]>
Commit | Line | Data |
---|---|---|
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- | |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 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 "PG.h" | |
16 | // #include "msg/Messenger.h" | |
17 | #include "messages/MOSDRepScrub.h" | |
18 | // #include "common/cmdparse.h" | |
19 | // #include "common/ceph_context.h" | |
20 | ||
21 | #include "common/errno.h" | |
22 | #include "common/config.h" | |
23 | #include "OSD.h" | |
24 | #include "OpRequest.h" | |
25 | #include "ScrubStore.h" | |
26 | #include "Session.h" | |
27 | ||
28 | #include "common/Timer.h" | |
29 | #include "common/perf_counters.h" | |
30 | ||
31 | #include "messages/MOSDOp.h" | |
32 | #include "messages/MOSDPGNotify.h" | |
33 | // #include "messages/MOSDPGLog.h" | |
34 | #include "messages/MOSDPGRemove.h" | |
35 | #include "messages/MOSDPGInfo.h" | |
36 | #include "messages/MOSDPGTrim.h" | |
37 | #include "messages/MOSDPGScan.h" | |
38 | #include "messages/MOSDPGBackfill.h" | |
39 | #include "messages/MOSDPGBackfillRemove.h" | |
40 | #include "messages/MBackfillReserve.h" | |
41 | #include "messages/MRecoveryReserve.h" | |
42 | #include "messages/MOSDPGPush.h" | |
43 | #include "messages/MOSDPGPushReply.h" | |
44 | #include "messages/MOSDPGPull.h" | |
45 | #include "messages/MOSDECSubOpWrite.h" | |
46 | #include "messages/MOSDECSubOpWriteReply.h" | |
47 | #include "messages/MOSDECSubOpRead.h" | |
48 | #include "messages/MOSDECSubOpReadReply.h" | |
49 | #include "messages/MOSDPGUpdateLogMissing.h" | |
50 | #include "messages/MOSDPGUpdateLogMissingReply.h" | |
51 | #include "messages/MOSDBackoff.h" | |
52 | #include "messages/MOSDScrubReserve.h" | |
53 | #include "messages/MOSDSubOp.h" | |
54 | #include "messages/MOSDRepOp.h" | |
55 | #include "messages/MOSDSubOpReply.h" | |
56 | #include "messages/MOSDRepOpReply.h" | |
57 | #include "messages/MOSDRepScrubMap.h" | |
58 | ||
59 | #include "common/BackTrace.h" | |
60 | #include "common/EventTrace.h" | |
61 | ||
62 | #ifdef WITH_LTTNG | |
63 | #define TRACEPOINT_DEFINE | |
64 | #define TRACEPOINT_PROBE_DYNAMIC_LINKAGE | |
65 | #include "tracing/pg.h" | |
66 | #undef TRACEPOINT_PROBE_DYNAMIC_LINKAGE | |
67 | #undef TRACEPOINT_DEFINE | |
68 | #else | |
69 | #define tracepoint(...) | |
70 | #endif | |
71 | ||
72 | #include <sstream> | |
73 | ||
74 | #define dout_context cct | |
75 | #define dout_subsys ceph_subsys_osd | |
76 | #undef dout_prefix | |
77 | #define dout_prefix _prefix(_dout, this) | |
78 | ||
79 | // prefix pgmeta_oid keys with _ so that PGLog::read_log_and_missing() can | |
80 | // easily skip them | |
81 | const string infover_key("_infover"); | |
82 | const string info_key("_info"); | |
83 | const string biginfo_key("_biginfo"); | |
84 | const string epoch_key("_epoch"); | |
85 | const string fastinfo_key("_fastinfo"); | |
86 | ||
87 | template <class T> | |
88 | static ostream& _prefix(std::ostream *_dout, T *t) | |
89 | { | |
90 | return *_dout << t->gen_prefix(); | |
91 | } | |
92 | ||
93 | MEMPOOL_DEFINE_OBJECT_FACTORY(PG::CephPeeringEvt, pg_peering_evt, osd); | |
94 | ||
95 | void PGStateHistory::enter(PG* pg, const utime_t entime, const char* state) | |
96 | { | |
97 | // Ignore trimming state machine for now | |
98 | if (::strstr(state, "Trimming") != NULL) { | |
99 | return; | |
100 | } else if (pi != nullptr) { | |
101 | pi->enter_state(entime, state); | |
102 | } else { | |
103 | // Store current state since we can't reliably take the PG lock here | |
104 | if ( tmppi == nullptr) { | |
105 | tmppi = std::unique_ptr<PGStateInstance>(new PGStateInstance); | |
106 | } | |
107 | ||
108 | thispg = pg; | |
109 | tmppi->enter_state(entime, state); | |
110 | } | |
111 | } | |
112 | ||
113 | void PGStateHistory::exit(const char* state) { | |
114 | // Ignore trimming state machine for now | |
115 | // Do nothing if PG is being destroyed! | |
116 | if (::strstr(state, "Trimming") != NULL || pg_in_destructor) { | |
117 | return; | |
118 | } else { | |
119 | bool ilocked = false; | |
120 | if(!thispg->is_locked()) { | |
121 | thispg->lock(); | |
122 | ilocked = true; | |
123 | } | |
124 | if (pi == nullptr) { | |
125 | buffer.push_back(std::unique_ptr<PGStateInstance>(tmppi.release())); | |
126 | pi = buffer.back().get(); | |
127 | pi->setepoch(thispg->get_osdmap()->get_epoch()); | |
128 | } | |
129 | ||
130 | pi->exit_state(ceph_clock_now()); | |
131 | if (::strcmp(state, "Reset") == 0) { | |
132 | this->reset(); | |
133 | } | |
134 | if(ilocked) { | |
135 | thispg->unlock(); | |
136 | } | |
137 | } | |
138 | } | |
139 | ||
140 | void PGStateHistory::dump(Formatter* f) const { | |
141 | f->open_array_section("history"); | |
142 | for (auto pi = buffer.begin(); pi != buffer.end(); ++pi) { | |
143 | f->open_object_section("states"); | |
144 | f->dump_stream("epoch") << (*pi)->this_epoch; | |
145 | for (auto she : (*pi)->state_history) { | |
146 | f->dump_string("state", std::get<2>(she)); | |
147 | f->dump_stream("enter") << std::get<0>(she); | |
148 | f->dump_stream("exit") << std::get<1>(she); | |
149 | } | |
150 | f->close_section(); | |
151 | } | |
152 | f->close_section(); | |
153 | } | |
154 | ||
155 | void PG::get(const char* tag) | |
156 | { | |
157 | ref++; | |
158 | #ifdef PG_DEBUG_REFS | |
159 | Mutex::Locker l(_ref_id_lock); | |
160 | _tag_counts[tag]++; | |
161 | #endif | |
162 | } | |
163 | ||
164 | void PG::put(const char* tag) | |
165 | { | |
166 | #ifdef PG_DEBUG_REFS | |
167 | { | |
168 | Mutex::Locker l(_ref_id_lock); | |
169 | auto tag_counts_entry = _tag_counts.find(tag); | |
170 | assert(tag_counts_entry != _tag_counts.end()); | |
171 | --tag_counts_entry->second; | |
172 | if (tag_counts_entry->second == 0) { | |
173 | _tag_counts.erase(tag_counts_entry); | |
174 | } | |
175 | } | |
176 | #endif | |
177 | if (--ref== 0) | |
178 | delete this; | |
179 | } | |
180 | ||
181 | #ifdef PG_DEBUG_REFS | |
182 | uint64_t PG::get_with_id() | |
183 | { | |
184 | ref++; | |
185 | Mutex::Locker l(_ref_id_lock); | |
186 | uint64_t id = ++_ref_id; | |
187 | BackTrace bt(0); | |
188 | stringstream ss; | |
189 | bt.print(ss); | |
190 | dout(20) << __func__ << ": " << info.pgid << " got id " << id << " (new) ref==" << ref << dendl; | |
191 | assert(!_live_ids.count(id)); | |
192 | _live_ids.insert(make_pair(id, ss.str())); | |
193 | return id; | |
194 | } | |
195 | ||
196 | void PG::put_with_id(uint64_t id) | |
197 | { | |
198 | dout(20) << __func__ << ": " << info.pgid << " put id " << id << " (current) ref==" << ref << dendl; | |
199 | { | |
200 | Mutex::Locker l(_ref_id_lock); | |
201 | assert(_live_ids.count(id)); | |
202 | _live_ids.erase(id); | |
203 | } | |
204 | if (--ref == 0) | |
205 | delete this; | |
206 | } | |
207 | ||
208 | void PG::dump_live_ids() | |
209 | { | |
210 | Mutex::Locker l(_ref_id_lock); | |
211 | dout(0) << "\t" << __func__ << ": " << info.pgid << " live ids:" << dendl; | |
212 | for (map<uint64_t, string>::iterator i = _live_ids.begin(); | |
213 | i != _live_ids.end(); | |
214 | ++i) { | |
215 | dout(0) << "\t\tid: " << *i << dendl; | |
216 | } | |
217 | dout(0) << "\t" << __func__ << ": " << info.pgid << " live tags:" << dendl; | |
218 | for (map<string, uint64_t>::iterator i = _tag_counts.begin(); | |
219 | i != _tag_counts.end(); | |
220 | ++i) { | |
221 | dout(0) << "\t\tid: " << *i << dendl; | |
222 | } | |
223 | } | |
224 | #endif | |
225 | ||
226 | void PGPool::update(OSDMapRef map) | |
227 | { | |
228 | const pg_pool_t *pi = map->get_pg_pool(id); | |
229 | assert(pi); | |
230 | info = *pi; | |
231 | auid = pi->auid; | |
232 | name = map->get_pool_name(id); | |
233 | bool updated = false; | |
234 | if ((map->get_epoch() != cached_epoch + 1) || | |
235 | (pi->get_snap_epoch() == map->get_epoch())) { | |
236 | updated = true; | |
237 | pi->build_removed_snaps(newly_removed_snaps); | |
238 | interval_set<snapid_t> intersection; | |
239 | intersection.intersection_of(newly_removed_snaps, cached_removed_snaps); | |
240 | if (intersection == cached_removed_snaps) { | |
241 | newly_removed_snaps.subtract(cached_removed_snaps); | |
242 | cached_removed_snaps.union_of(newly_removed_snaps); | |
243 | } else { | |
244 | lgeneric_subdout(cct, osd, 0) << __func__ | |
245 | << " cached_removed_snaps shrank from " << cached_removed_snaps | |
246 | << " to " << newly_removed_snaps << dendl; | |
247 | cached_removed_snaps = newly_removed_snaps; | |
248 | newly_removed_snaps.clear(); | |
249 | } | |
250 | snapc = pi->get_snap_context(); | |
251 | } else { | |
252 | /* 1) map->get_epoch() == cached_epoch + 1 && | |
253 | * 2) pi->get_snap_epoch() != map->get_epoch() | |
254 | * | |
255 | * From the if branch, 1 && 2 must be true. From 2, we know that | |
256 | * this map didn't change the set of removed snaps. From 1, we | |
257 | * know that our cached_removed_snaps matches the previous map. | |
258 | * Thus, from 1 && 2, cached_removed snaps matches the current | |
259 | * set of removed snaps and all we have to do is clear | |
260 | * newly_removed_snaps. | |
261 | */ | |
262 | newly_removed_snaps.clear(); | |
263 | } | |
264 | cached_epoch = map->get_epoch(); | |
265 | lgeneric_subdout(cct, osd, 20) | |
266 | << "PGPool::update cached_removed_snaps " | |
267 | << cached_removed_snaps | |
268 | << " newly_removed_snaps " | |
269 | << newly_removed_snaps | |
270 | << " snapc " << snapc | |
271 | << (updated ? " (updated)":" (no change)") | |
272 | << dendl; | |
273 | } | |
274 | ||
275 | PG::PG(OSDService *o, OSDMapRef curmap, | |
276 | const PGPool &_pool, spg_t p) : | |
277 | osd(o), | |
278 | cct(o->cct), | |
279 | osdriver(osd->store, coll_t(), OSD::make_snapmapper_oid()), | |
280 | snap_mapper( | |
281 | cct, | |
282 | &osdriver, | |
283 | p.ps(), | |
284 | p.get_split_bits(curmap->get_pg_num(_pool.id)), | |
285 | _pool.id, | |
286 | p.shard), | |
287 | osdmap_ref(curmap), last_persisted_osdmap_ref(curmap), pool(_pool), | |
288 | _lock("PG::_lock"), | |
289 | #ifdef PG_DEBUG_REFS | |
290 | _ref_id_lock("PG::_ref_id_lock"), _ref_id(0), | |
291 | #endif | |
292 | deleting(false), | |
293 | trace_endpoint("0.0.0.0", 0, "PG"), | |
294 | dirty_info(false), dirty_big_info(false), | |
295 | info(p), | |
296 | info_struct_v(0), | |
297 | coll(p), pg_log(cct), | |
298 | pgmeta_oid(p.make_pgmeta_oid()), | |
299 | missing_loc(this), | |
300 | past_intervals( | |
301 | curmap->get_pools().at(p.pgid.pool()).ec_pool(), | |
302 | *curmap), | |
303 | stat_queue_item(this), | |
304 | scrub_queued(false), | |
305 | recovery_queued(false), | |
306 | recovery_ops_active(0), | |
307 | role(-1), | |
308 | state(0), | |
309 | send_notify(false), | |
310 | pg_whoami(osd->whoami, p.shard), | |
311 | need_up_thru(false), | |
312 | last_peering_reset(0), | |
313 | heartbeat_peer_lock("PG::heartbeat_peer_lock"), | |
314 | backfill_reserved(false), | |
315 | backfill_reserving(false), | |
316 | flushes_in_progress(0), | |
317 | pg_stats_publish_lock("PG::pg_stats_publish_lock"), | |
318 | pg_stats_publish_valid(false), | |
319 | osr(osd->osr_registry.lookup_or_create(p, (stringify(p)))), | |
320 | finish_sync_event(NULL), | |
321 | backoff_lock("PG::backoff_lock"), | |
322 | scrub_after_recovery(false), | |
323 | active_pushes(0), | |
324 | recovery_state(this), | |
325 | pg_id(p), | |
326 | peer_features(CEPH_FEATURES_SUPPORTED_DEFAULT), | |
327 | acting_features(CEPH_FEATURES_SUPPORTED_DEFAULT), | |
328 | upacting_features(CEPH_FEATURES_SUPPORTED_DEFAULT), | |
329 | last_epoch(0) | |
330 | { | |
331 | #ifdef PG_DEBUG_REFS | |
332 | osd->add_pgid(p, this); | |
333 | #endif | |
334 | #ifdef WITH_BLKIN | |
335 | std::stringstream ss; | |
336 | ss << "PG " << info.pgid; | |
337 | trace_endpoint.copy_name(ss.str()); | |
338 | #endif | |
339 | osr->shard_hint = p; | |
340 | } | |
341 | ||
342 | PG::~PG() | |
343 | { | |
344 | pgstate_history.set_pg_in_destructor(); | |
345 | #ifdef PG_DEBUG_REFS | |
346 | osd->remove_pgid(info.pgid, this); | |
347 | #endif | |
348 | } | |
349 | ||
350 | void PG::lock_suspend_timeout(ThreadPool::TPHandle &handle) | |
351 | { | |
352 | handle.suspend_tp_timeout(); | |
353 | lock(); | |
354 | handle.reset_tp_timeout(); | |
355 | } | |
356 | ||
357 | void PG::lock(bool no_lockdep) const | |
358 | { | |
359 | _lock.Lock(no_lockdep); | |
360 | // if we have unrecorded dirty state with the lock dropped, there is a bug | |
361 | assert(!dirty_info); | |
362 | assert(!dirty_big_info); | |
363 | ||
364 | dout(30) << "lock" << dendl; | |
365 | } | |
366 | ||
367 | std::string PG::gen_prefix() const | |
368 | { | |
369 | stringstream out; | |
370 | OSDMapRef mapref = osdmap_ref; | |
371 | if (_lock.is_locked_by_me()) { | |
372 | out << "osd." << osd->whoami | |
373 | << " pg_epoch: " << (mapref ? mapref->get_epoch():0) | |
374 | << " " << *this << " "; | |
375 | } else { | |
376 | out << "osd." << osd->whoami | |
377 | << " pg_epoch: " << (mapref ? mapref->get_epoch():0) | |
378 | << " pg[" << info.pgid << "(unlocked)] "; | |
379 | } | |
380 | return out.str(); | |
381 | } | |
382 | ||
383 | /********* PG **********/ | |
384 | ||
385 | void PG::proc_master_log( | |
386 | ObjectStore::Transaction& t, pg_info_t &oinfo, | |
387 | pg_log_t &olog, pg_missing_t& omissing, pg_shard_t from) | |
388 | { | |
389 | dout(10) << "proc_master_log for osd." << from << ": " | |
390 | << olog << " " << omissing << dendl; | |
391 | assert(!is_peered() && is_primary()); | |
392 | ||
393 | // merge log into our own log to build master log. no need to | |
394 | // make any adjustments to their missing map; we are taking their | |
395 | // log to be authoritative (i.e., their entries are by definitely | |
396 | // non-divergent). | |
397 | merge_log(t, oinfo, olog, from); | |
398 | peer_info[from] = oinfo; | |
399 | dout(10) << " peer osd." << from << " now " << oinfo << " " << omissing << dendl; | |
400 | might_have_unfound.insert(from); | |
401 | ||
402 | // See doc/dev/osd_internals/last_epoch_started | |
403 | if (oinfo.last_epoch_started > info.last_epoch_started) { | |
404 | info.last_epoch_started = oinfo.last_epoch_started; | |
405 | dirty_info = true; | |
406 | } | |
407 | if (oinfo.last_interval_started > info.last_interval_started) { | |
408 | info.last_interval_started = oinfo.last_interval_started; | |
409 | dirty_info = true; | |
410 | } | |
411 | update_history(oinfo.history); | |
412 | assert(cct->_conf->osd_find_best_info_ignore_history_les || | |
413 | info.last_epoch_started >= info.history.last_epoch_started); | |
414 | ||
415 | peer_missing[from].claim(omissing); | |
416 | } | |
417 | ||
418 | void PG::proc_replica_log( | |
419 | pg_info_t &oinfo, | |
420 | const pg_log_t &olog, | |
421 | pg_missing_t& omissing, | |
422 | pg_shard_t from) | |
423 | { | |
424 | dout(10) << "proc_replica_log for osd." << from << ": " | |
425 | << oinfo << " " << olog << " " << omissing << dendl; | |
426 | ||
427 | pg_log.proc_replica_log(oinfo, olog, omissing, from); | |
428 | ||
429 | peer_info[from] = oinfo; | |
430 | dout(10) << " peer osd." << from << " now " << oinfo << " " << omissing << dendl; | |
431 | might_have_unfound.insert(from); | |
432 | ||
433 | for (map<hobject_t, pg_missing_item>::const_iterator i = | |
434 | omissing.get_items().begin(); | |
435 | i != omissing.get_items().end(); | |
436 | ++i) { | |
437 | dout(20) << " after missing " << i->first << " need " << i->second.need | |
438 | << " have " << i->second.have << dendl; | |
439 | } | |
440 | peer_missing[from].claim(omissing); | |
441 | } | |
442 | ||
443 | bool PG::proc_replica_info( | |
444 | pg_shard_t from, const pg_info_t &oinfo, epoch_t send_epoch) | |
445 | { | |
446 | map<pg_shard_t, pg_info_t>::iterator p = peer_info.find(from); | |
447 | if (p != peer_info.end() && p->second.last_update == oinfo.last_update) { | |
448 | dout(10) << " got dup osd." << from << " info " << oinfo << ", identical to ours" << dendl; | |
449 | return false; | |
450 | } | |
451 | ||
452 | if (!get_osdmap()->has_been_up_since(from.osd, send_epoch)) { | |
453 | dout(10) << " got info " << oinfo << " from down osd." << from | |
454 | << " discarding" << dendl; | |
455 | return false; | |
456 | } | |
457 | ||
458 | dout(10) << " got osd." << from << " " << oinfo << dendl; | |
459 | assert(is_primary()); | |
460 | peer_info[from] = oinfo; | |
461 | might_have_unfound.insert(from); | |
462 | ||
463 | update_history(oinfo.history); | |
464 | ||
465 | // stray? | |
466 | if (!is_up(from) && !is_acting(from)) { | |
467 | dout(10) << " osd." << from << " has stray content: " << oinfo << dendl; | |
468 | stray_set.insert(from); | |
469 | if (is_clean()) { | |
470 | purge_strays(); | |
471 | } | |
472 | } | |
473 | ||
474 | // was this a new info? if so, update peers! | |
475 | if (p == peer_info.end()) | |
476 | update_heartbeat_peers(); | |
477 | ||
478 | return true; | |
479 | } | |
480 | ||
481 | void PG::remove_snap_mapped_object( | |
482 | ObjectStore::Transaction &t, const hobject_t &soid) | |
483 | { | |
484 | t.remove( | |
485 | coll, | |
486 | ghobject_t(soid, ghobject_t::NO_GEN, pg_whoami.shard)); | |
487 | clear_object_snap_mapping(&t, soid); | |
488 | } | |
489 | ||
490 | void PG::clear_object_snap_mapping( | |
491 | ObjectStore::Transaction *t, const hobject_t &soid) | |
492 | { | |
493 | OSDriver::OSTransaction _t(osdriver.get_transaction(t)); | |
494 | if (soid.snap < CEPH_MAXSNAP) { | |
495 | int r = snap_mapper.remove_oid( | |
496 | soid, | |
497 | &_t); | |
498 | if (!(r == 0 || r == -ENOENT)) { | |
499 | derr << __func__ << ": remove_oid returned " << cpp_strerror(r) << dendl; | |
500 | ceph_abort(); | |
501 | } | |
502 | } | |
503 | } | |
504 | ||
505 | void PG::update_object_snap_mapping( | |
506 | ObjectStore::Transaction *t, const hobject_t &soid, const set<snapid_t> &snaps) | |
507 | { | |
508 | OSDriver::OSTransaction _t(osdriver.get_transaction(t)); | |
509 | assert(soid.snap < CEPH_MAXSNAP); | |
510 | int r = snap_mapper.remove_oid( | |
511 | soid, | |
512 | &_t); | |
513 | if (!(r == 0 || r == -ENOENT)) { | |
514 | derr << __func__ << ": remove_oid returned " << cpp_strerror(r) << dendl; | |
515 | ceph_abort(); | |
516 | } | |
517 | snap_mapper.add_oid( | |
518 | soid, | |
519 | snaps, | |
520 | &_t); | |
521 | } | |
522 | ||
523 | void PG::merge_log( | |
524 | ObjectStore::Transaction& t, pg_info_t &oinfo, pg_log_t &olog, pg_shard_t from) | |
525 | { | |
526 | PGLogEntryHandler rollbacker{this, &t}; | |
527 | pg_log.merge_log( | |
528 | oinfo, olog, from, info, &rollbacker, dirty_info, dirty_big_info); | |
529 | } | |
530 | ||
531 | void PG::rewind_divergent_log(ObjectStore::Transaction& t, eversion_t newhead) | |
532 | { | |
533 | PGLogEntryHandler rollbacker{this, &t}; | |
534 | pg_log.rewind_divergent_log( | |
535 | newhead, info, &rollbacker, dirty_info, dirty_big_info); | |
536 | } | |
537 | ||
538 | /* | |
539 | * Process information from a replica to determine if it could have any | |
540 | * objects that i need. | |
541 | * | |
542 | * TODO: if the missing set becomes very large, this could get expensive. | |
543 | * Instead, we probably want to just iterate over our unfound set. | |
544 | */ | |
545 | bool PG::search_for_missing( | |
546 | const pg_info_t &oinfo, const pg_missing_t &omissing, | |
547 | pg_shard_t from, | |
548 | RecoveryCtx *ctx) | |
549 | { | |
550 | uint64_t num_unfound_before = missing_loc.num_unfound(); | |
551 | bool found_missing = missing_loc.add_source_info( | |
552 | from, oinfo, omissing, ctx->handle); | |
553 | if (found_missing && num_unfound_before != missing_loc.num_unfound()) | |
554 | publish_stats_to_osd(); | |
555 | if (found_missing && | |
556 | (get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD, NULL) & | |
557 | CEPH_FEATURE_OSD_ERASURE_CODES)) { | |
558 | pg_info_t tinfo(oinfo); | |
559 | tinfo.pgid.shard = pg_whoami.shard; | |
560 | (*(ctx->info_map))[from.osd].push_back( | |
561 | make_pair( | |
562 | pg_notify_t( | |
563 | from.shard, pg_whoami.shard, | |
564 | get_osdmap()->get_epoch(), | |
565 | get_osdmap()->get_epoch(), | |
566 | tinfo), | |
567 | past_intervals)); | |
568 | } | |
569 | return found_missing; | |
570 | } | |
571 | ||
572 | bool PG::MissingLoc::readable_with_acting( | |
573 | const hobject_t &hoid, | |
574 | const set<pg_shard_t> &acting) const { | |
575 | if (!needs_recovery(hoid)) return true; | |
576 | auto missing_loc_entry = missing_loc.find(hoid); | |
577 | if (missing_loc_entry == missing_loc.end()) return false; | |
578 | const set<pg_shard_t> &locs = missing_loc_entry->second; | |
579 | ldout(pg->cct, 10) << __func__ << ": locs:" << locs << dendl; | |
580 | set<pg_shard_t> have_acting; | |
581 | for (set<pg_shard_t>::const_iterator i = locs.begin(); | |
582 | i != locs.end(); | |
583 | ++i) { | |
584 | if (acting.count(*i)) | |
585 | have_acting.insert(*i); | |
586 | } | |
587 | return (*is_readable)(have_acting); | |
588 | } | |
589 | ||
590 | void PG::MissingLoc::add_batch_sources_info( | |
591 | const set<pg_shard_t> &sources, ThreadPool::TPHandle* handle) | |
592 | { | |
593 | ldout(pg->cct, 10) << __func__ << ": adding sources in batch " | |
594 | << sources.size() << dendl; | |
595 | unsigned loop = 0; | |
596 | for (map<hobject_t, pg_missing_item>::const_iterator i = needs_recovery_map.begin(); | |
597 | i != needs_recovery_map.end(); | |
598 | ++i) { | |
599 | if (handle && ++loop >= pg->cct->_conf->osd_loop_before_reset_tphandle) { | |
600 | handle->reset_tp_timeout(); | |
601 | loop = 0; | |
602 | } | |
603 | missing_loc[i->first].insert(sources.begin(), sources.end()); | |
604 | missing_loc_sources.insert(sources.begin(), sources.end()); | |
605 | } | |
606 | } | |
607 | ||
608 | bool PG::MissingLoc::add_source_info( | |
609 | pg_shard_t fromosd, | |
610 | const pg_info_t &oinfo, | |
611 | const pg_missing_t &omissing, | |
612 | ThreadPool::TPHandle* handle) | |
613 | { | |
614 | bool found_missing = false; | |
615 | unsigned loop = 0; | |
616 | // found items? | |
617 | for (map<hobject_t,pg_missing_item>::const_iterator p = needs_recovery_map.begin(); | |
618 | p != needs_recovery_map.end(); | |
619 | ++p) { | |
620 | const hobject_t &soid(p->first); | |
621 | eversion_t need = p->second.need; | |
622 | if (handle && ++loop >= pg->cct->_conf->osd_loop_before_reset_tphandle) { | |
623 | handle->reset_tp_timeout(); | |
624 | loop = 0; | |
625 | } | |
626 | if (oinfo.last_update < need) { | |
627 | ldout(pg->cct, 10) << "search_for_missing " << soid << " " << need | |
628 | << " also missing on osd." << fromosd | |
629 | << " (last_update " << oinfo.last_update | |
630 | << " < needed " << need << ")" << dendl; | |
631 | continue; | |
632 | } | |
633 | if (!oinfo.last_backfill.is_max() && | |
634 | !oinfo.last_backfill_bitwise) { | |
635 | ldout(pg->cct, 10) << "search_for_missing " << soid << " " << need | |
636 | << " also missing on osd." << fromosd | |
637 | << " (last_backfill " << oinfo.last_backfill | |
638 | << " but with wrong sort order)" | |
639 | << dendl; | |
640 | continue; | |
641 | } | |
642 | if (p->first >= oinfo.last_backfill) { | |
643 | // FIXME: this is _probably_ true, although it could conceivably | |
644 | // be in the undefined region! Hmm! | |
645 | ldout(pg->cct, 10) << "search_for_missing " << soid << " " << need | |
646 | << " also missing on osd." << fromosd | |
647 | << " (past last_backfill " << oinfo.last_backfill | |
648 | << ")" << dendl; | |
649 | continue; | |
650 | } | |
651 | if (oinfo.last_complete < need) { | |
652 | if (omissing.is_missing(soid)) { | |
653 | ldout(pg->cct, 10) << "search_for_missing " << soid << " " << need | |
654 | << " also missing on osd." << fromosd << dendl; | |
655 | continue; | |
656 | } | |
657 | } | |
658 | ||
659 | ldout(pg->cct, 10) << "search_for_missing " << soid << " " << need | |
660 | << " is on osd." << fromosd << dendl; | |
661 | ||
662 | missing_loc[soid].insert(fromosd); | |
663 | missing_loc_sources.insert(fromosd); | |
664 | found_missing = true; | |
665 | } | |
666 | ||
667 | ldout(pg->cct, 20) << "needs_recovery_map missing " << needs_recovery_map | |
668 | << dendl; | |
669 | return found_missing; | |
670 | } | |
671 | ||
672 | void PG::discover_all_missing(map<int, map<spg_t,pg_query_t> > &query_map) | |
673 | { | |
674 | auto &missing = pg_log.get_missing(); | |
675 | uint64_t unfound = get_num_unfound(); | |
676 | assert(unfound > 0); | |
677 | ||
678 | dout(10) << __func__ << " " | |
679 | << missing.num_missing() << " missing, " | |
680 | << unfound << " unfound" | |
681 | << dendl; | |
682 | ||
683 | std::set<pg_shard_t>::const_iterator m = might_have_unfound.begin(); | |
684 | std::set<pg_shard_t>::const_iterator mend = might_have_unfound.end(); | |
685 | for (; m != mend; ++m) { | |
686 | pg_shard_t peer(*m); | |
687 | ||
688 | if (!get_osdmap()->is_up(peer.osd)) { | |
689 | dout(20) << __func__ << " skipping down osd." << peer << dendl; | |
690 | continue; | |
691 | } | |
692 | ||
693 | map<pg_shard_t, pg_info_t>::const_iterator iter = peer_info.find(peer); | |
694 | if (iter != peer_info.end() && | |
695 | (iter->second.is_empty() || iter->second.dne())) { | |
696 | // ignore empty peers | |
697 | continue; | |
698 | } | |
699 | ||
700 | // If we've requested any of this stuff, the pg_missing_t information | |
701 | // should be on its way. | |
702 | // TODO: coalsce requested_* into a single data structure | |
703 | if (peer_missing.find(peer) != peer_missing.end()) { | |
704 | dout(20) << __func__ << ": osd." << peer | |
705 | << ": we already have pg_missing_t" << dendl; | |
706 | continue; | |
707 | } | |
708 | if (peer_log_requested.find(peer) != peer_log_requested.end()) { | |
709 | dout(20) << __func__ << ": osd." << peer | |
710 | << ": in peer_log_requested" << dendl; | |
711 | continue; | |
712 | } | |
713 | if (peer_missing_requested.find(peer) != peer_missing_requested.end()) { | |
714 | dout(20) << __func__ << ": osd." << peer | |
715 | << ": in peer_missing_requested" << dendl; | |
716 | continue; | |
717 | } | |
718 | ||
719 | // Request missing | |
720 | dout(10) << __func__ << ": osd." << peer << ": requesting pg_missing_t" | |
721 | << dendl; | |
722 | peer_missing_requested.insert(peer); | |
723 | query_map[peer.osd][spg_t(info.pgid.pgid, peer.shard)] = | |
724 | pg_query_t( | |
725 | pg_query_t::FULLLOG, | |
726 | peer.shard, pg_whoami.shard, | |
727 | info.history, get_osdmap()->get_epoch()); | |
728 | } | |
729 | } | |
730 | ||
731 | /******* PG ***********/ | |
732 | bool PG::needs_recovery() const | |
733 | { | |
734 | assert(is_primary()); | |
735 | ||
736 | auto &missing = pg_log.get_missing(); | |
737 | ||
738 | if (missing.num_missing()) { | |
739 | dout(10) << __func__ << " primary has " << missing.num_missing() | |
740 | << " missing" << dendl; | |
741 | return true; | |
742 | } | |
743 | ||
744 | assert(!actingbackfill.empty()); | |
745 | set<pg_shard_t>::const_iterator end = actingbackfill.end(); | |
746 | set<pg_shard_t>::const_iterator a = actingbackfill.begin(); | |
747 | for (; a != end; ++a) { | |
748 | if (*a == get_primary()) continue; | |
749 | pg_shard_t peer = *a; | |
750 | map<pg_shard_t, pg_missing_t>::const_iterator pm = peer_missing.find(peer); | |
751 | if (pm == peer_missing.end()) { | |
752 | dout(10) << __func__ << " osd." << peer << " doesn't have missing set" | |
753 | << dendl; | |
754 | continue; | |
755 | } | |
756 | if (pm->second.num_missing()) { | |
757 | dout(10) << __func__ << " osd." << peer << " has " | |
758 | << pm->second.num_missing() << " missing" << dendl; | |
759 | return true; | |
760 | } | |
761 | } | |
762 | ||
763 | dout(10) << __func__ << " is recovered" << dendl; | |
764 | return false; | |
765 | } | |
766 | ||
767 | bool PG::needs_backfill() const | |
768 | { | |
769 | assert(is_primary()); | |
770 | ||
771 | // We can assume that only possible osds that need backfill | |
772 | // are on the backfill_targets vector nodes. | |
773 | set<pg_shard_t>::const_iterator end = backfill_targets.end(); | |
774 | set<pg_shard_t>::const_iterator a = backfill_targets.begin(); | |
775 | for (; a != end; ++a) { | |
776 | pg_shard_t peer = *a; | |
777 | map<pg_shard_t, pg_info_t>::const_iterator pi = peer_info.find(peer); | |
778 | if (!pi->second.last_backfill.is_max()) { | |
779 | dout(10) << __func__ << " osd." << peer << " has last_backfill " << pi->second.last_backfill << dendl; | |
780 | return true; | |
781 | } | |
782 | } | |
783 | ||
784 | dout(10) << __func__ << " does not need backfill" << dendl; | |
785 | return false; | |
786 | } | |
787 | ||
788 | ||
789 | void PG::check_past_interval_bounds() const | |
790 | { | |
791 | auto rpib = get_required_past_interval_bounds( | |
792 | info, | |
793 | osd->get_superblock().oldest_map); | |
794 | if (rpib.first >= rpib.second) { | |
795 | if (!past_intervals.empty()) { | |
796 | osd->clog->error() << info.pgid << " required past_interval bounds are" | |
797 | << " empty [" << rpib << ") but past_intervals is not: " | |
798 | << past_intervals; | |
799 | derr << info.pgid << " required past_interval bounds are" | |
800 | << " empty [" << rpib << ") but past_intervals is not: " | |
801 | << past_intervals << dendl; | |
802 | } | |
803 | } else { | |
804 | if (past_intervals.empty()) { | |
805 | osd->clog->error() << info.pgid << " required past_interval bounds are" | |
806 | << " not empty [" << rpib << ") but past_intervals " | |
807 | << past_intervals << " is empty"; | |
808 | derr << info.pgid << " required past_interval bounds are" | |
809 | << " not empty [" << rpib << ") but past_intervals " | |
810 | << past_intervals << " is empty" << dendl; | |
811 | assert(!past_intervals.empty()); | |
812 | } | |
813 | ||
814 | auto apib = past_intervals.get_bounds(); | |
815 | if (apib.first > rpib.first) { | |
816 | osd->clog->error() << info.pgid << " past_intervals [" << apib | |
817 | << ") start interval does not contain the required" | |
818 | << " bound [" << rpib << ") start"; | |
819 | derr << info.pgid << " past_intervals [" << apib | |
820 | << ") start interval does not contain the required" | |
821 | << " bound [" << rpib << ") start" << dendl; | |
822 | assert(0 == "past_interval start interval mismatch"); | |
823 | } | |
824 | if (apib.second != rpib.second) { | |
825 | osd->clog->error() << info.pgid << " past_interal bound [" << apib | |
826 | << ") end does not match required [" << rpib | |
827 | << ") end"; | |
828 | derr << info.pgid << " past_interal bound [" << apib | |
829 | << ") end does not match required [" << rpib | |
830 | << ") end" << dendl; | |
831 | assert(0 == "past_interval end mismatch"); | |
832 | } | |
833 | } | |
834 | } | |
835 | ||
836 | bool PG::adjust_need_up_thru(const OSDMapRef osdmap) | |
837 | { | |
838 | epoch_t up_thru = osdmap->get_up_thru(osd->whoami); | |
839 | if (need_up_thru && | |
840 | up_thru >= info.history.same_interval_since) { | |
841 | dout(10) << "adjust_need_up_thru now " << up_thru << ", need_up_thru now false" << dendl; | |
842 | need_up_thru = false; | |
843 | return true; | |
844 | } | |
845 | return false; | |
846 | } | |
847 | ||
848 | void PG::remove_down_peer_info(const OSDMapRef osdmap) | |
849 | { | |
850 | // Remove any downed osds from peer_info | |
851 | bool removed = false; | |
852 | map<pg_shard_t, pg_info_t>::iterator p = peer_info.begin(); | |
853 | while (p != peer_info.end()) { | |
854 | if (!osdmap->is_up(p->first.osd)) { | |
855 | dout(10) << " dropping down osd." << p->first << " info " << p->second << dendl; | |
856 | peer_missing.erase(p->first); | |
857 | peer_log_requested.erase(p->first); | |
858 | peer_missing_requested.erase(p->first); | |
859 | peer_info.erase(p++); | |
860 | removed = true; | |
861 | } else | |
862 | ++p; | |
863 | } | |
864 | ||
865 | // if we removed anyone, update peers (which include peer_info) | |
866 | if (removed) | |
867 | update_heartbeat_peers(); | |
868 | check_recovery_sources(osdmap); | |
869 | } | |
870 | ||
871 | /* | |
872 | * Returns true unless there is a non-lost OSD in might_have_unfound. | |
873 | */ | |
874 | bool PG::all_unfound_are_queried_or_lost(const OSDMapRef osdmap) const | |
875 | { | |
876 | assert(is_primary()); | |
877 | ||
878 | set<pg_shard_t>::const_iterator peer = might_have_unfound.begin(); | |
879 | set<pg_shard_t>::const_iterator mend = might_have_unfound.end(); | |
880 | for (; peer != mend; ++peer) { | |
881 | if (peer_missing.count(*peer)) | |
882 | continue; | |
883 | map<pg_shard_t, pg_info_t>::const_iterator iter = peer_info.find(*peer); | |
884 | if (iter != peer_info.end() && | |
885 | (iter->second.is_empty() || iter->second.dne())) | |
886 | continue; | |
887 | if (!osdmap->exists(peer->osd)) | |
888 | continue; | |
889 | const osd_info_t &osd_info(osdmap->get_info(peer->osd)); | |
890 | if (osd_info.lost_at <= osd_info.up_from) { | |
891 | // If there is even one OSD in might_have_unfound that isn't lost, we | |
892 | // still might retrieve our unfound. | |
893 | return false; | |
894 | } | |
895 | } | |
896 | dout(10) << "all_unfound_are_queried_or_lost all of might_have_unfound " << might_have_unfound | |
897 | << " have been queried or are marked lost" << dendl; | |
898 | return true; | |
899 | } | |
900 | ||
901 | PastIntervals::PriorSet PG::build_prior() | |
902 | { | |
903 | if (1) { | |
904 | // sanity check | |
905 | for (map<pg_shard_t,pg_info_t>::iterator it = peer_info.begin(); | |
906 | it != peer_info.end(); | |
907 | ++it) { | |
908 | assert(info.history.last_epoch_started >= it->second.history.last_epoch_started); | |
909 | } | |
910 | } | |
911 | ||
912 | const OSDMap &osdmap = *get_osdmap(); | |
913 | PastIntervals::PriorSet prior = past_intervals.get_prior_set( | |
914 | pool.info.ec_pool(), | |
915 | info.history.last_epoch_started, | |
916 | get_pgbackend()->get_is_recoverable_predicate(), | |
917 | [&](epoch_t start, int osd, epoch_t *lost_at) { | |
918 | const osd_info_t *pinfo = 0; | |
919 | if (osdmap.exists(osd)) { | |
920 | pinfo = &osdmap.get_info(osd); | |
921 | if (lost_at) | |
922 | *lost_at = pinfo->lost_at; | |
923 | } | |
924 | ||
925 | if (osdmap.is_up(osd)) { | |
926 | return PastIntervals::UP; | |
927 | } else if (!pinfo) { | |
928 | return PastIntervals::DNE; | |
929 | } else if (pinfo->lost_at > start) { | |
930 | return PastIntervals::LOST; | |
931 | } else { | |
932 | return PastIntervals::DOWN; | |
933 | } | |
934 | }, | |
935 | up, | |
936 | acting, | |
937 | this); | |
938 | ||
939 | if (prior.pg_down) { | |
940 | state_set(PG_STATE_DOWN); | |
941 | } | |
942 | ||
943 | if (get_osdmap()->get_up_thru(osd->whoami) < info.history.same_interval_since) { | |
944 | dout(10) << "up_thru " << get_osdmap()->get_up_thru(osd->whoami) | |
945 | << " < same_since " << info.history.same_interval_since | |
946 | << ", must notify monitor" << dendl; | |
947 | need_up_thru = true; | |
948 | } else { | |
949 | dout(10) << "up_thru " << get_osdmap()->get_up_thru(osd->whoami) | |
950 | << " >= same_since " << info.history.same_interval_since | |
951 | << ", all is well" << dendl; | |
952 | need_up_thru = false; | |
953 | } | |
954 | set_probe_targets(prior.probe); | |
955 | return prior; | |
956 | } | |
957 | ||
958 | void PG::clear_primary_state() | |
959 | { | |
960 | dout(10) << "clear_primary_state" << dendl; | |
961 | ||
962 | // clear peering state | |
963 | stray_set.clear(); | |
964 | peer_log_requested.clear(); | |
965 | peer_missing_requested.clear(); | |
966 | peer_info.clear(); | |
967 | peer_missing.clear(); | |
968 | need_up_thru = false; | |
969 | peer_last_complete_ondisk.clear(); | |
970 | peer_activated.clear(); | |
971 | min_last_complete_ondisk = eversion_t(); | |
972 | pg_trim_to = eversion_t(); | |
973 | might_have_unfound.clear(); | |
974 | projected_log = PGLog::IndexedLog(); | |
975 | ||
976 | last_update_ondisk = eversion_t(); | |
977 | ||
978 | snap_trimq.clear(); | |
979 | ||
980 | finish_sync_event = 0; // so that _finish_recovery doesn't go off in another thread | |
981 | ||
982 | missing_loc.clear(); | |
983 | ||
984 | release_pg_backoffs(); | |
985 | ||
986 | pg_log.reset_recovery_pointers(); | |
987 | ||
988 | scrubber.reserved_peers.clear(); | |
989 | scrub_after_recovery = false; | |
990 | ||
991 | agent_clear(); | |
992 | } | |
993 | ||
994 | PG::Scrubber::Scrubber() | |
995 | : reserved(false), reserve_failed(false), | |
996 | epoch_start(0), | |
997 | active(false), queue_snap_trim(false), | |
998 | waiting_on(0), shallow_errors(0), deep_errors(0), fixed(0), | |
999 | must_scrub(false), must_deep_scrub(false), must_repair(false), | |
1000 | auto_repair(false), | |
1001 | num_digest_updates_pending(0), | |
1002 | state(INACTIVE), | |
1003 | deep(false), | |
1004 | seed(0) | |
1005 | {} | |
1006 | ||
1007 | PG::Scrubber::~Scrubber() {} | |
1008 | ||
1009 | /** | |
1010 | * find_best_info | |
1011 | * | |
1012 | * Returns an iterator to the best info in infos sorted by: | |
1013 | * 1) Prefer newer last_update | |
1014 | * 2) Prefer longer tail if it brings another info into contiguity | |
1015 | * 3) Prefer current primary | |
1016 | */ | |
1017 | map<pg_shard_t, pg_info_t>::const_iterator PG::find_best_info( | |
1018 | const map<pg_shard_t, pg_info_t> &infos, | |
1019 | bool restrict_to_up_acting, | |
1020 | bool *history_les_bound) const | |
1021 | { | |
1022 | assert(history_les_bound); | |
1023 | /* See doc/dev/osd_internals/last_epoch_started.rst before attempting | |
1024 | * to make changes to this process. Also, make sure to update it | |
1025 | * when you find bugs! */ | |
1026 | eversion_t min_last_update_acceptable = eversion_t::max(); | |
1027 | epoch_t max_last_epoch_started_found = 0; | |
1028 | for (map<pg_shard_t, pg_info_t>::const_iterator i = infos.begin(); | |
1029 | i != infos.end(); | |
1030 | ++i) { | |
1031 | if (!cct->_conf->osd_find_best_info_ignore_history_les && | |
1032 | max_last_epoch_started_found < i->second.history.last_epoch_started) { | |
1033 | *history_les_bound = true; | |
1034 | max_last_epoch_started_found = i->second.history.last_epoch_started; | |
1035 | } | |
1036 | if (!i->second.is_incomplete() && | |
1037 | max_last_epoch_started_found < i->second.last_epoch_started) { | |
1038 | max_last_epoch_started_found = i->second.last_epoch_started; | |
1039 | } | |
1040 | } | |
1041 | for (map<pg_shard_t, pg_info_t>::const_iterator i = infos.begin(); | |
1042 | i != infos.end(); | |
1043 | ++i) { | |
1044 | if (max_last_epoch_started_found <= i->second.last_epoch_started) { | |
1045 | if (min_last_update_acceptable > i->second.last_update) | |
1046 | min_last_update_acceptable = i->second.last_update; | |
1047 | } | |
1048 | } | |
1049 | if (min_last_update_acceptable == eversion_t::max()) | |
1050 | return infos.end(); | |
1051 | ||
1052 | map<pg_shard_t, pg_info_t>::const_iterator best = infos.end(); | |
1053 | // find osd with newest last_update (oldest for ec_pool). | |
1054 | // if there are multiples, prefer | |
1055 | // - a longer tail, if it brings another peer into log contiguity | |
1056 | // - the current primary | |
1057 | for (map<pg_shard_t, pg_info_t>::const_iterator p = infos.begin(); | |
1058 | p != infos.end(); | |
1059 | ++p) { | |
1060 | if (restrict_to_up_acting && !is_up(p->first) && | |
1061 | !is_acting(p->first)) | |
1062 | continue; | |
1063 | // Only consider peers with last_update >= min_last_update_acceptable | |
1064 | if (p->second.last_update < min_last_update_acceptable) | |
1065 | continue; | |
1066 | // Disqualify anyone with a too old last_epoch_started | |
1067 | if (p->second.last_epoch_started < max_last_epoch_started_found) | |
1068 | continue; | |
1069 | // Disqualify anyone who is incomplete (not fully backfilled) | |
1070 | if (p->second.is_incomplete()) | |
1071 | continue; | |
1072 | if (best == infos.end()) { | |
1073 | best = p; | |
1074 | continue; | |
1075 | } | |
1076 | // Prefer newer last_update | |
1077 | if (pool.info.require_rollback()) { | |
1078 | if (p->second.last_update > best->second.last_update) | |
1079 | continue; | |
1080 | if (p->second.last_update < best->second.last_update) { | |
1081 | best = p; | |
1082 | continue; | |
1083 | } | |
1084 | } else { | |
1085 | if (p->second.last_update < best->second.last_update) | |
1086 | continue; | |
1087 | if (p->second.last_update > best->second.last_update) { | |
1088 | best = p; | |
1089 | continue; | |
1090 | } | |
1091 | } | |
1092 | ||
1093 | // Prefer longer tail | |
1094 | if (p->second.log_tail > best->second.log_tail) { | |
1095 | continue; | |
1096 | } else if (p->second.log_tail < best->second.log_tail) { | |
1097 | best = p; | |
1098 | continue; | |
1099 | } | |
1100 | ||
1101 | // prefer current primary (usually the caller), all things being equal | |
1102 | if (p->first == pg_whoami) { | |
1103 | dout(10) << "calc_acting prefer osd." << p->first | |
1104 | << " because it is current primary" << dendl; | |
1105 | best = p; | |
1106 | continue; | |
1107 | } | |
1108 | } | |
1109 | return best; | |
1110 | } | |
1111 | ||
1112 | void PG::calc_ec_acting( | |
1113 | map<pg_shard_t, pg_info_t>::const_iterator auth_log_shard, | |
1114 | unsigned size, | |
1115 | const vector<int> &acting, | |
1116 | pg_shard_t acting_primary, | |
1117 | const vector<int> &up, | |
1118 | pg_shard_t up_primary, | |
1119 | const map<pg_shard_t, pg_info_t> &all_info, | |
1120 | bool restrict_to_up_acting, | |
1121 | vector<int> *_want, | |
1122 | set<pg_shard_t> *backfill, | |
1123 | set<pg_shard_t> *acting_backfill, | |
1124 | pg_shard_t *want_primary, | |
1125 | ostream &ss) | |
1126 | { | |
1127 | vector<int> want(size, CRUSH_ITEM_NONE); | |
1128 | map<shard_id_t, set<pg_shard_t> > all_info_by_shard; | |
1129 | unsigned usable = 0; | |
1130 | for (map<pg_shard_t, pg_info_t>::const_iterator i = all_info.begin(); | |
1131 | i != all_info.end(); | |
1132 | ++i) { | |
1133 | all_info_by_shard[i->first.shard].insert(i->first); | |
1134 | } | |
1135 | for (uint8_t i = 0; i < want.size(); ++i) { | |
1136 | ss << "For position " << (unsigned)i << ": "; | |
1137 | if (up.size() > (unsigned)i && up[i] != CRUSH_ITEM_NONE && | |
1138 | !all_info.find(pg_shard_t(up[i], shard_id_t(i)))->second.is_incomplete() && | |
1139 | all_info.find(pg_shard_t(up[i], shard_id_t(i)))->second.last_update >= | |
1140 | auth_log_shard->second.log_tail) { | |
1141 | ss << " selecting up[i]: " << pg_shard_t(up[i], shard_id_t(i)) << std::endl; | |
1142 | want[i] = up[i]; | |
1143 | ++usable; | |
1144 | continue; | |
1145 | } | |
1146 | if (up.size() > (unsigned)i && up[i] != CRUSH_ITEM_NONE) { | |
1147 | ss << " backfilling up[i]: " << pg_shard_t(up[i], shard_id_t(i)) | |
1148 | << " and "; | |
1149 | backfill->insert(pg_shard_t(up[i], shard_id_t(i))); | |
1150 | } | |
1151 | ||
1152 | if (acting.size() > (unsigned)i && acting[i] != CRUSH_ITEM_NONE && | |
1153 | !all_info.find(pg_shard_t(acting[i], shard_id_t(i)))->second.is_incomplete() && | |
1154 | all_info.find(pg_shard_t(acting[i], shard_id_t(i)))->second.last_update >= | |
1155 | auth_log_shard->second.log_tail) { | |
1156 | ss << " selecting acting[i]: " << pg_shard_t(acting[i], shard_id_t(i)) << std::endl; | |
1157 | want[i] = acting[i]; | |
1158 | ++usable; | |
1159 | } else if (!restrict_to_up_acting) { | |
1160 | for (set<pg_shard_t>::iterator j = all_info_by_shard[shard_id_t(i)].begin(); | |
1161 | j != all_info_by_shard[shard_id_t(i)].end(); | |
1162 | ++j) { | |
1163 | assert(j->shard == i); | |
1164 | if (!all_info.find(*j)->second.is_incomplete() && | |
1165 | all_info.find(*j)->second.last_update >= | |
1166 | auth_log_shard->second.log_tail) { | |
1167 | ss << " selecting stray: " << *j << std::endl; | |
1168 | want[i] = j->osd; | |
1169 | ++usable; | |
1170 | break; | |
1171 | } | |
1172 | } | |
1173 | if (want[i] == CRUSH_ITEM_NONE) | |
1174 | ss << " failed to fill position " << (int)i << std::endl; | |
1175 | } | |
1176 | } | |
1177 | ||
1178 | bool found_primary = false; | |
1179 | for (uint8_t i = 0; i < want.size(); ++i) { | |
1180 | if (want[i] != CRUSH_ITEM_NONE) { | |
1181 | acting_backfill->insert(pg_shard_t(want[i], shard_id_t(i))); | |
1182 | if (!found_primary) { | |
1183 | *want_primary = pg_shard_t(want[i], shard_id_t(i)); | |
1184 | found_primary = true; | |
1185 | } | |
1186 | } | |
1187 | } | |
1188 | acting_backfill->insert(backfill->begin(), backfill->end()); | |
1189 | _want->swap(want); | |
1190 | } | |
1191 | ||
1192 | /** | |
1193 | * calculate the desired acting set. | |
1194 | * | |
1195 | * Choose an appropriate acting set. Prefer up[0], unless it is | |
1196 | * incomplete, or another osd has a longer tail that allows us to | |
1197 | * bring other up nodes up to date. | |
1198 | */ | |
1199 | void PG::calc_replicated_acting( | |
1200 | map<pg_shard_t, pg_info_t>::const_iterator auth_log_shard, | |
1201 | unsigned size, | |
1202 | const vector<int> &acting, | |
1203 | pg_shard_t acting_primary, | |
1204 | const vector<int> &up, | |
1205 | pg_shard_t up_primary, | |
1206 | const map<pg_shard_t, pg_info_t> &all_info, | |
1207 | bool restrict_to_up_acting, | |
1208 | vector<int> *want, | |
1209 | set<pg_shard_t> *backfill, | |
1210 | set<pg_shard_t> *acting_backfill, | |
1211 | pg_shard_t *want_primary, | |
1212 | ostream &ss) | |
1213 | { | |
1214 | ss << "calc_acting newest update on osd." << auth_log_shard->first | |
1215 | << " with " << auth_log_shard->second | |
1216 | << (restrict_to_up_acting ? " restrict_to_up_acting" : "") << std::endl; | |
1217 | pg_shard_t auth_log_shard_id = auth_log_shard->first; | |
1218 | ||
1219 | // select primary | |
1220 | map<pg_shard_t,pg_info_t>::const_iterator primary; | |
1221 | if (up.size() && | |
1222 | !all_info.find(up_primary)->second.is_incomplete() && | |
1223 | all_info.find(up_primary)->second.last_update >= | |
1224 | auth_log_shard->second.log_tail) { | |
1225 | ss << "up_primary: " << up_primary << ") selected as primary" << std::endl; | |
1226 | primary = all_info.find(up_primary); // prefer up[0], all thing being equal | |
1227 | } else { | |
1228 | assert(!auth_log_shard->second.is_incomplete()); | |
1229 | ss << "up[0] needs backfill, osd." << auth_log_shard_id | |
1230 | << " selected as primary instead" << std::endl; | |
1231 | primary = auth_log_shard; | |
1232 | } | |
1233 | ||
1234 | ss << "calc_acting primary is osd." << primary->first | |
1235 | << " with " << primary->second << std::endl; | |
1236 | *want_primary = primary->first; | |
1237 | want->push_back(primary->first.osd); | |
1238 | acting_backfill->insert(primary->first); | |
1239 | unsigned usable = 1; | |
1240 | ||
1241 | // select replicas that have log contiguity with primary. | |
1242 | // prefer up, then acting, then any peer_info osds | |
1243 | for (vector<int>::const_iterator i = up.begin(); | |
1244 | i != up.end(); | |
1245 | ++i) { | |
1246 | pg_shard_t up_cand = pg_shard_t(*i, shard_id_t::NO_SHARD); | |
1247 | if (up_cand == primary->first) | |
1248 | continue; | |
1249 | const pg_info_t &cur_info = all_info.find(up_cand)->second; | |
1250 | if (cur_info.is_incomplete() || | |
1251 | cur_info.last_update < MIN( | |
1252 | primary->second.log_tail, | |
1253 | auth_log_shard->second.log_tail)) { | |
1254 | /* We include auth_log_shard->second.log_tail because in GetLog, | |
1255 | * we will request logs back to the min last_update over our | |
1256 | * acting_backfill set, which will result in our log being extended | |
1257 | * as far backwards as necessary to pick up any peers which can | |
1258 | * be log recovered by auth_log_shard's log */ | |
1259 | ss << " shard " << up_cand << " (up) backfill " << cur_info << std::endl; | |
1260 | backfill->insert(up_cand); | |
1261 | acting_backfill->insert(up_cand); | |
1262 | } else { | |
1263 | want->push_back(*i); | |
1264 | acting_backfill->insert(up_cand); | |
1265 | usable++; | |
1266 | ss << " osd." << *i << " (up) accepted " << cur_info << std::endl; | |
1267 | } | |
1268 | } | |
1269 | ||
1270 | // This no longer has backfill OSDs, but they are covered above. | |
1271 | for (vector<int>::const_iterator i = acting.begin(); | |
1272 | i != acting.end(); | |
1273 | ++i) { | |
1274 | pg_shard_t acting_cand(*i, shard_id_t::NO_SHARD); | |
1275 | if (usable >= size) | |
1276 | break; | |
1277 | ||
1278 | // skip up osds we already considered above | |
1279 | if (acting_cand == primary->first) | |
1280 | continue; | |
1281 | vector<int>::const_iterator up_it = find(up.begin(), up.end(), acting_cand.osd); | |
1282 | if (up_it != up.end()) | |
1283 | continue; | |
1284 | ||
1285 | const pg_info_t &cur_info = all_info.find(acting_cand)->second; | |
1286 | if (cur_info.is_incomplete() || | |
1287 | cur_info.last_update < primary->second.log_tail) { | |
1288 | ss << " shard " << acting_cand << " (stray) REJECTED " | |
1289 | << cur_info << std::endl; | |
1290 | } else { | |
1291 | want->push_back(*i); | |
1292 | acting_backfill->insert(acting_cand); | |
1293 | ss << " shard " << acting_cand << " (stray) accepted " | |
1294 | << cur_info << std::endl; | |
1295 | usable++; | |
1296 | } | |
1297 | } | |
1298 | ||
1299 | if (restrict_to_up_acting) { | |
1300 | return; | |
1301 | } | |
1302 | for (map<pg_shard_t,pg_info_t>::const_iterator i = all_info.begin(); | |
1303 | i != all_info.end(); | |
1304 | ++i) { | |
1305 | if (usable >= size) | |
1306 | break; | |
1307 | ||
1308 | // skip up osds we already considered above | |
1309 | if (i->first == primary->first) | |
1310 | continue; | |
1311 | vector<int>::const_iterator up_it = find(up.begin(), up.end(), i->first.osd); | |
1312 | if (up_it != up.end()) | |
1313 | continue; | |
1314 | vector<int>::const_iterator acting_it = find( | |
1315 | acting.begin(), acting.end(), i->first.osd); | |
1316 | if (acting_it != acting.end()) | |
1317 | continue; | |
1318 | ||
1319 | if (i->second.is_incomplete() || | |
1320 | i->second.last_update < primary->second.log_tail) { | |
1321 | ss << " shard " << i->first << " (stray) REJECTED " | |
1322 | << i->second << std::endl; | |
1323 | } else { | |
1324 | want->push_back(i->first.osd); | |
1325 | acting_backfill->insert(i->first); | |
1326 | ss << " shard " << i->first << " (stray) accepted " | |
1327 | << i->second << std::endl; | |
1328 | usable++; | |
1329 | } | |
1330 | } | |
1331 | } | |
1332 | ||
1333 | /** | |
1334 | * choose acting | |
1335 | * | |
1336 | * calculate the desired acting, and request a change with the monitor | |
1337 | * if it differs from the current acting. | |
1338 | * | |
1339 | * if restrict_to_up_acting=true, we filter out anything that's not in | |
1340 | * up/acting. in order to lift this restriction, we need to | |
1341 | * 1) check whether it's worth switching the acting set any time we get | |
1342 | * a new pg info (not just here, when recovery finishes) | |
1343 | * 2) check whether anything in want_acting went down on each new map | |
1344 | * (and, if so, calculate a new want_acting) | |
1345 | * 3) remove the assertion in PG::RecoveryState::Active::react(const AdvMap) | |
1346 | * TODO! | |
1347 | */ | |
1348 | bool PG::choose_acting(pg_shard_t &auth_log_shard_id, | |
1349 | bool restrict_to_up_acting, | |
1350 | bool *history_les_bound) | |
1351 | { | |
1352 | map<pg_shard_t, pg_info_t> all_info(peer_info.begin(), peer_info.end()); | |
1353 | all_info[pg_whoami] = info; | |
1354 | ||
1355 | for (map<pg_shard_t, pg_info_t>::iterator p = all_info.begin(); | |
1356 | p != all_info.end(); | |
1357 | ++p) { | |
1358 | dout(10) << "calc_acting osd." << p->first << " " << p->second << dendl; | |
1359 | } | |
1360 | ||
1361 | map<pg_shard_t, pg_info_t>::const_iterator auth_log_shard = | |
1362 | find_best_info(all_info, restrict_to_up_acting, history_les_bound); | |
1363 | ||
1364 | if (auth_log_shard == all_info.end()) { | |
1365 | if (up != acting) { | |
1366 | dout(10) << "choose_acting no suitable info found (incomplete backfills?)," | |
1367 | << " reverting to up" << dendl; | |
1368 | want_acting = up; | |
1369 | vector<int> empty; | |
1370 | osd->queue_want_pg_temp(info.pgid.pgid, empty); | |
1371 | } else { | |
1372 | dout(10) << "choose_acting failed" << dendl; | |
1373 | assert(want_acting.empty()); | |
1374 | } | |
1375 | return false; | |
1376 | } | |
1377 | ||
1378 | assert(!auth_log_shard->second.is_incomplete()); | |
1379 | auth_log_shard_id = auth_log_shard->first; | |
1380 | ||
1381 | set<pg_shard_t> want_backfill, want_acting_backfill; | |
1382 | vector<int> want; | |
1383 | pg_shard_t want_primary; | |
1384 | stringstream ss; | |
1385 | if (!pool.info.ec_pool()) | |
1386 | calc_replicated_acting( | |
1387 | auth_log_shard, | |
1388 | get_osdmap()->get_pg_size(info.pgid.pgid), | |
1389 | acting, | |
1390 | primary, | |
1391 | up, | |
1392 | up_primary, | |
1393 | all_info, | |
1394 | restrict_to_up_acting, | |
1395 | &want, | |
1396 | &want_backfill, | |
1397 | &want_acting_backfill, | |
1398 | &want_primary, | |
1399 | ss); | |
1400 | else | |
1401 | calc_ec_acting( | |
1402 | auth_log_shard, | |
1403 | get_osdmap()->get_pg_size(info.pgid.pgid), | |
1404 | acting, | |
1405 | primary, | |
1406 | up, | |
1407 | up_primary, | |
1408 | all_info, | |
1409 | restrict_to_up_acting, | |
1410 | &want, | |
1411 | &want_backfill, | |
1412 | &want_acting_backfill, | |
1413 | &want_primary, | |
1414 | ss); | |
1415 | dout(10) << ss.str() << dendl; | |
1416 | ||
1417 | unsigned num_want_acting = 0; | |
1418 | set<pg_shard_t> have; | |
1419 | for (int i = 0; i < (int)want.size(); ++i) { | |
1420 | if (want[i] != CRUSH_ITEM_NONE) { | |
1421 | ++num_want_acting; | |
1422 | have.insert( | |
1423 | pg_shard_t( | |
1424 | want[i], | |
1425 | pool.info.ec_pool() ? shard_id_t(i) : shard_id_t::NO_SHARD)); | |
1426 | } | |
1427 | } | |
1428 | ||
1429 | // We go incomplete if below min_size for ec_pools since backfill | |
1430 | // does not currently maintain rollbackability | |
1431 | // Otherwise, we will go "peered", but not "active" | |
1432 | if (num_want_acting < pool.info.min_size && | |
1433 | (pool.info.ec_pool() || | |
1434 | !cct->_conf->osd_allow_recovery_below_min_size)) { | |
1435 | want_acting.clear(); | |
1436 | dout(10) << "choose_acting failed, below min size" << dendl; | |
1437 | return false; | |
1438 | } | |
1439 | ||
1440 | /* Check whether we have enough acting shards to later perform recovery */ | |
1441 | boost::scoped_ptr<IsPGRecoverablePredicate> recoverable_predicate( | |
1442 | get_pgbackend()->get_is_recoverable_predicate()); | |
1443 | if (!(*recoverable_predicate)(have)) { | |
1444 | want_acting.clear(); | |
1445 | dout(10) << "choose_acting failed, not recoverable" << dendl; | |
1446 | return false; | |
1447 | } | |
1448 | ||
1449 | if (want != acting) { | |
1450 | dout(10) << "choose_acting want " << want << " != acting " << acting | |
1451 | << ", requesting pg_temp change" << dendl; | |
1452 | want_acting = want; | |
1453 | ||
1454 | if (want_acting == up) { | |
1455 | // There can't be any pending backfill if | |
1456 | // want is the same as crush map up OSDs. | |
1457 | assert(want_backfill.empty()); | |
1458 | vector<int> empty; | |
1459 | osd->queue_want_pg_temp(info.pgid.pgid, empty); | |
1460 | } else | |
1461 | osd->queue_want_pg_temp(info.pgid.pgid, want); | |
1462 | return false; | |
1463 | } | |
1464 | want_acting.clear(); | |
1465 | actingbackfill = want_acting_backfill; | |
1466 | dout(10) << "actingbackfill is " << actingbackfill << dendl; | |
1467 | assert(backfill_targets.empty() || backfill_targets == want_backfill); | |
1468 | if (backfill_targets.empty()) { | |
1469 | // Caller is GetInfo | |
1470 | backfill_targets = want_backfill; | |
1471 | } | |
1472 | // Will not change if already set because up would have had to change | |
1473 | // Verify that nothing in backfill is in stray_set | |
1474 | for (set<pg_shard_t>::iterator i = want_backfill.begin(); | |
1475 | i != want_backfill.end(); | |
1476 | ++i) { | |
1477 | assert(stray_set.find(*i) == stray_set.end()); | |
1478 | } | |
1479 | dout(10) << "choose_acting want " << want << " (== acting) backfill_targets " | |
1480 | << want_backfill << dendl; | |
1481 | return true; | |
1482 | } | |
1483 | ||
1484 | /* Build the might_have_unfound set. | |
1485 | * | |
1486 | * This is used by the primary OSD during recovery. | |
1487 | * | |
1488 | * This set tracks the OSDs which might have unfound objects that the primary | |
1489 | * OSD needs. As we receive pg_missing_t from each OSD in might_have_unfound, we | |
1490 | * will remove the OSD from the set. | |
1491 | */ | |
1492 | void PG::build_might_have_unfound() | |
1493 | { | |
1494 | assert(might_have_unfound.empty()); | |
1495 | assert(is_primary()); | |
1496 | ||
1497 | dout(10) << __func__ << dendl; | |
1498 | ||
1499 | check_past_interval_bounds(); | |
1500 | ||
1501 | might_have_unfound = past_intervals.get_might_have_unfound( | |
1502 | pg_whoami, | |
1503 | pool.info.ec_pool()); | |
1504 | ||
1505 | // include any (stray) peers | |
1506 | for (map<pg_shard_t, pg_info_t>::iterator p = peer_info.begin(); | |
1507 | p != peer_info.end(); | |
1508 | ++p) | |
1509 | might_have_unfound.insert(p->first); | |
1510 | ||
1511 | dout(15) << __func__ << ": built " << might_have_unfound << dendl; | |
1512 | } | |
1513 | ||
1514 | struct C_PG_ActivateCommitted : public Context { | |
1515 | PGRef pg; | |
1516 | epoch_t epoch; | |
1517 | epoch_t activation_epoch; | |
1518 | C_PG_ActivateCommitted(PG *p, epoch_t e, epoch_t ae) | |
1519 | : pg(p), epoch(e), activation_epoch(ae) {} | |
1520 | void finish(int r) override { | |
1521 | pg->_activate_committed(epoch, activation_epoch); | |
1522 | } | |
1523 | }; | |
1524 | ||
1525 | void PG::activate(ObjectStore::Transaction& t, | |
1526 | epoch_t activation_epoch, | |
1527 | list<Context*>& tfin, | |
1528 | map<int, map<spg_t,pg_query_t> >& query_map, | |
1529 | map<int, | |
1530 | vector< | |
1531 | pair<pg_notify_t, | |
1532 | PastIntervals> > > *activator_map, | |
1533 | RecoveryCtx *ctx) | |
1534 | { | |
1535 | assert(!is_peered()); | |
1536 | assert(scrubber.callbacks.empty()); | |
1537 | assert(callbacks_for_degraded_object.empty()); | |
1538 | ||
1539 | // twiddle pg state | |
1540 | state_clear(PG_STATE_DOWN); | |
1541 | ||
1542 | send_notify = false; | |
1543 | ||
1544 | if (is_primary()) { | |
1545 | // only update primary last_epoch_started if we will go active | |
1546 | if (acting.size() >= pool.info.min_size) { | |
1547 | assert(cct->_conf->osd_find_best_info_ignore_history_les || | |
1548 | info.last_epoch_started <= activation_epoch); | |
1549 | info.last_epoch_started = activation_epoch; | |
1550 | info.last_interval_started = info.history.same_interval_since; | |
1551 | } | |
1552 | } else if (is_acting(pg_whoami)) { | |
1553 | /* update last_epoch_started on acting replica to whatever the primary sent | |
1554 | * unless it's smaller (could happen if we are going peered rather than | |
1555 | * active, see doc/dev/osd_internals/last_epoch_started.rst) */ | |
1556 | if (info.last_epoch_started < activation_epoch) { | |
1557 | info.last_epoch_started = activation_epoch; | |
1558 | info.last_interval_started = info.history.same_interval_since; | |
1559 | } | |
1560 | } | |
1561 | ||
1562 | auto &missing = pg_log.get_missing(); | |
1563 | ||
1564 | if (is_primary()) { | |
1565 | last_update_ondisk = info.last_update; | |
1566 | min_last_complete_ondisk = eversion_t(0,0); // we don't know (yet)! | |
1567 | } | |
1568 | last_update_applied = info.last_update; | |
1569 | last_rollback_info_trimmed_to_applied = pg_log.get_can_rollback_to(); | |
1570 | ||
1571 | need_up_thru = false; | |
1572 | ||
1573 | // write pg info, log | |
1574 | dirty_info = true; | |
1575 | dirty_big_info = true; // maybe | |
1576 | ||
1577 | // find out when we commit | |
1578 | t.register_on_complete( | |
1579 | new C_PG_ActivateCommitted( | |
1580 | this, | |
1581 | get_osdmap()->get_epoch(), | |
1582 | activation_epoch)); | |
1583 | ||
1584 | // initialize snap_trimq | |
1585 | if (is_primary()) { | |
1586 | dout(20) << "activate - purged_snaps " << info.purged_snaps | |
1587 | << " cached_removed_snaps " << pool.cached_removed_snaps << dendl; | |
1588 | snap_trimq = pool.cached_removed_snaps; | |
1589 | interval_set<snapid_t> intersection; | |
1590 | intersection.intersection_of(snap_trimq, info.purged_snaps); | |
1591 | if (intersection == info.purged_snaps) { | |
1592 | snap_trimq.subtract(info.purged_snaps); | |
1593 | } else { | |
1594 | dout(0) << "warning: info.purged_snaps (" << info.purged_snaps | |
1595 | << ") is not a subset of pool.cached_removed_snaps (" | |
1596 | << pool.cached_removed_snaps << ")" << dendl; | |
1597 | snap_trimq.subtract(intersection); | |
1598 | } | |
1599 | } | |
1600 | ||
1601 | // init complete pointer | |
1602 | if (missing.num_missing() == 0) { | |
1603 | dout(10) << "activate - no missing, moving last_complete " << info.last_complete | |
1604 | << " -> " << info.last_update << dendl; | |
1605 | info.last_complete = info.last_update; | |
1606 | pg_log.reset_recovery_pointers(); | |
1607 | } else { | |
1608 | dout(10) << "activate - not complete, " << missing << dendl; | |
1609 | pg_log.activate_not_complete(info); | |
1610 | } | |
1611 | ||
1612 | log_weirdness(); | |
1613 | ||
1614 | // if primary.. | |
1615 | if (is_primary()) { | |
1616 | assert(ctx); | |
1617 | // start up replicas | |
1618 | ||
1619 | assert(!actingbackfill.empty()); | |
1620 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
1621 | i != actingbackfill.end(); | |
1622 | ++i) { | |
1623 | if (*i == pg_whoami) continue; | |
1624 | pg_shard_t peer = *i; | |
1625 | assert(peer_info.count(peer)); | |
1626 | pg_info_t& pi = peer_info[peer]; | |
1627 | ||
1628 | dout(10) << "activate peer osd." << peer << " " << pi << dendl; | |
1629 | ||
1630 | MOSDPGLog *m = 0; | |
1631 | pg_missing_t& pm = peer_missing[peer]; | |
1632 | ||
1633 | bool needs_past_intervals = pi.dne(); | |
1634 | ||
1635 | /* | |
1636 | * cover case where peer sort order was different and | |
1637 | * last_backfill cannot be interpreted | |
1638 | */ | |
1639 | bool force_restart_backfill = | |
1640 | !pi.last_backfill.is_max() && | |
1641 | !pi.last_backfill_bitwise; | |
1642 | ||
1643 | if (pi.last_update == info.last_update && !force_restart_backfill) { | |
1644 | // empty log | |
1645 | if (!pi.last_backfill.is_max()) | |
1646 | osd->clog->info() << info.pgid << " continuing backfill to osd." | |
1647 | << peer | |
1648 | << " from (" << pi.log_tail << "," << pi.last_update | |
1649 | << "] " << pi.last_backfill | |
1650 | << " to " << info.last_update; | |
1651 | if (!pi.is_empty() && activator_map) { | |
1652 | dout(10) << "activate peer osd." << peer << " is up to date, queueing in pending_activators" << dendl; | |
1653 | (*activator_map)[peer.osd].push_back( | |
1654 | make_pair( | |
1655 | pg_notify_t( | |
1656 | peer.shard, pg_whoami.shard, | |
1657 | get_osdmap()->get_epoch(), | |
1658 | get_osdmap()->get_epoch(), | |
1659 | info), | |
1660 | past_intervals)); | |
1661 | } else { | |
1662 | dout(10) << "activate peer osd." << peer << " is up to date, but sending pg_log anyway" << dendl; | |
1663 | m = new MOSDPGLog( | |
1664 | i->shard, pg_whoami.shard, | |
1665 | get_osdmap()->get_epoch(), info); | |
1666 | } | |
1667 | } else if ( | |
1668 | pg_log.get_tail() > pi.last_update || | |
1669 | pi.last_backfill == hobject_t() || | |
1670 | force_restart_backfill || | |
1671 | (backfill_targets.count(*i) && pi.last_backfill.is_max())) { | |
1672 | /* ^ This last case covers a situation where a replica is not contiguous | |
1673 | * with the auth_log, but is contiguous with this replica. Reshuffling | |
1674 | * the active set to handle this would be tricky, so instead we just go | |
1675 | * ahead and backfill it anyway. This is probably preferrable in any | |
1676 | * case since the replica in question would have to be significantly | |
1677 | * behind. | |
1678 | */ | |
1679 | // backfill | |
1680 | osd->clog->info() << info.pgid << " starting backfill to osd." << peer | |
1681 | << " from (" << pi.log_tail << "," << pi.last_update | |
1682 | << "] " << pi.last_backfill | |
1683 | << " to " << info.last_update; | |
1684 | ||
1685 | pi.last_update = info.last_update; | |
1686 | pi.last_complete = info.last_update; | |
1687 | pi.set_last_backfill(hobject_t()); | |
1688 | pi.last_epoch_started = info.last_epoch_started; | |
1689 | pi.last_interval_started = info.last_interval_started; | |
1690 | pi.history = info.history; | |
1691 | pi.hit_set = info.hit_set; | |
1692 | pi.stats.stats.clear(); | |
1693 | ||
1694 | // initialize peer with our purged_snaps. | |
1695 | pi.purged_snaps = info.purged_snaps; | |
1696 | ||
1697 | m = new MOSDPGLog( | |
1698 | i->shard, pg_whoami.shard, | |
1699 | get_osdmap()->get_epoch(), pi); | |
1700 | ||
1701 | // send some recent log, so that op dup detection works well. | |
1702 | m->log.copy_up_to(pg_log.get_log(), cct->_conf->osd_min_pg_log_entries); | |
1703 | m->info.log_tail = m->log.tail; | |
1704 | pi.log_tail = m->log.tail; // sigh... | |
1705 | ||
1706 | pm.clear(); | |
1707 | } else { | |
1708 | // catch up | |
1709 | assert(pg_log.get_tail() <= pi.last_update); | |
1710 | m = new MOSDPGLog( | |
1711 | i->shard, pg_whoami.shard, | |
1712 | get_osdmap()->get_epoch(), info); | |
1713 | // send new stuff to append to replicas log | |
1714 | m->log.copy_after(pg_log.get_log(), pi.last_update); | |
1715 | } | |
1716 | ||
1717 | // share past_intervals if we are creating the pg on the replica | |
1718 | // based on whether our info for that peer was dne() *before* | |
1719 | // updating pi.history in the backfill block above. | |
1720 | if (m && needs_past_intervals) | |
1721 | m->past_intervals = past_intervals; | |
1722 | ||
1723 | // update local version of peer's missing list! | |
1724 | if (m && pi.last_backfill != hobject_t()) { | |
1725 | for (list<pg_log_entry_t>::iterator p = m->log.log.begin(); | |
1726 | p != m->log.log.end(); | |
1727 | ++p) | |
1728 | if (p->soid <= pi.last_backfill && | |
1729 | !p->is_error()) | |
1730 | pm.add_next_event(*p); | |
1731 | } | |
1732 | ||
1733 | if (m) { | |
1734 | dout(10) << "activate peer osd." << peer << " sending " << m->log << dendl; | |
1735 | //m->log.print(cout); | |
1736 | osd->send_message_osd_cluster(peer.osd, m, get_osdmap()->get_epoch()); | |
1737 | } | |
1738 | ||
1739 | // peer now has | |
1740 | pi.last_update = info.last_update; | |
1741 | ||
1742 | // update our missing | |
1743 | if (pm.num_missing() == 0) { | |
1744 | pi.last_complete = pi.last_update; | |
1745 | dout(10) << "activate peer osd." << peer << " " << pi << " uptodate" << dendl; | |
1746 | } else { | |
1747 | dout(10) << "activate peer osd." << peer << " " << pi << " missing " << pm << dendl; | |
1748 | } | |
1749 | } | |
1750 | ||
1751 | // Set up missing_loc | |
1752 | set<pg_shard_t> complete_shards; | |
1753 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
1754 | i != actingbackfill.end(); | |
1755 | ++i) { | |
1756 | if (*i == get_primary()) { | |
1757 | missing_loc.add_active_missing(missing); | |
1758 | if (!missing.have_missing()) | |
1759 | complete_shards.insert(*i); | |
1760 | } else { | |
1761 | auto peer_missing_entry = peer_missing.find(*i); | |
1762 | assert(peer_missing_entry != peer_missing.end()); | |
1763 | missing_loc.add_active_missing(peer_missing_entry->second); | |
1764 | if (!peer_missing_entry->second.have_missing() && | |
1765 | peer_info[*i].last_backfill.is_max()) | |
1766 | complete_shards.insert(*i); | |
1767 | } | |
1768 | } | |
1769 | // If necessary, create might_have_unfound to help us find our unfound objects. | |
1770 | // NOTE: It's important that we build might_have_unfound before trimming the | |
1771 | // past intervals. | |
1772 | might_have_unfound.clear(); | |
1773 | if (needs_recovery()) { | |
1774 | // If only one shard has missing, we do a trick to add all others as recovery | |
1775 | // source, this is considered safe since the PGLogs have been merged locally, | |
1776 | // and covers vast majority of the use cases, like one OSD/host is down for | |
1777 | // a while for hardware repairing | |
1778 | if (complete_shards.size() + 1 == actingbackfill.size()) { | |
1779 | missing_loc.add_batch_sources_info(complete_shards, ctx->handle); | |
1780 | } else { | |
1781 | missing_loc.add_source_info(pg_whoami, info, pg_log.get_missing(), | |
1782 | ctx->handle); | |
1783 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
1784 | i != actingbackfill.end(); | |
1785 | ++i) { | |
1786 | if (*i == pg_whoami) continue; | |
1787 | dout(10) << __func__ << ": adding " << *i << " as a source" << dendl; | |
1788 | assert(peer_missing.count(*i)); | |
1789 | assert(peer_info.count(*i)); | |
1790 | missing_loc.add_source_info( | |
1791 | *i, | |
1792 | peer_info[*i], | |
1793 | peer_missing[*i], | |
1794 | ctx->handle); | |
1795 | } | |
1796 | } | |
1797 | for (map<pg_shard_t, pg_missing_t>::iterator i = peer_missing.begin(); | |
1798 | i != peer_missing.end(); | |
1799 | ++i) { | |
1800 | if (is_actingbackfill(i->first)) | |
1801 | continue; | |
1802 | assert(peer_info.count(i->first)); | |
1803 | search_for_missing( | |
1804 | peer_info[i->first], | |
1805 | i->second, | |
1806 | i->first, | |
1807 | ctx); | |
1808 | } | |
1809 | ||
1810 | build_might_have_unfound(); | |
1811 | ||
1812 | state_set(PG_STATE_DEGRADED); | |
1813 | if (have_unfound()) | |
1814 | discover_all_missing(query_map); | |
1815 | } | |
1816 | ||
1817 | // degraded? | |
1818 | if (get_osdmap()->get_pg_size(info.pgid.pgid) > actingset.size()) { | |
1819 | state_set(PG_STATE_DEGRADED); | |
1820 | state_set(PG_STATE_UNDERSIZED); | |
1821 | } | |
1822 | ||
1823 | state_set(PG_STATE_ACTIVATING); | |
1824 | release_pg_backoffs(); | |
1825 | projected_last_update = info.last_update; | |
1826 | } | |
1827 | if (acting.size() >= pool.info.min_size) { | |
1828 | PGLogEntryHandler handler{this, &t}; | |
1829 | pg_log.roll_forward(&handler); | |
1830 | } | |
1831 | } | |
1832 | ||
1833 | bool PG::op_has_sufficient_caps(OpRequestRef& op) | |
1834 | { | |
1835 | // only check MOSDOp | |
1836 | if (op->get_req()->get_type() != CEPH_MSG_OSD_OP) | |
1837 | return true; | |
1838 | ||
1839 | const MOSDOp *req = static_cast<const MOSDOp*>(op->get_req()); | |
1840 | ||
1841 | Session *session = static_cast<Session*>(req->get_connection()->get_priv()); | |
1842 | if (!session) { | |
1843 | dout(0) << "op_has_sufficient_caps: no session for op " << *req << dendl; | |
1844 | return false; | |
1845 | } | |
1846 | OSDCap& caps = session->caps; | |
1847 | session->put(); | |
1848 | ||
1849 | const string &key = req->get_hobj().get_key().empty() ? | |
1850 | req->get_oid().name : | |
1851 | req->get_hobj().get_key(); | |
1852 | ||
1853 | bool cap = caps.is_capable(pool.name, req->get_hobj().nspace, | |
1854 | pool.auid, key, | |
1855 | op->need_read_cap(), | |
1856 | op->need_write_cap(), | |
1857 | op->classes()); | |
1858 | ||
1859 | dout(20) << "op_has_sufficient_caps pool=" << pool.id << " (" << pool.name | |
1860 | << " " << req->get_hobj().nspace | |
1861 | << ") owner=" << pool.auid | |
1862 | << " need_read_cap=" << op->need_read_cap() | |
1863 | << " need_write_cap=" << op->need_write_cap() | |
1864 | << " classes=" << op->classes() | |
1865 | << " -> " << (cap ? "yes" : "NO") | |
1866 | << dendl; | |
1867 | return cap; | |
1868 | } | |
1869 | ||
1870 | void PG::_activate_committed(epoch_t epoch, epoch_t activation_epoch) | |
1871 | { | |
1872 | lock(); | |
1873 | if (pg_has_reset_since(epoch)) { | |
1874 | dout(10) << "_activate_committed " << epoch | |
1875 | << ", that was an old interval" << dendl; | |
1876 | } else if (is_primary()) { | |
1877 | peer_activated.insert(pg_whoami); | |
1878 | dout(10) << "_activate_committed " << epoch | |
1879 | << " peer_activated now " << peer_activated | |
1880 | << " last_interval_started " << info.history.last_interval_started | |
1881 | << " last_epoch_started " << info.history.last_epoch_started | |
1882 | << " same_interval_since " << info.history.same_interval_since << dendl; | |
1883 | assert(!actingbackfill.empty()); | |
1884 | if (peer_activated.size() == actingbackfill.size()) | |
1885 | all_activated_and_committed(); | |
1886 | } else { | |
1887 | dout(10) << "_activate_committed " << epoch << " telling primary" << dendl; | |
1888 | MOSDPGInfo *m = new MOSDPGInfo(epoch); | |
1889 | pg_notify_t i = pg_notify_t( | |
1890 | get_primary().shard, pg_whoami.shard, | |
1891 | get_osdmap()->get_epoch(), | |
1892 | get_osdmap()->get_epoch(), | |
1893 | info); | |
1894 | ||
1895 | i.info.history.last_epoch_started = activation_epoch; | |
1896 | i.info.history.last_interval_started = i.info.history.same_interval_since; | |
1897 | if (acting.size() >= pool.info.min_size) { | |
1898 | state_set(PG_STATE_ACTIVE); | |
1899 | } else { | |
1900 | state_set(PG_STATE_PEERED); | |
1901 | } | |
1902 | ||
1903 | m->pg_list.push_back(make_pair(i, PastIntervals())); | |
1904 | osd->send_message_osd_cluster(get_primary().osd, m, get_osdmap()->get_epoch()); | |
1905 | ||
1906 | // waiters | |
1907 | if (flushes_in_progress == 0) { | |
1908 | requeue_ops(waiting_for_peered); | |
1909 | } | |
1910 | } | |
1911 | ||
1912 | assert(!dirty_info); | |
1913 | ||
1914 | unlock(); | |
1915 | } | |
1916 | ||
1917 | /* | |
1918 | * update info.history.last_epoch_started ONLY after we and all | |
1919 | * replicas have activated AND committed the activate transaction | |
1920 | * (i.e. the peering results are stable on disk). | |
1921 | */ | |
1922 | void PG::all_activated_and_committed() | |
1923 | { | |
1924 | dout(10) << "all_activated_and_committed" << dendl; | |
1925 | assert(is_primary()); | |
1926 | assert(peer_activated.size() == actingbackfill.size()); | |
1927 | assert(!actingbackfill.empty()); | |
1928 | assert(blocked_by.empty()); | |
1929 | ||
1930 | queue_peering_event( | |
1931 | CephPeeringEvtRef( | |
1932 | std::make_shared<CephPeeringEvt>( | |
1933 | get_osdmap()->get_epoch(), | |
1934 | get_osdmap()->get_epoch(), | |
1935 | AllReplicasActivated()))); | |
1936 | } | |
1937 | ||
1938 | bool PG::requeue_scrub(bool high_priority) | |
1939 | { | |
1940 | assert(is_locked()); | |
1941 | if (scrub_queued) { | |
1942 | dout(10) << __func__ << ": already queued" << dendl; | |
1943 | return false; | |
1944 | } else { | |
1945 | dout(10) << __func__ << ": queueing" << dendl; | |
1946 | scrub_queued = true; | |
1947 | osd->queue_for_scrub(this, high_priority); | |
1948 | return true; | |
1949 | } | |
1950 | } | |
1951 | ||
1952 | void PG::queue_recovery(bool front) | |
1953 | { | |
1954 | if (!is_primary() || !is_peered()) { | |
1955 | dout(10) << "queue_recovery -- not primary or not peered " << dendl; | |
1956 | assert(!recovery_queued); | |
1957 | } else if (recovery_queued) { | |
1958 | dout(10) << "queue_recovery -- already queued" << dendl; | |
1959 | } else { | |
1960 | dout(10) << "queue_recovery -- queuing" << dendl; | |
1961 | recovery_queued = true; | |
1962 | osd->queue_for_recovery(this, front); | |
1963 | } | |
1964 | } | |
1965 | ||
1966 | bool PG::queue_scrub() | |
1967 | { | |
1968 | assert(is_locked()); | |
1969 | if (is_scrubbing()) { | |
1970 | return false; | |
1971 | } | |
1972 | scrubber.priority = scrubber.must_scrub ? | |
1973 | cct->_conf->osd_requested_scrub_priority : get_scrub_priority(); | |
1974 | scrubber.must_scrub = false; | |
1975 | state_set(PG_STATE_SCRUBBING); | |
1976 | if (scrubber.must_deep_scrub) { | |
1977 | state_set(PG_STATE_DEEP_SCRUB); | |
1978 | scrubber.must_deep_scrub = false; | |
1979 | } | |
1980 | if (scrubber.must_repair || scrubber.auto_repair) { | |
1981 | state_set(PG_STATE_REPAIR); | |
1982 | scrubber.must_repair = false; | |
1983 | } | |
1984 | requeue_scrub(); | |
1985 | return true; | |
1986 | } | |
1987 | ||
1988 | unsigned PG::get_scrub_priority() | |
1989 | { | |
1990 | // a higher value -> a higher priority | |
1991 | int pool_scrub_priority = 0; | |
1992 | pool.info.opts.get(pool_opts_t::SCRUB_PRIORITY, &pool_scrub_priority); | |
1993 | return pool_scrub_priority > 0 ? pool_scrub_priority : cct->_conf->osd_scrub_priority; | |
1994 | } | |
1995 | ||
1996 | struct C_PG_FinishRecovery : public Context { | |
1997 | PGRef pg; | |
1998 | explicit C_PG_FinishRecovery(PG *p) : pg(p) {} | |
1999 | void finish(int r) override { | |
2000 | pg->_finish_recovery(this); | |
2001 | } | |
2002 | }; | |
2003 | ||
2004 | void PG::mark_clean() | |
2005 | { | |
2006 | // only mark CLEAN if we have the desired number of replicas AND we | |
2007 | // are not remapped. | |
2008 | if (actingset.size() == get_osdmap()->get_pg_size(info.pgid.pgid) && | |
2009 | up == acting) | |
2010 | state_set(PG_STATE_CLEAN); | |
2011 | ||
2012 | // NOTE: this is actually a bit premature: we haven't purged the | |
2013 | // strays yet. | |
2014 | info.history.last_epoch_clean = get_osdmap()->get_epoch(); | |
2015 | info.history.last_interval_clean = info.history.same_interval_since; | |
2016 | ||
2017 | past_intervals.clear(); | |
2018 | dirty_big_info = true; | |
2019 | ||
2020 | if (is_active()) { | |
2021 | /* The check is needed because if we are below min_size we're not | |
2022 | * actually active */ | |
2023 | kick_snap_trim(); | |
2024 | } | |
2025 | ||
2026 | dirty_info = true; | |
2027 | } | |
2028 | ||
2029 | unsigned PG::get_recovery_priority() | |
2030 | { | |
2031 | // a higher value -> a higher priority | |
2032 | ||
2033 | int pool_recovery_priority = 0; | |
2034 | pool.info.opts.get(pool_opts_t::RECOVERY_PRIORITY, &pool_recovery_priority); | |
2035 | ||
2036 | int ret = OSD_RECOVERY_PRIORITY_BASE + pool_recovery_priority; | |
2037 | ||
2038 | // Clamp to valid range | |
2039 | if (ret > OSD_RECOVERY_PRIORITY_MAX) { | |
2040 | ret = OSD_RECOVERY_PRIORITY_MAX; | |
2041 | } else if (ret < OSD_RECOVERY_PRIORITY_MIN) { | |
2042 | ret = OSD_RECOVERY_PRIORITY_MIN; | |
2043 | } | |
2044 | ||
2045 | static_assert(OSD_RECOVERY_PRIORITY_MIN < OSD_RECOVERY_PRIORITY_MAX, "Invalid priority range"); | |
2046 | static_assert(OSD_RECOVERY_PRIORITY_MIN >= 0, "Priority range must match unsigned type"); | |
2047 | ||
2048 | return static_cast<unsigned>(ret); | |
2049 | } | |
2050 | ||
2051 | unsigned PG::get_backfill_priority() | |
2052 | { | |
2053 | // a higher value -> a higher priority | |
2054 | ||
2055 | int ret = OSD_BACKFILL_PRIORITY_BASE; | |
2056 | if (acting.size() < pool.info.min_size) { | |
2057 | // inactive: no. of replicas < min_size, highest priority since it blocks IO | |
2058 | ret = OSD_BACKFILL_INACTIVE_PRIORITY_BASE + (pool.info.min_size - acting.size()); | |
2059 | ||
2060 | } else if (is_undersized()) { | |
2061 | // undersized: OSD_BACKFILL_DEGRADED_PRIORITY_BASE + num missing replicas | |
2062 | assert(pool.info.size > actingset.size()); | |
2063 | ret = OSD_BACKFILL_DEGRADED_PRIORITY_BASE + (pool.info.size - actingset.size()); | |
2064 | ||
2065 | } else if (is_degraded()) { | |
2066 | // degraded: baseline degraded | |
2067 | ret = OSD_BACKFILL_DEGRADED_PRIORITY_BASE; | |
2068 | } | |
2069 | ||
2070 | // Adjust with pool's recovery priority | |
2071 | int pool_recovery_priority = 0; | |
2072 | pool.info.opts.get(pool_opts_t::RECOVERY_PRIORITY, &pool_recovery_priority); | |
2073 | ret += pool_recovery_priority; | |
2074 | ||
2075 | // Clamp to valid range | |
2076 | if (ret > OSD_RECOVERY_PRIORITY_MAX) { | |
2077 | ret = OSD_RECOVERY_PRIORITY_MAX; | |
2078 | } else if (ret < OSD_RECOVERY_PRIORITY_MIN) { | |
2079 | ret = OSD_RECOVERY_PRIORITY_MIN; | |
2080 | } | |
2081 | ||
2082 | return static_cast<unsigned>(ret); | |
2083 | } | |
2084 | ||
2085 | void PG::finish_recovery(list<Context*>& tfin) | |
2086 | { | |
2087 | dout(10) << "finish_recovery" << dendl; | |
2088 | assert(info.last_complete == info.last_update); | |
2089 | ||
2090 | clear_recovery_state(); | |
2091 | ||
2092 | /* | |
2093 | * sync all this before purging strays. but don't block! | |
2094 | */ | |
2095 | finish_sync_event = new C_PG_FinishRecovery(this); | |
2096 | tfin.push_back(finish_sync_event); | |
2097 | } | |
2098 | ||
2099 | void PG::_finish_recovery(Context *c) | |
2100 | { | |
2101 | lock(); | |
2102 | if (deleting) { | |
2103 | unlock(); | |
2104 | return; | |
2105 | } | |
2106 | if (c == finish_sync_event) { | |
2107 | dout(10) << "_finish_recovery" << dendl; | |
2108 | finish_sync_event = 0; | |
2109 | purge_strays(); | |
2110 | ||
2111 | publish_stats_to_osd(); | |
2112 | ||
2113 | if (scrub_after_recovery) { | |
2114 | dout(10) << "_finish_recovery requeueing for scrub" << dendl; | |
2115 | scrub_after_recovery = false; | |
2116 | scrubber.must_deep_scrub = true; | |
2117 | queue_scrub(); | |
2118 | } | |
2119 | } else { | |
2120 | dout(10) << "_finish_recovery -- stale" << dendl; | |
2121 | } | |
2122 | unlock(); | |
2123 | } | |
2124 | ||
2125 | void PG::start_recovery_op(const hobject_t& soid) | |
2126 | { | |
2127 | dout(10) << "start_recovery_op " << soid | |
2128 | #ifdef DEBUG_RECOVERY_OIDS | |
2129 | << " (" << recovering_oids << ")" | |
2130 | #endif | |
2131 | << dendl; | |
2132 | assert(recovery_ops_active >= 0); | |
2133 | recovery_ops_active++; | |
2134 | #ifdef DEBUG_RECOVERY_OIDS | |
2135 | assert(recovering_oids.count(soid) == 0); | |
2136 | recovering_oids.insert(soid); | |
2137 | #endif | |
2138 | osd->start_recovery_op(this, soid); | |
2139 | } | |
2140 | ||
2141 | void PG::finish_recovery_op(const hobject_t& soid, bool dequeue) | |
2142 | { | |
2143 | dout(10) << "finish_recovery_op " << soid | |
2144 | #ifdef DEBUG_RECOVERY_OIDS | |
2145 | << " (" << recovering_oids << ")" | |
2146 | #endif | |
2147 | << dendl; | |
2148 | assert(recovery_ops_active > 0); | |
2149 | recovery_ops_active--; | |
2150 | #ifdef DEBUG_RECOVERY_OIDS | |
2151 | assert(recovering_oids.count(soid)); | |
2152 | recovering_oids.erase(soid); | |
2153 | #endif | |
2154 | osd->finish_recovery_op(this, soid, dequeue); | |
2155 | ||
2156 | if (!dequeue) { | |
2157 | queue_recovery(); | |
2158 | } | |
2159 | } | |
2160 | ||
2161 | void PG::split_into(pg_t child_pgid, PG *child, unsigned split_bits) | |
2162 | { | |
2163 | child->update_snap_mapper_bits(split_bits); | |
2164 | child->update_osdmap_ref(get_osdmap()); | |
2165 | ||
2166 | child->pool = pool; | |
2167 | ||
2168 | // Log | |
2169 | pg_log.split_into(child_pgid, split_bits, &(child->pg_log)); | |
2170 | child->info.last_complete = info.last_complete; | |
2171 | ||
2172 | info.last_update = pg_log.get_head(); | |
2173 | child->info.last_update = child->pg_log.get_head(); | |
2174 | ||
2175 | child->info.last_user_version = info.last_user_version; | |
2176 | ||
2177 | info.log_tail = pg_log.get_tail(); | |
2178 | child->info.log_tail = child->pg_log.get_tail(); | |
2179 | ||
2180 | if (info.last_complete < pg_log.get_tail()) | |
2181 | info.last_complete = pg_log.get_tail(); | |
2182 | if (child->info.last_complete < child->pg_log.get_tail()) | |
2183 | child->info.last_complete = child->pg_log.get_tail(); | |
2184 | ||
2185 | // Info | |
2186 | child->info.history = info.history; | |
2187 | child->info.history.epoch_created = get_osdmap()->get_epoch(); | |
2188 | child->info.purged_snaps = info.purged_snaps; | |
2189 | ||
2190 | if (info.last_backfill.is_max()) { | |
2191 | child->info.set_last_backfill(hobject_t::get_max()); | |
2192 | } else { | |
2193 | // restart backfill on parent and child to be safe. we could | |
2194 | // probably do better in the bitwise sort case, but it's more | |
2195 | // fragile (there may be special work to do on backfill completion | |
2196 | // in the future). | |
2197 | info.set_last_backfill(hobject_t()); | |
2198 | child->info.set_last_backfill(hobject_t()); | |
2199 | } | |
2200 | ||
2201 | child->info.stats = info.stats; | |
2202 | child->info.stats.parent_split_bits = split_bits; | |
2203 | info.stats.stats_invalid = true; | |
2204 | child->info.stats.stats_invalid = true; | |
2205 | child->info.last_epoch_started = info.last_epoch_started; | |
2206 | child->info.last_interval_started = info.last_interval_started; | |
2207 | ||
2208 | child->snap_trimq = snap_trimq; | |
2209 | ||
2210 | // There can't be recovery/backfill going on now | |
2211 | int primary, up_primary; | |
2212 | vector<int> newup, newacting; | |
2213 | get_osdmap()->pg_to_up_acting_osds( | |
2214 | child->info.pgid.pgid, &newup, &up_primary, &newacting, &primary); | |
2215 | child->init_primary_up_acting( | |
2216 | newup, | |
2217 | newacting, | |
2218 | up_primary, | |
2219 | primary); | |
2220 | child->role = OSDMap::calc_pg_role(osd->whoami, child->acting); | |
2221 | ||
2222 | // this comparison includes primary rank via pg_shard_t | |
2223 | if (get_primary() != child->get_primary()) | |
2224 | child->info.history.same_primary_since = get_osdmap()->get_epoch(); | |
2225 | ||
2226 | child->info.stats.up = up; | |
2227 | child->info.stats.up_primary = up_primary; | |
2228 | child->info.stats.acting = acting; | |
2229 | child->info.stats.acting_primary = primary; | |
2230 | child->info.stats.mapping_epoch = get_osdmap()->get_epoch(); | |
2231 | ||
2232 | // History | |
2233 | child->past_intervals = past_intervals; | |
2234 | ||
2235 | _split_into(child_pgid, child, split_bits); | |
2236 | ||
2237 | // release all backoffs for simplicity | |
2238 | release_backoffs(hobject_t(), hobject_t::get_max()); | |
2239 | ||
2240 | child->on_new_interval(); | |
2241 | ||
2242 | child->dirty_info = true; | |
2243 | child->dirty_big_info = true; | |
2244 | dirty_info = true; | |
2245 | dirty_big_info = true; | |
2246 | } | |
2247 | ||
2248 | void PG::add_backoff(SessionRef s, const hobject_t& begin, const hobject_t& end) | |
2249 | { | |
2250 | ConnectionRef con = s->con; | |
2251 | if (!con) // OSD::ms_handle_reset clears s->con without a lock | |
2252 | return; | |
2253 | BackoffRef b(s->have_backoff(info.pgid, begin)); | |
2254 | if (b) { | |
2255 | derr << __func__ << " already have backoff for " << s << " begin " << begin | |
2256 | << " " << *b << dendl; | |
2257 | ceph_abort(); | |
2258 | } | |
2259 | Mutex::Locker l(backoff_lock); | |
2260 | { | |
2261 | b = new Backoff(info.pgid, this, s, ++s->backoff_seq, begin, end); | |
2262 | backoffs[begin].insert(b); | |
2263 | s->add_backoff(b); | |
2264 | dout(10) << __func__ << " session " << s << " added " << *b << dendl; | |
2265 | } | |
2266 | con->send_message( | |
2267 | new MOSDBackoff( | |
2268 | info.pgid, | |
2269 | get_osdmap()->get_epoch(), | |
2270 | CEPH_OSD_BACKOFF_OP_BLOCK, | |
2271 | b->id, | |
2272 | begin, | |
2273 | end)); | |
2274 | } | |
2275 | ||
2276 | void PG::release_backoffs(const hobject_t& begin, const hobject_t& end) | |
2277 | { | |
2278 | dout(10) << __func__ << " [" << begin << "," << end << ")" << dendl; | |
2279 | vector<BackoffRef> bv; | |
2280 | { | |
2281 | Mutex::Locker l(backoff_lock); | |
2282 | auto p = backoffs.lower_bound(begin); | |
2283 | while (p != backoffs.end()) { | |
2284 | int r = cmp(p->first, end); | |
2285 | dout(20) << __func__ << " ? " << r << " " << p->first | |
2286 | << " " << p->second << dendl; | |
2287 | // note: must still examine begin=end=p->first case | |
2288 | if (r > 0 || (r == 0 && begin < end)) { | |
2289 | break; | |
2290 | } | |
2291 | dout(20) << __func__ << " checking " << p->first | |
2292 | << " " << p->second << dendl; | |
2293 | auto q = p->second.begin(); | |
2294 | while (q != p->second.end()) { | |
2295 | dout(20) << __func__ << " checking " << *q << dendl; | |
2296 | int r = cmp((*q)->begin, begin); | |
2297 | if (r == 0 || (r > 0 && (*q)->end < end)) { | |
2298 | bv.push_back(*q); | |
2299 | q = p->second.erase(q); | |
2300 | } else { | |
2301 | ++q; | |
2302 | } | |
2303 | } | |
2304 | if (p->second.empty()) { | |
2305 | p = backoffs.erase(p); | |
2306 | } else { | |
2307 | ++p; | |
2308 | } | |
2309 | } | |
2310 | } | |
2311 | for (auto b : bv) { | |
2312 | Mutex::Locker l(b->lock); | |
2313 | dout(10) << __func__ << " " << *b << dendl; | |
2314 | if (b->session) { | |
2315 | assert(b->pg == this); | |
2316 | ConnectionRef con = b->session->con; | |
2317 | if (con) { // OSD::ms_handle_reset clears s->con without a lock | |
2318 | con->send_message( | |
2319 | new MOSDBackoff( | |
2320 | info.pgid, | |
2321 | get_osdmap()->get_epoch(), | |
2322 | CEPH_OSD_BACKOFF_OP_UNBLOCK, | |
2323 | b->id, | |
2324 | b->begin, | |
2325 | b->end)); | |
2326 | } | |
2327 | if (b->is_new()) { | |
2328 | b->state = Backoff::STATE_DELETING; | |
2329 | } else { | |
2330 | b->session->rm_backoff(b); | |
2331 | b->session.reset(); | |
2332 | } | |
2333 | b->pg.reset(); | |
2334 | } | |
2335 | } | |
2336 | } | |
2337 | ||
2338 | void PG::clear_backoffs() | |
2339 | { | |
2340 | dout(10) << __func__ << " " << dendl; | |
2341 | map<hobject_t,set<BackoffRef>> ls; | |
2342 | { | |
2343 | Mutex::Locker l(backoff_lock); | |
2344 | ls.swap(backoffs); | |
2345 | } | |
2346 | for (auto& p : ls) { | |
2347 | for (auto& b : p.second) { | |
2348 | Mutex::Locker l(b->lock); | |
2349 | dout(10) << __func__ << " " << *b << dendl; | |
2350 | if (b->session) { | |
2351 | assert(b->pg == this); | |
2352 | if (b->is_new()) { | |
2353 | b->state = Backoff::STATE_DELETING; | |
2354 | } else { | |
2355 | b->session->rm_backoff(b); | |
2356 | b->session.reset(); | |
2357 | } | |
2358 | b->pg.reset(); | |
2359 | } | |
2360 | } | |
2361 | } | |
2362 | } | |
2363 | ||
2364 | // called by Session::clear_backoffs() | |
2365 | void PG::rm_backoff(BackoffRef b) | |
2366 | { | |
2367 | dout(10) << __func__ << " " << *b << dendl; | |
2368 | Mutex::Locker l(backoff_lock); | |
2369 | assert(b->lock.is_locked_by_me()); | |
2370 | assert(b->pg == this); | |
2371 | auto p = backoffs.find(b->begin); | |
2372 | // may race with release_backoffs() | |
2373 | if (p != backoffs.end()) { | |
2374 | auto q = p->second.find(b); | |
2375 | if (q != p->second.end()) { | |
2376 | p->second.erase(q); | |
2377 | if (p->second.empty()) { | |
2378 | backoffs.erase(p); | |
2379 | } | |
2380 | } | |
2381 | } | |
2382 | } | |
2383 | ||
2384 | void PG::clear_recovery_state() | |
2385 | { | |
2386 | dout(10) << "clear_recovery_state" << dendl; | |
2387 | ||
2388 | pg_log.reset_recovery_pointers(); | |
2389 | finish_sync_event = 0; | |
2390 | ||
2391 | hobject_t soid; | |
2392 | while (recovery_ops_active > 0) { | |
2393 | #ifdef DEBUG_RECOVERY_OIDS | |
2394 | soid = *recovering_oids.begin(); | |
2395 | #endif | |
2396 | finish_recovery_op(soid, true); | |
2397 | } | |
2398 | ||
2399 | backfill_targets.clear(); | |
2400 | backfill_info.clear(); | |
2401 | peer_backfill_info.clear(); | |
2402 | waiting_on_backfill.clear(); | |
2403 | _clear_recovery_state(); // pg impl specific hook | |
2404 | } | |
2405 | ||
2406 | void PG::cancel_recovery() | |
2407 | { | |
2408 | dout(10) << "cancel_recovery" << dendl; | |
2409 | clear_recovery_state(); | |
2410 | } | |
2411 | ||
2412 | ||
2413 | void PG::purge_strays() | |
2414 | { | |
2415 | dout(10) << "purge_strays " << stray_set << dendl; | |
2416 | ||
2417 | bool removed = false; | |
2418 | for (set<pg_shard_t>::iterator p = stray_set.begin(); | |
2419 | p != stray_set.end(); | |
2420 | ++p) { | |
2421 | assert(!is_actingbackfill(*p)); | |
2422 | if (get_osdmap()->is_up(p->osd)) { | |
2423 | dout(10) << "sending PGRemove to osd." << *p << dendl; | |
2424 | vector<spg_t> to_remove; | |
2425 | to_remove.push_back(spg_t(info.pgid.pgid, p->shard)); | |
2426 | MOSDPGRemove *m = new MOSDPGRemove( | |
2427 | get_osdmap()->get_epoch(), | |
2428 | to_remove); | |
2429 | osd->send_message_osd_cluster(p->osd, m, get_osdmap()->get_epoch()); | |
2430 | } else { | |
2431 | dout(10) << "not sending PGRemove to down osd." << *p << dendl; | |
2432 | } | |
2433 | peer_missing.erase(*p); | |
2434 | peer_info.erase(*p); | |
2435 | peer_purged.insert(*p); | |
2436 | removed = true; | |
2437 | } | |
2438 | ||
2439 | // if we removed anyone, update peers (which include peer_info) | |
2440 | if (removed) | |
2441 | update_heartbeat_peers(); | |
2442 | ||
2443 | stray_set.clear(); | |
2444 | ||
2445 | // clear _requested maps; we may have to peer() again if we discover | |
2446 | // (more) stray content | |
2447 | peer_log_requested.clear(); | |
2448 | peer_missing_requested.clear(); | |
2449 | } | |
2450 | ||
2451 | void PG::set_probe_targets(const set<pg_shard_t> &probe_set) | |
2452 | { | |
2453 | Mutex::Locker l(heartbeat_peer_lock); | |
2454 | probe_targets.clear(); | |
2455 | for (set<pg_shard_t>::iterator i = probe_set.begin(); | |
2456 | i != probe_set.end(); | |
2457 | ++i) { | |
2458 | probe_targets.insert(i->osd); | |
2459 | } | |
2460 | } | |
2461 | ||
2462 | void PG::clear_probe_targets() | |
2463 | { | |
2464 | Mutex::Locker l(heartbeat_peer_lock); | |
2465 | probe_targets.clear(); | |
2466 | } | |
2467 | ||
2468 | void PG::update_heartbeat_peers() | |
2469 | { | |
2470 | assert(is_locked()); | |
2471 | ||
2472 | if (!is_primary()) | |
2473 | return; | |
2474 | ||
2475 | set<int> new_peers; | |
2476 | for (unsigned i=0; i<acting.size(); i++) { | |
2477 | if (acting[i] != CRUSH_ITEM_NONE) | |
2478 | new_peers.insert(acting[i]); | |
2479 | } | |
2480 | for (unsigned i=0; i<up.size(); i++) { | |
2481 | if (up[i] != CRUSH_ITEM_NONE) | |
2482 | new_peers.insert(up[i]); | |
2483 | } | |
2484 | for (map<pg_shard_t,pg_info_t>::iterator p = peer_info.begin(); | |
2485 | p != peer_info.end(); | |
2486 | ++p) | |
2487 | new_peers.insert(p->first.osd); | |
2488 | ||
2489 | bool need_update = false; | |
2490 | heartbeat_peer_lock.Lock(); | |
2491 | if (new_peers == heartbeat_peers) { | |
2492 | dout(10) << "update_heartbeat_peers " << heartbeat_peers << " unchanged" << dendl; | |
2493 | } else { | |
2494 | dout(10) << "update_heartbeat_peers " << heartbeat_peers << " -> " << new_peers << dendl; | |
2495 | heartbeat_peers.swap(new_peers); | |
2496 | need_update = true; | |
2497 | } | |
2498 | heartbeat_peer_lock.Unlock(); | |
2499 | ||
2500 | if (need_update) | |
2501 | osd->need_heartbeat_peer_update(); | |
2502 | } | |
2503 | ||
2504 | ||
2505 | bool PG::check_in_progress_op( | |
2506 | const osd_reqid_t &r, | |
2507 | eversion_t *version, | |
2508 | version_t *user_version, | |
2509 | int *return_code) const | |
2510 | { | |
2511 | return ( | |
2512 | projected_log.get_request(r, version, user_version, return_code) || | |
2513 | pg_log.get_log().get_request(r, version, user_version, return_code)); | |
2514 | } | |
2515 | ||
2516 | void PG::_update_calc_stats() | |
2517 | { | |
2518 | info.stats.version = info.last_update; | |
2519 | info.stats.created = info.history.epoch_created; | |
2520 | info.stats.last_scrub = info.history.last_scrub; | |
2521 | info.stats.last_scrub_stamp = info.history.last_scrub_stamp; | |
2522 | info.stats.last_deep_scrub = info.history.last_deep_scrub; | |
2523 | info.stats.last_deep_scrub_stamp = info.history.last_deep_scrub_stamp; | |
2524 | info.stats.last_clean_scrub_stamp = info.history.last_clean_scrub_stamp; | |
2525 | info.stats.last_epoch_clean = info.history.last_epoch_clean; | |
2526 | ||
2527 | info.stats.log_size = pg_log.get_head().version - pg_log.get_tail().version; | |
2528 | info.stats.ondisk_log_size = info.stats.log_size; | |
2529 | info.stats.log_start = pg_log.get_tail(); | |
2530 | info.stats.ondisk_log_start = pg_log.get_tail(); | |
2531 | ||
2532 | // If actingset is larger then upset we will have misplaced, | |
2533 | // so we will report based on actingset size. | |
2534 | ||
2535 | // If upset is larger then we will have degraded, | |
2536 | // so we will report based on upset size. | |
2537 | ||
2538 | // If target is the largest of them all, it will contribute to | |
2539 | // the degraded count because num_object_copies is | |
2540 | // computed using target and eventual used to get degraded total. | |
2541 | ||
2542 | unsigned target = get_osdmap()->get_pg_size(info.pgid.pgid); | |
2543 | unsigned nrep = MAX(actingset.size(), upset.size()); | |
2544 | // calc num_object_copies | |
2545 | info.stats.stats.calc_copies(MAX(target, nrep)); | |
2546 | info.stats.stats.sum.num_objects_degraded = 0; | |
2547 | info.stats.stats.sum.num_objects_unfound = 0; | |
2548 | info.stats.stats.sum.num_objects_misplaced = 0; | |
2549 | if ((is_degraded() || is_undersized() || !is_clean()) && is_peered()) { | |
2550 | // NOTE: we only generate copies, degraded, misplaced and unfound | |
2551 | // values for the summation, not individual stat categories. | |
2552 | int64_t num_objects = info.stats.stats.sum.num_objects; | |
2553 | ||
2554 | // Total sum of all missing | |
2555 | int64_t missing = 0; | |
2556 | // Objects that have arrived backfilled to up OSDs (not in acting) | |
2557 | int64_t backfilled = 0; | |
2558 | // A misplaced object is not stored on the correct OSD | |
2559 | int64_t misplaced = 0; | |
2560 | // Total of object copies/shards found | |
2561 | int64_t object_copies = 0; | |
2562 | ||
2563 | // num_objects_missing on each peer | |
2564 | for (map<pg_shard_t, pg_info_t>::iterator pi = | |
2565 | peer_info.begin(); | |
2566 | pi != peer_info.end(); | |
2567 | ++pi) { | |
2568 | map<pg_shard_t, pg_missing_t>::const_iterator pm = | |
2569 | peer_missing.find(pi->first); | |
2570 | if (pm != peer_missing.end()) { | |
2571 | pi->second.stats.stats.sum.num_objects_missing = | |
2572 | pm->second.num_missing(); | |
2573 | } | |
2574 | } | |
2575 | ||
2576 | assert(!actingbackfill.empty()); | |
2577 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
2578 | i != actingbackfill.end(); | |
2579 | ++i) { | |
2580 | const pg_shard_t &p = *i; | |
2581 | ||
2582 | bool in_up = (upset.find(p) != upset.end()); | |
2583 | bool in_acting = (actingset.find(p) != actingset.end()); | |
2584 | assert(in_up || in_acting); | |
2585 | ||
2586 | // in acting Compute total objects excluding num_missing | |
2587 | // in acting and not in up Compute misplaced objects excluding num_missing | |
2588 | // in up and not in acting Compute total objects already backfilled | |
2589 | if (in_acting) { | |
2590 | unsigned osd_missing; | |
2591 | // primary handling | |
2592 | if (p == pg_whoami) { | |
2593 | osd_missing = pg_log.get_missing().num_missing(); | |
2594 | info.stats.stats.sum.num_objects_missing_on_primary = | |
2595 | osd_missing; | |
2596 | object_copies += num_objects; // My local (primary) count | |
2597 | } else { | |
2598 | assert(peer_missing.count(p)); | |
2599 | osd_missing = peer_missing[p].num_missing(); | |
2600 | object_copies += peer_info[p].stats.stats.sum.num_objects; | |
2601 | } | |
2602 | missing += osd_missing; | |
2603 | // Count non-missing objects not in up as misplaced | |
2604 | if (!in_up && num_objects > osd_missing) | |
2605 | misplaced += num_objects - osd_missing; | |
2606 | } else { | |
2607 | assert(in_up && !in_acting); | |
2608 | ||
2609 | // If this peer has more objects then it should, ignore them | |
2610 | backfilled += MIN(num_objects, peer_info[p].stats.stats.sum.num_objects); | |
2611 | } | |
2612 | } | |
2613 | ||
2614 | // Any objects that have been backfilled to up OSDs can deducted from misplaced | |
2615 | misplaced = MAX(0, misplaced - backfilled); | |
2616 | ||
2617 | // Deduct computed total missing on acting nodes | |
2618 | object_copies -= missing; | |
2619 | // Include computed backfilled objects on up nodes | |
2620 | object_copies += backfilled; | |
2621 | // a degraded objects has fewer replicas or EC shards than the | |
2622 | // pool specifies. num_object_copies will never be smaller than target * num_copies. | |
2623 | int64_t degraded = MAX(0, info.stats.stats.sum.num_object_copies - object_copies); | |
2624 | ||
2625 | info.stats.stats.sum.num_objects_degraded = degraded; | |
2626 | info.stats.stats.sum.num_objects_unfound = get_num_unfound(); | |
2627 | info.stats.stats.sum.num_objects_misplaced = misplaced; | |
2628 | } | |
2629 | } | |
2630 | ||
2631 | void PG::_update_blocked_by() | |
2632 | { | |
2633 | // set a max on the number of blocking peers we report. if we go | |
2634 | // over, report a random subset. keep the result sorted. | |
2635 | unsigned keep = MIN(blocked_by.size(), cct->_conf->osd_max_pg_blocked_by); | |
2636 | unsigned skip = blocked_by.size() - keep; | |
2637 | info.stats.blocked_by.clear(); | |
2638 | info.stats.blocked_by.resize(keep); | |
2639 | unsigned pos = 0; | |
2640 | for (set<int>::iterator p = blocked_by.begin(); | |
2641 | p != blocked_by.end() && keep > 0; | |
2642 | ++p) { | |
2643 | if (skip > 0 && (rand() % (skip + keep) < skip)) { | |
2644 | --skip; | |
2645 | } else { | |
2646 | info.stats.blocked_by[pos++] = *p; | |
2647 | --keep; | |
2648 | } | |
2649 | } | |
2650 | } | |
2651 | ||
2652 | void PG::publish_stats_to_osd() | |
2653 | { | |
2654 | if (!is_primary()) | |
2655 | return; | |
2656 | ||
2657 | pg_stats_publish_lock.Lock(); | |
2658 | ||
2659 | if (info.stats.stats.sum.num_scrub_errors) | |
2660 | state_set(PG_STATE_INCONSISTENT); | |
2661 | else | |
2662 | state_clear(PG_STATE_INCONSISTENT); | |
2663 | ||
2664 | utime_t now = ceph_clock_now(); | |
2665 | if (info.stats.state != state) { | |
2666 | info.stats.last_change = now; | |
2667 | // Optimistic estimation, if we just find out an inactive PG, | |
2668 | // assumt it is active till now. | |
2669 | if (!(state & PG_STATE_ACTIVE) && | |
2670 | (info.stats.state & PG_STATE_ACTIVE)) | |
2671 | info.stats.last_active = now; | |
2672 | ||
2673 | if ((state & PG_STATE_ACTIVE) && | |
2674 | !(info.stats.state & PG_STATE_ACTIVE)) | |
2675 | info.stats.last_became_active = now; | |
2676 | if ((state & (PG_STATE_ACTIVE|PG_STATE_PEERED)) && | |
2677 | !(info.stats.state & (PG_STATE_ACTIVE|PG_STATE_PEERED))) | |
2678 | info.stats.last_became_peered = now; | |
2679 | if (!(state & PG_STATE_CREATING) && | |
2680 | (info.stats.state & PG_STATE_CREATING)) { | |
2681 | osd->send_pg_created(get_pgid().pgid); | |
2682 | } | |
2683 | info.stats.state = state; | |
2684 | } | |
2685 | ||
2686 | _update_calc_stats(); | |
2687 | _update_blocked_by(); | |
2688 | ||
2689 | bool publish = false; | |
2690 | pg_stat_t pre_publish = info.stats; | |
2691 | pre_publish.stats.add(unstable_stats); | |
2692 | utime_t cutoff = now; | |
2693 | cutoff -= cct->_conf->osd_pg_stat_report_interval_max; | |
2694 | if (pg_stats_publish_valid && pre_publish == pg_stats_publish && | |
2695 | info.stats.last_fresh > cutoff) { | |
2696 | dout(15) << "publish_stats_to_osd " << pg_stats_publish.reported_epoch | |
2697 | << ": no change since " << info.stats.last_fresh << dendl; | |
2698 | } else { | |
2699 | // update our stat summary and timestamps | |
2700 | info.stats.reported_epoch = get_osdmap()->get_epoch(); | |
2701 | ++info.stats.reported_seq; | |
2702 | ||
2703 | info.stats.last_fresh = now; | |
2704 | ||
2705 | if (info.stats.state & PG_STATE_CLEAN) | |
2706 | info.stats.last_clean = now; | |
2707 | if (info.stats.state & PG_STATE_ACTIVE) | |
2708 | info.stats.last_active = now; | |
2709 | if (info.stats.state & (PG_STATE_ACTIVE|PG_STATE_PEERED)) | |
2710 | info.stats.last_peered = now; | |
2711 | info.stats.last_unstale = now; | |
2712 | if ((info.stats.state & PG_STATE_DEGRADED) == 0) | |
2713 | info.stats.last_undegraded = now; | |
2714 | if ((info.stats.state & PG_STATE_UNDERSIZED) == 0) | |
2715 | info.stats.last_fullsized = now; | |
2716 | ||
2717 | // do not send pgstat to mon anymore once we are luminous, since mgr takes | |
2718 | // care of this by sending MMonMgrReport to mon. | |
2719 | publish = | |
2720 | osd->osd->get_osdmap()->require_osd_release < CEPH_RELEASE_LUMINOUS; | |
2721 | pg_stats_publish_valid = true; | |
2722 | pg_stats_publish = pre_publish; | |
2723 | ||
2724 | dout(15) << "publish_stats_to_osd " << pg_stats_publish.reported_epoch | |
2725 | << ":" << pg_stats_publish.reported_seq << dendl; | |
2726 | } | |
2727 | pg_stats_publish_lock.Unlock(); | |
2728 | ||
2729 | if (publish) | |
2730 | osd->pg_stat_queue_enqueue(this); | |
2731 | } | |
2732 | ||
2733 | void PG::clear_publish_stats() | |
2734 | { | |
2735 | dout(15) << "clear_stats" << dendl; | |
2736 | pg_stats_publish_lock.Lock(); | |
2737 | pg_stats_publish_valid = false; | |
2738 | pg_stats_publish_lock.Unlock(); | |
2739 | ||
2740 | osd->pg_stat_queue_dequeue(this); | |
2741 | } | |
2742 | ||
2743 | /** | |
2744 | * initialize a newly instantiated pg | |
2745 | * | |
2746 | * Initialize PG state, as when a PG is initially created, or when it | |
2747 | * is first instantiated on the current node. | |
2748 | * | |
2749 | * @param role our role/rank | |
2750 | * @param newup up set | |
2751 | * @param newacting acting set | |
2752 | * @param history pg history | |
2753 | * @param pi past_intervals | |
2754 | * @param backfill true if info should be marked as backfill | |
2755 | * @param t transaction to write out our new state in | |
2756 | */ | |
2757 | void PG::init( | |
2758 | int role, | |
2759 | const vector<int>& newup, int new_up_primary, | |
2760 | const vector<int>& newacting, int new_acting_primary, | |
2761 | const pg_history_t& history, | |
2762 | const PastIntervals& pi, | |
2763 | bool backfill, | |
2764 | ObjectStore::Transaction *t) | |
2765 | { | |
2766 | dout(10) << "init role " << role << " up " << newup << " acting " << newacting | |
2767 | << " history " << history | |
2768 | << " past_intervals " << pi | |
2769 | << dendl; | |
2770 | ||
2771 | set_role(role); | |
2772 | acting = newacting; | |
2773 | up = newup; | |
2774 | init_primary_up_acting( | |
2775 | newup, | |
2776 | newacting, | |
2777 | new_up_primary, | |
2778 | new_acting_primary); | |
2779 | ||
2780 | info.history = history; | |
2781 | past_intervals = pi; | |
2782 | ||
2783 | info.stats.up = up; | |
2784 | info.stats.up_primary = new_up_primary; | |
2785 | info.stats.acting = acting; | |
2786 | info.stats.acting_primary = new_acting_primary; | |
2787 | info.stats.mapping_epoch = info.history.same_interval_since; | |
2788 | ||
2789 | if (backfill) { | |
2790 | dout(10) << __func__ << ": Setting backfill" << dendl; | |
2791 | info.set_last_backfill(hobject_t()); | |
2792 | info.last_complete = info.last_update; | |
2793 | pg_log.mark_log_for_rewrite(); | |
2794 | } | |
2795 | ||
2796 | on_new_interval(); | |
2797 | ||
2798 | dirty_info = true; | |
2799 | dirty_big_info = true; | |
2800 | write_if_dirty(*t); | |
2801 | } | |
2802 | ||
2803 | #pragma GCC diagnostic ignored "-Wpragmas" | |
2804 | #pragma GCC diagnostic push | |
2805 | #pragma GCC diagnostic ignored "-Wdeprecated-declarations" | |
2806 | ||
2807 | void PG::upgrade(ObjectStore *store) | |
2808 | { | |
2809 | assert(info_struct_v <= 10); | |
2810 | ObjectStore::Transaction t; | |
2811 | ||
2812 | assert(info_struct_v >= 7); | |
2813 | ||
2814 | // 7 -> 8 | |
2815 | if (info_struct_v <= 7) { | |
2816 | pg_log.mark_log_for_rewrite(); | |
2817 | ghobject_t log_oid(OSD::make_pg_log_oid(pg_id)); | |
2818 | ghobject_t biginfo_oid(OSD::make_pg_biginfo_oid(pg_id)); | |
2819 | t.remove(coll_t::meta(), log_oid); | |
2820 | t.remove(coll_t::meta(), biginfo_oid); | |
2821 | t.touch(coll, pgmeta_oid); | |
2822 | } | |
2823 | ||
2824 | // 8 -> 9 | |
2825 | if (info_struct_v <= 8) { | |
2826 | // no special action needed. | |
2827 | } | |
2828 | ||
2829 | // 9 -> 10 | |
2830 | if (info_struct_v <= 9) { | |
2831 | // previous versions weren't (as) aggressively clearing past_intervals | |
2832 | if (info.history.last_epoch_clean >= info.history.same_interval_since) { | |
2833 | dout(20) << __func__ << " clearing past_intervals" << dendl; | |
2834 | past_intervals.clear(); | |
2835 | } | |
2836 | } | |
2837 | ||
2838 | // update infover_key | |
2839 | if (info_struct_v < cur_struct_v) { | |
2840 | map<string,bufferlist> v; | |
2841 | __u8 ver = cur_struct_v; | |
2842 | ::encode(ver, v[infover_key]); | |
2843 | t.omap_setkeys(coll, pgmeta_oid, v); | |
2844 | } | |
2845 | ||
2846 | dirty_info = true; | |
2847 | dirty_big_info = true; | |
2848 | write_if_dirty(t); | |
2849 | ||
2850 | ceph::shared_ptr<ObjectStore::Sequencer> osr (std::make_shared< | |
2851 | ObjectStore::Sequencer>("upgrade")); | |
2852 | int r = store->apply_transaction(osr.get(), std::move(t)); | |
2853 | if (r != 0) { | |
2854 | derr << __func__ << ": apply_transaction returned " | |
2855 | << cpp_strerror(r) << dendl; | |
2856 | ceph_abort(); | |
2857 | } | |
2858 | assert(r == 0); | |
2859 | ||
2860 | C_SaferCond waiter; | |
2861 | if (!osr->flush_commit(&waiter)) { | |
2862 | waiter.wait(); | |
2863 | } | |
2864 | } | |
2865 | ||
2866 | #pragma GCC diagnostic pop | |
2867 | #pragma GCC diagnostic warning "-Wpragmas" | |
2868 | ||
2869 | int PG::_prepare_write_info(CephContext* cct, | |
2870 | map<string,bufferlist> *km, | |
2871 | epoch_t epoch, | |
2872 | pg_info_t &info, pg_info_t &last_written_info, | |
2873 | PastIntervals &past_intervals, | |
2874 | bool dirty_big_info, | |
2875 | bool dirty_epoch, | |
2876 | bool try_fast_info, | |
2877 | PerfCounters *logger) | |
2878 | { | |
2879 | if (dirty_epoch) { | |
2880 | ::encode(epoch, (*km)[epoch_key]); | |
2881 | } | |
2882 | ||
2883 | if (logger) | |
2884 | logger->inc(l_osd_pg_info); | |
2885 | ||
2886 | // try to do info efficiently? | |
2887 | if (!dirty_big_info && try_fast_info && | |
2888 | info.last_update > last_written_info.last_update) { | |
2889 | pg_fast_info_t fast; | |
2890 | fast.populate_from(info); | |
2891 | bool did = fast.try_apply_to(&last_written_info); | |
2892 | assert(did); // we verified last_update increased above | |
2893 | if (info == last_written_info) { | |
2894 | ::encode(fast, (*km)[fastinfo_key]); | |
2895 | if (logger) | |
2896 | logger->inc(l_osd_pg_fastinfo); | |
2897 | return 0; | |
2898 | } | |
2899 | generic_dout(30) << __func__ << " fastinfo failed, info:\n"; | |
2900 | { | |
2901 | JSONFormatter jf(true); | |
2902 | jf.dump_object("info", info); | |
2903 | jf.flush(*_dout); | |
2904 | } | |
2905 | { | |
2906 | *_dout << "\nlast_written_info:\n"; | |
2907 | JSONFormatter jf(true); | |
2908 | jf.dump_object("last_written_info", last_written_info); | |
2909 | jf.flush(*_dout); | |
2910 | } | |
2911 | *_dout << dendl; | |
2912 | } | |
2913 | last_written_info = info; | |
2914 | ||
2915 | // info. store purged_snaps separately. | |
2916 | interval_set<snapid_t> purged_snaps; | |
2917 | purged_snaps.swap(info.purged_snaps); | |
2918 | ::encode(info, (*km)[info_key]); | |
2919 | purged_snaps.swap(info.purged_snaps); | |
2920 | ||
2921 | if (dirty_big_info) { | |
2922 | // potentially big stuff | |
2923 | bufferlist& bigbl = (*km)[biginfo_key]; | |
2924 | ::encode(past_intervals, bigbl); | |
2925 | ::encode(info.purged_snaps, bigbl); | |
2926 | //dout(20) << "write_info bigbl " << bigbl.length() << dendl; | |
2927 | if (logger) | |
2928 | logger->inc(l_osd_pg_biginfo); | |
2929 | } | |
2930 | ||
2931 | return 0; | |
2932 | } | |
2933 | ||
2934 | void PG::_create(ObjectStore::Transaction& t, spg_t pgid, int bits) | |
2935 | { | |
2936 | coll_t coll(pgid); | |
2937 | t.create_collection(coll, bits); | |
2938 | } | |
2939 | ||
2940 | void PG::_init(ObjectStore::Transaction& t, spg_t pgid, const pg_pool_t *pool) | |
2941 | { | |
2942 | coll_t coll(pgid); | |
2943 | ||
2944 | if (pool) { | |
2945 | // Give a hint to the PG collection | |
2946 | bufferlist hint; | |
2947 | uint32_t pg_num = pool->get_pg_num(); | |
2948 | uint64_t expected_num_objects_pg = pool->expected_num_objects / pg_num; | |
2949 | ::encode(pg_num, hint); | |
2950 | ::encode(expected_num_objects_pg, hint); | |
2951 | uint32_t hint_type = ObjectStore::Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS; | |
2952 | t.collection_hint(coll, hint_type, hint); | |
2953 | } | |
2954 | ||
2955 | ghobject_t pgmeta_oid(pgid.make_pgmeta_oid()); | |
2956 | t.touch(coll, pgmeta_oid); | |
2957 | map<string,bufferlist> values; | |
2958 | __u8 struct_v = cur_struct_v; | |
2959 | ::encode(struct_v, values[infover_key]); | |
2960 | t.omap_setkeys(coll, pgmeta_oid, values); | |
2961 | } | |
2962 | ||
2963 | void PG::prepare_write_info(map<string,bufferlist> *km) | |
2964 | { | |
2965 | info.stats.stats.add(unstable_stats); | |
2966 | unstable_stats.clear(); | |
2967 | ||
2968 | bool need_update_epoch = last_epoch < get_osdmap()->get_epoch(); | |
2969 | int ret = _prepare_write_info(cct, km, get_osdmap()->get_epoch(), | |
2970 | info, | |
2971 | last_written_info, | |
2972 | past_intervals, | |
2973 | dirty_big_info, need_update_epoch, | |
2974 | cct->_conf->osd_fast_info, | |
2975 | osd->logger); | |
2976 | assert(ret == 0); | |
2977 | if (need_update_epoch) | |
2978 | last_epoch = get_osdmap()->get_epoch(); | |
2979 | last_persisted_osdmap_ref = osdmap_ref; | |
2980 | ||
2981 | dirty_info = false; | |
2982 | dirty_big_info = false; | |
2983 | } | |
2984 | ||
2985 | #pragma GCC diagnostic ignored "-Wpragmas" | |
2986 | #pragma GCC diagnostic push | |
2987 | #pragma GCC diagnostic ignored "-Wdeprecated-declarations" | |
2988 | ||
2989 | bool PG::_has_removal_flag(ObjectStore *store, | |
2990 | spg_t pgid) | |
2991 | { | |
2992 | coll_t coll(pgid); | |
2993 | ghobject_t pgmeta_oid(pgid.make_pgmeta_oid()); | |
2994 | ||
2995 | // first try new way | |
2996 | set<string> keys; | |
2997 | keys.insert("_remove"); | |
2998 | map<string,bufferlist> values; | |
2999 | if (store->omap_get_values(coll, pgmeta_oid, keys, &values) == 0 && | |
3000 | values.size() == 1) | |
3001 | return true; | |
3002 | ||
3003 | return false; | |
3004 | } | |
3005 | ||
3006 | int PG::peek_map_epoch(ObjectStore *store, | |
3007 | spg_t pgid, | |
3008 | epoch_t *pepoch, | |
3009 | bufferlist *bl) | |
3010 | { | |
3011 | coll_t coll(pgid); | |
3012 | ghobject_t legacy_infos_oid(OSD::make_infos_oid()); | |
3013 | ghobject_t pgmeta_oid(pgid.make_pgmeta_oid()); | |
3014 | epoch_t cur_epoch = 0; | |
3015 | ||
3016 | assert(bl); | |
3017 | { | |
3018 | // validate collection name | |
3019 | assert(coll.is_pg()); | |
3020 | } | |
3021 | ||
3022 | // try for v8 | |
3023 | set<string> keys; | |
3024 | keys.insert(infover_key); | |
3025 | keys.insert(epoch_key); | |
3026 | map<string,bufferlist> values; | |
3027 | int r = store->omap_get_values(coll, pgmeta_oid, keys, &values); | |
3028 | if (r == 0) { | |
3029 | assert(values.size() == 2); | |
3030 | ||
3031 | // sanity check version | |
3032 | bufferlist::iterator bp = values[infover_key].begin(); | |
3033 | __u8 struct_v = 0; | |
3034 | ::decode(struct_v, bp); | |
3035 | assert(struct_v >= 8); | |
3036 | ||
3037 | // get epoch | |
3038 | bp = values[epoch_key].begin(); | |
3039 | ::decode(cur_epoch, bp); | |
3040 | } else { | |
3041 | // probably bug 10617; see OSD::load_pgs() | |
3042 | return -1; | |
3043 | } | |
3044 | ||
3045 | *pepoch = cur_epoch; | |
3046 | return 0; | |
3047 | } | |
3048 | ||
3049 | #pragma GCC diagnostic pop | |
3050 | #pragma GCC diagnostic warning "-Wpragmas" | |
3051 | ||
3052 | void PG::write_if_dirty(ObjectStore::Transaction& t) | |
3053 | { | |
3054 | map<string,bufferlist> km; | |
3055 | if (dirty_big_info || dirty_info) | |
3056 | prepare_write_info(&km); | |
3057 | pg_log.write_log_and_missing(t, &km, coll, pgmeta_oid, pool.info.require_rollback()); | |
3058 | if (!km.empty()) | |
3059 | t.omap_setkeys(coll, pgmeta_oid, km); | |
3060 | } | |
3061 | ||
3062 | void PG::trim_log() | |
3063 | { | |
3064 | assert(is_primary()); | |
3065 | calc_trim_to(); | |
3066 | dout(10) << __func__ << " to " << pg_trim_to << dendl; | |
3067 | if (pg_trim_to != eversion_t()) { | |
3068 | // inform peers to trim log | |
3069 | assert(!actingbackfill.empty()); | |
3070 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
3071 | i != actingbackfill.end(); | |
3072 | ++i) { | |
3073 | if (*i == pg_whoami) continue; | |
3074 | osd->send_message_osd_cluster( | |
3075 | i->osd, | |
3076 | new MOSDPGTrim( | |
3077 | get_osdmap()->get_epoch(), | |
3078 | spg_t(info.pgid.pgid, i->shard), | |
3079 | pg_trim_to), | |
3080 | get_osdmap()->get_epoch()); | |
3081 | } | |
3082 | ||
3083 | // trim primary as well | |
3084 | pg_log.trim(pg_trim_to, info); | |
3085 | dirty_info = true; | |
3086 | } | |
3087 | } | |
3088 | ||
3089 | void PG::add_log_entry(const pg_log_entry_t& e, bool applied) | |
3090 | { | |
3091 | // raise last_complete only if we were previously up to date | |
3092 | if (info.last_complete == info.last_update) | |
3093 | info.last_complete = e.version; | |
3094 | ||
3095 | // raise last_update. | |
3096 | assert(e.version > info.last_update); | |
3097 | info.last_update = e.version; | |
3098 | ||
3099 | // raise user_version, if it increased (it may have not get bumped | |
3100 | // by all logged updates) | |
3101 | if (e.user_version > info.last_user_version) | |
3102 | info.last_user_version = e.user_version; | |
3103 | ||
3104 | // log mutation | |
3105 | pg_log.add(e, applied); | |
3106 | dout(10) << "add_log_entry " << e << dendl; | |
3107 | } | |
3108 | ||
3109 | ||
3110 | void PG::append_log( | |
3111 | const vector<pg_log_entry_t>& logv, | |
3112 | eversion_t trim_to, | |
3113 | eversion_t roll_forward_to, | |
3114 | ObjectStore::Transaction &t, | |
3115 | bool transaction_applied) | |
3116 | { | |
3117 | if (transaction_applied) | |
3118 | update_snap_map(logv, t); | |
3119 | ||
3120 | /* The primary has sent an info updating the history, but it may not | |
3121 | * have arrived yet. We want to make sure that we cannot remember this | |
3122 | * write without remembering that it happened in an interval which went | |
3123 | * active in epoch history.last_epoch_started. | |
3124 | */ | |
3125 | if (info.last_epoch_started != info.history.last_epoch_started) { | |
3126 | info.history.last_epoch_started = info.last_epoch_started; | |
3127 | } | |
3128 | if (info.last_interval_started != info.history.last_interval_started) { | |
3129 | info.history.last_interval_started = info.last_interval_started; | |
3130 | } | |
3131 | dout(10) << "append_log " << pg_log.get_log() << " " << logv << dendl; | |
3132 | ||
3133 | PGLogEntryHandler handler{this, &t}; | |
3134 | if (!transaction_applied) { | |
3135 | /* We must be a backfill peer, so it's ok if we apply | |
3136 | * out-of-turn since we won't be considered when | |
3137 | * determining a min possible last_update. | |
3138 | */ | |
3139 | pg_log.roll_forward(&handler); | |
3140 | } | |
3141 | ||
3142 | for (vector<pg_log_entry_t>::const_iterator p = logv.begin(); | |
3143 | p != logv.end(); | |
3144 | ++p) { | |
3145 | add_log_entry(*p, transaction_applied); | |
3146 | ||
3147 | /* We don't want to leave the rollforward artifacts around | |
3148 | * here past last_backfill. It's ok for the same reason as | |
3149 | * above */ | |
3150 | if (transaction_applied && | |
3151 | p->soid > info.last_backfill) { | |
3152 | pg_log.roll_forward(&handler); | |
3153 | } | |
3154 | } | |
3155 | auto last = logv.rbegin(); | |
3156 | if (is_primary() && last != logv.rend()) { | |
3157 | projected_log.skip_can_rollback_to_to_head(); | |
3158 | projected_log.trim(cct, last->version, nullptr); | |
3159 | } | |
3160 | ||
3161 | if (transaction_applied && roll_forward_to > pg_log.get_can_rollback_to()) { | |
3162 | pg_log.roll_forward_to( | |
3163 | roll_forward_to, | |
3164 | &handler); | |
3165 | t.register_on_applied( | |
3166 | new C_UpdateLastRollbackInfoTrimmedToApplied( | |
3167 | this, | |
3168 | get_osdmap()->get_epoch(), | |
3169 | roll_forward_to)); | |
3170 | } | |
3171 | ||
3172 | pg_log.trim(trim_to, info); | |
3173 | ||
3174 | // update the local pg, pg log | |
3175 | dirty_info = true; | |
3176 | write_if_dirty(t); | |
3177 | } | |
3178 | ||
3179 | bool PG::check_log_for_corruption(ObjectStore *store) | |
3180 | { | |
3181 | /// TODO: this method needs to work with the omap log | |
3182 | return true; | |
3183 | } | |
3184 | ||
3185 | //! Get the name we're going to save our corrupt page log as | |
3186 | std::string PG::get_corrupt_pg_log_name() const | |
3187 | { | |
3188 | const int MAX_BUF = 512; | |
3189 | char buf[MAX_BUF]; | |
3190 | struct tm tm_buf; | |
3191 | time_t my_time(time(NULL)); | |
3192 | const struct tm *t = localtime_r(&my_time, &tm_buf); | |
3193 | int ret = strftime(buf, sizeof(buf), "corrupt_log_%Y-%m-%d_%k:%M_", t); | |
3194 | if (ret == 0) { | |
3195 | dout(0) << "strftime failed" << dendl; | |
3196 | return "corrupt_log_unknown_time"; | |
3197 | } | |
3198 | string out(buf); | |
3199 | out += stringify(info.pgid); | |
3200 | return out; | |
3201 | } | |
3202 | ||
3203 | int PG::read_info( | |
3204 | ObjectStore *store, spg_t pgid, const coll_t &coll, bufferlist &bl, | |
3205 | pg_info_t &info, PastIntervals &past_intervals, | |
3206 | __u8 &struct_v) | |
3207 | { | |
3208 | // try for v8 or later | |
3209 | set<string> keys; | |
3210 | keys.insert(infover_key); | |
3211 | keys.insert(info_key); | |
3212 | keys.insert(biginfo_key); | |
3213 | keys.insert(fastinfo_key); | |
3214 | ghobject_t pgmeta_oid(pgid.make_pgmeta_oid()); | |
3215 | map<string,bufferlist> values; | |
3216 | int r = store->omap_get_values(coll, pgmeta_oid, keys, &values); | |
3217 | if (r == 0) { | |
3218 | assert(values.size() == 3 || | |
3219 | values.size() == 4); | |
3220 | ||
3221 | bufferlist::iterator p = values[infover_key].begin(); | |
3222 | ::decode(struct_v, p); | |
3223 | assert(struct_v >= 8); | |
3224 | ||
3225 | p = values[info_key].begin(); | |
3226 | ::decode(info, p); | |
3227 | ||
3228 | p = values[biginfo_key].begin(); | |
3229 | if (struct_v >= 10) { | |
3230 | ::decode(past_intervals, p); | |
3231 | } else { | |
3232 | past_intervals.decode_classic(p); | |
3233 | } | |
3234 | ::decode(info.purged_snaps, p); | |
3235 | ||
3236 | p = values[fastinfo_key].begin(); | |
3237 | if (!p.end()) { | |
3238 | pg_fast_info_t fast; | |
3239 | ::decode(fast, p); | |
3240 | fast.try_apply_to(&info); | |
3241 | } | |
3242 | return 0; | |
3243 | } | |
3244 | ||
3245 | // legacy (ver < 8) | |
3246 | ghobject_t infos_oid(OSD::make_infos_oid()); | |
3247 | bufferlist::iterator p = bl.begin(); | |
3248 | ::decode(struct_v, p); | |
3249 | assert(struct_v == 7); | |
3250 | ||
3251 | // get info out of leveldb | |
3252 | string k = get_info_key(info.pgid); | |
3253 | string bk = get_biginfo_key(info.pgid); | |
3254 | keys.clear(); | |
3255 | keys.insert(k); | |
3256 | keys.insert(bk); | |
3257 | values.clear(); | |
3258 | store->omap_get_values(coll_t::meta(), ghobject_t(infos_oid), keys, &values); | |
3259 | assert(values.size() == 2); | |
3260 | ||
3261 | p = values[k].begin(); | |
3262 | ::decode(info, p); | |
3263 | ||
3264 | p = values[bk].begin(); | |
3265 | ::decode(past_intervals, p); | |
3266 | interval_set<snapid_t> snap_collections; // obsolete | |
3267 | ::decode(snap_collections, p); | |
3268 | ::decode(info.purged_snaps, p); | |
3269 | return 0; | |
3270 | } | |
3271 | ||
3272 | void PG::read_state(ObjectStore *store, bufferlist &bl) | |
3273 | { | |
3274 | int r = read_info(store, pg_id, coll, bl, info, past_intervals, | |
3275 | info_struct_v); | |
3276 | assert(r >= 0); | |
3277 | ||
3278 | last_written_info = info; | |
3279 | ||
3280 | ostringstream oss; | |
3281 | pg_log.read_log_and_missing( | |
3282 | store, | |
3283 | coll, | |
3284 | info_struct_v < 8 ? coll_t::meta() : coll, | |
3285 | ghobject_t(info_struct_v < 8 ? OSD::make_pg_log_oid(pg_id) : pgmeta_oid), | |
3286 | info, | |
3287 | oss, | |
3288 | cct->_conf->osd_ignore_stale_divergent_priors, | |
3289 | cct->_conf->osd_debug_verify_missing_on_start); | |
3290 | if (oss.tellp()) | |
3291 | osd->clog->error() << oss.rdbuf(); | |
3292 | ||
3293 | // log any weirdness | |
3294 | log_weirdness(); | |
3295 | } | |
3296 | ||
3297 | void PG::log_weirdness() | |
3298 | { | |
3299 | if (pg_log.get_tail() != info.log_tail) | |
3300 | osd->clog->error() << info.pgid | |
3301 | << " info mismatch, log.tail " << pg_log.get_tail() | |
3302 | << " != info.log_tail " << info.log_tail; | |
3303 | if (pg_log.get_head() != info.last_update) | |
3304 | osd->clog->error() << info.pgid | |
3305 | << " info mismatch, log.head " << pg_log.get_head() | |
3306 | << " != info.last_update " << info.last_update; | |
3307 | ||
3308 | if (!pg_log.get_log().empty()) { | |
3309 | // sloppy check | |
3310 | if ((pg_log.get_log().log.begin()->version <= pg_log.get_tail())) | |
3311 | osd->clog->error() << info.pgid | |
3312 | << " log bound mismatch, info (" << pg_log.get_tail() << "," | |
3313 | << pg_log.get_head() << "]" | |
3314 | << " actual [" | |
3315 | << pg_log.get_log().log.begin()->version << "," | |
3316 | << pg_log.get_log().log.rbegin()->version << "]"; | |
3317 | } | |
3318 | ||
3319 | if (pg_log.get_log().caller_ops.size() > pg_log.get_log().log.size()) { | |
3320 | osd->clog->error() << info.pgid | |
3321 | << " caller_ops.size " << pg_log.get_log().caller_ops.size() | |
3322 | << " > log size " << pg_log.get_log().log.size(); | |
3323 | } | |
3324 | } | |
3325 | ||
3326 | void PG::update_snap_map( | |
3327 | const vector<pg_log_entry_t> &log_entries, | |
3328 | ObjectStore::Transaction &t) | |
3329 | { | |
3330 | for (vector<pg_log_entry_t>::const_iterator i = log_entries.begin(); | |
3331 | i != log_entries.end(); | |
3332 | ++i) { | |
3333 | OSDriver::OSTransaction _t(osdriver.get_transaction(&t)); | |
3334 | if (i->soid.snap < CEPH_MAXSNAP) { | |
3335 | if (i->is_delete()) { | |
3336 | int r = snap_mapper.remove_oid( | |
3337 | i->soid, | |
3338 | &_t); | |
3339 | assert(r == 0); | |
3340 | } else if (i->is_update()) { | |
3341 | assert(i->snaps.length() > 0); | |
3342 | vector<snapid_t> snaps; | |
3343 | bufferlist snapbl = i->snaps; | |
3344 | bufferlist::iterator p = snapbl.begin(); | |
3345 | try { | |
3346 | ::decode(snaps, p); | |
3347 | } catch (...) { | |
3348 | snaps.clear(); | |
3349 | } | |
3350 | set<snapid_t> _snaps(snaps.begin(), snaps.end()); | |
3351 | ||
3352 | if (i->is_clone() || i->is_promote()) { | |
3353 | snap_mapper.add_oid( | |
3354 | i->soid, | |
3355 | _snaps, | |
3356 | &_t); | |
3357 | } else if (i->is_modify()) { | |
3358 | assert(i->is_modify()); | |
3359 | int r = snap_mapper.update_snaps( | |
3360 | i->soid, | |
3361 | _snaps, | |
3362 | 0, | |
3363 | &_t); | |
3364 | assert(r == 0); | |
3365 | } else { | |
3366 | assert(i->is_clean()); | |
3367 | } | |
3368 | } | |
3369 | } | |
3370 | } | |
3371 | } | |
3372 | ||
3373 | /** | |
3374 | * filter trimming|trimmed snaps out of snapcontext | |
3375 | */ | |
3376 | void PG::filter_snapc(vector<snapid_t> &snaps) | |
3377 | { | |
3378 | //nothing needs to trim, we can return immediately | |
3379 | if(snap_trimq.empty() && info.purged_snaps.empty()) | |
3380 | return; | |
3381 | ||
3382 | bool filtering = false; | |
3383 | vector<snapid_t> newsnaps; | |
3384 | for (vector<snapid_t>::iterator p = snaps.begin(); | |
3385 | p != snaps.end(); | |
3386 | ++p) { | |
3387 | if (snap_trimq.contains(*p) || info.purged_snaps.contains(*p)) { | |
3388 | if (!filtering) { | |
3389 | // start building a new vector with what we've seen so far | |
3390 | dout(10) << "filter_snapc filtering " << snaps << dendl; | |
3391 | newsnaps.insert(newsnaps.begin(), snaps.begin(), p); | |
3392 | filtering = true; | |
3393 | } | |
3394 | dout(20) << "filter_snapc removing trimq|purged snap " << *p << dendl; | |
3395 | } else { | |
3396 | if (filtering) | |
3397 | newsnaps.push_back(*p); // continue building new vector | |
3398 | } | |
3399 | } | |
3400 | if (filtering) { | |
3401 | snaps.swap(newsnaps); | |
3402 | dout(10) << "filter_snapc result " << snaps << dendl; | |
3403 | } | |
3404 | } | |
3405 | ||
3406 | void PG::requeue_object_waiters(map<hobject_t, list<OpRequestRef>>& m) | |
3407 | { | |
3408 | for (map<hobject_t, list<OpRequestRef>>::iterator it = m.begin(); | |
3409 | it != m.end(); | |
3410 | ++it) | |
3411 | requeue_ops(it->second); | |
3412 | m.clear(); | |
3413 | } | |
3414 | ||
3415 | void PG::requeue_op(OpRequestRef op) | |
3416 | { | |
3417 | auto p = waiting_for_map.find(op->get_source()); | |
3418 | if (p != waiting_for_map.end()) { | |
3419 | dout(20) << __func__ << " " << op << " (waiting_for_map " << p->first << ")" | |
3420 | << dendl; | |
3421 | p->second.push_front(op); | |
3422 | } else { | |
3423 | dout(20) << __func__ << " " << op << dendl; | |
3424 | osd->enqueue_front(info.pgid, PGQueueable(op, get_osdmap()->get_epoch())); | |
3425 | } | |
3426 | } | |
3427 | ||
3428 | void PG::requeue_ops(list<OpRequestRef> &ls) | |
3429 | { | |
3430 | for (list<OpRequestRef>::reverse_iterator i = ls.rbegin(); | |
3431 | i != ls.rend(); | |
3432 | ++i) { | |
3433 | auto p = waiting_for_map.find((*i)->get_source()); | |
3434 | if (p != waiting_for_map.end()) { | |
3435 | dout(20) << __func__ << " " << *i << " (waiting_for_map " << p->first | |
3436 | << ")" << dendl; | |
3437 | p->second.push_front(*i); | |
3438 | } else { | |
3439 | dout(20) << __func__ << " " << *i << dendl; | |
3440 | osd->enqueue_front(info.pgid, PGQueueable(*i, get_osdmap()->get_epoch())); | |
3441 | } | |
3442 | } | |
3443 | ls.clear(); | |
3444 | } | |
3445 | ||
3446 | void PG::requeue_map_waiters() | |
3447 | { | |
3448 | epoch_t epoch = get_osdmap()->get_epoch(); | |
3449 | auto p = waiting_for_map.begin(); | |
3450 | while (p != waiting_for_map.end()) { | |
3451 | if (epoch < p->second.front()->min_epoch) { | |
3452 | dout(20) << __func__ << " " << p->first << " front op " | |
3453 | << p->second.front() << " must still wait, doing nothing" | |
3454 | << dendl; | |
3455 | ++p; | |
3456 | } else { | |
3457 | dout(20) << __func__ << " " << p->first << " " << p->second << dendl; | |
3458 | for (auto q = p->second.rbegin(); q != p->second.rend(); ++q) { | |
3459 | osd->enqueue_front(info.pgid, PGQueueable(*q, epoch)); | |
3460 | } | |
3461 | p = waiting_for_map.erase(p); | |
3462 | } | |
3463 | } | |
3464 | } | |
3465 | ||
3466 | ||
3467 | // ========================================================================================== | |
3468 | // SCRUB | |
3469 | ||
3470 | /* | |
3471 | * when holding pg and sched_scrub_lock, then the states are: | |
3472 | * scheduling: | |
3473 | * scrubber.reserved = true | |
3474 | * scrub_rserved_peers includes whoami | |
3475 | * osd->scrub_pending++ | |
3476 | * scheduling, replica declined: | |
3477 | * scrubber.reserved = true | |
3478 | * scrubber.reserved_peers includes -1 | |
3479 | * osd->scrub_pending++ | |
3480 | * pending: | |
3481 | * scrubber.reserved = true | |
3482 | * scrubber.reserved_peers.size() == acting.size(); | |
3483 | * pg on scrub_wq | |
3484 | * osd->scrub_pending++ | |
3485 | * scrubbing: | |
3486 | * scrubber.reserved = false; | |
3487 | * scrubber.reserved_peers empty | |
3488 | * osd->scrubber.active++ | |
3489 | */ | |
3490 | ||
3491 | // returns true if a scrub has been newly kicked off | |
3492 | bool PG::sched_scrub() | |
3493 | { | |
3494 | bool nodeep_scrub = false; | |
3495 | assert(is_locked()); | |
3496 | if (!(is_primary() && is_active() && is_clean() && !is_scrubbing())) { | |
3497 | return false; | |
3498 | } | |
3499 | ||
3500 | double deep_scrub_interval = 0; | |
3501 | pool.info.opts.get(pool_opts_t::DEEP_SCRUB_INTERVAL, &deep_scrub_interval); | |
3502 | if (deep_scrub_interval <= 0) { | |
3503 | deep_scrub_interval = cct->_conf->osd_deep_scrub_interval; | |
3504 | } | |
3505 | bool time_for_deep = ceph_clock_now() >= | |
3506 | info.history.last_deep_scrub_stamp + deep_scrub_interval; | |
3507 | ||
3508 | bool deep_coin_flip = false; | |
3509 | // Only add random deep scrubs when NOT user initiated scrub | |
3510 | if (!scrubber.must_scrub) | |
3511 | deep_coin_flip = (rand() % 100) < cct->_conf->osd_deep_scrub_randomize_ratio * 100; | |
3512 | dout(20) << __func__ << ": time_for_deep=" << time_for_deep << " deep_coin_flip=" << deep_coin_flip << dendl; | |
3513 | ||
3514 | time_for_deep = (time_for_deep || deep_coin_flip); | |
3515 | ||
3516 | //NODEEP_SCRUB so ignore time initiated deep-scrub | |
3517 | if (osd->osd->get_osdmap()->test_flag(CEPH_OSDMAP_NODEEP_SCRUB) || | |
3518 | pool.info.has_flag(pg_pool_t::FLAG_NODEEP_SCRUB)) { | |
3519 | time_for_deep = false; | |
3520 | nodeep_scrub = true; | |
3521 | } | |
3522 | ||
3523 | if (!scrubber.must_scrub) { | |
3524 | assert(!scrubber.must_deep_scrub); | |
3525 | ||
3526 | //NOSCRUB so skip regular scrubs | |
3527 | if ((osd->osd->get_osdmap()->test_flag(CEPH_OSDMAP_NOSCRUB) || | |
3528 | pool.info.has_flag(pg_pool_t::FLAG_NOSCRUB)) && !time_for_deep) { | |
3529 | if (scrubber.reserved) { | |
3530 | // cancel scrub if it is still in scheduling, | |
3531 | // so pgs from other pools where scrub are still legal | |
3532 | // have a chance to go ahead with scrubbing. | |
3533 | clear_scrub_reserved(); | |
3534 | scrub_unreserve_replicas(); | |
3535 | } | |
3536 | return false; | |
3537 | } | |
3538 | } | |
3539 | ||
3540 | if (cct->_conf->osd_scrub_auto_repair | |
3541 | && get_pgbackend()->auto_repair_supported() | |
3542 | && time_for_deep | |
3543 | // respect the command from user, and not do auto-repair | |
3544 | && !scrubber.must_repair | |
3545 | && !scrubber.must_scrub | |
3546 | && !scrubber.must_deep_scrub) { | |
3547 | dout(20) << __func__ << ": auto repair with deep scrubbing" << dendl; | |
3548 | scrubber.auto_repair = true; | |
3549 | } else { | |
3550 | // this happens when user issue the scrub/repair command during | |
3551 | // the scheduling of the scrub/repair (e.g. request reservation) | |
3552 | scrubber.auto_repair = false; | |
3553 | } | |
3554 | ||
3555 | bool ret = true; | |
3556 | if (!scrubber.reserved) { | |
3557 | assert(scrubber.reserved_peers.empty()); | |
3558 | if (osd->inc_scrubs_pending()) { | |
3559 | dout(20) << "sched_scrub: reserved locally, reserving replicas" << dendl; | |
3560 | scrubber.reserved = true; | |
3561 | scrubber.reserved_peers.insert(pg_whoami); | |
3562 | scrub_reserve_replicas(); | |
3563 | } else { | |
3564 | dout(20) << "sched_scrub: failed to reserve locally" << dendl; | |
3565 | ret = false; | |
3566 | } | |
3567 | } | |
3568 | if (scrubber.reserved) { | |
3569 | if (scrubber.reserve_failed) { | |
3570 | dout(20) << "sched_scrub: failed, a peer declined" << dendl; | |
3571 | clear_scrub_reserved(); | |
3572 | scrub_unreserve_replicas(); | |
3573 | ret = false; | |
3574 | } else if (scrubber.reserved_peers.size() == acting.size()) { | |
3575 | dout(20) << "sched_scrub: success, reserved self and replicas" << dendl; | |
3576 | if (time_for_deep) { | |
3577 | dout(10) << "sched_scrub: scrub will be deep" << dendl; | |
3578 | state_set(PG_STATE_DEEP_SCRUB); | |
3579 | } else if (!scrubber.must_deep_scrub && info.stats.stats.sum.num_deep_scrub_errors) { | |
3580 | if (!nodeep_scrub) { | |
3581 | osd->clog->info() << "osd." << osd->whoami | |
3582 | << " pg " << info.pgid | |
3583 | << " Deep scrub errors, upgrading scrub to deep-scrub"; | |
3584 | state_set(PG_STATE_DEEP_SCRUB); | |
3585 | } else if (!scrubber.must_scrub) { | |
3586 | osd->clog->error() << "osd." << osd->whoami | |
3587 | << " pg " << info.pgid | |
3588 | << " Regular scrub skipped due to deep-scrub errors and nodeep-scrub set"; | |
3589 | clear_scrub_reserved(); | |
3590 | scrub_unreserve_replicas(); | |
3591 | return false; | |
3592 | } else { | |
3593 | osd->clog->error() << "osd." << osd->whoami | |
3594 | << " pg " << info.pgid | |
3595 | << " Regular scrub request, losing deep-scrub details"; | |
3596 | } | |
3597 | } | |
3598 | queue_scrub(); | |
3599 | } else { | |
3600 | // none declined, since scrubber.reserved is set | |
3601 | dout(20) << "sched_scrub: reserved " << scrubber.reserved_peers << ", waiting for replicas" << dendl; | |
3602 | } | |
3603 | } | |
3604 | ||
3605 | return ret; | |
3606 | } | |
3607 | ||
3608 | void PG::reg_next_scrub() | |
3609 | { | |
3610 | if (!is_primary()) | |
3611 | return; | |
3612 | ||
3613 | utime_t reg_stamp; | |
3614 | if (scrubber.must_scrub || | |
3615 | (info.stats.stats_invalid && cct->_conf->osd_scrub_invalid_stats)) { | |
3616 | reg_stamp = ceph_clock_now(); | |
3617 | } else { | |
3618 | reg_stamp = info.history.last_scrub_stamp; | |
3619 | } | |
3620 | // note down the sched_time, so we can locate this scrub, and remove it | |
3621 | // later on. | |
3622 | double scrub_min_interval = 0, scrub_max_interval = 0; | |
3623 | pool.info.opts.get(pool_opts_t::SCRUB_MIN_INTERVAL, &scrub_min_interval); | |
3624 | pool.info.opts.get(pool_opts_t::SCRUB_MAX_INTERVAL, &scrub_max_interval); | |
3625 | assert(scrubber.scrub_reg_stamp == utime_t()); | |
3626 | scrubber.scrub_reg_stamp = osd->reg_pg_scrub(info.pgid, | |
3627 | reg_stamp, | |
3628 | scrub_min_interval, | |
3629 | scrub_max_interval, | |
3630 | scrubber.must_scrub); | |
3631 | } | |
3632 | ||
3633 | void PG::unreg_next_scrub() | |
3634 | { | |
3635 | if (is_primary()) { | |
3636 | osd->unreg_pg_scrub(info.pgid, scrubber.scrub_reg_stamp); | |
3637 | scrubber.scrub_reg_stamp = utime_t(); | |
3638 | } | |
3639 | } | |
3640 | ||
3641 | void PG::do_replica_scrub_map(OpRequestRef op) | |
3642 | { | |
3643 | const MOSDRepScrubMap *m = static_cast<const MOSDRepScrubMap*>(op->get_req()); | |
3644 | dout(7) << __func__ << " " << *m << dendl; | |
3645 | if (m->map_epoch < info.history.same_interval_since) { | |
3646 | dout(10) << __func__ << " discarding old from " | |
3647 | << m->map_epoch << " < " << info.history.same_interval_since | |
3648 | << dendl; | |
3649 | return; | |
3650 | } | |
3651 | if (!scrubber.is_chunky_scrub_active()) { | |
3652 | dout(10) << __func__ << " scrub isn't active" << dendl; | |
3653 | return; | |
3654 | } | |
3655 | ||
3656 | op->mark_started(); | |
3657 | ||
3658 | bufferlist::iterator p = const_cast<bufferlist&>(m->get_data()).begin(); | |
3659 | scrubber.received_maps[m->from].decode(p, info.pgid.pool()); | |
3660 | dout(10) << "map version is " | |
3661 | << scrubber.received_maps[m->from].valid_through | |
3662 | << dendl; | |
3663 | ||
3664 | --scrubber.waiting_on; | |
3665 | scrubber.waiting_on_whom.erase(m->from); | |
3666 | if (scrubber.waiting_on == 0) { | |
3667 | if (ops_blocked_by_scrub()) { | |
3668 | requeue_scrub(true); | |
3669 | } else { | |
3670 | requeue_scrub(false); | |
3671 | } | |
3672 | } | |
3673 | } | |
3674 | ||
3675 | void PG::sub_op_scrub_map(OpRequestRef op) | |
3676 | { | |
3677 | // for legacy jewel compatibility only | |
3678 | const MOSDSubOp *m = static_cast<const MOSDSubOp *>(op->get_req()); | |
3679 | assert(m->get_type() == MSG_OSD_SUBOP); | |
3680 | dout(7) << "sub_op_scrub_map" << dendl; | |
3681 | ||
3682 | if (m->map_epoch < info.history.same_interval_since) { | |
3683 | dout(10) << "sub_op_scrub discarding old sub_op from " | |
3684 | << m->map_epoch << " < " << info.history.same_interval_since << dendl; | |
3685 | return; | |
3686 | } | |
3687 | ||
3688 | if (!scrubber.is_chunky_scrub_active()) { | |
3689 | dout(10) << "sub_op_scrub_map scrub isn't active" << dendl; | |
3690 | return; | |
3691 | } | |
3692 | ||
3693 | op->mark_started(); | |
3694 | ||
3695 | dout(10) << " got " << m->from << " scrub map" << dendl; | |
3696 | bufferlist::iterator p = const_cast<bufferlist&>(m->get_data()).begin(); | |
3697 | ||
3698 | scrubber.received_maps[m->from].decode(p, info.pgid.pool()); | |
3699 | dout(10) << "map version is " | |
3700 | << scrubber.received_maps[m->from].valid_through | |
3701 | << dendl; | |
3702 | ||
3703 | --scrubber.waiting_on; | |
3704 | scrubber.waiting_on_whom.erase(m->from); | |
3705 | ||
3706 | if (scrubber.waiting_on == 0) { | |
3707 | if (ops_blocked_by_scrub()) { | |
3708 | requeue_scrub(true); | |
3709 | } else { | |
3710 | requeue_scrub(false); | |
3711 | } | |
3712 | } | |
3713 | } | |
3714 | ||
3715 | // send scrub v3 messages (chunky scrub) | |
3716 | void PG::_request_scrub_map( | |
3717 | pg_shard_t replica, eversion_t version, | |
3718 | hobject_t start, hobject_t end, | |
3719 | bool deep, uint32_t seed) | |
3720 | { | |
3721 | assert(replica != pg_whoami); | |
3722 | dout(10) << "scrub requesting scrubmap from osd." << replica | |
3723 | << " deep " << (int)deep << " seed " << seed << dendl; | |
3724 | MOSDRepScrub *repscrubop = new MOSDRepScrub( | |
3725 | spg_t(info.pgid.pgid, replica.shard), version, | |
3726 | get_osdmap()->get_epoch(), | |
3727 | get_last_peering_reset(), | |
3728 | start, end, deep, seed); | |
3729 | // default priority, we want the rep scrub processed prior to any recovery | |
3730 | // or client io messages (we are holding a lock!) | |
3731 | osd->send_message_osd_cluster( | |
3732 | replica.osd, repscrubop, get_osdmap()->get_epoch()); | |
3733 | } | |
3734 | ||
3735 | void PG::handle_scrub_reserve_request(OpRequestRef op) | |
3736 | { | |
3737 | dout(7) << __func__ << " " << *op->get_req() << dendl; | |
3738 | op->mark_started(); | |
3739 | if (scrubber.reserved) { | |
3740 | dout(10) << __func__ << " ignoring reserve request: Already reserved" | |
3741 | << dendl; | |
3742 | return; | |
3743 | } | |
3744 | scrubber.reserved = osd->inc_scrubs_pending(); | |
3745 | if (op->get_req()->get_type() == MSG_OSD_SCRUB_RESERVE) { | |
3746 | const MOSDScrubReserve *m = | |
3747 | static_cast<const MOSDScrubReserve*>(op->get_req()); | |
3748 | Message *reply = new MOSDScrubReserve( | |
3749 | spg_t(info.pgid.pgid, primary.shard), | |
3750 | m->map_epoch, | |
3751 | scrubber.reserved ? MOSDScrubReserve::GRANT : MOSDScrubReserve::REJECT, | |
3752 | pg_whoami); | |
3753 | osd->send_message_osd_cluster(reply, op->get_req()->get_connection()); | |
3754 | } else { | |
3755 | // for jewel compat only | |
3756 | const MOSDSubOp *req = static_cast<const MOSDSubOp*>(op->get_req()); | |
3757 | assert(req->get_type() == MSG_OSD_SUBOP); | |
3758 | MOSDSubOpReply *reply = new MOSDSubOpReply( | |
3759 | req, pg_whoami, 0, get_osdmap()->get_epoch(), CEPH_OSD_FLAG_ACK); | |
3760 | ::encode(scrubber.reserved, reply->get_data()); | |
3761 | osd->send_message_osd_cluster(reply, op->get_req()->get_connection()); | |
3762 | } | |
3763 | } | |
3764 | ||
3765 | void PG::handle_scrub_reserve_grant(OpRequestRef op, pg_shard_t from) | |
3766 | { | |
3767 | dout(7) << __func__ << " " << *op->get_req() << dendl; | |
3768 | op->mark_started(); | |
3769 | if (!scrubber.reserved) { | |
3770 | dout(10) << "ignoring obsolete scrub reserve reply" << dendl; | |
3771 | return; | |
3772 | } | |
3773 | if (scrubber.reserved_peers.find(from) != scrubber.reserved_peers.end()) { | |
3774 | dout(10) << " already had osd." << from << " reserved" << dendl; | |
3775 | } else { | |
3776 | dout(10) << " osd." << from << " scrub reserve = success" << dendl; | |
3777 | scrubber.reserved_peers.insert(from); | |
3778 | sched_scrub(); | |
3779 | } | |
3780 | } | |
3781 | ||
3782 | void PG::handle_scrub_reserve_reject(OpRequestRef op, pg_shard_t from) | |
3783 | { | |
3784 | dout(7) << __func__ << " " << *op->get_req() << dendl; | |
3785 | op->mark_started(); | |
3786 | if (!scrubber.reserved) { | |
3787 | dout(10) << "ignoring obsolete scrub reserve reply" << dendl; | |
3788 | return; | |
3789 | } | |
3790 | if (scrubber.reserved_peers.find(from) != scrubber.reserved_peers.end()) { | |
3791 | dout(10) << " already had osd." << from << " reserved" << dendl; | |
3792 | } else { | |
3793 | /* One decline stops this pg from being scheduled for scrubbing. */ | |
3794 | dout(10) << " osd." << from << " scrub reserve = fail" << dendl; | |
3795 | scrubber.reserve_failed = true; | |
3796 | sched_scrub(); | |
3797 | } | |
3798 | } | |
3799 | ||
3800 | void PG::handle_scrub_reserve_release(OpRequestRef op) | |
3801 | { | |
3802 | dout(7) << __func__ << " " << *op->get_req() << dendl; | |
3803 | op->mark_started(); | |
3804 | clear_scrub_reserved(); | |
3805 | } | |
3806 | ||
3807 | void PG::reject_reservation() | |
3808 | { | |
3809 | osd->send_message_osd_cluster( | |
3810 | primary.osd, | |
3811 | new MBackfillReserve( | |
3812 | MBackfillReserve::REJECT, | |
3813 | spg_t(info.pgid.pgid, primary.shard), | |
3814 | get_osdmap()->get_epoch()), | |
3815 | get_osdmap()->get_epoch()); | |
3816 | } | |
3817 | ||
3818 | void PG::schedule_backfill_full_retry() | |
3819 | { | |
3820 | Mutex::Locker lock(osd->recovery_request_lock); | |
3821 | osd->recovery_request_timer.add_event_after( | |
3822 | cct->_conf->osd_backfill_retry_interval, | |
3823 | new QueuePeeringEvt<RequestBackfill>( | |
3824 | this, get_osdmap()->get_epoch(), | |
3825 | RequestBackfill())); | |
3826 | } | |
3827 | ||
3828 | void PG::schedule_recovery_full_retry() | |
3829 | { | |
3830 | Mutex::Locker lock(osd->recovery_request_lock); | |
3831 | osd->recovery_request_timer.add_event_after( | |
3832 | cct->_conf->osd_recovery_retry_interval, | |
3833 | new QueuePeeringEvt<DoRecovery>( | |
3834 | this, get_osdmap()->get_epoch(), | |
3835 | DoRecovery())); | |
3836 | } | |
3837 | ||
3838 | void PG::clear_scrub_reserved() | |
3839 | { | |
3840 | scrubber.reserved_peers.clear(); | |
3841 | scrubber.reserve_failed = false; | |
3842 | ||
3843 | if (scrubber.reserved) { | |
3844 | scrubber.reserved = false; | |
3845 | osd->dec_scrubs_pending(); | |
3846 | } | |
3847 | } | |
3848 | ||
3849 | void PG::scrub_reserve_replicas() | |
3850 | { | |
3851 | assert(backfill_targets.empty()); | |
3852 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
3853 | i != actingbackfill.end(); | |
3854 | ++i) { | |
3855 | if (*i == pg_whoami) continue; | |
3856 | dout(10) << "scrub requesting reserve from osd." << *i << dendl; | |
3857 | if (HAVE_FEATURE(get_min_acting_features(), SERVER_LUMINOUS)) { | |
3858 | osd->send_message_osd_cluster( | |
3859 | i->osd, | |
3860 | new MOSDScrubReserve(spg_t(info.pgid.pgid, i->shard), | |
3861 | get_osdmap()->get_epoch(), | |
3862 | MOSDScrubReserve::REQUEST, pg_whoami), | |
3863 | get_osdmap()->get_epoch()); | |
3864 | } else { | |
3865 | // for jewel compat only | |
3866 | vector<OSDOp> scrub(1); | |
3867 | scrub[0].op.op = CEPH_OSD_OP_SCRUB_RESERVE; | |
3868 | hobject_t poid; | |
3869 | eversion_t v; | |
3870 | osd_reqid_t reqid; | |
3871 | MOSDSubOp *subop = new MOSDSubOp( | |
3872 | reqid, pg_whoami, spg_t(info.pgid.pgid, i->shard), poid, 0, | |
3873 | get_osdmap()->get_epoch(), osd->get_tid(), v); | |
3874 | subop->ops = scrub; | |
3875 | osd->send_message_osd_cluster( | |
3876 | i->osd, subop, get_osdmap()->get_epoch()); | |
3877 | } | |
3878 | } | |
3879 | } | |
3880 | ||
3881 | void PG::scrub_unreserve_replicas() | |
3882 | { | |
3883 | assert(backfill_targets.empty()); | |
3884 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
3885 | i != actingbackfill.end(); | |
3886 | ++i) { | |
3887 | if (*i == pg_whoami) continue; | |
3888 | dout(10) << "scrub requesting unreserve from osd." << *i << dendl; | |
3889 | if (HAVE_FEATURE(get_min_acting_features(), SERVER_LUMINOUS)) { | |
3890 | osd->send_message_osd_cluster( | |
3891 | i->osd, | |
3892 | new MOSDScrubReserve(spg_t(info.pgid.pgid, i->shard), | |
3893 | get_osdmap()->get_epoch(), | |
3894 | MOSDScrubReserve::RELEASE, pg_whoami), | |
3895 | get_osdmap()->get_epoch()); | |
3896 | } else { | |
3897 | // for jewel compat only | |
3898 | vector<OSDOp> scrub(1); | |
3899 | scrub[0].op.op = CEPH_OSD_OP_SCRUB_UNRESERVE; | |
3900 | hobject_t poid; | |
3901 | eversion_t v; | |
3902 | osd_reqid_t reqid; | |
3903 | MOSDSubOp *subop = new MOSDSubOp( | |
3904 | reqid, pg_whoami, spg_t(info.pgid.pgid, i->shard), poid, 0, | |
3905 | get_osdmap()->get_epoch(), osd->get_tid(), v); | |
3906 | subop->ops = scrub; | |
3907 | osd->send_message_osd_cluster(i->osd, subop, get_osdmap()->get_epoch()); | |
3908 | } | |
3909 | } | |
3910 | } | |
3911 | ||
3912 | void PG::_scan_rollback_obs( | |
3913 | const vector<ghobject_t> &rollback_obs, | |
3914 | ThreadPool::TPHandle &handle) | |
3915 | { | |
3916 | ObjectStore::Transaction t; | |
3917 | eversion_t trimmed_to = last_rollback_info_trimmed_to_applied; | |
3918 | for (vector<ghobject_t>::const_iterator i = rollback_obs.begin(); | |
3919 | i != rollback_obs.end(); | |
3920 | ++i) { | |
3921 | if (i->generation < trimmed_to.version) { | |
3922 | osd->clog->error() << "osd." << osd->whoami | |
3923 | << " pg " << info.pgid | |
3924 | << " found obsolete rollback obj " | |
3925 | << *i << " generation < trimmed_to " | |
3926 | << trimmed_to | |
3927 | << "...repaired"; | |
3928 | t.remove(coll, *i); | |
3929 | } | |
3930 | } | |
3931 | if (!t.empty()) { | |
3932 | derr << __func__ << ": queueing trans to clean up obsolete rollback objs" | |
3933 | << dendl; | |
3934 | osd->store->queue_transaction(osr.get(), std::move(t), NULL); | |
3935 | } | |
3936 | } | |
3937 | ||
3938 | void PG::_scan_snaps(ScrubMap &smap) | |
3939 | { | |
3940 | hobject_t head; | |
3941 | SnapSet snapset; | |
3942 | for (map<hobject_t, ScrubMap::object>::reverse_iterator i = smap.objects.rbegin(); | |
3943 | i != smap.objects.rend(); | |
3944 | ++i) { | |
3945 | const hobject_t &hoid = i->first; | |
3946 | ScrubMap::object &o = i->second; | |
3947 | ||
3948 | if (hoid.is_head() || hoid.is_snapdir()) { | |
3949 | // parse the SnapSet | |
3950 | bufferlist bl; | |
3951 | if (o.attrs.find(SS_ATTR) == o.attrs.end()) { | |
3952 | continue; | |
3953 | } | |
3954 | bl.push_back(o.attrs[SS_ATTR]); | |
3955 | auto p = bl.begin(); | |
3956 | try { | |
3957 | ::decode(snapset, p); | |
3958 | } catch(...) { | |
3959 | continue; | |
3960 | } | |
3961 | head = hoid.get_head(); | |
3962 | continue; | |
3963 | } | |
3964 | if (hoid.snap < CEPH_MAXSNAP) { | |
3965 | // check and if necessary fix snap_mapper | |
3966 | if (hoid.get_head() != head) { | |
3967 | derr << __func__ << " no head for " << hoid << " (have " << head << ")" | |
3968 | << dendl; | |
3969 | continue; | |
3970 | } | |
3971 | set<snapid_t> obj_snaps; | |
3972 | if (!snapset.is_legacy()) { | |
3973 | auto p = snapset.clone_snaps.find(hoid.snap); | |
3974 | if (p == snapset.clone_snaps.end()) { | |
3975 | derr << __func__ << " no clone_snaps for " << hoid << " in " << snapset | |
3976 | << dendl; | |
3977 | continue; | |
3978 | } | |
3979 | obj_snaps.insert(p->second.begin(), p->second.end()); | |
3980 | } else { | |
3981 | bufferlist bl; | |
3982 | if (o.attrs.find(OI_ATTR) == o.attrs.end()) { | |
3983 | continue; | |
3984 | } | |
3985 | bl.push_back(o.attrs[OI_ATTR]); | |
3986 | object_info_t oi; | |
3987 | try { | |
3988 | oi.decode(bl); | |
3989 | } catch(...) { | |
3990 | continue; | |
3991 | } | |
3992 | obj_snaps.insert(oi.legacy_snaps.begin(), oi.legacy_snaps.end()); | |
3993 | } | |
3994 | set<snapid_t> cur_snaps; | |
3995 | int r = snap_mapper.get_snaps(hoid, &cur_snaps); | |
3996 | if (r != 0 && r != -ENOENT) { | |
3997 | derr << __func__ << ": get_snaps returned " << cpp_strerror(r) << dendl; | |
3998 | ceph_abort(); | |
3999 | } | |
4000 | if (r == -ENOENT || cur_snaps != obj_snaps) { | |
4001 | ObjectStore::Transaction t; | |
4002 | OSDriver::OSTransaction _t(osdriver.get_transaction(&t)); | |
4003 | if (r == 0) { | |
4004 | r = snap_mapper.remove_oid(hoid, &_t); | |
4005 | if (r != 0) { | |
4006 | derr << __func__ << ": remove_oid returned " << cpp_strerror(r) | |
4007 | << dendl; | |
4008 | ceph_abort(); | |
4009 | } | |
4010 | osd->clog->error() << "osd." << osd->whoami | |
4011 | << " found snap mapper error on pg " | |
4012 | << info.pgid | |
4013 | << " oid " << hoid << " snaps in mapper: " | |
4014 | << cur_snaps << ", oi: " | |
4015 | << obj_snaps | |
4016 | << "...repaired"; | |
4017 | } else { | |
4018 | osd->clog->error() << "osd." << osd->whoami | |
4019 | << " found snap mapper error on pg " | |
4020 | << info.pgid | |
4021 | << " oid " << hoid << " snaps missing in mapper" | |
4022 | << ", should be: " | |
4023 | << obj_snaps | |
4024 | << "...repaired"; | |
4025 | } | |
4026 | snap_mapper.add_oid(hoid, obj_snaps, &_t); | |
4027 | r = osd->store->apply_transaction(osr.get(), std::move(t)); | |
4028 | if (r != 0) { | |
4029 | derr << __func__ << ": apply_transaction got " << cpp_strerror(r) | |
4030 | << dendl; | |
4031 | } | |
4032 | } | |
4033 | } | |
4034 | } | |
4035 | } | |
4036 | ||
4037 | /* | |
4038 | * build a scrub map over a chunk without releasing the lock | |
4039 | * only used by chunky scrub | |
4040 | */ | |
4041 | int PG::build_scrub_map_chunk( | |
4042 | ScrubMap &map, | |
4043 | hobject_t start, hobject_t end, bool deep, uint32_t seed, | |
4044 | ThreadPool::TPHandle &handle) | |
4045 | { | |
4046 | dout(10) << __func__ << " [" << start << "," << end << ") " | |
4047 | << " seed " << seed << dendl; | |
4048 | ||
4049 | map.valid_through = info.last_update; | |
4050 | ||
4051 | // objects | |
4052 | vector<hobject_t> ls; | |
4053 | vector<ghobject_t> rollback_obs; | |
4054 | int ret = get_pgbackend()->objects_list_range( | |
4055 | start, | |
4056 | end, | |
4057 | 0, | |
4058 | &ls, | |
4059 | &rollback_obs); | |
4060 | if (ret < 0) { | |
4061 | dout(5) << "objects_list_range error: " << ret << dendl; | |
4062 | return ret; | |
4063 | } | |
4064 | ||
4065 | ||
4066 | get_pgbackend()->be_scan_list(map, ls, deep, seed, handle); | |
4067 | _scan_rollback_obs(rollback_obs, handle); | |
4068 | _scan_snaps(map); | |
4069 | ||
4070 | dout(20) << __func__ << " done" << dendl; | |
4071 | return 0; | |
4072 | } | |
4073 | ||
4074 | void PG::Scrubber::cleanup_store(ObjectStore::Transaction *t) { | |
4075 | if (!store) | |
4076 | return; | |
4077 | struct OnComplete : Context { | |
4078 | std::unique_ptr<Scrub::Store> store; | |
4079 | OnComplete( | |
4080 | std::unique_ptr<Scrub::Store> &&store) | |
4081 | : store(std::move(store)) {} | |
4082 | void finish(int) override {} | |
4083 | }; | |
4084 | store->cleanup(t); | |
4085 | t->register_on_complete(new OnComplete(std::move(store))); | |
4086 | assert(!store); | |
4087 | } | |
4088 | ||
4089 | void PG::repair_object( | |
4090 | const hobject_t& soid, list<pair<ScrubMap::object, pg_shard_t> > *ok_peers, | |
4091 | pg_shard_t bad_peer) | |
4092 | { | |
4093 | list<pg_shard_t> op_shards; | |
4094 | for (auto i : *ok_peers) { | |
4095 | op_shards.push_back(i.second); | |
4096 | } | |
4097 | dout(10) << "repair_object " << soid << " bad_peer osd." | |
4098 | << bad_peer << " ok_peers osd.{" << op_shards << "}" << dendl; | |
4099 | ScrubMap::object &po = ok_peers->back().first; | |
4100 | eversion_t v; | |
4101 | bufferlist bv; | |
4102 | bv.push_back(po.attrs[OI_ATTR]); | |
4103 | object_info_t oi(bv); | |
4104 | if (bad_peer != primary) { | |
4105 | peer_missing[bad_peer].add(soid, oi.version, eversion_t()); | |
4106 | } else { | |
4107 | // We should only be scrubbing if the PG is clean. | |
4108 | assert(waiting_for_unreadable_object.empty()); | |
4109 | ||
4110 | pg_log.missing_add(soid, oi.version, eversion_t()); | |
4111 | ||
4112 | pg_log.set_last_requested(0); | |
4113 | dout(10) << __func__ << ": primary = " << primary << dendl; | |
4114 | } | |
4115 | ||
4116 | if (is_ec_pg() || bad_peer == primary) { | |
4117 | // we'd better collect all shard for EC pg, and prepare good peers as the | |
4118 | // source of pull in the case of replicated pg. | |
4119 | missing_loc.add_missing(soid, oi.version, eversion_t()); | |
4120 | list<pair<ScrubMap::object, pg_shard_t> >::iterator i; | |
4121 | for (i = ok_peers->begin(); | |
4122 | i != ok_peers->end(); | |
4123 | ++i) | |
4124 | missing_loc.add_location(soid, i->second); | |
4125 | } | |
4126 | } | |
4127 | ||
4128 | /* replica_scrub | |
4129 | * | |
4130 | * Wait for last_update_applied to match msg->scrub_to as above. Wait | |
4131 | * for pushes to complete in case of recent recovery. Build a single | |
4132 | * scrubmap of objects that are in the range [msg->start, msg->end). | |
4133 | */ | |
4134 | void PG::replica_scrub( | |
4135 | OpRequestRef op, | |
4136 | ThreadPool::TPHandle &handle) | |
4137 | { | |
4138 | const MOSDRepScrub *msg = static_cast<const MOSDRepScrub *>(op->get_req()); | |
4139 | assert(!scrubber.active_rep_scrub); | |
4140 | dout(7) << "replica_scrub" << dendl; | |
4141 | ||
4142 | if (msg->map_epoch < info.history.same_interval_since) { | |
4143 | dout(10) << "replica_scrub discarding old replica_scrub from " | |
4144 | << msg->map_epoch << " < " << info.history.same_interval_since | |
4145 | << dendl; | |
4146 | return; | |
4147 | } | |
4148 | ||
4149 | ScrubMap map; | |
4150 | ||
4151 | assert(msg->chunky); | |
4152 | if (last_update_applied < msg->scrub_to) { | |
4153 | dout(10) << "waiting for last_update_applied to catch up" << dendl; | |
4154 | scrubber.active_rep_scrub = op; | |
4155 | return; | |
4156 | } | |
4157 | ||
4158 | if (active_pushes > 0) { | |
4159 | dout(10) << "waiting for active pushes to finish" << dendl; | |
4160 | scrubber.active_rep_scrub = op; | |
4161 | return; | |
4162 | } | |
4163 | ||
4164 | // compensate for hobject_t's with wrong pool from sloppy hammer OSDs | |
4165 | hobject_t start = msg->start; | |
4166 | hobject_t end = msg->end; | |
4167 | if (!start.is_max()) | |
4168 | start.pool = info.pgid.pool(); | |
4169 | if (!end.is_max()) | |
4170 | end.pool = info.pgid.pool(); | |
4171 | ||
4172 | build_scrub_map_chunk( | |
4173 | map, start, end, msg->deep, msg->seed, | |
4174 | handle); | |
4175 | ||
4176 | if (HAVE_FEATURE(acting_features, SERVER_LUMINOUS)) { | |
4177 | MOSDRepScrubMap *reply = new MOSDRepScrubMap( | |
4178 | spg_t(info.pgid.pgid, get_primary().shard), | |
4179 | msg->map_epoch, | |
4180 | pg_whoami); | |
4181 | ::encode(map, reply->get_data()); | |
4182 | osd->send_message_osd_cluster(reply, msg->get_connection()); | |
4183 | } else { | |
4184 | // for jewel compatibility | |
4185 | vector<OSDOp> scrub(1); | |
4186 | scrub[0].op.op = CEPH_OSD_OP_SCRUB_MAP; | |
4187 | hobject_t poid; | |
4188 | eversion_t v; | |
4189 | osd_reqid_t reqid; | |
4190 | MOSDSubOp *subop = new MOSDSubOp( | |
4191 | reqid, | |
4192 | pg_whoami, | |
4193 | spg_t(info.pgid.pgid, get_primary().shard), | |
4194 | poid, | |
4195 | 0, | |
4196 | msg->map_epoch, | |
4197 | osd->get_tid(), | |
4198 | v); | |
4199 | ::encode(map, subop->get_data()); | |
4200 | subop->ops = scrub; | |
4201 | osd->send_message_osd_cluster(subop, msg->get_connection()); | |
4202 | } | |
4203 | } | |
4204 | ||
4205 | /* Scrub: | |
4206 | * PG_STATE_SCRUBBING is set when the scrub is queued | |
4207 | * | |
4208 | * scrub will be chunky if all OSDs in PG support chunky scrub | |
4209 | * scrub will fail if OSDs are too old. | |
4210 | */ | |
4211 | void PG::scrub(epoch_t queued, ThreadPool::TPHandle &handle) | |
4212 | { | |
4213 | if (cct->_conf->osd_scrub_sleep > 0 && | |
4214 | (scrubber.state == PG::Scrubber::NEW_CHUNK || | |
4215 | scrubber.state == PG::Scrubber::INACTIVE) && | |
4216 | scrubber.needs_sleep) { | |
4217 | ceph_assert(!scrubber.sleeping); | |
4218 | dout(20) << __func__ << " state is INACTIVE|NEW_CHUNK, sleeping" << dendl; | |
4219 | ||
4220 | // Do an async sleep so we don't block the op queue | |
4221 | OSDService *osds = osd; | |
4222 | spg_t pgid = get_pgid(); | |
4223 | int state = scrubber.state; | |
4224 | auto scrub_requeue_callback = | |
4225 | new FunctionContext([osds, pgid, state](int r) { | |
4226 | PG *pg = osds->osd->lookup_lock_pg(pgid); | |
4227 | if (pg == nullptr) { | |
4228 | lgeneric_dout(osds->osd->cct, 20) | |
4229 | << "scrub_requeue_callback: Could not find " | |
4230 | << "PG " << pgid << " can't complete scrub requeue after sleep" | |
4231 | << dendl; | |
4232 | return; | |
4233 | } | |
4234 | pg->scrubber.sleeping = false; | |
4235 | pg->scrubber.needs_sleep = false; | |
4236 | lgeneric_dout(pg->cct, 20) | |
4237 | << "scrub_requeue_callback: slept for " | |
4238 | << ceph_clock_now() - pg->scrubber.sleep_start | |
4239 | << ", re-queuing scrub with state " << state << dendl; | |
4240 | pg->scrub_queued = false; | |
4241 | pg->requeue_scrub(); | |
4242 | pg->scrubber.sleep_start = utime_t(); | |
4243 | pg->unlock(); | |
4244 | }); | |
4245 | Mutex::Locker l(osd->scrub_sleep_lock); | |
4246 | osd->scrub_sleep_timer.add_event_after(cct->_conf->osd_scrub_sleep, | |
4247 | scrub_requeue_callback); | |
4248 | scrubber.sleeping = true; | |
4249 | scrubber.sleep_start = ceph_clock_now(); | |
4250 | return; | |
4251 | } | |
4252 | if (pg_has_reset_since(queued)) { | |
4253 | return; | |
4254 | } | |
4255 | assert(scrub_queued); | |
4256 | scrub_queued = false; | |
4257 | scrubber.needs_sleep = true; | |
4258 | ||
4259 | if (!is_primary() || !is_active() || !is_clean() || !is_scrubbing()) { | |
4260 | dout(10) << "scrub -- not primary or active or not clean" << dendl; | |
4261 | state_clear(PG_STATE_SCRUBBING); | |
4262 | state_clear(PG_STATE_REPAIR); | |
4263 | state_clear(PG_STATE_DEEP_SCRUB); | |
4264 | publish_stats_to_osd(); | |
4265 | return; | |
4266 | } | |
4267 | ||
4268 | if (!scrubber.active) { | |
4269 | assert(backfill_targets.empty()); | |
4270 | ||
4271 | scrubber.deep = state_test(PG_STATE_DEEP_SCRUB); | |
4272 | ||
4273 | dout(10) << "starting a new chunky scrub" << dendl; | |
4274 | } | |
4275 | ||
4276 | chunky_scrub(handle); | |
4277 | } | |
4278 | ||
4279 | /* | |
4280 | * Chunky scrub scrubs objects one chunk at a time with writes blocked for that | |
4281 | * chunk. | |
4282 | * | |
4283 | * The object store is partitioned into chunks which end on hash boundaries. For | |
4284 | * each chunk, the following logic is performed: | |
4285 | * | |
4286 | * (1) Block writes on the chunk | |
4287 | * (2) Request maps from replicas | |
4288 | * (3) Wait for pushes to be applied (after recovery) | |
4289 | * (4) Wait for writes to flush on the chunk | |
4290 | * (5) Wait for maps from replicas | |
4291 | * (6) Compare / repair all scrub maps | |
4292 | * (7) Wait for digest updates to apply | |
4293 | * | |
4294 | * This logic is encoded in the mostly linear state machine: | |
4295 | * | |
4296 | * +------------------+ | |
4297 | * _________v__________ | | |
4298 | * | | | | |
4299 | * | INACTIVE | | | |
4300 | * |____________________| | | |
4301 | * | | | |
4302 | * | +----------+ | | |
4303 | * _________v___v______ | | | |
4304 | * | | | | | |
4305 | * | NEW_CHUNK | | | | |
4306 | * |____________________| | | | |
4307 | * | | | | |
4308 | * _________v__________ | | | |
4309 | * | | | | | |
4310 | * | WAIT_PUSHES | | | | |
4311 | * |____________________| | | | |
4312 | * | | | | |
4313 | * _________v__________ | | | |
4314 | * | | | | | |
4315 | * | WAIT_LAST_UPDATE | | | | |
4316 | * |____________________| | | | |
4317 | * | | | | |
4318 | * _________v__________ | | | |
4319 | * | | | | | |
4320 | * | BUILD_MAP | | | | |
4321 | * |____________________| | | | |
4322 | * | | | | |
4323 | * _________v__________ | | | |
4324 | * | | | | | |
4325 | * | WAIT_REPLICAS | | | | |
4326 | * |____________________| | | | |
4327 | * | | | | |
4328 | * _________v__________ | | | |
4329 | * | | | | | |
4330 | * | COMPARE_MAPS | | | | |
4331 | * |____________________| | | | |
4332 | * | | | | |
4333 | * | | | | |
4334 | * _________v__________ | | | |
4335 | * | | | | | |
4336 | * |WAIT_DIGEST_UPDATES | | | | |
4337 | * |____________________| | | | |
4338 | * | | | | | |
4339 | * | +----------+ | | |
4340 | * _________v__________ | | |
4341 | * | | | | |
4342 | * | FINISH | | | |
4343 | * |____________________| | | |
4344 | * | | | |
4345 | * +------------------+ | |
4346 | * | |
4347 | * The primary determines the last update from the subset by walking the log. If | |
4348 | * it sees a log entry pertaining to a file in the chunk, it tells the replicas | |
4349 | * to wait until that update is applied before building a scrub map. Both the | |
4350 | * primary and replicas will wait for any active pushes to be applied. | |
4351 | * | |
4352 | * In contrast to classic_scrub, chunky_scrub is entirely handled by scrub_wq. | |
4353 | * | |
4354 | * scrubber.state encodes the current state of the scrub (refer to state diagram | |
4355 | * for details). | |
4356 | */ | |
4357 | void PG::chunky_scrub(ThreadPool::TPHandle &handle) | |
4358 | { | |
4359 | // check for map changes | |
4360 | if (scrubber.is_chunky_scrub_active()) { | |
4361 | if (scrubber.epoch_start != info.history.same_interval_since) { | |
4362 | dout(10) << "scrub pg changed, aborting" << dendl; | |
4363 | scrub_clear_state(); | |
4364 | scrub_unreserve_replicas(); | |
4365 | return; | |
4366 | } | |
4367 | } | |
4368 | ||
4369 | bool done = false; | |
4370 | int ret; | |
4371 | ||
4372 | while (!done) { | |
4373 | dout(20) << "scrub state " << Scrubber::state_string(scrubber.state) | |
4374 | << " [" << scrubber.start << "," << scrubber.end << ")" << dendl; | |
4375 | ||
4376 | switch (scrubber.state) { | |
4377 | case PG::Scrubber::INACTIVE: | |
4378 | dout(10) << "scrub start" << dendl; | |
4379 | ||
4380 | publish_stats_to_osd(); | |
4381 | scrubber.epoch_start = info.history.same_interval_since; | |
4382 | scrubber.active = true; | |
4383 | ||
4384 | osd->inc_scrubs_active(scrubber.reserved); | |
4385 | if (scrubber.reserved) { | |
4386 | scrubber.reserved = false; | |
4387 | scrubber.reserved_peers.clear(); | |
4388 | } | |
4389 | ||
4390 | { | |
4391 | ObjectStore::Transaction t; | |
4392 | scrubber.cleanup_store(&t); | |
4393 | scrubber.store.reset(Scrub::Store::create(osd->store, &t, | |
4394 | info.pgid, coll)); | |
4395 | osd->store->queue_transaction(osr.get(), std::move(t), nullptr); | |
4396 | } | |
4397 | ||
4398 | // Don't include temporary objects when scrubbing | |
4399 | scrubber.start = info.pgid.pgid.get_hobj_start(); | |
4400 | scrubber.state = PG::Scrubber::NEW_CHUNK; | |
4401 | ||
4402 | { | |
4403 | bool repair = state_test(PG_STATE_REPAIR); | |
4404 | bool deep_scrub = state_test(PG_STATE_DEEP_SCRUB); | |
4405 | const char *mode = (repair ? "repair": (deep_scrub ? "deep-scrub" : "scrub")); | |
4406 | stringstream oss; | |
4407 | oss << info.pgid.pgid << " " << mode << " starts" << std::endl; | |
4408 | osd->clog->info(oss); | |
4409 | } | |
4410 | ||
4411 | scrubber.seed = -1; | |
4412 | ||
4413 | break; | |
4414 | ||
4415 | case PG::Scrubber::NEW_CHUNK: | |
4416 | scrubber.primary_scrubmap = ScrubMap(); | |
4417 | scrubber.received_maps.clear(); | |
4418 | ||
4419 | { | |
4420 | /* get the start and end of our scrub chunk | |
4421 | * | |
4422 | * Our scrub chunk has an important restriction we're going to need to | |
4423 | * respect. We can't let head or snapdir be start or end. | |
4424 | * Using a half-open interval means that if end == head|snapdir, | |
4425 | * we'd scrub/lock head and the clone right next to head in different | |
4426 | * chunks which would allow us to miss clones created between | |
4427 | * scrubbing that chunk and scrubbing the chunk including head. | |
4428 | * This isn't true for any of the other clones since clones can | |
4429 | * only be created "just to the left of" head. There is one exception | |
4430 | * to this: promotion of clones which always happens to the left of the | |
4431 | * left-most clone, but promote_object checks the scrubber in that | |
4432 | * case, so it should be ok. Also, it's ok to "miss" clones at the | |
4433 | * left end of the range if we are a tier because they may legitimately | |
4434 | * not exist (see _scrub). | |
4435 | */ | |
4436 | int min = MAX(3, cct->_conf->osd_scrub_chunk_min); | |
4437 | hobject_t start = scrubber.start; | |
4438 | hobject_t candidate_end; | |
4439 | vector<hobject_t> objects; | |
4440 | ret = get_pgbackend()->objects_list_partial( | |
4441 | start, | |
4442 | min, | |
4443 | MAX(min, cct->_conf->osd_scrub_chunk_max), | |
4444 | &objects, | |
4445 | &candidate_end); | |
4446 | assert(ret >= 0); | |
4447 | ||
4448 | if (!objects.empty()) { | |
4449 | hobject_t back = objects.back(); | |
4450 | while (candidate_end.has_snapset() && | |
4451 | candidate_end.get_head() == back.get_head()) { | |
4452 | candidate_end = back; | |
4453 | objects.pop_back(); | |
4454 | if (objects.empty()) { | |
4455 | assert(0 == | |
4456 | "Somehow we got more than 2 objects which" | |
4457 | "have the same head but are not clones"); | |
4458 | } | |
4459 | back = objects.back(); | |
4460 | } | |
4461 | if (candidate_end.has_snapset()) { | |
4462 | assert(candidate_end.get_head() != back.get_head()); | |
4463 | candidate_end = candidate_end.get_object_boundary(); | |
4464 | } | |
4465 | } else { | |
4466 | assert(candidate_end.is_max()); | |
4467 | } | |
4468 | ||
4469 | if (!_range_available_for_scrub(scrubber.start, candidate_end)) { | |
4470 | // we'll be requeued by whatever made us unavailable for scrub | |
4471 | dout(10) << __func__ << ": scrub blocked somewhere in range " | |
4472 | << "[" << scrubber.start << ", " << candidate_end << ")" | |
4473 | << dendl; | |
4474 | done = true; | |
4475 | break; | |
4476 | } | |
4477 | scrubber.end = candidate_end; | |
4478 | } | |
4479 | ||
4480 | // walk the log to find the latest update that affects our chunk | |
4481 | scrubber.subset_last_update = eversion_t(); | |
4482 | for (auto p = projected_log.log.rbegin(); | |
4483 | p != projected_log.log.rend(); | |
4484 | ++p) { | |
4485 | if (p->soid >= scrubber.start && | |
4486 | p->soid < scrubber.end) { | |
4487 | scrubber.subset_last_update = p->version; | |
4488 | break; | |
4489 | } | |
4490 | } | |
4491 | if (scrubber.subset_last_update == eversion_t()) { | |
4492 | for (list<pg_log_entry_t>::const_reverse_iterator p = | |
4493 | pg_log.get_log().log.rbegin(); | |
4494 | p != pg_log.get_log().log.rend(); | |
4495 | ++p) { | |
4496 | if (p->soid >= scrubber.start && | |
4497 | p->soid < scrubber.end) { | |
4498 | scrubber.subset_last_update = p->version; | |
4499 | break; | |
4500 | } | |
4501 | } | |
4502 | } | |
4503 | ||
4504 | // ask replicas to wait until | |
4505 | // last_update_applied >= scrubber.subset_last_update and then scan | |
4506 | scrubber.waiting_on_whom.insert(pg_whoami); | |
4507 | ++scrubber.waiting_on; | |
4508 | ||
4509 | // request maps from replicas | |
4510 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
4511 | i != actingbackfill.end(); | |
4512 | ++i) { | |
4513 | if (*i == pg_whoami) continue; | |
4514 | _request_scrub_map(*i, scrubber.subset_last_update, | |
4515 | scrubber.start, scrubber.end, scrubber.deep, | |
4516 | scrubber.seed); | |
4517 | scrubber.waiting_on_whom.insert(*i); | |
4518 | ++scrubber.waiting_on; | |
4519 | } | |
4520 | ||
4521 | scrubber.state = PG::Scrubber::WAIT_PUSHES; | |
4522 | ||
4523 | break; | |
4524 | ||
4525 | case PG::Scrubber::WAIT_PUSHES: | |
4526 | if (active_pushes == 0) { | |
4527 | scrubber.state = PG::Scrubber::WAIT_LAST_UPDATE; | |
4528 | } else { | |
4529 | dout(15) << "wait for pushes to apply" << dendl; | |
4530 | done = true; | |
4531 | } | |
4532 | break; | |
4533 | ||
4534 | case PG::Scrubber::WAIT_LAST_UPDATE: | |
4535 | if (last_update_applied >= scrubber.subset_last_update) { | |
4536 | scrubber.state = PG::Scrubber::BUILD_MAP; | |
4537 | } else { | |
4538 | // will be requeued by op_applied | |
4539 | dout(15) << "wait for writes to flush" << dendl; | |
4540 | done = true; | |
4541 | } | |
4542 | break; | |
4543 | ||
4544 | case PG::Scrubber::BUILD_MAP: | |
4545 | assert(last_update_applied >= scrubber.subset_last_update); | |
4546 | ||
4547 | // build my own scrub map | |
4548 | ret = build_scrub_map_chunk(scrubber.primary_scrubmap, | |
4549 | scrubber.start, scrubber.end, | |
4550 | scrubber.deep, scrubber.seed, | |
4551 | handle); | |
4552 | if (ret < 0) { | |
4553 | dout(5) << "error building scrub map: " << ret << ", aborting" << dendl; | |
4554 | scrub_clear_state(); | |
4555 | scrub_unreserve_replicas(); | |
4556 | return; | |
4557 | } | |
4558 | ||
4559 | --scrubber.waiting_on; | |
4560 | scrubber.waiting_on_whom.erase(pg_whoami); | |
4561 | ||
4562 | scrubber.state = PG::Scrubber::WAIT_REPLICAS; | |
4563 | break; | |
4564 | ||
4565 | case PG::Scrubber::WAIT_REPLICAS: | |
4566 | if (scrubber.waiting_on > 0) { | |
4567 | // will be requeued by sub_op_scrub_map | |
4568 | dout(10) << "wait for replicas to build scrub map" << dendl; | |
4569 | done = true; | |
4570 | } else { | |
4571 | scrubber.state = PG::Scrubber::COMPARE_MAPS; | |
4572 | } | |
4573 | break; | |
4574 | ||
4575 | case PG::Scrubber::COMPARE_MAPS: | |
4576 | assert(last_update_applied >= scrubber.subset_last_update); | |
4577 | assert(scrubber.waiting_on == 0); | |
4578 | ||
4579 | scrub_compare_maps(); | |
4580 | scrubber.start = scrubber.end; | |
4581 | scrubber.run_callbacks(); | |
4582 | ||
4583 | // requeue the writes from the chunk that just finished | |
4584 | requeue_ops(waiting_for_scrub); | |
4585 | ||
4586 | scrubber.state = PG::Scrubber::WAIT_DIGEST_UPDATES; | |
4587 | ||
4588 | // fall-thru | |
4589 | ||
4590 | case PG::Scrubber::WAIT_DIGEST_UPDATES: | |
4591 | if (scrubber.num_digest_updates_pending) { | |
4592 | dout(10) << __func__ << " waiting on " | |
4593 | << scrubber.num_digest_updates_pending | |
4594 | << " digest updates" << dendl; | |
4595 | done = true; | |
4596 | break; | |
4597 | } | |
4598 | ||
4599 | if (!(scrubber.end.is_max())) { | |
4600 | scrubber.state = PG::Scrubber::NEW_CHUNK; | |
4601 | requeue_scrub(); | |
4602 | done = true; | |
4603 | } else { | |
4604 | scrubber.state = PG::Scrubber::FINISH; | |
4605 | } | |
4606 | ||
4607 | break; | |
4608 | ||
4609 | case PG::Scrubber::FINISH: | |
4610 | scrub_finish(); | |
4611 | scrubber.state = PG::Scrubber::INACTIVE; | |
4612 | done = true; | |
4613 | ||
4614 | break; | |
4615 | ||
4616 | default: | |
4617 | ceph_abort(); | |
4618 | } | |
4619 | } | |
4620 | dout(20) << "scrub final state " << Scrubber::state_string(scrubber.state) | |
4621 | << " [" << scrubber.start << "," << scrubber.end << ")" << dendl; | |
4622 | } | |
4623 | ||
4624 | void PG::scrub_clear_state() | |
4625 | { | |
4626 | assert(is_locked()); | |
4627 | state_clear(PG_STATE_SCRUBBING); | |
4628 | state_clear(PG_STATE_REPAIR); | |
4629 | state_clear(PG_STATE_DEEP_SCRUB); | |
4630 | publish_stats_to_osd(); | |
4631 | ||
4632 | // active -> nothing. | |
4633 | if (scrubber.active) | |
4634 | osd->dec_scrubs_active(); | |
4635 | ||
4636 | requeue_ops(waiting_for_scrub); | |
4637 | ||
4638 | if (scrubber.queue_snap_trim) { | |
4639 | dout(10) << "scrub finished, requeuing snap_trimmer" << dendl; | |
4640 | snap_trimmer_scrub_complete(); | |
4641 | } | |
4642 | ||
4643 | scrubber.reset(); | |
4644 | ||
4645 | // type-specific state clear | |
4646 | _scrub_clear_state(); | |
4647 | } | |
4648 | ||
4649 | void PG::scrub_compare_maps() | |
4650 | { | |
4651 | dout(10) << __func__ << " has maps, analyzing" << dendl; | |
4652 | ||
4653 | // construct authoritative scrub map for type specific scrubbing | |
4654 | scrubber.cleaned_meta_map.insert(scrubber.primary_scrubmap); | |
4655 | map<hobject_t, pair<uint32_t, uint32_t>> missing_digest; | |
4656 | ||
4657 | if (acting.size() > 1) { | |
4658 | dout(10) << __func__ << " comparing replica scrub maps" << dendl; | |
4659 | ||
4660 | stringstream ss; | |
4661 | ||
4662 | // Map from object with errors to good peer | |
4663 | map<hobject_t, list<pg_shard_t>> authoritative; | |
4664 | map<pg_shard_t, ScrubMap *> maps; | |
4665 | ||
4666 | dout(2) << __func__ << " osd." << acting[0] << " has " | |
4667 | << scrubber.primary_scrubmap.objects.size() << " items" << dendl; | |
4668 | maps[pg_whoami] = &scrubber.primary_scrubmap; | |
4669 | ||
4670 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
4671 | i != actingbackfill.end(); | |
4672 | ++i) { | |
4673 | if (*i == pg_whoami) continue; | |
4674 | dout(2) << __func__ << " replica " << *i << " has " | |
4675 | << scrubber.received_maps[*i].objects.size() | |
4676 | << " items" << dendl; | |
4677 | maps[*i] = &scrubber.received_maps[*i]; | |
4678 | } | |
4679 | ||
4680 | get_pgbackend()->be_compare_scrubmaps( | |
4681 | maps, | |
4682 | state_test(PG_STATE_REPAIR), | |
4683 | scrubber.missing, | |
4684 | scrubber.inconsistent, | |
4685 | authoritative, | |
4686 | missing_digest, | |
4687 | scrubber.shallow_errors, | |
4688 | scrubber.deep_errors, | |
4689 | scrubber.store.get(), | |
4690 | info.pgid, acting, | |
4691 | ss); | |
4692 | dout(2) << ss.str() << dendl; | |
4693 | ||
4694 | if (!ss.str().empty()) { | |
4695 | osd->clog->error(ss); | |
4696 | } | |
4697 | ||
4698 | for (map<hobject_t, list<pg_shard_t>>::iterator i = authoritative.begin(); | |
4699 | i != authoritative.end(); | |
4700 | ++i) { | |
4701 | list<pair<ScrubMap::object, pg_shard_t> > good_peers; | |
4702 | for (list<pg_shard_t>::const_iterator j = i->second.begin(); | |
4703 | j != i->second.end(); | |
4704 | ++j) { | |
4705 | good_peers.push_back(make_pair(maps[*j]->objects[i->first], *j)); | |
4706 | } | |
4707 | scrubber.authoritative.insert( | |
4708 | make_pair( | |
4709 | i->first, | |
4710 | good_peers)); | |
4711 | } | |
4712 | ||
4713 | for (map<hobject_t, list<pg_shard_t>>::iterator i = authoritative.begin(); | |
4714 | i != authoritative.end(); | |
4715 | ++i) { | |
4716 | scrubber.cleaned_meta_map.objects.erase(i->first); | |
4717 | scrubber.cleaned_meta_map.objects.insert( | |
4718 | *(maps[i->second.back()]->objects.find(i->first)) | |
4719 | ); | |
4720 | } | |
4721 | } | |
4722 | ||
4723 | ScrubMap for_meta_scrub; | |
4724 | if (scrubber.end.is_max() || | |
4725 | scrubber.cleaned_meta_map.objects.empty()) { | |
4726 | scrubber.cleaned_meta_map.swap(for_meta_scrub); | |
4727 | } else { | |
4728 | auto iter = scrubber.cleaned_meta_map.objects.end(); | |
4729 | --iter; // not empty, see if clause | |
4730 | auto begin = scrubber.cleaned_meta_map.objects.begin(); | |
4731 | while (iter != begin) { | |
4732 | auto next = iter--; | |
4733 | if (next->first.get_head() != iter->first.get_head()) { | |
4734 | ++iter; | |
4735 | break; | |
4736 | } | |
4737 | } | |
4738 | for_meta_scrub.objects.insert(begin, iter); | |
4739 | scrubber.cleaned_meta_map.objects.erase(begin, iter); | |
4740 | } | |
4741 | ||
4742 | // ok, do the pg-type specific scrubbing | |
4743 | scrub_snapshot_metadata(for_meta_scrub, missing_digest); | |
4744 | if (!scrubber.store->empty()) { | |
4745 | if (state_test(PG_STATE_REPAIR)) { | |
4746 | dout(10) << __func__ << ": discarding scrub results" << dendl; | |
4747 | scrubber.store->flush(nullptr); | |
4748 | } else { | |
4749 | dout(10) << __func__ << ": updating scrub object" << dendl; | |
4750 | ObjectStore::Transaction t; | |
4751 | scrubber.store->flush(&t); | |
4752 | osd->store->queue_transaction(osr.get(), std::move(t), nullptr); | |
4753 | } | |
4754 | } | |
4755 | } | |
4756 | ||
4757 | bool PG::scrub_process_inconsistent() | |
4758 | { | |
4759 | dout(10) << __func__ << ": checking authoritative" << dendl; | |
4760 | bool repair = state_test(PG_STATE_REPAIR); | |
4761 | bool deep_scrub = state_test(PG_STATE_DEEP_SCRUB); | |
4762 | const char *mode = (repair ? "repair": (deep_scrub ? "deep-scrub" : "scrub")); | |
4763 | ||
4764 | // authoriative only store objects which missing or inconsistent. | |
4765 | if (!scrubber.authoritative.empty()) { | |
4766 | stringstream ss; | |
4767 | ss << info.pgid << " " << mode << " " | |
4768 | << scrubber.missing.size() << " missing, " | |
4769 | << scrubber.inconsistent.size() << " inconsistent objects"; | |
4770 | dout(2) << ss.str() << dendl; | |
4771 | osd->clog->error(ss); | |
4772 | if (repair) { | |
4773 | state_clear(PG_STATE_CLEAN); | |
4774 | for (map<hobject_t, list<pair<ScrubMap::object, pg_shard_t> >>::iterator i = | |
4775 | scrubber.authoritative.begin(); | |
4776 | i != scrubber.authoritative.end(); | |
4777 | ++i) { | |
4778 | set<pg_shard_t>::iterator j; | |
4779 | ||
4780 | auto missing_entry = scrubber.missing.find(i->first); | |
4781 | if (missing_entry != scrubber.missing.end()) { | |
4782 | for (j = missing_entry->second.begin(); | |
4783 | j != missing_entry->second.end(); | |
4784 | ++j) { | |
4785 | repair_object( | |
4786 | i->first, | |
4787 | &(i->second), | |
4788 | *j); | |
4789 | ++scrubber.fixed; | |
4790 | } | |
4791 | } | |
4792 | if (scrubber.inconsistent.count(i->first)) { | |
4793 | for (j = scrubber.inconsistent[i->first].begin(); | |
4794 | j != scrubber.inconsistent[i->first].end(); | |
4795 | ++j) { | |
4796 | repair_object(i->first, | |
4797 | &(i->second), | |
4798 | *j); | |
4799 | ++scrubber.fixed; | |
4800 | } | |
4801 | } | |
4802 | } | |
4803 | } | |
4804 | } | |
4805 | return (!scrubber.authoritative.empty() && repair); | |
4806 | } | |
4807 | ||
4808 | bool PG::ops_blocked_by_scrub() const { | |
4809 | return (waiting_for_scrub.size() != 0); | |
4810 | } | |
4811 | ||
4812 | // the part that actually finalizes a scrub | |
4813 | void PG::scrub_finish() | |
4814 | { | |
4815 | bool repair = state_test(PG_STATE_REPAIR); | |
4816 | // if the repair request comes from auto-repair and large number of errors, | |
4817 | // we would like to cancel auto-repair | |
4818 | if (repair && scrubber.auto_repair | |
4819 | && scrubber.authoritative.size() > cct->_conf->osd_scrub_auto_repair_num_errors) { | |
4820 | state_clear(PG_STATE_REPAIR); | |
4821 | repair = false; | |
4822 | } | |
4823 | bool deep_scrub = state_test(PG_STATE_DEEP_SCRUB); | |
4824 | const char *mode = (repair ? "repair": (deep_scrub ? "deep-scrub" : "scrub")); | |
4825 | ||
4826 | // type-specific finish (can tally more errors) | |
4827 | _scrub_finish(); | |
4828 | ||
4829 | bool has_error = scrub_process_inconsistent(); | |
4830 | ||
4831 | { | |
4832 | stringstream oss; | |
4833 | oss << info.pgid.pgid << " " << mode << " "; | |
4834 | int total_errors = scrubber.shallow_errors + scrubber.deep_errors; | |
4835 | if (total_errors) | |
4836 | oss << total_errors << " errors"; | |
4837 | else | |
4838 | oss << "ok"; | |
4839 | if (!deep_scrub && info.stats.stats.sum.num_deep_scrub_errors) | |
4840 | oss << " ( " << info.stats.stats.sum.num_deep_scrub_errors | |
4841 | << " remaining deep scrub error details lost)"; | |
4842 | if (repair) | |
4843 | oss << ", " << scrubber.fixed << " fixed"; | |
4844 | if (total_errors) | |
4845 | osd->clog->error(oss); | |
4846 | else | |
4847 | osd->clog->info(oss); | |
4848 | } | |
4849 | ||
4850 | // finish up | |
4851 | unreg_next_scrub(); | |
4852 | utime_t now = ceph_clock_now(); | |
4853 | info.history.last_scrub = info.last_update; | |
4854 | info.history.last_scrub_stamp = now; | |
4855 | if (scrubber.deep) { | |
4856 | info.history.last_deep_scrub = info.last_update; | |
4857 | info.history.last_deep_scrub_stamp = now; | |
4858 | } | |
4859 | // Since we don't know which errors were fixed, we can only clear them | |
4860 | // when every one has been fixed. | |
4861 | if (repair) { | |
4862 | if (scrubber.fixed == scrubber.shallow_errors + scrubber.deep_errors) { | |
4863 | assert(deep_scrub); | |
4864 | scrubber.shallow_errors = scrubber.deep_errors = 0; | |
4865 | } else { | |
4866 | // Deep scrub in order to get corrected error counts | |
4867 | scrub_after_recovery = true; | |
4868 | } | |
4869 | } | |
4870 | if (deep_scrub) { | |
4871 | if ((scrubber.shallow_errors == 0) && (scrubber.deep_errors == 0)) | |
4872 | info.history.last_clean_scrub_stamp = now; | |
4873 | info.stats.stats.sum.num_shallow_scrub_errors = scrubber.shallow_errors; | |
4874 | info.stats.stats.sum.num_deep_scrub_errors = scrubber.deep_errors; | |
4875 | } else { | |
4876 | info.stats.stats.sum.num_shallow_scrub_errors = scrubber.shallow_errors; | |
4877 | // XXX: last_clean_scrub_stamp doesn't mean the pg is not inconsistent | |
4878 | // because of deep-scrub errors | |
4879 | if (scrubber.shallow_errors == 0) | |
4880 | info.history.last_clean_scrub_stamp = now; | |
4881 | } | |
4882 | info.stats.stats.sum.num_scrub_errors = | |
4883 | info.stats.stats.sum.num_shallow_scrub_errors + | |
4884 | info.stats.stats.sum.num_deep_scrub_errors; | |
4885 | reg_next_scrub(); | |
4886 | ||
4887 | { | |
4888 | ObjectStore::Transaction t; | |
4889 | dirty_info = true; | |
4890 | write_if_dirty(t); | |
4891 | int tr = osd->store->queue_transaction(osr.get(), std::move(t), NULL); | |
4892 | assert(tr == 0); | |
4893 | } | |
4894 | ||
4895 | ||
4896 | if (has_error) { | |
4897 | queue_peering_event( | |
4898 | CephPeeringEvtRef( | |
4899 | std::make_shared<CephPeeringEvt>( | |
4900 | get_osdmap()->get_epoch(), | |
4901 | get_osdmap()->get_epoch(), | |
4902 | DoRecovery()))); | |
4903 | } | |
4904 | ||
4905 | scrub_clear_state(); | |
4906 | scrub_unreserve_replicas(); | |
4907 | ||
4908 | if (is_active() && is_primary()) { | |
4909 | share_pg_info(); | |
4910 | } | |
4911 | } | |
4912 | ||
4913 | void PG::share_pg_info() | |
4914 | { | |
4915 | dout(10) << "share_pg_info" << dendl; | |
4916 | ||
4917 | // share new pg_info_t with replicas | |
4918 | assert(!actingbackfill.empty()); | |
4919 | for (set<pg_shard_t>::iterator i = actingbackfill.begin(); | |
4920 | i != actingbackfill.end(); | |
4921 | ++i) { | |
4922 | if (*i == pg_whoami) continue; | |
4923 | pg_shard_t peer = *i; | |
4924 | if (peer_info.count(peer)) { | |
4925 | peer_info[peer].last_epoch_started = info.last_epoch_started; | |
4926 | peer_info[peer].last_interval_started = info.last_interval_started; | |
4927 | peer_info[peer].history.merge(info.history); | |
4928 | } | |
4929 | MOSDPGInfo *m = new MOSDPGInfo(get_osdmap()->get_epoch()); | |
4930 | m->pg_list.push_back( | |
4931 | make_pair( | |
4932 | pg_notify_t( | |
4933 | peer.shard, pg_whoami.shard, | |
4934 | get_osdmap()->get_epoch(), | |
4935 | get_osdmap()->get_epoch(), | |
4936 | info), | |
4937 | PastIntervals())); | |
4938 | osd->send_message_osd_cluster(peer.osd, m, get_osdmap()->get_epoch()); | |
4939 | } | |
4940 | } | |
4941 | ||
4942 | bool PG::append_log_entries_update_missing( | |
4943 | const mempool::osd_pglog::list<pg_log_entry_t> &entries, | |
4944 | ObjectStore::Transaction &t) | |
4945 | { | |
4946 | assert(!entries.empty()); | |
4947 | assert(entries.begin()->version > info.last_update); | |
4948 | ||
4949 | PGLogEntryHandler rollbacker{this, &t}; | |
4950 | bool invalidate_stats = | |
4951 | pg_log.append_new_log_entries(info.last_backfill, | |
4952 | info.last_backfill_bitwise, | |
4953 | entries, | |
4954 | &rollbacker); | |
4955 | info.last_update = pg_log.get_head(); | |
4956 | ||
4957 | if (pg_log.get_missing().num_missing() == 0) { | |
4958 | // advance last_complete since nothing else is missing! | |
4959 | info.last_complete = info.last_update; | |
4960 | } | |
4961 | ||
4962 | info.stats.stats_invalid = info.stats.stats_invalid || invalidate_stats; | |
4963 | dirty_info = true; | |
4964 | write_if_dirty(t); | |
4965 | return invalidate_stats; | |
4966 | } | |
4967 | ||
4968 | ||
4969 | void PG::merge_new_log_entries( | |
4970 | const mempool::osd_pglog::list<pg_log_entry_t> &entries, | |
4971 | ObjectStore::Transaction &t) | |
4972 | { | |
4973 | dout(10) << __func__ << " " << entries << dendl; | |
4974 | assert(is_primary()); | |
4975 | ||
4976 | bool rebuild_missing = append_log_entries_update_missing(entries, t); | |
4977 | for (set<pg_shard_t>::const_iterator i = actingbackfill.begin(); | |
4978 | i != actingbackfill.end(); | |
4979 | ++i) { | |
4980 | pg_shard_t peer(*i); | |
4981 | if (peer == pg_whoami) continue; | |
4982 | assert(peer_missing.count(peer)); | |
4983 | assert(peer_info.count(peer)); | |
4984 | pg_missing_t& pmissing(peer_missing[peer]); | |
4985 | pg_info_t& pinfo(peer_info[peer]); | |
4986 | bool invalidate_stats = PGLog::append_log_entries_update_missing( | |
4987 | pinfo.last_backfill, | |
4988 | info.last_backfill_bitwise, | |
4989 | entries, | |
4990 | true, | |
4991 | NULL, | |
4992 | pmissing, | |
4993 | NULL, | |
4994 | this); | |
4995 | pinfo.last_update = info.last_update; | |
4996 | pinfo.stats.stats_invalid = pinfo.stats.stats_invalid || invalidate_stats; | |
4997 | rebuild_missing = rebuild_missing || invalidate_stats; | |
4998 | } | |
4999 | ||
5000 | if (!rebuild_missing) { | |
5001 | return; | |
5002 | } | |
5003 | ||
5004 | for (auto &&i: entries) { | |
5005 | missing_loc.rebuild( | |
5006 | i.soid, | |
5007 | pg_whoami, | |
5008 | actingbackfill, | |
5009 | info, | |
5010 | pg_log.get_missing(), | |
5011 | peer_missing, | |
5012 | peer_info); | |
5013 | } | |
5014 | } | |
5015 | ||
5016 | void PG::update_history(const pg_history_t& new_history) | |
5017 | { | |
5018 | unreg_next_scrub(); | |
5019 | if (info.history.merge(new_history)) { | |
5020 | dout(20) << __func__ << " advanced history from " << new_history << dendl; | |
5021 | dirty_info = true; | |
5022 | if (info.history.last_epoch_clean >= info.history.same_interval_since) { | |
5023 | dout(20) << __func__ << " clearing past_intervals" << dendl; | |
5024 | past_intervals.clear(); | |
5025 | dirty_big_info = true; | |
5026 | } | |
5027 | } | |
5028 | reg_next_scrub(); | |
5029 | } | |
5030 | ||
5031 | void PG::fulfill_info( | |
5032 | pg_shard_t from, const pg_query_t &query, | |
5033 | pair<pg_shard_t, pg_info_t> ¬ify_info) | |
5034 | { | |
5035 | assert(from == primary); | |
5036 | assert(query.type == pg_query_t::INFO); | |
5037 | ||
5038 | // info | |
5039 | dout(10) << "sending info" << dendl; | |
5040 | notify_info = make_pair(from, info); | |
5041 | } | |
5042 | ||
5043 | void PG::fulfill_log( | |
5044 | pg_shard_t from, const pg_query_t &query, epoch_t query_epoch) | |
5045 | { | |
5046 | dout(10) << "log request from " << from << dendl; | |
5047 | assert(from == primary); | |
5048 | assert(query.type != pg_query_t::INFO); | |
5049 | ConnectionRef con = osd->get_con_osd_cluster( | |
5050 | from.osd, get_osdmap()->get_epoch()); | |
5051 | if (!con) return; | |
5052 | ||
5053 | MOSDPGLog *mlog = new MOSDPGLog( | |
5054 | from.shard, pg_whoami.shard, | |
5055 | get_osdmap()->get_epoch(), | |
5056 | info, query_epoch); | |
5057 | mlog->missing = pg_log.get_missing(); | |
5058 | ||
5059 | // primary -> other, when building master log | |
5060 | if (query.type == pg_query_t::LOG) { | |
5061 | dout(10) << " sending info+missing+log since " << query.since | |
5062 | << dendl; | |
5063 | if (query.since != eversion_t() && query.since < pg_log.get_tail()) { | |
5064 | osd->clog->error() << info.pgid << " got broken pg_query_t::LOG since " << query.since | |
5065 | << " when my log.tail is " << pg_log.get_tail() | |
5066 | << ", sending full log instead"; | |
5067 | mlog->log = pg_log.get_log(); // primary should not have requested this!! | |
5068 | } else | |
5069 | mlog->log.copy_after(pg_log.get_log(), query.since); | |
5070 | } | |
5071 | else if (query.type == pg_query_t::FULLLOG) { | |
5072 | dout(10) << " sending info+missing+full log" << dendl; | |
5073 | mlog->log = pg_log.get_log(); | |
5074 | } | |
5075 | ||
5076 | dout(10) << " sending " << mlog->log << " " << mlog->missing << dendl; | |
5077 | ||
5078 | osd->share_map_peer(from.osd, con.get(), get_osdmap()); | |
5079 | osd->send_message_osd_cluster(mlog, con.get()); | |
5080 | } | |
5081 | ||
5082 | void PG::check_full_transition(OSDMapRef lastmap, OSDMapRef osdmap) | |
5083 | { | |
5084 | bool changed = false; | |
5085 | if (osdmap->test_flag(CEPH_OSDMAP_FULL) && | |
5086 | !lastmap->test_flag(CEPH_OSDMAP_FULL)) { | |
5087 | dout(10) << " cluster was marked full in " << osdmap->get_epoch() << dendl; | |
5088 | changed = true; | |
5089 | } | |
5090 | const pg_pool_t *pi = osdmap->get_pg_pool(info.pgid.pool()); | |
5091 | assert(pi); | |
5092 | if (pi->has_flag(pg_pool_t::FLAG_FULL)) { | |
5093 | const pg_pool_t *opi = lastmap->get_pg_pool(info.pgid.pool()); | |
5094 | if (!opi || !opi->has_flag(pg_pool_t::FLAG_FULL)) { | |
5095 | dout(10) << " pool was marked full in " << osdmap->get_epoch() << dendl; | |
5096 | changed = true; | |
5097 | } | |
5098 | } | |
5099 | if (changed) { | |
5100 | info.history.last_epoch_marked_full = osdmap->get_epoch(); | |
5101 | dirty_info = true; | |
5102 | } | |
5103 | } | |
5104 | ||
5105 | bool PG::should_restart_peering( | |
5106 | int newupprimary, | |
5107 | int newactingprimary, | |
5108 | const vector<int>& newup, | |
5109 | const vector<int>& newacting, | |
5110 | OSDMapRef lastmap, | |
5111 | OSDMapRef osdmap) | |
5112 | { | |
5113 | if (PastIntervals::is_new_interval( | |
5114 | primary.osd, | |
5115 | newactingprimary, | |
5116 | acting, | |
5117 | newacting, | |
5118 | up_primary.osd, | |
5119 | newupprimary, | |
5120 | up, | |
5121 | newup, | |
5122 | osdmap, | |
5123 | lastmap, | |
5124 | info.pgid.pgid)) { | |
5125 | dout(20) << "new interval newup " << newup | |
5126 | << " newacting " << newacting << dendl; | |
5127 | return true; | |
5128 | } else { | |
5129 | return false; | |
5130 | } | |
5131 | } | |
5132 | ||
5133 | bool PG::old_peering_msg(epoch_t reply_epoch, epoch_t query_epoch) | |
5134 | { | |
5135 | if (last_peering_reset > reply_epoch || | |
5136 | last_peering_reset > query_epoch) { | |
5137 | dout(10) << "old_peering_msg reply_epoch " << reply_epoch << " query_epoch " << query_epoch | |
5138 | << " last_peering_reset " << last_peering_reset | |
5139 | << dendl; | |
5140 | return true; | |
5141 | } | |
5142 | return false; | |
5143 | } | |
5144 | ||
5145 | void PG::set_last_peering_reset() | |
5146 | { | |
5147 | dout(20) << "set_last_peering_reset " << get_osdmap()->get_epoch() << dendl; | |
5148 | if (last_peering_reset != get_osdmap()->get_epoch()) { | |
5149 | last_peering_reset = get_osdmap()->get_epoch(); | |
5150 | reset_interval_flush(); | |
5151 | } | |
5152 | } | |
5153 | ||
5154 | struct FlushState { | |
5155 | PGRef pg; | |
5156 | epoch_t epoch; | |
5157 | FlushState(PG *pg, epoch_t epoch) : pg(pg), epoch(epoch) {} | |
5158 | ~FlushState() { | |
5159 | pg->lock(); | |
5160 | if (!pg->pg_has_reset_since(epoch)) | |
5161 | pg->queue_flushed(epoch); | |
5162 | pg->unlock(); | |
5163 | } | |
5164 | }; | |
5165 | typedef ceph::shared_ptr<FlushState> FlushStateRef; | |
5166 | ||
5167 | void PG::start_flush(ObjectStore::Transaction *t, | |
5168 | list<Context *> *on_applied, | |
5169 | list<Context *> *on_safe) | |
5170 | { | |
5171 | // flush in progress ops | |
5172 | FlushStateRef flush_trigger (std::make_shared<FlushState>( | |
5173 | this, get_osdmap()->get_epoch())); | |
5174 | t->nop(); | |
5175 | flushes_in_progress++; | |
5176 | on_applied->push_back(new ContainerContext<FlushStateRef>(flush_trigger)); | |
5177 | on_safe->push_back(new ContainerContext<FlushStateRef>(flush_trigger)); | |
5178 | } | |
5179 | ||
5180 | void PG::reset_interval_flush() | |
5181 | { | |
5182 | dout(10) << "Clearing blocked outgoing recovery messages" << dendl; | |
5183 | recovery_state.clear_blocked_outgoing(); | |
5184 | ||
5185 | Context *c = new QueuePeeringEvt<IntervalFlush>( | |
5186 | this, get_osdmap()->get_epoch(), IntervalFlush()); | |
5187 | if (!osr->flush_commit(c)) { | |
5188 | dout(10) << "Beginning to block outgoing recovery messages" << dendl; | |
5189 | recovery_state.begin_block_outgoing(); | |
5190 | } else { | |
5191 | dout(10) << "Not blocking outgoing recovery messages" << dendl; | |
5192 | delete c; | |
5193 | } | |
5194 | } | |
5195 | ||
5196 | /* Called before initializing peering during advance_map */ | |
5197 | void PG::start_peering_interval( | |
5198 | const OSDMapRef lastmap, | |
5199 | const vector<int>& newup, int new_up_primary, | |
5200 | const vector<int>& newacting, int new_acting_primary, | |
5201 | ObjectStore::Transaction *t) | |
5202 | { | |
5203 | const OSDMapRef osdmap = get_osdmap(); | |
5204 | ||
5205 | set_last_peering_reset(); | |
5206 | ||
5207 | vector<int> oldacting, oldup; | |
5208 | int oldrole = get_role(); | |
5209 | ||
5210 | unreg_next_scrub(); | |
5211 | ||
5212 | pg_shard_t old_acting_primary = get_primary(); | |
5213 | pg_shard_t old_up_primary = up_primary; | |
5214 | bool was_old_primary = is_primary(); | |
5215 | ||
5216 | acting.swap(oldacting); | |
5217 | up.swap(oldup); | |
5218 | init_primary_up_acting( | |
5219 | newup, | |
5220 | newacting, | |
5221 | new_up_primary, | |
5222 | new_acting_primary); | |
5223 | ||
5224 | if (info.stats.up != up || | |
5225 | info.stats.acting != acting || | |
5226 | info.stats.up_primary != new_up_primary || | |
5227 | info.stats.acting_primary != new_acting_primary) { | |
5228 | info.stats.up = up; | |
5229 | info.stats.up_primary = new_up_primary; | |
5230 | info.stats.acting = acting; | |
5231 | info.stats.acting_primary = new_acting_primary; | |
5232 | info.stats.mapping_epoch = osdmap->get_epoch(); | |
5233 | } | |
5234 | ||
5235 | pg_stats_publish_lock.Lock(); | |
5236 | pg_stats_publish_valid = false; | |
5237 | pg_stats_publish_lock.Unlock(); | |
5238 | ||
5239 | // This will now be remapped during a backfill in cases | |
5240 | // that it would not have been before. | |
5241 | if (up != acting) | |
5242 | state_set(PG_STATE_REMAPPED); | |
5243 | else | |
5244 | state_clear(PG_STATE_REMAPPED); | |
5245 | ||
5246 | int role = osdmap->calc_pg_role(osd->whoami, acting, acting.size()); | |
5247 | if (pool.info.is_replicated() || role == pg_whoami.shard) | |
5248 | set_role(role); | |
5249 | else | |
5250 | set_role(-1); | |
5251 | ||
5252 | // did acting, up, primary|acker change? | |
5253 | if (!lastmap) { | |
5254 | dout(10) << " no lastmap" << dendl; | |
5255 | dirty_info = true; | |
5256 | dirty_big_info = true; | |
5257 | info.history.same_interval_since = osdmap->get_epoch(); | |
5258 | } else { | |
5259 | std::stringstream debug; | |
5260 | assert(info.history.same_interval_since != 0); | |
5261 | boost::scoped_ptr<IsPGRecoverablePredicate> recoverable( | |
5262 | get_is_recoverable_predicate()); | |
5263 | bool new_interval = PastIntervals::check_new_interval( | |
5264 | old_acting_primary.osd, | |
5265 | new_acting_primary, | |
5266 | oldacting, newacting, | |
5267 | old_up_primary.osd, | |
5268 | new_up_primary, | |
5269 | oldup, newup, | |
5270 | info.history.same_interval_since, | |
5271 | info.history.last_epoch_clean, | |
5272 | osdmap, | |
5273 | lastmap, | |
5274 | info.pgid.pgid, | |
5275 | recoverable.get(), | |
5276 | &past_intervals, | |
5277 | &debug); | |
5278 | dout(10) << __func__ << ": check_new_interval output: " | |
5279 | << debug.str() << dendl; | |
5280 | if (new_interval) { | |
5281 | if (osdmap->get_epoch() == osd->get_superblock().oldest_map && | |
5282 | info.history.last_epoch_clean < osdmap->get_epoch()) { | |
5283 | dout(10) << " map gap, clearing past_intervals and faking" << dendl; | |
5284 | // our information is incomplete and useless; someone else was clean | |
5285 | // after everything we know if osdmaps were trimmed. | |
5286 | past_intervals.clear(); | |
5287 | } else { | |
5288 | dout(10) << " noting past " << past_intervals << dendl; | |
5289 | } | |
5290 | dirty_info = true; | |
5291 | dirty_big_info = true; | |
5292 | info.history.same_interval_since = osdmap->get_epoch(); | |
5293 | if (info.pgid.pgid.is_split(lastmap->get_pg_num(info.pgid.pgid.pool()), | |
5294 | osdmap->get_pg_num(info.pgid.pgid.pool()), | |
5295 | nullptr)) { | |
5296 | info.history.last_epoch_split = osdmap->get_epoch(); | |
5297 | } | |
5298 | } | |
5299 | } | |
5300 | ||
5301 | if (old_up_primary != up_primary || | |
5302 | oldup != up) { | |
5303 | info.history.same_up_since = osdmap->get_epoch(); | |
5304 | } | |
5305 | // this comparison includes primary rank via pg_shard_t | |
5306 | if (old_acting_primary != get_primary()) { | |
5307 | info.history.same_primary_since = osdmap->get_epoch(); | |
5308 | } | |
5309 | ||
5310 | on_new_interval(); | |
5311 | ||
5312 | dout(1) << __func__ << " up " << oldup << " -> " << up | |
5313 | << ", acting " << oldacting << " -> " << acting | |
5314 | << ", acting_primary " << old_acting_primary << " -> " << new_acting_primary | |
5315 | << ", up_primary " << old_up_primary << " -> " << new_up_primary | |
5316 | << ", role " << oldrole << " -> " << role | |
5317 | << ", features acting " << acting_features | |
5318 | << " upacting " << upacting_features | |
5319 | << dendl; | |
5320 | ||
5321 | // deactivate. | |
5322 | state_clear(PG_STATE_ACTIVE); | |
5323 | state_clear(PG_STATE_PEERED); | |
5324 | state_clear(PG_STATE_DOWN); | |
5325 | state_clear(PG_STATE_RECOVERY_WAIT); | |
5326 | state_clear(PG_STATE_RECOVERY_TOOFULL); | |
5327 | state_clear(PG_STATE_RECOVERING); | |
5328 | ||
5329 | peer_purged.clear(); | |
5330 | actingbackfill.clear(); | |
5331 | scrub_queued = false; | |
5332 | ||
5333 | // reset primary state? | |
5334 | if (was_old_primary || is_primary()) { | |
5335 | osd->remove_want_pg_temp(info.pgid.pgid); | |
5336 | } | |
5337 | clear_primary_state(); | |
5338 | ||
5339 | ||
5340 | // pg->on_* | |
5341 | on_change(t); | |
5342 | ||
5343 | projected_last_update = eversion_t(); | |
5344 | ||
5345 | assert(!deleting); | |
5346 | ||
5347 | // should we tell the primary we are here? | |
5348 | send_notify = !is_primary(); | |
5349 | ||
5350 | if (role != oldrole || | |
5351 | was_old_primary != is_primary()) { | |
5352 | // did primary change? | |
5353 | if (was_old_primary != is_primary()) { | |
5354 | state_clear(PG_STATE_CLEAN); | |
5355 | clear_publish_stats(); | |
5356 | } | |
5357 | ||
5358 | on_role_change(); | |
5359 | ||
5360 | // take active waiters | |
5361 | requeue_ops(waiting_for_peered); | |
5362 | ||
5363 | } else { | |
5364 | // no role change. | |
5365 | // did primary change? | |
5366 | if (get_primary() != old_acting_primary) { | |
5367 | dout(10) << *this << " " << oldacting << " -> " << acting | |
5368 | << ", acting primary " | |
5369 | << old_acting_primary << " -> " << get_primary() | |
5370 | << dendl; | |
5371 | } else { | |
5372 | // primary is the same. | |
5373 | if (is_primary()) { | |
5374 | // i am (still) primary. but my replica set changed. | |
5375 | state_clear(PG_STATE_CLEAN); | |
5376 | ||
5377 | dout(10) << oldacting << " -> " << acting | |
5378 | << ", replicas changed" << dendl; | |
5379 | } | |
5380 | } | |
5381 | } | |
5382 | cancel_recovery(); | |
5383 | ||
5384 | if (acting.empty() && !up.empty() && up_primary == pg_whoami) { | |
5385 | dout(10) << " acting empty, but i am up[0], clearing pg_temp" << dendl; | |
5386 | osd->queue_want_pg_temp(info.pgid.pgid, acting); | |
5387 | } | |
5388 | } | |
5389 | ||
5390 | void PG::on_new_interval() | |
5391 | { | |
5392 | const OSDMapRef osdmap = get_osdmap(); | |
5393 | ||
5394 | reg_next_scrub(); | |
5395 | ||
5396 | // initialize features | |
5397 | acting_features = CEPH_FEATURES_SUPPORTED_DEFAULT; | |
5398 | upacting_features = CEPH_FEATURES_SUPPORTED_DEFAULT; | |
5399 | for (vector<int>::iterator p = acting.begin(); p != acting.end(); ++p) { | |
5400 | if (*p == CRUSH_ITEM_NONE) | |
5401 | continue; | |
5402 | uint64_t f = osdmap->get_xinfo(*p).features; | |
5403 | acting_features &= f; | |
5404 | upacting_features &= f; | |
5405 | } | |
5406 | for (vector<int>::iterator p = up.begin(); p != up.end(); ++p) { | |
5407 | if (*p == CRUSH_ITEM_NONE) | |
5408 | continue; | |
5409 | upacting_features &= osdmap->get_xinfo(*p).features; | |
5410 | } | |
5411 | ||
5412 | assert(osdmap->test_flag(CEPH_OSDMAP_SORTBITWISE)); | |
5413 | ||
5414 | _on_new_interval(); | |
5415 | } | |
5416 | ||
5417 | void PG::proc_primary_info(ObjectStore::Transaction &t, const pg_info_t &oinfo) | |
5418 | { | |
5419 | assert(!is_primary()); | |
5420 | ||
5421 | update_history(oinfo.history); | |
5422 | ||
5423 | if (last_complete_ondisk.epoch >= info.history.last_epoch_started) { | |
5424 | // DEBUG: verify that the snaps are empty in snap_mapper | |
5425 | if (cct->_conf->osd_debug_verify_snaps_on_info) { | |
5426 | interval_set<snapid_t> p; | |
5427 | p.union_of(oinfo.purged_snaps, info.purged_snaps); | |
5428 | p.subtract(info.purged_snaps); | |
5429 | if (!p.empty()) { | |
5430 | for (interval_set<snapid_t>::iterator i = p.begin(); | |
5431 | i != p.end(); | |
5432 | ++i) { | |
5433 | for (snapid_t snap = i.get_start(); | |
5434 | snap != i.get_len() + i.get_start(); | |
5435 | ++snap) { | |
5436 | vector<hobject_t> hoids; | |
5437 | int r = snap_mapper.get_next_objects_to_trim(snap, 1, &hoids); | |
5438 | if (r != 0 && r != -ENOENT) { | |
5439 | derr << __func__ << ": snap_mapper get_next_object_to_trim returned " | |
5440 | << cpp_strerror(r) << dendl; | |
5441 | ceph_abort(); | |
5442 | } else if (r != -ENOENT) { | |
5443 | assert(!hoids.empty()); | |
5444 | derr << __func__ << ": snap_mapper get_next_object_to_trim returned " | |
5445 | << cpp_strerror(r) << " for object " | |
5446 | << hoids[0] << " on snap " << snap | |
5447 | << " which should have been fully trimmed " << dendl; | |
5448 | ceph_abort(); | |
5449 | } | |
5450 | } | |
5451 | } | |
5452 | } | |
5453 | } | |
5454 | info.purged_snaps = oinfo.purged_snaps; | |
5455 | dirty_info = true; | |
5456 | dirty_big_info = true; | |
5457 | } | |
5458 | } | |
5459 | ||
5460 | ostream& operator<<(ostream& out, const PG& pg) | |
5461 | { | |
5462 | out << "pg[" << pg.info | |
5463 | << " " << pg.up; | |
5464 | if (pg.acting != pg.up) | |
5465 | out << "/" << pg.acting; | |
5466 | out << " r=" << pg.get_role(); | |
5467 | out << " lpr=" << pg.get_last_peering_reset(); | |
5468 | ||
5469 | if (!pg.past_intervals.empty()) { | |
5470 | out << " pi=[" << pg.past_intervals.get_bounds() | |
5471 | << ")/" << pg.past_intervals.size(); | |
5472 | } | |
5473 | ||
5474 | if (pg.is_peered()) { | |
5475 | if (pg.last_update_ondisk != pg.info.last_update) | |
5476 | out << " luod=" << pg.last_update_ondisk; | |
5477 | if (pg.last_update_applied != pg.info.last_update) | |
5478 | out << " lua=" << pg.last_update_applied; | |
5479 | } | |
5480 | ||
5481 | if (pg.recovery_ops_active) | |
5482 | out << " rops=" << pg.recovery_ops_active; | |
5483 | ||
5484 | if (pg.pg_log.get_tail() != pg.info.log_tail || | |
5485 | pg.pg_log.get_head() != pg.info.last_update) | |
5486 | out << " (info mismatch, " << pg.pg_log.get_log() << ")"; | |
5487 | ||
5488 | if (!pg.pg_log.get_log().empty()) { | |
5489 | if ((pg.pg_log.get_log().log.begin()->version <= pg.pg_log.get_tail())) { | |
5490 | out << " (log bound mismatch, actual=[" | |
5491 | << pg.pg_log.get_log().log.begin()->version << "," | |
5492 | << pg.pg_log.get_log().log.rbegin()->version << "]"; | |
5493 | out << ")"; | |
5494 | } | |
5495 | } | |
5496 | ||
5497 | if (!pg.backfill_targets.empty()) | |
5498 | out << " bft=" << pg.backfill_targets; | |
5499 | out << " crt=" << pg.pg_log.get_can_rollback_to(); | |
5500 | ||
5501 | if (pg.last_complete_ondisk != pg.info.last_complete) | |
5502 | out << " lcod " << pg.last_complete_ondisk; | |
5503 | ||
5504 | if (pg.is_primary()) { | |
5505 | out << " mlcod " << pg.min_last_complete_ondisk; | |
5506 | } | |
5507 | ||
5508 | out << " " << pg_state_string(pg.get_state()); | |
5509 | if (pg.should_send_notify()) | |
5510 | out << " NOTIFY"; | |
5511 | ||
5512 | if (pg.scrubber.must_repair) | |
5513 | out << " MUST_REPAIR"; | |
5514 | if (pg.scrubber.auto_repair) | |
5515 | out << " AUTO_REPAIR"; | |
5516 | if (pg.scrubber.must_deep_scrub) | |
5517 | out << " MUST_DEEP_SCRUB"; | |
5518 | if (pg.scrubber.must_scrub) | |
5519 | out << " MUST_SCRUB"; | |
5520 | ||
5521 | //out << " (" << pg.pg_log.get_tail() << "," << pg.pg_log.get_head() << "]"; | |
5522 | if (pg.pg_log.get_missing().num_missing()) { | |
5523 | out << " m=" << pg.pg_log.get_missing().num_missing(); | |
5524 | if (pg.is_primary()) { | |
5525 | uint64_t unfound = pg.get_num_unfound(); | |
5526 | if (unfound) | |
5527 | out << " u=" << unfound; | |
5528 | } | |
5529 | } | |
5530 | if (pg.snap_trimq.size()) | |
5531 | out << " snaptrimq=" << pg.snap_trimq; | |
5532 | ||
5533 | out << "]"; | |
5534 | ||
5535 | ||
5536 | return out; | |
5537 | } | |
5538 | ||
5539 | bool PG::can_discard_op(OpRequestRef& op) | |
5540 | { | |
5541 | const MOSDOp *m = static_cast<const MOSDOp*>(op->get_req()); | |
5542 | if (cct->_conf->osd_discard_disconnected_ops && OSD::op_is_discardable(m)) { | |
5543 | dout(20) << " discard " << *m << dendl; | |
5544 | return true; | |
5545 | } | |
5546 | ||
5547 | if (m->get_map_epoch() < info.history.same_primary_since) { | |
5548 | dout(7) << " changed after " << m->get_map_epoch() | |
5549 | << ", dropping " << *m << dendl; | |
5550 | return true; | |
5551 | } | |
5552 | ||
5553 | if (m->get_connection()->has_feature(CEPH_FEATURE_RESEND_ON_SPLIT)) { | |
5554 | if (m->get_map_epoch() < pool.info.get_last_force_op_resend()) { | |
5555 | dout(7) << __func__ << " sent before last_force_op_resend " | |
5556 | << pool.info.last_force_op_resend << ", dropping" << *m << dendl; | |
5557 | return true; | |
5558 | } | |
5559 | if (m->get_map_epoch() < info.history.last_epoch_split) { | |
5560 | dout(7) << __func__ << " pg split in " | |
5561 | << info.history.last_epoch_split << ", dropping" << dendl; | |
5562 | return true; | |
5563 | } | |
5564 | } else if (m->get_connection()->has_feature(CEPH_FEATURE_OSD_POOLRESEND)) { | |
5565 | if (m->get_map_epoch() < pool.info.get_last_force_op_resend_preluminous()) { | |
5566 | dout(7) << __func__ << " sent before last_force_op_resend_preluminous " | |
5567 | << pool.info.last_force_op_resend_preluminous | |
5568 | << ", dropping" << *m << dendl; | |
5569 | return true; | |
5570 | } | |
5571 | } | |
5572 | ||
5573 | return false; | |
5574 | } | |
5575 | ||
5576 | template<typename T, int MSGTYPE> | |
5577 | bool PG::can_discard_replica_op(OpRequestRef& op) | |
5578 | { | |
5579 | const T *m = static_cast<const T *>(op->get_req()); | |
5580 | assert(m->get_type() == MSGTYPE); | |
5581 | ||
5582 | /* Mostly, this overlaps with the old_peering_msg | |
5583 | * condition. An important exception is pushes | |
5584 | * sent by replicas not in the acting set, since | |
5585 | * if such a replica goes down it does not cause | |
5586 | * a new interval. */ | |
5587 | int from = m->get_source().num(); | |
5588 | if (get_osdmap()->get_down_at(from) >= m->map_epoch) | |
5589 | return true; | |
5590 | ||
5591 | // same pg? | |
5592 | // if pg changes _at all_, we reset and repeer! | |
5593 | if (old_peering_msg(m->map_epoch, m->map_epoch)) { | |
5594 | dout(10) << "can_discard_replica_op pg changed " << info.history | |
5595 | << " after " << m->map_epoch | |
5596 | << ", dropping" << dendl; | |
5597 | return true; | |
5598 | } | |
5599 | return false; | |
5600 | } | |
5601 | ||
5602 | bool PG::can_discard_scan(OpRequestRef op) | |
5603 | { | |
5604 | const MOSDPGScan *m = static_cast<const MOSDPGScan *>(op->get_req()); | |
5605 | assert(m->get_type() == MSG_OSD_PG_SCAN); | |
5606 | ||
5607 | if (old_peering_msg(m->map_epoch, m->query_epoch)) { | |
5608 | dout(10) << " got old scan, ignoring" << dendl; | |
5609 | return true; | |
5610 | } | |
5611 | return false; | |
5612 | } | |
5613 | ||
5614 | bool PG::can_discard_backfill(OpRequestRef op) | |
5615 | { | |
5616 | const MOSDPGBackfill *m = static_cast<const MOSDPGBackfill *>(op->get_req()); | |
5617 | assert(m->get_type() == MSG_OSD_PG_BACKFILL); | |
5618 | ||
5619 | if (old_peering_msg(m->map_epoch, m->query_epoch)) { | |
5620 | dout(10) << " got old backfill, ignoring" << dendl; | |
5621 | return true; | |
5622 | } | |
5623 | ||
5624 | return false; | |
5625 | ||
5626 | } | |
5627 | ||
5628 | bool PG::can_discard_request(OpRequestRef& op) | |
5629 | { | |
5630 | switch (op->get_req()->get_type()) { | |
5631 | case CEPH_MSG_OSD_OP: | |
5632 | return can_discard_op(op); | |
5633 | case CEPH_MSG_OSD_BACKOFF: | |
5634 | return false; // never discard | |
5635 | case MSG_OSD_SUBOP: | |
5636 | return can_discard_replica_op<MOSDSubOp, MSG_OSD_SUBOP>(op); | |
5637 | case MSG_OSD_REPOP: | |
5638 | return can_discard_replica_op<MOSDRepOp, MSG_OSD_REPOP>(op); | |
5639 | case MSG_OSD_PG_PUSH: | |
5640 | return can_discard_replica_op<MOSDPGPush, MSG_OSD_PG_PUSH>(op); | |
5641 | case MSG_OSD_PG_PULL: | |
5642 | return can_discard_replica_op<MOSDPGPull, MSG_OSD_PG_PULL>(op); | |
5643 | case MSG_OSD_PG_PUSH_REPLY: | |
5644 | return can_discard_replica_op<MOSDPGPushReply, MSG_OSD_PG_PUSH_REPLY>(op); | |
5645 | case MSG_OSD_SUBOPREPLY: | |
5646 | return can_discard_replica_op<MOSDSubOpReply, MSG_OSD_SUBOPREPLY>(op); | |
5647 | case MSG_OSD_REPOPREPLY: | |
5648 | return can_discard_replica_op<MOSDRepOpReply, MSG_OSD_REPOPREPLY>(op); | |
5649 | ||
5650 | case MSG_OSD_EC_WRITE: | |
5651 | return can_discard_replica_op<MOSDECSubOpWrite, MSG_OSD_EC_WRITE>(op); | |
5652 | case MSG_OSD_EC_WRITE_REPLY: | |
5653 | return can_discard_replica_op<MOSDECSubOpWriteReply, MSG_OSD_EC_WRITE_REPLY>(op); | |
5654 | case MSG_OSD_EC_READ: | |
5655 | return can_discard_replica_op<MOSDECSubOpRead, MSG_OSD_EC_READ>(op); | |
5656 | case MSG_OSD_EC_READ_REPLY: | |
5657 | return can_discard_replica_op<MOSDECSubOpReadReply, MSG_OSD_EC_READ_REPLY>(op); | |
5658 | case MSG_OSD_REP_SCRUB: | |
5659 | return can_discard_replica_op<MOSDRepScrub, MSG_OSD_REP_SCRUB>(op); | |
5660 | case MSG_OSD_SCRUB_RESERVE: | |
5661 | return can_discard_replica_op<MOSDScrubReserve, MSG_OSD_SCRUB_RESERVE>(op); | |
5662 | case MSG_OSD_REP_SCRUBMAP: | |
5663 | return can_discard_replica_op<MOSDRepScrubMap, MSG_OSD_REP_SCRUBMAP>(op); | |
5664 | case MSG_OSD_PG_UPDATE_LOG_MISSING: | |
5665 | return can_discard_replica_op< | |
5666 | MOSDPGUpdateLogMissing, MSG_OSD_PG_UPDATE_LOG_MISSING>(op); | |
5667 | case MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY: | |
5668 | return can_discard_replica_op< | |
5669 | MOSDPGUpdateLogMissingReply, MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY>(op); | |
5670 | ||
5671 | case MSG_OSD_PG_SCAN: | |
5672 | return can_discard_scan(op); | |
5673 | case MSG_OSD_PG_BACKFILL: | |
5674 | return can_discard_backfill(op); | |
5675 | case MSG_OSD_PG_BACKFILL_REMOVE: | |
5676 | return can_discard_replica_op<MOSDPGBackfillRemove, | |
5677 | MSG_OSD_PG_BACKFILL_REMOVE>(op); | |
5678 | } | |
5679 | return true; | |
5680 | } | |
5681 | ||
5682 | void PG::take_waiters() | |
5683 | { | |
5684 | dout(10) << "take_waiters" << dendl; | |
5685 | requeue_map_waiters(); | |
5686 | for (list<CephPeeringEvtRef>::iterator i = peering_waiters.begin(); | |
5687 | i != peering_waiters.end(); | |
5688 | ++i) osd->queue_for_peering(this); | |
5689 | peering_queue.splice(peering_queue.begin(), peering_waiters, | |
5690 | peering_waiters.begin(), peering_waiters.end()); | |
5691 | } | |
5692 | ||
5693 | void PG::handle_peering_event(CephPeeringEvtRef evt, RecoveryCtx *rctx) | |
5694 | { | |
5695 | dout(10) << "handle_peering_event: " << evt->get_desc() << dendl; | |
5696 | if (!have_same_or_newer_map(evt->get_epoch_sent())) { | |
5697 | dout(10) << "deferring event " << evt->get_desc() << dendl; | |
5698 | peering_waiters.push_back(evt); | |
5699 | return; | |
5700 | } | |
5701 | if (old_peering_evt(evt)) | |
5702 | return; | |
5703 | recovery_state.handle_event(evt, rctx); | |
5704 | } | |
5705 | ||
5706 | void PG::queue_peering_event(CephPeeringEvtRef evt) | |
5707 | { | |
5708 | if (old_peering_evt(evt)) | |
5709 | return; | |
5710 | peering_queue.push_back(evt); | |
5711 | osd->queue_for_peering(this); | |
5712 | } | |
5713 | ||
5714 | void PG::queue_null(epoch_t msg_epoch, | |
5715 | epoch_t query_epoch) | |
5716 | { | |
5717 | dout(10) << "null" << dendl; | |
5718 | queue_peering_event( | |
5719 | CephPeeringEvtRef(std::make_shared<CephPeeringEvt>(msg_epoch, query_epoch, | |
5720 | NullEvt()))); | |
5721 | } | |
5722 | ||
5723 | void PG::queue_flushed(epoch_t e) | |
5724 | { | |
5725 | dout(10) << "flushed" << dendl; | |
5726 | queue_peering_event( | |
5727 | CephPeeringEvtRef(std::make_shared<CephPeeringEvt>(e, e, | |
5728 | FlushedEvt()))); | |
5729 | } | |
5730 | ||
5731 | void PG::queue_query(epoch_t msg_epoch, | |
5732 | epoch_t query_epoch, | |
5733 | pg_shard_t from, const pg_query_t& q) | |
5734 | { | |
5735 | dout(10) << "handle_query " << q << " from replica " << from << dendl; | |
5736 | queue_peering_event( | |
5737 | CephPeeringEvtRef(std::make_shared<CephPeeringEvt>(msg_epoch, query_epoch, | |
5738 | MQuery(from, q, query_epoch)))); | |
5739 | } | |
5740 | ||
5741 | void PG::handle_advance_map( | |
5742 | OSDMapRef osdmap, OSDMapRef lastmap, | |
5743 | vector<int>& newup, int up_primary, | |
5744 | vector<int>& newacting, int acting_primary, | |
5745 | RecoveryCtx *rctx) | |
5746 | { | |
5747 | assert(lastmap->get_epoch() == osdmap_ref->get_epoch()); | |
5748 | assert(lastmap == osdmap_ref); | |
5749 | dout(10) << "handle_advance_map " | |
5750 | << newup << "/" << newacting | |
5751 | << " -- " << up_primary << "/" << acting_primary | |
5752 | << dendl; | |
5753 | update_osdmap_ref(osdmap); | |
5754 | pool.update(osdmap); | |
5755 | past_intervals.update_type_from_map(pool.info.ec_pool(), *osdmap); | |
5756 | if (cct->_conf->osd_debug_verify_cached_snaps) { | |
5757 | interval_set<snapid_t> actual_removed_snaps; | |
5758 | const pg_pool_t *pi = osdmap->get_pg_pool(info.pgid.pool()); | |
5759 | assert(pi); | |
5760 | pi->build_removed_snaps(actual_removed_snaps); | |
5761 | if (!(actual_removed_snaps == pool.cached_removed_snaps)) { | |
5762 | derr << __func__ << ": mismatch between the actual removed snaps " | |
5763 | << actual_removed_snaps << " and pool.cached_removed_snaps " | |
5764 | << " pool.cached_removed_snaps " << pool.cached_removed_snaps | |
5765 | << dendl; | |
5766 | } | |
5767 | assert(actual_removed_snaps == pool.cached_removed_snaps); | |
5768 | } | |
5769 | AdvMap evt( | |
5770 | osdmap, lastmap, newup, up_primary, | |
5771 | newacting, acting_primary); | |
5772 | recovery_state.handle_event(evt, rctx); | |
5773 | if (pool.info.last_change == osdmap_ref->get_epoch()) { | |
5774 | on_pool_change(); | |
5775 | update_store_with_options(); | |
5776 | } | |
5777 | } | |
5778 | ||
5779 | void PG::handle_activate_map(RecoveryCtx *rctx) | |
5780 | { | |
5781 | dout(10) << "handle_activate_map " << dendl; | |
5782 | ActMap evt; | |
5783 | recovery_state.handle_event(evt, rctx); | |
5784 | if (osdmap_ref->get_epoch() - last_persisted_osdmap_ref->get_epoch() > | |
5785 | cct->_conf->osd_pg_epoch_persisted_max_stale) { | |
5786 | dout(20) << __func__ << ": Dirtying info: last_persisted is " | |
5787 | << last_persisted_osdmap_ref->get_epoch() | |
5788 | << " while current is " << osdmap_ref->get_epoch() << dendl; | |
5789 | dirty_info = true; | |
5790 | } else { | |
5791 | dout(20) << __func__ << ": Not dirtying info: last_persisted is " | |
5792 | << last_persisted_osdmap_ref->get_epoch() | |
5793 | << " while current is " << osdmap_ref->get_epoch() << dendl; | |
5794 | } | |
5795 | if (osdmap_ref->check_new_blacklist_entries()) check_blacklisted_watchers(); | |
5796 | } | |
5797 | ||
5798 | void PG::handle_loaded(RecoveryCtx *rctx) | |
5799 | { | |
5800 | dout(10) << "handle_loaded" << dendl; | |
5801 | Load evt; | |
5802 | recovery_state.handle_event(evt, rctx); | |
5803 | } | |
5804 | ||
5805 | void PG::handle_create(RecoveryCtx *rctx) | |
5806 | { | |
5807 | dout(10) << "handle_create" << dendl; | |
5808 | rctx->created_pgs.insert(this); | |
5809 | Initialize evt; | |
5810 | recovery_state.handle_event(evt, rctx); | |
5811 | ActMap evt2; | |
5812 | recovery_state.handle_event(evt2, rctx); | |
5813 | } | |
5814 | ||
5815 | void PG::handle_query_state(Formatter *f) | |
5816 | { | |
5817 | dout(10) << "handle_query_state" << dendl; | |
5818 | QueryState q(f); | |
5819 | recovery_state.handle_event(q, 0); | |
5820 | } | |
5821 | ||
5822 | void PG::update_store_with_options() | |
5823 | { | |
5824 | auto r = osd->store->set_collection_opts(coll, pool.info.opts); | |
5825 | if(r < 0 && r != -EOPNOTSUPP) { | |
5826 | derr << __func__ << "set_collection_opts returns error:" << r << dendl; | |
5827 | } | |
5828 | } | |
5829 | ||
5830 | void PG::update_store_on_load() | |
5831 | { | |
5832 | if (osd->store->get_type() == "filestore") { | |
5833 | // legacy filestore didn't store collection bit width; fix. | |
5834 | int bits = osd->store->collection_bits(coll); | |
5835 | if (bits < 0) { | |
5836 | if (coll.is_meta()) | |
5837 | bits = 0; | |
5838 | else | |
5839 | bits = info.pgid.get_split_bits(pool.info.get_pg_num()); | |
5840 | lderr(cct) << __func__ << " setting bit width to " << bits << dendl; | |
5841 | ObjectStore::Transaction t; | |
5842 | t.collection_set_bits(coll, bits); | |
5843 | osd->store->apply_transaction(osr.get(), std::move(t)); | |
5844 | } | |
5845 | } | |
5846 | } | |
5847 | ||
5848 | /*------------ Recovery State Machine----------------*/ | |
5849 | #undef dout_prefix | |
5850 | #define dout_prefix (*_dout << context< RecoveryMachine >().pg->gen_prefix() \ | |
5851 | << "state<" << get_state_name() << ">: ") | |
5852 | ||
5853 | /*------Crashed-------*/ | |
5854 | PG::RecoveryState::Crashed::Crashed(my_context ctx) | |
5855 | : my_base(ctx), | |
5856 | NamedState(context< RecoveryMachine >().pg, "Crashed") | |
5857 | { | |
5858 | context< RecoveryMachine >().log_enter(state_name); | |
5859 | assert(0 == "we got a bad state machine event"); | |
5860 | } | |
5861 | ||
5862 | ||
5863 | /*------Initial-------*/ | |
5864 | PG::RecoveryState::Initial::Initial(my_context ctx) | |
5865 | : my_base(ctx), | |
5866 | NamedState(context< RecoveryMachine >().pg, "Initial") | |
5867 | { | |
5868 | context< RecoveryMachine >().log_enter(state_name); | |
5869 | } | |
5870 | ||
5871 | boost::statechart::result PG::RecoveryState::Initial::react(const Load& l) | |
5872 | { | |
5873 | PG *pg = context< RecoveryMachine >().pg; | |
5874 | ||
5875 | // do we tell someone we're here? | |
5876 | pg->send_notify = (!pg->is_primary()); | |
5877 | pg->update_store_with_options(); | |
5878 | ||
5879 | pg->update_store_on_load(); | |
5880 | ||
5881 | return transit< Reset >(); | |
5882 | } | |
5883 | ||
5884 | boost::statechart::result PG::RecoveryState::Initial::react(const MNotifyRec& notify) | |
5885 | { | |
5886 | PG *pg = context< RecoveryMachine >().pg; | |
5887 | pg->proc_replica_info( | |
5888 | notify.from, notify.notify.info, notify.notify.epoch_sent); | |
5889 | pg->set_last_peering_reset(); | |
5890 | return transit< Primary >(); | |
5891 | } | |
5892 | ||
5893 | boost::statechart::result PG::RecoveryState::Initial::react(const MInfoRec& i) | |
5894 | { | |
5895 | PG *pg = context< RecoveryMachine >().pg; | |
5896 | assert(!pg->is_primary()); | |
5897 | post_event(i); | |
5898 | return transit< Stray >(); | |
5899 | } | |
5900 | ||
5901 | boost::statechart::result PG::RecoveryState::Initial::react(const MLogRec& i) | |
5902 | { | |
5903 | PG *pg = context< RecoveryMachine >().pg; | |
5904 | assert(!pg->is_primary()); | |
5905 | post_event(i); | |
5906 | return transit< Stray >(); | |
5907 | } | |
5908 | ||
5909 | void PG::RecoveryState::Initial::exit() | |
5910 | { | |
5911 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
5912 | PG *pg = context< RecoveryMachine >().pg; | |
5913 | utime_t dur = ceph_clock_now() - enter_time; | |
5914 | pg->osd->recoverystate_perf->tinc(rs_initial_latency, dur); | |
5915 | } | |
5916 | ||
5917 | /*------Started-------*/ | |
5918 | PG::RecoveryState::Started::Started(my_context ctx) | |
5919 | : my_base(ctx), | |
5920 | NamedState(context< RecoveryMachine >().pg, "Started") | |
5921 | { | |
5922 | context< RecoveryMachine >().log_enter(state_name); | |
5923 | } | |
5924 | ||
5925 | boost::statechart::result | |
5926 | PG::RecoveryState::Started::react(const IntervalFlush&) | |
5927 | { | |
5928 | PG *pg = context< RecoveryMachine >().pg; | |
5929 | ldout(pg->cct, 10) << "Ending blocked outgoing recovery messages" << dendl; | |
5930 | context< RecoveryMachine >().pg->recovery_state.end_block_outgoing(); | |
5931 | return discard_event(); | |
5932 | } | |
5933 | ||
5934 | ||
5935 | boost::statechart::result | |
5936 | PG::RecoveryState::Started::react(const FlushedEvt&) | |
5937 | { | |
5938 | PG *pg = context< RecoveryMachine >().pg; | |
5939 | pg->on_flushed(); | |
5940 | return discard_event(); | |
5941 | } | |
5942 | ||
5943 | ||
5944 | boost::statechart::result PG::RecoveryState::Started::react(const AdvMap& advmap) | |
5945 | { | |
5946 | PG *pg = context< RecoveryMachine >().pg; | |
5947 | ldout(pg->cct, 10) << "Started advmap" << dendl; | |
5948 | pg->check_full_transition(advmap.lastmap, advmap.osdmap); | |
5949 | if (pg->should_restart_peering( | |
5950 | advmap.up_primary, | |
5951 | advmap.acting_primary, | |
5952 | advmap.newup, | |
5953 | advmap.newacting, | |
5954 | advmap.lastmap, | |
5955 | advmap.osdmap)) { | |
5956 | ldout(pg->cct, 10) << "should_restart_peering, transitioning to Reset" | |
5957 | << dendl; | |
5958 | post_event(advmap); | |
5959 | return transit< Reset >(); | |
5960 | } | |
5961 | pg->remove_down_peer_info(advmap.osdmap); | |
5962 | return discard_event(); | |
5963 | } | |
5964 | ||
5965 | boost::statechart::result PG::RecoveryState::Started::react(const QueryState& q) | |
5966 | { | |
5967 | q.f->open_object_section("state"); | |
5968 | q.f->dump_string("name", state_name); | |
5969 | q.f->dump_stream("enter_time") << enter_time; | |
5970 | q.f->close_section(); | |
5971 | return discard_event(); | |
5972 | } | |
5973 | ||
5974 | void PG::RecoveryState::Started::exit() | |
5975 | { | |
5976 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
5977 | PG *pg = context< RecoveryMachine >().pg; | |
5978 | utime_t dur = ceph_clock_now() - enter_time; | |
5979 | pg->osd->recoverystate_perf->tinc(rs_started_latency, dur); | |
5980 | } | |
5981 | ||
5982 | /*--------Reset---------*/ | |
5983 | PG::RecoveryState::Reset::Reset(my_context ctx) | |
5984 | : my_base(ctx), | |
5985 | NamedState(context< RecoveryMachine >().pg, "Reset") | |
5986 | { | |
5987 | context< RecoveryMachine >().log_enter(state_name); | |
5988 | PG *pg = context< RecoveryMachine >().pg; | |
5989 | ||
5990 | pg->flushes_in_progress = 0; | |
5991 | pg->set_last_peering_reset(); | |
5992 | } | |
5993 | ||
5994 | boost::statechart::result | |
5995 | PG::RecoveryState::Reset::react(const FlushedEvt&) | |
5996 | { | |
5997 | PG *pg = context< RecoveryMachine >().pg; | |
5998 | pg->on_flushed(); | |
5999 | return discard_event(); | |
6000 | } | |
6001 | ||
6002 | boost::statechart::result | |
6003 | PG::RecoveryState::Reset::react(const IntervalFlush&) | |
6004 | { | |
6005 | PG *pg = context< RecoveryMachine >().pg; | |
6006 | ldout(pg->cct, 10) << "Ending blocked outgoing recovery messages" << dendl; | |
6007 | context< RecoveryMachine >().pg->recovery_state.end_block_outgoing(); | |
6008 | return discard_event(); | |
6009 | } | |
6010 | ||
6011 | boost::statechart::result PG::RecoveryState::Reset::react(const AdvMap& advmap) | |
6012 | { | |
6013 | PG *pg = context< RecoveryMachine >().pg; | |
6014 | ldout(pg->cct, 10) << "Reset advmap" << dendl; | |
6015 | ||
6016 | pg->check_full_transition(advmap.lastmap, advmap.osdmap); | |
6017 | ||
6018 | if (pg->should_restart_peering( | |
6019 | advmap.up_primary, | |
6020 | advmap.acting_primary, | |
6021 | advmap.newup, | |
6022 | advmap.newacting, | |
6023 | advmap.lastmap, | |
6024 | advmap.osdmap)) { | |
6025 | ldout(pg->cct, 10) << "should restart peering, calling start_peering_interval again" | |
6026 | << dendl; | |
6027 | pg->start_peering_interval( | |
6028 | advmap.lastmap, | |
6029 | advmap.newup, advmap.up_primary, | |
6030 | advmap.newacting, advmap.acting_primary, | |
6031 | context< RecoveryMachine >().get_cur_transaction()); | |
6032 | } | |
6033 | pg->remove_down_peer_info(advmap.osdmap); | |
6034 | pg->check_past_interval_bounds(); | |
6035 | return discard_event(); | |
6036 | } | |
6037 | ||
6038 | boost::statechart::result PG::RecoveryState::Reset::react(const ActMap&) | |
6039 | { | |
6040 | PG *pg = context< RecoveryMachine >().pg; | |
6041 | if (pg->should_send_notify() && pg->get_primary().osd >= 0) { | |
6042 | context< RecoveryMachine >().send_notify( | |
6043 | pg->get_primary(), | |
6044 | pg_notify_t( | |
6045 | pg->get_primary().shard, pg->pg_whoami.shard, | |
6046 | pg->get_osdmap()->get_epoch(), | |
6047 | pg->get_osdmap()->get_epoch(), | |
6048 | pg->info), | |
6049 | pg->past_intervals); | |
6050 | } | |
6051 | ||
6052 | pg->update_heartbeat_peers(); | |
6053 | pg->take_waiters(); | |
6054 | ||
6055 | return transit< Started >(); | |
6056 | } | |
6057 | ||
6058 | boost::statechart::result PG::RecoveryState::Reset::react(const QueryState& q) | |
6059 | { | |
6060 | q.f->open_object_section("state"); | |
6061 | q.f->dump_string("name", state_name); | |
6062 | q.f->dump_stream("enter_time") << enter_time; | |
6063 | q.f->close_section(); | |
6064 | return discard_event(); | |
6065 | } | |
6066 | ||
6067 | void PG::RecoveryState::Reset::exit() | |
6068 | { | |
6069 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6070 | PG *pg = context< RecoveryMachine >().pg; | |
6071 | utime_t dur = ceph_clock_now() - enter_time; | |
6072 | pg->osd->recoverystate_perf->tinc(rs_reset_latency, dur); | |
6073 | } | |
6074 | ||
6075 | /*-------Start---------*/ | |
6076 | PG::RecoveryState::Start::Start(my_context ctx) | |
6077 | : my_base(ctx), | |
6078 | NamedState(context< RecoveryMachine >().pg, "Start") | |
6079 | { | |
6080 | context< RecoveryMachine >().log_enter(state_name); | |
6081 | ||
6082 | PG *pg = context< RecoveryMachine >().pg; | |
6083 | if (pg->is_primary()) { | |
6084 | ldout(pg->cct, 1) << "transitioning to Primary" << dendl; | |
6085 | post_event(MakePrimary()); | |
6086 | } else { //is_stray | |
6087 | ldout(pg->cct, 1) << "transitioning to Stray" << dendl; | |
6088 | post_event(MakeStray()); | |
6089 | } | |
6090 | } | |
6091 | ||
6092 | void PG::RecoveryState::Start::exit() | |
6093 | { | |
6094 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6095 | PG *pg = context< RecoveryMachine >().pg; | |
6096 | utime_t dur = ceph_clock_now() - enter_time; | |
6097 | pg->osd->recoverystate_perf->tinc(rs_start_latency, dur); | |
6098 | } | |
6099 | ||
6100 | /*---------Primary--------*/ | |
6101 | PG::RecoveryState::Primary::Primary(my_context ctx) | |
6102 | : my_base(ctx), | |
6103 | NamedState(context< RecoveryMachine >().pg, "Started/Primary") | |
6104 | { | |
6105 | context< RecoveryMachine >().log_enter(state_name); | |
6106 | PG *pg = context< RecoveryMachine >().pg; | |
6107 | assert(pg->want_acting.empty()); | |
6108 | ||
6109 | // set CREATING bit until we have peered for the first time. | |
6110 | if (pg->info.history.last_epoch_started == 0) { | |
6111 | pg->state_set(PG_STATE_CREATING); | |
6112 | // use the history timestamp, which ultimately comes from the | |
6113 | // monitor in the create case. | |
6114 | utime_t t = pg->info.history.last_scrub_stamp; | |
6115 | pg->info.stats.last_fresh = t; | |
6116 | pg->info.stats.last_active = t; | |
6117 | pg->info.stats.last_change = t; | |
6118 | pg->info.stats.last_peered = t; | |
6119 | pg->info.stats.last_clean = t; | |
6120 | pg->info.stats.last_unstale = t; | |
6121 | pg->info.stats.last_undegraded = t; | |
6122 | pg->info.stats.last_fullsized = t; | |
6123 | pg->info.stats.last_scrub_stamp = t; | |
6124 | pg->info.stats.last_deep_scrub_stamp = t; | |
6125 | pg->info.stats.last_clean_scrub_stamp = t; | |
6126 | } | |
6127 | } | |
6128 | ||
6129 | boost::statechart::result PG::RecoveryState::Primary::react(const MNotifyRec& notevt) | |
6130 | { | |
6131 | PG *pg = context< RecoveryMachine >().pg; | |
6132 | ldout(pg->cct, 7) << "handle_pg_notify from osd." << notevt.from << dendl; | |
6133 | pg->proc_replica_info( | |
6134 | notevt.from, notevt.notify.info, notevt.notify.epoch_sent); | |
6135 | return discard_event(); | |
6136 | } | |
6137 | ||
6138 | boost::statechart::result PG::RecoveryState::Primary::react(const ActMap&) | |
6139 | { | |
6140 | PG *pg = context< RecoveryMachine >().pg; | |
6141 | ldout(pg->cct, 7) << "handle ActMap primary" << dendl; | |
6142 | pg->publish_stats_to_osd(); | |
6143 | pg->take_waiters(); | |
6144 | return discard_event(); | |
6145 | } | |
6146 | ||
6147 | void PG::RecoveryState::Primary::exit() | |
6148 | { | |
6149 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6150 | PG *pg = context< RecoveryMachine >().pg; | |
6151 | pg->want_acting.clear(); | |
6152 | utime_t dur = ceph_clock_now() - enter_time; | |
6153 | pg->osd->recoverystate_perf->tinc(rs_primary_latency, dur); | |
6154 | pg->clear_primary_state(); | |
6155 | pg->state_clear(PG_STATE_CREATING); | |
6156 | } | |
6157 | ||
6158 | /*---------Peering--------*/ | |
6159 | PG::RecoveryState::Peering::Peering(my_context ctx) | |
6160 | : my_base(ctx), | |
6161 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Peering"), | |
6162 | history_les_bound(false) | |
6163 | { | |
6164 | context< RecoveryMachine >().log_enter(state_name); | |
6165 | ||
6166 | PG *pg = context< RecoveryMachine >().pg; | |
6167 | assert(!pg->is_peered()); | |
6168 | assert(!pg->is_peering()); | |
6169 | assert(pg->is_primary()); | |
6170 | pg->state_set(PG_STATE_PEERING); | |
6171 | } | |
6172 | ||
6173 | boost::statechart::result PG::RecoveryState::Peering::react(const AdvMap& advmap) | |
6174 | { | |
6175 | PG *pg = context< RecoveryMachine >().pg; | |
6176 | ldout(pg->cct, 10) << "Peering advmap" << dendl; | |
6177 | if (prior_set.affected_by_map(*(advmap.osdmap), pg)) { | |
6178 | ldout(pg->cct, 1) << "Peering, affected_by_map, going to Reset" << dendl; | |
6179 | post_event(advmap); | |
6180 | return transit< Reset >(); | |
6181 | } | |
6182 | ||
6183 | pg->adjust_need_up_thru(advmap.osdmap); | |
6184 | ||
6185 | return forward_event(); | |
6186 | } | |
6187 | ||
6188 | boost::statechart::result PG::RecoveryState::Peering::react(const QueryState& q) | |
6189 | { | |
6190 | PG *pg = context< RecoveryMachine >().pg; | |
6191 | ||
6192 | q.f->open_object_section("state"); | |
6193 | q.f->dump_string("name", state_name); | |
6194 | q.f->dump_stream("enter_time") << enter_time; | |
6195 | ||
6196 | q.f->open_array_section("past_intervals"); | |
6197 | pg->past_intervals.dump(q.f); | |
6198 | q.f->close_section(); | |
6199 | ||
6200 | q.f->open_array_section("probing_osds"); | |
6201 | for (set<pg_shard_t>::iterator p = prior_set.probe.begin(); | |
6202 | p != prior_set.probe.end(); | |
6203 | ++p) | |
6204 | q.f->dump_stream("osd") << *p; | |
6205 | q.f->close_section(); | |
6206 | ||
6207 | if (prior_set.pg_down) | |
6208 | q.f->dump_string("blocked", "peering is blocked due to down osds"); | |
6209 | ||
6210 | q.f->open_array_section("down_osds_we_would_probe"); | |
6211 | for (set<int>::iterator p = prior_set.down.begin(); | |
6212 | p != prior_set.down.end(); | |
6213 | ++p) | |
6214 | q.f->dump_int("osd", *p); | |
6215 | q.f->close_section(); | |
6216 | ||
6217 | q.f->open_array_section("peering_blocked_by"); | |
6218 | for (map<int,epoch_t>::iterator p = prior_set.blocked_by.begin(); | |
6219 | p != prior_set.blocked_by.end(); | |
6220 | ++p) { | |
6221 | q.f->open_object_section("osd"); | |
6222 | q.f->dump_int("osd", p->first); | |
6223 | q.f->dump_int("current_lost_at", p->second); | |
6224 | q.f->dump_string("comment", "starting or marking this osd lost may let us proceed"); | |
6225 | q.f->close_section(); | |
6226 | } | |
6227 | q.f->close_section(); | |
6228 | ||
6229 | if (history_les_bound) { | |
6230 | q.f->open_array_section("peering_blocked_by_detail"); | |
6231 | q.f->open_object_section("item"); | |
6232 | q.f->dump_string("detail","peering_blocked_by_history_les_bound"); | |
6233 | q.f->close_section(); | |
6234 | q.f->close_section(); | |
6235 | } | |
6236 | ||
6237 | q.f->close_section(); | |
6238 | return forward_event(); | |
6239 | } | |
6240 | ||
6241 | void PG::RecoveryState::Peering::exit() | |
6242 | { | |
6243 | PG *pg = context< RecoveryMachine >().pg; | |
6244 | ldout(pg->cct, 10) << "Leaving Peering" << dendl; | |
6245 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6246 | pg->state_clear(PG_STATE_PEERING); | |
6247 | pg->clear_probe_targets(); | |
6248 | ||
6249 | utime_t dur = ceph_clock_now() - enter_time; | |
6250 | pg->osd->recoverystate_perf->tinc(rs_peering_latency, dur); | |
6251 | } | |
6252 | ||
6253 | ||
6254 | /*------Backfilling-------*/ | |
6255 | PG::RecoveryState::Backfilling::Backfilling(my_context ctx) | |
6256 | : my_base(ctx), | |
6257 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/Backfilling") | |
6258 | { | |
6259 | context< RecoveryMachine >().log_enter(state_name); | |
6260 | PG *pg = context< RecoveryMachine >().pg; | |
6261 | pg->backfill_reserved = true; | |
6262 | pg->queue_recovery(); | |
6263 | pg->state_clear(PG_STATE_BACKFILL_TOOFULL); | |
6264 | pg->state_clear(PG_STATE_BACKFILL_WAIT); | |
6265 | pg->state_set(PG_STATE_BACKFILL); | |
6266 | pg->publish_stats_to_osd(); | |
6267 | } | |
6268 | ||
6269 | boost::statechart::result | |
6270 | PG::RecoveryState::Backfilling::react(const RemoteReservationRejected &) | |
6271 | { | |
6272 | PG *pg = context< RecoveryMachine >().pg; | |
6273 | pg->osd->local_reserver.cancel_reservation(pg->info.pgid); | |
6274 | pg->state_set(PG_STATE_BACKFILL_TOOFULL); | |
6275 | ||
6276 | for (set<pg_shard_t>::iterator it = pg->backfill_targets.begin(); | |
6277 | it != pg->backfill_targets.end(); | |
6278 | ++it) { | |
6279 | assert(*it != pg->pg_whoami); | |
6280 | ConnectionRef con = pg->osd->get_con_osd_cluster( | |
6281 | it->osd, pg->get_osdmap()->get_epoch()); | |
6282 | if (con) { | |
6283 | pg->osd->send_message_osd_cluster( | |
6284 | new MBackfillReserve( | |
6285 | MBackfillReserve::REJECT, | |
6286 | spg_t(pg->info.pgid.pgid, it->shard), | |
6287 | pg->get_osdmap()->get_epoch()), | |
6288 | con.get()); | |
6289 | } | |
6290 | } | |
6291 | ||
6292 | pg->waiting_on_backfill.clear(); | |
6293 | pg->finish_recovery_op(hobject_t::get_max()); | |
6294 | ||
6295 | pg->schedule_backfill_full_retry(); | |
6296 | return transit<NotBackfilling>(); | |
6297 | } | |
6298 | ||
6299 | void PG::RecoveryState::Backfilling::exit() | |
6300 | { | |
6301 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6302 | PG *pg = context< RecoveryMachine >().pg; | |
6303 | pg->backfill_reserved = false; | |
6304 | pg->backfill_reserving = false; | |
6305 | pg->state_clear(PG_STATE_BACKFILL); | |
6306 | utime_t dur = ceph_clock_now() - enter_time; | |
6307 | pg->osd->recoverystate_perf->tinc(rs_backfilling_latency, dur); | |
6308 | } | |
6309 | ||
6310 | /*--WaitRemoteBackfillReserved--*/ | |
6311 | ||
6312 | PG::RecoveryState::WaitRemoteBackfillReserved::WaitRemoteBackfillReserved(my_context ctx) | |
6313 | : my_base(ctx), | |
6314 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/WaitRemoteBackfillReserved"), | |
6315 | backfill_osd_it(context< Active >().remote_shards_to_reserve_backfill.begin()) | |
6316 | { | |
6317 | context< RecoveryMachine >().log_enter(state_name); | |
6318 | PG *pg = context< RecoveryMachine >().pg; | |
6319 | pg->state_set(PG_STATE_BACKFILL_WAIT); | |
6320 | pg->publish_stats_to_osd(); | |
6321 | post_event(RemoteBackfillReserved()); | |
6322 | } | |
6323 | ||
6324 | boost::statechart::result | |
6325 | PG::RecoveryState::WaitRemoteBackfillReserved::react(const RemoteBackfillReserved &evt) | |
6326 | { | |
6327 | PG *pg = context< RecoveryMachine >().pg; | |
6328 | ||
6329 | if (backfill_osd_it != context< Active >().remote_shards_to_reserve_backfill.end()) { | |
6330 | //The primary never backfills itself | |
6331 | assert(*backfill_osd_it != pg->pg_whoami); | |
6332 | ConnectionRef con = pg->osd->get_con_osd_cluster( | |
6333 | backfill_osd_it->osd, pg->get_osdmap()->get_epoch()); | |
6334 | if (con) { | |
6335 | pg->osd->send_message_osd_cluster( | |
6336 | new MBackfillReserve( | |
6337 | MBackfillReserve::REQUEST, | |
6338 | spg_t(pg->info.pgid.pgid, backfill_osd_it->shard), | |
6339 | pg->get_osdmap()->get_epoch(), | |
6340 | pg->get_backfill_priority()), | |
6341 | con.get()); | |
6342 | } | |
6343 | ++backfill_osd_it; | |
6344 | } else { | |
6345 | post_event(AllBackfillsReserved()); | |
6346 | } | |
6347 | return discard_event(); | |
6348 | } | |
6349 | ||
6350 | void PG::RecoveryState::WaitRemoteBackfillReserved::exit() | |
6351 | { | |
6352 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6353 | PG *pg = context< RecoveryMachine >().pg; | |
6354 | utime_t dur = ceph_clock_now() - enter_time; | |
6355 | pg->osd->recoverystate_perf->tinc(rs_waitremotebackfillreserved_latency, dur); | |
6356 | } | |
6357 | ||
6358 | boost::statechart::result | |
6359 | PG::RecoveryState::WaitRemoteBackfillReserved::react(const RemoteReservationRejected &evt) | |
6360 | { | |
6361 | PG *pg = context< RecoveryMachine >().pg; | |
6362 | pg->osd->local_reserver.cancel_reservation(pg->info.pgid); | |
6363 | ||
6364 | // Send REJECT to all previously acquired reservations | |
6365 | set<pg_shard_t>::const_iterator it, begin, end, next; | |
6366 | begin = context< Active >().remote_shards_to_reserve_backfill.begin(); | |
6367 | end = context< Active >().remote_shards_to_reserve_backfill.end(); | |
6368 | assert(begin != end); | |
6369 | for (next = it = begin, ++next ; next != backfill_osd_it; ++it, ++next) { | |
6370 | //The primary never backfills itself | |
6371 | assert(*it != pg->pg_whoami); | |
6372 | ConnectionRef con = pg->osd->get_con_osd_cluster( | |
6373 | it->osd, pg->get_osdmap()->get_epoch()); | |
6374 | if (con) { | |
6375 | pg->osd->send_message_osd_cluster( | |
6376 | new MBackfillReserve( | |
6377 | MBackfillReserve::REJECT, | |
6378 | spg_t(pg->info.pgid.pgid, it->shard), | |
6379 | pg->get_osdmap()->get_epoch()), | |
6380 | con.get()); | |
6381 | } | |
6382 | } | |
6383 | ||
6384 | pg->state_clear(PG_STATE_BACKFILL_WAIT); | |
6385 | pg->state_set(PG_STATE_BACKFILL_TOOFULL); | |
6386 | pg->publish_stats_to_osd(); | |
6387 | ||
6388 | pg->schedule_backfill_full_retry(); | |
6389 | ||
6390 | return transit<NotBackfilling>(); | |
6391 | } | |
6392 | ||
6393 | /*--WaitLocalBackfillReserved--*/ | |
6394 | PG::RecoveryState::WaitLocalBackfillReserved::WaitLocalBackfillReserved(my_context ctx) | |
6395 | : my_base(ctx), | |
6396 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/WaitLocalBackfillReserved") | |
6397 | { | |
6398 | context< RecoveryMachine >().log_enter(state_name); | |
6399 | PG *pg = context< RecoveryMachine >().pg; | |
6400 | pg->state_set(PG_STATE_BACKFILL_WAIT); | |
6401 | pg->osd->local_reserver.request_reservation( | |
6402 | pg->info.pgid, | |
6403 | new QueuePeeringEvt<LocalBackfillReserved>( | |
6404 | pg, pg->get_osdmap()->get_epoch(), | |
6405 | LocalBackfillReserved()), | |
6406 | pg->get_backfill_priority()); | |
6407 | pg->publish_stats_to_osd(); | |
6408 | } | |
6409 | ||
6410 | void PG::RecoveryState::WaitLocalBackfillReserved::exit() | |
6411 | { | |
6412 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6413 | PG *pg = context< RecoveryMachine >().pg; | |
6414 | utime_t dur = ceph_clock_now() - enter_time; | |
6415 | pg->osd->recoverystate_perf->tinc(rs_waitlocalbackfillreserved_latency, dur); | |
6416 | } | |
6417 | ||
6418 | /*----NotBackfilling------*/ | |
6419 | PG::RecoveryState::NotBackfilling::NotBackfilling(my_context ctx) | |
6420 | : my_base(ctx), | |
6421 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/NotBackfilling") | |
6422 | { | |
6423 | context< RecoveryMachine >().log_enter(state_name); | |
6424 | PG *pg = context< RecoveryMachine >().pg; | |
6425 | pg->publish_stats_to_osd(); | |
6426 | } | |
6427 | ||
6428 | boost::statechart::result | |
6429 | PG::RecoveryState::NotBackfilling::react(const RemoteBackfillReserved &evt) | |
6430 | { | |
6431 | return discard_event(); | |
6432 | } | |
6433 | ||
6434 | boost::statechart::result | |
6435 | PG::RecoveryState::NotBackfilling::react(const RemoteReservationRejected &evt) | |
6436 | { | |
6437 | return discard_event(); | |
6438 | } | |
6439 | ||
6440 | void PG::RecoveryState::NotBackfilling::exit() | |
6441 | { | |
6442 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6443 | PG *pg = context< RecoveryMachine >().pg; | |
6444 | utime_t dur = ceph_clock_now() - enter_time; | |
6445 | pg->osd->recoverystate_perf->tinc(rs_notbackfilling_latency, dur); | |
6446 | } | |
6447 | ||
6448 | /*----NotRecovering------*/ | |
6449 | PG::RecoveryState::NotRecovering::NotRecovering(my_context ctx) | |
6450 | : my_base(ctx), | |
6451 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/NotRecovering") | |
6452 | { | |
6453 | context< RecoveryMachine >().log_enter(state_name); | |
6454 | PG *pg = context< RecoveryMachine >().pg; | |
6455 | pg->publish_stats_to_osd(); | |
6456 | } | |
6457 | ||
6458 | void PG::RecoveryState::NotRecovering::exit() | |
6459 | { | |
6460 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6461 | PG *pg = context< RecoveryMachine >().pg; | |
6462 | utime_t dur = ceph_clock_now() - enter_time; | |
6463 | pg->osd->recoverystate_perf->tinc(rs_notrecovering_latency, dur); | |
6464 | } | |
6465 | ||
6466 | /*---RepNotRecovering----*/ | |
6467 | PG::RecoveryState::RepNotRecovering::RepNotRecovering(my_context ctx) | |
6468 | : my_base(ctx), | |
6469 | NamedState(context< RecoveryMachine >().pg, "Started/ReplicaActive/RepNotRecovering") | |
6470 | { | |
6471 | context< RecoveryMachine >().log_enter(state_name); | |
6472 | } | |
6473 | ||
6474 | void PG::RecoveryState::RepNotRecovering::exit() | |
6475 | { | |
6476 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6477 | PG *pg = context< RecoveryMachine >().pg; | |
6478 | utime_t dur = ceph_clock_now() - enter_time; | |
6479 | pg->osd->recoverystate_perf->tinc(rs_repnotrecovering_latency, dur); | |
6480 | } | |
6481 | ||
6482 | /*---RepWaitRecoveryReserved--*/ | |
6483 | PG::RecoveryState::RepWaitRecoveryReserved::RepWaitRecoveryReserved(my_context ctx) | |
6484 | : my_base(ctx), | |
6485 | NamedState(context< RecoveryMachine >().pg, "Started/ReplicaActive/RepWaitRecoveryReserved") | |
6486 | { | |
6487 | context< RecoveryMachine >().log_enter(state_name); | |
6488 | PG *pg = context< RecoveryMachine >().pg; | |
6489 | ||
6490 | pg->osd->remote_reserver.request_reservation( | |
6491 | pg->info.pgid, | |
6492 | new QueuePeeringEvt<RemoteRecoveryReserved>( | |
6493 | pg, pg->get_osdmap()->get_epoch(), | |
6494 | RemoteRecoveryReserved()), | |
6495 | pg->get_recovery_priority()); | |
6496 | } | |
6497 | ||
6498 | boost::statechart::result | |
6499 | PG::RecoveryState::RepWaitRecoveryReserved::react(const RemoteRecoveryReserved &evt) | |
6500 | { | |
6501 | PG *pg = context< RecoveryMachine >().pg; | |
6502 | pg->osd->send_message_osd_cluster( | |
6503 | pg->primary.osd, | |
6504 | new MRecoveryReserve( | |
6505 | MRecoveryReserve::GRANT, | |
6506 | spg_t(pg->info.pgid.pgid, pg->primary.shard), | |
6507 | pg->get_osdmap()->get_epoch()), | |
6508 | pg->get_osdmap()->get_epoch()); | |
6509 | return transit<RepRecovering>(); | |
6510 | } | |
6511 | ||
6512 | void PG::RecoveryState::RepWaitRecoveryReserved::exit() | |
6513 | { | |
6514 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6515 | PG *pg = context< RecoveryMachine >().pg; | |
6516 | utime_t dur = ceph_clock_now() - enter_time; | |
6517 | pg->osd->recoverystate_perf->tinc(rs_repwaitrecoveryreserved_latency, dur); | |
6518 | } | |
6519 | ||
6520 | /*-RepWaitBackfillReserved*/ | |
6521 | PG::RecoveryState::RepWaitBackfillReserved::RepWaitBackfillReserved(my_context ctx) | |
6522 | : my_base(ctx), | |
6523 | NamedState(context< RecoveryMachine >().pg, "Started/ReplicaActive/RepWaitBackfillReserved") | |
6524 | { | |
6525 | context< RecoveryMachine >().log_enter(state_name); | |
6526 | } | |
6527 | ||
6528 | boost::statechart::result | |
6529 | PG::RecoveryState::RepNotRecovering::react(const RequestBackfillPrio &evt) | |
6530 | { | |
6531 | PG *pg = context< RecoveryMachine >().pg; | |
6532 | ostringstream ss; | |
6533 | ||
6534 | if (pg->cct->_conf->osd_debug_reject_backfill_probability > 0 && | |
6535 | (rand()%1000 < (pg->cct->_conf->osd_debug_reject_backfill_probability*1000.0))) { | |
6536 | ldout(pg->cct, 10) << "backfill reservation rejected: failure injection" | |
6537 | << dendl; | |
6538 | post_event(RemoteReservationRejected()); | |
6539 | } else if (!pg->cct->_conf->osd_debug_skip_full_check_in_backfill_reservation && | |
6540 | pg->osd->check_backfill_full(ss)) { | |
6541 | ldout(pg->cct, 10) << "backfill reservation rejected: " | |
6542 | << ss.str() << dendl; | |
6543 | post_event(RemoteReservationRejected()); | |
6544 | } else { | |
6545 | pg->osd->remote_reserver.request_reservation( | |
6546 | pg->info.pgid, | |
6547 | new QueuePeeringEvt<RemoteBackfillReserved>( | |
6548 | pg, pg->get_osdmap()->get_epoch(), | |
6549 | RemoteBackfillReserved()), evt.priority); | |
6550 | } | |
6551 | return transit<RepWaitBackfillReserved>(); | |
6552 | } | |
6553 | ||
6554 | void PG::RecoveryState::RepWaitBackfillReserved::exit() | |
6555 | { | |
6556 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6557 | PG *pg = context< RecoveryMachine >().pg; | |
6558 | utime_t dur = ceph_clock_now() - enter_time; | |
6559 | pg->osd->recoverystate_perf->tinc(rs_repwaitbackfillreserved_latency, dur); | |
6560 | } | |
6561 | ||
6562 | boost::statechart::result | |
6563 | PG::RecoveryState::RepWaitBackfillReserved::react(const RemoteBackfillReserved &evt) | |
6564 | { | |
6565 | PG *pg = context< RecoveryMachine >().pg; | |
6566 | ||
6567 | ostringstream ss; | |
6568 | if (pg->cct->_conf->osd_debug_reject_backfill_probability > 0 && | |
6569 | (rand()%1000 < (pg->cct->_conf->osd_debug_reject_backfill_probability*1000.0))) { | |
6570 | ldout(pg->cct, 10) << "backfill reservation rejected after reservation: " | |
6571 | << "failure injection" << dendl; | |
6572 | pg->osd->remote_reserver.cancel_reservation(pg->info.pgid); | |
6573 | post_event(RemoteReservationRejected()); | |
6574 | return discard_event(); | |
6575 | } else if (!pg->cct->_conf->osd_debug_skip_full_check_in_backfill_reservation && | |
6576 | pg->osd->check_backfill_full(ss)) { | |
6577 | ldout(pg->cct, 10) << "backfill reservation rejected after reservation: " | |
6578 | << ss.str() << dendl; | |
6579 | pg->osd->remote_reserver.cancel_reservation(pg->info.pgid); | |
6580 | post_event(RemoteReservationRejected()); | |
6581 | return discard_event(); | |
6582 | } else { | |
6583 | pg->osd->send_message_osd_cluster( | |
6584 | pg->primary.osd, | |
6585 | new MBackfillReserve( | |
6586 | MBackfillReserve::GRANT, | |
6587 | spg_t(pg->info.pgid.pgid, pg->primary.shard), | |
6588 | pg->get_osdmap()->get_epoch()), | |
6589 | pg->get_osdmap()->get_epoch()); | |
6590 | return transit<RepRecovering>(); | |
6591 | } | |
6592 | } | |
6593 | ||
6594 | boost::statechart::result | |
6595 | PG::RecoveryState::RepWaitBackfillReserved::react(const RemoteReservationRejected &evt) | |
6596 | { | |
6597 | PG *pg = context< RecoveryMachine >().pg; | |
6598 | pg->reject_reservation(); | |
6599 | return transit<RepNotRecovering>(); | |
6600 | } | |
6601 | ||
6602 | /*---RepRecovering-------*/ | |
6603 | PG::RecoveryState::RepRecovering::RepRecovering(my_context ctx) | |
6604 | : my_base(ctx), | |
6605 | NamedState(context< RecoveryMachine >().pg, "Started/ReplicaActive/RepRecovering") | |
6606 | { | |
6607 | context< RecoveryMachine >().log_enter(state_name); | |
6608 | } | |
6609 | ||
6610 | boost::statechart::result | |
6611 | PG::RecoveryState::RepRecovering::react(const BackfillTooFull &) | |
6612 | { | |
6613 | PG *pg = context< RecoveryMachine >().pg; | |
6614 | pg->reject_reservation(); | |
6615 | return discard_event(); | |
6616 | } | |
6617 | ||
6618 | void PG::RecoveryState::RepRecovering::exit() | |
6619 | { | |
6620 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6621 | PG *pg = context< RecoveryMachine >().pg; | |
6622 | pg->osd->remote_reserver.cancel_reservation(pg->info.pgid); | |
6623 | utime_t dur = ceph_clock_now() - enter_time; | |
6624 | pg->osd->recoverystate_perf->tinc(rs_reprecovering_latency, dur); | |
6625 | } | |
6626 | ||
6627 | /*------Activating--------*/ | |
6628 | PG::RecoveryState::Activating::Activating(my_context ctx) | |
6629 | : my_base(ctx), | |
6630 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/Activating") | |
6631 | { | |
6632 | context< RecoveryMachine >().log_enter(state_name); | |
6633 | } | |
6634 | ||
6635 | void PG::RecoveryState::Activating::exit() | |
6636 | { | |
6637 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6638 | PG *pg = context< RecoveryMachine >().pg; | |
6639 | utime_t dur = ceph_clock_now() - enter_time; | |
6640 | pg->osd->recoverystate_perf->tinc(rs_activating_latency, dur); | |
6641 | } | |
6642 | ||
6643 | PG::RecoveryState::WaitLocalRecoveryReserved::WaitLocalRecoveryReserved(my_context ctx) | |
6644 | : my_base(ctx), | |
6645 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/WaitLocalRecoveryReserved") | |
6646 | { | |
6647 | context< RecoveryMachine >().log_enter(state_name); | |
6648 | PG *pg = context< RecoveryMachine >().pg; | |
6649 | ||
6650 | // Make sure all nodes that part of the recovery aren't full | |
6651 | if (!pg->cct->_conf->osd_debug_skip_full_check_in_recovery && | |
6652 | pg->osd->check_osdmap_full(pg->actingbackfill)) { | |
6653 | post_event(RecoveryTooFull()); | |
6654 | return; | |
6655 | } | |
6656 | ||
6657 | pg->state_clear(PG_STATE_RECOVERY_TOOFULL); | |
6658 | pg->state_set(PG_STATE_RECOVERY_WAIT); | |
6659 | pg->osd->local_reserver.request_reservation( | |
6660 | pg->info.pgid, | |
6661 | new QueuePeeringEvt<LocalRecoveryReserved>( | |
6662 | pg, pg->get_osdmap()->get_epoch(), | |
6663 | LocalRecoveryReserved()), | |
6664 | pg->get_recovery_priority()); | |
6665 | pg->publish_stats_to_osd(); | |
6666 | } | |
6667 | ||
6668 | boost::statechart::result | |
6669 | PG::RecoveryState::WaitLocalRecoveryReserved::react(const RecoveryTooFull &evt) | |
6670 | { | |
6671 | PG *pg = context< RecoveryMachine >().pg; | |
6672 | pg->state_set(PG_STATE_RECOVERY_TOOFULL); | |
6673 | pg->schedule_recovery_full_retry(); | |
6674 | return transit<NotRecovering>(); | |
6675 | } | |
6676 | ||
6677 | void PG::RecoveryState::WaitLocalRecoveryReserved::exit() | |
6678 | { | |
6679 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6680 | PG *pg = context< RecoveryMachine >().pg; | |
6681 | utime_t dur = ceph_clock_now() - enter_time; | |
6682 | pg->osd->recoverystate_perf->tinc(rs_waitlocalrecoveryreserved_latency, dur); | |
6683 | } | |
6684 | ||
6685 | PG::RecoveryState::WaitRemoteRecoveryReserved::WaitRemoteRecoveryReserved(my_context ctx) | |
6686 | : my_base(ctx), | |
6687 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/WaitRemoteRecoveryReserved"), | |
6688 | remote_recovery_reservation_it(context< Active >().remote_shards_to_reserve_recovery.begin()) | |
6689 | { | |
6690 | context< RecoveryMachine >().log_enter(state_name); | |
6691 | post_event(RemoteRecoveryReserved()); | |
6692 | } | |
6693 | ||
6694 | boost::statechart::result | |
6695 | PG::RecoveryState::WaitRemoteRecoveryReserved::react(const RemoteRecoveryReserved &evt) { | |
6696 | PG *pg = context< RecoveryMachine >().pg; | |
6697 | ||
6698 | if (remote_recovery_reservation_it != context< Active >().remote_shards_to_reserve_recovery.end()) { | |
6699 | assert(*remote_recovery_reservation_it != pg->pg_whoami); | |
6700 | ConnectionRef con = pg->osd->get_con_osd_cluster( | |
6701 | remote_recovery_reservation_it->osd, pg->get_osdmap()->get_epoch()); | |
6702 | if (con) { | |
6703 | pg->osd->send_message_osd_cluster( | |
6704 | new MRecoveryReserve( | |
6705 | MRecoveryReserve::REQUEST, | |
6706 | spg_t(pg->info.pgid.pgid, remote_recovery_reservation_it->shard), | |
6707 | pg->get_osdmap()->get_epoch()), | |
6708 | con.get()); | |
6709 | } | |
6710 | ++remote_recovery_reservation_it; | |
6711 | } else { | |
6712 | post_event(AllRemotesReserved()); | |
6713 | } | |
6714 | return discard_event(); | |
6715 | } | |
6716 | ||
6717 | void PG::RecoveryState::WaitRemoteRecoveryReserved::exit() | |
6718 | { | |
6719 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6720 | PG *pg = context< RecoveryMachine >().pg; | |
6721 | utime_t dur = ceph_clock_now() - enter_time; | |
6722 | pg->osd->recoverystate_perf->tinc(rs_waitremoterecoveryreserved_latency, dur); | |
6723 | } | |
6724 | ||
6725 | PG::RecoveryState::Recovering::Recovering(my_context ctx) | |
6726 | : my_base(ctx), | |
6727 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/Recovering") | |
6728 | { | |
6729 | context< RecoveryMachine >().log_enter(state_name); | |
6730 | ||
6731 | PG *pg = context< RecoveryMachine >().pg; | |
6732 | pg->state_clear(PG_STATE_RECOVERY_WAIT); | |
6733 | pg->state_clear(PG_STATE_RECOVERY_TOOFULL); | |
6734 | pg->state_set(PG_STATE_RECOVERING); | |
6735 | pg->publish_stats_to_osd(); | |
6736 | pg->queue_recovery(); | |
6737 | } | |
6738 | ||
6739 | void PG::RecoveryState::Recovering::release_reservations() | |
6740 | { | |
6741 | PG *pg = context< RecoveryMachine >().pg; | |
6742 | assert(!pg->pg_log.get_missing().have_missing()); | |
6743 | ||
6744 | // release remote reservations | |
6745 | for (set<pg_shard_t>::const_iterator i = | |
6746 | context< Active >().remote_shards_to_reserve_recovery.begin(); | |
6747 | i != context< Active >().remote_shards_to_reserve_recovery.end(); | |
6748 | ++i) { | |
6749 | if (*i == pg->pg_whoami) // skip myself | |
6750 | continue; | |
6751 | ConnectionRef con = pg->osd->get_con_osd_cluster( | |
6752 | i->osd, pg->get_osdmap()->get_epoch()); | |
6753 | if (con) { | |
6754 | pg->osd->send_message_osd_cluster( | |
6755 | new MRecoveryReserve( | |
6756 | MRecoveryReserve::RELEASE, | |
6757 | spg_t(pg->info.pgid.pgid, i->shard), | |
6758 | pg->get_osdmap()->get_epoch()), | |
6759 | con.get()); | |
6760 | } | |
6761 | } | |
6762 | } | |
6763 | ||
6764 | boost::statechart::result | |
6765 | PG::RecoveryState::Recovering::react(const AllReplicasRecovered &evt) | |
6766 | { | |
6767 | PG *pg = context< RecoveryMachine >().pg; | |
6768 | pg->state_clear(PG_STATE_RECOVERING); | |
6769 | release_reservations(); | |
6770 | return transit<Recovered>(); | |
6771 | } | |
6772 | ||
6773 | boost::statechart::result | |
6774 | PG::RecoveryState::Recovering::react(const RequestBackfill &evt) | |
6775 | { | |
6776 | PG *pg = context< RecoveryMachine >().pg; | |
6777 | pg->state_clear(PG_STATE_RECOVERING); | |
6778 | release_reservations(); | |
6779 | return transit<WaitRemoteBackfillReserved>(); | |
6780 | } | |
6781 | ||
6782 | void PG::RecoveryState::Recovering::exit() | |
6783 | { | |
6784 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6785 | PG *pg = context< RecoveryMachine >().pg; | |
6786 | utime_t dur = ceph_clock_now() - enter_time; | |
6787 | pg->osd->recoverystate_perf->tinc(rs_recovering_latency, dur); | |
6788 | } | |
6789 | ||
6790 | PG::RecoveryState::Recovered::Recovered(my_context ctx) | |
6791 | : my_base(ctx), | |
6792 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/Recovered") | |
6793 | { | |
6794 | pg_shard_t auth_log_shard; | |
6795 | ||
6796 | context< RecoveryMachine >().log_enter(state_name); | |
6797 | ||
6798 | PG *pg = context< RecoveryMachine >().pg; | |
6799 | pg->osd->local_reserver.cancel_reservation(pg->info.pgid); | |
6800 | ||
6801 | assert(!pg->needs_recovery()); | |
6802 | ||
6803 | // if we finished backfill, all acting are active; recheck if | |
6804 | // DEGRADED | UNDERSIZED is appropriate. | |
6805 | assert(!pg->actingbackfill.empty()); | |
6806 | if (pg->get_osdmap()->get_pg_size(pg->info.pgid.pgid) <= | |
6807 | pg->actingbackfill.size()) { | |
6808 | pg->state_clear(PG_STATE_DEGRADED); | |
6809 | pg->publish_stats_to_osd(); | |
6810 | } | |
6811 | ||
6812 | // trim pglog on recovered | |
6813 | pg->trim_log(); | |
6814 | ||
6815 | // adjust acting set? (e.g. because backfill completed...) | |
6816 | bool history_les_bound = false; | |
6817 | if (pg->acting != pg->up && !pg->choose_acting(auth_log_shard, | |
6818 | true, &history_les_bound)) | |
6819 | assert(pg->want_acting.size()); | |
6820 | ||
6821 | if (context< Active >().all_replicas_activated) | |
6822 | post_event(GoClean()); | |
6823 | } | |
6824 | ||
6825 | void PG::RecoveryState::Recovered::exit() | |
6826 | { | |
6827 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6828 | PG *pg = context< RecoveryMachine >().pg; | |
6829 | utime_t dur = ceph_clock_now() - enter_time; | |
6830 | pg->osd->recoverystate_perf->tinc(rs_recovered_latency, dur); | |
6831 | } | |
6832 | ||
6833 | PG::RecoveryState::Clean::Clean(my_context ctx) | |
6834 | : my_base(ctx), | |
6835 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active/Clean") | |
6836 | { | |
6837 | context< RecoveryMachine >().log_enter(state_name); | |
6838 | ||
6839 | PG *pg = context< RecoveryMachine >().pg; | |
6840 | ||
6841 | if (pg->info.last_complete != pg->info.last_update) { | |
6842 | ceph_abort(); | |
6843 | } | |
6844 | pg->finish_recovery(*context< RecoveryMachine >().get_on_safe_context_list()); | |
6845 | pg->mark_clean(); | |
6846 | ||
6847 | pg->share_pg_info(); | |
6848 | pg->publish_stats_to_osd(); | |
6849 | ||
6850 | } | |
6851 | ||
6852 | void PG::RecoveryState::Clean::exit() | |
6853 | { | |
6854 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
6855 | PG *pg = context< RecoveryMachine >().pg; | |
6856 | pg->state_clear(PG_STATE_CLEAN); | |
6857 | utime_t dur = ceph_clock_now() - enter_time; | |
6858 | pg->osd->recoverystate_perf->tinc(rs_clean_latency, dur); | |
6859 | } | |
6860 | ||
6861 | template <typename T> | |
6862 | set<pg_shard_t> unique_osd_shard_set(const pg_shard_t & skip, const T &in) | |
6863 | { | |
6864 | set<int> osds_found; | |
6865 | set<pg_shard_t> out; | |
6866 | for (typename T::const_iterator i = in.begin(); | |
6867 | i != in.end(); | |
6868 | ++i) { | |
6869 | if (*i != skip && !osds_found.count(i->osd)) { | |
6870 | osds_found.insert(i->osd); | |
6871 | out.insert(*i); | |
6872 | } | |
6873 | } | |
6874 | return out; | |
6875 | } | |
6876 | ||
6877 | /*---------Active---------*/ | |
6878 | PG::RecoveryState::Active::Active(my_context ctx) | |
6879 | : my_base(ctx), | |
6880 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Active"), | |
6881 | remote_shards_to_reserve_recovery( | |
6882 | unique_osd_shard_set( | |
6883 | context< RecoveryMachine >().pg->pg_whoami, | |
6884 | context< RecoveryMachine >().pg->actingbackfill)), | |
6885 | remote_shards_to_reserve_backfill( | |
6886 | unique_osd_shard_set( | |
6887 | context< RecoveryMachine >().pg->pg_whoami, | |
6888 | context< RecoveryMachine >().pg->backfill_targets)), | |
6889 | all_replicas_activated(false) | |
6890 | { | |
6891 | context< RecoveryMachine >().log_enter(state_name); | |
6892 | ||
6893 | PG *pg = context< RecoveryMachine >().pg; | |
6894 | ||
6895 | assert(!pg->backfill_reserving); | |
6896 | assert(!pg->backfill_reserved); | |
6897 | assert(pg->is_primary()); | |
6898 | ldout(pg->cct, 10) << "In Active, about to call activate" << dendl; | |
6899 | pg->start_flush( | |
6900 | context< RecoveryMachine >().get_cur_transaction(), | |
6901 | context< RecoveryMachine >().get_on_applied_context_list(), | |
6902 | context< RecoveryMachine >().get_on_safe_context_list()); | |
6903 | pg->activate(*context< RecoveryMachine >().get_cur_transaction(), | |
6904 | pg->get_osdmap()->get_epoch(), | |
6905 | *context< RecoveryMachine >().get_on_safe_context_list(), | |
6906 | *context< RecoveryMachine >().get_query_map(), | |
6907 | context< RecoveryMachine >().get_info_map(), | |
6908 | context< RecoveryMachine >().get_recovery_ctx()); | |
6909 | ||
6910 | // everyone has to commit/ack before we are truly active | |
6911 | pg->blocked_by.clear(); | |
6912 | for (set<pg_shard_t>::iterator p = pg->actingbackfill.begin(); | |
6913 | p != pg->actingbackfill.end(); | |
6914 | ++p) { | |
6915 | if (p->shard != pg->pg_whoami.shard) { | |
6916 | pg->blocked_by.insert(p->shard); | |
6917 | } | |
6918 | } | |
6919 | pg->publish_stats_to_osd(); | |
6920 | ldout(pg->cct, 10) << "Activate Finished" << dendl; | |
6921 | } | |
6922 | ||
6923 | boost::statechart::result PG::RecoveryState::Active::react(const AdvMap& advmap) | |
6924 | { | |
6925 | PG *pg = context< RecoveryMachine >().pg; | |
6926 | ldout(pg->cct, 10) << "Active advmap" << dendl; | |
6927 | if (!pg->pool.newly_removed_snaps.empty()) { | |
6928 | pg->snap_trimq.union_of(pg->pool.newly_removed_snaps); | |
6929 | ldout(pg->cct, 10) << *pg << " snap_trimq now " << pg->snap_trimq << dendl; | |
6930 | pg->dirty_info = true; | |
6931 | pg->dirty_big_info = true; | |
6932 | } | |
6933 | ||
6934 | for (size_t i = 0; i < pg->want_acting.size(); i++) { | |
6935 | int osd = pg->want_acting[i]; | |
6936 | if (!advmap.osdmap->is_up(osd)) { | |
6937 | pg_shard_t osd_with_shard(osd, shard_id_t(i)); | |
6938 | assert(pg->is_acting(osd_with_shard) || pg->is_up(osd_with_shard)); | |
6939 | } | |
6940 | } | |
6941 | ||
6942 | bool need_publish = false; | |
6943 | /* Check for changes in pool size (if the acting set changed as a result, | |
6944 | * this does not matter) */ | |
6945 | if (advmap.lastmap->get_pg_size(pg->info.pgid.pgid) != | |
6946 | pg->get_osdmap()->get_pg_size(pg->info.pgid.pgid)) { | |
6947 | if (pg->get_osdmap()->get_pg_size(pg->info.pgid.pgid) <= pg->actingset.size()) { | |
6948 | pg->state_clear(PG_STATE_UNDERSIZED); | |
6949 | if (pg->needs_recovery()) { | |
6950 | pg->state_set(PG_STATE_DEGRADED); | |
6951 | } else { | |
6952 | pg->state_clear(PG_STATE_DEGRADED); | |
6953 | } | |
6954 | } else { | |
6955 | pg->state_set(PG_STATE_UNDERSIZED); | |
6956 | pg->state_set(PG_STATE_DEGRADED); | |
6957 | } | |
6958 | need_publish = true; // degraded may have changed | |
6959 | } | |
6960 | ||
6961 | // if we haven't reported our PG stats in a long time, do so now. | |
6962 | if (pg->info.stats.reported_epoch + pg->cct->_conf->osd_pg_stat_report_interval_max < advmap.osdmap->get_epoch()) { | |
6963 | ldout(pg->cct, 20) << "reporting stats to osd after " << (advmap.osdmap->get_epoch() - pg->info.stats.reported_epoch) | |
6964 | << " epochs" << dendl; | |
6965 | need_publish = true; | |
6966 | } | |
6967 | ||
6968 | if (need_publish) | |
6969 | pg->publish_stats_to_osd(); | |
6970 | ||
6971 | return forward_event(); | |
6972 | } | |
6973 | ||
6974 | boost::statechart::result PG::RecoveryState::Active::react(const ActMap&) | |
6975 | { | |
6976 | PG *pg = context< RecoveryMachine >().pg; | |
6977 | ldout(pg->cct, 10) << "Active: handling ActMap" << dendl; | |
6978 | assert(pg->is_primary()); | |
6979 | ||
6980 | if (pg->have_unfound()) { | |
6981 | // object may have become unfound | |
6982 | pg->discover_all_missing(*context< RecoveryMachine >().get_query_map()); | |
6983 | } | |
6984 | ||
6985 | if (pg->cct->_conf->osd_check_for_log_corruption) | |
6986 | pg->check_log_for_corruption(pg->osd->store); | |
6987 | ||
6988 | uint64_t unfound = pg->missing_loc.num_unfound(); | |
6989 | if (unfound > 0 && | |
6990 | pg->all_unfound_are_queried_or_lost(pg->get_osdmap())) { | |
6991 | if (pg->cct->_conf->osd_auto_mark_unfound_lost) { | |
6992 | pg->osd->clog->error() << pg->info.pgid.pgid << " has " << unfound | |
6993 | << " objects unfound and apparently lost, would automatically marking lost but NOT IMPLEMENTED"; | |
6994 | } else | |
6995 | pg->osd->clog->error() << pg->info.pgid.pgid << " has " << unfound << " objects unfound and apparently lost"; | |
6996 | } | |
6997 | ||
6998 | if (pg->is_active()) { | |
6999 | ldout(pg->cct, 10) << "Active: kicking snap trim" << dendl; | |
7000 | pg->kick_snap_trim(); | |
7001 | } | |
7002 | ||
7003 | if (pg->is_peered() && | |
7004 | !pg->is_clean() && | |
7005 | !pg->get_osdmap()->test_flag(CEPH_OSDMAP_NOBACKFILL) && | |
7006 | (!pg->get_osdmap()->test_flag(CEPH_OSDMAP_NOREBALANCE) || pg->is_degraded())) { | |
7007 | pg->queue_recovery(); | |
7008 | } | |
7009 | return forward_event(); | |
7010 | } | |
7011 | ||
7012 | boost::statechart::result PG::RecoveryState::Active::react(const MNotifyRec& notevt) | |
7013 | { | |
7014 | PG *pg = context< RecoveryMachine >().pg; | |
7015 | assert(pg->is_primary()); | |
7016 | if (pg->peer_info.count(notevt.from)) { | |
7017 | ldout(pg->cct, 10) << "Active: got notify from " << notevt.from | |
7018 | << ", already have info from that osd, ignoring" | |
7019 | << dendl; | |
7020 | } else if (pg->peer_purged.count(notevt.from)) { | |
7021 | ldout(pg->cct, 10) << "Active: got notify from " << notevt.from | |
7022 | << ", already purged that peer, ignoring" | |
7023 | << dendl; | |
7024 | } else { | |
7025 | ldout(pg->cct, 10) << "Active: got notify from " << notevt.from | |
7026 | << ", calling proc_replica_info and discover_all_missing" | |
7027 | << dendl; | |
7028 | pg->proc_replica_info( | |
7029 | notevt.from, notevt.notify.info, notevt.notify.epoch_sent); | |
7030 | if (pg->have_unfound()) { | |
7031 | pg->discover_all_missing(*context< RecoveryMachine >().get_query_map()); | |
7032 | } | |
7033 | } | |
7034 | return discard_event(); | |
7035 | } | |
7036 | ||
7037 | boost::statechart::result PG::RecoveryState::Active::react(const MInfoRec& infoevt) | |
7038 | { | |
7039 | PG *pg = context< RecoveryMachine >().pg; | |
7040 | assert(pg->is_primary()); | |
7041 | ||
7042 | assert(!pg->actingbackfill.empty()); | |
7043 | // don't update history (yet) if we are active and primary; the replica | |
7044 | // may be telling us they have activated (and committed) but we can't | |
7045 | // share that until _everyone_ does the same. | |
7046 | if (pg->is_actingbackfill(infoevt.from)) { | |
7047 | ldout(pg->cct, 10) << " peer osd." << infoevt.from | |
7048 | << " activated and committed" << dendl; | |
7049 | pg->peer_activated.insert(infoevt.from); | |
7050 | pg->blocked_by.erase(infoevt.from.shard); | |
7051 | pg->publish_stats_to_osd(); | |
7052 | if (pg->peer_activated.size() == pg->actingbackfill.size()) { | |
7053 | pg->all_activated_and_committed(); | |
7054 | } | |
7055 | } | |
7056 | return discard_event(); | |
7057 | } | |
7058 | ||
7059 | boost::statechart::result PG::RecoveryState::Active::react(const MLogRec& logevt) | |
7060 | { | |
7061 | PG *pg = context< RecoveryMachine >().pg; | |
7062 | ldout(pg->cct, 10) << "searching osd." << logevt.from | |
7063 | << " log for unfound items" << dendl; | |
7064 | pg->proc_replica_log( | |
7065 | logevt.msg->info, logevt.msg->log, logevt.msg->missing, logevt.from); | |
7066 | bool got_missing = pg->search_for_missing( | |
7067 | pg->peer_info[logevt.from], | |
7068 | pg->peer_missing[logevt.from], | |
7069 | logevt.from, | |
7070 | context< RecoveryMachine >().get_recovery_ctx()); | |
7071 | if (pg->is_peered() && | |
7072 | got_missing) | |
7073 | pg->queue_recovery(); | |
7074 | return discard_event(); | |
7075 | } | |
7076 | ||
7077 | boost::statechart::result PG::RecoveryState::Active::react(const QueryState& q) | |
7078 | { | |
7079 | PG *pg = context< RecoveryMachine >().pg; | |
7080 | ||
7081 | q.f->open_object_section("state"); | |
7082 | q.f->dump_string("name", state_name); | |
7083 | q.f->dump_stream("enter_time") << enter_time; | |
7084 | ||
7085 | { | |
7086 | q.f->open_array_section("might_have_unfound"); | |
7087 | for (set<pg_shard_t>::iterator p = pg->might_have_unfound.begin(); | |
7088 | p != pg->might_have_unfound.end(); | |
7089 | ++p) { | |
7090 | q.f->open_object_section("osd"); | |
7091 | q.f->dump_stream("osd") << *p; | |
7092 | if (pg->peer_missing.count(*p)) { | |
7093 | q.f->dump_string("status", "already probed"); | |
7094 | } else if (pg->peer_missing_requested.count(*p)) { | |
7095 | q.f->dump_string("status", "querying"); | |
7096 | } else if (!pg->get_osdmap()->is_up(p->osd)) { | |
7097 | q.f->dump_string("status", "osd is down"); | |
7098 | } else { | |
7099 | q.f->dump_string("status", "not queried"); | |
7100 | } | |
7101 | q.f->close_section(); | |
7102 | } | |
7103 | q.f->close_section(); | |
7104 | } | |
7105 | { | |
7106 | q.f->open_object_section("recovery_progress"); | |
7107 | pg->dump_recovery_info(q.f); | |
7108 | q.f->close_section(); | |
7109 | } | |
7110 | ||
7111 | { | |
7112 | q.f->open_object_section("scrub"); | |
7113 | q.f->dump_stream("scrubber.epoch_start") << pg->scrubber.epoch_start; | |
7114 | q.f->dump_bool("scrubber.active", pg->scrubber.active); | |
7115 | q.f->dump_string("scrubber.state", Scrubber::state_string(pg->scrubber.state)); | |
7116 | q.f->dump_stream("scrubber.start") << pg->scrubber.start; | |
7117 | q.f->dump_stream("scrubber.end") << pg->scrubber.end; | |
7118 | q.f->dump_stream("scrubber.subset_last_update") << pg->scrubber.subset_last_update; | |
7119 | q.f->dump_bool("scrubber.deep", pg->scrubber.deep); | |
7120 | q.f->dump_unsigned("scrubber.seed", pg->scrubber.seed); | |
7121 | q.f->dump_int("scrubber.waiting_on", pg->scrubber.waiting_on); | |
7122 | { | |
7123 | q.f->open_array_section("scrubber.waiting_on_whom"); | |
7124 | for (set<pg_shard_t>::iterator p = pg->scrubber.waiting_on_whom.begin(); | |
7125 | p != pg->scrubber.waiting_on_whom.end(); | |
7126 | ++p) { | |
7127 | q.f->dump_stream("shard") << *p; | |
7128 | } | |
7129 | q.f->close_section(); | |
7130 | } | |
7131 | q.f->close_section(); | |
7132 | } | |
7133 | ||
7134 | q.f->close_section(); | |
7135 | return forward_event(); | |
7136 | } | |
7137 | ||
7138 | boost::statechart::result PG::RecoveryState::Active::react(const AllReplicasActivated &evt) | |
7139 | { | |
7140 | PG *pg = context< RecoveryMachine >().pg; | |
7141 | all_replicas_activated = true; | |
7142 | ||
7143 | pg->state_clear(PG_STATE_ACTIVATING); | |
7144 | pg->state_clear(PG_STATE_CREATING); | |
7145 | if (pg->acting.size() >= pg->pool.info.min_size) { | |
7146 | pg->state_set(PG_STATE_ACTIVE); | |
7147 | } else { | |
7148 | pg->state_set(PG_STATE_PEERED); | |
7149 | } | |
7150 | ||
7151 | // info.last_epoch_started is set during activate() | |
7152 | pg->info.history.last_epoch_started = pg->info.last_epoch_started; | |
7153 | pg->info.history.last_interval_started = pg->info.last_interval_started; | |
7154 | pg->dirty_info = true; | |
7155 | ||
7156 | pg->share_pg_info(); | |
7157 | pg->publish_stats_to_osd(); | |
7158 | ||
7159 | pg->check_local(); | |
7160 | ||
7161 | // waiters | |
7162 | if (pg->flushes_in_progress == 0) { | |
7163 | pg->requeue_ops(pg->waiting_for_peered); | |
7164 | } | |
7165 | ||
7166 | pg->on_activate(); | |
7167 | ||
7168 | return discard_event(); | |
7169 | } | |
7170 | ||
7171 | void PG::RecoveryState::Active::exit() | |
7172 | { | |
7173 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7174 | PG *pg = context< RecoveryMachine >().pg; | |
7175 | pg->osd->local_reserver.cancel_reservation(pg->info.pgid); | |
7176 | ||
7177 | pg->blocked_by.clear(); | |
7178 | pg->backfill_reserved = false; | |
7179 | pg->backfill_reserving = false; | |
7180 | pg->state_clear(PG_STATE_ACTIVATING); | |
7181 | pg->state_clear(PG_STATE_DEGRADED); | |
7182 | pg->state_clear(PG_STATE_UNDERSIZED); | |
7183 | pg->state_clear(PG_STATE_BACKFILL_TOOFULL); | |
7184 | pg->state_clear(PG_STATE_BACKFILL_WAIT); | |
7185 | pg->state_clear(PG_STATE_RECOVERY_WAIT); | |
7186 | pg->state_clear(PG_STATE_RECOVERY_TOOFULL); | |
7187 | utime_t dur = ceph_clock_now() - enter_time; | |
7188 | pg->osd->recoverystate_perf->tinc(rs_active_latency, dur); | |
7189 | pg->agent_stop(); | |
7190 | } | |
7191 | ||
7192 | /*------ReplicaActive-----*/ | |
7193 | PG::RecoveryState::ReplicaActive::ReplicaActive(my_context ctx) | |
7194 | : my_base(ctx), | |
7195 | NamedState(context< RecoveryMachine >().pg, "Started/ReplicaActive") | |
7196 | { | |
7197 | context< RecoveryMachine >().log_enter(state_name); | |
7198 | ||
7199 | PG *pg = context< RecoveryMachine >().pg; | |
7200 | pg->start_flush( | |
7201 | context< RecoveryMachine >().get_cur_transaction(), | |
7202 | context< RecoveryMachine >().get_on_applied_context_list(), | |
7203 | context< RecoveryMachine >().get_on_safe_context_list()); | |
7204 | } | |
7205 | ||
7206 | ||
7207 | boost::statechart::result PG::RecoveryState::ReplicaActive::react( | |
7208 | const Activate& actevt) { | |
7209 | PG *pg = context< RecoveryMachine >().pg; | |
7210 | ldout(pg->cct, 10) << "In ReplicaActive, about to call activate" << dendl; | |
7211 | map<int, map<spg_t, pg_query_t> > query_map; | |
7212 | pg->activate(*context< RecoveryMachine >().get_cur_transaction(), | |
7213 | actevt.activation_epoch, | |
7214 | *context< RecoveryMachine >().get_on_safe_context_list(), | |
7215 | query_map, NULL, NULL); | |
7216 | ldout(pg->cct, 10) << "Activate Finished" << dendl; | |
7217 | return discard_event(); | |
7218 | } | |
7219 | ||
7220 | boost::statechart::result PG::RecoveryState::ReplicaActive::react(const MInfoRec& infoevt) | |
7221 | { | |
7222 | PG *pg = context< RecoveryMachine >().pg; | |
7223 | pg->proc_primary_info(*context<RecoveryMachine>().get_cur_transaction(), | |
7224 | infoevt.info); | |
7225 | return discard_event(); | |
7226 | } | |
7227 | ||
7228 | boost::statechart::result PG::RecoveryState::ReplicaActive::react(const MLogRec& logevt) | |
7229 | { | |
7230 | PG *pg = context< RecoveryMachine >().pg; | |
7231 | ldout(pg->cct, 10) << "received log from " << logevt.from << dendl; | |
7232 | ObjectStore::Transaction* t = context<RecoveryMachine>().get_cur_transaction(); | |
7233 | pg->merge_log(*t, logevt.msg->info, logevt.msg->log, logevt.from); | |
7234 | assert(pg->pg_log.get_head() == pg->info.last_update); | |
7235 | ||
7236 | return discard_event(); | |
7237 | } | |
7238 | ||
7239 | boost::statechart::result PG::RecoveryState::ReplicaActive::react(const ActMap&) | |
7240 | { | |
7241 | PG *pg = context< RecoveryMachine >().pg; | |
7242 | if (pg->should_send_notify() && pg->get_primary().osd >= 0) { | |
7243 | context< RecoveryMachine >().send_notify( | |
7244 | pg->get_primary(), | |
7245 | pg_notify_t( | |
7246 | pg->get_primary().shard, pg->pg_whoami.shard, | |
7247 | pg->get_osdmap()->get_epoch(), | |
7248 | pg->get_osdmap()->get_epoch(), | |
7249 | pg->info), | |
7250 | pg->past_intervals); | |
7251 | } | |
7252 | pg->take_waiters(); | |
7253 | return discard_event(); | |
7254 | } | |
7255 | ||
7256 | boost::statechart::result PG::RecoveryState::ReplicaActive::react(const MQuery& query) | |
7257 | { | |
7258 | PG *pg = context< RecoveryMachine >().pg; | |
7259 | if (query.query.type == pg_query_t::MISSING) { | |
7260 | pg->update_history(query.query.history); | |
7261 | pg->fulfill_log(query.from, query.query, query.query_epoch); | |
7262 | } // else: from prior to activation, safe to ignore | |
7263 | return discard_event(); | |
7264 | } | |
7265 | ||
7266 | boost::statechart::result PG::RecoveryState::ReplicaActive::react(const QueryState& q) | |
7267 | { | |
7268 | q.f->open_object_section("state"); | |
7269 | q.f->dump_string("name", state_name); | |
7270 | q.f->dump_stream("enter_time") << enter_time; | |
7271 | q.f->close_section(); | |
7272 | return forward_event(); | |
7273 | } | |
7274 | ||
7275 | void PG::RecoveryState::ReplicaActive::exit() | |
7276 | { | |
7277 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7278 | PG *pg = context< RecoveryMachine >().pg; | |
7279 | pg->osd->remote_reserver.cancel_reservation(pg->info.pgid); | |
7280 | utime_t dur = ceph_clock_now() - enter_time; | |
7281 | pg->osd->recoverystate_perf->tinc(rs_replicaactive_latency, dur); | |
7282 | } | |
7283 | ||
7284 | /*-------Stray---*/ | |
7285 | PG::RecoveryState::Stray::Stray(my_context ctx) | |
7286 | : my_base(ctx), | |
7287 | NamedState(context< RecoveryMachine >().pg, "Started/Stray") | |
7288 | { | |
7289 | context< RecoveryMachine >().log_enter(state_name); | |
7290 | ||
7291 | PG *pg = context< RecoveryMachine >().pg; | |
7292 | assert(!pg->is_peered()); | |
7293 | assert(!pg->is_peering()); | |
7294 | assert(!pg->is_primary()); | |
7295 | pg->start_flush( | |
7296 | context< RecoveryMachine >().get_cur_transaction(), | |
7297 | context< RecoveryMachine >().get_on_applied_context_list(), | |
7298 | context< RecoveryMachine >().get_on_safe_context_list()); | |
7299 | } | |
7300 | ||
7301 | boost::statechart::result PG::RecoveryState::Stray::react(const MLogRec& logevt) | |
7302 | { | |
7303 | PG *pg = context< RecoveryMachine >().pg; | |
7304 | MOSDPGLog *msg = logevt.msg.get(); | |
7305 | ldout(pg->cct, 10) << "got info+log from osd." << logevt.from << " " << msg->info << " " << msg->log << dendl; | |
7306 | ||
7307 | ObjectStore::Transaction* t = context<RecoveryMachine>().get_cur_transaction(); | |
7308 | if (msg->info.last_backfill == hobject_t()) { | |
7309 | // restart backfill | |
7310 | pg->unreg_next_scrub(); | |
7311 | pg->info = msg->info; | |
7312 | pg->reg_next_scrub(); | |
7313 | pg->dirty_info = true; | |
7314 | pg->dirty_big_info = true; // maybe. | |
7315 | ||
7316 | PGLogEntryHandler rollbacker{pg, t}; | |
7317 | pg->pg_log.reset_backfill_claim_log(msg->log, &rollbacker); | |
7318 | ||
7319 | pg->pg_log.reset_backfill(); | |
7320 | } else { | |
7321 | pg->merge_log(*t, msg->info, msg->log, logevt.from); | |
7322 | } | |
7323 | ||
7324 | assert(pg->pg_log.get_head() == pg->info.last_update); | |
7325 | ||
7326 | post_event(Activate(logevt.msg->info.last_epoch_started)); | |
7327 | return transit<ReplicaActive>(); | |
7328 | } | |
7329 | ||
7330 | boost::statechart::result PG::RecoveryState::Stray::react(const MInfoRec& infoevt) | |
7331 | { | |
7332 | PG *pg = context< RecoveryMachine >().pg; | |
7333 | ldout(pg->cct, 10) << "got info from osd." << infoevt.from << " " << infoevt.info << dendl; | |
7334 | ||
7335 | if (pg->info.last_update > infoevt.info.last_update) { | |
7336 | // rewind divergent log entries | |
7337 | ObjectStore::Transaction* t = context<RecoveryMachine>().get_cur_transaction(); | |
7338 | pg->rewind_divergent_log(*t, infoevt.info.last_update); | |
7339 | pg->info.stats = infoevt.info.stats; | |
7340 | pg->info.hit_set = infoevt.info.hit_set; | |
7341 | } | |
7342 | ||
7343 | assert(infoevt.info.last_update == pg->info.last_update); | |
7344 | assert(pg->pg_log.get_head() == pg->info.last_update); | |
7345 | ||
7346 | post_event(Activate(infoevt.info.last_epoch_started)); | |
7347 | return transit<ReplicaActive>(); | |
7348 | } | |
7349 | ||
7350 | boost::statechart::result PG::RecoveryState::Stray::react(const MQuery& query) | |
7351 | { | |
7352 | PG *pg = context< RecoveryMachine >().pg; | |
7353 | if (query.query.type == pg_query_t::INFO) { | |
7354 | pair<pg_shard_t, pg_info_t> notify_info; | |
7355 | pg->update_history(query.query.history); | |
7356 | pg->fulfill_info(query.from, query.query, notify_info); | |
7357 | context< RecoveryMachine >().send_notify( | |
7358 | notify_info.first, | |
7359 | pg_notify_t( | |
7360 | notify_info.first.shard, pg->pg_whoami.shard, | |
7361 | query.query_epoch, | |
7362 | pg->get_osdmap()->get_epoch(), | |
7363 | notify_info.second), | |
7364 | pg->past_intervals); | |
7365 | } else { | |
7366 | pg->fulfill_log(query.from, query.query, query.query_epoch); | |
7367 | } | |
7368 | return discard_event(); | |
7369 | } | |
7370 | ||
7371 | boost::statechart::result PG::RecoveryState::Stray::react(const ActMap&) | |
7372 | { | |
7373 | PG *pg = context< RecoveryMachine >().pg; | |
7374 | if (pg->should_send_notify() && pg->get_primary().osd >= 0) { | |
7375 | context< RecoveryMachine >().send_notify( | |
7376 | pg->get_primary(), | |
7377 | pg_notify_t( | |
7378 | pg->get_primary().shard, pg->pg_whoami.shard, | |
7379 | pg->get_osdmap()->get_epoch(), | |
7380 | pg->get_osdmap()->get_epoch(), | |
7381 | pg->info), | |
7382 | pg->past_intervals); | |
7383 | } | |
7384 | pg->take_waiters(); | |
7385 | return discard_event(); | |
7386 | } | |
7387 | ||
7388 | void PG::RecoveryState::Stray::exit() | |
7389 | { | |
7390 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7391 | PG *pg = context< RecoveryMachine >().pg; | |
7392 | utime_t dur = ceph_clock_now() - enter_time; | |
7393 | pg->osd->recoverystate_perf->tinc(rs_stray_latency, dur); | |
7394 | } | |
7395 | ||
7396 | /*--------GetInfo---------*/ | |
7397 | PG::RecoveryState::GetInfo::GetInfo(my_context ctx) | |
7398 | : my_base(ctx), | |
7399 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Peering/GetInfo") | |
7400 | { | |
7401 | context< RecoveryMachine >().log_enter(state_name); | |
7402 | ||
7403 | PG *pg = context< RecoveryMachine >().pg; | |
7404 | pg->check_past_interval_bounds(); | |
7405 | PastIntervals::PriorSet &prior_set = context< Peering >().prior_set; | |
7406 | ||
7407 | assert(pg->blocked_by.empty()); | |
7408 | ||
7409 | prior_set = pg->build_prior(); | |
7410 | ||
7411 | pg->reset_min_peer_features(); | |
7412 | get_infos(); | |
7413 | if (prior_set.pg_down) { | |
7414 | post_event(IsDown()); | |
7415 | } else if (peer_info_requested.empty()) { | |
7416 | post_event(GotInfo()); | |
7417 | } | |
7418 | } | |
7419 | ||
7420 | void PG::RecoveryState::GetInfo::get_infos() | |
7421 | { | |
7422 | PG *pg = context< RecoveryMachine >().pg; | |
7423 | PastIntervals::PriorSet &prior_set = context< Peering >().prior_set; | |
7424 | ||
7425 | pg->blocked_by.clear(); | |
7426 | for (set<pg_shard_t>::const_iterator it = prior_set.probe.begin(); | |
7427 | it != prior_set.probe.end(); | |
7428 | ++it) { | |
7429 | pg_shard_t peer = *it; | |
7430 | if (peer == pg->pg_whoami) { | |
7431 | continue; | |
7432 | } | |
7433 | if (pg->peer_info.count(peer)) { | |
7434 | ldout(pg->cct, 10) << " have osd." << peer << " info " << pg->peer_info[peer] << dendl; | |
7435 | continue; | |
7436 | } | |
7437 | if (peer_info_requested.count(peer)) { | |
7438 | ldout(pg->cct, 10) << " already requested info from osd." << peer << dendl; | |
7439 | pg->blocked_by.insert(peer.osd); | |
7440 | } else if (!pg->get_osdmap()->is_up(peer.osd)) { | |
7441 | ldout(pg->cct, 10) << " not querying info from down osd." << peer << dendl; | |
7442 | } else { | |
7443 | ldout(pg->cct, 10) << " querying info from osd." << peer << dendl; | |
7444 | context< RecoveryMachine >().send_query( | |
7445 | peer, pg_query_t(pg_query_t::INFO, | |
7446 | it->shard, pg->pg_whoami.shard, | |
7447 | pg->info.history, | |
7448 | pg->get_osdmap()->get_epoch())); | |
7449 | peer_info_requested.insert(peer); | |
7450 | pg->blocked_by.insert(peer.osd); | |
7451 | } | |
7452 | } | |
7453 | ||
7454 | pg->publish_stats_to_osd(); | |
7455 | } | |
7456 | ||
7457 | boost::statechart::result PG::RecoveryState::GetInfo::react(const MNotifyRec& infoevt) | |
7458 | { | |
7459 | PG *pg = context< RecoveryMachine >().pg; | |
7460 | ||
7461 | set<pg_shard_t>::iterator p = peer_info_requested.find(infoevt.from); | |
7462 | if (p != peer_info_requested.end()) { | |
7463 | peer_info_requested.erase(p); | |
7464 | pg->blocked_by.erase(infoevt.from.osd); | |
7465 | } | |
7466 | ||
7467 | epoch_t old_start = pg->info.history.last_epoch_started; | |
7468 | if (pg->proc_replica_info( | |
7469 | infoevt.from, infoevt.notify.info, infoevt.notify.epoch_sent)) { | |
7470 | // we got something new ... | |
7471 | PastIntervals::PriorSet &prior_set = context< Peering >().prior_set; | |
7472 | if (old_start < pg->info.history.last_epoch_started) { | |
7473 | ldout(pg->cct, 10) << " last_epoch_started moved forward, rebuilding prior" << dendl; | |
7474 | prior_set = pg->build_prior(); | |
7475 | ||
7476 | // filter out any osds that got dropped from the probe set from | |
7477 | // peer_info_requested. this is less expensive than restarting | |
7478 | // peering (which would re-probe everyone). | |
7479 | set<pg_shard_t>::iterator p = peer_info_requested.begin(); | |
7480 | while (p != peer_info_requested.end()) { | |
7481 | if (prior_set.probe.count(*p) == 0) { | |
7482 | ldout(pg->cct, 20) << " dropping osd." << *p << " from info_requested, no longer in probe set" << dendl; | |
7483 | peer_info_requested.erase(p++); | |
7484 | } else { | |
7485 | ++p; | |
7486 | } | |
7487 | } | |
7488 | get_infos(); | |
7489 | } | |
7490 | ldout(pg->cct, 20) << "Adding osd: " << infoevt.from.osd << " peer features: " | |
7491 | << hex << infoevt.features << dec << dendl; | |
7492 | pg->apply_peer_features(infoevt.features); | |
7493 | ||
7494 | // are we done getting everything? | |
7495 | if (peer_info_requested.empty() && !prior_set.pg_down) { | |
7496 | ldout(pg->cct, 20) << "Common peer features: " << hex << pg->get_min_peer_features() << dec << dendl; | |
7497 | ldout(pg->cct, 20) << "Common acting features: " << hex << pg->get_min_acting_features() << dec << dendl; | |
7498 | ldout(pg->cct, 20) << "Common upacting features: " << hex << pg->get_min_upacting_features() << dec << dendl; | |
7499 | post_event(GotInfo()); | |
7500 | } | |
7501 | } | |
7502 | return discard_event(); | |
7503 | } | |
7504 | ||
7505 | boost::statechart::result PG::RecoveryState::GetInfo::react(const QueryState& q) | |
7506 | { | |
7507 | PG *pg = context< RecoveryMachine >().pg; | |
7508 | q.f->open_object_section("state"); | |
7509 | q.f->dump_string("name", state_name); | |
7510 | q.f->dump_stream("enter_time") << enter_time; | |
7511 | ||
7512 | q.f->open_array_section("requested_info_from"); | |
7513 | for (set<pg_shard_t>::iterator p = peer_info_requested.begin(); | |
7514 | p != peer_info_requested.end(); | |
7515 | ++p) { | |
7516 | q.f->open_object_section("osd"); | |
7517 | q.f->dump_stream("osd") << *p; | |
7518 | if (pg->peer_info.count(*p)) { | |
7519 | q.f->open_object_section("got_info"); | |
7520 | pg->peer_info[*p].dump(q.f); | |
7521 | q.f->close_section(); | |
7522 | } | |
7523 | q.f->close_section(); | |
7524 | } | |
7525 | q.f->close_section(); | |
7526 | ||
7527 | q.f->close_section(); | |
7528 | return forward_event(); | |
7529 | } | |
7530 | ||
7531 | void PG::RecoveryState::GetInfo::exit() | |
7532 | { | |
7533 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7534 | PG *pg = context< RecoveryMachine >().pg; | |
7535 | utime_t dur = ceph_clock_now() - enter_time; | |
7536 | pg->osd->recoverystate_perf->tinc(rs_getinfo_latency, dur); | |
7537 | pg->blocked_by.clear(); | |
7538 | pg->publish_stats_to_osd(); | |
7539 | } | |
7540 | ||
7541 | /*------GetLog------------*/ | |
7542 | PG::RecoveryState::GetLog::GetLog(my_context ctx) | |
7543 | : my_base(ctx), | |
7544 | NamedState( | |
7545 | context< RecoveryMachine >().pg, "Started/Primary/Peering/GetLog"), | |
7546 | msg(0) | |
7547 | { | |
7548 | context< RecoveryMachine >().log_enter(state_name); | |
7549 | ||
7550 | PG *pg = context< RecoveryMachine >().pg; | |
7551 | ||
7552 | // adjust acting? | |
7553 | if (!pg->choose_acting(auth_log_shard, false, | |
7554 | &context< Peering >().history_les_bound)) { | |
7555 | if (!pg->want_acting.empty()) { | |
7556 | post_event(NeedActingChange()); | |
7557 | } else { | |
7558 | post_event(IsIncomplete()); | |
7559 | } | |
7560 | return; | |
7561 | } | |
7562 | ||
7563 | // am i the best? | |
7564 | if (auth_log_shard == pg->pg_whoami) { | |
7565 | post_event(GotLog()); | |
7566 | return; | |
7567 | } | |
7568 | ||
7569 | const pg_info_t& best = pg->peer_info[auth_log_shard]; | |
7570 | ||
7571 | // am i broken? | |
7572 | if (pg->info.last_update < best.log_tail) { | |
7573 | ldout(pg->cct, 10) << " not contiguous with osd." << auth_log_shard << ", down" << dendl; | |
7574 | post_event(IsIncomplete()); | |
7575 | return; | |
7576 | } | |
7577 | ||
7578 | // how much log to request? | |
7579 | eversion_t request_log_from = pg->info.last_update; | |
7580 | assert(!pg->actingbackfill.empty()); | |
7581 | for (set<pg_shard_t>::iterator p = pg->actingbackfill.begin(); | |
7582 | p != pg->actingbackfill.end(); | |
7583 | ++p) { | |
7584 | if (*p == pg->pg_whoami) continue; | |
7585 | pg_info_t& ri = pg->peer_info[*p]; | |
7586 | if (ri.last_update < pg->info.log_tail && ri.last_update >= best.log_tail && | |
7587 | ri.last_update < request_log_from) | |
7588 | request_log_from = ri.last_update; | |
7589 | } | |
7590 | ||
7591 | // how much? | |
7592 | ldout(pg->cct, 10) << " requesting log from osd." << auth_log_shard << dendl; | |
7593 | context<RecoveryMachine>().send_query( | |
7594 | auth_log_shard, | |
7595 | pg_query_t( | |
7596 | pg_query_t::LOG, | |
7597 | auth_log_shard.shard, pg->pg_whoami.shard, | |
7598 | request_log_from, pg->info.history, | |
7599 | pg->get_osdmap()->get_epoch())); | |
7600 | ||
7601 | assert(pg->blocked_by.empty()); | |
7602 | pg->blocked_by.insert(auth_log_shard.osd); | |
7603 | pg->publish_stats_to_osd(); | |
7604 | } | |
7605 | ||
7606 | boost::statechart::result PG::RecoveryState::GetLog::react(const AdvMap& advmap) | |
7607 | { | |
7608 | PG *pg = context< RecoveryMachine >().pg; | |
7609 | // make sure our log source didn't go down. we need to check | |
7610 | // explicitly because it may not be part of the prior set, which | |
7611 | // means the Peering state check won't catch it going down. | |
7612 | if (!advmap.osdmap->is_up(auth_log_shard.osd)) { | |
7613 | ldout(pg->cct, 10) << "GetLog: auth_log_shard osd." | |
7614 | << auth_log_shard.osd << " went down" << dendl; | |
7615 | post_event(advmap); | |
7616 | return transit< Reset >(); | |
7617 | } | |
7618 | ||
7619 | // let the Peering state do its checks. | |
7620 | return forward_event(); | |
7621 | } | |
7622 | ||
7623 | boost::statechart::result PG::RecoveryState::GetLog::react(const MLogRec& logevt) | |
7624 | { | |
7625 | PG *pg = context< RecoveryMachine >().pg; | |
7626 | assert(!msg); | |
7627 | if (logevt.from != auth_log_shard) { | |
7628 | ldout(pg->cct, 10) << "GetLog: discarding log from " | |
7629 | << "non-auth_log_shard osd." << logevt.from << dendl; | |
7630 | return discard_event(); | |
7631 | } | |
7632 | ldout(pg->cct, 10) << "GetLog: received master log from osd" | |
7633 | << logevt.from << dendl; | |
7634 | msg = logevt.msg; | |
7635 | post_event(GotLog()); | |
7636 | return discard_event(); | |
7637 | } | |
7638 | ||
7639 | boost::statechart::result PG::RecoveryState::GetLog::react(const GotLog&) | |
7640 | { | |
7641 | PG *pg = context< RecoveryMachine >().pg; | |
7642 | ldout(pg->cct, 10) << "leaving GetLog" << dendl; | |
7643 | if (msg) { | |
7644 | ldout(pg->cct, 10) << "processing master log" << dendl; | |
7645 | pg->proc_master_log(*context<RecoveryMachine>().get_cur_transaction(), | |
7646 | msg->info, msg->log, msg->missing, | |
7647 | auth_log_shard); | |
7648 | } | |
7649 | pg->start_flush( | |
7650 | context< RecoveryMachine >().get_cur_transaction(), | |
7651 | context< RecoveryMachine >().get_on_applied_context_list(), | |
7652 | context< RecoveryMachine >().get_on_safe_context_list()); | |
7653 | return transit< GetMissing >(); | |
7654 | } | |
7655 | ||
7656 | boost::statechart::result PG::RecoveryState::GetLog::react(const QueryState& q) | |
7657 | { | |
7658 | q.f->open_object_section("state"); | |
7659 | q.f->dump_string("name", state_name); | |
7660 | q.f->dump_stream("enter_time") << enter_time; | |
7661 | q.f->dump_stream("auth_log_shard") << auth_log_shard; | |
7662 | q.f->close_section(); | |
7663 | return forward_event(); | |
7664 | } | |
7665 | ||
7666 | void PG::RecoveryState::GetLog::exit() | |
7667 | { | |
7668 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7669 | PG *pg = context< RecoveryMachine >().pg; | |
7670 | utime_t dur = ceph_clock_now() - enter_time; | |
7671 | pg->osd->recoverystate_perf->tinc(rs_getlog_latency, dur); | |
7672 | pg->blocked_by.clear(); | |
7673 | pg->publish_stats_to_osd(); | |
7674 | } | |
7675 | ||
7676 | /*------WaitActingChange--------*/ | |
7677 | PG::RecoveryState::WaitActingChange::WaitActingChange(my_context ctx) | |
7678 | : my_base(ctx), | |
7679 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Peering/WaitActingChange") | |
7680 | { | |
7681 | context< RecoveryMachine >().log_enter(state_name); | |
7682 | } | |
7683 | ||
7684 | boost::statechart::result PG::RecoveryState::WaitActingChange::react(const AdvMap& advmap) | |
7685 | { | |
7686 | PG *pg = context< RecoveryMachine >().pg; | |
7687 | OSDMapRef osdmap = advmap.osdmap; | |
7688 | ||
7689 | ldout(pg->cct, 10) << "verifying no want_acting " << pg->want_acting << " targets didn't go down" << dendl; | |
7690 | for (vector<int>::iterator p = pg->want_acting.begin(); p != pg->want_acting.end(); ++p) { | |
7691 | if (!osdmap->is_up(*p)) { | |
7692 | ldout(pg->cct, 10) << " want_acting target osd." << *p << " went down, resetting" << dendl; | |
7693 | post_event(advmap); | |
7694 | return transit< Reset >(); | |
7695 | } | |
7696 | } | |
7697 | return forward_event(); | |
7698 | } | |
7699 | ||
7700 | boost::statechart::result PG::RecoveryState::WaitActingChange::react(const MLogRec& logevt) | |
7701 | { | |
7702 | PG *pg = context< RecoveryMachine >().pg; | |
7703 | ldout(pg->cct, 10) << "In WaitActingChange, ignoring MLocRec" << dendl; | |
7704 | return discard_event(); | |
7705 | } | |
7706 | ||
7707 | boost::statechart::result PG::RecoveryState::WaitActingChange::react(const MInfoRec& evt) | |
7708 | { | |
7709 | PG *pg = context< RecoveryMachine >().pg; | |
7710 | ldout(pg->cct, 10) << "In WaitActingChange, ignoring MInfoRec" << dendl; | |
7711 | return discard_event(); | |
7712 | } | |
7713 | ||
7714 | boost::statechart::result PG::RecoveryState::WaitActingChange::react(const MNotifyRec& evt) | |
7715 | { | |
7716 | PG *pg = context< RecoveryMachine >().pg; | |
7717 | ldout(pg->cct, 10) << "In WaitActingChange, ignoring MNotifyRec" << dendl; | |
7718 | return discard_event(); | |
7719 | } | |
7720 | ||
7721 | boost::statechart::result PG::RecoveryState::WaitActingChange::react(const QueryState& q) | |
7722 | { | |
7723 | q.f->open_object_section("state"); | |
7724 | q.f->dump_string("name", state_name); | |
7725 | q.f->dump_stream("enter_time") << enter_time; | |
7726 | q.f->dump_string("comment", "waiting for pg acting set to change"); | |
7727 | q.f->close_section(); | |
7728 | return forward_event(); | |
7729 | } | |
7730 | ||
7731 | void PG::RecoveryState::WaitActingChange::exit() | |
7732 | { | |
7733 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7734 | PG *pg = context< RecoveryMachine >().pg; | |
7735 | utime_t dur = ceph_clock_now() - enter_time; | |
7736 | pg->osd->recoverystate_perf->tinc(rs_waitactingchange_latency, dur); | |
7737 | } | |
7738 | ||
7739 | /*------Down--------*/ | |
7740 | PG::RecoveryState::Down::Down(my_context ctx) | |
7741 | : my_base(ctx), | |
7742 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Peering/Down") | |
7743 | { | |
7744 | context< RecoveryMachine >().log_enter(state_name); | |
7745 | PG *pg = context< RecoveryMachine >().pg; | |
7746 | ||
7747 | pg->state_clear(PG_STATE_PEERING); | |
7748 | pg->state_set(PG_STATE_DOWN); | |
7749 | ||
7750 | auto &prior_set = context< Peering >().prior_set; | |
7751 | assert(pg->blocked_by.empty()); | |
7752 | pg->blocked_by.insert(prior_set.down.begin(), prior_set.down.end()); | |
7753 | pg->publish_stats_to_osd(); | |
7754 | } | |
7755 | ||
7756 | void PG::RecoveryState::Down::exit() | |
7757 | { | |
7758 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7759 | PG *pg = context< RecoveryMachine >().pg; | |
7760 | ||
7761 | pg->state_clear(PG_STATE_DOWN); | |
7762 | utime_t dur = ceph_clock_now() - enter_time; | |
7763 | pg->osd->recoverystate_perf->tinc(rs_down_latency, dur); | |
7764 | ||
7765 | pg->blocked_by.clear(); | |
7766 | pg->publish_stats_to_osd(); | |
7767 | } | |
7768 | ||
7769 | boost::statechart::result PG::RecoveryState::Down::react(const QueryState& q) | |
7770 | { | |
7771 | q.f->open_object_section("state"); | |
7772 | q.f->dump_string("name", state_name); | |
7773 | q.f->dump_stream("enter_time") << enter_time; | |
7774 | q.f->dump_string("comment", | |
7775 | "not enough up instances of this PG to go active"); | |
7776 | q.f->close_section(); | |
7777 | return forward_event(); | |
7778 | } | |
7779 | ||
7780 | /*------Incomplete--------*/ | |
7781 | PG::RecoveryState::Incomplete::Incomplete(my_context ctx) | |
7782 | : my_base(ctx), | |
7783 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Peering/Incomplete") | |
7784 | { | |
7785 | context< RecoveryMachine >().log_enter(state_name); | |
7786 | PG *pg = context< RecoveryMachine >().pg; | |
7787 | ||
7788 | pg->state_clear(PG_STATE_PEERING); | |
7789 | pg->state_set(PG_STATE_INCOMPLETE); | |
7790 | ||
7791 | PastIntervals::PriorSet &prior_set = context< Peering >().prior_set; | |
7792 | assert(pg->blocked_by.empty()); | |
7793 | pg->blocked_by.insert(prior_set.down.begin(), prior_set.down.end()); | |
7794 | pg->publish_stats_to_osd(); | |
7795 | } | |
7796 | ||
7797 | boost::statechart::result PG::RecoveryState::Incomplete::react(const AdvMap &advmap) { | |
7798 | PG *pg = context< RecoveryMachine >().pg; | |
7799 | int64_t poolnum = pg->info.pgid.pool(); | |
7800 | ||
7801 | // Reset if min_size turn smaller than previous value, pg might now be able to go active | |
7802 | if (advmap.lastmap->get_pools().find(poolnum)->second.min_size > | |
7803 | advmap.osdmap->get_pools().find(poolnum)->second.min_size) { | |
7804 | post_event(advmap); | |
7805 | return transit< Reset >(); | |
7806 | } | |
7807 | ||
7808 | return forward_event(); | |
7809 | } | |
7810 | ||
7811 | boost::statechart::result PG::RecoveryState::Incomplete::react(const MNotifyRec& notevt) { | |
7812 | PG *pg = context< RecoveryMachine >().pg; | |
7813 | ldout(pg->cct, 7) << "handle_pg_notify from osd." << notevt.from << dendl; | |
7814 | if (pg->proc_replica_info( | |
7815 | notevt.from, notevt.notify.info, notevt.notify.epoch_sent)) { | |
7816 | // We got something new, try again! | |
7817 | return transit< GetLog >(); | |
7818 | } else { | |
7819 | return discard_event(); | |
7820 | } | |
7821 | } | |
7822 | ||
7823 | boost::statechart::result PG::RecoveryState::Incomplete::react( | |
7824 | const QueryState& q) | |
7825 | { | |
7826 | q.f->open_object_section("state"); | |
7827 | q.f->dump_string("name", state_name); | |
7828 | q.f->dump_stream("enter_time") << enter_time; | |
7829 | q.f->dump_string("comment", "not enough complete instances of this PG"); | |
7830 | q.f->close_section(); | |
7831 | return forward_event(); | |
7832 | } | |
7833 | ||
7834 | void PG::RecoveryState::Incomplete::exit() | |
7835 | { | |
7836 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7837 | PG *pg = context< RecoveryMachine >().pg; | |
7838 | ||
7839 | pg->state_clear(PG_STATE_INCOMPLETE); | |
7840 | utime_t dur = ceph_clock_now() - enter_time; | |
7841 | pg->osd->recoverystate_perf->tinc(rs_incomplete_latency, dur); | |
7842 | ||
7843 | pg->blocked_by.clear(); | |
7844 | pg->publish_stats_to_osd(); | |
7845 | } | |
7846 | ||
7847 | /*------GetMissing--------*/ | |
7848 | PG::RecoveryState::GetMissing::GetMissing(my_context ctx) | |
7849 | : my_base(ctx), | |
7850 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Peering/GetMissing") | |
7851 | { | |
7852 | context< RecoveryMachine >().log_enter(state_name); | |
7853 | ||
7854 | PG *pg = context< RecoveryMachine >().pg; | |
7855 | assert(!pg->actingbackfill.empty()); | |
7856 | eversion_t since; | |
7857 | for (set<pg_shard_t>::iterator i = pg->actingbackfill.begin(); | |
7858 | i != pg->actingbackfill.end(); | |
7859 | ++i) { | |
7860 | if (*i == pg->get_primary()) continue; | |
7861 | const pg_info_t& pi = pg->peer_info[*i]; | |
7862 | ||
7863 | if (pi.is_empty()) | |
7864 | continue; // no pg data, nothing divergent | |
7865 | ||
7866 | if (pi.last_update < pg->pg_log.get_tail()) { | |
7867 | ldout(pg->cct, 10) << " osd." << *i << " is not contiguous, will restart backfill" << dendl; | |
7868 | pg->peer_missing[*i]; | |
7869 | continue; | |
7870 | } | |
7871 | if (pi.last_backfill == hobject_t()) { | |
7872 | ldout(pg->cct, 10) << " osd." << *i << " will fully backfill; can infer empty missing set" << dendl; | |
7873 | pg->peer_missing[*i]; | |
7874 | continue; | |
7875 | } | |
7876 | ||
7877 | if (pi.last_update == pi.last_complete && // peer has no missing | |
7878 | pi.last_update == pg->info.last_update) { // peer is up to date | |
7879 | // replica has no missing and identical log as us. no need to | |
7880 | // pull anything. | |
7881 | // FIXME: we can do better here. if last_update==last_complete we | |
7882 | // can infer the rest! | |
7883 | ldout(pg->cct, 10) << " osd." << *i << " has no missing, identical log" << dendl; | |
7884 | pg->peer_missing[*i]; | |
7885 | continue; | |
7886 | } | |
7887 | ||
7888 | // We pull the log from the peer's last_epoch_started to ensure we | |
7889 | // get enough log to detect divergent updates. | |
7890 | since.epoch = pi.last_epoch_started; | |
7891 | assert(pi.last_update >= pg->info.log_tail); // or else choose_acting() did a bad thing | |
7892 | if (pi.log_tail <= since) { | |
7893 | ldout(pg->cct, 10) << " requesting log+missing since " << since << " from osd." << *i << dendl; | |
7894 | context< RecoveryMachine >().send_query( | |
7895 | *i, | |
7896 | pg_query_t( | |
7897 | pg_query_t::LOG, | |
7898 | i->shard, pg->pg_whoami.shard, | |
7899 | since, pg->info.history, | |
7900 | pg->get_osdmap()->get_epoch())); | |
7901 | } else { | |
7902 | ldout(pg->cct, 10) << " requesting fulllog+missing from osd." << *i | |
7903 | << " (want since " << since << " < log.tail " | |
7904 | << pi.log_tail << ")" << dendl; | |
7905 | context< RecoveryMachine >().send_query( | |
7906 | *i, pg_query_t( | |
7907 | pg_query_t::FULLLOG, | |
7908 | i->shard, pg->pg_whoami.shard, | |
7909 | pg->info.history, pg->get_osdmap()->get_epoch())); | |
7910 | } | |
7911 | peer_missing_requested.insert(*i); | |
7912 | pg->blocked_by.insert(i->osd); | |
7913 | } | |
7914 | ||
7915 | if (peer_missing_requested.empty()) { | |
7916 | if (pg->need_up_thru) { | |
7917 | ldout(pg->cct, 10) << " still need up_thru update before going active" | |
7918 | << dendl; | |
7919 | post_event(NeedUpThru()); | |
7920 | return; | |
7921 | } | |
7922 | ||
7923 | // all good! | |
7924 | post_event(Activate(pg->get_osdmap()->get_epoch())); | |
7925 | } else { | |
7926 | pg->publish_stats_to_osd(); | |
7927 | } | |
7928 | } | |
7929 | ||
7930 | boost::statechart::result PG::RecoveryState::GetMissing::react(const MLogRec& logevt) | |
7931 | { | |
7932 | PG *pg = context< RecoveryMachine >().pg; | |
7933 | ||
7934 | peer_missing_requested.erase(logevt.from); | |
7935 | pg->proc_replica_log(logevt.msg->info, logevt.msg->log, logevt.msg->missing, logevt.from); | |
7936 | ||
7937 | if (peer_missing_requested.empty()) { | |
7938 | if (pg->need_up_thru) { | |
7939 | ldout(pg->cct, 10) << " still need up_thru update before going active" | |
7940 | << dendl; | |
7941 | post_event(NeedUpThru()); | |
7942 | } else { | |
7943 | ldout(pg->cct, 10) << "Got last missing, don't need missing " | |
7944 | << "posting Activate" << dendl; | |
7945 | post_event(Activate(pg->get_osdmap()->get_epoch())); | |
7946 | } | |
7947 | } | |
7948 | return discard_event(); | |
7949 | } | |
7950 | ||
7951 | boost::statechart::result PG::RecoveryState::GetMissing::react(const QueryState& q) | |
7952 | { | |
7953 | PG *pg = context< RecoveryMachine >().pg; | |
7954 | q.f->open_object_section("state"); | |
7955 | q.f->dump_string("name", state_name); | |
7956 | q.f->dump_stream("enter_time") << enter_time; | |
7957 | ||
7958 | q.f->open_array_section("peer_missing_requested"); | |
7959 | for (set<pg_shard_t>::iterator p = peer_missing_requested.begin(); | |
7960 | p != peer_missing_requested.end(); | |
7961 | ++p) { | |
7962 | q.f->open_object_section("osd"); | |
7963 | q.f->dump_stream("osd") << *p; | |
7964 | if (pg->peer_missing.count(*p)) { | |
7965 | q.f->open_object_section("got_missing"); | |
7966 | pg->peer_missing[*p].dump(q.f); | |
7967 | q.f->close_section(); | |
7968 | } | |
7969 | q.f->close_section(); | |
7970 | } | |
7971 | q.f->close_section(); | |
7972 | ||
7973 | q.f->close_section(); | |
7974 | return forward_event(); | |
7975 | } | |
7976 | ||
7977 | void PG::RecoveryState::GetMissing::exit() | |
7978 | { | |
7979 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
7980 | PG *pg = context< RecoveryMachine >().pg; | |
7981 | utime_t dur = ceph_clock_now() - enter_time; | |
7982 | pg->osd->recoverystate_perf->tinc(rs_getmissing_latency, dur); | |
7983 | pg->blocked_by.clear(); | |
7984 | pg->publish_stats_to_osd(); | |
7985 | } | |
7986 | ||
7987 | /*------WaitUpThru--------*/ | |
7988 | PG::RecoveryState::WaitUpThru::WaitUpThru(my_context ctx) | |
7989 | : my_base(ctx), | |
7990 | NamedState(context< RecoveryMachine >().pg, "Started/Primary/Peering/WaitUpThru") | |
7991 | { | |
7992 | context< RecoveryMachine >().log_enter(state_name); | |
7993 | } | |
7994 | ||
7995 | boost::statechart::result PG::RecoveryState::WaitUpThru::react(const ActMap& am) | |
7996 | { | |
7997 | PG *pg = context< RecoveryMachine >().pg; | |
7998 | if (!pg->need_up_thru) { | |
7999 | post_event(Activate(pg->get_osdmap()->get_epoch())); | |
8000 | } | |
8001 | return forward_event(); | |
8002 | } | |
8003 | ||
8004 | boost::statechart::result PG::RecoveryState::WaitUpThru::react(const MLogRec& logevt) | |
8005 | { | |
8006 | PG *pg = context< RecoveryMachine >().pg; | |
8007 | ldout(pg->cct, 10) << "Noting missing from osd." << logevt.from << dendl; | |
8008 | pg->peer_missing[logevt.from].claim(logevt.msg->missing); | |
8009 | pg->peer_info[logevt.from] = logevt.msg->info; | |
8010 | return discard_event(); | |
8011 | } | |
8012 | ||
8013 | boost::statechart::result PG::RecoveryState::WaitUpThru::react(const QueryState& q) | |
8014 | { | |
8015 | q.f->open_object_section("state"); | |
8016 | q.f->dump_string("name", state_name); | |
8017 | q.f->dump_stream("enter_time") << enter_time; | |
8018 | q.f->dump_string("comment", "waiting for osdmap to reflect a new up_thru for this osd"); | |
8019 | q.f->close_section(); | |
8020 | return forward_event(); | |
8021 | } | |
8022 | ||
8023 | void PG::RecoveryState::WaitUpThru::exit() | |
8024 | { | |
8025 | context< RecoveryMachine >().log_exit(state_name, enter_time); | |
8026 | PG *pg = context< RecoveryMachine >().pg; | |
8027 | utime_t dur = ceph_clock_now() - enter_time; | |
8028 | pg->osd->recoverystate_perf->tinc(rs_waitupthru_latency, dur); | |
8029 | } | |
8030 | ||
8031 | /*----RecoveryState::RecoveryMachine Methods-----*/ | |
8032 | #undef dout_prefix | |
8033 | #define dout_prefix *_dout << pg->gen_prefix() | |
8034 | ||
8035 | void PG::RecoveryState::RecoveryMachine::log_enter(const char *state_name) | |
8036 | { | |
8037 | PG *pg = context< RecoveryMachine >().pg; | |
8038 | ldout(pg->cct, 5) << "enter " << state_name << dendl; | |
8039 | pg->osd->pg_recovery_stats.log_enter(state_name); | |
8040 | } | |
8041 | ||
8042 | void PG::RecoveryState::RecoveryMachine::log_exit(const char *state_name, utime_t enter_time) | |
8043 | { | |
8044 | utime_t dur = ceph_clock_now() - enter_time; | |
8045 | PG *pg = context< RecoveryMachine >().pg; | |
8046 | ldout(pg->cct, 5) << "exit " << state_name << " " << dur << " " << event_count << " " << event_time << dendl; | |
8047 | pg->osd->pg_recovery_stats.log_exit(state_name, ceph_clock_now() - enter_time, | |
8048 | event_count, event_time); | |
8049 | event_count = 0; | |
8050 | event_time = utime_t(); | |
8051 | } | |
8052 | ||
8053 | ||
8054 | /*---------------------------------------------------*/ | |
8055 | #undef dout_prefix | |
8056 | #define dout_prefix (*_dout << (debug_pg ? debug_pg->gen_prefix() : string()) << " PriorSet: ") | |
8057 | ||
8058 | void PG::RecoveryState::start_handle(RecoveryCtx *new_ctx) { | |
8059 | assert(!rctx); | |
8060 | assert(!orig_ctx); | |
8061 | orig_ctx = new_ctx; | |
8062 | if (new_ctx) { | |
8063 | if (messages_pending_flush) { | |
8064 | rctx = RecoveryCtx(*messages_pending_flush, *new_ctx); | |
8065 | } else { | |
8066 | rctx = *new_ctx; | |
8067 | } | |
8068 | rctx->start_time = ceph_clock_now(); | |
8069 | } | |
8070 | } | |
8071 | ||
8072 | void PG::RecoveryState::begin_block_outgoing() { | |
8073 | assert(!messages_pending_flush); | |
8074 | assert(orig_ctx); | |
8075 | assert(rctx); | |
8076 | messages_pending_flush = BufferedRecoveryMessages(); | |
8077 | rctx = RecoveryCtx(*messages_pending_flush, *orig_ctx); | |
8078 | } | |
8079 | ||
8080 | void PG::RecoveryState::clear_blocked_outgoing() { | |
8081 | assert(orig_ctx); | |
8082 | assert(rctx); | |
8083 | messages_pending_flush = boost::optional<BufferedRecoveryMessages>(); | |
8084 | } | |
8085 | ||
8086 | void PG::RecoveryState::end_block_outgoing() { | |
8087 | assert(messages_pending_flush); | |
8088 | assert(orig_ctx); | |
8089 | assert(rctx); | |
8090 | ||
8091 | rctx = RecoveryCtx(*orig_ctx); | |
8092 | rctx->accept_buffered_messages(*messages_pending_flush); | |
8093 | messages_pending_flush = boost::optional<BufferedRecoveryMessages>(); | |
8094 | } | |
8095 | ||
8096 | void PG::RecoveryState::end_handle() { | |
8097 | if (rctx) { | |
8098 | utime_t dur = ceph_clock_now() - rctx->start_time; | |
8099 | machine.event_time += dur; | |
8100 | } | |
8101 | ||
8102 | machine.event_count++; | |
8103 | rctx = boost::optional<RecoveryCtx>(); | |
8104 | orig_ctx = NULL; | |
8105 | } | |
8106 | ||
8107 | ostream& operator<<(ostream& out, const PG::BackfillInterval& bi) | |
8108 | { | |
8109 | out << "BackfillInfo(" << bi.begin << "-" << bi.end | |
8110 | << " " << bi.objects.size() << " objects"; | |
8111 | if (!bi.objects.empty()) | |
8112 | out << " " << bi.objects; | |
8113 | out << ")"; | |
8114 | return out; | |
8115 | } | |
8116 | ||
8117 | void intrusive_ptr_add_ref(PG *pg) { pg->get("intptr"); } | |
8118 | void intrusive_ptr_release(PG *pg) { pg->put("intptr"); } | |
8119 | ||
8120 | #ifdef PG_DEBUG_REFS | |
8121 | uint64_t get_with_id(PG *pg) { return pg->get_with_id(); } | |
8122 | void put_with_id(PG *pg, uint64_t id) { return pg->put_with_id(id); } | |
8123 | #endif |