]>
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) 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 | ||
11fdf7f2 | 58 | class StrayManagerContext : public virtual MDSContext { |
7c673cae FG |
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; | |
7c673cae FG |
76 | public: |
77 | C_IO_PurgeStrayPurged(StrayManager *sm_, CDentry *d, bool oh) : | |
78 | StrayManagerIOContext(sm_), dn(d), only_head(oh) { } | |
79 | void finish(int r) override { | |
f67539c2 | 80 | ceph_assert(r == 0 || r == -CEPHFS_ENOENT); |
d2e6a577 | 81 | sm->_purge_stray_purged(dn, only_head); |
7c673cae | 82 | } |
91327a77 AA |
83 | void print(ostream& out) const override { |
84 | CInode *in = dn->get_projected_linkage()->get_inode(); | |
85 | out << "purge_stray(" << in->ino() << ")"; | |
86 | } | |
7c673cae FG |
87 | }; |
88 | ||
89 | ||
90 | void StrayManager::purge(CDentry *dn) | |
91 | { | |
92 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
93 | CInode *in = dnl->get_inode(); | |
94 | dout(10) << __func__ << " " << *dn << " " << *in << dendl; | |
11fdf7f2 | 95 | ceph_assert(!dn->is_replicated()); |
7c673cae FG |
96 | |
97 | // CHEAT. there's no real need to journal our intent to purge, since | |
98 | // that is implicit in the dentry's presence and non-use in the stray | |
99 | // dir. on recovery, we'll need to re-eval all strays anyway. | |
100 | ||
101 | SnapContext nullsnapc; | |
102 | ||
103 | PurgeItem item; | |
f67539c2 | 104 | item.ino = in->ino(); |
11fdf7f2 | 105 | item.stamp = ceph_clock_now(); |
7c673cae FG |
106 | if (in->is_dir()) { |
107 | item.action = PurgeItem::PURGE_DIR; | |
108 | item.fragtree = in->dirfragtree; | |
109 | } else { | |
110 | item.action = PurgeItem::PURGE_FILE; | |
111 | ||
112 | const SnapContext *snapc; | |
113 | SnapRealm *realm = in->find_snaprealm(); | |
114 | if (realm) { | |
115 | dout(10) << " realm " << *realm << dendl; | |
116 | snapc = &realm->get_snap_context(); | |
117 | } else { | |
118 | dout(10) << " NO realm, using null context" << dendl; | |
119 | snapc = &nullsnapc; | |
11fdf7f2 | 120 | ceph_assert(in->last == CEPH_NOSNAP); |
7c673cae FG |
121 | } |
122 | ||
f67539c2 TL |
123 | const auto& pi = in->get_projected_inode(); |
124 | ||
7c673cae FG |
125 | uint64_t to = 0; |
126 | if (in->is_file()) { | |
f67539c2 | 127 | to = std::max(pi->size, pi->get_max_size()); |
7c673cae FG |
128 | // when truncating a file, the filer does not delete stripe objects that are |
129 | // truncated to zero. so we need to purge stripe objects up to the max size | |
130 | // the file has ever been. | |
f67539c2 | 131 | to = std::max(pi->max_size_ever, to); |
7c673cae FG |
132 | } |
133 | ||
7c673cae FG |
134 | item.size = to; |
135 | item.layout = pi->layout; | |
f67539c2 TL |
136 | item.old_pools.reserve(pi->old_pools.size()); |
137 | for (const auto &p : pi->old_pools) { | |
138 | if (p != pi->layout.pool_id) | |
139 | item.old_pools.push_back(p); | |
140 | } | |
7c673cae FG |
141 | item.snapc = *snapc; |
142 | } | |
143 | ||
144 | purge_queue.push(item, new C_IO_PurgeStrayPurged( | |
145 | this, dn, false)); | |
146 | } | |
147 | ||
148 | class C_PurgeStrayLogged : public StrayManagerLogContext { | |
149 | CDentry *dn; | |
150 | version_t pdv; | |
f67539c2 | 151 | MutationRef mut; |
7c673cae | 152 | public: |
f67539c2 TL |
153 | C_PurgeStrayLogged(StrayManager *sm_, CDentry *d, version_t v, MutationRef& m) : |
154 | StrayManagerLogContext(sm_), dn(d), pdv(v), mut(m) { } | |
7c673cae | 155 | void finish(int r) override { |
f67539c2 | 156 | sm->_purge_stray_logged(dn, pdv, mut); |
7c673cae FG |
157 | } |
158 | }; | |
159 | ||
160 | class C_TruncateStrayLogged : public StrayManagerLogContext { | |
161 | CDentry *dn; | |
f67539c2 | 162 | MutationRef mut; |
7c673cae | 163 | public: |
f67539c2 TL |
164 | C_TruncateStrayLogged(StrayManager *sm, CDentry *d, MutationRef& m) : |
165 | StrayManagerLogContext(sm), dn(d), mut(m) {} | |
7c673cae | 166 | void finish(int r) override { |
f67539c2 | 167 | sm->_truncate_stray_logged(dn, mut); |
7c673cae FG |
168 | } |
169 | }; | |
170 | ||
171 | void StrayManager::_purge_stray_purged( | |
d2e6a577 | 172 | CDentry *dn, bool only_head) |
7c673cae FG |
173 | { |
174 | CInode *in = dn->get_projected_linkage()->get_inode(); | |
175 | dout(10) << "_purge_stray_purged " << *dn << " " << *in << dendl; | |
176 | ||
177 | logger->inc(l_mdc_strays_enqueued); | |
178 | num_strays_enqueuing--; | |
179 | logger->set(l_mdc_num_strays_enqueuing, num_strays_enqueuing); | |
180 | ||
181 | if (only_head) { | |
182 | /* This was a ::truncate */ | |
f67539c2 TL |
183 | MutationRef mut(new MutationImpl()); |
184 | mut->ls = mds->mdlog->get_current_segment(); | |
7c673cae | 185 | |
f67539c2 TL |
186 | auto pi = in->project_inode(mut); |
187 | pi.inode->size = 0; | |
188 | pi.inode->max_size_ever = 0; | |
189 | pi.inode->client_ranges.clear(); | |
190 | pi.inode->truncate_size = 0; | |
191 | pi.inode->truncate_from = 0; | |
192 | pi.inode->version = in->pre_dirty(); | |
193 | pi.inode->client_ranges.clear(); | |
f91f0fd5 | 194 | in->clear_clientwriteable(); |
7c673cae | 195 | |
f67539c2 TL |
196 | CDir *dir = dn->get_dir(); |
197 | auto pf = dir->project_fnode(mut); | |
198 | pf->version = dir->pre_dirty(); | |
199 | ||
200 | EUpdate *le = new EUpdate(mds->mdlog, "purge_stray truncate"); | |
201 | mds->mdlog->start_entry(le); | |
202 | ||
203 | le->metablob.add_dir_context(dir); | |
204 | auto& dl = le->metablob.add_dir(dn->dir, true); | |
205 | le->metablob.add_primary_dentry(dl, dn, in, EMetaBlob::fullbit::STATE_DIRTY); | |
7c673cae | 206 | |
f67539c2 | 207 | mds->mdlog->submit_entry(le, new C_TruncateStrayLogged(this, dn, mut)); |
7c673cae FG |
208 | } else { |
209 | if (in->get_num_ref() != (int)in->is_dirty() || | |
f67539c2 TL |
210 | dn->get_num_ref() != |
211 | (int)dn->is_dirty() + | |
212 | !!dn->state_test(CDentry::STATE_FRAGMENTING) + | |
213 | !!in->get_num_ref() + 1 /* PIN_PURGING */) { | |
7c673cae FG |
214 | // Nobody should be taking new references to an inode when it |
215 | // is being purged (aside from it were | |
216 | ||
217 | derr << "Rogue reference after purge to " << *dn << dendl; | |
11fdf7f2 | 218 | ceph_abort_msg("rogue reference to purging inode"); |
7c673cae FG |
219 | } |
220 | ||
f67539c2 TL |
221 | MutationRef mut(new MutationImpl()); |
222 | mut->ls = mds->mdlog->get_current_segment(); | |
223 | ||
7c673cae FG |
224 | // kill dentry. |
225 | version_t pdv = dn->pre_dirty(); | |
226 | dn->push_projected_linkage(); // NULL | |
227 | ||
228 | EUpdate *le = new EUpdate(mds->mdlog, "purge_stray"); | |
229 | mds->mdlog->start_entry(le); | |
230 | ||
231 | // update dirfrag fragstat, rstat | |
232 | CDir *dir = dn->get_dir(); | |
f67539c2 | 233 | auto pf = dir->project_fnode(mut); |
7c673cae FG |
234 | pf->version = dir->pre_dirty(); |
235 | if (in->is_dir()) | |
236 | pf->fragstat.nsubdirs--; | |
237 | else | |
238 | pf->fragstat.nfiles--; | |
f67539c2 | 239 | pf->rstat.sub(in->get_inode()->accounted_rstat); |
7c673cae FG |
240 | |
241 | le->metablob.add_dir_context(dn->dir); | |
f67539c2 | 242 | auto& dl = le->metablob.add_dir(dn->dir, true); |
7c673cae FG |
243 | le->metablob.add_null_dentry(dl, dn, true); |
244 | le->metablob.add_destroyed_inode(in->ino()); | |
245 | ||
f67539c2 | 246 | mds->mdlog->submit_entry(le, new C_PurgeStrayLogged(this, dn, pdv, mut)); |
7c673cae FG |
247 | } |
248 | } | |
249 | ||
f67539c2 | 250 | void StrayManager::_purge_stray_logged(CDentry *dn, version_t pdv, MutationRef& mut) |
7c673cae FG |
251 | { |
252 | CInode *in = dn->get_linkage()->get_inode(); | |
f67539c2 | 253 | CDir *dir = dn->get_dir(); |
7c673cae FG |
254 | dout(10) << "_purge_stray_logged " << *dn << " " << *in << dendl; |
255 | ||
11fdf7f2 | 256 | ceph_assert(!in->state_test(CInode::STATE_RECOVERING)); |
f67539c2 | 257 | ceph_assert(!dir->is_frozen_dir()); |
7c673cae | 258 | |
31f18b77 FG |
259 | bool new_dn = dn->is_new(); |
260 | ||
7c673cae | 261 | // unlink |
11fdf7f2 | 262 | ceph_assert(dn->get_projected_linkage()->is_null()); |
f67539c2 | 263 | dir->unlink_inode(dn, !new_dn); |
7c673cae | 264 | dn->pop_projected_linkage(); |
f67539c2 | 265 | dn->mark_dirty(pdv, mut->ls); |
7c673cae | 266 | |
f67539c2 | 267 | mut->apply(); |
7c673cae FG |
268 | |
269 | in->state_clear(CInode::STATE_ORPHAN); | |
270 | dn->state_clear(CDentry::STATE_PURGING | CDentry::STATE_PURGINGPINNED); | |
271 | dn->put(CDentry::PIN_PURGING); | |
272 | ||
f67539c2 | 273 | |
7c673cae | 274 | // drop dentry? |
31f18b77 | 275 | if (new_dn) { |
7c673cae FG |
276 | dout(20) << " dn is new, removing" << dendl; |
277 | dn->mark_clean(); | |
f67539c2 | 278 | dir->remove_dentry(dn); |
7c673cae FG |
279 | } |
280 | ||
281 | // drop inode | |
f64942e4 | 282 | inodeno_t ino = in->ino(); |
7c673cae FG |
283 | if (in->is_dirty()) |
284 | in->mark_clean(); | |
f64942e4 AA |
285 | mds->mdcache->remove_inode(in); |
286 | ||
f67539c2 TL |
287 | dir->auth_unpin(this); |
288 | ||
f64942e4 AA |
289 | if (mds->is_stopping()) |
290 | mds->mdcache->shutdown_export_stray_finish(ino); | |
7c673cae FG |
291 | } |
292 | ||
293 | void StrayManager::enqueue(CDentry *dn, bool trunc) | |
294 | { | |
295 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
11fdf7f2 | 296 | ceph_assert(dnl); |
7c673cae | 297 | CInode *in = dnl->get_inode(); |
11fdf7f2 | 298 | ceph_assert(in); |
7c673cae FG |
299 | |
300 | /* We consider a stray to be purging as soon as it is enqueued, to avoid | |
301 | * enqueing it twice */ | |
302 | dn->state_set(CDentry::STATE_PURGING); | |
303 | in->state_set(CInode::STATE_PURGING); | |
304 | ||
305 | /* We must clear this as soon as enqueuing it, to prevent the journal | |
306 | * expiry code from seeing a dirty parent and trying to write a backtrace */ | |
307 | if (!trunc) { | |
308 | if (in->is_dirty_parent()) { | |
309 | in->clear_dirty_parent(); | |
310 | } | |
311 | } | |
312 | ||
313 | dout(20) << __func__ << ": purging dn: " << *dn << dendl; | |
314 | ||
315 | if (!dn->state_test(CDentry::STATE_PURGINGPINNED)) { | |
316 | dn->get(CDentry::PIN_PURGING); | |
317 | dn->state_set(CDentry::STATE_PURGINGPINNED); | |
318 | } | |
319 | ||
320 | ++num_strays_enqueuing; | |
321 | logger->set(l_mdc_num_strays_enqueuing, num_strays_enqueuing); | |
322 | ||
323 | // Resources are available, acquire them and execute the purge | |
324 | _enqueue(dn, trunc); | |
325 | ||
326 | dout(10) << __func__ << ": purging this dentry immediately: " | |
327 | << *dn << dendl; | |
328 | } | |
329 | ||
f67539c2 | 330 | class C_RetryEnqueue : public StrayManagerContext { |
7c673cae FG |
331 | CDentry *dn; |
332 | bool trunc; | |
333 | public: | |
f67539c2 | 334 | C_RetryEnqueue(StrayManager *sm_, CDentry *dn_, bool t) : |
7c673cae FG |
335 | StrayManagerContext(sm_), dn(dn_), trunc(t) { } |
336 | void finish(int r) override { | |
337 | sm->_enqueue(dn, trunc); | |
338 | } | |
339 | }; | |
340 | ||
341 | void StrayManager::_enqueue(CDentry *dn, bool trunc) | |
342 | { | |
11fdf7f2 | 343 | ceph_assert(started); |
c07f9fc5 | 344 | |
f67539c2 TL |
345 | CDir *dir = dn->get_dir(); |
346 | if (!dir->can_auth_pin()) { | |
347 | dout(10) << " can't auth_pin (freezing?) " << *dir << ", waiting" << dendl; | |
348 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_RetryEnqueue(this, dn, trunc)); | |
7c673cae FG |
349 | return; |
350 | } | |
351 | ||
f67539c2 | 352 | dn->get_dir()->auth_pin(this); |
7c673cae FG |
353 | if (trunc) { |
354 | truncate(dn); | |
355 | } else { | |
356 | purge(dn); | |
357 | } | |
358 | } | |
359 | ||
11fdf7f2 TL |
360 | void StrayManager::queue_delayed(CDentry *dn) |
361 | { | |
362 | if (!started) | |
363 | return; | |
364 | ||
365 | if (dn->state_test(CDentry::STATE_EVALUATINGSTRAY)) | |
366 | return; | |
367 | ||
368 | if (!dn->item_stray.is_on_list()) { | |
369 | delayed_eval_stray.push_back(&dn->item_stray); | |
370 | num_strays_delayed++; | |
371 | logger->set(l_mdc_num_strays_delayed, num_strays_delayed); | |
372 | } | |
373 | } | |
7c673cae FG |
374 | |
375 | void StrayManager::advance_delayed() | |
376 | { | |
c07f9fc5 FG |
377 | if (!started) |
378 | return; | |
379 | ||
11fdf7f2 TL |
380 | while (!delayed_eval_stray.empty()) { |
381 | CDentry *dn = delayed_eval_stray.front(); | |
7c673cae FG |
382 | dn->item_stray.remove_myself(); |
383 | num_strays_delayed--; | |
384 | ||
385 | if (dn->get_projected_linkage()->is_null()) { | |
386 | /* A special case: a stray dentry can go null if its inode is being | |
387 | * re-linked into another MDS's stray dir during a shutdown migration. */ | |
388 | dout(4) << __func__ << ": delayed dentry is now null: " << *dn << dendl; | |
389 | continue; | |
390 | } | |
391 | ||
81eedcae | 392 | eval_stray(dn); |
7c673cae FG |
393 | } |
394 | logger->set(l_mdc_num_strays_delayed, num_strays_delayed); | |
395 | } | |
396 | ||
397 | void StrayManager::set_num_strays(uint64_t num) | |
398 | { | |
11fdf7f2 | 399 | ceph_assert(!started); |
7c673cae FG |
400 | num_strays = num; |
401 | logger->set(l_mdc_num_strays, num_strays); | |
402 | } | |
403 | ||
404 | void StrayManager::notify_stray_created() | |
405 | { | |
406 | num_strays++; | |
407 | logger->set(l_mdc_num_strays, num_strays); | |
408 | logger->inc(l_mdc_strays_created); | |
409 | } | |
410 | ||
411 | void StrayManager::notify_stray_removed() | |
412 | { | |
413 | num_strays--; | |
414 | logger->set(l_mdc_num_strays, num_strays); | |
415 | } | |
416 | ||
417 | struct C_EvalStray : public StrayManagerContext { | |
418 | CDentry *dn; | |
419 | C_EvalStray(StrayManager *sm_, CDentry *d) : StrayManagerContext(sm_), dn(d) {} | |
420 | void finish(int r) override { | |
421 | sm->eval_stray(dn); | |
422 | } | |
423 | }; | |
424 | ||
425 | struct C_MDC_EvalStray : public StrayManagerContext { | |
426 | CDentry *dn; | |
427 | C_MDC_EvalStray(StrayManager *sm_, CDentry *d) : StrayManagerContext(sm_), dn(d) {} | |
428 | void finish(int r) override { | |
429 | sm->eval_stray(dn); | |
430 | } | |
431 | }; | |
432 | ||
11fdf7f2 | 433 | bool StrayManager::_eval_stray(CDentry *dn) |
7c673cae FG |
434 | { |
435 | dout(10) << "eval_stray " << *dn << dendl; | |
436 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
11fdf7f2 | 437 | ceph_assert(dnl->is_primary()); |
7c673cae FG |
438 | dout(10) << " inode is " << *dnl->get_inode() << dendl; |
439 | CInode *in = dnl->get_inode(); | |
11fdf7f2 TL |
440 | ceph_assert(in); |
441 | ceph_assert(!in->state_test(CInode::STATE_REJOINUNDEF)); | |
7c673cae FG |
442 | |
443 | // The only dentries elegible for purging are those | |
444 | // in the stray directories | |
11fdf7f2 | 445 | ceph_assert(dn->get_dir()->get_inode()->is_stray()); |
7c673cae FG |
446 | |
447 | // Inode may not pass through this function if it | |
448 | // was already identified for purging (i.e. cannot | |
449 | // call eval_stray() after purge() | |
11fdf7f2 | 450 | ceph_assert(!dn->state_test(CDentry::STATE_PURGING)); |
7c673cae | 451 | |
11fdf7f2 | 452 | if (!dn->is_auth()) |
7c673cae | 453 | return false; |
7c673cae | 454 | |
c07f9fc5 | 455 | if (!started) |
11fdf7f2 | 456 | return false; |
c07f9fc5 | 457 | |
7c673cae | 458 | if (dn->item_stray.is_on_list()) { |
7c673cae FG |
459 | dn->item_stray.remove_myself(); |
460 | num_strays_delayed--; | |
461 | logger->set(l_mdc_num_strays_delayed, num_strays_delayed); | |
462 | } | |
463 | ||
464 | // purge? | |
f67539c2 | 465 | if (in->get_inode()->nlink == 0) { |
7c673cae FG |
466 | // past snaprealm parents imply snapped dentry remote links. |
467 | // only important for directories. normal file data snaps are handled | |
468 | // by the object store. | |
469 | if (in->snaprealm) { | |
f67539c2 | 470 | in->snaprealm->prune_past_parent_snaps(); |
7c673cae FG |
471 | in->purge_stale_snap_data(in->snaprealm->get_snaps()); |
472 | } | |
473 | if (in->is_dir()) { | |
f67539c2 | 474 | if (in->snaprealm && in->snaprealm->has_past_parent_snaps()) { |
7c673cae | 475 | dout(20) << " directory has past parents " |
11fdf7f2 | 476 | << in->snaprealm << dendl; |
7c673cae FG |
477 | if (in->state_test(CInode::STATE_MISSINGOBJS)) { |
478 | mds->clog->error() << "previous attempt at committing dirfrag of ino " | |
479 | << in->ino() << " has failed, missing object"; | |
f67539c2 | 480 | mds->handle_write_error(-CEPHFS_ENOENT); |
7c673cae FG |
481 | } |
482 | return false; // not until some snaps are deleted. | |
483 | } | |
484 | ||
f64942e4 | 485 | mds->mdcache->clear_dirty_bits_for_stray(in); |
7c673cae FG |
486 | |
487 | if (!in->remote_parents.empty()) { | |
488 | // unlink any stale remote snap dentry. | |
94b18763 FG |
489 | for (auto it = in->remote_parents.begin(); it != in->remote_parents.end(); ) { |
490 | CDentry *remote_dn = *it; | |
491 | ++it; | |
11fdf7f2 | 492 | ceph_assert(remote_dn->last != CEPH_NOSNAP); |
7c673cae FG |
493 | remote_dn->unlink_remote(remote_dn->get_linkage()); |
494 | } | |
495 | } | |
496 | } | |
497 | if (dn->is_replicated()) { | |
498 | dout(20) << " replicated" << dendl; | |
499 | return false; | |
500 | } | |
501 | if (dn->is_any_leases() || in->is_any_caps()) { | |
502 | dout(20) << " caps | leases" << dendl; | |
503 | return false; // wait | |
504 | } | |
505 | if (in->state_test(CInode::STATE_NEEDSRECOVER) || | |
506 | in->state_test(CInode::STATE_RECOVERING)) { | |
507 | dout(20) << " pending recovery" << dendl; | |
508 | return false; // don't mess with file size probing | |
509 | } | |
510 | if (in->get_num_ref() > (int)in->is_dirty() + (int)in->is_dirty_parent()) { | |
511 | dout(20) << " too many inode refs" << dendl; | |
512 | return false; | |
513 | } | |
514 | if (dn->get_num_ref() > (int)dn->is_dirty() + !!in->get_num_ref()) { | |
515 | dout(20) << " too many dn refs" << dendl; | |
516 | return false; | |
517 | } | |
7c673cae | 518 | // don't purge multiversion inode with snap data |
f67539c2 TL |
519 | if (in->snaprealm && in->snaprealm->has_past_parent_snaps() && |
520 | in->is_any_old_inodes()) { | |
7c673cae FG |
521 | // A file with snapshots: we will truncate the HEAD revision |
522 | // but leave the metadata intact. | |
11fdf7f2 | 523 | ceph_assert(!in->is_dir()); |
7c673cae | 524 | dout(20) << " file has past parents " |
11fdf7f2 | 525 | << in->snaprealm << dendl; |
7c673cae FG |
526 | if (in->is_file() && in->get_projected_inode()->size > 0) { |
527 | enqueue(dn, true); // truncate head objects | |
528 | } | |
529 | } else { | |
530 | // A straightforward file, ready to be purged. Enqueue it. | |
531 | if (in->is_dir()) { | |
532 | in->close_dirfrags(); | |
533 | } | |
534 | ||
535 | enqueue(dn, false); | |
536 | } | |
537 | ||
538 | return true; | |
539 | } else { | |
540 | /* | |
541 | * Where a stray has some links, they should be remotes, check | |
542 | * if we can do anything with them if we happen to have them in | |
543 | * cache. | |
544 | */ | |
31f18b77 | 545 | _eval_stray_remote(dn, NULL); |
7c673cae FG |
546 | return false; |
547 | } | |
548 | } | |
549 | ||
550 | void StrayManager::activate() | |
551 | { | |
552 | dout(10) << __func__ << dendl; | |
553 | started = true; | |
c07f9fc5 | 554 | purge_queue.activate(); |
7c673cae FG |
555 | } |
556 | ||
11fdf7f2 | 557 | bool StrayManager::eval_stray(CDentry *dn) |
7c673cae FG |
558 | { |
559 | // avoid nested eval_stray | |
560 | if (dn->state_test(CDentry::STATE_EVALUATINGSTRAY)) | |
561 | return false; | |
562 | ||
563 | dn->state_set(CDentry::STATE_EVALUATINGSTRAY); | |
11fdf7f2 | 564 | bool ret = _eval_stray(dn); |
7c673cae FG |
565 | dn->state_clear(CDentry::STATE_EVALUATINGSTRAY); |
566 | return ret; | |
567 | } | |
568 | ||
31f18b77 | 569 | void StrayManager::eval_remote(CDentry *remote_dn) |
7c673cae | 570 | { |
31f18b77 FG |
571 | dout(10) << __func__ << " " << *remote_dn << dendl; |
572 | ||
573 | CDentry::linkage_t *dnl = remote_dn->get_projected_linkage(); | |
11fdf7f2 | 574 | ceph_assert(dnl->is_remote()); |
31f18b77 FG |
575 | CInode *in = dnl->get_inode(); |
576 | ||
577 | if (!in) { | |
578 | dout(20) << __func__ << ": no inode, cannot evaluate" << dendl; | |
579 | return; | |
580 | } | |
581 | ||
582 | if (remote_dn->last != CEPH_NOSNAP) { | |
583 | dout(20) << __func__ << ": snap dentry, cannot evaluate" << dendl; | |
584 | return; | |
585 | } | |
586 | ||
587 | // refers to stray? | |
588 | CDentry *primary_dn = in->get_projected_parent_dn(); | |
11fdf7f2 | 589 | ceph_assert(primary_dn != NULL); |
31f18b77 FG |
590 | if (primary_dn->get_dir()->get_inode()->is_stray()) { |
591 | _eval_stray_remote(primary_dn, remote_dn); | |
592 | } else { | |
593 | dout(20) << __func__ << ": inode's primary dn not stray" << dendl; | |
594 | } | |
595 | } | |
596 | ||
597 | class C_RetryEvalRemote : public StrayManagerContext { | |
598 | CDentry *dn; | |
599 | public: | |
600 | C_RetryEvalRemote(StrayManager *sm_, CDentry *dn_) : | |
601 | StrayManagerContext(sm_), dn(dn_) { | |
602 | dn->get(CDentry::PIN_PTRWAITER); | |
603 | } | |
604 | void finish(int r) override { | |
605 | if (dn->get_projected_linkage()->is_remote()) | |
606 | sm->eval_remote(dn); | |
607 | dn->put(CDentry::PIN_PTRWAITER); | |
608 | } | |
609 | }; | |
610 | ||
611 | void StrayManager::_eval_stray_remote(CDentry *stray_dn, CDentry *remote_dn) | |
612 | { | |
613 | dout(20) << __func__ << " " << *stray_dn << dendl; | |
11fdf7f2 TL |
614 | ceph_assert(stray_dn != NULL); |
615 | ceph_assert(stray_dn->get_dir()->get_inode()->is_stray()); | |
7c673cae | 616 | CDentry::linkage_t *stray_dnl = stray_dn->get_projected_linkage(); |
11fdf7f2 | 617 | ceph_assert(stray_dnl->is_primary()); |
7c673cae | 618 | CInode *stray_in = stray_dnl->get_inode(); |
f67539c2 | 619 | ceph_assert(stray_in->get_inode()->nlink >= 1); |
11fdf7f2 | 620 | ceph_assert(stray_in->last == CEPH_NOSNAP); |
7c673cae FG |
621 | |
622 | /* If no remote_dn hinted, pick one arbitrarily */ | |
623 | if (remote_dn == NULL) { | |
624 | if (!stray_in->remote_parents.empty()) { | |
94b18763 FG |
625 | for (const auto &dn : stray_in->remote_parents) { |
626 | if (dn->last == CEPH_NOSNAP && !dn->is_projected()) { | |
627 | if (dn->is_auth()) { | |
628 | remote_dn = dn; | |
31f18b77 FG |
629 | if (remote_dn->dir->can_auth_pin()) |
630 | break; | |
631 | } else if (!remote_dn) { | |
94b18763 | 632 | remote_dn = dn; |
31f18b77 | 633 | } |
7c673cae | 634 | } |
94b18763 | 635 | } |
7c673cae FG |
636 | } |
637 | if (!remote_dn) { | |
638 | dout(20) << __func__ << ": not reintegrating (no remote parents in cache)" << dendl; | |
639 | return; | |
640 | } | |
641 | } | |
11fdf7f2 | 642 | ceph_assert(remote_dn->last == CEPH_NOSNAP); |
7c673cae FG |
643 | // NOTE: we repeat this check in _rename(), since our submission path is racey. |
644 | if (!remote_dn->is_projected()) { | |
31f18b77 FG |
645 | if (remote_dn->is_auth()) { |
646 | if (remote_dn->dir->can_auth_pin()) { | |
647 | reintegrate_stray(stray_dn, remote_dn); | |
648 | } else { | |
649 | remote_dn->dir->add_waiter(CDir::WAIT_UNFREEZE, new C_RetryEvalRemote(this, remote_dn)); | |
650 | dout(20) << __func__ << ": not reintegrating (can't authpin remote parent)" << dendl; | |
651 | } | |
652 | ||
7c673cae FG |
653 | } else if (!remote_dn->is_auth() && stray_dn->is_auth()) { |
654 | migrate_stray(stray_dn, remote_dn->authority().first); | |
655 | } else { | |
656 | dout(20) << __func__ << ": not reintegrating" << dendl; | |
657 | } | |
658 | } else { | |
659 | // don't do anything if the remote parent is projected, or we may | |
660 | // break user-visible semantics! | |
661 | dout(20) << __func__ << ": not reintegrating (projected)" << dendl; | |
662 | } | |
663 | } | |
664 | ||
665 | void StrayManager::reintegrate_stray(CDentry *straydn, CDentry *rdn) | |
666 | { | |
9f95a23c | 667 | dout(10) << __func__ << " " << *straydn << " to " << *rdn << dendl; |
7c673cae FG |
668 | |
669 | logger->inc(l_mdc_strays_reintegrated); | |
670 | ||
9f95a23c TL |
671 | // rename it to remote linkage . |
672 | filepath src(straydn->get_name(), straydn->get_dir()->ino()); | |
673 | filepath dst(rdn->get_name(), rdn->get_dir()->ino()); | |
7c673cae | 674 | |
9f95a23c | 675 | auto req = make_message<MClientRequest>(CEPH_MDS_OP_RENAME); |
7c673cae FG |
676 | req->set_filepath(dst); |
677 | req->set_filepath2(src); | |
678 | req->set_tid(mds->issue_tid()); | |
679 | ||
680 | mds->send_message_mds(req, rdn->authority().first); | |
681 | } | |
682 | ||
683 | void StrayManager::migrate_stray(CDentry *dn, mds_rank_t to) | |
684 | { | |
9f95a23c | 685 | dout(10) << __func__ << " " << *dn << " to mds." << to << dendl; |
7c673cae FG |
686 | |
687 | logger->inc(l_mdc_strays_migrated); | |
688 | ||
689 | // rename it to another mds. | |
9f95a23c TL |
690 | inodeno_t dirino = dn->get_dir()->ino(); |
691 | ceph_assert(MDS_INO_IS_STRAY(dirino)); | |
7c673cae | 692 | |
9f95a23c TL |
693 | filepath src(dn->get_name(), dirino); |
694 | filepath dst(dn->get_name(), MDS_INO_STRAY(to, MDS_INO_STRAY_INDEX(dirino))); | |
7c673cae | 695 | |
9f95a23c | 696 | auto req = make_message<MClientRequest>(CEPH_MDS_OP_RENAME); |
7c673cae FG |
697 | req->set_filepath(dst); |
698 | req->set_filepath2(src); | |
699 | req->set_tid(mds->issue_tid()); | |
700 | ||
701 | mds->send_message_mds(req, to); | |
702 | } | |
703 | ||
704 | StrayManager::StrayManager(MDSRank *mds, PurgeQueue &purge_queue_) | |
705 | : delayed_eval_stray(member_offset(CDentry, item_stray)), | |
9f95a23c | 706 | mds(mds), purge_queue(purge_queue_) |
7c673cae | 707 | { |
11fdf7f2 | 708 | ceph_assert(mds != NULL); |
7c673cae FG |
709 | } |
710 | ||
711 | void StrayManager::truncate(CDentry *dn) | |
712 | { | |
713 | const CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
714 | const CInode *in = dnl->get_inode(); | |
11fdf7f2 | 715 | ceph_assert(in); |
7c673cae | 716 | dout(10) << __func__ << ": " << *dn << " " << *in << dendl; |
11fdf7f2 | 717 | ceph_assert(!dn->is_replicated()); |
7c673cae FG |
718 | |
719 | const SnapRealm *realm = in->find_snaprealm(); | |
11fdf7f2 | 720 | ceph_assert(realm); |
7c673cae FG |
721 | dout(10) << " realm " << *realm << dendl; |
722 | const SnapContext *snapc = &realm->get_snap_context(); | |
723 | ||
f67539c2 | 724 | uint64_t to = std::max(in->get_inode()->size, in->get_inode()->get_max_size()); |
7c673cae FG |
725 | // when truncating a file, the filer does not delete stripe objects that are |
726 | // truncated to zero. so we need to purge stripe objects up to the max size | |
727 | // the file has ever been. | |
f67539c2 | 728 | to = std::max(in->get_inode()->max_size_ever, to); |
7c673cae | 729 | |
11fdf7f2 | 730 | ceph_assert(to > 0); |
7c673cae FG |
731 | |
732 | PurgeItem item; | |
b32b8144 | 733 | item.action = PurgeItem::TRUNCATE_FILE; |
f67539c2 TL |
734 | item.ino = in->ino(); |
735 | item.layout = in->get_inode()->layout; | |
7c673cae FG |
736 | item.snapc = *snapc; |
737 | item.size = to; | |
11fdf7f2 | 738 | item.stamp = ceph_clock_now(); |
7c673cae FG |
739 | |
740 | purge_queue.push(item, new C_IO_PurgeStrayPurged( | |
741 | this, dn, true)); | |
742 | } | |
743 | ||
f67539c2 | 744 | void StrayManager::_truncate_stray_logged(CDentry *dn, MutationRef& mut) |
7c673cae FG |
745 | { |
746 | CInode *in = dn->get_projected_linkage()->get_inode(); | |
747 | ||
748 | dout(10) << __func__ << ": " << *dn << " " << *in << dendl; | |
749 | ||
f67539c2 | 750 | mut->apply(); |
f64942e4 AA |
751 | |
752 | in->state_clear(CInode::STATE_PURGING); | |
7c673cae FG |
753 | dn->state_clear(CDentry::STATE_PURGING | CDentry::STATE_PURGINGPINNED); |
754 | dn->put(CDentry::PIN_PURGING); | |
755 | ||
f67539c2 TL |
756 | dn->get_dir()->auth_unpin(this); |
757 | ||
7c673cae | 758 | eval_stray(dn); |
f64942e4 AA |
759 | |
760 | if (!dn->state_test(CDentry::STATE_PURGING) && mds->is_stopping()) | |
761 | mds->mdcache->shutdown_export_stray_finish(in->ino()); | |
7c673cae FG |
762 | } |
763 |