]>
Commit | Line | Data |
---|---|---|
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) 2015 Red Hat | |
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 | ||
16 | #include "common/perf_counters.h" | |
17 | ||
18 | #include "mds/MDSRank.h" | |
19 | #include "mds/MDCache.h" | |
20 | #include "mds/MDLog.h" | |
21 | #include "mds/CDir.h" | |
22 | #include "mds/CDentry.h" | |
23 | #include "events/EUpdate.h" | |
24 | #include "messages/MClientRequest.h" | |
25 | ||
26 | #include "StrayManager.h" | |
27 | ||
28 | #define dout_context g_ceph_context | |
29 | #define dout_subsys ceph_subsys_mds | |
30 | #undef dout_prefix | |
31 | #define dout_prefix _prefix(_dout, mds) | |
32 | static ostream& _prefix(std::ostream *_dout, MDSRank *mds) { | |
33 | return *_dout << "mds." << mds->get_nodeid() << ".cache.strays "; | |
34 | } | |
35 | ||
36 | class StrayManagerIOContext : public virtual MDSIOContextBase { | |
37 | protected: | |
38 | StrayManager *sm; | |
39 | MDSRank *get_mds() override | |
40 | { | |
41 | return sm->mds; | |
42 | } | |
43 | public: | |
44 | explicit StrayManagerIOContext(StrayManager *sm_) : sm(sm_) {} | |
45 | }; | |
46 | ||
47 | class StrayManagerLogContext : public virtual MDSLogContextBase { | |
48 | protected: | |
49 | StrayManager *sm; | |
50 | MDSRank *get_mds() override | |
51 | { | |
52 | return sm->mds; | |
53 | } | |
54 | public: | |
55 | explicit StrayManagerLogContext(StrayManager *sm_) : sm(sm_) {} | |
56 | }; | |
57 | ||
58 | class StrayManagerContext : public virtual MDSInternalContextBase { | |
59 | protected: | |
60 | StrayManager *sm; | |
61 | MDSRank *get_mds() override | |
62 | { | |
63 | return sm->mds; | |
64 | } | |
65 | public: | |
66 | explicit StrayManagerContext(StrayManager *sm_) : sm(sm_) {} | |
67 | }; | |
68 | ||
69 | ||
70 | /** | |
71 | * Context wrapper for _purge_stray_purged completion | |
72 | */ | |
73 | class C_IO_PurgeStrayPurged : public StrayManagerIOContext { | |
74 | CDentry *dn; | |
75 | bool only_head; | |
76 | // How many ops_in_flight were allocated to this purge? | |
77 | uint32_t ops_allowance; | |
78 | public: | |
79 | C_IO_PurgeStrayPurged(StrayManager *sm_, CDentry *d, bool oh) : | |
80 | StrayManagerIOContext(sm_), dn(d), only_head(oh) { } | |
81 | void finish(int r) override { | |
82 | assert(r == 0 || r == -ENOENT); | |
83 | sm->_purge_stray_purged(dn, ops_allowance, only_head); | |
84 | } | |
85 | }; | |
86 | ||
87 | ||
88 | void StrayManager::purge(CDentry *dn) | |
89 | { | |
90 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
91 | CInode *in = dnl->get_inode(); | |
92 | dout(10) << __func__ << " " << *dn << " " << *in << dendl; | |
93 | assert(!dn->is_replicated()); | |
94 | ||
95 | // CHEAT. there's no real need to journal our intent to purge, since | |
96 | // that is implicit in the dentry's presence and non-use in the stray | |
97 | // dir. on recovery, we'll need to re-eval all strays anyway. | |
98 | ||
99 | SnapContext nullsnapc; | |
100 | ||
101 | PurgeItem item; | |
102 | item.ino = in->inode.ino; | |
103 | if (in->is_dir()) { | |
104 | item.action = PurgeItem::PURGE_DIR; | |
105 | item.fragtree = in->dirfragtree; | |
106 | } else { | |
107 | item.action = PurgeItem::PURGE_FILE; | |
108 | ||
109 | const SnapContext *snapc; | |
110 | SnapRealm *realm = in->find_snaprealm(); | |
111 | if (realm) { | |
112 | dout(10) << " realm " << *realm << dendl; | |
113 | snapc = &realm->get_snap_context(); | |
114 | } else { | |
115 | dout(10) << " NO realm, using null context" << dendl; | |
116 | snapc = &nullsnapc; | |
117 | assert(in->last == CEPH_NOSNAP); | |
118 | } | |
119 | ||
120 | uint64_t to = 0; | |
121 | if (in->is_file()) { | |
122 | to = in->inode.get_max_size(); | |
123 | to = MAX(in->inode.size, to); | |
124 | // when truncating a file, the filer does not delete stripe objects that are | |
125 | // truncated to zero. so we need to purge stripe objects up to the max size | |
126 | // the file has ever been. | |
127 | to = MAX(in->inode.max_size_ever, to); | |
128 | } | |
129 | ||
130 | inode_t *pi = in->get_projected_inode(); | |
131 | ||
132 | item.size = to; | |
133 | item.layout = pi->layout; | |
134 | item.old_pools = pi->old_pools; | |
135 | item.snapc = *snapc; | |
136 | } | |
137 | ||
138 | purge_queue.push(item, new C_IO_PurgeStrayPurged( | |
139 | this, dn, false)); | |
140 | } | |
141 | ||
142 | class C_PurgeStrayLogged : public StrayManagerLogContext { | |
143 | CDentry *dn; | |
144 | version_t pdv; | |
145 | LogSegment *ls; | |
146 | public: | |
147 | C_PurgeStrayLogged(StrayManager *sm_, CDentry *d, version_t v, LogSegment *s) : | |
148 | StrayManagerLogContext(sm_), dn(d), pdv(v), ls(s) { } | |
149 | void finish(int r) override { | |
150 | sm->_purge_stray_logged(dn, pdv, ls); | |
151 | } | |
152 | }; | |
153 | ||
154 | class C_TruncateStrayLogged : public StrayManagerLogContext { | |
155 | CDentry *dn; | |
156 | LogSegment *ls; | |
157 | public: | |
158 | C_TruncateStrayLogged(StrayManager *sm, CDentry *d, LogSegment *s) : | |
159 | StrayManagerLogContext(sm), dn(d), ls(s) { } | |
160 | void finish(int r) override { | |
161 | sm->_truncate_stray_logged(dn, ls); | |
162 | } | |
163 | }; | |
164 | ||
165 | void StrayManager::_purge_stray_purged( | |
166 | CDentry *dn, uint32_t ops_allowance, bool only_head) | |
167 | { | |
168 | CInode *in = dn->get_projected_linkage()->get_inode(); | |
169 | dout(10) << "_purge_stray_purged " << *dn << " " << *in << dendl; | |
170 | ||
171 | logger->inc(l_mdc_strays_enqueued); | |
172 | num_strays_enqueuing--; | |
173 | logger->set(l_mdc_num_strays_enqueuing, num_strays_enqueuing); | |
174 | ||
175 | if (only_head) { | |
176 | /* This was a ::truncate */ | |
177 | EUpdate *le = new EUpdate(mds->mdlog, "purge_stray truncate"); | |
178 | mds->mdlog->start_entry(le); | |
179 | ||
180 | inode_t *pi = in->project_inode(); | |
181 | pi->size = 0; | |
182 | pi->max_size_ever = 0; | |
183 | pi->client_ranges.clear(); | |
184 | pi->truncate_size = 0; | |
185 | pi->truncate_from = 0; | |
186 | pi->version = in->pre_dirty(); | |
187 | ||
188 | le->metablob.add_dir_context(dn->dir); | |
189 | le->metablob.add_primary_dentry(dn, in, true); | |
190 | ||
191 | mds->mdlog->submit_entry(le, | |
192 | new C_TruncateStrayLogged( | |
193 | this, dn, mds->mdlog->get_current_segment())); | |
194 | } else { | |
195 | if (in->get_num_ref() != (int)in->is_dirty() || | |
196 | dn->get_num_ref() != (int)dn->is_dirty() + !!in->get_num_ref() + 1/*PIN_PURGING*/) { | |
197 | // Nobody should be taking new references to an inode when it | |
198 | // is being purged (aside from it were | |
199 | ||
200 | derr << "Rogue reference after purge to " << *dn << dendl; | |
201 | assert(0 == "rogue reference to purging inode"); | |
202 | } | |
203 | ||
204 | // kill dentry. | |
205 | version_t pdv = dn->pre_dirty(); | |
206 | dn->push_projected_linkage(); // NULL | |
207 | ||
208 | EUpdate *le = new EUpdate(mds->mdlog, "purge_stray"); | |
209 | mds->mdlog->start_entry(le); | |
210 | ||
211 | // update dirfrag fragstat, rstat | |
212 | CDir *dir = dn->get_dir(); | |
213 | fnode_t *pf = dir->project_fnode(); | |
214 | pf->version = dir->pre_dirty(); | |
215 | if (in->is_dir()) | |
216 | pf->fragstat.nsubdirs--; | |
217 | else | |
218 | pf->fragstat.nfiles--; | |
219 | pf->rstat.sub(in->inode.accounted_rstat); | |
220 | ||
221 | le->metablob.add_dir_context(dn->dir); | |
222 | EMetaBlob::dirlump& dl = le->metablob.add_dir(dn->dir, true); | |
223 | le->metablob.add_null_dentry(dl, dn, true); | |
224 | le->metablob.add_destroyed_inode(in->ino()); | |
225 | ||
226 | mds->mdlog->submit_entry(le, new C_PurgeStrayLogged(this, dn, pdv, | |
227 | mds->mdlog->get_current_segment())); | |
228 | ||
229 | logger->set(l_mdc_num_strays, num_strays); | |
230 | } | |
231 | } | |
232 | ||
233 | void StrayManager::_purge_stray_logged(CDentry *dn, version_t pdv, LogSegment *ls) | |
234 | { | |
235 | CInode *in = dn->get_linkage()->get_inode(); | |
236 | dout(10) << "_purge_stray_logged " << *dn << " " << *in << dendl; | |
237 | ||
238 | assert(!in->state_test(CInode::STATE_RECOVERING)); | |
239 | ||
240 | bool new_dn = dn->is_new(); | |
241 | ||
242 | // unlink | |
243 | assert(dn->get_projected_linkage()->is_null()); | |
244 | dn->dir->unlink_inode(dn, !new_dn); | |
245 | dn->pop_projected_linkage(); | |
246 | dn->mark_dirty(pdv, ls); | |
247 | ||
248 | dn->dir->pop_and_dirty_projected_fnode(ls); | |
249 | ||
250 | in->state_clear(CInode::STATE_ORPHAN); | |
251 | dn->state_clear(CDentry::STATE_PURGING | CDentry::STATE_PURGINGPINNED); | |
252 | dn->put(CDentry::PIN_PURGING); | |
253 | ||
254 | // drop dentry? | |
255 | if (new_dn) { | |
256 | dout(20) << " dn is new, removing" << dendl; | |
257 | dn->mark_clean(); | |
258 | dn->dir->remove_dentry(dn); | |
259 | } | |
260 | ||
261 | // drop inode | |
262 | if (in->is_dirty()) | |
263 | in->mark_clean(); | |
264 | in->mdcache->remove_inode(in); | |
265 | } | |
266 | ||
267 | void StrayManager::enqueue(CDentry *dn, bool trunc) | |
268 | { | |
269 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
270 | assert(dnl); | |
271 | CInode *in = dnl->get_inode(); | |
272 | assert(in); | |
273 | ||
274 | /* We consider a stray to be purging as soon as it is enqueued, to avoid | |
275 | * enqueing it twice */ | |
276 | dn->state_set(CDentry::STATE_PURGING); | |
277 | in->state_set(CInode::STATE_PURGING); | |
278 | ||
279 | /* We must clear this as soon as enqueuing it, to prevent the journal | |
280 | * expiry code from seeing a dirty parent and trying to write a backtrace */ | |
281 | if (!trunc) { | |
282 | if (in->is_dirty_parent()) { | |
283 | in->clear_dirty_parent(); | |
284 | } | |
285 | } | |
286 | ||
287 | dout(20) << __func__ << ": purging dn: " << *dn << dendl; | |
288 | ||
289 | if (!dn->state_test(CDentry::STATE_PURGINGPINNED)) { | |
290 | dn->get(CDentry::PIN_PURGING); | |
291 | dn->state_set(CDentry::STATE_PURGINGPINNED); | |
292 | } | |
293 | ||
294 | ++num_strays_enqueuing; | |
295 | logger->set(l_mdc_num_strays_enqueuing, num_strays_enqueuing); | |
296 | ||
297 | // Resources are available, acquire them and execute the purge | |
298 | _enqueue(dn, trunc); | |
299 | ||
300 | dout(10) << __func__ << ": purging this dentry immediately: " | |
301 | << *dn << dendl; | |
302 | } | |
303 | ||
304 | class C_OpenSnapParents : public StrayManagerContext { | |
305 | CDentry *dn; | |
306 | bool trunc; | |
307 | public: | |
308 | C_OpenSnapParents(StrayManager *sm_, CDentry *dn_, bool t) : | |
309 | StrayManagerContext(sm_), dn(dn_), trunc(t) { } | |
310 | void finish(int r) override { | |
311 | sm->_enqueue(dn, trunc); | |
312 | } | |
313 | }; | |
314 | ||
315 | void StrayManager::_enqueue(CDentry *dn, bool trunc) | |
316 | { | |
317 | CInode *in = dn->get_linkage()->get_inode(); | |
318 | if (in->snaprealm && | |
319 | !in->snaprealm->have_past_parents_open() && | |
320 | !in->snaprealm->open_parents(new C_OpenSnapParents(this, dn, trunc))) { | |
321 | // this can happen if the dentry had been trimmed from cache. | |
322 | return; | |
323 | } | |
324 | ||
325 | if (!started) { | |
326 | // If the MDS is not yet active, defer executing this purge | |
327 | // in order to avoid the mdlog writes we do on purge completion. | |
328 | mds->wait_for_active( | |
329 | new MDSInternalContextWrapper(mds, | |
330 | new FunctionContext([this, dn, trunc](int r){ | |
331 | // It is safe to hold on to this CDentry* pointer | |
332 | // because the dentry is pinned with PIN_PURGING | |
333 | _enqueue(dn, trunc); | |
334 | }) | |
335 | ) | |
336 | ); | |
337 | ||
338 | return; | |
339 | } | |
340 | ||
341 | if (trunc) { | |
342 | truncate(dn); | |
343 | } else { | |
344 | purge(dn); | |
345 | } | |
346 | } | |
347 | ||
348 | ||
349 | void StrayManager::advance_delayed() | |
350 | { | |
351 | for (elist<CDentry*>::iterator p = delayed_eval_stray.begin(); !p.end(); ) { | |
352 | CDentry *dn = *p; | |
353 | ++p; | |
354 | dn->item_stray.remove_myself(); | |
355 | num_strays_delayed--; | |
356 | ||
357 | if (dn->get_projected_linkage()->is_null()) { | |
358 | /* A special case: a stray dentry can go null if its inode is being | |
359 | * re-linked into another MDS's stray dir during a shutdown migration. */ | |
360 | dout(4) << __func__ << ": delayed dentry is now null: " << *dn << dendl; | |
361 | continue; | |
362 | } | |
363 | ||
364 | const bool purging = eval_stray(dn); | |
365 | if (!purging) { | |
366 | derr << "Dentry " << *dn << " was purgeable but no longer is!" << dendl; | |
367 | /* | |
368 | * This can happen if a stray is purgeable, but has gained an extra | |
369 | * reference by virtue of having its backtrace updated. | |
370 | * FIXME perhaps we could simplify this further by | |
371 | * avoiding writing the backtrace of purge-ready strays, so | |
372 | * that this code could be more rigid? | |
373 | */ | |
374 | } | |
375 | } | |
376 | logger->set(l_mdc_num_strays_delayed, num_strays_delayed); | |
377 | } | |
378 | ||
379 | void StrayManager::set_num_strays(uint64_t num) | |
380 | { | |
381 | assert(!started); | |
382 | num_strays = num; | |
383 | logger->set(l_mdc_num_strays, num_strays); | |
384 | } | |
385 | ||
386 | void StrayManager::notify_stray_created() | |
387 | { | |
388 | num_strays++; | |
389 | logger->set(l_mdc_num_strays, num_strays); | |
390 | logger->inc(l_mdc_strays_created); | |
391 | } | |
392 | ||
393 | void StrayManager::notify_stray_removed() | |
394 | { | |
395 | num_strays--; | |
396 | logger->set(l_mdc_num_strays, num_strays); | |
397 | } | |
398 | ||
399 | struct C_EvalStray : public StrayManagerContext { | |
400 | CDentry *dn; | |
401 | C_EvalStray(StrayManager *sm_, CDentry *d) : StrayManagerContext(sm_), dn(d) {} | |
402 | void finish(int r) override { | |
403 | sm->eval_stray(dn); | |
404 | } | |
405 | }; | |
406 | ||
407 | struct C_MDC_EvalStray : public StrayManagerContext { | |
408 | CDentry *dn; | |
409 | C_MDC_EvalStray(StrayManager *sm_, CDentry *d) : StrayManagerContext(sm_), dn(d) {} | |
410 | void finish(int r) override { | |
411 | sm->eval_stray(dn); | |
412 | } | |
413 | }; | |
414 | ||
415 | bool StrayManager::_eval_stray(CDentry *dn, bool delay) | |
416 | { | |
417 | dout(10) << "eval_stray " << *dn << dendl; | |
418 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
419 | assert(dnl->is_primary()); | |
420 | dout(10) << " inode is " << *dnl->get_inode() << dendl; | |
421 | CInode *in = dnl->get_inode(); | |
422 | assert(in); | |
423 | assert(!in->state_test(CInode::STATE_REJOINUNDEF)); | |
424 | ||
425 | // The only dentries elegible for purging are those | |
426 | // in the stray directories | |
427 | assert(dn->get_dir()->get_inode()->is_stray()); | |
428 | ||
429 | // Inode may not pass through this function if it | |
430 | // was already identified for purging (i.e. cannot | |
431 | // call eval_stray() after purge() | |
432 | assert(!dn->state_test(CDentry::STATE_PURGING)); | |
433 | ||
434 | if (!dn->is_auth()) { | |
435 | return false; | |
436 | } | |
437 | ||
438 | if (dn->item_stray.is_on_list()) { | |
439 | if (delay) | |
440 | return false; | |
441 | ||
442 | dn->item_stray.remove_myself(); | |
443 | num_strays_delayed--; | |
444 | logger->set(l_mdc_num_strays_delayed, num_strays_delayed); | |
445 | } | |
446 | ||
447 | // purge? | |
448 | if (in->inode.nlink == 0) { | |
449 | // past snaprealm parents imply snapped dentry remote links. | |
450 | // only important for directories. normal file data snaps are handled | |
451 | // by the object store. | |
452 | if (in->snaprealm) { | |
453 | if (!in->snaprealm->have_past_parents_open() && | |
454 | !in->snaprealm->open_parents(new C_MDC_EvalStray(this, dn))) { | |
455 | return false; | |
456 | } | |
457 | in->snaprealm->prune_past_parents(); | |
458 | in->purge_stale_snap_data(in->snaprealm->get_snaps()); | |
459 | } | |
460 | if (in->is_dir()) { | |
461 | if (in->snaprealm && in->snaprealm->has_past_parents()) { | |
462 | dout(20) << " directory has past parents " | |
463 | << in->snaprealm->srnode.past_parents << dendl; | |
464 | if (in->state_test(CInode::STATE_MISSINGOBJS)) { | |
465 | mds->clog->error() << "previous attempt at committing dirfrag of ino " | |
466 | << in->ino() << " has failed, missing object"; | |
467 | mds->handle_write_error(-ENOENT); | |
468 | } | |
469 | return false; // not until some snaps are deleted. | |
470 | } | |
471 | ||
472 | in->mdcache->clear_dirty_bits_for_stray(in); | |
473 | ||
474 | if (!in->remote_parents.empty()) { | |
475 | // unlink any stale remote snap dentry. | |
476 | for (compact_set<CDentry*>::iterator p = in->remote_parents.begin(); | |
477 | p != in->remote_parents.end(); ) { | |
478 | CDentry *remote_dn = *p; | |
479 | ++p; | |
480 | assert(remote_dn->last != CEPH_NOSNAP); | |
481 | remote_dn->unlink_remote(remote_dn->get_linkage()); | |
482 | } | |
483 | } | |
484 | } | |
485 | if (dn->is_replicated()) { | |
486 | dout(20) << " replicated" << dendl; | |
487 | return false; | |
488 | } | |
489 | if (dn->is_any_leases() || in->is_any_caps()) { | |
490 | dout(20) << " caps | leases" << dendl; | |
491 | return false; // wait | |
492 | } | |
493 | if (in->state_test(CInode::STATE_NEEDSRECOVER) || | |
494 | in->state_test(CInode::STATE_RECOVERING)) { | |
495 | dout(20) << " pending recovery" << dendl; | |
496 | return false; // don't mess with file size probing | |
497 | } | |
498 | if (in->get_num_ref() > (int)in->is_dirty() + (int)in->is_dirty_parent()) { | |
499 | dout(20) << " too many inode refs" << dendl; | |
500 | return false; | |
501 | } | |
502 | if (dn->get_num_ref() > (int)dn->is_dirty() + !!in->get_num_ref()) { | |
503 | dout(20) << " too many dn refs" << dendl; | |
504 | return false; | |
505 | } | |
506 | if (delay) { | |
507 | if (!dn->item_stray.is_on_list()) { | |
508 | delayed_eval_stray.push_back(&dn->item_stray); | |
509 | num_strays_delayed++; | |
510 | logger->set(l_mdc_num_strays_delayed, num_strays_delayed); | |
511 | } | |
512 | // don't purge multiversion inode with snap data | |
513 | } else if (in->snaprealm && in->snaprealm->has_past_parents() && | |
514 | !in->old_inodes.empty()) { | |
515 | // A file with snapshots: we will truncate the HEAD revision | |
516 | // but leave the metadata intact. | |
517 | assert(!in->is_dir()); | |
518 | dout(20) << " file has past parents " | |
519 | << in->snaprealm->srnode.past_parents << dendl; | |
520 | if (in->is_file() && in->get_projected_inode()->size > 0) { | |
521 | enqueue(dn, true); // truncate head objects | |
522 | } | |
523 | } else { | |
524 | // A straightforward file, ready to be purged. Enqueue it. | |
525 | if (in->is_dir()) { | |
526 | in->close_dirfrags(); | |
527 | } | |
528 | ||
529 | enqueue(dn, false); | |
530 | } | |
531 | ||
532 | return true; | |
533 | } else { | |
534 | /* | |
535 | * Where a stray has some links, they should be remotes, check | |
536 | * if we can do anything with them if we happen to have them in | |
537 | * cache. | |
538 | */ | |
539 | _eval_stray_remote(dn, NULL); | |
540 | return false; | |
541 | } | |
542 | } | |
543 | ||
544 | void StrayManager::activate() | |
545 | { | |
546 | dout(10) << __func__ << dendl; | |
547 | started = true; | |
548 | } | |
549 | ||
550 | bool StrayManager::eval_stray(CDentry *dn, bool delay) | |
551 | { | |
552 | // avoid nested eval_stray | |
553 | if (dn->state_test(CDentry::STATE_EVALUATINGSTRAY)) | |
554 | return false; | |
555 | ||
556 | dn->state_set(CDentry::STATE_EVALUATINGSTRAY); | |
557 | bool ret = _eval_stray(dn, delay); | |
558 | dn->state_clear(CDentry::STATE_EVALUATINGSTRAY); | |
559 | return ret; | |
560 | } | |
561 | ||
562 | void StrayManager::eval_remote(CDentry *remote_dn) | |
563 | { | |
564 | dout(10) << __func__ << " " << *remote_dn << dendl; | |
565 | ||
566 | CDentry::linkage_t *dnl = remote_dn->get_projected_linkage(); | |
567 | assert(dnl->is_remote()); | |
568 | CInode *in = dnl->get_inode(); | |
569 | ||
570 | if (!in) { | |
571 | dout(20) << __func__ << ": no inode, cannot evaluate" << dendl; | |
572 | return; | |
573 | } | |
574 | ||
575 | if (remote_dn->last != CEPH_NOSNAP) { | |
576 | dout(20) << __func__ << ": snap dentry, cannot evaluate" << dendl; | |
577 | return; | |
578 | } | |
579 | ||
580 | // refers to stray? | |
581 | CDentry *primary_dn = in->get_projected_parent_dn(); | |
582 | assert(primary_dn != NULL); | |
583 | if (primary_dn->get_dir()->get_inode()->is_stray()) { | |
584 | _eval_stray_remote(primary_dn, remote_dn); | |
585 | } else { | |
586 | dout(20) << __func__ << ": inode's primary dn not stray" << dendl; | |
587 | } | |
588 | } | |
589 | ||
590 | class C_RetryEvalRemote : public StrayManagerContext { | |
591 | CDentry *dn; | |
592 | public: | |
593 | C_RetryEvalRemote(StrayManager *sm_, CDentry *dn_) : | |
594 | StrayManagerContext(sm_), dn(dn_) { | |
595 | dn->get(CDentry::PIN_PTRWAITER); | |
596 | } | |
597 | void finish(int r) override { | |
598 | if (dn->get_projected_linkage()->is_remote()) | |
599 | sm->eval_remote(dn); | |
600 | dn->put(CDentry::PIN_PTRWAITER); | |
601 | } | |
602 | }; | |
603 | ||
604 | void StrayManager::_eval_stray_remote(CDentry *stray_dn, CDentry *remote_dn) | |
605 | { | |
606 | dout(20) << __func__ << " " << *stray_dn << dendl; | |
607 | assert(stray_dn != NULL); | |
608 | assert(stray_dn->get_dir()->get_inode()->is_stray()); | |
609 | CDentry::linkage_t *stray_dnl = stray_dn->get_projected_linkage(); | |
610 | assert(stray_dnl->is_primary()); | |
611 | CInode *stray_in = stray_dnl->get_inode(); | |
612 | assert(stray_in->inode.nlink >= 1); | |
613 | assert(stray_in->last == CEPH_NOSNAP); | |
614 | ||
615 | /* If no remote_dn hinted, pick one arbitrarily */ | |
616 | if (remote_dn == NULL) { | |
617 | if (!stray_in->remote_parents.empty()) { | |
618 | for (compact_set<CDentry*>::iterator p = stray_in->remote_parents.begin(); | |
619 | p != stray_in->remote_parents.end(); | |
620 | ++p) | |
621 | if ((*p)->last == CEPH_NOSNAP && !(*p)->is_projected()) { | |
622 | if ((*p)->is_auth()) { | |
623 | remote_dn = *p; | |
624 | if (remote_dn->dir->can_auth_pin()) | |
625 | break; | |
626 | } else if (!remote_dn) { | |
627 | remote_dn = *p; | |
628 | } | |
629 | } | |
630 | } | |
631 | if (!remote_dn) { | |
632 | dout(20) << __func__ << ": not reintegrating (no remote parents in cache)" << dendl; | |
633 | return; | |
634 | } | |
635 | } | |
636 | assert(remote_dn->last == CEPH_NOSNAP); | |
637 | // NOTE: we repeat this check in _rename(), since our submission path is racey. | |
638 | if (!remote_dn->is_projected()) { | |
639 | if (remote_dn->is_auth()) { | |
640 | if (remote_dn->dir->can_auth_pin()) { | |
641 | reintegrate_stray(stray_dn, remote_dn); | |
642 | } else { | |
643 | remote_dn->dir->add_waiter(CDir::WAIT_UNFREEZE, new C_RetryEvalRemote(this, remote_dn)); | |
644 | dout(20) << __func__ << ": not reintegrating (can't authpin remote parent)" << dendl; | |
645 | } | |
646 | ||
647 | } else if (!remote_dn->is_auth() && stray_dn->is_auth()) { | |
648 | migrate_stray(stray_dn, remote_dn->authority().first); | |
649 | } else { | |
650 | dout(20) << __func__ << ": not reintegrating" << dendl; | |
651 | } | |
652 | } else { | |
653 | // don't do anything if the remote parent is projected, or we may | |
654 | // break user-visible semantics! | |
655 | dout(20) << __func__ << ": not reintegrating (projected)" << dendl; | |
656 | } | |
657 | } | |
658 | ||
659 | void StrayManager::reintegrate_stray(CDentry *straydn, CDentry *rdn) | |
660 | { | |
661 | dout(10) << __func__ << " " << *straydn << " into " << *rdn << dendl; | |
662 | ||
663 | logger->inc(l_mdc_strays_reintegrated); | |
664 | ||
665 | // rename it to another mds. | |
666 | filepath src; | |
667 | straydn->make_path(src); | |
668 | filepath dst; | |
669 | rdn->make_path(dst); | |
670 | ||
671 | MClientRequest *req = new MClientRequest(CEPH_MDS_OP_RENAME); | |
672 | req->set_filepath(dst); | |
673 | req->set_filepath2(src); | |
674 | req->set_tid(mds->issue_tid()); | |
675 | ||
676 | mds->send_message_mds(req, rdn->authority().first); | |
677 | } | |
678 | ||
679 | void StrayManager::migrate_stray(CDentry *dn, mds_rank_t to) | |
680 | { | |
681 | CInode *in = dn->get_projected_linkage()->get_inode(); | |
682 | assert(in); | |
683 | CInode *diri = dn->dir->get_inode(); | |
684 | assert(diri->is_stray()); | |
685 | dout(10) << "migrate_stray from mds." << MDS_INO_STRAY_OWNER(diri->inode.ino) | |
686 | << " to mds." << to | |
687 | << " " << *dn << " " << *in << dendl; | |
688 | ||
689 | logger->inc(l_mdc_strays_migrated); | |
690 | ||
691 | // rename it to another mds. | |
692 | filepath src; | |
693 | dn->make_path(src); | |
694 | assert(src.depth() == 2); | |
695 | ||
696 | filepath dst(MDS_INO_MDSDIR(to)); | |
697 | dst.push_dentry(src[0]); | |
698 | dst.push_dentry(src[1]); | |
699 | ||
700 | MClientRequest *req = new MClientRequest(CEPH_MDS_OP_RENAME); | |
701 | req->set_filepath(dst); | |
702 | req->set_filepath2(src); | |
703 | req->set_tid(mds->issue_tid()); | |
704 | ||
705 | mds->send_message_mds(req, to); | |
706 | } | |
707 | ||
708 | StrayManager::StrayManager(MDSRank *mds, PurgeQueue &purge_queue_) | |
709 | : delayed_eval_stray(member_offset(CDentry, item_stray)), | |
710 | mds(mds), logger(NULL), started(false), num_strays(0), | |
711 | num_strays_delayed(0), num_strays_enqueuing(0), | |
712 | purge_queue(purge_queue_) | |
713 | { | |
714 | assert(mds != NULL); | |
715 | } | |
716 | ||
717 | void StrayManager::truncate(CDentry *dn) | |
718 | { | |
719 | const CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
720 | const CInode *in = dnl->get_inode(); | |
721 | assert(in); | |
722 | dout(10) << __func__ << ": " << *dn << " " << *in << dendl; | |
723 | assert(!dn->is_replicated()); | |
724 | ||
725 | const SnapRealm *realm = in->find_snaprealm(); | |
726 | assert(realm); | |
727 | dout(10) << " realm " << *realm << dendl; | |
728 | const SnapContext *snapc = &realm->get_snap_context(); | |
729 | ||
730 | uint64_t to = in->inode.get_max_size(); | |
731 | to = MAX(in->inode.size, to); | |
732 | // when truncating a file, the filer does not delete stripe objects that are | |
733 | // truncated to zero. so we need to purge stripe objects up to the max size | |
734 | // the file has ever been. | |
735 | to = MAX(in->inode.max_size_ever, to); | |
736 | ||
737 | assert(to > 0); | |
738 | ||
739 | PurgeItem item; | |
740 | item.ino = in->inode.ino; | |
741 | item.layout = in->inode.layout; | |
742 | item.snapc = *snapc; | |
743 | item.size = to; | |
744 | ||
745 | purge_queue.push(item, new C_IO_PurgeStrayPurged( | |
746 | this, dn, true)); | |
747 | } | |
748 | ||
749 | void StrayManager::_truncate_stray_logged(CDentry *dn, LogSegment *ls) | |
750 | { | |
751 | CInode *in = dn->get_projected_linkage()->get_inode(); | |
752 | ||
753 | dout(10) << __func__ << ": " << *dn << " " << *in << dendl; | |
754 | ||
755 | dn->state_clear(CDentry::STATE_PURGING | CDentry::STATE_PURGINGPINNED); | |
756 | dn->put(CDentry::PIN_PURGING); | |
757 | ||
758 | in->pop_and_dirty_projected_inode(ls); | |
759 | ||
760 | eval_stray(dn); | |
761 | } | |
762 |