
1422 | { | |
1423 | int n = 0; | |
1424 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
1425 | p != subtrees.end(); | |
1426 | ++p) { | |
1427 | CDir *root = p->first; | |
1428 | if (root->is_full_dir_auth()) | |
1429 | n++; | |
1430 | } | |
1431 | return n; | |
1432 | } | |
1433 | ||
1434 | int MDCache::num_subtrees_fullnonauth() | |
1435 | { | |
1436 | int n = 0; | |
1437 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
1438 | p != subtrees.end(); | |
1439 | ++p) { | |
1440 | CDir *root = p->first; | |
1441 | if (root->is_full_dir_nonauth()) | |
1442 | n++; | |
1443 | } | |
1444 | return n; | |
1445 | } | |
1446 | ||
1447 | ||
1448 | ||
1449 | // =================================== | |
1450 | // journal and snap/cow helpers | |
1451 | ||
1452 | ||
1453 | /* | |
1454 | * find first inode in cache that follows given snapid. otherwise, return current. | |
1455 | */ | |
1456 | CInode *MDCache::pick_inode_snap(CInode *in, snapid_t follows) | |
1457 | { | |
1458 | dout(10) << "pick_inode_snap follows " << follows << " on " << *in << dendl; | |
1459 | assert(in->last == CEPH_NOSNAP); | |
1460 | ||
1461 | SnapRealm *realm = in->find_snaprealm(); | |
1462 | const set<snapid_t>& snaps = realm->get_snaps(); | |
1463 | dout(10) << " realm " << *realm << " " << *realm->inode << dendl; | |
1464 | dout(10) << " snaps " << snaps << dendl; | |
1465 | ||
1466 | if (snaps.empty()) | |
1467 | return in; | |
1468 | ||
1469 | for (set<snapid_t>::const_iterator p = snaps.upper_bound(follows); // first item > follows | |
1470 | p != snaps.end(); | |
1471 | ++p) { | |
1472 | CInode *t = get_inode(in->ino(), *p); | |
1473 | if (t) { | |
1474 | in = t; | |
1475 | dout(10) << "pick_inode_snap snap " << *p << " found " << *in << dendl; | |
1476 | break; | |
1477 | } | |
1478 | } | |
1479 | return in; | |
1480 | } | |
1481 | ||
1482 | ||
1483 | /* | |
1484 | * note: i'm currently cheating wrt dirty and inode.version on cow | |
1485 | * items. instead of doing a full dir predirty, i just take the | |
1486 | * original item's version, and set the dirty flag (via | |
1487 | * mutation::add_cow_{inode,dentry}() and mutation::apply(). that | |
1488 | * means a special case in the dir commit clean sweep assertions. | |
1489 | * bah. | |
1490 | */ | |
1491 | CInode *MDCache::cow_inode(CInode *in, snapid_t last) | |
1492 | { | |
1493 | assert(last >= in->first); | |
1494 | ||
1495 | SnapRealm *realm = in->find_snaprealm(); | |
1496 | const set<snapid_t>& snaps = realm->get_snaps(); | |
1497 | ||
1498 | // make sure snap inode's last match existing snapshots. | |
1499 | // MDCache::pick_inode_snap() requires this. | |
1500 | snapid_t last_snap = last; | |
1501 | if (snaps.count(last) == 0) { | |
1502 | set<snapid_t>::const_iterator p = snaps.upper_bound(last); | |
1503 | if (p != snaps.begin()) { | |
1504 | --p; | |
1505 | if (*p >= in->first) | |
1506 | last_snap = *p; | |
1507 | } | |
1508 | } | |
1509 | ||
1510 | CInode *oldin = new CInode(this, true, in->first, last_snap); | |
1511 | oldin->inode = *in->get_previous_projected_inode(); | |
1512 | oldin->symlink = in->symlink; | |
1513 | oldin->xattrs = *in->get_previous_projected_xattrs(); | |
1514 | oldin->inode.trim_client_ranges(last); | |
1515 | ||
1516 | if (in->first < in->oldest_snap) | |
1517 | in->oldest_snap = in->first; | |
1518 | ||
1519 | in->first = last+1; | |
1520 | ||
1521 | dout(10) << "cow_inode " << *in << " to " << *oldin << dendl; | |
1522 | add_inode(oldin); | |
1523 | ||
1524 | if (in->last != CEPH_NOSNAP) { | |
1525 | CInode *head_in = get_inode(in->ino()); | |
1526 | assert(head_in); | |
1527 | if (head_in->split_need_snapflush(oldin, in)) { | |
1528 | oldin->client_snap_caps = in->client_snap_caps; | |
1529 | for (compact_map<int,set<client_t> >::iterator p = in->client_snap_caps.begin(); | |
1530 | p != in->client_snap_caps.end(); | |
1531 | ++p) { | |
1532 | SimpleLock *lock = oldin->get_lock(p->first); | |
1533 | assert(lock); | |
1534 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
1535 | oldin->auth_pin(lock); | |
1536 | lock->set_state(LOCK_SNAP_SYNC); // gathering | |
1537 | lock->get_wrlock(true); | |
1538 | } | |
1539 | } | |
1540 | } | |
1541 | return oldin; | |
1542 | } | |
1543 | ||
1544 | // clone caps? | |
1545 | for (map<client_t,Capability*>::iterator p = in->client_caps.begin(); | |
1546 | p != in->client_caps.end(); | |
1547 | ++p) { | |
1548 | client_t client = p->first; | |
1549 | Capability *cap = p->second; | |
1550 | int issued = cap->issued(); | |
1551 | if ((issued & CEPH_CAP_ANY_WR) && | |
1552 | cap->client_follows < last) { | |
1553 | // note in oldin | |
1554 | for (int i = 0; i < num_cinode_locks; i++) { | |
1555 | if (issued & cinode_lock_info[i].wr_caps) { | |
1556 | int lockid = cinode_lock_info[i].lock; | |
1557 | SimpleLock *lock = oldin->get_lock(lockid); | |
1558 | assert(lock); | |
1559 | oldin->client_snap_caps[lockid].insert(client); | |
1560 | oldin->auth_pin(lock); | |
1561 | lock->set_state(LOCK_SNAP_SYNC); // gathering | |
1562 | lock->get_wrlock(true); | |
1563 | dout(10) << " client." << client << " cap " << ccap_string(issued & cinode_lock_info[i].wr_caps) | |
1564 | << " wrlock lock " << *lock << " on " << *oldin << dendl; | |
1565 | } | |
1566 | } | |
1567 | cap->client_follows = last; | |
1568 | ||
1569 | // we need snapflushes for any intervening snaps | |
1570 | dout(10) << " snaps " << snaps << dendl; | |
1571 | for (set<snapid_t>::const_iterator q = snaps.lower_bound(oldin->first); | |
1572 | q != snaps.end() && *q <= last; | |
1573 | ++q) { | |
1574 | in->add_need_snapflush(oldin, *q, client); | |
1575 | } | |
1576 | } else { | |
1577 | dout(10) << " ignoring client." << client << " cap follows " << cap->client_follows << dendl; | |
1578 | } | |
1579 | } | |
1580 | ||
1581 | return oldin; | |
1582 | } | |
1583 | ||
1584 | void MDCache::journal_cow_dentry(MutationImpl *mut, EMetaBlob *metablob, | |
1585 | CDentry *dn, snapid_t follows, | |
1586 | CInode **pcow_inode, CDentry::linkage_t *dnl) | |
1587 | { | |
1588 | if (!dn) { | |
1589 | dout(10) << "journal_cow_dentry got null CDentry, returning" << dendl; | |
1590 | return; | |
1591 | } | |
1592 | dout(10) << "journal_cow_dentry follows " << follows << " on " << *dn << dendl; | |
1593 | assert(dn->is_auth()); | |
1594 | ||
1595 | // nothing to cow on a null dentry, fix caller | |
1596 | if (!dnl) | |
1597 | dnl = dn->get_projected_linkage(); | |
1598 | assert(!dnl->is_null()); | |
1599 | ||
1600 | if (dnl->is_primary() && dnl->get_inode()->is_multiversion()) { | |
1601 | // multiversion inode. | |
1602 | CInode *in = dnl->get_inode(); | |
1603 | SnapRealm *realm = NULL; | |
1604 | ||
1605 | if (in->get_projected_parent_dn() != dn) { | |
1606 | assert(follows == CEPH_NOSNAP); | |
1607 | realm = dn->dir->inode->find_snaprealm(); | |
1608 | snapid_t dir_follows = realm->get_newest_snap(); | |
1609 | ||
1610 | if (dir_follows+1 > dn->first) { | |
1611 | snapid_t oldfirst = dn->first; | |
1612 | dn->first = dir_follows+1; | |
1613 | if (realm->has_snaps_in_range(oldfirst, dir_follows)) { | |
1614 | CDentry *olddn = dn->dir->add_remote_dentry(dn->name, in->ino(), in->d_type(), | |
1615 | oldfirst, dir_follows); | |
1616 | olddn->pre_dirty(); | |
1617 | dout(10) << " olddn " << *olddn << dendl; | |
1618 | metablob->add_remote_dentry(olddn, true); | |
1619 | mut->add_cow_dentry(olddn); | |
1620 | // FIXME: adjust link count here? hmm. | |
1621 | ||
1622 | if (dir_follows+1 > in->first) | |
1623 | in->cow_old_inode(dir_follows, false); | |
1624 | } | |
1625 | } | |
1626 | ||
1627 | if (in->snaprealm) { | |
1628 | realm = in->snaprealm; | |
1629 | follows = realm->get_newest_seq(); | |
1630 | } else | |
1631 | follows = dir_follows; | |
1632 | } else { | |
1633 | realm = in->find_snaprealm(); | |
1634 | if (follows == CEPH_NOSNAP) | |
1635 | follows = realm->get_newest_seq(); | |
1636 | } | |
1637 | ||
1638 | // already cloned? | |
1639 | if (follows < in->first) { | |
1640 | dout(10) << "journal_cow_dentry follows " << follows << " < first on " << *in << dendl; | |
1641 | return; | |
1642 | } | |
1643 | ||
1644 | if (!realm->has_snaps_in_range(in->first, follows)) { | |
1645 | dout(10) << "journal_cow_dentry no snapshot follows " << follows << " on " << *in << dendl; | |
1646 | in->first = follows + 1; | |
1647 | return; | |
1648 | } | |
1649 | ||
1650 | in->cow_old_inode(follows, false); | |
1651 | ||
1652 | } else { | |
1653 | SnapRealm *realm = dn->dir->inode->find_snaprealm(); | |
1654 | if (follows == CEPH_NOSNAP) | |
1655 | follows = realm->get_newest_seq(); | |
1656 | ||
1657 | // already cloned? | |
1658 | if (follows < dn->first) { | |
1659 | dout(10) << "journal_cow_dentry follows " << follows << " < first on " << *dn << dendl; | |
1660 | return; | |
1661 | } | |
1662 | ||
1663 | // update dn.first before adding old dentry to cdir's map | |
1664 | snapid_t oldfirst = dn->first; | |
1665 | dn->first = follows+1; | |
1666 | ||
1667 | CInode *in = dnl->is_primary() ? dnl->get_inode() : NULL; | |
1668 | ||
1669 | if (!realm->has_snaps_in_range(oldfirst, follows)) { | |
1670 | dout(10) << "journal_cow_dentry no snapshot follows " << follows << " on " << *dn << dendl; | |
1671 | if (in) | |
1672 | in->first = follows+1; | |
1673 | return; | |
1674 | } | |
1675 | ||
1676 | dout(10) << " dn " << *dn << dendl; | |
1677 | if (in) { | |
1678 | CInode *oldin = cow_inode(in, follows); | |
1679 | mut->add_cow_inode(oldin); | |
1680 | if (pcow_inode) | |
1681 | *pcow_inode = oldin; | |
1682 | CDentry *olddn = dn->dir->add_primary_dentry(dn->name, oldin, oldfirst, follows); | |
1683 | oldin->inode.version = olddn->pre_dirty(); | |
1684 | dout(10) << " olddn " << *olddn << dendl; | |
1685 | bool need_snapflush = !oldin->client_snap_caps.empty(); | |
1686 | if (need_snapflush) | |
1687 | mut->ls->open_files.push_back(&oldin->item_open_file); | |
1688 | metablob->add_primary_dentry(olddn, 0, true, false, false, need_snapflush); | |
1689 | mut->add_cow_dentry(olddn); | |
1690 | } else { | |
1691 | assert(dnl->is_remote()); | |
1692 | CDentry *olddn = dn->dir->add_remote_dentry(dn->name, dnl->get_remote_ino(), dnl->get_remote_d_type(), | |
1693 | oldfirst, follows); | |
1694 | olddn->pre_dirty(); | |
1695 | dout(10) << " olddn " << *olddn << dendl; | |
1696 | metablob->add_remote_dentry(olddn, true); | |
1697 | mut->add_cow_dentry(olddn); | |
1698 | } | |
1699 | } | |
1700 | } | |
1701 | ||
1702 | ||
1703 | void MDCache::journal_cow_inode(MutationRef& mut, EMetaBlob *metablob, | |
1704 | CInode *in, snapid_t follows, | |
1705 | CInode **pcow_inode) | |
1706 | { | |
1707 | dout(10) << "journal_cow_inode follows " << follows << " on " << *in << dendl; | |
1708 | CDentry *dn = in->get_projected_parent_dn(); | |
1709 | journal_cow_dentry(mut.get(), metablob, dn, follows, pcow_inode); | |
1710 | } | |
1711 | ||
1712 | void MDCache::journal_dirty_inode(MutationImpl *mut, EMetaBlob *metablob, CInode *in, snapid_t follows) | |
1713 | { | |
1714 | if (in->is_base()) { | |
1715 | metablob->add_root(true, in, in->get_projected_inode()); | |
1716 | } else { | |
1717 | if (follows == CEPH_NOSNAP && in->last != CEPH_NOSNAP) | |
1718 | follows = in->first - 1; | |
1719 | CDentry *dn = in->get_projected_parent_dn(); | |
1720 | if (!dn->get_projected_linkage()->is_null()) // no need to cow a null dentry | |
1721 | journal_cow_dentry(mut, metablob, dn, follows); | |
1722 | if (in->get_projected_inode()->is_backtrace_updated()) { | |
1723 | bool dirty_pool = in->get_projected_inode()->layout.pool_id != | |
1724 | in->get_previous_projected_inode()->layout.pool_id; | |
1725 | metablob->add_primary_dentry(dn, in, true, true, dirty_pool); | |
1726 | } else { | |
1727 | metablob->add_primary_dentry(dn, in, true); | |
1728 | } | |
1729 | } | |
1730 | } | |
1731 | ||
1732 | ||
1733 | ||
1734 | // nested --------------------------------------------------------------- | |
1735 | ||
1736 | void MDCache::project_rstat_inode_to_frag(CInode *cur, CDir *parent, snapid_t first, | |
1737 | int linkunlink, SnapRealm *prealm) | |
1738 | { | |
1739 | CDentry *parentdn = cur->get_projected_parent_dn(); | |
1740 | inode_t *curi = cur->get_projected_inode(); | |
1741 | ||
1742 | if (cur->first > first) | |
1743 | first = cur->first; | |
1744 | ||
1745 | dout(10) << "projected_rstat_inode_to_frag first " << first << " linkunlink " << linkunlink | |
1746 | << " " << *cur << dendl; | |
1747 | dout(20) << " frag head is [" << parent->first << ",head] " << dendl; | |
1748 | dout(20) << " inode update is [" << first << "," << cur->last << "]" << dendl; | |
1749 | ||
1750 | /* | |
1751 | * FIXME. this incompletely propagates rstats to _old_ parents | |
1752 | * (i.e. shortly after a directory rename). but we need full | |
1753 | * blown hard link backpointers to make this work properly... | |
1754 | */ | |
1755 | snapid_t floor = parentdn->first; | |
1756 | dout(20) << " floor of " << floor << " from parent dn " << *parentdn << dendl; | |
1757 | ||
1758 | if (!prealm) | |
1759 | prealm = parent->inode->find_snaprealm(); | |
1760 | const set<snapid_t> snaps = prealm->get_snaps(); | |
1761 | ||
1762 | if (cur->last != CEPH_NOSNAP) { | |
1763 | assert(cur->dirty_old_rstats.empty()); | |
1764 | set<snapid_t>::const_iterator q = snaps.lower_bound(MAX(first, floor)); | |
1765 | if (q == snaps.end() || *q > cur->last) | |
1766 | return; | |
1767 | } | |
1768 | ||
1769 | if (cur->last >= floor) { | |
1770 | bool update = true; | |
1771 | if (cur->state_test(CInode::STATE_AMBIGUOUSAUTH) && cur->is_auth()) { | |
1772 | // rename src inode is not projected in the slave rename prep case. so we should | |
1773 | // avoid updateing the inode. | |
1774 | assert(linkunlink < 0); | |
1775 | assert(cur->is_frozen_inode()); | |
1776 | update = false; | |
1777 | } | |
1778 | _project_rstat_inode_to_frag(*curi, MAX(first, floor), cur->last, parent, | |
1779 | linkunlink, update); | |
1780 | } | |
1781 | ||
1782 | if (g_conf->mds_snap_rstat) { | |
1783 | for (compact_set<snapid_t>::iterator p = cur->dirty_old_rstats.begin(); | |
1784 | p != cur->dirty_old_rstats.end(); | |
1785 | ++p) { | |
1786 | old_inode_t& old = cur->old_inodes[*p]; | |
1787 | snapid_t ofirst = MAX(old.first, floor); | |
1788 | set<snapid_t>::const_iterator q = snaps.lower_bound(ofirst); | |
1789 | if (q == snaps.end() || *q > *p) | |
1790 | continue; | |
1791 | if (*p >= floor) | |
1792 | _project_rstat_inode_to_frag(old.inode, ofirst, *p, parent, 0, false); | |
1793 | } | |
1794 | } | |
1795 | cur->dirty_old_rstats.clear(); | |
1796 | } | |
1797 | ||
1798 | ||
1799 | void MDCache::_project_rstat_inode_to_frag(inode_t& inode, snapid_t ofirst, snapid_t last, | |
1800 | CDir *parent, int linkunlink, bool update_inode) | |
1801 | { | |
1802 | dout(10) << "_project_rstat_inode_to_frag [" << ofirst << "," << last << "]" << dendl; | |
1803 | dout(20) << " inode rstat " << inode.rstat << dendl; | |
1804 | dout(20) << " inode accounted_rstat " << inode.accounted_rstat << dendl; | |
1805 | nest_info_t delta; | |
1806 | if (linkunlink == 0) { | |
1807 | delta.add(inode.rstat); | |
1808 | delta.sub(inode.accounted_rstat); | |
1809 | } else if (linkunlink < 0) { | |
1810 | delta.sub(inode.accounted_rstat); | |
1811 | } else { | |
1812 | delta.add(inode.rstat); | |
1813 | } | |
1814 | dout(20) << " delta " << delta << dendl; | |
1815 | ||
1816 | if (update_inode) | |
1817 | inode.accounted_rstat = inode.rstat; | |
1818 | ||
1819 | while (last >= ofirst) { | |
1820 | /* | |
1821 | * pick fnode version to update. at each iteration, we want to | |
1822 | * pick a segment ending in 'last' to update. split as necessary | |
1823 | * to make that work. then, adjust first up so that we only | |
1824 | * update one segment at a time. then loop to cover the whole | |
1825 | * [ofirst,last] interval. | |
1826 | */ | |
1827 | nest_info_t *prstat; | |
1828 | snapid_t first; | |
1829 | fnode_t *pf = parent->get_projected_fnode(); | |
1830 | if (last == CEPH_NOSNAP) { | |
1831 | if (g_conf->mds_snap_rstat) | |
1832 | first = MAX(ofirst, parent->first); | |
1833 | else | |
1834 | first = parent->first; | |
1835 | prstat = &pf->rstat; | |
1836 | dout(20) << " projecting to head [" << first << "," << last << "] " << *prstat << dendl; | |
1837 | ||
1838 | if (first > parent->first && | |
1839 | !(pf->rstat == pf->accounted_rstat)) { | |
1840 | dout(10) << " target snapped and not fully accounted, cow to dirty_old_rstat [" | |
1841 | << parent->first << "," << (first-1) << "] " | |
1842 | << " " << *prstat << "/" << pf->accounted_rstat | |
1843 | << dendl; | |
1844 | parent->dirty_old_rstat[first-1].first = parent->first; | |
1845 | parent->dirty_old_rstat[first-1].rstat = pf->rstat; | |
1846 | parent->dirty_old_rstat[first-1].accounted_rstat = pf->accounted_rstat; | |
1847 | } | |
1848 | parent->first = first; | |
1849 | } else if (!g_conf->mds_snap_rstat) { | |
1850 | // drop snapshots' rstats | |
1851 | break; | |
1852 | } else if (last >= parent->first) { | |
1853 | first = parent->first; | |
1854 | parent->dirty_old_rstat[last].first = first; | |
1855 | parent->dirty_old_rstat[last].rstat = pf->rstat; | |
1856 | parent->dirty_old_rstat[last].accounted_rstat = pf->accounted_rstat; | |
1857 | prstat = &parent->dirty_old_rstat[last].rstat; | |
1858 | dout(10) << " projecting to newly split dirty_old_fnode [" << first << "," << last << "] " | |
1859 | << " " << *prstat << "/" << pf->accounted_rstat << dendl; | |
1860 | } else { | |
1861 | // be careful, dirty_old_rstat is a _sparse_ map. | |
1862 | // sorry, this is ugly. | |
1863 | first = ofirst; | |
1864 | ||
1865 | // find any intersection with last | |
1866 | compact_map<snapid_t,old_rstat_t>::iterator p = parent->dirty_old_rstat.lower_bound(last); | |
1867 | if (p == parent->dirty_old_rstat.end()) { | |
1868 | dout(20) << " no dirty_old_rstat with last >= last " << last << dendl; | |
1869 | if (!parent->dirty_old_rstat.empty() && parent->dirty_old_rstat.rbegin()->first >= first) { | |
1870 | dout(20) << " last dirty_old_rstat ends at " << parent->dirty_old_rstat.rbegin()->first << dendl; | |
1871 | first = parent->dirty_old_rstat.rbegin()->first+1; | |
1872 | } | |
1873 | } else { | |
1874 | // *p last is >= last | |
1875 | if (p->second.first <= last) { | |
1876 | // *p intersects [first,last] | |
1877 | if (p->second.first < first) { | |
1878 | dout(10) << " splitting off left bit [" << p->second.first << "," << first-1 << "]" << dendl; | |
1879 | parent->dirty_old_rstat[first-1] = p->second; | |
1880 | p->second.first = first; | |
1881 | } | |
1882 | if (p->second.first > first) | |
1883 | first = p->second.first; | |
1884 | if (last < p->first) { | |
1885 | dout(10) << " splitting off right bit [" << last+1 << "," << p->first << "]" << dendl; | |
1886 | parent->dirty_old_rstat[last] = p->second; | |
1887 | p->second.first = last+1; | |
1888 | } | |
1889 | } else { | |
1890 | // *p is to the _right_ of [first,last] | |
1891 | p = parent->dirty_old_rstat.lower_bound(first); | |
1892 | // new *p last is >= first | |
1893 | if (p->second.first <= last && // new *p isn't also to the right, and | |
1894 | p->first >= first) { // it intersects our first bit, | |
1895 | dout(10) << " staying to the right of [" << p->second.first << "," << p->first << "]..." << dendl; | |
1896 | first = p->first+1; | |
1897 | } | |
1898 | dout(10) << " projecting to new dirty_old_rstat [" << first << "," << last << "]" << dendl; | |
1899 | } | |
1900 | } | |
1901 | dout(20) << " projecting to dirty_old_rstat [" << first << "," << last << "]" << dendl; | |
1902 | parent->dirty_old_rstat[last].first = first; | |
1903 | prstat = &parent->dirty_old_rstat[last].rstat; | |
1904 | } | |
1905 | ||
1906 | // apply | |
1907 | dout(20) << " project to [" << first << "," << last << "] " << *prstat << dendl; | |
1908 | assert(last >= first); | |
1909 | prstat->add(delta); | |
1910 | if (update_inode) | |
1911 | inode.accounted_rstat = inode.rstat; | |
1912 | dout(20) << " result [" << first << "," << last << "] " << *prstat << " " << *parent << dendl; | |
1913 | ||
1914 | last = first-1; | |
1915 | } | |
1916 | } | |
1917 | ||
1918 | void MDCache::project_rstat_frag_to_inode(nest_info_t& rstat, nest_info_t& accounted_rstat, | |
1919 | snapid_t ofirst, snapid_t last, | |
1920 | CInode *pin, bool cow_head) | |
1921 | { | |
1922 | dout(10) << "project_rstat_frag_to_inode [" << ofirst << "," << last << "]" << dendl; | |
1923 | dout(20) << " frag rstat " << rstat << dendl; | |
1924 | dout(20) << " frag accounted_rstat " << accounted_rstat << dendl; | |
1925 | nest_info_t delta = rstat; | |
1926 | delta.sub(accounted_rstat); | |
1927 | dout(20) << " delta " << delta << dendl; | |
1928 | ||
1929 | while (last >= ofirst) { | |
1930 | inode_t *pi; | |
1931 | snapid_t first; | |
1932 | if (last == pin->last) { | |
1933 | pi = pin->get_projected_inode(); | |
1934 | first = MAX(ofirst, pin->first); | |
1935 | if (first > pin->first) { | |
1936 | old_inode_t& old = pin->cow_old_inode(first-1, cow_head); | |
1937 | dout(20) << " cloned old_inode rstat is " << old.inode.rstat << dendl; | |
1938 | } | |
1939 | } else { | |
1940 | if (last >= pin->first) { | |
1941 | first = pin->first; | |
1942 | pin->cow_old_inode(last, cow_head); | |
1943 | } else { | |
1944 | // our life is easier here because old_inodes is not sparse | |
1945 | // (although it may not begin at snapid 1) | |
1946 | compact_map<snapid_t,old_inode_t>::iterator p = pin->old_inodes.lower_bound(last); | |
1947 | if (p == pin->old_inodes.end()) { | |
1948 | dout(10) << " no old_inode <= " << last << ", done." << dendl; | |
1949 | break; | |
1950 | } | |
1951 | first = p->second.first; | |
1952 | if (first > last) { | |
1953 | dout(10) << " oldest old_inode is [" << first << "," << p->first << "], done." << dendl; | |
1954 | //assert(p == pin->old_inodes.begin()); | |
1955 | break; | |
1956 | } | |
1957 | if (p->first > last) { | |
1958 | dout(10) << " splitting right old_inode [" << first << "," << p->first << "] to [" | |
1959 | << (last+1) << "," << p->first << "]" << dendl; | |
1960 | pin->old_inodes[last] = p->second; | |
1961 | p->second.first = last+1; | |
1962 | pin->dirty_old_rstats.insert(p->first); | |
1963 | } | |
1964 | } | |
1965 | if (first < ofirst) { | |
1966 | dout(10) << " splitting left old_inode [" << first << "," << last << "] to [" | |
1967 | << first << "," << ofirst-1 << "]" << dendl; | |
1968 | pin->old_inodes[ofirst-1] = pin->old_inodes[last]; | |
1969 | pin->dirty_old_rstats.insert(ofirst-1); | |
1970 | pin->old_inodes[last].first = first = ofirst; | |
1971 | } | |
1972 | pi = &pin->old_inodes[last].inode; | |
1973 | pin->dirty_old_rstats.insert(last); | |
1974 | } | |
1975 | dout(20) << " projecting to [" << first << "," << last << "] " << pi->rstat << dendl; | |
1976 | pi->rstat.add(delta); | |
1977 | dout(20) << " result [" << first << "," << last << "] " << pi->rstat << dendl; | |
1978 | ||
1979 | last = first-1; | |
1980 | } | |
1981 | } | |
1982 | ||
1983 | void MDCache::broadcast_quota_to_client(CInode *in) | |
1984 | { | |
1985 | if (!in->is_auth() || in->is_frozen()) | |
1986 | return; | |
1987 | ||
1988 | inode_t *i = in->get_projected_inode(); | |
1989 | ||
1990 | if (!i->quota.is_enable()) | |
1991 | return; | |
1992 | ||
1993 | for (map<client_t,Capability*>::iterator it = in->client_caps.begin(); | |
1994 | it != in->client_caps.end(); | |
1995 | ++it) { | |
1996 | Session *session = mds->get_session(it->first); | |
1997 | if (!session || !session->connection || | |
1998 | !session->connection->has_feature(CEPH_FEATURE_MDS_QUOTA)) | |
1999 | continue; | |
2000 | ||
2001 | Capability *cap = it->second; | |
2002 | if (cap->last_rbytes == i->rstat.rbytes && | |
2003 | cap->last_rsize == i->rstat.rsize()) | |
2004 | continue; | |
2005 | ||
2006 | if (i->quota.max_files > 0) { | |
2007 | if (i->rstat.rsize() >= i->quota.max_files) | |
2008 | goto update; | |
2009 | ||
2010 | if ((abs(cap->last_rsize - i->quota.max_files) >> 4) < | |
2011 | abs(cap->last_rsize - i->rstat.rsize())) | |
2012 | goto update; | |
2013 | } | |
2014 | ||
2015 | if (i->quota.max_bytes > 0) { | |
2016 | if (i->rstat.rbytes > i->quota.max_bytes - (i->quota.max_bytes >> 3)) | |
2017 | goto update; | |
2018 | ||
2019 | if ((abs(cap->last_rbytes - i->quota.max_bytes) >> 4) < | |
2020 | abs(cap->last_rbytes - i->rstat.rbytes)) | |
2021 | goto update; | |
2022 | } | |
2023 | ||
2024 | continue; | |
2025 | ||
2026 | update: | |
2027 | cap->last_rsize = i->rstat.rsize(); | |
2028 | cap->last_rbytes = i->rstat.rbytes; | |
2029 | ||
2030 | MClientQuota *msg = new MClientQuota(); | |
2031 | msg->ino = in->ino(); | |
2032 | msg->rstat = i->rstat; | |
2033 | msg->quota = i->quota; | |
2034 | mds->send_message_client_counted(msg, session->connection); | |
2035 | } | |
181888fb | 2036 | for (const auto &it : in->get_replicas()) { |
7c673cae FG |
2037 | MGatherCaps *msg = new MGatherCaps; |
2038 | msg->ino = in->ino(); | |
181888fb | 2039 | mds->send_message_mds(msg, it.first); |
7c673cae FG |
2040 | } |
2041 | } | |
2042 | ||
2043 | /* | |
2044 | * NOTE: we _have_ to delay the scatter if we are called during a | |
2045 | * rejoin, because we can't twiddle locks between when the | |
2046 | * rejoin_(weak|strong) is received and when we send the rejoin_ack. | |
2047 | * normally, this isn't a problem: a recover mds doesn't twiddle locks | |
2048 | * (no requests), and a survivor acks immediately. _except_ that | |
2049 | * during rejoin_(weak|strong) processing, we may complete a lock | |
2050 | * gather, and do a scatter_writebehind.. and we _can't_ twiddle the | |
2051 | * scatterlock state in that case or the lock states will get out of | |
2052 | * sync between the auth and replica. | |
2053 | * | |
2054 | * the simple solution is to never do the scatter here. instead, put | |
2055 | * the scatterlock on a list if it isn't already wrlockable. this is | |
2056 | * probably the best plan anyway, since we avoid too many | |
2057 | * scatters/locks under normal usage. | |
2058 | */ | |
2059 | /* | |
2060 | * some notes on dirlock/nestlock scatterlock semantics: | |
2061 | * | |
2062 | * the fragstat (dirlock) will never be updated without | |
2063 | * dirlock+nestlock wrlock held by the caller. | |
2064 | * | |
2065 | * the rstat (nestlock) _may_ get updated without a wrlock when nested | |
2066 | * data is pushed up the tree. this could be changed with some | |
2067 | * restructuring here, but in its current form we ensure that the | |
2068 | * fragstat+rstat _always_ reflect an accurrate summation over the dir | |
2069 | * frag, which is nice. and, we only need to track frags that need to | |
2070 | * be nudged (and not inodes with pending rstat changes that need to | |
2071 | * be pushed into the frag). a consequence of this is that the | |
2072 | * accounted_rstat on scatterlock sync may not match our current | |
2073 | * rstat. this is normal and expected. | |
2074 | */ | |
2075 | void MDCache::predirty_journal_parents(MutationRef mut, EMetaBlob *blob, | |
2076 | CInode *in, CDir *parent, | |
2077 | int flags, int linkunlink, | |
2078 | snapid_t cfollows) | |
2079 | { | |
2080 | bool primary_dn = flags & PREDIRTY_PRIMARY; | |
2081 | bool do_parent_mtime = flags & PREDIRTY_DIR; | |
2082 | bool shallow = flags & PREDIRTY_SHALLOW; | |
2083 | ||
2084 | assert(mds->mdlog->entry_is_open()); | |
2085 | ||
2086 | // make sure stamp is set | |
2087 | if (mut->get_mds_stamp() == utime_t()) | |
2088 | mut->set_mds_stamp(ceph_clock_now()); | |
2089 | ||
2090 | if (in->is_base()) | |
2091 | return; | |
2092 | ||
2093 | dout(10) << "predirty_journal_parents" | |
2094 | << (do_parent_mtime ? " do_parent_mtime":"") | |
2095 | << " linkunlink=" << linkunlink | |
2096 | << (primary_dn ? " primary_dn":" remote_dn") | |
2097 | << (shallow ? " SHALLOW":"") | |
2098 | << " follows " << cfollows | |
2099 | << " " << *in << dendl; | |
2100 | ||
2101 | if (!parent) { | |
2102 | assert(primary_dn); | |
2103 | parent = in->get_projected_parent_dn()->get_dir(); | |
2104 | } | |
2105 | ||
2106 | if (flags == 0 && linkunlink == 0) { | |
2107 | dout(10) << " no flags/linkunlink, just adding dir context to blob(s)" << dendl; | |
2108 | blob->add_dir_context(parent); | |
2109 | return; | |
2110 | } | |
2111 | ||
2112 | // build list of inodes to wrlock, dirty, and update | |
2113 | list<CInode*> lsi; | |
2114 | CInode *cur = in; | |
2115 | CDentry *parentdn = NULL; | |
2116 | bool first = true; | |
2117 | while (parent) { | |
2118 | //assert(cur->is_auth() || !primary_dn); // this breaks the rename auth twiddle hack | |
2119 | assert(parent->is_auth()); | |
2120 | ||
2121 | // opportunistically adjust parent dirfrag | |
2122 | CInode *pin = parent->get_inode(); | |
2123 | ||
2124 | // inode -> dirfrag | |
2125 | mut->auth_pin(parent); | |
2126 | mut->add_projected_fnode(parent); | |
2127 | ||
2128 | fnode_t *pf = parent->project_fnode(); | |
2129 | pf->version = parent->pre_dirty(); | |
2130 | ||
2131 | if (do_parent_mtime || linkunlink) { | |
2132 | assert(mut->wrlocks.count(&pin->filelock)); | |
2133 | assert(mut->wrlocks.count(&pin->nestlock)); | |
2134 | assert(cfollows == CEPH_NOSNAP); | |
2135 | ||
2136 | // update stale fragstat/rstat? | |
2137 | parent->resync_accounted_fragstat(); | |
2138 | parent->resync_accounted_rstat(); | |
2139 | ||
2140 | if (do_parent_mtime) { | |
2141 | pf->fragstat.mtime = mut->get_op_stamp(); | |
2142 | pf->fragstat.change_attr++; | |
2143 | dout(10) << "predirty_journal_parents bumping change_attr to " << pf->fragstat.change_attr << " on " << parent << dendl; | |
2144 | if (pf->fragstat.mtime > pf->rstat.rctime) { | |
2145 | dout(10) << "predirty_journal_parents updating mtime on " << *parent << dendl; | |
2146 | pf->rstat.rctime = pf->fragstat.mtime; | |
2147 | } else { | |
2148 | dout(10) << "predirty_journal_parents updating mtime UNDERWATER on " << *parent << dendl; | |
2149 | } | |
2150 | } | |
2151 | if (linkunlink) { | |
2152 | dout(10) << "predirty_journal_parents updating size on " << *parent << dendl; | |
2153 | if (in->is_dir()) { | |
2154 | pf->fragstat.nsubdirs += linkunlink; | |
2155 | //pf->rstat.rsubdirs += linkunlink; | |
2156 | } else { | |
2157 | pf->fragstat.nfiles += linkunlink; | |
2158 | //pf->rstat.rfiles += linkunlink; | |
2159 | } | |
2160 | } | |
2161 | } | |
2162 | ||
2163 | // rstat | |
2164 | if (!primary_dn) { | |
2165 | // don't update parent this pass | |
2166 | } else if (!linkunlink && !(pin->nestlock.can_wrlock(-1) && | |
2167 | pin->versionlock.can_wrlock())) { | |
2168 | dout(20) << " unwritable parent nestlock " << pin->nestlock | |
2169 | << ", marking dirty rstat on " << *cur << dendl; | |
2170 | cur->mark_dirty_rstat(); | |
2171 | } else { | |
2172 | // if we don't hold a wrlock reference on this nestlock, take one, | |
2173 | // because we are about to write into the dirfrag fnode and that needs | |
2174 | // to commit before the lock can cycle. | |
2175 | if (linkunlink) { | |
2176 | assert(pin->nestlock.get_num_wrlocks() || mut->is_slave()); | |
2177 | } | |
2178 | ||
2179 | if (mut->wrlocks.count(&pin->nestlock) == 0) { | |
2180 | dout(10) << " taking wrlock on " << pin->nestlock << " on " << *pin << dendl; | |
2181 | mds->locker->wrlock_force(&pin->nestlock, mut); | |
2182 | } | |
2183 | ||
2184 | // now we can project the inode rstat diff the dirfrag | |
2185 | SnapRealm *prealm = pin->find_snaprealm(); | |
2186 | ||
2187 | snapid_t follows = cfollows; | |
2188 | if (follows == CEPH_NOSNAP) | |
2189 | follows = prealm->get_newest_seq(); | |
2190 | ||
2191 | snapid_t first = follows+1; | |
2192 | ||
2193 | // first, if the frag is stale, bring it back in sync. | |
2194 | parent->resync_accounted_rstat(); | |
2195 | ||
2196 | // now push inode rstats into frag | |
2197 | project_rstat_inode_to_frag(cur, parent, first, linkunlink, prealm); | |
2198 | cur->clear_dirty_rstat(); | |
2199 | } | |
2200 | ||
2201 | bool stop = false; | |
2202 | if (!pin->is_auth() || (!mut->is_auth_pinned(pin) && !pin->can_auth_pin())) { | |
2203 | dout(10) << "predirty_journal_parents !auth or ambig or can't authpin on " << *pin << dendl; | |
2204 | stop = true; | |
2205 | } | |
2206 | ||
2207 | // delay propagating until later? | |
2208 | if (!stop && !first && | |
2209 | g_conf->mds_dirstat_min_interval > 0) { | |
2210 | double since_last_prop = mut->get_mds_stamp() - pin->last_dirstat_prop; | |
2211 | if (since_last_prop < g_conf->mds_dirstat_min_interval) { | |
2212 | dout(10) << "predirty_journal_parents last prop " << since_last_prop | |
2213 | << " < " << g_conf->mds_dirstat_min_interval | |
2214 | << ", stopping" << dendl; | |
2215 | stop = true; | |
2216 | } else { | |
2217 | dout(10) << "predirty_journal_parents last prop " << since_last_prop << " ago, continuing" << dendl; | |
2218 | } | |
2219 | } | |
2220 | ||
2221 | // can cast only because i'm passing nowait=true in the sole user | |
2222 | MDRequestRef mdmut = static_cast<MDRequestImpl*>(mut.get()); | |
2223 | if (!stop && | |
2224 | mut->wrlocks.count(&pin->nestlock) == 0 && | |
2225 | (!pin->versionlock.can_wrlock() || // make sure we can take versionlock, too | |
2226 | //true | |
2227 | !mds->locker->wrlock_start(&pin->nestlock, mdmut, true) | |
2228 | )) { // ** do not initiate.. see above comment ** | |
2229 | dout(10) << "predirty_journal_parents can't wrlock one of " << pin->versionlock << " or " << pin->nestlock | |
2230 | << " on " << *pin << dendl; | |
2231 | stop = true; | |
2232 | } | |
2233 | if (stop) { | |
2234 | dout(10) << "predirty_journal_parents stop. marking nestlock on " << *pin << dendl; | |
2235 | mds->locker->mark_updated_scatterlock(&pin->nestlock); | |
2236 | mut->ls->dirty_dirfrag_nest.push_back(&pin->item_dirty_dirfrag_nest); | |
2237 | mut->add_updated_lock(&pin->nestlock); | |
2238 | if (do_parent_mtime || linkunlink) { | |
2239 | mds->locker->mark_updated_scatterlock(&pin->filelock); | |
2240 | mut->ls->dirty_dirfrag_dir.push_back(&pin->item_dirty_dirfrag_dir); | |
2241 | mut->add_updated_lock(&pin->filelock); | |
2242 | } | |
2243 | break; | |
2244 | } | |
2245 | if (!mut->wrlocks.count(&pin->versionlock)) | |
2246 | mds->locker->local_wrlock_grab(&pin->versionlock, mut); | |
2247 | ||
2248 | assert(mut->wrlocks.count(&pin->nestlock) || | |
2249 | mut->is_slave()); | |
2250 | ||
2251 | pin->last_dirstat_prop = mut->get_mds_stamp(); | |
2252 | ||
2253 | // dirfrag -> diri | |
2254 | mut->auth_pin(pin); | |
2255 | mut->add_projected_inode(pin); | |
2256 | lsi.push_front(pin); | |
2257 | ||
2258 | pin->pre_cow_old_inode(); // avoid cow mayhem! | |
2259 | ||
2260 | inode_t *pi = pin->project_inode(); | |
2261 | pi->version = pin->pre_dirty(); | |
2262 | ||
2263 | // dirstat | |
2264 | if (do_parent_mtime || linkunlink) { | |
2265 | dout(20) << "predirty_journal_parents add_delta " << pf->fragstat << dendl; | |
2266 | dout(20) << "predirty_journal_parents - " << pf->accounted_fragstat << dendl; | |
2267 | bool touched_mtime = false, touched_chattr = false; | |
2268 | pi->dirstat.add_delta(pf->fragstat, pf->accounted_fragstat, &touched_mtime, &touched_chattr); | |
2269 | pf->accounted_fragstat = pf->fragstat; | |
2270 | if (touched_mtime) | |
2271 | pi->mtime = pi->ctime = pi->dirstat.mtime; | |
2272 | if (touched_chattr) | |
2273 | pi->change_attr = pi->dirstat.change_attr; | |
2274 | dout(20) << "predirty_journal_parents gives " << pi->dirstat << " on " << *pin << dendl; | |
2275 | ||
2276 | if (parent->get_frag() == frag_t()) { // i.e., we are the only frag | |
2277 | if (pi->dirstat.size() < 0) | |
2278 | assert(!"negative dirstat size" == g_conf->mds_verify_scatter); | |
2279 | if (pi->dirstat.size() != pf->fragstat.size()) { | |
2280 | mds->clog->error() << "unmatched fragstat size on single dirfrag " | |
2281 | << parent->dirfrag() << ", inode has " << pi->dirstat | |
2282 | << ", dirfrag has " << pf->fragstat; | |
2283 | ||
2284 | // trust the dirfrag for now | |
2285 | pi->dirstat = pf->fragstat; | |
2286 | ||
2287 | assert(!"unmatched fragstat size" == g_conf->mds_verify_scatter); | |
2288 | } | |
2289 | } | |
2290 | } | |
2291 | ||
2292 | /* | |
2293 | * the rule here is to follow the _oldest_ parent with dirty rstat | |
2294 | * data. if we don't propagate all data, we add ourselves to the | |
2295 | * nudge list. that way all rstat data will (eventually) get | |
2296 | * pushed up the tree. | |
2297 | * | |
2298 | * actually, no. for now, silently drop rstats for old parents. we need | |
2299 | * hard link backpointers to do the above properly. | |
2300 | */ | |
2301 | ||
2302 | // stop? | |
2303 | if (pin->is_base()) | |
2304 | break; | |
2305 | parentdn = pin->get_projected_parent_dn(); | |
2306 | assert(parentdn); | |
2307 | ||
2308 | // rstat | |
2309 | dout(10) << "predirty_journal_parents frag->inode on " << *parent << dendl; | |
2310 | ||
2311 | // first, if the frag is stale, bring it back in sync. | |
2312 | parent->resync_accounted_rstat(); | |
2313 | ||
2314 | if (g_conf->mds_snap_rstat) { | |
2315 | for (compact_map<snapid_t,old_rstat_t>::iterator p = parent->dirty_old_rstat.begin(); | |
2316 | p != parent->dirty_old_rstat.end(); | |
2317 | ++p) | |
2318 | project_rstat_frag_to_inode(p->second.rstat, p->second.accounted_rstat, p->second.first, | |
2319 | p->first, pin, true);//false); | |
2320 | } | |
2321 | parent->dirty_old_rstat.clear(); | |
2322 | project_rstat_frag_to_inode(pf->rstat, pf->accounted_rstat, parent->first, CEPH_NOSNAP, pin, true);//false); | |
2323 | ||
2324 | pf->accounted_rstat = pf->rstat; | |
2325 | ||
2326 | if (parent->get_frag() == frag_t()) { // i.e., we are the only frag | |
2327 | if (pi->rstat.rbytes != pf->rstat.rbytes) { | |
2328 | mds->clog->error() << "unmatched rstat rbytes on single dirfrag " | |
2329 | << parent->dirfrag() << ", inode has " << pi->rstat | |
2330 | << ", dirfrag has " << pf->rstat; | |
2331 | ||
2332 | // trust the dirfrag for now | |
2333 | pi->rstat = pf->rstat; | |
2334 | ||
2335 | assert(!"unmatched rstat rbytes" == g_conf->mds_verify_scatter); | |
2336 | } | |
2337 | } | |
2338 | ||
2339 | parent->check_rstats(); | |
2340 | broadcast_quota_to_client(pin); | |
2341 | // next parent! | |
2342 | cur = pin; | |
2343 | parent = parentdn->get_dir(); | |
2344 | linkunlink = 0; | |
2345 | do_parent_mtime = false; | |
2346 | primary_dn = true; | |
2347 | first = false; | |
2348 | } | |
2349 | ||
2350 | // now, stick it in the blob | |
2351 | assert(parent); | |
2352 | assert(parent->is_auth()); | |
2353 | blob->add_dir_context(parent); | |
2354 | blob->add_dir(parent, true); | |
2355 | for (list<CInode*>::iterator p = lsi.begin(); | |
2356 | p != lsi.end(); | |
2357 | ++p) { | |
2358 | CInode *cur = *p; | |
2359 | journal_dirty_inode(mut.get(), blob, cur); | |
2360 | } | |
2361 | ||
2362 | } | |
2363 | ||
2364 | ||
2365 | ||
2366 | ||
2367 | ||
2368 | // =================================== | |
2369 | // slave requests | |
2370 | ||
2371 | ||
2372 | /* | |
2373 | * some handlers for master requests with slaves. we need to make | |
2374 | * sure slaves journal commits before we forget we mastered them and | |
2375 | * remove them from the uncommitted_masters map (used during recovery | |
2376 | * to commit|abort slaves). | |
2377 | */ | |
2378 | struct C_MDC_CommittedMaster : public MDCacheLogContext { | |
2379 | metareqid_t reqid; | |
2380 | C_MDC_CommittedMaster(MDCache *s, metareqid_t r) : MDCacheLogContext(s), reqid(r) {} | |
2381 | void finish(int r) override { | |
2382 | mdcache->_logged_master_commit(reqid); | |
2383 | } | |
2384 | }; | |
2385 | ||
2386 | void MDCache::log_master_commit(metareqid_t reqid) | |
2387 | { | |
2388 | dout(10) << "log_master_commit " << reqid << dendl; | |
2389 | uncommitted_masters[reqid].committing = true; | |
2390 | mds->mdlog->start_submit_entry(new ECommitted(reqid), | |
2391 | new C_MDC_CommittedMaster(this, reqid)); | |
2392 | } | |
2393 | ||
2394 | void MDCache::_logged_master_commit(metareqid_t reqid) | |
2395 | { | |
2396 | dout(10) << "_logged_master_commit " << reqid << dendl; | |
2397 | assert(uncommitted_masters.count(reqid)); | |
2398 | uncommitted_masters[reqid].ls->uncommitted_masters.erase(reqid); | |
2399 | mds->queue_waiters(uncommitted_masters[reqid].waiters); | |
2400 | uncommitted_masters.erase(reqid); | |
2401 | } | |
2402 | ||
2403 | // while active... | |
2404 | ||
2405 | void MDCache::committed_master_slave(metareqid_t r, mds_rank_t from) | |
2406 | { | |
2407 | dout(10) << "committed_master_slave mds." << from << " on " << r << dendl; | |
2408 | assert(uncommitted_masters.count(r)); | |
2409 | uncommitted_masters[r].slaves.erase(from); | |
2410 | if (!uncommitted_masters[r].recovering && uncommitted_masters[r].slaves.empty()) | |
2411 | log_master_commit(r); | |
2412 | } | |
2413 | ||
2414 | void MDCache::logged_master_update(metareqid_t reqid) | |
2415 | { | |
2416 | dout(10) << "logged_master_update " << reqid << dendl; | |
2417 | assert(uncommitted_masters.count(reqid)); | |
2418 | uncommitted_masters[reqid].safe = true; | |
2419 | if (pending_masters.count(reqid)) { | |
2420 | pending_masters.erase(reqid); | |
2421 | if (pending_masters.empty()) | |
2422 | process_delayed_resolve(); | |
2423 | } | |
2424 | } | |
2425 | ||
2426 | /* | |
2427 | * Master may crash after receiving all slaves' commit acks, but before journalling | |
2428 | * the final commit. Slaves may crash after journalling the slave commit, but before | |
2429 | * sending commit ack to the master. Commit masters with no uncommitted slave when | |
2430 | * resolve finishes. | |
2431 | */ | |
2432 | void MDCache::finish_committed_masters() | |
2433 | { | |
2434 | for (map<metareqid_t, umaster>::iterator p = uncommitted_masters.begin(); | |
2435 | p != uncommitted_masters.end(); | |
2436 | ++p) { | |
2437 | p->second.recovering = false; | |
2438 | if (!p->second.committing && p->second.slaves.empty()) { | |
2439 | dout(10) << "finish_committed_masters " << p->first << dendl; | |
2440 | log_master_commit(p->first); | |
2441 | } | |
2442 | } | |
2443 | } | |
2444 | ||
2445 | /* | |
2446 | * at end of resolve... we must journal a commit|abort for all slave | |
2447 | * updates, before moving on. | |
2448 | * | |
2449 | * this is so that the master can safely journal ECommitted on ops it | |
2450 | * masters when it reaches up:active (all other recovering nodes must | |
2451 | * complete resolve before that happens). | |
2452 | */ | |
2453 | struct C_MDC_SlaveCommit : public MDCacheLogContext { | |
2454 | mds_rank_t from; | |
2455 | metareqid_t reqid; | |
2456 | C_MDC_SlaveCommit(MDCache *c, int f, metareqid_t r) : MDCacheLogContext(c), from(f), reqid(r) {} | |
2457 | void finish(int r) override { | |
2458 | mdcache->_logged_slave_commit(from, reqid); | |
2459 | } | |
2460 | }; | |
2461 | ||
2462 | void MDCache::_logged_slave_commit(mds_rank_t from, metareqid_t reqid) | |
2463 | { | |
2464 | dout(10) << "_logged_slave_commit from mds." << from << " " << reqid << dendl; | |
2465 | ||
2466 | // send a message | |
2467 | MMDSSlaveRequest *req = new MMDSSlaveRequest(reqid, 0, MMDSSlaveRequest::OP_COMMITTED); | |
2468 | mds->send_message_mds(req, from); | |
2469 | } | |
2470 | ||
2471 | ||
2472 | ||
2473 | ||
2474 | ||
2475 | ||
2476 | // ==================================================================== | |
2477 | // import map, recovery | |
2478 | ||
2479 | void MDCache::_move_subtree_map_bound(dirfrag_t df, dirfrag_t oldparent, dirfrag_t newparent, | |
2480 | map<dirfrag_t,vector<dirfrag_t> >& subtrees) | |
2481 | { | |
2482 | if (subtrees.count(oldparent)) { | |
2483 | vector<dirfrag_t>& v = subtrees[oldparent]; | |
2484 | dout(10) << " removing " << df << " from " << oldparent << " bounds " << v << dendl; | |
2485 | for (vector<dirfrag_t>::iterator it = v.begin(); it != v.end(); ++it) | |
2486 | if (*it == df) { | |
2487 | v.erase(it); | |
2488 | break; | |
2489 | } | |
2490 | } | |
2491 | if (subtrees.count(newparent)) { | |
2492 | vector<dirfrag_t>& v = subtrees[newparent]; | |
2493 | dout(10) << " adding " << df << " to " << newparent << " bounds " << v << dendl; | |
2494 | v.push_back(df); | |
2495 | } | |
2496 | } | |
2497 | ||
2498 | ESubtreeMap *MDCache::create_subtree_map() | |
2499 | { | |
2500 | dout(10) << "create_subtree_map " << num_subtrees() << " subtrees, " | |
2501 | << num_subtrees_fullauth() << " fullauth" | |
2502 | << dendl; | |
2503 | ||
2504 | show_subtrees(); | |
2505 | ||
2506 | ESubtreeMap *le = new ESubtreeMap(); | |
2507 | mds->mdlog->_start_entry(le); | |
2508 | ||
2509 | map<dirfrag_t, CDir*> dirs_to_add; | |
2510 | ||
2511 | if (myin) { | |
2512 | CDir* mydir = myin->get_dirfrag(frag_t()); | |
2513 | dirs_to_add[mydir->dirfrag()] = mydir; | |
2514 | } | |
2515 | ||
2516 | // include all auth subtrees, and their bounds. | |
2517 | // and a spanning tree to tie it to the root. | |
2518 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
2519 | p != subtrees.end(); | |
2520 | ++p) { | |
2521 | CDir *dir = p->first; | |
2522 | ||
2523 | // journal subtree as "ours" if we are | |
2524 | // me, -2 | |
2525 | // me, me | |
2526 | // me, !me (may be importing and ambiguous!) | |
2527 | ||
2528 | // so not | |
2529 | // !me, * | |
2530 | if (dir->get_dir_auth().first != mds->get_nodeid()) | |
2531 | continue; | |
2532 | ||
2533 | if (migrator->is_ambiguous_import(dir->dirfrag()) || | |
2534 | my_ambiguous_imports.count(dir->dirfrag())) { | |
2535 | dout(15) << " ambig subtree " << *dir << dendl; | |
2536 | le->ambiguous_subtrees.insert(dir->dirfrag()); | |
2537 | } else { | |
2538 | dout(15) << " subtree " << *dir << dendl; | |
2539 | } | |
2540 | ||
2541 | dirs_to_add[dir->dirfrag()] = dir; | |
2542 | le->subtrees[dir->dirfrag()].clear(); | |
2543 | ||
2544 | ||
2545 | // bounds | |
2546 | for (set<CDir*>::iterator q = p->second.begin(); | |
2547 | q != p->second.end(); | |
2548 | ++q) { | |
2549 | CDir *bound = *q; | |
2550 | dout(15) << " subtree bound " << *bound << dendl; | |
2551 | dirs_to_add[bound->dirfrag()] = bound; | |
2552 | le->subtrees[dir->dirfrag()].push_back(bound->dirfrag()); | |
2553 | } | |
2554 | } | |
2555 | ||
2556 | // apply projected renames | |
2557 | for (map<CInode*,list<pair<CDir*,CDir*> > >::iterator p = projected_subtree_renames.begin(); | |
2558 | p != projected_subtree_renames.end(); | |
2559 | ++p) { | |
2560 | for (list<pair<CDir*,CDir*> >::iterator q = p->second.begin(); q != p->second.end(); ++q) { | |
2561 | CInode *diri = p->first; | |
2562 | CDir *olddir = q->first; | |
2563 | CDir *newdir = q->second; | |
2564 | dout(10) << " adjusting for projected rename of " << *diri << " to " << *newdir << dendl; | |
2565 | ||
2566 | list<CDir*> dfls; | |
2567 | diri->get_dirfrags(dfls); | |
2568 | for (list<CDir*>::iterator p = dfls.begin(); p != dfls.end(); ++p) { | |
2569 | CDir *dir = *p; | |
2570 | dout(10) << "dirfrag " << dir->dirfrag() << " " << *dir << dendl; | |
2571 | CDir *oldparent = get_projected_subtree_root(olddir); | |
2572 | dout(10) << " old parent " << oldparent->dirfrag() << " " << *oldparent << dendl; | |
2573 | CDir *newparent = get_projected_subtree_root(newdir); | |
2574 | dout(10) << " new parent " << newparent->dirfrag() << " " << *newparent << dendl; | |
2575 | ||
2576 | if (oldparent == newparent) { | |
2577 | dout(10) << "parent unchanged for " << dir->dirfrag() << " at " | |
2578 | << oldparent->dirfrag() << dendl; | |
2579 | continue; | |
2580 | } | |
2581 | ||
2582 | if (dir->is_subtree_root()) { | |
2583 | if (le->subtrees.count(newparent->dirfrag()) && | |
2584 | oldparent->get_dir_auth() != newparent->get_dir_auth()) | |
2585 | dirs_to_add[dir->dirfrag()] = dir; | |
2586 | // children are fine. change parent. | |
2587 | _move_subtree_map_bound(dir->dirfrag(), oldparent->dirfrag(), newparent->dirfrag(), | |
2588 | le->subtrees); | |
2589 | } else { | |
2590 | // mid-subtree. | |
2591 | ||
2592 | if (oldparent->get_dir_auth() != newparent->get_dir_auth()) { | |
2593 | dout(10) << " creating subtree for " << dir->dirfrag() << dendl; | |
2594 | // if oldparent is auth, subtree is mine; include it. | |
2595 | if (le->subtrees.count(oldparent->dirfrag())) { | |
2596 | dirs_to_add[dir->dirfrag()] = dir; | |
2597 | le->subtrees[dir->dirfrag()].clear(); | |
2598 | } | |
2599 | // if newparent is auth, subtree is a new bound | |
2600 | if (le->subtrees.count(newparent->dirfrag())) { | |
2601 | dirs_to_add[dir->dirfrag()] = dir; | |
2602 | le->subtrees[newparent->dirfrag()].push_back(dir->dirfrag()); // newparent is auth; new bound | |
2603 | } | |
2604 | newparent = dir; | |
2605 | } | |
2606 | ||
2607 | // see if any old bounds move to the new parent. | |
2608 | for (set<CDir*>::iterator p = subtrees[oldparent].begin(); | |
2609 | p != subtrees[oldparent].end(); | |
2610 | ++p) { | |
2611 | CDir *bound = *p; | |
2612 | if (dir->contains(bound->get_parent_dir())) | |
2613 | _move_subtree_map_bound(bound->dirfrag(), oldparent->dirfrag(), newparent->dirfrag(), | |
2614 | le->subtrees); | |
2615 | } | |
2616 | } | |
2617 | } | |
2618 | } | |
2619 | } | |
2620 | ||
2621 | // simplify the journaled map. our in memory map may have more | |
2622 | // subtrees than needed due to migrations that are just getting | |
2623 | // started or just completing. but on replay, the "live" map will | |
2624 | // be simple and we can do a straight comparison. | |
2625 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator p = le->subtrees.begin(); p != le->subtrees.end(); ++p) { | |
2626 | if (le->ambiguous_subtrees.count(p->first)) | |
2627 | continue; | |
2628 | unsigned i = 0; | |
2629 | while (i < p->second.size()) { | |
2630 | dirfrag_t b = p->second[i]; | |
2631 | if (le->subtrees.count(b) && | |
2632 | le->ambiguous_subtrees.count(b) == 0) { | |
2633 | vector<dirfrag_t>& bb = le->subtrees[b]; | |
2634 | dout(10) << "simplify: " << p->first << " swallowing " << b << " with bounds " << bb << dendl; | |
2635 | for (vector<dirfrag_t>::iterator r = bb.begin(); r != bb.end(); ++r) | |
2636 | p->second.push_back(*r); | |
2637 | dirs_to_add.erase(b); | |
2638 | le->subtrees.erase(b); | |
2639 | p->second.erase(p->second.begin() + i); | |
2640 | } else { | |
2641 | ++i; | |
2642 | } | |
2643 | } | |
2644 | } | |
2645 | ||
2646 | for (auto p : dirs_to_add) { | |
2647 | CDir *dir = p.second; | |
2648 | le->metablob.add_dir_context(dir, EMetaBlob::TO_ROOT); | |
2649 | le->metablob.add_dir(dir, false); | |
2650 | } | |
2651 | ||
2652 | dout(15) << " subtrees " << le->subtrees << dendl; | |
2653 | dout(15) << " ambiguous_subtrees " << le->ambiguous_subtrees << dendl; | |
2654 | ||
2655 | //le->metablob.print(cout); | |
2656 | le->expire_pos = mds->mdlog->journaler->get_expire_pos(); | |
2657 | return le; | |
2658 | } | |
2659 | ||
2660 | void MDCache::dump_resolve_status(Formatter *f) const | |
2661 | { | |
2662 | f->open_object_section("resolve_status"); | |
2663 | f->dump_stream("resolve_gather") << resolve_gather; | |
2664 | f->dump_stream("resolve_ack_gather") << resolve_gather; | |
2665 | f->close_section(); | |
2666 | } | |
2667 | ||
2668 | void MDCache::resolve_start(MDSInternalContext *resolve_done_) | |
2669 | { | |
2670 | dout(10) << "resolve_start" << dendl; | |
2671 | assert(!resolve_done); | |
2672 | resolve_done.reset(resolve_done_); | |
2673 | ||
2674 | if (mds->mdsmap->get_root() != mds->get_nodeid()) { | |
2675 | // if we don't have the root dir, adjust it to UNKNOWN. during | |
2676 | // resolve we want mds0 to explicit claim the portion of it that | |
2677 | // it owns, so that anything beyond its bounds get left as | |
2678 | // unknown. | |
2679 | CDir *rootdir = root->get_dirfrag(frag_t()); | |
2680 | if (rootdir) | |
2681 | adjust_subtree_auth(rootdir, CDIR_AUTH_UNKNOWN); | |
2682 | } | |
2683 | resolve_gather = recovery_set; | |
2684 | } | |
2685 | ||
2686 | void MDCache::send_resolves() | |
2687 | { | |
2688 | send_slave_resolves(); | |
2689 | if (!resolve_ack_gather.empty()) { | |
2690 | dout(10) << "send_resolves still waiting for resolve ack from (" | |
2691 | << resolve_ack_gather << ")" << dendl; | |
2692 | return; | |
2693 | } | |
2694 | if (!need_resolve_rollback.empty()) { | |
2695 | dout(10) << "send_resolves still waiting for rollback to commit on (" | |
2696 | << need_resolve_rollback << ")" << dendl; | |
2697 | return; | |
2698 | } | |
2699 | send_subtree_resolves(); | |
2700 | } | |
2701 | ||
2702 | void MDCache::send_slave_resolves() | |
2703 | { | |
2704 | dout(10) << "send_slave_resolves" << dendl; | |
2705 | ||
2706 | map<mds_rank_t, MMDSResolve*> resolves; | |
2707 | ||
2708 | if (mds->is_resolve()) { | |
2709 | for (map<mds_rank_t, map<metareqid_t, MDSlaveUpdate*> >::iterator p = uncommitted_slave_updates.begin(); | |
2710 | p != uncommitted_slave_updates.end(); | |
2711 | ++p) { | |
2712 | resolves[p->first] = new MMDSResolve; | |
2713 | for (map<metareqid_t, MDSlaveUpdate*>::iterator q = p->second.begin(); | |
2714 | q != p->second.end(); | |
2715 | ++q) { | |
2716 | dout(10) << " including uncommitted " << q->first << dendl; | |
2717 | resolves[p->first]->add_slave_request(q->first, false); | |
2718 | } | |
2719 | } | |
2720 | } else { | |
2721 | set<mds_rank_t> resolve_set; | |
2722 | mds->mdsmap->get_mds_set(resolve_set, MDSMap::STATE_RESOLVE); | |
2723 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
2724 | p != active_requests.end(); | |
2725 | ++p) { | |
2726 | MDRequestRef& mdr = p->second; | |
2727 | if (!mdr->is_slave()) | |
2728 | continue; | |
2729 | if (!mdr->slave_did_prepare() && !mdr->committing) { | |
2730 | continue; | |
2731 | } | |
2732 | mds_rank_t master = mdr->slave_to_mds; | |
2733 | if (resolve_set.count(master) || is_ambiguous_slave_update(p->first, master)) { | |
2734 | dout(10) << " including uncommitted " << *mdr << dendl; | |
2735 | if (!resolves.count(master)) | |
2736 | resolves[master] = new MMDSResolve; | |
2737 | if (!mdr->committing && | |
2738 | mdr->has_more() && mdr->more()->is_inode_exporter) { | |
2739 | // re-send cap exports | |
2740 | CInode *in = mdr->more()->rename_inode; | |
2741 | map<client_t, Capability::Export> cap_map; | |
2742 | in->export_client_caps(cap_map); | |
2743 | bufferlist bl; | |
2744 | ::encode(in->ino(), bl); | |
2745 | ::encode(cap_map, bl); | |
2746 | resolves[master]->add_slave_request(p->first, bl); | |
2747 | } else { | |
2748 | resolves[master]->add_slave_request(p->first, mdr->committing); | |
2749 | } | |
2750 | } | |
2751 | } | |
2752 | } | |
2753 | ||
2754 | for (map<mds_rank_t, MMDSResolve*>::iterator p = resolves.begin(); | |
2755 | p != resolves.end(); | |
2756 | ++p) { | |
2757 | dout(10) << "sending slave resolve to mds." << p->first << dendl; | |
2758 | mds->send_message_mds(p->second, p->first); | |
2759 | resolve_ack_gather.insert(p->first); | |
2760 | } | |
2761 | } | |
2762 | ||
2763 | void MDCache::send_subtree_resolves() | |
2764 | { | |
2765 | dout(10) << "send_subtree_resolves" << dendl; | |
2766 | ||
2767 | if (migrator->is_exporting() || migrator->is_importing()) { | |
2768 | dout(7) << "send_subtree_resolves waiting, imports/exports still in progress" << dendl; | |
2769 | migrator->show_importing(); | |
2770 | migrator->show_exporting(); | |
2771 | resolves_pending = true; | |
2772 | return; // not now | |
2773 | } | |
2774 | ||
2775 | map<mds_rank_t, MMDSResolve*> resolves; | |
2776 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); | |
2777 | p != recovery_set.end(); | |
2778 | ++p) { | |
2779 | if (*p == mds->get_nodeid()) | |
2780 | continue; | |
2781 | if (mds->is_resolve() || mds->mdsmap->is_resolve(*p)) | |
2782 | resolves[*p] = new MMDSResolve; | |
2783 | } | |
2784 | ||
2785 | map<dirfrag_t, vector<dirfrag_t> > my_subtrees; | |
2786 | map<dirfrag_t, vector<dirfrag_t> > my_ambig_imports; | |
2787 | ||
2788 | // known | |
2789 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
2790 | p != subtrees.end(); | |
2791 | ++p) { | |
2792 | CDir *dir = p->first; | |
2793 | ||
2794 | // only our subtrees | |
2795 | if (dir->authority().first != mds->get_nodeid()) | |
2796 | continue; | |
2797 | ||
2798 | if (mds->is_resolve() && my_ambiguous_imports.count(dir->dirfrag())) | |
2799 | continue; // we'll add it below | |
2800 | ||
2801 | if (migrator->is_ambiguous_import(dir->dirfrag())) { | |
2802 | // ambiguous (mid-import) | |
2803 | set<CDir*> bounds; | |
2804 | get_subtree_bounds(dir, bounds); | |
2805 | vector<dirfrag_t> dfls; | |
2806 | for (set<CDir*>::iterator q = bounds.begin(); q != bounds.end(); ++q) | |
2807 | dfls.push_back((*q)->dirfrag()); | |
2808 | ||
2809 | my_ambig_imports[dir->dirfrag()] = dfls; | |
2810 | dout(10) << " ambig " << dir->dirfrag() << " " << dfls << dendl; | |
2811 | } else { | |
2812 | // not ambiguous. | |
2813 | for (map<mds_rank_t, MMDSResolve*>::iterator q = resolves.begin(); | |
2814 | q != resolves.end(); | |
2815 | ++q) | |
2816 | resolves[q->first]->add_subtree(dir->dirfrag()); | |
2817 | // bounds too | |
2818 | vector<dirfrag_t> dfls; | |
2819 | for (set<CDir*>::iterator q = subtrees[dir].begin(); | |
2820 | q != subtrees[dir].end(); | |
2821 | ++q) { | |
2822 | CDir *bound = *q; | |
2823 | dfls.push_back(bound->dirfrag()); | |
2824 | } | |
2825 | ||
2826 | my_subtrees[dir->dirfrag()] = dfls; | |
2827 | dout(10) << " claim " << dir->dirfrag() << " " << dfls << dendl; | |
2828 | } | |
2829 | } | |
2830 | ||
2831 | // ambiguous | |
2832 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator p = my_ambiguous_imports.begin(); | |
2833 | p != my_ambiguous_imports.end(); | |
2834 | ++p) { | |
2835 | my_ambig_imports[p->first] = p->second; | |
2836 | dout(10) << " ambig " << p->first << " " << p->second << dendl; | |
2837 | } | |
2838 | ||
2839 | // simplify the claimed subtree. | |
2840 | for (auto p = my_subtrees.begin(); p != my_subtrees.end(); ++p) { | |
2841 | unsigned i = 0; | |
2842 | while (i < p->second.size()) { | |
2843 | dirfrag_t b = p->second[i]; | |
2844 | if (my_subtrees.count(b)) { | |
2845 | vector<dirfrag_t>& bb = my_subtrees[b]; | |
2846 | dout(10) << " simplify: " << p->first << " swallowing " << b << " with bounds " << bb << dendl; | |
2847 | for (vector<dirfrag_t>::iterator r = bb.begin(); r != bb.end(); ++r) | |
2848 | p->second.push_back(*r); | |
2849 | my_subtrees.erase(b); | |
2850 | p->second.erase(p->second.begin() + i); | |
2851 | } else { | |
2852 | ++i; | |
2853 | } | |
2854 | } | |
2855 | } | |
2856 | ||
2857 | // send | |
2858 | for (map<mds_rank_t, MMDSResolve*>::iterator p = resolves.begin(); | |
2859 | p != resolves.end(); | |
2860 | ++p) { | |
2861 | MMDSResolve* m = p->second; | |
2862 | m->subtrees = my_subtrees; | |
2863 | m->ambiguous_imports = my_ambig_imports; | |
2864 | dout(10) << "sending subtee resolve to mds." << p->first << dendl; | |
2865 | mds->send_message_mds(m, p->first); | |
2866 | } | |
2867 | resolves_pending = false; | |
2868 | } | |
2869 | ||
2870 | void MDCache::handle_mds_failure(mds_rank_t who) | |
2871 | { | |
2872 | dout(7) << "handle_mds_failure mds." << who << dendl; | |
2873 | ||
2874 | dout(1) << "handle_mds_failure mds." << who << " : recovery peers are " << recovery_set << dendl; | |
2875 | ||
2876 | resolve_gather.insert(who); | |
2877 | discard_delayed_resolve(who); | |
2878 | ambiguous_slave_updates.erase(who); | |
2879 | ||
2880 | rejoin_gather.insert(who); | |
2881 | rejoin_sent.erase(who); // i need to send another | |
31f18b77 | 2882 | rejoin_ack_sent.erase(who); // i need to send another |
7c673cae FG |
2883 | rejoin_ack_gather.erase(who); // i'll need/get another. |
2884 | ||
2885 | dout(10) << " resolve_gather " << resolve_gather << dendl; | |
2886 | dout(10) << " resolve_ack_gather " << resolve_ack_gather << dendl; | |
2887 | dout(10) << " rejoin_sent " << rejoin_sent << dendl; | |
2888 | dout(10) << " rejoin_gather " << rejoin_gather << dendl; | |
2889 | dout(10) << " rejoin_ack_gather " << rejoin_ack_gather << dendl; | |
2890 | ||
2891 | ||
2892 | // tell the migrator too. | |
2893 | migrator->handle_mds_failure_or_stop(who); | |
2894 | ||
224ce89b WB |
2895 | // tell the balancer too. |
2896 | mds->balancer->handle_mds_failure(who); | |
2897 | ||
7c673cae FG |
2898 | // clean up any requests slave to/from this node |
2899 | list<MDRequestRef> finish; | |
2900 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
2901 | p != active_requests.end(); | |
2902 | ++p) { | |
2903 | MDRequestRef& mdr = p->second; | |
2904 | // slave to the failed node? | |
2905 | if (mdr->slave_to_mds == who) { | |
2906 | if (mdr->slave_did_prepare()) { | |
2907 | dout(10) << " slave request " << *mdr << " uncommitted, will resolve shortly" << dendl; | |
2908 | if (is_ambiguous_slave_update(p->first, mdr->slave_to_mds)) | |
2909 | remove_ambiguous_slave_update(p->first, mdr->slave_to_mds); | |
2910 | ||
2911 | if (!mdr->more()->waiting_on_slave.empty()) { | |
2912 | assert(mdr->more()->srcdn_auth_mds == mds->get_nodeid()); | |
2913 | // will rollback, no need to wait | |
2914 | if (mdr->slave_request) { | |
2915 | mdr->slave_request->put(); | |
2916 | mdr->slave_request = 0; | |
2917 | } | |
2918 | mdr->more()->waiting_on_slave.clear(); | |
2919 | } | |
2920 | } else if (!mdr->committing) { | |
2921 | dout(10) << " slave request " << *mdr << " has no prepare, finishing up" << dendl; | |
2922 | if (mdr->slave_request || mdr->slave_rolling_back()) | |
2923 | mdr->aborted = true; | |
2924 | else | |
2925 | finish.push_back(mdr); | |
2926 | } | |
2927 | } | |
2928 | ||
2929 | if (mdr->is_slave() && mdr->slave_did_prepare()) { | |
2930 | if (mdr->more()->waiting_on_slave.count(who)) { | |
2931 | assert(mdr->more()->srcdn_auth_mds == mds->get_nodeid()); | |
2932 | dout(10) << " slave request " << *mdr << " no longer need rename notity ack from mds." | |
2933 | << who << dendl; | |
2934 | mdr->more()->waiting_on_slave.erase(who); | |
2935 | if (mdr->more()->waiting_on_slave.empty() && mdr->slave_request) | |
2936 | mds->queue_waiter(new C_MDS_RetryRequest(this, mdr)); | |
2937 | } | |
2938 | ||
2939 | if (mdr->more()->srcdn_auth_mds == who && | |
2940 | mds->mdsmap->is_clientreplay_or_active_or_stopping(mdr->slave_to_mds)) { | |
2941 | // rename srcdn's auth mds failed, resolve even I'm a survivor. | |
2942 | dout(10) << " slave request " << *mdr << " uncommitted, will resolve shortly" << dendl; | |
2943 | add_ambiguous_slave_update(p->first, mdr->slave_to_mds); | |
2944 | } | |
31f18b77 FG |
2945 | } else if (mdr->slave_request) { |
2946 | MMDSSlaveRequest *slave_req = mdr->slave_request; | |
2947 | // FIXME: Slave rename request can arrive after we notice mds failure. | |
2948 | // This can cause mds to crash (does not affect integrity of FS). | |
2949 | if (slave_req->get_op() == MMDSSlaveRequest::OP_RENAMEPREP && | |
2950 | slave_req->srcdn_auth == who) | |
2951 | slave_req->mark_interrupted(); | |
7c673cae FG |
2952 | } |
2953 | ||
2954 | // failed node is slave? | |
2955 | if (mdr->is_master() && !mdr->committing) { | |
2956 | if (mdr->more()->srcdn_auth_mds == who) { | |
2957 | dout(10) << " master request " << *mdr << " waiting for rename srcdn's auth mds." | |
2958 | << who << " to recover" << dendl; | |
2959 | assert(mdr->more()->witnessed.count(who) == 0); | |
2960 | if (mdr->more()->is_ambiguous_auth) | |
2961 | mdr->clear_ambiguous_auth(); | |
2962 | // rename srcdn's auth mds failed, all witnesses will rollback | |
2963 | mdr->more()->witnessed.clear(); | |
2964 | pending_masters.erase(p->first); | |
2965 | } | |
2966 | ||
2967 | if (mdr->more()->witnessed.count(who)) { | |
2968 | mds_rank_t srcdn_auth = mdr->more()->srcdn_auth_mds; | |
2969 | if (srcdn_auth >= 0 && mdr->more()->waiting_on_slave.count(srcdn_auth)) { | |
2970 | dout(10) << " master request " << *mdr << " waiting for rename srcdn's auth mds." | |
2971 | << mdr->more()->srcdn_auth_mds << " to reply" << dendl; | |
2972 | // waiting for the slave (rename srcdn's auth mds), delay sending resolve ack | |
2973 | // until either the request is committing or the slave also fails. | |
2974 | assert(mdr->more()->waiting_on_slave.size() == 1); | |
2975 | pending_masters.insert(p->first); | |
2976 | } else { | |
2977 | dout(10) << " master request " << *mdr << " no longer witnessed by slave mds." | |
2978 | << who << " to recover" << dendl; | |
2979 | if (srcdn_auth >= 0) | |
2980 | assert(mdr->more()->witnessed.count(srcdn_auth) == 0); | |
2981 | ||
2982 | // discard this peer's prepare (if any) | |
2983 | mdr->more()->witnessed.erase(who); | |
2984 | } | |
2985 | } | |
2986 | ||
2987 | if (mdr->more()->waiting_on_slave.count(who)) { | |
2988 | dout(10) << " master request " << *mdr << " waiting for slave mds." << who | |
2989 | << " to recover" << dendl; | |
2990 | // retry request when peer recovers | |
2991 | mdr->more()->waiting_on_slave.erase(who); | |
2992 | if (mdr->more()->waiting_on_slave.empty()) | |
2993 | mds->wait_for_active_peer(who, new C_MDS_RetryRequest(this, mdr)); | |
2994 | } | |
2995 | ||
2996 | if (mdr->locking && mdr->locking_target_mds == who) | |
2997 | mdr->finish_locking(mdr->locking); | |
2998 | } | |
2999 | } | |
3000 | ||
3001 | for (map<metareqid_t, umaster>::iterator p = uncommitted_masters.begin(); | |
3002 | p != uncommitted_masters.end(); | |
3003 | ++p) { | |
3004 | // The failed MDS may have already committed the slave update | |
3005 | if (p->second.slaves.count(who)) { | |
3006 | p->second.recovering = true; | |
3007 | p->second.slaves.erase(who); | |
3008 | } | |
3009 | } | |
3010 | ||
3011 | while (!finish.empty()) { | |
3012 | dout(10) << "cleaning up slave request " << *finish.front() << dendl; | |
3013 | request_finish(finish.front()); | |
3014 | finish.pop_front(); | |
3015 | } | |
3016 | ||
3017 | kick_find_ino_peers(who); | |
3018 | kick_open_ino_peers(who); | |
3019 | ||
3020 | for (map<dirfrag_t,fragment_info_t>::iterator p = fragments.begin(); | |
3021 | p != fragments.end(); ) { | |
3022 | dirfrag_t df = p->first; | |
3023 | fragment_info_t& info = p->second; | |
3024 | ++p; | |
3025 | if (info.is_fragmenting()) | |
3026 | continue; | |
3027 | dout(10) << "cancelling fragment " << df << " bit " << info.bits << dendl; | |
3028 | list<CDir*> dirs; | |
3029 | info.dirs.swap(dirs); | |
3030 | fragments.erase(df); | |
3031 | fragment_unmark_unfreeze_dirs(dirs); | |
3032 | } | |
3033 | ||
3034 | // MDCache::shutdown_export_strays() always exports strays to mds.0 | |
3035 | if (who == mds_rank_t(0)) | |
3036 | shutdown_exported_strays.clear(); | |
3037 | ||
3038 | show_subtrees(); | |
3039 | } | |
3040 | ||
3041 | /* | |
3042 | * handle_mds_recovery - called on another node's transition | |
3043 | * from resolve -> active. | |
3044 | */ | |
3045 | void MDCache::handle_mds_recovery(mds_rank_t who) | |
3046 | { | |
3047 | dout(7) << "handle_mds_recovery mds." << who << dendl; | |
3048 | ||
3049 | // exclude all discover waiters. kick_discovers() will do the job | |
3050 | static const uint64_t i_mask = CInode::WAIT_ANY_MASK & ~CInode::WAIT_DIR; | |
3051 | static const uint64_t d_mask = CDir::WAIT_ANY_MASK & ~CDir::WAIT_DENTRY; | |
3052 | ||
3053 | list<MDSInternalContextBase*> waiters; | |
3054 | ||
3055 | // wake up any waiters in their subtrees | |
3056 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3057 | p != subtrees.end(); | |
3058 | ++p) { | |
3059 | CDir *dir = p->first; | |
3060 | ||
3061 | if (dir->authority().first != who || | |
3062 | dir->authority().second == mds->get_nodeid()) | |
3063 | continue; | |
3064 | assert(!dir->is_auth()); | |
3065 | ||
3066 | // wake any waiters | |
3067 | list<CDir*> q; | |
3068 | q.push_back(dir); | |
3069 | ||
3070 | while (!q.empty()) { | |
3071 | CDir *d = q.front(); | |
3072 | q.pop_front(); | |
3073 | d->take_waiting(d_mask, waiters); | |
3074 | ||
3075 | // inode waiters too | |
3076 | for (CDir::map_t::iterator p = d->items.begin(); | |
3077 | p != d->items.end(); | |
3078 | ++p) { | |
3079 | CDentry *dn = p->second; | |
3080 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
3081 | if (dnl->is_primary()) { | |
3082 | dnl->get_inode()->take_waiting(i_mask, waiters); | |
3083 | ||
3084 | // recurse? | |
3085 | list<CDir*> ls; | |
3086 | dnl->get_inode()->get_dirfrags(ls); | |
3087 | for (list<CDir*>::iterator p = ls.begin(); | |
3088 | p != ls.end(); | |
3089 | ++p) { | |
3090 | CDir *subdir = *p; | |
3091 | if (!subdir->is_subtree_root()) | |
3092 | q.push_back(subdir); | |
3093 | } | |
3094 | } | |
3095 | } | |
3096 | } | |
3097 | } | |
3098 | ||
3099 | kick_open_ino_peers(who); | |
3100 | kick_find_ino_peers(who); | |
3101 | ||
3102 | // queue them up. | |
3103 | mds->queue_waiters(waiters); | |
3104 | } | |
3105 | ||
3106 | void MDCache::set_recovery_set(set<mds_rank_t>& s) | |
3107 | { | |
3108 | dout(7) << "set_recovery_set " << s << dendl; | |
3109 | recovery_set = s; | |
3110 | } | |
3111 | ||
3112 | ||
3113 | /* | |
3114 | * during resolve state, we share resolves to determine who | |
3115 | * is authoritative for which trees. we expect to get an resolve | |
3116 | * from _everyone_ in the recovery_set (the mds cluster at the time of | |
3117 | * the first failure). | |
3118 | * | |
3119 | * This functions puts the passed message before returning | |
3120 | */ | |
3121 | void MDCache::handle_resolve(MMDSResolve *m) | |
3122 | { | |
3123 | dout(7) << "handle_resolve from " << m->get_source() << dendl; | |
3124 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
3125 | ||
3126 | if (mds->get_state() < MDSMap::STATE_RESOLVE) { | |
3127 | if (mds->get_want_state() == CEPH_MDS_STATE_RESOLVE) { | |
3128 | mds->wait_for_resolve(new C_MDS_RetryMessage(mds, m)); | |
3129 | return; | |
3130 | } | |
3131 | // wait until we reach the resolve stage! | |
3132 | m->put(); | |
3133 | return; | |
3134 | } | |
3135 | ||
3136 | discard_delayed_resolve(from); | |
3137 | ||
3138 | // ambiguous slave requests? | |
3139 | if (!m->slave_requests.empty()) { | |
3140 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
3141 | for (auto p = m->slave_requests.begin(); p != m->slave_requests.end(); ++p) { | |
3142 | if (uncommitted_masters.count(p->first) && !uncommitted_masters[p->first].safe) { | |
3143 | assert(!p->second.committing); | |
3144 | pending_masters.insert(p->first); | |
3145 | } | |
3146 | } | |
3147 | ||
3148 | if (!pending_masters.empty()) { | |
3149 | dout(10) << " still have pending updates, delay processing slave resolve" << dendl; | |
3150 | delayed_resolve[from] = m; | |
3151 | return; | |
3152 | } | |
3153 | } | |
3154 | ||
3155 | MMDSResolveAck *ack = new MMDSResolveAck; | |
3156 | for (auto p = m->slave_requests.begin(); p != m->slave_requests.end(); ++p) { | |
3157 | if (uncommitted_masters.count(p->first)) { //mds->sessionmap.have_completed_request(p->first)) { | |
3158 | // COMMIT | |
3159 | if (p->second.committing) { | |
3160 | // already committing, waiting for the OP_COMMITTED slave reply | |
3161 | dout(10) << " already committing slave request " << *p << " noop "<< dendl; | |
3162 | } else { | |
3163 | dout(10) << " ambiguous slave request " << *p << " will COMMIT" << dendl; | |
3164 | ack->add_commit(p->first); | |
3165 | } | |
3166 | uncommitted_masters[p->first].slaves.insert(from); // wait for slave OP_COMMITTED before we log ECommitted | |
3167 | ||
3168 | if (p->second.inode_caps.length() > 0) { | |
3169 | // slave wants to export caps (rename) | |
3170 | assert(mds->is_resolve()); | |
3171 | ||
3172 | inodeno_t ino; | |
3173 | map<client_t,Capability::Export> cap_exports; | |
3174 | bufferlist::iterator q = p->second.inode_caps.begin(); | |
3175 | ::decode(ino, q); | |
3176 | ::decode(cap_exports, q); | |
3177 | ||
3178 | assert(get_inode(ino)); | |
3179 | ||
3180 | for (map<client_t,Capability::Export>::iterator q = cap_exports.begin(); | |
3181 | q != cap_exports.end(); | |
3182 | ++q) { | |
3183 | Capability::Import& im = rejoin_imported_caps[from][ino][q->first]; | |
3184 | im.cap_id = ++last_cap_id; // assign a new cap ID | |
3185 | im.issue_seq = 1; | |
3186 | im.mseq = q->second.mseq; | |
3187 | } | |
3188 | ||
3189 | // will process these caps in rejoin stage | |
3190 | rejoin_slave_exports[ino].first = from; | |
3191 | rejoin_slave_exports[ino].second.swap(cap_exports); | |
3192 | ||
3193 | // send information of imported caps back to slave | |
3194 | ::encode(rejoin_imported_caps[from][ino], ack->commit[p->first]); | |
3195 | } | |
3196 | } else { | |
3197 | // ABORT | |
3198 | dout(10) << " ambiguous slave request " << *p << " will ABORT" << dendl; | |
3199 | assert(!p->second.committing); | |
3200 | ack->add_abort(p->first); | |
3201 | } | |
3202 | } | |
3203 | mds->send_message(ack, m->get_connection()); | |
3204 | m->put(); | |
3205 | return; | |
3206 | } | |
3207 | ||
3208 | if (!resolve_ack_gather.empty() || !need_resolve_rollback.empty()) { | |
3209 | dout(10) << "delay processing subtree resolve" << dendl; | |
3210 | delayed_resolve[from] = m; | |
3211 | return; | |
3212 | } | |
3213 | ||
3214 | bool survivor = false; | |
3215 | // am i a surviving ambiguous importer? | |
3216 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
3217 | survivor = true; | |
3218 | // check for any import success/failure (from this node) | |
3219 | map<dirfrag_t, vector<dirfrag_t> >::iterator p = my_ambiguous_imports.begin(); | |
3220 | while (p != my_ambiguous_imports.end()) { | |
3221 | map<dirfrag_t, vector<dirfrag_t> >::iterator next = p; | |
3222 | ++next; | |
3223 | CDir *dir = get_dirfrag(p->first); | |
3224 | assert(dir); | |
3225 | dout(10) << "checking ambiguous import " << *dir << dendl; | |
3226 | if (migrator->is_importing(dir->dirfrag()) && | |
3227 | migrator->get_import_peer(dir->dirfrag()) == from) { | |
3228 | assert(migrator->get_import_state(dir->dirfrag()) == Migrator::IMPORT_ACKING); | |
3229 | ||
3230 | // check if sender claims the subtree | |
3231 | bool claimed_by_sender = false; | |
3232 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator q = m->subtrees.begin(); | |
3233 | q != m->subtrees.end(); | |
3234 | ++q) { | |
3235 | // an ambiguous import won't race with a refragmentation; it's appropriate to force here. | |
3236 | CDir *base = get_force_dirfrag(q->first, false); | |
3237 | if (!base || !base->contains(dir)) | |
3238 | continue; // base not dir or an ancestor of dir, clearly doesn't claim dir. | |
3239 | ||
3240 | bool inside = true; | |
3241 | set<CDir*> bounds; | |
3242 | get_force_dirfrag_bound_set(q->second, bounds); | |
3243 | for (set<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) { | |
3244 | CDir *bound = *p; | |
3245 | if (bound->contains(dir)) { | |
3246 | inside = false; // nope, bound is dir or parent of dir, not inside. | |
3247 | break; | |
3248 | } | |
3249 | } | |
3250 | if (inside) | |
3251 | claimed_by_sender = true; | |
3252 | } | |
3253 | ||
3254 | my_ambiguous_imports.erase(p); // no longer ambiguous. | |
3255 | if (claimed_by_sender) { | |
3256 | dout(7) << "ambiguous import failed on " << *dir << dendl; | |
3257 | migrator->import_reverse(dir); | |
3258 | } else { | |
3259 | dout(7) << "ambiguous import succeeded on " << *dir << dendl; | |
3260 | migrator->import_finish(dir, true); | |
3261 | } | |
3262 | } | |
3263 | p = next; | |
3264 | } | |
3265 | } | |
3266 | ||
3267 | // update my dir_auth values | |
3268 | // need to do this on recoverying nodes _and_ bystanders (to resolve ambiguous | |
3269 | // migrations between other nodes) | |
3270 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator pi = m->subtrees.begin(); | |
3271 | pi != m->subtrees.end(); | |
3272 | ++pi) { | |
3273 | dout(10) << "peer claims " << pi->first << " bounds " << pi->second << dendl; | |
3274 | CDir *dir = get_force_dirfrag(pi->first, !survivor); | |
3275 | if (!dir) | |
3276 | continue; | |
3277 | adjust_bounded_subtree_auth(dir, pi->second, from); | |
3278 | try_subtree_merge(dir); | |
3279 | } | |
3280 | ||
3281 | show_subtrees(); | |
3282 | ||
3283 | // note ambiguous imports too | |
3284 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator pi = m->ambiguous_imports.begin(); | |
3285 | pi != m->ambiguous_imports.end(); | |
3286 | ++pi) { | |
3287 | dout(10) << "noting ambiguous import on " << pi->first << " bounds " << pi->second << dendl; | |
3288 | other_ambiguous_imports[from][pi->first].swap( pi->second ); | |
3289 | } | |
3290 | ||
3291 | // did i get them all? | |
3292 | resolve_gather.erase(from); | |
3293 | ||
3294 | maybe_resolve_finish(); | |
3295 | ||
3296 | m->put(); | |
3297 | } | |
3298 | ||
3299 | void MDCache::process_delayed_resolve() | |
3300 | { | |
3301 | dout(10) << "process_delayed_resolve" << dendl; | |
3302 | map<mds_rank_t, MMDSResolve*> tmp; | |
3303 | tmp.swap(delayed_resolve); | |
3304 | for (map<mds_rank_t, MMDSResolve*>::iterator p = tmp.begin(); p != tmp.end(); ++p) | |
3305 | handle_resolve(p->second); | |
3306 | } | |
3307 | ||
3308 | void MDCache::discard_delayed_resolve(mds_rank_t who) | |
3309 | { | |
3310 | if (delayed_resolve.count(who)) { | |
3311 | delayed_resolve[who]->put(); | |
3312 | delayed_resolve.erase(who); | |
3313 | } | |
3314 | } | |
3315 | ||
3316 | void MDCache::maybe_resolve_finish() | |
3317 | { | |
3318 | assert(resolve_ack_gather.empty()); | |
3319 | assert(need_resolve_rollback.empty()); | |
3320 | ||
3321 | if (!resolve_gather.empty()) { | |
3322 | dout(10) << "maybe_resolve_finish still waiting for resolves (" | |
3323 | << resolve_gather << ")" << dendl; | |
3324 | return; | |
3325 | } | |
3326 | ||
3327 | dout(10) << "maybe_resolve_finish got all resolves+resolve_acks, done." << dendl; | |
3328 | disambiguate_my_imports(); | |
3329 | finish_committed_masters(); | |
3330 | ||
3331 | if (resolve_done) { | |
3332 | assert(mds->is_resolve()); | |
3333 | trim_unlinked_inodes(); | |
3334 | recalc_auth_bits(false); | |
3335 | resolve_done.release()->complete(0); | |
3336 | } else { | |
3337 | maybe_send_pending_rejoins(); | |
3338 | } | |
3339 | } | |
3340 | ||
3341 | /* This functions puts the passed message before returning */ | |
3342 | void MDCache::handle_resolve_ack(MMDSResolveAck *ack) | |
3343 | { | |
3344 | dout(10) << "handle_resolve_ack " << *ack << " from " << ack->get_source() << dendl; | |
3345 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
3346 | ||
3347 | if (!resolve_ack_gather.count(from) || | |
3348 | mds->mdsmap->get_state(from) < MDSMap::STATE_RESOLVE) { | |
3349 | ack->put(); | |
3350 | return; | |
3351 | } | |
3352 | ||
3353 | if (ambiguous_slave_updates.count(from)) { | |
3354 | assert(mds->mdsmap->is_clientreplay_or_active_or_stopping(from)); | |
3355 | assert(mds->is_clientreplay() || mds->is_active() || mds->is_stopping()); | |
3356 | } | |
3357 | ||
3358 | for (map<metareqid_t, bufferlist>::iterator p = ack->commit.begin(); | |
3359 | p != ack->commit.end(); | |
3360 | ++p) { | |
3361 | dout(10) << " commit on slave " << p->first << dendl; | |
3362 | ||
3363 | if (ambiguous_slave_updates.count(from)) { | |
3364 | remove_ambiguous_slave_update(p->first, from); | |
3365 | continue; | |
3366 | } | |
3367 | ||
3368 | if (mds->is_resolve()) { | |
3369 | // replay | |
3370 | MDSlaveUpdate *su = get_uncommitted_slave_update(p->first, from); | |
3371 | assert(su); | |
3372 | ||
3373 | // log commit | |
3374 | mds->mdlog->start_submit_entry(new ESlaveUpdate(mds->mdlog, "unknown", p->first, from, | |
3375 | ESlaveUpdate::OP_COMMIT, su->origop), | |
3376 | new C_MDC_SlaveCommit(this, from, p->first)); | |
3377 | mds->mdlog->flush(); | |
3378 | ||
3379 | finish_uncommitted_slave_update(p->first, from); | |
3380 | } else { | |
3381 | MDRequestRef mdr = request_get(p->first); | |
3382 | // information about master imported caps | |
3383 | if (p->second.length() > 0) | |
3384 | mdr->more()->inode_import.claim(p->second); | |
3385 | ||
3386 | assert(mdr->slave_request == 0); // shouldn't be doing anything! | |
3387 | request_finish(mdr); | |
3388 | } | |
3389 | } | |
3390 | ||
3391 | for (vector<metareqid_t>::iterator p = ack->abort.begin(); | |
3392 | p != ack->abort.end(); | |
3393 | ++p) { | |
3394 | dout(10) << " abort on slave " << *p << dendl; | |
3395 | ||
3396 | if (mds->is_resolve()) { | |
3397 | MDSlaveUpdate *su = get_uncommitted_slave_update(*p, from); | |
3398 | assert(su); | |
3399 | ||
3400 | // perform rollback (and journal a rollback entry) | |
3401 | // note: this will hold up the resolve a bit, until the rollback entries journal. | |
3402 | MDRequestRef null_ref; | |
3403 | switch (su->origop) { | |
3404 | case ESlaveUpdate::LINK: | |
3405 | mds->server->do_link_rollback(su->rollback, from, null_ref); | |
3406 | break; | |
3407 | case ESlaveUpdate::RENAME: | |
3408 | mds->server->do_rename_rollback(su->rollback, from, null_ref); | |
3409 | break; | |
3410 | case ESlaveUpdate::RMDIR: | |
3411 | mds->server->do_rmdir_rollback(su->rollback, from, null_ref); | |
3412 | break; | |
3413 | default: | |
3414 | ceph_abort(); | |
3415 | } | |
3416 | } else { | |
3417 | MDRequestRef mdr = request_get(*p); | |
3418 | mdr->aborted = true; | |
3419 | if (mdr->slave_request) { | |
3420 | if (mdr->slave_did_prepare()) // journaling slave prepare ? | |
3421 | add_rollback(*p, from); | |
3422 | } else { | |
3423 | request_finish(mdr); | |
3424 | } | |
3425 | } | |
3426 | } | |
3427 | ||
3428 | if (!ambiguous_slave_updates.count(from)) | |
3429 | resolve_ack_gather.erase(from); | |
3430 | if (resolve_ack_gather.empty() && need_resolve_rollback.empty()) { | |
3431 | send_subtree_resolves(); | |
3432 | process_delayed_resolve(); | |
3433 | } | |
3434 | ||
3435 | ack->put(); | |
3436 | } | |
3437 | ||
3438 | void MDCache::add_uncommitted_slave_update(metareqid_t reqid, mds_rank_t master, MDSlaveUpdate *su) | |
3439 | { | |
3440 | assert(uncommitted_slave_updates[master].count(reqid) == 0); | |
3441 | uncommitted_slave_updates[master][reqid] = su; | |
3442 | for(set<CInode*>::iterator p = su->olddirs.begin(); p != su->olddirs.end(); ++p) | |
3443 | uncommitted_slave_rename_olddir[*p]++; | |
3444 | for(set<CInode*>::iterator p = su->unlinked.begin(); p != su->unlinked.end(); ++p) | |
3445 | uncommitted_slave_unlink[*p]++; | |
3446 | } | |
3447 | ||
3448 | void MDCache::finish_uncommitted_slave_update(metareqid_t reqid, mds_rank_t master) | |
3449 | { | |
3450 | assert(uncommitted_slave_updates[master].count(reqid)); | |
3451 | MDSlaveUpdate* su = uncommitted_slave_updates[master][reqid]; | |
3452 | ||
3453 | uncommitted_slave_updates[master].erase(reqid); | |
3454 | if (uncommitted_slave_updates[master].empty()) | |
3455 | uncommitted_slave_updates.erase(master); | |
3456 | // discard the non-auth subtree we renamed out of | |
3457 | for(set<CInode*>::iterator p = su->olddirs.begin(); p != su->olddirs.end(); ++p) { | |
3458 | CInode *diri = *p; | |
3459 | map<CInode*, int>::iterator it = uncommitted_slave_rename_olddir.find(diri); | |
3460 | assert(it != uncommitted_slave_rename_olddir.end()); | |
3461 | it->second--; | |
3462 | if (it->second == 0) { | |
3463 | uncommitted_slave_rename_olddir.erase(it); | |
3464 | list<CDir*> ls; | |
3465 | diri->get_dirfrags(ls); | |
3466 | for (list<CDir*>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
3467 | CDir *root = get_subtree_root(*q); | |
3468 | if (root->get_dir_auth() == CDIR_AUTH_UNDEF) { | |
3469 | try_trim_non_auth_subtree(root); | |
3470 | if (*q != root) | |
3471 | break; | |
3472 | } | |
3473 | } | |
3474 | } else | |
3475 | assert(it->second > 0); | |
3476 | } | |
3477 | // removed the inodes that were unlinked by slave update | |
3478 | for(set<CInode*>::iterator p = su->unlinked.begin(); p != su->unlinked.end(); ++p) { | |
3479 | CInode *in = *p; | |
3480 | map<CInode*, int>::iterator it = uncommitted_slave_unlink.find(in); | |
3481 | assert(it != uncommitted_slave_unlink.end()); | |
3482 | it->second--; | |
3483 | if (it->second == 0) { | |
3484 | uncommitted_slave_unlink.erase(it); | |
3485 | if (!in->get_projected_parent_dn()) | |
3486 | mds->mdcache->remove_inode_recursive(in); | |
3487 | } else | |
3488 | assert(it->second > 0); | |
3489 | } | |
3490 | delete su; | |
3491 | } | |
3492 | ||
3493 | MDSlaveUpdate* MDCache::get_uncommitted_slave_update(metareqid_t reqid, mds_rank_t master) | |
3494 | { | |
3495 | ||
3496 | MDSlaveUpdate* su = NULL; | |
3497 | if (uncommitted_slave_updates.count(master) && | |
3498 | uncommitted_slave_updates[master].count(reqid)) { | |
3499 | su = uncommitted_slave_updates[master][reqid]; | |
3500 | assert(su); | |
3501 | } | |
3502 | return su; | |
3503 | } | |
3504 | ||
3505 | void MDCache::finish_rollback(metareqid_t reqid) { | |
3506 | assert(need_resolve_rollback.count(reqid)); | |
3507 | if (mds->is_resolve()) | |
3508 | finish_uncommitted_slave_update(reqid, need_resolve_rollback[reqid]); | |
3509 | need_resolve_rollback.erase(reqid); | |
3510 | if (resolve_ack_gather.empty() && need_resolve_rollback.empty()) { | |
3511 | send_subtree_resolves(); | |
3512 | process_delayed_resolve(); | |
3513 | } | |
3514 | } | |
3515 | ||
3516 | void MDCache::disambiguate_other_imports() | |
3517 | { | |
3518 | dout(10) << "disambiguate_other_imports" << dendl; | |
3519 | ||
3520 | bool recovering = !(mds->is_clientreplay() || mds->is_active() || mds->is_stopping()); | |
3521 | // other nodes' ambiguous imports | |
3522 | for (map<mds_rank_t, map<dirfrag_t, vector<dirfrag_t> > >::iterator p = other_ambiguous_imports.begin(); | |
3523 | p != other_ambiguous_imports.end(); | |
3524 | ++p) { | |
3525 | mds_rank_t who = p->first; | |
3526 | dout(10) << "ambiguous imports for mds." << who << dendl; | |
3527 | ||
3528 | for (map<dirfrag_t, vector<dirfrag_t> >::iterator q = p->second.begin(); | |
3529 | q != p->second.end(); | |
3530 | ++q) { | |
3531 | dout(10) << " ambiguous import " << q->first << " bounds " << q->second << dendl; | |
3532 | // an ambiguous import will not race with a refragmentation; it's appropriate to force here. | |
3533 | CDir *dir = get_force_dirfrag(q->first, recovering); | |
3534 | if (!dir) continue; | |
3535 | ||
3536 | if (dir->is_ambiguous_auth() || // works for me_ambig or if i am a surviving bystander | |
3537 | dir->authority() == CDIR_AUTH_UNDEF) { // resolving | |
3538 | dout(10) << " mds." << who << " did import " << *dir << dendl; | |
3539 | adjust_bounded_subtree_auth(dir, q->second, who); | |
3540 | try_subtree_merge(dir); | |
3541 | } else { | |
3542 | dout(10) << " mds." << who << " did not import " << *dir << dendl; | |
3543 | } | |
3544 | } | |
3545 | } | |
3546 | other_ambiguous_imports.clear(); | |
3547 | } | |
3548 | ||
3549 | void MDCache::disambiguate_my_imports() | |
3550 | { | |
3551 | dout(10) << "disambiguate_my_imports" << dendl; | |
3552 | ||
3553 | if (!mds->is_resolve()) { | |
3554 | assert(my_ambiguous_imports.empty()); | |
3555 | return; | |
3556 | } | |
3557 | ||
3558 | disambiguate_other_imports(); | |
3559 | ||
3560 | // my ambiguous imports | |
3561 | mds_authority_t me_ambig(mds->get_nodeid(), mds->get_nodeid()); | |
3562 | while (!my_ambiguous_imports.empty()) { | |
3563 | map<dirfrag_t, vector<dirfrag_t> >::iterator q = my_ambiguous_imports.begin(); | |
3564 | ||
3565 | CDir *dir = get_dirfrag(q->first); | |
3566 | assert(dir); | |
3567 | ||
3568 | if (dir->authority() != me_ambig) { | |
3569 | dout(10) << "ambiguous import auth known, must not be me " << *dir << dendl; | |
3570 | cancel_ambiguous_import(dir); | |
3571 | ||
3572 | mds->mdlog->start_submit_entry(new EImportFinish(dir, false)); | |
3573 | ||
3574 | // subtree may have been swallowed by another node claiming dir | |
3575 | // as their own. | |
3576 | CDir *root = get_subtree_root(dir); | |
3577 | if (root != dir) | |
3578 | dout(10) << " subtree root is " << *root << dendl; | |
3579 | assert(root->dir_auth.first != mds->get_nodeid()); // no us! | |
3580 | try_trim_non_auth_subtree(root); | |
3581 | } else { | |
3582 | dout(10) << "ambiguous import auth unclaimed, must be me " << *dir << dendl; | |
3583 | finish_ambiguous_import(q->first); | |
3584 | mds->mdlog->start_submit_entry(new EImportFinish(dir, true)); | |
3585 | } | |
3586 | } | |
3587 | assert(my_ambiguous_imports.empty()); | |
3588 | mds->mdlog->flush(); | |
3589 | ||
3590 | // verify all my subtrees are unambiguous! | |
3591 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3592 | p != subtrees.end(); | |
3593 | ++p) { | |
3594 | CDir *dir = p->first; | |
3595 | if (dir->is_ambiguous_dir_auth()) { | |
3596 | dout(0) << "disambiguate_imports uh oh, dir_auth is still ambiguous for " << *dir << dendl; | |
3597 | } | |
3598 | assert(!dir->is_ambiguous_dir_auth()); | |
3599 | } | |
3600 | ||
3601 | show_subtrees(); | |
3602 | } | |
3603 | ||
3604 | ||
3605 | void MDCache::add_ambiguous_import(dirfrag_t base, const vector<dirfrag_t>& bounds) | |
3606 | { | |
3607 | assert(my_ambiguous_imports.count(base) == 0); | |
3608 | my_ambiguous_imports[base] = bounds; | |
3609 | } | |
3610 | ||
3611 | ||
3612 | void MDCache::add_ambiguous_import(CDir *base, const set<CDir*>& bounds) | |
3613 | { | |
3614 | // make a list | |
3615 | vector<dirfrag_t> binos; | |
3616 | for (set<CDir*>::iterator p = bounds.begin(); | |
3617 | p != bounds.end(); | |
3618 | ++p) | |
3619 | binos.push_back((*p)->dirfrag()); | |
3620 | ||
3621 | // note: this can get called twice if the exporter fails during recovery | |
3622 | if (my_ambiguous_imports.count(base->dirfrag())) | |
3623 | my_ambiguous_imports.erase(base->dirfrag()); | |
3624 | ||
3625 | add_ambiguous_import(base->dirfrag(), binos); | |
3626 | } | |
3627 | ||
3628 | void MDCache::cancel_ambiguous_import(CDir *dir) | |
3629 | { | |
3630 | dirfrag_t df = dir->dirfrag(); | |
3631 | assert(my_ambiguous_imports.count(df)); | |
3632 | dout(10) << "cancel_ambiguous_import " << df | |
3633 | << " bounds " << my_ambiguous_imports[df] | |
3634 | << " " << *dir | |
3635 | << dendl; | |
3636 | my_ambiguous_imports.erase(df); | |
3637 | } | |
3638 | ||
3639 | void MDCache::finish_ambiguous_import(dirfrag_t df) | |
3640 | { | |
3641 | assert(my_ambiguous_imports.count(df)); | |
3642 | vector<dirfrag_t> bounds; | |
3643 | bounds.swap(my_ambiguous_imports[df]); | |
3644 | my_ambiguous_imports.erase(df); | |
3645 | ||
3646 | dout(10) << "finish_ambiguous_import " << df | |
3647 | << " bounds " << bounds | |
3648 | << dendl; | |
3649 | CDir *dir = get_dirfrag(df); | |
3650 | assert(dir); | |
3651 | ||
3652 | // adjust dir_auth, import maps | |
3653 | adjust_bounded_subtree_auth(dir, bounds, mds->get_nodeid()); | |
3654 | try_subtree_merge(dir); | |
3655 | } | |
3656 | ||
3657 | void MDCache::remove_inode_recursive(CInode *in) | |
3658 | { | |
3659 | dout(10) << "remove_inode_recursive " << *in << dendl; | |
3660 | list<CDir*> ls; | |
3661 | in->get_dirfrags(ls); | |
3662 | list<CDir*>::iterator p = ls.begin(); | |
3663 | while (p != ls.end()) { | |
3664 | CDir *subdir = *p++; | |
3665 | ||
3666 | dout(10) << " removing dirfrag " << subdir << dendl; | |
3667 | CDir::map_t::iterator q = subdir->items.begin(); | |
3668 | while (q != subdir->items.end()) { | |
3669 | CDentry *dn = q->second; | |
3670 | ++q; | |
3671 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
3672 | if (dnl->is_primary()) { | |
3673 | CInode *tin = dnl->get_inode(); | |
31f18b77 | 3674 | subdir->unlink_inode(dn, false); |
7c673cae FG |
3675 | remove_inode_recursive(tin); |
3676 | } | |
3677 | subdir->remove_dentry(dn); | |
3678 | } | |
3679 | ||
3680 | if (subdir->is_subtree_root()) | |
3681 | remove_subtree(subdir); | |
3682 | in->close_dirfrag(subdir->dirfrag().frag); | |
3683 | } | |
3684 | remove_inode(in); | |
3685 | } | |
3686 | ||
3687 | bool MDCache::expire_recursive( | |
3688 | CInode *in, | |
3689 | map<mds_rank_t, MCacheExpire*>& expiremap) | |
3690 | { | |
3691 | assert(!in->is_auth()); | |
3692 | ||
3693 | dout(10) << __func__ << ":" << *in << dendl; | |
3694 | ||
3695 | // Recurse into any dirfrags beneath this inode | |
3696 | list<CDir*> ls; | |
3697 | in->get_dirfrags(ls); | |
3698 | for (auto subdir : ls) { | |
3699 | if (!in->is_mdsdir() && subdir->is_subtree_root()) { | |
3700 | dout(10) << __func__ << ": stray still has subtree " << *in << dendl; | |
3701 | return true; | |
3702 | } | |
3703 | ||
3704 | for (auto &it : subdir->items) { | |
3705 | CDentry *dn = it.second; | |
3706 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
3707 | if (dnl->is_primary()) { | |
3708 | CInode *tin = dnl->get_inode(); | |
3709 | ||
3710 | /* Remote strays with linkage (i.e. hardlinks) should not be | |
3711 | * expired, because they may be the target of | |
3712 | * a rename() as the owning MDS shuts down */ | |
3713 | if (!tin->is_stray() && tin->inode.nlink) { | |
3714 | dout(10) << __func__ << ": stray still has linkage " << *tin << dendl; | |
3715 | return true; | |
3716 | } | |
3717 | ||
3718 | const bool abort = expire_recursive(tin, expiremap); | |
3719 | if (abort) { | |
3720 | return true; | |
3721 | } | |
3722 | } | |
3723 | if (dn->lru_is_expireable()) { | |
3724 | trim_dentry(dn, expiremap); | |
3725 | } else { | |
3726 | dout(10) << __func__ << ": stray dn is not expireable " << *dn << dendl; | |
3727 | return true; | |
3728 | } | |
3729 | } | |
3730 | } | |
3731 | ||
3732 | return false; | |
3733 | } | |
3734 | ||
3735 | void MDCache::trim_unlinked_inodes() | |
3736 | { | |
3737 | dout(7) << "trim_unlinked_inodes" << dendl; | |
3738 | list<CInode*> q; | |
3739 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator p = inode_map.begin(); | |
3740 | p != inode_map.end(); | |
3741 | ++p) { | |
3742 | CInode *in = p->second; | |
3743 | if (in->get_parent_dn() == NULL && !in->is_base()) { | |
3744 | dout(7) << " will trim from " << *in << dendl; | |
3745 | q.push_back(in); | |
3746 | } | |
3747 | } | |
3748 | for (list<CInode*>::iterator p = q.begin(); p != q.end(); ++p) | |
3749 | remove_inode_recursive(*p); | |
3750 | } | |
3751 | ||
3752 | /** recalc_auth_bits() | |
3753 | * once subtree auth is disambiguated, we need to adjust all the | |
3754 | * auth and dirty bits in our cache before moving on. | |
3755 | */ | |
3756 | void MDCache::recalc_auth_bits(bool replay) | |
3757 | { | |
3758 | dout(7) << "recalc_auth_bits " << (replay ? "(replay)" : "") << dendl; | |
3759 | ||
3760 | if (root) { | |
3761 | root->inode_auth.first = mds->mdsmap->get_root(); | |
3762 | bool auth = mds->get_nodeid() == root->inode_auth.first; | |
3763 | if (auth) { | |
3764 | root->state_set(CInode::STATE_AUTH); | |
3765 | } else { | |
3766 | root->state_clear(CInode::STATE_AUTH); | |
3767 | if (!replay) | |
3768 | root->state_set(CInode::STATE_REJOINING); | |
3769 | } | |
3770 | } | |
3771 | ||
3772 | set<CInode*> subtree_inodes; | |
3773 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3774 | p != subtrees.end(); | |
3775 | ++p) { | |
3776 | if (p->first->dir_auth.first == mds->get_nodeid()) | |
3777 | subtree_inodes.insert(p->first->inode); | |
3778 | } | |
3779 | ||
3780 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
3781 | p != subtrees.end(); | |
3782 | ++p) { | |
3783 | if (p->first->inode->is_mdsdir()) { | |
3784 | CInode *in = p->first->inode; | |
3785 | bool auth = in->ino() == MDS_INO_MDSDIR(mds->get_nodeid()); | |
3786 | if (auth) { | |
3787 | in->state_set(CInode::STATE_AUTH); | |
3788 | } else { | |
3789 | in->state_clear(CInode::STATE_AUTH); | |
3790 | if (!replay) | |
3791 | in->state_set(CInode::STATE_REJOINING); | |
3792 | } | |
3793 | } | |
3794 | ||
3795 | list<CDir*> dfq; // dirfrag queue | |
3796 | dfq.push_back(p->first); | |
3797 | ||
3798 | bool auth = p->first->authority().first == mds->get_nodeid(); | |
3799 | dout(10) << " subtree auth=" << auth << " for " << *p->first << dendl; | |
3800 | ||
3801 | while (!dfq.empty()) { | |
3802 | CDir *dir = dfq.front(); | |
3803 | dfq.pop_front(); | |
3804 | ||
3805 | // dir | |
3806 | if (auth) { | |
3807 | dir->state_set(CDir::STATE_AUTH); | |
3808 | } else { | |
3809 | dir->state_clear(CDir::STATE_AUTH); | |
3810 | if (!replay) { | |
3811 | // close empty non-auth dirfrag | |
3812 | if (!dir->is_subtree_root() && dir->get_num_any() == 0) { | |
3813 | dir->inode->close_dirfrag(dir->get_frag()); | |
3814 | continue; | |
3815 | } | |
3816 | dir->state_set(CDir::STATE_REJOINING); | |
3817 | dir->state_clear(CDir::STATE_COMPLETE); | |
3818 | if (dir->is_dirty()) | |
3819 | dir->mark_clean(); | |
3820 | } | |
3821 | } | |
3822 | ||
3823 | // dentries in this dir | |
3824 | for (CDir::map_t::iterator q = dir->items.begin(); | |
3825 | q != dir->items.end(); | |
3826 | ++q) { | |
3827 | // dn | |
3828 | CDentry *dn = q->second; | |
3829 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
3830 | if (auth) { | |
3831 | dn->state_set(CDentry::STATE_AUTH); | |
3832 | } else { | |
3833 | dn->state_clear(CDentry::STATE_AUTH); | |
3834 | if (!replay) { | |
3835 | dn->state_set(CDentry::STATE_REJOINING); | |
3836 | if (dn->is_dirty()) | |
3837 | dn->mark_clean(); | |
3838 | } | |
3839 | } | |
3840 | ||
3841 | if (dnl->is_primary()) { | |
3842 | // inode | |
3843 | CInode *in = dnl->get_inode(); | |
3844 | if (auth) { | |
3845 | in->state_set(CInode::STATE_AUTH); | |
3846 | } else { | |
3847 | in->state_clear(CInode::STATE_AUTH); | |
3848 | if (!replay) { | |
3849 | in->state_set(CInode::STATE_REJOINING); | |
3850 | if (in->is_dirty()) | |
3851 | in->mark_clean(); | |
3852 | if (in->is_dirty_parent()) | |
3853 | in->clear_dirty_parent(); | |
3854 | // avoid touching scatterlocks for our subtree roots! | |
3855 | if (subtree_inodes.count(in) == 0) | |
3856 | in->clear_scatter_dirty(); | |
3857 | } | |
3858 | } | |
3859 | // recurse? | |
3860 | if (in->is_dir()) | |
3861 | in->get_nested_dirfrags(dfq); | |
3862 | } | |
3863 | } | |
3864 | } | |
3865 | } | |
3866 | ||
3867 | show_subtrees(); | |
3868 | show_cache(); | |
3869 | } | |
3870 | ||
3871 | ||
3872 | ||
3873 | // =========================================================================== | |
3874 | // REJOIN | |
3875 | ||
3876 | /* | |
3877 | * notes on scatterlock recovery: | |
3878 | * | |
3879 | * - recovering inode replica sends scatterlock data for any subtree | |
3880 | * roots (the only ones that are possibly dirty). | |
3881 | * | |
3882 | * - surviving auth incorporates any provided scatterlock data. any | |
3883 | * pending gathers are then finished, as with the other lock types. | |
3884 | * | |
3885 | * that takes care of surviving auth + (recovering replica)*. | |
3886 | * | |
3887 | * - surviving replica sends strong_inode, which includes current | |
3888 | * scatterlock state, AND any dirty scatterlock data. this | |
3889 | * provides the recovering auth with everything it might need. | |
3890 | * | |
3891 | * - recovering auth must pick initial scatterlock state based on | |
3892 | * (weak|strong) rejoins. | |
3893 | * - always assimilate scatterlock data (it can't hurt) | |
3894 | * - any surviving replica in SCATTER state -> SCATTER. otherwise, SYNC. | |
3895 | * - include base inode in ack for all inodes that saw scatterlock content | |
3896 | * | |
3897 | * also, for scatter gather, | |
3898 | * | |
3899 | * - auth increments {frag,r}stat.version on completion of any gather. | |
3900 | * | |
3901 | * - auth incorporates changes in a gather _only_ if the version | |
3902 | * matches. | |
3903 | * | |
3904 | * - replica discards changes any time the scatterlock syncs, and | |
3905 | * after recovery. | |
3906 | */ | |
3907 | ||
3908 | void MDCache::dump_rejoin_status(Formatter *f) const | |
3909 | { | |
3910 | f->open_object_section("rejoin_status"); | |
3911 | f->dump_stream("rejoin_gather") << rejoin_gather; | |
3912 | f->dump_stream("rejoin_ack_gather") << rejoin_ack_gather; | |
3913 | f->dump_unsigned("num_opening_inodes", cap_imports_num_opening); | |
3914 | f->close_section(); | |
3915 | } | |
3916 | ||
3917 | void MDCache::rejoin_start(MDSInternalContext *rejoin_done_) | |
3918 | { | |
3919 | dout(10) << "rejoin_start" << dendl; | |
3920 | assert(!rejoin_done); | |
3921 | rejoin_done.reset(rejoin_done_); | |
3922 | ||
3923 | rejoin_gather = recovery_set; | |
3924 | // need finish opening cap inodes before sending cache rejoins | |
3925 | rejoin_gather.insert(mds->get_nodeid()); | |
3926 | process_imported_caps(); | |
3927 | } | |
3928 | ||
3929 | /* | |
3930 | * rejoin phase! | |
3931 | * | |
3932 | * this initiates rejoin. it shoudl be called before we get any | |
3933 | * rejoin or rejoin_ack messages (or else mdsmap distribution is broken). | |
3934 | * | |
3935 | * we start out by sending rejoins to everyone in the recovery set. | |
3936 | * | |
3937 | * if we are rejoin, send for all regions in our cache. | |
3938 | * if we are active|stopping, send only to nodes that are are rejoining. | |
3939 | */ | |
3940 | void MDCache::rejoin_send_rejoins() | |
3941 | { | |
3942 | dout(10) << "rejoin_send_rejoins with recovery_set " << recovery_set << dendl; | |
3943 | ||
3944 | if (rejoin_gather.count(mds->get_nodeid())) { | |
3945 | dout(7) << "rejoin_send_rejoins still processing imported caps, delaying" << dendl; | |
3946 | rejoins_pending = true; | |
3947 | return; | |
3948 | } | |
3949 | if (!resolve_gather.empty()) { | |
3950 | dout(7) << "rejoin_send_rejoins still waiting for resolves (" | |
3951 | << resolve_gather << ")" << dendl; | |
3952 | rejoins_pending = true; | |
3953 | return; | |
3954 | } | |
3955 | ||
3956 | assert(!migrator->is_importing()); | |
3957 | assert(!migrator->is_exporting()); | |
3958 | ||
3959 | if (!mds->is_rejoin()) { | |
3960 | disambiguate_other_imports(); | |
3961 | } | |
3962 | ||
3963 | map<mds_rank_t, MMDSCacheRejoin*> rejoins; | |
3964 | ||
3965 | ||
3966 | // if i am rejoining, send a rejoin to everyone. | |
3967 | // otherwise, just send to others who are rejoining. | |
3968 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); | |
3969 | p != recovery_set.end(); | |
3970 | ++p) { | |
3971 | if (*p == mds->get_nodeid()) continue; // nothing to myself! | |
3972 | if (rejoin_sent.count(*p)) continue; // already sent a rejoin to this node! | |
3973 | if (mds->is_rejoin()) | |
3974 | rejoins[*p] = new MMDSCacheRejoin(MMDSCacheRejoin::OP_WEAK); | |
3975 | else if (mds->mdsmap->is_rejoin(*p)) | |
3976 | rejoins[*p] = new MMDSCacheRejoin(MMDSCacheRejoin::OP_STRONG); | |
3977 | } | |
3978 | ||
3979 | if (mds->is_rejoin()) { | |
3980 | map<client_t, set<mds_rank_t> > client_exports; | |
3981 | for (auto p = cap_exports.begin(); p != cap_exports.end(); ++p) { | |
3982 | assert(cap_export_targets.count(p->first)); | |
3983 | mds_rank_t target = cap_export_targets[p->first]; | |
3984 | if (rejoins.count(target) == 0) | |
3985 | continue; | |
3986 | rejoins[target]->cap_exports[p->first] = p->second; | |
3987 | for (auto q = p->second.begin(); q != p->second.end(); ++q) | |
3988 | client_exports[q->first].insert(target); | |
3989 | } | |
3990 | for (map<client_t, set<mds_rank_t> >::iterator p = client_exports.begin(); | |
3991 | p != client_exports.end(); | |
3992 | ++p) { | |
3993 | entity_inst_t inst = mds->sessionmap.get_inst(entity_name_t::CLIENT(p->first.v)); | |
3994 | for (set<mds_rank_t>::iterator q = p->second.begin(); q != p->second.end(); ++q) | |
3995 | rejoins[*q]->client_map[p->first] = inst; | |
3996 | } | |
3997 | } | |
3998 | ||
3999 | ||
4000 | // check all subtrees | |
4001 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
4002 | p != subtrees.end(); | |
4003 | ++p) { | |
4004 | CDir *dir = p->first; | |
4005 | assert(dir->is_subtree_root()); | |
4006 | if (dir->is_ambiguous_dir_auth()) { | |
4007 | // exporter is recovering, importer is survivor. | |
4008 | assert(rejoins.count(dir->authority().first)); | |
4009 | assert(!rejoins.count(dir->authority().second)); | |
4010 | continue; | |
4011 | } | |
4012 | ||
4013 | // my subtree? | |
4014 | if (dir->is_auth()) | |
4015 | continue; // skip my own regions! | |
4016 | ||
4017 | mds_rank_t auth = dir->get_dir_auth().first; | |
4018 | assert(auth >= 0); | |
4019 | if (rejoins.count(auth) == 0) | |
4020 | continue; // don't care about this node's subtrees | |
4021 | ||
4022 | rejoin_walk(dir, rejoins[auth]); | |
4023 | } | |
4024 | ||
4025 | // rejoin root inodes, too | |
4026 | for (map<mds_rank_t, MMDSCacheRejoin*>::iterator p = rejoins.begin(); | |
4027 | p != rejoins.end(); | |
4028 | ++p) { | |
4029 | if (mds->is_rejoin()) { | |
4030 | // weak | |
4031 | if (p->first == 0 && root) { | |
4032 | p->second->add_weak_inode(root->vino()); | |
4033 | if (root->is_dirty_scattered()) { | |
4034 | dout(10) << " sending scatterlock state on root " << *root << dendl; | |
4035 | p->second->add_scatterlock_state(root); | |
4036 | } | |
4037 | } | |
4038 | if (CInode *in = get_inode(MDS_INO_MDSDIR(p->first))) { | |
4039 | if (in) | |
4040 | p->second->add_weak_inode(in->vino()); | |
4041 | } | |
4042 | } else { | |
4043 | // strong | |
4044 | if (p->first == 0 && root) { | |
4045 | p->second->add_strong_inode(root->vino(), | |
4046 | root->get_replica_nonce(), | |
4047 | root->get_caps_wanted(), | |
4048 | root->filelock.get_state(), | |
4049 | root->nestlock.get_state(), | |
4050 | root->dirfragtreelock.get_state()); | |
4051 | root->state_set(CInode::STATE_REJOINING); | |
4052 | if (root->is_dirty_scattered()) { | |
4053 | dout(10) << " sending scatterlock state on root " << *root << dendl; | |
4054 | p->second->add_scatterlock_state(root); | |
4055 | } | |
4056 | } | |
4057 | ||
4058 | if (CInode *in = get_inode(MDS_INO_MDSDIR(p->first))) { | |
4059 | p->second->add_strong_inode(in->vino(), | |
4060 | in->get_replica_nonce(), | |
4061 | in->get_caps_wanted(), | |
4062 | in->filelock.get_state(), | |
4063 | in->nestlock.get_state(), | |
4064 | in->dirfragtreelock.get_state()); | |
4065 | in->state_set(CInode::STATE_REJOINING); | |
4066 | } | |
4067 | } | |
4068 | } | |
4069 | ||
4070 | if (!mds->is_rejoin()) { | |
4071 | // i am survivor. send strong rejoin. | |
4072 | // note request remote_auth_pins, xlocks | |
4073 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
4074 | p != active_requests.end(); | |
4075 | ++p) { | |
4076 | MDRequestRef& mdr = p->second; | |
4077 | if (mdr->is_slave()) | |
4078 | continue; | |
4079 | // auth pins | |
4080 | for (map<MDSCacheObject*,mds_rank_t>::iterator q = mdr->remote_auth_pins.begin(); | |
4081 | q != mdr->remote_auth_pins.end(); | |
4082 | ++q) { | |
4083 | if (!q->first->is_auth()) { | |
4084 | assert(q->second == q->first->authority().first); | |
4085 | if (rejoins.count(q->second) == 0) continue; | |
4086 | MMDSCacheRejoin *rejoin = rejoins[q->second]; | |
4087 | ||
4088 | dout(15) << " " << *mdr << " authpin on " << *q->first << dendl; | |
4089 | MDSCacheObjectInfo i; | |
4090 | q->first->set_object_info(i); | |
4091 | if (i.ino) | |
4092 | rejoin->add_inode_authpin(vinodeno_t(i.ino, i.snapid), mdr->reqid, mdr->attempt); | |
4093 | else | |
4094 | rejoin->add_dentry_authpin(i.dirfrag, i.dname, i.snapid, mdr->reqid, mdr->attempt); | |
4095 | ||
4096 | if (mdr->has_more() && mdr->more()->is_remote_frozen_authpin && | |
4097 | mdr->more()->rename_inode == q->first) | |
4098 | rejoin->add_inode_frozen_authpin(vinodeno_t(i.ino, i.snapid), | |
4099 | mdr->reqid, mdr->attempt); | |
4100 | } | |
4101 | } | |
4102 | // xlocks | |
4103 | for (set<SimpleLock*>::iterator q = mdr->xlocks.begin(); | |
4104 | q != mdr->xlocks.end(); | |
4105 | ++q) { | |
4106 | if (!(*q)->get_parent()->is_auth()) { | |
4107 | mds_rank_t who = (*q)->get_parent()->authority().first; | |
4108 | if (rejoins.count(who) == 0) continue; | |
4109 | MMDSCacheRejoin *rejoin = rejoins[who]; | |
4110 | ||
4111 | dout(15) << " " << *mdr << " xlock on " << **q << " " << *(*q)->get_parent() << dendl; | |
4112 | MDSCacheObjectInfo i; | |
4113 | (*q)->get_parent()->set_object_info(i); | |
4114 | if (i.ino) | |
4115 | rejoin->add_inode_xlock(vinodeno_t(i.ino, i.snapid), (*q)->get_type(), | |
4116 | mdr->reqid, mdr->attempt); | |
4117 | else | |
4118 | rejoin->add_dentry_xlock(i.dirfrag, i.dname, i.snapid, | |
4119 | mdr->reqid, mdr->attempt); | |
4120 | } | |
4121 | } | |
4122 | // remote wrlocks | |
4123 | for (map<SimpleLock*, mds_rank_t>::iterator q = mdr->remote_wrlocks.begin(); | |
4124 | q != mdr->remote_wrlocks.end(); | |
4125 | ++q) { | |
4126 | mds_rank_t who = q->second; | |
4127 | if (rejoins.count(who) == 0) continue; | |
4128 | MMDSCacheRejoin *rejoin = rejoins[who]; | |
4129 | ||
4130 | dout(15) << " " << *mdr << " wrlock on " << q->second | |
4131 | << " " << q->first->get_parent() << dendl; | |
4132 | MDSCacheObjectInfo i; | |
4133 | q->first->get_parent()->set_object_info(i); | |
4134 | assert(i.ino); | |
4135 | rejoin->add_inode_wrlock(vinodeno_t(i.ino, i.snapid), q->first->get_type(), | |
4136 | mdr->reqid, mdr->attempt); | |
4137 | } | |
4138 | } | |
4139 | } | |
4140 | ||
4141 | // send the messages | |
4142 | for (map<mds_rank_t,MMDSCacheRejoin*>::iterator p = rejoins.begin(); | |
4143 | p != rejoins.end(); | |
4144 | ++p) { | |
4145 | assert(rejoin_sent.count(p->first) == 0); | |
4146 | assert(rejoin_ack_gather.count(p->first) == 0); | |
4147 | rejoin_sent.insert(p->first); | |
4148 | rejoin_ack_gather.insert(p->first); | |
4149 | mds->send_message_mds(p->second, p->first); | |
4150 | } | |
4151 | rejoin_ack_gather.insert(mds->get_nodeid()); // we need to complete rejoin_gather_finish, too | |
4152 | rejoins_pending = false; | |
4153 | ||
4154 | // nothing? | |
4155 | if (mds->is_rejoin() && rejoins.empty()) { | |
4156 | dout(10) << "nothing to rejoin" << dendl; | |
4157 | rejoin_gather_finish(); | |
4158 | } | |
4159 | } | |
4160 | ||
4161 | ||
4162 | /** | |
4163 | * rejoin_walk - build rejoin declarations for a subtree | |
4164 | * | |
4165 | * @param dir subtree root | |
4166 | * @param rejoin rejoin message | |
4167 | * | |
4168 | * from a rejoining node: | |
4169 | * weak dirfrag | |
4170 | * weak dentries (w/ connectivity) | |
4171 | * | |
4172 | * from a surviving node: | |
4173 | * strong dirfrag | |
4174 | * strong dentries (no connectivity!) | |
4175 | * strong inodes | |
4176 | */ | |
4177 | void MDCache::rejoin_walk(CDir *dir, MMDSCacheRejoin *rejoin) | |
4178 | { | |
4179 | dout(10) << "rejoin_walk " << *dir << dendl; | |
4180 | ||
4181 | list<CDir*> nested; // finish this dir, then do nested items | |
4182 | ||
4183 | if (mds->is_rejoin()) { | |
4184 | // WEAK | |
4185 | rejoin->add_weak_dirfrag(dir->dirfrag()); | |
4186 | for (CDir::map_t::iterator p = dir->items.begin(); | |
4187 | p != dir->items.end(); | |
4188 | ++p) { | |
4189 | CDentry *dn = p->second; | |
4190 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4191 | dout(15) << " add_weak_primary_dentry " << *dn << dendl; | |
4192 | assert(dnl->is_primary()); | |
4193 | CInode *in = dnl->get_inode(); | |
4194 | assert(dnl->get_inode()->is_dir()); | |
4195 | rejoin->add_weak_primary_dentry(dir->ino(), dn->name.c_str(), dn->first, dn->last, in->ino()); | |
4196 | in->get_nested_dirfrags(nested); | |
4197 | if (in->is_dirty_scattered()) { | |
4198 | dout(10) << " sending scatterlock state on " << *in << dendl; | |
4199 | rejoin->add_scatterlock_state(in); | |
4200 | } | |
4201 | } | |
4202 | } else { | |
4203 | // STRONG | |
4204 | dout(15) << " add_strong_dirfrag " << *dir << dendl; | |
4205 | rejoin->add_strong_dirfrag(dir->dirfrag(), dir->get_replica_nonce(), dir->get_dir_rep()); | |
4206 | dir->state_set(CDir::STATE_REJOINING); | |
4207 | ||
4208 | for (CDir::map_t::iterator p = dir->items.begin(); | |
4209 | p != dir->items.end(); | |
4210 | ++p) { | |
4211 | CDentry *dn = p->second; | |
4212 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4213 | dout(15) << " add_strong_dentry " << *dn << dendl; | |
4214 | rejoin->add_strong_dentry(dir->dirfrag(), dn->name, dn->first, dn->last, | |
4215 | dnl->is_primary() ? dnl->get_inode()->ino():inodeno_t(0), | |
4216 | dnl->is_remote() ? dnl->get_remote_ino():inodeno_t(0), | |
4217 | dnl->is_remote() ? dnl->get_remote_d_type():0, | |
4218 | dn->get_replica_nonce(), | |
4219 | dn->lock.get_state()); | |
4220 | dn->state_set(CDentry::STATE_REJOINING); | |
4221 | if (dnl->is_primary()) { | |
4222 | CInode *in = dnl->get_inode(); | |
4223 | dout(15) << " add_strong_inode " << *in << dendl; | |
4224 | rejoin->add_strong_inode(in->vino(), | |
4225 | in->get_replica_nonce(), | |
4226 | in->get_caps_wanted(), | |
4227 | in->filelock.get_state(), | |
4228 | in->nestlock.get_state(), | |
4229 | in->dirfragtreelock.get_state()); | |
4230 | in->state_set(CInode::STATE_REJOINING); | |
4231 | in->get_nested_dirfrags(nested); | |
4232 | if (in->is_dirty_scattered()) { | |
4233 | dout(10) << " sending scatterlock state on " << *in << dendl; | |
4234 | rejoin->add_scatterlock_state(in); | |
4235 | } | |
4236 | } | |
4237 | } | |
4238 | } | |
4239 | ||
4240 | // recurse into nested dirs | |
4241 | for (list<CDir*>::iterator p = nested.begin(); | |
4242 | p != nested.end(); | |
4243 | ++p) | |
4244 | rejoin_walk(*p, rejoin); | |
4245 | } | |
4246 | ||
4247 | ||
4248 | /* | |
4249 | * i got a rejoin. | |
4250 | * - reply with the lockstate | |
4251 | * | |
4252 | * if i am active|stopping, | |
4253 | * - remove source from replica list for everything not referenced here. | |
4254 | * This function puts the passed message before returning. | |
4255 | */ | |
4256 | void MDCache::handle_cache_rejoin(MMDSCacheRejoin *m) | |
4257 | { | |
4258 | dout(7) << "handle_cache_rejoin " << *m << " from " << m->get_source() | |
4259 | << " (" << m->get_payload().length() << " bytes)" | |
4260 | << dendl; | |
4261 | ||
4262 | switch (m->op) { | |
4263 | case MMDSCacheRejoin::OP_WEAK: | |
4264 | handle_cache_rejoin_weak(m); | |
4265 | break; | |
4266 | case MMDSCacheRejoin::OP_STRONG: | |
4267 | handle_cache_rejoin_strong(m); | |
4268 | break; | |
4269 | case MMDSCacheRejoin::OP_ACK: | |
4270 | handle_cache_rejoin_ack(m); | |
4271 | break; | |
4272 | ||
4273 | default: | |
4274 | ceph_abort(); | |
4275 | } | |
4276 | m->put(); | |
4277 | } | |
4278 | ||
4279 | ||
4280 | /* | |
4281 | * handle_cache_rejoin_weak | |
4282 | * | |
4283 | * the sender | |
4284 | * - is recovering from their journal. | |
4285 | * - may have incorrect (out of date) inode contents | |
4286 | * - will include weak dirfrag if sender is dirfrag auth and parent inode auth is recipient | |
4287 | * | |
4288 | * if the sender didn't trim_non_auth(), they | |
4289 | * - may have incorrect (out of date) dentry/inode linkage | |
4290 | * - may have deleted/purged inodes | |
4291 | * and i may have to go to disk to get accurate inode contents. yuck. | |
4292 | * This functions DOES NOT put the passed message before returning | |
4293 | */ | |
4294 | void MDCache::handle_cache_rejoin_weak(MMDSCacheRejoin *weak) | |
4295 | { | |
4296 | mds_rank_t from = mds_rank_t(weak->get_source().num()); | |
4297 | ||
4298 | // possible response(s) | |
4299 | MMDSCacheRejoin *ack = 0; // if survivor | |
4300 | set<vinodeno_t> acked_inodes; // if survivor | |
4301 | set<SimpleLock *> gather_locks; // if survivor | |
4302 | bool survivor = false; // am i a survivor? | |
4303 | ||
4304 | if (mds->is_clientreplay() || mds->is_active() || mds->is_stopping()) { | |
4305 | survivor = true; | |
4306 | dout(10) << "i am a surivivor, and will ack immediately" << dendl; | |
4307 | ack = new MMDSCacheRejoin(MMDSCacheRejoin::OP_ACK); | |
4308 | ||
4309 | map<inodeno_t,map<client_t,Capability::Import> > imported_caps; | |
4310 | ||
4311 | // check cap exports | |
4312 | for (auto p = weak->cap_exports.begin(); p != weak->cap_exports.end(); ++p) { | |
4313 | CInode *in = get_inode(p->first); | |
4314 | assert(!in || in->is_auth()); | |
4315 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
4316 | dout(10) << " claiming cap import " << p->first << " client." << q->first << " on " << *in << dendl; | |
4317 | Capability *cap = rejoin_import_cap(in, q->first, q->second, from); | |
4318 | Capability::Import& im = imported_caps[p->first][q->first]; | |
4319 | if (cap) { | |
4320 | im.cap_id = cap->get_cap_id(); | |
4321 | im.issue_seq = cap->get_last_seq(); | |
4322 | im.mseq = cap->get_mseq(); | |
4323 | } else { | |
4324 | // all are zero | |
4325 | } | |
4326 | } | |
4327 | mds->locker->eval(in, CEPH_CAP_LOCKS, true); | |
4328 | } | |
4329 | ||
4330 | ::encode(imported_caps, ack->imported_caps); | |
4331 | } else { | |
4332 | assert(mds->is_rejoin()); | |
4333 | ||
4334 | // we may have already received a strong rejoin from the sender. | |
4335 | rejoin_scour_survivor_replicas(from, NULL, acked_inodes, gather_locks); | |
4336 | assert(gather_locks.empty()); | |
4337 | ||
4338 | // check cap exports. | |
4339 | rejoin_client_map.insert(weak->client_map.begin(), weak->client_map.end()); | |
4340 | ||
4341 | for (auto p = weak->cap_exports.begin(); p != weak->cap_exports.end(); ++p) { | |
4342 | CInode *in = get_inode(p->first); | |
4343 | assert(in && in->is_auth()); | |
4344 | // note | |
4345 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
4346 | dout(10) << " claiming cap import " << p->first << " client." << q->first << dendl; | |
4347 | cap_imports[p->first][q->first][from] = q->second; | |
4348 | } | |
4349 | } | |
4350 | } | |
4351 | ||
4352 | // assimilate any potentially dirty scatterlock state | |
4353 | for (map<inodeno_t,MMDSCacheRejoin::lock_bls>::iterator p = weak->inode_scatterlocks.begin(); | |
4354 | p != weak->inode_scatterlocks.end(); | |
4355 | ++p) { | |
4356 | CInode *in = get_inode(p->first); | |
4357 | assert(in); | |
4358 | in->decode_lock_state(CEPH_LOCK_IFILE, p->second.file); | |
4359 | in->decode_lock_state(CEPH_LOCK_INEST, p->second.nest); | |
4360 | in->decode_lock_state(CEPH_LOCK_IDFT, p->second.dft); | |
4361 | if (!survivor) | |
4362 | rejoin_potential_updated_scatterlocks.insert(in); | |
4363 | } | |
4364 | ||
4365 | // recovering peer may send incorrect dirfrags here. we need to | |
4366 | // infer which dirfrag they meant. the ack will include a | |
4367 | // strong_dirfrag that will set them straight on the fragmentation. | |
4368 | ||
4369 | // walk weak map | |
4370 | set<CDir*> dirs_to_share; | |
4371 | for (set<dirfrag_t>::iterator p = weak->weak_dirfrags.begin(); | |
4372 | p != weak->weak_dirfrags.end(); | |
4373 | ++p) { | |
4374 | CInode *diri = get_inode(p->ino); | |
4375 | if (!diri) | |
4376 | dout(0) << " missing dir ino " << p->ino << dendl; | |
4377 | assert(diri); | |
4378 | ||
4379 | list<frag_t> ls; | |
4380 | if (diri->dirfragtree.is_leaf(p->frag)) { | |
4381 | ls.push_back(p->frag); | |
4382 | } else { | |
4383 | diri->dirfragtree.get_leaves_under(p->frag, ls); | |
4384 | if (ls.empty()) | |
4385 | ls.push_back(diri->dirfragtree[p->frag.value()]); | |
4386 | } | |
4387 | for (list<frag_t>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
4388 | frag_t fg = *q; | |
4389 | CDir *dir = diri->get_dirfrag(fg); | |
4390 | if (!dir) { | |
4391 | dout(0) << " missing dir for " << p->frag << " (which maps to " << fg << ") on " << *diri << dendl; | |
4392 | continue; | |
4393 | } | |
4394 | assert(dir); | |
4395 | if (dirs_to_share.count(dir)) { | |
4396 | dout(10) << " already have " << p->frag << " -> " << fg << " " << *dir << dendl; | |
4397 | } else { | |
4398 | dirs_to_share.insert(dir); | |
4399 | unsigned nonce = dir->add_replica(from); | |
4400 | dout(10) << " have " << p->frag << " -> " << fg << " " << *dir << dendl; | |
4401 | if (ack) { | |
4402 | ack->add_strong_dirfrag(dir->dirfrag(), nonce, dir->dir_rep); | |
4403 | ack->add_dirfrag_base(dir); | |
4404 | } | |
4405 | } | |
4406 | } | |
4407 | } | |
4408 | ||
4409 | for (map<inodeno_t,map<string_snap_t,MMDSCacheRejoin::dn_weak> >::iterator p = weak->weak.begin(); | |
4410 | p != weak->weak.end(); | |
4411 | ++p) { | |
4412 | CInode *diri = get_inode(p->first); | |
4413 | if (!diri) | |
4414 | dout(0) << " missing dir ino " << p->first << dendl; | |
4415 | assert(diri); | |
4416 | ||
4417 | // weak dentries | |
4418 | CDir *dir = 0; | |
4419 | for (map<string_snap_t,MMDSCacheRejoin::dn_weak>::iterator q = p->second.begin(); | |
4420 | q != p->second.end(); | |
4421 | ++q) { | |
4422 | // locate proper dirfrag. | |
4423 | // optimize for common case (one dirfrag) to avoid dirs_to_share set check | |
4424 | frag_t fg = diri->pick_dirfrag(q->first.name); | |
4425 | if (!dir || dir->get_frag() != fg) { | |
4426 | dir = diri->get_dirfrag(fg); | |
4427 | if (!dir) | |
4428 | dout(0) << " missing dir frag " << fg << " on " << *diri << dendl; | |
4429 | assert(dir); | |
4430 | assert(dirs_to_share.count(dir)); | |
4431 | } | |
4432 | ||
4433 | // and dentry | |
4434 | CDentry *dn = dir->lookup(q->first.name, q->first.snapid); | |
4435 | assert(dn); | |
4436 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4437 | assert(dnl->is_primary()); | |
4438 | ||
4439 | if (survivor && dn->is_replica(from)) | |
4440 | dentry_remove_replica(dn, from, gather_locks); | |
4441 | unsigned dnonce = dn->add_replica(from); | |
4442 | dout(10) << " have " << *dn << dendl; | |
4443 | if (ack) | |
4444 | ack->add_strong_dentry(dir->dirfrag(), dn->name, dn->first, dn->last, | |
4445 | dnl->get_inode()->ino(), inodeno_t(0), 0, | |
4446 | dnonce, dn->lock.get_replica_state()); | |
4447 | ||
4448 | // inode | |
4449 | CInode *in = dnl->get_inode(); | |
4450 | assert(in); | |
4451 | ||
4452 | if (survivor && in->is_replica(from)) | |
4453 | inode_remove_replica(in, from, true, gather_locks); | |
4454 | unsigned inonce = in->add_replica(from); | |
4455 | dout(10) << " have " << *in << dendl; | |
4456 | ||
4457 | // scatter the dirlock, just in case? | |
4458 | if (!survivor && in->is_dir() && in->has_subtree_root_dirfrag()) | |
4459 | in->filelock.set_state(LOCK_MIX); | |
4460 | ||
4461 | if (ack) { | |
4462 | acked_inodes.insert(in->vino()); | |
4463 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4464 | bufferlist bl; | |
4465 | in->_encode_locks_state_for_rejoin(bl, from); | |
4466 | ack->add_inode_locks(in, inonce, bl); | |
4467 | } | |
4468 | } | |
4469 | } | |
4470 | ||
4471 | // weak base inodes? (root, stray, etc.) | |
4472 | for (set<vinodeno_t>::iterator p = weak->weak_inodes.begin(); | |
4473 | p != weak->weak_inodes.end(); | |
4474 | ++p) { | |
4475 | CInode *in = get_inode(*p); | |
4476 | assert(in); // hmm fixme wrt stray? | |
4477 | if (survivor && in->is_replica(from)) | |
4478 | inode_remove_replica(in, from, true, gather_locks); | |
4479 | unsigned inonce = in->add_replica(from); | |
4480 | dout(10) << " have base " << *in << dendl; | |
4481 | ||
4482 | if (ack) { | |
4483 | acked_inodes.insert(in->vino()); | |
4484 | ack->add_inode_base(in, mds->mdsmap->get_up_features()); | |
4485 | bufferlist bl; | |
4486 | in->_encode_locks_state_for_rejoin(bl, from); | |
4487 | ack->add_inode_locks(in, inonce, bl); | |
4488 | } | |
4489 | } | |
4490 | ||
4491 | assert(rejoin_gather.count(from)); | |
4492 | rejoin_gather.erase(from); | |
4493 | if (survivor) { | |
4494 | // survivor. do everything now. | |
4495 | for (map<inodeno_t,MMDSCacheRejoin::lock_bls>::iterator p = weak->inode_scatterlocks.begin(); | |
4496 | p != weak->inode_scatterlocks.end(); | |
4497 | ++p) { | |
4498 | CInode *in = get_inode(p->first); | |
4499 | assert(in); | |
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? | |
4514 | if (rejoin_gather.empty()) { | |
4515 | rejoin_gather_finish(); | |
4516 | } else { | |
4517 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" << dendl; | |
4518 | } | |
4519 | } | |
4520 | } | |
4521 | ||
4522 | class C_MDC_RejoinGatherFinish : public MDCacheContext { | |
4523 | public: | |
4524 | explicit C_MDC_RejoinGatherFinish(MDCache *c) : MDCacheContext(c) {} | |
4525 | void finish(int r) override { | |
4526 | mdcache->rejoin_gather_finish(); | |
4527 | } | |
4528 | }; | |
4529 | ||
4530 | /* | |
4531 | * rejoin_scour_survivor_replica - remove source from replica list on unmentioned objects | |
4532 | * | |
4533 | * all validated replicas are acked with a strong nonce, etc. if that isn't in the | |
4534 | * ack, the replica dne, and we can remove it from our replica maps. | |
4535 | */ | |
4536 | void MDCache::rejoin_scour_survivor_replicas(mds_rank_t from, MMDSCacheRejoin *ack, | |
4537 | set<vinodeno_t>& acked_inodes, | |
4538 | set<SimpleLock *>& gather_locks) | |
4539 | { | |
4540 | dout(10) << "rejoin_scour_survivor_replicas from mds." << from << dendl; | |
4541 | ||
4542 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator p = inode_map.begin(); | |
4543 | p != inode_map.end(); | |
4544 | ++p) { | |
4545 | CInode *in = p->second; | |
4546 | ||
4547 | // inode? | |
4548 | if (in->is_auth() && | |
4549 | in->is_replica(from) && | |
4550 | (ack == NULL || acked_inodes.count(p->second->vino()) == 0)) { | |
4551 | inode_remove_replica(in, from, false, gather_locks); | |
4552 | dout(10) << " rem " << *in << dendl; | |
4553 | } | |
4554 | ||
4555 | if (!in->is_dir()) continue; | |
4556 | ||
4557 | list<CDir*> dfs; | |
4558 | in->get_dirfrags(dfs); | |
4559 | for (list<CDir*>::iterator p = dfs.begin(); | |
4560 | p != dfs.end(); | |
4561 | ++p) { | |
4562 | CDir *dir = *p; | |
181888fb FG |
4563 | if (!dir->is_auth()) |
4564 | continue; | |
7c673cae | 4565 | |
181888fb | 4566 | if (dir->is_replica(from) && |
7c673cae FG |
4567 | (ack == NULL || ack->strong_dirfrags.count(dir->dirfrag()) == 0)) { |
4568 | dir->remove_replica(from); | |
4569 | dout(10) << " rem " << *dir << dendl; | |
4570 | } | |
4571 | ||
4572 | // dentries | |
4573 | for (CDir::map_t::iterator p = dir->items.begin(); | |
4574 | p != dir->items.end(); | |
4575 | ++p) { | |
4576 | CDentry *dn = p->second; | |
4577 | ||
4578 | if (dn->is_replica(from) && | |
4579 | (ack == NULL || | |
4580 | ack->strong_dentries.count(dir->dirfrag()) == 0 || | |
4581 | ack->strong_dentries[dir->dirfrag()].count(string_snap_t(dn->name, dn->last)) == 0)) { | |
4582 | dentry_remove_replica(dn, from, gather_locks); | |
4583 | dout(10) << " rem " << *dn << dendl; | |
4584 | } | |
4585 | } | |
4586 | } | |
4587 | } | |
4588 | } | |
4589 | ||
4590 | ||
4591 | CInode *MDCache::rejoin_invent_inode(inodeno_t ino, snapid_t last) | |
4592 | { | |
4593 | CInode *in = new CInode(this, true, 1, last); | |
4594 | in->inode.ino = ino; | |
4595 | in->state_set(CInode::STATE_REJOINUNDEF); | |
4596 | add_inode(in); | |
4597 | rejoin_undef_inodes.insert(in); | |
4598 | dout(10) << " invented " << *in << dendl; | |
4599 | return in; | |
4600 | } | |
4601 | ||
4602 | CDir *MDCache::rejoin_invent_dirfrag(dirfrag_t df) | |
4603 | { | |
4604 | CInode *in = get_inode(df.ino); | |
4605 | if (!in) | |
4606 | in = rejoin_invent_inode(df.ino, CEPH_NOSNAP); | |
4607 | if (!in->is_dir()) { | |
4608 | assert(in->state_test(CInode::STATE_REJOINUNDEF)); | |
4609 | in->inode.mode = S_IFDIR; | |
4610 | in->inode.dir_layout.dl_dir_hash = g_conf->mds_default_dir_hash; | |
4611 | } | |
4612 | CDir *dir = in->get_or_open_dirfrag(this, df.frag); | |
4613 | dir->state_set(CDir::STATE_REJOINUNDEF); | |
4614 | rejoin_undef_dirfrags.insert(dir); | |
4615 | dout(10) << " invented " << *dir << dendl; | |
4616 | return dir; | |
4617 | } | |
4618 | ||
4619 | /* This functions DOES NOT put the passed message before returning */ | |
4620 | void MDCache::handle_cache_rejoin_strong(MMDSCacheRejoin *strong) | |
4621 | { | |
4622 | mds_rank_t from = mds_rank_t(strong->get_source().num()); | |
4623 | ||
4624 | // only a recovering node will get a strong rejoin. | |
4625 | assert(mds->is_rejoin()); | |
4626 | ||
4627 | // assimilate any potentially dirty scatterlock state | |
4628 | for (map<inodeno_t,MMDSCacheRejoin::lock_bls>::iterator p = strong->inode_scatterlocks.begin(); | |
4629 | p != strong->inode_scatterlocks.end(); | |
4630 | ++p) { | |
4631 | CInode *in = get_inode(p->first); | |
4632 | assert(in); | |
4633 | in->decode_lock_state(CEPH_LOCK_IFILE, p->second.file); | |
4634 | in->decode_lock_state(CEPH_LOCK_INEST, p->second.nest); | |
4635 | in->decode_lock_state(CEPH_LOCK_IDFT, p->second.dft); | |
4636 | rejoin_potential_updated_scatterlocks.insert(in); | |
4637 | } | |
4638 | ||
4639 | rejoin_unlinked_inodes[from].clear(); | |
4640 | ||
4641 | // surviving peer may send incorrect dirfrag here (maybe they didn't | |
4642 | // get the fragment notify, or maybe we rolled back?). we need to | |
4643 | // infer the right frag and get them with the program. somehow. | |
4644 | // we don't normally send ACK.. so we'll need to bundle this with | |
4645 | // MISSING or something. | |
4646 | ||
4647 | // strong dirfrags/dentries. | |
4648 | // also process auth_pins, xlocks. | |
4649 | for (map<dirfrag_t, MMDSCacheRejoin::dirfrag_strong>::iterator p = strong->strong_dirfrags.begin(); | |
4650 | p != strong->strong_dirfrags.end(); | |
4651 | ++p) { | |
4652 | CInode *diri = get_inode(p->first.ino); | |
4653 | if (!diri) | |
4654 | diri = rejoin_invent_inode(p->first.ino, CEPH_NOSNAP); | |
4655 | CDir *dir = diri->get_dirfrag(p->first.frag); | |
4656 | bool refragged = false; | |
4657 | if (dir) { | |
4658 | dout(10) << " have " << *dir << dendl; | |
4659 | } else { | |
4660 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) | |
4661 | dir = rejoin_invent_dirfrag(dirfrag_t(diri->ino(), frag_t())); | |
4662 | else if (diri->dirfragtree.is_leaf(p->first.frag)) | |
4663 | dir = rejoin_invent_dirfrag(p->first); | |
4664 | } | |
4665 | if (dir) { | |
4666 | dir->add_replica(from, p->second.nonce); | |
4667 | dir->dir_rep = p->second.dir_rep; | |
4668 | } else { | |
4669 | dout(10) << " frag " << p->first << " doesn't match dirfragtree " << *diri << dendl; | |
4670 | list<frag_t> ls; | |
4671 | diri->dirfragtree.get_leaves_under(p->first.frag, ls); | |
4672 | if (ls.empty()) | |
4673 | ls.push_back(diri->dirfragtree[p->first.frag.value()]); | |
4674 | dout(10) << " maps to frag(s) " << ls << dendl; | |
4675 | for (list<frag_t>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
4676 | CDir *dir = diri->get_dirfrag(*q); | |
4677 | if (!dir) | |
4678 | dir = rejoin_invent_dirfrag(dirfrag_t(diri->ino(), *q)); | |
4679 | else | |
4680 | dout(10) << " have(approx) " << *dir << dendl; | |
4681 | dir->add_replica(from, p->second.nonce); | |
4682 | dir->dir_rep = p->second.dir_rep; | |
4683 | } | |
4684 | refragged = true; | |
4685 | } | |
4686 | ||
4687 | map<string_snap_t,MMDSCacheRejoin::dn_strong>& dmap = strong->strong_dentries[p->first]; | |
4688 | for (map<string_snap_t,MMDSCacheRejoin::dn_strong>::iterator q = dmap.begin(); | |
4689 | q != dmap.end(); | |
4690 | ++q) { | |
4691 | CDentry *dn; | |
4692 | if (!refragged) | |
4693 | dn = dir->lookup(q->first.name, q->first.snapid); | |
4694 | else { | |
4695 | frag_t fg = diri->pick_dirfrag(q->first.name); | |
4696 | dir = diri->get_dirfrag(fg); | |
4697 | assert(dir); | |
4698 | dn = dir->lookup(q->first.name, q->first.snapid); | |
4699 | } | |
4700 | if (!dn) { | |
4701 | if (q->second.is_remote()) { | |
4702 | dn = dir->add_remote_dentry(q->first.name, q->second.remote_ino, q->second.remote_d_type, | |
4703 | q->second.first, q->first.snapid); | |
4704 | } else if (q->second.is_null()) { | |
4705 | dn = dir->add_null_dentry(q->first.name, q->second.first, q->first.snapid); | |
4706 | } else { | |
4707 | CInode *in = get_inode(q->second.ino, q->first.snapid); | |
4708 | if (!in) in = rejoin_invent_inode(q->second.ino, q->first.snapid); | |
4709 | dn = dir->add_primary_dentry(q->first.name, in, q->second.first, q->first.snapid); | |
4710 | } | |
4711 | dout(10) << " invented " << *dn << dendl; | |
4712 | } | |
4713 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4714 | ||
4715 | // dn auth_pin? | |
4716 | if (strong->authpinned_dentries.count(p->first) && | |
4717 | strong->authpinned_dentries[p->first].count(q->first)) { | |
4718 | for (list<MMDSCacheRejoin::slave_reqid>::iterator r = strong->authpinned_dentries[p->first][q->first].begin(); | |
4719 | r != strong->authpinned_dentries[p->first][q->first].end(); | |
4720 | ++r) { | |
4721 | dout(10) << " dn authpin by " << *r << " on " << *dn << dendl; | |
4722 | ||
4723 | // get/create slave mdrequest | |
4724 | MDRequestRef mdr; | |
4725 | if (have_request(r->reqid)) | |
4726 | mdr = request_get(r->reqid); | |
4727 | else | |
4728 | mdr = request_start_slave(r->reqid, r->attempt, strong); | |
4729 | mdr->auth_pin(dn); | |
4730 | } | |
4731 | } | |
4732 | ||
4733 | // dn xlock? | |
4734 | if (strong->xlocked_dentries.count(p->first) && | |
4735 | strong->xlocked_dentries[p->first].count(q->first)) { | |
4736 | MMDSCacheRejoin::slave_reqid r = strong->xlocked_dentries[p->first][q->first]; | |
4737 | dout(10) << " dn xlock by " << r << " on " << *dn << dendl; | |
4738 | MDRequestRef mdr = request_get(r.reqid); // should have this from auth_pin above. | |
4739 | assert(mdr->is_auth_pinned(dn)); | |
4740 | if (!mdr->xlocks.count(&dn->versionlock)) { | |
4741 | assert(dn->versionlock.can_xlock_local()); | |
4742 | dn->versionlock.get_xlock(mdr, mdr->get_client()); | |
4743 | mdr->xlocks.insert(&dn->versionlock); | |
4744 | mdr->locks.insert(&dn->versionlock); | |
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->xlocks.insert(&dn->lock); | |
4751 | mdr->locks.insert(&dn->lock); | |
4752 | } | |
4753 | ||
4754 | dn->add_replica(from, q->second.nonce); | |
4755 | dout(10) << " have " << *dn << dendl; | |
4756 | ||
4757 | if (dnl->is_primary()) { | |
4758 | if (q->second.is_primary()) { | |
4759 | if (vinodeno_t(q->second.ino, q->first.snapid) != dnl->get_inode()->vino()) { | |
4760 | // the survivor missed MDentryUnlink+MDentryLink messages ? | |
4761 | assert(strong->strong_inodes.count(dnl->get_inode()->vino()) == 0); | |
4762 | CInode *in = get_inode(q->second.ino, q->first.snapid); | |
4763 | assert(in); | |
4764 | 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 | 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 (q->second.is_primary()) { | |
4775 | // the survivor missed MDentryUnlink message ? | |
4776 | CInode *in = get_inode(q->second.ino, q->first.snapid); | |
4777 | assert(in); | |
4778 | assert(in->get_parent_dn()); | |
4779 | rejoin_unlinked_inodes[from].insert(in); | |
4780 | dout(7) << " sender has primary dentry but we don't" << dendl; | |
4781 | } | |
4782 | } | |
4783 | } | |
4784 | } | |
4785 | ||
4786 | for (map<vinodeno_t, MMDSCacheRejoin::inode_strong>::iterator p = strong->strong_inodes.begin(); | |
4787 | p != strong->strong_inodes.end(); | |
4788 | ++p) { | |
4789 | CInode *in = get_inode(p->first); | |
4790 | assert(in); | |
4791 | in->add_replica(from, p->second.nonce); | |
4792 | dout(10) << " have " << *in << dendl; | |
4793 | ||
4794 | MMDSCacheRejoin::inode_strong &is = p->second; | |
4795 | ||
4796 | // caps_wanted | |
4797 | if (is.caps_wanted) { | |
4798 | in->mds_caps_wanted[from] = is.caps_wanted; | |
4799 | dout(15) << " inode caps_wanted " << ccap_string(is.caps_wanted) | |
4800 | << " on " << *in << dendl; | |
4801 | } | |
4802 | ||
4803 | // scatterlocks? | |
4804 | // infer state from replica state: | |
4805 | // * go to MIX if they might have wrlocks | |
4806 | // * go to LOCK if they are LOCK (just bc identify_files_to_recover might start twiddling filelock) | |
4807 | in->filelock.infer_state_from_strong_rejoin(is.filelock, !in->is_dir()); // maybe also go to LOCK | |
4808 | in->nestlock.infer_state_from_strong_rejoin(is.nestlock, false); | |
4809 | in->dirfragtreelock.infer_state_from_strong_rejoin(is.dftlock, false); | |
4810 | ||
4811 | // auth pin? | |
4812 | if (strong->authpinned_inodes.count(in->vino())) { | |
4813 | for (list<MMDSCacheRejoin::slave_reqid>::iterator r = strong->authpinned_inodes[in->vino()].begin(); | |
4814 | r != strong->authpinned_inodes[in->vino()].end(); | |
4815 | ++r) { | |
4816 | dout(10) << " inode authpin by " << *r << " on " << *in << dendl; | |
4817 | ||
4818 | // get/create slave mdrequest | |
4819 | MDRequestRef mdr; | |
4820 | if (have_request(r->reqid)) | |
4821 | mdr = request_get(r->reqid); | |
4822 | else | |
4823 | mdr = request_start_slave(r->reqid, r->attempt, strong); | |
4824 | if (strong->frozen_authpin_inodes.count(in->vino())) { | |
4825 | assert(!in->get_num_auth_pins()); | |
4826 | mdr->freeze_auth_pin(in); | |
4827 | } else { | |
4828 | assert(!in->is_frozen_auth_pin()); | |
4829 | } | |
4830 | mdr->auth_pin(in); | |
4831 | } | |
4832 | } | |
4833 | // xlock(s)? | |
4834 | if (strong->xlocked_inodes.count(in->vino())) { | |
4835 | for (map<int,MMDSCacheRejoin::slave_reqid>::iterator q = strong->xlocked_inodes[in->vino()].begin(); | |
4836 | q != strong->xlocked_inodes[in->vino()].end(); | |
4837 | ++q) { | |
4838 | SimpleLock *lock = in->get_lock(q->first); | |
4839 | dout(10) << " inode xlock by " << q->second << " on " << *lock << " on " << *in << dendl; | |
4840 | MDRequestRef mdr = request_get(q->second.reqid); // should have this from auth_pin above. | |
4841 | assert(mdr->is_auth_pinned(in)); | |
4842 | if (!mdr->xlocks.count(&in->versionlock)) { | |
4843 | assert(in->versionlock.can_xlock_local()); | |
4844 | in->versionlock.get_xlock(mdr, mdr->get_client()); | |
4845 | mdr->xlocks.insert(&in->versionlock); | |
4846 | mdr->locks.insert(&in->versionlock); | |
4847 | } | |
4848 | if (lock->is_stable()) | |
4849 | in->auth_pin(lock); | |
4850 | lock->set_state(LOCK_XLOCK); | |
4851 | if (lock == &in->filelock) | |
4852 | in->loner_cap = -1; | |
4853 | lock->get_xlock(mdr, mdr->get_client()); | |
4854 | mdr->xlocks.insert(lock); | |
4855 | mdr->locks.insert(lock); | |
4856 | } | |
4857 | } | |
4858 | } | |
4859 | // wrlock(s)? | |
4860 | for (map<vinodeno_t, map<int, list<MMDSCacheRejoin::slave_reqid> > >::iterator p = strong->wrlocked_inodes.begin(); | |
4861 | p != strong->wrlocked_inodes.end(); | |
4862 | ++p) { | |
4863 | CInode *in = get_inode(p->first); | |
4864 | for (map<int, list<MMDSCacheRejoin::slave_reqid> >::iterator q = p->second.begin(); | |
4865 | q != p->second.end(); | |
4866 | ++q) { | |
4867 | SimpleLock *lock = in->get_lock(q->first); | |
4868 | for (list<MMDSCacheRejoin::slave_reqid>::iterator r = q->second.begin(); | |
4869 | r != q->second.end(); | |
4870 | ++r) { | |
4871 | dout(10) << " inode wrlock by " << *r << " on " << *lock << " on " << *in << dendl; | |
4872 | MDRequestRef mdr = request_get(r->reqid); // should have this from auth_pin above. | |
4873 | if (in->is_auth()) | |
4874 | assert(mdr->is_auth_pinned(in)); | |
4875 | lock->set_state(LOCK_MIX); | |
4876 | if (lock == &in->filelock) | |
4877 | in->loner_cap = -1; | |
4878 | lock->get_wrlock(true); | |
4879 | mdr->wrlocks.insert(lock); | |
4880 | mdr->locks.insert(lock); | |
4881 | } | |
4882 | } | |
4883 | } | |
4884 | ||
4885 | // done? | |
4886 | assert(rejoin_gather.count(from)); | |
4887 | rejoin_gather.erase(from); | |
4888 | if (rejoin_gather.empty()) { | |
4889 | rejoin_gather_finish(); | |
4890 | } else { | |
4891 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" << dendl; | |
4892 | } | |
4893 | } | |
4894 | ||
4895 | /* This functions DOES NOT put the passed message before returning */ | |
4896 | void MDCache::handle_cache_rejoin_ack(MMDSCacheRejoin *ack) | |
4897 | { | |
4898 | dout(7) << "handle_cache_rejoin_ack from " << ack->get_source() << dendl; | |
4899 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
4900 | ||
4901 | // for sending cache expire message | |
4902 | set<CInode*> isolated_inodes; | |
4903 | set<CInode*> refragged_inodes; | |
4904 | ||
4905 | // dirs | |
4906 | for (map<dirfrag_t, MMDSCacheRejoin::dirfrag_strong>::iterator p = ack->strong_dirfrags.begin(); | |
4907 | p != ack->strong_dirfrags.end(); | |
4908 | ++p) { | |
4909 | // we may have had incorrect dir fragmentation; refragment based | |
4910 | // on what they auth tells us. | |
4911 | CDir *dir = get_dirfrag(p->first); | |
4912 | if (!dir) { | |
4913 | dir = get_force_dirfrag(p->first, false); | |
4914 | if (dir) | |
4915 | refragged_inodes.insert(dir->get_inode()); | |
4916 | } | |
4917 | if (!dir) { | |
4918 | CInode *diri = get_inode(p->first.ino); | |
4919 | if (!diri) { | |
4920 | // barebones inode; the full inode loop below will clean up. | |
4921 | diri = new CInode(this, false); | |
4922 | diri->inode.ino = p->first.ino; | |
4923 | diri->inode.mode = S_IFDIR; | |
4924 | diri->inode.dir_layout.dl_dir_hash = g_conf->mds_default_dir_hash; | |
4925 | add_inode(diri); | |
4926 | if (MDS_INO_MDSDIR(from) == p->first.ino) { | |
4927 | diri->inode_auth = mds_authority_t(from, CDIR_AUTH_UNKNOWN); | |
4928 | dout(10) << " add inode " << *diri << dendl; | |
4929 | } else { | |
4930 | diri->inode_auth = CDIR_AUTH_DEFAULT; | |
4931 | isolated_inodes.insert(diri); | |
4932 | dout(10) << " unconnected dirfrag " << p->first << dendl; | |
4933 | } | |
4934 | } | |
4935 | // barebones dirfrag; the full dirfrag loop below will clean up. | |
4936 | dir = diri->add_dirfrag(new CDir(diri, p->first.frag, this, false)); | |
4937 | if (MDS_INO_MDSDIR(from) == p->first.ino || | |
4938 | (dir->authority() != CDIR_AUTH_UNDEF && | |
4939 | dir->authority().first != from)) | |
4940 | adjust_subtree_auth(dir, from); | |
4941 | dout(10) << " add dirfrag " << *dir << dendl; | |
4942 | } | |
4943 | ||
4944 | dir->set_replica_nonce(p->second.nonce); | |
4945 | dir->state_clear(CDir::STATE_REJOINING); | |
4946 | dout(10) << " got " << *dir << dendl; | |
4947 | ||
4948 | // dentries | |
4949 | map<string_snap_t,MMDSCacheRejoin::dn_strong>& dmap = ack->strong_dentries[p->first]; | |
4950 | for (map<string_snap_t,MMDSCacheRejoin::dn_strong>::iterator q = dmap.begin(); | |
4951 | q != dmap.end(); | |
4952 | ++q) { | |
4953 | CDentry *dn = dir->lookup(q->first.name, q->first.snapid); | |
4954 | if(!dn) | |
4955 | dn = dir->add_null_dentry(q->first.name, q->second.first, q->first.snapid); | |
4956 | ||
4957 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
4958 | ||
4959 | assert(dn->last == q->first.snapid); | |
4960 | if (dn->first != q->second.first) { | |
4961 | dout(10) << " adjust dn.first " << dn->first << " -> " << q->second.first << " on " << *dn << dendl; | |
4962 | dn->first = q->second.first; | |
4963 | } | |
4964 | ||
4965 | // may have bad linkage if we missed dentry link/unlink messages | |
4966 | if (dnl->is_primary()) { | |
4967 | CInode *in = dnl->get_inode(); | |
4968 | if (!q->second.is_primary() || | |
4969 | vinodeno_t(q->second.ino, q->first.snapid) != in->vino()) { | |
4970 | dout(10) << " had bad linkage for " << *dn << ", unlinking " << *in << dendl; | |
4971 | dir->unlink_inode(dn); | |
4972 | } | |
4973 | } else if (dnl->is_remote()) { | |
4974 | if (!q->second.is_remote() || | |
4975 | q->second.remote_ino != dnl->get_remote_ino() || | |
4976 | q->second.remote_d_type != dnl->get_remote_d_type()) { | |
4977 | dout(10) << " had bad linkage for " << *dn << dendl; | |
4978 | dir->unlink_inode(dn); | |
4979 | } | |
4980 | } else { | |
4981 | if (!q->second.is_null()) | |
4982 | dout(10) << " had bad linkage for " << *dn << dendl; | |
4983 | } | |
4984 | ||
4985 | // hmm, did we have the proper linkage here? | |
4986 | if (dnl->is_null() && !q->second.is_null()) { | |
4987 | if (q->second.is_remote()) { | |
4988 | dn->dir->link_remote_inode(dn, q->second.remote_ino, q->second.remote_d_type); | |
4989 | } else { | |
4990 | CInode *in = get_inode(q->second.ino, q->first.snapid); | |
4991 | if (!in) { | |
4992 | // barebones inode; assume it's dir, the full inode loop below will clean up. | |
4993 | in = new CInode(this, false, q->second.first, q->first.snapid); | |
4994 | in->inode.ino = q->second.ino; | |
4995 | in->inode.mode = S_IFDIR; | |
4996 | in->inode.dir_layout.dl_dir_hash = g_conf->mds_default_dir_hash; | |
4997 | add_inode(in); | |
4998 | dout(10) << " add inode " << *in << dendl; | |
4999 | } else if (in->get_parent_dn()) { | |
5000 | dout(10) << " had bad linkage for " << *(in->get_parent_dn()) | |
5001 | << ", unlinking " << *in << dendl; | |
5002 | in->get_parent_dir()->unlink_inode(in->get_parent_dn()); | |
5003 | } | |
5004 | dn->dir->link_primary_inode(dn, in); | |
5005 | isolated_inodes.erase(in); | |
5006 | } | |
5007 | } | |
5008 | ||
5009 | dn->set_replica_nonce(q->second.nonce); | |
5010 | dn->lock.set_state_rejoin(q->second.lock, rejoin_waiters); | |
5011 | dn->state_clear(CDentry::STATE_REJOINING); | |
5012 | dout(10) << " got " << *dn << dendl; | |
5013 | } | |
5014 | } | |
5015 | ||
5016 | for (set<CInode*>::iterator p = refragged_inodes.begin(); | |
5017 | p != refragged_inodes.end(); | |
5018 | ++p) { | |
5019 | list<CDir*> ls; | |
5020 | (*p)->get_nested_dirfrags(ls); | |
5021 | for (list<CDir*>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
5022 | if ((*q)->is_auth() || ack->strong_dirfrags.count((*q)->dirfrag())) | |
5023 | continue; | |
5024 | assert((*q)->get_num_any() == 0); | |
5025 | (*p)->close_dirfrag((*q)->get_frag()); | |
5026 | } | |
5027 | } | |
5028 | ||
5029 | // full dirfrags | |
5030 | for (map<dirfrag_t, bufferlist>::iterator p = ack->dirfrag_bases.begin(); | |
5031 | p != ack->dirfrag_bases.end(); | |
5032 | ++p) { | |
5033 | CDir *dir = get_dirfrag(p->first); | |
5034 | assert(dir); | |
5035 | bufferlist::iterator q = p->second.begin(); | |
5036 | dir->_decode_base(q); | |
5037 | dout(10) << " got dir replica " << *dir << dendl; | |
5038 | } | |
5039 | ||
5040 | // full inodes | |
5041 | bufferlist::iterator p = ack->inode_base.begin(); | |
5042 | while (!p.end()) { | |
5043 | inodeno_t ino; | |
5044 | snapid_t last; | |
5045 | bufferlist basebl; | |
5046 | ::decode(ino, p); | |
5047 | ::decode(last, p); | |
5048 | ::decode(basebl, p); | |
5049 | CInode *in = get_inode(ino, last); | |
5050 | assert(in); | |
5051 | bufferlist::iterator q = basebl.begin(); | |
5052 | in->_decode_base(q); | |
5053 | dout(10) << " got inode base " << *in << dendl; | |
5054 | } | |
5055 | ||
5056 | // inodes | |
5057 | p = ack->inode_locks.begin(); | |
5058 | //dout(10) << "inode_locks len " << ack->inode_locks.length() << " is " << ack->inode_locks << dendl; | |
5059 | while (!p.end()) { | |
5060 | inodeno_t ino; | |
5061 | snapid_t last; | |
5062 | __u32 nonce; | |
5063 | bufferlist lockbl; | |
5064 | ::decode(ino, p); | |
5065 | ::decode(last, p); | |
5066 | ::decode(nonce, p); | |
5067 | ::decode(lockbl, p); | |
5068 | ||
5069 | CInode *in = get_inode(ino, last); | |
5070 | assert(in); | |
5071 | in->set_replica_nonce(nonce); | |
5072 | bufferlist::iterator q = lockbl.begin(); | |
5073 | in->_decode_locks_rejoin(q, rejoin_waiters, rejoin_eval_locks); | |
5074 | in->state_clear(CInode::STATE_REJOINING); | |
5075 | dout(10) << " got inode locks " << *in << dendl; | |
5076 | } | |
5077 | ||
5078 | // FIXME: This can happen if entire subtree, together with the inode subtree root | |
5079 | // belongs to, were trimmed between sending cache rejoin and receiving rejoin ack. | |
5080 | assert(isolated_inodes.empty()); | |
5081 | ||
5082 | map<inodeno_t,map<client_t,Capability::Import> > peer_imported; | |
5083 | bufferlist::iterator bp = ack->imported_caps.begin(); | |
5084 | ::decode(peer_imported, bp); | |
5085 | ||
5086 | for (map<inodeno_t,map<client_t,Capability::Import> >::iterator p = peer_imported.begin(); | |
5087 | p != peer_imported.end(); | |
5088 | ++p) { | |
5089 | assert(cap_exports.count(p->first)); | |
5090 | assert(cap_export_targets.count(p->first)); | |
5091 | assert(cap_export_targets[p->first] == from); | |
5092 | for (map<client_t,Capability::Import>::iterator q = p->second.begin(); | |
5093 | q != p->second.end(); | |
5094 | ++q) { | |
5095 | assert(cap_exports[p->first].count(q->first)); | |
5096 | ||
5097 | dout(10) << " exporting caps for client." << q->first << " ino " << p->first << dendl; | |
5098 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
5099 | assert(session); | |
5100 | ||
5101 | // mark client caps stale. | |
5102 | MClientCaps *m = new MClientCaps(CEPH_CAP_OP_EXPORT, p->first, 0, | |
5103 | cap_exports[p->first][q->first].capinfo.cap_id, 0, | |
5104 | mds->get_osd_epoch_barrier()); | |
5105 | m->set_cap_peer(q->second.cap_id, q->second.issue_seq, q->second.mseq, | |
5106 | (q->second.cap_id > 0 ? from : -1), 0); | |
5107 | mds->send_message_client_counted(m, session); | |
5108 | ||
5109 | cap_exports[p->first].erase(q->first); | |
5110 | } | |
5111 | assert(cap_exports[p->first].empty()); | |
5112 | } | |
5113 | ||
5114 | // done? | |
5115 | assert(rejoin_ack_gather.count(from)); | |
5116 | rejoin_ack_gather.erase(from); | |
5117 | if (mds->is_rejoin()) { | |
5118 | ||
5119 | if (rejoin_gather.empty()) { | |
5120 | // eval unstable scatter locks after all wrlocks are rejoined. | |
5121 | while (!rejoin_eval_locks.empty()) { | |
5122 | SimpleLock *lock = rejoin_eval_locks.front(); | |
5123 | rejoin_eval_locks.pop_front(); | |
5124 | if (!lock->is_stable()) | |
5125 | mds->locker->eval_gather(lock); | |
5126 | } | |
5127 | } | |
5128 | ||
5129 | if (rejoin_gather.empty() && // make sure we've gotten our FULL inodes, too. | |
5130 | rejoin_ack_gather.empty()) { | |
5131 | // finally, kickstart past snap parent opens | |
5132 | open_snap_parents(); | |
5133 | } else { | |
5134 | dout(7) << "still need rejoin from (" << rejoin_gather << ")" | |
5135 | << ", rejoin_ack from (" << rejoin_ack_gather << ")" << dendl; | |
5136 | } | |
5137 | } else { | |
5138 | // survivor. | |
5139 | mds->queue_waiters(rejoin_waiters); | |
5140 | } | |
5141 | } | |
5142 | ||
5143 | /** | |
5144 | * rejoin_trim_undef_inodes() -- remove REJOINUNDEF flagged inodes | |
5145 | * | |
5146 | * FIXME: wait, can this actually happen? a survivor should generate cache trim | |
5147 | * messages that clean these guys up... | |
5148 | */ | |
5149 | void MDCache::rejoin_trim_undef_inodes() | |
5150 | { | |
5151 | dout(10) << "rejoin_trim_undef_inodes" << dendl; | |
5152 | ||
5153 | while (!rejoin_undef_inodes.empty()) { | |
5154 | set<CInode*>::iterator p = rejoin_undef_inodes.begin(); | |
5155 | CInode *in = *p; | |
5156 | rejoin_undef_inodes.erase(p); | |
5157 | ||
5158 | in->clear_replica_map(); | |
5159 | ||
5160 | // close out dirfrags | |
5161 | if (in->is_dir()) { | |
5162 | list<CDir*> dfls; | |
5163 | in->get_dirfrags(dfls); | |
5164 | for (list<CDir*>::iterator p = dfls.begin(); | |
5165 | p != dfls.end(); | |
5166 | ++p) { | |
5167 | CDir *dir = *p; | |
5168 | dir->clear_replica_map(); | |
5169 | ||
5170 | for (CDir::map_t::iterator p = dir->items.begin(); | |
5171 | p != dir->items.end(); | |
5172 | ++p) { | |
5173 | CDentry *dn = p->second; | |
5174 | dn->clear_replica_map(); | |
5175 | ||
5176 | dout(10) << " trimming " << *dn << dendl; | |
5177 | dir->remove_dentry(dn); | |
5178 | } | |
5179 | ||
5180 | dout(10) << " trimming " << *dir << dendl; | |
5181 | in->close_dirfrag(dir->dirfrag().frag); | |
5182 | } | |
5183 | } | |
5184 | ||
5185 | CDentry *dn = in->get_parent_dn(); | |
5186 | if (dn) { | |
5187 | dn->clear_replica_map(); | |
5188 | dout(10) << " trimming " << *dn << dendl; | |
5189 | dn->dir->remove_dentry(dn); | |
5190 | } else { | |
5191 | dout(10) << " trimming " << *in << dendl; | |
5192 | remove_inode(in); | |
5193 | } | |
5194 | } | |
5195 | ||
5196 | assert(rejoin_undef_inodes.empty()); | |
5197 | } | |
5198 | ||
5199 | void MDCache::rejoin_gather_finish() | |
5200 | { | |
5201 | dout(10) << "rejoin_gather_finish" << dendl; | |
5202 | assert(mds->is_rejoin()); | |
5203 | ||
5204 | if (open_undef_inodes_dirfrags()) | |
5205 | return; | |
5206 | ||
5207 | if (process_imported_caps()) | |
5208 | return; | |
5209 | ||
5210 | choose_lock_states_and_reconnect_caps(); | |
5211 | ||
5212 | identify_files_to_recover(); | |
5213 | rejoin_send_acks(); | |
5214 | ||
5215 | // signal completion of fetches, rejoin_gather_finish, etc. | |
5216 | assert(rejoin_ack_gather.count(mds->get_nodeid())); | |
5217 | rejoin_ack_gather.erase(mds->get_nodeid()); | |
5218 | ||
5219 | // did we already get our acks too? | |
5220 | if (rejoin_ack_gather.empty()) { | |
5221 | // finally, kickstart past snap parent opens | |
5222 | open_snap_parents(); | |
5223 | } | |
5224 | } | |
5225 | ||
5226 | class C_MDC_RejoinOpenInoFinish: public MDCacheContext { | |
5227 | inodeno_t ino; | |
5228 | public: | |
5229 | C_MDC_RejoinOpenInoFinish(MDCache *c, inodeno_t i) : MDCacheContext(c), ino(i) {} | |
5230 | void finish(int r) override { | |
5231 | mdcache->rejoin_open_ino_finish(ino, r); | |
5232 | } | |
5233 | }; | |
5234 | ||
5235 | void MDCache::rejoin_open_ino_finish(inodeno_t ino, int ret) | |
5236 | { | |
5237 | dout(10) << "open_caps_inode_finish ino " << ino << " ret " << ret << dendl; | |
5238 | ||
5239 | if (ret < 0) { | |
5240 | cap_imports_missing.insert(ino); | |
5241 | } else if (ret == mds->get_nodeid()) { | |
5242 | assert(get_inode(ino)); | |
5243 | } else { | |
5244 | auto p = cap_imports.find(ino); | |
5245 | assert(p != cap_imports.end()); | |
5246 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
5247 | assert(q->second.count(MDS_RANK_NONE)); | |
5248 | assert(q->second.size() == 1); | |
5249 | rejoin_export_caps(p->first, q->first, q->second[MDS_RANK_NONE], ret); | |
5250 | } | |
5251 | cap_imports.erase(p); | |
5252 | } | |
5253 | ||
5254 | assert(cap_imports_num_opening > 0); | |
5255 | cap_imports_num_opening--; | |
5256 | ||
5257 | if (cap_imports_num_opening == 0) { | |
5258 | if (rejoin_gather.empty()) | |
5259 | rejoin_gather_finish(); | |
5260 | else if (rejoin_gather.count(mds->get_nodeid())) | |
5261 | process_imported_caps(); | |
5262 | } | |
5263 | } | |
5264 | ||
5265 | class C_MDC_RejoinSessionsOpened : public MDCacheLogContext { | |
5266 | public: | |
5267 | map<client_t,entity_inst_t> client_map; | |
5268 | map<client_t,uint64_t> sseqmap; | |
5269 | ||
5270 | C_MDC_RejoinSessionsOpened(MDCache *c, map<client_t,entity_inst_t>& cm) : | |
5271 | MDCacheLogContext(c), client_map(cm) {} | |
5272 | void finish(int r) override { | |
5273 | assert(r == 0); | |
5274 | mdcache->rejoin_open_sessions_finish(client_map, sseqmap); | |
5275 | } | |
5276 | }; | |
5277 | ||
5278 | void MDCache::rejoin_open_sessions_finish(map<client_t,entity_inst_t> client_map, | |
5279 | map<client_t,uint64_t>& sseqmap) | |
5280 | { | |
5281 | dout(10) << "rejoin_open_sessions_finish" << dendl; | |
5282 | mds->server->finish_force_open_sessions(client_map, sseqmap); | |
5283 | if (rejoin_gather.empty()) | |
5284 | rejoin_gather_finish(); | |
5285 | } | |
5286 | ||
5287 | bool MDCache::process_imported_caps() | |
5288 | { | |
5289 | dout(10) << "process_imported_caps" << dendl; | |
5290 | ||
5291 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ++p) { | |
5292 | CInode *in = get_inode(p->first); | |
5293 | if (in) { | |
5294 | assert(in->is_auth()); | |
5295 | cap_imports_missing.erase(p->first); | |
5296 | continue; | |
5297 | } | |
5298 | if (cap_imports_missing.count(p->first) > 0) | |
5299 | continue; | |
5300 | ||
5301 | cap_imports_num_opening++; | |
5302 | dout(10) << " opening missing ino " << p->first << dendl; | |
5303 | open_ino(p->first, (int64_t)-1, new C_MDC_RejoinOpenInoFinish(this, p->first), false); | |
5304 | } | |
5305 | ||
5306 | if (cap_imports_num_opening > 0) | |
5307 | return true; | |
5308 | ||
5309 | // called by rejoin_gather_finish() ? | |
5310 | if (rejoin_gather.count(mds->get_nodeid()) == 0) { | |
5311 | // if sessions for imported caps are all open ? | |
5312 | for (map<client_t,entity_inst_t>::iterator p = rejoin_client_map.begin(); | |
5313 | p != rejoin_client_map.end(); | |
5314 | ++p) { | |
5315 | if (!mds->sessionmap.have_session(entity_name_t::CLIENT(p->first.v))) { | |
5316 | C_MDC_RejoinSessionsOpened *finish = new C_MDC_RejoinSessionsOpened(this, rejoin_client_map); | |
5317 | version_t pv = mds->server->prepare_force_open_sessions(rejoin_client_map, finish->sseqmap); | |
5318 | ESessions *le = new ESessions(pv, rejoin_client_map); | |
5319 | mds->mdlog->start_submit_entry(le, finish); | |
5320 | mds->mdlog->flush(); | |
5321 | rejoin_client_map.clear(); | |
5322 | return true; | |
5323 | } | |
5324 | } | |
5325 | rejoin_client_map.clear(); | |
5326 | ||
5327 | // process caps that were exported by slave rename | |
5328 | for (map<inodeno_t,pair<mds_rank_t,map<client_t,Capability::Export> > >::iterator p = rejoin_slave_exports.begin(); | |
5329 | p != rejoin_slave_exports.end(); | |
5330 | ++p) { | |
5331 | CInode *in = get_inode(p->first); | |
5332 | assert(in); | |
5333 | for (map<client_t,Capability::Export>::iterator q = p->second.second.begin(); | |
5334 | q != p->second.second.end(); | |
5335 | ++q) { | |
5336 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
5337 | assert(session); | |
5338 | ||
5339 | Capability *cap = in->get_client_cap(q->first); | |
5340 | if (!cap) | |
5341 | cap = in->add_client_cap(q->first, session); | |
5342 | cap->merge(q->second, true); | |
5343 | ||
5344 | Capability::Import& im = rejoin_imported_caps[p->second.first][p->first][q->first]; | |
5345 | assert(cap->get_last_seq() == im.issue_seq); | |
5346 | assert(cap->get_mseq() == im.mseq); | |
5347 | cap->set_cap_id(im.cap_id); | |
5348 | // send cap import because we assigned a new cap ID | |
5349 | do_cap_import(session, in, cap, q->second.cap_id, q->second.seq, q->second.mseq - 1, | |
5350 | p->second.first, CEPH_CAP_FLAG_AUTH); | |
5351 | } | |
5352 | } | |
5353 | rejoin_slave_exports.clear(); | |
5354 | rejoin_imported_caps.clear(); | |
5355 | ||
5356 | // process cap imports | |
5357 | // ino -> client -> frommds -> capex | |
5358 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ) { | |
5359 | CInode *in = get_inode(p->first); | |
5360 | if (!in) { | |
5361 | dout(10) << " still missing ino " << p->first | |
5362 | << ", will try again after replayed client requests" << dendl; | |
5363 | ++p; | |
5364 | continue; | |
5365 | } | |
5366 | assert(in->is_auth()); | |
5367 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
5368 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
5369 | assert(session); | |
5370 | for (auto r = q->second.begin(); r != q->second.end(); ++r) { | |
5371 | Capability *cap = in->reconnect_cap(q->first, r->second, session); | |
5372 | add_reconnected_cap(q->first, in->ino(), r->second); | |
5373 | if (r->first >= 0) { | |
5374 | if (cap->get_last_seq() == 0) // don't increase mseq if cap already exists | |
5375 | cap->inc_mseq(); | |
5376 | do_cap_import(session, in, cap, r->second.capinfo.cap_id, 0, 0, r->first, 0); | |
5377 | ||
5378 | Capability::Import& im = rejoin_imported_caps[r->first][p->first][q->first]; | |
5379 | im.cap_id = cap->get_cap_id(); | |
5380 | im.issue_seq = cap->get_last_seq(); | |
5381 | im.mseq = cap->get_mseq(); | |
5382 | } | |
5383 | } | |
5384 | } | |
5385 | cap_imports.erase(p++); // remove and move on | |
5386 | } | |
5387 | } else { | |
5388 | trim_non_auth(); | |
5389 | ||
5390 | rejoin_gather.erase(mds->get_nodeid()); | |
5391 | maybe_send_pending_rejoins(); | |
5392 | ||
5393 | if (rejoin_gather.empty() && rejoin_ack_gather.count(mds->get_nodeid())) | |
5394 | rejoin_gather_finish(); | |
5395 | } | |
5396 | return false; | |
5397 | } | |
5398 | ||
5399 | void MDCache::check_realm_past_parents(SnapRealm *realm, bool reconnect) | |
5400 | { | |
5401 | // are this realm's parents fully open? | |
5402 | if (realm->have_past_parents_open()) { | |
5403 | dout(10) << " have past snap parents for realm " << *realm | |
5404 | << " on " << *realm->inode << dendl; | |
5405 | if (reconnect) { | |
5406 | // finish off client snaprealm reconnects? | |
5407 | auto p = reconnected_snaprealms.find(realm->inode->ino()); | |
5408 | if (p != reconnected_snaprealms.end()) { | |
5409 | for (auto q = p->second.begin(); q != p->second.end(); ++q) | |
5410 | finish_snaprealm_reconnect(q->first, realm, q->second); | |
5411 | reconnected_snaprealms.erase(p); | |
5412 | } | |
5413 | } | |
5414 | } else { | |
5415 | if (!missing_snap_parents.count(realm->inode)) { | |
5416 | dout(10) << " MISSING past snap parents for realm " << *realm | |
5417 | << " on " << *realm->inode << dendl; | |
5418 | realm->inode->get(CInode::PIN_OPENINGSNAPPARENTS); | |
5419 | missing_snap_parents[realm->inode].size(); // just to get it into the map! | |
5420 | } else { | |
5421 | dout(10) << " (already) MISSING past snap parents for realm " << *realm | |
5422 | << " on " << *realm->inode << dendl; | |
5423 | } | |
5424 | } | |
5425 | } | |
5426 | ||
5427 | void MDCache::rebuild_need_snapflush(CInode *head_in, SnapRealm *realm, | |
5428 | client_t client, snapid_t snap_follows) | |
5429 | { | |
5430 | dout(10) << "rebuild_need_snapflush " << snap_follows << " on " << *head_in << dendl; | |
5431 | ||
5432 | const set<snapid_t>& snaps = realm->get_snaps(); | |
5433 | snapid_t follows = snap_follows; | |
5434 | ||
5435 | while (true) { | |
5436 | CInode *in = pick_inode_snap(head_in, follows); | |
5437 | if (in == head_in) | |
5438 | break; | |
5439 | dout(10) << " need snapflush from client." << client << " on " << *in << dendl; | |
5440 | ||
5441 | /* TODO: we can check the reconnected/flushing caps to find | |
5442 | * which locks need gathering */ | |
5443 | for (int i = 0; i < num_cinode_locks; i++) { | |
5444 | int lockid = cinode_lock_info[i].lock; | |
5445 | SimpleLock *lock = in->get_lock(lockid); | |
5446 | assert(lock); | |
5447 | in->client_snap_caps[lockid].insert(client); | |
5448 | in->auth_pin(lock); | |
5449 | lock->set_state(LOCK_SNAP_SYNC); | |
5450 | lock->get_wrlock(true); | |
5451 | } | |
5452 | ||
5453 | for (auto p = snaps.lower_bound(in->first); | |
5454 | p != snaps.end() && *p <= in->last; | |
5455 | ++p) { | |
5456 | head_in->add_need_snapflush(in, *p, client); | |
5457 | } | |
5458 | ||
5459 | follows = in->last; | |
5460 | } | |
5461 | } | |
5462 | ||
5463 | /* | |
5464 | * choose lock states based on reconnected caps | |
5465 | */ | |
5466 | void MDCache::choose_lock_states_and_reconnect_caps() | |
5467 | { | |
5468 | dout(10) << "choose_lock_states_and_reconnect_caps" << dendl; | |
5469 | ||
5470 | map<client_t,MClientSnap*> splits; | |
5471 | ||
5472 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator i = inode_map.begin(); | |
5473 | i != inode_map.end(); | |
5474 | ++i) { | |
5475 | CInode *in = i->second; | |
5476 | ||
5477 | if (in->last != CEPH_NOSNAP) | |
5478 | continue; | |
5479 | ||
5480 | if (in->is_auth() && !in->is_base() && in->inode.is_dirty_rstat()) | |
5481 | in->mark_dirty_rstat(); | |
5482 | ||
5483 | auto p = reconnected_caps.find(in->ino()); | |
5484 | ||
5485 | int dirty_caps = 0; | |
5486 | if (p != reconnected_caps.end()) { | |
5487 | for (const auto &it : p->second) | |
5488 | dirty_caps |= it.second.dirty_caps; | |
5489 | } | |
5490 | in->choose_lock_states(dirty_caps); | |
5491 | dout(15) << " chose lock states on " << *in << dendl; | |
5492 | ||
5493 | SnapRealm *realm = in->find_snaprealm(); | |
5494 | ||
5495 | check_realm_past_parents(realm, realm == in->snaprealm); | |
5496 | ||
5497 | if (p != reconnected_caps.end()) { | |
5498 | bool missing_snap_parent = false; | |
5499 | // also, make sure client's cap is in the correct snaprealm. | |
5500 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
5501 | if (q->second.snap_follows > 0 && q->second.snap_follows < in->first - 1) { | |
5502 | if (realm->have_past_parents_open()) { | |
5503 | rebuild_need_snapflush(in, realm, q->first, q->second.snap_follows); | |
5504 | } else { | |
5505 | missing_snap_parent = true; | |
5506 | } | |
5507 | } | |
5508 | ||
5509 | if (q->second.realm_ino == realm->inode->ino()) { | |
5510 | dout(15) << " client." << q->first << " has correct realm " << q->second.realm_ino << dendl; | |
5511 | } else { | |
5512 | dout(15) << " client." << q->first << " has wrong realm " << q->second.realm_ino | |
5513 | << " != " << realm->inode->ino() << dendl; | |
5514 | if (realm->have_past_parents_open()) { | |
5515 | // ok, include in a split message _now_. | |
5516 | prepare_realm_split(realm, q->first, in->ino(), splits); | |
5517 | } else { | |
5518 | // send the split later. | |
5519 | missing_snap_parent = true; | |
5520 | } | |
5521 | } | |
5522 | } | |
5523 | if (missing_snap_parent) | |
5524 | missing_snap_parents[realm->inode].insert(in); | |
5525 | } | |
5526 | } | |
5527 | ||
5528 | send_snaps(splits); | |
5529 | } | |
5530 | ||
5531 | void MDCache::prepare_realm_split(SnapRealm *realm, client_t client, inodeno_t ino, | |
5532 | map<client_t,MClientSnap*>& splits) | |
5533 | { | |
5534 | MClientSnap *snap; | |
5535 | if (splits.count(client) == 0) { | |
5536 | splits[client] = snap = new MClientSnap(CEPH_SNAP_OP_SPLIT); | |
5537 | snap->head.split = realm->inode->ino(); | |
5538 | realm->build_snap_trace(snap->bl); | |
5539 | ||
5540 | for (set<SnapRealm*>::iterator p = realm->open_children.begin(); | |
5541 | p != realm->open_children.end(); | |
5542 | ++p) | |
5543 | snap->split_realms.push_back((*p)->inode->ino()); | |
5544 | ||
5545 | } else | |
5546 | snap = splits[client]; | |
5547 | snap->split_inos.push_back(ino); | |
5548 | } | |
5549 | ||
5550 | void MDCache::send_snaps(map<client_t,MClientSnap*>& splits) | |
5551 | { | |
5552 | dout(10) << "send_snaps" << dendl; | |
5553 | ||
5554 | for (map<client_t,MClientSnap*>::iterator p = splits.begin(); | |
5555 | p != splits.end(); | |
5556 | ++p) { | |
5557 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(p->first.v)); | |
5558 | if (session) { | |
5559 | dout(10) << " client." << p->first | |
5560 | << " split " << p->second->head.split | |
5561 | << " inos " << p->second->split_inos | |
5562 | << dendl; | |
5563 | mds->send_message_client_counted(p->second, session); | |
5564 | } else { | |
5565 | dout(10) << " no session for client." << p->first << dendl; | |
5566 | p->second->put(); | |
5567 | } | |
5568 | } | |
5569 | splits.clear(); | |
5570 | } | |
5571 | ||
5572 | ||
5573 | /* | |
5574 | * remove any items from logsegment open_file lists that don't have | |
5575 | * any caps | |
5576 | */ | |
5577 | void MDCache::clean_open_file_lists() | |
5578 | { | |
5579 | dout(10) << "clean_open_file_lists" << dendl; | |
5580 | ||
5581 | for (map<uint64_t,LogSegment*>::iterator p = mds->mdlog->segments.begin(); | |
5582 | p != mds->mdlog->segments.end(); | |
5583 | ++p) { | |
5584 | LogSegment *ls = p->second; | |
5585 | ||
5586 | elist<CInode*>::iterator q = ls->open_files.begin(member_offset(CInode, item_open_file)); | |
5587 | while (!q.end()) { | |
5588 | CInode *in = *q; | |
5589 | ++q; | |
5590 | if (in->last == CEPH_NOSNAP) { | |
5591 | if (!in->is_any_caps_wanted()) { | |
5592 | dout(10) << " unlisting unwanted/capless inode " << *in << dendl; | |
5593 | in->item_open_file.remove_myself(); | |
5594 | } | |
5595 | } else if (in->last != CEPH_NOSNAP) { | |
5596 | if (in->client_snap_caps.empty()) { | |
5597 | dout(10) << " unlisting flushed snap inode " << *in << dendl; | |
5598 | in->item_open_file.remove_myself(); | |
5599 | } | |
5600 | } | |
5601 | } | |
5602 | } | |
5603 | } | |
5604 | ||
5605 | ||
5606 | ||
5607 | Capability* MDCache::rejoin_import_cap(CInode *in, client_t client, const cap_reconnect_t& icr, mds_rank_t frommds) | |
5608 | { | |
5609 | dout(10) << "rejoin_import_cap for client." << client << " from mds." << frommds | |
5610 | << " on " << *in << dendl; | |
5611 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(client.v)); | |
5612 | if (!session) { | |
5613 | dout(10) << " no session for client." << client << dendl; | |
5614 | return NULL; | |
5615 | } | |
5616 | ||
5617 | Capability *cap = in->reconnect_cap(client, icr, session); | |
5618 | ||
5619 | if (frommds >= 0) { | |
5620 | if (cap->get_last_seq() == 0) // don't increase mseq if cap already exists | |
5621 | cap->inc_mseq(); | |
5622 | do_cap_import(session, in, cap, icr.capinfo.cap_id, 0, 0, frommds, 0); | |
5623 | } | |
5624 | ||
5625 | return cap; | |
5626 | } | |
5627 | ||
5628 | void MDCache::export_remaining_imported_caps() | |
5629 | { | |
5630 | dout(10) << "export_remaining_imported_caps" << dendl; | |
5631 | ||
5632 | stringstream warn_str; | |
5633 | ||
5634 | for (auto p = cap_imports.begin(); p != cap_imports.end(); ++p) { | |
5635 | warn_str << " ino " << p->first << "\n"; | |
5636 | for (auto q = p->second.begin(); q != p->second.end(); ++q) { | |
5637 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(q->first.v)); | |
5638 | if (session) { | |
5639 | // mark client caps stale. | |
5640 | MClientCaps *stale = new MClientCaps(CEPH_CAP_OP_EXPORT, p->first, 0, 0, 0, mds->get_osd_epoch_barrier()); | |
5641 | stale->set_cap_peer(0, 0, 0, -1, 0); | |
5642 | mds->send_message_client_counted(stale, q->first); | |
5643 | } | |
5644 | } | |
5645 | ||
5646 | mds->heartbeat_reset(); | |
5647 | } | |
5648 | ||
5649 | for (map<inodeno_t, list<MDSInternalContextBase*> >::iterator p = cap_reconnect_waiters.begin(); | |
5650 | p != cap_reconnect_waiters.end(); | |
5651 | ++p) | |
5652 | mds->queue_waiters(p->second); | |
5653 | ||
5654 | cap_imports.clear(); | |
5655 | cap_reconnect_waiters.clear(); | |
5656 | ||
5657 | if (warn_str.peek() != EOF) { | |
5658 | mds->clog->warn() << "failed to reconnect caps for missing inodes:"; | |
5659 | mds->clog->warn(warn_str); | |
5660 | } | |
5661 | } | |
5662 | ||
5663 | void MDCache::try_reconnect_cap(CInode *in, Session *session) | |
5664 | { | |
5665 | client_t client = session->info.get_client(); | |
5666 | const cap_reconnect_t *rc = get_replay_cap_reconnect(in->ino(), client); | |
5667 | if (rc) { | |
5668 | in->reconnect_cap(client, *rc, session); | |
5669 | dout(10) << "try_reconnect_cap client." << client | |
5670 | << " reconnect wanted " << ccap_string(rc->capinfo.wanted) | |
5671 | << " issue " << ccap_string(rc->capinfo.issued) | |
5672 | << " on " << *in << dendl; | |
5673 | remove_replay_cap_reconnect(in->ino(), client); | |
5674 | ||
5675 | if (in->is_replicated()) { | |
5676 | mds->locker->try_eval(in, CEPH_CAP_LOCKS); | |
5677 | } else { | |
5678 | int dirty_caps = 0; | |
5679 | auto p = reconnected_caps.find(in->ino()); | |
5680 | if (p != reconnected_caps.end()) { | |
5681 | auto q = p->second.find(client); | |
5682 | if (q != p->second.end()) | |
5683 | dirty_caps = q->second.dirty_caps; | |
5684 | } | |
5685 | in->choose_lock_states(dirty_caps); | |
5686 | dout(15) << " chose lock states on " << *in << dendl; | |
5687 | } | |
5688 | ||
5689 | map<inodeno_t, list<MDSInternalContextBase*> >::iterator it = | |
5690 | cap_reconnect_waiters.find(in->ino()); | |
5691 | if (it != cap_reconnect_waiters.end()) { | |
5692 | mds->queue_waiters(it->second); | |
5693 | cap_reconnect_waiters.erase(it); | |
5694 | } | |
5695 | } | |
5696 | } | |
5697 | ||
5698 | ||
5699 | ||
5700 | // ------- | |
5701 | // cap imports and delayed snap parent opens | |
5702 | ||
5703 | void MDCache::do_cap_import(Session *session, CInode *in, Capability *cap, | |
5704 | uint64_t p_cap_id, ceph_seq_t p_seq, ceph_seq_t p_mseq, | |
5705 | int peer, int p_flags) | |
5706 | { | |
5707 | client_t client = session->info.inst.name.num(); | |
5708 | SnapRealm *realm = in->find_snaprealm(); | |
5709 | if (realm->have_past_parents_open()) { | |
5710 | dout(10) << "do_cap_import " << session->info.inst.name << " mseq " << cap->get_mseq() << " on " << *in << dendl; | |
5711 | if (cap->get_last_seq() == 0) // reconnected cap | |
5712 | cap->inc_last_seq(); | |
5713 | cap->set_last_issue(); | |
5714 | cap->set_last_issue_stamp(ceph_clock_now()); | |
5715 | cap->clear_new(); | |
5716 | MClientCaps *reap = new MClientCaps(CEPH_CAP_OP_IMPORT, | |
5717 | in->ino(), | |
5718 | realm->inode->ino(), | |
5719 | cap->get_cap_id(), cap->get_last_seq(), | |
5720 | cap->pending(), cap->wanted(), 0, | |
5721 | cap->get_mseq(), mds->get_osd_epoch_barrier()); | |
5722 | in->encode_cap_message(reap, cap); | |
5723 | realm->build_snap_trace(reap->snapbl); | |
5724 | reap->set_cap_peer(p_cap_id, p_seq, p_mseq, peer, p_flags); | |
5725 | mds->send_message_client_counted(reap, session); | |
5726 | } else { | |
5727 | dout(10) << "do_cap_import missing past snap parents, delaying " << session->info.inst.name << " mseq " | |
5728 | << cap->get_mseq() << " on " << *in << dendl; | |
5729 | in->auth_pin(this); | |
5730 | cap->inc_suppress(); | |
5731 | delayed_imported_caps[client].insert(in); | |
5732 | missing_snap_parents[in].size(); | |
5733 | } | |
5734 | } | |
5735 | ||
5736 | void MDCache::do_delayed_cap_imports() | |
5737 | { | |
5738 | dout(10) << "do_delayed_cap_imports" << dendl; | |
5739 | ||
5740 | assert(delayed_imported_caps.empty()); | |
5741 | } | |
5742 | ||
5743 | struct C_MDC_OpenSnapParents : public MDCacheContext { | |
5744 | explicit C_MDC_OpenSnapParents(MDCache *c) : MDCacheContext(c) {} | |
5745 | void finish(int r) override { | |
5746 | mdcache->open_snap_parents(); | |
5747 | } | |
5748 | }; | |
5749 | ||
5750 | void MDCache::open_snap_parents() | |
5751 | { | |
5752 | dout(10) << "open_snap_parents" << dendl; | |
5753 | ||
5754 | map<client_t,MClientSnap*> splits; | |
5755 | MDSGatherBuilder gather(g_ceph_context); | |
5756 | ||
5757 | auto p = missing_snap_parents.begin(); | |
5758 | while (p != missing_snap_parents.end()) { | |
5759 | CInode *in = p->first; | |
5760 | assert(in->snaprealm); | |
5761 | if (in->snaprealm->open_parents(gather.new_sub())) { | |
5762 | dout(10) << " past parents now open on " << *in << dendl; | |
5763 | ||
5764 | for (CInode *child : p->second) { | |
5765 | auto q = reconnected_caps.find(child->ino()); | |
5766 | assert(q != reconnected_caps.end()); | |
5767 | for (auto r = q->second.begin(); r != q->second.end(); ++r) { | |
5768 | if (r->second.snap_follows > 0 && r->second.snap_follows < in->first - 1) { | |
5769 | rebuild_need_snapflush(child, in->snaprealm, r->first, r->second.snap_follows); | |
5770 | } | |
5771 | // make sure client's cap is in the correct snaprealm. | |
5772 | if (r->second.realm_ino != in->ino()) { | |
5773 | prepare_realm_split(in->snaprealm, r->first, child->ino(), splits); | |
5774 | } | |
5775 | } | |
5776 | } | |
5777 | ||
5778 | missing_snap_parents.erase(p++); | |
5779 | ||
5780 | in->put(CInode::PIN_OPENINGSNAPPARENTS); | |
5781 | ||
5782 | // finish off client snaprealm reconnects? | |
5783 | map<inodeno_t,map<client_t,snapid_t> >::iterator q = reconnected_snaprealms.find(in->ino()); | |
5784 | if (q != reconnected_snaprealms.end()) { | |
5785 | for (map<client_t,snapid_t>::iterator r = q->second.begin(); | |
5786 | r != q->second.end(); | |
5787 | ++r) | |
5788 | finish_snaprealm_reconnect(r->first, in->snaprealm, r->second); | |
5789 | reconnected_snaprealms.erase(q); | |
5790 | } | |
5791 | } else { | |
5792 | dout(10) << " opening past parents on " << *in << dendl; | |
5793 | ++p; | |
5794 | } | |
5795 | } | |
5796 | ||
5797 | send_snaps(splits); | |
5798 | ||
5799 | if (gather.has_subs()) { | |
5800 | dout(10) << "open_snap_parents - waiting for " | |
5801 | << gather.num_subs_remaining() << dendl; | |
5802 | gather.set_finisher(new C_MDC_OpenSnapParents(this)); | |
5803 | gather.activate(); | |
5804 | } else { | |
5805 | if (!reconnected_snaprealms.empty()) { | |
5806 | stringstream warn_str; | |
5807 | for (map<inodeno_t,map<client_t,snapid_t> >::iterator p = reconnected_snaprealms.begin(); | |
5808 | p != reconnected_snaprealms.end(); | |
5809 | ++p) { | |
5810 | warn_str << " unconnected snaprealm " << p->first << "\n"; | |
5811 | for (map<client_t,snapid_t>::iterator q = p->second.begin(); | |
5812 | q != p->second.end(); | |
5813 | ++q) | |
5814 | warn_str << " client." << q->first << " snapid " << q->second << "\n"; | |
5815 | } | |
5816 | mds->clog->warn() << "open_snap_parents has:"; | |
5817 | mds->clog->warn(warn_str); | |
5818 | } | |
5819 | assert(rejoin_waiters.empty()); | |
5820 | assert(missing_snap_parents.empty()); | |
5821 | dout(10) << "open_snap_parents - all open" << dendl; | |
5822 | do_delayed_cap_imports(); | |
5823 | ||
5824 | assert(rejoin_done); | |
5825 | rejoin_done.release()->complete(0); | |
5826 | reconnected_caps.clear(); | |
5827 | } | |
5828 | } | |
5829 | ||
5830 | bool MDCache::open_undef_inodes_dirfrags() | |
5831 | { | |
5832 | dout(10) << "open_undef_inodes_dirfrags " | |
5833 | << rejoin_undef_inodes.size() << " inodes " | |
5834 | << rejoin_undef_dirfrags.size() << " dirfrags" << dendl; | |
5835 | ||
5836 | set<CDir*> fetch_queue = rejoin_undef_dirfrags; | |
5837 | ||
5838 | for (set<CInode*>::iterator p = rejoin_undef_inodes.begin(); | |
5839 | p != rejoin_undef_inodes.end(); | |
5840 | ++p) { | |
5841 | CInode *in = *p; | |
5842 | assert(!in->is_base()); | |
5843 | fetch_queue.insert(in->get_parent_dir()); | |
5844 | } | |
5845 | ||
5846 | if (fetch_queue.empty()) | |
5847 | return false; | |
5848 | ||
5849 | MDSGatherBuilder gather(g_ceph_context, new C_MDC_RejoinGatherFinish(this)); | |
5850 | for (set<CDir*>::iterator p = fetch_queue.begin(); | |
5851 | p != fetch_queue.end(); | |
5852 | ++p) { | |
5853 | CDir *dir = *p; | |
5854 | CInode *diri = dir->get_inode(); | |
5855 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) | |
5856 | continue; | |
5857 | if (dir->state_test(CDir::STATE_REJOINUNDEF)) | |
5858 | assert(diri->dirfragtree.is_leaf(dir->get_frag())); | |
5859 | dir->fetch(gather.new_sub()); | |
5860 | } | |
5861 | assert(gather.has_subs()); | |
5862 | gather.activate(); | |
5863 | return true; | |
5864 | } | |
5865 | ||
5866 | void MDCache::opened_undef_inode(CInode *in) { | |
5867 | dout(10) << "opened_undef_inode " << *in << dendl; | |
5868 | rejoin_undef_inodes.erase(in); | |
5869 | if (in->is_dir()) { | |
5870 | // FIXME: re-hash dentries if necessary | |
5871 | assert(in->inode.dir_layout.dl_dir_hash == g_conf->mds_default_dir_hash); | |
5872 | if (in->has_dirfrags() && !in->dirfragtree.is_leaf(frag_t())) { | |
5873 | CDir *dir = in->get_dirfrag(frag_t()); | |
5874 | assert(dir); | |
5875 | rejoin_undef_dirfrags.erase(dir); | |
5876 | in->force_dirfrags(); | |
5877 | list<CDir*> ls; | |
5878 | in->get_dirfrags(ls); | |
5879 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) | |
5880 | rejoin_undef_dirfrags.insert(*p); | |
5881 | } | |
5882 | } | |
5883 | } | |
5884 | ||
5885 | void MDCache::finish_snaprealm_reconnect(client_t client, SnapRealm *realm, snapid_t seq) | |
5886 | { | |
5887 | if (seq < realm->get_newest_seq()) { | |
5888 | dout(10) << "finish_snaprealm_reconnect client." << client << " has old seq " << seq << " < " | |
5889 | << realm->get_newest_seq() | |
5890 | << " on " << *realm << dendl; | |
5891 | // send an update | |
5892 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(client.v)); | |
5893 | if (session) { | |
5894 | MClientSnap *snap = new MClientSnap(CEPH_SNAP_OP_UPDATE); | |
5895 | realm->build_snap_trace(snap->bl); | |
5896 | mds->send_message_client_counted(snap, session); | |
5897 | } else { | |
5898 | dout(10) << " ...or not, no session for this client!" << dendl; | |
5899 | } | |
5900 | } else { | |
5901 | dout(10) << "finish_snaprealm_reconnect client." << client << " up to date" | |
5902 | << " on " << *realm << dendl; | |
5903 | } | |
5904 | } | |
5905 | ||
5906 | ||
5907 | ||
5908 | void MDCache::rejoin_send_acks() | |
5909 | { | |
5910 | dout(7) << "rejoin_send_acks" << dendl; | |
5911 | ||
5912 | // replicate stray | |
5913 | for (map<mds_rank_t, set<CInode*> >::iterator p = rejoin_unlinked_inodes.begin(); | |
5914 | p != rejoin_unlinked_inodes.end(); | |
5915 | ++p) { | |
5916 | for (set<CInode*>::iterator q = p->second.begin(); | |
5917 | q != p->second.end(); | |
5918 | ++q) { | |
5919 | CInode *in = *q; | |
5920 | dout(7) << " unlinked inode " << *in << dendl; | |
5921 | // inode expired | |
5922 | if (!in->is_replica(p->first)) | |
5923 | continue; | |
5924 | while (1) { | |
5925 | CDentry *dn = in->get_parent_dn(); | |
5926 | if (dn->is_replica(p->first)) | |
5927 | break; | |
5928 | dn->add_replica(p->first); | |
5929 | CDir *dir = dn->get_dir(); | |
5930 | if (dir->is_replica(p->first)) | |
5931 | break; | |
5932 | dir->add_replica(p->first); | |
5933 | in = dir->get_inode(); | |
5934 | if (in->is_replica(p->first)) | |
5935 | break; | |
224ce89b | 5936 | in->add_replica(p->first); |
7c673cae FG |
5937 | if (in->is_base()) |
5938 | break; | |
5939 | } | |
5940 | } | |
5941 | } | |
5942 | rejoin_unlinked_inodes.clear(); | |
5943 | ||
5944 | // send acks to everyone in the recovery set | |
31f18b77 | 5945 | map<mds_rank_t,MMDSCacheRejoin*> acks; |
7c673cae FG |
5946 | for (set<mds_rank_t>::iterator p = recovery_set.begin(); |
5947 | p != recovery_set.end(); | |
31f18b77 FG |
5948 | ++p) { |
5949 | if (rejoin_ack_sent.count(*p)) | |
5950 | continue; | |
5951 | acks[*p] = new MMDSCacheRejoin(MMDSCacheRejoin::OP_ACK); | |
5952 | } | |
5953 | ||
5954 | rejoin_ack_sent = recovery_set; | |
7c673cae FG |
5955 | |
5956 | // walk subtrees | |
5957 | for (map<CDir*,set<CDir*> >::iterator p = subtrees.begin(); | |
5958 | p != subtrees.end(); | |
5959 | ++p) { | |
5960 | CDir *dir = p->first; | |
5961 | if (!dir->is_auth()) | |
5962 | continue; | |
5963 | dout(10) << "subtree " << *dir << dendl; | |
5964 | ||
5965 | // auth items in this subtree | |
5966 | list<CDir*> dq; | |
5967 | dq.push_back(dir); | |
5968 | ||
5969 | while (!dq.empty()) { | |
5970 | CDir *dir = dq.front(); | |
5971 | dq.pop_front(); | |
5972 | ||
5973 | // dir | |
181888fb FG |
5974 | for (auto &r : dir->get_replicas()) { |
5975 | auto it = acks.find(r.first); | |
31f18b77 FG |
5976 | if (it == acks.end()) |
5977 | continue; | |
181888fb | 5978 | it->second->add_strong_dirfrag(dir->dirfrag(), ++r.second, dir->dir_rep); |
31f18b77 | 5979 | it->second->add_dirfrag_base(dir); |
7c673cae FG |
5980 | } |
5981 | ||
5982 | for (CDir::map_t::iterator q = dir->items.begin(); | |
5983 | q != dir->items.end(); | |
5984 | ++q) { | |
5985 | CDentry *dn = q->second; | |
5986 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
5987 | ||
5988 | // inode | |
5989 | CInode *in = NULL; | |
5990 | if (dnl->is_primary()) | |
5991 | in = dnl->get_inode(); | |
5992 | ||
5993 | // dentry | |
181888fb FG |
5994 | for (auto &r : dn->get_replicas()) { |
5995 | auto it = acks.find(r.first); | |
31f18b77 FG |
5996 | if (it == acks.end()) |
5997 | continue; | |
5998 | it->second->add_strong_dentry(dir->dirfrag(), dn->name, dn->first, dn->last, | |
7c673cae FG |
5999 | dnl->is_primary() ? dnl->get_inode()->ino():inodeno_t(0), |
6000 | dnl->is_remote() ? dnl->get_remote_ino():inodeno_t(0), | |
6001 | dnl->is_remote() ? dnl->get_remote_d_type():0, | |
181888fb | 6002 | ++r.second, |
7c673cae FG |
6003 | dn->lock.get_replica_state()); |
6004 | // peer missed MDentrylink message ? | |
181888fb FG |
6005 | if (in && !in->is_replica(r.first)) |
6006 | in->add_replica(r.first); | |
7c673cae FG |
6007 | } |
6008 | ||
6009 | if (!in) | |
6010 | continue; | |
6011 | ||
181888fb FG |
6012 | for (auto &r : in->get_replicas()) { |
6013 | auto it = acks.find(r.first); | |
31f18b77 FG |
6014 | if (it == acks.end()) |
6015 | continue; | |
6016 | it->second->add_inode_base(in, mds->mdsmap->get_up_features()); | |
7c673cae | 6017 | bufferlist bl; |
181888fb FG |
6018 | in->_encode_locks_state_for_rejoin(bl, r.first); |
6019 | it->second->add_inode_locks(in, ++r.second, bl); | |
7c673cae FG |
6020 | } |
6021 | ||
6022 | // subdirs in this subtree? | |
6023 | in->get_nested_dirfrags(dq); | |
6024 | } | |
6025 | } | |
6026 | } | |
6027 | ||
6028 | // base inodes too | |
6029 | if (root && root->is_auth()) | |
181888fb FG |
6030 | for (auto &r : root->get_replicas()) { |
6031 | auto it = acks.find(r.first); | |
31f18b77 FG |
6032 | if (it == acks.end()) |
6033 | continue; | |
6034 | it->second->add_inode_base(root, mds->mdsmap->get_up_features()); | |
7c673cae | 6035 | bufferlist bl; |
181888fb FG |
6036 | root->_encode_locks_state_for_rejoin(bl, r.first); |
6037 | it->second->add_inode_locks(root, ++r.second, bl); | |
7c673cae FG |
6038 | } |
6039 | if (myin) | |
181888fb FG |
6040 | for (auto &r : myin->get_replicas()) { |
6041 | auto it = acks.find(r.first); | |
31f18b77 FG |
6042 | if (it == acks.end()) |
6043 | continue; | |
6044 | it->second->add_inode_base(myin, mds->mdsmap->get_up_features()); | |
7c673cae | 6045 | bufferlist bl; |
181888fb FG |
6046 | myin->_encode_locks_state_for_rejoin(bl, r.first); |
6047 | it->second->add_inode_locks(myin, ++r.second, bl); | |
7c673cae FG |
6048 | } |
6049 | ||
6050 | // include inode base for any inodes whose scatterlocks may have updated | |
6051 | for (set<CInode*>::iterator p = rejoin_potential_updated_scatterlocks.begin(); | |
6052 | p != rejoin_potential_updated_scatterlocks.end(); | |
6053 | ++p) { | |
6054 | CInode *in = *p; | |
181888fb FG |
6055 | for (const auto &r : in->get_replicas()) { |
6056 | auto it = acks.find(r.first); | |
31f18b77 FG |
6057 | if (it == acks.end()) |
6058 | continue; | |
6059 | it->second->add_inode_base(in, mds->mdsmap->get_up_features()); | |
6060 | } | |
7c673cae FG |
6061 | } |
6062 | ||
6063 | // send acks | |
31f18b77 | 6064 | for (auto p = acks.begin(); p != acks.end(); ++p) { |
7c673cae FG |
6065 | ::encode(rejoin_imported_caps[p->first], p->second->imported_caps); |
6066 | mds->send_message_mds(p->second, p->first); | |
6067 | } | |
6068 | ||
6069 | rejoin_imported_caps.clear(); | |
6070 | } | |
6071 | ||
c07f9fc5 FG |
6072 | class C_MDC_ReIssueCaps : public MDCacheContext { |
6073 | CInode *in; | |
6074 | public: | |
6075 | C_MDC_ReIssueCaps(MDCache *mdc, CInode *i) : | |
6076 | MDCacheContext(mdc), in(i) | |
6077 | { | |
6078 | in->get(CInode::PIN_PTRWAITER); | |
6079 | } | |
6080 | void finish(int r) override { | |
6081 | if (!mdcache->mds->locker->eval(in, CEPH_CAP_LOCKS)) | |
6082 | mdcache->mds->locker->issue_caps(in); | |
6083 | in->put(CInode::PIN_PTRWAITER); | |
6084 | } | |
6085 | }; | |
7c673cae FG |
6086 | |
6087 | void MDCache::reissue_all_caps() | |
6088 | { | |
6089 | dout(10) << "reissue_all_caps" << dendl; | |
6090 | ||
6091 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator p = inode_map.begin(); | |
6092 | p != inode_map.end(); | |
6093 | ++p) { | |
6094 | CInode *in = p->second; | |
6095 | if (in->is_head() && in->is_any_caps()) { | |
c07f9fc5 FG |
6096 | // called by MDSRank::active_start(). There shouldn't be any frozen subtree. |
6097 | if (in->is_frozen_inode()) { | |
6098 | in->add_waiter(CInode::WAIT_UNFREEZE, new C_MDC_ReIssueCaps(this, in)); | |
6099 | continue; | |
6100 | } | |
7c673cae FG |
6101 | if (!mds->locker->eval(in, CEPH_CAP_LOCKS)) |
6102 | mds->locker->issue_caps(in); | |
6103 | } | |
6104 | } | |
6105 | } | |
6106 | ||
6107 | ||
6108 | // =============================================================================== | |
6109 | ||
6110 | struct C_MDC_QueuedCow : public MDCacheContext { | |
6111 | CInode *in; | |
6112 | MutationRef mut; | |
6113 | C_MDC_QueuedCow(MDCache *mdc, CInode *i, MutationRef& m) : | |
6114 | MDCacheContext(mdc), in(i), mut(m) {} | |
6115 | void finish(int r) override { | |
6116 | mdcache->_queued_file_recover_cow(in, mut); | |
6117 | } | |
6118 | }; | |
6119 | ||
6120 | ||
6121 | void MDCache::queue_file_recover(CInode *in) | |
6122 | { | |
6123 | dout(10) << "queue_file_recover " << *in << dendl; | |
6124 | assert(in->is_auth()); | |
6125 | ||
6126 | // cow? | |
6127 | /* | |
6128 | SnapRealm *realm = in->find_snaprealm(); | |
6129 | set<snapid_t> s = realm->get_snaps(); | |
6130 | while (!s.empty() && *s.begin() < in->first) | |
6131 | s.erase(s.begin()); | |
6132 | while (!s.empty() && *s.rbegin() > in->last) | |
6133 | s.erase(*s.rbegin()); | |
6134 | dout(10) << " snaps in [" << in->first << "," << in->last << "] are " << s << dendl; | |
6135 | if (s.size() > 1) { | |
6136 | inode_t *pi = in->project_inode(); | |
6137 | pi->version = in->pre_dirty(); | |
6138 | ||
6139 | auto mut(std::make_shared<MutationImpl>()); | |
6140 | mut->ls = mds->mdlog->get_current_segment(); | |
6141 | EUpdate *le = new EUpdate(mds->mdlog, "queue_file_recover cow"); | |
6142 | mds->mdlog->start_entry(le); | |
6143 | predirty_journal_parents(mut, &le->metablob, in, 0, PREDIRTY_PRIMARY); | |
6144 | ||
6145 | s.erase(*s.begin()); | |
6146 | while (!s.empty()) { | |
6147 | snapid_t snapid = *s.begin(); | |
6148 | CInode *cow_inode = 0; | |
6149 | journal_cow_inode(mut, &le->metablob, in, snapid-1, &cow_inode); | |
6150 | assert(cow_inode); | |
6151 | recovery_queue.enqueue(cow_inode); | |
6152 | s.erase(*s.begin()); | |
6153 | } | |
6154 | ||
6155 | in->parent->first = in->first; | |
6156 | le->metablob.add_primary_dentry(in->parent, in, true); | |
6157 | mds->mdlog->submit_entry(le, new C_MDC_QueuedCow(this, in, mut)); | |
6158 | mds->mdlog->flush(); | |
6159 | } | |
6160 | */ | |
6161 | ||
6162 | recovery_queue.enqueue(in); | |
6163 | } | |
6164 | ||
6165 | void MDCache::_queued_file_recover_cow(CInode *in, MutationRef& mut) | |
6166 | { | |
6167 | in->pop_and_dirty_projected_inode(mut->ls); | |
6168 | mut->apply(); | |
6169 | mds->locker->drop_locks(mut.get()); | |
6170 | mut->cleanup(); | |
6171 | } | |
6172 | ||
6173 | ||
6174 | /* | |
6175 | * called after recovery to recover file sizes for previously opened (for write) | |
6176 | * files. that is, those where max_size > size. | |
6177 | */ | |
6178 | void MDCache::identify_files_to_recover() | |
6179 | { | |
6180 | dout(10) << "identify_files_to_recover" << dendl; | |
6181 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator p = inode_map.begin(); | |
6182 | p != inode_map.end(); | |
6183 | ++p) { | |
6184 | CInode *in = p->second; | |
6185 | if (!in->is_auth()) | |
6186 | continue; | |
6187 | ||
6188 | if (in->last != CEPH_NOSNAP) | |
6189 | continue; | |
6190 | ||
6191 | // Only normal files need file size recovery | |
6192 | if (!in->is_file()) { | |
6193 | continue; | |
6194 | } | |
6195 | ||
6196 | bool recover = false; | |
6197 | for (map<client_t,client_writeable_range_t>::iterator p = in->inode.client_ranges.begin(); | |
6198 | p != in->inode.client_ranges.end(); | |
6199 | ++p) { | |
6200 | Capability *cap = in->get_client_cap(p->first); | |
6201 | if (!cap) { | |
6202 | dout(10) << " client." << p->first << " has range " << p->second << " but no cap on " << *in << dendl; | |
6203 | recover = true; | |
6204 | break; | |
6205 | } | |
6206 | } | |
6207 | ||
6208 | if (recover) { | |
6209 | if (in->filelock.is_stable()) { | |
6210 | in->auth_pin(&in->filelock); | |
6211 | } else { | |
6212 | assert(in->filelock.get_state() == LOCK_XLOCKSNAP); | |
6213 | } | |
6214 | in->filelock.set_state(LOCK_PRE_SCAN); | |
6215 | rejoin_recover_q.push_back(in); | |
6216 | } else { | |
6217 | rejoin_check_q.push_back(in); | |
6218 | } | |
6219 | } | |
6220 | } | |
6221 | ||
6222 | void MDCache::start_files_to_recover() | |
6223 | { | |
6224 | for (CInode *in : rejoin_check_q) { | |
6225 | if (in->filelock.get_state() == LOCK_XLOCKSNAP) | |
6226 | mds->locker->issue_caps(in); | |
6227 | mds->locker->check_inode_max_size(in); | |
6228 | } | |
6229 | rejoin_check_q.clear(); | |
6230 | for (CInode *in : rejoin_recover_q) { | |
6231 | mds->locker->file_recover(&in->filelock); | |
6232 | } | |
6233 | if (!rejoin_recover_q.empty()) { | |
6234 | rejoin_recover_q.clear(); | |
6235 | do_file_recover(); | |
6236 | } | |
6237 | } | |
6238 | ||
6239 | void MDCache::do_file_recover() | |
6240 | { | |
6241 | recovery_queue.advance(); | |
6242 | } | |
6243 | ||
6244 | // =============================================================================== | |
6245 | ||
6246 | ||
6247 | // ---------------------------- | |
6248 | // truncate | |
6249 | ||
6250 | class C_MDC_RetryTruncate : public MDCacheContext { | |
6251 | CInode *in; | |
6252 | LogSegment *ls; | |
6253 | public: | |
6254 | C_MDC_RetryTruncate(MDCache *c, CInode *i, LogSegment *l) : | |
6255 | MDCacheContext(c), in(i), ls(l) {} | |
6256 | void finish(int r) override { | |
6257 | mdcache->_truncate_inode(in, ls); | |
6258 | } | |
6259 | }; | |
6260 | ||
6261 | void MDCache::truncate_inode(CInode *in, LogSegment *ls) | |
6262 | { | |
6263 | inode_t *pi = in->get_projected_inode(); | |
6264 | dout(10) << "truncate_inode " | |
6265 | << pi->truncate_from << " -> " << pi->truncate_size | |
6266 | << " on " << *in | |
6267 | << dendl; | |
6268 | ||
6269 | ls->truncating_inodes.insert(in); | |
6270 | in->get(CInode::PIN_TRUNCATING); | |
6271 | in->auth_pin(this); | |
6272 | ||
6273 | if (!in->client_need_snapflush.empty() && | |
6274 | (in->get_caps_issued() & CEPH_CAP_FILE_BUFFER)) { | |
6275 | assert(in->filelock.is_xlocked()); | |
6276 | in->filelock.set_xlock_snap_sync(new C_MDC_RetryTruncate(this, in, ls)); | |
6277 | mds->locker->issue_caps(in); | |
6278 | return; | |
6279 | } | |
6280 | ||
6281 | _truncate_inode(in, ls); | |
6282 | } | |
6283 | ||
6284 | struct C_IO_MDC_TruncateFinish : public MDCacheIOContext { | |
6285 | CInode *in; | |
6286 | LogSegment *ls; | |
6287 | C_IO_MDC_TruncateFinish(MDCache *c, CInode *i, LogSegment *l) : | |
6288 | MDCacheIOContext(c), in(i), ls(l) {} | |
6289 | void finish(int r) override { | |
6290 | assert(r == 0 || r == -ENOENT); | |
6291 | mdcache->truncate_inode_finish(in, ls); | |
6292 | } | |
6293 | }; | |
6294 | ||
6295 | void MDCache::_truncate_inode(CInode *in, LogSegment *ls) | |
6296 | { | |
6297 | inode_t *pi = &in->inode; | |
6298 | dout(10) << "_truncate_inode " | |
6299 | << pi->truncate_from << " -> " << pi->truncate_size | |
6300 | << " on " << *in << dendl; | |
6301 | ||
6302 | assert(pi->is_truncating()); | |
6303 | assert(pi->truncate_size < (1ULL << 63)); | |
6304 | assert(pi->truncate_from < (1ULL << 63)); | |
6305 | assert(pi->truncate_size < pi->truncate_from); | |
6306 | ||
6307 | ||
6308 | SnapRealm *realm = in->find_snaprealm(); | |
6309 | SnapContext nullsnap; | |
6310 | const SnapContext *snapc; | |
6311 | if (realm) { | |
6312 | dout(10) << " realm " << *realm << dendl; | |
6313 | snapc = &realm->get_snap_context(); | |
6314 | } else { | |
6315 | dout(10) << " NO realm, using null context" << dendl; | |
6316 | snapc = &nullsnap; | |
6317 | assert(in->last == CEPH_NOSNAP); | |
6318 | } | |
6319 | dout(10) << "_truncate_inode snapc " << snapc << " on " << *in << dendl; | |
6320 | filer.truncate(in->inode.ino, &in->inode.layout, *snapc, | |
6321 | pi->truncate_size, pi->truncate_from-pi->truncate_size, | |
6322 | pi->truncate_seq, ceph::real_time::min(), 0, | |
6323 | new C_OnFinisher(new C_IO_MDC_TruncateFinish(this, in, ls), | |
6324 | mds->finisher)); | |
6325 | } | |
6326 | ||
6327 | struct C_MDC_TruncateLogged : public MDCacheLogContext { | |
6328 | CInode *in; | |
6329 | MutationRef mut; | |
6330 | C_MDC_TruncateLogged(MDCache *m, CInode *i, MutationRef& mu) : | |
6331 | MDCacheLogContext(m), in(i), mut(mu) {} | |
6332 | void finish(int r) override { | |
6333 | mdcache->truncate_inode_logged(in, mut); | |
6334 | } | |
6335 | }; | |
6336 | ||
6337 | void MDCache::truncate_inode_finish(CInode *in, LogSegment *ls) | |
6338 | { | |
6339 | dout(10) << "truncate_inode_finish " << *in << dendl; | |
6340 | ||
6341 | set<CInode*>::iterator p = ls->truncating_inodes.find(in); | |
6342 | assert(p != ls->truncating_inodes.end()); | |
6343 | ls->truncating_inodes.erase(p); | |
6344 | ||
6345 | // update | |
6346 | inode_t *pi = in->project_inode(); | |
6347 | pi->version = in->pre_dirty(); | |
6348 | pi->truncate_from = 0; | |
6349 | pi->truncate_pending--; | |
6350 | ||
6351 | MutationRef mut(new MutationImpl()); | |
6352 | mut->ls = mds->mdlog->get_current_segment(); | |
6353 | mut->add_projected_inode(in); | |
6354 | ||
6355 | EUpdate *le = new EUpdate(mds->mdlog, "truncate finish"); | |
6356 | mds->mdlog->start_entry(le); | |
6357 | CDentry *dn = in->get_projected_parent_dn(); | |
6358 | le->metablob.add_dir_context(dn->get_dir()); | |
6359 | le->metablob.add_primary_dentry(dn, in, true); | |
6360 | le->metablob.add_truncate_finish(in->ino(), ls->seq); | |
6361 | ||
6362 | journal_dirty_inode(mut.get(), &le->metablob, in); | |
6363 | mds->mdlog->submit_entry(le, new C_MDC_TruncateLogged(this, in, mut)); | |
6364 | ||
6365 | // flush immediately if there are readers/writers waiting | |
6366 | if (in->is_waiter_for(CInode::WAIT_TRUNC) || | |
6367 | (in->get_caps_wanted() & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR))) | |
6368 | mds->mdlog->flush(); | |
6369 | } | |
6370 | ||
6371 | void MDCache::truncate_inode_logged(CInode *in, MutationRef& mut) | |
6372 | { | |
6373 | dout(10) << "truncate_inode_logged " << *in << dendl; | |
6374 | mut->apply(); | |
6375 | mds->locker->drop_locks(mut.get()); | |
6376 | mut->cleanup(); | |
6377 | ||
6378 | in->put(CInode::PIN_TRUNCATING); | |
6379 | in->auth_unpin(this); | |
6380 | ||
6381 | list<MDSInternalContextBase*> waiters; | |
6382 | in->take_waiting(CInode::WAIT_TRUNC, waiters); | |
6383 | mds->queue_waiters(waiters); | |
6384 | } | |
6385 | ||
6386 | ||
6387 | void MDCache::add_recovered_truncate(CInode *in, LogSegment *ls) | |
6388 | { | |
6389 | dout(20) << "add_recovered_truncate " << *in << " in log segment " | |
6390 | << ls->seq << "/" << ls->offset << dendl; | |
6391 | ls->truncating_inodes.insert(in); | |
6392 | in->get(CInode::PIN_TRUNCATING); | |
6393 | } | |
6394 | ||
6395 | void MDCache::remove_recovered_truncate(CInode *in, LogSegment *ls) | |
6396 | { | |
6397 | dout(20) << "remove_recovered_truncate " << *in << " in log segment " | |
6398 | << ls->seq << "/" << ls->offset << dendl; | |
6399 | // if we have the logseg the truncate started in, it must be in our list. | |
6400 | set<CInode*>::iterator p = ls->truncating_inodes.find(in); | |
6401 | assert(p != ls->truncating_inodes.end()); | |
6402 | ls->truncating_inodes.erase(p); | |
6403 | in->put(CInode::PIN_TRUNCATING); | |
6404 | } | |
6405 | ||
6406 | void MDCache::start_recovered_truncates() | |
6407 | { | |
6408 | dout(10) << "start_recovered_truncates" << dendl; | |
6409 | for (map<uint64_t,LogSegment*>::iterator p = mds->mdlog->segments.begin(); | |
6410 | p != mds->mdlog->segments.end(); | |
6411 | ++p) { | |
6412 | LogSegment *ls = p->second; | |
6413 | for (set<CInode*>::iterator q = ls->truncating_inodes.begin(); | |
6414 | q != ls->truncating_inodes.end(); | |
6415 | ++q) { | |
6416 | CInode *in = *q; | |
6417 | in->auth_pin(this); | |
6418 | ||
6419 | if (!in->client_need_snapflush.empty() && | |
6420 | (in->get_caps_issued() & CEPH_CAP_FILE_BUFFER)) { | |
6421 | assert(in->filelock.is_stable()); | |
6422 | in->filelock.set_state(LOCK_XLOCKDONE); | |
6423 | in->auth_pin(&in->filelock); | |
6424 | in->filelock.set_xlock_snap_sync(new C_MDC_RetryTruncate(this, in, ls)); | |
6425 | // start_files_to_recover will revoke caps | |
6426 | continue; | |
6427 | } | |
6428 | _truncate_inode(in, ls); | |
6429 | } | |
6430 | } | |
6431 | } | |
6432 | ||
6433 | ||
6434 | ||
6435 | ||
6436 | ||
6437 | ||
6438 | // ================================================================================ | |
6439 | // cache trimming | |
6440 | ||
181888fb FG |
6441 | void MDCache::trim_lru(uint64_t count, map<mds_rank_t, MCacheExpire*> &expiremap) |
6442 | { | |
7c673cae | 6443 | bool is_standby_replay = mds->is_standby_replay(); |
181888fb FG |
6444 | std::vector<CDentry *> unexpirables; |
6445 | uint64_t trimmed = 0; | |
6446 | ||
6447 | dout(7) << "trim_lru trimming " << count | |
6448 | << " items from LRU" | |
6449 | << " size=" << lru.lru_get_size() | |
6450 | << " mid=" << lru.lru_get_top() | |
6451 | << " pintail=" << lru.lru_get_pintail() | |
6452 | << " pinned=" << lru.lru_get_num_pinned() | |
6453 | << dendl; | |
7c673cae | 6454 | |
31f18b77 FG |
6455 | for (;;) { |
6456 | CDentry *dn = static_cast<CDentry*>(bottom_lru.lru_expire()); | |
6457 | if (!dn) | |
6458 | break; | |
6459 | if (trim_dentry(dn, expiremap)) { | |
6460 | unexpirables.push_back(dn); | |
181888fb FG |
6461 | } else { |
6462 | trimmed++; | |
31f18b77 FG |
6463 | } |
6464 | } | |
6465 | ||
181888fb | 6466 | for (auto &dn : unexpirables) { |
31f18b77 | 6467 | bottom_lru.lru_insert_mid(dn); |
181888fb | 6468 | } |
31f18b77 FG |
6469 | unexpirables.clear(); |
6470 | ||
181888fb FG |
6471 | // trim dentries from the LRU until count is reached |
6472 | while (cache_toofull() || count > 0) { | |
7c673cae FG |
6473 | CDentry *dn = static_cast<CDentry*>(lru.lru_expire()); |
6474 | if (!dn) { | |
6475 | break; | |
6476 | } | |
7c673cae | 6477 | if ((is_standby_replay && dn->get_linkage()->inode && |
181888fb | 6478 | dn->get_linkage()->inode->item_open_file.is_on_list())) { |
7c673cae | 6479 | unexpirables.push_back(dn); |
181888fb FG |
6480 | } else if (trim_dentry(dn, expiremap)) { |
6481 | unexpirables.push_back(dn); | |
6482 | } else { | |
6483 | trimmed++; | |
7c673cae | 6484 | } |
181888fb | 6485 | count--; |
7c673cae | 6486 | } |
181888fb FG |
6487 | |
6488 | for (auto &dn : unexpirables) { | |
31f18b77 | 6489 | lru.lru_insert_mid(dn); |
181888fb | 6490 | } |
31f18b77 | 6491 | unexpirables.clear(); |
7c673cae | 6492 | |
181888fb FG |
6493 | dout(7) << "trim_lru trimmed " << trimmed << " items" << dendl; |
6494 | } | |
6495 | ||
6496 | /* | |
6497 | * note: only called while MDS is active or stopping... NOT during recovery. | |
6498 | * however, we may expire a replica whose authority is recovering. | |
6499 | * | |
6500 | * @param count is number of dentries to try to expire | |
6501 | */ | |
6502 | bool MDCache::trim(uint64_t count) | |
6503 | { | |
6504 | uint64_t used = cache_size(); | |
6505 | uint64_t limit = cache_limit_memory(); | |
6506 | map<mds_rank_t, MCacheExpire*> expiremap; | |
6507 | ||
6508 | dout(7) << "trim bytes_used=" << bytes2str(used) | |
6509 | << " limit=" << bytes2str(limit) | |
6510 | << " reservation=" << cache_reservation() | |
6511 | << "% count=" << count << dendl; | |
6512 | ||
6513 | // process delayed eval_stray() | |
6514 | stray_manager.advance_delayed(); | |
6515 | ||
6516 | trim_lru(count, expiremap); | |
6517 | ||
7c673cae | 6518 | // trim non-auth, non-bound subtrees |
181888fb | 6519 | for (auto p = subtrees.begin(); p != subtrees.end();) { |
7c673cae FG |
6520 | CDir *dir = p->first; |
6521 | ++p; | |
31f18b77 FG |
6522 | CInode *diri = dir->get_inode(); |
6523 | if (dir->is_auth()) { | |
6524 | if (!diri->is_auth() && !diri->is_base() && | |
6525 | dir->get_num_head_items() == 0) { | |
6526 | if (dir->state_test(CDir::STATE_EXPORTING) || | |
181888fb | 6527 | !(mds->is_active() || mds->is_stopping()) || |
31f18b77 FG |
6528 | dir->is_freezing() || dir->is_frozen()) |
6529 | continue; | |
6530 | ||
6531 | migrator->export_empty_import(dir); | |
6532 | } | |
6533 | } else { | |
6534 | if (!diri->is_auth()) { | |
6535 | if (dir->get_num_ref() > 1) // only subtree pin | |
6536 | continue; | |
6537 | list<CDir*> ls; | |
6538 | diri->get_subtree_dirfrags(ls); | |
6539 | if (diri->get_num_ref() > (int)ls.size()) // only pinned by subtrees | |
6540 | continue; | |
6541 | ||
6542 | // don't trim subtree root if its auth MDS is recovering. | |
6543 | // This simplify the cache rejoin code. | |
6544 | if (dir->is_subtree_root() && | |
6545 | rejoin_ack_gather.count(dir->get_dir_auth().first)) | |
6546 | continue; | |
7c673cae | 6547 | trim_dirfrag(dir, 0, expiremap); |
31f18b77 | 6548 | } |
7c673cae FG |
6549 | } |
6550 | } | |
6551 | ||
6552 | // trim root? | |
181888fb | 6553 | if (mds->is_stopping() && root) { |
7c673cae FG |
6554 | list<CDir*> ls; |
6555 | root->get_dirfrags(ls); | |
6556 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
6557 | CDir *dir = *p; | |
6558 | if (dir->get_num_ref() == 1) // subtree pin | |
6559 | trim_dirfrag(dir, 0, expiremap); | |
6560 | } | |
6561 | if (root->get_num_ref() == 0) | |
6562 | trim_inode(0, root, 0, expiremap); | |
6563 | } | |
6564 | ||
6565 | std::set<mds_rank_t> stopping; | |
6566 | mds->mdsmap->get_mds_set(stopping, MDSMap::STATE_STOPPING); | |
6567 | stopping.erase(mds->get_nodeid()); | |
6568 | for (auto rank : stopping) { | |
6569 | CInode* mdsdir_in = get_inode(MDS_INO_MDSDIR(rank)); | |
6570 | if (!mdsdir_in) | |
6571 | continue; | |
6572 | ||
6573 | if (expiremap.count(rank) == 0) { | |
6574 | expiremap[rank] = new MCacheExpire(mds->get_nodeid()); | |
6575 | } | |
6576 | ||
6577 | dout(20) << __func__ << ": try expiring " << *mdsdir_in << " for stopping mds." << mds << dendl; | |
6578 | ||
6579 | const bool aborted = expire_recursive(mdsdir_in, expiremap); | |
6580 | if (!aborted) { | |
6581 | dout(20) << __func__ << ": successfully expired mdsdir" << dendl; | |
6582 | list<CDir*> ls; | |
6583 | mdsdir_in->get_dirfrags(ls); | |
6584 | for (auto dir : ls) { | |
6585 | if (dir->get_num_ref() == 1) // subtree pin | |
6586 | trim_dirfrag(dir, dir, expiremap); | |
6587 | } | |
6588 | if (mdsdir_in->get_num_ref() == 0) | |
6589 | trim_inode(NULL, mdsdir_in, NULL, expiremap); | |
6590 | } else { | |
6591 | dout(20) << __func__ << ": some unexpirable contents in mdsdir" << dendl; | |
6592 | } | |
6593 | } | |
6594 | ||
6595 | // Other rank's base inodes (when I'm stopping) | |
181888fb | 6596 | if (mds->is_stopping()) { |
7c673cae FG |
6597 | for (set<CInode*>::iterator p = base_inodes.begin(); |
6598 | p != base_inodes.end(); ++p) { | |
6599 | if (MDS_INO_MDSDIR_OWNER((*p)->ino()) != mds->get_nodeid()) { | |
6600 | dout(20) << __func__ << ": maybe trimming base: " << *(*p) << dendl; | |
6601 | if ((*p)->get_num_ref() == 0) { | |
6602 | trim_inode(NULL, *p, NULL, expiremap); | |
6603 | } | |
6604 | } | |
6605 | } | |
6606 | } | |
6607 | ||
6608 | // send any expire messages | |
6609 | send_expire_messages(expiremap); | |
6610 | ||
6611 | return true; | |
6612 | } | |
6613 | ||
6614 | void MDCache::send_expire_messages(map<mds_rank_t, MCacheExpire*>& expiremap) | |
6615 | { | |
6616 | // send expires | |
6617 | for (map<mds_rank_t, MCacheExpire*>::iterator it = expiremap.begin(); | |
6618 | it != expiremap.end(); | |
6619 | ++it) { | |
6620 | if (mds->is_cluster_degraded() && | |
6621 | (mds->mdsmap->get_state(it->first) < MDSMap::STATE_REJOIN || | |
6622 | (mds->mdsmap->get_state(it->first) == MDSMap::STATE_REJOIN && | |
6623 | rejoin_sent.count(it->first) == 0))) { | |
6624 | it->second->put(); | |
6625 | continue; | |
6626 | } | |
6627 | dout(7) << "sending cache_expire to " << it->first << dendl; | |
6628 | mds->send_message_mds(it->second, it->first); | |
6629 | } | |
6630 | } | |
6631 | ||
6632 | ||
6633 | bool MDCache::trim_dentry(CDentry *dn, map<mds_rank_t, MCacheExpire*>& expiremap) | |
6634 | { | |
6635 | dout(12) << "trim_dentry " << *dn << dendl; | |
6636 | ||
6637 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
6638 | ||
6639 | CDir *dir = dn->get_dir(); | |
6640 | assert(dir); | |
6641 | ||
6642 | CDir *con = get_subtree_root(dir); | |
6643 | if (con) | |
6644 | dout(12) << " in container " << *con << dendl; | |
6645 | else { | |
6646 | dout(12) << " no container; under a not-yet-linked dir" << dendl; | |
6647 | assert(dn->is_auth()); | |
6648 | } | |
6649 | ||
6650 | // If replica dentry is not readable, it's likely we will receive | |
6651 | // MDentryLink/MDentryUnlink message soon (It's possible we first | |
6652 | // receive a MDentryUnlink message, then MDentryLink message) | |
6653 | // MDentryLink message only replicates an inode, so we should | |
6654 | // avoid trimming the inode's parent dentry. This is because that | |
6655 | // unconnected replicas are problematic for subtree migration. | |
6656 | if (!dn->is_auth() && !dn->lock.can_read(-1) && | |
6657 | !dn->get_dir()->get_inode()->is_stray()) | |
6658 | return true; | |
6659 | ||
6660 | // adjust the dir state | |
6661 | // NOTE: we can safely remove a clean, null dentry without effecting | |
6662 | // directory completeness. | |
6663 | // (check this _before_ we unlink the inode, below!) | |
6664 | bool clear_complete = false; | |
6665 | if (!(dnl->is_null() && dn->is_clean())) | |
6666 | clear_complete = true; | |
6667 | ||
6668 | // unlink the dentry | |
6669 | if (dnl->is_remote()) { | |
6670 | // just unlink. | |
31f18b77 | 6671 | dir->unlink_inode(dn, false); |
7c673cae FG |
6672 | } else if (dnl->is_primary()) { |
6673 | // expire the inode, too. | |
6674 | CInode *in = dnl->get_inode(); | |
6675 | assert(in); | |
6676 | if (trim_inode(dn, in, con, expiremap)) | |
6677 | return true; // purging stray instead of trimming | |
6678 | } else { | |
6679 | assert(dnl->is_null()); | |
6680 | } | |
6681 | ||
6682 | if (!dn->is_auth()) { | |
6683 | // notify dentry authority. | |
6684 | mds_authority_t auth = dn->authority(); | |
6685 | ||
6686 | for (int p=0; p<2; p++) { | |
6687 | mds_rank_t a = auth.first; | |
6688 | if (p) a = auth.second; | |
6689 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6690 | if (mds->get_nodeid() == auth.second && | |
6691 | con->is_importing()) break; // don't send any expire while importing. | |
6692 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6693 | ||
6694 | dout(12) << " sending expire to mds." << a << " on " << *dn << dendl; | |
6695 | assert(a != mds->get_nodeid()); | |
6696 | if (expiremap.count(a) == 0) | |
6697 | expiremap[a] = new MCacheExpire(mds->get_nodeid()); | |
6698 | expiremap[a]->add_dentry(con->dirfrag(), dir->dirfrag(), dn->name, dn->last, dn->get_replica_nonce()); | |
6699 | } | |
6700 | } | |
6701 | ||
6702 | // remove dentry | |
6703 | if (dn->last == CEPH_NOSNAP && dir->is_auth()) | |
6704 | dir->add_to_bloom(dn); | |
6705 | dir->remove_dentry(dn); | |
6706 | ||
6707 | if (clear_complete) | |
6708 | dir->state_clear(CDir::STATE_COMPLETE); | |
6709 | ||
7c673cae FG |
6710 | if (mds->logger) mds->logger->inc(l_mds_inodes_expired); |
6711 | return false; | |
6712 | } | |
6713 | ||
6714 | ||
6715 | void MDCache::trim_dirfrag(CDir *dir, CDir *con, map<mds_rank_t, MCacheExpire*>& expiremap) | |
6716 | { | |
6717 | dout(15) << "trim_dirfrag " << *dir << dendl; | |
6718 | ||
6719 | if (dir->is_subtree_root()) { | |
6720 | assert(!dir->is_auth() || | |
6721 | (!dir->is_replicated() && dir->inode->is_base())); | |
6722 | remove_subtree(dir); // remove from subtree map | |
6723 | } | |
6724 | assert(dir->get_num_ref() == 0); | |
6725 | ||
6726 | CInode *in = dir->get_inode(); | |
6727 | ||
6728 | if (!dir->is_auth()) { | |
6729 | mds_authority_t auth = dir->authority(); | |
6730 | ||
6731 | // was this an auth delegation? (if so, slightly modified container) | |
6732 | dirfrag_t condf; | |
6733 | if (dir->is_subtree_root()) { | |
6734 | dout(12) << " subtree root, container is " << *dir << dendl; | |
6735 | con = dir; | |
6736 | condf = dir->dirfrag(); | |
6737 | } else { | |
6738 | condf = con->dirfrag(); | |
6739 | } | |
6740 | ||
6741 | for (int p=0; p<2; p++) { | |
6742 | mds_rank_t a = auth.first; | |
6743 | if (p) a = auth.second; | |
6744 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6745 | if (mds->get_nodeid() == auth.second && | |
6746 | con->is_importing()) break; // don't send any expire while importing. | |
6747 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6748 | ||
6749 | dout(12) << " sending expire to mds." << a << " on " << *dir << dendl; | |
6750 | assert(a != mds->get_nodeid()); | |
6751 | if (expiremap.count(a) == 0) | |
6752 | expiremap[a] = new MCacheExpire(mds->get_nodeid()); | |
6753 | expiremap[a]->add_dir(condf, dir->dirfrag(), dir->replica_nonce); | |
6754 | } | |
6755 | } | |
6756 | ||
6757 | in->close_dirfrag(dir->dirfrag().frag); | |
6758 | } | |
6759 | ||
6760 | /** | |
6761 | * Try trimming an inode from the cache | |
6762 | * | |
6763 | * @return true if the inode is still in cache, else false if it was trimmed | |
6764 | */ | |
6765 | bool MDCache::trim_inode(CDentry *dn, CInode *in, CDir *con, map<mds_rank_t, MCacheExpire*>& expiremap) | |
6766 | { | |
6767 | dout(15) << "trim_inode " << *in << dendl; | |
6768 | assert(in->get_num_ref() == 0); | |
6769 | ||
6770 | if (in->is_dir()) { | |
6771 | // If replica inode's dirfragtreelock is not readable, it's likely | |
6772 | // some dirfrags of the inode are being fragmented and we will receive | |
6773 | // MMDSFragmentNotify soon. MMDSFragmentNotify only replicates the new | |
6774 | // dirfrags, so we should avoid trimming these dirfrags' parent inode. | |
6775 | // This is because that unconnected replicas are problematic for | |
6776 | // subtree migration. | |
6777 | // | |
6778 | if (!in->is_auth() && !in->dirfragtreelock.can_read(-1)) | |
6779 | return true; | |
6780 | ||
6781 | // DIR | |
6782 | list<CDir*> dfls; | |
6783 | in->get_dirfrags(dfls); | |
6784 | for (list<CDir*>::iterator p = dfls.begin(); p != dfls.end(); ++p) { | |
6785 | CDir *dir = *p; | |
6786 | assert(!dir->is_subtree_root()); | |
6787 | trim_dirfrag(dir, con ? con:dir, expiremap); // if no container (e.g. root dirfrag), use *p | |
6788 | } | |
6789 | } | |
6790 | ||
6791 | // INODE | |
6792 | if (in->is_auth()) { | |
6793 | // eval stray after closing dirfrags | |
6794 | if (dn && !dn->state_test(CDentry::STATE_PURGING)) { | |
6795 | maybe_eval_stray(in); | |
6796 | if (dn->state_test(CDentry::STATE_PURGING) || dn->get_num_ref() > 0) | |
6797 | return true; | |
6798 | } | |
6799 | } else { | |
6800 | mds_authority_t auth = in->authority(); | |
6801 | ||
6802 | dirfrag_t df; | |
6803 | if (con) | |
6804 | df = con->dirfrag(); | |
6805 | else | |
6806 | df = dirfrag_t(0,frag_t()); // must be a root or stray inode. | |
6807 | ||
6808 | for (int p=0; p<2; p++) { | |
6809 | mds_rank_t a = auth.first; | |
6810 | if (p) a = auth.second; | |
6811 | if (a < 0 || (p == 1 && auth.second == auth.first)) break; | |
6812 | if (con && mds->get_nodeid() == auth.second && | |
6813 | con->is_importing()) break; // don't send any expire while importing. | |
6814 | if (a == mds->get_nodeid()) continue; // on export, ignore myself. | |
6815 | ||
6816 | dout(12) << " sending expire to mds." << a << " on " << *in << dendl; | |
6817 | assert(a != mds->get_nodeid()); | |
6818 | if (expiremap.count(a) == 0) | |
6819 | expiremap[a] = new MCacheExpire(mds->get_nodeid()); | |
6820 | expiremap[a]->add_inode(df, in->vino(), in->get_replica_nonce()); | |
6821 | } | |
6822 | } | |
6823 | ||
6824 | /* | |
6825 | if (in->is_auth()) { | |
6826 | if (in->hack_accessed) | |
6827 | mds->logger->inc("outt"); | |
6828 | else { | |
6829 | mds->logger->inc("outut"); | |
6830 | mds->logger->fset("oututl", ceph_clock_now() - in->hack_load_stamp); | |
6831 | } | |
6832 | } | |
6833 | */ | |
6834 | ||
6835 | // unlink | |
6836 | if (dn) | |
31f18b77 | 6837 | dn->get_dir()->unlink_inode(dn, false); |
7c673cae FG |
6838 | remove_inode(in); |
6839 | return false; | |
6840 | } | |
6841 | ||
6842 | ||
6843 | /** | |
6844 | * trim_non_auth - remove any non-auth items from our cache | |
6845 | * | |
6846 | * this reduces the amount of non-auth metadata in our cache, reducing the | |
6847 | * load incurred by the rejoin phase. | |
6848 | * | |
6849 | * the only non-auth items that remain are those that are needed to | |
6850 | * attach our own subtrees to the root. | |
6851 | * | |
6852 | * when we are done, all dentries will be in the top bit of the lru. | |
6853 | * | |
6854 | * why we have to do this: | |
6855 | * we may not have accurate linkage for non-auth items. which means we will | |
6856 | * know which subtree it falls into, and can not be sure to declare it to the | |
6857 | * correct authority. | |
6858 | */ | |
6859 | void MDCache::trim_non_auth() | |
6860 | { | |
6861 | dout(7) << "trim_non_auth" << dendl; | |
6862 | ||
6863 | // temporarily pin all subtree roots | |
6864 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
6865 | p != subtrees.end(); | |
6866 | ++p) | |
6867 | p->first->get(CDir::PIN_SUBTREETEMP); | |
6868 | ||
31f18b77 | 6869 | list<CDentry*> auth_list; |
7c673cae FG |
6870 | |
6871 | // trim non-auth items from the lru | |
31f18b77 FG |
6872 | for (;;) { |
6873 | CDentry *dn = NULL; | |
6874 | if (bottom_lru.lru_get_size() > 0) | |
6875 | dn = static_cast<CDentry*>(bottom_lru.lru_expire()); | |
6876 | if (!dn && lru.lru_get_size() > 0) | |
6877 | dn = static_cast<CDentry*>(lru.lru_expire()); | |
6878 | if (!dn) | |
6879 | break; | |
6880 | ||
7c673cae FG |
6881 | CDentry::linkage_t *dnl = dn->get_linkage(); |
6882 | ||
6883 | if (dn->is_auth()) { | |
6884 | // add back into lru (at the top) | |
31f18b77 | 6885 | auth_list.push_back(dn); |
7c673cae FG |
6886 | |
6887 | if (dnl->is_remote() && dnl->get_inode() && !dnl->get_inode()->is_auth()) | |
6888 | dn->unlink_remote(dnl); | |
7c673cae FG |
6889 | } else { |
6890 | // non-auth. expire. | |
6891 | CDir *dir = dn->get_dir(); | |
6892 | assert(dir); | |
6893 | ||
6894 | // unlink the dentry | |
6895 | dout(10) << " removing " << *dn << dendl; | |
6896 | if (dnl->is_remote()) { | |
31f18b77 | 6897 | dir->unlink_inode(dn, false); |
7c673cae FG |
6898 | } |
6899 | else if (dnl->is_primary()) { | |
6900 | CInode *in = dnl->get_inode(); | |
6901 | dout(10) << " removing " << *in << dendl; | |
6902 | list<CDir*> ls; | |
6903 | in->get_dirfrags(ls); | |
6904 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
6905 | CDir *subdir = *p; | |
6906 | assert(!subdir->is_subtree_root()); | |
6907 | in->close_dirfrag(subdir->dirfrag().frag); | |
6908 | } | |
31f18b77 | 6909 | dir->unlink_inode(dn, false); |
7c673cae FG |
6910 | remove_inode(in); |
6911 | } | |
6912 | else { | |
6913 | assert(dnl->is_null()); | |
6914 | } | |
6915 | ||
6916 | assert(!dir->has_bloom()); | |
6917 | dir->remove_dentry(dn); | |
6918 | // adjust the dir state | |
6919 | dir->state_clear(CDir::STATE_COMPLETE); // dir incomplete! | |
6920 | // close empty non-auth dirfrag | |
6921 | if (!dir->is_subtree_root() && dir->get_num_any() == 0) | |
6922 | dir->inode->close_dirfrag(dir->get_frag()); | |
6923 | } | |
6924 | } | |
6925 | ||
31f18b77 FG |
6926 | for (auto dn : auth_list) { |
6927 | if (dn->state_test(CDentry::STATE_BOTTOMLRU)) | |
6928 | bottom_lru.lru_insert_mid(dn); | |
6929 | else | |
6930 | lru.lru_insert_top(dn); | |
6931 | } | |
6932 | ||
7c673cae FG |
6933 | // move everything in the pintail to the top bit of the lru. |
6934 | lru.lru_touch_entire_pintail(); | |
6935 | ||
6936 | // unpin all subtrees | |
6937 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
6938 | p != subtrees.end(); | |
6939 | ++p) | |
6940 | p->first->put(CDir::PIN_SUBTREETEMP); | |
6941 | ||
31f18b77 FG |
6942 | if (lru.lru_get_size() == 0 && |
6943 | bottom_lru.lru_get_size() == 0) { | |
7c673cae FG |
6944 | // root, stray, etc.? |
6945 | ceph::unordered_map<vinodeno_t,CInode*>::iterator p = inode_map.begin(); | |
6946 | while (p != inode_map.end()) { | |
6947 | ceph::unordered_map<vinodeno_t,CInode*>::iterator next = p; | |
6948 | ++next; | |
6949 | CInode *in = p->second; | |
6950 | if (!in->is_auth()) { | |
6951 | list<CDir*> ls; | |
6952 | in->get_dirfrags(ls); | |
6953 | for (list<CDir*>::iterator p = ls.begin(); | |
6954 | p != ls.end(); | |
6955 | ++p) { | |
6956 | dout(10) << " removing " << **p << dendl; | |
6957 | assert((*p)->get_num_ref() == 1); // SUBTREE | |
6958 | remove_subtree((*p)); | |
6959 | in->close_dirfrag((*p)->dirfrag().frag); | |
6960 | } | |
6961 | dout(10) << " removing " << *in << dendl; | |
6962 | assert(!in->get_parent_dn()); | |
6963 | assert(in->get_num_ref() == 0); | |
6964 | remove_inode(in); | |
6965 | } | |
6966 | p = next; | |
6967 | } | |
6968 | } | |
6969 | ||
6970 | show_subtrees(); | |
6971 | } | |
6972 | ||
6973 | /** | |
6974 | * Recursively trim the subtree rooted at directory to remove all | |
6975 | * CInodes/CDentrys/CDirs that aren't links to remote MDSes, or ancestors | |
6976 | * of those links. This is used to clear invalid data out of the cache. | |
6977 | * Note that it doesn't clear the passed-in directory, since that's not | |
6978 | * always safe. | |
6979 | */ | |
6980 | bool MDCache::trim_non_auth_subtree(CDir *dir) | |
6981 | { | |
6982 | dout(10) << "trim_non_auth_subtree(" << dir << ") " << *dir << dendl; | |
6983 | ||
6984 | bool keep_dir = !can_trim_non_auth_dirfrag(dir); | |
6985 | ||
6986 | CDir::map_t::iterator j = dir->begin(); | |
6987 | CDir::map_t::iterator i = j; | |
6988 | while (j != dir->end()) { | |
6989 | i = j++; | |
6990 | CDentry *dn = i->second; | |
6991 | dout(10) << "trim_non_auth_subtree(" << dir << ") Checking dentry " << dn << dendl; | |
6992 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
6993 | if (dnl->is_primary()) { // check for subdirectories, etc | |
6994 | CInode *in = dnl->get_inode(); | |
6995 | bool keep_inode = false; | |
6996 | if (in->is_dir()) { | |
6997 | list<CDir*> subdirs; | |
6998 | in->get_dirfrags(subdirs); | |
6999 | for (list<CDir*>::iterator subdir = subdirs.begin(); | |
7000 | subdir != subdirs.end(); | |
7001 | ++subdir) { | |
7002 | if ((*subdir)->is_subtree_root()) { | |
7003 | keep_inode = true; | |
7004 | dout(10) << "trim_non_auth_subtree(" << dir << ") keeping " << **subdir << dendl; | |
7005 | } else { | |
7006 | if (trim_non_auth_subtree(*subdir)) | |
7007 | keep_inode = true; | |
7008 | else { | |
7009 | in->close_dirfrag((*subdir)->get_frag()); | |
7010 | dir->state_clear(CDir::STATE_COMPLETE); // now incomplete! | |
7011 | } | |
7012 | } | |
7013 | } | |
7014 | ||
7015 | } | |
7016 | if (!keep_inode) { // remove it! | |
7017 | dout(20) << "trim_non_auth_subtree(" << dir << ") removing inode " << in << " with dentry" << dn << dendl; | |
31f18b77 | 7018 | dir->unlink_inode(dn, false); |
7c673cae FG |
7019 | remove_inode(in); |
7020 | assert(!dir->has_bloom()); | |
7021 | dir->remove_dentry(dn); | |
7022 | } else { | |
7023 | dout(20) << "trim_non_auth_subtree(" << dir << ") keeping inode " << in << " with dentry " << dn <<dendl; | |
7024 | dn->state_clear(CDentry::STATE_AUTH); | |
7025 | in->state_clear(CInode::STATE_AUTH); | |
7026 | } | |
7027 | } else if (keep_dir && dnl->is_null()) { // keep null dentry for slave rollback | |
7028 | dout(20) << "trim_non_auth_subtree(" << dir << ") keeping dentry " << dn <<dendl; | |
7029 | } else { // just remove it | |
7030 | dout(20) << "trim_non_auth_subtree(" << dir << ") removing dentry " << dn << dendl; | |
7031 | if (dnl->is_remote()) | |
31f18b77 | 7032 | dir->unlink_inode(dn, false); |
7c673cae FG |
7033 | dir->remove_dentry(dn); |
7034 | } | |
7035 | } | |
7036 | dir->state_clear(CDir::STATE_AUTH); | |
7037 | /** | |
7038 | * We've now checked all our children and deleted those that need it. | |
7039 | * Now return to caller, and tell them if *we're* a keeper. | |
7040 | */ | |
7041 | return keep_dir || dir->get_num_any(); | |
7042 | } | |
7043 | ||
7044 | /* | |
7045 | * during replay, when we determine a subtree is no longer ours, we | |
7046 | * try to trim it from our cache. because subtrees must be connected | |
7047 | * to the root, the fact that we can trim this tree may mean that our | |
7048 | * children or parents can also be trimmed. | |
7049 | */ | |
7050 | void MDCache::try_trim_non_auth_subtree(CDir *dir) | |
7051 | { | |
7052 | dout(10) << "try_trim_nonauth_subtree " << *dir << dendl; | |
7053 | ||
7054 | // can we now trim child subtrees? | |
7055 | set<CDir*> bounds; | |
7056 | get_subtree_bounds(dir, bounds); | |
7057 | for (set<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) { | |
7058 | CDir *bd = *p; | |
7059 | if (bd->get_dir_auth().first != mds->get_nodeid() && // we are not auth | |
7060 | bd->get_num_any() == 0 && // and empty | |
7061 | can_trim_non_auth_dirfrag(bd)) { | |
7062 | CInode *bi = bd->get_inode(); | |
7063 | dout(10) << " closing empty non-auth child subtree " << *bd << dendl; | |
7064 | remove_subtree(bd); | |
7065 | bd->mark_clean(); | |
7066 | bi->close_dirfrag(bd->get_frag()); | |
7067 | } | |
7068 | } | |
7069 | ||
7070 | if (trim_non_auth_subtree(dir)) { | |
7071 | // keep | |
7072 | try_subtree_merge(dir); | |
7073 | } else { | |
7074 | // can we trim this subtree (and possibly our ancestors) too? | |
7075 | while (true) { | |
7076 | CInode *diri = dir->get_inode(); | |
7077 | if (diri->is_base()) { | |
7078 | if (!diri->is_root() && diri->authority().first != mds->get_nodeid()) { | |
7079 | dout(10) << " closing empty non-auth subtree " << *dir << dendl; | |
7080 | remove_subtree(dir); | |
7081 | dir->mark_clean(); | |
7082 | diri->close_dirfrag(dir->get_frag()); | |
7083 | ||
7084 | dout(10) << " removing " << *diri << dendl; | |
7085 | assert(!diri->get_parent_dn()); | |
7086 | assert(diri->get_num_ref() == 0); | |
7087 | remove_inode(diri); | |
7088 | } | |
7089 | break; | |
7090 | } | |
7091 | ||
7092 | CDir *psub = get_subtree_root(diri->get_parent_dir()); | |
7093 | dout(10) << " parent subtree is " << *psub << dendl; | |
7094 | if (psub->get_dir_auth().first == mds->get_nodeid()) | |
7095 | break; // we are auth, keep. | |
7096 | ||
7097 | dout(10) << " closing empty non-auth subtree " << *dir << dendl; | |
7098 | remove_subtree(dir); | |
7099 | dir->mark_clean(); | |
7100 | diri->close_dirfrag(dir->get_frag()); | |
7101 | ||
7102 | dout(10) << " parent subtree also non-auth: " << *psub << dendl; | |
7103 | if (trim_non_auth_subtree(psub)) | |
7104 | break; | |
7105 | dir = psub; | |
7106 | } | |
7107 | } | |
7108 | ||
7109 | show_subtrees(); | |
7110 | } | |
7111 | ||
7112 | void MDCache::standby_trim_segment(LogSegment *ls) | |
7113 | { | |
7114 | ls->new_dirfrags.clear_list(); | |
7115 | ls->open_files.clear_list(); | |
7116 | ||
7117 | while (!ls->dirty_dirfrags.empty()) { | |
7118 | CDir *dir = ls->dirty_dirfrags.front(); | |
7119 | dir->mark_clean(); | |
7120 | } | |
7121 | while (!ls->dirty_inodes.empty()) { | |
7122 | CInode *in = ls->dirty_inodes.front(); | |
7123 | in->mark_clean(); | |
7124 | } | |
7125 | while (!ls->dirty_dentries.empty()) { | |
7126 | CDentry *dn = ls->dirty_dentries.front(); | |
7127 | dn->mark_clean(); | |
7128 | } | |
7129 | while (!ls->dirty_parent_inodes.empty()) { | |
7130 | CInode *in = ls->dirty_parent_inodes.front(); | |
7131 | in->clear_dirty_parent(); | |
7132 | } | |
7133 | while (!ls->dirty_dirfrag_dir.empty()) { | |
7134 | CInode *in = ls->dirty_dirfrag_dir.front(); | |
7135 | in->filelock.remove_dirty(); | |
7136 | } | |
7137 | while (!ls->dirty_dirfrag_nest.empty()) { | |
7138 | CInode *in = ls->dirty_dirfrag_nest.front(); | |
7139 | in->nestlock.remove_dirty(); | |
7140 | } | |
7141 | while (!ls->dirty_dirfrag_dirfragtree.empty()) { | |
7142 | CInode *in = ls->dirty_dirfrag_dirfragtree.front(); | |
7143 | in->dirfragtreelock.remove_dirty(); | |
7144 | } | |
7145 | } | |
7146 | ||
7147 | /* This function DOES put the passed message before returning */ | |
7148 | void MDCache::handle_cache_expire(MCacheExpire *m) | |
7149 | { | |
7150 | mds_rank_t from = mds_rank_t(m->get_from()); | |
7151 | ||
7152 | dout(7) << "cache_expire from mds." << from << dendl; | |
7153 | ||
7154 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
7155 | m->put(); | |
7156 | return; | |
7157 | } | |
7158 | ||
7159 | set<SimpleLock *> gather_locks; | |
7160 | // loop over realms | |
7161 | for (map<dirfrag_t,MCacheExpire::realm>::iterator p = m->realms.begin(); | |
7162 | p != m->realms.end(); | |
7163 | ++p) { | |
7164 | // check container? | |
7165 | if (p->first.ino > 0) { | |
7166 | CInode *expired_inode = get_inode(p->first.ino); | |
7167 | assert(expired_inode); // we had better have this. | |
7168 | CDir *parent_dir = expired_inode->get_approx_dirfrag(p->first.frag); | |
7169 | assert(parent_dir); | |
7170 | ||
7171 | int export_state = -1; | |
7172 | if (parent_dir->is_auth() && parent_dir->is_exporting()) { | |
7173 | export_state = migrator->get_export_state(parent_dir); | |
7174 | assert(export_state >= 0); | |
7175 | } | |
7176 | ||
7177 | if (!parent_dir->is_auth() || | |
7178 | (export_state != -1 && | |
7179 | ((export_state == Migrator::EXPORT_WARNING && | |
7180 | migrator->export_has_warned(parent_dir,from)) || | |
7181 | export_state == Migrator::EXPORT_EXPORTING || | |
7182 | export_state == Migrator::EXPORT_LOGGINGFINISH || | |
7183 | (export_state == Migrator::EXPORT_NOTIFYING && | |
7184 | !migrator->export_has_notified(parent_dir,from))))) { | |
7185 | ||
7186 | // not auth. | |
7187 | dout(7) << "delaying nonauth|warned expires for " << *parent_dir << dendl; | |
7188 | assert(parent_dir->is_frozen_tree_root()); | |
7189 | ||
7190 | // make a message container | |
7191 | if (delayed_expire[parent_dir].count(from) == 0) | |
7192 | delayed_expire[parent_dir][from] = new MCacheExpire(from); | |
7193 | ||
7194 | // merge these expires into it | |
7195 | delayed_expire[parent_dir][from]->add_realm(p->first, p->second); | |
7196 | continue; | |
7197 | } | |
7198 | assert(export_state <= Migrator::EXPORT_PREPPING || | |
7199 | (export_state == Migrator::EXPORT_WARNING && | |
7200 | !migrator->export_has_warned(parent_dir, from))); | |
7201 | ||
7202 | dout(7) << "expires for " << *parent_dir << dendl; | |
7203 | } else { | |
7204 | dout(7) << "containerless expires (root, stray inodes)" << dendl; | |
7205 | } | |
7206 | ||
7207 | // INODES | |
7208 | for (map<vinodeno_t,uint32_t>::iterator it = p->second.inodes.begin(); | |
7209 | it != p->second.inodes.end(); | |
7210 | ++it) { | |
7211 | CInode *in = get_inode(it->first); | |
7212 | unsigned nonce = it->second; | |
7213 | ||
7214 | if (!in) { | |
7215 | dout(0) << " inode expire on " << it->first << " from " << from | |
7216 | << ", don't have it" << dendl; | |
7217 | assert(in); | |
7218 | } | |
7219 | assert(in->is_auth()); | |
7220 | dout(20) << __func__ << ": expiring inode " << *in << dendl; | |
7221 | ||
7222 | // check nonce | |
7223 | if (nonce == in->get_replica_nonce(from)) { | |
7224 | // remove from our cached_by | |
7225 | dout(7) << " inode expire on " << *in << " from mds." << from | |
7226 | << " cached_by was " << in->get_replicas() << dendl; | |
7227 | inode_remove_replica(in, from, false, gather_locks); | |
7228 | } | |
7229 | else { | |
7230 | // this is an old nonce, ignore expire. | |
7231 | dout(7) << " inode expire on " << *in << " from mds." << from | |
7232 | << " with old nonce " << nonce | |
7233 | << " (current " << in->get_replica_nonce(from) << "), dropping" | |
7234 | << dendl; | |
7235 | } | |
7236 | } | |
7237 | ||
7238 | // DIRS | |
7239 | for (map<dirfrag_t,uint32_t>::iterator it = p->second.dirs.begin(); | |
7240 | it != p->second.dirs.end(); | |
7241 | ++it) { | |
7242 | CDir *dir = get_dirfrag(it->first); | |
7243 | unsigned nonce = it->second; | |
7244 | ||
7245 | if (!dir) { | |
7246 | CInode *diri = get_inode(it->first.ino); | |
7247 | if (diri) { | |
7248 | if (mds->is_rejoin() && | |
7249 | rejoin_ack_gather.count(mds->get_nodeid()) && // haven't sent rejoin ack yet | |
7250 | !diri->is_replica(from)) { | |
7251 | list<CDir*> ls; | |
7252 | diri->get_nested_dirfrags(ls); | |
7253 | dout(7) << " dir expire on dirfrag " << it->first << " from mds." << from | |
7254 | << " while rejoining, inode isn't replicated" << dendl; | |
7255 | for (list<CDir*>::iterator q = ls.begin(); q != ls.end(); ++q) { | |
7256 | dir = *q; | |
7257 | if (dir->is_replica(from)) { | |
7258 | dout(7) << " dir expire on " << *dir << " from mds." << from << dendl; | |
7259 | dir->remove_replica(from); | |
7260 | } | |
7261 | } | |
7262 | continue; | |
7263 | } | |
7264 | CDir *other = diri->get_approx_dirfrag(it->first.frag); | |
7265 | if (other) { | |
7266 | dout(7) << " dir expire on dirfrag " << it->first << " from mds." << from | |
7267 | << " have " << *other << ", mismatched frags, dropping" << dendl; | |
7268 | continue; | |
7269 | } | |
7270 | } | |
7271 | dout(0) << " dir expire on " << it->first << " from " << from | |
7272 | << ", don't have it" << dendl; | |
7273 | assert(dir); | |
7274 | } | |
7275 | dout(20) << __func__ << ": expiring dirfrag " << *dir << dendl; | |
7276 | ||
7277 | assert(dir->is_auth()); | |
7278 | ||
7279 | // check nonce | |
7280 | if (nonce == dir->get_replica_nonce(from)) { | |
7281 | // remove from our cached_by | |
7282 | dout(7) << " dir expire on " << *dir << " from mds." << from | |
181888fb | 7283 | << " replicas was " << dir->get_replicas() << dendl; |
7c673cae FG |
7284 | dir->remove_replica(from); |
7285 | } | |
7286 | else { | |
7287 | // this is an old nonce, ignore expire. | |
7288 | dout(7) << " dir expire on " << *dir << " from mds." << from | |
7289 | << " with old nonce " << nonce << " (current " << dir->get_replica_nonce(from) | |
7290 | << "), dropping" << dendl; | |
7291 | } | |
7292 | } | |
7293 | ||
7294 | // DENTRIES | |
7295 | for (map<dirfrag_t, map<pair<string,snapid_t>,uint32_t> >::iterator pd = p->second.dentries.begin(); | |
7296 | pd != p->second.dentries.end(); | |
7297 | ++pd) { | |
7298 | dout(10) << " dn expires in dir " << pd->first << dendl; | |
7299 | CInode *diri = get_inode(pd->first.ino); | |
7300 | assert(diri); | |
7301 | CDir *dir = diri->get_dirfrag(pd->first.frag); | |
7302 | ||
7303 | if (!dir) { | |
7304 | dout(0) << " dn expires on " << pd->first << " from " << from | |
7305 | << ", must have refragmented" << dendl; | |
7306 | } else { | |
7307 | assert(dir->is_auth()); | |
7308 | } | |
7309 | ||
7310 | for (map<pair<string,snapid_t>,uint32_t>::iterator p = pd->second.begin(); | |
7311 | p != pd->second.end(); | |
7312 | ++p) { | |
7313 | unsigned nonce = p->second; | |
7314 | CDentry *dn; | |
7315 | ||
7316 | if (dir) { | |
7317 | dn = dir->lookup(p->first.first, p->first.second); | |
7318 | } else { | |
7319 | // which dirfrag for this dentry? | |
7320 | CDir *dir = diri->get_dirfrag(diri->pick_dirfrag(p->first.first)); | |
7321 | assert(dir); | |
7322 | assert(dir->is_auth()); | |
7323 | dn = dir->lookup(p->first.first, p->first.second); | |
7324 | } | |
7325 | ||
7326 | if (!dn) { | |
7327 | if (dir) | |
7328 | dout(0) << " missing dentry for " << p->first.first << " snap " << p->first.second << " in " << *dir << dendl; | |
7329 | else | |
7330 | dout(0) << " missing dentry for " << p->first.first << " snap " << p->first.second << dendl; | |
7331 | } | |
7332 | assert(dn); | |
7333 | ||
7334 | if (nonce == dn->get_replica_nonce(from)) { | |
7335 | dout(7) << " dentry_expire on " << *dn << " from mds." << from << dendl; | |
7336 | dentry_remove_replica(dn, from, gather_locks); | |
7337 | } | |
7338 | else { | |
7339 | dout(7) << " dentry_expire on " << *dn << " from mds." << from | |
7340 | << " with old nonce " << nonce << " (current " << dn->get_replica_nonce(from) | |
7341 | << "), dropping" << dendl; | |
7342 | } | |
7343 | } | |
7344 | } | |
7345 | } | |
7346 | ||
7347 | // done | |
7348 | m->put(); | |
7349 | ||
7350 | for (set<SimpleLock*>::iterator p = gather_locks.begin(); p != gather_locks.end(); ++p) { | |
7351 | if (!(*p)->is_stable()) | |
7352 | mds->locker->eval_gather(*p); | |
7353 | } | |
7354 | } | |
7355 | ||
7356 | void MDCache::process_delayed_expire(CDir *dir) | |
7357 | { | |
7358 | dout(7) << "process_delayed_expire on " << *dir << dendl; | |
7359 | for (map<mds_rank_t,MCacheExpire*>::iterator p = delayed_expire[dir].begin(); | |
7360 | p != delayed_expire[dir].end(); | |
7361 | ++p) | |
7362 | handle_cache_expire(p->second); | |
7363 | delayed_expire.erase(dir); | |
7364 | } | |
7365 | ||
7366 | void MDCache::discard_delayed_expire(CDir *dir) | |
7367 | { | |
7368 | dout(7) << "discard_delayed_expire on " << *dir << dendl; | |
7369 | for (map<mds_rank_t,MCacheExpire*>::iterator p = delayed_expire[dir].begin(); | |
7370 | p != delayed_expire[dir].end(); | |
7371 | ++p) | |
7372 | p->second->put(); | |
7373 | delayed_expire.erase(dir); | |
7374 | } | |
7375 | ||
7376 | void MDCache::inode_remove_replica(CInode *in, mds_rank_t from, bool rejoin, | |
7377 | set<SimpleLock *>& gather_locks) | |
7378 | { | |
7379 | in->remove_replica(from); | |
7380 | in->mds_caps_wanted.erase(from); | |
7381 | ||
7382 | // note: this code calls _eval more often than it needs to! | |
7383 | // fix lock | |
7384 | if (in->authlock.remove_replica(from)) gather_locks.insert(&in->authlock); | |
7385 | if (in->linklock.remove_replica(from)) gather_locks.insert(&in->linklock); | |
7386 | if (in->snaplock.remove_replica(from)) gather_locks.insert(&in->snaplock); | |
7387 | if (in->xattrlock.remove_replica(from)) gather_locks.insert(&in->xattrlock); | |
7388 | if (in->flocklock.remove_replica(from)) gather_locks.insert(&in->flocklock); | |
7389 | if (in->policylock.remove_replica(from)) gather_locks.insert(&in->policylock); | |
7390 | ||
7391 | // If 'rejoin' is true and the scatter lock is in LOCK_MIX_* state. | |
7392 | // Don't remove the recovering mds from lock's gathering list because | |
7393 | // it may hold rejoined wrlocks. | |
7394 | if (in->dirfragtreelock.remove_replica(from, rejoin)) gather_locks.insert(&in->dirfragtreelock); | |
7395 | if (in->filelock.remove_replica(from, rejoin)) gather_locks.insert(&in->filelock); | |
7396 | if (in->nestlock.remove_replica(from, rejoin)) gather_locks.insert(&in->nestlock); | |
7397 | } | |
7398 | ||
7399 | void MDCache::dentry_remove_replica(CDentry *dn, mds_rank_t from, set<SimpleLock *>& gather_locks) | |
7400 | { | |
7401 | dn->remove_replica(from); | |
7402 | ||
7403 | // fix lock | |
7404 | if (dn->lock.remove_replica(from)) | |
7405 | gather_locks.insert(&dn->lock); | |
7406 | ||
7407 | // Replicated strays might now be elegible for purge | |
7408 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
7409 | if (dnl->is_primary()) { | |
7410 | maybe_eval_stray(dnl->get_inode()); | |
7411 | } | |
7412 | } | |
7413 | ||
7414 | void MDCache::trim_client_leases() | |
7415 | { | |
7416 | utime_t now = ceph_clock_now(); | |
7417 | ||
7418 | dout(10) << "trim_client_leases" << dendl; | |
7419 | ||
7420 | for (int pool=0; pool<client_lease_pools; pool++) { | |
7421 | int before = client_leases[pool].size(); | |
7422 | if (client_leases[pool].empty()) | |
7423 | continue; | |
7424 | ||
7425 | while (!client_leases[pool].empty()) { | |
7426 | ClientLease *r = client_leases[pool].front(); | |
7427 | if (r->ttl > now) break; | |
7428 | CDentry *dn = static_cast<CDentry*>(r->parent); | |
7429 | dout(10) << " expiring client." << r->client << " lease of " << *dn << dendl; | |
7430 | dn->remove_client_lease(r, mds->locker); | |
7431 | } | |
7432 | int after = client_leases[pool].size(); | |
7433 | dout(10) << "trim_client_leases pool " << pool << " trimmed " | |
7434 | << (before-after) << " leases, " << after << " left" << dendl; | |
7435 | } | |
7436 | } | |
7437 | ||
7438 | ||
7439 | void MDCache::check_memory_usage() | |
7440 | { | |
7441 | static MemoryModel mm(g_ceph_context); | |
7442 | static MemoryModel::snap last; | |
7443 | mm.sample(&last); | |
7444 | static MemoryModel::snap baseline = last; | |
7445 | ||
7446 | // check client caps | |
7447 | assert(CInode::count() == inode_map.size()); | |
181888fb | 7448 | double caps_per_inode = 0.0; |
7c673cae | 7449 | if (CInode::count()) |
181888fb | 7450 | caps_per_inode = (double)Capability::count() / (double)CInode::count(); |
7c673cae FG |
7451 | |
7452 | dout(2) << "check_memory_usage" | |
7453 | << " total " << last.get_total() | |
7454 | << ", rss " << last.get_rss() | |
7455 | << ", heap " << last.get_heap() | |
7456 | << ", baseline " << baseline.get_heap() | |
7457 | << ", buffers " << (buffer::get_total_alloc() >> 10) | |
7458 | << ", " << num_inodes_with_caps << " / " << CInode::count() << " inodes have caps" | |
7459 | << ", " << Capability::count() << " caps, " << caps_per_inode << " caps per inode" | |
7460 | << dendl; | |
7461 | ||
c07f9fc5 | 7462 | mds->update_mlogger(); |
7c673cae FG |
7463 | mds->mlogger->set(l_mdm_rss, last.get_rss()); |
7464 | mds->mlogger->set(l_mdm_heap, last.get_heap()); | |
7465 | ||
181888fb FG |
7466 | if (cache_toofull()) { |
7467 | last_recall_state = ceph_clock_now(); | |
7468 | mds->server->recall_client_state(); | |
7c673cae FG |
7469 | } |
7470 | ||
7471 | // If the cache size had exceeded its limit, but we're back in bounds | |
7472 | // now, free any unused pool memory so that our memory usage isn't | |
7473 | // permanently bloated. | |
181888fb | 7474 | if (exceeded_size_limit && !cache_toofull()) { |
7c673cae FG |
7475 | // Only do this once we are back in bounds: otherwise the releases would |
7476 | // slow down whatever process caused us to exceed bounds to begin with | |
7477 | if (ceph_using_tcmalloc()) { | |
7478 | dout(2) << "check_memory_usage: releasing unused space from tcmalloc" | |
7479 | << dendl; | |
7480 | ceph_heap_release_free_memory(); | |
7481 | } | |
7482 | exceeded_size_limit = false; | |
7483 | } | |
7484 | } | |
7485 | ||
7486 | ||
7487 | ||
7488 | // ========================================================================================= | |
7489 | // shutdown | |
7490 | ||
7491 | class C_MDC_ShutdownCheck : public MDCacheContext { | |
7492 | public: | |
7493 | explicit C_MDC_ShutdownCheck(MDCache *m) : MDCacheContext(m) {} | |
7494 | void finish(int) override { | |
7495 | mdcache->shutdown_check(); | |
7496 | } | |
7497 | }; | |
7498 | ||
7499 | void MDCache::shutdown_check() | |
7500 | { | |
7501 | dout(0) << "shutdown_check at " << ceph_clock_now() << dendl; | |
7502 | ||
7503 | // cache | |
7504 | char old_val[32] = { 0 }; | |
7505 | char *o = old_val; | |
7506 | g_conf->get_val("debug_mds", &o, sizeof(old_val)); | |
7507 | g_conf->set_val("debug_mds", "10"); | |
7508 | g_conf->apply_changes(NULL); | |
7509 | show_cache(); | |
7510 | g_conf->set_val("debug_mds", old_val); | |
7511 | g_conf->apply_changes(NULL); | |
7512 | mds->timer.add_event_after(g_conf->mds_shutdown_check, new C_MDC_ShutdownCheck(this)); | |
7513 | ||
7514 | // this | |
31f18b77 | 7515 | dout(0) << "lru size now " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; |
7c673cae FG |
7516 | dout(0) << "log len " << mds->mdlog->get_num_events() << dendl; |
7517 | ||
7518 | ||
7519 | if (mds->objecter->is_active()) { | |
7520 | dout(0) << "objecter still active" << dendl; | |
7521 | mds->objecter->dump_active(); | |
7522 | } | |
7523 | } | |
7524 | ||
7525 | ||
7526 | void MDCache::shutdown_start() | |
7527 | { | |
7528 | dout(2) << "shutdown_start" << dendl; | |
7529 | ||
7530 | if (g_conf->mds_shutdown_check) | |
7531 | mds->timer.add_event_after(g_conf->mds_shutdown_check, new C_MDC_ShutdownCheck(this)); | |
7532 | ||
7533 | // g_conf->debug_mds = 10; | |
7534 | } | |
7535 | ||
7536 | ||
7537 | ||
7538 | bool MDCache::shutdown_pass() | |
7539 | { | |
7540 | dout(7) << "shutdown_pass" << dendl; | |
7541 | ||
7542 | if (mds->is_stopped()) { | |
7543 | dout(7) << " already shut down" << dendl; | |
7544 | show_cache(); | |
7545 | show_subtrees(); | |
7546 | return true; | |
7547 | } | |
7548 | ||
7549 | // empty stray dir | |
7550 | if (!shutdown_export_strays()) { | |
7551 | dout(7) << "waiting for strays to migrate" << dendl; | |
7552 | return false; | |
7553 | } | |
7554 | ||
7555 | // drop our reference to our stray dir inode | |
7556 | for (int i = 0; i < NUM_STRAY; ++i) { | |
7557 | if (strays[i] && | |
7558 | strays[i]->state_test(CInode::STATE_STRAYPINNED)) { | |
7559 | strays[i]->state_clear(CInode::STATE_STRAYPINNED); | |
7560 | strays[i]->put(CInode::PIN_STRAY); | |
7561 | strays[i]->put_stickydirs(); | |
7562 | } | |
7563 | } | |
7564 | ||
7565 | // trim cache | |
181888fb | 7566 | trim(UINT64_MAX); |
31f18b77 | 7567 | dout(5) << "lru size now " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; |
7c673cae FG |
7568 | |
7569 | // SUBTREES | |
7570 | int num_auth_subtree = 0; | |
7571 | if (!subtrees.empty() && | |
7572 | mds->get_nodeid() != 0 && | |
7573 | migrator->get_export_queue_size() == 0) { | |
7574 | dout(7) << "looking for subtrees to export to mds0" << dendl; | |
7575 | list<CDir*> ls; | |
7576 | for (map<CDir*, set<CDir*> >::iterator it = subtrees.begin(); | |
7577 | it != subtrees.end(); | |
7578 | ++it) { | |
7579 | CDir *dir = it->first; | |
7580 | if (dir->get_inode()->is_mdsdir()) | |
7581 | continue; | |
7582 | if (dir->is_auth()) { | |
7583 | num_auth_subtree++; | |
7584 | if (dir->is_frozen() || | |
7585 | dir->is_freezing() || | |
7586 | dir->is_ambiguous_dir_auth() || | |
7587 | dir->state_test(CDir::STATE_EXPORTING)) | |
7588 | continue; | |
7589 | ls.push_back(dir); | |
7590 | } | |
7591 | } | |
7592 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
7593 | CDir *dir = *p; | |
7594 | mds_rank_t dest = dir->get_inode()->authority().first; | |
7595 | if (dest > 0 && !mds->mdsmap->is_active(dest)) | |
7596 | dest = 0; | |
7597 | dout(7) << "sending " << *dir << " back to mds." << dest << dendl; | |
7598 | migrator->export_dir_nicely(dir, dest); | |
7599 | } | |
7600 | } | |
7601 | ||
7602 | if (num_auth_subtree > 0) { | |
7603 | dout(7) << "still have " << num_auth_subtree << " auth subtrees" << dendl; | |
7604 | show_subtrees(); | |
7605 | return false; | |
7606 | } | |
7607 | ||
7608 | // close out any sessions (and open files!) before we try to trim the log, etc. | |
7609 | if (mds->sessionmap.have_unclosed_sessions()) { | |
7610 | if (!mds->server->terminating_sessions) | |
7611 | mds->server->terminate_sessions(); | |
7612 | return false; | |
7613 | } | |
7614 | ||
7615 | CDir *mydir = myin ? myin->get_dirfrag(frag_t()) : NULL; | |
7616 | if (mydir && !mydir->is_subtree_root()) | |
7617 | mydir = NULL; | |
7618 | ||
7619 | // subtrees map not empty yet? | |
7620 | if (subtrees.size() > (mydir ? 1 : 0)) { | |
7621 | dout(7) << "still have " << num_subtrees() << " subtrees" << dendl; | |
7622 | show_subtrees(); | |
7623 | migrator->show_importing(); | |
7624 | migrator->show_exporting(); | |
7625 | if (!migrator->is_importing() && !migrator->is_exporting()) | |
7626 | show_cache(); | |
7627 | return false; | |
7628 | } | |
7629 | assert(!migrator->is_exporting()); | |
7630 | assert(!migrator->is_importing()); | |
7631 | ||
7c673cae FG |
7632 | // flush what we can from the log |
7633 | mds->mdlog->trim(0); | |
7634 | if (mds->mdlog->get_num_segments() > 1) { | |
7635 | dout(7) << "still >1 segments, waiting for log to trim" << dendl; | |
7636 | return false; | |
7637 | } | |
7638 | ||
181888fb FG |
7639 | if ((myin && myin->is_auth_pinned()) || |
7640 | (mydir && mydir->is_auth_pinned())) { | |
7641 | dout(7) << "still have auth pinned objects" << dendl; | |
7642 | return false; | |
7643 | } | |
7644 | ||
7c673cae FG |
7645 | // (only do this once!) |
7646 | if (!mds->mdlog->is_capped()) { | |
7647 | dout(7) << "capping the log" << dendl; | |
7648 | mds->mdlog->cap(); | |
7649 | mds->mdlog->trim(); | |
7650 | } | |
7651 | ||
7652 | if (!mds->mdlog->empty()) { | |
7653 | dout(7) << "waiting for log to flush.. " << mds->mdlog->get_num_events() | |
7654 | << " in " << mds->mdlog->get_num_segments() << " segments" << dendl; | |
7655 | return false; | |
7656 | } | |
7657 | ||
7658 | if (!did_shutdown_log_cap) { | |
7659 | // flush journal header | |
7660 | dout(7) << "writing header for (now-empty) journal" << dendl; | |
7661 | assert(mds->mdlog->empty()); | |
7662 | mds->mdlog->write_head(0); | |
7663 | // NOTE: filer active checker below will block us until this completes. | |
7664 | did_shutdown_log_cap = true; | |
7665 | return false; | |
7666 | } | |
7667 | ||
7668 | // filer active? | |
7669 | if (mds->objecter->is_active()) { | |
7670 | dout(7) << "objecter still active" << dendl; | |
7671 | mds->objecter->dump_active(); | |
7672 | return false; | |
7673 | } | |
7674 | ||
7675 | // trim what we can from the cache | |
31f18b77 FG |
7676 | if (lru.lru_get_size() > 0 || bottom_lru.lru_get_size() > 0) { |
7677 | dout(7) << "there's still stuff in the cache: " << lru.lru_get_size() << "/" << bottom_lru.lru_get_size() << dendl; | |
7c673cae FG |
7678 | show_cache(); |
7679 | //dump(); | |
7680 | return false; | |
7681 | } | |
31f18b77 FG |
7682 | |
7683 | // make mydir subtree go away | |
7684 | if (mydir) { | |
7685 | if (mydir->get_num_ref() > 1) { // subtree pin | |
7686 | dout(7) << "there's still reference to mydir " << *mydir << dendl; | |
7687 | show_cache(); | |
7688 | return false; | |
7689 | } | |
7690 | ||
7691 | remove_subtree(mydir); | |
7692 | myin->close_dirfrag(mydir->get_frag()); | |
7693 | } | |
7694 | assert(subtrees.empty()); | |
7695 | ||
7696 | if (myin) | |
7697 | remove_inode(myin); | |
7c673cae FG |
7698 | |
7699 | // done! | |
7700 | dout(2) << "shutdown done." << dendl; | |
7701 | return true; | |
7702 | } | |
7703 | ||
7704 | bool MDCache::shutdown_export_strays() | |
7705 | { | |
7706 | if (mds->get_nodeid() == 0) | |
7707 | return true; | |
7708 | ||
7709 | dout(10) << "shutdown_export_strays" << dendl; | |
7710 | ||
7711 | bool mds0_active = mds->mdsmap->is_active(mds_rank_t(0)); | |
7712 | ||
7713 | bool done = true; | |
7714 | ||
7715 | list<CDir*> dfs; | |
7716 | for (int i = 0; i < NUM_STRAY; ++i) { | |
7717 | if (!strays[i]) { | |
7718 | continue; | |
7719 | } | |
7720 | strays[i]->get_dirfrags(dfs); | |
7721 | } | |
7722 | ||
7723 | for (std::list<CDir*>::iterator dfs_i = dfs.begin(); | |
7724 | dfs_i != dfs.end(); ++dfs_i) | |
7725 | { | |
7726 | CDir *dir = *dfs_i; | |
7727 | ||
7728 | if (!dir->is_complete()) { | |
7729 | dir->fetch(0); | |
7730 | done = false; | |
7731 | if (!mds0_active) | |
7732 | break; | |
7733 | } | |
7734 | ||
7735 | for (CDir::map_t::iterator p = dir->items.begin(); | |
7736 | p != dir->items.end(); | |
7737 | ++p) { | |
7738 | CDentry *dn = p->second; | |
7739 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
7740 | if (dnl->is_null()) | |
7741 | continue; | |
7742 | done = false; | |
7743 | if (!mds0_active) | |
7744 | break; | |
7745 | ||
7746 | if (dn->state_test(CDentry::STATE_PURGING)) { | |
7747 | // Don't try to migrate anything that is actually | |
7748 | // being purged right now | |
7749 | continue; | |
7750 | } | |
7751 | ||
7752 | if (shutdown_exported_strays.count(dnl->get_inode()->ino()) == 0) { | |
7753 | shutdown_exported_strays.insert(dnl->get_inode()->ino()); | |
7754 | stray_manager.migrate_stray(dn, mds_rank_t(0)); // send to root! | |
7755 | } else { | |
7756 | dout(10) << "already exporting " << *dn << dendl; | |
7757 | } | |
7758 | } | |
7759 | } | |
7760 | ||
7761 | return done; | |
7762 | } | |
7763 | ||
7764 | // ========= messaging ============== | |
7765 | ||
7766 | /* This function DOES put the passed message before returning */ | |
7767 | void MDCache::dispatch(Message *m) | |
7768 | { | |
7769 | switch (m->get_type()) { | |
7770 | ||
7771 | // RESOLVE | |
7772 | case MSG_MDS_RESOLVE: | |
7773 | handle_resolve(static_cast<MMDSResolve*>(m)); | |
7774 | break; | |
7775 | case MSG_MDS_RESOLVEACK: | |
7776 | handle_resolve_ack(static_cast<MMDSResolveAck*>(m)); | |
7777 | break; | |
7778 | ||
7779 | // REJOIN | |
7780 | case MSG_MDS_CACHEREJOIN: | |
7781 | handle_cache_rejoin(static_cast<MMDSCacheRejoin*>(m)); | |
7782 | break; | |
7783 | ||
7784 | case MSG_MDS_DISCOVER: | |
7785 | handle_discover(static_cast<MDiscover*>(m)); | |
7786 | break; | |
7787 | case MSG_MDS_DISCOVERREPLY: | |
7788 | handle_discover_reply(static_cast<MDiscoverReply*>(m)); | |
7789 | break; | |
7790 | ||
7791 | case MSG_MDS_DIRUPDATE: | |
7792 | handle_dir_update(static_cast<MDirUpdate*>(m)); | |
7793 | break; | |
7794 | ||
7795 | case MSG_MDS_CACHEEXPIRE: | |
7796 | handle_cache_expire(static_cast<MCacheExpire*>(m)); | |
7797 | break; | |
7798 | ||
7799 | case MSG_MDS_DENTRYLINK: | |
7800 | handle_dentry_link(static_cast<MDentryLink*>(m)); | |
7801 | break; | |
7802 | case MSG_MDS_DENTRYUNLINK: | |
7803 | handle_dentry_unlink(static_cast<MDentryUnlink*>(m)); | |
7804 | break; | |
7805 | ||
7806 | case MSG_MDS_FRAGMENTNOTIFY: | |
7807 | handle_fragment_notify(static_cast<MMDSFragmentNotify*>(m)); | |
7808 | break; | |
7809 | ||
7810 | case MSG_MDS_FINDINO: | |
7811 | handle_find_ino(static_cast<MMDSFindIno *>(m)); | |
7812 | break; | |
7813 | case MSG_MDS_FINDINOREPLY: | |
7814 | handle_find_ino_reply(static_cast<MMDSFindInoReply *>(m)); | |
7815 | break; | |
7816 | ||
7817 | case MSG_MDS_OPENINO: | |
7818 | handle_open_ino(static_cast<MMDSOpenIno *>(m)); | |
7819 | break; | |
7820 | case MSG_MDS_OPENINOREPLY: | |
7821 | handle_open_ino_reply(static_cast<MMDSOpenInoReply *>(m)); | |
7822 | break; | |
7823 | ||
7824 | default: | |
7825 | derr << "cache unknown message " << m->get_type() << dendl; | |
7826 | assert(0 == "cache unknown message"); | |
7827 | } | |
7828 | } | |
7829 | ||
7830 | MDSInternalContextBase *MDCache::_get_waiter(MDRequestRef& mdr, Message *req, MDSInternalContextBase *fin) | |
7831 | { | |
7832 | if (mdr) { | |
7833 | dout(20) << "_get_waiter retryrequest" << dendl; | |
7834 | return new C_MDS_RetryRequest(this, mdr); | |
7835 | } else if (req) { | |
7836 | dout(20) << "_get_waiter retrymessage" << dendl; | |
7837 | return new C_MDS_RetryMessage(mds, req); | |
7838 | } else { | |
7839 | return fin; | |
7840 | } | |
7841 | } | |
7842 | ||
7843 | int MDCache::path_traverse(MDRequestRef& mdr, Message *req, MDSInternalContextBase *fin, // who | |
7844 | const filepath& path, // what | |
7845 | vector<CDentry*> *pdnvec, // result | |
7846 | CInode **pin, | |
7847 | int onfail) | |
7848 | { | |
7849 | bool discover = (onfail == MDS_TRAVERSE_DISCOVER); | |
7850 | bool null_okay = (onfail == MDS_TRAVERSE_DISCOVERXLOCK); | |
7851 | bool forward = (onfail == MDS_TRAVERSE_FORWARD); | |
7852 | ||
7853 | assert(mdr || req || fin); | |
7854 | assert(!forward || mdr || req); // forward requires a request | |
7855 | ||
7856 | snapid_t snapid = CEPH_NOSNAP; | |
7857 | if (mdr) | |
7858 | mdr->snapid = snapid; | |
7859 | ||
7860 | client_t client = (mdr && mdr->reqid.name.is_client()) ? mdr->reqid.name.num() : -1; | |
7861 | ||
7862 | if (mds->logger) mds->logger->inc(l_mds_traverse); | |
7863 | ||
7864 | dout(7) << "traverse: opening base ino " << path.get_ino() << " snap " << snapid << dendl; | |
7865 | CInode *cur = get_inode(path.get_ino()); | |
7866 | if (cur == NULL) { | |
7867 | if (MDS_INO_IS_MDSDIR(path.get_ino())) | |
7868 | open_foreign_mdsdir(path.get_ino(), _get_waiter(mdr, req, fin)); | |
7869 | else { | |
7870 | //ceph_abort(); // hrm.. broken | |
7871 | return -ESTALE; | |
7872 | } | |
7873 | return 1; | |
7874 | } | |
7875 | if (cur->state_test(CInode::STATE_PURGING)) | |
7876 | return -ESTALE; | |
7877 | ||
7878 | // make sure snaprealm are open... | |
7879 | if (mdr && cur->snaprealm && !cur->snaprealm->is_open() && | |
7880 | !cur->snaprealm->open_parents(_get_waiter(mdr, req, fin))) { | |
7881 | return 1; | |
7882 | } | |
7883 | ||
7884 | // start trace | |
7885 | if (pdnvec) | |
7886 | pdnvec->clear(); | |
7887 | if (pin) | |
7888 | *pin = cur; | |
7889 | ||
7890 | unsigned depth = 0; | |
7891 | while (depth < path.depth()) { | |
7892 | dout(12) << "traverse: path seg depth " << depth << " '" << path[depth] | |
7893 | << "' snapid " << snapid << dendl; | |
7894 | ||
7895 | if (!cur->is_dir()) { | |
7896 | dout(7) << "traverse: " << *cur << " not a dir " << dendl; | |
7897 | return -ENOTDIR; | |
7898 | } | |
7899 | ||
7900 | // walk into snapdir? | |
7901 | if (path[depth].length() == 0) { | |
7902 | dout(10) << "traverse: snapdir" << dendl; | |
7903 | if (!mdr) | |
7904 | return -EINVAL; | |
7905 | snapid = CEPH_SNAPDIR; | |
7906 | mdr->snapid = snapid; | |
7907 | depth++; | |
7908 | continue; | |
7909 | } | |
7910 | // walk thru snapdir? | |
7911 | if (snapid == CEPH_SNAPDIR) { | |
7912 | if (!mdr) | |
7913 | return -EINVAL; | |
7914 | SnapRealm *realm = cur->find_snaprealm(); | |
7915 | snapid = realm->resolve_snapname(path[depth], cur->ino()); | |
7916 | dout(10) << "traverse: snap " << path[depth] << " -> " << snapid << dendl; | |
7917 | if (!snapid) | |
7918 | return -ENOENT; | |
7919 | mdr->snapid = snapid; | |
7920 | depth++; | |
7921 | continue; | |
7922 | } | |
7923 | ||
7924 | // open dir | |
7925 | frag_t fg = cur->pick_dirfrag(path[depth]); | |
7926 | CDir *curdir = cur->get_dirfrag(fg); | |
7927 | if (!curdir) { | |
7928 | if (cur->is_auth()) { | |
7929 | // parent dir frozen_dir? | |
7930 | if (cur->is_frozen()) { | |
7931 | dout(7) << "traverse: " << *cur << " is frozen, waiting" << dendl; | |
7932 | cur->add_waiter(CDir::WAIT_UNFREEZE, _get_waiter(mdr, req, fin)); | |
7933 | return 1; | |
7934 | } | |
7935 | curdir = cur->get_or_open_dirfrag(this, fg); | |
7936 | } else { | |
7937 | // discover? | |
7938 | dout(10) << "traverse: need dirfrag " << fg << ", doing discover from " << *cur << dendl; | |
7939 | discover_path(cur, snapid, path.postfixpath(depth), _get_waiter(mdr, req, fin), | |
7940 | null_okay); | |
7941 | if (mds->logger) mds->logger->inc(l_mds_traverse_discover); | |
7942 | return 1; | |
7943 | } | |
7944 | } | |
7945 | assert(curdir); | |
7946 | ||
7947 | #ifdef MDS_VERIFY_FRAGSTAT | |
7948 | if (curdir->is_complete()) | |
7949 | curdir->verify_fragstat(); | |
7950 | #endif | |
7951 | ||
7952 | // frozen? | |
7953 | /* | |
7954 | if (curdir->is_frozen()) { | |
7955 | // doh! | |
7956 | // FIXME: traverse is allowed? | |
7957 | dout(7) << "traverse: " << *curdir << " is frozen, waiting" << dendl; | |
7958 | curdir->add_waiter(CDir::WAIT_UNFREEZE, _get_waiter(mdr, req, fin)); | |
7959 | if (onfinish) delete onfinish; | |
7960 | return 1; | |
7961 | } | |
7962 | */ | |
7963 | ||
7964 | // Before doing dirfrag->dn lookup, compare with DamageTable's | |
7965 | // record of which dentries were unreadable | |
7966 | if (mds->damage_table.is_dentry_damaged(curdir, path[depth], snapid)) { | |
7967 | dout(4) << "traverse: stopped lookup at damaged dentry " | |
7968 | << *curdir << "/" << path[depth] << " snap=" << snapid << dendl; | |
7969 | return -EIO; | |
7970 | } | |
7971 | ||
7972 | // dentry | |
7973 | CDentry *dn = curdir->lookup(path[depth], snapid); | |
7974 | CDentry::linkage_t *dnl = dn ? dn->get_projected_linkage() : 0; | |
7975 | ||
7976 | // null and last_bit and xlocked by me? | |
7977 | if (dnl && dnl->is_null() && null_okay) { | |
7978 | dout(10) << "traverse: hit null dentry at tail of traverse, succeeding" << dendl; | |
7979 | if (pdnvec) | |
7980 | pdnvec->push_back(dn); | |
7981 | if (pin) | |
7982 | *pin = 0; | |
7983 | break; // done! | |
7984 | } | |
7985 | ||
7986 | if (dnl && | |
7987 | dn->lock.is_xlocked() && | |
7988 | dn->lock.get_xlock_by() != mdr && | |
7989 | !dn->lock.can_read(client) && | |
7990 | (dnl->is_null() || forward)) { | |
7991 | dout(10) << "traverse: xlocked dentry at " << *dn << dendl; | |
7992 | dn->lock.add_waiter(SimpleLock::WAIT_RD, _get_waiter(mdr, req, fin)); | |
7993 | if (mds->logger) mds->logger->inc(l_mds_traverse_lock); | |
7994 | mds->mdlog->flush(); | |
7995 | return 1; | |
7996 | } | |
7997 | ||
7998 | // can we conclude ENOENT? | |
7999 | if (dnl && dnl->is_null()) { | |
8000 | if (dn->lock.can_read(client) || | |
8001 | (dn->lock.is_xlocked() && dn->lock.get_xlock_by() == mdr)) { | |
8002 | dout(10) << "traverse: miss on null+readable dentry " << path[depth] << " " << *dn << dendl; | |
8003 | if (pdnvec) { | |
8004 | if (depth == path.depth() - 1) | |
8005 | pdnvec->push_back(dn); | |
8006 | else | |
8007 | pdnvec->clear(); // do not confuse likes of rdlock_path_pin_ref(); | |
8008 | } | |
8009 | return -ENOENT; | |
8010 | } else { | |
8011 | dout(10) << "miss on dentry " << *dn << ", can't read due to lock" << dendl; | |
8012 | dn->lock.add_waiter(SimpleLock::WAIT_RD, _get_waiter(mdr, req, fin)); | |
8013 | return 1; | |
8014 | } | |
8015 | } | |
8016 | ||
8017 | if (dnl && !dnl->is_null()) { | |
8018 | CInode *in = dnl->get_inode(); | |
8019 | ||
8020 | // do we have inode? | |
8021 | if (!in) { | |
8022 | assert(dnl->is_remote()); | |
8023 | // do i have it? | |
8024 | in = get_inode(dnl->get_remote_ino()); | |
8025 | if (in) { | |
8026 | dout(7) << "linking in remote in " << *in << dendl; | |
8027 | dn->link_remote(dnl, in); | |
8028 | } else { | |
8029 | dout(7) << "remote link to " << dnl->get_remote_ino() << ", which i don't have" << dendl; | |
8030 | assert(mdr); // we shouldn't hit non-primary dentries doing a non-mdr traversal! | |
8031 | if (mds->damage_table.is_remote_damaged(dnl->get_remote_ino())) { | |
8032 | dout(4) << "traverse: remote dentry points to damaged ino " | |
8033 | << *dn << dendl; | |
8034 | return -EIO; | |
8035 | } | |
8036 | open_remote_dentry(dn, true, _get_waiter(mdr, req, fin), | |
8037 | (null_okay && depth == path.depth() - 1)); | |
8038 | if (mds->logger) mds->logger->inc(l_mds_traverse_remote_ino); | |
8039 | return 1; | |
8040 | } | |
8041 | } | |
8042 | ||
8043 | cur = in; | |
8044 | // make sure snaprealm are open... | |
8045 | if (mdr && cur->snaprealm && !cur->snaprealm->is_open() && | |
8046 | !cur->snaprealm->open_parents(_get_waiter(mdr, req, fin))) { | |
8047 | return 1; | |
8048 | } | |
8049 | ||
8050 | // add to trace, continue. | |
8051 | touch_inode(cur); | |
8052 | if (pdnvec) | |
8053 | pdnvec->push_back(dn); | |
8054 | if (pin) | |
8055 | *pin = cur; | |
8056 | depth++; | |
8057 | continue; | |
8058 | } | |
8059 | ||
8060 | ||
8061 | // MISS. dentry doesn't exist. | |
8062 | dout(12) << "traverse: miss on dentry " << path[depth] << " in " << *curdir << dendl; | |
8063 | ||
8064 | if (curdir->is_auth()) { | |
8065 | // dentry is mine. | |
8066 | if (curdir->is_complete() || | |
8067 | (snapid == CEPH_NOSNAP && | |
8068 | curdir->has_bloom() && | |
8069 | !curdir->is_in_bloom(path[depth]))){ | |
8070 | // file not found | |
8071 | if (pdnvec) { | |
8072 | // instantiate a null dn? | |
8073 | if (depth < path.depth()-1){ | |
8074 | dout(20) << " didn't traverse full path; not returning pdnvec" << dendl; | |
8075 | dn = NULL; | |
8076 | } else if (dn) { | |
8077 | ceph_abort(); // should have fallen out in ->is_null() check above | |
8078 | } else if (curdir->is_frozen()) { | |
8079 | dout(20) << " not adding null to frozen dir " << dendl; | |
8080 | } else if (snapid < CEPH_MAXSNAP) { | |
8081 | dout(20) << " not adding null for snapid " << snapid << dendl; | |
8082 | } else { | |
8083 | // create a null dentry | |
8084 | dn = curdir->add_null_dentry(path[depth]); | |
8085 | dout(20) << " added null " << *dn << dendl; | |
8086 | } | |
8087 | if (dn) | |
8088 | pdnvec->push_back(dn); | |
8089 | else | |
8090 | pdnvec->clear(); // do not confuse likes of rdlock_path_pin_ref(); | |
8091 | } | |
8092 | return -ENOENT; | |
8093 | } else { | |
8094 | ||
8095 | // Check DamageTable for missing fragments before trying to fetch | |
8096 | // this | |
8097 | if (mds->damage_table.is_dirfrag_damaged(curdir)) { | |
8098 | dout(4) << "traverse: damaged dirfrag " << *curdir | |
8099 | << ", blocking fetch" << dendl; | |
8100 | return -EIO; | |
8101 | } | |
8102 | ||
8103 | // directory isn't complete; reload | |
8104 | dout(7) << "traverse: incomplete dir contents for " << *cur << ", fetching" << dendl; | |
8105 | touch_inode(cur); | |
8106 | curdir->fetch(_get_waiter(mdr, req, fin), path[depth]); | |
8107 | if (mds->logger) mds->logger->inc(l_mds_traverse_dir_fetch); | |
8108 | return 1; | |
8109 | } | |
8110 | } else { | |
8111 | // dirfrag/dentry is not mine. | |
8112 | mds_authority_t dauth = curdir->authority(); | |
8113 | ||
8114 | if (forward && | |
8115 | snapid && mdr && mdr->client_request && | |
8116 | (int)depth < mdr->client_request->get_num_fwd()) { | |
8117 | dout(7) << "traverse: snap " << snapid << " and depth " << depth | |
8118 | << " < fwd " << mdr->client_request->get_num_fwd() | |
8119 | << ", discovering instead of forwarding" << dendl; | |
8120 | discover = true; | |
8121 | } | |
8122 | ||
8123 | if ((discover || null_okay)) { | |
8124 | dout(7) << "traverse: discover from " << path[depth] << " from " << *curdir << dendl; | |
8125 | discover_path(curdir, snapid, path.postfixpath(depth), _get_waiter(mdr, req, fin), | |
8126 | null_okay); | |
8127 | if (mds->logger) mds->logger->inc(l_mds_traverse_discover); | |
8128 | return 1; | |
8129 | } | |
8130 | if (forward) { | |
8131 | // forward | |
8132 | dout(7) << "traverse: not auth for " << path << " in " << *curdir << dendl; | |
8133 | ||
8134 | if (curdir->is_ambiguous_auth()) { | |
8135 | // wait | |
8136 | dout(7) << "traverse: waiting for single auth in " << *curdir << dendl; | |
8137 | curdir->add_waiter(CDir::WAIT_SINGLEAUTH, _get_waiter(mdr, req, fin)); | |
8138 | return 1; | |
8139 | } | |
8140 | ||
8141 | dout(7) << "traverse: forwarding, not auth for " << *curdir << dendl; | |
8142 | ||
8143 | if (mdr) | |
8144 | request_forward(mdr, dauth.first); | |
8145 | else | |
8146 | mds->forward_message_mds(req, dauth.first); | |
8147 | ||
8148 | if (mds->logger) mds->logger->inc(l_mds_traverse_forward); | |
8149 | assert(fin == NULL); | |
8150 | return 2; | |
8151 | } | |
8152 | } | |
8153 | ||
8154 | ceph_abort(); // i shouldn't get here | |
8155 | } | |
8156 | ||
8157 | // success. | |
8158 | if (mds->logger) mds->logger->inc(l_mds_traverse_hit); | |
8159 | dout(10) << "path_traverse finish on snapid " << snapid << dendl; | |
8160 | if (mdr) | |
8161 | assert(mdr->snapid == snapid); | |
8162 | return 0; | |
8163 | } | |
8164 | ||
8165 | CInode *MDCache::cache_traverse(const filepath& fp) | |
8166 | { | |
8167 | dout(10) << "cache_traverse " << fp << dendl; | |
8168 | ||
8169 | CInode *in; | |
8170 | if (fp.get_ino()) | |
8171 | in = get_inode(fp.get_ino()); | |
8172 | else | |
8173 | in = root; | |
8174 | if (!in) | |
8175 | return NULL; | |
8176 | ||
8177 | for (unsigned i = 0; i < fp.depth(); i++) { | |
8178 | const string& dname = fp[i]; | |
8179 | frag_t fg = in->pick_dirfrag(dname); | |
8180 | dout(20) << " " << i << " " << dname << " frag " << fg << " from " << *in << dendl; | |
8181 | CDir *curdir = in->get_dirfrag(fg); | |
8182 | if (!curdir) | |
8183 | return NULL; | |
8184 | CDentry *dn = curdir->lookup(dname, CEPH_NOSNAP); | |
8185 | if (!dn) | |
8186 | return NULL; | |
8187 | in = dn->get_linkage()->get_inode(); | |
8188 | if (!in) | |
8189 | return NULL; | |
8190 | } | |
8191 | dout(10) << " got " << *in << dendl; | |
8192 | return in; | |
8193 | } | |
8194 | ||
8195 | ||
8196 | /** | |
8197 | * open_remote_dir -- open up a remote dirfrag | |
8198 | * | |
8199 | * @param diri base inode | |
8200 | * @param approxfg approximate fragment. | |
8201 | * @param fin completion callback | |
8202 | */ | |
8203 | void MDCache::open_remote_dirfrag(CInode *diri, frag_t approxfg, MDSInternalContextBase *fin) | |
8204 | { | |
8205 | dout(10) << "open_remote_dir on " << *diri << dendl; | |
7c673cae FG |
8206 | assert(diri->is_dir()); |
8207 | assert(!diri->is_auth()); | |
8208 | assert(diri->get_dirfrag(approxfg) == 0); | |
8209 | ||
224ce89b | 8210 | discover_dir_frag(diri, approxfg, fin); |
7c673cae FG |
8211 | } |
8212 | ||
8213 | ||
8214 | /** | |
8215 | * get_dentry_inode - get or open inode | |
8216 | * | |
8217 | * @param dn the dentry | |
8218 | * @param mdr current request | |
8219 | * | |
8220 | * will return inode for primary, or link up/open up remote link's inode as necessary. | |
8221 | * If it's not available right now, puts mdr on wait list and returns null. | |
8222 | */ | |
8223 | CInode *MDCache::get_dentry_inode(CDentry *dn, MDRequestRef& mdr, bool projected) | |
8224 | { | |
8225 | CDentry::linkage_t *dnl; | |
8226 | if (projected) | |
8227 | dnl = dn->get_projected_linkage(); | |
8228 | else | |
8229 | dnl = dn->get_linkage(); | |
8230 | ||
8231 | assert(!dnl->is_null()); | |
8232 | ||
8233 | if (dnl->is_primary()) | |
8234 | return dnl->inode; | |
8235 | ||
8236 | assert(dnl->is_remote()); | |
8237 | CInode *in = get_inode(dnl->get_remote_ino()); | |
8238 | if (in) { | |
8239 | dout(7) << "get_dentry_inode linking in remote in " << *in << dendl; | |
8240 | dn->link_remote(dnl, in); | |
8241 | return in; | |
8242 | } else { | |
8243 | dout(10) << "get_dentry_inode on remote dn, opening inode for " << *dn << dendl; | |
8244 | open_remote_dentry(dn, projected, new C_MDS_RetryRequest(this, mdr)); | |
8245 | return 0; | |
8246 | } | |
8247 | } | |
8248 | ||
8249 | struct C_MDC_OpenRemoteDentry : public MDCacheContext { | |
8250 | CDentry *dn; | |
8251 | inodeno_t ino; | |
8252 | MDSInternalContextBase *onfinish; | |
8253 | bool want_xlocked; | |
8254 | C_MDC_OpenRemoteDentry(MDCache *m, CDentry *d, inodeno_t i, MDSInternalContextBase *f, bool wx) : | |
31f18b77 FG |
8255 | MDCacheContext(m), dn(d), ino(i), onfinish(f), want_xlocked(wx) { |
8256 | dn->get(MDSCacheObject::PIN_PTRWAITER); | |
8257 | } | |
7c673cae FG |
8258 | void finish(int r) override { |
8259 | mdcache->_open_remote_dentry_finish(dn, ino, onfinish, want_xlocked, r); | |
31f18b77 | 8260 | dn->put(MDSCacheObject::PIN_PTRWAITER); |
7c673cae FG |
8261 | } |
8262 | }; | |
8263 | ||
8264 | void MDCache::open_remote_dentry(CDentry *dn, bool projected, MDSInternalContextBase *fin, bool want_xlocked) | |
8265 | { | |
8266 | dout(10) << "open_remote_dentry " << *dn << dendl; | |
8267 | CDentry::linkage_t *dnl = projected ? dn->get_projected_linkage() : dn->get_linkage(); | |
8268 | inodeno_t ino = dnl->get_remote_ino(); | |
8269 | int64_t pool = dnl->get_remote_d_type() == DT_DIR ? mds->mdsmap->get_metadata_pool() : -1; | |
8270 | open_ino(ino, pool, | |
8271 | new C_MDC_OpenRemoteDentry(this, dn, ino, fin, want_xlocked), true, want_xlocked); // backtrace | |
8272 | } | |
8273 | ||
8274 | void MDCache::_open_remote_dentry_finish(CDentry *dn, inodeno_t ino, MDSInternalContextBase *fin, | |
8275 | bool want_xlocked, int r) | |
8276 | { | |
8277 | if (r < 0) { | |
31f18b77 FG |
8278 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); |
8279 | if (dnl->is_remote() && dnl->get_remote_ino() == ino) { | |
7c673cae FG |
8280 | dout(0) << "open_remote_dentry_finish bad remote dentry " << *dn << dendl; |
8281 | dn->state_set(CDentry::STATE_BADREMOTEINO); | |
8282 | ||
8283 | std::string path; | |
8284 | CDir *dir = dn->get_dir(); | |
8285 | if (dir) { | |
31f18b77 FG |
8286 | dir->get_inode()->make_path_string(path); |
8287 | path = path + "/" + dn->get_name(); | |
7c673cae FG |
8288 | } |
8289 | ||
31f18b77 | 8290 | bool fatal = mds->damage_table.notify_remote_damaged(ino, path); |
7c673cae | 8291 | if (fatal) { |
31f18b77 FG |
8292 | mds->damaged(); |
8293 | ceph_abort(); // unreachable, damaged() respawns us | |
7c673cae | 8294 | } |
31f18b77 FG |
8295 | } else { |
8296 | r = 0; | |
8297 | } | |
7c673cae FG |
8298 | } |
8299 | fin->complete(r < 0 ? r : 0); | |
8300 | } | |
8301 | ||
8302 | ||
8303 | void MDCache::make_trace(vector<CDentry*>& trace, CInode *in) | |
8304 | { | |
8305 | // empty trace if we're a base inode | |
8306 | if (in->is_base()) | |
8307 | return; | |
8308 | ||
8309 | CInode *parent = in->get_parent_inode(); | |
8310 | assert(parent); | |
8311 | make_trace(trace, parent); | |
8312 | ||
8313 | CDentry *dn = in->get_parent_dn(); | |
8314 | dout(15) << "make_trace adding " << *dn << dendl; | |
8315 | trace.push_back(dn); | |
8316 | } | |
8317 | ||
8318 | ||
8319 | // ------------------------------------------------------------------------------- | |
8320 | // Open inode by inode number | |
8321 | ||
8322 | class C_IO_MDC_OpenInoBacktraceFetched : public MDCacheIOContext { | |
8323 | inodeno_t ino; | |
8324 | public: | |
8325 | bufferlist bl; | |
8326 | C_IO_MDC_OpenInoBacktraceFetched(MDCache *c, inodeno_t i) : | |
8327 | MDCacheIOContext(c), ino(i) {} | |
8328 | void finish(int r) override { | |
8329 | mdcache->_open_ino_backtrace_fetched(ino, bl, r); | |
8330 | } | |
8331 | }; | |
8332 | ||
8333 | struct C_MDC_OpenInoTraverseDir : public MDCacheContext { | |
8334 | inodeno_t ino; | |
8335 | MMDSOpenIno *msg; | |
8336 | bool parent; | |
8337 | public: | |
8338 | C_MDC_OpenInoTraverseDir(MDCache *c, inodeno_t i, MMDSOpenIno *m, bool p) : | |
8339 | MDCacheContext(c), ino(i), msg(m), parent(p) {} | |
8340 | void finish(int r) override { | |
8341 | if (r < 0 && !parent) | |
8342 | r = -EAGAIN; | |
8343 | if (msg) { | |
8344 | mdcache->handle_open_ino(msg, r); | |
8345 | return; | |
8346 | } | |
8347 | assert(mdcache->opening_inodes.count(ino)); | |
8348 | mdcache->_open_ino_traverse_dir(ino, mdcache->opening_inodes[ino], r); | |
8349 | } | |
8350 | }; | |
8351 | ||
8352 | struct C_MDC_OpenInoParentOpened : public MDCacheContext { | |
8353 | inodeno_t ino; | |
8354 | public: | |
8355 | C_MDC_OpenInoParentOpened(MDCache *c, inodeno_t i) : MDCacheContext(c), ino(i) {} | |
8356 | void finish(int r) override { | |
8357 | mdcache->_open_ino_parent_opened(ino, r); | |
8358 | } | |
8359 | }; | |
8360 | ||
8361 | void MDCache::_open_ino_backtrace_fetched(inodeno_t ino, bufferlist& bl, int err) | |
8362 | { | |
8363 | dout(10) << "_open_ino_backtrace_fetched ino " << ino << " errno " << err << dendl; | |
8364 | ||
8365 | assert(opening_inodes.count(ino)); | |
8366 | open_ino_info_t& info = opening_inodes[ino]; | |
8367 | ||
8368 | CInode *in = get_inode(ino); | |
8369 | if (in) { | |
8370 | dout(10) << " found cached " << *in << dendl; | |
8371 | open_ino_finish(ino, info, in->authority().first); | |
8372 | return; | |
8373 | } | |
8374 | ||
8375 | inode_backtrace_t backtrace; | |
8376 | if (err == 0) { | |
8377 | try { | |
8378 | ::decode(backtrace, bl); | |
8379 | } catch (const buffer::error &decode_exc) { | |
8380 | derr << "corrupt backtrace on ino x0" << std::hex << ino | |
8381 | << std::dec << ": " << decode_exc << dendl; | |
8382 | open_ino_finish(ino, info, -EIO); | |
8383 | return; | |
8384 | } | |
8385 | if (backtrace.pool != info.pool && backtrace.pool != -1) { | |
8386 | dout(10) << " old object in pool " << info.pool | |
8387 | << ", retrying pool " << backtrace.pool << dendl; | |
8388 | info.pool = backtrace.pool; | |
8389 | C_IO_MDC_OpenInoBacktraceFetched *fin = | |
8390 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8391 | fetch_backtrace(ino, info.pool, fin->bl, | |
8392 | new C_OnFinisher(fin, mds->finisher)); | |
8393 | return; | |
8394 | } | |
8395 | } else if (err == -ENOENT) { | |
8396 | int64_t meta_pool = mds->mdsmap->get_metadata_pool(); | |
8397 | if (info.pool != meta_pool) { | |
8398 | dout(10) << " no object in pool " << info.pool | |
8399 | << ", retrying pool " << meta_pool << dendl; | |
8400 | info.pool = meta_pool; | |
8401 | C_IO_MDC_OpenInoBacktraceFetched *fin = | |
8402 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8403 | fetch_backtrace(ino, info.pool, fin->bl, | |
8404 | new C_OnFinisher(fin, mds->finisher)); | |
8405 | return; | |
8406 | } | |
8407 | err = 0; // backtrace.ancestors.empty() is checked below | |
8408 | } | |
8409 | ||
8410 | if (err == 0) { | |
8411 | if (backtrace.ancestors.empty()) { | |
8412 | dout(10) << " got empty backtrace " << dendl; | |
8413 | err = -EIO; | |
8414 | } else if (!info.ancestors.empty()) { | |
8415 | if (info.ancestors[0] == backtrace.ancestors[0]) { | |
8416 | dout(10) << " got same parents " << info.ancestors[0] << " 2 times" << dendl; | |
8417 | err = -EINVAL; | |
8418 | } else { | |
8419 | info.last_err = 0; | |
8420 | } | |
8421 | } | |
8422 | } | |
8423 | if (err) { | |
8424 | dout(0) << " failed to open ino " << ino << " err " << err << "/" << info.last_err << dendl; | |
8425 | if (info.last_err) | |
8426 | err = info.last_err; | |
8427 | open_ino_finish(ino, info, err); | |
8428 | return; | |
8429 | } | |
8430 | ||
8431 | dout(10) << " got backtrace " << backtrace << dendl; | |
8432 | info.ancestors = backtrace.ancestors; | |
8433 | ||
8434 | _open_ino_traverse_dir(ino, info, 0); | |
8435 | } | |
8436 | ||
8437 | void MDCache::_open_ino_parent_opened(inodeno_t ino, int ret) | |
8438 | { | |
8439 | dout(10) << "_open_ino_parent_opened ino " << ino << " ret " << ret << dendl; | |
8440 | ||
8441 | assert(opening_inodes.count(ino)); | |
8442 | open_ino_info_t& info = opening_inodes[ino]; | |
8443 | ||
8444 | CInode *in = get_inode(ino); | |
8445 | if (in) { | |
8446 | dout(10) << " found cached " << *in << dendl; | |
8447 | open_ino_finish(ino, info, in->authority().first); | |
8448 | return; | |
8449 | } | |
8450 | ||
8451 | if (ret == mds->get_nodeid()) { | |
8452 | _open_ino_traverse_dir(ino, info, 0); | |
8453 | } else { | |
8454 | if (ret >= 0) { | |
8455 | mds_rank_t checked_rank = mds_rank_t(ret); | |
8456 | info.check_peers = true; | |
8457 | info.auth_hint = checked_rank; | |
8458 | info.checked.erase(checked_rank); | |
8459 | } | |
8460 | do_open_ino(ino, info, ret); | |
8461 | } | |
8462 | } | |
8463 | ||
8464 | void MDCache::_open_ino_traverse_dir(inodeno_t ino, open_ino_info_t& info, int ret) | |
8465 | { | |
8466 | dout(10) << __func__ << ": ino " << ino << " ret " << ret << dendl; | |
8467 | ||
8468 | CInode *in = get_inode(ino); | |
8469 | if (in) { | |
8470 | dout(10) << " found cached " << *in << dendl; | |
8471 | open_ino_finish(ino, info, in->authority().first); | |
8472 | return; | |
8473 | } | |
8474 | ||
8475 | if (ret) { | |
8476 | do_open_ino(ino, info, ret); | |
8477 | return; | |
8478 | } | |
8479 | ||
8480 | mds_rank_t hint = info.auth_hint; | |
8481 | ret = open_ino_traverse_dir(ino, NULL, info.ancestors, | |
8482 | info.discover, info.want_xlocked, &hint); | |
8483 | if (ret > 0) | |
8484 | return; | |
8485 | if (hint != mds->get_nodeid()) | |
8486 | info.auth_hint = hint; | |
8487 | do_open_ino(ino, info, ret); | |
8488 | } | |
8489 | ||
8490 | void MDCache::_open_ino_fetch_dir(inodeno_t ino, MMDSOpenIno *m, CDir *dir, bool parent) | |
8491 | { | |
8492 | if (dir->state_test(CDir::STATE_REJOINUNDEF)) | |
8493 | assert(dir->get_inode()->dirfragtree.is_leaf(dir->get_frag())); | |
8494 | dir->fetch(new C_MDC_OpenInoTraverseDir(this, ino, m, parent)); | |
8495 | } | |
8496 | ||
8497 | int MDCache::open_ino_traverse_dir(inodeno_t ino, MMDSOpenIno *m, | |
8498 | vector<inode_backpointer_t>& ancestors, | |
8499 | bool discover, bool want_xlocked, mds_rank_t *hint) | |
8500 | { | |
8501 | dout(10) << "open_ino_traverse_dir ino " << ino << " " << ancestors << dendl; | |
8502 | int err = 0; | |
8503 | for (unsigned i = 0; i < ancestors.size(); i++) { | |
8504 | CInode *diri = get_inode(ancestors[i].dirino); | |
8505 | ||
8506 | if (!diri) { | |
8507 | if (discover && MDS_INO_IS_MDSDIR(ancestors[i].dirino)) { | |
8508 | open_foreign_mdsdir(ancestors[i].dirino, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8509 | return 1; | |
8510 | } | |
8511 | continue; | |
8512 | } | |
8513 | ||
8514 | if (diri->state_test(CInode::STATE_REJOINUNDEF)) { | |
8515 | CDir *dir = diri->get_parent_dir(); | |
8516 | while (dir->state_test(CDir::STATE_REJOINUNDEF) && | |
8517 | dir->get_inode()->state_test(CInode::STATE_REJOINUNDEF)) | |
8518 | dir = dir->get_inode()->get_parent_dir(); | |
8519 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8520 | return 1; | |
8521 | } | |
8522 | ||
8523 | if (!diri->is_dir()) { | |
8524 | dout(10) << " " << *diri << " is not dir" << dendl; | |
8525 | if (i == 0) | |
8526 | err = -ENOTDIR; | |
8527 | break; | |
8528 | } | |
8529 | ||
8530 | string &name = ancestors[i].dname; | |
8531 | frag_t fg = diri->pick_dirfrag(name); | |
8532 | CDir *dir = diri->get_dirfrag(fg); | |
8533 | if (!dir) { | |
8534 | if (diri->is_auth()) { | |
8535 | if (diri->is_frozen()) { | |
8536 | dout(10) << " " << *diri << " is frozen, waiting " << dendl; | |
8537 | diri->add_waiter(CDir::WAIT_UNFREEZE, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8538 | return 1; | |
8539 | } | |
8540 | dir = diri->get_or_open_dirfrag(this, fg); | |
8541 | } else if (discover) { | |
8542 | open_remote_dirfrag(diri, fg, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8543 | return 1; | |
8544 | } | |
8545 | } | |
8546 | if (dir) { | |
8547 | inodeno_t next_ino = i > 0 ? ancestors[i - 1].dirino : ino; | |
8548 | CDentry *dn = dir->lookup(name); | |
8549 | CDentry::linkage_t *dnl = dn ? dn->get_linkage() : NULL; | |
8550 | if (dir->is_auth()) { | |
8551 | if (dnl && dnl->is_primary() && | |
8552 | dnl->get_inode()->state_test(CInode::STATE_REJOINUNDEF)) { | |
8553 | dout(10) << " fetching undef " << *dnl->get_inode() << dendl; | |
8554 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8555 | return 1; | |
8556 | } | |
8557 | ||
8558 | if (!dnl && !dir->is_complete() && | |
8559 | (!dir->has_bloom() || dir->is_in_bloom(name))) { | |
8560 | dout(10) << " fetching incomplete " << *dir << dendl; | |
8561 | _open_ino_fetch_dir(ino, m, dir, i == 0); | |
8562 | return 1; | |
8563 | } | |
8564 | ||
8565 | dout(10) << " no ino " << next_ino << " in " << *dir << dendl; | |
8566 | if (i == 0) | |
8567 | err = -ENOENT; | |
8568 | } else if (discover) { | |
8569 | if (!dnl) { | |
8570 | filepath path(name, 0); | |
8571 | discover_path(dir, CEPH_NOSNAP, path, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0), | |
8572 | (i == 0 && want_xlocked)); | |
8573 | return 1; | |
8574 | } | |
8575 | if (dnl->is_null() && !dn->lock.can_read(-1)) { | |
8576 | dout(10) << " null " << *dn << " is not readable, waiting" << dendl; | |
8577 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDC_OpenInoTraverseDir(this, ino, m, i == 0)); | |
8578 | return 1; | |
8579 | } | |
8580 | dout(10) << " no ino " << next_ino << " in " << *dir << dendl; | |
8581 | if (i == 0) | |
8582 | err = -ENOENT; | |
8583 | } | |
8584 | } | |
8585 | if (hint && i == 0) | |
8586 | *hint = dir ? dir->authority().first : diri->authority().first; | |
8587 | break; | |
8588 | } | |
8589 | return err; | |
8590 | } | |
8591 | ||
8592 | void MDCache::open_ino_finish(inodeno_t ino, open_ino_info_t& info, int ret) | |
8593 | { | |
8594 | dout(10) << "open_ino_finish ino " << ino << " ret " << ret << dendl; | |
8595 | ||
8596 | list<MDSInternalContextBase*> waiters; | |
8597 | waiters.swap(info.waiters); | |
8598 | opening_inodes.erase(ino); | |
8599 | finish_contexts(g_ceph_context, waiters, ret); | |
8600 | } | |
8601 | ||
8602 | void MDCache::do_open_ino(inodeno_t ino, open_ino_info_t& info, int err) | |
8603 | { | |
8604 | if (err < 0 && err != -EAGAIN) { | |
8605 | info.checked.clear(); | |
7c673cae FG |
8606 | info.checking = MDS_RANK_NONE; |
8607 | info.check_peers = true; | |
8608 | info.fetch_backtrace = true; | |
8609 | if (info.discover) { | |
8610 | info.discover = false; | |
8611 | info.ancestors.clear(); | |
8612 | } | |
8613 | if (err != -ENOENT && err != -ENOTDIR) | |
8614 | info.last_err = err; | |
8615 | } | |
8616 | ||
d2e6a577 FG |
8617 | if (info.check_peers || info.discover) { |
8618 | if (info.discover) { | |
8619 | // got backtrace from peer, but failed to find inode. re-check peers | |
8620 | info.discover = false; | |
8621 | info.ancestors.clear(); | |
8622 | info.checked.clear(); | |
8623 | } | |
7c673cae FG |
8624 | info.check_peers = false; |
8625 | info.checking = MDS_RANK_NONE; | |
8626 | do_open_ino_peer(ino, info); | |
8627 | } else if (info.fetch_backtrace) { | |
8628 | info.check_peers = true; | |
8629 | info.fetch_backtrace = false; | |
8630 | info.checking = mds->get_nodeid(); | |
8631 | info.checked.clear(); | |
7c673cae FG |
8632 | C_IO_MDC_OpenInoBacktraceFetched *fin = |
8633 | new C_IO_MDC_OpenInoBacktraceFetched(this, ino); | |
8634 | fetch_backtrace(ino, info.pool, fin->bl, | |
8635 | new C_OnFinisher(fin, mds->finisher)); | |
8636 | } else { | |
8637 | assert(!info.ancestors.empty()); | |
8638 | info.checking = mds->get_nodeid(); | |
8639 | open_ino(info.ancestors[0].dirino, mds->mdsmap->get_metadata_pool(), | |
8640 | new C_MDC_OpenInoParentOpened(this, ino), info.want_replica); | |
8641 | } | |
8642 | } | |
8643 | ||
8644 | void MDCache::do_open_ino_peer(inodeno_t ino, open_ino_info_t& info) | |
8645 | { | |
8646 | set<mds_rank_t> all, active; | |
8647 | mds->mdsmap->get_mds_set(all); | |
8648 | mds->mdsmap->get_clientreplay_or_active_or_stopping_mds_set(active); | |
8649 | if (mds->get_state() == MDSMap::STATE_REJOIN) | |
8650 | mds->mdsmap->get_mds_set(active, MDSMap::STATE_REJOIN); | |
8651 | ||
8652 | dout(10) << "do_open_ino_peer " << ino << " active " << active | |
8653 | << " all " << all << " checked " << info.checked << dendl; | |
8654 | ||
8655 | mds_rank_t peer = MDS_RANK_NONE; | |
8656 | if (info.auth_hint >= 0) { | |
8657 | if (active.count(info.auth_hint)) { | |
8658 | peer = info.auth_hint; | |
8659 | info.auth_hint = MDS_RANK_NONE; | |
8660 | } | |
8661 | } else { | |
8662 | for (set<mds_rank_t>::iterator p = active.begin(); p != active.end(); ++p) | |
8663 | if (*p != mds->get_nodeid() && info.checked.count(*p) == 0) { | |
8664 | peer = *p; | |
8665 | break; | |
8666 | } | |
8667 | } | |
8668 | if (peer < 0) { | |
d2e6a577 FG |
8669 | all.erase(mds->get_nodeid()); |
8670 | if (all != info.checked) { | |
7c673cae FG |
8671 | dout(10) << " waiting for more peers to be active" << dendl; |
8672 | } else { | |
8673 | dout(10) << " all MDS peers have been checked " << dendl; | |
8674 | do_open_ino(ino, info, 0); | |
8675 | } | |
8676 | } else { | |
8677 | info.checking = peer; | |
8678 | vector<inode_backpointer_t> *pa = NULL; | |
8679 | // got backtrace from peer or backtrace just fetched | |
8680 | if (info.discover || !info.fetch_backtrace) | |
8681 | pa = &info.ancestors; | |
8682 | mds->send_message_mds(new MMDSOpenIno(info.tid, ino, pa), peer); | |
8683 | } | |
8684 | } | |
8685 | ||
8686 | void MDCache::handle_open_ino(MMDSOpenIno *m, int err) | |
8687 | { | |
8688 | if (mds->get_state() < MDSMap::STATE_REJOIN && | |
8689 | mds->get_want_state() != CEPH_MDS_STATE_REJOIN) { | |
8690 | m->put(); | |
8691 | return; | |
8692 | } | |
8693 | ||
8694 | dout(10) << "handle_open_ino " << *m << " err " << err << dendl; | |
8695 | ||
8696 | inodeno_t ino = m->ino; | |
8697 | MMDSOpenInoReply *reply; | |
8698 | CInode *in = get_inode(ino); | |
8699 | if (in) { | |
8700 | dout(10) << " have " << *in << dendl; | |
8701 | reply = new MMDSOpenInoReply(m->get_tid(), ino, mds_rank_t(0)); | |
8702 | if (in->is_auth()) { | |
8703 | touch_inode(in); | |
8704 | while (1) { | |
8705 | CDentry *pdn = in->get_parent_dn(); | |
8706 | if (!pdn) | |
8707 | break; | |
8708 | CInode *diri = pdn->get_dir()->get_inode(); | |
8709 | reply->ancestors.push_back(inode_backpointer_t(diri->ino(), pdn->name, | |
8710 | in->inode.version)); | |
8711 | in = diri; | |
8712 | } | |
8713 | } else { | |
8714 | reply->hint = in->authority().first; | |
8715 | } | |
8716 | } else if (err < 0) { | |
8717 | reply = new MMDSOpenInoReply(m->get_tid(), ino, MDS_RANK_NONE, err); | |
8718 | } else { | |
8719 | mds_rank_t hint = MDS_RANK_NONE; | |
8720 | int ret = open_ino_traverse_dir(ino, m, m->ancestors, false, false, &hint); | |
8721 | if (ret > 0) | |
8722 | return; | |
8723 | reply = new MMDSOpenInoReply(m->get_tid(), ino, hint, ret); | |
8724 | } | |
8725 | m->get_connection()->send_message(reply); | |
8726 | m->put(); | |
8727 | } | |
8728 | ||
8729 | void MDCache::handle_open_ino_reply(MMDSOpenInoReply *m) | |
8730 | { | |
8731 | dout(10) << "handle_open_ino_reply " << *m << dendl; | |
8732 | ||
8733 | inodeno_t ino = m->ino; | |
8734 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
8735 | auto it = opening_inodes.find(ino); | |
8736 | if (it != opening_inodes.end() && it->second.checking == from) { | |
8737 | open_ino_info_t& info = it->second; | |
8738 | info.checking = MDS_RANK_NONE; | |
8739 | info.checked.insert(from); | |
8740 | ||
8741 | CInode *in = get_inode(ino); | |
8742 | if (in) { | |
8743 | dout(10) << " found cached " << *in << dendl; | |
8744 | open_ino_finish(ino, info, in->authority().first); | |
8745 | } else if (!m->ancestors.empty()) { | |
8746 | dout(10) << " found ino " << ino << " on mds." << from << dendl; | |
8747 | if (!info.want_replica) { | |
8748 | open_ino_finish(ino, info, from); | |
8749 | m->put(); | |
8750 | return; | |
8751 | } | |
8752 | ||
8753 | info.ancestors = m->ancestors; | |
8754 | info.auth_hint = from; | |
8755 | info.checking = mds->get_nodeid(); | |
8756 | info.discover = true; | |
8757 | _open_ino_traverse_dir(ino, info, 0); | |
8758 | } else if (m->error) { | |
8759 | dout(10) << " error " << m->error << " from mds." << from << dendl; | |
8760 | do_open_ino(ino, info, m->error); | |
8761 | } else { | |
8762 | if (m->hint >= 0 && m->hint != mds->get_nodeid()) { | |
8763 | info.auth_hint = m->hint; | |
8764 | info.checked.erase(m->hint); | |
8765 | } | |
8766 | do_open_ino_peer(ino, info); | |
8767 | } | |
8768 | } | |
8769 | m->put(); | |
8770 | } | |
8771 | ||
8772 | void MDCache::kick_open_ino_peers(mds_rank_t who) | |
8773 | { | |
8774 | dout(10) << "kick_open_ino_peers mds." << who << dendl; | |
8775 | ||
8776 | for (map<inodeno_t, open_ino_info_t>::iterator p = opening_inodes.begin(); | |
8777 | p != opening_inodes.end(); | |
8778 | ++p) { | |
8779 | open_ino_info_t& info = p->second; | |
8780 | if (info.checking == who) { | |
8781 | dout(10) << " kicking ino " << p->first << " who was checking mds." << who << dendl; | |
8782 | info.checking = MDS_RANK_NONE; | |
8783 | do_open_ino_peer(p->first, info); | |
8784 | } else if (info.checking == MDS_RANK_NONE) { | |
8785 | dout(10) << " kicking ino " << p->first << " who was waiting" << dendl; | |
8786 | do_open_ino_peer(p->first, info); | |
8787 | } | |
8788 | } | |
8789 | } | |
8790 | ||
8791 | void MDCache::open_ino(inodeno_t ino, int64_t pool, MDSInternalContextBase* fin, | |
8792 | bool want_replica, bool want_xlocked) | |
8793 | { | |
8794 | dout(10) << "open_ino " << ino << " pool " << pool << " want_replica " | |
8795 | << want_replica << dendl; | |
8796 | ||
8797 | if (opening_inodes.count(ino)) { | |
8798 | open_ino_info_t& info = opening_inodes[ino]; | |
8799 | if (want_replica) { | |
8800 | info.want_replica = true; | |
8801 | if (want_xlocked && !info.want_xlocked) { | |
8802 | if (!info.ancestors.empty()) { | |
8803 | CInode *diri = get_inode(info.ancestors[0].dirino); | |
8804 | if (diri) { | |
8805 | frag_t fg = diri->pick_dirfrag(info.ancestors[0].dname); | |
8806 | CDir *dir = diri->get_dirfrag(fg); | |
8807 | if (dir && !dir->is_auth()) { | |
8808 | filepath path(info.ancestors[0].dname, 0); | |
8809 | discover_path(dir, CEPH_NOSNAP, path, NULL, true); | |
8810 | } | |
8811 | } | |
8812 | } | |
8813 | info.want_xlocked = true; | |
8814 | } | |
8815 | } | |
8816 | info.waiters.push_back(fin); | |
8817 | } else { | |
8818 | open_ino_info_t& info = opening_inodes[ino]; | |
7c673cae FG |
8819 | info.want_replica = want_replica; |
8820 | info.want_xlocked = want_xlocked; | |
8821 | info.tid = ++open_ino_last_tid; | |
8822 | info.pool = pool >= 0 ? pool : default_file_layout.pool_id; | |
8823 | info.waiters.push_back(fin); | |
8824 | do_open_ino(ino, info, 0); | |
8825 | } | |
8826 | } | |
8827 | ||
8828 | /* ---------------------------- */ | |
8829 | ||
8830 | /* | |
8831 | * search for a given inode on MDS peers. optionally start with the given node. | |
8832 | ||
8833 | ||
8834 | TODO | |
8835 | - recover from mds node failure, recovery | |
8836 | - traverse path | |
8837 | ||
8838 | */ | |
8839 | void MDCache::find_ino_peers(inodeno_t ino, MDSInternalContextBase *c, mds_rank_t hint) | |
8840 | { | |
8841 | dout(5) << "find_ino_peers " << ino << " hint " << hint << dendl; | |
8842 | assert(!have_inode(ino)); | |
8843 | ||
8844 | ceph_tid_t tid = ++find_ino_peer_last_tid; | |
8845 | find_ino_peer_info_t& fip = find_ino_peer[tid]; | |
8846 | fip.ino = ino; | |
8847 | fip.tid = tid; | |
8848 | fip.fin = c; | |
8849 | fip.hint = hint; | |
7c673cae FG |
8850 | _do_find_ino_peer(fip); |
8851 | } | |
8852 | ||
8853 | void MDCache::_do_find_ino_peer(find_ino_peer_info_t& fip) | |
8854 | { | |
8855 | set<mds_rank_t> all, active; | |
8856 | mds->mdsmap->get_mds_set(all); | |
8857 | mds->mdsmap->get_clientreplay_or_active_or_stopping_mds_set(active); | |
8858 | ||
8859 | dout(10) << "_do_find_ino_peer " << fip.tid << " " << fip.ino | |
8860 | << " active " << active << " all " << all | |
8861 | << " checked " << fip.checked | |
8862 | << dendl; | |
8863 | ||
8864 | mds_rank_t m = MDS_RANK_NONE; | |
8865 | if (fip.hint >= 0) { | |
8866 | m = fip.hint; | |
8867 | fip.hint = MDS_RANK_NONE; | |
8868 | } else { | |
8869 | for (set<mds_rank_t>::iterator p = active.begin(); p != active.end(); ++p) | |
8870 | if (*p != mds->get_nodeid() && | |
8871 | fip.checked.count(*p) == 0) { | |
8872 | m = *p; | |
8873 | break; | |
8874 | } | |
8875 | } | |
8876 | if (m == MDS_RANK_NONE) { | |
d2e6a577 FG |
8877 | all.erase(mds->get_nodeid()); |
8878 | if (all != fip.checked) { | |
7c673cae FG |
8879 | dout(10) << "_do_find_ino_peer waiting for more peers to be active" << dendl; |
8880 | } else { | |
8881 | dout(10) << "_do_find_ino_peer failed on " << fip.ino << dendl; | |
8882 | fip.fin->complete(-ESTALE); | |
8883 | find_ino_peer.erase(fip.tid); | |
8884 | } | |
8885 | } else { | |
8886 | fip.checking = m; | |
8887 | mds->send_message_mds(new MMDSFindIno(fip.tid, fip.ino), m); | |
8888 | } | |
8889 | } | |
8890 | ||
8891 | void MDCache::handle_find_ino(MMDSFindIno *m) | |
8892 | { | |
8893 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
8894 | m->put(); | |
8895 | return; | |
8896 | } | |
8897 | ||
8898 | dout(10) << "handle_find_ino " << *m << dendl; | |
8899 | MMDSFindInoReply *r = new MMDSFindInoReply(m->tid); | |
8900 | CInode *in = get_inode(m->ino); | |
8901 | if (in) { | |
8902 | in->make_path(r->path); | |
8903 | dout(10) << " have " << r->path << " " << *in << dendl; | |
8904 | } | |
8905 | m->get_connection()->send_message(r); | |
8906 | m->put(); | |
8907 | } | |
8908 | ||
8909 | ||
8910 | void MDCache::handle_find_ino_reply(MMDSFindInoReply *m) | |
8911 | { | |
8912 | map<ceph_tid_t, find_ino_peer_info_t>::iterator p = find_ino_peer.find(m->tid); | |
8913 | if (p != find_ino_peer.end()) { | |
8914 | dout(10) << "handle_find_ino_reply " << *m << dendl; | |
8915 | find_ino_peer_info_t& fip = p->second; | |
8916 | ||
8917 | // success? | |
8918 | if (get_inode(fip.ino)) { | |
8919 | dout(10) << "handle_find_ino_reply successfully found " << fip.ino << dendl; | |
8920 | mds->queue_waiter(fip.fin); | |
8921 | find_ino_peer.erase(p); | |
8922 | m->put(); | |
8923 | return; | |
8924 | } | |
8925 | ||
8926 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
8927 | if (fip.checking == from) | |
8928 | fip.checking = MDS_RANK_NONE; | |
8929 | fip.checked.insert(from); | |
8930 | ||
8931 | if (!m->path.empty()) { | |
8932 | // we got a path! | |
8933 | vector<CDentry*> trace; | |
8934 | MDRequestRef null_ref; | |
8935 | int r = path_traverse(null_ref, m, NULL, m->path, &trace, NULL, MDS_TRAVERSE_DISCOVER); | |
8936 | if (r > 0) | |
8937 | return; | |
8938 | dout(0) << "handle_find_ino_reply failed with " << r << " on " << m->path | |
8939 | << ", retrying" << dendl; | |
8940 | fip.checked.clear(); | |
8941 | _do_find_ino_peer(fip); | |
8942 | } else { | |
8943 | // nope, continue. | |
8944 | _do_find_ino_peer(fip); | |
8945 | } | |
8946 | } else { | |
8947 | dout(10) << "handle_find_ino_reply tid " << m->tid << " dne" << dendl; | |
8948 | } | |
8949 | m->put(); | |
8950 | } | |
8951 | ||
8952 | void MDCache::kick_find_ino_peers(mds_rank_t who) | |
8953 | { | |
8954 | // find_ino_peers requests we should move on from | |
8955 | for (map<ceph_tid_t,find_ino_peer_info_t>::iterator p = find_ino_peer.begin(); | |
8956 | p != find_ino_peer.end(); | |
8957 | ++p) { | |
8958 | find_ino_peer_info_t& fip = p->second; | |
8959 | if (fip.checking == who) { | |
8960 | dout(10) << "kicking find_ino_peer " << fip.tid << " who was checking mds." << who << dendl; | |
8961 | fip.checking = MDS_RANK_NONE; | |
8962 | _do_find_ino_peer(fip); | |
8963 | } else if (fip.checking == MDS_RANK_NONE) { | |
8964 | dout(10) << "kicking find_ino_peer " << fip.tid << " who was waiting" << dendl; | |
8965 | _do_find_ino_peer(fip); | |
8966 | } | |
8967 | } | |
8968 | } | |
8969 | ||
8970 | /* ---------------------------- */ | |
8971 | ||
8972 | int MDCache::get_num_client_requests() | |
8973 | { | |
8974 | int count = 0; | |
8975 | for (ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.begin(); | |
8976 | p != active_requests.end(); | |
8977 | ++p) { | |
8978 | MDRequestRef& mdr = p->second; | |
8979 | if (mdr->reqid.name.is_client() && !mdr->is_slave()) | |
8980 | count++; | |
8981 | } | |
8982 | return count; | |
8983 | } | |
8984 | ||
8985 | /* This function takes over the reference to the passed Message */ | |
8986 | MDRequestRef MDCache::request_start(MClientRequest *req) | |
8987 | { | |
8988 | // did we win a forward race against a slave? | |
8989 | if (active_requests.count(req->get_reqid())) { | |
8990 | MDRequestRef& mdr = active_requests[req->get_reqid()]; | |
8991 | assert(mdr); | |
8992 | if (mdr->is_slave()) { | |
8993 | dout(10) << "request_start already had " << *mdr << ", waiting for finish" << dendl; | |
8994 | mdr->more()->waiting_for_finish.push_back(new C_MDS_RetryMessage(mds, req)); | |
8995 | } else { | |
8996 | dout(10) << "request_start already processing " << *mdr << ", dropping new msg" << dendl; | |
8997 | req->put(); | |
8998 | } | |
8999 | return MDRequestRef(); | |
9000 | } | |
9001 | ||
9002 | // register new client request | |
9003 | MDRequestImpl::Params params; | |
9004 | params.reqid = req->get_reqid(); | |
9005 | params.attempt = req->get_num_fwd(); | |
9006 | params.client_req = req; | |
9007 | params.initiated = req->get_recv_stamp(); | |
9008 | params.throttled = req->get_throttle_stamp(); | |
9009 | params.all_read = req->get_recv_complete_stamp(); | |
9010 | params.dispatched = req->get_dispatch_stamp(); | |
9011 | ||
9012 | MDRequestRef mdr = | |
9013 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params>(params); | |
9014 | active_requests[params.reqid] = mdr; | |
9015 | mdr->set_op_stamp(req->get_stamp()); | |
9016 | dout(7) << "request_start " << *mdr << dendl; | |
9017 | return mdr; | |
9018 | } | |
9019 | ||
9020 | MDRequestRef MDCache::request_start_slave(metareqid_t ri, __u32 attempt, Message *m) | |
9021 | { | |
9022 | int by = m->get_source().num(); | |
9023 | MDRequestImpl::Params params; | |
9024 | params.reqid = ri; | |
9025 | params.attempt = attempt; | |
9026 | params.triggering_slave_req = m; | |
9027 | params.slave_to = by; | |
9028 | params.initiated = m->get_recv_stamp(); | |
9029 | params.throttled = m->get_throttle_stamp(); | |
9030 | params.all_read = m->get_recv_complete_stamp(); | |
9031 | params.dispatched = m->get_dispatch_stamp(); | |
9032 | MDRequestRef mdr = | |
9033 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params>(params); | |
9034 | assert(active_requests.count(mdr->reqid) == 0); | |
9035 | active_requests[mdr->reqid] = mdr; | |
9036 | dout(7) << "request_start_slave " << *mdr << " by mds." << by << dendl; | |
9037 | return mdr; | |
9038 | } | |
9039 | ||
9040 | MDRequestRef MDCache::request_start_internal(int op) | |
9041 | { | |
9042 | MDRequestImpl::Params params; | |
9043 | params.reqid.name = entity_name_t::MDS(mds->get_nodeid()); | |
9044 | params.reqid.tid = mds->issue_tid(); | |
9045 | params.initiated = ceph_clock_now(); | |
9046 | params.internal_op = op; | |
9047 | MDRequestRef mdr = | |
9048 | mds->op_tracker.create_request<MDRequestImpl,MDRequestImpl::Params>(params); | |
9049 | ||
9050 | assert(active_requests.count(mdr->reqid) == 0); | |
9051 | active_requests[mdr->reqid] = mdr; | |
9052 | dout(7) << "request_start_internal " << *mdr << " op " << op << dendl; | |
9053 | return mdr; | |
9054 | } | |
9055 | ||
9056 | MDRequestRef MDCache::request_get(metareqid_t rid) | |
9057 | { | |
9058 | ceph::unordered_map<metareqid_t, MDRequestRef>::iterator p = active_requests.find(rid); | |
9059 | assert(p != active_requests.end()); | |
9060 | dout(7) << "request_get " << rid << " " << *p->second << dendl; | |
9061 | return p->second; | |
9062 | } | |
9063 | ||
9064 | void MDCache::request_finish(MDRequestRef& mdr) | |
9065 | { | |
9066 | dout(7) << "request_finish " << *mdr << dendl; | |
9067 | mdr->mark_event("finishing request"); | |
9068 | ||
9069 | // slave finisher? | |
9070 | if (mdr->has_more() && mdr->more()->slave_commit) { | |
9071 | Context *fin = mdr->more()->slave_commit; | |
9072 | mdr->more()->slave_commit = 0; | |
9073 | int ret; | |
9074 | if (mdr->aborted) { | |
9075 | mdr->aborted = false; | |
9076 | ret = -1; | |
9077 | mdr->more()->slave_rolling_back = true; | |
9078 | } else { | |
9079 | ret = 0; | |
9080 | mdr->committing = true; | |
9081 | } | |
9082 | fin->complete(ret); // this must re-call request_finish. | |
9083 | return; | |
9084 | } | |
9085 | ||
d2e6a577 FG |
9086 | switch(mdr->internal_op) { |
9087 | case CEPH_MDS_OP_FRAGMENTDIR: | |
9088 | logger->inc(l_mdss_ireq_fragmentdir); | |
9089 | break; | |
9090 | case CEPH_MDS_OP_EXPORTDIR: | |
9091 | logger->inc(l_mdss_ireq_exportdir); | |
9092 | break; | |
9093 | case CEPH_MDS_OP_ENQUEUE_SCRUB: | |
9094 | logger->inc(l_mdss_ireq_enqueue_scrub); | |
9095 | break; | |
9096 | case CEPH_MDS_OP_FLUSH: | |
9097 | logger->inc(l_mdss_ireq_flush); | |
9098 | break; | |
9099 | case CEPH_MDS_OP_REPAIR_FRAGSTATS: | |
9100 | logger->inc(l_mdss_ireq_fragstats); | |
9101 | break; | |
9102 | case CEPH_MDS_OP_REPAIR_INODESTATS: | |
9103 | logger->inc(l_mdss_ireq_inodestats); | |
9104 | break; | |
9105 | } | |
9106 | ||
7c673cae FG |
9107 | request_cleanup(mdr); |
9108 | } | |
9109 | ||
9110 | ||
9111 | void MDCache::request_forward(MDRequestRef& mdr, mds_rank_t who, int port) | |
9112 | { | |
9113 | mdr->mark_event("forwarding request"); | |
9114 | if (mdr->client_request && mdr->client_request->get_source().is_client()) { | |
9115 | dout(7) << "request_forward " << *mdr << " to mds." << who << " req " | |
9116 | << *mdr->client_request << dendl; | |
9117 | mds->forward_message_mds(mdr->client_request, who); | |
9118 | mdr->client_request = 0; | |
9119 | if (mds->logger) mds->logger->inc(l_mds_forward); | |
9120 | } else if (mdr->internal_op >= 0) { | |
9121 | dout(10) << "request_forward on internal op; cancelling" << dendl; | |
9122 | mdr->internal_op_finish->complete(-EXDEV); | |
9123 | } else { | |
9124 | dout(7) << "request_forward drop " << *mdr << " req " << *mdr->client_request | |
9125 | << " was from mds" << dendl; | |
9126 | } | |
9127 | request_cleanup(mdr); | |
9128 | } | |
9129 | ||
9130 | ||
9131 | void MDCache::dispatch_request(MDRequestRef& mdr) | |
9132 | { | |
9133 | if (mdr->client_request) { | |
9134 | mds->server->dispatch_client_request(mdr); | |
9135 | } else if (mdr->slave_request) { | |
9136 | mds->server->dispatch_slave_request(mdr); | |
9137 | } else { | |
9138 | switch (mdr->internal_op) { | |
9139 | case CEPH_MDS_OP_FRAGMENTDIR: | |
9140 | dispatch_fragment_dir(mdr); | |
9141 | break; | |
9142 | case CEPH_MDS_OP_EXPORTDIR: | |
9143 | migrator->dispatch_export_dir(mdr, 0); | |
9144 | break; | |
9145 | case CEPH_MDS_OP_ENQUEUE_SCRUB: | |
9146 | enqueue_scrub_work(mdr); | |
9147 | break; | |
9148 | case CEPH_MDS_OP_FLUSH: | |
9149 | flush_dentry_work(mdr); | |
9150 | break; | |
9151 | case CEPH_MDS_OP_REPAIR_FRAGSTATS: | |
9152 | repair_dirfrag_stats_work(mdr); | |
9153 | break; | |
9154 | case CEPH_MDS_OP_REPAIR_INODESTATS: | |
9155 | repair_inode_stats_work(mdr); | |
9156 | break; | |
9157 | default: | |
9158 | ceph_abort(); | |
9159 | } | |
9160 | } | |
9161 | } | |
9162 | ||
9163 | ||
9164 | void MDCache::request_drop_foreign_locks(MDRequestRef& mdr) | |
9165 | { | |
9166 | if (!mdr->has_more()) | |
9167 | return; | |
9168 | ||
9169 | // clean up slaves | |
9170 | // (will implicitly drop remote dn pins) | |
9171 | for (set<mds_rank_t>::iterator p = mdr->more()->slaves.begin(); | |
9172 | p != mdr->more()->slaves.end(); | |
9173 | ++p) { | |
9174 | MMDSSlaveRequest *r = new MMDSSlaveRequest(mdr->reqid, mdr->attempt, | |
9175 | MMDSSlaveRequest::OP_FINISH); | |
9176 | ||
9177 | if (mdr->killed && !mdr->committing) { | |
9178 | r->mark_abort(); | |
9179 | } else if (mdr->more()->srcdn_auth_mds == *p && | |
9180 | mdr->more()->inode_import.length() > 0) { | |
9181 | // information about rename imported caps | |
9182 | r->inode_export.claim(mdr->more()->inode_import); | |
9183 | } | |
9184 | ||
9185 | mds->send_message_mds(r, *p); | |
9186 | } | |
9187 | ||
9188 | /* strip foreign xlocks out of lock lists, since the OP_FINISH drops them | |
9189 | * implicitly. Note that we don't call the finishers -- there shouldn't | |
9190 | * be any on a remote lock and the request finish wakes up all | |
9191 | * the waiters anyway! */ | |
9192 | set<SimpleLock*>::iterator p = mdr->xlocks.begin(); | |
9193 | while (p != mdr->xlocks.end()) { | |
9194 | if ((*p)->get_parent()->is_auth()) | |
9195 | ++p; | |
9196 | else { | |
9197 | dout(10) << "request_drop_foreign_locks forgetting lock " << **p | |
9198 | << " on " << *(*p)->get_parent() << dendl; | |
9199 | (*p)->put_xlock(); | |
9200 | mdr->locks.erase(*p); | |
9201 | mdr->xlocks.erase(p++); | |
9202 | } | |
9203 | } | |
9204 | ||
9205 | map<SimpleLock*, mds_rank_t>::iterator q = mdr->remote_wrlocks.begin(); | |
9206 | while (q != mdr->remote_wrlocks.end()) { | |
9207 | dout(10) << "request_drop_foreign_locks forgetting remote_wrlock " << *q->first | |
9208 | << " on mds." << q->second | |
9209 | << " on " << *(q->first)->get_parent() << dendl; | |
9210 | mdr->locks.erase(q->first); | |
9211 | mdr->remote_wrlocks.erase(q++); | |
9212 | } | |
9213 | ||
9214 | mdr->more()->slaves.clear(); /* we no longer have requests out to them, and | |
9215 | * leaving them in can cause double-notifies as | |
9216 | * this function can get called more than once */ | |
9217 | } | |
9218 | ||
9219 | void MDCache::request_drop_non_rdlocks(MDRequestRef& mdr) | |
9220 | { | |
9221 | request_drop_foreign_locks(mdr); | |
9222 | mds->locker->drop_non_rdlocks(mdr.get()); | |
9223 | } | |
9224 | ||
9225 | void MDCache::request_drop_locks(MDRequestRef& mdr) | |
9226 | { | |
9227 | request_drop_foreign_locks(mdr); | |
9228 | mds->locker->drop_locks(mdr.get()); | |
9229 | } | |
9230 | ||
9231 | void MDCache::request_cleanup(MDRequestRef& mdr) | |
9232 | { | |
9233 | dout(15) << "request_cleanup " << *mdr << dendl; | |
9234 | ||
9235 | if (mdr->has_more()) { | |
9236 | if (mdr->more()->is_ambiguous_auth) | |
9237 | mdr->clear_ambiguous_auth(); | |
9238 | if (!mdr->more()->waiting_for_finish.empty()) | |
9239 | mds->queue_waiters(mdr->more()->waiting_for_finish); | |
9240 | } | |
9241 | ||
9242 | request_drop_locks(mdr); | |
9243 | ||
9244 | // drop (local) auth pins | |
9245 | mdr->drop_local_auth_pins(); | |
9246 | ||
9247 | // drop stickydirs | |
9248 | for (set<CInode*>::iterator p = mdr->stickydirs.begin(); | |
9249 | p != mdr->stickydirs.end(); | |
9250 | ++p) | |
9251 | (*p)->put_stickydirs(); | |
9252 | ||
9253 | mds->locker->kick_cap_releases(mdr); | |
9254 | ||
9255 | // drop cache pins | |
9256 | mdr->drop_pins(); | |
9257 | ||
9258 | // remove from session | |
9259 | mdr->item_session_request.remove_myself(); | |
9260 | ||
9261 | // remove from map | |
9262 | active_requests.erase(mdr->reqid); | |
9263 | ||
9264 | if (mds->logger) | |
9265 | log_stat(); | |
9266 | ||
9267 | mdr->mark_event("cleaned up request"); | |
9268 | } | |
9269 | ||
9270 | void MDCache::request_kill(MDRequestRef& mdr) | |
9271 | { | |
9272 | // rollback slave requests is tricky. just let the request proceed. | |
9273 | if (mdr->done_locking && mdr->has_more() && | |
9274 | (!mdr->more()->witnessed.empty() || !mdr->more()->waiting_on_slave.empty())) { | |
9275 | dout(10) << "request_kill " << *mdr << " -- already started slave requests, no-op" << dendl; | |
9276 | ||
9277 | assert(mdr->used_prealloc_ino == 0); | |
9278 | assert(mdr->prealloc_inos.empty()); | |
9279 | ||
9280 | mdr->session = NULL; | |
9281 | mdr->item_session_request.remove_myself(); | |
9282 | return; | |
9283 | } | |
9284 | ||
9285 | mdr->killed = true; | |
9286 | mdr->mark_event("killing request"); | |
9287 | ||
9288 | if (mdr->committing) { | |
9289 | dout(10) << "request_kill " << *mdr << " -- already committing, no-op" << dendl; | |
9290 | } else { | |
9291 | dout(10) << "request_kill " << *mdr << dendl; | |
9292 | request_cleanup(mdr); | |
9293 | } | |
9294 | } | |
9295 | ||
9296 | // ------------------------------------------------------------------------------- | |
9297 | // SNAPREALMS | |
9298 | ||
9299 | struct C_MDC_snaprealm_create_finish : public MDCacheLogContext { | |
9300 | MDRequestRef mdr; | |
9301 | MutationRef mut; | |
9302 | CInode *in; | |
9303 | C_MDC_snaprealm_create_finish(MDCache *c, MDRequestRef& m, | |
9304 | MutationRef& mu, CInode *i) : | |
9305 | MDCacheLogContext(c), mdr(m), mut(mu), in(i) {} | |
9306 | void finish(int r) override { | |
9307 | mdcache->_snaprealm_create_finish(mdr, mut, in); | |
9308 | } | |
9309 | }; | |
9310 | ||
9311 | void MDCache::snaprealm_create(MDRequestRef& mdr, CInode *in) | |
9312 | { | |
9313 | dout(10) << "snaprealm_create " << *in << dendl; | |
9314 | assert(!in->snaprealm); | |
9315 | ||
9316 | // allocate an id.. | |
9317 | if (!mdr->more()->stid) { | |
9318 | mds->snapclient->prepare_create_realm(in->ino(), &mdr->more()->stid, &mdr->more()->snapidbl, | |
9319 | new C_MDS_RetryRequest(this, mdr)); | |
9320 | return; | |
9321 | } | |
9322 | ||
9323 | MutationRef mut(new MutationImpl()); | |
9324 | mut->ls = mds->mdlog->get_current_segment(); | |
9325 | EUpdate *le = new EUpdate(mds->mdlog, "snaprealm_create"); | |
9326 | mds->mdlog->start_entry(le); | |
9327 | ||
9328 | le->metablob.add_table_transaction(TABLE_SNAP, mdr->more()->stid); | |
9329 | ||
9330 | inode_t *pi = in->project_inode(); | |
9331 | pi->version = in->pre_dirty(); | |
9332 | pi->rstat.rsnaprealms++; | |
9333 | ||
9334 | bufferlist::iterator p = mdr->more()->snapidbl.begin(); | |
9335 | snapid_t seq; | |
9336 | ::decode(seq, p); | |
9337 | ||
9338 | sr_t *newsnap = in->project_snaprealm(seq); | |
9339 | newsnap->seq = seq; | |
9340 | newsnap->last_created = seq; | |
9341 | ||
9342 | predirty_journal_parents(mut, &le->metablob, in, 0, PREDIRTY_PRIMARY); | |
9343 | journal_cow_inode(mut, &le->metablob, in); | |
9344 | le->metablob.add_primary_dentry(in->get_projected_parent_dn(), in, true); | |
9345 | ||
9346 | mds->server->submit_mdlog_entry(le, | |
9347 | new C_MDC_snaprealm_create_finish(this, mdr, | |
9348 | mut, in), | |
9349 | mdr, __func__); | |
9350 | mds->mdlog->flush(); | |
9351 | } | |
9352 | ||
9353 | ||
9354 | void MDCache::do_realm_invalidate_and_update_notify(CInode *in, int snapop, bool nosend) | |
9355 | { | |
9356 | dout(10) << "do_realm_invalidate_and_update_notify " << *in->snaprealm << " " << *in << dendl; | |
9357 | ||
9358 | vector<inodeno_t> split_inos; | |
9359 | vector<inodeno_t> split_realms; | |
9360 | ||
9361 | if (snapop == CEPH_SNAP_OP_SPLIT) { | |
9362 | // notify clients of update|split | |
9363 | for (elist<CInode*>::iterator p = in->snaprealm->inodes_with_caps.begin(member_offset(CInode, item_caps)); | |
9364 | !p.end(); ++p) | |
9365 | split_inos.push_back((*p)->ino()); | |
9366 | ||
9367 | for (set<SnapRealm*>::iterator p = in->snaprealm->open_children.begin(); | |
9368 | p != in->snaprealm->open_children.end(); | |
9369 | ++p) | |
9370 | split_realms.push_back((*p)->inode->ino()); | |
9371 | } | |
9372 | ||
9373 | bufferlist snapbl; | |
9374 | in->snaprealm->build_snap_trace(snapbl); | |
9375 | ||
9376 | set<SnapRealm*> past_children; | |
9377 | map<client_t, MClientSnap*> updates; | |
9378 | list<SnapRealm*> q; | |
9379 | q.push_back(in->snaprealm); | |
9380 | while (!q.empty()) { | |
9381 | SnapRealm *realm = q.front(); | |
9382 | q.pop_front(); | |
9383 | ||
9384 | dout(10) << " realm " << *realm << " on " << *realm->inode << dendl; | |
9385 | realm->invalidate_cached_snaps(); | |
9386 | ||
9387 | for (map<client_t, xlist<Capability*>* >::iterator p = realm->client_caps.begin(); | |
9388 | p != realm->client_caps.end(); | |
9389 | ++p) { | |
9390 | assert(!p->second->empty()); | |
9391 | if (!nosend && updates.count(p->first) == 0) { | |
9392 | MClientSnap *update = new MClientSnap(snapop); | |
9393 | update->head.split = in->ino(); | |
9394 | update->split_inos = split_inos; | |
9395 | update->split_realms = split_realms; | |
9396 | update->bl = snapbl; | |
9397 | updates[p->first] = update; | |
9398 | } | |
9399 | } | |
9400 | ||
9401 | if (snapop == CEPH_SNAP_OP_UPDATE || snapop == CEPH_SNAP_OP_DESTROY) { | |
9402 | for (set<SnapRealm*>::iterator p = realm->open_past_children.begin(); | |
9403 | p != realm->open_past_children.end(); | |
9404 | ++p) | |
9405 | past_children.insert(*p); | |
9406 | } | |
9407 | ||
9408 | // notify for active children, too. | |
9409 | dout(10) << " " << realm << " open_children are " << realm->open_children << dendl; | |
9410 | for (set<SnapRealm*>::iterator p = realm->open_children.begin(); | |
9411 | p != realm->open_children.end(); | |
9412 | ++p) | |
9413 | q.push_back(*p); | |
9414 | } | |
9415 | ||
9416 | if (!nosend) | |
9417 | send_snaps(updates); | |
9418 | ||
9419 | // notify past children and their descendants if we update/delete old snapshots | |
9420 | for (set<SnapRealm*>::iterator p = past_children.begin(); | |
9421 | p != past_children.end(); | |
9422 | ++p) | |
9423 | q.push_back(*p); | |
9424 | ||
9425 | while (!q.empty()) { | |
9426 | SnapRealm *realm = q.front(); | |
9427 | q.pop_front(); | |
9428 | ||
9429 | realm->invalidate_cached_snaps(); | |
9430 | ||
9431 | for (set<SnapRealm*>::iterator p = realm->open_children.begin(); | |
9432 | p != realm->open_children.end(); | |
9433 | ++p) { | |
9434 | if (past_children.count(*p) == 0) | |
9435 | q.push_back(*p); | |
9436 | } | |
9437 | ||
9438 | for (set<SnapRealm*>::iterator p = realm->open_past_children.begin(); | |
9439 | p != realm->open_past_children.end(); | |
9440 | ++p) { | |
9441 | if (past_children.count(*p) == 0) { | |
9442 | q.push_back(*p); | |
9443 | past_children.insert(*p); | |
9444 | } | |
9445 | } | |
9446 | } | |
9447 | ||
9448 | if (snapop == CEPH_SNAP_OP_DESTROY) { | |
9449 | // eval stray inodes if we delete snapshot from their past ancestor snaprealm | |
9450 | for (set<SnapRealm*>::iterator p = past_children.begin(); | |
9451 | p != past_children.end(); | |
9452 | ++p) | |
9453 | maybe_eval_stray((*p)->inode, true); | |
9454 | } | |
9455 | } | |
9456 | ||
9457 | void MDCache::_snaprealm_create_finish(MDRequestRef& mdr, MutationRef& mut, CInode *in) | |
9458 | { | |
9459 | dout(10) << "_snaprealm_create_finish " << *in << dendl; | |
9460 | ||
9461 | // apply | |
9462 | in->pop_and_dirty_projected_inode(mut->ls); | |
9463 | mut->apply(); | |
9464 | mds->locker->drop_locks(mut.get()); | |
9465 | mut->cleanup(); | |
9466 | ||
9467 | // tell table we've committed | |
9468 | mds->snapclient->commit(mdr->more()->stid, mut->ls); | |
9469 | ||
9470 | // create | |
9471 | bufferlist::iterator p = mdr->more()->snapidbl.begin(); | |
9472 | snapid_t seq; | |
9473 | ::decode(seq, p); | |
9474 | ||
9475 | in->open_snaprealm(); | |
9476 | in->snaprealm->srnode.seq = seq; | |
9477 | in->snaprealm->srnode.created = seq; | |
9478 | bool ok = in->snaprealm->_open_parents(NULL); | |
9479 | assert(ok); | |
9480 | ||
9481 | do_realm_invalidate_and_update_notify(in, CEPH_SNAP_OP_SPLIT); | |
9482 | ||
9483 | /* | |
9484 | static int count = 5; | |
9485 | if (--count == 0) | |
9486 | ceph_abort(); // hack test test ********** | |
9487 | */ | |
9488 | ||
9489 | // done. | |
9490 | mdr->more()->stid = 0; // caller will likely need to reuse this | |
9491 | dispatch_request(mdr); | |
9492 | } | |
9493 | ||
9494 | ||
9495 | // ------------------------------------------------------------------------------- | |
9496 | // STRAYS | |
9497 | ||
9498 | struct C_MDC_RetryScanStray : public MDCacheContext { | |
9499 | dirfrag_t next; | |
9500 | C_MDC_RetryScanStray(MDCache *c, dirfrag_t n) : MDCacheContext(c), next(n) { } | |
9501 | void finish(int r) override { | |
9502 | mdcache->scan_stray_dir(next); | |
9503 | } | |
9504 | }; | |
9505 | ||
9506 | void MDCache::scan_stray_dir(dirfrag_t next) | |
9507 | { | |
9508 | dout(10) << "scan_stray_dir " << next << dendl; | |
9509 | ||
9510 | list<CDir*> ls; | |
9511 | for (int i = 0; i < NUM_STRAY; ++i) { | |
9512 | if (strays[i]->ino() < next.ino) | |
9513 | continue; | |
9514 | strays[i]->get_dirfrags(ls); | |
9515 | } | |
9516 | ||
9517 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
9518 | CDir *dir = *p; | |
9519 | if (dir->dirfrag() < next) | |
9520 | continue; | |
9521 | if (!dir->is_complete()) { | |
9522 | dir->fetch(new C_MDC_RetryScanStray(this, dir->dirfrag())); | |
9523 | return; | |
9524 | } | |
9525 | for (CDir::map_t::iterator q = dir->items.begin(); q != dir->items.end(); ++q) { | |
9526 | CDentry *dn = q->second; | |
9527 | dn->state_set(CDentry::STATE_STRAY); | |
9528 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
9529 | if (dnl->is_primary()) { | |
9530 | CInode *in = dnl->get_inode(); | |
9531 | if (in->inode.nlink == 0) | |
9532 | in->state_set(CInode::STATE_ORPHAN); | |
9533 | maybe_eval_stray(in); | |
9534 | } | |
9535 | } | |
9536 | } | |
9537 | } | |
9538 | ||
7c673cae FG |
9539 | void MDCache::fetch_backtrace(inodeno_t ino, int64_t pool, bufferlist& bl, Context *fin) |
9540 | { | |
9541 | object_t oid = CInode::get_object_name(ino, frag_t(), ""); | |
9542 | mds->objecter->getxattr(oid, object_locator_t(pool), "parent", CEPH_NOSNAP, &bl, 0, fin); | |
9543 | } | |
9544 | ||
9545 | ||
9546 | ||
9547 | ||
9548 | ||
9549 | // ======================================================================================== | |
9550 | // DISCOVER | |
9551 | /* | |
9552 | ||
9553 | - for all discovers (except base_inos, e.g. root, stray), waiters are attached | |
9554 | to the parent metadata object in the cache (pinning it). | |
9555 | ||
9556 | - all discovers are tracked by tid, so that we can ignore potentially dup replies. | |
9557 | ||
9558 | */ | |
9559 | ||
9560 | void MDCache::_send_discover(discover_info_t& d) | |
9561 | { | |
9562 | MDiscover *dis = new MDiscover(d.ino, d.frag, d.snap, d.want_path, | |
9563 | d.want_base_dir, d.want_xlocked); | |
9564 | dis->set_tid(d.tid); | |
9565 | mds->send_message_mds(dis, d.mds); | |
9566 | } | |
9567 | ||
9568 | void MDCache::discover_base_ino(inodeno_t want_ino, | |
9569 | MDSInternalContextBase *onfinish, | |
9570 | mds_rank_t from) | |
9571 | { | |
9572 | dout(7) << "discover_base_ino " << want_ino << " from mds." << from << dendl; | |
9573 | if (waiting_for_base_ino[from].count(want_ino) == 0) { | |
9574 | discover_info_t& d = _create_discover(from); | |
9575 | d.ino = want_ino; | |
9576 | _send_discover(d); | |
9577 | } | |
9578 | waiting_for_base_ino[from][want_ino].push_back(onfinish); | |
9579 | } | |
9580 | ||
9581 | ||
9582 | void MDCache::discover_dir_frag(CInode *base, | |
9583 | frag_t approx_fg, | |
9584 | MDSInternalContextBase *onfinish, | |
9585 | mds_rank_t from) | |
9586 | { | |
9587 | if (from < 0) | |
9588 | from = base->authority().first; | |
9589 | ||
9590 | dirfrag_t df(base->ino(), approx_fg); | |
9591 | dout(7) << "discover_dir_frag " << df | |
9592 | << " from mds." << from << dendl; | |
9593 | ||
9594 | if (!base->is_waiting_for_dir(approx_fg) || !onfinish) { | |
9595 | discover_info_t& d = _create_discover(from); | |
9596 | d.pin_base(base); | |
9597 | d.ino = base->ino(); | |
9598 | d.frag = approx_fg; | |
9599 | d.want_base_dir = true; | |
9600 | _send_discover(d); | |
9601 | } | |
9602 | ||
9603 | if (onfinish) | |
9604 | base->add_dir_waiter(approx_fg, onfinish); | |
9605 | } | |
9606 | ||
9607 | struct C_MDC_RetryDiscoverPath : public MDCacheContext { | |
9608 | CInode *base; | |
9609 | snapid_t snapid; | |
9610 | filepath path; | |
9611 | mds_rank_t from; | |
9612 | C_MDC_RetryDiscoverPath(MDCache *c, CInode *b, snapid_t s, filepath &p, mds_rank_t f) : | |
9613 | MDCacheContext(c), base(b), snapid(s), path(p), from(f) {} | |
9614 | void finish(int r) override { | |
9615 | mdcache->discover_path(base, snapid, path, 0, from); | |
9616 | } | |
9617 | }; | |
9618 | ||
9619 | void MDCache::discover_path(CInode *base, | |
9620 | snapid_t snap, | |
9621 | filepath want_path, | |
9622 | MDSInternalContextBase *onfinish, | |
9623 | bool want_xlocked, | |
9624 | mds_rank_t from) | |
9625 | { | |
9626 | if (from < 0) | |
9627 | from = base->authority().first; | |
9628 | ||
9629 | dout(7) << "discover_path " << base->ino() << " " << want_path << " snap " << snap << " from mds." << from | |
9630 | << (want_xlocked ? " want_xlocked":"") | |
9631 | << dendl; | |
9632 | ||
9633 | if (base->is_ambiguous_auth()) { | |
9634 | dout(10) << " waiting for single auth on " << *base << dendl; | |
9635 | if (!onfinish) | |
9636 | onfinish = new C_MDC_RetryDiscoverPath(this, base, snap, want_path, from); | |
9637 | base->add_waiter(CInode::WAIT_SINGLEAUTH, onfinish); | |
9638 | return; | |
9639 | } else if (from == mds->get_nodeid()) { | |
9640 | list<MDSInternalContextBase*> finished; | |
9641 | base->take_waiting(CInode::WAIT_DIR, finished); | |
9642 | mds->queue_waiters(finished); | |
9643 | return; | |
9644 | } | |
9645 | ||
9646 | frag_t fg = base->pick_dirfrag(want_path[0]); | |
9647 | if ((want_xlocked && want_path.depth() == 1) || | |
9648 | !base->is_waiting_for_dir(fg) || !onfinish) { | |
9649 | discover_info_t& d = _create_discover(from); | |
9650 | d.ino = base->ino(); | |
9651 | d.pin_base(base); | |
9652 | d.frag = fg; | |
9653 | d.snap = snap; | |
9654 | d.want_path = want_path; | |
9655 | d.want_base_dir = true; | |
9656 | d.want_xlocked = want_xlocked; | |
9657 | _send_discover(d); | |
9658 | } | |
9659 | ||
9660 | // register + wait | |
9661 | if (onfinish) | |
9662 | base->add_dir_waiter(fg, onfinish); | |
9663 | } | |
9664 | ||
9665 | struct C_MDC_RetryDiscoverPath2 : public MDCacheContext { | |
9666 | CDir *base; | |
9667 | snapid_t snapid; | |
9668 | filepath path; | |
9669 | C_MDC_RetryDiscoverPath2(MDCache *c, CDir *b, snapid_t s, filepath &p) : | |
9670 | MDCacheContext(c), base(b), snapid(s), path(p) {} | |
9671 | void finish(int r) override { | |
9672 | mdcache->discover_path(base, snapid, path, 0); | |
9673 | } | |
9674 | }; | |
9675 | ||
9676 | void MDCache::discover_path(CDir *base, | |
9677 | snapid_t snap, | |
9678 | filepath want_path, | |
9679 | MDSInternalContextBase *onfinish, | |
9680 | bool want_xlocked) | |
9681 | { | |
9682 | mds_rank_t from = base->authority().first; | |
9683 | ||
9684 | dout(7) << "discover_path " << base->dirfrag() << " " << want_path << " snap " << snap << " from mds." << from | |
9685 | << (want_xlocked ? " want_xlocked":"") | |
9686 | << dendl; | |
9687 | ||
9688 | if (base->is_ambiguous_auth()) { | |
9689 | dout(7) << " waiting for single auth on " << *base << dendl; | |
9690 | if (!onfinish) | |
9691 | onfinish = new C_MDC_RetryDiscoverPath2(this, base, snap, want_path); | |
9692 | base->add_waiter(CDir::WAIT_SINGLEAUTH, onfinish); | |
9693 | return; | |
9694 | } else if (from == mds->get_nodeid()) { | |
9695 | list<MDSInternalContextBase*> finished; | |
9696 | base->take_sub_waiting(finished); | |
9697 | mds->queue_waiters(finished); | |
9698 | return; | |
9699 | } | |
9700 | ||
9701 | if ((want_xlocked && want_path.depth() == 1) || | |
9702 | !base->is_waiting_for_dentry(want_path[0].c_str(), snap) || !onfinish) { | |
9703 | discover_info_t& d = _create_discover(from); | |
9704 | d.ino = base->ino(); | |
31f18b77 | 9705 | d.pin_base(base->inode); |
7c673cae FG |
9706 | d.frag = base->get_frag(); |
9707 | d.snap = snap; | |
9708 | d.want_path = want_path; | |
9709 | d.want_base_dir = false; | |
9710 | d.want_xlocked = want_xlocked; | |
9711 | _send_discover(d); | |
9712 | } | |
9713 | ||
9714 | // register + wait | |
9715 | if (onfinish) | |
9716 | base->add_dentry_waiter(want_path[0], snap, onfinish); | |
9717 | } | |
9718 | ||
9719 | void MDCache::kick_discovers(mds_rank_t who) | |
9720 | { | |
9721 | for (map<ceph_tid_t,discover_info_t>::iterator p = discovers.begin(); | |
9722 | p != discovers.end(); | |
9723 | ++p) { | |
9724 | if (p->second.mds != who) | |
9725 | continue; | |
9726 | _send_discover(p->second); | |
9727 | } | |
9728 | } | |
9729 | ||
9730 | ||
9731 | /* This function DOES put the passed message before returning */ | |
9732 | void MDCache::handle_discover(MDiscover *dis) | |
9733 | { | |
9734 | mds_rank_t whoami = mds->get_nodeid(); | |
9735 | mds_rank_t from = mds_rank_t(dis->get_source().num()); | |
9736 | ||
9737 | assert(from != whoami); | |
9738 | ||
9739 | if (mds->get_state() <= MDSMap::STATE_REJOIN) { | |
9740 | if (mds->get_state() < MDSMap::STATE_REJOIN && | |
d2e6a577 | 9741 | mds->get_want_state() < CEPH_MDS_STATE_REJOIN) { |
7c673cae FG |
9742 | dis->put(); |
9743 | return; | |
9744 | } | |
9745 | ||
9746 | // proceed if requester is in the REJOIN stage, the request is from parallel_fetch(). | |
9747 | // delay processing request from survivor because we may not yet choose lock states. | |
9748 | if (!mds->mdsmap->is_rejoin(from)) { | |
9749 | dout(0) << "discover_reply not yet active(|still rejoining), delaying" << dendl; | |
9750 | mds->wait_for_replay(new C_MDS_RetryMessage(mds, dis)); | |
9751 | return; | |
9752 | } | |
9753 | } | |
9754 | ||
9755 | ||
9756 | CInode *cur = 0; | |
9757 | MDiscoverReply *reply = new MDiscoverReply(dis); | |
9758 | ||
9759 | snapid_t snapid = dis->get_snapid(); | |
9760 | ||
9761 | // get started. | |
9762 | if (MDS_INO_IS_BASE(dis->get_base_ino()) && | |
9763 | !dis->wants_base_dir() && dis->get_want().depth() == 0) { | |
9764 | // wants root | |
9765 | dout(7) << "handle_discover from mds." << from | |
9766 | << " wants base + " << dis->get_want().get_path() | |
9767 | << " snap " << snapid | |
9768 | << dendl; | |
9769 | ||
9770 | cur = get_inode(dis->get_base_ino()); | |
9771 | assert(cur); | |
9772 | ||
9773 | // add root | |
9774 | reply->starts_with = MDiscoverReply::INODE; | |
9775 | replicate_inode(cur, from, reply->trace, mds->mdsmap->get_up_features()); | |
9776 | dout(10) << "added base " << *cur << dendl; | |
9777 | } | |
9778 | else { | |
9779 | // there's a base inode | |
9780 | cur = get_inode(dis->get_base_ino(), snapid); | |
9781 | if (!cur && snapid != CEPH_NOSNAP) { | |
9782 | cur = get_inode(dis->get_base_ino()); | |
9783 | if (cur && !cur->is_multiversion()) | |
9784 | cur = NULL; // nope! | |
9785 | } | |
9786 | ||
9787 | if (!cur) { | |
9788 | dout(7) << "handle_discover mds." << from | |
9789 | << " don't have base ino " << dis->get_base_ino() << "." << snapid | |
9790 | << dendl; | |
9791 | if (!dis->wants_base_dir() && dis->get_want().depth() > 0) | |
9792 | reply->set_error_dentry(dis->get_dentry(0)); | |
9793 | reply->set_flag_error_dir(); | |
9794 | } else if (dis->wants_base_dir()) { | |
9795 | dout(7) << "handle_discover mds." << from | |
9796 | << " wants basedir+" << dis->get_want().get_path() | |
9797 | << " has " << *cur | |
9798 | << dendl; | |
9799 | } else { | |
9800 | dout(7) << "handle_discover mds." << from | |
9801 | << " wants " << dis->get_want().get_path() | |
9802 | << " has " << *cur | |
9803 | << dendl; | |
9804 | } | |
9805 | } | |
9806 | ||
9807 | assert(reply); | |
9808 | ||
9809 | // add content | |
9810 | // do some fidgeting to include a dir if they asked for the base dir, or just root. | |
9811 | for (unsigned i = 0; | |
9812 | cur && (i < dis->get_want().depth() || dis->get_want().depth() == 0); | |
9813 | i++) { | |
9814 | ||
9815 | // -- figure out the dir | |
9816 | ||
9817 | // is *cur even a dir at all? | |
9818 | if (!cur->is_dir()) { | |
9819 | dout(7) << *cur << " not a dir" << dendl; | |
9820 | reply->set_flag_error_dir(); | |
9821 | break; | |
9822 | } | |
9823 | ||
9824 | // pick frag | |
9825 | frag_t fg; | |
9826 | if (dis->get_want().depth()) { | |
9827 | // dentry specifies | |
9828 | fg = cur->pick_dirfrag(dis->get_dentry(i)); | |
9829 | } else { | |
9830 | // requester explicity specified the frag | |
9831 | assert(dis->wants_base_dir() || MDS_INO_IS_BASE(dis->get_base_ino())); | |
9832 | fg = dis->get_base_dir_frag(); | |
9833 | if (!cur->dirfragtree.is_leaf(fg)) | |
9834 | fg = cur->dirfragtree[fg.value()]; | |
9835 | } | |
9836 | CDir *curdir = cur->get_dirfrag(fg); | |
9837 | ||
9838 | if ((!curdir && !cur->is_auth()) || | |
9839 | (curdir && !curdir->is_auth())) { | |
9840 | ||
9841 | /* before: | |
9842 | * ONLY set flag if empty!! | |
9843 | * otherwise requester will wake up waiter(s) _and_ continue with discover, | |
9844 | * resulting in duplicate discovers in flight, | |
9845 | * which can wreak havoc when discovering rename srcdn (which may move) | |
9846 | */ | |
9847 | ||
9848 | if (reply->is_empty()) { | |
9849 | // only hint if empty. | |
9850 | // someday this could be better, but right now the waiter logic isn't smart enough. | |
9851 | ||
9852 | // hint | |
9853 | if (curdir) { | |
9854 | dout(7) << " not dirfrag auth, setting dir_auth_hint for " << *curdir << dendl; | |
9855 | reply->set_dir_auth_hint(curdir->authority().first); | |
9856 | } else { | |
9857 | dout(7) << " dirfrag not open, not inode auth, setting dir_auth_hint for " | |
9858 | << *cur << dendl; | |
9859 | reply->set_dir_auth_hint(cur->authority().first); | |
9860 | } | |
9861 | ||
9862 | // note error dentry, if any | |
9863 | // NOTE: important, as it allows requester to issue an equivalent discover | |
9864 | // to whomever we hint at. | |
9865 | if (dis->get_want().depth() > i) | |
9866 | reply->set_error_dentry(dis->get_dentry(i)); | |
9867 | } | |
9868 | ||
9869 | break; | |
9870 | } | |
9871 | ||
9872 | if (!curdir) { // open dir? | |
9873 | if (cur->is_frozen()) { | |
9874 | if (!reply->is_empty()) { | |
9875 | dout(7) << *cur << " is frozen, non-empty reply, stopping" << dendl; | |
9876 | break; | |
9877 | } | |
9878 | dout(7) << *cur << " is frozen, empty reply, waiting" << dendl; | |
9879 | cur->add_waiter(CInode::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
9880 | reply->put(); | |
9881 | return; | |
9882 | } | |
9883 | curdir = cur->get_or_open_dirfrag(this, fg); | |
9884 | } else if (curdir->is_frozen_tree() || | |
9885 | (curdir->is_frozen_dir() && fragment_are_all_frozen(curdir))) { | |
31f18b77 FG |
9886 | if (!reply->is_empty()) { |
9887 | dout(7) << *curdir << " is frozen, non-empty reply, stopping" << dendl; | |
9888 | break; | |
9889 | } | |
7c673cae FG |
9890 | if (dis->wants_base_dir() && dis->get_base_dir_frag() != curdir->get_frag()) { |
9891 | dout(7) << *curdir << " is frozen, dirfrag mismatch, stopping" << dendl; | |
9892 | reply->set_flag_error_dir(); | |
9893 | break; | |
9894 | } | |
7c673cae FG |
9895 | dout(7) << *curdir << " is frozen, empty reply, waiting" << dendl; |
9896 | curdir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
9897 | reply->put(); | |
9898 | return; | |
9899 | } | |
9900 | ||
9901 | // add dir | |
9902 | if (curdir->get_version() == 0) { | |
9903 | // fetch newly opened dir | |
9904 | } else if (reply->is_empty() && !dis->wants_base_dir()) { | |
9905 | dout(7) << "handle_discover not adding unwanted base dir " << *curdir << dendl; | |
9906 | // make sure the base frag is correct, though, in there was a refragment since the | |
9907 | // original request was sent. | |
9908 | reply->set_base_dir_frag(curdir->get_frag()); | |
9909 | } else { | |
9910 | assert(!curdir->is_ambiguous_auth()); // would be frozen. | |
9911 | if (!reply->trace.length()) | |
9912 | reply->starts_with = MDiscoverReply::DIR; | |
9913 | replicate_dir(curdir, from, reply->trace); | |
9914 | dout(7) << "handle_discover added dir " << *curdir << dendl; | |
9915 | } | |
9916 | ||
9917 | // lookup | |
9918 | CDentry *dn = 0; | |
9919 | if (curdir->get_version() == 0) { | |
9920 | // fetch newly opened dir | |
31f18b77 | 9921 | assert(!curdir->has_bloom()); |
7c673cae FG |
9922 | } else if (dis->get_want().depth() > 0) { |
9923 | // lookup dentry | |
9924 | dn = curdir->lookup(dis->get_dentry(i), snapid); | |
9925 | } else | |
9926 | break; // done! | |
9927 | ||
9928 | // incomplete dir? | |
9929 | if (!dn) { | |
31f18b77 FG |
9930 | if (!curdir->is_complete() && |
9931 | (!curdir->has_bloom() || curdir->is_in_bloom(dis->get_dentry(i)))) { | |
7c673cae FG |
9932 | // readdir |
9933 | dout(7) << "incomplete dir contents for " << *curdir << ", fetching" << dendl; | |
9934 | if (reply->is_empty()) { | |
9935 | // fetch and wait | |
9936 | curdir->fetch(new C_MDS_RetryMessage(mds, dis), | |
9937 | dis->wants_base_dir() && curdir->get_version() == 0); | |
9938 | reply->put(); | |
9939 | return; | |
9940 | } else { | |
9941 | // initiate fetch, but send what we have so far | |
9942 | curdir->fetch(0); | |
9943 | break; | |
9944 | } | |
9945 | } | |
9946 | ||
9947 | // send null dentry | |
9948 | dout(7) << "dentry " << dis->get_dentry(i) << " dne, returning null in " | |
9949 | << *curdir << dendl; | |
9950 | dn = curdir->add_null_dentry(dis->get_dentry(i)); | |
9951 | } | |
9952 | assert(dn); | |
9953 | ||
31f18b77 FG |
9954 | // don't add replica to purging dentry/inode |
9955 | if (dn->state_test(CDentry::STATE_PURGING)) { | |
9956 | if (reply->is_empty()) | |
9957 | reply->set_flag_error_dn(dis->get_dentry(i)); | |
9958 | break; | |
9959 | } | |
9960 | ||
7c673cae FG |
9961 | CDentry::linkage_t *dnl = dn->get_linkage(); |
9962 | ||
9963 | // xlocked dentry? | |
9964 | // ...always block on non-tail items (they are unrelated) | |
9965 | // ...allow xlocked tail disocvery _only_ if explicitly requested | |
9966 | bool tailitem = (dis->get_want().depth() == 0) || (i == dis->get_want().depth() - 1); | |
9967 | if (dn->lock.is_xlocked()) { | |
9968 | // is this the last (tail) item in the discover traversal? | |
9969 | if (tailitem && dis->wants_xlocked()) { | |
9970 | dout(7) << "handle_discover allowing discovery of xlocked tail " << *dn << dendl; | |
9971 | } else if (reply->is_empty()) { | |
9972 | dout(7) << "handle_discover blocking on xlocked " << *dn << dendl; | |
9973 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDS_RetryMessage(mds, dis)); | |
9974 | reply->put(); | |
9975 | return; | |
9976 | } else { | |
9977 | dout(7) << "handle_discover non-empty reply, xlocked tail " << *dn << dendl; | |
9978 | break; | |
9979 | } | |
9980 | } | |
9981 | ||
9982 | // frozen inode? | |
9983 | if (dnl->is_primary() && dnl->get_inode()->is_frozen_inode()) { | |
9984 | if (tailitem && dis->wants_xlocked()) { | |
9985 | dout(7) << "handle_discover allowing discovery of frozen tail " << *dnl->get_inode() << dendl; | |
9986 | } else if (reply->is_empty()) { | |
9987 | dout(7) << *dnl->get_inode() << " is frozen, empty reply, waiting" << dendl; | |
9988 | dnl->get_inode()->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryMessage(mds, dis)); | |
9989 | reply->put(); | |
9990 | return; | |
9991 | } else { | |
9992 | dout(7) << *dnl->get_inode() << " is frozen, non-empty reply, stopping" << dendl; | |
9993 | break; | |
9994 | } | |
9995 | } | |
9996 | ||
9997 | // add dentry | |
9998 | if (!reply->trace.length()) | |
9999 | reply->starts_with = MDiscoverReply::DENTRY; | |
10000 | replicate_dentry(dn, from, reply->trace); | |
10001 | dout(7) << "handle_discover added dentry " << *dn << dendl; | |
10002 | ||
10003 | if (!dnl->is_primary()) break; // stop on null or remote link. | |
10004 | ||
10005 | // add inode | |
10006 | CInode *next = dnl->get_inode(); | |
10007 | assert(next->is_auth()); | |
10008 | ||
10009 | replicate_inode(next, from, reply->trace, mds->mdsmap->get_up_features()); | |
10010 | dout(7) << "handle_discover added inode " << *next << dendl; | |
10011 | ||
10012 | // descend, keep going. | |
10013 | cur = next; | |
10014 | continue; | |
10015 | } | |
10016 | ||
10017 | // how did we do? | |
10018 | assert(!reply->is_empty()); | |
10019 | dout(7) << "handle_discover sending result back to asker mds." << from << dendl; | |
10020 | mds->send_message(reply, dis->get_connection()); | |
10021 | ||
10022 | dis->put(); | |
10023 | } | |
10024 | ||
10025 | /* This function DOES put the passed message before returning */ | |
10026 | void MDCache::handle_discover_reply(MDiscoverReply *m) | |
10027 | { | |
10028 | /* | |
10029 | if (mds->get_state() < MDSMap::STATE_ACTIVE) { | |
10030 | dout(0) << "discover_reply NOT ACTIVE YET" << dendl; | |
10031 | m->put(); | |
10032 | return; | |
10033 | } | |
10034 | */ | |
10035 | dout(7) << "discover_reply " << *m << dendl; | |
10036 | if (m->is_flag_error_dir()) | |
10037 | dout(7) << " flag error, dir" << dendl; | |
10038 | if (m->is_flag_error_dn()) | |
10039 | dout(7) << " flag error, dentry = " << m->get_error_dentry() << dendl; | |
10040 | ||
10041 | list<MDSInternalContextBase*> finished, error; | |
10042 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
10043 | ||
10044 | // starting point | |
10045 | CInode *cur = get_inode(m->get_base_ino()); | |
10046 | bufferlist::iterator p = m->trace.begin(); | |
10047 | ||
10048 | int next = m->starts_with; | |
10049 | ||
10050 | // decrement discover counters | |
10051 | if (m->get_tid()) { | |
10052 | map<ceph_tid_t,discover_info_t>::iterator p = discovers.find(m->get_tid()); | |
10053 | if (p != discovers.end()) { | |
10054 | dout(10) << " found tid " << m->get_tid() << dendl; | |
10055 | discovers.erase(p); | |
10056 | } else { | |
10057 | dout(10) << " tid " << m->get_tid() << " not found, must be dup reply" << dendl; | |
10058 | } | |
10059 | } | |
10060 | ||
10061 | // discover may start with an inode | |
10062 | if (!p.end() && next == MDiscoverReply::INODE) { | |
10063 | cur = add_replica_inode(p, NULL, finished); | |
10064 | dout(7) << "discover_reply got base inode " << *cur << dendl; | |
10065 | assert(cur->is_base()); | |
10066 | ||
10067 | next = MDiscoverReply::DIR; | |
10068 | ||
10069 | // take waiters? | |
10070 | if (cur->is_base() && | |
10071 | waiting_for_base_ino[from].count(cur->ino())) { | |
10072 | finished.swap(waiting_for_base_ino[from][cur->ino()]); | |
10073 | waiting_for_base_ino[from].erase(cur->ino()); | |
10074 | } | |
10075 | } | |
10076 | assert(cur); | |
10077 | ||
10078 | // loop over discover results. | |
10079 | // indexes follow each ([[dir] dentry] inode) | |
10080 | // can start, end with any type. | |
10081 | while (!p.end()) { | |
10082 | // dir | |
10083 | frag_t fg; | |
10084 | CDir *curdir = 0; | |
10085 | if (next == MDiscoverReply::DIR) { | |
10086 | curdir = add_replica_dir(p, cur, mds_rank_t(m->get_source().num()), finished); | |
10087 | if (cur->ino() == m->get_base_ino() && curdir->get_frag() != m->get_base_dir_frag()) { | |
10088 | assert(m->get_wanted_base_dir()); | |
10089 | cur->take_dir_waiting(m->get_base_dir_frag(), finished); | |
10090 | } | |
10091 | } else { | |
10092 | // note: this can only happen our first way around this loop. | |
10093 | if (p.end() && m->is_flag_error_dn()) { | |
10094 | fg = cur->pick_dirfrag(m->get_error_dentry()); | |
10095 | curdir = cur->get_dirfrag(fg); | |
10096 | } else | |
10097 | curdir = cur->get_dirfrag(m->get_base_dir_frag()); | |
10098 | } | |
10099 | ||
10100 | if (p.end()) | |
10101 | break; | |
10102 | ||
10103 | // dentry | |
10104 | CDentry *dn = add_replica_dentry(p, curdir, finished); | |
10105 | ||
10106 | if (p.end()) | |
10107 | break; | |
10108 | ||
10109 | // inode | |
10110 | cur = add_replica_inode(p, dn, finished); | |
10111 | ||
10112 | next = MDiscoverReply::DIR; | |
10113 | } | |
10114 | ||
10115 | // dir error? | |
10116 | // or dir_auth hint? | |
10117 | if (m->is_flag_error_dir() && !cur->is_dir()) { | |
10118 | // not a dir. | |
10119 | cur->take_waiting(CInode::WAIT_DIR, error); | |
10120 | } else if (m->is_flag_error_dir() || m->get_dir_auth_hint() != CDIR_AUTH_UNKNOWN) { | |
10121 | mds_rank_t who = m->get_dir_auth_hint(); | |
10122 | if (who == mds->get_nodeid()) who = -1; | |
10123 | if (who >= 0) | |
10124 | dout(7) << " dir_auth_hint is " << m->get_dir_auth_hint() << dendl; | |
10125 | ||
7c673cae FG |
10126 | |
10127 | if (m->get_wanted_base_dir()) { | |
31f18b77 FG |
10128 | frag_t fg = m->get_base_dir_frag(); |
10129 | CDir *dir = cur->get_dirfrag(fg); | |
10130 | ||
7c673cae FG |
10131 | if (cur->is_waiting_for_dir(fg)) { |
10132 | if (cur->is_auth()) | |
10133 | cur->take_waiting(CInode::WAIT_DIR, finished); | |
10134 | else if (dir || !cur->dirfragtree.is_leaf(fg)) | |
10135 | cur->take_dir_waiting(fg, finished); | |
10136 | else | |
10137 | discover_dir_frag(cur, fg, 0, who); | |
10138 | } else | |
10139 | dout(7) << " doing nothing, nobody is waiting for dir" << dendl; | |
10140 | } | |
10141 | ||
10142 | // try again? | |
10143 | if (m->get_error_dentry().length()) { | |
31f18b77 FG |
10144 | frag_t fg = cur->pick_dirfrag(m->get_error_dentry()); |
10145 | CDir *dir = cur->get_dirfrag(fg); | |
7c673cae FG |
10146 | // wanted a dentry |
10147 | if (dir && dir->is_waiting_for_dentry(m->get_error_dentry(), m->get_wanted_snapid())) { | |
10148 | if (dir->is_auth() || dir->lookup(m->get_error_dentry())) { | |
10149 | dir->take_dentry_waiting(m->get_error_dentry(), m->get_wanted_snapid(), | |
10150 | m->get_wanted_snapid(), finished); | |
10151 | } else { | |
10152 | filepath relpath(m->get_error_dentry(), 0); | |
10153 | discover_path(dir, m->get_wanted_snapid(), relpath, 0, m->get_wanted_xlocked()); | |
10154 | } | |
10155 | } else | |
10156 | dout(7) << " doing nothing, have dir but nobody is waiting on dentry " | |
10157 | << m->get_error_dentry() << dendl; | |
10158 | } | |
31f18b77 FG |
10159 | } else if (m->is_flag_error_dn()) { |
10160 | frag_t fg = cur->pick_dirfrag(m->get_error_dentry()); | |
10161 | CDir *dir = cur->get_dirfrag(fg); | |
10162 | if (dir) { | |
10163 | if (dir->is_auth()) { | |
10164 | dir->take_sub_waiting(finished); | |
10165 | } else { | |
10166 | dir->take_dentry_waiting(m->get_error_dentry(), m->get_wanted_snapid(), | |
10167 | m->get_wanted_snapid(), error); | |
10168 | } | |
10169 | } | |
7c673cae FG |
10170 | } |
10171 | ||
10172 | // waiters | |
10173 | finish_contexts(g_ceph_context, error, -ENOENT); // finish errors directly | |
10174 | mds->queue_waiters(finished); | |
10175 | ||
10176 | // done | |
10177 | m->put(); | |
10178 | } | |
10179 | ||
10180 | ||
10181 | ||
10182 | // ---------------------------- | |
10183 | // REPLICAS | |
10184 | ||
10185 | CDir *MDCache::add_replica_dir(bufferlist::iterator& p, CInode *diri, mds_rank_t from, | |
10186 | list<MDSInternalContextBase*>& finished) | |
10187 | { | |
10188 | dirfrag_t df; | |
10189 | ::decode(df, p); | |
10190 | ||
10191 | assert(diri->ino() == df.ino); | |
10192 | ||
10193 | // add it (_replica_) | |
10194 | CDir *dir = diri->get_dirfrag(df.frag); | |
10195 | ||
10196 | if (dir) { | |
10197 | // had replica. update w/ new nonce. | |
10198 | dir->decode_replica(p); | |
10199 | dout(7) << "add_replica_dir had " << *dir << " nonce " << dir->replica_nonce << dendl; | |
10200 | } else { | |
10201 | // force frag to leaf in the diri tree | |
10202 | if (!diri->dirfragtree.is_leaf(df.frag)) { | |
10203 | dout(7) << "add_replica_dir forcing frag " << df.frag << " to leaf in the fragtree " | |
10204 | << diri->dirfragtree << dendl; | |
10205 | diri->dirfragtree.force_to_leaf(g_ceph_context, df.frag); | |
10206 | } | |
10207 | ||
10208 | // add replica. | |
10209 | dir = diri->add_dirfrag( new CDir(diri, df.frag, this, false) ); | |
10210 | dir->decode_replica(p); | |
10211 | ||
10212 | // is this a dir_auth delegation boundary? | |
10213 | if (from != diri->authority().first || | |
10214 | diri->is_ambiguous_auth() || | |
10215 | diri->is_base()) | |
10216 | adjust_subtree_auth(dir, from); | |
10217 | ||
10218 | dout(7) << "add_replica_dir added " << *dir << " nonce " << dir->replica_nonce << dendl; | |
10219 | ||
10220 | // get waiters | |
10221 | diri->take_dir_waiting(df.frag, finished); | |
10222 | } | |
10223 | ||
10224 | return dir; | |
10225 | } | |
10226 | ||
7c673cae FG |
10227 | CDentry *MDCache::add_replica_dentry(bufferlist::iterator& p, CDir *dir, list<MDSInternalContextBase*>& finished) |
10228 | { | |
10229 | string name; | |
10230 | snapid_t last; | |
10231 | ::decode(name, p); | |
10232 | ::decode(last, p); | |
10233 | ||
10234 | CDentry *dn = dir->lookup(name, last); | |
10235 | ||
10236 | // have it? | |
10237 | if (dn) { | |
10238 | dn->decode_replica(p, false); | |
10239 | dout(7) << "add_replica_dentry had " << *dn << dendl; | |
10240 | } else { | |
10241 | dn = dir->add_null_dentry(name, 1 /* this will get updated below */, last); | |
10242 | dn->decode_replica(p, true); | |
10243 | dout(7) << "add_replica_dentry added " << *dn << dendl; | |
10244 | } | |
10245 | ||
10246 | dir->take_dentry_waiting(name, dn->first, dn->last, finished); | |
10247 | ||
10248 | return dn; | |
10249 | } | |
10250 | ||
10251 | CInode *MDCache::add_replica_inode(bufferlist::iterator& p, CDentry *dn, list<MDSInternalContextBase*>& finished) | |
10252 | { | |
10253 | inodeno_t ino; | |
10254 | snapid_t last; | |
10255 | ::decode(ino, p); | |
10256 | ::decode(last, p); | |
10257 | CInode *in = get_inode(ino, last); | |
10258 | if (!in) { | |
10259 | in = new CInode(this, false, 1, last); | |
10260 | in->decode_replica(p, true); | |
10261 | add_inode(in); | |
10262 | if (in->ino() == MDS_INO_ROOT) | |
10263 | in->inode_auth.first = 0; | |
10264 | else if (in->is_mdsdir()) | |
10265 | in->inode_auth.first = in->ino() - MDS_INO_MDSDIR_OFFSET; | |
10266 | dout(10) << "add_replica_inode added " << *in << dendl; | |
10267 | if (dn) { | |
10268 | assert(dn->get_linkage()->is_null()); | |
10269 | dn->dir->link_primary_inode(dn, in); | |
10270 | } | |
10271 | } else { | |
10272 | in->decode_replica(p, false); | |
10273 | dout(10) << "add_replica_inode had " << *in << dendl; | |
10274 | } | |
10275 | ||
10276 | if (dn) { | |
10277 | if (!dn->get_linkage()->is_primary() || dn->get_linkage()->get_inode() != in) | |
10278 | dout(10) << "add_replica_inode different linkage in dentry " << *dn << dendl; | |
10279 | } | |
10280 | ||
10281 | return in; | |
10282 | } | |
10283 | ||
10284 | ||
10285 | void MDCache::replicate_stray(CDentry *straydn, mds_rank_t who, bufferlist& bl) | |
10286 | { | |
10287 | uint64_t features = mds->mdsmap->get_up_features(); | |
10288 | replicate_inode(get_myin(), who, bl, features); | |
10289 | replicate_dir(straydn->get_dir()->inode->get_parent_dn()->get_dir(), who, bl); | |
10290 | replicate_dentry(straydn->get_dir()->inode->get_parent_dn(), who, bl); | |
10291 | replicate_inode(straydn->get_dir()->inode, who, bl, features); | |
10292 | replicate_dir(straydn->get_dir(), who, bl); | |
10293 | replicate_dentry(straydn, who, bl); | |
10294 | } | |
10295 | ||
10296 | CDentry *MDCache::add_replica_stray(bufferlist &bl, mds_rank_t from) | |
10297 | { | |
10298 | list<MDSInternalContextBase*> finished; | |
10299 | bufferlist::iterator p = bl.begin(); | |
10300 | ||
10301 | CInode *mdsin = add_replica_inode(p, NULL, finished); | |
10302 | CDir *mdsdir = add_replica_dir(p, mdsin, from, finished); | |
10303 | CDentry *straydirdn = add_replica_dentry(p, mdsdir, finished); | |
10304 | CInode *strayin = add_replica_inode(p, straydirdn, finished); | |
10305 | CDir *straydir = add_replica_dir(p, strayin, from, finished); | |
10306 | CDentry *straydn = add_replica_dentry(p, straydir, finished); | |
10307 | if (!finished.empty()) | |
10308 | mds->queue_waiters(finished); | |
10309 | ||
10310 | return straydn; | |
10311 | } | |
10312 | ||
10313 | ||
10314 | int MDCache::send_dir_updates(CDir *dir, bool bcast) | |
10315 | { | |
10316 | // this is an FYI, re: replication | |
10317 | ||
10318 | set<mds_rank_t> who; | |
10319 | if (bcast) { | |
10320 | mds->get_mds_map()->get_active_mds_set(who); | |
10321 | } else { | |
181888fb FG |
10322 | for (const auto &p : dir->get_replicas()) { |
10323 | who.insert(p.first); | |
10324 | } | |
7c673cae FG |
10325 | } |
10326 | ||
10327 | dout(7) << "sending dir_update on " << *dir << " bcast " << bcast << " to " << who << dendl; | |
10328 | ||
10329 | filepath path; | |
10330 | dir->inode->make_path(path); | |
10331 | ||
10332 | mds_rank_t whoami = mds->get_nodeid(); | |
10333 | for (set<mds_rank_t>::iterator it = who.begin(); | |
10334 | it != who.end(); | |
10335 | ++it) { | |
10336 | if (*it == whoami) continue; | |
10337 | //if (*it == except) continue; | |
10338 | dout(7) << "sending dir_update on " << *dir << " to " << *it << dendl; | |
10339 | ||
10340 | mds->send_message_mds(new MDirUpdate(mds->get_nodeid(), | |
10341 | dir->dirfrag(), | |
10342 | dir->dir_rep, | |
10343 | dir->dir_rep_by, | |
10344 | path, | |
10345 | bcast), | |
10346 | *it); | |
10347 | } | |
10348 | ||
10349 | return 0; | |
10350 | } | |
10351 | ||
10352 | /* This function DOES put the passed message before returning */ | |
10353 | void MDCache::handle_dir_update(MDirUpdate *m) | |
10354 | { | |
224ce89b WB |
10355 | dirfrag_t df = m->get_dirfrag(); |
10356 | CDir *dir = get_dirfrag(df); | |
7c673cae | 10357 | if (!dir) { |
224ce89b | 10358 | dout(5) << "dir_update on " << df << ", don't have it" << dendl; |
7c673cae FG |
10359 | |
10360 | // discover it? | |
10361 | if (m->should_discover()) { | |
10362 | // only try once! | |
10363 | // this is key to avoid a fragtree update race, among other things. | |
224ce89b | 10364 | m->inc_tried_discover(); |
7c673cae FG |
10365 | vector<CDentry*> trace; |
10366 | CInode *in; | |
10367 | filepath path = m->get_path(); | |
10368 | dout(5) << "trying discover on dir_update for " << path << dendl; | |
10369 | MDRequestRef null_ref; | |
10370 | int r = path_traverse(null_ref, m, NULL, path, &trace, &in, MDS_TRAVERSE_DISCOVER); | |
10371 | if (r > 0) | |
10372 | return; | |
224ce89b WB |
10373 | if (r == 0 && |
10374 | in->ino() == df.ino && | |
10375 | in->get_approx_dirfrag(df.frag) == NULL) { | |
10376 | open_remote_dirfrag(in, df.frag, new C_MDS_RetryMessage(mds, m)); | |
10377 | return; | |
10378 | } | |
7c673cae FG |
10379 | } |
10380 | ||
10381 | m->put(); | |
10382 | return; | |
10383 | } | |
10384 | ||
224ce89b WB |
10385 | if (!m->has_tried_discover()) { |
10386 | // Update if it already exists. Othwerwise it got updated by discover reply. | |
10387 | dout(5) << "dir_update on " << *dir << dendl; | |
10388 | dir->dir_rep = m->get_dir_rep(); | |
10389 | dir->dir_rep_by = m->get_dir_rep_by(); | |
10390 | } | |
10391 | ||
7c673cae FG |
10392 | // done |
10393 | m->put(); | |
10394 | } | |
10395 | ||
10396 | ||
10397 | ||
10398 | ||
10399 | ||
10400 | // LINK | |
10401 | ||
10402 | void MDCache::send_dentry_link(CDentry *dn, MDRequestRef& mdr) | |
10403 | { | |
10404 | dout(7) << "send_dentry_link " << *dn << dendl; | |
10405 | ||
10406 | CDir *subtree = get_subtree_root(dn->get_dir()); | |
181888fb | 10407 | for (const auto &p : dn->get_replicas()) { |
7c673cae | 10408 | // don't tell (rename) witnesses; they already know |
181888fb | 10409 | if (mdr.get() && mdr->more()->witnessed.count(p.first)) |
7c673cae | 10410 | continue; |
181888fb FG |
10411 | if (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || |
10412 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
10413 | rejoin_gather.count(p.first))) | |
7c673cae FG |
10414 | continue; |
10415 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
10416 | MDentryLink *m = new MDentryLink(subtree->dirfrag(), dn->get_dir()->dirfrag(), | |
10417 | dn->name, dnl->is_primary()); | |
10418 | if (dnl->is_primary()) { | |
10419 | dout(10) << " primary " << *dnl->get_inode() << dendl; | |
181888fb | 10420 | replicate_inode(dnl->get_inode(), p.first, m->bl, |
7c673cae FG |
10421 | mds->mdsmap->get_up_features()); |
10422 | } else if (dnl->is_remote()) { | |
10423 | inodeno_t ino = dnl->get_remote_ino(); | |
10424 | __u8 d_type = dnl->get_remote_d_type(); | |
10425 | dout(10) << " remote " << ino << " " << d_type << dendl; | |
10426 | ::encode(ino, m->bl); | |
10427 | ::encode(d_type, m->bl); | |
10428 | } else | |
10429 | ceph_abort(); // aie, bad caller! | |
181888fb | 10430 | mds->send_message_mds(m, p.first); |
7c673cae FG |
10431 | } |
10432 | } | |
10433 | ||
10434 | /* This function DOES put the passed message before returning */ | |
10435 | void MDCache::handle_dentry_link(MDentryLink *m) | |
10436 | { | |
10437 | ||
10438 | CDentry *dn = NULL; | |
10439 | CDir *dir = get_dirfrag(m->get_dirfrag()); | |
10440 | if (!dir) { | |
10441 | dout(7) << "handle_dentry_link don't have dirfrag " << m->get_dirfrag() << dendl; | |
10442 | } else { | |
10443 | dn = dir->lookup(m->get_dn()); | |
10444 | if (!dn) { | |
10445 | dout(7) << "handle_dentry_link don't have dentry " << *dir << " dn " << m->get_dn() << dendl; | |
10446 | } else { | |
10447 | dout(7) << "handle_dentry_link on " << *dn << dendl; | |
10448 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
10449 | ||
10450 | assert(!dn->is_auth()); | |
10451 | assert(dnl->is_null()); | |
10452 | } | |
10453 | } | |
10454 | ||
10455 | bufferlist::iterator p = m->bl.begin(); | |
10456 | list<MDSInternalContextBase*> finished; | |
10457 | if (dn) { | |
10458 | if (m->get_is_primary()) { | |
10459 | // primary link. | |
10460 | add_replica_inode(p, dn, finished); | |
10461 | } else { | |
10462 | // remote link, easy enough. | |
10463 | inodeno_t ino; | |
10464 | __u8 d_type; | |
10465 | ::decode(ino, p); | |
10466 | ::decode(d_type, p); | |
10467 | dir->link_remote_inode(dn, ino, d_type); | |
10468 | } | |
10469 | } else { | |
10470 | ceph_abort(); | |
10471 | } | |
10472 | ||
10473 | if (!finished.empty()) | |
10474 | mds->queue_waiters(finished); | |
10475 | ||
10476 | m->put(); | |
10477 | return; | |
10478 | } | |
10479 | ||
10480 | ||
10481 | // UNLINK | |
10482 | ||
10483 | void MDCache::send_dentry_unlink(CDentry *dn, CDentry *straydn, MDRequestRef& mdr) | |
10484 | { | |
10485 | dout(10) << "send_dentry_unlink " << *dn << dendl; | |
10486 | // share unlink news with replicas | |
10487 | set<mds_rank_t> replicas; | |
10488 | dn->list_replicas(replicas); | |
10489 | if (straydn) | |
10490 | straydn->list_replicas(replicas); | |
10491 | for (set<mds_rank_t>::iterator it = replicas.begin(); | |
10492 | it != replicas.end(); | |
10493 | ++it) { | |
10494 | // don't tell (rmdir) witnesses; they already know | |
10495 | if (mdr.get() && mdr->more()->witnessed.count(*it)) | |
10496 | continue; | |
10497 | ||
10498 | if (mds->mdsmap->get_state(*it) < MDSMap::STATE_REJOIN || | |
10499 | (mds->mdsmap->get_state(*it) == MDSMap::STATE_REJOIN && | |
10500 | rejoin_gather.count(*it))) | |
10501 | continue; | |
10502 | ||
10503 | MDentryUnlink *unlink = new MDentryUnlink(dn->get_dir()->dirfrag(), dn->name); | |
10504 | if (straydn) | |
10505 | replicate_stray(straydn, *it, unlink->straybl); | |
10506 | mds->send_message_mds(unlink, *it); | |
10507 | } | |
10508 | } | |
10509 | ||
10510 | /* This function DOES put the passed message before returning */ | |
10511 | void MDCache::handle_dentry_unlink(MDentryUnlink *m) | |
10512 | { | |
10513 | // straydn | |
10514 | CDentry *straydn = NULL; | |
10515 | if (m->straybl.length()) | |
10516 | straydn = add_replica_stray(m->straybl, mds_rank_t(m->get_source().num())); | |
10517 | ||
10518 | CDir *dir = get_dirfrag(m->get_dirfrag()); | |
10519 | if (!dir) { | |
10520 | dout(7) << "handle_dentry_unlink don't have dirfrag " << m->get_dirfrag() << dendl; | |
10521 | } else { | |
10522 | CDentry *dn = dir->lookup(m->get_dn()); | |
10523 | if (!dn) { | |
10524 | dout(7) << "handle_dentry_unlink don't have dentry " << *dir << " dn " << m->get_dn() << dendl; | |
10525 | } else { | |
10526 | dout(7) << "handle_dentry_unlink on " << *dn << dendl; | |
10527 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
10528 | ||
10529 | // open inode? | |
10530 | if (dnl->is_primary()) { | |
10531 | CInode *in = dnl->get_inode(); | |
10532 | dn->dir->unlink_inode(dn); | |
10533 | assert(straydn); | |
10534 | straydn->dir->link_primary_inode(straydn, in); | |
10535 | ||
10536 | // in->first is lazily updated on replica; drag it forward so | |
10537 | // that we always keep it in sync with the dnq | |
10538 | assert(straydn->first >= in->first); | |
10539 | in->first = straydn->first; | |
10540 | ||
10541 | // update subtree map? | |
10542 | if (in->is_dir()) | |
10543 | adjust_subtree_after_rename(in, dir, false); | |
10544 | ||
10545 | // send caps to auth (if we're not already) | |
10546 | if (in->is_any_caps() && | |
10547 | !in->state_test(CInode::STATE_EXPORTINGCAPS)) | |
10548 | migrator->export_caps(in); | |
10549 | ||
7c673cae FG |
10550 | straydn = NULL; |
10551 | } else { | |
10552 | assert(!straydn); | |
10553 | assert(dnl->is_remote()); | |
10554 | dn->dir->unlink_inode(dn); | |
10555 | } | |
10556 | assert(dnl->is_null()); | |
7c673cae FG |
10557 | } |
10558 | } | |
10559 | ||
10560 | // race with trim_dentry() | |
10561 | if (straydn) { | |
10562 | assert(straydn->get_num_ref() == 0); | |
10563 | assert(straydn->get_linkage()->is_null()); | |
10564 | map<mds_rank_t, MCacheExpire*> expiremap; | |
10565 | trim_dentry(straydn, expiremap); | |
10566 | send_expire_messages(expiremap); | |
10567 | } | |
10568 | ||
10569 | m->put(); | |
10570 | return; | |
10571 | } | |
10572 | ||
10573 | ||
10574 | ||
10575 | ||
10576 | ||
10577 | ||
10578 | // =================================================================== | |
10579 | ||
10580 | ||
10581 | ||
10582 | // =================================================================== | |
10583 | // FRAGMENT | |
10584 | ||
10585 | ||
10586 | /** | |
10587 | * adjust_dir_fragments -- adjust fragmentation for a directory | |
10588 | * | |
10589 | * @param diri directory inode | |
10590 | * @param basefrag base fragment | |
10591 | * @param bits bit adjustment. positive for split, negative for merge. | |
10592 | */ | |
10593 | void MDCache::adjust_dir_fragments(CInode *diri, frag_t basefrag, int bits, | |
10594 | list<CDir*>& resultfrags, | |
10595 | list<MDSInternalContextBase*>& waiters, | |
10596 | bool replay) | |
10597 | { | |
10598 | dout(10) << "adjust_dir_fragments " << basefrag << " " << bits | |
10599 | << " on " << *diri << dendl; | |
10600 | ||
10601 | list<CDir*> srcfrags; | |
10602 | diri->get_dirfrags_under(basefrag, srcfrags); | |
10603 | ||
10604 | adjust_dir_fragments(diri, srcfrags, basefrag, bits, resultfrags, waiters, replay); | |
10605 | } | |
10606 | ||
10607 | CDir *MDCache::force_dir_fragment(CInode *diri, frag_t fg, bool replay) | |
10608 | { | |
10609 | CDir *dir = diri->get_dirfrag(fg); | |
10610 | if (dir) | |
10611 | return dir; | |
10612 | ||
10613 | dout(10) << "force_dir_fragment " << fg << " on " << *diri << dendl; | |
10614 | ||
10615 | list<CDir*> src, result; | |
10616 | list<MDSInternalContextBase*> waiters; | |
10617 | ||
10618 | // split a parent? | |
10619 | frag_t parent = diri->dirfragtree.get_branch_or_leaf(fg); | |
10620 | while (1) { | |
10621 | CDir *pdir = diri->get_dirfrag(parent); | |
10622 | if (pdir) { | |
10623 | int split = fg.bits() - parent.bits(); | |
10624 | dout(10) << " splitting parent by " << split << " " << *pdir << dendl; | |
10625 | src.push_back(pdir); | |
10626 | adjust_dir_fragments(diri, src, parent, split, result, waiters, replay); | |
10627 | dir = diri->get_dirfrag(fg); | |
10628 | if (dir) { | |
10629 | dout(10) << "force_dir_fragment result " << *dir << dendl; | |
10630 | break; | |
10631 | } | |
10632 | } | |
10633 | if (parent == frag_t()) | |
10634 | break; | |
10635 | frag_t last = parent; | |
10636 | parent = parent.parent(); | |
10637 | dout(10) << " " << last << " parent is " << parent << dendl; | |
10638 | } | |
10639 | ||
10640 | if (!dir) { | |
10641 | // hoover up things under fg? | |
10642 | diri->get_dirfrags_under(fg, src); | |
10643 | if (src.empty()) { | |
10644 | dout(10) << "force_dir_fragment no frags under " << fg << dendl; | |
10645 | } else { | |
10646 | dout(10) << " will combine frags under " << fg << ": " << src << dendl; | |
10647 | adjust_dir_fragments(diri, src, fg, 0, result, waiters, replay); | |
10648 | dir = result.front(); | |
10649 | dout(10) << "force_dir_fragment result " << *dir << dendl; | |
10650 | } | |
10651 | } | |
10652 | if (!replay) | |
10653 | mds->queue_waiters(waiters); | |
10654 | return dir; | |
10655 | } | |
10656 | ||
10657 | void MDCache::adjust_dir_fragments(CInode *diri, | |
10658 | list<CDir*>& srcfrags, | |
10659 | frag_t basefrag, int bits, | |
10660 | list<CDir*>& resultfrags, | |
10661 | list<MDSInternalContextBase*>& waiters, | |
10662 | bool replay) | |
10663 | { | |
10664 | dout(10) << "adjust_dir_fragments " << basefrag << " bits " << bits | |
10665 | << " srcfrags " << srcfrags | |
10666 | << " on " << *diri << dendl; | |
10667 | ||
10668 | // adjust fragtree | |
10669 | // yuck. we may have discovered the inode while it was being fragmented. | |
10670 | if (!diri->dirfragtree.is_leaf(basefrag)) | |
10671 | diri->dirfragtree.force_to_leaf(g_ceph_context, basefrag); | |
10672 | ||
10673 | if (bits > 0) | |
10674 | diri->dirfragtree.split(basefrag, bits); | |
10675 | dout(10) << " new fragtree is " << diri->dirfragtree << dendl; | |
10676 | ||
10677 | if (srcfrags.empty()) | |
10678 | return; | |
10679 | ||
10680 | // split | |
10681 | CDir *parent_dir = diri->get_parent_dir(); | |
10682 | CDir *parent_subtree = 0; | |
10683 | if (parent_dir) | |
10684 | parent_subtree = get_subtree_root(parent_dir); | |
10685 | ||
10686 | if (bits > 0) { | |
10687 | // SPLIT | |
10688 | assert(srcfrags.size() == 1); | |
10689 | CDir *dir = srcfrags.front(); | |
10690 | ||
10691 | dir->split(bits, resultfrags, waiters, replay); | |
10692 | ||
10693 | // did i change the subtree map? | |
10694 | if (dir->is_subtree_root()) { | |
10695 | // new frags are now separate subtrees | |
10696 | for (list<CDir*>::iterator p = resultfrags.begin(); | |
10697 | p != resultfrags.end(); | |
10698 | ++p) | |
10699 | subtrees[*p].clear(); // new frag is now its own subtree | |
10700 | ||
10701 | // was i a bound? | |
10702 | if (parent_subtree) { | |
10703 | assert(subtrees[parent_subtree].count(dir)); | |
10704 | subtrees[parent_subtree].erase(dir); | |
10705 | for (list<CDir*>::iterator p = resultfrags.begin(); | |
10706 | p != resultfrags.end(); | |
10707 | ++p) { | |
10708 | assert((*p)->is_subtree_root()); | |
10709 | subtrees[parent_subtree].insert(*p); | |
10710 | } | |
10711 | } | |
10712 | ||
10713 | // adjust my bounds. | |
10714 | set<CDir*> bounds; | |
10715 | bounds.swap(subtrees[dir]); | |
10716 | subtrees.erase(dir); | |
10717 | for (set<CDir*>::iterator p = bounds.begin(); | |
10718 | p != bounds.end(); | |
10719 | ++p) { | |
10720 | CDir *frag = get_subtree_root((*p)->get_parent_dir()); | |
10721 | subtrees[frag].insert(*p); | |
10722 | } | |
10723 | ||
10724 | show_subtrees(10); | |
10725 | ||
10726 | // dir has no PIN_SUBTREE; CDir::purge_stolen() drops it. | |
10727 | dir->dir_auth = CDIR_AUTH_DEFAULT; | |
10728 | } | |
10729 | ||
10730 | diri->close_dirfrag(dir->get_frag()); | |
10731 | ||
10732 | } else { | |
10733 | // MERGE | |
10734 | ||
10735 | // are my constituent bits subtrees? if so, i will be too. | |
10736 | // (it's all or none, actually.) | |
31f18b77 FG |
10737 | bool any_subtree = false; |
10738 | for (CDir *dir : srcfrags) { | |
7c673cae | 10739 | if (dir->is_subtree_root()) { |
31f18b77 FG |
10740 | any_subtree = true; |
10741 | break; | |
10742 | } | |
10743 | } | |
10744 | set<CDir*> new_bounds; | |
10745 | if (any_subtree) { | |
10746 | for (CDir *dir : srcfrags) { | |
10747 | // this simplifies the code that find subtrees underneath the dirfrag | |
10748 | if (!dir->is_subtree_root()) { | |
10749 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
10750 | adjust_subtree_auth(dir, mds->get_nodeid()); | |
10751 | } | |
10752 | } | |
10753 | ||
10754 | for (CDir *dir : srcfrags) { | |
10755 | assert(dir->is_subtree_root()); | |
7c673cae | 10756 | dout(10) << " taking srcfrag subtree bounds from " << *dir << dendl; |
7c673cae FG |
10757 | map<CDir*, set<CDir*> >::iterator q = subtrees.find(dir); |
10758 | set<CDir*>::iterator r = q->second.begin(); | |
10759 | while (r != subtrees[dir].end()) { | |
10760 | new_bounds.insert(*r); | |
10761 | subtrees[dir].erase(r++); | |
10762 | } | |
10763 | subtrees.erase(q); | |
31f18b77 | 10764 | |
7c673cae FG |
10765 | // remove myself as my parent's bound |
10766 | if (parent_subtree) | |
10767 | subtrees[parent_subtree].erase(dir); | |
10768 | } | |
10769 | } | |
10770 | ||
10771 | // merge | |
10772 | CDir *f = new CDir(diri, basefrag, this, srcfrags.front()->is_auth()); | |
10773 | f->merge(srcfrags, waiters, replay); | |
7c673cae | 10774 | |
31f18b77 | 10775 | if (any_subtree) { |
7c673cae FG |
10776 | assert(f->is_subtree_root()); |
10777 | subtrees[f].swap(new_bounds); | |
10778 | if (parent_subtree) | |
10779 | subtrees[parent_subtree].insert(f); | |
10780 | ||
10781 | show_subtrees(10); | |
10782 | } | |
10783 | ||
10784 | resultfrags.push_back(f); | |
10785 | } | |
10786 | } | |
10787 | ||
10788 | ||
10789 | class C_MDC_FragmentFrozen : public MDSInternalContext { | |
10790 | MDCache *mdcache; | |
10791 | MDRequestRef mdr; | |
10792 | public: | |
10793 | C_MDC_FragmentFrozen(MDCache *m, MDRequestRef& r) : | |
10794 | MDSInternalContext(m->mds), mdcache(m), mdr(r) {} | |
10795 | void finish(int r) override { | |
10796 | mdcache->fragment_frozen(mdr, r); | |
10797 | } | |
10798 | }; | |
10799 | ||
10800 | bool MDCache::can_fragment(CInode *diri, list<CDir*>& dirs) | |
10801 | { | |
10802 | if (is_readonly()) { | |
10803 | dout(7) << "can_fragment: read-only FS, no fragmenting for now" << dendl; | |
10804 | return false; | |
10805 | } | |
10806 | if (mds->is_cluster_degraded()) { | |
10807 | dout(7) << "can_fragment: cluster degraded, no fragmenting for now" << dendl; | |
10808 | return false; | |
10809 | } | |
10810 | if (diri->get_parent_dir() && | |
10811 | diri->get_parent_dir()->get_inode()->is_stray()) { | |
10812 | dout(7) << "can_fragment: i won't merge|split anything in stray" << dendl; | |
10813 | return false; | |
10814 | } | |
10815 | if (diri->is_mdsdir() || diri->is_stray() || diri->ino() == MDS_INO_CEPH) { | |
10816 | dout(7) << "can_fragment: i won't fragment the mdsdir or straydir or .ceph" << dendl; | |
10817 | return false; | |
10818 | } | |
10819 | ||
10820 | if (diri->scrub_is_in_progress()) { | |
10821 | dout(7) << "can_fragment: scrub in progress" << dendl; | |
10822 | return false; | |
10823 | } | |
10824 | ||
10825 | for (list<CDir*>::iterator p = dirs.begin(); p != dirs.end(); ++p) { | |
10826 | CDir *dir = *p; | |
10827 | if (dir->state_test(CDir::STATE_FRAGMENTING)) { | |
10828 | dout(7) << "can_fragment: already fragmenting " << *dir << dendl; | |
10829 | return false; | |
10830 | } | |
10831 | if (!dir->is_auth()) { | |
10832 | dout(7) << "can_fragment: not auth on " << *dir << dendl; | |
10833 | return false; | |
10834 | } | |
10835 | if (dir->is_bad()) { | |
10836 | dout(7) << "can_fragment: bad dirfrag " << *dir << dendl; | |
10837 | return false; | |
10838 | } | |
10839 | if (dir->is_frozen() || | |
10840 | dir->is_freezing()) { | |
10841 | dout(7) << "can_fragment: can't merge, freezing|frozen. wait for other exports to finish first." << dendl; | |
10842 | return false; | |
10843 | } | |
10844 | } | |
10845 | ||
10846 | return true; | |
10847 | } | |
10848 | ||
10849 | void MDCache::split_dir(CDir *dir, int bits) | |
10850 | { | |
10851 | dout(7) << __func__ << " " << *dir << " bits " << bits << dendl; | |
10852 | assert(dir->is_auth()); | |
10853 | CInode *diri = dir->inode; | |
10854 | ||
10855 | list<CDir*> dirs; | |
10856 | dirs.push_back(dir); | |
10857 | ||
10858 | if (!can_fragment(diri, dirs)) { | |
10859 | dout(7) << __func__ << " cannot fragment right now, dropping" << dendl; | |
10860 | return; | |
10861 | } | |
10862 | ||
31f18b77 FG |
10863 | if (dir->frag.bits() + bits > 24) { |
10864 | dout(7) << __func__ << " frag bits > 24, dropping" << dendl; | |
10865 | return; | |
10866 | } | |
10867 | ||
7c673cae FG |
10868 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FRAGMENTDIR); |
10869 | mdr->more()->fragment_base = dir->dirfrag(); | |
10870 | ||
10871 | assert(fragments.count(dir->dirfrag()) == 0); | |
10872 | fragment_info_t& info = fragments[dir->dirfrag()]; | |
10873 | info.mdr = mdr; | |
10874 | info.dirs.push_back(dir); | |
10875 | info.bits = bits; | |
10876 | info.last_cum_auth_pins_change = ceph_clock_now(); | |
10877 | ||
10878 | fragment_freeze_dirs(dirs); | |
10879 | // initial mark+complete pass | |
10880 | fragment_mark_and_complete(mdr); | |
10881 | } | |
10882 | ||
10883 | void MDCache::merge_dir(CInode *diri, frag_t frag) | |
10884 | { | |
10885 | dout(7) << "merge_dir to " << frag << " on " << *diri << dendl; | |
10886 | ||
10887 | list<CDir*> dirs; | |
10888 | if (!diri->get_dirfrags_under(frag, dirs)) { | |
10889 | dout(7) << "don't have all frags under " << frag << " for " << *diri << dendl; | |
10890 | return; | |
10891 | } | |
10892 | ||
10893 | if (diri->dirfragtree.is_leaf(frag)) { | |
10894 | dout(10) << " " << frag << " already a leaf for " << *diri << dendl; | |
10895 | return; | |
10896 | } | |
10897 | ||
10898 | if (!can_fragment(diri, dirs)) | |
10899 | return; | |
10900 | ||
10901 | CDir *first = dirs.front(); | |
10902 | int bits = first->get_frag().bits() - frag.bits(); | |
10903 | dout(10) << " we are merginb by " << bits << " bits" << dendl; | |
10904 | ||
10905 | dirfrag_t basedirfrag(diri->ino(), frag); | |
10906 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FRAGMENTDIR); | |
10907 | mdr->more()->fragment_base = basedirfrag; | |
10908 | ||
10909 | assert(fragments.count(basedirfrag) == 0); | |
10910 | fragment_info_t& info = fragments[basedirfrag]; | |
10911 | info.mdr = mdr; | |
10912 | info.dirs = dirs; | |
10913 | info.bits = -bits; | |
10914 | info.last_cum_auth_pins_change = ceph_clock_now(); | |
10915 | ||
10916 | fragment_freeze_dirs(dirs); | |
10917 | // initial mark+complete pass | |
10918 | fragment_mark_and_complete(mdr); | |
10919 | } | |
10920 | ||
10921 | void MDCache::fragment_freeze_dirs(list<CDir*>& dirs) | |
10922 | { | |
10923 | for (list<CDir*>::iterator p = dirs.begin(); p != dirs.end(); ++p) { | |
10924 | CDir *dir = *p; | |
10925 | dir->auth_pin(dir); // until we mark and complete them | |
10926 | dir->state_set(CDir::STATE_FRAGMENTING); | |
10927 | dir->freeze_dir(); | |
10928 | assert(dir->is_freezing_dir()); | |
10929 | } | |
10930 | } | |
10931 | ||
10932 | class C_MDC_FragmentMarking : public MDCacheContext { | |
10933 | MDRequestRef mdr; | |
10934 | public: | |
10935 | C_MDC_FragmentMarking(MDCache *m, MDRequestRef& r) : MDCacheContext(m), mdr(r) {} | |
10936 | void finish(int r) override { | |
10937 | mdcache->fragment_mark_and_complete(mdr); | |
10938 | } | |
10939 | }; | |
10940 | ||
10941 | void MDCache::fragment_mark_and_complete(MDRequestRef& mdr) | |
10942 | { | |
10943 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
10944 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
10945 | if (it == fragments.end() || it->second.mdr != mdr) { | |
10946 | dout(7) << "fragment_mark_and_complete " << basedirfrag << " must have aborted" << dendl; | |
10947 | request_finish(mdr); | |
10948 | return; | |
10949 | } | |
10950 | ||
10951 | fragment_info_t& info = it->second; | |
10952 | CInode *diri = info.dirs.front()->get_inode(); | |
10953 | dout(10) << "fragment_mark_and_complete " << info.dirs << " on " << *diri << dendl; | |
10954 | ||
10955 | MDSGatherBuilder gather(g_ceph_context); | |
10956 | ||
10957 | for (list<CDir*>::iterator p = info.dirs.begin(); | |
10958 | p != info.dirs.end(); | |
10959 | ++p) { | |
10960 | CDir *dir = *p; | |
10961 | ||
10962 | bool ready = true; | |
10963 | if (!dir->is_complete()) { | |
10964 | dout(15) << " fetching incomplete " << *dir << dendl; | |
10965 | dir->fetch(gather.new_sub(), true); // ignore authpinnability | |
10966 | ready = false; | |
10967 | } else if (dir->get_frag() == frag_t()) { | |
10968 | // The COMPLETE flag gets lost if we fragment a new dirfrag, then rollback | |
10969 | // the operation. To avoid CDir::fetch() complaining about missing object, | |
10970 | // we commit new dirfrag first. | |
10971 | if (dir->state_test(CDir::STATE_CREATING)) { | |
10972 | dout(15) << " waiting until new dir gets journaled " << *dir << dendl; | |
10973 | dir->add_waiter(CDir::WAIT_CREATED, gather.new_sub()); | |
10974 | ready = false; | |
10975 | } else if (dir->is_new()) { | |
10976 | dout(15) << " committing new " << *dir << dendl; | |
10977 | assert(dir->is_dirty()); | |
10978 | dir->commit(0, gather.new_sub(), true); | |
10979 | ready = false; | |
10980 | } | |
10981 | } | |
10982 | if (!ready) | |
10983 | continue; | |
10984 | ||
10985 | if (!dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
10986 | dout(15) << " marking " << *dir << dendl; | |
10987 | for (CDir::map_t::iterator p = dir->items.begin(); | |
10988 | p != dir->items.end(); | |
10989 | ++p) { | |
10990 | CDentry *dn = p->second; | |
10991 | dn->get(CDentry::PIN_FRAGMENTING); | |
10992 | assert(!dn->state_test(CDentry::STATE_FRAGMENTING)); | |
10993 | dn->state_set(CDentry::STATE_FRAGMENTING); | |
10994 | } | |
10995 | dir->state_set(CDir::STATE_DNPINNEDFRAG); | |
10996 | dir->auth_unpin(dir); | |
10997 | } else { | |
10998 | dout(15) << " already marked " << *dir << dendl; | |
10999 | } | |
11000 | } | |
11001 | if (gather.has_subs()) { | |
11002 | gather.set_finisher(new C_MDC_FragmentMarking(this, mdr)); | |
11003 | gather.activate(); | |
11004 | return; | |
11005 | } | |
11006 | ||
11007 | for (list<CDir*>::iterator p = info.dirs.begin(); | |
11008 | p != info.dirs.end(); | |
11009 | ++p) { | |
11010 | CDir *dir = *p; | |
11011 | if (!dir->is_frozen_dir()) { | |
11012 | assert(dir->is_freezing_dir()); | |
11013 | dir->add_waiter(CDir::WAIT_FROZEN, gather.new_sub()); | |
11014 | } | |
11015 | } | |
11016 | if (gather.has_subs()) { | |
11017 | gather.set_finisher(new C_MDC_FragmentFrozen(this, mdr)); | |
11018 | gather.activate(); | |
11019 | // flush log so that request auth_pins are retired | |
11020 | mds->mdlog->flush(); | |
11021 | return; | |
11022 | } | |
11023 | ||
11024 | fragment_frozen(mdr, 0); | |
11025 | } | |
11026 | ||
11027 | void MDCache::fragment_unmark_unfreeze_dirs(list<CDir*>& dirs) | |
11028 | { | |
11029 | dout(10) << "fragment_unmark_unfreeze_dirs " << dirs << dendl; | |
11030 | for (list<CDir*>::iterator p = dirs.begin(); p != dirs.end(); ++p) { | |
11031 | CDir *dir = *p; | |
11032 | dout(10) << " frag " << *dir << dendl; | |
11033 | ||
11034 | assert(dir->state_test(CDir::STATE_FRAGMENTING)); | |
11035 | dir->state_clear(CDir::STATE_FRAGMENTING); | |
11036 | ||
11037 | if (dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
11038 | dir->state_clear(CDir::STATE_DNPINNEDFRAG); | |
11039 | ||
11040 | for (CDir::map_t::iterator p = dir->items.begin(); | |
11041 | p != dir->items.end(); | |
11042 | ++p) { | |
11043 | CDentry *dn = p->second; | |
11044 | assert(dn->state_test(CDentry::STATE_FRAGMENTING)); | |
11045 | dn->state_clear(CDentry::STATE_FRAGMENTING); | |
11046 | dn->put(CDentry::PIN_FRAGMENTING); | |
11047 | } | |
11048 | } else { | |
11049 | dir->auth_unpin(dir); | |
11050 | } | |
11051 | ||
11052 | dir->unfreeze_dir(); | |
11053 | } | |
11054 | } | |
11055 | ||
11056 | bool MDCache::fragment_are_all_frozen(CDir *dir) | |
11057 | { | |
11058 | assert(dir->is_frozen_dir()); | |
11059 | map<dirfrag_t,fragment_info_t>::iterator p; | |
11060 | for (p = fragments.lower_bound(dirfrag_t(dir->ino(), 0)); | |
11061 | p != fragments.end() && p->first.ino == dir->ino(); | |
11062 | ++p) { | |
11063 | if (p->first.frag.contains(dir->get_frag())) | |
11064 | return p->second.all_frozen; | |
11065 | } | |
11066 | ceph_abort(); | |
11067 | return false; | |
11068 | } | |
11069 | ||
11070 | void MDCache::fragment_freeze_inc_num_waiters(CDir *dir) | |
11071 | { | |
11072 | map<dirfrag_t,fragment_info_t>::iterator p; | |
11073 | for (p = fragments.lower_bound(dirfrag_t(dir->ino(), 0)); | |
11074 | p != fragments.end() && p->first.ino == dir->ino(); | |
11075 | ++p) { | |
11076 | if (p->first.frag.contains(dir->get_frag())) { | |
11077 | p->second.num_remote_waiters++; | |
11078 | return; | |
11079 | } | |
11080 | } | |
11081 | ceph_abort(); | |
11082 | } | |
11083 | ||
11084 | void MDCache::find_stale_fragment_freeze() | |
11085 | { | |
11086 | dout(10) << "find_stale_fragment_freeze" << dendl; | |
11087 | // see comment in Migrator::find_stale_export_freeze() | |
11088 | utime_t now = ceph_clock_now(); | |
11089 | utime_t cutoff = now; | |
11090 | cutoff -= g_conf->mds_freeze_tree_timeout; | |
11091 | ||
11092 | for (map<dirfrag_t,fragment_info_t>::iterator p = fragments.begin(); | |
11093 | p != fragments.end(); ) { | |
11094 | dirfrag_t df = p->first; | |
11095 | fragment_info_t& info = p->second; | |
11096 | ++p; | |
11097 | if (info.all_frozen) | |
11098 | continue; | |
11099 | CDir *dir; | |
11100 | int total_auth_pins = 0; | |
11101 | for (list<CDir*>::iterator q = info.dirs.begin(); | |
11102 | q != info.dirs.end(); | |
11103 | ++q) { | |
11104 | dir = *q; | |
11105 | if (!dir->state_test(CDir::STATE_DNPINNEDFRAG)) { | |
11106 | total_auth_pins = -1; | |
11107 | break; | |
11108 | } | |
11109 | if (dir->is_frozen_dir()) | |
11110 | continue; | |
11111 | total_auth_pins += dir->get_auth_pins() + dir->get_dir_auth_pins(); | |
11112 | } | |
11113 | if (total_auth_pins < 0) | |
11114 | continue; | |
11115 | if (info.last_cum_auth_pins != total_auth_pins) { | |
11116 | info.last_cum_auth_pins = total_auth_pins; | |
11117 | info.last_cum_auth_pins_change = now; | |
11118 | continue; | |
11119 | } | |
11120 | if (info.last_cum_auth_pins_change >= cutoff) | |
11121 | continue; | |
11122 | dir = info.dirs.front(); | |
11123 | if (info.num_remote_waiters > 0 || | |
11124 | (!dir->inode->is_root() && dir->get_parent_dir()->is_freezing())) { | |
11125 | dout(10) << " cancel fragmenting " << df << " bit " << info.bits << dendl; | |
11126 | list<CDir*> dirs; | |
11127 | info.dirs.swap(dirs); | |
11128 | fragments.erase(df); | |
11129 | fragment_unmark_unfreeze_dirs(dirs); | |
11130 | } | |
11131 | } | |
11132 | } | |
11133 | ||
11134 | class C_MDC_FragmentPrep : public MDCacheLogContext { | |
11135 | MDRequestRef mdr; | |
11136 | public: | |
11137 | C_MDC_FragmentPrep(MDCache *m, MDRequestRef& r) : MDCacheLogContext(m), mdr(r) {} | |
11138 | void finish(int r) override { | |
11139 | mdcache->_fragment_logged(mdr); | |
11140 | } | |
11141 | }; | |
11142 | ||
11143 | class C_MDC_FragmentStore : public MDCacheContext { | |
11144 | MDRequestRef mdr; | |
11145 | public: | |
11146 | C_MDC_FragmentStore(MDCache *m, MDRequestRef& r) : MDCacheContext(m), mdr(r) {} | |
11147 | void finish(int r) override { | |
11148 | mdcache->_fragment_stored(mdr); | |
11149 | } | |
11150 | }; | |
11151 | ||
11152 | class C_MDC_FragmentCommit : public MDCacheLogContext { | |
11153 | dirfrag_t basedirfrag; | |
11154 | list<CDir*> resultfrags; | |
11155 | public: | |
11156 | C_MDC_FragmentCommit(MDCache *m, dirfrag_t df, list<CDir*>& l) : | |
11157 | MDCacheLogContext(m), basedirfrag(df), resultfrags(l) {} | |
11158 | void finish(int r) override { | |
11159 | mdcache->_fragment_committed(basedirfrag, resultfrags); | |
11160 | } | |
11161 | }; | |
11162 | ||
11163 | class C_IO_MDC_FragmentFinish : public MDCacheIOContext { | |
11164 | dirfrag_t basedirfrag; | |
11165 | list<CDir*> resultfrags; | |
11166 | public: | |
11167 | C_IO_MDC_FragmentFinish(MDCache *m, dirfrag_t f, list<CDir*>& l) : | |
11168 | MDCacheIOContext(m), basedirfrag(f) { | |
11169 | resultfrags.swap(l); | |
11170 | } | |
11171 | void finish(int r) override { | |
11172 | assert(r == 0 || r == -ENOENT); | |
11173 | mdcache->_fragment_finish(basedirfrag, resultfrags); | |
11174 | } | |
11175 | }; | |
11176 | ||
11177 | void MDCache::fragment_frozen(MDRequestRef& mdr, int r) | |
11178 | { | |
11179 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11180 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11181 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11182 | dout(7) << "fragment_frozen " << basedirfrag << " must have aborted" << dendl; | |
11183 | request_finish(mdr); | |
11184 | return; | |
11185 | } | |
11186 | ||
11187 | assert(r == 0); | |
11188 | fragment_info_t& info = it->second; | |
11189 | dout(10) << "fragment_frozen " << basedirfrag.frag << " by " << info.bits | |
11190 | << " on " << info.dirs.front()->get_inode() << dendl; | |
11191 | ||
11192 | info.all_frozen = true; | |
11193 | dispatch_fragment_dir(mdr); | |
11194 | } | |
11195 | ||
11196 | void MDCache::dispatch_fragment_dir(MDRequestRef& mdr) | |
11197 | { | |
11198 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11199 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11200 | if (it == fragments.end() || it->second.mdr != mdr) { | |
11201 | dout(7) << "dispatch_fragment_dir " << basedirfrag << " must have aborted" << dendl; | |
11202 | request_finish(mdr); | |
11203 | return; | |
11204 | } | |
11205 | ||
11206 | fragment_info_t& info = it->second; | |
11207 | CInode *diri = info.dirs.front()->get_inode(); | |
11208 | ||
11209 | dout(10) << "dispatch_fragment_dir " << basedirfrag << " bits " << info.bits | |
11210 | << " on " << *diri << dendl; | |
11211 | if (!mdr->aborted) { | |
11212 | set<SimpleLock*> rdlocks, wrlocks, xlocks; | |
11213 | wrlocks.insert(&diri->dirfragtreelock); | |
11214 | // prevent a racing gather on any other scatterlocks too | |
11215 | wrlocks.insert(&diri->nestlock); | |
11216 | wrlocks.insert(&diri->filelock); | |
11217 | if (!mds->locker->acquire_locks(mdr, rdlocks, wrlocks, xlocks, NULL, NULL, true)) | |
11218 | if (!mdr->aborted) | |
11219 | return; | |
11220 | } | |
11221 | ||
11222 | if (mdr->aborted) { | |
11223 | dout(10) << " can't auth_pin " << *diri << ", requeuing dir " | |
11224 | << info.dirs.front()->dirfrag() << dendl; | |
11225 | if (info.bits > 0) | |
11226 | mds->balancer->queue_split(info.dirs.front(), false); | |
11227 | else | |
11228 | mds->balancer->queue_merge(info.dirs.front()); | |
11229 | fragment_unmark_unfreeze_dirs(info.dirs); | |
11230 | fragments.erase(it); | |
11231 | request_finish(mdr); | |
11232 | return; | |
11233 | } | |
11234 | ||
11235 | mdr->ls = mds->mdlog->get_current_segment(); | |
11236 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_PREPARE, basedirfrag, info.bits); | |
11237 | mds->mdlog->start_entry(le); | |
11238 | ||
11239 | for (list<CDir*>::iterator p = info.dirs.begin(); p != info.dirs.end(); ++p) { | |
11240 | CDir *dir = *p; | |
11241 | dirfrag_rollback rollback; | |
11242 | rollback.fnode = dir->fnode; | |
11243 | le->add_orig_frag(dir->get_frag(), &rollback); | |
11244 | } | |
11245 | ||
11246 | // refragment | |
11247 | list<MDSInternalContextBase*> waiters; | |
11248 | adjust_dir_fragments(diri, info.dirs, basedirfrag.frag, info.bits, | |
11249 | info.resultfrags, waiters, false); | |
11250 | if (g_conf->mds_debug_frag) | |
11251 | diri->verify_dirfrags(); | |
11252 | mds->queue_waiters(waiters); | |
11253 | ||
11254 | for (list<frag_t>::iterator p = le->orig_frags.begin(); p != le->orig_frags.end(); ++p) | |
11255 | assert(!diri->dirfragtree.is_leaf(*p)); | |
11256 | ||
11257 | le->metablob.add_dir_context(*info.resultfrags.begin()); | |
11258 | for (list<CDir*>::iterator p = info.resultfrags.begin(); | |
11259 | p != info.resultfrags.end(); | |
11260 | ++p) { | |
11261 | if (diri->is_auth()) { | |
11262 | le->metablob.add_fragmented_dir(*p, false, false); | |
11263 | } else { | |
11264 | (*p)->state_set(CDir::STATE_DIRTYDFT); | |
11265 | le->metablob.add_fragmented_dir(*p, false, true); | |
11266 | } | |
11267 | } | |
11268 | ||
11269 | // dft lock | |
11270 | if (diri->is_auth()) { | |
11271 | // journal dirfragtree | |
11272 | inode_t *pi = diri->project_inode(); | |
11273 | pi->version = diri->pre_dirty(); | |
11274 | journal_dirty_inode(mdr.get(), &le->metablob, diri); | |
11275 | } else { | |
11276 | mds->locker->mark_updated_scatterlock(&diri->dirfragtreelock); | |
11277 | mdr->ls->dirty_dirfrag_dirfragtree.push_back(&diri->item_dirty_dirfrag_dirfragtree); | |
11278 | mdr->add_updated_lock(&diri->dirfragtreelock); | |
11279 | } | |
11280 | ||
11281 | /* | |
11282 | // filelock | |
11283 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
11284 | mut->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
11285 | mut->add_updated_lock(&diri->filelock); | |
11286 | ||
11287 | // dirlock | |
11288 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
11289 | mut->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
11290 | mut->add_updated_lock(&diri->nestlock); | |
11291 | */ | |
11292 | ||
11293 | add_uncommitted_fragment(basedirfrag, info.bits, le->orig_frags, mdr->ls); | |
11294 | mds->server->submit_mdlog_entry(le, new C_MDC_FragmentPrep(this, mdr), | |
11295 | mdr, __func__); | |
11296 | mds->mdlog->flush(); | |
11297 | } | |
11298 | ||
11299 | void MDCache::_fragment_logged(MDRequestRef& mdr) | |
11300 | { | |
11301 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11302 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11303 | assert(it != fragments.end()); | |
11304 | fragment_info_t &info = it->second; | |
11305 | CInode *diri = info.resultfrags.front()->get_inode(); | |
11306 | ||
11307 | dout(10) << "fragment_logged " << basedirfrag << " bits " << info.bits | |
11308 | << " on " << *diri << dendl; | |
11309 | ||
11310 | if (diri->is_auth()) | |
11311 | diri->pop_and_dirty_projected_inode(mdr->ls); | |
11312 | ||
11313 | mdr->apply(); // mark scatterlock | |
11314 | ||
11315 | // store resulting frags | |
11316 | MDSGatherBuilder gather(g_ceph_context, new C_MDC_FragmentStore(this, mdr)); | |
11317 | ||
11318 | for (list<CDir*>::iterator p = info.resultfrags.begin(); | |
11319 | p != info.resultfrags.end(); | |
11320 | ++p) { | |
11321 | CDir *dir = *p; | |
11322 | dout(10) << " storing result frag " << *dir << dendl; | |
11323 | ||
11324 | // freeze and store them too | |
11325 | dir->auth_pin(this); | |
11326 | dir->state_set(CDir::STATE_FRAGMENTING); | |
11327 | dir->commit(0, gather.new_sub(), true); // ignore authpinnability | |
11328 | } | |
11329 | ||
11330 | gather.activate(); | |
11331 | } | |
11332 | ||
11333 | void MDCache::_fragment_stored(MDRequestRef& mdr) | |
11334 | { | |
11335 | dirfrag_t basedirfrag = mdr->more()->fragment_base; | |
11336 | map<dirfrag_t,fragment_info_t>::iterator it = fragments.find(basedirfrag); | |
11337 | assert(it != fragments.end()); | |
11338 | fragment_info_t &info = it->second; | |
11339 | CInode *diri = info.resultfrags.front()->get_inode(); | |
11340 | ||
11341 | dout(10) << "fragment_stored " << basedirfrag << " bits " << info.bits | |
11342 | << " on " << *diri << dendl; | |
11343 | ||
11344 | // tell peers | |
11345 | CDir *first = *info.resultfrags.begin(); | |
181888fb FG |
11346 | for (const auto &p : first->get_replicas()) { |
11347 | if (mds->mdsmap->get_state(p.first) < MDSMap::STATE_REJOIN || | |
11348 | (mds->mdsmap->get_state(p.first) == MDSMap::STATE_REJOIN && | |
11349 | rejoin_gather.count(p.first))) | |
7c673cae FG |
11350 | continue; |
11351 | ||
11352 | MMDSFragmentNotify *notify = new MMDSFragmentNotify(basedirfrag, info.bits); | |
11353 | ||
11354 | // freshly replicate new dirs to peers | |
11355 | for (list<CDir*>::iterator q = info.resultfrags.begin(); | |
11356 | q != info.resultfrags.end(); | |
11357 | ++q) | |
181888fb | 11358 | replicate_dir(*q, p.first, notify->basebl); |
7c673cae | 11359 | |
181888fb | 11360 | mds->send_message_mds(notify, p.first); |
7c673cae FG |
11361 | } |
11362 | ||
11363 | // journal commit | |
11364 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_COMMIT, basedirfrag, info.bits); | |
11365 | mds->mdlog->start_submit_entry(le, new C_MDC_FragmentCommit(this, basedirfrag, | |
11366 | info.resultfrags)); | |
11367 | ||
11368 | mds->locker->drop_locks(mdr.get()); | |
11369 | ||
11370 | // unfreeze resulting frags | |
11371 | for (list<CDir*>::iterator p = info.resultfrags.begin(); | |
11372 | p != info.resultfrags.end(); | |
11373 | ++p) { | |
11374 | CDir *dir = *p; | |
11375 | dout(10) << " result frag " << *dir << dendl; | |
11376 | ||
11377 | for (CDir::map_t::iterator p = dir->items.begin(); | |
11378 | p != dir->items.end(); | |
11379 | ++p) { | |
11380 | CDentry *dn = p->second; | |
11381 | assert(dn->state_test(CDentry::STATE_FRAGMENTING)); | |
11382 | dn->state_clear(CDentry::STATE_FRAGMENTING); | |
11383 | dn->put(CDentry::PIN_FRAGMENTING); | |
11384 | } | |
11385 | ||
11386 | // unfreeze | |
11387 | dir->unfreeze_dir(); | |
11388 | } | |
11389 | ||
11390 | fragments.erase(it); | |
11391 | request_finish(mdr); | |
11392 | } | |
11393 | ||
11394 | void MDCache::_fragment_committed(dirfrag_t basedirfrag, list<CDir*>& resultfrags) | |
11395 | { | |
11396 | dout(10) << "fragment_committed " << basedirfrag << dendl; | |
11397 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
11398 | assert(it != uncommitted_fragments.end()); | |
11399 | ufragment &uf = it->second; | |
11400 | ||
11401 | // remove old frags | |
11402 | C_GatherBuilder gather( | |
11403 | g_ceph_context, | |
11404 | new C_OnFinisher( | |
11405 | new C_IO_MDC_FragmentFinish(this, basedirfrag, resultfrags), | |
11406 | mds->finisher)); | |
11407 | ||
11408 | SnapContext nullsnapc; | |
11409 | object_locator_t oloc(mds->mdsmap->get_metadata_pool()); | |
11410 | for (list<frag_t>::iterator p = uf.old_frags.begin(); | |
11411 | p != uf.old_frags.end(); | |
11412 | ++p) { | |
11413 | object_t oid = CInode::get_object_name(basedirfrag.ino, *p, ""); | |
11414 | ObjectOperation op; | |
11415 | if (*p == frag_t()) { | |
11416 | // backtrace object | |
11417 | dout(10) << " truncate orphan dirfrag " << oid << dendl; | |
11418 | op.truncate(0); | |
11419 | op.omap_clear(); | |
11420 | } else { | |
11421 | dout(10) << " removing orphan dirfrag " << oid << dendl; | |
11422 | op.remove(); | |
11423 | } | |
11424 | mds->objecter->mutate(oid, oloc, op, nullsnapc, | |
11425 | ceph::real_clock::now(), | |
11426 | 0, gather.new_sub()); | |
11427 | } | |
11428 | ||
11429 | assert(gather.has_subs()); | |
11430 | gather.activate(); | |
11431 | } | |
11432 | ||
11433 | void MDCache::_fragment_finish(dirfrag_t basedirfrag, list<CDir*>& resultfrags) | |
11434 | { | |
11435 | dout(10) << "fragment_finish " << basedirfrag << "resultfrags.size=" | |
11436 | << resultfrags.size() << dendl; | |
11437 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
11438 | assert(it != uncommitted_fragments.end()); | |
11439 | ufragment &uf = it->second; | |
11440 | ||
11441 | // unmark & auth_unpin | |
11442 | for (const auto &dir : resultfrags) { | |
11443 | dir->state_clear(CDir::STATE_FRAGMENTING); | |
11444 | dir->auth_unpin(this); | |
11445 | ||
11446 | // In case the resulting fragments are beyond the split size, | |
11447 | // we might need to split them again right away (they could | |
11448 | // have been taking inserts between unfreezing and getting | |
11449 | // here) | |
11450 | mds->balancer->maybe_fragment(dir, false); | |
11451 | } | |
11452 | ||
11453 | if (mds->logger) { | |
11454 | if (resultfrags.size() > 1) { | |
11455 | mds->logger->inc(l_mds_dir_split); | |
11456 | } else { | |
11457 | mds->logger->inc(l_mds_dir_merge); | |
11458 | } | |
11459 | } | |
11460 | ||
11461 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_FINISH, basedirfrag, uf.bits); | |
11462 | mds->mdlog->start_submit_entry(le); | |
11463 | ||
11464 | finish_uncommitted_fragment(basedirfrag, EFragment::OP_FINISH); | |
11465 | } | |
11466 | ||
11467 | /* This function DOES put the passed message before returning */ | |
11468 | void MDCache::handle_fragment_notify(MMDSFragmentNotify *notify) | |
11469 | { | |
11470 | dout(10) << "handle_fragment_notify " << *notify << " from " << notify->get_source() << dendl; | |
11471 | ||
11472 | if (mds->get_state() < MDSMap::STATE_REJOIN) { | |
11473 | notify->put(); | |
11474 | return; | |
11475 | } | |
11476 | ||
11477 | CInode *diri = get_inode(notify->get_ino()); | |
11478 | if (diri) { | |
11479 | frag_t base = notify->get_basefrag(); | |
11480 | int bits = notify->get_bits(); | |
11481 | ||
11482 | /* | |
11483 | if ((bits < 0 && diri->dirfragtree.is_leaf(base)) || | |
11484 | (bits > 0 && !diri->dirfragtree.is_leaf(base))) { | |
11485 | dout(10) << " dft " << diri->dirfragtree << " state doesn't match " << base << " by " << bits | |
11486 | << ", must have found out during resolve/rejoin? ignoring. " << *diri << dendl; | |
11487 | notify->put(); | |
11488 | return; | |
11489 | } | |
11490 | */ | |
11491 | ||
11492 | // refragment | |
11493 | list<MDSInternalContextBase*> waiters; | |
11494 | list<CDir*> resultfrags; | |
11495 | adjust_dir_fragments(diri, base, bits, resultfrags, waiters, false); | |
11496 | if (g_conf->mds_debug_frag) | |
11497 | diri->verify_dirfrags(); | |
11498 | ||
11499 | for (list<CDir*>::iterator p = resultfrags.begin(); p != resultfrags.end(); ++p) | |
11500 | diri->take_dir_waiting((*p)->get_frag(), waiters); | |
11501 | ||
11502 | // add new replica dirs values | |
11503 | bufferlist::iterator p = notify->basebl.begin(); | |
11504 | while (!p.end()) | |
11505 | add_replica_dir(p, diri, mds_rank_t(notify->get_source().num()), waiters); | |
11506 | ||
11507 | mds->queue_waiters(waiters); | |
11508 | } else { | |
11509 | ceph_abort(); | |
11510 | } | |
11511 | ||
11512 | notify->put(); | |
11513 | } | |
11514 | ||
11515 | void MDCache::add_uncommitted_fragment(dirfrag_t basedirfrag, int bits, list<frag_t>& old_frags, | |
11516 | LogSegment *ls, bufferlist *rollback) | |
11517 | { | |
11518 | dout(10) << "add_uncommitted_fragment: base dirfrag " << basedirfrag << " bits " << bits << dendl; | |
11519 | assert(!uncommitted_fragments.count(basedirfrag)); | |
11520 | ufragment& uf = uncommitted_fragments[basedirfrag]; | |
11521 | uf.old_frags = old_frags; | |
11522 | uf.bits = bits; | |
11523 | uf.ls = ls; | |
11524 | ls->uncommitted_fragments.insert(basedirfrag); | |
11525 | if (rollback) | |
11526 | uf.rollback.swap(*rollback); | |
11527 | } | |
11528 | ||
11529 | void MDCache::finish_uncommitted_fragment(dirfrag_t basedirfrag, int op) | |
11530 | { | |
11531 | dout(10) << "finish_uncommitted_fragments: base dirfrag " << basedirfrag | |
11532 | << " op " << EFragment::op_name(op) << dendl; | |
11533 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
11534 | if (it != uncommitted_fragments.end()) { | |
11535 | ufragment& uf = it->second; | |
11536 | if (op != EFragment::OP_FINISH && !uf.old_frags.empty()) { | |
11537 | uf.committed = true; | |
11538 | } else { | |
11539 | uf.ls->uncommitted_fragments.erase(basedirfrag); | |
11540 | mds->queue_waiters(uf.waiters); | |
11541 | uncommitted_fragments.erase(it); | |
11542 | } | |
11543 | } | |
11544 | } | |
11545 | ||
11546 | void MDCache::rollback_uncommitted_fragment(dirfrag_t basedirfrag, list<frag_t>& old_frags) | |
11547 | { | |
11548 | dout(10) << "rollback_uncommitted_fragment: base dirfrag " << basedirfrag | |
11549 | << " old_frags (" << old_frags << ")" << dendl; | |
11550 | map<dirfrag_t, ufragment>::iterator it = uncommitted_fragments.find(basedirfrag); | |
11551 | if (it != uncommitted_fragments.end()) { | |
11552 | ufragment& uf = it->second; | |
11553 | if (!uf.old_frags.empty()) { | |
11554 | uf.old_frags.swap(old_frags); | |
11555 | uf.committed = true; | |
11556 | } else { | |
11557 | uf.ls->uncommitted_fragments.erase(basedirfrag); | |
11558 | uncommitted_fragments.erase(it); | |
11559 | } | |
11560 | } | |
11561 | } | |
11562 | ||
11563 | void MDCache::rollback_uncommitted_fragments() | |
11564 | { | |
11565 | dout(10) << "rollback_uncommitted_fragments: " << uncommitted_fragments.size() << " pending" << dendl; | |
11566 | for (map<dirfrag_t, ufragment>::iterator p = uncommitted_fragments.begin(); | |
11567 | p != uncommitted_fragments.end(); | |
11568 | ++p) { | |
11569 | ufragment &uf = p->second; | |
11570 | CInode *diri = get_inode(p->first.ino); | |
11571 | assert(diri); | |
11572 | ||
11573 | if (uf.committed) { | |
11574 | list<CDir*> frags; | |
11575 | diri->get_dirfrags_under(p->first.frag, frags); | |
11576 | for (list<CDir*>::iterator q = frags.begin(); q != frags.end(); ++q) { | |
11577 | CDir *dir = *q; | |
11578 | dir->auth_pin(this); | |
11579 | dir->state_set(CDir::STATE_FRAGMENTING); | |
11580 | } | |
11581 | _fragment_committed(p->first, frags); | |
11582 | continue; | |
11583 | } | |
11584 | ||
11585 | dout(10) << " rolling back " << p->first << " refragment by " << uf.bits << " bits" << dendl; | |
11586 | ||
11587 | LogSegment *ls = mds->mdlog->get_current_segment(); | |
11588 | EFragment *le = new EFragment(mds->mdlog, EFragment::OP_ROLLBACK, p->first, uf.bits); | |
11589 | mds->mdlog->start_entry(le); | |
11590 | bool diri_auth = (diri->authority() != CDIR_AUTH_UNDEF); | |
11591 | ||
11592 | list<frag_t> old_frags; | |
11593 | diri->dirfragtree.get_leaves_under(p->first.frag, old_frags); | |
11594 | ||
11595 | list<CDir*> resultfrags; | |
11596 | if (uf.old_frags.empty()) { | |
11597 | // created by old format EFragment | |
11598 | list<MDSInternalContextBase*> waiters; | |
11599 | adjust_dir_fragments(diri, p->first.frag, -uf.bits, resultfrags, waiters, true); | |
11600 | } else { | |
11601 | bufferlist::iterator bp = uf.rollback.begin(); | |
11602 | for (list<frag_t>::iterator q = uf.old_frags.begin(); q != uf.old_frags.end(); ++q) { | |
11603 | CDir *dir = force_dir_fragment(diri, *q); | |
11604 | resultfrags.push_back(dir); | |
11605 | ||
11606 | dirfrag_rollback rollback; | |
11607 | ::decode(rollback, bp); | |
11608 | ||
11609 | dir->set_version(rollback.fnode.version); | |
11610 | dir->fnode = rollback.fnode; | |
11611 | ||
11612 | dir->_mark_dirty(ls); | |
11613 | ||
11614 | if (!(dir->fnode.rstat == dir->fnode.accounted_rstat)) { | |
11615 | dout(10) << " dirty nestinfo on " << *dir << dendl; | |
11616 | mds->locker->mark_updated_scatterlock(&dir->inode->nestlock); | |
11617 | ls->dirty_dirfrag_nest.push_back(&dir->inode->item_dirty_dirfrag_nest); | |
11618 | } | |
11619 | if (!(dir->fnode.fragstat == dir->fnode.accounted_fragstat)) { | |
11620 | dout(10) << " dirty fragstat on " << *dir << dendl; | |
11621 | mds->locker->mark_updated_scatterlock(&dir->inode->filelock); | |
11622 | ls->dirty_dirfrag_dir.push_back(&dir->inode->item_dirty_dirfrag_dir); | |
11623 | } | |
11624 | ||
11625 | le->add_orig_frag(dir->get_frag()); | |
11626 | le->metablob.add_dir_context(dir); | |
11627 | if (diri_auth) { | |
11628 | le->metablob.add_fragmented_dir(dir, true, false); | |
11629 | } else { | |
11630 | dout(10) << " dirty dirfragtree on " << *dir << dendl; | |
11631 | dir->state_set(CDir::STATE_DIRTYDFT); | |
11632 | le->metablob.add_fragmented_dir(dir, true, true); | |
11633 | } | |
11634 | } | |
11635 | } | |
11636 | ||
11637 | if (diri_auth) { | |
11638 | diri->project_inode()->version = diri->pre_dirty(); | |
11639 | diri->pop_and_dirty_projected_inode(ls); // hacky | |
11640 | le->metablob.add_primary_dentry(diri->get_projected_parent_dn(), diri, true); | |
11641 | } else { | |
11642 | mds->locker->mark_updated_scatterlock(&diri->dirfragtreelock); | |
11643 | ls->dirty_dirfrag_dirfragtree.push_back(&diri->item_dirty_dirfrag_dirfragtree); | |
11644 | } | |
11645 | ||
11646 | if (g_conf->mds_debug_frag) | |
11647 | diri->verify_dirfrags(); | |
11648 | ||
11649 | for (list<frag_t>::iterator q = old_frags.begin(); q != old_frags.end(); ++q) | |
11650 | assert(!diri->dirfragtree.is_leaf(*q)); | |
11651 | ||
11652 | for (list<CDir*>::iterator q = resultfrags.begin(); q != resultfrags.end(); ++q) { | |
11653 | CDir *dir = *q; | |
11654 | dir->auth_pin(this); | |
11655 | dir->state_set(CDir::STATE_FRAGMENTING); | |
11656 | } | |
11657 | ||
11658 | mds->mdlog->submit_entry(le); | |
11659 | ||
11660 | uf.old_frags.swap(old_frags); | |
11661 | _fragment_committed(p->first, resultfrags); | |
11662 | } | |
11663 | } | |
11664 | ||
11665 | void MDCache::force_readonly() | |
11666 | { | |
11667 | if (is_readonly()) | |
11668 | return; | |
11669 | ||
11670 | dout(1) << "force file system read-only" << dendl; | |
11671 | mds->clog->warn() << "force file system read-only"; | |
11672 | ||
11673 | set_readonly(); | |
11674 | ||
11675 | mds->server->force_clients_readonly(); | |
11676 | ||
11677 | // revoke write caps | |
11678 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator p = inode_map.begin(); | |
11679 | p != inode_map.end(); | |
11680 | ++p) { | |
11681 | CInode *in = p->second; | |
11682 | if (in->is_head()) | |
11683 | mds->locker->eval(in, CEPH_CAP_LOCKS); | |
11684 | } | |
11685 | ||
11686 | mds->mdlog->flush(); | |
11687 | } | |
11688 | ||
11689 | ||
11690 | // ============================================================== | |
11691 | // debug crap | |
11692 | ||
11693 | void MDCache::show_subtrees(int dbl) | |
11694 | { | |
11695 | if (g_conf->mds_thrash_exports) | |
11696 | dbl += 15; | |
11697 | ||
11698 | //dout(10) << "show_subtrees" << dendl; | |
11699 | ||
11700 | if (!g_conf->subsys.should_gather(ceph_subsys_mds, dbl)) | |
11701 | return; // i won't print anything. | |
11702 | ||
11703 | if (subtrees.empty()) { | |
11704 | dout(dbl) << "show_subtrees - no subtrees" << dendl; | |
11705 | return; | |
11706 | } | |
11707 | ||
11708 | // root frags | |
11709 | list<CDir*> basefrags; | |
11710 | for (set<CInode*>::iterator p = base_inodes.begin(); | |
11711 | p != base_inodes.end(); | |
11712 | ++p) | |
11713 | (*p)->get_dirfrags(basefrags); | |
11714 | //dout(15) << "show_subtrees, base dirfrags " << basefrags << dendl; | |
11715 | dout(15) << "show_subtrees" << dendl; | |
11716 | ||
11717 | // queue stuff | |
11718 | list<pair<CDir*,int> > q; | |
11719 | string indent; | |
11720 | set<CDir*> seen; | |
11721 | ||
11722 | // calc max depth | |
11723 | for (list<CDir*>::iterator p = basefrags.begin(); p != basefrags.end(); ++p) | |
11724 | q.push_back(pair<CDir*,int>(*p, 0)); | |
11725 | ||
11726 | set<CDir*> subtrees_seen; | |
11727 | ||
11728 | int depth = 0; | |
11729 | while (!q.empty()) { | |
11730 | CDir *dir = q.front().first; | |
11731 | int d = q.front().second; | |
11732 | q.pop_front(); | |
11733 | ||
11734 | if (subtrees.count(dir) == 0) continue; | |
11735 | ||
11736 | subtrees_seen.insert(dir); | |
11737 | ||
11738 | if (d > depth) depth = d; | |
11739 | ||
11740 | // sanity check | |
11741 | //dout(25) << "saw depth " << d << " " << *dir << dendl; | |
11742 | if (seen.count(dir)) dout(0) << "aah, already seen " << *dir << dendl; | |
11743 | assert(seen.count(dir) == 0); | |
11744 | seen.insert(dir); | |
11745 | ||
11746 | // nested items? | |
11747 | if (!subtrees[dir].empty()) { | |
11748 | for (set<CDir*>::iterator p = subtrees[dir].begin(); | |
11749 | p != subtrees[dir].end(); | |
11750 | ++p) { | |
11751 | //dout(25) << " saw sub " << **p << dendl; | |
11752 | q.push_front(pair<CDir*,int>(*p, d+1)); | |
11753 | } | |
11754 | } | |
11755 | } | |
11756 | ||
11757 | ||
11758 | // print tree | |
11759 | for (list<CDir*>::iterator p = basefrags.begin(); p != basefrags.end(); ++p) | |
11760 | q.push_back(pair<CDir*,int>(*p, 0)); | |
11761 | ||
11762 | while (!q.empty()) { | |
11763 | CDir *dir = q.front().first; | |
11764 | int d = q.front().second; | |
11765 | q.pop_front(); | |
11766 | ||
11767 | if (subtrees.count(dir) == 0) continue; | |
11768 | ||
11769 | // adjust indenter | |
11770 | while ((unsigned)d < indent.size()) | |
11771 | indent.resize(d); | |
11772 | ||
11773 | // pad | |
11774 | string pad = "______________________________________"; | |
11775 | pad.resize(depth*2+1-indent.size()); | |
11776 | if (!subtrees[dir].empty()) | |
11777 | pad[0] = '.'; // parent | |
11778 | ||
11779 | ||
11780 | string auth; | |
11781 | if (dir->is_auth()) | |
11782 | auth = "auth "; | |
11783 | else | |
11784 | auth = " rep "; | |
11785 | ||
11786 | char s[10]; | |
11787 | if (dir->get_dir_auth().second == CDIR_AUTH_UNKNOWN) | |
11788 | snprintf(s, sizeof(s), "%2d ", int(dir->get_dir_auth().first)); | |
11789 | else | |
11790 | snprintf(s, sizeof(s), "%2d,%2d", int(dir->get_dir_auth().first), int(dir->get_dir_auth().second)); | |
11791 | ||
11792 | ||
11793 | dout(dbl) << indent << "|_" << pad << s << " " << auth << *dir << dendl; | |
11794 | ||
11795 | if (dir->ino() == MDS_INO_ROOT) | |
11796 | assert(dir->inode == root); | |
11797 | if (dir->ino() == MDS_INO_MDSDIR(mds->get_nodeid())) | |
11798 | assert(dir->inode == myin); | |
11799 | if (dir->inode->is_stray() && (MDS_INO_STRAY_OWNER(dir->ino()) == mds->get_nodeid())) | |
11800 | assert(strays[MDS_INO_STRAY_INDEX(dir->ino())] == dir->inode); | |
11801 | ||
11802 | // nested items? | |
11803 | if (!subtrees[dir].empty()) { | |
11804 | // more at my level? | |
11805 | if (!q.empty() && q.front().second == d) | |
11806 | indent += "| "; | |
11807 | else | |
11808 | indent += " "; | |
11809 | ||
11810 | for (set<CDir*>::iterator p = subtrees[dir].begin(); | |
11811 | p != subtrees[dir].end(); | |
11812 | ++p) | |
11813 | q.push_front(pair<CDir*,int>(*p, d+2)); | |
11814 | } | |
11815 | } | |
11816 | ||
11817 | // verify there isn't stray crap in subtree map | |
11818 | int lost = 0; | |
11819 | for (map<CDir*, set<CDir*> >::iterator p = subtrees.begin(); | |
11820 | p != subtrees.end(); | |
11821 | ++p) { | |
11822 | if (subtrees_seen.count(p->first)) continue; | |
11823 | dout(10) << "*** stray/lost entry in subtree map: " << *p->first << dendl; | |
11824 | lost++; | |
11825 | } | |
11826 | assert(lost == 0); | |
11827 | } | |
11828 | ||
11829 | ||
11830 | void MDCache::show_cache() | |
11831 | { | |
11832 | dout(7) << "show_cache" << dendl; | |
11833 | ||
11834 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator it = inode_map.begin(); | |
11835 | it != inode_map.end(); | |
11836 | ++it) { | |
11837 | // unlinked? | |
11838 | if (!it->second->parent) | |
11839 | dout(7) << " unlinked " << *it->second << dendl; | |
11840 | ||
11841 | // dirfrags? | |
11842 | list<CDir*> dfs; | |
11843 | it->second->get_dirfrags(dfs); | |
11844 | for (list<CDir*>::iterator p = dfs.begin(); p != dfs.end(); ++p) { | |
11845 | CDir *dir = *p; | |
11846 | dout(7) << " dirfrag " << *dir << dendl; | |
11847 | ||
11848 | for (CDir::map_t::iterator p = dir->items.begin(); | |
11849 | p != dir->items.end(); | |
11850 | ++p) { | |
11851 | CDentry *dn = p->second; | |
11852 | dout(7) << " dentry " << *dn << dendl; | |
11853 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
11854 | if (dnl->is_primary() && dnl->get_inode()) | |
11855 | dout(7) << " inode " << *dnl->get_inode() << dendl; | |
11856 | } | |
11857 | } | |
11858 | } | |
11859 | } | |
11860 | ||
181888fb FG |
11861 | int MDCache::cache_status(Formatter *f) |
11862 | { | |
11863 | f->open_object_section("cache"); | |
11864 | ||
11865 | f->open_object_section("pool"); | |
11866 | mempool::get_pool(mempool::mds_co::id).dump(f); | |
11867 | f->close_section(); | |
11868 | ||
11869 | f->close_section(); | |
11870 | return 0; | |
11871 | } | |
11872 | ||
31f18b77 | 11873 | int MDCache::dump_cache(std::string const &file_name) |
7c673cae | 11874 | { |
31f18b77 | 11875 | return dump_cache(file_name.c_str(), NULL); |
7c673cae FG |
11876 | } |
11877 | ||
31f18b77 | 11878 | int MDCache::dump_cache(Formatter *f) |
7c673cae | 11879 | { |
31f18b77 | 11880 | return dump_cache(NULL, f); |
7c673cae FG |
11881 | } |
11882 | ||
31f18b77 | 11883 | int MDCache::dump_cache(const string& dump_root, int depth, Formatter *f) |
7c673cae | 11884 | { |
31f18b77 | 11885 | return dump_cache(NULL, f, dump_root, depth); |
7c673cae FG |
11886 | } |
11887 | ||
11888 | /** | |
11889 | * Dump the metadata cache, either to a Formatter, if | |
11890 | * provided, else to a plain text file. | |
11891 | */ | |
31f18b77 | 11892 | int MDCache::dump_cache(const char *fn, Formatter *f, |
7c673cae FG |
11893 | const string& dump_root, int depth) |
11894 | { | |
11895 | int r = 0; | |
11896 | int fd = -1; | |
11897 | ||
11898 | if (f) { | |
11899 | f->open_array_section("inodes"); | |
11900 | } else { | |
11901 | char deffn[200]; | |
11902 | if (!fn) { | |
11903 | snprintf(deffn, sizeof(deffn), "cachedump.%d.mds%d", (int)mds->mdsmap->get_epoch(), int(mds->get_nodeid())); | |
11904 | fn = deffn; | |
11905 | } | |
11906 | ||
11907 | dout(1) << "dump_cache to " << fn << dendl; | |
11908 | ||
11909 | fd = ::open(fn, O_WRONLY|O_CREAT|O_EXCL, 0600); | |
11910 | if (fd < 0) { | |
11911 | derr << "failed to open " << fn << ": " << cpp_strerror(errno) << dendl; | |
31f18b77 | 11912 | return errno; |
7c673cae FG |
11913 | } |
11914 | } | |
11915 | ||
11916 | for (ceph::unordered_map<vinodeno_t,CInode*>::iterator it = inode_map.begin(); | |
11917 | it != inode_map.end(); | |
11918 | ++it) { | |
11919 | CInode *in = it->second; | |
11920 | ||
11921 | if (!dump_root.empty()) { | |
11922 | string ipath; | |
11923 | if (in->is_root()) | |
11924 | ipath = "/"; | |
11925 | else | |
11926 | in->make_path_string(ipath); | |
11927 | ||
11928 | if (dump_root.length() > ipath.length() || | |
11929 | !equal(dump_root.begin(), dump_root.end(), ipath.begin())) | |
11930 | continue; | |
11931 | ||
11932 | if (depth >= 0 && | |
11933 | count(ipath.begin() + dump_root.length(), ipath.end(), '/') > depth) | |
11934 | continue; | |
11935 | } | |
11936 | ||
11937 | if (f) { | |
11938 | f->open_object_section("inode"); | |
11939 | in->dump(f); | |
11940 | } else { | |
11941 | ostringstream ss; | |
11942 | ss << *in << std::endl; | |
11943 | std::string s = ss.str(); | |
11944 | r = safe_write(fd, s.c_str(), s.length()); | |
11945 | if (r < 0) { | |
11946 | goto out; | |
11947 | } | |
11948 | } | |
11949 | ||
11950 | list<CDir*> dfs; | |
11951 | in->get_dirfrags(dfs); | |
11952 | if (f) { | |
11953 | f->open_array_section("dirfrags"); | |
11954 | } | |
11955 | for (list<CDir*>::iterator p = dfs.begin(); p != dfs.end(); ++p) { | |
11956 | CDir *dir = *p; | |
11957 | if (f) { | |
11958 | f->open_object_section("dir"); | |
11959 | dir->dump(f); | |
11960 | } else { | |
11961 | ostringstream tt; | |
11962 | tt << " " << *dir << std::endl; | |
11963 | string t = tt.str(); | |
11964 | r = safe_write(fd, t.c_str(), t.length()); | |
11965 | if (r < 0) { | |
11966 | goto out; | |
11967 | } | |
11968 | } | |
11969 | ||
11970 | if (f) { | |
11971 | f->open_array_section("dentries"); | |
11972 | } | |
11973 | for (CDir::map_t::iterator q = dir->items.begin(); | |
11974 | q != dir->items.end(); | |
11975 | ++q) { | |
11976 | CDentry *dn = q->second; | |
11977 | if (f) { | |
11978 | f->open_object_section("dentry"); | |
11979 | dn->dump(f); | |
11980 | f->close_section(); | |
11981 | } else { | |
11982 | ostringstream uu; | |
11983 | uu << " " << *dn << std::endl; | |
11984 | string u = uu.str(); | |
11985 | r = safe_write(fd, u.c_str(), u.length()); | |
11986 | if (r < 0) { | |
11987 | goto out; | |
11988 | } | |
11989 | } | |
11990 | } | |
11991 | if (f) { | |
11992 | f->close_section(); //dentries | |
11993 | } | |
11994 | dir->check_rstats(); | |
11995 | if (f) { | |
11996 | f->close_section(); //dir | |
11997 | } | |
11998 | } | |
11999 | if (f) { | |
12000 | f->close_section(); // dirfrags | |
12001 | } | |
12002 | ||
12003 | if (f) { | |
12004 | f->close_section(); // inode | |
12005 | } | |
12006 | } | |
12007 | ||
12008 | out: | |
12009 | if (f) { | |
12010 | f->close_section(); // inodes | |
12011 | } else { | |
12012 | ::close(fd); | |
12013 | } | |
31f18b77 | 12014 | return r; |
7c673cae FG |
12015 | } |
12016 | ||
12017 | ||
12018 | ||
12019 | C_MDS_RetryRequest::C_MDS_RetryRequest(MDCache *c, MDRequestRef& r) | |
12020 | : MDSInternalContext(c->mds), cache(c), mdr(r) | |
12021 | {} | |
12022 | ||
12023 | void C_MDS_RetryRequest::finish(int r) | |
12024 | { | |
12025 | mdr->retry++; | |
12026 | cache->dispatch_request(mdr); | |
12027 | } | |
12028 | ||
12029 | ||
12030 | class C_MDS_EnqueueScrub : public Context | |
12031 | { | |
12032 | Formatter *formatter; | |
12033 | Context *on_finish; | |
12034 | public: | |
12035 | ScrubHeaderRef header; | |
12036 | C_MDS_EnqueueScrub(Formatter *f, Context *fin) : | |
12037 | formatter(f), on_finish(fin), header(nullptr) {} | |
12038 | ||
12039 | Context *take_finisher() { | |
12040 | Context *fin = on_finish; | |
12041 | on_finish = NULL; | |
12042 | return fin; | |
12043 | } | |
12044 | ||
12045 | void finish(int r) override { | |
12046 | if (r < 0) { // we failed the lookup or something; dump ourselves | |
12047 | formatter->open_object_section("results"); | |
12048 | formatter->dump_int("return_code", r); | |
12049 | formatter->close_section(); // results | |
12050 | } | |
12051 | if (on_finish) | |
12052 | on_finish->complete(r); | |
12053 | } | |
12054 | }; | |
12055 | ||
12056 | void MDCache::enqueue_scrub( | |
12057 | const string& path, | |
12058 | const std::string &tag, | |
12059 | bool force, bool recursive, bool repair, | |
12060 | Formatter *f, Context *fin) | |
12061 | { | |
12062 | dout(10) << __func__ << path << dendl; | |
12063 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_ENQUEUE_SCRUB); | |
12064 | filepath fp(path.c_str()); | |
12065 | mdr->set_filepath(fp); | |
12066 | ||
12067 | C_MDS_EnqueueScrub *cs = new C_MDS_EnqueueScrub(f, fin); | |
12068 | cs->header = std::make_shared<ScrubHeader>( | |
12069 | tag, force, recursive, repair, f); | |
12070 | ||
12071 | mdr->internal_op_finish = cs; | |
12072 | enqueue_scrub_work(mdr); | |
12073 | } | |
12074 | ||
12075 | void MDCache::enqueue_scrub_work(MDRequestRef& mdr) | |
12076 | { | |
12077 | set<SimpleLock*> rdlocks, wrlocks, xlocks; | |
12078 | CInode *in = mds->server->rdlock_path_pin_ref(mdr, 0, rdlocks, true); | |
12079 | if (NULL == in) | |
12080 | return; | |
12081 | ||
12082 | // TODO: Remove this restriction | |
12083 | assert(in->is_auth()); | |
12084 | ||
12085 | bool locked = mds->locker->acquire_locks(mdr, rdlocks, wrlocks, xlocks); | |
12086 | if (!locked) | |
12087 | return; | |
12088 | ||
12089 | C_MDS_EnqueueScrub *cs = static_cast<C_MDS_EnqueueScrub*>(mdr->internal_op_finish); | |
12090 | ScrubHeaderRef &header = cs->header; | |
12091 | ||
12092 | // Cannot scrub same dentry twice at same time | |
12093 | if (in->scrub_infop && in->scrub_infop->scrub_in_progress) { | |
12094 | mds->server->respond_to_request(mdr, -EBUSY); | |
12095 | return; | |
12096 | } else { | |
12097 | in->scrub_info(); | |
12098 | } | |
12099 | ||
12100 | header->set_origin(in); | |
12101 | ||
12102 | // only set completion context for non-recursive scrub, because we don't | |
12103 | // want to block asok caller on long running scrub | |
12104 | if (!header->get_recursive()) { | |
12105 | Context *fin = cs->take_finisher(); | |
12106 | mds->scrubstack->enqueue_inode_top(in, header, | |
12107 | new MDSInternalContextWrapper(mds, fin)); | |
12108 | } else | |
12109 | mds->scrubstack->enqueue_inode_bottom(in, header, NULL); | |
12110 | ||
12111 | mds->server->respond_to_request(mdr, 0); | |
12112 | return; | |
12113 | } | |
12114 | ||
12115 | struct C_MDC_RepairDirfragStats : public MDCacheLogContext { | |
12116 | MDRequestRef mdr; | |
12117 | C_MDC_RepairDirfragStats(MDCache *c, MDRequestRef& m) : | |
12118 | MDCacheLogContext(c), mdr(m) {} | |
12119 | void finish(int r) override { | |
12120 | mdr->apply(); | |
12121 | get_mds()->server->respond_to_request(mdr, r); | |
12122 | } | |
12123 | }; | |
12124 | ||
12125 | void MDCache::repair_dirfrag_stats(CDir *dir) | |
12126 | { | |
12127 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_REPAIR_FRAGSTATS); | |
12128 | mdr->pin(dir); | |
12129 | mdr->internal_op_private = dir; | |
12130 | mdr->internal_op_finish = new C_MDSInternalNoop; | |
12131 | repair_dirfrag_stats_work(mdr); | |
12132 | } | |
12133 | ||
12134 | void MDCache::repair_dirfrag_stats_work(MDRequestRef& mdr) | |
12135 | { | |
12136 | CDir *dir = static_cast<CDir*>(mdr->internal_op_private); | |
12137 | dout(10) << __func__ << " " << *dir << dendl; | |
12138 | ||
12139 | if (!dir->is_auth()) { | |
12140 | mds->server->respond_to_request(mdr, -ESTALE); | |
12141 | return; | |
12142 | } | |
12143 | ||
12144 | if (!mdr->is_auth_pinned(dir) && !dir->can_auth_pin()) { | |
224ce89b WB |
12145 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryRequest(this, mdr)); |
12146 | ||
7c673cae FG |
12147 | mds->locker->drop_locks(mdr.get()); |
12148 | mdr->drop_local_auth_pins(); | |
224ce89b WB |
12149 | if (!mdr->remote_auth_pins.empty()) |
12150 | mds->locker->notify_freeze_waiter(dir); | |
7c673cae FG |
12151 | return; |
12152 | } | |
12153 | ||
12154 | mdr->auth_pin(dir); | |
12155 | ||
12156 | set<SimpleLock*> rdlocks, wrlocks, xlocks; | |
12157 | CInode *diri = dir->inode; | |
12158 | rdlocks.insert(&diri->dirfragtreelock); | |
12159 | wrlocks.insert(&diri->nestlock); | |
12160 | wrlocks.insert(&diri->filelock); | |
12161 | if (!mds->locker->acquire_locks(mdr, rdlocks, wrlocks, xlocks)) | |
12162 | return; | |
12163 | ||
12164 | if (!dir->is_complete()) { | |
12165 | dir->fetch(new C_MDS_RetryRequest(this, mdr)); | |
12166 | return; | |
12167 | } | |
12168 | ||
12169 | frag_info_t frag_info; | |
12170 | nest_info_t nest_info; | |
12171 | for (CDir::map_t::iterator it = dir->begin(); it != dir->end(); ++it) { | |
12172 | CDentry *dn = it->second; | |
12173 | if (dn->last != CEPH_NOSNAP) | |
12174 | continue; | |
12175 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
12176 | if (dnl->is_primary()) { | |
12177 | CInode *in = dnl->get_inode(); | |
12178 | nest_info.add(in->get_projected_inode()->accounted_rstat); | |
12179 | if (in->is_dir()) | |
12180 | frag_info.nsubdirs++; | |
12181 | else | |
12182 | frag_info.nfiles++; | |
12183 | } else if (dnl->is_remote()) | |
12184 | frag_info.nfiles++; | |
12185 | } | |
12186 | ||
12187 | fnode_t *pf = dir->get_projected_fnode(); | |
12188 | bool good_fragstat = frag_info.same_sums(pf->fragstat); | |
12189 | bool good_rstat = nest_info.same_sums(pf->rstat); | |
12190 | if (good_fragstat && good_rstat) { | |
12191 | dout(10) << __func__ << " no corruption found" << dendl; | |
12192 | mds->server->respond_to_request(mdr, 0); | |
12193 | return; | |
12194 | } | |
12195 | ||
12196 | pf = dir->project_fnode(); | |
12197 | pf->version = dir->pre_dirty(); | |
12198 | mdr->add_projected_fnode(dir); | |
12199 | ||
12200 | mdr->ls = mds->mdlog->get_current_segment(); | |
12201 | EUpdate *le = new EUpdate(mds->mdlog, "repair_dirfrag"); | |
12202 | mds->mdlog->start_entry(le); | |
12203 | ||
12204 | if (!good_fragstat) { | |
12205 | if (pf->fragstat.mtime > frag_info.mtime) | |
12206 | frag_info.mtime = pf->fragstat.mtime; | |
12207 | if (pf->fragstat.change_attr > frag_info.change_attr) | |
12208 | frag_info.change_attr = pf->fragstat.change_attr; | |
12209 | pf->fragstat = frag_info; | |
12210 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
12211 | mdr->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
12212 | mdr->add_updated_lock(&diri->filelock); | |
12213 | } | |
12214 | ||
12215 | if (!good_rstat) { | |
12216 | if (pf->rstat.rctime > nest_info.rctime) | |
12217 | nest_info.rctime = pf->rstat.rctime; | |
12218 | pf->rstat = nest_info; | |
12219 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
12220 | mdr->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
12221 | mdr->add_updated_lock(&diri->nestlock); | |
12222 | } | |
12223 | ||
12224 | le->metablob.add_dir_context(dir); | |
12225 | le->metablob.add_dir(dir, true); | |
12226 | ||
12227 | mds->mdlog->submit_entry(le, new C_MDC_RepairDirfragStats(this, mdr)); | |
12228 | } | |
12229 | ||
12230 | void MDCache::repair_inode_stats(CInode *diri) | |
12231 | { | |
12232 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_REPAIR_INODESTATS); | |
12233 | mdr->pin(diri); | |
12234 | mdr->internal_op_private = diri; | |
12235 | mdr->internal_op_finish = new C_MDSInternalNoop; | |
12236 | repair_inode_stats_work(mdr); | |
12237 | } | |
12238 | ||
12239 | void MDCache::repair_inode_stats_work(MDRequestRef& mdr) | |
12240 | { | |
12241 | CInode *diri = static_cast<CInode*>(mdr->internal_op_private); | |
12242 | dout(10) << __func__ << " " << *diri << dendl; | |
12243 | ||
12244 | if (!diri->is_auth()) { | |
12245 | mds->server->respond_to_request(mdr, -ESTALE); | |
12246 | return; | |
12247 | } | |
12248 | if (!diri->is_dir()) { | |
12249 | mds->server->respond_to_request(mdr, -ENOTDIR); | |
12250 | return; | |
12251 | } | |
12252 | ||
12253 | set<SimpleLock*> rdlocks, wrlocks, xlocks; | |
12254 | std::list<frag_t> frags; | |
12255 | ||
12256 | if (mdr->ls) // already marked filelock/nestlock dirty ? | |
12257 | goto do_rdlocks; | |
12258 | ||
12259 | rdlocks.insert(&diri->dirfragtreelock); | |
12260 | wrlocks.insert(&diri->nestlock); | |
12261 | wrlocks.insert(&diri->filelock); | |
12262 | if (!mds->locker->acquire_locks(mdr, rdlocks, wrlocks, xlocks)) | |
12263 | return; | |
12264 | ||
12265 | // Fetch all dirfrags and mark filelock/nestlock dirty. This will tirgger | |
12266 | // the scatter-gather process, which will fix any fragstat/rstat errors. | |
12267 | diri->dirfragtree.get_leaves(frags); | |
12268 | for (list<frag_t>::iterator p = frags.begin(); p != frags.end(); ++p) { | |
12269 | CDir *dir = diri->get_dirfrag(*p); | |
12270 | if (!dir) { | |
12271 | assert(mdr->is_auth_pinned(diri)); | |
12272 | dir = diri->get_or_open_dirfrag(this, *p); | |
12273 | } | |
12274 | if (dir->get_version() == 0) { | |
12275 | assert(dir->is_auth()); | |
12276 | dir->fetch(new C_MDS_RetryRequest(this, mdr)); | |
12277 | return; | |
12278 | } | |
12279 | } | |
12280 | ||
12281 | diri->state_set(CInode::STATE_REPAIRSTATS); | |
12282 | mdr->ls = mds->mdlog->get_current_segment(); | |
12283 | mds->locker->mark_updated_scatterlock(&diri->filelock); | |
12284 | mdr->ls->dirty_dirfrag_dir.push_back(&diri->item_dirty_dirfrag_dir); | |
12285 | mds->locker->mark_updated_scatterlock(&diri->nestlock); | |
12286 | mdr->ls->dirty_dirfrag_nest.push_back(&diri->item_dirty_dirfrag_nest); | |
12287 | ||
12288 | mds->locker->drop_locks(mdr.get()); | |
12289 | ||
12290 | do_rdlocks: | |
12291 | // force the scatter-gather process | |
12292 | rdlocks.insert(&diri->dirfragtreelock); | |
12293 | rdlocks.insert(&diri->nestlock); | |
12294 | rdlocks.insert(&diri->filelock); | |
12295 | wrlocks.clear(); | |
12296 | if (!mds->locker->acquire_locks(mdr, rdlocks, wrlocks, xlocks)) | |
12297 | return; | |
12298 | ||
12299 | diri->state_clear(CInode::STATE_REPAIRSTATS); | |
12300 | ||
12301 | frag_info_t dir_info; | |
12302 | nest_info_t nest_info; | |
12303 | nest_info.rsubdirs++; // it gets one to account for self | |
12304 | ||
12305 | diri->dirfragtree.get_leaves(frags); | |
12306 | for (list<frag_t>::iterator p = frags.begin(); p != frags.end(); ++p) { | |
12307 | CDir *dir = diri->get_dirfrag(*p); | |
12308 | assert(dir); | |
12309 | assert(dir->get_version() > 0); | |
12310 | dir_info.add(dir->fnode.accounted_fragstat); | |
12311 | nest_info.add(dir->fnode.accounted_rstat); | |
12312 | } | |
12313 | ||
12314 | if (!dir_info.same_sums(diri->inode.dirstat) || | |
12315 | !nest_info.same_sums(diri->inode.rstat)) { | |
12316 | dout(10) << __func__ << " failed to fix fragstat/rstat on " | |
12317 | << *diri << dendl; | |
12318 | } | |
12319 | ||
12320 | mds->server->respond_to_request(mdr, 0); | |
12321 | } | |
12322 | ||
12323 | void MDCache::flush_dentry(const string& path, Context *fin) | |
12324 | { | |
12325 | if (is_readonly()) { | |
12326 | dout(10) << __func__ << ": read-only FS" << dendl; | |
12327 | fin->complete(-EROFS); | |
12328 | return; | |
12329 | } | |
12330 | dout(10) << "flush_dentry " << path << dendl; | |
12331 | MDRequestRef mdr = request_start_internal(CEPH_MDS_OP_FLUSH); | |
12332 | filepath fp(path.c_str()); | |
12333 | mdr->set_filepath(fp); | |
12334 | mdr->internal_op_finish = fin; | |
12335 | flush_dentry_work(mdr); | |
12336 | } | |
12337 | ||
12338 | class C_FinishIOMDR : public MDSInternalContextBase { | |
12339 | protected: | |
12340 | MDSRank *mds; | |
12341 | MDRequestRef mdr; | |
12342 | MDSRank *get_mds() override { return mds; } | |
12343 | public: | |
12344 | C_FinishIOMDR(MDSRank *mds_, MDRequestRef& mdr_) : mds(mds_), mdr(mdr_) {} | |
12345 | void finish(int r) override { mds->server->respond_to_request(mdr, r); } | |
12346 | }; | |
12347 | ||
12348 | void MDCache::flush_dentry_work(MDRequestRef& mdr) | |
12349 | { | |
12350 | set<SimpleLock*> rdlocks, wrlocks, xlocks; | |
12351 | CInode *in = mds->server->rdlock_path_pin_ref(mdr, 0, rdlocks, true); | |
12352 | if (NULL == in) | |
12353 | return; | |
12354 | ||
12355 | // TODO: Is this necessary? Fix it if so | |
12356 | assert(in->is_auth()); | |
12357 | bool locked = mds->locker->acquire_locks(mdr, rdlocks, wrlocks, xlocks); | |
12358 | if (!locked) | |
12359 | return; | |
12360 | in->flush(new C_FinishIOMDR(mds, mdr)); | |
12361 | } | |
12362 | ||
12363 | ||
12364 | /** | |
12365 | * Initialize performance counters with global perfcounter | |
12366 | * collection. | |
12367 | */ | |
12368 | void MDCache::register_perfcounters() | |
12369 | { | |
12370 | PerfCountersBuilder pcb(g_ceph_context, | |
12371 | "mds_cache", l_mdc_first, l_mdc_last); | |
12372 | ||
12373 | /* Stray/purge statistics */ | |
12374 | pcb.add_u64(l_mdc_num_strays, "num_strays", | |
c07f9fc5 | 12375 | "Stray dentries", "stry", PerfCountersBuilder::PRIO_INTERESTING); |
7c673cae FG |
12376 | pcb.add_u64(l_mdc_num_strays_delayed, "num_strays_delayed", "Stray dentries delayed"); |
12377 | pcb.add_u64(l_mdc_num_strays_enqueuing, "num_strays_enqueuing", "Stray dentries enqueuing for purge"); | |
12378 | ||
12379 | pcb.add_u64_counter(l_mdc_strays_created, "strays_created", "Stray dentries created"); | |
12380 | pcb.add_u64_counter(l_mdc_strays_enqueued, "strays_enqueued", | |
12381 | "Stray dentries enqueued for purge"); | |
12382 | pcb.add_u64_counter(l_mdc_strays_reintegrated, "strays_reintegrated", "Stray dentries reintegrated"); | |
12383 | pcb.add_u64_counter(l_mdc_strays_migrated, "strays_migrated", "Stray dentries migrated"); | |
12384 | ||
12385 | ||
12386 | /* Recovery queue statistics */ | |
12387 | pcb.add_u64(l_mdc_num_recovering_processing, "num_recovering_processing", "Files currently being recovered"); | |
12388 | pcb.add_u64(l_mdc_num_recovering_enqueued, "num_recovering_enqueued", | |
c07f9fc5 | 12389 | "Files waiting for recovery", "recy", PerfCountersBuilder::PRIO_INTERESTING); |
7c673cae FG |
12390 | pcb.add_u64(l_mdc_num_recovering_prioritized, "num_recovering_prioritized", "Files waiting for recovery with elevated priority"); |
12391 | pcb.add_u64_counter(l_mdc_recovery_started, "recovery_started", "File recoveries started"); | |
12392 | pcb.add_u64_counter(l_mdc_recovery_completed, "recovery_completed", | |
c07f9fc5 | 12393 | "File recoveries completed", "recd", PerfCountersBuilder::PRIO_INTERESTING); |
7c673cae | 12394 | |
d2e6a577 FG |
12395 | pcb.add_u64_counter(l_mdss_ireq_enqueue_scrub, "ireq_enqueue_scrub", |
12396 | "Internal Request type enqueue scrub"); | |
12397 | pcb.add_u64_counter(l_mdss_ireq_exportdir, "ireq_exportdir", | |
12398 | "Internal Request type export dir"); | |
12399 | pcb.add_u64_counter(l_mdss_ireq_flush, "ireq_flush", | |
12400 | "Internal Request type flush"); | |
12401 | pcb.add_u64_counter(l_mdss_ireq_fragmentdir, "ireq_fragmentdir", | |
12402 | "Internal Request type fragmentdir"); | |
12403 | pcb.add_u64_counter(l_mdss_ireq_fragstats, "ireq_fragstats", | |
12404 | "Internal Request type frag stats"); | |
12405 | pcb.add_u64_counter(l_mdss_ireq_inodestats, "ireq_inodestats", | |
12406 | "Internal Request type inode stats"); | |
12407 | ||
7c673cae FG |
12408 | logger.reset(pcb.create_perf_counters()); |
12409 | g_ceph_context->get_perfcounters_collection()->add(logger.get()); | |
12410 | recovery_queue.set_logger(logger.get()); | |
12411 | stray_manager.set_logger(logger.get()); | |
12412 | } | |
12413 | ||
12414 | void MDCache::activate_stray_manager() | |
12415 | { | |
12416 | if (open) { | |
12417 | stray_manager.activate(); | |
12418 | } else { | |
12419 | wait_for_open( | |
12420 | new MDSInternalContextWrapper(mds, | |
12421 | new FunctionContext([this](int r){ | |
12422 | stray_manager.activate(); | |
12423 | }) | |
12424 | ) | |
12425 | ); | |
12426 | } | |
12427 | } | |
12428 | ||
12429 | /** | |
12430 | * Call this when putting references to an inode/dentry or | |
12431 | * when attempting to trim it. | |
12432 | * | |
12433 | * If this inode is no longer linked by anyone, and this MDS | |
12434 | * rank holds the primary dentry, and that dentry is in a stray | |
12435 | * directory, then give up the dentry to the StrayManager, never | |
12436 | * to be seen again by MDCache. | |
12437 | * | |
12438 | * @param delay if true, then purgeable inodes are stashed til | |
12439 | * the next trim(), rather than being purged right | |
12440 | * away. | |
12441 | */ | |
12442 | void MDCache::maybe_eval_stray(CInode *in, bool delay) { | |
224ce89b WB |
12443 | if (in->inode.nlink > 0 || in->is_base() || is_readonly() || |
12444 | mds->get_state() <= MDSMap::STATE_REJOIN) | |
7c673cae | 12445 | return; |
224ce89b | 12446 | |
7c673cae FG |
12447 | CDentry *dn = in->get_projected_parent_dn(); |
12448 | ||
12449 | if (dn->state_test(CDentry::STATE_PURGING)) { | |
12450 | /* We have already entered the purging process, no need | |
12451 | * to re-evaluate me ! */ | |
12452 | return; | |
12453 | } | |
12454 | ||
12455 | if (dn->get_projected_linkage()->is_primary() && | |
12456 | dn->get_dir()->get_inode()->is_stray()) { | |
12457 | stray_manager.eval_stray(dn, delay); | |
12458 | } | |
12459 | } | |
12460 | ||
31f18b77 FG |
12461 | void MDCache::clear_dirty_bits_for_stray(CInode* diri) { |
12462 | dout(10) << __func__ << " " << *diri << dendl; | |
12463 | assert(diri->get_projected_parent_dir()->inode->is_stray()); | |
12464 | list<CDir*> ls; | |
12465 | diri->get_dirfrags(ls); | |
12466 | for (auto p : ls) { | |
12467 | if (p->is_auth() && !(p->is_frozen() || p->is_freezing())) | |
12468 | p->try_remove_dentries_for_stray(); | |
12469 | } | |
12470 | if (!diri->snaprealm) { | |
12471 | if (diri->is_auth()) | |
12472 | diri->clear_dirty_rstat(); | |
12473 | diri->clear_scatter_dirty(); | |
12474 | } | |
12475 | } | |
12476 |