]>
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 "MDSRank.h" | |
16 | #include "MDCache.h" | |
17 | #include "CInode.h" | |
18 | #include "CDir.h" | |
19 | #include "CDentry.h" | |
20 | #include "Migrator.h" | |
21 | #include "Locker.h" | |
22 | #include "Server.h" | |
23 | ||
24 | #include "MDBalancer.h" | |
25 | #include "MDLog.h" | |
26 | #include "MDSMap.h" | |
27 | #include "Mutation.h" | |
28 | ||
29 | #include "include/filepath.h" | |
28e407b8 | 30 | #include "common/likely.h" |
7c673cae FG |
31 | |
32 | #include "events/EExport.h" | |
33 | #include "events/EImportStart.h" | |
34 | #include "events/EImportFinish.h" | |
35 | #include "events/ESessions.h" | |
36 | ||
37 | #include "msg/Messenger.h" | |
38 | ||
39 | #include "messages/MClientCaps.h" | |
40 | ||
41 | #include "messages/MExportDirDiscover.h" | |
42 | #include "messages/MExportDirDiscoverAck.h" | |
43 | #include "messages/MExportDirCancel.h" | |
44 | #include "messages/MExportDirPrep.h" | |
45 | #include "messages/MExportDirPrepAck.h" | |
46 | #include "messages/MExportDir.h" | |
47 | #include "messages/MExportDirAck.h" | |
48 | #include "messages/MExportDirNotify.h" | |
49 | #include "messages/MExportDirNotifyAck.h" | |
50 | #include "messages/MExportDirFinish.h" | |
51 | ||
52 | #include "messages/MExportCaps.h" | |
53 | #include "messages/MExportCapsAck.h" | |
54 | #include "messages/MGatherCaps.h" | |
55 | ||
56 | ||
57 | /* | |
58 | * this is what the dir->dir_auth values look like | |
59 | * | |
60 | * dir_auth authbits | |
61 | * export | |
62 | * me me - before | |
63 | * me, me me - still me, but preparing for export | |
64 | * me, them me - send MExportDir (peer is preparing) | |
65 | * them, me me - journaled EExport | |
66 | * them them - done | |
67 | * | |
68 | * import: | |
69 | * them them - before | |
70 | * me, them me - journaled EImportStart | |
71 | * me me - done | |
72 | * | |
73 | * which implies: | |
74 | * - auth bit is set if i am listed as first _or_ second dir_auth. | |
75 | */ | |
76 | ||
77 | #include "common/config.h" | |
78 | ||
79 | ||
80 | #define dout_context g_ceph_context | |
81 | #define dout_subsys ceph_subsys_mds | |
82 | #undef dout_prefix | |
83 | #define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".migrator " | |
84 | ||
85 | ||
86 | class MigratorContext : public MDSInternalContextBase { | |
87 | protected: | |
88 | Migrator *mig; | |
89 | MDSRank *get_mds() override { | |
90 | return mig->mds; | |
91 | } | |
92 | public: | |
93 | explicit MigratorContext(Migrator *mig_) : mig(mig_) { | |
94 | assert(mig != NULL); | |
95 | } | |
96 | }; | |
97 | ||
98 | class MigratorLogContext : public MDSLogContextBase { | |
99 | protected: | |
100 | Migrator *mig; | |
101 | MDSRank *get_mds() override { | |
102 | return mig->mds; | |
103 | } | |
104 | public: | |
105 | explicit MigratorLogContext(Migrator *mig_) : mig(mig_) { | |
106 | assert(mig != NULL); | |
107 | } | |
108 | }; | |
109 | ||
110 | /* This function DOES put the passed message before returning*/ | |
111 | void Migrator::dispatch(Message *m) | |
112 | { | |
113 | switch (m->get_type()) { | |
114 | // import | |
115 | case MSG_MDS_EXPORTDIRDISCOVER: | |
116 | handle_export_discover(static_cast<MExportDirDiscover*>(m)); | |
117 | break; | |
118 | case MSG_MDS_EXPORTDIRPREP: | |
119 | handle_export_prep(static_cast<MExportDirPrep*>(m)); | |
120 | break; | |
121 | case MSG_MDS_EXPORTDIR: | |
28e407b8 AA |
122 | if (unlikely(inject_session_race)) { |
123 | dout(0) << "waiting for inject_session_race" << dendl; | |
124 | mds->wait_for_any_client_connection(new C_MDS_RetryMessage(mds, m)); | |
125 | } else { | |
126 | handle_export_dir(static_cast<MExportDir*>(m)); | |
127 | } | |
7c673cae FG |
128 | break; |
129 | case MSG_MDS_EXPORTDIRFINISH: | |
130 | handle_export_finish(static_cast<MExportDirFinish*>(m)); | |
131 | break; | |
132 | case MSG_MDS_EXPORTDIRCANCEL: | |
133 | handle_export_cancel(static_cast<MExportDirCancel*>(m)); | |
134 | break; | |
135 | ||
136 | // export | |
137 | case MSG_MDS_EXPORTDIRDISCOVERACK: | |
138 | handle_export_discover_ack(static_cast<MExportDirDiscoverAck*>(m)); | |
139 | break; | |
140 | case MSG_MDS_EXPORTDIRPREPACK: | |
141 | handle_export_prep_ack(static_cast<MExportDirPrepAck*>(m)); | |
142 | break; | |
143 | case MSG_MDS_EXPORTDIRACK: | |
144 | handle_export_ack(static_cast<MExportDirAck*>(m)); | |
145 | break; | |
146 | case MSG_MDS_EXPORTDIRNOTIFYACK: | |
147 | handle_export_notify_ack(static_cast<MExportDirNotifyAck*>(m)); | |
148 | break; | |
149 | ||
150 | // export 3rd party (dir_auth adjustments) | |
151 | case MSG_MDS_EXPORTDIRNOTIFY: | |
152 | handle_export_notify(static_cast<MExportDirNotify*>(m)); | |
153 | break; | |
154 | ||
155 | // caps | |
156 | case MSG_MDS_EXPORTCAPS: | |
157 | handle_export_caps(static_cast<MExportCaps*>(m)); | |
158 | break; | |
159 | case MSG_MDS_GATHERCAPS: | |
160 | handle_gather_caps(static_cast<MGatherCaps*>(m)); | |
161 | break; | |
162 | ||
163 | default: | |
164 | derr << "migrator unknown message " << m->get_type() << dendl; | |
165 | assert(0 == "migrator unknown message"); | |
166 | } | |
167 | } | |
168 | ||
169 | ||
170 | class C_MDC_EmptyImport : public MigratorContext { | |
171 | CDir *dir; | |
172 | public: | |
173 | C_MDC_EmptyImport(Migrator *m, CDir *d) : MigratorContext(m), dir(d) {} | |
174 | void finish(int r) override { | |
175 | mig->export_empty_import(dir); | |
176 | } | |
177 | }; | |
178 | ||
179 | ||
180 | void Migrator::export_empty_import(CDir *dir) | |
181 | { | |
182 | dout(7) << "export_empty_import " << *dir << dendl; | |
183 | assert(dir->is_subtree_root()); | |
184 | ||
185 | if (dir->inode->is_auth()) { | |
186 | dout(7) << " inode is auth" << dendl; | |
187 | return; | |
188 | } | |
189 | if (!dir->is_auth()) { | |
190 | dout(7) << " not auth" << dendl; | |
191 | return; | |
192 | } | |
193 | if (dir->is_freezing() || dir->is_frozen()) { | |
194 | dout(7) << " freezing or frozen" << dendl; | |
195 | return; | |
196 | } | |
197 | if (dir->get_num_head_items() > 0) { | |
198 | dout(7) << " not actually empty" << dendl; | |
199 | return; | |
200 | } | |
201 | if (dir->inode->is_root()) { | |
202 | dout(7) << " root" << dendl; | |
203 | return; | |
204 | } | |
205 | ||
206 | mds_rank_t dest = dir->inode->authority().first; | |
207 | //if (mds->is_shutting_down()) dest = 0; // this is more efficient. | |
208 | ||
209 | dout(7) << " really empty, exporting to " << dest << dendl; | |
210 | assert (dest != mds->get_nodeid()); | |
211 | ||
212 | dout(7) << "exporting to mds." << dest | |
213 | << " empty import " << *dir << dendl; | |
214 | export_dir( dir, dest ); | |
215 | } | |
216 | ||
217 | void Migrator::find_stale_export_freeze() | |
218 | { | |
219 | utime_t now = ceph_clock_now(); | |
220 | utime_t cutoff = now; | |
221 | cutoff -= g_conf->mds_freeze_tree_timeout; | |
222 | ||
223 | ||
224 | /* | |
225 | * We could have situations like: | |
226 | * | |
227 | * - mds.0 authpins an item in subtree A | |
228 | * - mds.0 sends request to mds.1 to authpin an item in subtree B | |
229 | * - mds.0 freezes subtree A | |
230 | * - mds.1 authpins an item in subtree B | |
231 | * - mds.1 sends request to mds.0 to authpin an item in subtree A | |
232 | * - mds.1 freezes subtree B | |
233 | * - mds.1 receives the remote authpin request from mds.0 | |
234 | * (wait because subtree B is freezing) | |
235 | * - mds.0 receives the remote authpin request from mds.1 | |
236 | * (wait because subtree A is freezing) | |
237 | * | |
238 | * | |
239 | * - client request authpins items in subtree B | |
240 | * - freeze subtree B | |
241 | * - import subtree A which is parent of subtree B | |
242 | * (authpins parent inode of subtree B, see CDir::set_dir_auth()) | |
243 | * - freeze subtree A | |
244 | * - client request tries authpinning items in subtree A | |
245 | * (wait because subtree A is freezing) | |
246 | */ | |
247 | for (map<CDir*,export_state_t>::iterator p = export_state.begin(); | |
248 | p != export_state.end(); ) { | |
249 | CDir* dir = p->first; | |
250 | export_state_t& stat = p->second; | |
251 | ++p; | |
252 | if (stat.state != EXPORT_DISCOVERING && stat.state != EXPORT_FREEZING) | |
253 | continue; | |
254 | if (stat.last_cum_auth_pins != dir->get_cum_auth_pins()) { | |
255 | stat.last_cum_auth_pins = dir->get_cum_auth_pins(); | |
256 | stat.last_cum_auth_pins_change = now; | |
257 | continue; | |
258 | } | |
259 | if (stat.last_cum_auth_pins_change >= cutoff) | |
260 | continue; | |
261 | if (stat.num_remote_waiters > 0 || | |
262 | (!dir->inode->is_root() && dir->get_parent_dir()->is_freezing())) { | |
263 | export_try_cancel(dir); | |
264 | } | |
265 | } | |
266 | } | |
267 | ||
268 | void Migrator::export_try_cancel(CDir *dir, bool notify_peer) | |
269 | { | |
270 | dout(10) << "export_try_cancel " << *dir << dendl; | |
271 | ||
272 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
273 | assert(it != export_state.end()); | |
274 | ||
275 | int state = it->second.state; | |
276 | switch (state) { | |
277 | case EXPORT_LOCKING: | |
278 | dout(10) << "export state=locking : dropping locks and removing auth_pin" << dendl; | |
279 | it->second.state = EXPORT_CANCELLED; | |
280 | dir->auth_unpin(this); | |
281 | break; | |
282 | case EXPORT_DISCOVERING: | |
283 | dout(10) << "export state=discovering : canceling freeze and removing auth_pin" << dendl; | |
284 | it->second.state = EXPORT_CANCELLED; | |
285 | dir->unfreeze_tree(); // cancel the freeze | |
286 | dir->auth_unpin(this); | |
287 | if (notify_peer && | |
288 | (!mds->is_cluster_degraded() || | |
289 | mds->mdsmap->is_clientreplay_or_active_or_stopping(it->second.peer))) // tell them. | |
290 | mds->send_message_mds(new MExportDirCancel(dir->dirfrag(), it->second.tid), it->second.peer); | |
291 | break; | |
292 | ||
293 | case EXPORT_FREEZING: | |
294 | dout(10) << "export state=freezing : canceling freeze" << dendl; | |
295 | it->second.state = EXPORT_CANCELLED; | |
296 | dir->unfreeze_tree(); // cancel the freeze | |
224ce89b WB |
297 | if (dir->is_subtree_root()) |
298 | cache->try_subtree_merge(dir); | |
7c673cae FG |
299 | if (notify_peer && |
300 | (!mds->is_cluster_degraded() || | |
301 | mds->mdsmap->is_clientreplay_or_active_or_stopping(it->second.peer))) // tell them. | |
302 | mds->send_message_mds(new MExportDirCancel(dir->dirfrag(), it->second.tid), it->second.peer); | |
303 | break; | |
304 | ||
305 | // NOTE: state order reversal, warning comes after prepping | |
306 | case EXPORT_WARNING: | |
307 | dout(10) << "export state=warning : unpinning bounds, unfreezing, notifying" << dendl; | |
308 | it->second.state = EXPORT_CANCELLING; | |
309 | // fall-thru | |
310 | ||
311 | case EXPORT_PREPPING: | |
312 | if (state != EXPORT_WARNING) { | |
313 | dout(10) << "export state=prepping : unpinning bounds, unfreezing" << dendl; | |
314 | it->second.state = EXPORT_CANCELLED; | |
315 | } | |
316 | ||
317 | { | |
318 | // unpin bounds | |
319 | set<CDir*> bounds; | |
320 | cache->get_subtree_bounds(dir, bounds); | |
321 | for (set<CDir*>::iterator q = bounds.begin(); | |
322 | q != bounds.end(); | |
323 | ++q) { | |
324 | CDir *bd = *q; | |
325 | bd->put(CDir::PIN_EXPORTBOUND); | |
326 | bd->state_clear(CDir::STATE_EXPORTBOUND); | |
327 | } | |
328 | if (state == EXPORT_WARNING) { | |
329 | // notify bystanders | |
b32b8144 | 330 | export_notify_abort(dir, it->second, bounds); |
7c673cae FG |
331 | // process delayed expires |
332 | cache->process_delayed_expire(dir); | |
333 | } | |
334 | } | |
335 | dir->unfreeze_tree(); | |
7c673cae | 336 | cache->try_subtree_merge(dir); |
b32b8144 FG |
337 | for (auto bd : it->second.residual_dirs) { |
338 | bd->unfreeze_tree(); | |
339 | cache->try_subtree_merge(bd); | |
340 | } | |
7c673cae FG |
341 | if (notify_peer && |
342 | (!mds->is_cluster_degraded() || | |
343 | mds->mdsmap->is_clientreplay_or_active_or_stopping(it->second.peer))) // tell them. | |
344 | mds->send_message_mds(new MExportDirCancel(dir->dirfrag(), it->second.tid), it->second.peer); | |
345 | break; | |
346 | ||
347 | case EXPORT_EXPORTING: | |
348 | dout(10) << "export state=exporting : reversing, and unfreezing" << dendl; | |
349 | it->second.state = EXPORT_CANCELLING; | |
b32b8144 | 350 | export_reverse(dir, it->second); |
7c673cae FG |
351 | break; |
352 | ||
353 | case EXPORT_LOGGINGFINISH: | |
354 | case EXPORT_NOTIFYING: | |
355 | dout(10) << "export state=loggingfinish|notifying : ignoring dest failure, we were successful." << dendl; | |
356 | // leave export_state, don't clean up now. | |
357 | break; | |
358 | case EXPORT_CANCELLING: | |
359 | break; | |
360 | ||
361 | default: | |
362 | ceph_abort(); | |
363 | } | |
364 | ||
365 | // finish clean-up? | |
366 | if (it->second.state == EXPORT_CANCELLING || | |
367 | it->second.state == EXPORT_CANCELLED) { | |
368 | MutationRef mut; | |
369 | mut.swap(it->second.mut); | |
370 | ||
371 | if (it->second.state == EXPORT_CANCELLED) { | |
372 | export_state.erase(it); | |
373 | dir->state_clear(CDir::STATE_EXPORTING); | |
374 | // send pending import_maps? | |
375 | cache->maybe_send_pending_resolves(); | |
376 | } | |
377 | ||
378 | // drop locks | |
379 | if (state == EXPORT_LOCKING || state == EXPORT_DISCOVERING) { | |
380 | MDRequestRef mdr = static_cast<MDRequestImpl*>(mut.get()); | |
381 | assert(mdr); | |
382 | if (mdr->more()->waiting_on_slave.empty()) | |
383 | mds->mdcache->request_finish(mdr); | |
384 | } else if (mut) { | |
385 | mds->locker->drop_locks(mut.get()); | |
386 | mut->cleanup(); | |
387 | } | |
388 | ||
389 | cache->show_subtrees(); | |
390 | ||
391 | maybe_do_queued_export(); | |
392 | } | |
393 | } | |
394 | ||
395 | void Migrator::export_cancel_finish(CDir *dir) | |
396 | { | |
397 | assert(dir->state_test(CDir::STATE_EXPORTING)); | |
398 | dir->state_clear(CDir::STATE_EXPORTING); | |
399 | ||
400 | // pinned by Migrator::export_notify_abort() | |
401 | dir->auth_unpin(this); | |
402 | // send pending import_maps? (these need to go out when all exports have finished.) | |
403 | cache->maybe_send_pending_resolves(); | |
404 | } | |
405 | ||
406 | // ========================================================== | |
407 | // mds failure handling | |
408 | ||
409 | void Migrator::handle_mds_failure_or_stop(mds_rank_t who) | |
410 | { | |
411 | dout(5) << "handle_mds_failure_or_stop mds." << who << dendl; | |
412 | ||
413 | // check my exports | |
414 | ||
415 | // first add an extra auth_pin on any freezes, so that canceling a | |
416 | // nested freeze doesn't complete one further up the hierarchy and | |
417 | // confuse the shit out of us. we'll remove it after canceling the | |
418 | // freeze. this way no freeze completions run before we want them | |
419 | // to. | |
420 | list<CDir*> pinned_dirs; | |
421 | for (map<CDir*,export_state_t>::iterator p = export_state.begin(); | |
422 | p != export_state.end(); | |
423 | ++p) { | |
424 | if (p->second.state == EXPORT_FREEZING) { | |
425 | CDir *dir = p->first; | |
426 | dout(10) << "adding temp auth_pin on freezing " << *dir << dendl; | |
427 | dir->auth_pin(this); | |
428 | pinned_dirs.push_back(dir); | |
429 | } | |
430 | } | |
431 | ||
432 | map<CDir*,export_state_t>::iterator p = export_state.begin(); | |
433 | while (p != export_state.end()) { | |
434 | map<CDir*,export_state_t>::iterator next = p; | |
435 | ++next; | |
436 | CDir *dir = p->first; | |
437 | ||
438 | // abort exports: | |
439 | // - that are going to the failed node | |
440 | // - that aren't frozen yet (to avoid auth_pin deadlock) | |
441 | // - they havne't prepped yet (they may need to discover bounds to do that) | |
442 | if ((p->second.peer == who && | |
443 | p->second.state != EXPORT_CANCELLING) || | |
444 | p->second.state == EXPORT_LOCKING || | |
445 | p->second.state == EXPORT_DISCOVERING || | |
446 | p->second.state == EXPORT_FREEZING || | |
447 | p->second.state == EXPORT_PREPPING) { | |
448 | // the guy i'm exporting to failed, or we're just freezing. | |
449 | dout(10) << "cleaning up export state (" << p->second.state << ")" | |
450 | << get_export_statename(p->second.state) << " of " << *dir << dendl; | |
451 | export_try_cancel(dir); | |
452 | } else if (p->second.peer != who) { | |
453 | // bystander failed. | |
454 | if (p->second.warning_ack_waiting.erase(who)) { | |
455 | if (p->second.state == EXPORT_WARNING) { | |
456 | p->second.notify_ack_waiting.erase(who); // they won't get a notify either. | |
457 | // exporter waiting for warning acks, let's fake theirs. | |
458 | dout(10) << "faking export_warning_ack from mds." << who | |
459 | << " on " << *dir << " to mds." << p->second.peer | |
460 | << dendl; | |
461 | if (p->second.warning_ack_waiting.empty()) | |
462 | export_go(dir); | |
463 | } | |
464 | } | |
465 | if (p->second.notify_ack_waiting.erase(who)) { | |
466 | // exporter is waiting for notify acks, fake it | |
467 | dout(10) << "faking export_notify_ack from mds." << who | |
468 | << " on " << *dir << " to mds." << p->second.peer | |
469 | << dendl; | |
470 | if (p->second.state == EXPORT_NOTIFYING) { | |
471 | if (p->second.notify_ack_waiting.empty()) | |
472 | export_finish(dir); | |
473 | } else if (p->second.state == EXPORT_CANCELLING) { | |
474 | if (p->second.notify_ack_waiting.empty()) { | |
475 | export_state.erase(p); | |
476 | export_cancel_finish(dir); | |
477 | } | |
478 | } | |
479 | } | |
480 | } | |
481 | ||
482 | // next! | |
483 | p = next; | |
484 | } | |
485 | ||
486 | ||
487 | // check my imports | |
488 | map<dirfrag_t,import_state_t>::iterator q = import_state.begin(); | |
489 | while (q != import_state.end()) { | |
490 | map<dirfrag_t,import_state_t>::iterator next = q; | |
491 | ++next; | |
492 | dirfrag_t df = q->first; | |
493 | CInode *diri = mds->mdcache->get_inode(df.ino); | |
494 | CDir *dir = mds->mdcache->get_dirfrag(df); | |
495 | ||
496 | if (q->second.peer == who) { | |
497 | if (dir) | |
498 | dout(10) << "cleaning up import state (" << q->second.state << ")" | |
499 | << get_import_statename(q->second.state) << " of " << *dir << dendl; | |
500 | else | |
501 | dout(10) << "cleaning up import state (" << q->second.state << ")" | |
502 | << get_import_statename(q->second.state) << " of " << df << dendl; | |
503 | ||
504 | switch (q->second.state) { | |
505 | case IMPORT_DISCOVERING: | |
506 | dout(10) << "import state=discovering : clearing state" << dendl; | |
507 | import_reverse_discovering(df); | |
508 | break; | |
509 | ||
510 | case IMPORT_DISCOVERED: | |
511 | assert(diri); | |
512 | dout(10) << "import state=discovered : unpinning inode " << *diri << dendl; | |
513 | import_reverse_discovered(df, diri); | |
514 | break; | |
515 | ||
516 | case IMPORT_PREPPING: | |
517 | assert(dir); | |
518 | dout(10) << "import state=prepping : unpinning base+bounds " << *dir << dendl; | |
b32b8144 | 519 | import_reverse_prepping(dir, q->second); |
7c673cae FG |
520 | break; |
521 | ||
522 | case IMPORT_PREPPED: | |
523 | assert(dir); | |
524 | dout(10) << "import state=prepped : unpinning base+bounds, unfreezing " << *dir << dendl; | |
525 | { | |
526 | set<CDir*> bounds; | |
527 | cache->get_subtree_bounds(dir, bounds); | |
528 | import_remove_pins(dir, bounds); | |
529 | ||
530 | // adjust auth back to the exporter | |
531 | cache->adjust_subtree_auth(dir, q->second.peer); | |
7c673cae FG |
532 | |
533 | // notify bystanders ; wait in aborting state | |
b32b8144 | 534 | q->second.state = IMPORT_ABORTING; |
7c673cae FG |
535 | import_notify_abort(dir, bounds); |
536 | assert(g_conf->mds_kill_import_at != 10); | |
537 | } | |
538 | break; | |
539 | ||
540 | case IMPORT_LOGGINGSTART: | |
541 | assert(dir); | |
542 | dout(10) << "import state=loggingstart : reversing import on " << *dir << dendl; | |
543 | import_reverse(dir); | |
544 | break; | |
545 | ||
546 | case IMPORT_ACKING: | |
547 | assert(dir); | |
548 | // hrm. make this an ambiguous import, and wait for exporter recovery to disambiguate | |
549 | dout(10) << "import state=acking : noting ambiguous import " << *dir << dendl; | |
550 | { | |
551 | set<CDir*> bounds; | |
552 | cache->get_subtree_bounds(dir, bounds); | |
553 | cache->add_ambiguous_import(dir, bounds); | |
554 | } | |
555 | break; | |
556 | ||
557 | case IMPORT_FINISHING: | |
558 | assert(dir); | |
559 | dout(10) << "import state=finishing : finishing import on " << *dir << dendl; | |
560 | import_finish(dir, true); | |
561 | break; | |
562 | ||
563 | case IMPORT_ABORTING: | |
564 | assert(dir); | |
565 | dout(10) << "import state=aborting : ignoring repeat failure " << *dir << dendl; | |
566 | break; | |
567 | } | |
568 | } else { | |
569 | auto bystanders_entry = q->second.bystanders.find(who); | |
570 | if (bystanders_entry != q->second.bystanders.end()) { | |
571 | q->second.bystanders.erase(bystanders_entry); | |
572 | if (q->second.state == IMPORT_ABORTING) { | |
573 | assert(dir); | |
574 | dout(10) << "faking export_notify_ack from mds." << who | |
575 | << " on aborting import " << *dir << " from mds." << q->second.peer | |
576 | << dendl; | |
224ce89b | 577 | if (q->second.bystanders.empty()) |
7c673cae | 578 | import_reverse_unfreeze(dir); |
7c673cae FG |
579 | } |
580 | } | |
581 | } | |
582 | ||
583 | // next! | |
584 | q = next; | |
585 | } | |
586 | ||
587 | while (!pinned_dirs.empty()) { | |
588 | CDir *dir = pinned_dirs.front(); | |
589 | dout(10) << "removing temp auth_pin on " << *dir << dendl; | |
590 | dir->auth_unpin(this); | |
591 | pinned_dirs.pop_front(); | |
592 | } | |
593 | } | |
594 | ||
595 | ||
596 | ||
597 | void Migrator::show_importing() | |
598 | { | |
599 | dout(10) << "show_importing" << dendl; | |
600 | for (map<dirfrag_t,import_state_t>::iterator p = import_state.begin(); | |
601 | p != import_state.end(); | |
602 | ++p) { | |
603 | CDir *dir = mds->mdcache->get_dirfrag(p->first); | |
604 | if (dir) { | |
605 | dout(10) << " importing from " << p->second.peer | |
606 | << ": (" << p->second.state << ") " << get_import_statename(p->second.state) | |
607 | << " " << p->first << " " << *dir << dendl; | |
608 | } else { | |
609 | dout(10) << " importing from " << p->second.peer | |
610 | << ": (" << p->second.state << ") " << get_import_statename(p->second.state) | |
611 | << " " << p->first << dendl; | |
612 | } | |
613 | } | |
614 | } | |
615 | ||
616 | void Migrator::show_exporting() | |
617 | { | |
618 | dout(10) << "show_exporting" << dendl; | |
619 | for (map<CDir*,export_state_t>::iterator p = export_state.begin(); | |
620 | p != export_state.end(); | |
621 | ++p) | |
622 | dout(10) << " exporting to " << p->second.peer | |
623 | << ": (" << p->second.state << ") " << get_export_statename(p->second.state) | |
624 | << " " << p->first->dirfrag() << " " << *p->first << dendl; | |
625 | } | |
626 | ||
627 | ||
628 | ||
629 | void Migrator::audit() | |
630 | { | |
631 | if (!g_conf->subsys.should_gather(ceph_subsys_mds, 5)) | |
632 | return; // hrm. | |
633 | ||
634 | // import_state | |
635 | show_importing(); | |
636 | for (map<dirfrag_t,import_state_t>::iterator p = import_state.begin(); | |
637 | p != import_state.end(); | |
638 | ++p) { | |
639 | if (p->second.state == IMPORT_DISCOVERING) | |
640 | continue; | |
641 | if (p->second.state == IMPORT_DISCOVERED) { | |
642 | CInode *in = cache->get_inode(p->first.ino); | |
643 | assert(in); | |
644 | continue; | |
645 | } | |
646 | CDir *dir = cache->get_dirfrag(p->first); | |
647 | assert(dir); | |
648 | if (p->second.state == IMPORT_PREPPING) | |
649 | continue; | |
650 | if (p->second.state == IMPORT_ABORTING) { | |
651 | assert(!dir->is_ambiguous_dir_auth()); | |
652 | assert(dir->get_dir_auth().first != mds->get_nodeid()); | |
653 | continue; | |
654 | } | |
655 | assert(dir->is_ambiguous_dir_auth()); | |
656 | assert(dir->authority().first == mds->get_nodeid() || | |
657 | dir->authority().second == mds->get_nodeid()); | |
658 | } | |
659 | ||
660 | // export_state | |
661 | show_exporting(); | |
662 | for (map<CDir*,export_state_t>::iterator p = export_state.begin(); | |
663 | p != export_state.end(); | |
664 | ++p) { | |
665 | CDir *dir = p->first; | |
666 | if (p->second.state == EXPORT_LOCKING || | |
667 | p->second.state == EXPORT_DISCOVERING || | |
668 | p->second.state == EXPORT_FREEZING || | |
669 | p->second.state == EXPORT_CANCELLING) | |
670 | continue; | |
671 | assert(dir->is_ambiguous_dir_auth()); | |
672 | assert(dir->authority().first == mds->get_nodeid() || | |
673 | dir->authority().second == mds->get_nodeid()); | |
674 | } | |
675 | ||
676 | // ambiguous+me subtrees should be importing|exporting | |
677 | ||
678 | // write me | |
679 | } | |
680 | ||
681 | ||
682 | ||
683 | ||
684 | ||
685 | // ========================================================== | |
686 | // EXPORT | |
687 | ||
688 | void Migrator::export_dir_nicely(CDir *dir, mds_rank_t dest) | |
689 | { | |
690 | // enqueue | |
691 | dout(7) << "export_dir_nicely " << *dir << " to " << dest << dendl; | |
692 | export_queue.push_back(pair<dirfrag_t,mds_rank_t>(dir->dirfrag(), dest)); | |
693 | ||
694 | maybe_do_queued_export(); | |
695 | } | |
696 | ||
697 | void Migrator::maybe_do_queued_export() | |
698 | { | |
699 | static bool running; | |
700 | if (running) | |
701 | return; | |
702 | running = true; | |
703 | while (!export_queue.empty() && | |
704 | export_state.size() <= 4) { | |
705 | dirfrag_t df = export_queue.front().first; | |
706 | mds_rank_t dest = export_queue.front().second; | |
707 | export_queue.pop_front(); | |
708 | ||
709 | CDir *dir = mds->mdcache->get_dirfrag(df); | |
710 | if (!dir) continue; | |
711 | if (!dir->is_auth()) continue; | |
712 | ||
713 | dout(0) << "nicely exporting to mds." << dest << " " << *dir << dendl; | |
714 | ||
715 | export_dir(dir, dest); | |
716 | } | |
717 | running = false; | |
718 | } | |
719 | ||
720 | ||
721 | ||
722 | ||
723 | class C_MDC_ExportFreeze : public MigratorContext { | |
724 | CDir *ex; // dir i'm exporting | |
725 | uint64_t tid; | |
726 | public: | |
727 | C_MDC_ExportFreeze(Migrator *m, CDir *e, uint64_t t) : | |
728 | MigratorContext(m), ex(e), tid(t) { | |
729 | assert(ex != NULL); | |
730 | } | |
731 | void finish(int r) override { | |
732 | if (r >= 0) | |
733 | mig->export_frozen(ex, tid); | |
734 | } | |
735 | }; | |
736 | ||
737 | ||
738 | void Migrator::get_export_lock_set(CDir *dir, set<SimpleLock*>& locks) | |
739 | { | |
740 | // path | |
741 | vector<CDentry*> trace; | |
742 | cache->make_trace(trace, dir->inode); | |
743 | for (vector<CDentry*>::iterator it = trace.begin(); | |
744 | it != trace.end(); | |
745 | ++it) | |
746 | locks.insert(&(*it)->lock); | |
747 | ||
748 | // prevent scatter gather race | |
749 | locks.insert(&dir->get_inode()->dirfragtreelock); | |
750 | ||
751 | // bound dftlocks: | |
752 | // NOTE: We need to take an rdlock on bounding dirfrags during | |
753 | // migration for a rather irritating reason: when we export the | |
754 | // bound inode, we need to send scatterlock state for the dirfrags | |
755 | // as well, so that the new auth also gets the correct info. If we | |
756 | // race with a refragment, this info is useless, as we can't | |
757 | // redivvy it up. And it's needed for the scatterlocks to work | |
758 | // properly: when the auth is in a sync/lock state it keeps each | |
759 | // dirfrag's portion in the local (auth OR replica) dirfrag. | |
760 | set<CDir*> wouldbe_bounds; | |
761 | cache->get_wouldbe_subtree_bounds(dir, wouldbe_bounds); | |
762 | for (set<CDir*>::iterator p = wouldbe_bounds.begin(); p != wouldbe_bounds.end(); ++p) | |
763 | locks.insert(&(*p)->get_inode()->dirfragtreelock); | |
764 | } | |
765 | ||
766 | ||
31f18b77 | 767 | class C_M_ExportDirWait : public MigratorContext { |
7c673cae FG |
768 | MDRequestRef mdr; |
769 | int count; | |
770 | public: | |
31f18b77 | 771 | C_M_ExportDirWait(Migrator *m, MDRequestRef mdr, int count) |
7c673cae FG |
772 | : MigratorContext(m), mdr(mdr), count(count) {} |
773 | void finish(int r) override { | |
774 | mig->dispatch_export_dir(mdr, count); | |
775 | } | |
776 | }; | |
777 | ||
778 | ||
779 | /** export_dir(dir, dest) | |
780 | * public method to initiate an export. | |
781 | * will fail if the directory is freezing, frozen, unpinnable, or root. | |
782 | */ | |
783 | void Migrator::export_dir(CDir *dir, mds_rank_t dest) | |
784 | { | |
785 | dout(7) << "export_dir " << *dir << " to " << dest << dendl; | |
786 | assert(dir->is_auth()); | |
787 | assert(dest != mds->get_nodeid()); | |
788 | ||
181888fb FG |
789 | if (!(mds->is_active() || mds->is_stopping())) { |
790 | dout(7) << "i'm not active, no exports for now" << dendl; | |
791 | return; | |
792 | } | |
7c673cae FG |
793 | if (mds->mdcache->is_readonly()) { |
794 | dout(7) << "read-only FS, no exports for now" << dendl; | |
795 | return; | |
796 | } | |
31f18b77 FG |
797 | if (!mds->mdsmap->is_active(dest)) { |
798 | dout(7) << "dest not active, no exports for now" << dendl; | |
799 | return; | |
800 | } | |
7c673cae FG |
801 | if (mds->is_cluster_degraded()) { |
802 | dout(7) << "cluster degraded, no exports for now" << dendl; | |
803 | return; | |
804 | } | |
805 | if (dir->inode->is_system()) { | |
806 | dout(7) << "i won't export system dirs (root, mdsdirs, stray, /.ceph, etc.)" << dendl; | |
807 | //ceph_abort(); | |
808 | return; | |
809 | } | |
810 | ||
b32b8144 FG |
811 | CDir* parent_dir = dir->inode->get_projected_parent_dir(); |
812 | if (parent_dir && parent_dir->inode->is_stray()) { | |
813 | if (parent_dir->get_parent_dir()->ino() != MDS_INO_MDSDIR(dest)) { | |
814 | dout(7) << "i won't export anything in stray" << dendl; | |
815 | return; | |
816 | } | |
817 | } else { | |
818 | if (!mds->is_stopping() && !dir->inode->is_exportable(dest)) { | |
819 | dout(7) << "dir is export pinned" << dendl; | |
820 | return; | |
821 | } | |
7c673cae FG |
822 | } |
823 | ||
824 | if (dir->is_frozen() || | |
825 | dir->is_freezing()) { | |
826 | dout(7) << " can't export, freezing|frozen. wait for other exports to finish first." << dendl; | |
827 | return; | |
828 | } | |
829 | if (dir->state_test(CDir::STATE_EXPORTING)) { | |
830 | dout(7) << "already exporting" << dendl; | |
831 | return; | |
832 | } | |
833 | ||
7c673cae FG |
834 | if (g_conf->mds_thrash_exports) { |
835 | // create random subtree bound (which will not be exported) | |
836 | list<CDir*> ls; | |
837 | for (auto p = dir->begin(); p != dir->end(); ++p) { | |
838 | auto dn = p->second; | |
839 | CDentry::linkage_t *dnl= dn->get_linkage(); | |
840 | if (dnl->is_primary()) { | |
841 | CInode *in = dnl->get_inode(); | |
842 | if (in->is_dir()) | |
843 | in->get_nested_dirfrags(ls); | |
844 | } | |
845 | } | |
846 | if (ls.size() > 0) { | |
847 | int n = rand() % ls.size(); | |
848 | auto p = ls.begin(); | |
849 | while (n--) ++p; | |
850 | CDir *bd = *p; | |
851 | if (!(bd->is_frozen() || bd->is_freezing())) { | |
852 | assert(bd->is_auth()); | |
853 | dir->state_set(CDir::STATE_AUXSUBTREE); | |
854 | mds->mdcache->adjust_subtree_auth(dir, mds->get_nodeid()); | |
855 | dout(0) << "export_dir: create aux subtree " << *bd << " under " << *dir << dendl; | |
856 | } | |
857 | } | |
858 | } | |
859 | ||
860 | mds->hit_export_target(ceph_clock_now(), dest, -1); | |
861 | ||
862 | dir->auth_pin(this); | |
863 | dir->state_set(CDir::STATE_EXPORTING); | |
864 | ||
865 | MDRequestRef mdr = mds->mdcache->request_start_internal(CEPH_MDS_OP_EXPORTDIR); | |
866 | mdr->more()->export_dir = dir; | |
867 | ||
868 | assert(export_state.count(dir) == 0); | |
869 | export_state_t& stat = export_state[dir]; | |
870 | stat.state = EXPORT_LOCKING; | |
871 | stat.peer = dest; | |
872 | stat.tid = mdr->reqid.tid; | |
873 | stat.mut = mdr; | |
874 | ||
875 | return mds->mdcache->dispatch_request(mdr); | |
876 | } | |
877 | ||
878 | void Migrator::dispatch_export_dir(MDRequestRef& mdr, int count) | |
879 | { | |
880 | dout(7) << "dispatch_export_dir " << *mdr << dendl; | |
881 | ||
882 | CDir *dir = mdr->more()->export_dir; | |
883 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
884 | if (it == export_state.end() || it->second.tid != mdr->reqid.tid) { | |
885 | // export must have aborted. | |
886 | dout(7) << "export must have aborted " << *mdr << dendl; | |
887 | mds->mdcache->request_finish(mdr); | |
888 | return; | |
889 | } | |
890 | assert(it->second.state == EXPORT_LOCKING); | |
891 | ||
892 | mds_rank_t dest = it->second.peer; | |
893 | ||
894 | if (!mds->is_export_target(dest)) { | |
895 | dout(7) << "dest is not yet an export target" << dendl; | |
896 | if (count > 3) { | |
897 | dout(5) << "dest has not been added as export target after three MDSMap epochs, canceling export" << dendl; | |
898 | export_try_cancel(dir); | |
899 | return; | |
900 | } | |
224ce89b WB |
901 | |
902 | mds->locker->drop_locks(mdr.get()); | |
903 | mdr->drop_local_auth_pins(); | |
904 | ||
31f18b77 | 905 | mds->wait_for_mdsmap(mds->mdsmap->get_epoch(), new C_M_ExportDirWait(this, mdr, count+1)); |
7c673cae FG |
906 | return; |
907 | } | |
908 | ||
909 | if (!dir->inode->get_parent_dn()) { | |
910 | dout(7) << "waiting for dir to become stable before export: " << *dir << dendl; | |
31f18b77 | 911 | dir->add_waiter(CDir::WAIT_CREATED, new C_M_ExportDirWait(this, mdr, 1)); |
7c673cae FG |
912 | return; |
913 | } | |
914 | ||
915 | if (mdr->aborted || dir->is_frozen() || dir->is_freezing()) { | |
916 | dout(7) << "wouldblock|freezing|frozen, canceling export" << dendl; | |
917 | export_try_cancel(dir); | |
918 | return; | |
919 | } | |
920 | ||
921 | // locks? | |
922 | set<SimpleLock*> rdlocks; | |
923 | set<SimpleLock*> xlocks; | |
924 | set<SimpleLock*> wrlocks; | |
925 | get_export_lock_set(dir, rdlocks); | |
926 | // If auth MDS of the subtree root inode is neither the exporter MDS | |
927 | // nor the importer MDS and it gathers subtree root's fragstat/neststat | |
928 | // while the subtree is exporting. It's possible that the exporter MDS | |
929 | // and the importer MDS both are auth MDS of the subtree root or both | |
930 | // are not auth MDS of the subtree root at the time they receive the | |
931 | // lock messages. So the auth MDS of the subtree root inode may get no | |
932 | // or duplicated fragstat/neststat for the subtree root dirfrag. | |
933 | wrlocks.insert(&dir->get_inode()->filelock); | |
934 | wrlocks.insert(&dir->get_inode()->nestlock); | |
935 | if (dir->get_inode()->is_auth()) { | |
936 | dir->get_inode()->filelock.set_scatter_wanted(); | |
937 | dir->get_inode()->nestlock.set_scatter_wanted(); | |
938 | } | |
939 | ||
940 | if (!mds->locker->acquire_locks(mdr, rdlocks, wrlocks, xlocks, NULL, NULL, true)) { | |
941 | if (mdr->aborted) | |
942 | export_try_cancel(dir); | |
943 | return; | |
944 | } | |
945 | ||
946 | assert(g_conf->mds_kill_export_at != 1); | |
947 | it->second.state = EXPORT_DISCOVERING; | |
948 | ||
949 | // send ExportDirDiscover (ask target) | |
950 | filepath path; | |
951 | dir->inode->make_path(path); | |
952 | MExportDirDiscover *discover = new MExportDirDiscover(dir->dirfrag(), path, | |
953 | mds->get_nodeid(), | |
954 | it->second.tid); | |
955 | mds->send_message_mds(discover, dest); | |
956 | assert(g_conf->mds_kill_export_at != 2); | |
957 | ||
958 | it->second.last_cum_auth_pins_change = ceph_clock_now(); | |
959 | ||
960 | // start the freeze, but hold it up with an auth_pin. | |
961 | dir->freeze_tree(); | |
962 | assert(dir->is_freezing_tree()); | |
963 | dir->add_waiter(CDir::WAIT_FROZEN, new C_MDC_ExportFreeze(this, dir, it->second.tid)); | |
964 | } | |
965 | ||
966 | /* | |
967 | * called on receipt of MExportDirDiscoverAck | |
968 | * the importer now has the directory's _inode_ in memory, and pinned. | |
969 | * | |
970 | * This function DOES put the passed message before returning | |
971 | */ | |
972 | void Migrator::handle_export_discover_ack(MExportDirDiscoverAck *m) | |
973 | { | |
974 | CDir *dir = cache->get_dirfrag(m->get_dirfrag()); | |
975 | mds_rank_t dest(m->get_source().num()); | |
976 | utime_t now = ceph_clock_now(); | |
977 | assert(dir); | |
978 | ||
979 | dout(7) << "export_discover_ack from " << m->get_source() | |
980 | << " on " << *dir << dendl; | |
981 | ||
982 | mds->hit_export_target(now, dest, -1); | |
983 | ||
984 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
985 | if (it == export_state.end() || | |
986 | it->second.tid != m->get_tid() || | |
987 | it->second.peer != dest) { | |
988 | dout(7) << "must have aborted" << dendl; | |
989 | } else { | |
990 | assert(it->second.state == EXPORT_DISCOVERING); | |
c07f9fc5 FG |
991 | |
992 | if (m->is_success()) { | |
993 | // release locks to avoid deadlock | |
994 | MDRequestRef mdr = static_cast<MDRequestImpl*>(it->second.mut.get()); | |
995 | assert(mdr); | |
996 | mds->mdcache->request_finish(mdr); | |
997 | it->second.mut.reset(); | |
998 | // freeze the subtree | |
999 | it->second.state = EXPORT_FREEZING; | |
1000 | dir->auth_unpin(this); | |
1001 | assert(g_conf->mds_kill_export_at != 3); | |
1002 | ||
1003 | } else { | |
1004 | dout(7) << "peer failed to discover (not active?), canceling" << dendl; | |
1005 | export_try_cancel(dir, false); | |
1006 | } | |
7c673cae FG |
1007 | } |
1008 | ||
1009 | m->put(); // done | |
1010 | } | |
1011 | ||
1012 | class C_M_ExportSessionsFlushed : public MigratorContext { | |
1013 | CDir *dir; | |
1014 | uint64_t tid; | |
1015 | public: | |
1016 | C_M_ExportSessionsFlushed(Migrator *m, CDir *d, uint64_t t) | |
1017 | : MigratorContext(m), dir(d), tid(t) { | |
1018 | assert(dir != NULL); | |
1019 | } | |
1020 | void finish(int r) override { | |
1021 | mig->export_sessions_flushed(dir, tid); | |
1022 | } | |
1023 | }; | |
1024 | ||
1025 | void Migrator::export_sessions_flushed(CDir *dir, uint64_t tid) | |
1026 | { | |
1027 | dout(7) << "export_sessions_flushed " << *dir << dendl; | |
1028 | ||
1029 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
1030 | if (it == export_state.end() || | |
1031 | it->second.state == EXPORT_CANCELLING || | |
1032 | it->second.tid != tid) { | |
1033 | // export must have aborted. | |
1034 | dout(7) << "export must have aborted on " << dir << dendl; | |
1035 | return; | |
1036 | } | |
1037 | ||
1038 | assert(it->second.state == EXPORT_PREPPING || it->second.state == EXPORT_WARNING); | |
1039 | assert(it->second.warning_ack_waiting.count(MDS_RANK_NONE) > 0); | |
1040 | it->second.warning_ack_waiting.erase(MDS_RANK_NONE); | |
1041 | if (it->second.state == EXPORT_WARNING && it->second.warning_ack_waiting.empty()) | |
1042 | export_go(dir); // start export. | |
1043 | } | |
1044 | ||
1045 | void Migrator::export_frozen(CDir *dir, uint64_t tid) | |
1046 | { | |
1047 | dout(7) << "export_frozen on " << *dir << dendl; | |
1048 | ||
1049 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
1050 | if (it == export_state.end() || it->second.tid != tid) { | |
1051 | dout(7) << "export must have aborted" << dendl; | |
1052 | return; | |
1053 | } | |
1054 | ||
1055 | assert(it->second.state == EXPORT_FREEZING); | |
1056 | assert(dir->is_frozen_tree_root()); | |
1057 | assert(dir->get_cum_auth_pins() == 0); | |
1058 | ||
1059 | CInode *diri = dir->get_inode(); | |
1060 | ||
1061 | // ok, try to grab all my locks. | |
1062 | set<SimpleLock*> rdlocks; | |
1063 | get_export_lock_set(dir, rdlocks); | |
1064 | if ((diri->is_auth() && diri->is_frozen()) || | |
1065 | !mds->locker->can_rdlock_set(rdlocks) || | |
1066 | !diri->filelock.can_wrlock(-1) || | |
1067 | !diri->nestlock.can_wrlock(-1)) { | |
1068 | dout(7) << "export_dir couldn't acquire all needed locks, failing. " | |
1069 | << *dir << dendl; | |
7c673cae FG |
1070 | // .. unwind .. |
1071 | dir->unfreeze_tree(); | |
224ce89b | 1072 | cache->try_subtree_merge(dir); |
7c673cae FG |
1073 | |
1074 | mds->send_message_mds(new MExportDirCancel(dir->dirfrag(), it->second.tid), it->second.peer); | |
7c673cae | 1075 | export_state.erase(it); |
224ce89b WB |
1076 | |
1077 | dir->state_clear(CDir::STATE_EXPORTING); | |
1078 | cache->maybe_send_pending_resolves(); | |
7c673cae FG |
1079 | return; |
1080 | } | |
1081 | ||
1082 | it->second.mut = new MutationImpl(); | |
1083 | if (diri->is_auth()) | |
1084 | it->second.mut->auth_pin(diri); | |
1085 | mds->locker->rdlock_take_set(rdlocks, it->second.mut); | |
1086 | mds->locker->wrlock_force(&diri->filelock, it->second.mut); | |
1087 | mds->locker->wrlock_force(&diri->nestlock, it->second.mut); | |
1088 | ||
1089 | cache->show_subtrees(); | |
1090 | ||
224ce89b WB |
1091 | // CDir::_freeze_tree() should have forced it into subtree. |
1092 | assert(dir->get_dir_auth() == mds_authority_t(mds->get_nodeid(), mds->get_nodeid())); | |
b32b8144 FG |
1093 | |
1094 | set<client_t> export_client_set; | |
1095 | check_export_size(dir, it->second, export_client_set); | |
1096 | ||
7c673cae | 1097 | // note the bounds. |
7c673cae FG |
1098 | set<CDir*> bounds; |
1099 | cache->get_subtree_bounds(dir, bounds); | |
1100 | ||
1101 | // generate prep message, log entry. | |
1102 | MExportDirPrep *prep = new MExportDirPrep(dir->dirfrag(), it->second.tid); | |
1103 | ||
1104 | // include list of bystanders | |
181888fb FG |
1105 | for (const auto &p : dir->get_replicas()) { |
1106 | if (p.first != it->second.peer) { | |
1107 | dout(10) << "bystander mds." << p.first << dendl; | |
1108 | prep->add_bystander(p.first); | |
7c673cae FG |
1109 | } |
1110 | } | |
1111 | ||
1112 | // include base dirfrag | |
1113 | cache->replicate_dir(dir, it->second.peer, prep->basedir); | |
1114 | ||
1115 | /* | |
1116 | * include spanning tree for all nested exports. | |
1117 | * these need to be on the destination _before_ the final export so that | |
1118 | * dir_auth updates on any nested exports are properly absorbed. | |
1119 | * this includes inodes and dirfrags included in the subtree, but | |
1120 | * only the inodes at the bounds. | |
1121 | * | |
1122 | * each trace is: df ('-' | ('f' dir | 'd') dentry inode (dir dentry inode)*) | |
1123 | */ | |
1124 | set<inodeno_t> inodes_added; | |
1125 | set<dirfrag_t> dirfrags_added; | |
1126 | ||
1127 | // check bounds | |
1128 | for (set<CDir*>::iterator p = bounds.begin(); | |
1129 | p != bounds.end(); | |
1130 | ++p) { | |
1131 | CDir *bound = *p; | |
1132 | ||
1133 | // pin it. | |
b32b8144 | 1134 | assert(bound->state_test(CDir::STATE_EXPORTBOUND)); |
7c673cae FG |
1135 | |
1136 | dout(7) << " export bound " << *bound << dendl; | |
1137 | prep->add_bound( bound->dirfrag() ); | |
1138 | ||
1139 | // trace to bound | |
1140 | bufferlist tracebl; | |
1141 | CDir *cur = bound; | |
b32b8144 | 1142 | |
7c673cae | 1143 | char start = '-'; |
b32b8144 FG |
1144 | if (it->second.residual_dirs.count(bound)) { |
1145 | start = 'f'; | |
1146 | cache->replicate_dir(bound, it->second.peer, tracebl); | |
1147 | dout(7) << " added " << *bound << dendl; | |
1148 | } | |
1149 | ||
7c673cae FG |
1150 | while (1) { |
1151 | // don't repeat inodes | |
1152 | if (inodes_added.count(cur->inode->ino())) | |
1153 | break; | |
1154 | inodes_added.insert(cur->inode->ino()); | |
1155 | ||
1156 | // prepend dentry + inode | |
1157 | assert(cur->inode->is_auth()); | |
1158 | bufferlist bl; | |
1159 | cache->replicate_dentry(cur->inode->parent, it->second.peer, bl); | |
1160 | dout(7) << " added " << *cur->inode->parent << dendl; | |
1161 | cache->replicate_inode(cur->inode, it->second.peer, bl, | |
1162 | mds->mdsmap->get_up_features()); | |
1163 | dout(7) << " added " << *cur->inode << dendl; | |
1164 | bl.claim_append(tracebl); | |
1165 | tracebl.claim(bl); | |
1166 | ||
1167 | cur = cur->get_parent_dir(); | |
1168 | ||
1169 | // don't repeat dirfrags | |
1170 | if (dirfrags_added.count(cur->dirfrag()) || | |
1171 | cur == dir) { | |
1172 | start = 'd'; // start with dentry | |
1173 | break; | |
1174 | } | |
1175 | dirfrags_added.insert(cur->dirfrag()); | |
1176 | ||
1177 | // prepend dir | |
1178 | cache->replicate_dir(cur, it->second.peer, bl); | |
1179 | dout(7) << " added " << *cur << dendl; | |
1180 | bl.claim_append(tracebl); | |
1181 | tracebl.claim(bl); | |
1182 | ||
1183 | start = 'f'; // start with dirfrag | |
1184 | } | |
1185 | bufferlist final_bl; | |
1186 | dirfrag_t df = cur->dirfrag(); | |
1187 | ::encode(df, final_bl); | |
1188 | ::encode(start, final_bl); | |
1189 | final_bl.claim_append(tracebl); | |
1190 | prep->add_trace(final_bl); | |
1191 | } | |
1192 | ||
1193 | // send. | |
1194 | it->second.state = EXPORT_PREPPING; | |
1195 | mds->send_message_mds(prep, it->second.peer); | |
1196 | assert (g_conf->mds_kill_export_at != 4); | |
1197 | ||
1198 | // make sure any new instantiations of caps are flushed out | |
1199 | assert(it->second.warning_ack_waiting.empty()); | |
1200 | ||
7c673cae FG |
1201 | MDSGatherBuilder gather(g_ceph_context); |
1202 | mds->server->flush_client_sessions(export_client_set, gather); | |
1203 | if (gather.has_subs()) { | |
1204 | it->second.warning_ack_waiting.insert(MDS_RANK_NONE); | |
1205 | gather.set_finisher(new C_M_ExportSessionsFlushed(this, dir, it->second.tid)); | |
1206 | gather.activate(); | |
1207 | } | |
1208 | } | |
1209 | ||
b32b8144 | 1210 | void Migrator::check_export_size(CDir *dir, export_state_t& stat, set<client_t>& client_set) |
7c673cae | 1211 | { |
b32b8144 FG |
1212 | const unsigned frag_size = 800; |
1213 | const unsigned inode_size = 1000; | |
1214 | const unsigned cap_size = 80; | |
1215 | const unsigned link_size = 10; | |
1216 | const unsigned null_size = 1; | |
1217 | ||
1218 | uint64_t max_size = g_conf->get_val<uint64_t>("mds_max_export_size"); | |
1219 | uint64_t approx_size = 0; | |
1220 | ||
7c673cae FG |
1221 | list<CDir*> dfs; |
1222 | dfs.push_back(dir); | |
1223 | while (!dfs.empty()) { | |
1224 | CDir *dir = dfs.front(); | |
1225 | dfs.pop_front(); | |
b32b8144 FG |
1226 | |
1227 | approx_size += frag_size; | |
94b18763 FG |
1228 | for (auto &p : *dir) { |
1229 | CDentry *dn = p.second; | |
b32b8144 FG |
1230 | if (dn->get_linkage()->is_null()) { |
1231 | approx_size += null_size; | |
1232 | continue; | |
1233 | } | |
1234 | if (dn->get_linkage()->is_remote()) { | |
1235 | approx_size += link_size; | |
7c673cae | 1236 | continue; |
b32b8144 FG |
1237 | } |
1238 | ||
1239 | approx_size += inode_size; | |
7c673cae FG |
1240 | CInode *in = dn->get_linkage()->get_inode(); |
1241 | if (in->is_dir()) { | |
1242 | // directory? | |
1243 | list<CDir*> ls; | |
1244 | in->get_dirfrags(ls); | |
b32b8144 FG |
1245 | for (auto q : ls) { |
1246 | if (q->is_subtree_root()) { | |
1247 | q->state_set(CDir::STATE_EXPORTBOUND); | |
1248 | q->get(CDir::PIN_EXPORTBOUND); | |
1249 | } else { | |
7c673cae | 1250 | // include nested dirfrag |
b32b8144 FG |
1251 | assert(q->get_dir_auth().first == CDIR_AUTH_PARENT); |
1252 | dfs.push_front(q); | |
7c673cae FG |
1253 | } |
1254 | } | |
1255 | } | |
1256 | for (map<client_t, Capability*>::iterator q = in->client_caps.begin(); | |
1257 | q != in->client_caps.end(); | |
b32b8144 FG |
1258 | ++q) { |
1259 | approx_size += cap_size; | |
7c673cae | 1260 | client_set.insert(q->first); |
b32b8144 | 1261 | } |
7c673cae | 1262 | } |
b32b8144 FG |
1263 | |
1264 | if (approx_size >= max_size) | |
1265 | break; | |
1266 | } | |
1267 | ||
1268 | while (!dfs.empty()) { | |
1269 | CDir *dir = dfs.front(); | |
1270 | dfs.pop_front(); | |
1271 | ||
1272 | dout(7) << "check_export_size: creating bound " << *dir << dendl; | |
1273 | assert(dir->is_auth()); | |
1274 | dir->state_set(CDir::STATE_EXPORTBOUND); | |
1275 | dir->get(CDir::PIN_EXPORTBOUND); | |
1276 | ||
1277 | mds->mdcache->adjust_subtree_auth(dir, mds->get_nodeid()); | |
1278 | // Another choice here is finishing all WAIT_UNFREEZE contexts and keeping | |
1279 | // the newly created subtree unfreeze. | |
1280 | dir->_freeze_tree(); | |
1281 | ||
1282 | stat.residual_dirs.insert(dir); | |
7c673cae FG |
1283 | } |
1284 | } | |
1285 | ||
1286 | void Migrator::get_export_client_set(CInode *in, set<client_t>& client_set) | |
1287 | { | |
1288 | for (map<client_t, Capability*>::iterator q = in->client_caps.begin(); | |
1289 | q != in->client_caps.end(); | |
1290 | ++q) | |
1291 | client_set.insert(q->first); | |
1292 | } | |
1293 | ||
1294 | /* This function DOES put the passed message before returning*/ | |
1295 | void Migrator::handle_export_prep_ack(MExportDirPrepAck *m) | |
1296 | { | |
1297 | CDir *dir = cache->get_dirfrag(m->get_dirfrag()); | |
1298 | mds_rank_t dest(m->get_source().num()); | |
1299 | utime_t now = ceph_clock_now(); | |
1300 | assert(dir); | |
1301 | ||
1302 | dout(7) << "export_prep_ack " << *dir << dendl; | |
1303 | ||
1304 | mds->hit_export_target(now, dest, -1); | |
1305 | ||
1306 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
1307 | if (it == export_state.end() || | |
1308 | it->second.tid != m->get_tid() || | |
1309 | it->second.peer != mds_rank_t(m->get_source().num())) { | |
1310 | // export must have aborted. | |
1311 | dout(7) << "export must have aborted" << dendl; | |
1312 | m->put(); | |
1313 | return; | |
1314 | } | |
1315 | assert(it->second.state == EXPORT_PREPPING); | |
1316 | ||
1317 | if (!m->is_success()) { | |
c07f9fc5 | 1318 | dout(7) << "peer couldn't acquire all needed locks or wasn't active, canceling" << dendl; |
7c673cae FG |
1319 | export_try_cancel(dir, false); |
1320 | m->put(); | |
1321 | return; | |
1322 | } | |
1323 | ||
1324 | assert (g_conf->mds_kill_export_at != 5); | |
1325 | // send warnings | |
1326 | set<CDir*> bounds; | |
1327 | cache->get_subtree_bounds(dir, bounds); | |
1328 | ||
1329 | assert(it->second.warning_ack_waiting.empty() || | |
1330 | (it->second.warning_ack_waiting.size() == 1 && | |
1331 | it->second.warning_ack_waiting.count(MDS_RANK_NONE) > 0)); | |
1332 | assert(it->second.notify_ack_waiting.empty()); | |
1333 | ||
181888fb FG |
1334 | for (const auto &p : dir->get_replicas()) { |
1335 | if (p.first == it->second.peer) continue; | |
7c673cae | 1336 | if (mds->is_cluster_degraded() && |
181888fb | 1337 | !mds->mdsmap->is_clientreplay_or_active_or_stopping(p.first)) |
7c673cae | 1338 | continue; // only if active |
181888fb FG |
1339 | it->second.warning_ack_waiting.insert(p.first); |
1340 | it->second.notify_ack_waiting.insert(p.first); // we'll eventually get a notifyack, too! | |
7c673cae FG |
1341 | |
1342 | MExportDirNotify *notify = new MExportDirNotify(dir->dirfrag(), it->second.tid, true, | |
1343 | mds_authority_t(mds->get_nodeid(),CDIR_AUTH_UNKNOWN), | |
1344 | mds_authority_t(mds->get_nodeid(),it->second.peer)); | |
1345 | for (set<CDir*>::iterator q = bounds.begin(); q != bounds.end(); ++q) | |
1346 | notify->get_bounds().push_back((*q)->dirfrag()); | |
181888fb | 1347 | mds->send_message_mds(notify, p.first); |
7c673cae FG |
1348 | |
1349 | } | |
1350 | ||
1351 | it->second.state = EXPORT_WARNING; | |
1352 | ||
1353 | assert(g_conf->mds_kill_export_at != 6); | |
1354 | // nobody to warn? | |
1355 | if (it->second.warning_ack_waiting.empty()) | |
1356 | export_go(dir); // start export. | |
1357 | ||
1358 | // done. | |
1359 | m->put(); | |
1360 | } | |
1361 | ||
1362 | ||
1363 | class C_M_ExportGo : public MigratorContext { | |
1364 | CDir *dir; | |
1365 | uint64_t tid; | |
1366 | public: | |
1367 | C_M_ExportGo(Migrator *m, CDir *d, uint64_t t) : | |
1368 | MigratorContext(m), dir(d), tid(t) { | |
1369 | assert(dir != NULL); | |
1370 | } | |
1371 | void finish(int r) override { | |
1372 | mig->export_go_synced(dir, tid); | |
1373 | } | |
1374 | }; | |
1375 | ||
1376 | void Migrator::export_go(CDir *dir) | |
1377 | { | |
b32b8144 FG |
1378 | auto it = export_state.find(dir); |
1379 | assert(it != export_state.end()); | |
1380 | dout(7) << "export_go " << *dir << " to " << it->second.peer << dendl; | |
7c673cae FG |
1381 | |
1382 | // first sync log to flush out e.g. any cap imports | |
b32b8144 | 1383 | mds->mdlog->wait_for_safe(new C_M_ExportGo(this, dir, it->second.tid)); |
7c673cae FG |
1384 | mds->mdlog->flush(); |
1385 | } | |
1386 | ||
1387 | void Migrator::export_go_synced(CDir *dir, uint64_t tid) | |
1388 | { | |
1389 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
1390 | if (it == export_state.end() || | |
1391 | it->second.state == EXPORT_CANCELLING || | |
1392 | it->second.tid != tid) { | |
1393 | // export must have aborted. | |
1394 | dout(7) << "export must have aborted on " << dir << dendl; | |
1395 | return; | |
1396 | } | |
1397 | assert(it->second.state == EXPORT_WARNING); | |
1398 | mds_rank_t dest = it->second.peer; | |
1399 | ||
1400 | dout(7) << "export_go_synced " << *dir << " to " << dest << dendl; | |
1401 | ||
1402 | cache->show_subtrees(); | |
1403 | ||
1404 | it->second.state = EXPORT_EXPORTING; | |
1405 | assert(g_conf->mds_kill_export_at != 7); | |
1406 | ||
1407 | assert(dir->is_frozen_tree_root()); | |
1408 | assert(dir->get_cum_auth_pins() == 0); | |
1409 | ||
1410 | // set ambiguous auth | |
1411 | cache->adjust_subtree_auth(dir, mds->get_nodeid(), dest); | |
1412 | ||
1413 | // take away the popularity we're sending. | |
1414 | utime_t now = ceph_clock_now(); | |
1415 | mds->balancer->subtract_export(dir, now); | |
1416 | ||
1417 | // fill export message with cache data | |
1418 | MExportDir *req = new MExportDir(dir->dirfrag(), it->second.tid); | |
1419 | map<client_t,entity_inst_t> exported_client_map; | |
1420 | uint64_t num_exported_inodes = encode_export_dir(req->export_data, | |
1421 | dir, // recur start point | |
1422 | exported_client_map, | |
1423 | now); | |
1424 | ::encode(exported_client_map, req->client_map, | |
1425 | mds->mdsmap->get_up_features()); | |
1426 | ||
1427 | // add bounds to message | |
1428 | set<CDir*> bounds; | |
1429 | cache->get_subtree_bounds(dir, bounds); | |
1430 | for (set<CDir*>::iterator p = bounds.begin(); | |
1431 | p != bounds.end(); | |
1432 | ++p) | |
1433 | req->add_export((*p)->dirfrag()); | |
1434 | ||
1435 | // send | |
1436 | mds->send_message_mds(req, dest); | |
1437 | assert(g_conf->mds_kill_export_at != 8); | |
1438 | ||
1439 | mds->hit_export_target(now, dest, num_exported_inodes+1); | |
1440 | ||
1441 | // stats | |
1442 | if (mds->logger) mds->logger->inc(l_mds_exported); | |
1443 | if (mds->logger) mds->logger->inc(l_mds_exported_inodes, num_exported_inodes); | |
1444 | ||
1445 | cache->show_subtrees(); | |
1446 | } | |
1447 | ||
1448 | ||
1449 | /** encode_export_inode | |
1450 | * update our local state for this inode to export. | |
1451 | * encode relevant state to be sent over the wire. | |
1452 | * used by: encode_export_dir, file_rename (if foreign) | |
1453 | * | |
1454 | * FIXME: the separation between CInode.encode_export and these methods | |
1455 | * is pretty arbitrary and dumb. | |
1456 | */ | |
1457 | void Migrator::encode_export_inode(CInode *in, bufferlist& enc_state, | |
1458 | map<client_t,entity_inst_t>& exported_client_map) | |
1459 | { | |
1460 | dout(7) << "encode_export_inode " << *in << dendl; | |
1461 | assert(!in->is_replica(mds->get_nodeid())); | |
1462 | ||
1463 | // relax locks? | |
1464 | if (!in->is_replicated()) { | |
1465 | in->replicate_relax_locks(); | |
1466 | dout(20) << " did replicate_relax_locks, now " << *in << dendl; | |
1467 | } | |
1468 | ||
1469 | ::encode(in->inode.ino, enc_state); | |
1470 | ::encode(in->last, enc_state); | |
1471 | in->encode_export(enc_state); | |
1472 | ||
1473 | // caps | |
1474 | encode_export_inode_caps(in, true, enc_state, exported_client_map); | |
1475 | } | |
1476 | ||
1477 | void Migrator::encode_export_inode_caps(CInode *in, bool auth_cap, bufferlist& bl, | |
1478 | map<client_t,entity_inst_t>& exported_client_map) | |
1479 | { | |
1480 | dout(20) << "encode_export_inode_caps " << *in << dendl; | |
1481 | ||
1482 | // encode caps | |
1483 | map<client_t,Capability::Export> cap_map; | |
1484 | in->export_client_caps(cap_map); | |
1485 | ::encode(cap_map, bl); | |
1486 | if (auth_cap) { | |
1487 | ::encode(in->get_mds_caps_wanted(), bl); | |
1488 | ||
1489 | in->state_set(CInode::STATE_EXPORTINGCAPS); | |
1490 | in->get(CInode::PIN_EXPORTINGCAPS); | |
1491 | } | |
1492 | ||
1493 | // make note of clients named by exported capabilities | |
1494 | for (map<client_t, Capability*>::iterator it = in->client_caps.begin(); | |
1495 | it != in->client_caps.end(); | |
1496 | ++it) | |
1497 | exported_client_map[it->first] = mds->sessionmap.get_inst(entity_name_t::CLIENT(it->first.v)); | |
1498 | } | |
1499 | ||
1500 | void Migrator::finish_export_inode_caps(CInode *in, mds_rank_t peer, | |
1501 | map<client_t,Capability::Import>& peer_imported) | |
1502 | { | |
1503 | dout(20) << "finish_export_inode_caps " << *in << dendl; | |
1504 | ||
1505 | in->state_clear(CInode::STATE_EXPORTINGCAPS); | |
1506 | in->put(CInode::PIN_EXPORTINGCAPS); | |
1507 | ||
1508 | // tell (all) clients about migrating caps.. | |
1509 | for (map<client_t, Capability*>::iterator it = in->client_caps.begin(); | |
1510 | it != in->client_caps.end(); | |
1511 | ++it) { | |
1512 | Capability *cap = it->second; | |
1513 | dout(7) << "finish_export_inode_caps telling client." << it->first | |
1514 | << " exported caps on " << *in << dendl; | |
1515 | MClientCaps *m = new MClientCaps(CEPH_CAP_OP_EXPORT, in->ino(), 0, | |
1516 | cap->get_cap_id(), cap->get_mseq(), mds->get_osd_epoch_barrier()); | |
1517 | ||
1518 | map<client_t,Capability::Import>::iterator q = peer_imported.find(it->first); | |
1519 | assert(q != peer_imported.end()); | |
28e407b8 AA |
1520 | m->set_cap_peer(q->second.cap_id, q->second.issue_seq, q->second.mseq, |
1521 | (q->second.cap_id > 0 ? peer : -1), 0); | |
7c673cae FG |
1522 | mds->send_message_client_counted(m, it->first); |
1523 | } | |
1524 | in->clear_client_caps_after_export(); | |
1525 | mds->locker->eval(in, CEPH_CAP_LOCKS); | |
1526 | } | |
1527 | ||
1528 | void Migrator::finish_export_inode(CInode *in, utime_t now, mds_rank_t peer, | |
1529 | map<client_t,Capability::Import>& peer_imported, | |
1530 | list<MDSInternalContextBase*>& finished) | |
1531 | { | |
1532 | dout(12) << "finish_export_inode " << *in << dendl; | |
1533 | ||
1534 | // clean | |
1535 | if (in->is_dirty()) | |
1536 | in->mark_clean(); | |
1537 | ||
1538 | // clear/unpin cached_by (we're no longer the authority) | |
1539 | in->clear_replica_map(); | |
1540 | ||
1541 | // twiddle lock states for auth -> replica transition | |
1542 | in->authlock.export_twiddle(); | |
1543 | in->linklock.export_twiddle(); | |
1544 | in->dirfragtreelock.export_twiddle(); | |
1545 | in->filelock.export_twiddle(); | |
1546 | in->nestlock.export_twiddle(); | |
1547 | in->xattrlock.export_twiddle(); | |
1548 | in->snaplock.export_twiddle(); | |
1549 | in->flocklock.export_twiddle(); | |
1550 | in->policylock.export_twiddle(); | |
1551 | ||
1552 | // mark auth | |
1553 | assert(in->is_auth()); | |
1554 | in->state_clear(CInode::STATE_AUTH); | |
1555 | in->replica_nonce = CInode::EXPORT_NONCE; | |
1556 | ||
1557 | in->clear_dirty_rstat(); | |
1558 | ||
1559 | // no more auth subtree? clear scatter dirty | |
1560 | if (!in->has_subtree_root_dirfrag(mds->get_nodeid())) | |
1561 | in->clear_scatter_dirty(); | |
1562 | ||
1563 | in->item_open_file.remove_myself(); | |
1564 | ||
1565 | in->clear_dirty_parent(); | |
1566 | ||
1567 | in->clear_file_locks(); | |
1568 | ||
1569 | // waiters | |
1570 | in->take_waiting(CInode::WAIT_ANY_MASK, finished); | |
1571 | ||
1572 | in->finish_export(now); | |
1573 | ||
1574 | finish_export_inode_caps(in, peer, peer_imported); | |
7c673cae FG |
1575 | } |
1576 | ||
1577 | uint64_t Migrator::encode_export_dir(bufferlist& exportbl, | |
1578 | CDir *dir, | |
1579 | map<client_t,entity_inst_t>& exported_client_map, | |
1580 | utime_t now) | |
1581 | { | |
1582 | uint64_t num_exported = 0; | |
1583 | ||
1584 | dout(7) << "encode_export_dir " << *dir << " " << dir->get_num_head_items() << " head items" << dendl; | |
1585 | ||
1586 | assert(dir->get_projected_version() == dir->get_version()); | |
1587 | ||
1588 | #ifdef MDS_VERIFY_FRAGSTAT | |
1589 | if (dir->is_complete()) | |
1590 | dir->verify_fragstat(); | |
1591 | #endif | |
1592 | ||
1593 | // dir | |
1594 | dirfrag_t df = dir->dirfrag(); | |
1595 | ::encode(df, exportbl); | |
1596 | dir->encode_export(exportbl); | |
1597 | ||
1598 | __u32 nden = dir->items.size(); | |
1599 | ::encode(nden, exportbl); | |
1600 | ||
1601 | // dentries | |
1602 | list<CDir*> subdirs; | |
94b18763 FG |
1603 | for (auto &p : *dir) { |
1604 | CDentry *dn = p.second; | |
7c673cae FG |
1605 | CInode *in = dn->get_linkage()->get_inode(); |
1606 | ||
1607 | if (!dn->is_replicated()) | |
1608 | dn->lock.replicate_relax(); | |
1609 | ||
1610 | num_exported++; | |
1611 | ||
1612 | // -- dentry | |
1613 | dout(7) << "encode_export_dir exporting " << *dn << dendl; | |
1614 | ||
1615 | // dn name | |
94b18763 | 1616 | ::encode(dn->get_name(), exportbl); |
7c673cae FG |
1617 | ::encode(dn->last, exportbl); |
1618 | ||
1619 | // state | |
1620 | dn->encode_export(exportbl); | |
1621 | ||
1622 | // points to... | |
1623 | ||
1624 | // null dentry? | |
1625 | if (dn->get_linkage()->is_null()) { | |
1626 | exportbl.append("N", 1); // null dentry | |
1627 | continue; | |
1628 | } | |
1629 | ||
1630 | if (dn->get_linkage()->is_remote()) { | |
1631 | // remote link | |
1632 | exportbl.append("L", 1); // remote link | |
1633 | ||
1634 | inodeno_t ino = dn->get_linkage()->get_remote_ino(); | |
1635 | unsigned char d_type = dn->get_linkage()->get_remote_d_type(); | |
1636 | ::encode(ino, exportbl); | |
1637 | ::encode(d_type, exportbl); | |
1638 | continue; | |
1639 | } | |
1640 | ||
1641 | // primary link | |
1642 | // -- inode | |
1643 | exportbl.append("I", 1); // inode dentry | |
1644 | ||
1645 | encode_export_inode(in, exportbl, exported_client_map); // encode, and (update state for) export | |
1646 | ||
1647 | // directory? | |
1648 | list<CDir*> dfs; | |
1649 | in->get_dirfrags(dfs); | |
1650 | for (list<CDir*>::iterator p = dfs.begin(); p != dfs.end(); ++p) { | |
1651 | CDir *t = *p; | |
1652 | if (!t->state_test(CDir::STATE_EXPORTBOUND)) { | |
1653 | // include nested dirfrag | |
1654 | assert(t->get_dir_auth().first == CDIR_AUTH_PARENT); | |
b32b8144 | 1655 | subdirs.push_front(t); // it's ours, recurse (later) |
7c673cae FG |
1656 | } |
1657 | } | |
1658 | } | |
1659 | ||
1660 | // subdirs | |
94b18763 FG |
1661 | for (auto &dir : subdirs) |
1662 | num_exported += encode_export_dir(exportbl, dir, exported_client_map, now); | |
7c673cae FG |
1663 | |
1664 | return num_exported; | |
1665 | } | |
1666 | ||
1667 | void Migrator::finish_export_dir(CDir *dir, utime_t now, mds_rank_t peer, | |
1668 | map<inodeno_t,map<client_t,Capability::Import> >& peer_imported, | |
1669 | list<MDSInternalContextBase*>& finished, int *num_dentries) | |
1670 | { | |
1671 | dout(10) << "finish_export_dir " << *dir << dendl; | |
1672 | ||
1673 | // release open_by | |
1674 | dir->clear_replica_map(); | |
1675 | ||
1676 | // mark | |
1677 | assert(dir->is_auth()); | |
1678 | dir->state_clear(CDir::STATE_AUTH); | |
1679 | dir->remove_bloom(); | |
1680 | dir->replica_nonce = CDir::EXPORT_NONCE; | |
1681 | ||
1682 | if (dir->is_dirty()) | |
1683 | dir->mark_clean(); | |
1684 | ||
1685 | // suck up all waiters | |
1686 | dir->take_waiting(CDir::WAIT_ANY_MASK, finished); // all dir waiters | |
1687 | ||
1688 | // pop | |
1689 | dir->finish_export(now); | |
1690 | ||
1691 | // dentries | |
1692 | list<CDir*> subdirs; | |
94b18763 FG |
1693 | for (auto &p : *dir) { |
1694 | CDentry *dn = p.second; | |
7c673cae FG |
1695 | CInode *in = dn->get_linkage()->get_inode(); |
1696 | ||
1697 | // dentry | |
1698 | dn->finish_export(); | |
1699 | ||
1700 | // inode? | |
1701 | if (dn->get_linkage()->is_primary()) { | |
1702 | finish_export_inode(in, now, peer, peer_imported[in->ino()], finished); | |
1703 | ||
1704 | // subdirs? | |
1705 | in->get_nested_dirfrags(subdirs); | |
1706 | } | |
1707 | ||
1708 | cache->touch_dentry_bottom(dn); // move dentry to tail of LRU | |
1709 | ++(*num_dentries); | |
1710 | } | |
1711 | ||
1712 | // subdirs | |
1713 | for (list<CDir*>::iterator it = subdirs.begin(); it != subdirs.end(); ++it) | |
1714 | finish_export_dir(*it, now, peer, peer_imported, finished, num_dentries); | |
1715 | } | |
1716 | ||
1717 | class C_MDS_ExportFinishLogged : public MigratorLogContext { | |
1718 | CDir *dir; | |
1719 | public: | |
1720 | C_MDS_ExportFinishLogged(Migrator *m, CDir *d) : MigratorLogContext(m), dir(d) {} | |
1721 | void finish(int r) override { | |
1722 | mig->export_logged_finish(dir); | |
1723 | } | |
1724 | }; | |
1725 | ||
1726 | ||
1727 | /* | |
1728 | * i should get an export_ack from the export target. | |
1729 | * | |
1730 | * This function DOES put the passed message before returning | |
1731 | */ | |
1732 | void Migrator::handle_export_ack(MExportDirAck *m) | |
1733 | { | |
1734 | CDir *dir = cache->get_dirfrag(m->get_dirfrag()); | |
1735 | mds_rank_t dest(m->get_source().num()); | |
1736 | utime_t now = ceph_clock_now(); | |
1737 | assert(dir); | |
1738 | assert(dir->is_frozen_tree_root()); // i'm exporting! | |
1739 | ||
1740 | // yay! | |
1741 | dout(7) << "handle_export_ack " << *dir << dendl; | |
1742 | ||
1743 | mds->hit_export_target(now, dest, -1); | |
1744 | ||
1745 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
1746 | assert(it != export_state.end()); | |
1747 | assert(it->second.state == EXPORT_EXPORTING); | |
1748 | assert(it->second.tid == m->get_tid()); | |
1749 | ||
1750 | bufferlist::iterator bp = m->imported_caps.begin(); | |
1751 | ::decode(it->second.peer_imported, bp); | |
1752 | ||
1753 | it->second.state = EXPORT_LOGGINGFINISH; | |
1754 | assert (g_conf->mds_kill_export_at != 9); | |
1755 | set<CDir*> bounds; | |
1756 | cache->get_subtree_bounds(dir, bounds); | |
1757 | ||
7c673cae FG |
1758 | // log completion. |
1759 | // include export bounds, to ensure they're in the journal. | |
31f18b77 | 1760 | EExport *le = new EExport(mds->mdlog, dir, it->second.peer);; |
7c673cae FG |
1761 | mds->mdlog->start_entry(le); |
1762 | ||
1763 | le->metablob.add_dir_context(dir, EMetaBlob::TO_ROOT); | |
31f18b77 | 1764 | le->metablob.add_dir(dir, false); |
7c673cae FG |
1765 | for (set<CDir*>::iterator p = bounds.begin(); |
1766 | p != bounds.end(); | |
1767 | ++p) { | |
1768 | CDir *bound = *p; | |
1769 | le->get_bounds().insert(bound->dirfrag()); | |
1770 | le->metablob.add_dir_context(bound); | |
1771 | le->metablob.add_dir(bound, false); | |
1772 | } | |
1773 | ||
31f18b77 FG |
1774 | // list us second, them first. |
1775 | // this keeps authority().first in sync with subtree auth state in the journal. | |
1776 | cache->adjust_subtree_auth(dir, it->second.peer, mds->get_nodeid()); | |
1777 | ||
7c673cae FG |
1778 | // log export completion, then finish (unfreeze, trigger finish context, etc.) |
1779 | mds->mdlog->submit_entry(le, new C_MDS_ExportFinishLogged(this, dir)); | |
1780 | mds->mdlog->flush(); | |
1781 | assert (g_conf->mds_kill_export_at != 10); | |
1782 | ||
1783 | m->put(); | |
1784 | } | |
1785 | ||
b32b8144 | 1786 | void Migrator::export_notify_abort(CDir *dir, export_state_t& stat, set<CDir*>& bounds) |
7c673cae FG |
1787 | { |
1788 | dout(7) << "export_notify_abort " << *dir << dendl; | |
1789 | ||
7c673cae FG |
1790 | assert(stat.state == EXPORT_CANCELLING); |
1791 | ||
1792 | if (stat.notify_ack_waiting.empty()) { | |
1793 | stat.state = EXPORT_CANCELLED; | |
1794 | return; | |
1795 | } | |
1796 | ||
1797 | dir->auth_pin(this); | |
1798 | ||
1799 | for (set<mds_rank_t>::iterator p = stat.notify_ack_waiting.begin(); | |
1800 | p != stat.notify_ack_waiting.end(); | |
1801 | ++p) { | |
b32b8144 FG |
1802 | MExportDirNotify *notify = new MExportDirNotify(dir->dirfrag(), stat.tid, true, |
1803 | pair<int,int>(mds->get_nodeid(), stat.peer), | |
1804 | pair<int,int>(mds->get_nodeid(), CDIR_AUTH_UNKNOWN)); | |
7c673cae FG |
1805 | for (set<CDir*>::iterator i = bounds.begin(); i != bounds.end(); ++i) |
1806 | notify->get_bounds().push_back((*i)->dirfrag()); | |
1807 | mds->send_message_mds(notify, *p); | |
1808 | } | |
1809 | } | |
1810 | ||
1811 | /* | |
1812 | * this happens if hte dest failes after i send teh export data but before it is acked | |
1813 | * that is, we don't know they safely received and logged it, so we reverse our changes | |
1814 | * and go on. | |
1815 | */ | |
b32b8144 | 1816 | void Migrator::export_reverse(CDir *dir, export_state_t& stat) |
7c673cae FG |
1817 | { |
1818 | dout(7) << "export_reverse " << *dir << dendl; | |
1819 | ||
1820 | set<CInode*> to_eval; | |
1821 | ||
1822 | set<CDir*> bounds; | |
1823 | cache->get_subtree_bounds(dir, bounds); | |
1824 | ||
1825 | // remove exporting pins | |
1826 | list<CDir*> rq; | |
1827 | rq.push_back(dir); | |
1828 | while (!rq.empty()) { | |
1829 | CDir *t = rq.front(); | |
1830 | rq.pop_front(); | |
1831 | t->abort_export(); | |
94b18763 FG |
1832 | for (auto &p : *t) { |
1833 | CDentry *dn = p.second; | |
1834 | dn->abort_export(); | |
1835 | if (!dn->get_linkage()->is_primary()) | |
7c673cae | 1836 | continue; |
94b18763 | 1837 | CInode *in = dn->get_linkage()->get_inode(); |
7c673cae FG |
1838 | in->abort_export(); |
1839 | if (in->state_test(CInode::STATE_EVALSTALECAPS)) { | |
1840 | in->state_clear(CInode::STATE_EVALSTALECAPS); | |
1841 | to_eval.insert(in); | |
1842 | } | |
1843 | if (in->is_dir()) | |
1844 | in->get_nested_dirfrags(rq); | |
1845 | } | |
1846 | } | |
1847 | ||
1848 | // unpin bounds | |
b32b8144 | 1849 | for (auto bd : bounds) { |
7c673cae FG |
1850 | bd->put(CDir::PIN_EXPORTBOUND); |
1851 | bd->state_clear(CDir::STATE_EXPORTBOUND); | |
1852 | } | |
1853 | ||
7c673cae | 1854 | // notify bystanders |
b32b8144 | 1855 | export_notify_abort(dir, stat, bounds); |
7c673cae | 1856 | |
224ce89b WB |
1857 | // unfreeze tree, with possible subtree merge. |
1858 | cache->adjust_subtree_auth(dir, mds->get_nodeid(), mds->get_nodeid()); | |
1859 | ||
7c673cae FG |
1860 | // process delayed expires |
1861 | cache->process_delayed_expire(dir); | |
224ce89b | 1862 | |
7c673cae | 1863 | dir->unfreeze_tree(); |
224ce89b | 1864 | cache->try_subtree_merge(dir); |
b32b8144 FG |
1865 | for (auto bd : stat.residual_dirs) { |
1866 | bd->unfreeze_tree(); | |
1867 | cache->try_subtree_merge(bd); | |
1868 | } | |
7c673cae FG |
1869 | |
1870 | // revoke/resume stale caps | |
1871 | for (auto in : to_eval) { | |
1872 | bool need_issue = false; | |
1873 | for (auto& p : in->get_client_caps()) { | |
1874 | Capability *cap = p.second; | |
1875 | if (cap->is_stale()) { | |
1876 | mds->locker->revoke_stale_caps(cap); | |
1877 | } else { | |
1878 | need_issue = true; | |
1879 | } | |
1880 | } | |
1881 | if (need_issue && | |
1882 | (!in->is_auth() || !mds->locker->eval(in, CEPH_CAP_LOCKS))) | |
1883 | mds->locker->issue_caps(in); | |
1884 | } | |
1885 | ||
1886 | cache->show_cache(); | |
1887 | } | |
1888 | ||
1889 | ||
1890 | /* | |
1891 | * once i get the ack, and logged the EExportFinish(true), | |
1892 | * send notifies (if any), otherwise go straight to finish. | |
1893 | * | |
1894 | */ | |
1895 | void Migrator::export_logged_finish(CDir *dir) | |
1896 | { | |
1897 | dout(7) << "export_logged_finish " << *dir << dendl; | |
1898 | ||
1899 | export_state_t& stat = export_state[dir]; | |
1900 | ||
1901 | // send notifies | |
1902 | set<CDir*> bounds; | |
1903 | cache->get_subtree_bounds(dir, bounds); | |
1904 | ||
1905 | for (set<mds_rank_t>::iterator p = stat.notify_ack_waiting.begin(); | |
1906 | p != stat.notify_ack_waiting.end(); | |
1907 | ++p) { | |
1908 | MExportDirNotify *notify = new MExportDirNotify(dir->dirfrag(), stat.tid, true, | |
1909 | pair<int,int>(mds->get_nodeid(), stat.peer), | |
1910 | pair<int,int>(stat.peer, CDIR_AUTH_UNKNOWN)); | |
1911 | ||
1912 | for (set<CDir*>::iterator i = bounds.begin(); i != bounds.end(); ++i) | |
1913 | notify->get_bounds().push_back((*i)->dirfrag()); | |
1914 | ||
1915 | mds->send_message_mds(notify, *p); | |
1916 | } | |
1917 | ||
1918 | // wait for notifyacks | |
1919 | stat.state = EXPORT_NOTIFYING; | |
1920 | assert (g_conf->mds_kill_export_at != 11); | |
1921 | ||
1922 | // no notifies to wait for? | |
1923 | if (stat.notify_ack_waiting.empty()) { | |
1924 | export_finish(dir); // skip notify/notify_ack stage. | |
1925 | } else { | |
1926 | // notify peer to send cap import messages to clients | |
1927 | if (!mds->is_cluster_degraded() || | |
1928 | mds->mdsmap->is_clientreplay_or_active_or_stopping(stat.peer)) { | |
1929 | mds->send_message_mds(new MExportDirFinish(dir->dirfrag(), false, stat.tid), stat.peer); | |
1930 | } else { | |
1931 | dout(7) << "not sending MExportDirFinish, dest has failed" << dendl; | |
1932 | } | |
1933 | } | |
1934 | } | |
1935 | ||
1936 | /* | |
1937 | * warning: | |
1938 | * i'll get an ack from each bystander. | |
1939 | * when i get them all, do the export. | |
1940 | * notify: | |
1941 | * i'll get an ack from each bystander. | |
1942 | * when i get them all, unfreeze and send the finish. | |
1943 | * | |
1944 | * This function DOES put the passed message before returning | |
1945 | */ | |
1946 | void Migrator::handle_export_notify_ack(MExportDirNotifyAck *m) | |
1947 | { | |
1948 | CDir *dir = cache->get_dirfrag(m->get_dirfrag()); | |
1949 | mds_rank_t dest(m->get_source().num()); | |
1950 | utime_t now = ceph_clock_now(); | |
1951 | assert(dir); | |
1952 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
1953 | ||
1954 | mds->hit_export_target(now, dest, -1); | |
1955 | ||
1956 | auto export_state_entry = export_state.find(dir); | |
1957 | if (export_state_entry != export_state.end()) { | |
1958 | export_state_t& stat = export_state_entry->second; | |
1959 | if (stat.state == EXPORT_WARNING && | |
1960 | stat.warning_ack_waiting.erase(from)) { | |
1961 | // exporting. process warning. | |
1962 | dout(7) << "handle_export_notify_ack from " << m->get_source() | |
1963 | << ": exporting, processing warning on " << *dir << dendl; | |
1964 | if (stat.warning_ack_waiting.empty()) | |
1965 | export_go(dir); // start export. | |
1966 | } else if (stat.state == EXPORT_NOTIFYING && | |
1967 | stat.notify_ack_waiting.erase(from)) { | |
1968 | // exporting. process notify. | |
1969 | dout(7) << "handle_export_notify_ack from " << m->get_source() | |
1970 | << ": exporting, processing notify on " << *dir << dendl; | |
1971 | if (stat.notify_ack_waiting.empty()) | |
1972 | export_finish(dir); | |
1973 | } else if (stat.state == EXPORT_CANCELLING && | |
1974 | m->get_new_auth().second == CDIR_AUTH_UNKNOWN && // not warning ack | |
1975 | stat.notify_ack_waiting.erase(from)) { | |
1976 | dout(7) << "handle_export_notify_ack from " << m->get_source() | |
1977 | << ": cancelling export, processing notify on " << *dir << dendl; | |
1978 | if (stat.notify_ack_waiting.empty()) { | |
1979 | export_state.erase(export_state_entry); | |
1980 | export_cancel_finish(dir); | |
1981 | } | |
1982 | } | |
1983 | } | |
1984 | else { | |
1985 | auto import_state_entry = import_state.find(dir->dirfrag()); | |
1986 | if (import_state_entry != import_state.end()) { | |
1987 | import_state_t& stat = import_state_entry->second; | |
1988 | if (stat.state == IMPORT_ABORTING) { | |
1989 | // reversing import | |
1990 | dout(7) << "handle_export_notify_ack from " << m->get_source() | |
1991 | << ": aborting import on " << *dir << dendl; | |
1992 | assert(stat.bystanders.count(from)); | |
1993 | stat.bystanders.erase(from); | |
1994 | if (stat.bystanders.empty()) | |
1995 | import_reverse_unfreeze(dir); | |
1996 | } | |
1997 | } | |
1998 | } | |
1999 | ||
2000 | m->put(); | |
2001 | } | |
2002 | ||
2003 | void Migrator::export_finish(CDir *dir) | |
2004 | { | |
2005 | dout(5) << "export_finish " << *dir << dendl; | |
2006 | ||
2007 | assert (g_conf->mds_kill_export_at != 12); | |
2008 | map<CDir*,export_state_t>::iterator it = export_state.find(dir); | |
2009 | if (it == export_state.end()) { | |
2010 | dout(7) << "target must have failed, not sending final commit message. export succeeded anyway." << dendl; | |
2011 | return; | |
2012 | } | |
2013 | ||
2014 | // send finish/commit to new auth | |
2015 | if (!mds->is_cluster_degraded() || | |
2016 | mds->mdsmap->is_clientreplay_or_active_or_stopping(it->second.peer)) { | |
2017 | mds->send_message_mds(new MExportDirFinish(dir->dirfrag(), true, it->second.tid), it->second.peer); | |
2018 | } else { | |
2019 | dout(7) << "not sending MExportDirFinish last, dest has failed" << dendl; | |
2020 | } | |
2021 | assert(g_conf->mds_kill_export_at != 13); | |
2022 | ||
2023 | // finish export (adjust local cache state) | |
2024 | int num_dentries = 0; | |
224ce89b | 2025 | list<MDSInternalContextBase*> finished; |
7c673cae | 2026 | finish_export_dir(dir, ceph_clock_now(), it->second.peer, |
224ce89b WB |
2027 | it->second.peer_imported, finished, &num_dentries); |
2028 | ||
2029 | assert(!dir->is_auth()); | |
2030 | cache->adjust_subtree_auth(dir, it->second.peer); | |
2031 | ||
7c673cae FG |
2032 | // unpin bounds |
2033 | set<CDir*> bounds; | |
2034 | cache->get_subtree_bounds(dir, bounds); | |
2035 | for (set<CDir*>::iterator p = bounds.begin(); | |
2036 | p != bounds.end(); | |
2037 | ++p) { | |
2038 | CDir *bd = *p; | |
2039 | bd->put(CDir::PIN_EXPORTBOUND); | |
2040 | bd->state_clear(CDir::STATE_EXPORTBOUND); | |
2041 | } | |
2042 | ||
2043 | if (dir->state_test(CDir::STATE_AUXSUBTREE)) | |
2044 | dir->state_clear(CDir::STATE_AUXSUBTREE); | |
2045 | ||
224ce89b WB |
2046 | // discard delayed expires |
2047 | cache->discard_delayed_expire(dir); | |
2048 | ||
2049 | dout(7) << "export_finish unfreezing" << dendl; | |
2050 | ||
2051 | // unfreeze tree, with possible subtree merge. | |
7c673cae | 2052 | // (we do this _after_ removing EXPORTBOUND pins, to allow merges) |
224ce89b | 2053 | dir->unfreeze_tree(); |
7c673cae | 2054 | cache->try_subtree_merge(dir); |
b32b8144 FG |
2055 | for (auto bd : it->second.residual_dirs) { |
2056 | export_queue.push_front(pair<dirfrag_t,mds_rank_t>(bd->dirfrag(), it->second.peer)); | |
2057 | bd->take_waiting(CDir::WAIT_ANY_MASK, finished); | |
2058 | bd->unfreeze_tree(); | |
2059 | cache->try_subtree_merge(bd); | |
2060 | } | |
7c673cae FG |
2061 | |
2062 | // no more auth subtree? clear scatter dirty | |
2063 | if (!dir->get_inode()->is_auth() && | |
2064 | !dir->get_inode()->has_subtree_root_dirfrag(mds->get_nodeid())) { | |
2065 | dir->get_inode()->clear_scatter_dirty(); | |
2066 | // wake up scatter_nudge waiters | |
224ce89b | 2067 | dir->get_inode()->take_waiting(CInode::WAIT_ANY_MASK, finished); |
7c673cae FG |
2068 | } |
2069 | ||
224ce89b WB |
2070 | if (!finished.empty()) |
2071 | mds->queue_waiters(finished); | |
7c673cae FG |
2072 | |
2073 | MutationRef mut = it->second.mut; | |
2074 | // remove from exporting list, clean up state | |
2075 | export_state.erase(it); | |
2076 | dir->state_clear(CDir::STATE_EXPORTING); | |
2077 | ||
2078 | cache->show_subtrees(); | |
2079 | audit(); | |
2080 | ||
181888fb | 2081 | cache->trim(num_dentries); // try trimming exported dentries |
7c673cae FG |
2082 | |
2083 | // send pending import_maps? | |
2084 | mds->mdcache->maybe_send_pending_resolves(); | |
2085 | ||
2086 | // drop locks, unpin path | |
2087 | if (mut) { | |
2088 | mds->locker->drop_locks(mut.get()); | |
2089 | mut->cleanup(); | |
2090 | } | |
2091 | ||
2092 | maybe_do_queued_export(); | |
2093 | } | |
2094 | ||
2095 | ||
2096 | ||
2097 | ||
2098 | ||
2099 | ||
2100 | ||
2101 | ||
2102 | // ========================================================== | |
2103 | // IMPORT | |
2104 | ||
2105 | void Migrator::handle_export_discover(MExportDirDiscover *m) | |
2106 | { | |
2107 | mds_rank_t from = m->get_source_mds(); | |
2108 | assert(from != mds->get_nodeid()); | |
2109 | ||
2110 | dout(7) << "handle_export_discover on " << m->get_path() << dendl; | |
2111 | ||
2112 | // note import state | |
2113 | dirfrag_t df = m->get_dirfrag(); | |
c07f9fc5 FG |
2114 | |
2115 | if (!mds->is_active()) { | |
2116 | dout(7) << " not active, send NACK " << dendl; | |
2117 | mds->send_message_mds(new MExportDirDiscoverAck(df, m->get_tid(), false), from); | |
2118 | m->put(); | |
2119 | return; | |
2120 | } | |
2121 | ||
7c673cae | 2122 | // only start discovering on this message once. |
b32b8144 | 2123 | import_state_t *p_state; |
7c673cae FG |
2124 | map<dirfrag_t,import_state_t>::iterator it = import_state.find(df); |
2125 | if (!m->started) { | |
2126 | assert(it == import_state.end()); | |
2127 | m->started = true; | |
b32b8144 FG |
2128 | p_state = &import_state[df]; |
2129 | p_state->state = IMPORT_DISCOVERING; | |
2130 | p_state->peer = from; | |
2131 | p_state->tid = m->get_tid(); | |
7c673cae FG |
2132 | } else { |
2133 | // am i retrying after ancient path_traverse results? | |
2134 | if (it == import_state.end() || | |
2135 | it->second.peer != from || | |
2136 | it->second.tid != m->get_tid()) { | |
2137 | dout(7) << " dropping obsolete message" << dendl; | |
2138 | m->put(); | |
2139 | return; | |
2140 | } | |
2141 | assert(it->second.state == IMPORT_DISCOVERING); | |
b32b8144 | 2142 | p_state = &it->second; |
7c673cae FG |
2143 | } |
2144 | ||
2145 | if (!mds->mdcache->is_open()) { | |
2146 | dout(5) << " waiting for root" << dendl; | |
2147 | mds->mdcache->wait_for_open(new C_MDS_RetryMessage(mds, m)); | |
2148 | return; | |
2149 | } | |
2150 | ||
2151 | assert (g_conf->mds_kill_import_at != 1); | |
2152 | ||
2153 | // do we have it? | |
2154 | CInode *in = cache->get_inode(m->get_dirfrag().ino); | |
2155 | if (!in) { | |
2156 | // must discover it! | |
2157 | filepath fpath(m->get_path()); | |
2158 | vector<CDentry*> trace; | |
2159 | MDRequestRef null_ref; | |
2160 | int r = cache->path_traverse(null_ref, m, NULL, fpath, &trace, NULL, MDS_TRAVERSE_DISCOVER); | |
2161 | if (r > 0) return; | |
2162 | if (r < 0) { | |
2163 | dout(7) << "handle_export_discover_2 failed to discover or not dir " << m->get_path() << ", NAK" << dendl; | |
2164 | ceph_abort(); // this shouldn't happen if the auth pins its path properly!!!! | |
2165 | } | |
2166 | ||
2167 | ceph_abort(); // this shouldn't happen; the get_inode above would have succeeded. | |
2168 | } | |
2169 | ||
2170 | // yay | |
2171 | dout(7) << "handle_export_discover have " << df << " inode " << *in << dendl; | |
2172 | ||
b32b8144 | 2173 | p_state->state = IMPORT_DISCOVERED; |
7c673cae FG |
2174 | |
2175 | // pin inode in the cache (for now) | |
2176 | assert(in->is_dir()); | |
2177 | in->get(CInode::PIN_IMPORTING); | |
2178 | ||
2179 | // reply | |
2180 | dout(7) << " sending export_discover_ack on " << *in << dendl; | |
b32b8144 | 2181 | mds->send_message_mds(new MExportDirDiscoverAck(df, m->get_tid()), p_state->peer); |
7c673cae FG |
2182 | m->put(); |
2183 | assert (g_conf->mds_kill_import_at != 2); | |
2184 | } | |
2185 | ||
2186 | void Migrator::import_reverse_discovering(dirfrag_t df) | |
2187 | { | |
2188 | import_state.erase(df); | |
2189 | } | |
2190 | ||
2191 | void Migrator::import_reverse_discovered(dirfrag_t df, CInode *diri) | |
2192 | { | |
2193 | // unpin base | |
2194 | diri->put(CInode::PIN_IMPORTING); | |
2195 | import_state.erase(df); | |
2196 | } | |
2197 | ||
b32b8144 | 2198 | void Migrator::import_reverse_prepping(CDir *dir, import_state_t& stat) |
7c673cae FG |
2199 | { |
2200 | set<CDir*> bounds; | |
b32b8144 | 2201 | cache->map_dirfrag_set(stat.bound_ls, bounds); |
7c673cae FG |
2202 | import_remove_pins(dir, bounds); |
2203 | import_reverse_final(dir); | |
2204 | } | |
2205 | ||
2206 | /* This function DOES put the passed message before returning*/ | |
2207 | void Migrator::handle_export_cancel(MExportDirCancel *m) | |
2208 | { | |
2209 | dout(7) << "handle_export_cancel on " << m->get_dirfrag() << dendl; | |
2210 | dirfrag_t df = m->get_dirfrag(); | |
2211 | map<dirfrag_t,import_state_t>::iterator it = import_state.find(df); | |
2212 | if (it == import_state.end()) { | |
2213 | assert(0 == "got export_cancel in weird state"); | |
2214 | } else if (it->second.state == IMPORT_DISCOVERING) { | |
2215 | import_reverse_discovering(df); | |
2216 | } else if (it->second.state == IMPORT_DISCOVERED) { | |
2217 | CInode *in = cache->get_inode(df.ino); | |
2218 | assert(in); | |
2219 | import_reverse_discovered(df, in); | |
2220 | } else if (it->second.state == IMPORT_PREPPING) { | |
2221 | CDir *dir = mds->mdcache->get_dirfrag(df); | |
2222 | assert(dir); | |
b32b8144 | 2223 | import_reverse_prepping(dir, it->second); |
7c673cae FG |
2224 | } else if (it->second.state == IMPORT_PREPPED) { |
2225 | CDir *dir = mds->mdcache->get_dirfrag(df); | |
2226 | assert(dir); | |
2227 | set<CDir*> bounds; | |
2228 | cache->get_subtree_bounds(dir, bounds); | |
2229 | import_remove_pins(dir, bounds); | |
2230 | // adjust auth back to the exportor | |
2231 | cache->adjust_subtree_auth(dir, it->second.peer); | |
7c673cae FG |
2232 | import_reverse_unfreeze(dir); |
2233 | } else { | |
2234 | assert(0 == "got export_cancel in weird state"); | |
2235 | } | |
2236 | m->put(); | |
2237 | } | |
2238 | ||
2239 | /* This function DOES put the passed message before returning*/ | |
2240 | void Migrator::handle_export_prep(MExportDirPrep *m) | |
2241 | { | |
2242 | mds_rank_t oldauth = mds_rank_t(m->get_source().num()); | |
2243 | assert(oldauth != mds->get_nodeid()); | |
2244 | ||
2245 | CDir *dir; | |
2246 | CInode *diri; | |
2247 | list<MDSInternalContextBase*> finished; | |
2248 | ||
2249 | // assimilate root dir. | |
2250 | map<dirfrag_t,import_state_t>::iterator it = import_state.find(m->get_dirfrag()); | |
2251 | if (!m->did_assim()) { | |
2252 | assert(it != import_state.end()); | |
2253 | assert(it->second.state == IMPORT_DISCOVERED); | |
31f18b77 | 2254 | assert(it->second.peer == oldauth); |
7c673cae FG |
2255 | diri = cache->get_inode(m->get_dirfrag().ino); |
2256 | assert(diri); | |
2257 | bufferlist::iterator p = m->basedir.begin(); | |
2258 | dir = cache->add_replica_dir(p, diri, oldauth, finished); | |
2259 | dout(7) << "handle_export_prep on " << *dir << " (first pass)" << dendl; | |
2260 | } else { | |
2261 | if (it == import_state.end() || | |
2262 | it->second.peer != oldauth || | |
2263 | it->second.tid != m->get_tid()) { | |
2264 | dout(7) << "handle_export_prep obsolete message, dropping" << dendl; | |
2265 | m->put(); | |
2266 | return; | |
2267 | } | |
2268 | assert(it->second.state == IMPORT_PREPPING); | |
31f18b77 | 2269 | assert(it->second.peer == oldauth); |
7c673cae FG |
2270 | |
2271 | dir = cache->get_dirfrag(m->get_dirfrag()); | |
2272 | assert(dir); | |
2273 | dout(7) << "handle_export_prep on " << *dir << " (subsequent pass)" << dendl; | |
2274 | diri = dir->get_inode(); | |
2275 | } | |
2276 | assert(dir->is_auth() == false); | |
2277 | ||
2278 | cache->show_subtrees(); | |
2279 | ||
2280 | // build import bound map | |
2281 | map<inodeno_t, fragset_t> import_bound_fragset; | |
2282 | for (list<dirfrag_t>::iterator p = m->get_bounds().begin(); | |
2283 | p != m->get_bounds().end(); | |
2284 | ++p) { | |
2285 | dout(10) << " bound " << *p << dendl; | |
2286 | import_bound_fragset[p->ino].insert(p->frag); | |
2287 | } | |
2288 | ||
2289 | // assimilate contents? | |
2290 | if (!m->did_assim()) { | |
2291 | dout(7) << "doing assim on " << *dir << dendl; | |
2292 | m->mark_assim(); // only do this the first time! | |
2293 | ||
2294 | // change import state | |
2295 | it->second.state = IMPORT_PREPPING; | |
2296 | it->second.bound_ls = m->get_bounds(); | |
2297 | it->second.bystanders = m->get_bystanders(); | |
2298 | assert(g_conf->mds_kill_import_at != 3); | |
2299 | ||
2300 | // bystander list | |
2301 | dout(7) << "bystanders are " << it->second.bystanders << dendl; | |
2302 | ||
2303 | // move pin to dir | |
2304 | diri->put(CInode::PIN_IMPORTING); | |
2305 | dir->get(CDir::PIN_IMPORTING); | |
2306 | dir->state_set(CDir::STATE_IMPORTING); | |
2307 | ||
2308 | // assimilate traces to exports | |
2309 | // each trace is: df ('-' | ('f' dir | 'd') dentry inode (dir dentry inode)*) | |
2310 | for (list<bufferlist>::iterator p = m->traces.begin(); | |
2311 | p != m->traces.end(); | |
2312 | ++p) { | |
2313 | bufferlist::iterator q = p->begin(); | |
2314 | dirfrag_t df; | |
2315 | ::decode(df, q); | |
2316 | char start; | |
2317 | ::decode(start, q); | |
2318 | dout(10) << " trace from " << df << " start " << start << " len " << p->length() << dendl; | |
2319 | ||
2320 | CDir *cur = 0; | |
2321 | if (start == 'd') { | |
2322 | cur = cache->get_dirfrag(df); | |
2323 | assert(cur); | |
2324 | dout(10) << " had " << *cur << dendl; | |
2325 | } else if (start == 'f') { | |
2326 | CInode *in = cache->get_inode(df.ino); | |
2327 | assert(in); | |
2328 | dout(10) << " had " << *in << dendl; | |
2329 | cur = cache->add_replica_dir(q, in, oldauth, finished); | |
2330 | dout(10) << " added " << *cur << dendl; | |
2331 | } else if (start == '-') { | |
2332 | // nothing | |
2333 | } else | |
2334 | assert(0 == "unrecognized start char"); | |
2335 | ||
b32b8144 | 2336 | while (!q.end()) { |
7c673cae FG |
2337 | CDentry *dn = cache->add_replica_dentry(q, cur, finished); |
2338 | dout(10) << " added " << *dn << dendl; | |
2339 | CInode *in = cache->add_replica_inode(q, dn, finished); | |
2340 | dout(10) << " added " << *in << dendl; | |
2341 | if (q.end()) | |
2342 | break; | |
2343 | cur = cache->add_replica_dir(q, in, oldauth, finished); | |
2344 | dout(10) << " added " << *cur << dendl; | |
2345 | } | |
2346 | } | |
2347 | ||
2348 | // make bound sticky | |
2349 | for (map<inodeno_t,fragset_t>::iterator p = import_bound_fragset.begin(); | |
2350 | p != import_bound_fragset.end(); | |
2351 | ++p) { | |
2352 | CInode *in = cache->get_inode(p->first); | |
2353 | assert(in); | |
2354 | in->get_stickydirs(); | |
2355 | dout(7) << " set stickydirs on bound inode " << *in << dendl; | |
2356 | } | |
2357 | ||
2358 | } else { | |
2359 | dout(7) << " not doing assim on " << *dir << dendl; | |
2360 | } | |
2361 | ||
2362 | if (!finished.empty()) | |
2363 | mds->queue_waiters(finished); | |
2364 | ||
2365 | ||
c07f9fc5 FG |
2366 | bool success = true; |
2367 | if (mds->is_active()) { | |
2368 | // open all bounds | |
2369 | set<CDir*> import_bounds; | |
2370 | for (map<inodeno_t,fragset_t>::iterator p = import_bound_fragset.begin(); | |
2371 | p != import_bound_fragset.end(); | |
2372 | ++p) { | |
2373 | CInode *in = cache->get_inode(p->first); | |
2374 | assert(in); | |
7c673cae | 2375 | |
c07f9fc5 FG |
2376 | // map fragset into a frag_t list, based on the inode fragtree |
2377 | list<frag_t> fglist; | |
2378 | for (set<frag_t>::iterator q = p->second.begin(); q != p->second.end(); ++q) | |
2379 | in->dirfragtree.get_leaves_under(*q, fglist); | |
2380 | dout(10) << " bound inode " << p->first << " fragset " << p->second << " maps to " << fglist << dendl; | |
2381 | ||
2382 | for (list<frag_t>::iterator q = fglist.begin(); | |
2383 | q != fglist.end(); | |
2384 | ++q) { | |
2385 | CDir *bound = cache->get_dirfrag(dirfrag_t(p->first, *q)); | |
2386 | if (!bound) { | |
2387 | dout(7) << " opening bounding dirfrag " << *q << " on " << *in << dendl; | |
2388 | cache->open_remote_dirfrag(in, *q, | |
2389 | new C_MDS_RetryMessage(mds, m)); | |
2390 | return; | |
2391 | } | |
7c673cae | 2392 | |
c07f9fc5 FG |
2393 | if (!bound->state_test(CDir::STATE_IMPORTBOUND)) { |
2394 | dout(7) << " pinning import bound " << *bound << dendl; | |
2395 | bound->get(CDir::PIN_IMPORTBOUND); | |
2396 | bound->state_set(CDir::STATE_IMPORTBOUND); | |
2397 | } else { | |
2398 | dout(7) << " already pinned import bound " << *bound << dendl; | |
2399 | } | |
2400 | import_bounds.insert(bound); | |
7c673cae | 2401 | } |
7c673cae | 2402 | } |
7c673cae | 2403 | |
c07f9fc5 FG |
2404 | dout(7) << " all ready, noting auth and freezing import region" << dendl; |
2405 | ||
2406 | if (!mds->mdcache->is_readonly() && | |
2407 | dir->get_inode()->filelock.can_wrlock(-1) && | |
2408 | dir->get_inode()->nestlock.can_wrlock(-1)) { | |
2409 | it->second.mut = new MutationImpl(); | |
2410 | // force some locks. hacky. | |
2411 | mds->locker->wrlock_force(&dir->inode->filelock, it->second.mut); | |
2412 | mds->locker->wrlock_force(&dir->inode->nestlock, it->second.mut); | |
2413 | ||
2414 | // note that i am an ambiguous auth for this subtree. | |
2415 | // specify bounds, since the exporter explicitly defines the region. | |
2416 | cache->adjust_bounded_subtree_auth(dir, import_bounds, | |
2417 | pair<int,int>(oldauth, mds->get_nodeid())); | |
2418 | cache->verify_subtree_bounds(dir, import_bounds); | |
2419 | // freeze. | |
2420 | dir->_freeze_tree(); | |
2421 | // note new state | |
2422 | it->second.state = IMPORT_PREPPED; | |
2423 | } else { | |
2424 | dout(7) << " couldn't acquire all needed locks, failing. " << *dir << dendl; | |
2425 | success = false; | |
2426 | } | |
7c673cae | 2427 | } else { |
c07f9fc5 | 2428 | dout(7) << " not active, failing. " << *dir << dendl; |
7c673cae | 2429 | success = false; |
7c673cae FG |
2430 | } |
2431 | ||
c07f9fc5 | 2432 | if (!success) |
b32b8144 | 2433 | import_reverse_prepping(dir, it->second); |
c07f9fc5 | 2434 | |
7c673cae FG |
2435 | // ok! |
2436 | dout(7) << " sending export_prep_ack on " << *dir << dendl; | |
2437 | mds->send_message(new MExportDirPrepAck(dir->dirfrag(), success, m->get_tid()), m->get_connection()); | |
2438 | ||
2439 | assert(g_conf->mds_kill_import_at != 4); | |
2440 | // done | |
2441 | m->put(); | |
2442 | } | |
2443 | ||
2444 | ||
2445 | ||
2446 | ||
2447 | class C_MDS_ImportDirLoggedStart : public MigratorLogContext { | |
2448 | dirfrag_t df; | |
2449 | CDir *dir; | |
2450 | mds_rank_t from; | |
2451 | public: | |
28e407b8 | 2452 | map<client_t,pair<Session*,uint64_t> > imported_session_map; |
7c673cae FG |
2453 | |
2454 | C_MDS_ImportDirLoggedStart(Migrator *m, CDir *d, mds_rank_t f) : | |
2455 | MigratorLogContext(m), df(d->dirfrag()), dir(d), from(f) { | |
2456 | } | |
2457 | void finish(int r) override { | |
28e407b8 | 2458 | mig->import_logged_start(df, dir, from, imported_session_map); |
7c673cae FG |
2459 | } |
2460 | }; | |
2461 | ||
2462 | /* This function DOES put the passed message before returning*/ | |
2463 | void Migrator::handle_export_dir(MExportDir *m) | |
2464 | { | |
2465 | assert (g_conf->mds_kill_import_at != 5); | |
2466 | CDir *dir = cache->get_dirfrag(m->dirfrag); | |
2467 | assert(dir); | |
31f18b77 FG |
2468 | |
2469 | mds_rank_t oldauth = mds_rank_t(m->get_source().num()); | |
2470 | dout(7) << "handle_export_dir importing " << *dir << " from " << oldauth << dendl; | |
2471 | ||
2472 | assert(!dir->is_auth()); | |
7c673cae FG |
2473 | |
2474 | map<dirfrag_t,import_state_t>::iterator it = import_state.find(m->dirfrag); | |
2475 | assert(it != import_state.end()); | |
2476 | assert(it->second.state == IMPORT_PREPPED); | |
2477 | assert(it->second.tid == m->get_tid()); | |
31f18b77 | 2478 | assert(it->second.peer == oldauth); |
7c673cae FG |
2479 | |
2480 | utime_t now = ceph_clock_now(); | |
7c673cae FG |
2481 | |
2482 | if (!dir->get_inode()->dirfragtree.is_leaf(dir->get_frag())) | |
2483 | dir->get_inode()->dirfragtree.force_to_leaf(g_ceph_context, dir->get_frag()); | |
2484 | ||
2485 | cache->show_subtrees(); | |
2486 | ||
31f18b77 | 2487 | C_MDS_ImportDirLoggedStart *onlogged = new C_MDS_ImportDirLoggedStart(this, dir, oldauth); |
7c673cae FG |
2488 | |
2489 | // start the journal entry | |
31f18b77 | 2490 | EImportStart *le = new EImportStart(mds->mdlog, dir->dirfrag(), m->bounds, oldauth); |
7c673cae FG |
2491 | mds->mdlog->start_entry(le); |
2492 | ||
2493 | le->metablob.add_dir_context(dir); | |
2494 | ||
2495 | // adjust auth (list us _first_) | |
2496 | cache->adjust_subtree_auth(dir, mds->get_nodeid(), oldauth); | |
2497 | ||
2498 | // new client sessions, open these after we journal | |
2499 | // include imported sessions in EImportStart | |
2500 | bufferlist::iterator cmp = m->client_map.begin(); | |
28e407b8 AA |
2501 | map<client_t,entity_inst_t> client_map; |
2502 | decode(client_map, cmp); | |
7c673cae | 2503 | assert(cmp.end()); |
28e407b8 AA |
2504 | le->cmapv = mds->server->prepare_force_open_sessions(client_map, onlogged->imported_session_map); |
2505 | encode(client_map, le->client_map, mds->mdsmap->get_up_features()); | |
7c673cae FG |
2506 | |
2507 | bufferlist::iterator blp = m->export_data.begin(); | |
2508 | int num_imported_inodes = 0; | |
2509 | while (!blp.end()) { | |
2510 | num_imported_inodes += | |
2511 | decode_import_dir(blp, | |
2512 | oldauth, | |
2513 | dir, // import root | |
2514 | le, | |
2515 | mds->mdlog->get_current_segment(), | |
2516 | it->second.peer_exports, | |
2517 | it->second.updated_scatterlocks, | |
2518 | now); | |
2519 | } | |
2520 | dout(10) << " " << m->bounds.size() << " imported bounds" << dendl; | |
2521 | ||
2522 | // include bounds in EImportStart | |
2523 | set<CDir*> import_bounds; | |
2524 | for (vector<dirfrag_t>::iterator p = m->bounds.begin(); | |
2525 | p != m->bounds.end(); | |
2526 | ++p) { | |
2527 | CDir *bd = cache->get_dirfrag(*p); | |
2528 | assert(bd); | |
2529 | le->metablob.add_dir(bd, false); // note that parent metadata is already in the event | |
2530 | import_bounds.insert(bd); | |
2531 | } | |
2532 | cache->verify_subtree_bounds(dir, import_bounds); | |
2533 | ||
2534 | // adjust popularity | |
2535 | mds->balancer->add_import(dir, now); | |
2536 | ||
2537 | dout(7) << "handle_export_dir did " << *dir << dendl; | |
2538 | ||
2539 | // note state | |
2540 | it->second.state = IMPORT_LOGGINGSTART; | |
2541 | assert (g_conf->mds_kill_import_at != 6); | |
2542 | ||
2543 | // log it | |
2544 | mds->mdlog->submit_entry(le, onlogged); | |
2545 | mds->mdlog->flush(); | |
2546 | ||
2547 | // some stats | |
2548 | if (mds->logger) { | |
2549 | mds->logger->inc(l_mds_imported); | |
2550 | mds->logger->inc(l_mds_imported_inodes, num_imported_inodes); | |
2551 | } | |
2552 | ||
2553 | m->put(); | |
2554 | } | |
2555 | ||
2556 | ||
2557 | /* | |
2558 | * this is an import helper | |
2559 | * called by import_finish, and import_reverse and friends. | |
2560 | */ | |
2561 | void Migrator::import_remove_pins(CDir *dir, set<CDir*>& bounds) | |
2562 | { | |
2563 | import_state_t& stat = import_state[dir->dirfrag()]; | |
2564 | // root | |
2565 | dir->put(CDir::PIN_IMPORTING); | |
2566 | dir->state_clear(CDir::STATE_IMPORTING); | |
2567 | ||
2568 | // bounding inodes | |
2569 | set<inodeno_t> did; | |
2570 | for (list<dirfrag_t>::iterator p = stat.bound_ls.begin(); | |
2571 | p != stat.bound_ls.end(); | |
2572 | ++p) { | |
2573 | if (did.count(p->ino)) | |
2574 | continue; | |
2575 | did.insert(p->ino); | |
2576 | CInode *in = cache->get_inode(p->ino); | |
2577 | assert(in); | |
2578 | in->put_stickydirs(); | |
2579 | } | |
2580 | ||
2581 | if (stat.state == IMPORT_PREPPING) { | |
2582 | for (auto bd : bounds) { | |
2583 | if (bd->state_test(CDir::STATE_IMPORTBOUND)) { | |
2584 | bd->put(CDir::PIN_IMPORTBOUND); | |
2585 | bd->state_clear(CDir::STATE_IMPORTBOUND); | |
2586 | } | |
2587 | } | |
2588 | } else if (stat.state >= IMPORT_PREPPED) { | |
2589 | // bounding dirfrags | |
2590 | for (auto bd : bounds) { | |
2591 | assert(bd->state_test(CDir::STATE_IMPORTBOUND)); | |
2592 | bd->put(CDir::PIN_IMPORTBOUND); | |
2593 | bd->state_clear(CDir::STATE_IMPORTBOUND); | |
2594 | } | |
2595 | } | |
2596 | } | |
2597 | ||
2598 | ||
2599 | /* | |
2600 | * note: this does teh full work of reversing and import and cleaning up | |
2601 | * state. | |
2602 | * called by both handle_mds_failure and by handle_resolve (if we are | |
2603 | * a survivor coping with an exporter failure+recovery). | |
2604 | */ | |
2605 | void Migrator::import_reverse(CDir *dir) | |
2606 | { | |
2607 | dout(7) << "import_reverse " << *dir << dendl; | |
2608 | ||
2609 | import_state_t& stat = import_state[dir->dirfrag()]; | |
2610 | stat.state = IMPORT_ABORTING; | |
2611 | ||
2612 | set<CDir*> bounds; | |
2613 | cache->get_subtree_bounds(dir, bounds); | |
2614 | ||
2615 | // remove pins | |
2616 | import_remove_pins(dir, bounds); | |
2617 | ||
2618 | // update auth, with possible subtree merge. | |
2619 | assert(dir->is_subtree_root()); | |
2620 | if (mds->is_resolve()) | |
2621 | cache->trim_non_auth_subtree(dir); | |
2622 | ||
2623 | cache->adjust_subtree_auth(dir, stat.peer); | |
2624 | ||
2625 | C_ContextsBase<MDSInternalContextBase, MDSInternalContextGather> *fin = new C_ContextsBase<MDSInternalContextBase, MDSInternalContextGather>(g_ceph_context); | |
2626 | if (!dir->get_inode()->is_auth() && | |
2627 | !dir->get_inode()->has_subtree_root_dirfrag(mds->get_nodeid())) { | |
2628 | dir->get_inode()->clear_scatter_dirty(); | |
2629 | // wake up scatter_nudge waiters | |
2630 | dir->get_inode()->take_waiting(CInode::WAIT_ANY_MASK, fin->contexts); | |
2631 | } | |
2632 | ||
2633 | int num_dentries = 0; | |
2634 | // adjust auth bits. | |
2635 | list<CDir*> q; | |
2636 | q.push_back(dir); | |
2637 | while (!q.empty()) { | |
2638 | CDir *cur = q.front(); | |
2639 | q.pop_front(); | |
2640 | ||
2641 | // dir | |
2642 | assert(cur->is_auth()); | |
2643 | cur->state_clear(CDir::STATE_AUTH); | |
2644 | cur->remove_bloom(); | |
2645 | cur->clear_replica_map(); | |
2646 | cur->set_replica_nonce(CDir::EXPORT_NONCE); | |
2647 | if (cur->is_dirty()) | |
2648 | cur->mark_clean(); | |
2649 | ||
94b18763 FG |
2650 | for (auto &p : *cur) { |
2651 | CDentry *dn = p.second; | |
7c673cae FG |
2652 | |
2653 | // dentry | |
2654 | dn->state_clear(CDentry::STATE_AUTH); | |
2655 | dn->clear_replica_map(); | |
2656 | dn->set_replica_nonce(CDentry::EXPORT_NONCE); | |
2657 | if (dn->is_dirty()) | |
2658 | dn->mark_clean(); | |
2659 | ||
2660 | // inode? | |
2661 | if (dn->get_linkage()->is_primary()) { | |
2662 | CInode *in = dn->get_linkage()->get_inode(); | |
2663 | in->state_clear(CDentry::STATE_AUTH); | |
2664 | in->clear_replica_map(); | |
2665 | in->set_replica_nonce(CInode::EXPORT_NONCE); | |
2666 | if (in->is_dirty()) | |
2667 | in->mark_clean(); | |
2668 | in->clear_dirty_rstat(); | |
2669 | if (!in->has_subtree_root_dirfrag(mds->get_nodeid())) { | |
2670 | in->clear_scatter_dirty(); | |
2671 | in->take_waiting(CInode::WAIT_ANY_MASK, fin->contexts); | |
2672 | } | |
2673 | ||
2674 | in->clear_dirty_parent(); | |
2675 | ||
2676 | in->authlock.clear_gather(); | |
2677 | in->linklock.clear_gather(); | |
2678 | in->dirfragtreelock.clear_gather(); | |
2679 | in->filelock.clear_gather(); | |
2680 | ||
2681 | in->clear_file_locks(); | |
2682 | ||
2683 | // non-bounding dir? | |
2684 | list<CDir*> dfs; | |
2685 | in->get_dirfrags(dfs); | |
2686 | for (list<CDir*>::iterator p = dfs.begin(); p != dfs.end(); ++p) | |
2687 | if (bounds.count(*p) == 0) | |
2688 | q.push_back(*p); | |
2689 | } | |
2690 | ||
2691 | cache->touch_dentry_bottom(dn); // move dentry to tail of LRU | |
2692 | ++num_dentries; | |
2693 | } | |
2694 | } | |
2695 | ||
2696 | dir->add_waiter(CDir::WAIT_UNFREEZE, fin); | |
2697 | ||
2698 | if (stat.state == IMPORT_ACKING) { | |
2699 | // remove imported caps | |
2700 | for (map<CInode*,map<client_t,Capability::Export> >::iterator p = stat.peer_exports.begin(); | |
28e407b8 AA |
2701 | p != stat.peer_exports.end(); |
2702 | ++p) { | |
7c673cae FG |
2703 | CInode *in = p->first; |
2704 | for (map<client_t,Capability::Export>::iterator q = p->second.begin(); | |
28e407b8 AA |
2705 | q != p->second.end(); |
2706 | ++q) { | |
7c673cae | 2707 | Capability *cap = in->get_client_cap(q->first); |
28e407b8 AA |
2708 | if (!cap) { |
2709 | assert(!stat.session_map.count(q->first)); | |
2710 | continue; | |
2711 | } | |
7c673cae FG |
2712 | if (cap->is_importing()) |
2713 | in->remove_client_cap(q->first); | |
2714 | } | |
2715 | in->put(CInode::PIN_IMPORTINGCAPS); | |
2716 | } | |
28e407b8 AA |
2717 | for (auto& p : stat.session_map) { |
2718 | Session *session = p.second.first; | |
7c673cae FG |
2719 | session->dec_importing(); |
2720 | } | |
2721 | } | |
2722 | ||
2723 | // log our failure | |
2724 | mds->mdlog->start_submit_entry(new EImportFinish(dir, false)); // log failure | |
2725 | ||
181888fb | 2726 | cache->trim(num_dentries); // try trimming dentries |
7c673cae FG |
2727 | |
2728 | // notify bystanders; wait in aborting state | |
2729 | import_notify_abort(dir, bounds); | |
2730 | } | |
2731 | ||
2732 | void Migrator::import_notify_finish(CDir *dir, set<CDir*>& bounds) | |
2733 | { | |
2734 | dout(7) << "import_notify_finish " << *dir << dendl; | |
2735 | ||
2736 | import_state_t& stat = import_state[dir->dirfrag()]; | |
2737 | for (set<mds_rank_t>::iterator p = stat.bystanders.begin(); | |
2738 | p != stat.bystanders.end(); | |
2739 | ++p) { | |
2740 | MExportDirNotify *notify = | |
2741 | new MExportDirNotify(dir->dirfrag(), stat.tid, false, | |
2742 | pair<int,int>(stat.peer, mds->get_nodeid()), | |
2743 | pair<int,int>(mds->get_nodeid(), CDIR_AUTH_UNKNOWN)); | |
2744 | for (set<CDir*>::iterator i = bounds.begin(); i != bounds.end(); ++i) | |
2745 | notify->get_bounds().push_back((*i)->dirfrag()); | |
2746 | mds->send_message_mds(notify, *p); | |
2747 | } | |
2748 | } | |
2749 | ||
2750 | void Migrator::import_notify_abort(CDir *dir, set<CDir*>& bounds) | |
2751 | { | |
2752 | dout(7) << "import_notify_abort " << *dir << dendl; | |
2753 | ||
2754 | import_state_t& stat = import_state[dir->dirfrag()]; | |
2755 | for (set<mds_rank_t>::iterator p = stat.bystanders.begin(); | |
2756 | p != stat.bystanders.end(); ) { | |
2757 | if (mds->is_cluster_degraded() && | |
2758 | !mds->mdsmap->is_clientreplay_or_active_or_stopping(*p)) { | |
2759 | // this can happen if both exporter and bystander fail in the same mdsmap epoch | |
2760 | stat.bystanders.erase(p++); | |
2761 | continue; | |
2762 | } | |
2763 | MExportDirNotify *notify = | |
2764 | new MExportDirNotify(dir->dirfrag(), stat.tid, true, | |
2765 | mds_authority_t(stat.peer, mds->get_nodeid()), | |
2766 | mds_authority_t(stat.peer, CDIR_AUTH_UNKNOWN)); | |
2767 | for (set<CDir*>::iterator i = bounds.begin(); i != bounds.end(); ++i) | |
2768 | notify->get_bounds().push_back((*i)->dirfrag()); | |
2769 | mds->send_message_mds(notify, *p); | |
2770 | ++p; | |
2771 | } | |
2772 | if (stat.bystanders.empty()) { | |
2773 | dout(7) << "no bystanders, finishing reverse now" << dendl; | |
2774 | import_reverse_unfreeze(dir); | |
2775 | } else { | |
2776 | assert (g_conf->mds_kill_import_at != 10); | |
2777 | } | |
2778 | } | |
2779 | ||
2780 | void Migrator::import_reverse_unfreeze(CDir *dir) | |
2781 | { | |
7c673cae | 2782 | dout(7) << "import_reverse_unfreeze " << *dir << dendl; |
224ce89b | 2783 | assert(!dir->is_auth()); |
7c673cae | 2784 | cache->discard_delayed_expire(dir); |
224ce89b WB |
2785 | dir->unfreeze_tree(); |
2786 | if (dir->is_subtree_root()) | |
2787 | cache->try_subtree_merge(dir); | |
7c673cae FG |
2788 | import_reverse_final(dir); |
2789 | } | |
2790 | ||
2791 | void Migrator::import_reverse_final(CDir *dir) | |
2792 | { | |
2793 | dout(7) << "import_reverse_final " << *dir << dendl; | |
2794 | ||
2795 | // clean up | |
2796 | map<dirfrag_t, import_state_t>::iterator it = import_state.find(dir->dirfrag()); | |
2797 | assert(it != import_state.end()); | |
2798 | ||
2799 | MutationRef mut = it->second.mut; | |
2800 | import_state.erase(it); | |
2801 | ||
2802 | // send pending import_maps? | |
2803 | mds->mdcache->maybe_send_pending_resolves(); | |
2804 | ||
2805 | if (mut) { | |
2806 | mds->locker->drop_locks(mut.get()); | |
2807 | mut->cleanup(); | |
2808 | } | |
2809 | ||
2810 | cache->show_subtrees(); | |
2811 | //audit(); // this fails, bc we munge up the subtree map during handle_import_map (resolve phase) | |
2812 | } | |
2813 | ||
2814 | ||
2815 | ||
2816 | ||
2817 | void Migrator::import_logged_start(dirfrag_t df, CDir *dir, mds_rank_t from, | |
28e407b8 | 2818 | map<client_t,pair<Session*,uint64_t> >& imported_session_map) |
7c673cae FG |
2819 | { |
2820 | map<dirfrag_t, import_state_t>::iterator it = import_state.find(dir->dirfrag()); | |
2821 | if (it == import_state.end() || | |
2822 | it->second.state != IMPORT_LOGGINGSTART) { | |
2823 | dout(7) << "import " << df << " must have aborted" << dendl; | |
28e407b8 | 2824 | mds->server->finish_force_open_sessions(imported_session_map); |
7c673cae FG |
2825 | return; |
2826 | } | |
2827 | ||
2828 | dout(7) << "import_logged " << *dir << dendl; | |
2829 | ||
2830 | // note state | |
2831 | it->second.state = IMPORT_ACKING; | |
2832 | ||
2833 | assert (g_conf->mds_kill_import_at != 7); | |
2834 | ||
2835 | // force open client sessions and finish cap import | |
28e407b8 | 2836 | mds->server->finish_force_open_sessions(imported_session_map, false); |
7c673cae FG |
2837 | |
2838 | map<inodeno_t,map<client_t,Capability::Import> > imported_caps; | |
2839 | for (map<CInode*, map<client_t,Capability::Export> >::iterator p = it->second.peer_exports.begin(); | |
2840 | p != it->second.peer_exports.end(); | |
2841 | ++p) { | |
2842 | // parameter 'peer' is NONE, delay sending cap import messages to client | |
28e407b8 AA |
2843 | finish_import_inode_caps(p->first, MDS_RANK_NONE, true, imported_session_map, |
2844 | p->second, imported_caps[p->first->ino()]); | |
7c673cae | 2845 | } |
28e407b8 AA |
2846 | |
2847 | it->second.session_map.swap(imported_session_map); | |
7c673cae FG |
2848 | |
2849 | // send notify's etc. | |
2850 | dout(7) << "sending ack for " << *dir << " to old auth mds." << from << dendl; | |
2851 | ||
2852 | // test surviving observer of a failed migration that did not complete | |
2853 | //assert(dir->replica_map.size() < 2 || mds->get_nodeid() != 0); | |
2854 | ||
2855 | MExportDirAck *ack = new MExportDirAck(dir->dirfrag(), it->second.tid); | |
2856 | ::encode(imported_caps, ack->imported_caps); | |
2857 | ||
2858 | mds->send_message_mds(ack, from); | |
2859 | assert (g_conf->mds_kill_import_at != 8); | |
2860 | ||
2861 | cache->show_subtrees(); | |
2862 | } | |
2863 | ||
2864 | /* This function DOES put the passed message before returning*/ | |
2865 | void Migrator::handle_export_finish(MExportDirFinish *m) | |
2866 | { | |
2867 | CDir *dir = cache->get_dirfrag(m->get_dirfrag()); | |
2868 | assert(dir); | |
2869 | dout(7) << "handle_export_finish on " << *dir << (m->is_last() ? " last" : "") << dendl; | |
2870 | ||
2871 | map<dirfrag_t,import_state_t>::iterator it = import_state.find(m->get_dirfrag()); | |
2872 | assert(it != import_state.end()); | |
2873 | assert(it->second.tid == m->get_tid()); | |
2874 | ||
2875 | import_finish(dir, false, m->is_last()); | |
2876 | ||
2877 | m->put(); | |
2878 | } | |
2879 | ||
2880 | void Migrator::import_finish(CDir *dir, bool notify, bool last) | |
2881 | { | |
2882 | dout(7) << "import_finish on " << *dir << dendl; | |
2883 | ||
2884 | map<dirfrag_t,import_state_t>::iterator it = import_state.find(dir->dirfrag()); | |
2885 | assert(it != import_state.end()); | |
2886 | assert(it->second.state == IMPORT_ACKING || it->second.state == IMPORT_FINISHING); | |
2887 | ||
224ce89b WB |
2888 | if (it->second.state == IMPORT_ACKING) { |
2889 | assert(dir->is_auth()); | |
2890 | cache->adjust_subtree_auth(dir, mds->get_nodeid(), mds->get_nodeid()); | |
2891 | } | |
2892 | ||
7c673cae FG |
2893 | // log finish |
2894 | assert(g_conf->mds_kill_import_at != 9); | |
2895 | ||
2896 | if (it->second.state == IMPORT_ACKING) { | |
2897 | for (map<CInode*, map<client_t,Capability::Export> >::iterator p = it->second.peer_exports.begin(); | |
2898 | p != it->second.peer_exports.end(); | |
2899 | ++p) { | |
2900 | CInode *in = p->first; | |
2901 | assert(in->is_auth()); | |
2902 | for (map<client_t,Capability::Export>::iterator q = p->second.begin(); | |
2903 | q != p->second.end(); | |
2904 | ++q) { | |
28e407b8 AA |
2905 | auto r = it->second.session_map.find(q->first); |
2906 | if (r == it->second.session_map.end()) | |
2907 | continue; | |
2908 | ||
2909 | Session *session = r->second.first; | |
7c673cae FG |
2910 | Capability *cap = in->get_client_cap(q->first); |
2911 | assert(cap); | |
2912 | cap->merge(q->second, true); | |
2913 | cap->clear_importing(); | |
2914 | mds->mdcache->do_cap_import(session, in, cap, q->second.cap_id, q->second.seq, | |
2915 | q->second.mseq - 1, it->second.peer, CEPH_CAP_FLAG_AUTH); | |
2916 | } | |
2917 | p->second.clear(); | |
2918 | in->replica_caps_wanted = 0; | |
2919 | } | |
28e407b8 AA |
2920 | for (auto& p : it->second.session_map) { |
2921 | Session *session = p.second.first; | |
7c673cae FG |
2922 | session->dec_importing(); |
2923 | } | |
2924 | } | |
2925 | ||
2926 | if (!last) { | |
2927 | assert(it->second.state == IMPORT_ACKING); | |
2928 | it->second.state = IMPORT_FINISHING; | |
2929 | return; | |
2930 | } | |
2931 | ||
2932 | // remove pins | |
2933 | set<CDir*> bounds; | |
2934 | cache->get_subtree_bounds(dir, bounds); | |
2935 | ||
2936 | if (notify) | |
2937 | import_notify_finish(dir, bounds); | |
2938 | ||
2939 | import_remove_pins(dir, bounds); | |
2940 | ||
2941 | map<CInode*, map<client_t,Capability::Export> > peer_exports; | |
2942 | it->second.peer_exports.swap(peer_exports); | |
2943 | ||
2944 | // clear import state (we're done!) | |
2945 | MutationRef mut = it->second.mut; | |
2946 | import_state.erase(it); | |
2947 | ||
7c673cae FG |
2948 | mds->mdlog->start_submit_entry(new EImportFinish(dir, true)); |
2949 | ||
7c673cae FG |
2950 | // process delayed expires |
2951 | cache->process_delayed_expire(dir); | |
2952 | ||
224ce89b | 2953 | // unfreeze tree, with possible subtree merge. |
7c673cae | 2954 | dir->unfreeze_tree(); |
224ce89b WB |
2955 | cache->try_subtree_merge(dir); |
2956 | ||
7c673cae FG |
2957 | cache->show_subtrees(); |
2958 | //audit(); // this fails, bc we munge up the subtree map during handle_import_map (resolve phase) | |
2959 | ||
2960 | if (mut) { | |
2961 | mds->locker->drop_locks(mut.get()); | |
2962 | mut->cleanup(); | |
2963 | } | |
2964 | ||
2965 | // re-eval imported caps | |
2966 | for (map<CInode*, map<client_t,Capability::Export> >::iterator p = peer_exports.begin(); | |
2967 | p != peer_exports.end(); | |
2968 | ++p) { | |
2969 | if (p->first->is_auth()) | |
2970 | mds->locker->eval(p->first, CEPH_CAP_LOCKS, true); | |
2971 | p->first->put(CInode::PIN_IMPORTINGCAPS); | |
2972 | } | |
2973 | ||
2974 | // send pending import_maps? | |
2975 | mds->mdcache->maybe_send_pending_resolves(); | |
2976 | ||
2977 | // did i just import mydir? | |
2978 | if (dir->ino() == MDS_INO_MDSDIR(mds->get_nodeid())) | |
2979 | cache->populate_mydir(); | |
2980 | ||
2981 | // is it empty? | |
2982 | if (dir->get_num_head_items() == 0 && | |
2983 | !dir->inode->is_auth()) { | |
2984 | // reexport! | |
2985 | export_empty_import(dir); | |
2986 | } | |
2987 | } | |
2988 | ||
2989 | ||
2990 | void Migrator::decode_import_inode(CDentry *dn, bufferlist::iterator& blp, | |
2991 | mds_rank_t oldauth, LogSegment *ls, | |
2992 | map<CInode*, map<client_t,Capability::Export> >& peer_exports, | |
2993 | list<ScatterLock*>& updated_scatterlocks) | |
2994 | { | |
2995 | dout(15) << "decode_import_inode on " << *dn << dendl; | |
2996 | ||
2997 | inodeno_t ino; | |
2998 | snapid_t last; | |
2999 | ::decode(ino, blp); | |
3000 | ::decode(last, blp); | |
3001 | ||
3002 | bool added = false; | |
3003 | CInode *in = cache->get_inode(ino, last); | |
3004 | if (!in) { | |
3005 | in = new CInode(mds->mdcache, true, 1, last); | |
3006 | added = true; | |
3007 | } | |
3008 | ||
3009 | // state after link -- or not! -sage | |
3010 | in->decode_import(blp, ls); // cap imports are noted for later action | |
3011 | ||
3012 | // caps | |
3013 | decode_import_inode_caps(in, true, blp, peer_exports); | |
3014 | ||
3015 | // link before state -- or not! -sage | |
3016 | if (dn->get_linkage()->get_inode() != in) { | |
3017 | assert(!dn->get_linkage()->get_inode()); | |
3018 | dn->dir->link_primary_inode(dn, in); | |
3019 | } | |
28e407b8 AA |
3020 | |
3021 | if (in->is_dir()) | |
3022 | dn->dir->pop_lru_subdirs.push_back(&in->item_pop_lru); | |
7c673cae FG |
3023 | |
3024 | // add inode? | |
3025 | if (added) { | |
3026 | cache->add_inode(in); | |
3027 | dout(10) << "added " << *in << dendl; | |
3028 | } else { | |
3029 | dout(10) << " had " << *in << dendl; | |
3030 | } | |
3031 | ||
3032 | if (in->inode.is_dirty_rstat()) | |
3033 | in->mark_dirty_rstat(); | |
3034 | ||
3035 | // clear if dirtyscattered, since we're going to journal this | |
3036 | // but not until we _actually_ finish the import... | |
3037 | if (in->filelock.is_dirty()) { | |
3038 | updated_scatterlocks.push_back(&in->filelock); | |
3039 | mds->locker->mark_updated_scatterlock(&in->filelock); | |
3040 | } | |
3041 | ||
3042 | if (in->dirfragtreelock.is_dirty()) { | |
3043 | updated_scatterlocks.push_back(&in->dirfragtreelock); | |
3044 | mds->locker->mark_updated_scatterlock(&in->dirfragtreelock); | |
3045 | } | |
3046 | ||
3047 | // adjust replica list | |
3048 | //assert(!in->is_replica(oldauth)); // not true on failed export | |
3049 | in->add_replica(oldauth, CInode::EXPORT_NONCE); | |
3050 | if (in->is_replica(mds->get_nodeid())) | |
3051 | in->remove_replica(mds->get_nodeid()); | |
3052 | } | |
3053 | ||
3054 | void Migrator::decode_import_inode_caps(CInode *in, bool auth_cap, | |
3055 | bufferlist::iterator &blp, | |
3056 | map<CInode*, map<client_t,Capability::Export> >& peer_exports) | |
3057 | { | |
3058 | map<client_t,Capability::Export> cap_map; | |
3059 | ::decode(cap_map, blp); | |
3060 | if (auth_cap) | |
3061 | ::decode(in->get_mds_caps_wanted(), blp); | |
3062 | if (!cap_map.empty() || | |
b32b8144 | 3063 | (auth_cap && (in->get_caps_wanted() & ~CEPH_CAP_PIN))) { |
7c673cae FG |
3064 | peer_exports[in].swap(cap_map); |
3065 | in->get(CInode::PIN_IMPORTINGCAPS); | |
3066 | } | |
3067 | } | |
3068 | ||
3069 | void Migrator::finish_import_inode_caps(CInode *in, mds_rank_t peer, bool auth_cap, | |
28e407b8 AA |
3070 | const map<client_t,pair<Session*,uint64_t> >& session_map, |
3071 | const map<client_t,Capability::Export> &export_map, | |
7c673cae FG |
3072 | map<client_t,Capability::Import> &import_map) |
3073 | { | |
28e407b8 AA |
3074 | for (auto& it : export_map) { |
3075 | dout(10) << "finish_import_inode_caps for client." << it.first << " on " << *in << dendl; | |
3076 | ||
3077 | auto p = session_map.find(it.first); | |
3078 | if (p == session_map.end()) { | |
3079 | dout(10) << " no session for client." << it.first << dendl; | |
3080 | (void)import_map[it.first]; | |
3081 | continue; | |
3082 | } | |
7c673cae | 3083 | |
28e407b8 AA |
3084 | Session *session = p->second.first; |
3085 | ||
3086 | Capability *cap = in->get_client_cap(it.first); | |
7c673cae | 3087 | if (!cap) { |
28e407b8 | 3088 | cap = in->add_client_cap(it.first, session); |
7c673cae FG |
3089 | if (peer < 0) |
3090 | cap->mark_importing(); | |
3091 | } | |
3092 | ||
28e407b8 | 3093 | Capability::Import& im = import_map[it.first]; |
7c673cae | 3094 | im.cap_id = cap->get_cap_id(); |
28e407b8 | 3095 | im.mseq = auth_cap ? it.second.mseq : cap->get_mseq(); |
7c673cae FG |
3096 | im.issue_seq = cap->get_last_seq() + 1; |
3097 | ||
3098 | if (peer >= 0) { | |
28e407b8 AA |
3099 | cap->merge(it.second, auth_cap); |
3100 | mds->mdcache->do_cap_import(session, in, cap, it.second.cap_id, | |
3101 | it.second.seq, it.second.mseq - 1, peer, | |
7c673cae FG |
3102 | auth_cap ? CEPH_CAP_FLAG_AUTH : CEPH_CAP_FLAG_RELEASE); |
3103 | } | |
3104 | } | |
3105 | ||
3106 | if (peer >= 0) { | |
3107 | in->replica_caps_wanted = 0; | |
3108 | in->put(CInode::PIN_IMPORTINGCAPS); | |
3109 | } | |
3110 | } | |
3111 | ||
3112 | int Migrator::decode_import_dir(bufferlist::iterator& blp, | |
3113 | mds_rank_t oldauth, | |
3114 | CDir *import_root, | |
3115 | EImportStart *le, | |
3116 | LogSegment *ls, | |
3117 | map<CInode*,map<client_t,Capability::Export> >& peer_exports, | |
3118 | list<ScatterLock*>& updated_scatterlocks, utime_t now) | |
3119 | { | |
3120 | // set up dir | |
3121 | dirfrag_t df; | |
3122 | ::decode(df, blp); | |
3123 | ||
3124 | CInode *diri = cache->get_inode(df.ino); | |
3125 | assert(diri); | |
3126 | CDir *dir = diri->get_or_open_dirfrag(mds->mdcache, df.frag); | |
3127 | assert(dir); | |
3128 | ||
3129 | dout(7) << "decode_import_dir " << *dir << dendl; | |
3130 | ||
3131 | // assimilate state | |
3132 | dir->decode_import(blp, now, ls); | |
3133 | ||
3134 | // adjust replica list | |
3135 | //assert(!dir->is_replica(oldauth)); // not true on failed export | |
3136 | dir->add_replica(oldauth, CDir::EXPORT_NONCE); | |
3137 | if (dir->is_replica(mds->get_nodeid())) | |
3138 | dir->remove_replica(mds->get_nodeid()); | |
3139 | ||
3140 | // add to journal entry | |
3141 | if (le) | |
3142 | le->metablob.add_import_dir(dir); | |
3143 | ||
3144 | int num_imported = 0; | |
3145 | ||
3146 | // take all waiters on this dir | |
3147 | // NOTE: a pass of imported data is guaranteed to get all of my waiters because | |
3148 | // a replica's presense in my cache implies/forces it's presense in authority's. | |
3149 | list<MDSInternalContextBase*> waiters; | |
3150 | ||
3151 | dir->take_waiting(CDir::WAIT_ANY_MASK, waiters); | |
3152 | for (list<MDSInternalContextBase*>::iterator it = waiters.begin(); | |
3153 | it != waiters.end(); | |
3154 | ++it) | |
3155 | import_root->add_waiter(CDir::WAIT_UNFREEZE, *it); // UNFREEZE will get kicked both on success or failure | |
3156 | ||
3157 | dout(15) << "doing contents" << dendl; | |
3158 | ||
3159 | // contents | |
3160 | __u32 nden; | |
3161 | ::decode(nden, blp); | |
3162 | ||
3163 | for (; nden>0; nden--) { | |
3164 | num_imported++; | |
3165 | ||
3166 | // dentry | |
3167 | string dname; | |
3168 | snapid_t last; | |
3169 | ::decode(dname, blp); | |
3170 | ::decode(last, blp); | |
3171 | ||
3172 | CDentry *dn = dir->lookup_exact_snap(dname, last); | |
3173 | if (!dn) | |
3174 | dn = dir->add_null_dentry(dname, 1, last); | |
3175 | ||
3176 | dn->decode_import(blp, ls); | |
3177 | ||
3178 | dn->add_replica(oldauth, CDentry::EXPORT_NONCE); | |
3179 | if (dn->is_replica(mds->get_nodeid())) | |
3180 | dn->remove_replica(mds->get_nodeid()); | |
3181 | ||
3182 | // dentry lock in unreadable state can block path traverse | |
3183 | if (dn->lock.get_state() != LOCK_SYNC) | |
3184 | mds->locker->try_eval(&dn->lock, NULL); | |
3185 | ||
3186 | dout(15) << "decode_import_dir got " << *dn << dendl; | |
3187 | ||
3188 | // points to... | |
3189 | char icode; | |
3190 | ::decode(icode, blp); | |
3191 | ||
3192 | if (icode == 'N') { | |
3193 | // null dentry | |
3194 | assert(dn->get_linkage()->is_null()); | |
3195 | ||
3196 | // fall thru | |
3197 | } | |
3198 | else if (icode == 'L') { | |
3199 | // remote link | |
3200 | inodeno_t ino; | |
3201 | unsigned char d_type; | |
3202 | ::decode(ino, blp); | |
3203 | ::decode(d_type, blp); | |
3204 | if (dn->get_linkage()->is_remote()) { | |
3205 | assert(dn->get_linkage()->get_remote_ino() == ino); | |
3206 | } else { | |
3207 | dir->link_remote_inode(dn, ino, d_type); | |
3208 | } | |
3209 | } | |
3210 | else if (icode == 'I') { | |
3211 | // inode | |
3212 | assert(le); | |
3213 | decode_import_inode(dn, blp, oldauth, ls, | |
3214 | peer_exports, updated_scatterlocks); | |
3215 | } | |
3216 | ||
3217 | // add dentry to journal entry | |
3218 | if (le) | |
3219 | le->metablob.add_import_dentry(dn); | |
3220 | } | |
3221 | ||
3222 | #ifdef MDS_VERIFY_FRAGSTAT | |
3223 | if (dir->is_complete()) | |
3224 | dir->verify_fragstat(); | |
3225 | #endif | |
3226 | ||
3227 | dir->inode->maybe_export_pin(); | |
3228 | ||
3229 | dout(7) << "decode_import_dir done " << *dir << dendl; | |
3230 | return num_imported; | |
3231 | } | |
3232 | ||
3233 | ||
3234 | ||
3235 | ||
3236 | ||
3237 | // authority bystander | |
3238 | ||
3239 | /* This function DOES put the passed message before returning*/ | |
3240 | void Migrator::handle_export_notify(MExportDirNotify *m) | |
3241 | { | |
3242 | if (!(mds->is_clientreplay() || mds->is_active() || mds->is_stopping())) { | |
3243 | m->put(); | |
3244 | return; | |
3245 | } | |
3246 | ||
3247 | CDir *dir = cache->get_dirfrag(m->get_dirfrag()); | |
3248 | ||
3249 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
3250 | mds_authority_t old_auth = m->get_old_auth(); | |
3251 | mds_authority_t new_auth = m->get_new_auth(); | |
3252 | ||
3253 | if (!dir) { | |
3254 | dout(7) << "handle_export_notify " << old_auth << " -> " << new_auth | |
3255 | << " on missing dir " << m->get_dirfrag() << dendl; | |
3256 | } else if (dir->authority() != old_auth) { | |
3257 | dout(7) << "handle_export_notify old_auth was " << dir->authority() | |
3258 | << " != " << old_auth << " -> " << new_auth | |
3259 | << " on " << *dir << dendl; | |
3260 | } else { | |
3261 | dout(7) << "handle_export_notify " << old_auth << " -> " << new_auth | |
3262 | << " on " << *dir << dendl; | |
3263 | // adjust auth | |
3264 | set<CDir*> have; | |
3265 | cache->map_dirfrag_set(m->get_bounds(), have); | |
3266 | cache->adjust_bounded_subtree_auth(dir, have, new_auth); | |
3267 | ||
3268 | // induce a merge? | |
3269 | cache->try_subtree_merge(dir); | |
3270 | } | |
3271 | ||
3272 | // send ack | |
3273 | if (m->wants_ack()) { | |
3274 | mds->send_message_mds(new MExportDirNotifyAck(m->get_dirfrag(), m->get_tid(), m->get_new_auth()), from); | |
3275 | } else { | |
3276 | // aborted. no ack. | |
3277 | dout(7) << "handle_export_notify no ack requested" << dendl; | |
3278 | } | |
3279 | ||
3280 | m->put(); | |
3281 | } | |
3282 | ||
3283 | /** cap exports **/ | |
3284 | void Migrator::export_caps(CInode *in) | |
3285 | { | |
3286 | mds_rank_t dest = in->authority().first; | |
3287 | dout(7) << "export_caps to mds." << dest << " " << *in << dendl; | |
3288 | ||
3289 | assert(in->is_any_caps()); | |
3290 | assert(!in->is_auth()); | |
3291 | assert(!in->is_ambiguous_auth()); | |
3292 | assert(!in->state_test(CInode::STATE_EXPORTINGCAPS)); | |
3293 | ||
3294 | MExportCaps *ex = new MExportCaps; | |
3295 | ex->ino = in->ino(); | |
3296 | ||
3297 | encode_export_inode_caps(in, false, ex->cap_bl, ex->client_map); | |
3298 | ||
3299 | mds->send_message_mds(ex, dest); | |
3300 | } | |
3301 | ||
3302 | void Migrator::handle_gather_caps(MGatherCaps *m) | |
3303 | { | |
3304 | CInode *in = cache->get_inode(m->ino); | |
3305 | ||
3306 | if (!in) | |
3307 | goto out; | |
3308 | ||
3309 | dout(10) << "handle_gather_caps " << *m << " from " << m->get_source() | |
3310 | << " on " << *in | |
3311 | << dendl; | |
3312 | if (in->is_any_caps() && | |
3313 | !in->is_auth() && | |
3314 | !in->is_ambiguous_auth() && | |
3315 | !in->state_test(CInode::STATE_EXPORTINGCAPS)) | |
3316 | export_caps(in); | |
3317 | ||
3318 | out: | |
3319 | m->put(); | |
3320 | } | |
3321 | ||
3322 | class C_M_LoggedImportCaps : public MigratorLogContext { | |
3323 | CInode *in; | |
3324 | mds_rank_t from; | |
3325 | public: | |
28e407b8 | 3326 | map<client_t,pair<Session*,uint64_t> > imported_session_map; |
7c673cae | 3327 | map<CInode*, map<client_t,Capability::Export> > peer_exports; |
7c673cae FG |
3328 | |
3329 | C_M_LoggedImportCaps(Migrator *m, CInode *i, mds_rank_t f) : MigratorLogContext(m), in(i), from(f) {} | |
3330 | void finish(int r) override { | |
28e407b8 | 3331 | mig->logged_import_caps(in, from, imported_session_map, peer_exports); |
7c673cae FG |
3332 | } |
3333 | }; | |
3334 | ||
3335 | /* This function DOES put the passed message before returning*/ | |
3336 | void Migrator::handle_export_caps(MExportCaps *ex) | |
3337 | { | |
3338 | dout(10) << "handle_export_caps " << *ex << " from " << ex->get_source() << dendl; | |
3339 | CInode *in = cache->get_inode(ex->ino); | |
3340 | ||
3341 | assert(in); | |
3342 | assert(in->is_auth()); | |
3343 | ||
3344 | // FIXME | |
28e407b8 AA |
3345 | if (!in->can_auth_pin()) { |
3346 | ex->put(); | |
7c673cae | 3347 | return; |
28e407b8 AA |
3348 | } |
3349 | ||
181888fb | 3350 | in->auth_pin(this); |
7c673cae | 3351 | |
28e407b8 AA |
3352 | map<client_t,entity_inst_t> client_map; |
3353 | client_map.swap(ex->client_map); | |
3354 | ||
7c673cae FG |
3355 | C_M_LoggedImportCaps *finish = new C_M_LoggedImportCaps( |
3356 | this, in, mds_rank_t(ex->get_source().num())); | |
7c673cae | 3357 | |
28e407b8 AA |
3358 | version_t pv = mds->server->prepare_force_open_sessions(client_map, |
3359 | finish->imported_session_map); | |
7c673cae FG |
3360 | // decode new caps |
3361 | bufferlist::iterator blp = ex->cap_bl.begin(); | |
3362 | decode_import_inode_caps(in, false, blp, finish->peer_exports); | |
3363 | assert(!finish->peer_exports.empty()); // thus, inode is pinned. | |
3364 | ||
3365 | // journal open client sessions | |
7c673cae | 3366 | |
28e407b8 | 3367 | ESessions *le = new ESessions(pv, client_map); |
7c673cae FG |
3368 | mds->mdlog->start_submit_entry(le, finish); |
3369 | mds->mdlog->flush(); | |
3370 | ||
3371 | ex->put(); | |
3372 | } | |
3373 | ||
3374 | ||
3375 | void Migrator::logged_import_caps(CInode *in, | |
3376 | mds_rank_t from, | |
28e407b8 AA |
3377 | map<client_t,pair<Session*,uint64_t> >& imported_session_map, |
3378 | map<CInode*, map<client_t,Capability::Export> >& peer_exports) | |
7c673cae FG |
3379 | { |
3380 | dout(10) << "logged_import_caps on " << *in << dendl; | |
3381 | // see export_go() vs export_go_synced() | |
3382 | assert(in->is_auth()); | |
3383 | ||
3384 | // force open client sessions and finish cap import | |
28e407b8 | 3385 | mds->server->finish_force_open_sessions(imported_session_map); |
7c673cae FG |
3386 | |
3387 | map<client_t,Capability::Import> imported_caps; | |
3388 | ||
28e407b8 AA |
3389 | auto it = peer_exports.find(in); |
3390 | assert(it != peer_exports.end()); | |
3391 | ||
7c673cae | 3392 | // clients will release caps from the exporter when they receive the cap import message. |
28e407b8 | 3393 | finish_import_inode_caps(in, from, false, imported_session_map, it->second, imported_caps); |
7c673cae | 3394 | mds->locker->eval(in, CEPH_CAP_LOCKS, true); |
181888fb | 3395 | in->auth_unpin(this); |
7c673cae | 3396 | } |
28e407b8 AA |
3397 | |
3398 | void Migrator::handle_conf_change(const struct md_config_t *conf, | |
3399 | const std::set <std::string> &changed, | |
3400 | const MDSMap &mds_map) | |
3401 | { | |
3402 | if (changed.count("mds_inject_migrator_session_race")) { | |
3403 | inject_session_race = conf->get_val<bool>("mds_inject_migrator_session_race"); | |
3404 | dout(0) << "mds_inject_migrator_session_race is " << inject_session_race << dendl; | |
3405 | } | |
3406 | } |