1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 * Ceph - scalable distributed file system
6 * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
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.
16 // #include "msg/Messenger.h"
17 #include "messages/MOSDRepScrub.h"
18 // #include "common/cmdparse.h"
19 // #include "common/ceph_context.h"
21 #include "common/errno.h"
22 #include "common/config.h"
24 #include "OpRequest.h"
25 #include "ScrubStore.h"
28 #include "common/Timer.h"
29 #include "common/perf_counters.h"
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 #include "messages/MOSDPGRecoveryDelete.h"
59 #include "messages/MOSDPGRecoveryDeleteReply.h"
61 #include "common/BackTrace.h"
62 #include "common/EventTrace.h"
65 #define TRACEPOINT_DEFINE
66 #define TRACEPOINT_PROBE_DYNAMIC_LINKAGE
67 #include "tracing/pg.h"
68 #undef TRACEPOINT_PROBE_DYNAMIC_LINKAGE
69 #undef TRACEPOINT_DEFINE
71 #define tracepoint(...)
76 #define dout_context cct
77 #define dout_subsys ceph_subsys_osd
79 #define dout_prefix _prefix(_dout, this)
81 // prefix pgmeta_oid keys with _ so that PGLog::read_log_and_missing() can
83 const string
infover_key("_infover");
84 const string
info_key("_info");
85 const string
biginfo_key("_biginfo");
86 const string
epoch_key("_epoch");
87 const string
fastinfo_key("_fastinfo");
90 static ostream
& _prefix(std::ostream
*_dout
, T
*t
)
92 return *_dout
<< t
->gen_prefix();
95 MEMPOOL_DEFINE_OBJECT_FACTORY(PG::CephPeeringEvt
, pg_peering_evt
, osd
);
97 void PGStateHistory::enter(PG
* pg
, const utime_t entime
, const char* state
)
99 // Ignore trimming state machine for now
100 if (::strstr(state
, "Trimming") != NULL
) {
102 } else if (pi
!= nullptr) {
103 pi
->enter_state(entime
, state
);
105 // Store current state since we can't reliably take the PG lock here
106 if ( tmppi
== nullptr) {
107 tmppi
= std::unique_ptr
<PGStateInstance
>(new PGStateInstance
);
111 tmppi
->enter_state(entime
, state
);
115 void PGStateHistory::exit(const char* state
) {
116 // Ignore trimming state machine for now
117 // Do nothing if PG is being destroyed!
118 if (::strstr(state
, "Trimming") != NULL
|| pg_in_destructor
) {
121 bool ilocked
= false;
122 if(!thispg
->is_locked()) {
127 buffer
.push_back(std::unique_ptr
<PGStateInstance
>(tmppi
.release()));
128 pi
= buffer
.back().get();
129 pi
->setepoch(thispg
->get_osdmap()->get_epoch());
132 pi
->exit_state(ceph_clock_now());
133 if (::strcmp(state
, "Reset") == 0) {
142 void PGStateHistory::dump(Formatter
* f
) const {
143 f
->open_array_section("history");
144 for (auto pi
= buffer
.begin(); pi
!= buffer
.end(); ++pi
) {
145 f
->open_object_section("states");
146 f
->dump_stream("epoch") << (*pi
)->this_epoch
;
147 for (auto she
: (*pi
)->state_history
) {
148 f
->dump_string("state", std::get
<2>(she
));
149 f
->dump_stream("enter") << std::get
<0>(she
);
150 f
->dump_stream("exit") << std::get
<1>(she
);
157 void PG::get(const char* tag
)
161 Mutex::Locker
l(_ref_id_lock
);
166 void PG::put(const char* tag
)
170 Mutex::Locker
l(_ref_id_lock
);
171 auto tag_counts_entry
= _tag_counts
.find(tag
);
172 assert(tag_counts_entry
!= _tag_counts
.end());
173 --tag_counts_entry
->second
;
174 if (tag_counts_entry
->second
== 0) {
175 _tag_counts
.erase(tag_counts_entry
);
184 uint64_t PG::get_with_id()
187 Mutex::Locker
l(_ref_id_lock
);
188 uint64_t id
= ++_ref_id
;
192 dout(20) << __func__
<< ": " << info
.pgid
<< " got id " << id
<< " (new) ref==" << ref
<< dendl
;
193 assert(!_live_ids
.count(id
));
194 _live_ids
.insert(make_pair(id
, ss
.str()));
198 void PG::put_with_id(uint64_t id
)
200 dout(20) << __func__
<< ": " << info
.pgid
<< " put id " << id
<< " (current) ref==" << ref
<< dendl
;
202 Mutex::Locker
l(_ref_id_lock
);
203 assert(_live_ids
.count(id
));
210 void PG::dump_live_ids()
212 Mutex::Locker
l(_ref_id_lock
);
213 dout(0) << "\t" << __func__
<< ": " << info
.pgid
<< " live ids:" << dendl
;
214 for (map
<uint64_t, string
>::iterator i
= _live_ids
.begin();
215 i
!= _live_ids
.end();
217 dout(0) << "\t\tid: " << *i
<< dendl
;
219 dout(0) << "\t" << __func__
<< ": " << info
.pgid
<< " live tags:" << dendl
;
220 for (map
<string
, uint64_t>::iterator i
= _tag_counts
.begin();
221 i
!= _tag_counts
.end();
223 dout(0) << "\t\tid: " << *i
<< dendl
;
229 void PGPool::update(OSDMapRef map
)
231 const pg_pool_t
*pi
= map
->get_pg_pool(id
);
235 name
= map
->get_pool_name(id
);
236 bool updated
= false;
237 if ((map
->get_epoch() != cached_epoch
+ 1) ||
238 (pi
->get_snap_epoch() == map
->get_epoch())) {
240 if (pi
->maybe_updated_removed_snaps(cached_removed_snaps
)) {
241 pi
->build_removed_snaps(newly_removed_snaps
);
242 if (cached_removed_snaps
.subset_of(newly_removed_snaps
)) {
243 interval_set
<snapid_t
> removed_snaps
= newly_removed_snaps
;
244 newly_removed_snaps
.subtract(cached_removed_snaps
);
245 cached_removed_snaps
.swap(removed_snaps
);
247 lgeneric_subdout(cct
, osd
, 0) << __func__
248 << " cached_removed_snaps shrank from " << cached_removed_snaps
249 << " to " << newly_removed_snaps
<< dendl
;
250 cached_removed_snaps
.swap(newly_removed_snaps
);
251 newly_removed_snaps
.clear();
254 newly_removed_snaps
.clear();
255 snapc
= pi
->get_snap_context();
257 /* 1) map->get_epoch() == cached_epoch + 1 &&
258 * 2) pi->get_snap_epoch() != map->get_epoch()
260 * From the if branch, 1 && 2 must be true. From 2, we know that
261 * this map didn't change the set of removed snaps. From 1, we
262 * know that our cached_removed_snaps matches the previous map.
263 * Thus, from 1 && 2, cached_removed snaps matches the current
264 * set of removed snaps and all we have to do is clear
265 * newly_removed_snaps.
267 newly_removed_snaps
.clear();
269 cached_epoch
= map
->get_epoch();
270 lgeneric_subdout(cct
, osd
, 20)
271 << "PGPool::update cached_removed_snaps "
272 << cached_removed_snaps
273 << " newly_removed_snaps "
274 << newly_removed_snaps
275 << " snapc " << snapc
276 << (updated
? " (updated)":" (no change)")
280 PG::PG(OSDService
*o
, OSDMapRef curmap
,
281 const PGPool
&_pool
, spg_t p
) :
284 osdriver(osd
->store
, coll_t(), OSD::make_snapmapper_oid()),
289 p
.get_split_bits(curmap
->get_pg_num(_pool
.id
)),
292 osdmap_ref(curmap
), last_persisted_osdmap_ref(curmap
), pool(_pool
),
295 _ref_id_lock("PG::_ref_id_lock"), _ref_id(0),
298 trace_endpoint("0.0.0.0", 0, "PG"),
299 dirty_info(false), dirty_big_info(false),
304 pgmeta_oid(p
.make_pgmeta_oid()),
307 curmap
->get_pools().at(p
.pgid
.pool()).ec_pool(),
309 stat_queue_item(this),
311 recovery_queued(false),
312 recovery_ops_active(0),
316 pg_whoami(osd
->whoami
, p
.shard
),
318 last_peering_reset(0),
319 heartbeat_peer_lock("PG::heartbeat_peer_lock"),
320 backfill_reserved(false),
321 backfill_reserving(false),
322 flushes_in_progress(0),
323 pg_stats_publish_lock("PG::pg_stats_publish_lock"),
324 pg_stats_publish_valid(false),
325 osr(osd
->osr_registry
.lookup_or_create(p
, (stringify(p
)))),
326 finish_sync_event(NULL
),
327 backoff_lock("PG::backoff_lock"),
328 scrub_after_recovery(false),
330 recovery_state(this),
332 peer_features(CEPH_FEATURES_SUPPORTED_DEFAULT
),
333 acting_features(CEPH_FEATURES_SUPPORTED_DEFAULT
),
334 upacting_features(CEPH_FEATURES_SUPPORTED_DEFAULT
),
338 osd
->add_pgid(p
, this);
341 std::stringstream ss
;
342 ss
<< "PG " << info
.pgid
;
343 trace_endpoint
.copy_name(ss
.str());
350 pgstate_history
.set_pg_in_destructor();
352 osd
->remove_pgid(info
.pgid
, this);
356 void PG::lock_suspend_timeout(ThreadPool::TPHandle
&handle
)
358 handle
.suspend_tp_timeout();
360 handle
.reset_tp_timeout();
363 void PG::lock(bool no_lockdep
) const
365 _lock
.Lock(no_lockdep
);
366 // if we have unrecorded dirty state with the lock dropped, there is a bug
368 assert(!dirty_big_info
);
370 dout(30) << "lock" << dendl
;
373 std::string
PG::gen_prefix() const
376 OSDMapRef mapref
= osdmap_ref
;
377 if (_lock
.is_locked_by_me()) {
378 out
<< "osd." << osd
->whoami
379 << " pg_epoch: " << (mapref
? mapref
->get_epoch():0)
380 << " " << *this << " ";
382 out
<< "osd." << osd
->whoami
383 << " pg_epoch: " << (mapref
? mapref
->get_epoch():0)
384 << " pg[" << info
.pgid
<< "(unlocked)] ";
389 /********* PG **********/
391 void PG::proc_master_log(
392 ObjectStore::Transaction
& t
, pg_info_t
&oinfo
,
393 pg_log_t
&olog
, pg_missing_t
& omissing
, pg_shard_t from
)
395 dout(10) << "proc_master_log for osd." << from
<< ": "
396 << olog
<< " " << omissing
<< dendl
;
397 assert(!is_peered() && is_primary());
399 // merge log into our own log to build master log. no need to
400 // make any adjustments to their missing map; we are taking their
401 // log to be authoritative (i.e., their entries are by definitely
403 merge_log(t
, oinfo
, olog
, from
);
404 peer_info
[from
] = oinfo
;
405 dout(10) << " peer osd." << from
<< " now " << oinfo
<< " " << omissing
<< dendl
;
406 might_have_unfound
.insert(from
);
408 // See doc/dev/osd_internals/last_epoch_started
409 if (oinfo
.last_epoch_started
> info
.last_epoch_started
) {
410 info
.last_epoch_started
= oinfo
.last_epoch_started
;
413 if (oinfo
.last_interval_started
> info
.last_interval_started
) {
414 info
.last_interval_started
= oinfo
.last_interval_started
;
417 update_history(oinfo
.history
);
418 assert(cct
->_conf
->osd_find_best_info_ignore_history_les
||
419 info
.last_epoch_started
>= info
.history
.last_epoch_started
);
421 peer_missing
[from
].claim(omissing
);
424 void PG::proc_replica_log(
426 const pg_log_t
&olog
,
427 pg_missing_t
& omissing
,
430 dout(10) << "proc_replica_log for osd." << from
<< ": "
431 << oinfo
<< " " << olog
<< " " << omissing
<< dendl
;
433 pg_log
.proc_replica_log(oinfo
, olog
, omissing
, from
);
435 peer_info
[from
] = oinfo
;
436 dout(10) << " peer osd." << from
<< " now " << oinfo
<< " " << omissing
<< dendl
;
437 might_have_unfound
.insert(from
);
439 for (map
<hobject_t
, pg_missing_item
>::const_iterator i
=
440 omissing
.get_items().begin();
441 i
!= omissing
.get_items().end();
443 dout(20) << " after missing " << i
->first
<< " need " << i
->second
.need
444 << " have " << i
->second
.have
<< dendl
;
446 peer_missing
[from
].claim(omissing
);
449 bool PG::proc_replica_info(
450 pg_shard_t from
, const pg_info_t
&oinfo
, epoch_t send_epoch
)
452 map
<pg_shard_t
, pg_info_t
>::iterator p
= peer_info
.find(from
);
453 if (p
!= peer_info
.end() && p
->second
.last_update
== oinfo
.last_update
) {
454 dout(10) << " got dup osd." << from
<< " info " << oinfo
<< ", identical to ours" << dendl
;
458 if (!get_osdmap()->has_been_up_since(from
.osd
, send_epoch
)) {
459 dout(10) << " got info " << oinfo
<< " from down osd." << from
460 << " discarding" << dendl
;
464 dout(10) << " got osd." << from
<< " " << oinfo
<< dendl
;
465 assert(is_primary());
466 peer_info
[from
] = oinfo
;
467 might_have_unfound
.insert(from
);
469 update_history(oinfo
.history
);
472 if (!is_up(from
) && !is_acting(from
)) {
473 dout(10) << " osd." << from
<< " has stray content: " << oinfo
<< dendl
;
474 stray_set
.insert(from
);
480 // was this a new info? if so, update peers!
481 if (p
== peer_info
.end())
482 update_heartbeat_peers();
487 void PG::remove_snap_mapped_object(
488 ObjectStore::Transaction
&t
, const hobject_t
&soid
)
492 ghobject_t(soid
, ghobject_t::NO_GEN
, pg_whoami
.shard
));
493 clear_object_snap_mapping(&t
, soid
);
496 void PG::clear_object_snap_mapping(
497 ObjectStore::Transaction
*t
, const hobject_t
&soid
)
499 OSDriver::OSTransaction
_t(osdriver
.get_transaction(t
));
500 if (soid
.snap
< CEPH_MAXSNAP
) {
501 int r
= snap_mapper
.remove_oid(
504 if (!(r
== 0 || r
== -ENOENT
)) {
505 derr
<< __func__
<< ": remove_oid returned " << cpp_strerror(r
) << dendl
;
511 void PG::update_object_snap_mapping(
512 ObjectStore::Transaction
*t
, const hobject_t
&soid
, const set
<snapid_t
> &snaps
)
514 OSDriver::OSTransaction
_t(osdriver
.get_transaction(t
));
515 assert(soid
.snap
< CEPH_MAXSNAP
);
516 int r
= snap_mapper
.remove_oid(
519 if (!(r
== 0 || r
== -ENOENT
)) {
520 derr
<< __func__
<< ": remove_oid returned " << cpp_strerror(r
) << dendl
;
530 ObjectStore::Transaction
& t
, pg_info_t
&oinfo
, pg_log_t
&olog
, pg_shard_t from
)
532 PGLogEntryHandler rollbacker
{this, &t
};
534 oinfo
, olog
, from
, info
, &rollbacker
, dirty_info
, dirty_big_info
);
537 void PG::rewind_divergent_log(ObjectStore::Transaction
& t
, eversion_t newhead
)
539 PGLogEntryHandler rollbacker
{this, &t
};
540 pg_log
.rewind_divergent_log(
541 newhead
, info
, &rollbacker
, dirty_info
, dirty_big_info
);
545 * Process information from a replica to determine if it could have any
546 * objects that i need.
548 * TODO: if the missing set becomes very large, this could get expensive.
549 * Instead, we probably want to just iterate over our unfound set.
551 bool PG::search_for_missing(
552 const pg_info_t
&oinfo
, const pg_missing_t
&omissing
,
556 uint64_t num_unfound_before
= missing_loc
.num_unfound();
557 bool found_missing
= missing_loc
.add_source_info(
558 from
, oinfo
, omissing
, ctx
->handle
);
559 if (found_missing
&& num_unfound_before
!= missing_loc
.num_unfound())
560 publish_stats_to_osd();
561 // avoid doing this if the peer is empty. This is abit of paranoia
562 // to avoid doing something rash if add_source_info() above
563 // incorrectly decided we found something new. (if the peer has
564 // last_update=0'0 that's impossible.)
566 (get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, NULL
) &
567 CEPH_FEATURE_OSD_ERASURE_CODES
) &&
568 oinfo
.last_update
!= eversion_t()) {
569 pg_info_t
tinfo(oinfo
);
570 tinfo
.pgid
.shard
= pg_whoami
.shard
;
571 (*(ctx
->info_map
))[from
.osd
].push_back(
574 from
.shard
, pg_whoami
.shard
,
575 get_osdmap()->get_epoch(),
576 get_osdmap()->get_epoch(),
580 return found_missing
;
586 bool PG::MissingLoc::readable_with_acting(
587 const hobject_t
&hoid
,
588 const set
<pg_shard_t
> &acting
) const {
589 if (!needs_recovery(hoid
))
591 if (is_deleted(hoid
))
593 auto missing_loc_entry
= missing_loc
.find(hoid
);
594 if (missing_loc_entry
== missing_loc
.end())
596 const set
<pg_shard_t
> &locs
= missing_loc_entry
->second
;
597 ldout(pg
->cct
, 10) << __func__
<< ": locs:" << locs
<< dendl
;
598 set
<pg_shard_t
> have_acting
;
599 for (set
<pg_shard_t
>::const_iterator i
= locs
.begin();
602 if (acting
.count(*i
))
603 have_acting
.insert(*i
);
605 return (*is_readable
)(have_acting
);
608 void PG::MissingLoc::add_batch_sources_info(
609 const set
<pg_shard_t
> &sources
, ThreadPool::TPHandle
* handle
)
611 ldout(pg
->cct
, 10) << __func__
<< ": adding sources in batch "
612 << sources
.size() << dendl
;
614 for (map
<hobject_t
, pg_missing_item
>::const_iterator i
= needs_recovery_map
.begin();
615 i
!= needs_recovery_map
.end();
617 if (handle
&& ++loop
>= pg
->cct
->_conf
->osd_loop_before_reset_tphandle
) {
618 handle
->reset_tp_timeout();
621 if (i
->second
.is_delete())
624 auto p
= missing_loc
.find(i
->first
);
625 if (p
== missing_loc
.end()) {
626 p
= missing_loc
.emplace(i
->first
, set
<pg_shard_t
>()).first
;
628 _dec_count(p
->second
);
630 missing_loc
[i
->first
].insert(sources
.begin(), sources
.end());
631 missing_loc_sources
.insert(sources
.begin(), sources
.end());
632 _inc_count(p
->second
);
637 bool PG::MissingLoc::add_source_info(
639 const pg_info_t
&oinfo
,
640 const pg_missing_t
&omissing
,
641 ThreadPool::TPHandle
* handle
)
643 bool found_missing
= false;
646 for (map
<hobject_t
,pg_missing_item
>::const_iterator p
= needs_recovery_map
.begin();
647 p
!= needs_recovery_map
.end();
649 const hobject_t
&soid(p
->first
);
650 eversion_t need
= p
->second
.need
;
651 if (handle
&& ++loop
>= pg
->cct
->_conf
->osd_loop_before_reset_tphandle
) {
652 handle
->reset_tp_timeout();
655 if (p
->second
.is_delete()) {
656 ldout(pg
->cct
, 10) << __func__
<< " " << soid
657 << " delete, ignoring source" << dendl
;
660 if (oinfo
.last_update
< need
) {
661 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
662 << " also missing on osd." << fromosd
663 << " (last_update " << oinfo
.last_update
664 << " < needed " << need
<< ")" << dendl
;
667 if (!oinfo
.last_backfill
.is_max() &&
668 !oinfo
.last_backfill_bitwise
) {
669 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
670 << " also missing on osd." << fromosd
671 << " (last_backfill " << oinfo
.last_backfill
672 << " but with wrong sort order)"
676 if (p
->first
>= oinfo
.last_backfill
) {
677 // FIXME: this is _probably_ true, although it could conceivably
678 // be in the undefined region! Hmm!
679 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
680 << " also missing on osd." << fromosd
681 << " (past last_backfill " << oinfo
.last_backfill
685 if (oinfo
.last_complete
< need
) {
686 if (omissing
.is_missing(soid
)) {
687 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
688 << " also missing on osd." << fromosd
<< dendl
;
693 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
694 << " is on osd." << fromosd
<< dendl
;
696 missing_loc_sources
.insert(fromosd
);
698 auto p
= missing_loc
.find(soid
);
699 if (p
== missing_loc
.end()) {
700 p
= missing_loc
.emplace(soid
, set
<pg_shard_t
>()).first
;
702 _dec_count(p
->second
);
704 p
->second
.insert(fromosd
);
705 _inc_count(p
->second
);
708 found_missing
= true;
711 ldout(pg
->cct
, 20) << "needs_recovery_map missing " << needs_recovery_map
713 return found_missing
;
716 void PG::MissingLoc::check_recovery_sources(const OSDMapRef
& osdmap
)
718 set
<pg_shard_t
> now_down
;
719 for (set
<pg_shard_t
>::iterator p
= missing_loc_sources
.begin();
720 p
!= missing_loc_sources
.end();
722 if (osdmap
->is_up(p
->osd
)) {
726 ldout(pg
->cct
, 10) << __func__
<< " source osd." << *p
<< " now down" << dendl
;
728 missing_loc_sources
.erase(p
++);
731 if (now_down
.empty()) {
732 ldout(pg
->cct
, 10) << __func__
<< " no source osds (" << missing_loc_sources
<< ") went down" << dendl
;
734 ldout(pg
->cct
, 10) << __func__
<< " sources osds " << now_down
<< " now down, remaining sources are "
735 << missing_loc_sources
<< dendl
;
737 // filter missing_loc
738 map
<hobject_t
, set
<pg_shard_t
>>::iterator p
= missing_loc
.begin();
739 while (p
!= missing_loc
.end()) {
740 set
<pg_shard_t
>::iterator q
= p
->second
.begin();
741 bool changed
= false;
742 while (q
!= p
->second
.end()) {
743 if (now_down
.count(*q
)) {
746 _dec_count(p
->second
);
748 p
->second
.erase(q
++);
753 if (p
->second
.empty()) {
754 missing_loc
.erase(p
++);
757 _inc_count(p
->second
);
765 void PG::discover_all_missing(map
<int, map
<spg_t
,pg_query_t
> > &query_map
)
767 auto &missing
= pg_log
.get_missing();
768 uint64_t unfound
= get_num_unfound();
770 dout(10) << __func__
<< " "
771 << missing
.num_missing() << " missing, "
772 << unfound
<< " unfound"
775 std::set
<pg_shard_t
>::const_iterator m
= might_have_unfound
.begin();
776 std::set
<pg_shard_t
>::const_iterator mend
= might_have_unfound
.end();
777 for (; m
!= mend
; ++m
) {
780 if (!get_osdmap()->is_up(peer
.osd
)) {
781 dout(20) << __func__
<< " skipping down osd." << peer
<< dendl
;
785 map
<pg_shard_t
, pg_info_t
>::const_iterator iter
= peer_info
.find(peer
);
786 if (iter
!= peer_info
.end() &&
787 (iter
->second
.is_empty() || iter
->second
.dne())) {
788 // ignore empty peers
792 // If we've requested any of this stuff, the pg_missing_t information
793 // should be on its way.
794 // TODO: coalsce requested_* into a single data structure
795 if (peer_missing
.find(peer
) != peer_missing
.end()) {
796 dout(20) << __func__
<< ": osd." << peer
797 << ": we already have pg_missing_t" << dendl
;
800 if (peer_log_requested
.find(peer
) != peer_log_requested
.end()) {
801 dout(20) << __func__
<< ": osd." << peer
802 << ": in peer_log_requested" << dendl
;
805 if (peer_missing_requested
.find(peer
) != peer_missing_requested
.end()) {
806 dout(20) << __func__
<< ": osd." << peer
807 << ": in peer_missing_requested" << dendl
;
812 dout(10) << __func__
<< ": osd." << peer
<< ": requesting pg_missing_t"
814 peer_missing_requested
.insert(peer
);
815 query_map
[peer
.osd
][spg_t(info
.pgid
.pgid
, peer
.shard
)] =
818 peer
.shard
, pg_whoami
.shard
,
819 info
.history
, get_osdmap()->get_epoch());
823 /******* PG ***********/
824 bool PG::needs_recovery() const
826 assert(is_primary());
828 auto &missing
= pg_log
.get_missing();
830 if (missing
.num_missing()) {
831 dout(10) << __func__
<< " primary has " << missing
.num_missing()
832 << " missing" << dendl
;
836 assert(!actingbackfill
.empty());
837 set
<pg_shard_t
>::const_iterator end
= actingbackfill
.end();
838 set
<pg_shard_t
>::const_iterator a
= actingbackfill
.begin();
839 for (; a
!= end
; ++a
) {
840 if (*a
== get_primary()) continue;
841 pg_shard_t peer
= *a
;
842 map
<pg_shard_t
, pg_missing_t
>::const_iterator pm
= peer_missing
.find(peer
);
843 if (pm
== peer_missing
.end()) {
844 dout(10) << __func__
<< " osd." << peer
<< " doesn't have missing set"
848 if (pm
->second
.num_missing()) {
849 dout(10) << __func__
<< " osd." << peer
<< " has "
850 << pm
->second
.num_missing() << " missing" << dendl
;
855 dout(10) << __func__
<< " is recovered" << dendl
;
859 bool PG::needs_backfill() const
861 assert(is_primary());
863 // We can assume that only possible osds that need backfill
864 // are on the backfill_targets vector nodes.
865 set
<pg_shard_t
>::const_iterator end
= backfill_targets
.end();
866 set
<pg_shard_t
>::const_iterator a
= backfill_targets
.begin();
867 for (; a
!= end
; ++a
) {
868 pg_shard_t peer
= *a
;
869 map
<pg_shard_t
, pg_info_t
>::const_iterator pi
= peer_info
.find(peer
);
870 if (!pi
->second
.last_backfill
.is_max()) {
871 dout(10) << __func__
<< " osd." << peer
<< " has last_backfill " << pi
->second
.last_backfill
<< dendl
;
876 dout(10) << __func__
<< " does not need backfill" << dendl
;
881 void PG::check_past_interval_bounds() const
883 auto rpib
= get_required_past_interval_bounds(
885 osd
->get_superblock().oldest_map
);
886 if (rpib
.first
>= rpib
.second
) {
887 if (!past_intervals
.empty()) {
888 osd
->clog
->error() << info
.pgid
<< " required past_interval bounds are"
889 << " empty [" << rpib
<< ") but past_intervals is not: "
891 derr
<< info
.pgid
<< " required past_interval bounds are"
892 << " empty [" << rpib
<< ") but past_intervals is not: "
893 << past_intervals
<< dendl
;
896 if (past_intervals
.empty()) {
897 osd
->clog
->error() << info
.pgid
<< " required past_interval bounds are"
898 << " not empty [" << rpib
<< ") but past_intervals "
899 << past_intervals
<< " is empty";
900 derr
<< info
.pgid
<< " required past_interval bounds are"
901 << " not empty [" << rpib
<< ") but past_intervals "
902 << past_intervals
<< " is empty" << dendl
;
903 assert(!past_intervals
.empty());
906 auto apib
= past_intervals
.get_bounds();
907 if (apib
.first
> rpib
.first
) {
908 osd
->clog
->error() << info
.pgid
<< " past_intervals [" << apib
909 << ") start interval does not contain the required"
910 << " bound [" << rpib
<< ") start";
911 derr
<< info
.pgid
<< " past_intervals [" << apib
912 << ") start interval does not contain the required"
913 << " bound [" << rpib
<< ") start" << dendl
;
914 assert(0 == "past_interval start interval mismatch");
916 if (apib
.second
!= rpib
.second
) {
917 osd
->clog
->error() << info
.pgid
<< " past_interal bound [" << apib
918 << ") end does not match required [" << rpib
920 derr
<< info
.pgid
<< " past_interal bound [" << apib
921 << ") end does not match required [" << rpib
923 assert(0 == "past_interval end mismatch");
928 bool PG::adjust_need_up_thru(const OSDMapRef osdmap
)
930 epoch_t up_thru
= osdmap
->get_up_thru(osd
->whoami
);
932 up_thru
>= info
.history
.same_interval_since
) {
933 dout(10) << "adjust_need_up_thru now " << up_thru
<< ", need_up_thru now false" << dendl
;
934 need_up_thru
= false;
940 void PG::remove_down_peer_info(const OSDMapRef osdmap
)
942 // Remove any downed osds from peer_info
943 bool removed
= false;
944 map
<pg_shard_t
, pg_info_t
>::iterator p
= peer_info
.begin();
945 while (p
!= peer_info
.end()) {
946 if (!osdmap
->is_up(p
->first
.osd
)) {
947 dout(10) << " dropping down osd." << p
->first
<< " info " << p
->second
<< dendl
;
948 peer_missing
.erase(p
->first
);
949 peer_log_requested
.erase(p
->first
);
950 peer_missing_requested
.erase(p
->first
);
951 peer_info
.erase(p
++);
957 // if we removed anyone, update peers (which include peer_info)
959 update_heartbeat_peers();
960 check_recovery_sources(osdmap
);
964 * Returns true unless there is a non-lost OSD in might_have_unfound.
966 bool PG::all_unfound_are_queried_or_lost(const OSDMapRef osdmap
) const
968 assert(is_primary());
970 set
<pg_shard_t
>::const_iterator peer
= might_have_unfound
.begin();
971 set
<pg_shard_t
>::const_iterator mend
= might_have_unfound
.end();
972 for (; peer
!= mend
; ++peer
) {
973 if (peer_missing
.count(*peer
))
975 map
<pg_shard_t
, pg_info_t
>::const_iterator iter
= peer_info
.find(*peer
);
976 if (iter
!= peer_info
.end() &&
977 (iter
->second
.is_empty() || iter
->second
.dne()))
979 if (!osdmap
->exists(peer
->osd
))
981 const osd_info_t
&osd_info(osdmap
->get_info(peer
->osd
));
982 if (osd_info
.lost_at
<= osd_info
.up_from
) {
983 // If there is even one OSD in might_have_unfound that isn't lost, we
984 // still might retrieve our unfound.
988 dout(10) << "all_unfound_are_queried_or_lost all of might_have_unfound " << might_have_unfound
989 << " have been queried or are marked lost" << dendl
;
993 PastIntervals::PriorSet
PG::build_prior()
997 for (map
<pg_shard_t
,pg_info_t
>::iterator it
= peer_info
.begin();
998 it
!= peer_info
.end();
1000 assert(info
.history
.last_epoch_started
>= it
->second
.history
.last_epoch_started
);
1004 const OSDMap
&osdmap
= *get_osdmap();
1005 PastIntervals::PriorSet prior
= past_intervals
.get_prior_set(
1006 pool
.info
.ec_pool(),
1007 info
.history
.last_epoch_started
,
1008 get_pgbackend()->get_is_recoverable_predicate(),
1009 [&](epoch_t start
, int osd
, epoch_t
*lost_at
) {
1010 const osd_info_t
*pinfo
= 0;
1011 if (osdmap
.exists(osd
)) {
1012 pinfo
= &osdmap
.get_info(osd
);
1014 *lost_at
= pinfo
->lost_at
;
1017 if (osdmap
.is_up(osd
)) {
1018 return PastIntervals::UP
;
1019 } else if (!pinfo
) {
1020 return PastIntervals::DNE
;
1021 } else if (pinfo
->lost_at
> start
) {
1022 return PastIntervals::LOST
;
1024 return PastIntervals::DOWN
;
1031 if (prior
.pg_down
) {
1032 state_set(PG_STATE_DOWN
);
1035 if (get_osdmap()->get_up_thru(osd
->whoami
) < info
.history
.same_interval_since
) {
1036 dout(10) << "up_thru " << get_osdmap()->get_up_thru(osd
->whoami
)
1037 << " < same_since " << info
.history
.same_interval_since
1038 << ", must notify monitor" << dendl
;
1039 need_up_thru
= true;
1041 dout(10) << "up_thru " << get_osdmap()->get_up_thru(osd
->whoami
)
1042 << " >= same_since " << info
.history
.same_interval_since
1043 << ", all is well" << dendl
;
1044 need_up_thru
= false;
1046 set_probe_targets(prior
.probe
);
1050 void PG::clear_primary_state()
1052 dout(10) << "clear_primary_state" << dendl
;
1054 // clear peering state
1056 peer_log_requested
.clear();
1057 peer_missing_requested
.clear();
1059 peer_missing
.clear();
1060 need_up_thru
= false;
1061 peer_last_complete_ondisk
.clear();
1062 peer_activated
.clear();
1063 min_last_complete_ondisk
= eversion_t();
1064 pg_trim_to
= eversion_t();
1065 might_have_unfound
.clear();
1066 projected_log
= PGLog::IndexedLog();
1068 last_update_ondisk
= eversion_t();
1072 finish_sync_event
= 0; // so that _finish_recovery doesn't go off in another thread
1074 missing_loc
.clear();
1076 release_pg_backoffs();
1078 pg_log
.reset_recovery_pointers();
1080 scrubber
.reserved_peers
.clear();
1081 scrub_after_recovery
= false;
1086 PG::Scrubber::Scrubber()
1087 : reserved(false), reserve_failed(false),
1090 shallow_errors(0), deep_errors(0), fixed(0),
1091 must_scrub(false), must_deep_scrub(false), must_repair(false),
1093 num_digest_updates_pending(0),
1098 PG::Scrubber::~Scrubber() {}
1103 * Returns an iterator to the best info in infos sorted by:
1104 * 1) Prefer newer last_update
1105 * 2) Prefer longer tail if it brings another info into contiguity
1106 * 3) Prefer current primary
1108 map
<pg_shard_t
, pg_info_t
>::const_iterator
PG::find_best_info(
1109 const map
<pg_shard_t
, pg_info_t
> &infos
,
1110 bool restrict_to_up_acting
,
1111 bool *history_les_bound
) const
1113 assert(history_les_bound
);
1114 /* See doc/dev/osd_internals/last_epoch_started.rst before attempting
1115 * to make changes to this process. Also, make sure to update it
1116 * when you find bugs! */
1117 eversion_t min_last_update_acceptable
= eversion_t::max();
1118 epoch_t max_last_epoch_started_found
= 0;
1119 for (map
<pg_shard_t
, pg_info_t
>::const_iterator i
= infos
.begin();
1122 if (!cct
->_conf
->osd_find_best_info_ignore_history_les
&&
1123 max_last_epoch_started_found
< i
->second
.history
.last_epoch_started
) {
1124 *history_les_bound
= true;
1125 max_last_epoch_started_found
= i
->second
.history
.last_epoch_started
;
1127 if (!i
->second
.is_incomplete() &&
1128 max_last_epoch_started_found
< i
->second
.last_epoch_started
) {
1129 max_last_epoch_started_found
= i
->second
.last_epoch_started
;
1132 for (map
<pg_shard_t
, pg_info_t
>::const_iterator i
= infos
.begin();
1135 if (max_last_epoch_started_found
<= i
->second
.last_epoch_started
) {
1136 if (min_last_update_acceptable
> i
->second
.last_update
)
1137 min_last_update_acceptable
= i
->second
.last_update
;
1140 if (min_last_update_acceptable
== eversion_t::max())
1143 map
<pg_shard_t
, pg_info_t
>::const_iterator best
= infos
.end();
1144 // find osd with newest last_update (oldest for ec_pool).
1145 // if there are multiples, prefer
1146 // - a longer tail, if it brings another peer into log contiguity
1147 // - the current primary
1148 for (map
<pg_shard_t
, pg_info_t
>::const_iterator p
= infos
.begin();
1151 if (restrict_to_up_acting
&& !is_up(p
->first
) &&
1152 !is_acting(p
->first
))
1154 // Only consider peers with last_update >= min_last_update_acceptable
1155 if (p
->second
.last_update
< min_last_update_acceptable
)
1157 // Disqualify anyone with a too old last_epoch_started
1158 if (p
->second
.last_epoch_started
< max_last_epoch_started_found
)
1160 // Disqualify anyone who is incomplete (not fully backfilled)
1161 if (p
->second
.is_incomplete())
1163 if (best
== infos
.end()) {
1167 // Prefer newer last_update
1168 if (pool
.info
.require_rollback()) {
1169 if (p
->second
.last_update
> best
->second
.last_update
)
1171 if (p
->second
.last_update
< best
->second
.last_update
) {
1176 if (p
->second
.last_update
< best
->second
.last_update
)
1178 if (p
->second
.last_update
> best
->second
.last_update
) {
1184 // Prefer longer tail
1185 if (p
->second
.log_tail
> best
->second
.log_tail
) {
1187 } else if (p
->second
.log_tail
< best
->second
.log_tail
) {
1192 // prefer current primary (usually the caller), all things being equal
1193 if (p
->first
== pg_whoami
) {
1194 dout(10) << "calc_acting prefer osd." << p
->first
1195 << " because it is current primary" << dendl
;
1203 void PG::calc_ec_acting(
1204 map
<pg_shard_t
, pg_info_t
>::const_iterator auth_log_shard
,
1206 const vector
<int> &acting
,
1207 pg_shard_t acting_primary
,
1208 const vector
<int> &up
,
1209 pg_shard_t up_primary
,
1210 const map
<pg_shard_t
, pg_info_t
> &all_info
,
1211 bool restrict_to_up_acting
,
1213 set
<pg_shard_t
> *backfill
,
1214 set
<pg_shard_t
> *acting_backfill
,
1215 pg_shard_t
*want_primary
,
1218 vector
<int> want(size
, CRUSH_ITEM_NONE
);
1219 map
<shard_id_t
, set
<pg_shard_t
> > all_info_by_shard
;
1220 unsigned usable
= 0;
1221 for (map
<pg_shard_t
, pg_info_t
>::const_iterator i
= all_info
.begin();
1222 i
!= all_info
.end();
1224 all_info_by_shard
[i
->first
.shard
].insert(i
->first
);
1226 for (uint8_t i
= 0; i
< want
.size(); ++i
) {
1227 ss
<< "For position " << (unsigned)i
<< ": ";
1228 if (up
.size() > (unsigned)i
&& up
[i
] != CRUSH_ITEM_NONE
&&
1229 !all_info
.find(pg_shard_t(up
[i
], shard_id_t(i
)))->second
.is_incomplete() &&
1230 all_info
.find(pg_shard_t(up
[i
], shard_id_t(i
)))->second
.last_update
>=
1231 auth_log_shard
->second
.log_tail
) {
1232 ss
<< " selecting up[i]: " << pg_shard_t(up
[i
], shard_id_t(i
)) << std::endl
;
1237 if (up
.size() > (unsigned)i
&& up
[i
] != CRUSH_ITEM_NONE
) {
1238 ss
<< " backfilling up[i]: " << pg_shard_t(up
[i
], shard_id_t(i
))
1240 backfill
->insert(pg_shard_t(up
[i
], shard_id_t(i
)));
1243 if (acting
.size() > (unsigned)i
&& acting
[i
] != CRUSH_ITEM_NONE
&&
1244 !all_info
.find(pg_shard_t(acting
[i
], shard_id_t(i
)))->second
.is_incomplete() &&
1245 all_info
.find(pg_shard_t(acting
[i
], shard_id_t(i
)))->second
.last_update
>=
1246 auth_log_shard
->second
.log_tail
) {
1247 ss
<< " selecting acting[i]: " << pg_shard_t(acting
[i
], shard_id_t(i
)) << std::endl
;
1248 want
[i
] = acting
[i
];
1250 } else if (!restrict_to_up_acting
) {
1251 for (set
<pg_shard_t
>::iterator j
= all_info_by_shard
[shard_id_t(i
)].begin();
1252 j
!= all_info_by_shard
[shard_id_t(i
)].end();
1254 assert(j
->shard
== i
);
1255 if (!all_info
.find(*j
)->second
.is_incomplete() &&
1256 all_info
.find(*j
)->second
.last_update
>=
1257 auth_log_shard
->second
.log_tail
) {
1258 ss
<< " selecting stray: " << *j
<< std::endl
;
1264 if (want
[i
] == CRUSH_ITEM_NONE
)
1265 ss
<< " failed to fill position " << (int)i
<< std::endl
;
1269 bool found_primary
= false;
1270 for (uint8_t i
= 0; i
< want
.size(); ++i
) {
1271 if (want
[i
] != CRUSH_ITEM_NONE
) {
1272 acting_backfill
->insert(pg_shard_t(want
[i
], shard_id_t(i
)));
1273 if (!found_primary
) {
1274 *want_primary
= pg_shard_t(want
[i
], shard_id_t(i
));
1275 found_primary
= true;
1279 acting_backfill
->insert(backfill
->begin(), backfill
->end());
1284 * calculate the desired acting set.
1286 * Choose an appropriate acting set. Prefer up[0], unless it is
1287 * incomplete, or another osd has a longer tail that allows us to
1288 * bring other up nodes up to date.
1290 void PG::calc_replicated_acting(
1291 map
<pg_shard_t
, pg_info_t
>::const_iterator auth_log_shard
,
1293 const vector
<int> &acting
,
1294 pg_shard_t acting_primary
,
1295 const vector
<int> &up
,
1296 pg_shard_t up_primary
,
1297 const map
<pg_shard_t
, pg_info_t
> &all_info
,
1298 bool restrict_to_up_acting
,
1300 set
<pg_shard_t
> *backfill
,
1301 set
<pg_shard_t
> *acting_backfill
,
1302 pg_shard_t
*want_primary
,
1305 ss
<< "calc_acting newest update on osd." << auth_log_shard
->first
1306 << " with " << auth_log_shard
->second
1307 << (restrict_to_up_acting
? " restrict_to_up_acting" : "") << std::endl
;
1308 pg_shard_t auth_log_shard_id
= auth_log_shard
->first
;
1311 map
<pg_shard_t
,pg_info_t
>::const_iterator primary
;
1313 !all_info
.find(up_primary
)->second
.is_incomplete() &&
1314 all_info
.find(up_primary
)->second
.last_update
>=
1315 auth_log_shard
->second
.log_tail
) {
1316 ss
<< "up_primary: " << up_primary
<< ") selected as primary" << std::endl
;
1317 primary
= all_info
.find(up_primary
); // prefer up[0], all thing being equal
1319 assert(!auth_log_shard
->second
.is_incomplete());
1320 ss
<< "up[0] needs backfill, osd." << auth_log_shard_id
1321 << " selected as primary instead" << std::endl
;
1322 primary
= auth_log_shard
;
1325 ss
<< "calc_acting primary is osd." << primary
->first
1326 << " with " << primary
->second
<< std::endl
;
1327 *want_primary
= primary
->first
;
1328 want
->push_back(primary
->first
.osd
);
1329 acting_backfill
->insert(primary
->first
);
1330 unsigned usable
= 1;
1332 // select replicas that have log contiguity with primary.
1333 // prefer up, then acting, then any peer_info osds
1334 for (vector
<int>::const_iterator i
= up
.begin();
1337 pg_shard_t up_cand
= pg_shard_t(*i
, shard_id_t::NO_SHARD
);
1338 if (up_cand
== primary
->first
)
1340 const pg_info_t
&cur_info
= all_info
.find(up_cand
)->second
;
1341 if (cur_info
.is_incomplete() ||
1342 cur_info
.last_update
< MIN(
1343 primary
->second
.log_tail
,
1344 auth_log_shard
->second
.log_tail
)) {
1345 /* We include auth_log_shard->second.log_tail because in GetLog,
1346 * we will request logs back to the min last_update over our
1347 * acting_backfill set, which will result in our log being extended
1348 * as far backwards as necessary to pick up any peers which can
1349 * be log recovered by auth_log_shard's log */
1350 ss
<< " shard " << up_cand
<< " (up) backfill " << cur_info
<< std::endl
;
1351 backfill
->insert(up_cand
);
1352 acting_backfill
->insert(up_cand
);
1354 want
->push_back(*i
);
1355 acting_backfill
->insert(up_cand
);
1357 ss
<< " osd." << *i
<< " (up) accepted " << cur_info
<< std::endl
;
1359 if (want
->size() >= size
) {
1364 // This no longer has backfill OSDs, but they are covered above.
1365 for (vector
<int>::const_iterator i
= acting
.begin();
1368 pg_shard_t
acting_cand(*i
, shard_id_t::NO_SHARD
);
1372 // skip up osds we already considered above
1373 if (acting_cand
== primary
->first
)
1375 vector
<int>::const_iterator up_it
= find(up
.begin(), up
.end(), acting_cand
.osd
);
1376 if (up_it
!= up
.end())
1379 const pg_info_t
&cur_info
= all_info
.find(acting_cand
)->second
;
1380 if (cur_info
.is_incomplete() ||
1381 cur_info
.last_update
< primary
->second
.log_tail
) {
1382 ss
<< " shard " << acting_cand
<< " (stray) REJECTED "
1383 << cur_info
<< std::endl
;
1385 want
->push_back(*i
);
1386 acting_backfill
->insert(acting_cand
);
1387 ss
<< " shard " << acting_cand
<< " (stray) accepted "
1388 << cur_info
<< std::endl
;
1393 if (restrict_to_up_acting
) {
1396 for (map
<pg_shard_t
,pg_info_t
>::const_iterator i
= all_info
.begin();
1397 i
!= all_info
.end();
1402 // skip up osds we already considered above
1403 if (i
->first
== primary
->first
)
1405 vector
<int>::const_iterator up_it
= find(up
.begin(), up
.end(), i
->first
.osd
);
1406 if (up_it
!= up
.end())
1408 vector
<int>::const_iterator acting_it
= find(
1409 acting
.begin(), acting
.end(), i
->first
.osd
);
1410 if (acting_it
!= acting
.end())
1413 if (i
->second
.is_incomplete() ||
1414 i
->second
.last_update
< primary
->second
.log_tail
) {
1415 ss
<< " shard " << i
->first
<< " (stray) REJECTED "
1416 << i
->second
<< std::endl
;
1418 want
->push_back(i
->first
.osd
);
1419 acting_backfill
->insert(i
->first
);
1420 ss
<< " shard " << i
->first
<< " (stray) accepted "
1421 << i
->second
<< std::endl
;
1430 * calculate the desired acting, and request a change with the monitor
1431 * if it differs from the current acting.
1433 * if restrict_to_up_acting=true, we filter out anything that's not in
1434 * up/acting. in order to lift this restriction, we need to
1435 * 1) check whether it's worth switching the acting set any time we get
1436 * a new pg info (not just here, when recovery finishes)
1437 * 2) check whether anything in want_acting went down on each new map
1438 * (and, if so, calculate a new want_acting)
1439 * 3) remove the assertion in PG::RecoveryState::Active::react(const AdvMap)
1442 bool PG::choose_acting(pg_shard_t
&auth_log_shard_id
,
1443 bool restrict_to_up_acting
,
1444 bool *history_les_bound
)
1446 map
<pg_shard_t
, pg_info_t
> all_info(peer_info
.begin(), peer_info
.end());
1447 all_info
[pg_whoami
] = info
;
1449 for (map
<pg_shard_t
, pg_info_t
>::iterator p
= all_info
.begin();
1450 p
!= all_info
.end();
1452 dout(10) << __func__
<< " all_info osd." << p
->first
<< " " << p
->second
<< dendl
;
1455 map
<pg_shard_t
, pg_info_t
>::const_iterator auth_log_shard
=
1456 find_best_info(all_info
, restrict_to_up_acting
, history_les_bound
);
1458 if (auth_log_shard
== all_info
.end()) {
1460 dout(10) << "choose_acting no suitable info found (incomplete backfills?),"
1461 << " reverting to up" << dendl
;
1464 osd
->queue_want_pg_temp(info
.pgid
.pgid
, empty
);
1466 dout(10) << "choose_acting failed" << dendl
;
1467 assert(want_acting
.empty());
1472 assert(!auth_log_shard
->second
.is_incomplete());
1473 auth_log_shard_id
= auth_log_shard
->first
;
1475 set
<pg_shard_t
> want_backfill
, want_acting_backfill
;
1477 pg_shard_t want_primary
;
1479 if (!pool
.info
.ec_pool())
1480 calc_replicated_acting(
1482 get_osdmap()->get_pg_size(info
.pgid
.pgid
),
1488 restrict_to_up_acting
,
1491 &want_acting_backfill
,
1497 get_osdmap()->get_pg_size(info
.pgid
.pgid
),
1503 restrict_to_up_acting
,
1506 &want_acting_backfill
,
1509 dout(10) << ss
.str() << dendl
;
1511 unsigned num_want_acting
= 0;
1512 set
<pg_shard_t
> have
;
1513 for (int i
= 0; i
< (int)want
.size(); ++i
) {
1514 if (want
[i
] != CRUSH_ITEM_NONE
) {
1519 pool
.info
.ec_pool() ? shard_id_t(i
) : shard_id_t::NO_SHARD
));
1523 // We go incomplete if below min_size for ec_pools since backfill
1524 // does not currently maintain rollbackability
1525 // Otherwise, we will go "peered", but not "active"
1526 if (num_want_acting
< pool
.info
.min_size
&&
1527 (pool
.info
.ec_pool() ||
1528 !cct
->_conf
->osd_allow_recovery_below_min_size
)) {
1529 want_acting
.clear();
1530 dout(10) << "choose_acting failed, below min size" << dendl
;
1534 /* Check whether we have enough acting shards to later perform recovery */
1535 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable_predicate(
1536 get_pgbackend()->get_is_recoverable_predicate());
1537 if (!(*recoverable_predicate
)(have
)) {
1538 want_acting
.clear();
1539 dout(10) << "choose_acting failed, not recoverable" << dendl
;
1543 if (want
!= acting
) {
1544 dout(10) << "choose_acting want " << want
<< " != acting " << acting
1545 << ", requesting pg_temp change" << dendl
;
1548 if (want_acting
== up
) {
1549 // There can't be any pending backfill if
1550 // want is the same as crush map up OSDs.
1551 assert(want_backfill
.empty());
1553 osd
->queue_want_pg_temp(info
.pgid
.pgid
, empty
);
1555 osd
->queue_want_pg_temp(info
.pgid
.pgid
, want
);
1558 want_acting
.clear();
1559 actingbackfill
= want_acting_backfill
;
1560 dout(10) << "actingbackfill is " << actingbackfill
<< dendl
;
1561 assert(backfill_targets
.empty() || backfill_targets
== want_backfill
);
1562 if (backfill_targets
.empty()) {
1563 // Caller is GetInfo
1564 backfill_targets
= want_backfill
;
1566 // Will not change if already set because up would have had to change
1567 // Verify that nothing in backfill is in stray_set
1568 for (set
<pg_shard_t
>::iterator i
= want_backfill
.begin();
1569 i
!= want_backfill
.end();
1571 assert(stray_set
.find(*i
) == stray_set
.end());
1573 dout(10) << "choose_acting want " << want
<< " (== acting) backfill_targets "
1574 << want_backfill
<< dendl
;
1578 /* Build the might_have_unfound set.
1580 * This is used by the primary OSD during recovery.
1582 * This set tracks the OSDs which might have unfound objects that the primary
1583 * OSD needs. As we receive pg_missing_t from each OSD in might_have_unfound, we
1584 * will remove the OSD from the set.
1586 void PG::build_might_have_unfound()
1588 assert(might_have_unfound
.empty());
1589 assert(is_primary());
1591 dout(10) << __func__
<< dendl
;
1593 check_past_interval_bounds();
1595 might_have_unfound
= past_intervals
.get_might_have_unfound(
1597 pool
.info
.ec_pool());
1599 // include any (stray) peers
1600 for (map
<pg_shard_t
, pg_info_t
>::iterator p
= peer_info
.begin();
1601 p
!= peer_info
.end();
1603 might_have_unfound
.insert(p
->first
);
1605 dout(15) << __func__
<< ": built " << might_have_unfound
<< dendl
;
1608 struct C_PG_ActivateCommitted
: public Context
{
1611 epoch_t activation_epoch
;
1612 C_PG_ActivateCommitted(PG
*p
, epoch_t e
, epoch_t ae
)
1613 : pg(p
), epoch(e
), activation_epoch(ae
) {}
1614 void finish(int r
) override
{
1615 pg
->_activate_committed(epoch
, activation_epoch
);
1619 void PG::activate(ObjectStore::Transaction
& t
,
1620 epoch_t activation_epoch
,
1621 list
<Context
*>& tfin
,
1622 map
<int, map
<spg_t
,pg_query_t
> >& query_map
,
1626 PastIntervals
> > > *activator_map
,
1629 assert(!is_peered());
1630 assert(scrubber
.callbacks
.empty());
1631 assert(callbacks_for_degraded_object
.empty());
1634 state_clear(PG_STATE_DOWN
);
1636 send_notify
= false;
1639 // only update primary last_epoch_started if we will go active
1640 if (acting
.size() >= pool
.info
.min_size
) {
1641 assert(cct
->_conf
->osd_find_best_info_ignore_history_les
||
1642 info
.last_epoch_started
<= activation_epoch
);
1643 info
.last_epoch_started
= activation_epoch
;
1644 info
.last_interval_started
= info
.history
.same_interval_since
;
1646 } else if (is_acting(pg_whoami
)) {
1647 /* update last_epoch_started on acting replica to whatever the primary sent
1648 * unless it's smaller (could happen if we are going peered rather than
1649 * active, see doc/dev/osd_internals/last_epoch_started.rst) */
1650 if (info
.last_epoch_started
< activation_epoch
) {
1651 info
.last_epoch_started
= activation_epoch
;
1652 info
.last_interval_started
= info
.history
.same_interval_since
;
1656 auto &missing
= pg_log
.get_missing();
1659 last_update_ondisk
= info
.last_update
;
1660 min_last_complete_ondisk
= eversion_t(0,0); // we don't know (yet)!
1662 last_update_applied
= info
.last_update
;
1663 last_rollback_info_trimmed_to_applied
= pg_log
.get_can_rollback_to();
1665 need_up_thru
= false;
1667 // write pg info, log
1669 dirty_big_info
= true; // maybe
1671 // find out when we commit
1672 t
.register_on_complete(
1673 new C_PG_ActivateCommitted(
1675 get_osdmap()->get_epoch(),
1678 // initialize snap_trimq
1680 dout(20) << "activate - purged_snaps " << info
.purged_snaps
1681 << " cached_removed_snaps " << pool
.cached_removed_snaps
<< dendl
;
1682 snap_trimq
= pool
.cached_removed_snaps
;
1683 interval_set
<snapid_t
> intersection
;
1684 intersection
.intersection_of(snap_trimq
, info
.purged_snaps
);
1685 if (intersection
== info
.purged_snaps
) {
1686 snap_trimq
.subtract(info
.purged_snaps
);
1688 dout(0) << "warning: info.purged_snaps (" << info
.purged_snaps
1689 << ") is not a subset of pool.cached_removed_snaps ("
1690 << pool
.cached_removed_snaps
<< ")" << dendl
;
1691 snap_trimq
.subtract(intersection
);
1695 // init complete pointer
1696 if (missing
.num_missing() == 0) {
1697 dout(10) << "activate - no missing, moving last_complete " << info
.last_complete
1698 << " -> " << info
.last_update
<< dendl
;
1699 info
.last_complete
= info
.last_update
;
1700 pg_log
.reset_recovery_pointers();
1702 dout(10) << "activate - not complete, " << missing
<< dendl
;
1703 pg_log
.activate_not_complete(info
);
1711 // start up replicas
1713 assert(!actingbackfill
.empty());
1714 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
1715 i
!= actingbackfill
.end();
1717 if (*i
== pg_whoami
) continue;
1718 pg_shard_t peer
= *i
;
1719 assert(peer_info
.count(peer
));
1720 pg_info_t
& pi
= peer_info
[peer
];
1722 dout(10) << "activate peer osd." << peer
<< " " << pi
<< dendl
;
1725 assert(peer_missing
.count(peer
));
1726 pg_missing_t
& pm
= peer_missing
[peer
];
1728 bool needs_past_intervals
= pi
.dne();
1731 * cover case where peer sort order was different and
1732 * last_backfill cannot be interpreted
1734 bool force_restart_backfill
=
1735 !pi
.last_backfill
.is_max() &&
1736 !pi
.last_backfill_bitwise
;
1738 if (pi
.last_update
== info
.last_update
&& !force_restart_backfill
) {
1740 if (!pi
.last_backfill
.is_max())
1741 osd
->clog
->info() << info
.pgid
<< " continuing backfill to osd."
1743 << " from (" << pi
.log_tail
<< "," << pi
.last_update
1744 << "] " << pi
.last_backfill
1745 << " to " << info
.last_update
;
1746 if (!pi
.is_empty() && activator_map
) {
1747 dout(10) << "activate peer osd." << peer
<< " is up to date, queueing in pending_activators" << dendl
;
1748 (*activator_map
)[peer
.osd
].push_back(
1751 peer
.shard
, pg_whoami
.shard
,
1752 get_osdmap()->get_epoch(),
1753 get_osdmap()->get_epoch(),
1757 dout(10) << "activate peer osd." << peer
<< " is up to date, but sending pg_log anyway" << dendl
;
1759 i
->shard
, pg_whoami
.shard
,
1760 get_osdmap()->get_epoch(), info
);
1763 pg_log
.get_tail() > pi
.last_update
||
1764 pi
.last_backfill
== hobject_t() ||
1765 force_restart_backfill
||
1766 (backfill_targets
.count(*i
) && pi
.last_backfill
.is_max())) {
1767 /* ^ This last case covers a situation where a replica is not contiguous
1768 * with the auth_log, but is contiguous with this replica. Reshuffling
1769 * the active set to handle this would be tricky, so instead we just go
1770 * ahead and backfill it anyway. This is probably preferrable in any
1771 * case since the replica in question would have to be significantly
1775 osd
->clog
->debug() << info
.pgid
<< " starting backfill to osd." << peer
1776 << " from (" << pi
.log_tail
<< "," << pi
.last_update
1777 << "] " << pi
.last_backfill
1778 << " to " << info
.last_update
;
1780 pi
.last_update
= info
.last_update
;
1781 pi
.last_complete
= info
.last_update
;
1782 pi
.set_last_backfill(hobject_t());
1783 pi
.last_epoch_started
= info
.last_epoch_started
;
1784 pi
.last_interval_started
= info
.last_interval_started
;
1785 pi
.history
= info
.history
;
1786 pi
.hit_set
= info
.hit_set
;
1787 pi
.stats
.stats
.clear();
1789 // initialize peer with our purged_snaps.
1790 pi
.purged_snaps
= info
.purged_snaps
;
1793 i
->shard
, pg_whoami
.shard
,
1794 get_osdmap()->get_epoch(), pi
);
1796 // send some recent log, so that op dup detection works well.
1797 m
->log
.copy_up_to(pg_log
.get_log(), cct
->_conf
->osd_min_pg_log_entries
);
1798 m
->info
.log_tail
= m
->log
.tail
;
1799 pi
.log_tail
= m
->log
.tail
; // sigh...
1804 assert(pg_log
.get_tail() <= pi
.last_update
);
1806 i
->shard
, pg_whoami
.shard
,
1807 get_osdmap()->get_epoch(), info
);
1808 // send new stuff to append to replicas log
1809 m
->log
.copy_after(pg_log
.get_log(), pi
.last_update
);
1812 // share past_intervals if we are creating the pg on the replica
1813 // based on whether our info for that peer was dne() *before*
1814 // updating pi.history in the backfill block above.
1815 if (m
&& needs_past_intervals
)
1816 m
->past_intervals
= past_intervals
;
1818 // update local version of peer's missing list!
1819 if (m
&& pi
.last_backfill
!= hobject_t()) {
1820 for (list
<pg_log_entry_t
>::iterator p
= m
->log
.log
.begin();
1821 p
!= m
->log
.log
.end();
1823 if (p
->soid
<= pi
.last_backfill
&&
1825 if (perform_deletes_during_peering() && p
->is_delete()) {
1826 pm
.rm(p
->soid
, p
->version
);
1828 pm
.add_next_event(*p
);
1835 dout(10) << "activate peer osd." << peer
<< " sending " << m
->log
<< dendl
;
1836 //m->log.print(cout);
1837 osd
->send_message_osd_cluster(peer
.osd
, m
, get_osdmap()->get_epoch());
1841 pi
.last_update
= info
.last_update
;
1843 // update our missing
1844 if (pm
.num_missing() == 0) {
1845 pi
.last_complete
= pi
.last_update
;
1846 dout(10) << "activate peer osd." << peer
<< " " << pi
<< " uptodate" << dendl
;
1848 dout(10) << "activate peer osd." << peer
<< " " << pi
<< " missing " << pm
<< dendl
;
1852 // Set up missing_loc
1853 set
<pg_shard_t
> complete_shards
;
1854 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
1855 i
!= actingbackfill
.end();
1857 dout(20) << __func__
<< " setting up missing_loc from shard " << *i
<< " " << dendl
;
1858 if (*i
== get_primary()) {
1859 missing_loc
.add_active_missing(missing
);
1860 if (!missing
.have_missing())
1861 complete_shards
.insert(*i
);
1863 auto peer_missing_entry
= peer_missing
.find(*i
);
1864 assert(peer_missing_entry
!= peer_missing
.end());
1865 missing_loc
.add_active_missing(peer_missing_entry
->second
);
1866 if (!peer_missing_entry
->second
.have_missing() &&
1867 peer_info
[*i
].last_backfill
.is_max())
1868 complete_shards
.insert(*i
);
1872 // If necessary, create might_have_unfound to help us find our unfound objects.
1873 // NOTE: It's important that we build might_have_unfound before trimming the
1875 might_have_unfound
.clear();
1876 if (needs_recovery()) {
1877 // If only one shard has missing, we do a trick to add all others as recovery
1878 // source, this is considered safe since the PGLogs have been merged locally,
1879 // and covers vast majority of the use cases, like one OSD/host is down for
1880 // a while for hardware repairing
1881 if (complete_shards
.size() + 1 == actingbackfill
.size()) {
1882 missing_loc
.add_batch_sources_info(complete_shards
, ctx
->handle
);
1884 missing_loc
.add_source_info(pg_whoami
, info
, pg_log
.get_missing(),
1886 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
1887 i
!= actingbackfill
.end();
1889 if (*i
== pg_whoami
) continue;
1890 dout(10) << __func__
<< ": adding " << *i
<< " as a source" << dendl
;
1891 assert(peer_missing
.count(*i
));
1892 assert(peer_info
.count(*i
));
1893 missing_loc
.add_source_info(
1900 for (map
<pg_shard_t
, pg_missing_t
>::iterator i
= peer_missing
.begin();
1901 i
!= peer_missing
.end();
1903 if (is_actingbackfill(i
->first
))
1905 assert(peer_info
.count(i
->first
));
1907 peer_info
[i
->first
],
1913 build_might_have_unfound();
1915 // Always call now so _update_calc_stats() will be accurate
1916 discover_all_missing(query_map
);
1919 // num_objects_degraded if calculated should reflect this too, unless no
1920 // missing and we are about to go clean.
1921 if (get_osdmap()->get_pg_size(info
.pgid
.pgid
) > actingset
.size()) {
1922 state_set(PG_STATE_UNDERSIZED
);
1925 state_set(PG_STATE_ACTIVATING
);
1926 release_pg_backoffs();
1927 projected_last_update
= info
.last_update
;
1929 if (acting
.size() >= pool
.info
.min_size
) {
1930 PGLogEntryHandler handler
{this, &t
};
1931 pg_log
.roll_forward(&handler
);
1935 bool PG::op_has_sufficient_caps(OpRequestRef
& op
)
1937 // only check MOSDOp
1938 if (op
->get_req()->get_type() != CEPH_MSG_OSD_OP
)
1941 const MOSDOp
*req
= static_cast<const MOSDOp
*>(op
->get_req());
1943 Session
*session
= static_cast<Session
*>(req
->get_connection()->get_priv());
1945 dout(0) << "op_has_sufficient_caps: no session for op " << *req
<< dendl
;
1948 OSDCap
& caps
= session
->caps
;
1951 const string
&key
= req
->get_hobj().get_key().empty() ?
1952 req
->get_oid().name
:
1953 req
->get_hobj().get_key();
1955 bool cap
= caps
.is_capable(pool
.name
, req
->get_hobj().nspace
,
1957 op
->need_read_cap(),
1958 op
->need_write_cap(),
1961 dout(20) << "op_has_sufficient_caps "
1962 << "session=" << session
1963 << " pool=" << pool
.id
<< " (" << pool
.name
1964 << " " << req
->get_hobj().nspace
1965 << ") owner=" << pool
.auid
1966 << " need_read_cap=" << op
->need_read_cap()
1967 << " need_write_cap=" << op
->need_write_cap()
1968 << " classes=" << op
->classes()
1969 << " -> " << (cap
? "yes" : "NO")
1974 void PG::_activate_committed(epoch_t epoch
, epoch_t activation_epoch
)
1977 if (pg_has_reset_since(epoch
)) {
1978 dout(10) << "_activate_committed " << epoch
1979 << ", that was an old interval" << dendl
;
1980 } else if (is_primary()) {
1981 peer_activated
.insert(pg_whoami
);
1982 dout(10) << "_activate_committed " << epoch
1983 << " peer_activated now " << peer_activated
1984 << " last_interval_started " << info
.history
.last_interval_started
1985 << " last_epoch_started " << info
.history
.last_epoch_started
1986 << " same_interval_since " << info
.history
.same_interval_since
<< dendl
;
1987 assert(!actingbackfill
.empty());
1988 if (peer_activated
.size() == actingbackfill
.size())
1989 all_activated_and_committed();
1991 dout(10) << "_activate_committed " << epoch
<< " telling primary" << dendl
;
1992 MOSDPGInfo
*m
= new MOSDPGInfo(epoch
);
1993 pg_notify_t i
= pg_notify_t(
1994 get_primary().shard
, pg_whoami
.shard
,
1995 get_osdmap()->get_epoch(),
1996 get_osdmap()->get_epoch(),
1999 i
.info
.history
.last_epoch_started
= activation_epoch
;
2000 i
.info
.history
.last_interval_started
= i
.info
.history
.same_interval_since
;
2001 if (acting
.size() >= pool
.info
.min_size
) {
2002 state_set(PG_STATE_ACTIVE
);
2004 state_set(PG_STATE_PEERED
);
2007 m
->pg_list
.push_back(make_pair(i
, PastIntervals()));
2008 osd
->send_message_osd_cluster(get_primary().osd
, m
, get_osdmap()->get_epoch());
2011 if (flushes_in_progress
== 0) {
2012 requeue_ops(waiting_for_peered
);
2013 } else if (!waiting_for_peered
.empty()) {
2014 dout(10) << __func__
<< " flushes in progress, moving "
2015 << waiting_for_peered
.size() << " items to waiting_for_flush"
2017 assert(waiting_for_flush
.empty());
2018 waiting_for_flush
.swap(waiting_for_peered
);
2022 assert(!dirty_info
);
2028 * update info.history.last_epoch_started ONLY after we and all
2029 * replicas have activated AND committed the activate transaction
2030 * (i.e. the peering results are stable on disk).
2032 void PG::all_activated_and_committed()
2034 dout(10) << "all_activated_and_committed" << dendl
;
2035 assert(is_primary());
2036 assert(peer_activated
.size() == actingbackfill
.size());
2037 assert(!actingbackfill
.empty());
2038 assert(blocked_by
.empty());
2041 _update_calc_stats();
2042 if (info
.stats
.stats
.sum
.num_objects_degraded
) {
2043 state_set(PG_STATE_DEGRADED
);
2045 state_clear(PG_STATE_DEGRADED
);
2048 queue_peering_event(
2050 std::make_shared
<CephPeeringEvt
>(
2051 get_osdmap()->get_epoch(),
2052 get_osdmap()->get_epoch(),
2053 AllReplicasActivated())));
2056 bool PG::requeue_scrub(bool high_priority
)
2058 assert(is_locked());
2060 dout(10) << __func__
<< ": already queued" << dendl
;
2063 dout(10) << __func__
<< ": queueing" << dendl
;
2064 scrub_queued
= true;
2065 osd
->queue_for_scrub(this, high_priority
);
2070 void PG::queue_recovery()
2072 if (!is_primary() || !is_peered()) {
2073 dout(10) << "queue_recovery -- not primary or not peered " << dendl
;
2074 assert(!recovery_queued
);
2075 } else if (recovery_queued
) {
2076 dout(10) << "queue_recovery -- already queued" << dendl
;
2078 dout(10) << "queue_recovery -- queuing" << dendl
;
2079 recovery_queued
= true;
2080 osd
->queue_for_recovery(this);
2084 bool PG::queue_scrub()
2086 assert(is_locked());
2087 if (is_scrubbing()) {
2090 scrubber
.priority
= scrubber
.must_scrub
?
2091 cct
->_conf
->osd_requested_scrub_priority
: get_scrub_priority();
2092 scrubber
.must_scrub
= false;
2093 state_set(PG_STATE_SCRUBBING
);
2094 if (scrubber
.must_deep_scrub
) {
2095 state_set(PG_STATE_DEEP_SCRUB
);
2096 scrubber
.must_deep_scrub
= false;
2098 if (scrubber
.must_repair
|| scrubber
.auto_repair
) {
2099 state_set(PG_STATE_REPAIR
);
2100 scrubber
.must_repair
= false;
2106 unsigned PG::get_scrub_priority()
2108 // a higher value -> a higher priority
2109 int pool_scrub_priority
= 0;
2110 pool
.info
.opts
.get(pool_opts_t::SCRUB_PRIORITY
, &pool_scrub_priority
);
2111 return pool_scrub_priority
> 0 ? pool_scrub_priority
: cct
->_conf
->osd_scrub_priority
;
2114 struct C_PG_FinishRecovery
: public Context
{
2116 explicit C_PG_FinishRecovery(PG
*p
) : pg(p
) {}
2117 void finish(int r
) override
{
2118 pg
->_finish_recovery(this);
2122 void PG::mark_clean()
2124 if (actingset
.size() == get_osdmap()->get_pg_size(info
.pgid
.pgid
)) {
2125 state_clear(PG_STATE_FORCED_BACKFILL
| PG_STATE_FORCED_RECOVERY
);
2126 state_set(PG_STATE_CLEAN
);
2127 info
.history
.last_epoch_clean
= get_osdmap()->get_epoch();
2128 info
.history
.last_interval_clean
= info
.history
.same_interval_since
;
2129 past_intervals
.clear();
2130 dirty_big_info
= true;
2137 void PG::_change_recovery_force_mode(int new_mode
, bool clear
)
2140 // we can't and shouldn't do anything if the PG is being deleted locally
2142 state_clear(new_mode
);
2144 state_set(new_mode
);
2146 publish_stats_to_osd();
2150 inline int PG::clamp_recovery_priority(int priority
)
2152 static_assert(OSD_RECOVERY_PRIORITY_MIN
< OSD_RECOVERY_PRIORITY_MAX
, "Invalid priority range");
2153 static_assert(OSD_RECOVERY_PRIORITY_MIN
>= 0, "Priority range must match unsigned type");
2155 // Clamp to valid range
2156 if (priority
> OSD_RECOVERY_PRIORITY_MAX
) {
2157 return OSD_RECOVERY_PRIORITY_MAX
;
2158 } else if (priority
< OSD_RECOVERY_PRIORITY_MIN
) {
2159 return OSD_RECOVERY_PRIORITY_MIN
;
2165 unsigned PG::get_recovery_priority()
2167 // a higher value -> a higher priority
2170 if (state
& PG_STATE_FORCED_RECOVERY
) {
2171 ret
= OSD_RECOVERY_PRIORITY_FORCED
;
2173 pool
.info
.opts
.get(pool_opts_t::RECOVERY_PRIORITY
, &ret
);
2174 ret
= clamp_recovery_priority(OSD_RECOVERY_PRIORITY_BASE
+ ret
);
2176 dout(20) << __func__
<< " recovery priority for " << *this << " is " << ret
<< ", state is " << state
<< dendl
;
2177 return static_cast<unsigned>(ret
);
2180 unsigned PG::get_backfill_priority()
2182 // a higher value -> a higher priority
2183 int ret
= OSD_BACKFILL_PRIORITY_BASE
;
2184 if (state
& PG_STATE_FORCED_BACKFILL
) {
2185 ret
= OSD_RECOVERY_PRIORITY_FORCED
;
2187 if (acting
.size() < pool
.info
.min_size
) {
2188 // inactive: no. of replicas < min_size, highest priority since it blocks IO
2189 ret
= OSD_BACKFILL_INACTIVE_PRIORITY_BASE
+ (pool
.info
.min_size
- acting
.size());
2191 } else if (is_undersized()) {
2192 // undersized: OSD_BACKFILL_DEGRADED_PRIORITY_BASE + num missing replicas
2193 assert(pool
.info
.size
> actingset
.size());
2194 ret
= OSD_BACKFILL_DEGRADED_PRIORITY_BASE
+ (pool
.info
.size
- actingset
.size());
2196 } else if (is_degraded()) {
2197 // degraded: baseline degraded
2198 ret
= OSD_BACKFILL_DEGRADED_PRIORITY_BASE
;
2201 // Adjust with pool's recovery priority
2202 int pool_recovery_priority
= 0;
2203 pool
.info
.opts
.get(pool_opts_t::RECOVERY_PRIORITY
, &pool_recovery_priority
);
2205 ret
= clamp_recovery_priority(pool_recovery_priority
+ ret
);
2208 return static_cast<unsigned>(ret
);
2211 void PG::finish_recovery(list
<Context
*>& tfin
)
2213 dout(10) << "finish_recovery" << dendl
;
2214 assert(info
.last_complete
== info
.last_update
);
2216 clear_recovery_state();
2219 * sync all this before purging strays. but don't block!
2221 finish_sync_event
= new C_PG_FinishRecovery(this);
2222 tfin
.push_back(finish_sync_event
);
2225 void PG::_finish_recovery(Context
*c
)
2232 if (c
== finish_sync_event
) {
2233 dout(10) << "_finish_recovery" << dendl
;
2234 finish_sync_event
= 0;
2237 publish_stats_to_osd();
2239 if (scrub_after_recovery
) {
2240 dout(10) << "_finish_recovery requeueing for scrub" << dendl
;
2241 scrub_after_recovery
= false;
2242 scrubber
.must_deep_scrub
= true;
2246 dout(10) << "_finish_recovery -- stale" << dendl
;
2251 void PG::start_recovery_op(const hobject_t
& soid
)
2253 dout(10) << "start_recovery_op " << soid
2254 #ifdef DEBUG_RECOVERY_OIDS
2255 << " (" << recovering_oids
<< ")"
2258 assert(recovery_ops_active
>= 0);
2259 recovery_ops_active
++;
2260 #ifdef DEBUG_RECOVERY_OIDS
2261 assert(recovering_oids
.count(soid
) == 0);
2262 recovering_oids
.insert(soid
);
2264 osd
->start_recovery_op(this, soid
);
2267 void PG::finish_recovery_op(const hobject_t
& soid
, bool dequeue
)
2269 dout(10) << "finish_recovery_op " << soid
2270 #ifdef DEBUG_RECOVERY_OIDS
2271 << " (" << recovering_oids
<< ")"
2274 assert(recovery_ops_active
> 0);
2275 recovery_ops_active
--;
2276 #ifdef DEBUG_RECOVERY_OIDS
2277 assert(recovering_oids
.count(soid
));
2278 recovering_oids
.erase(soid
);
2280 osd
->finish_recovery_op(this, soid
, dequeue
);
2287 void PG::split_into(pg_t child_pgid
, PG
*child
, unsigned split_bits
)
2289 child
->update_snap_mapper_bits(split_bits
);
2290 child
->update_osdmap_ref(get_osdmap());
2295 pg_log
.split_into(child_pgid
, split_bits
, &(child
->pg_log
));
2296 child
->info
.last_complete
= info
.last_complete
;
2298 info
.last_update
= pg_log
.get_head();
2299 child
->info
.last_update
= child
->pg_log
.get_head();
2301 child
->info
.last_user_version
= info
.last_user_version
;
2303 info
.log_tail
= pg_log
.get_tail();
2304 child
->info
.log_tail
= child
->pg_log
.get_tail();
2306 if (info
.last_complete
< pg_log
.get_tail())
2307 info
.last_complete
= pg_log
.get_tail();
2308 if (child
->info
.last_complete
< child
->pg_log
.get_tail())
2309 child
->info
.last_complete
= child
->pg_log
.get_tail();
2312 child
->info
.history
= info
.history
;
2313 child
->info
.history
.epoch_created
= get_osdmap()->get_epoch();
2314 child
->info
.purged_snaps
= info
.purged_snaps
;
2316 if (info
.last_backfill
.is_max()) {
2317 child
->info
.set_last_backfill(hobject_t::get_max());
2319 // restart backfill on parent and child to be safe. we could
2320 // probably do better in the bitwise sort case, but it's more
2321 // fragile (there may be special work to do on backfill completion
2323 info
.set_last_backfill(hobject_t());
2324 child
->info
.set_last_backfill(hobject_t());
2325 // restarting backfill implies that the missing set is empty,
2326 // since it is only used for objects prior to last_backfill
2327 pg_log
.reset_backfill();
2328 child
->pg_log
.reset_backfill();
2331 child
->info
.stats
= info
.stats
;
2332 child
->info
.stats
.parent_split_bits
= split_bits
;
2333 info
.stats
.stats_invalid
= true;
2334 child
->info
.stats
.stats_invalid
= true;
2335 child
->info
.last_epoch_started
= info
.last_epoch_started
;
2336 child
->info
.last_interval_started
= info
.last_interval_started
;
2338 child
->snap_trimq
= snap_trimq
;
2340 // There can't be recovery/backfill going on now
2341 int primary
, up_primary
;
2342 vector
<int> newup
, newacting
;
2343 get_osdmap()->pg_to_up_acting_osds(
2344 child
->info
.pgid
.pgid
, &newup
, &up_primary
, &newacting
, &primary
);
2345 child
->init_primary_up_acting(
2350 child
->role
= OSDMap::calc_pg_role(osd
->whoami
, child
->acting
);
2352 // this comparison includes primary rank via pg_shard_t
2353 if (get_primary() != child
->get_primary())
2354 child
->info
.history
.same_primary_since
= get_osdmap()->get_epoch();
2356 child
->info
.stats
.up
= up
;
2357 child
->info
.stats
.up_primary
= up_primary
;
2358 child
->info
.stats
.acting
= acting
;
2359 child
->info
.stats
.acting_primary
= primary
;
2360 child
->info
.stats
.mapping_epoch
= get_osdmap()->get_epoch();
2363 child
->past_intervals
= past_intervals
;
2365 _split_into(child_pgid
, child
, split_bits
);
2367 // release all backoffs for simplicity
2368 release_backoffs(hobject_t(), hobject_t::get_max());
2370 child
->on_new_interval();
2372 child
->dirty_info
= true;
2373 child
->dirty_big_info
= true;
2375 dirty_big_info
= true;
2378 void PG::add_backoff(SessionRef s
, const hobject_t
& begin
, const hobject_t
& end
)
2380 ConnectionRef con
= s
->con
;
2381 if (!con
) // OSD::ms_handle_reset clears s->con without a lock
2383 BackoffRef
b(s
->have_backoff(info
.pgid
, begin
));
2385 derr
<< __func__
<< " already have backoff for " << s
<< " begin " << begin
2386 << " " << *b
<< dendl
;
2389 Mutex::Locker
l(backoff_lock
);
2391 b
= new Backoff(info
.pgid
, this, s
, ++s
->backoff_seq
, begin
, end
);
2392 backoffs
[begin
].insert(b
);
2394 dout(10) << __func__
<< " session " << s
<< " added " << *b
<< dendl
;
2399 get_osdmap()->get_epoch(),
2400 CEPH_OSD_BACKOFF_OP_BLOCK
,
2406 void PG::release_backoffs(const hobject_t
& begin
, const hobject_t
& end
)
2408 dout(10) << __func__
<< " [" << begin
<< "," << end
<< ")" << dendl
;
2409 vector
<BackoffRef
> bv
;
2411 Mutex::Locker
l(backoff_lock
);
2412 auto p
= backoffs
.lower_bound(begin
);
2413 while (p
!= backoffs
.end()) {
2414 int r
= cmp(p
->first
, end
);
2415 dout(20) << __func__
<< " ? " << r
<< " " << p
->first
2416 << " " << p
->second
<< dendl
;
2417 // note: must still examine begin=end=p->first case
2418 if (r
> 0 || (r
== 0 && begin
< end
)) {
2421 dout(20) << __func__
<< " checking " << p
->first
2422 << " " << p
->second
<< dendl
;
2423 auto q
= p
->second
.begin();
2424 while (q
!= p
->second
.end()) {
2425 dout(20) << __func__
<< " checking " << *q
<< dendl
;
2426 int r
= cmp((*q
)->begin
, begin
);
2427 if (r
== 0 || (r
> 0 && (*q
)->end
< end
)) {
2429 q
= p
->second
.erase(q
);
2434 if (p
->second
.empty()) {
2435 p
= backoffs
.erase(p
);
2442 Mutex::Locker
l(b
->lock
);
2443 dout(10) << __func__
<< " " << *b
<< dendl
;
2445 assert(b
->pg
== this);
2446 ConnectionRef con
= b
->session
->con
;
2447 if (con
) { // OSD::ms_handle_reset clears s->con without a lock
2451 get_osdmap()->get_epoch(),
2452 CEPH_OSD_BACKOFF_OP_UNBLOCK
,
2458 b
->state
= Backoff::STATE_DELETING
;
2460 b
->session
->rm_backoff(b
);
2468 void PG::clear_backoffs()
2470 dout(10) << __func__
<< " " << dendl
;
2471 map
<hobject_t
,set
<BackoffRef
>> ls
;
2473 Mutex::Locker
l(backoff_lock
);
2476 for (auto& p
: ls
) {
2477 for (auto& b
: p
.second
) {
2478 Mutex::Locker
l(b
->lock
);
2479 dout(10) << __func__
<< " " << *b
<< dendl
;
2481 assert(b
->pg
== this);
2483 b
->state
= Backoff::STATE_DELETING
;
2485 b
->session
->rm_backoff(b
);
2494 // called by Session::clear_backoffs()
2495 void PG::rm_backoff(BackoffRef b
)
2497 dout(10) << __func__
<< " " << *b
<< dendl
;
2498 Mutex::Locker
l(backoff_lock
);
2499 assert(b
->lock
.is_locked_by_me());
2500 assert(b
->pg
== this);
2501 auto p
= backoffs
.find(b
->begin
);
2502 // may race with release_backoffs()
2503 if (p
!= backoffs
.end()) {
2504 auto q
= p
->second
.find(b
);
2505 if (q
!= p
->second
.end()) {
2507 if (p
->second
.empty()) {
2514 void PG::clear_recovery_state()
2516 dout(10) << "clear_recovery_state" << dendl
;
2518 pg_log
.reset_recovery_pointers();
2519 finish_sync_event
= 0;
2522 while (recovery_ops_active
> 0) {
2523 #ifdef DEBUG_RECOVERY_OIDS
2524 soid
= *recovering_oids
.begin();
2526 finish_recovery_op(soid
, true);
2529 backfill_targets
.clear();
2530 backfill_info
.clear();
2531 peer_backfill_info
.clear();
2532 waiting_on_backfill
.clear();
2533 _clear_recovery_state(); // pg impl specific hook
2536 void PG::cancel_recovery()
2538 dout(10) << "cancel_recovery" << dendl
;
2539 clear_recovery_state();
2543 void PG::purge_strays()
2545 dout(10) << "purge_strays " << stray_set
<< dendl
;
2547 bool removed
= false;
2548 for (set
<pg_shard_t
>::iterator p
= stray_set
.begin();
2549 p
!= stray_set
.end();
2551 assert(!is_actingbackfill(*p
));
2552 if (get_osdmap()->is_up(p
->osd
)) {
2553 dout(10) << "sending PGRemove to osd." << *p
<< dendl
;
2554 vector
<spg_t
> to_remove
;
2555 to_remove
.push_back(spg_t(info
.pgid
.pgid
, p
->shard
));
2556 MOSDPGRemove
*m
= new MOSDPGRemove(
2557 get_osdmap()->get_epoch(),
2559 osd
->send_message_osd_cluster(p
->osd
, m
, get_osdmap()->get_epoch());
2561 dout(10) << "not sending PGRemove to down osd." << *p
<< dendl
;
2563 peer_missing
.erase(*p
);
2564 peer_info
.erase(*p
);
2565 peer_purged
.insert(*p
);
2569 // if we removed anyone, update peers (which include peer_info)
2571 update_heartbeat_peers();
2575 // clear _requested maps; we may have to peer() again if we discover
2576 // (more) stray content
2577 peer_log_requested
.clear();
2578 peer_missing_requested
.clear();
2581 void PG::set_probe_targets(const set
<pg_shard_t
> &probe_set
)
2583 Mutex::Locker
l(heartbeat_peer_lock
);
2584 probe_targets
.clear();
2585 for (set
<pg_shard_t
>::iterator i
= probe_set
.begin();
2586 i
!= probe_set
.end();
2588 probe_targets
.insert(i
->osd
);
2592 void PG::clear_probe_targets()
2594 Mutex::Locker
l(heartbeat_peer_lock
);
2595 probe_targets
.clear();
2598 void PG::update_heartbeat_peers()
2600 assert(is_locked());
2606 for (unsigned i
=0; i
<acting
.size(); i
++) {
2607 if (acting
[i
] != CRUSH_ITEM_NONE
)
2608 new_peers
.insert(acting
[i
]);
2610 for (unsigned i
=0; i
<up
.size(); i
++) {
2611 if (up
[i
] != CRUSH_ITEM_NONE
)
2612 new_peers
.insert(up
[i
]);
2614 for (map
<pg_shard_t
,pg_info_t
>::iterator p
= peer_info
.begin();
2615 p
!= peer_info
.end();
2617 new_peers
.insert(p
->first
.osd
);
2619 bool need_update
= false;
2620 heartbeat_peer_lock
.Lock();
2621 if (new_peers
== heartbeat_peers
) {
2622 dout(10) << "update_heartbeat_peers " << heartbeat_peers
<< " unchanged" << dendl
;
2624 dout(10) << "update_heartbeat_peers " << heartbeat_peers
<< " -> " << new_peers
<< dendl
;
2625 heartbeat_peers
.swap(new_peers
);
2628 heartbeat_peer_lock
.Unlock();
2631 osd
->need_heartbeat_peer_update();
2635 bool PG::check_in_progress_op(
2636 const osd_reqid_t
&r
,
2637 eversion_t
*version
,
2638 version_t
*user_version
,
2639 int *return_code
) const
2642 projected_log
.get_request(r
, version
, user_version
, return_code
) ||
2643 pg_log
.get_log().get_request(r
, version
, user_version
, return_code
));
2646 static bool find_shard(const set
<pg_shard_t
> & pgs
, shard_id_t shard
)
2649 if (p
.shard
== shard
)
2654 static pg_shard_t
get_another_shard(const set
<pg_shard_t
> & pgs
, pg_shard_t skip
, shard_id_t shard
)
2656 for (auto&p
: pgs
) {
2659 if (p
.shard
== shard
)
2662 return pg_shard_t();
2665 void PG::_update_calc_stats()
2667 info
.stats
.version
= info
.last_update
;
2668 info
.stats
.created
= info
.history
.epoch_created
;
2669 info
.stats
.last_scrub
= info
.history
.last_scrub
;
2670 info
.stats
.last_scrub_stamp
= info
.history
.last_scrub_stamp
;
2671 info
.stats
.last_deep_scrub
= info
.history
.last_deep_scrub
;
2672 info
.stats
.last_deep_scrub_stamp
= info
.history
.last_deep_scrub_stamp
;
2673 info
.stats
.last_clean_scrub_stamp
= info
.history
.last_clean_scrub_stamp
;
2674 info
.stats
.last_epoch_clean
= info
.history
.last_epoch_clean
;
2676 info
.stats
.log_size
= pg_log
.get_head().version
- pg_log
.get_tail().version
;
2677 info
.stats
.ondisk_log_size
= info
.stats
.log_size
;
2678 info
.stats
.log_start
= pg_log
.get_tail();
2679 info
.stats
.ondisk_log_start
= pg_log
.get_tail();
2680 info
.stats
.snaptrimq_len
= snap_trimq
.size();
2682 unsigned num_shards
= get_osdmap()->get_pg_size(info
.pgid
.pgid
);
2684 // In rare case that upset is too large (usually transient), use as target
2685 // for calculations below.
2686 unsigned target
= std::max(num_shards
, (unsigned)upset
.size());
2687 // For undersized actingset may be larger with OSDs out
2688 unsigned nrep
= std::max(actingset
.size(), upset
.size());
2689 // calc num_object_copies
2690 info
.stats
.stats
.calc_copies(MAX(target
, nrep
));
2691 info
.stats
.stats
.sum
.num_objects_degraded
= 0;
2692 info
.stats
.stats
.sum
.num_objects_unfound
= 0;
2693 info
.stats
.stats
.sum
.num_objects_misplaced
= 0;
2695 if ((is_remapped() || is_undersized() || !is_clean()) && (is_peered() || is_activating())) {
2696 dout(20) << __func__
<< " actingset " << actingset
<< " upset "
2697 << upset
<< " actingbackfill " << actingbackfill
<< dendl
;
2698 dout(20) << __func__
<< " acting " << acting
<< " up " << up
<< dendl
;
2700 assert(!actingbackfill
.empty());
2702 bool estimate
= false;
2704 // NOTE: we only generate degraded, misplaced and unfound
2705 // values for the summation, not individual stat categories.
2706 int64_t num_objects
= info
.stats
.stats
.sum
.num_objects
;
2708 // Objects missing from up nodes, sorted by # objects.
2709 boost::container::flat_set
<pair
<int64_t,pg_shard_t
>> missing_target_objects
;
2710 // Objects missing from nodes not in up, sort by # objects
2711 boost::container::flat_set
<pair
<int64_t,pg_shard_t
>> acting_source_objects
;
2713 // Fill missing_target_objects/acting_source_objects
2719 missing
= pg_log
.get_missing().num_missing();
2720 assert(actingbackfill
.count(pg_whoami
));
2721 if (upset
.count(pg_whoami
)) {
2722 missing_target_objects
.insert(make_pair(missing
, pg_whoami
));
2724 acting_source_objects
.insert(make_pair(missing
, pg_whoami
));
2726 info
.stats
.stats
.sum
.num_objects_missing_on_primary
= missing
;
2727 dout(20) << __func__
<< " shard " << pg_whoami
2728 << " primary objects " << num_objects
2729 << " missing " << missing
2735 for (auto& peer
: peer_info
) {
2736 // Primary should not be in the peer_info, skip if it is.
2737 if (peer
.first
== pg_whoami
) continue;
2738 int64_t missing
= 0;
2739 int64_t peer_num_objects
= peer
.second
.stats
.stats
.sum
.num_objects
;
2740 // Backfill targets always track num_objects accurately
2741 // all other peers track missing accurately.
2742 if (is_backfill_targets(peer
.first
)) {
2743 missing
= std::max((int64_t)0, num_objects
- peer_num_objects
);
2745 if (peer_missing
.count(peer
.first
)) {
2746 missing
= peer_missing
[peer
.first
].num_missing();
2748 dout(20) << __func__
<< " no peer_missing found for " << peer
.first
<< dendl
;
2749 if (is_recovering()) {
2752 missing
= std::max((int64_t)0, num_objects
- peer_num_objects
);
2755 if (upset
.count(peer
.first
)) {
2756 missing_target_objects
.insert(make_pair(missing
, peer
.first
));
2757 } else if (actingset
.count(peer
.first
)) {
2758 acting_source_objects
.insert(make_pair(missing
, peer
.first
));
2760 peer
.second
.stats
.stats
.sum
.num_objects_missing
= missing
;
2761 dout(20) << __func__
<< " shard " << peer
.first
2762 << " objects " << peer_num_objects
2763 << " missing " << missing
2767 // A misplaced object is not stored on the correct OSD
2768 int64_t misplaced
= 0;
2769 // a degraded objects has fewer replicas or EC shards than the pool specifies.
2770 int64_t degraded
= 0;
2772 if (is_recovering()) {
2773 for (auto& sml
: missing_loc
.get_missing_by_count()) {
2774 for (auto& ml
: sml
.second
) {
2776 if (sml
.first
== shard_id_t::NO_SHARD
) {
2777 dout(0) << __func__
<< " ml " << ml
.second
<< " upset size " << upset
.size() << " up " << ml
.first
.up
<< dendl
;
2778 missing_shards
= (int)upset
.size() - ml
.first
.up
;
2780 // Handle shards not even in upset below
2781 if (!find_shard(upset
, sml
.first
))
2783 missing_shards
= std::max(0, 1 - ml
.first
.up
);
2784 dout(0) << __func__
<< " shard " << sml
.first
<< " ml " << ml
.second
<< " missing shards " << missing_shards
<< dendl
;
2786 int odegraded
= ml
.second
* missing_shards
;
2787 // Copies on other osds but limited to the possible degraded
2788 int more_osds
= std::min(missing_shards
, ml
.first
.other
);
2789 int omisplaced
= ml
.second
* more_osds
;
2790 assert(omisplaced
<= odegraded
);
2791 odegraded
-= omisplaced
;
2793 misplaced
+= omisplaced
;
2794 degraded
+= odegraded
;
2798 dout(20) << __func__
<< " missing based degraded " << degraded
<< dendl
;
2799 dout(20) << __func__
<< " missing based misplaced " << misplaced
<< dendl
;
2801 // Handle undersized case
2802 if (pool
.info
.is_replicated()) {
2803 // Add degraded for missing targets (num_objects missing)
2804 assert(target
>= upset
.size());
2805 unsigned needed
= target
- upset
.size();
2806 degraded
+= num_objects
* needed
;
2808 for (unsigned i
= 0 ; i
< num_shards
; ++i
) {
2809 shard_id_t
shard(i
);
2811 if (!find_shard(upset
, shard
)) {
2812 pg_shard_t pgs
= get_another_shard(actingset
, pg_shard_t(), shard
);
2814 if (pgs
!= pg_shard_t()) {
2817 if (pgs
== pg_whoami
)
2818 missing
= info
.stats
.stats
.sum
.num_objects_missing_on_primary
;
2820 missing
= peer_info
[pgs
].stats
.stats
.sum
.num_objects_missing
;
2822 degraded
+= missing
;
2823 misplaced
+= std::max((int64_t)0, num_objects
- missing
);
2825 // No shard anywhere
2826 degraded
+= num_objects
;
2834 // Handle undersized case
2835 if (pool
.info
.is_replicated()) {
2836 // Add to missing_target_objects
2837 assert(target
>= missing_target_objects
.size());
2838 unsigned needed
= target
- missing_target_objects
.size();
2840 missing_target_objects
.insert(make_pair(num_objects
* needed
, pg_shard_t(pg_shard_t::NO_OSD
)));
2842 for (unsigned i
= 0 ; i
< num_shards
; ++i
) {
2843 shard_id_t
shard(i
);
2845 for (const auto& t
: missing_target_objects
) {
2846 if (std::get
<1>(t
).shard
== shard
) {
2852 missing_target_objects
.insert(make_pair(num_objects
, pg_shard_t(pg_shard_t::NO_OSD
,shard
)));
2856 for (const auto& item
: missing_target_objects
)
2857 dout(20) << __func__
<< " missing shard " << std::get
<1>(item
) << " missing= " << std::get
<0>(item
) << dendl
;
2858 for (const auto& item
: acting_source_objects
)
2859 dout(20) << __func__
<< " acting shard " << std::get
<1>(item
) << " missing= " << std::get
<0>(item
) << dendl
;
2861 // Handle all objects not in missing for remapped
2863 for (auto m
= missing_target_objects
.rbegin();
2864 m
!= missing_target_objects
.rend(); ++m
) {
2866 int64_t extra_missing
= -1;
2868 if (pool
.info
.is_replicated()) {
2869 if (!acting_source_objects
.empty()) {
2870 auto extra_copy
= acting_source_objects
.begin();
2871 extra_missing
= std::get
<0>(*extra_copy
);
2872 acting_source_objects
.erase(extra_copy
);
2874 } else { // Erasure coded
2875 // Use corresponding shard
2876 for (const auto& a
: acting_source_objects
) {
2877 if (std::get
<1>(a
).shard
== std::get
<1>(*m
).shard
) {
2878 extra_missing
= std::get
<0>(a
);
2879 acting_source_objects
.erase(a
);
2885 if (extra_missing
>= 0 && std::get
<0>(*m
) >= extra_missing
) {
2886 // We don't know which of the objects on the target
2887 // are part of extra_missing so assume are all degraded.
2888 misplaced
+= std::get
<0>(*m
) - extra_missing
;
2889 degraded
+= extra_missing
;
2891 // 1. extra_missing == -1, more targets than sources so degraded
2892 // 2. extra_missing > std::get<0>(m), so that we know that some extra_missing
2893 // previously degraded are now present on the target.
2894 degraded
+= std::get
<0>(*m
);
2897 // If there are still acting that haven't been accounted for
2898 // then they are misplaced
2899 for (const auto& a
: acting_source_objects
) {
2900 int64_t extra_misplaced
= std::max((int64_t)0, num_objects
- std::get
<0>(a
));
2901 dout(20) << __func__
<< " extra acting misplaced " << extra_misplaced
<< dendl
;
2902 misplaced
+= extra_misplaced
;
2905 // NOTE: Tests use these messages to verify this code
2906 dout(20) << __func__
<< " degraded " << degraded
<< (estimate
? " (est)": "") << dendl
;
2907 dout(20) << __func__
<< " misplaced " << misplaced
<< (estimate
? " (est)": "")<< dendl
;
2909 info
.stats
.stats
.sum
.num_objects_degraded
= degraded
;
2910 info
.stats
.stats
.sum
.num_objects_unfound
= get_num_unfound();
2911 info
.stats
.stats
.sum
.num_objects_misplaced
= misplaced
;
2915 void PG::_update_blocked_by()
2917 // set a max on the number of blocking peers we report. if we go
2918 // over, report a random subset. keep the result sorted.
2919 unsigned keep
= MIN(blocked_by
.size(), cct
->_conf
->osd_max_pg_blocked_by
);
2920 unsigned skip
= blocked_by
.size() - keep
;
2921 info
.stats
.blocked_by
.clear();
2922 info
.stats
.blocked_by
.resize(keep
);
2924 for (set
<int>::iterator p
= blocked_by
.begin();
2925 p
!= blocked_by
.end() && keep
> 0;
2927 if (skip
> 0 && (rand() % (skip
+ keep
) < skip
)) {
2930 info
.stats
.blocked_by
[pos
++] = *p
;
2936 void PG::publish_stats_to_osd()
2941 pg_stats_publish_lock
.Lock();
2943 if (info
.stats
.stats
.sum
.num_scrub_errors
)
2944 state_set(PG_STATE_INCONSISTENT
);
2946 state_clear(PG_STATE_INCONSISTENT
);
2948 utime_t now
= ceph_clock_now();
2949 if (info
.stats
.state
!= state
) {
2950 info
.stats
.last_change
= now
;
2951 // Optimistic estimation, if we just find out an inactive PG,
2952 // assumt it is active till now.
2953 if (!(state
& PG_STATE_ACTIVE
) &&
2954 (info
.stats
.state
& PG_STATE_ACTIVE
))
2955 info
.stats
.last_active
= now
;
2957 if ((state
& PG_STATE_ACTIVE
) &&
2958 !(info
.stats
.state
& PG_STATE_ACTIVE
))
2959 info
.stats
.last_became_active
= now
;
2960 if ((state
& (PG_STATE_ACTIVE
|PG_STATE_PEERED
)) &&
2961 !(info
.stats
.state
& (PG_STATE_ACTIVE
|PG_STATE_PEERED
)))
2962 info
.stats
.last_became_peered
= now
;
2963 if (!(state
& PG_STATE_CREATING
) &&
2964 (info
.stats
.state
& PG_STATE_CREATING
)) {
2965 osd
->send_pg_created(get_pgid().pgid
);
2967 info
.stats
.state
= state
;
2970 _update_calc_stats();
2971 if (info
.stats
.stats
.sum
.num_objects_degraded
) {
2972 state_set(PG_STATE_DEGRADED
);
2974 state_clear(PG_STATE_DEGRADED
);
2976 _update_blocked_by();
2978 bool publish
= false;
2979 pg_stat_t pre_publish
= info
.stats
;
2980 pre_publish
.stats
.add(unstable_stats
);
2981 utime_t cutoff
= now
;
2982 cutoff
-= cct
->_conf
->osd_pg_stat_report_interval_max
;
2983 if (pg_stats_publish_valid
&& pre_publish
== pg_stats_publish
&&
2984 info
.stats
.last_fresh
> cutoff
) {
2985 dout(15) << "publish_stats_to_osd " << pg_stats_publish
.reported_epoch
2986 << ": no change since " << info
.stats
.last_fresh
<< dendl
;
2988 // update our stat summary and timestamps
2989 info
.stats
.reported_epoch
= get_osdmap()->get_epoch();
2990 ++info
.stats
.reported_seq
;
2992 info
.stats
.last_fresh
= now
;
2994 if (info
.stats
.state
& PG_STATE_CLEAN
)
2995 info
.stats
.last_clean
= now
;
2996 if (info
.stats
.state
& PG_STATE_ACTIVE
)
2997 info
.stats
.last_active
= now
;
2998 if (info
.stats
.state
& (PG_STATE_ACTIVE
|PG_STATE_PEERED
))
2999 info
.stats
.last_peered
= now
;
3000 info
.stats
.last_unstale
= now
;
3001 if ((info
.stats
.state
& PG_STATE_DEGRADED
) == 0)
3002 info
.stats
.last_undegraded
= now
;
3003 if ((info
.stats
.state
& PG_STATE_UNDERSIZED
) == 0)
3004 info
.stats
.last_fullsized
= now
;
3006 // do not send pgstat to mon anymore once we are luminous, since mgr takes
3007 // care of this by sending MMonMgrReport to mon.
3009 osd
->osd
->get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
;
3010 pg_stats_publish_valid
= true;
3011 pg_stats_publish
= pre_publish
;
3013 dout(15) << "publish_stats_to_osd " << pg_stats_publish
.reported_epoch
3014 << ":" << pg_stats_publish
.reported_seq
<< dendl
;
3016 pg_stats_publish_lock
.Unlock();
3019 osd
->pg_stat_queue_enqueue(this);
3022 void PG::clear_publish_stats()
3024 dout(15) << "clear_stats" << dendl
;
3025 pg_stats_publish_lock
.Lock();
3026 pg_stats_publish_valid
= false;
3027 pg_stats_publish_lock
.Unlock();
3029 osd
->pg_stat_queue_dequeue(this);
3033 * initialize a newly instantiated pg
3035 * Initialize PG state, as when a PG is initially created, or when it
3036 * is first instantiated on the current node.
3038 * @param role our role/rank
3039 * @param newup up set
3040 * @param newacting acting set
3041 * @param history pg history
3042 * @param pi past_intervals
3043 * @param backfill true if info should be marked as backfill
3044 * @param t transaction to write out our new state in
3048 const vector
<int>& newup
, int new_up_primary
,
3049 const vector
<int>& newacting
, int new_acting_primary
,
3050 const pg_history_t
& history
,
3051 const PastIntervals
& pi
,
3053 ObjectStore::Transaction
*t
)
3055 dout(10) << "init role " << role
<< " up " << newup
<< " acting " << newacting
3056 << " history " << history
3057 << " past_intervals " << pi
3063 init_primary_up_acting(
3067 new_acting_primary
);
3069 info
.history
= history
;
3070 past_intervals
= pi
;
3073 info
.stats
.up_primary
= new_up_primary
;
3074 info
.stats
.acting
= acting
;
3075 info
.stats
.acting_primary
= new_acting_primary
;
3076 info
.stats
.mapping_epoch
= info
.history
.same_interval_since
;
3079 dout(10) << __func__
<< ": Setting backfill" << dendl
;
3080 info
.set_last_backfill(hobject_t());
3081 info
.last_complete
= info
.last_update
;
3082 pg_log
.mark_log_for_rewrite();
3088 dirty_big_info
= true;
3092 #pragma GCC diagnostic ignored "-Wpragmas"
3093 #pragma GCC diagnostic push
3094 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
3096 void PG::upgrade(ObjectStore
*store
)
3098 assert(info_struct_v
<= 10);
3099 ObjectStore::Transaction t
;
3101 assert(info_struct_v
>= 7);
3104 if (info_struct_v
<= 7) {
3105 pg_log
.mark_log_for_rewrite();
3106 ghobject_t
log_oid(OSD::make_pg_log_oid(pg_id
));
3107 ghobject_t
biginfo_oid(OSD::make_pg_biginfo_oid(pg_id
));
3108 t
.remove(coll_t::meta(), log_oid
);
3109 t
.remove(coll_t::meta(), biginfo_oid
);
3110 t
.touch(coll
, pgmeta_oid
);
3114 if (info_struct_v
<= 8) {
3115 // no special action needed.
3119 if (info_struct_v
<= 9) {
3120 // previous versions weren't (as) aggressively clearing past_intervals
3121 if (info
.history
.last_epoch_clean
>= info
.history
.same_interval_since
) {
3122 dout(20) << __func__
<< " clearing past_intervals" << dendl
;
3123 past_intervals
.clear();
3127 // update infover_key
3128 if (info_struct_v
< cur_struct_v
) {
3129 map
<string
,bufferlist
> v
;
3130 __u8 ver
= cur_struct_v
;
3131 ::encode(ver
, v
[infover_key
]);
3132 t
.omap_setkeys(coll
, pgmeta_oid
, v
);
3136 dirty_big_info
= true;
3139 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
3140 ObjectStore::Sequencer
>("upgrade"));
3141 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3143 derr
<< __func__
<< ": apply_transaction returned "
3144 << cpp_strerror(r
) << dendl
;
3150 if (!osr
->flush_commit(&waiter
)) {
3155 #pragma GCC diagnostic pop
3156 #pragma GCC diagnostic warning "-Wpragmas"
3158 int PG::_prepare_write_info(CephContext
* cct
,
3159 map
<string
,bufferlist
> *km
,
3161 pg_info_t
&info
, pg_info_t
&last_written_info
,
3162 PastIntervals
&past_intervals
,
3163 bool dirty_big_info
,
3166 PerfCounters
*logger
)
3169 ::encode(epoch
, (*km
)[epoch_key
]);
3173 logger
->inc(l_osd_pg_info
);
3175 // try to do info efficiently?
3176 if (!dirty_big_info
&& try_fast_info
&&
3177 info
.last_update
> last_written_info
.last_update
) {
3178 pg_fast_info_t fast
;
3179 fast
.populate_from(info
);
3180 bool did
= fast
.try_apply_to(&last_written_info
);
3181 assert(did
); // we verified last_update increased above
3182 if (info
== last_written_info
) {
3183 ::encode(fast
, (*km
)[fastinfo_key
]);
3185 logger
->inc(l_osd_pg_fastinfo
);
3188 generic_dout(30) << __func__
<< " fastinfo failed, info:\n";
3190 JSONFormatter
jf(true);
3191 jf
.dump_object("info", info
);
3195 *_dout
<< "\nlast_written_info:\n";
3196 JSONFormatter
jf(true);
3197 jf
.dump_object("last_written_info", last_written_info
);
3202 last_written_info
= info
;
3204 // info. store purged_snaps separately.
3205 interval_set
<snapid_t
> purged_snaps
;
3206 purged_snaps
.swap(info
.purged_snaps
);
3207 ::encode(info
, (*km
)[info_key
]);
3208 purged_snaps
.swap(info
.purged_snaps
);
3210 if (dirty_big_info
) {
3211 // potentially big stuff
3212 bufferlist
& bigbl
= (*km
)[biginfo_key
];
3213 ::encode(past_intervals
, bigbl
);
3214 ::encode(info
.purged_snaps
, bigbl
);
3215 //dout(20) << "write_info bigbl " << bigbl.length() << dendl;
3217 logger
->inc(l_osd_pg_biginfo
);
3223 void PG::_create(ObjectStore::Transaction
& t
, spg_t pgid
, int bits
)
3226 t
.create_collection(coll
, bits
);
3229 void PG::_init(ObjectStore::Transaction
& t
, spg_t pgid
, const pg_pool_t
*pool
)
3234 // Give a hint to the PG collection
3236 uint32_t pg_num
= pool
->get_pg_num();
3237 uint64_t expected_num_objects_pg
= pool
->expected_num_objects
/ pg_num
;
3238 ::encode(pg_num
, hint
);
3239 ::encode(expected_num_objects_pg
, hint
);
3240 uint32_t hint_type
= ObjectStore::Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS
;
3241 t
.collection_hint(coll
, hint_type
, hint
);
3244 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3245 t
.touch(coll
, pgmeta_oid
);
3246 map
<string
,bufferlist
> values
;
3247 __u8 struct_v
= cur_struct_v
;
3248 ::encode(struct_v
, values
[infover_key
]);
3249 t
.omap_setkeys(coll
, pgmeta_oid
, values
);
3252 void PG::prepare_write_info(map
<string
,bufferlist
> *km
)
3254 info
.stats
.stats
.add(unstable_stats
);
3255 unstable_stats
.clear();
3257 bool need_update_epoch
= last_epoch
< get_osdmap()->get_epoch();
3258 int ret
= _prepare_write_info(cct
, km
, get_osdmap()->get_epoch(),
3262 dirty_big_info
, need_update_epoch
,
3263 cct
->_conf
->osd_fast_info
,
3266 if (need_update_epoch
)
3267 last_epoch
= get_osdmap()->get_epoch();
3268 last_persisted_osdmap_ref
= osdmap_ref
;
3271 dirty_big_info
= false;
3274 #pragma GCC diagnostic ignored "-Wpragmas"
3275 #pragma GCC diagnostic push
3276 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
3278 bool PG::_has_removal_flag(ObjectStore
*store
,
3282 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3284 // first try new way
3286 keys
.insert("_remove");
3287 map
<string
,bufferlist
> values
;
3288 if (store
->omap_get_values(coll
, pgmeta_oid
, keys
, &values
) == 0 &&
3295 int PG::peek_map_epoch(ObjectStore
*store
,
3301 ghobject_t
legacy_infos_oid(OSD::make_infos_oid());
3302 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3303 epoch_t cur_epoch
= 0;
3307 // validate collection name
3308 assert(coll
.is_pg());
3313 keys
.insert(infover_key
);
3314 keys
.insert(epoch_key
);
3315 map
<string
,bufferlist
> values
;
3316 int r
= store
->omap_get_values(coll
, pgmeta_oid
, keys
, &values
);
3318 assert(values
.size() == 2);
3320 // sanity check version
3321 bufferlist::iterator bp
= values
[infover_key
].begin();
3323 ::decode(struct_v
, bp
);
3324 assert(struct_v
>= 8);
3327 bp
= values
[epoch_key
].begin();
3328 ::decode(cur_epoch
, bp
);
3330 // probably bug 10617; see OSD::load_pgs()
3334 *pepoch
= cur_epoch
;
3338 #pragma GCC diagnostic pop
3339 #pragma GCC diagnostic warning "-Wpragmas"
3341 void PG::write_if_dirty(ObjectStore::Transaction
& t
)
3343 map
<string
,bufferlist
> km
;
3344 if (dirty_big_info
|| dirty_info
)
3345 prepare_write_info(&km
);
3346 pg_log
.write_log_and_missing(t
, &km
, coll
, pgmeta_oid
, pool
.info
.require_rollback());
3348 t
.omap_setkeys(coll
, pgmeta_oid
, km
);
3353 assert(is_primary());
3355 dout(10) << __func__
<< " to " << pg_trim_to
<< dendl
;
3356 if (pg_trim_to
!= eversion_t()) {
3357 // inform peers to trim log
3358 assert(!actingbackfill
.empty());
3359 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
3360 i
!= actingbackfill
.end();
3362 if (*i
== pg_whoami
) continue;
3363 osd
->send_message_osd_cluster(
3366 get_osdmap()->get_epoch(),
3367 spg_t(info
.pgid
.pgid
, i
->shard
),
3369 get_osdmap()->get_epoch());
3372 // trim primary as well
3373 pg_log
.trim(pg_trim_to
, info
);
3378 void PG::add_log_entry(const pg_log_entry_t
& e
, bool applied
)
3380 // raise last_complete only if we were previously up to date
3381 if (info
.last_complete
== info
.last_update
)
3382 info
.last_complete
= e
.version
;
3384 // raise last_update.
3385 assert(e
.version
> info
.last_update
);
3386 info
.last_update
= e
.version
;
3388 // raise user_version, if it increased (it may have not get bumped
3389 // by all logged updates)
3390 if (e
.user_version
> info
.last_user_version
)
3391 info
.last_user_version
= e
.user_version
;
3394 pg_log
.add(e
, applied
);
3395 dout(10) << "add_log_entry " << e
<< dendl
;
3399 void PG::append_log(
3400 const vector
<pg_log_entry_t
>& logv
,
3402 eversion_t roll_forward_to
,
3403 ObjectStore::Transaction
&t
,
3404 bool transaction_applied
)
3406 if (transaction_applied
)
3407 update_snap_map(logv
, t
);
3409 /* The primary has sent an info updating the history, but it may not
3410 * have arrived yet. We want to make sure that we cannot remember this
3411 * write without remembering that it happened in an interval which went
3412 * active in epoch history.last_epoch_started.
3414 if (info
.last_epoch_started
!= info
.history
.last_epoch_started
) {
3415 info
.history
.last_epoch_started
= info
.last_epoch_started
;
3417 if (info
.last_interval_started
!= info
.history
.last_interval_started
) {
3418 info
.history
.last_interval_started
= info
.last_interval_started
;
3420 dout(10) << "append_log " << pg_log
.get_log() << " " << logv
<< dendl
;
3422 PGLogEntryHandler handler
{this, &t
};
3423 if (!transaction_applied
) {
3424 /* We must be a backfill peer, so it's ok if we apply
3425 * out-of-turn since we won't be considered when
3426 * determining a min possible last_update.
3428 pg_log
.roll_forward(&handler
);
3431 for (vector
<pg_log_entry_t
>::const_iterator p
= logv
.begin();
3434 add_log_entry(*p
, transaction_applied
);
3436 /* We don't want to leave the rollforward artifacts around
3437 * here past last_backfill. It's ok for the same reason as
3439 if (transaction_applied
&&
3440 p
->soid
> info
.last_backfill
) {
3441 pg_log
.roll_forward(&handler
);
3444 auto last
= logv
.rbegin();
3445 if (is_primary() && last
!= logv
.rend()) {
3446 projected_log
.skip_can_rollback_to_to_head();
3447 projected_log
.trim(cct
, last
->version
, nullptr, nullptr, nullptr);
3450 if (transaction_applied
&& roll_forward_to
> pg_log
.get_can_rollback_to()) {
3451 pg_log
.roll_forward_to(
3454 t
.register_on_applied(
3455 new C_UpdateLastRollbackInfoTrimmedToApplied(
3457 get_osdmap()->get_epoch(),
3461 pg_log
.trim(trim_to
, info
);
3463 // update the local pg, pg log
3468 bool PG::check_log_for_corruption(ObjectStore
*store
)
3470 /// TODO: this method needs to work with the omap log
3474 //! Get the name we're going to save our corrupt page log as
3475 std::string
PG::get_corrupt_pg_log_name() const
3477 const int MAX_BUF
= 512;
3480 time_t my_time(time(NULL
));
3481 const struct tm
*t
= localtime_r(&my_time
, &tm_buf
);
3482 int ret
= strftime(buf
, sizeof(buf
), "corrupt_log_%Y-%m-%d_%k:%M_", t
);
3484 dout(0) << "strftime failed" << dendl
;
3485 return "corrupt_log_unknown_time";
3488 out
+= stringify(info
.pgid
);
3493 ObjectStore
*store
, spg_t pgid
, const coll_t
&coll
, bufferlist
&bl
,
3494 pg_info_t
&info
, PastIntervals
&past_intervals
,
3497 // try for v8 or later
3499 keys
.insert(infover_key
);
3500 keys
.insert(info_key
);
3501 keys
.insert(biginfo_key
);
3502 keys
.insert(fastinfo_key
);
3503 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3504 map
<string
,bufferlist
> values
;
3505 int r
= store
->omap_get_values(coll
, pgmeta_oid
, keys
, &values
);
3507 assert(values
.size() == 3 ||
3508 values
.size() == 4);
3510 bufferlist::iterator p
= values
[infover_key
].begin();
3511 ::decode(struct_v
, p
);
3512 assert(struct_v
>= 8);
3514 p
= values
[info_key
].begin();
3517 p
= values
[biginfo_key
].begin();
3518 if (struct_v
>= 10) {
3519 ::decode(past_intervals
, p
);
3521 past_intervals
.decode_classic(p
);
3523 ::decode(info
.purged_snaps
, p
);
3525 p
= values
[fastinfo_key
].begin();
3527 pg_fast_info_t fast
;
3529 fast
.try_apply_to(&info
);
3535 ghobject_t
infos_oid(OSD::make_infos_oid());
3536 bufferlist::iterator p
= bl
.begin();
3537 ::decode(struct_v
, p
);
3538 assert(struct_v
== 7);
3540 // get info out of leveldb
3541 string k
= get_info_key(info
.pgid
);
3542 string bk
= get_biginfo_key(info
.pgid
);
3547 store
->omap_get_values(coll_t::meta(), ghobject_t(infos_oid
), keys
, &values
);
3548 assert(values
.size() == 2);
3550 p
= values
[k
].begin();
3553 p
= values
[bk
].begin();
3554 ::decode(past_intervals
, p
);
3555 interval_set
<snapid_t
> snap_collections
; // obsolete
3556 ::decode(snap_collections
, p
);
3557 ::decode(info
.purged_snaps
, p
);
3561 void PG::read_state(ObjectStore
*store
, bufferlist
&bl
)
3563 int r
= read_info(store
, pg_id
, coll
, bl
, info
, past_intervals
,
3567 last_written_info
= info
;
3569 // if we are upgrading from jewel, we need to force rebuild of
3570 // missing set. v9 was fastinfo, added v11.0.2-331-g1d5dc29a13
3571 // (before kraken). persisted missing set was circa
3572 // v11.0.0-866-gb0e239da95 (a bit earlier, also before kraken).
3573 // v8 was pre-jewel (per-pg meta object).
3574 bool force_rebuild_missing
= info_struct_v
< 9;
3575 if (force_rebuild_missing
) {
3576 dout(10) << __func__
<< " detected upgrade from jewel, force_rebuild_missing"
3581 pg_log
.read_log_and_missing(
3584 info_struct_v
< 8 ? coll_t::meta() : coll
,
3585 ghobject_t(info_struct_v
< 8 ? OSD::make_pg_log_oid(pg_id
) : pgmeta_oid
),
3587 force_rebuild_missing
,
3589 cct
->_conf
->osd_ignore_stale_divergent_priors
,
3590 cct
->_conf
->osd_debug_verify_missing_on_start
);
3592 osd
->clog
->error() << oss
.str();
3594 if (force_rebuild_missing
) {
3595 dout(10) << __func__
<< " forced rebuild of missing got "
3596 << pg_log
.get_missing()
3600 // log any weirdness
3604 void PG::log_weirdness()
3606 if (pg_log
.get_tail() != info
.log_tail
)
3607 osd
->clog
->error() << info
.pgid
3608 << " info mismatch, log.tail " << pg_log
.get_tail()
3609 << " != info.log_tail " << info
.log_tail
;
3610 if (pg_log
.get_head() != info
.last_update
)
3611 osd
->clog
->error() << info
.pgid
3612 << " info mismatch, log.head " << pg_log
.get_head()
3613 << " != info.last_update " << info
.last_update
;
3615 if (!pg_log
.get_log().empty()) {
3617 if ((pg_log
.get_log().log
.begin()->version
<= pg_log
.get_tail()))
3618 osd
->clog
->error() << info
.pgid
3619 << " log bound mismatch, info (tail,head] ("
3620 << pg_log
.get_tail() << "," << pg_log
.get_head() << "]"
3622 << pg_log
.get_log().log
.begin()->version
<< ","
3623 << pg_log
.get_log().log
.rbegin()->version
<< "]";
3626 if (pg_log
.get_log().caller_ops
.size() > pg_log
.get_log().log
.size()) {
3627 osd
->clog
->error() << info
.pgid
3628 << " caller_ops.size " << pg_log
.get_log().caller_ops
.size()
3629 << " > log size " << pg_log
.get_log().log
.size();
3633 void PG::update_snap_map(
3634 const vector
<pg_log_entry_t
> &log_entries
,
3635 ObjectStore::Transaction
&t
)
3637 for (vector
<pg_log_entry_t
>::const_iterator i
= log_entries
.begin();
3638 i
!= log_entries
.end();
3640 OSDriver::OSTransaction
_t(osdriver
.get_transaction(&t
));
3641 if (i
->soid
.snap
< CEPH_MAXSNAP
) {
3642 if (i
->is_delete()) {
3643 int r
= snap_mapper
.remove_oid(
3647 } else if (i
->is_update()) {
3648 assert(i
->snaps
.length() > 0);
3649 vector
<snapid_t
> snaps
;
3650 bufferlist snapbl
= i
->snaps
;
3651 bufferlist::iterator p
= snapbl
.begin();
3655 derr
<< __func__
<< " decode snaps failure on " << *i
<< dendl
;
3658 set
<snapid_t
> _snaps(snaps
.begin(), snaps
.end());
3660 if (i
->is_clone() || i
->is_promote()) {
3661 snap_mapper
.add_oid(
3665 } else if (i
->is_modify()) {
3666 assert(i
->is_modify());
3667 int r
= snap_mapper
.update_snaps(
3674 assert(i
->is_clean());
3682 * filter trimming|trimmed snaps out of snapcontext
3684 void PG::filter_snapc(vector
<snapid_t
> &snaps
)
3686 //nothing needs to trim, we can return immediately
3687 if(snap_trimq
.empty() && info
.purged_snaps
.empty())
3690 bool filtering
= false;
3691 vector
<snapid_t
> newsnaps
;
3692 for (vector
<snapid_t
>::iterator p
= snaps
.begin();
3695 if (snap_trimq
.contains(*p
) || info
.purged_snaps
.contains(*p
)) {
3697 // start building a new vector with what we've seen so far
3698 dout(10) << "filter_snapc filtering " << snaps
<< dendl
;
3699 newsnaps
.insert(newsnaps
.begin(), snaps
.begin(), p
);
3702 dout(20) << "filter_snapc removing trimq|purged snap " << *p
<< dendl
;
3705 newsnaps
.push_back(*p
); // continue building new vector
3709 snaps
.swap(newsnaps
);
3710 dout(10) << "filter_snapc result " << snaps
<< dendl
;
3714 void PG::requeue_object_waiters(map
<hobject_t
, list
<OpRequestRef
>>& m
)
3716 for (map
<hobject_t
, list
<OpRequestRef
>>::iterator it
= m
.begin();
3719 requeue_ops(it
->second
);
3723 void PG::requeue_op(OpRequestRef op
)
3725 auto p
= waiting_for_map
.find(op
->get_source());
3726 if (p
!= waiting_for_map
.end()) {
3727 dout(20) << __func__
<< " " << op
<< " (waiting_for_map " << p
->first
<< ")"
3729 p
->second
.push_front(op
);
3731 dout(20) << __func__
<< " " << op
<< dendl
;
3732 osd
->enqueue_front(info
.pgid
, PGQueueable(op
, get_osdmap()->get_epoch()));
3736 void PG::requeue_ops(list
<OpRequestRef
> &ls
)
3738 for (list
<OpRequestRef
>::reverse_iterator i
= ls
.rbegin();
3741 auto p
= waiting_for_map
.find((*i
)->get_source());
3742 if (p
!= waiting_for_map
.end()) {
3743 dout(20) << __func__
<< " " << *i
<< " (waiting_for_map " << p
->first
3745 p
->second
.push_front(*i
);
3747 dout(20) << __func__
<< " " << *i
<< dendl
;
3748 osd
->enqueue_front(info
.pgid
, PGQueueable(*i
, get_osdmap()->get_epoch()));
3754 void PG::requeue_map_waiters()
3756 epoch_t epoch
= get_osdmap()->get_epoch();
3757 auto p
= waiting_for_map
.begin();
3758 while (p
!= waiting_for_map
.end()) {
3759 if (epoch
< p
->second
.front()->min_epoch
) {
3760 dout(20) << __func__
<< " " << p
->first
<< " front op "
3761 << p
->second
.front() << " must still wait, doing nothing"
3765 dout(20) << __func__
<< " " << p
->first
<< " " << p
->second
<< dendl
;
3766 for (auto q
= p
->second
.rbegin(); q
!= p
->second
.rend(); ++q
) {
3767 osd
->enqueue_front(info
.pgid
, PGQueueable(*q
, epoch
));
3769 p
= waiting_for_map
.erase(p
);
3775 // ==========================================================================================
3779 * when holding pg and sched_scrub_lock, then the states are:
3781 * scrubber.reserved = true
3782 * scrub_rserved_peers includes whoami
3783 * osd->scrub_pending++
3784 * scheduling, replica declined:
3785 * scrubber.reserved = true
3786 * scrubber.reserved_peers includes -1
3787 * osd->scrub_pending++
3789 * scrubber.reserved = true
3790 * scrubber.reserved_peers.size() == acting.size();
3792 * osd->scrub_pending++
3794 * scrubber.reserved = false;
3795 * scrubber.reserved_peers empty
3796 * osd->scrubber.active++
3799 // returns true if a scrub has been newly kicked off
3800 bool PG::sched_scrub()
3802 bool nodeep_scrub
= false;
3803 assert(is_locked());
3804 if (!(is_primary() && is_active() && is_clean() && !is_scrubbing())) {
3808 double deep_scrub_interval
= 0;
3809 pool
.info
.opts
.get(pool_opts_t::DEEP_SCRUB_INTERVAL
, &deep_scrub_interval
);
3810 if (deep_scrub_interval
<= 0) {
3811 deep_scrub_interval
= cct
->_conf
->osd_deep_scrub_interval
;
3813 bool time_for_deep
= ceph_clock_now() >=
3814 info
.history
.last_deep_scrub_stamp
+ deep_scrub_interval
;
3816 bool deep_coin_flip
= false;
3817 // Only add random deep scrubs when NOT user initiated scrub
3818 if (!scrubber
.must_scrub
)
3819 deep_coin_flip
= (rand() % 100) < cct
->_conf
->osd_deep_scrub_randomize_ratio
* 100;
3820 dout(20) << __func__
<< ": time_for_deep=" << time_for_deep
<< " deep_coin_flip=" << deep_coin_flip
<< dendl
;
3822 time_for_deep
= (time_for_deep
|| deep_coin_flip
);
3824 //NODEEP_SCRUB so ignore time initiated deep-scrub
3825 if (osd
->osd
->get_osdmap()->test_flag(CEPH_OSDMAP_NODEEP_SCRUB
) ||
3826 pool
.info
.has_flag(pg_pool_t::FLAG_NODEEP_SCRUB
)) {
3827 time_for_deep
= false;
3828 nodeep_scrub
= true;
3831 if (!scrubber
.must_scrub
) {
3832 assert(!scrubber
.must_deep_scrub
);
3834 //NOSCRUB so skip regular scrubs
3835 if ((osd
->osd
->get_osdmap()->test_flag(CEPH_OSDMAP_NOSCRUB
) ||
3836 pool
.info
.has_flag(pg_pool_t::FLAG_NOSCRUB
)) && !time_for_deep
) {
3837 if (scrubber
.reserved
) {
3838 // cancel scrub if it is still in scheduling,
3839 // so pgs from other pools where scrub are still legal
3840 // have a chance to go ahead with scrubbing.
3841 clear_scrub_reserved();
3842 scrub_unreserve_replicas();
3848 if (cct
->_conf
->osd_scrub_auto_repair
3849 && get_pgbackend()->auto_repair_supported()
3851 // respect the command from user, and not do auto-repair
3852 && !scrubber
.must_repair
3853 && !scrubber
.must_scrub
3854 && !scrubber
.must_deep_scrub
) {
3855 dout(20) << __func__
<< ": auto repair with deep scrubbing" << dendl
;
3856 scrubber
.auto_repair
= true;
3858 // this happens when user issue the scrub/repair command during
3859 // the scheduling of the scrub/repair (e.g. request reservation)
3860 scrubber
.auto_repair
= false;
3864 if (!scrubber
.reserved
) {
3865 assert(scrubber
.reserved_peers
.empty());
3866 if ((cct
->_conf
->osd_scrub_during_recovery
|| !osd
->is_recovery_active()) &&
3867 osd
->inc_scrubs_pending()) {
3868 dout(20) << __func__
<< ": reserved locally, reserving replicas" << dendl
;
3869 scrubber
.reserved
= true;
3870 scrubber
.reserved_peers
.insert(pg_whoami
);
3871 scrub_reserve_replicas();
3873 dout(20) << __func__
<< ": failed to reserve locally" << dendl
;
3877 if (scrubber
.reserved
) {
3878 if (scrubber
.reserve_failed
) {
3879 dout(20) << "sched_scrub: failed, a peer declined" << dendl
;
3880 clear_scrub_reserved();
3881 scrub_unreserve_replicas();
3883 } else if (scrubber
.reserved_peers
.size() == acting
.size()) {
3884 dout(20) << "sched_scrub: success, reserved self and replicas" << dendl
;
3885 if (time_for_deep
) {
3886 dout(10) << "sched_scrub: scrub will be deep" << dendl
;
3887 state_set(PG_STATE_DEEP_SCRUB
);
3888 } else if (!scrubber
.must_deep_scrub
&& info
.stats
.stats
.sum
.num_deep_scrub_errors
) {
3889 if (!nodeep_scrub
) {
3890 osd
->clog
->info() << "osd." << osd
->whoami
3891 << " pg " << info
.pgid
3892 << " Deep scrub errors, upgrading scrub to deep-scrub";
3893 state_set(PG_STATE_DEEP_SCRUB
);
3894 } else if (!scrubber
.must_scrub
) {
3895 osd
->clog
->error() << "osd." << osd
->whoami
3896 << " pg " << info
.pgid
3897 << " Regular scrub skipped due to deep-scrub errors and nodeep-scrub set";
3898 clear_scrub_reserved();
3899 scrub_unreserve_replicas();
3902 osd
->clog
->error() << "osd." << osd
->whoami
3903 << " pg " << info
.pgid
3904 << " Regular scrub request, deep-scrub details will be lost";
3909 // none declined, since scrubber.reserved is set
3910 dout(20) << "sched_scrub: reserved " << scrubber
.reserved_peers
<< ", waiting for replicas" << dendl
;
3917 void PG::reg_next_scrub()
3923 if (scrubber
.must_scrub
||
3924 (info
.stats
.stats_invalid
&& cct
->_conf
->osd_scrub_invalid_stats
)) {
3925 reg_stamp
= ceph_clock_now();
3927 reg_stamp
= info
.history
.last_scrub_stamp
;
3929 // note down the sched_time, so we can locate this scrub, and remove it
3931 double scrub_min_interval
= 0, scrub_max_interval
= 0;
3932 pool
.info
.opts
.get(pool_opts_t::SCRUB_MIN_INTERVAL
, &scrub_min_interval
);
3933 pool
.info
.opts
.get(pool_opts_t::SCRUB_MAX_INTERVAL
, &scrub_max_interval
);
3934 assert(scrubber
.scrub_reg_stamp
== utime_t());
3935 scrubber
.scrub_reg_stamp
= osd
->reg_pg_scrub(info
.pgid
,
3939 scrubber
.must_scrub
);
3942 void PG::unreg_next_scrub()
3945 osd
->unreg_pg_scrub(info
.pgid
, scrubber
.scrub_reg_stamp
);
3946 scrubber
.scrub_reg_stamp
= utime_t();
3950 void PG::do_replica_scrub_map(OpRequestRef op
)
3952 const MOSDRepScrubMap
*m
= static_cast<const MOSDRepScrubMap
*>(op
->get_req());
3953 dout(7) << __func__
<< " " << *m
<< dendl
;
3954 if (m
->map_epoch
< info
.history
.same_interval_since
) {
3955 dout(10) << __func__
<< " discarding old from "
3956 << m
->map_epoch
<< " < " << info
.history
.same_interval_since
3960 if (!scrubber
.is_chunky_scrub_active()) {
3961 dout(10) << __func__
<< " scrub isn't active" << dendl
;
3967 bufferlist::iterator p
= const_cast<bufferlist
&>(m
->get_data()).begin();
3968 scrubber
.received_maps
[m
->from
].decode(p
, info
.pgid
.pool());
3969 dout(10) << "map version is "
3970 << scrubber
.received_maps
[m
->from
].valid_through
3973 dout(10) << __func__
<< " waiting_on_whom was " << scrubber
.waiting_on_whom
3975 assert(scrubber
.waiting_on_whom
.count(m
->from
));
3976 scrubber
.waiting_on_whom
.erase(m
->from
);
3978 dout(10) << __func__
<< " replica was preempted, setting flag" << dendl
;
3979 scrub_preempted
= true;
3981 if (scrubber
.waiting_on_whom
.empty()) {
3982 if (ops_blocked_by_scrub()) {
3983 requeue_scrub(true);
3985 requeue_scrub(false);
3990 void PG::sub_op_scrub_map(OpRequestRef op
)
3992 // for legacy jewel compatibility only
3993 const MOSDSubOp
*m
= static_cast<const MOSDSubOp
*>(op
->get_req());
3994 assert(m
->get_type() == MSG_OSD_SUBOP
);
3995 dout(7) << "sub_op_scrub_map" << dendl
;
3997 if (m
->map_epoch
< info
.history
.same_interval_since
) {
3998 dout(10) << "sub_op_scrub discarding old sub_op from "
3999 << m
->map_epoch
<< " < " << info
.history
.same_interval_since
<< dendl
;
4003 if (!scrubber
.is_chunky_scrub_active()) {
4004 dout(10) << "sub_op_scrub_map scrub isn't active" << dendl
;
4010 dout(10) << " got " << m
->from
<< " scrub map" << dendl
;
4011 bufferlist::iterator p
= const_cast<bufferlist
&>(m
->get_data()).begin();
4013 scrubber
.received_maps
[m
->from
].decode(p
, info
.pgid
.pool());
4014 dout(10) << "map version is "
4015 << scrubber
.received_maps
[m
->from
].valid_through
4018 scrubber
.waiting_on_whom
.erase(m
->from
);
4020 if (scrubber
.waiting_on_whom
.empty()) {
4021 if (ops_blocked_by_scrub()) {
4022 requeue_scrub(true);
4024 requeue_scrub(false);
4029 // send scrub v3 messages (chunky scrub)
4030 void PG::_request_scrub_map(
4031 pg_shard_t replica
, eversion_t version
,
4032 hobject_t start
, hobject_t end
,
4034 bool allow_preemption
)
4036 assert(replica
!= pg_whoami
);
4037 dout(10) << "scrub requesting scrubmap from osd." << replica
4038 << " deep " << (int)deep
<< dendl
;
4039 MOSDRepScrub
*repscrubop
= new MOSDRepScrub(
4040 spg_t(info
.pgid
.pgid
, replica
.shard
), version
,
4041 get_osdmap()->get_epoch(),
4042 get_last_peering_reset(),
4046 ops_blocked_by_scrub());
4047 // default priority, we want the rep scrub processed prior to any recovery
4048 // or client io messages (we are holding a lock!)
4049 osd
->send_message_osd_cluster(
4050 replica
.osd
, repscrubop
, get_osdmap()->get_epoch());
4053 void PG::handle_scrub_reserve_request(OpRequestRef op
)
4055 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
4057 if (scrubber
.reserved
) {
4058 dout(10) << __func__
<< " ignoring reserve request: Already reserved"
4062 if ((cct
->_conf
->osd_scrub_during_recovery
|| !osd
->is_recovery_active()) &&
4063 osd
->inc_scrubs_pending()) {
4064 scrubber
.reserved
= true;
4066 dout(20) << __func__
<< ": failed to reserve remotely" << dendl
;
4067 scrubber
.reserved
= false;
4069 if (op
->get_req()->get_type() == MSG_OSD_SCRUB_RESERVE
) {
4070 const MOSDScrubReserve
*m
=
4071 static_cast<const MOSDScrubReserve
*>(op
->get_req());
4072 Message
*reply
= new MOSDScrubReserve(
4073 spg_t(info
.pgid
.pgid
, primary
.shard
),
4075 scrubber
.reserved
? MOSDScrubReserve::GRANT
: MOSDScrubReserve::REJECT
,
4077 osd
->send_message_osd_cluster(reply
, op
->get_req()->get_connection());
4079 // for jewel compat only
4080 const MOSDSubOp
*req
= static_cast<const MOSDSubOp
*>(op
->get_req());
4081 assert(req
->get_type() == MSG_OSD_SUBOP
);
4082 MOSDSubOpReply
*reply
= new MOSDSubOpReply(
4083 req
, pg_whoami
, 0, get_osdmap()->get_epoch(), CEPH_OSD_FLAG_ACK
);
4084 ::encode(scrubber
.reserved
, reply
->get_data());
4085 osd
->send_message_osd_cluster(reply
, op
->get_req()->get_connection());
4089 void PG::handle_scrub_reserve_grant(OpRequestRef op
, pg_shard_t from
)
4091 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
4093 if (!scrubber
.reserved
) {
4094 dout(10) << "ignoring obsolete scrub reserve reply" << dendl
;
4097 if (scrubber
.reserved_peers
.find(from
) != scrubber
.reserved_peers
.end()) {
4098 dout(10) << " already had osd." << from
<< " reserved" << dendl
;
4100 dout(10) << " osd." << from
<< " scrub reserve = success" << dendl
;
4101 scrubber
.reserved_peers
.insert(from
);
4106 void PG::handle_scrub_reserve_reject(OpRequestRef op
, pg_shard_t from
)
4108 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
4110 if (!scrubber
.reserved
) {
4111 dout(10) << "ignoring obsolete scrub reserve reply" << dendl
;
4114 if (scrubber
.reserved_peers
.find(from
) != scrubber
.reserved_peers
.end()) {
4115 dout(10) << " already had osd." << from
<< " reserved" << dendl
;
4117 /* One decline stops this pg from being scheduled for scrubbing. */
4118 dout(10) << " osd." << from
<< " scrub reserve = fail" << dendl
;
4119 scrubber
.reserve_failed
= true;
4124 void PG::handle_scrub_reserve_release(OpRequestRef op
)
4126 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
4128 clear_scrub_reserved();
4131 void PG::reject_reservation()
4133 osd
->send_message_osd_cluster(
4135 new MBackfillReserve(
4136 MBackfillReserve::REJECT
,
4137 spg_t(info
.pgid
.pgid
, primary
.shard
),
4138 get_osdmap()->get_epoch()),
4139 get_osdmap()->get_epoch());
4142 void PG::schedule_backfill_retry(float delay
)
4144 Mutex::Locker
lock(osd
->recovery_request_lock
);
4145 osd
->recovery_request_timer
.add_event_after(
4147 new QueuePeeringEvt
<RequestBackfill
>(
4148 this, get_osdmap()->get_epoch(),
4149 RequestBackfill()));
4152 void PG::schedule_recovery_retry(float delay
)
4154 Mutex::Locker
lock(osd
->recovery_request_lock
);
4155 osd
->recovery_request_timer
.add_event_after(
4157 new QueuePeeringEvt
<DoRecovery
>(
4158 this, get_osdmap()->get_epoch(),
4162 void PG::clear_scrub_reserved()
4164 scrubber
.reserved_peers
.clear();
4165 scrubber
.reserve_failed
= false;
4167 if (scrubber
.reserved
) {
4168 scrubber
.reserved
= false;
4169 osd
->dec_scrubs_pending();
4173 void PG::scrub_reserve_replicas()
4175 assert(backfill_targets
.empty());
4176 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
4177 i
!= actingbackfill
.end();
4179 if (*i
== pg_whoami
) continue;
4180 dout(10) << "scrub requesting reserve from osd." << *i
<< dendl
;
4181 if (HAVE_FEATURE(get_min_acting_features(), SERVER_LUMINOUS
)) {
4182 osd
->send_message_osd_cluster(
4184 new MOSDScrubReserve(spg_t(info
.pgid
.pgid
, i
->shard
),
4185 get_osdmap()->get_epoch(),
4186 MOSDScrubReserve::REQUEST
, pg_whoami
),
4187 get_osdmap()->get_epoch());
4189 // for jewel compat only
4190 vector
<OSDOp
> scrub(1);
4191 scrub
[0].op
.op
= CEPH_OSD_OP_SCRUB_RESERVE
;
4195 MOSDSubOp
*subop
= new MOSDSubOp(
4196 reqid
, pg_whoami
, spg_t(info
.pgid
.pgid
, i
->shard
), poid
, 0,
4197 get_osdmap()->get_epoch(), osd
->get_tid(), v
);
4199 osd
->send_message_osd_cluster(
4200 i
->osd
, subop
, get_osdmap()->get_epoch());
4205 void PG::scrub_unreserve_replicas()
4207 assert(backfill_targets
.empty());
4208 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
4209 i
!= actingbackfill
.end();
4211 if (*i
== pg_whoami
) continue;
4212 dout(10) << "scrub requesting unreserve from osd." << *i
<< dendl
;
4213 if (HAVE_FEATURE(get_min_acting_features(), SERVER_LUMINOUS
)) {
4214 osd
->send_message_osd_cluster(
4216 new MOSDScrubReserve(spg_t(info
.pgid
.pgid
, i
->shard
),
4217 get_osdmap()->get_epoch(),
4218 MOSDScrubReserve::RELEASE
, pg_whoami
),
4219 get_osdmap()->get_epoch());
4221 // for jewel compat only
4222 vector
<OSDOp
> scrub(1);
4223 scrub
[0].op
.op
= CEPH_OSD_OP_SCRUB_UNRESERVE
;
4227 MOSDSubOp
*subop
= new MOSDSubOp(
4228 reqid
, pg_whoami
, spg_t(info
.pgid
.pgid
, i
->shard
), poid
, 0,
4229 get_osdmap()->get_epoch(), osd
->get_tid(), v
);
4231 osd
->send_message_osd_cluster(i
->osd
, subop
, get_osdmap()->get_epoch());
4236 void PG::_scan_rollback_obs(
4237 const vector
<ghobject_t
> &rollback_obs
,
4238 ThreadPool::TPHandle
&handle
)
4240 ObjectStore::Transaction t
;
4241 eversion_t trimmed_to
= last_rollback_info_trimmed_to_applied
;
4242 for (vector
<ghobject_t
>::const_iterator i
= rollback_obs
.begin();
4243 i
!= rollback_obs
.end();
4245 if (i
->generation
< trimmed_to
.version
) {
4246 osd
->clog
->error() << "osd." << osd
->whoami
4247 << " pg " << info
.pgid
4248 << " found obsolete rollback obj "
4249 << *i
<< " generation < trimmed_to "
4256 derr
<< __func__
<< ": queueing trans to clean up obsolete rollback objs"
4258 osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
4262 void PG::_scan_snaps(ScrubMap
&smap
)
4267 // Test qa/standalone/scrub/osd-scrub-snaps.sh uses this message to verify
4268 // caller using clean_meta_map(), and it works properly.
4269 dout(20) << __func__
<< " start" << dendl
;
4271 for (map
<hobject_t
, ScrubMap::object
>::reverse_iterator i
= smap
.objects
.rbegin();
4272 i
!= smap
.objects
.rend();
4274 const hobject_t
&hoid
= i
->first
;
4275 ScrubMap::object
&o
= i
->second
;
4277 dout(20) << __func__
<< " " << hoid
<< dendl
;
4279 if (hoid
.is_head() || hoid
.is_snapdir()) {
4280 // parse the SnapSet
4282 if (o
.attrs
.find(SS_ATTR
) == o
.attrs
.end()) {
4285 bl
.push_back(o
.attrs
[SS_ATTR
]);
4286 auto p
= bl
.begin();
4288 ::decode(snapset
, p
);
4292 head
= hoid
.get_head();
4293 // Make sure head_exists is correct for is_legacy() check
4295 snapset
.head_exists
= true;
4298 if (hoid
.snap
< CEPH_MAXSNAP
) {
4299 // check and if necessary fix snap_mapper
4300 if (hoid
.get_head() != head
) {
4301 derr
<< __func__
<< " no head for " << hoid
<< " (have " << head
<< ")"
4305 set
<snapid_t
> obj_snaps
;
4306 if (!snapset
.is_legacy()) {
4307 auto p
= snapset
.clone_snaps
.find(hoid
.snap
);
4308 if (p
== snapset
.clone_snaps
.end()) {
4309 derr
<< __func__
<< " no clone_snaps for " << hoid
<< " in " << snapset
4313 obj_snaps
.insert(p
->second
.begin(), p
->second
.end());
4316 if (o
.attrs
.find(OI_ATTR
) == o
.attrs
.end()) {
4319 bl
.push_back(o
.attrs
[OI_ATTR
]);
4326 obj_snaps
.insert(oi
.legacy_snaps
.begin(), oi
.legacy_snaps
.end());
4328 set
<snapid_t
> cur_snaps
;
4329 int r
= snap_mapper
.get_snaps(hoid
, &cur_snaps
);
4330 if (r
!= 0 && r
!= -ENOENT
) {
4331 derr
<< __func__
<< ": get_snaps returned " << cpp_strerror(r
) << dendl
;
4334 if (r
== -ENOENT
|| cur_snaps
!= obj_snaps
) {
4335 ObjectStore::Transaction t
;
4336 OSDriver::OSTransaction
_t(osdriver
.get_transaction(&t
));
4338 r
= snap_mapper
.remove_oid(hoid
, &_t
);
4340 derr
<< __func__
<< ": remove_oid returned " << cpp_strerror(r
)
4344 osd
->clog
->error() << "osd." << osd
->whoami
4345 << " found snap mapper error on pg "
4347 << " oid " << hoid
<< " snaps in mapper: "
4348 << cur_snaps
<< ", oi: "
4352 osd
->clog
->error() << "osd." << osd
->whoami
4353 << " found snap mapper error on pg "
4355 << " oid " << hoid
<< " snaps missing in mapper"
4360 snap_mapper
.add_oid(hoid
, obj_snaps
, &_t
);
4362 // wait for repair to apply to avoid confusing other bits of the system.
4365 Mutex
my_lock("PG::_scan_snaps my_lock");
4368 t
.register_on_applied_sync(
4369 new C_SafeCond(&my_lock
, &my_cond
, &done
, &r
));
4370 r
= osd
->store
->apply_transaction(osr
.get(), std::move(t
));
4372 derr
<< __func__
<< ": apply_transaction got " << cpp_strerror(r
)
4377 my_cond
.Wait(my_lock
);
4386 void PG::_repair_oinfo_oid(ScrubMap
&smap
)
4388 for (map
<hobject_t
, ScrubMap::object
>::reverse_iterator i
= smap
.objects
.rbegin();
4389 i
!= smap
.objects
.rend();
4391 const hobject_t
&hoid
= i
->first
;
4392 ScrubMap::object
&o
= i
->second
;
4395 if (o
.attrs
.find(OI_ATTR
) == o
.attrs
.end()) {
4398 bl
.push_back(o
.attrs
[OI_ATTR
]);
4405 if (oi
.soid
!= hoid
) {
4406 ObjectStore::Transaction t
;
4407 OSDriver::OSTransaction
_t(osdriver
.get_transaction(&t
));
4408 osd
->clog
->error() << "osd." << osd
->whoami
4409 << " found object info error on pg "
4411 << " oid " << hoid
<< " oid in object info: "
4417 ::encode(oi
, bl
, get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
4419 bufferptr
bp(bl
.c_str(), bl
.length());
4420 o
.attrs
[OI_ATTR
] = bp
;
4422 t
.setattr(coll
, ghobject_t(hoid
), OI_ATTR
, bl
);
4423 int r
= osd
->store
->apply_transaction(osr
.get(), std::move(t
));
4425 derr
<< __func__
<< ": apply_transaction got " << cpp_strerror(r
)
4431 int PG::build_scrub_map_chunk(
4433 ScrubMapBuilder
&pos
,
4437 ThreadPool::TPHandle
&handle
)
4439 dout(10) << __func__
<< " [" << start
<< "," << end
<< ") "
4444 while (pos
.empty()) {
4446 map
.valid_through
= info
.last_update
;
4450 vector
<ghobject_t
> rollback_obs
;
4451 pos
.ret
= get_pgbackend()->objects_list_range(
4458 dout(5) << "objects_list_range error: " << pos
.ret
<< dendl
;
4461 if (pos
.ls
.empty()) {
4464 _scan_rollback_obs(rollback_obs
, handle
);
4466 return -EINPROGRESS
;
4470 while (!pos
.done()) {
4471 int r
= get_pgbackend()->be_scan_list(map
, pos
);
4472 if (r
== -EINPROGRESS
) {
4478 dout(20) << __func__
<< " finishing" << dendl
;
4480 _repair_oinfo_oid(map
);
4481 if (!is_primary()) {
4482 ScrubMap for_meta_scrub
;
4483 // In case we restarted smaller chunk, clear old data
4484 scrubber
.cleaned_meta_map
.clear_from(scrubber
.start
);
4485 scrubber
.cleaned_meta_map
.insert(map
);
4486 scrubber
.clean_meta_map(for_meta_scrub
);
4487 _scan_snaps(for_meta_scrub
);
4490 dout(20) << __func__
<< " done, got " << map
.objects
.size() << " items"
4495 void PG::Scrubber::cleanup_store(ObjectStore::Transaction
*t
) {
4498 struct OnComplete
: Context
{
4499 std::unique_ptr
<Scrub::Store
> store
;
4501 std::unique_ptr
<Scrub::Store
> &&store
)
4502 : store(std::move(store
)) {}
4503 void finish(int) override
{}
4506 t
->register_on_complete(new OnComplete(std::move(store
)));
4510 void PG::repair_object(
4511 const hobject_t
& soid
, list
<pair
<ScrubMap::object
, pg_shard_t
> > *ok_peers
,
4512 pg_shard_t bad_peer
)
4514 list
<pg_shard_t
> op_shards
;
4515 for (auto i
: *ok_peers
) {
4516 op_shards
.push_back(i
.second
);
4518 dout(10) << "repair_object " << soid
<< " bad_peer osd."
4519 << bad_peer
<< " ok_peers osd.{" << op_shards
<< "}" << dendl
;
4520 ScrubMap::object
&po
= ok_peers
->back().first
;
4523 bv
.push_back(po
.attrs
[OI_ATTR
]);
4526 bufferlist::iterator bliter
= bv
.begin();
4527 ::decode(oi
, bliter
);
4529 dout(0) << __func__
<< ": Need version of replica, bad object_info_t: " << soid
<< dendl
;
4532 if (bad_peer
!= primary
) {
4533 peer_missing
[bad_peer
].add(soid
, oi
.version
, eversion_t(), false);
4535 // We should only be scrubbing if the PG is clean.
4536 assert(waiting_for_unreadable_object
.empty());
4538 pg_log
.missing_add(soid
, oi
.version
, eversion_t());
4540 pg_log
.set_last_requested(0);
4541 dout(10) << __func__
<< ": primary = " << primary
<< dendl
;
4544 if (is_ec_pg() || bad_peer
== primary
) {
4545 // we'd better collect all shard for EC pg, and prepare good peers as the
4546 // source of pull in the case of replicated pg.
4547 missing_loc
.add_missing(soid
, oi
.version
, eversion_t());
4548 list
<pair
<ScrubMap::object
, pg_shard_t
> >::iterator i
;
4549 for (i
= ok_peers
->begin();
4550 i
!= ok_peers
->end();
4552 missing_loc
.add_location(soid
, i
->second
);
4558 * Wait for last_update_applied to match msg->scrub_to as above. Wait
4559 * for pushes to complete in case of recent recovery. Build a single
4560 * scrubmap of objects that are in the range [msg->start, msg->end).
4562 void PG::replica_scrub(
4564 ThreadPool::TPHandle
&handle
)
4566 const MOSDRepScrub
*msg
= static_cast<const MOSDRepScrub
*>(op
->get_req());
4567 assert(!scrubber
.active_rep_scrub
);
4568 dout(7) << "replica_scrub" << dendl
;
4570 if (msg
->map_epoch
< info
.history
.same_interval_since
) {
4571 dout(10) << "replica_scrub discarding old replica_scrub from "
4572 << msg
->map_epoch
<< " < " << info
.history
.same_interval_since
4577 assert(msg
->chunky
);
4578 if (last_update_applied
< msg
->scrub_to
) {
4579 dout(10) << "waiting for last_update_applied to catch up" << dendl
;
4580 scrubber
.active_rep_scrub
= op
;
4584 if (active_pushes
> 0) {
4585 dout(10) << "waiting for active pushes to finish" << dendl
;
4586 scrubber
.active_rep_scrub
= op
;
4590 scrubber
.state
= Scrubber::BUILD_MAP_REPLICA
;
4591 scrubber
.replica_scrub_start
= msg
->min_epoch
;
4592 scrubber
.start
= msg
->start
;
4593 scrubber
.end
= msg
->end
;
4594 scrubber
.max_end
= msg
->end
;
4595 scrubber
.deep
= msg
->deep
;
4596 scrubber
.epoch_start
= info
.history
.same_interval_since
;
4597 if (msg
->priority
) {
4598 scrubber
.priority
= msg
->priority
;
4600 scrubber
.priority
= get_scrub_priority();
4603 scrub_can_preempt
= msg
->allow_preemption
;
4604 scrub_preempted
= false;
4605 scrubber
.replica_scrubmap_pos
.reset();
4607 requeue_scrub(msg
->high_priority
);
4611 * PG_STATE_SCRUBBING is set when the scrub is queued
4613 * scrub will be chunky if all OSDs in PG support chunky scrub
4614 * scrub will fail if OSDs are too old.
4616 void PG::scrub(epoch_t queued
, ThreadPool::TPHandle
&handle
)
4618 if (cct
->_conf
->osd_scrub_sleep
> 0 &&
4619 (scrubber
.state
== PG::Scrubber::NEW_CHUNK
||
4620 scrubber
.state
== PG::Scrubber::INACTIVE
) &&
4621 scrubber
.needs_sleep
) {
4622 ceph_assert(!scrubber
.sleeping
);
4623 dout(20) << __func__
<< " state is INACTIVE|NEW_CHUNK, sleeping" << dendl
;
4625 // Do an async sleep so we don't block the op queue
4626 OSDService
*osds
= osd
;
4627 spg_t pgid
= get_pgid();
4628 int state
= scrubber
.state
;
4629 auto scrub_requeue_callback
=
4630 new FunctionContext([osds
, pgid
, state
](int r
) {
4631 PG
*pg
= osds
->osd
->lookup_lock_pg(pgid
);
4632 if (pg
== nullptr) {
4633 lgeneric_dout(osds
->osd
->cct
, 20)
4634 << "scrub_requeue_callback: Could not find "
4635 << "PG " << pgid
<< " can't complete scrub requeue after sleep"
4639 pg
->scrubber
.sleeping
= false;
4640 pg
->scrubber
.needs_sleep
= false;
4641 lgeneric_dout(pg
->cct
, 20)
4642 << "scrub_requeue_callback: slept for "
4643 << ceph_clock_now() - pg
->scrubber
.sleep_start
4644 << ", re-queuing scrub with state " << state
<< dendl
;
4645 pg
->scrub_queued
= false;
4646 pg
->requeue_scrub();
4647 pg
->scrubber
.sleep_start
= utime_t();
4650 Mutex::Locker
l(osd
->scrub_sleep_lock
);
4651 osd
->scrub_sleep_timer
.add_event_after(cct
->_conf
->osd_scrub_sleep
,
4652 scrub_requeue_callback
);
4653 scrubber
.sleeping
= true;
4654 scrubber
.sleep_start
= ceph_clock_now();
4657 if (pg_has_reset_since(queued
)) {
4660 assert(scrub_queued
);
4661 scrub_queued
= false;
4662 scrubber
.needs_sleep
= true;
4665 if (!is_primary() &&
4666 scrubber
.state
== PG::Scrubber::BUILD_MAP_REPLICA
) {
4667 chunky_scrub(handle
);
4671 if (!is_primary() || !is_active() || !is_clean() || !is_scrubbing()) {
4672 dout(10) << "scrub -- not primary or active or not clean" << dendl
;
4673 state_clear(PG_STATE_SCRUBBING
);
4674 state_clear(PG_STATE_REPAIR
);
4675 state_clear(PG_STATE_DEEP_SCRUB
);
4676 publish_stats_to_osd();
4680 if (!scrubber
.active
) {
4681 assert(backfill_targets
.empty());
4683 scrubber
.deep
= state_test(PG_STATE_DEEP_SCRUB
);
4685 dout(10) << "starting a new chunky scrub" << dendl
;
4688 chunky_scrub(handle
);
4692 * Chunky scrub scrubs objects one chunk at a time with writes blocked for that
4695 * The object store is partitioned into chunks which end on hash boundaries. For
4696 * each chunk, the following logic is performed:
4698 * (1) Block writes on the chunk
4699 * (2) Request maps from replicas
4700 * (3) Wait for pushes to be applied (after recovery)
4701 * (4) Wait for writes to flush on the chunk
4702 * (5) Wait for maps from replicas
4703 * (6) Compare / repair all scrub maps
4704 * (7) Wait for digest updates to apply
4706 * This logic is encoded in the mostly linear state machine:
4708 * +------------------+
4709 * _________v__________ |
4712 * |____________________| |
4715 * _________v___v______ | |
4718 * |____________________| | |
4720 * _________v__________ | |
4722 * | WAIT_PUSHES | | |
4723 * |____________________| | |
4725 * _________v__________ | |
4727 * | WAIT_LAST_UPDATE | | |
4728 * |____________________| | |
4730 * _________v__________ | |
4733 * |____________________| | |
4735 * _________v__________ | |
4737 * | WAIT_REPLICAS | | |
4738 * |____________________| | |
4740 * _________v__________ | |
4742 * | COMPARE_MAPS | | |
4743 * |____________________| | |
4746 * _________v__________ | |
4748 * |WAIT_DIGEST_UPDATES | | |
4749 * |____________________| | |
4752 * _________v__________ |
4755 * |____________________| |
4757 * +------------------+
4759 * The primary determines the last update from the subset by walking the log. If
4760 * it sees a log entry pertaining to a file in the chunk, it tells the replicas
4761 * to wait until that update is applied before building a scrub map. Both the
4762 * primary and replicas will wait for any active pushes to be applied.
4764 * In contrast to classic_scrub, chunky_scrub is entirely handled by scrub_wq.
4766 * scrubber.state encodes the current state of the scrub (refer to state diagram
4769 void PG::chunky_scrub(ThreadPool::TPHandle
&handle
)
4771 // check for map changes
4772 if (scrubber
.is_chunky_scrub_active()) {
4773 if (scrubber
.epoch_start
!= info
.history
.same_interval_since
) {
4774 dout(10) << "scrub pg changed, aborting" << dendl
;
4775 scrub_clear_state();
4776 scrub_unreserve_replicas();
4785 dout(20) << "scrub state " << Scrubber::state_string(scrubber
.state
)
4786 << " [" << scrubber
.start
<< "," << scrubber
.end
<< ")"
4787 << " max_end " << scrubber
.max_end
<< dendl
;
4789 switch (scrubber
.state
) {
4790 case PG::Scrubber::INACTIVE
:
4791 dout(10) << "scrub start" << dendl
;
4792 assert(is_primary());
4794 publish_stats_to_osd();
4795 scrubber
.epoch_start
= info
.history
.same_interval_since
;
4796 scrubber
.active
= true;
4798 osd
->inc_scrubs_active(scrubber
.reserved
);
4799 if (scrubber
.reserved
) {
4800 scrubber
.reserved
= false;
4801 scrubber
.reserved_peers
.clear();
4805 ObjectStore::Transaction t
;
4806 scrubber
.cleanup_store(&t
);
4807 scrubber
.store
.reset(Scrub::Store::create(osd
->store
, &t
,
4809 osd
->store
->queue_transaction(osr
.get(), std::move(t
), nullptr);
4812 // Don't include temporary objects when scrubbing
4813 scrubber
.start
= info
.pgid
.pgid
.get_hobj_start();
4814 scrubber
.state
= PG::Scrubber::NEW_CHUNK
;
4817 bool repair
= state_test(PG_STATE_REPAIR
);
4818 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
4819 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
4821 oss
<< info
.pgid
.pgid
<< " " << mode
<< " starts" << std::endl
;
4822 osd
->clog
->debug(oss
);
4825 scrubber
.preempt_left
= cct
->_conf
->get_val
<uint64_t>(
4826 "osd_scrub_max_preemptions");
4827 scrubber
.preempt_divisor
= 1;
4830 case PG::Scrubber::NEW_CHUNK
:
4831 scrubber
.primary_scrubmap
= ScrubMap();
4832 scrubber
.received_maps
.clear();
4834 // begin (possible) preemption window
4835 if (scrub_preempted
) {
4836 scrubber
.preempt_left
--;
4837 scrubber
.preempt_divisor
*= 2;
4838 dout(10) << __func__
<< " preempted, " << scrubber
.preempt_left
4839 << " left" << dendl
;
4840 scrub_preempted
= false;
4842 scrub_can_preempt
= scrubber
.preempt_left
> 0;
4845 /* get the start and end of our scrub chunk
4847 * Our scrub chunk has an important restriction we're going to need to
4848 * respect. We can't let head or snapdir be start or end.
4849 * Using a half-open interval means that if end == head|snapdir,
4850 * we'd scrub/lock head and the clone right next to head in different
4851 * chunks which would allow us to miss clones created between
4852 * scrubbing that chunk and scrubbing the chunk including head.
4853 * This isn't true for any of the other clones since clones can
4854 * only be created "just to the left of" head. There is one exception
4855 * to this: promotion of clones which always happens to the left of the
4856 * left-most clone, but promote_object checks the scrubber in that
4857 * case, so it should be ok. Also, it's ok to "miss" clones at the
4858 * left end of the range if we are a tier because they may legitimately
4859 * not exist (see _scrub).
4861 int min
= std::max
<int64_t>(3, cct
->_conf
->osd_scrub_chunk_min
/
4862 scrubber
.preempt_divisor
);
4863 int max
= std::max
<int64_t>(min
, cct
->_conf
->osd_scrub_chunk_max
/
4864 scrubber
.preempt_divisor
);
4865 hobject_t start
= scrubber
.start
;
4866 hobject_t candidate_end
;
4867 vector
<hobject_t
> objects
;
4869 ret
= get_pgbackend()->objects_list_partial(
4877 if (!objects
.empty()) {
4878 hobject_t back
= objects
.back();
4879 while (candidate_end
.has_snapset() &&
4880 candidate_end
.get_head() == back
.get_head()) {
4881 candidate_end
= back
;
4883 if (objects
.empty()) {
4885 "Somehow we got more than 2 objects which"
4886 "have the same head but are not clones");
4888 back
= objects
.back();
4890 if (candidate_end
.has_snapset()) {
4891 assert(candidate_end
.get_head() != back
.get_head());
4892 candidate_end
= candidate_end
.get_object_boundary();
4895 assert(candidate_end
.is_max());
4898 if (!_range_available_for_scrub(scrubber
.start
, candidate_end
)) {
4899 // we'll be requeued by whatever made us unavailable for scrub
4900 dout(10) << __func__
<< ": scrub blocked somewhere in range "
4901 << "[" << scrubber
.start
<< ", " << candidate_end
<< ")"
4906 scrubber
.end
= candidate_end
;
4907 if (scrubber
.end
> scrubber
.max_end
)
4908 scrubber
.max_end
= scrubber
.end
;
4911 // walk the log to find the latest update that affects our chunk
4912 scrubber
.subset_last_update
= eversion_t();
4913 for (auto p
= projected_log
.log
.rbegin();
4914 p
!= projected_log
.log
.rend();
4916 if (p
->soid
>= scrubber
.start
&&
4917 p
->soid
< scrubber
.end
) {
4918 scrubber
.subset_last_update
= p
->version
;
4922 if (scrubber
.subset_last_update
== eversion_t()) {
4923 for (list
<pg_log_entry_t
>::const_reverse_iterator p
=
4924 pg_log
.get_log().log
.rbegin();
4925 p
!= pg_log
.get_log().log
.rend();
4927 if (p
->soid
>= scrubber
.start
&&
4928 p
->soid
< scrubber
.end
) {
4929 scrubber
.subset_last_update
= p
->version
;
4935 // ask replicas to wait until
4936 // last_update_applied >= scrubber.subset_last_update and then scan
4937 scrubber
.waiting_on_whom
.insert(pg_whoami
);
4939 // request maps from replicas
4940 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
4941 i
!= actingbackfill
.end();
4943 if (*i
== pg_whoami
) continue;
4944 _request_scrub_map(*i
, scrubber
.subset_last_update
,
4945 scrubber
.start
, scrubber
.end
, scrubber
.deep
,
4946 scrubber
.preempt_left
> 0);
4947 scrubber
.waiting_on_whom
.insert(*i
);
4949 dout(10) << __func__
<< " waiting_on_whom " << scrubber
.waiting_on_whom
4952 scrubber
.state
= PG::Scrubber::WAIT_PUSHES
;
4955 case PG::Scrubber::WAIT_PUSHES
:
4956 if (active_pushes
== 0) {
4957 scrubber
.state
= PG::Scrubber::WAIT_LAST_UPDATE
;
4959 dout(15) << "wait for pushes to apply" << dendl
;
4964 case PG::Scrubber::WAIT_LAST_UPDATE
:
4965 if (last_update_applied
< scrubber
.subset_last_update
) {
4966 // will be requeued by op_applied
4967 dout(15) << "wait for writes to flush" << dendl
;
4972 scrubber
.state
= PG::Scrubber::BUILD_MAP
;
4973 scrubber
.primary_scrubmap_pos
.reset();
4976 case PG::Scrubber::BUILD_MAP
:
4977 assert(last_update_applied
>= scrubber
.subset_last_update
);
4979 // build my own scrub map
4980 if (scrub_preempted
) {
4981 dout(10) << __func__
<< " preempted" << dendl
;
4982 scrubber
.state
= PG::Scrubber::BUILD_MAP_DONE
;
4985 ret
= build_scrub_map_chunk(
4986 scrubber
.primary_scrubmap
,
4987 scrubber
.primary_scrubmap_pos
,
4988 scrubber
.start
, scrubber
.end
,
4991 if (ret
== -EINPROGRESS
) {
4996 scrubber
.state
= PG::Scrubber::BUILD_MAP_DONE
;
4999 case PG::Scrubber::BUILD_MAP_DONE
:
5000 if (scrubber
.primary_scrubmap_pos
.ret
< 0) {
5001 dout(5) << "error: " << scrubber
.primary_scrubmap_pos
.ret
5002 << ", aborting" << dendl
;
5003 scrub_clear_state();
5004 scrub_unreserve_replicas();
5007 dout(10) << __func__
<< " waiting_on_whom was "
5008 << scrubber
.waiting_on_whom
<< dendl
;
5009 assert(scrubber
.waiting_on_whom
.count(pg_whoami
));
5010 scrubber
.waiting_on_whom
.erase(pg_whoami
);
5012 scrubber
.state
= PG::Scrubber::WAIT_REPLICAS
;
5015 case PG::Scrubber::WAIT_REPLICAS
:
5016 if (!scrubber
.waiting_on_whom
.empty()) {
5017 // will be requeued by sub_op_scrub_map
5018 dout(10) << "wait for replicas to build scrub map" << dendl
;
5022 // end (possible) preemption window
5023 scrub_can_preempt
= false;
5024 if (scrub_preempted
) {
5025 dout(10) << __func__
<< " preempted, restarting chunk" << dendl
;
5026 scrubber
.state
= PG::Scrubber::NEW_CHUNK
;
5028 scrubber
.state
= PG::Scrubber::COMPARE_MAPS
;
5032 case PG::Scrubber::COMPARE_MAPS
:
5033 assert(last_update_applied
>= scrubber
.subset_last_update
);
5034 assert(scrubber
.waiting_on_whom
.empty());
5036 scrub_compare_maps();
5037 scrubber
.start
= scrubber
.end
;
5038 scrubber
.run_callbacks();
5040 // requeue the writes from the chunk that just finished
5041 requeue_ops(waiting_for_scrub
);
5043 scrubber
.state
= PG::Scrubber::WAIT_DIGEST_UPDATES
;
5047 case PG::Scrubber::WAIT_DIGEST_UPDATES
:
5048 if (scrubber
.num_digest_updates_pending
) {
5049 dout(10) << __func__
<< " waiting on "
5050 << scrubber
.num_digest_updates_pending
5051 << " digest updates" << dendl
;
5056 scrubber
.preempt_left
= cct
->_conf
->get_val
<uint64_t>(
5057 "osd_scrub_max_preemptions");
5058 scrubber
.preempt_divisor
= 1;
5060 if (!(scrubber
.end
.is_max())) {
5061 scrubber
.state
= PG::Scrubber::NEW_CHUNK
;
5065 scrubber
.state
= PG::Scrubber::FINISH
;
5070 case PG::Scrubber::FINISH
:
5072 scrubber
.state
= PG::Scrubber::INACTIVE
;
5075 if (!snap_trimq
.empty()) {
5076 dout(10) << "scrub finished, requeuing snap_trimmer" << dendl
;
5077 snap_trimmer_scrub_complete();
5082 case PG::Scrubber::BUILD_MAP_REPLICA
:
5083 // build my own scrub map
5084 if (scrub_preempted
) {
5085 dout(10) << __func__
<< " preempted" << dendl
;
5088 ret
= build_scrub_map_chunk(
5089 scrubber
.replica_scrubmap
,
5090 scrubber
.replica_scrubmap_pos
,
5091 scrubber
.start
, scrubber
.end
,
5095 if (ret
== -EINPROGRESS
) {
5101 if (HAVE_FEATURE(acting_features
, SERVER_LUMINOUS
)) {
5102 MOSDRepScrubMap
*reply
= new MOSDRepScrubMap(
5103 spg_t(info
.pgid
.pgid
, get_primary().shard
),
5104 scrubber
.replica_scrub_start
,
5106 reply
->preempted
= scrub_preempted
;
5107 ::encode(scrubber
.replica_scrubmap
, reply
->get_data());
5108 osd
->send_message_osd_cluster(
5109 get_primary().osd
, reply
,
5110 scrubber
.replica_scrub_start
);
5112 // for jewel compatibility
5113 vector
<OSDOp
> scrub(1);
5114 scrub
[0].op
.op
= CEPH_OSD_OP_SCRUB_MAP
;
5118 MOSDSubOp
*subop
= new MOSDSubOp(
5121 spg_t(info
.pgid
.pgid
, get_primary().shard
),
5124 scrubber
.replica_scrub_start
,
5127 ::encode(scrubber
.replica_scrubmap
, subop
->get_data());
5129 osd
->send_message_osd_cluster(
5130 get_primary().osd
, subop
,
5131 scrubber
.replica_scrub_start
);
5133 scrub_preempted
= false;
5134 scrub_can_preempt
= false;
5135 scrubber
.state
= PG::Scrubber::INACTIVE
;
5136 scrubber
.replica_scrubmap
= ScrubMap();
5137 scrubber
.replica_scrubmap_pos
= ScrubMapBuilder();
5138 scrubber
.start
= hobject_t();
5139 scrubber
.end
= hobject_t();
5140 scrubber
.max_end
= hobject_t();
5148 dout(20) << "scrub final state " << Scrubber::state_string(scrubber
.state
)
5149 << " [" << scrubber
.start
<< "," << scrubber
.end
<< ")"
5150 << " max_end " << scrubber
.max_end
<< dendl
;
5153 bool PG::write_blocked_by_scrub(const hobject_t
& soid
)
5155 if (soid
< scrubber
.start
|| soid
>= scrubber
.end
) {
5158 if (scrub_can_preempt
) {
5159 if (!scrub_preempted
) {
5160 dout(10) << __func__
<< " " << soid
<< " preempted" << dendl
;
5161 scrub_preempted
= true;
5163 dout(10) << __func__
<< " " << soid
<< " already preempted" << dendl
;
5170 bool PG::range_intersects_scrub(const hobject_t
&start
, const hobject_t
& end
)
5172 // does [start, end] intersect [scrubber.start, scrubber.max_end)
5173 return (start
< scrubber
.max_end
&&
5174 end
>= scrubber
.start
);
5177 void PG::scrub_clear_state()
5179 assert(is_locked());
5180 state_clear(PG_STATE_SCRUBBING
);
5181 state_clear(PG_STATE_REPAIR
);
5182 state_clear(PG_STATE_DEEP_SCRUB
);
5183 publish_stats_to_osd();
5185 // active -> nothing.
5186 if (scrubber
.active
)
5187 osd
->dec_scrubs_active();
5189 requeue_ops(waiting_for_scrub
);
5193 // type-specific state clear
5194 _scrub_clear_state();
5197 void PG::scrub_compare_maps()
5199 dout(10) << __func__
<< " has maps, analyzing" << dendl
;
5201 // construct authoritative scrub map for type specific scrubbing
5202 scrubber
.cleaned_meta_map
.insert(scrubber
.primary_scrubmap
);
5204 pair
<boost::optional
<uint32_t>,
5205 boost::optional
<uint32_t>>> missing_digest
;
5207 map
<pg_shard_t
, ScrubMap
*> maps
;
5208 maps
[pg_whoami
] = &scrubber
.primary_scrubmap
;
5210 for (const auto& i
: actingbackfill
) {
5211 if (i
== pg_whoami
) continue;
5212 dout(2) << __func__
<< " replica " << i
<< " has "
5213 << scrubber
.received_maps
[i
].objects
.size()
5214 << " items" << dendl
;
5215 maps
[i
] = &scrubber
.received_maps
[i
];
5218 set
<hobject_t
> master_set
;
5220 // Construct master set
5221 for (const auto map
: maps
) {
5222 for (const auto i
: map
.second
->objects
) {
5223 master_set
.insert(i
.first
);
5228 get_pgbackend()->be_large_omap_check(maps
, master_set
,
5229 scrubber
.large_omap_objects
, ss
);
5230 if (!ss
.str().empty()) {
5231 osd
->clog
->warn(ss
);
5234 if (acting
.size() > 1) {
5235 dout(10) << __func__
<< " comparing replica scrub maps" << dendl
;
5237 // Map from object with errors to good peer
5238 map
<hobject_t
, list
<pg_shard_t
>> authoritative
;
5240 dout(2) << __func__
<< " osd." << acting
[0] << " has "
5241 << scrubber
.primary_scrubmap
.objects
.size() << " items" << dendl
;
5246 get_pgbackend()->be_compare_scrubmaps(
5249 state_test(PG_STATE_REPAIR
),
5251 scrubber
.inconsistent
,
5254 scrubber
.shallow_errors
,
5255 scrubber
.deep_errors
,
5256 scrubber
.store
.get(),
5259 dout(2) << ss
.str() << dendl
;
5261 if (!ss
.str().empty()) {
5262 osd
->clog
->error(ss
);
5265 for (map
<hobject_t
, list
<pg_shard_t
>>::iterator i
= authoritative
.begin();
5266 i
!= authoritative
.end();
5268 list
<pair
<ScrubMap::object
, pg_shard_t
> > good_peers
;
5269 for (list
<pg_shard_t
>::const_iterator j
= i
->second
.begin();
5270 j
!= i
->second
.end();
5272 good_peers
.push_back(make_pair(maps
[*j
]->objects
[i
->first
], *j
));
5274 scrubber
.authoritative
.insert(
5280 for (map
<hobject_t
, list
<pg_shard_t
>>::iterator i
= authoritative
.begin();
5281 i
!= authoritative
.end();
5283 scrubber
.cleaned_meta_map
.objects
.erase(i
->first
);
5284 scrubber
.cleaned_meta_map
.objects
.insert(
5285 *(maps
[i
->second
.back()]->objects
.find(i
->first
))
5290 ScrubMap for_meta_scrub
;
5291 scrubber
.clean_meta_map(for_meta_scrub
);
5293 // ok, do the pg-type specific scrubbing
5294 scrub_snapshot_metadata(for_meta_scrub
, missing_digest
);
5295 // Called here on the primary can use an authoritative map if it isn't the primary
5296 _scan_snaps(for_meta_scrub
);
5297 if (!scrubber
.store
->empty()) {
5298 if (state_test(PG_STATE_REPAIR
)) {
5299 dout(10) << __func__
<< ": discarding scrub results" << dendl
;
5300 scrubber
.store
->flush(nullptr);
5302 dout(10) << __func__
<< ": updating scrub object" << dendl
;
5303 ObjectStore::Transaction t
;
5304 scrubber
.store
->flush(&t
);
5305 osd
->store
->queue_transaction(osr
.get(), std::move(t
), nullptr);
5310 bool PG::scrub_process_inconsistent()
5312 dout(10) << __func__
<< ": checking authoritative" << dendl
;
5313 bool repair
= state_test(PG_STATE_REPAIR
);
5314 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
5315 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
5317 // authoriative only store objects which missing or inconsistent.
5318 if (!scrubber
.authoritative
.empty()) {
5320 ss
<< info
.pgid
<< " " << mode
<< " "
5321 << scrubber
.missing
.size() << " missing, "
5322 << scrubber
.inconsistent
.size() << " inconsistent objects";
5323 dout(2) << ss
.str() << dendl
;
5324 osd
->clog
->error(ss
);
5326 state_clear(PG_STATE_CLEAN
);
5327 for (map
<hobject_t
, list
<pair
<ScrubMap::object
, pg_shard_t
> >>::iterator i
=
5328 scrubber
.authoritative
.begin();
5329 i
!= scrubber
.authoritative
.end();
5331 set
<pg_shard_t
>::iterator j
;
5333 auto missing_entry
= scrubber
.missing
.find(i
->first
);
5334 if (missing_entry
!= scrubber
.missing
.end()) {
5335 for (j
= missing_entry
->second
.begin();
5336 j
!= missing_entry
->second
.end();
5345 if (scrubber
.inconsistent
.count(i
->first
)) {
5346 for (j
= scrubber
.inconsistent
[i
->first
].begin();
5347 j
!= scrubber
.inconsistent
[i
->first
].end();
5349 repair_object(i
->first
,
5358 return (!scrubber
.authoritative
.empty() && repair
);
5361 bool PG::ops_blocked_by_scrub() const {
5362 return (waiting_for_scrub
.size() != 0);
5365 // the part that actually finalizes a scrub
5366 void PG::scrub_finish()
5368 bool repair
= state_test(PG_STATE_REPAIR
);
5369 // if the repair request comes from auto-repair and large number of errors,
5370 // we would like to cancel auto-repair
5371 if (repair
&& scrubber
.auto_repair
5372 && scrubber
.authoritative
.size() > cct
->_conf
->osd_scrub_auto_repair_num_errors
) {
5373 state_clear(PG_STATE_REPAIR
);
5376 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
5377 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
5379 // type-specific finish (can tally more errors)
5382 bool has_error
= scrub_process_inconsistent();
5386 oss
<< info
.pgid
.pgid
<< " " << mode
<< " ";
5387 int total_errors
= scrubber
.shallow_errors
+ scrubber
.deep_errors
;
5389 oss
<< total_errors
<< " errors";
5392 if (!deep_scrub
&& info
.stats
.stats
.sum
.num_deep_scrub_errors
)
5393 oss
<< " ( " << info
.stats
.stats
.sum
.num_deep_scrub_errors
5394 << " remaining deep scrub error details lost)";
5396 oss
<< ", " << scrubber
.fixed
<< " fixed";
5398 osd
->clog
->error(oss
);
5400 osd
->clog
->debug(oss
);
5405 utime_t now
= ceph_clock_now();
5406 info
.history
.last_scrub
= info
.last_update
;
5407 info
.history
.last_scrub_stamp
= now
;
5408 if (scrubber
.deep
) {
5409 info
.history
.last_deep_scrub
= info
.last_update
;
5410 info
.history
.last_deep_scrub_stamp
= now
;
5412 // Since we don't know which errors were fixed, we can only clear them
5413 // when every one has been fixed.
5415 if (scrubber
.fixed
== scrubber
.shallow_errors
+ scrubber
.deep_errors
) {
5417 scrubber
.shallow_errors
= scrubber
.deep_errors
= 0;
5419 // Deep scrub in order to get corrected error counts
5420 scrub_after_recovery
= true;
5424 if ((scrubber
.shallow_errors
== 0) && (scrubber
.deep_errors
== 0))
5425 info
.history
.last_clean_scrub_stamp
= now
;
5426 info
.stats
.stats
.sum
.num_shallow_scrub_errors
= scrubber
.shallow_errors
;
5427 info
.stats
.stats
.sum
.num_deep_scrub_errors
= scrubber
.deep_errors
;
5428 info
.stats
.stats
.sum
.num_large_omap_objects
= scrubber
.large_omap_objects
;
5430 info
.stats
.stats
.sum
.num_shallow_scrub_errors
= scrubber
.shallow_errors
;
5431 // XXX: last_clean_scrub_stamp doesn't mean the pg is not inconsistent
5432 // because of deep-scrub errors
5433 if (scrubber
.shallow_errors
== 0)
5434 info
.history
.last_clean_scrub_stamp
= now
;
5436 info
.stats
.stats
.sum
.num_scrub_errors
=
5437 info
.stats
.stats
.sum
.num_shallow_scrub_errors
+
5438 info
.stats
.stats
.sum
.num_deep_scrub_errors
;
5442 ObjectStore::Transaction t
;
5445 int tr
= osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
5451 queue_peering_event(
5453 std::make_shared
<CephPeeringEvt
>(
5454 get_osdmap()->get_epoch(),
5455 get_osdmap()->get_epoch(),
5459 scrub_clear_state();
5460 scrub_unreserve_replicas();
5462 if (is_active() && is_primary()) {
5467 void PG::share_pg_info()
5469 dout(10) << "share_pg_info" << dendl
;
5471 // share new pg_info_t with replicas
5472 assert(!actingbackfill
.empty());
5473 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
5474 i
!= actingbackfill
.end();
5476 if (*i
== pg_whoami
) continue;
5477 pg_shard_t peer
= *i
;
5478 if (peer_info
.count(peer
)) {
5479 peer_info
[peer
].last_epoch_started
= info
.last_epoch_started
;
5480 peer_info
[peer
].last_interval_started
= info
.last_interval_started
;
5481 peer_info
[peer
].history
.merge(info
.history
);
5483 MOSDPGInfo
*m
= new MOSDPGInfo(get_osdmap()->get_epoch());
5484 m
->pg_list
.push_back(
5487 peer
.shard
, pg_whoami
.shard
,
5488 get_osdmap()->get_epoch(),
5489 get_osdmap()->get_epoch(),
5492 osd
->send_message_osd_cluster(peer
.osd
, m
, get_osdmap()->get_epoch());
5496 bool PG::append_log_entries_update_missing(
5497 const mempool::osd_pglog::list
<pg_log_entry_t
> &entries
,
5498 ObjectStore::Transaction
&t
, boost::optional
<eversion_t
> trim_to
,
5499 boost::optional
<eversion_t
> roll_forward_to
)
5501 assert(!entries
.empty());
5502 assert(entries
.begin()->version
> info
.last_update
);
5504 PGLogEntryHandler rollbacker
{this, &t
};
5505 bool invalidate_stats
=
5506 pg_log
.append_new_log_entries(info
.last_backfill
,
5507 info
.last_backfill_bitwise
,
5511 if (roll_forward_to
&& entries
.rbegin()->soid
> info
.last_backfill
) {
5512 pg_log
.roll_forward(&rollbacker
);
5514 if (roll_forward_to
&& *roll_forward_to
> pg_log
.get_can_rollback_to()) {
5515 pg_log
.roll_forward_to(*roll_forward_to
, &rollbacker
);
5516 last_rollback_info_trimmed_to_applied
= *roll_forward_to
;
5519 info
.last_update
= pg_log
.get_head();
5521 if (pg_log
.get_missing().num_missing() == 0) {
5522 // advance last_complete since nothing else is missing!
5523 info
.last_complete
= info
.last_update
;
5525 info
.stats
.stats_invalid
= info
.stats
.stats_invalid
|| invalidate_stats
;
5527 dout(20) << __func__
<< "trim_to bool = " << bool(trim_to
) << " trim_to = " << (trim_to
? *trim_to
: eversion_t()) << dendl
;
5529 pg_log
.trim(*trim_to
, info
);
5532 return invalidate_stats
;
5536 void PG::merge_new_log_entries(
5537 const mempool::osd_pglog::list
<pg_log_entry_t
> &entries
,
5538 ObjectStore::Transaction
&t
,
5539 boost::optional
<eversion_t
> trim_to
,
5540 boost::optional
<eversion_t
> roll_forward_to
)
5542 dout(10) << __func__
<< " " << entries
<< dendl
;
5543 assert(is_primary());
5545 bool rebuild_missing
= append_log_entries_update_missing(entries
, t
, trim_to
, roll_forward_to
);
5546 for (set
<pg_shard_t
>::const_iterator i
= actingbackfill
.begin();
5547 i
!= actingbackfill
.end();
5549 pg_shard_t
peer(*i
);
5550 if (peer
== pg_whoami
) continue;
5551 assert(peer_missing
.count(peer
));
5552 assert(peer_info
.count(peer
));
5553 pg_missing_t
& pmissing(peer_missing
[peer
]);
5554 dout(20) << __func__
<< " peer_missing for " << peer
<< " = " << pmissing
<< dendl
;
5555 pg_info_t
& pinfo(peer_info
[peer
]);
5556 bool invalidate_stats
= PGLog::append_log_entries_update_missing(
5557 pinfo
.last_backfill
,
5558 info
.last_backfill_bitwise
,
5565 pinfo
.last_update
= info
.last_update
;
5566 pinfo
.stats
.stats_invalid
= pinfo
.stats
.stats_invalid
|| invalidate_stats
;
5567 rebuild_missing
= rebuild_missing
|| invalidate_stats
;
5570 if (!rebuild_missing
) {
5574 for (auto &&i
: entries
) {
5575 missing_loc
.rebuild(
5580 pg_log
.get_missing(),
5586 void PG::update_history(const pg_history_t
& new_history
)
5589 if (info
.history
.merge(new_history
)) {
5590 dout(20) << __func__
<< " advanced history from " << new_history
<< dendl
;
5592 if (info
.history
.last_epoch_clean
>= info
.history
.same_interval_since
) {
5593 dout(20) << __func__
<< " clearing past_intervals" << dendl
;
5594 past_intervals
.clear();
5595 dirty_big_info
= true;
5601 void PG::fulfill_info(
5602 pg_shard_t from
, const pg_query_t
&query
,
5603 pair
<pg_shard_t
, pg_info_t
> ¬ify_info
)
5605 assert(from
== primary
);
5606 assert(query
.type
== pg_query_t::INFO
);
5609 dout(10) << "sending info" << dendl
;
5610 notify_info
= make_pair(from
, info
);
5613 void PG::fulfill_log(
5614 pg_shard_t from
, const pg_query_t
&query
, epoch_t query_epoch
)
5616 dout(10) << "log request from " << from
<< dendl
;
5617 assert(from
== primary
);
5618 assert(query
.type
!= pg_query_t::INFO
);
5619 ConnectionRef con
= osd
->get_con_osd_cluster(
5620 from
.osd
, get_osdmap()->get_epoch());
5623 MOSDPGLog
*mlog
= new MOSDPGLog(
5624 from
.shard
, pg_whoami
.shard
,
5625 get_osdmap()->get_epoch(),
5627 mlog
->missing
= pg_log
.get_missing();
5629 // primary -> other, when building master log
5630 if (query
.type
== pg_query_t::LOG
) {
5631 dout(10) << " sending info+missing+log since " << query
.since
5633 if (query
.since
!= eversion_t() && query
.since
< pg_log
.get_tail()) {
5634 osd
->clog
->error() << info
.pgid
<< " got broken pg_query_t::LOG since " << query
.since
5635 << " when my log.tail is " << pg_log
.get_tail()
5636 << ", sending full log instead";
5637 mlog
->log
= pg_log
.get_log(); // primary should not have requested this!!
5639 mlog
->log
.copy_after(pg_log
.get_log(), query
.since
);
5641 else if (query
.type
== pg_query_t::FULLLOG
) {
5642 dout(10) << " sending info+missing+full log" << dendl
;
5643 mlog
->log
= pg_log
.get_log();
5646 dout(10) << " sending " << mlog
->log
<< " " << mlog
->missing
<< dendl
;
5648 osd
->share_map_peer(from
.osd
, con
.get(), get_osdmap());
5649 osd
->send_message_osd_cluster(mlog
, con
.get());
5652 void PG::fulfill_query(const MQuery
& query
, RecoveryCtx
*rctx
)
5654 if (query
.query
.type
== pg_query_t::INFO
) {
5655 pair
<pg_shard_t
, pg_info_t
> notify_info
;
5656 update_history(query
.query
.history
);
5657 fulfill_info(query
.from
, query
.query
, notify_info
);
5661 notify_info
.first
.shard
, pg_whoami
.shard
,
5663 get_osdmap()->get_epoch(),
5664 notify_info
.second
),
5667 update_history(query
.query
.history
);
5668 fulfill_log(query
.from
, query
.query
, query
.query_epoch
);
5672 void PG::check_full_transition(OSDMapRef lastmap
, OSDMapRef osdmap
)
5674 bool changed
= false;
5675 if (osdmap
->test_flag(CEPH_OSDMAP_FULL
) &&
5676 !lastmap
->test_flag(CEPH_OSDMAP_FULL
)) {
5677 dout(10) << " cluster was marked full in " << osdmap
->get_epoch() << dendl
;
5680 const pg_pool_t
*pi
= osdmap
->get_pg_pool(info
.pgid
.pool());
5682 if (pi
->has_flag(pg_pool_t::FLAG_FULL
)) {
5683 const pg_pool_t
*opi
= lastmap
->get_pg_pool(info
.pgid
.pool());
5684 if (!opi
|| !opi
->has_flag(pg_pool_t::FLAG_FULL
)) {
5685 dout(10) << " pool was marked full in " << osdmap
->get_epoch() << dendl
;
5690 info
.history
.last_epoch_marked_full
= osdmap
->get_epoch();
5695 bool PG::should_restart_peering(
5697 int newactingprimary
,
5698 const vector
<int>& newup
,
5699 const vector
<int>& newacting
,
5703 if (PastIntervals::is_new_interval(
5715 dout(20) << "new interval newup " << newup
5716 << " newacting " << newacting
<< dendl
;
5719 if (!lastmap
->is_up(osd
->whoami
) && osdmap
->is_up(osd
->whoami
)) {
5720 dout(10) << __func__
<< " osd transitioned from down -> up" << dendl
;
5726 bool PG::old_peering_msg(epoch_t reply_epoch
, epoch_t query_epoch
)
5728 if (last_peering_reset
> reply_epoch
||
5729 last_peering_reset
> query_epoch
) {
5730 dout(10) << "old_peering_msg reply_epoch " << reply_epoch
<< " query_epoch " << query_epoch
5731 << " last_peering_reset " << last_peering_reset
5738 void PG::set_last_peering_reset()
5740 dout(20) << "set_last_peering_reset " << get_osdmap()->get_epoch() << dendl
;
5741 if (last_peering_reset
!= get_osdmap()->get_epoch()) {
5742 last_peering_reset
= get_osdmap()->get_epoch();
5743 reset_interval_flush();
5750 FlushState(PG
*pg
, epoch_t epoch
) : pg(pg
), epoch(epoch
) {}
5753 if (!pg
->pg_has_reset_since(epoch
))
5754 pg
->queue_flushed(epoch
);
5758 typedef ceph::shared_ptr
<FlushState
> FlushStateRef
;
5760 void PG::start_flush(ObjectStore::Transaction
*t
,
5761 list
<Context
*> *on_applied
,
5762 list
<Context
*> *on_safe
)
5764 // flush in progress ops
5765 FlushStateRef
flush_trigger (std::make_shared
<FlushState
>(
5766 this, get_osdmap()->get_epoch()));
5768 flushes_in_progress
++;
5769 on_applied
->push_back(new ContainerContext
<FlushStateRef
>(flush_trigger
));
5770 on_safe
->push_back(new ContainerContext
<FlushStateRef
>(flush_trigger
));
5773 void PG::reset_interval_flush()
5775 dout(10) << "Clearing blocked outgoing recovery messages" << dendl
;
5776 recovery_state
.clear_blocked_outgoing();
5778 Context
*c
= new QueuePeeringEvt
<IntervalFlush
>(
5779 this, get_osdmap()->get_epoch(), IntervalFlush());
5780 if (!osr
->flush_commit(c
)) {
5781 dout(10) << "Beginning to block outgoing recovery messages" << dendl
;
5782 recovery_state
.begin_block_outgoing();
5784 dout(10) << "Not blocking outgoing recovery messages" << dendl
;
5789 /* Called before initializing peering during advance_map */
5790 void PG::start_peering_interval(
5791 const OSDMapRef lastmap
,
5792 const vector
<int>& newup
, int new_up_primary
,
5793 const vector
<int>& newacting
, int new_acting_primary
,
5794 ObjectStore::Transaction
*t
)
5796 const OSDMapRef osdmap
= get_osdmap();
5798 set_last_peering_reset();
5800 vector
<int> oldacting
, oldup
;
5801 int oldrole
= get_role();
5805 pg_shard_t old_acting_primary
= get_primary();
5806 pg_shard_t old_up_primary
= up_primary
;
5807 bool was_old_primary
= is_primary();
5808 bool was_old_replica
= is_replica();
5810 acting
.swap(oldacting
);
5812 init_primary_up_acting(
5816 new_acting_primary
);
5818 if (info
.stats
.up
!= up
||
5819 info
.stats
.acting
!= acting
||
5820 info
.stats
.up_primary
!= new_up_primary
||
5821 info
.stats
.acting_primary
!= new_acting_primary
) {
5823 info
.stats
.up_primary
= new_up_primary
;
5824 info
.stats
.acting
= acting
;
5825 info
.stats
.acting_primary
= new_acting_primary
;
5826 info
.stats
.mapping_epoch
= osdmap
->get_epoch();
5829 pg_stats_publish_lock
.Lock();
5830 pg_stats_publish_valid
= false;
5831 pg_stats_publish_lock
.Unlock();
5833 // This will now be remapped during a backfill in cases
5834 // that it would not have been before.
5836 state_set(PG_STATE_REMAPPED
);
5838 state_clear(PG_STATE_REMAPPED
);
5840 int role
= osdmap
->calc_pg_role(osd
->whoami
, acting
, acting
.size());
5841 if (pool
.info
.is_replicated() || role
== pg_whoami
.shard
)
5846 // did acting, up, primary|acker change?
5848 dout(10) << " no lastmap" << dendl
;
5850 dirty_big_info
= true;
5851 info
.history
.same_interval_since
= osdmap
->get_epoch();
5853 std::stringstream debug
;
5854 assert(info
.history
.same_interval_since
!= 0);
5855 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable(
5856 get_is_recoverable_predicate());
5857 bool new_interval
= PastIntervals::check_new_interval(
5858 old_acting_primary
.osd
,
5860 oldacting
, newacting
,
5864 info
.history
.same_interval_since
,
5865 info
.history
.last_epoch_clean
,
5872 dout(10) << __func__
<< ": check_new_interval output: "
5873 << debug
.str() << dendl
;
5875 if (osdmap
->get_epoch() == osd
->get_superblock().oldest_map
&&
5876 info
.history
.last_epoch_clean
< osdmap
->get_epoch()) {
5877 dout(10) << " map gap, clearing past_intervals and faking" << dendl
;
5878 // our information is incomplete and useless; someone else was clean
5879 // after everything we know if osdmaps were trimmed.
5880 past_intervals
.clear();
5882 dout(10) << " noting past " << past_intervals
<< dendl
;
5885 dirty_big_info
= true;
5886 info
.history
.same_interval_since
= osdmap
->get_epoch();
5887 if (info
.pgid
.pgid
.is_split(lastmap
->get_pg_num(info
.pgid
.pgid
.pool()),
5888 osdmap
->get_pg_num(info
.pgid
.pgid
.pool()),
5890 info
.history
.last_epoch_split
= osdmap
->get_epoch();
5895 if (old_up_primary
!= up_primary
||
5897 info
.history
.same_up_since
= osdmap
->get_epoch();
5899 // this comparison includes primary rank via pg_shard_t
5900 if (old_acting_primary
!= get_primary()) {
5901 info
.history
.same_primary_since
= osdmap
->get_epoch();
5906 dout(1) << __func__
<< " up " << oldup
<< " -> " << up
5907 << ", acting " << oldacting
<< " -> " << acting
5908 << ", acting_primary " << old_acting_primary
<< " -> " << new_acting_primary
5909 << ", up_primary " << old_up_primary
<< " -> " << new_up_primary
5910 << ", role " << oldrole
<< " -> " << role
5911 << ", features acting " << acting_features
5912 << " upacting " << upacting_features
5916 state_clear(PG_STATE_ACTIVE
);
5917 state_clear(PG_STATE_PEERED
);
5918 state_clear(PG_STATE_DOWN
);
5919 state_clear(PG_STATE_RECOVERY_WAIT
);
5920 state_clear(PG_STATE_RECOVERY_TOOFULL
);
5921 state_clear(PG_STATE_RECOVERING
);
5923 peer_purged
.clear();
5924 actingbackfill
.clear();
5925 scrub_queued
= false;
5927 // reset primary/replica state?
5928 if (was_old_primary
|| is_primary()) {
5929 osd
->remove_want_pg_temp(info
.pgid
.pgid
);
5930 } else if (was_old_replica
|| is_replica()) {
5931 osd
->remove_want_pg_temp(info
.pgid
.pgid
);
5933 clear_primary_state();
5939 projected_last_update
= eversion_t();
5943 // should we tell the primary we are here?
5944 send_notify
= !is_primary();
5946 if (role
!= oldrole
||
5947 was_old_primary
!= is_primary()) {
5948 // did primary change?
5949 if (was_old_primary
!= is_primary()) {
5950 state_clear(PG_STATE_CLEAN
);
5951 clear_publish_stats();
5956 // take active waiters
5957 requeue_ops(waiting_for_peered
);
5961 // did primary change?
5962 if (get_primary() != old_acting_primary
) {
5963 dout(10) << *this << " " << oldacting
<< " -> " << acting
5964 << ", acting primary "
5965 << old_acting_primary
<< " -> " << get_primary()
5968 // primary is the same.
5970 // i am (still) primary. but my replica set changed.
5971 state_clear(PG_STATE_CLEAN
);
5973 dout(10) << oldacting
<< " -> " << acting
5974 << ", replicas changed" << dendl
;
5980 if (acting
.empty() && !up
.empty() && up_primary
== pg_whoami
) {
5981 dout(10) << " acting empty, but i am up[0], clearing pg_temp" << dendl
;
5982 osd
->queue_want_pg_temp(info
.pgid
.pgid
, acting
);
5986 void PG::on_new_interval()
5988 const OSDMapRef osdmap
= get_osdmap();
5992 // initialize features
5993 acting_features
= CEPH_FEATURES_SUPPORTED_DEFAULT
;
5994 upacting_features
= CEPH_FEATURES_SUPPORTED_DEFAULT
;
5995 for (vector
<int>::iterator p
= acting
.begin(); p
!= acting
.end(); ++p
) {
5996 if (*p
== CRUSH_ITEM_NONE
)
5998 uint64_t f
= osdmap
->get_xinfo(*p
).features
;
5999 acting_features
&= f
;
6000 upacting_features
&= f
;
6002 for (vector
<int>::iterator p
= up
.begin(); p
!= up
.end(); ++p
) {
6003 if (*p
== CRUSH_ITEM_NONE
)
6005 upacting_features
&= osdmap
->get_xinfo(*p
).features
;
6011 void PG::proc_primary_info(ObjectStore::Transaction
&t
, const pg_info_t
&oinfo
)
6013 assert(!is_primary());
6015 update_history(oinfo
.history
);
6016 if (!info
.stats
.stats_invalid
&& info
.stats
.stats
.sum
.num_scrub_errors
) {
6017 info
.stats
.stats
.sum
.num_scrub_errors
= 0;
6018 info
.stats
.stats
.sum
.num_shallow_scrub_errors
= 0;
6019 info
.stats
.stats
.sum
.num_deep_scrub_errors
= 0;
6023 if (!(info
.purged_snaps
== oinfo
.purged_snaps
)) {
6024 dout(10) << __func__
<< " updating purged_snaps to " << oinfo
.purged_snaps
6026 info
.purged_snaps
= oinfo
.purged_snaps
;
6028 dirty_big_info
= true;
6032 ostream
& operator<<(ostream
& out
, const PG
& pg
)
6034 out
<< "pg[" << pg
.info
6036 if (pg
.acting
!= pg
.up
)
6037 out
<< "/" << pg
.acting
;
6039 out
<< "p" << pg
.get_primary();
6040 out
<< " r=" << pg
.get_role();
6041 out
<< " lpr=" << pg
.get_last_peering_reset();
6043 if (!pg
.past_intervals
.empty()) {
6044 out
<< " pi=[" << pg
.past_intervals
.get_bounds()
6045 << ")/" << pg
.past_intervals
.size();
6048 if (pg
.is_peered()) {
6049 if (pg
.last_update_ondisk
!= pg
.info
.last_update
)
6050 out
<< " luod=" << pg
.last_update_ondisk
;
6051 if (pg
.last_update_applied
!= pg
.info
.last_update
)
6052 out
<< " lua=" << pg
.last_update_applied
;
6055 if (pg
.recovery_ops_active
)
6056 out
<< " rops=" << pg
.recovery_ops_active
;
6058 if (pg
.pg_log
.get_tail() != pg
.info
.log_tail
||
6059 pg
.pg_log
.get_head() != pg
.info
.last_update
)
6060 out
<< " (info mismatch, " << pg
.pg_log
.get_log() << ")";
6062 if (!pg
.pg_log
.get_log().empty()) {
6063 if ((pg
.pg_log
.get_log().log
.begin()->version
<= pg
.pg_log
.get_tail())) {
6064 out
<< " (log bound mismatch, actual=["
6065 << pg
.pg_log
.get_log().log
.begin()->version
<< ","
6066 << pg
.pg_log
.get_log().log
.rbegin()->version
<< "]";
6071 if (!pg
.backfill_targets
.empty())
6072 out
<< " bft=" << pg
.backfill_targets
;
6073 out
<< " crt=" << pg
.pg_log
.get_can_rollback_to();
6075 if (pg
.last_complete_ondisk
!= pg
.info
.last_complete
)
6076 out
<< " lcod " << pg
.last_complete_ondisk
;
6078 if (pg
.is_primary()) {
6079 out
<< " mlcod " << pg
.min_last_complete_ondisk
;
6082 out
<< " " << pg_state_string(pg
.get_state());
6083 if (pg
.should_send_notify())
6086 if (pg
.scrubber
.must_repair
)
6087 out
<< " MUST_REPAIR";
6088 if (pg
.scrubber
.auto_repair
)
6089 out
<< " AUTO_REPAIR";
6090 if (pg
.scrubber
.must_deep_scrub
)
6091 out
<< " MUST_DEEP_SCRUB";
6092 if (pg
.scrubber
.must_scrub
)
6093 out
<< " MUST_SCRUB";
6095 //out << " (" << pg.pg_log.get_tail() << "," << pg.pg_log.get_head() << "]";
6096 if (pg
.pg_log
.get_missing().num_missing()) {
6097 out
<< " m=" << pg
.pg_log
.get_missing().num_missing();
6098 if (pg
.is_primary()) {
6099 uint64_t unfound
= pg
.get_num_unfound();
6101 out
<< " u=" << unfound
;
6104 if (pg
.snap_trimq
.size())
6105 out
<< " snaptrimq=" << pg
.snap_trimq
;
6106 if (!pg
.is_clean()) {
6107 out
<< " mbc=" << pg
.missing_loc
.get_missing_by_count();
6116 bool PG::can_discard_op(OpRequestRef
& op
)
6118 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
6119 if (cct
->_conf
->osd_discard_disconnected_ops
&& OSD::op_is_discardable(m
)) {
6120 dout(20) << " discard " << *m
<< dendl
;
6124 if (m
->get_map_epoch() < info
.history
.same_primary_since
) {
6125 dout(7) << " changed after " << m
->get_map_epoch()
6126 << ", dropping " << *m
<< dendl
;
6130 if (m
->get_connection()->has_feature(CEPH_FEATURE_RESEND_ON_SPLIT
)) {
6131 if (m
->get_map_epoch() < pool
.info
.get_last_force_op_resend()) {
6132 dout(7) << __func__
<< " sent before last_force_op_resend "
6133 << pool
.info
.last_force_op_resend
<< ", dropping" << *m
<< dendl
;
6136 if (m
->get_map_epoch() < info
.history
.last_epoch_split
) {
6137 dout(7) << __func__
<< " pg split in "
6138 << info
.history
.last_epoch_split
<< ", dropping" << dendl
;
6141 } else if (m
->get_connection()->has_feature(CEPH_FEATURE_OSD_POOLRESEND
)) {
6142 if (m
->get_map_epoch() < pool
.info
.get_last_force_op_resend_preluminous()) {
6143 dout(7) << __func__
<< " sent before last_force_op_resend_preluminous "
6144 << pool
.info
.last_force_op_resend_preluminous
6145 << ", dropping" << *m
<< dendl
;
6153 template<typename T
, int MSGTYPE
>
6154 bool PG::can_discard_replica_op(OpRequestRef
& op
)
6156 const T
*m
= static_cast<const T
*>(op
->get_req());
6157 assert(m
->get_type() == MSGTYPE
);
6159 int from
= m
->get_source().num();
6161 // if a repop is replied after a replica goes down in a new osdmap, and
6162 // before the pg advances to this new osdmap, the repop replies before this
6163 // repop can be discarded by that replica OSD, because the primary resets the
6164 // connection to it when handling the new osdmap marking it down, and also
6165 // resets the messenger sesssion when the replica reconnects. to avoid the
6166 // out-of-order replies, the messages from that replica should be discarded.
6167 if (osd
->get_osdmap()->is_down(from
))
6169 /* Mostly, this overlaps with the old_peering_msg
6170 * condition. An important exception is pushes
6171 * sent by replicas not in the acting set, since
6172 * if such a replica goes down it does not cause
6173 * a new interval. */
6174 if (get_osdmap()->get_down_at(from
) >= m
->map_epoch
)
6178 // if pg changes _at all_, we reset and repeer!
6179 if (old_peering_msg(m
->map_epoch
, m
->map_epoch
)) {
6180 dout(10) << "can_discard_replica_op pg changed " << info
.history
6181 << " after " << m
->map_epoch
6182 << ", dropping" << dendl
;
6188 bool PG::can_discard_scan(OpRequestRef op
)
6190 const MOSDPGScan
*m
= static_cast<const MOSDPGScan
*>(op
->get_req());
6191 assert(m
->get_type() == MSG_OSD_PG_SCAN
);
6193 if (old_peering_msg(m
->map_epoch
, m
->query_epoch
)) {
6194 dout(10) << " got old scan, ignoring" << dendl
;
6200 bool PG::can_discard_backfill(OpRequestRef op
)
6202 const MOSDPGBackfill
*m
= static_cast<const MOSDPGBackfill
*>(op
->get_req());
6203 assert(m
->get_type() == MSG_OSD_PG_BACKFILL
);
6205 if (old_peering_msg(m
->map_epoch
, m
->query_epoch
)) {
6206 dout(10) << " got old backfill, ignoring" << dendl
;
6214 bool PG::can_discard_request(OpRequestRef
& op
)
6216 switch (op
->get_req()->get_type()) {
6217 case CEPH_MSG_OSD_OP
:
6218 return can_discard_op(op
);
6219 case CEPH_MSG_OSD_BACKOFF
:
6220 return false; // never discard
6222 return can_discard_replica_op
<MOSDSubOp
, MSG_OSD_SUBOP
>(op
);
6224 return can_discard_replica_op
<MOSDRepOp
, MSG_OSD_REPOP
>(op
);
6225 case MSG_OSD_PG_PUSH
:
6226 return can_discard_replica_op
<MOSDPGPush
, MSG_OSD_PG_PUSH
>(op
);
6227 case MSG_OSD_PG_PULL
:
6228 return can_discard_replica_op
<MOSDPGPull
, MSG_OSD_PG_PULL
>(op
);
6229 case MSG_OSD_PG_PUSH_REPLY
:
6230 return can_discard_replica_op
<MOSDPGPushReply
, MSG_OSD_PG_PUSH_REPLY
>(op
);
6231 case MSG_OSD_SUBOPREPLY
:
6232 return can_discard_replica_op
<MOSDSubOpReply
, MSG_OSD_SUBOPREPLY
>(op
);
6233 case MSG_OSD_REPOPREPLY
:
6234 return can_discard_replica_op
<MOSDRepOpReply
, MSG_OSD_REPOPREPLY
>(op
);
6235 case MSG_OSD_PG_RECOVERY_DELETE
:
6236 return can_discard_replica_op
<MOSDPGRecoveryDelete
, MSG_OSD_PG_RECOVERY_DELETE
>(op
);
6238 case MSG_OSD_PG_RECOVERY_DELETE_REPLY
:
6239 return can_discard_replica_op
<MOSDPGRecoveryDeleteReply
, MSG_OSD_PG_RECOVERY_DELETE_REPLY
>(op
);
6241 case MSG_OSD_EC_WRITE
:
6242 return can_discard_replica_op
<MOSDECSubOpWrite
, MSG_OSD_EC_WRITE
>(op
);
6243 case MSG_OSD_EC_WRITE_REPLY
:
6244 return can_discard_replica_op
<MOSDECSubOpWriteReply
, MSG_OSD_EC_WRITE_REPLY
>(op
);
6245 case MSG_OSD_EC_READ
:
6246 return can_discard_replica_op
<MOSDECSubOpRead
, MSG_OSD_EC_READ
>(op
);
6247 case MSG_OSD_EC_READ_REPLY
:
6248 return can_discard_replica_op
<MOSDECSubOpReadReply
, MSG_OSD_EC_READ_REPLY
>(op
);
6249 case MSG_OSD_REP_SCRUB
:
6250 return can_discard_replica_op
<MOSDRepScrub
, MSG_OSD_REP_SCRUB
>(op
);
6251 case MSG_OSD_SCRUB_RESERVE
:
6252 return can_discard_replica_op
<MOSDScrubReserve
, MSG_OSD_SCRUB_RESERVE
>(op
);
6253 case MSG_OSD_REP_SCRUBMAP
:
6254 return can_discard_replica_op
<MOSDRepScrubMap
, MSG_OSD_REP_SCRUBMAP
>(op
);
6255 case MSG_OSD_PG_UPDATE_LOG_MISSING
:
6256 return can_discard_replica_op
<
6257 MOSDPGUpdateLogMissing
, MSG_OSD_PG_UPDATE_LOG_MISSING
>(op
);
6258 case MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY
:
6259 return can_discard_replica_op
<
6260 MOSDPGUpdateLogMissingReply
, MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY
>(op
);
6262 case MSG_OSD_PG_SCAN
:
6263 return can_discard_scan(op
);
6264 case MSG_OSD_PG_BACKFILL
:
6265 return can_discard_backfill(op
);
6266 case MSG_OSD_PG_BACKFILL_REMOVE
:
6267 return can_discard_replica_op
<MOSDPGBackfillRemove
,
6268 MSG_OSD_PG_BACKFILL_REMOVE
>(op
);
6273 void PG::take_waiters()
6275 dout(10) << "take_waiters" << dendl
;
6276 requeue_map_waiters();
6277 for (list
<CephPeeringEvtRef
>::iterator i
= peering_waiters
.begin();
6278 i
!= peering_waiters
.end();
6279 ++i
) osd
->queue_for_peering(this);
6280 peering_queue
.splice(peering_queue
.begin(), peering_waiters
,
6281 peering_waiters
.begin(), peering_waiters
.end());
6284 void PG::handle_peering_event(CephPeeringEvtRef evt
, RecoveryCtx
*rctx
)
6286 dout(10) << "handle_peering_event: " << evt
->get_desc() << dendl
;
6287 if (!have_same_or_newer_map(evt
->get_epoch_sent())) {
6288 dout(10) << "deferring event " << evt
->get_desc() << dendl
;
6289 peering_waiters
.push_back(evt
);
6292 if (old_peering_evt(evt
))
6294 recovery_state
.handle_event(evt
, rctx
);
6297 void PG::queue_peering_event(CephPeeringEvtRef evt
)
6299 if (old_peering_evt(evt
))
6301 peering_queue
.push_back(evt
);
6302 osd
->queue_for_peering(this);
6305 void PG::queue_null(epoch_t msg_epoch
,
6306 epoch_t query_epoch
)
6308 dout(10) << "null" << dendl
;
6309 queue_peering_event(
6310 CephPeeringEvtRef(std::make_shared
<CephPeeringEvt
>(msg_epoch
, query_epoch
,
6314 void PG::queue_flushed(epoch_t e
)
6316 dout(10) << "flushed" << dendl
;
6317 queue_peering_event(
6318 CephPeeringEvtRef(std::make_shared
<CephPeeringEvt
>(e
, e
,
6322 void PG::queue_query(epoch_t msg_epoch
,
6323 epoch_t query_epoch
,
6324 pg_shard_t from
, const pg_query_t
& q
)
6326 dout(10) << "handle_query " << q
<< " from replica " << from
<< dendl
;
6327 queue_peering_event(
6328 CephPeeringEvtRef(std::make_shared
<CephPeeringEvt
>(msg_epoch
, query_epoch
,
6329 MQuery(from
, q
, query_epoch
))));
6332 void PG::handle_advance_map(
6333 OSDMapRef osdmap
, OSDMapRef lastmap
,
6334 vector
<int>& newup
, int up_primary
,
6335 vector
<int>& newacting
, int acting_primary
,
6338 assert(lastmap
->get_epoch() == osdmap_ref
->get_epoch());
6339 assert(lastmap
== osdmap_ref
);
6340 dout(10) << "handle_advance_map "
6341 << newup
<< "/" << newacting
6342 << " -- " << up_primary
<< "/" << acting_primary
6344 update_osdmap_ref(osdmap
);
6345 pool
.update(osdmap
);
6346 past_intervals
.update_type_from_map(pool
.info
.ec_pool(), *osdmap
);
6347 if (cct
->_conf
->osd_debug_verify_cached_snaps
) {
6348 interval_set
<snapid_t
> actual_removed_snaps
;
6349 const pg_pool_t
*pi
= osdmap
->get_pg_pool(info
.pgid
.pool());
6351 pi
->build_removed_snaps(actual_removed_snaps
);
6352 if (!(actual_removed_snaps
== pool
.cached_removed_snaps
)) {
6353 derr
<< __func__
<< ": mismatch between the actual removed snaps "
6354 << actual_removed_snaps
<< " and pool.cached_removed_snaps "
6355 << " pool.cached_removed_snaps " << pool
.cached_removed_snaps
6358 assert(actual_removed_snaps
== pool
.cached_removed_snaps
);
6361 osdmap
, lastmap
, newup
, up_primary
,
6362 newacting
, acting_primary
);
6363 recovery_state
.handle_event(evt
, rctx
);
6364 if (pool
.info
.last_change
== osdmap_ref
->get_epoch()) {
6366 update_store_with_options();
6370 void PG::handle_activate_map(RecoveryCtx
*rctx
)
6372 dout(10) << "handle_activate_map " << dendl
;
6374 recovery_state
.handle_event(evt
, rctx
);
6375 if (osdmap_ref
->get_epoch() - last_persisted_osdmap_ref
->get_epoch() >
6376 cct
->_conf
->osd_pg_epoch_persisted_max_stale
) {
6377 dout(20) << __func__
<< ": Dirtying info: last_persisted is "
6378 << last_persisted_osdmap_ref
->get_epoch()
6379 << " while current is " << osdmap_ref
->get_epoch() << dendl
;
6382 dout(20) << __func__
<< ": Not dirtying info: last_persisted is "
6383 << last_persisted_osdmap_ref
->get_epoch()
6384 << " while current is " << osdmap_ref
->get_epoch() << dendl
;
6386 if (osdmap_ref
->check_new_blacklist_entries()) check_blacklisted_watchers();
6389 void PG::handle_loaded(RecoveryCtx
*rctx
)
6391 dout(10) << "handle_loaded" << dendl
;
6393 recovery_state
.handle_event(evt
, rctx
);
6396 void PG::handle_create(RecoveryCtx
*rctx
)
6398 dout(10) << "handle_create" << dendl
;
6399 rctx
->created_pgs
.insert(this);
6401 recovery_state
.handle_event(evt
, rctx
);
6403 recovery_state
.handle_event(evt2
, rctx
);
6405 rctx
->on_applied
->add(make_lambda_context([this]() {
6406 update_store_with_options();
6410 void PG::handle_query_state(Formatter
*f
)
6412 dout(10) << "handle_query_state" << dendl
;
6414 recovery_state
.handle_event(q
, 0);
6417 void PG::update_store_with_options()
6419 auto r
= osd
->store
->set_collection_opts(coll
, pool
.info
.opts
);
6420 if(r
< 0 && r
!= -EOPNOTSUPP
) {
6421 derr
<< __func__
<< " set_collection_opts returns error:" << r
<< dendl
;
6425 void PG::update_store_on_load()
6427 if (osd
->store
->get_type() == "filestore") {
6428 // legacy filestore didn't store collection bit width; fix.
6429 int bits
= osd
->store
->collection_bits(coll
);
6431 assert(!coll
.is_meta()); // otherwise OSD::load_pgs() did a bad thing
6432 bits
= info
.pgid
.get_split_bits(pool
.info
.get_pg_num());
6433 lderr(cct
) << __func__
<< " setting bit width to " << bits
<< dendl
;
6434 ObjectStore::Transaction t
;
6435 t
.collection_set_bits(coll
, bits
);
6436 osd
->store
->apply_transaction(osr
.get(), std::move(t
));
6441 /*------------ Recovery State Machine----------------*/
6443 #define dout_prefix (*_dout << context< RecoveryMachine >().pg->gen_prefix() \
6444 << "state<" << get_state_name() << ">: ")
6446 /*------Crashed-------*/
6447 PG::RecoveryState::Crashed::Crashed(my_context ctx
)
6449 NamedState(context
< RecoveryMachine
>().pg
, "Crashed")
6451 context
< RecoveryMachine
>().log_enter(state_name
);
6452 assert(0 == "we got a bad state machine event");
6456 /*------Initial-------*/
6457 PG::RecoveryState::Initial::Initial(my_context ctx
)
6459 NamedState(context
< RecoveryMachine
>().pg
, "Initial")
6461 context
< RecoveryMachine
>().log_enter(state_name
);
6464 boost::statechart::result
PG::RecoveryState::Initial::react(const Load
& l
)
6466 PG
*pg
= context
< RecoveryMachine
>().pg
;
6468 // do we tell someone we're here?
6469 pg
->send_notify
= (!pg
->is_primary());
6470 pg
->update_store_with_options();
6472 pg
->update_store_on_load();
6474 return transit
< Reset
>();
6477 boost::statechart::result
PG::RecoveryState::Initial::react(const MNotifyRec
& notify
)
6479 PG
*pg
= context
< RecoveryMachine
>().pg
;
6480 pg
->proc_replica_info(
6481 notify
.from
, notify
.notify
.info
, notify
.notify
.epoch_sent
);
6482 pg
->set_last_peering_reset();
6483 return transit
< Primary
>();
6486 boost::statechart::result
PG::RecoveryState::Initial::react(const MInfoRec
& i
)
6488 PG
*pg
= context
< RecoveryMachine
>().pg
;
6489 assert(!pg
->is_primary());
6491 return transit
< Stray
>();
6494 boost::statechart::result
PG::RecoveryState::Initial::react(const MLogRec
& i
)
6496 PG
*pg
= context
< RecoveryMachine
>().pg
;
6497 assert(!pg
->is_primary());
6499 return transit
< Stray
>();
6502 void PG::RecoveryState::Initial::exit()
6504 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6505 PG
*pg
= context
< RecoveryMachine
>().pg
;
6506 utime_t dur
= ceph_clock_now() - enter_time
;
6507 pg
->osd
->recoverystate_perf
->tinc(rs_initial_latency
, dur
);
6510 /*------Started-------*/
6511 PG::RecoveryState::Started::Started(my_context ctx
)
6513 NamedState(context
< RecoveryMachine
>().pg
, "Started")
6515 context
< RecoveryMachine
>().log_enter(state_name
);
6518 boost::statechart::result
6519 PG::RecoveryState::Started::react(const IntervalFlush
&)
6521 PG
*pg
= context
< RecoveryMachine
>().pg
;
6522 ldout(pg
->cct
, 10) << "Ending blocked outgoing recovery messages" << dendl
;
6523 context
< RecoveryMachine
>().pg
->recovery_state
.end_block_outgoing();
6524 return discard_event();
6528 boost::statechart::result
6529 PG::RecoveryState::Started::react(const FlushedEvt
&)
6531 PG
*pg
= context
< RecoveryMachine
>().pg
;
6533 return discard_event();
6537 boost::statechart::result
PG::RecoveryState::Started::react(const AdvMap
& advmap
)
6539 PG
*pg
= context
< RecoveryMachine
>().pg
;
6540 ldout(pg
->cct
, 10) << "Started advmap" << dendl
;
6541 pg
->check_full_transition(advmap
.lastmap
, advmap
.osdmap
);
6542 if (pg
->should_restart_peering(
6544 advmap
.acting_primary
,
6549 ldout(pg
->cct
, 10) << "should_restart_peering, transitioning to Reset"
6552 return transit
< Reset
>();
6554 pg
->remove_down_peer_info(advmap
.osdmap
);
6555 return discard_event();
6558 boost::statechart::result
PG::RecoveryState::Started::react(const QueryState
& q
)
6560 q
.f
->open_object_section("state");
6561 q
.f
->dump_string("name", state_name
);
6562 q
.f
->dump_stream("enter_time") << enter_time
;
6563 q
.f
->close_section();
6564 return discard_event();
6567 void PG::RecoveryState::Started::exit()
6569 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6570 PG
*pg
= context
< RecoveryMachine
>().pg
;
6571 utime_t dur
= ceph_clock_now() - enter_time
;
6572 pg
->osd
->recoverystate_perf
->tinc(rs_started_latency
, dur
);
6575 /*--------Reset---------*/
6576 PG::RecoveryState::Reset::Reset(my_context ctx
)
6578 NamedState(context
< RecoveryMachine
>().pg
, "Reset")
6580 context
< RecoveryMachine
>().log_enter(state_name
);
6581 PG
*pg
= context
< RecoveryMachine
>().pg
;
6583 pg
->flushes_in_progress
= 0;
6584 pg
->set_last_peering_reset();
6587 boost::statechart::result
6588 PG::RecoveryState::Reset::react(const FlushedEvt
&)
6590 PG
*pg
= context
< RecoveryMachine
>().pg
;
6592 return discard_event();
6595 boost::statechart::result
6596 PG::RecoveryState::Reset::react(const IntervalFlush
&)
6598 PG
*pg
= context
< RecoveryMachine
>().pg
;
6599 ldout(pg
->cct
, 10) << "Ending blocked outgoing recovery messages" << dendl
;
6600 context
< RecoveryMachine
>().pg
->recovery_state
.end_block_outgoing();
6601 return discard_event();
6604 boost::statechart::result
PG::RecoveryState::Reset::react(const AdvMap
& advmap
)
6606 PG
*pg
= context
< RecoveryMachine
>().pg
;
6607 ldout(pg
->cct
, 10) << "Reset advmap" << dendl
;
6609 pg
->check_full_transition(advmap
.lastmap
, advmap
.osdmap
);
6611 if (pg
->should_restart_peering(
6613 advmap
.acting_primary
,
6618 ldout(pg
->cct
, 10) << "should restart peering, calling start_peering_interval again"
6620 pg
->start_peering_interval(
6622 advmap
.newup
, advmap
.up_primary
,
6623 advmap
.newacting
, advmap
.acting_primary
,
6624 context
< RecoveryMachine
>().get_cur_transaction());
6626 pg
->remove_down_peer_info(advmap
.osdmap
);
6627 pg
->check_past_interval_bounds();
6628 return discard_event();
6631 boost::statechart::result
PG::RecoveryState::Reset::react(const ActMap
&)
6633 PG
*pg
= context
< RecoveryMachine
>().pg
;
6634 if (pg
->should_send_notify() && pg
->get_primary().osd
>= 0) {
6635 context
< RecoveryMachine
>().send_notify(
6638 pg
->get_primary().shard
, pg
->pg_whoami
.shard
,
6639 pg
->get_osdmap()->get_epoch(),
6640 pg
->get_osdmap()->get_epoch(),
6642 pg
->past_intervals
);
6645 pg
->update_heartbeat_peers();
6648 return transit
< Started
>();
6651 boost::statechart::result
PG::RecoveryState::Reset::react(const QueryState
& q
)
6653 q
.f
->open_object_section("state");
6654 q
.f
->dump_string("name", state_name
);
6655 q
.f
->dump_stream("enter_time") << enter_time
;
6656 q
.f
->close_section();
6657 return discard_event();
6660 void PG::RecoveryState::Reset::exit()
6662 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6663 PG
*pg
= context
< RecoveryMachine
>().pg
;
6664 utime_t dur
= ceph_clock_now() - enter_time
;
6665 pg
->osd
->recoverystate_perf
->tinc(rs_reset_latency
, dur
);
6668 /*-------Start---------*/
6669 PG::RecoveryState::Start::Start(my_context ctx
)
6671 NamedState(context
< RecoveryMachine
>().pg
, "Start")
6673 context
< RecoveryMachine
>().log_enter(state_name
);
6675 PG
*pg
= context
< RecoveryMachine
>().pg
;
6676 if (pg
->is_primary()) {
6677 ldout(pg
->cct
, 1) << "transitioning to Primary" << dendl
;
6678 post_event(MakePrimary());
6680 ldout(pg
->cct
, 1) << "transitioning to Stray" << dendl
;
6681 post_event(MakeStray());
6685 void PG::RecoveryState::Start::exit()
6687 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6688 PG
*pg
= context
< RecoveryMachine
>().pg
;
6689 utime_t dur
= ceph_clock_now() - enter_time
;
6690 pg
->osd
->recoverystate_perf
->tinc(rs_start_latency
, dur
);
6693 /*---------Primary--------*/
6694 PG::RecoveryState::Primary::Primary(my_context ctx
)
6696 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary")
6698 context
< RecoveryMachine
>().log_enter(state_name
);
6699 PG
*pg
= context
< RecoveryMachine
>().pg
;
6700 assert(pg
->want_acting
.empty());
6702 // set CREATING bit until we have peered for the first time.
6703 if (pg
->info
.history
.last_epoch_started
== 0) {
6704 pg
->state_set(PG_STATE_CREATING
);
6705 // use the history timestamp, which ultimately comes from the
6706 // monitor in the create case.
6707 utime_t t
= pg
->info
.history
.last_scrub_stamp
;
6708 pg
->info
.stats
.last_fresh
= t
;
6709 pg
->info
.stats
.last_active
= t
;
6710 pg
->info
.stats
.last_change
= t
;
6711 pg
->info
.stats
.last_peered
= t
;
6712 pg
->info
.stats
.last_clean
= t
;
6713 pg
->info
.stats
.last_unstale
= t
;
6714 pg
->info
.stats
.last_undegraded
= t
;
6715 pg
->info
.stats
.last_fullsized
= t
;
6716 pg
->info
.stats
.last_scrub_stamp
= t
;
6717 pg
->info
.stats
.last_deep_scrub_stamp
= t
;
6718 pg
->info
.stats
.last_clean_scrub_stamp
= t
;
6722 boost::statechart::result
PG::RecoveryState::Primary::react(const MNotifyRec
& notevt
)
6724 PG
*pg
= context
< RecoveryMachine
>().pg
;
6725 ldout(pg
->cct
, 7) << "handle_pg_notify from osd." << notevt
.from
<< dendl
;
6726 pg
->proc_replica_info(
6727 notevt
.from
, notevt
.notify
.info
, notevt
.notify
.epoch_sent
);
6728 return discard_event();
6731 boost::statechart::result
PG::RecoveryState::Primary::react(const ActMap
&)
6733 PG
*pg
= context
< RecoveryMachine
>().pg
;
6734 ldout(pg
->cct
, 7) << "handle ActMap primary" << dendl
;
6735 pg
->publish_stats_to_osd();
6737 return discard_event();
6740 void PG::RecoveryState::Primary::exit()
6742 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6743 PG
*pg
= context
< RecoveryMachine
>().pg
;
6744 pg
->want_acting
.clear();
6745 utime_t dur
= ceph_clock_now() - enter_time
;
6746 pg
->osd
->recoverystate_perf
->tinc(rs_primary_latency
, dur
);
6747 pg
->clear_primary_state();
6748 pg
->state_clear(PG_STATE_CREATING
);
6751 /*---------Peering--------*/
6752 PG::RecoveryState::Peering::Peering(my_context ctx
)
6754 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering"),
6755 history_les_bound(false)
6757 context
< RecoveryMachine
>().log_enter(state_name
);
6759 PG
*pg
= context
< RecoveryMachine
>().pg
;
6760 assert(!pg
->is_peered());
6761 assert(!pg
->is_peering());
6762 assert(pg
->is_primary());
6763 pg
->state_set(PG_STATE_PEERING
);
6766 boost::statechart::result
PG::RecoveryState::Peering::react(const AdvMap
& advmap
)
6768 PG
*pg
= context
< RecoveryMachine
>().pg
;
6769 ldout(pg
->cct
, 10) << "Peering advmap" << dendl
;
6770 if (prior_set
.affected_by_map(*(advmap
.osdmap
), pg
)) {
6771 ldout(pg
->cct
, 1) << "Peering, affected_by_map, going to Reset" << dendl
;
6773 return transit
< Reset
>();
6776 pg
->adjust_need_up_thru(advmap
.osdmap
);
6778 return forward_event();
6781 boost::statechart::result
PG::RecoveryState::Peering::react(const QueryState
& q
)
6783 PG
*pg
= context
< RecoveryMachine
>().pg
;
6785 q
.f
->open_object_section("state");
6786 q
.f
->dump_string("name", state_name
);
6787 q
.f
->dump_stream("enter_time") << enter_time
;
6789 q
.f
->open_array_section("past_intervals");
6790 pg
->past_intervals
.dump(q
.f
);
6791 q
.f
->close_section();
6793 q
.f
->open_array_section("probing_osds");
6794 for (set
<pg_shard_t
>::iterator p
= prior_set
.probe
.begin();
6795 p
!= prior_set
.probe
.end();
6797 q
.f
->dump_stream("osd") << *p
;
6798 q
.f
->close_section();
6800 if (prior_set
.pg_down
)
6801 q
.f
->dump_string("blocked", "peering is blocked due to down osds");
6803 q
.f
->open_array_section("down_osds_we_would_probe");
6804 for (set
<int>::iterator p
= prior_set
.down
.begin();
6805 p
!= prior_set
.down
.end();
6807 q
.f
->dump_int("osd", *p
);
6808 q
.f
->close_section();
6810 q
.f
->open_array_section("peering_blocked_by");
6811 for (map
<int,epoch_t
>::iterator p
= prior_set
.blocked_by
.begin();
6812 p
!= prior_set
.blocked_by
.end();
6814 q
.f
->open_object_section("osd");
6815 q
.f
->dump_int("osd", p
->first
);
6816 q
.f
->dump_int("current_lost_at", p
->second
);
6817 q
.f
->dump_string("comment", "starting or marking this osd lost may let us proceed");
6818 q
.f
->close_section();
6820 q
.f
->close_section();
6822 if (history_les_bound
) {
6823 q
.f
->open_array_section("peering_blocked_by_detail");
6824 q
.f
->open_object_section("item");
6825 q
.f
->dump_string("detail","peering_blocked_by_history_les_bound");
6826 q
.f
->close_section();
6827 q
.f
->close_section();
6830 q
.f
->close_section();
6831 return forward_event();
6834 void PG::RecoveryState::Peering::exit()
6836 PG
*pg
= context
< RecoveryMachine
>().pg
;
6837 ldout(pg
->cct
, 10) << "Leaving Peering" << dendl
;
6838 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6839 pg
->state_clear(PG_STATE_PEERING
);
6840 pg
->clear_probe_targets();
6842 utime_t dur
= ceph_clock_now() - enter_time
;
6843 pg
->osd
->recoverystate_perf
->tinc(rs_peering_latency
, dur
);
6847 /*------Backfilling-------*/
6848 PG::RecoveryState::Backfilling::Backfilling(my_context ctx
)
6850 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Backfilling")
6852 context
< RecoveryMachine
>().log_enter(state_name
);
6853 PG
*pg
= context
< RecoveryMachine
>().pg
;
6854 pg
->backfill_reserved
= true;
6855 pg
->queue_recovery();
6856 pg
->state_clear(PG_STATE_BACKFILL_TOOFULL
);
6857 pg
->state_clear(PG_STATE_BACKFILL_WAIT
);
6858 pg
->state_set(PG_STATE_BACKFILLING
);
6859 pg
->publish_stats_to_osd();
6862 boost::statechart::result
6863 PG::RecoveryState::Backfilling::react(const DeferBackfill
&c
)
6865 PG
*pg
= context
< RecoveryMachine
>().pg
;
6866 ldout(pg
->cct
, 10) << "defer backfill, retry delay " << c
.delay
<< dendl
;
6867 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
6869 pg
->state_set(PG_STATE_BACKFILL_WAIT
);
6870 pg
->state_clear(PG_STATE_BACKFILLING
);
6872 for (set
<pg_shard_t
>::iterator it
= pg
->backfill_targets
.begin();
6873 it
!= pg
->backfill_targets
.end();
6875 assert(*it
!= pg
->pg_whoami
);
6876 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6877 it
->osd
, pg
->get_osdmap()->get_epoch());
6879 pg
->osd
->send_message_osd_cluster(
6880 new MBackfillReserve(
6881 MBackfillReserve::REJECT
,
6882 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
6883 pg
->get_osdmap()->get_epoch()),
6889 if (!pg
->waiting_on_backfill
.empty()) {
6890 pg
->waiting_on_backfill
.clear();
6891 pg
->finish_recovery_op(hobject_t::get_max());
6894 pg
->schedule_backfill_retry(c
.delay
);
6895 return transit
<NotBackfilling
>();
6898 boost::statechart::result
6899 PG::RecoveryState::Backfilling::react(const UnfoundBackfill
&c
)
6901 PG
*pg
= context
< RecoveryMachine
>().pg
;
6902 ldout(pg
->cct
, 10) << "backfill has unfound, can't continue" << dendl
;
6903 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
6905 pg
->state_set(PG_STATE_BACKFILL_UNFOUND
);
6906 pg
->state_clear(PG_STATE_BACKFILLING
);
6908 for (set
<pg_shard_t
>::iterator it
= pg
->backfill_targets
.begin();
6909 it
!= pg
->backfill_targets
.end();
6911 assert(*it
!= pg
->pg_whoami
);
6912 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6913 it
->osd
, pg
->get_osdmap()->get_epoch());
6915 pg
->osd
->send_message_osd_cluster(
6916 new MBackfillReserve(
6917 MBackfillReserve::REJECT
,
6918 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
6919 pg
->get_osdmap()->get_epoch()),
6924 pg
->waiting_on_backfill
.clear();
6926 return transit
<NotBackfilling
>();
6929 boost::statechart::result
6930 PG::RecoveryState::Backfilling::react(const RemoteReservationRejected
&)
6932 PG
*pg
= context
< RecoveryMachine
>().pg
;
6933 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
6934 pg
->state_set(PG_STATE_BACKFILL_TOOFULL
);
6936 for (set
<pg_shard_t
>::iterator it
= pg
->backfill_targets
.begin();
6937 it
!= pg
->backfill_targets
.end();
6939 assert(*it
!= pg
->pg_whoami
);
6940 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6941 it
->osd
, pg
->get_osdmap()->get_epoch());
6943 pg
->osd
->send_message_osd_cluster(
6944 new MBackfillReserve(
6945 MBackfillReserve::REJECT
,
6946 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
6947 pg
->get_osdmap()->get_epoch()),
6952 if (!pg
->waiting_on_backfill
.empty()) {
6953 pg
->waiting_on_backfill
.clear();
6954 pg
->finish_recovery_op(hobject_t::get_max());
6957 pg
->schedule_backfill_retry(pg
->cct
->_conf
->osd_recovery_retry_interval
);
6958 return transit
<NotBackfilling
>();
6961 void PG::RecoveryState::Backfilling::exit()
6963 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6964 PG
*pg
= context
< RecoveryMachine
>().pg
;
6965 pg
->backfill_reserved
= false;
6966 pg
->backfill_reserving
= false;
6967 pg
->state_clear(PG_STATE_BACKFILLING
);
6968 pg
->state_clear(PG_STATE_FORCED_BACKFILL
| PG_STATE_FORCED_RECOVERY
);
6969 utime_t dur
= ceph_clock_now() - enter_time
;
6970 pg
->osd
->recoverystate_perf
->tinc(rs_backfilling_latency
, dur
);
6973 /*--WaitRemoteBackfillReserved--*/
6975 PG::RecoveryState::WaitRemoteBackfillReserved::WaitRemoteBackfillReserved(my_context ctx
)
6977 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitRemoteBackfillReserved"),
6978 backfill_osd_it(context
< Active
>().remote_shards_to_reserve_backfill
.begin())
6980 context
< RecoveryMachine
>().log_enter(state_name
);
6981 PG
*pg
= context
< RecoveryMachine
>().pg
;
6982 pg
->state_set(PG_STATE_BACKFILL_WAIT
);
6983 pg
->publish_stats_to_osd();
6984 post_event(RemoteBackfillReserved());
6987 boost::statechart::result
6988 PG::RecoveryState::WaitRemoteBackfillReserved::react(const RemoteBackfillReserved
&evt
)
6990 PG
*pg
= context
< RecoveryMachine
>().pg
;
6992 if (backfill_osd_it
!= context
< Active
>().remote_shards_to_reserve_backfill
.end()) {
6993 //The primary never backfills itself
6994 assert(*backfill_osd_it
!= pg
->pg_whoami
);
6995 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6996 backfill_osd_it
->osd
, pg
->get_osdmap()->get_epoch());
6998 pg
->osd
->send_message_osd_cluster(
6999 new MBackfillReserve(
7000 MBackfillReserve::REQUEST
,
7001 spg_t(pg
->info
.pgid
.pgid
, backfill_osd_it
->shard
),
7002 pg
->get_osdmap()->get_epoch(),
7003 pg
->get_backfill_priority()),
7008 post_event(AllBackfillsReserved());
7010 return discard_event();
7013 void PG::RecoveryState::WaitRemoteBackfillReserved::exit()
7015 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7016 PG
*pg
= context
< RecoveryMachine
>().pg
;
7017 utime_t dur
= ceph_clock_now() - enter_time
;
7018 pg
->osd
->recoverystate_perf
->tinc(rs_waitremotebackfillreserved_latency
, dur
);
7021 boost::statechart::result
7022 PG::RecoveryState::WaitRemoteBackfillReserved::react(const RemoteReservationRejected
&evt
)
7024 PG
*pg
= context
< RecoveryMachine
>().pg
;
7025 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7027 // Send REJECT to all previously acquired reservations
7028 set
<pg_shard_t
>::const_iterator it
, begin
, end
, next
;
7029 begin
= context
< Active
>().remote_shards_to_reserve_backfill
.begin();
7030 end
= context
< Active
>().remote_shards_to_reserve_backfill
.end();
7031 assert(begin
!= end
);
7032 for (next
= it
= begin
, ++next
; next
!= backfill_osd_it
; ++it
, ++next
) {
7033 //The primary never backfills itself
7034 assert(*it
!= pg
->pg_whoami
);
7035 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
7036 it
->osd
, pg
->get_osdmap()->get_epoch());
7038 pg
->osd
->send_message_osd_cluster(
7039 new MBackfillReserve(
7040 MBackfillReserve::REJECT
,
7041 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
7042 pg
->get_osdmap()->get_epoch()),
7047 pg
->state_clear(PG_STATE_BACKFILL_WAIT
);
7048 pg
->state_set(PG_STATE_BACKFILL_TOOFULL
);
7049 pg
->publish_stats_to_osd();
7051 pg
->schedule_backfill_retry(pg
->cct
->_conf
->osd_recovery_retry_interval
);
7053 return transit
<NotBackfilling
>();
7056 /*--WaitLocalBackfillReserved--*/
7057 PG::RecoveryState::WaitLocalBackfillReserved::WaitLocalBackfillReserved(my_context ctx
)
7059 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitLocalBackfillReserved")
7061 context
< RecoveryMachine
>().log_enter(state_name
);
7062 PG
*pg
= context
< RecoveryMachine
>().pg
;
7063 pg
->state_set(PG_STATE_BACKFILL_WAIT
);
7064 pg
->osd
->local_reserver
.request_reservation(
7066 new QueuePeeringEvt
<LocalBackfillReserved
>(
7067 pg
, pg
->get_osdmap()->get_epoch(),
7068 LocalBackfillReserved()),
7069 pg
->get_backfill_priority(),
7070 new QueuePeeringEvt
<DeferBackfill
>(
7071 pg
, pg
->get_osdmap()->get_epoch(),
7072 DeferBackfill(0.0)));
7073 pg
->publish_stats_to_osd();
7076 void PG::RecoveryState::WaitLocalBackfillReserved::exit()
7078 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7079 PG
*pg
= context
< RecoveryMachine
>().pg
;
7080 utime_t dur
= ceph_clock_now() - enter_time
;
7081 pg
->osd
->recoverystate_perf
->tinc(rs_waitlocalbackfillreserved_latency
, dur
);
7084 /*----NotBackfilling------*/
7085 PG::RecoveryState::NotBackfilling::NotBackfilling(my_context ctx
)
7087 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/NotBackfilling")
7089 context
< RecoveryMachine
>().log_enter(state_name
);
7090 PG
*pg
= context
< RecoveryMachine
>().pg
;
7091 pg
->publish_stats_to_osd();
7094 boost::statechart::result
7095 PG::RecoveryState::NotBackfilling::react(const RemoteBackfillReserved
&evt
)
7097 return discard_event();
7100 boost::statechart::result
7101 PG::RecoveryState::NotBackfilling::react(const RemoteReservationRejected
&evt
)
7103 return discard_event();
7106 void PG::RecoveryState::NotBackfilling::exit()
7108 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7109 PG
*pg
= context
< RecoveryMachine
>().pg
;
7110 pg
->state_clear(PG_STATE_BACKFILL_UNFOUND
);
7111 utime_t dur
= ceph_clock_now() - enter_time
;
7112 pg
->osd
->recoverystate_perf
->tinc(rs_notbackfilling_latency
, dur
);
7115 /*----NotRecovering------*/
7116 PG::RecoveryState::NotRecovering::NotRecovering(my_context ctx
)
7118 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/NotRecovering")
7120 context
< RecoveryMachine
>().log_enter(state_name
);
7121 PG
*pg
= context
< RecoveryMachine
>().pg
;
7122 pg
->publish_stats_to_osd();
7125 void PG::RecoveryState::NotRecovering::exit()
7127 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7128 PG
*pg
= context
< RecoveryMachine
>().pg
;
7129 pg
->state_clear(PG_STATE_RECOVERY_UNFOUND
);
7130 utime_t dur
= ceph_clock_now() - enter_time
;
7131 pg
->osd
->recoverystate_perf
->tinc(rs_notrecovering_latency
, dur
);
7134 /*---RepNotRecovering----*/
7135 PG::RecoveryState::RepNotRecovering::RepNotRecovering(my_context ctx
)
7137 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepNotRecovering")
7139 context
< RecoveryMachine
>().log_enter(state_name
);
7142 boost::statechart::result
7143 PG::RecoveryState::RepNotRecovering::react(const RejectRemoteReservation
&evt
)
7145 PG
*pg
= context
< RecoveryMachine
>().pg
;
7146 pg
->reject_reservation();
7147 post_event(RemoteReservationRejected());
7148 return discard_event();
7151 void PG::RecoveryState::RepNotRecovering::exit()
7153 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7154 PG
*pg
= context
< RecoveryMachine
>().pg
;
7155 utime_t dur
= ceph_clock_now() - enter_time
;
7156 pg
->osd
->recoverystate_perf
->tinc(rs_repnotrecovering_latency
, dur
);
7159 /*---RepWaitRecoveryReserved--*/
7160 PG::RecoveryState::RepWaitRecoveryReserved::RepWaitRecoveryReserved(my_context ctx
)
7162 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepWaitRecoveryReserved")
7164 context
< RecoveryMachine
>().log_enter(state_name
);
7165 PG
*pg
= context
< RecoveryMachine
>().pg
;
7167 pg
->osd
->remote_reserver
.request_reservation(
7169 new QueuePeeringEvt
<RemoteRecoveryReserved
>(
7170 pg
, pg
->get_osdmap()->get_epoch(),
7171 RemoteRecoveryReserved()),
7172 pg
->get_recovery_priority());
7175 boost::statechart::result
7176 PG::RecoveryState::RepWaitRecoveryReserved::react(const RemoteRecoveryReserved
&evt
)
7178 PG
*pg
= context
< RecoveryMachine
>().pg
;
7179 pg
->osd
->send_message_osd_cluster(
7181 new MRecoveryReserve(
7182 MRecoveryReserve::GRANT
,
7183 spg_t(pg
->info
.pgid
.pgid
, pg
->primary
.shard
),
7184 pg
->get_osdmap()->get_epoch()),
7185 pg
->get_osdmap()->get_epoch());
7186 return transit
<RepRecovering
>();
7189 boost::statechart::result
7190 PG::RecoveryState::RepWaitRecoveryReserved::react(
7191 const RemoteReservationCanceled
&evt
)
7193 PG
*pg
= context
< RecoveryMachine
>().pg
;
7194 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
7195 return transit
<RepNotRecovering
>();
7198 void PG::RecoveryState::RepWaitRecoveryReserved::exit()
7200 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7201 PG
*pg
= context
< RecoveryMachine
>().pg
;
7202 utime_t dur
= ceph_clock_now() - enter_time
;
7203 pg
->osd
->recoverystate_perf
->tinc(rs_repwaitrecoveryreserved_latency
, dur
);
7206 /*-RepWaitBackfillReserved*/
7207 PG::RecoveryState::RepWaitBackfillReserved::RepWaitBackfillReserved(my_context ctx
)
7209 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepWaitBackfillReserved")
7211 context
< RecoveryMachine
>().log_enter(state_name
);
7214 boost::statechart::result
7215 PG::RecoveryState::RepNotRecovering::react(const RequestBackfillPrio
&evt
)
7217 PG
*pg
= context
< RecoveryMachine
>().pg
;
7220 if (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
> 0 &&
7221 (rand()%1000 < (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
*1000.0))) {
7222 ldout(pg
->cct
, 10) << "backfill reservation rejected: failure injection"
7224 post_event(RejectRemoteReservation());
7225 } else if (!pg
->cct
->_conf
->osd_debug_skip_full_check_in_backfill_reservation
&&
7226 pg
->osd
->check_backfill_full(ss
)) {
7227 ldout(pg
->cct
, 10) << "backfill reservation rejected: "
7228 << ss
.str() << dendl
;
7229 post_event(RejectRemoteReservation());
7231 pg
->osd
->remote_reserver
.request_reservation(
7233 new QueuePeeringEvt
<RemoteBackfillReserved
>(
7234 pg
, pg
->get_osdmap()->get_epoch(),
7235 RemoteBackfillReserved()), evt
.priority
);
7237 return transit
<RepWaitBackfillReserved
>();
7240 void PG::RecoveryState::RepWaitBackfillReserved::exit()
7242 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7243 PG
*pg
= context
< RecoveryMachine
>().pg
;
7244 utime_t dur
= ceph_clock_now() - enter_time
;
7245 pg
->osd
->recoverystate_perf
->tinc(rs_repwaitbackfillreserved_latency
, dur
);
7248 boost::statechart::result
7249 PG::RecoveryState::RepWaitBackfillReserved::react(const RemoteBackfillReserved
&evt
)
7251 PG
*pg
= context
< RecoveryMachine
>().pg
;
7254 if (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
> 0 &&
7255 (rand()%1000 < (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
*1000.0))) {
7256 ldout(pg
->cct
, 10) << "backfill reservation rejected after reservation: "
7257 << "failure injection" << dendl
;
7258 post_event(RejectRemoteReservation());
7259 return discard_event();
7260 } else if (!pg
->cct
->_conf
->osd_debug_skip_full_check_in_backfill_reservation
&&
7261 pg
->osd
->check_backfill_full(ss
)) {
7262 ldout(pg
->cct
, 10) << "backfill reservation rejected after reservation: "
7263 << ss
.str() << dendl
;
7264 post_event(RejectRemoteReservation());
7265 return discard_event();
7267 pg
->osd
->send_message_osd_cluster(
7269 new MBackfillReserve(
7270 MBackfillReserve::GRANT
,
7271 spg_t(pg
->info
.pgid
.pgid
, pg
->primary
.shard
),
7272 pg
->get_osdmap()->get_epoch()),
7273 pg
->get_osdmap()->get_epoch());
7274 return transit
<RepRecovering
>();
7278 boost::statechart::result
7279 PG::RecoveryState::RepWaitBackfillReserved::react(
7280 const RejectRemoteReservation
&evt
)
7282 PG
*pg
= context
< RecoveryMachine
>().pg
;
7283 pg
->reject_reservation();
7284 post_event(RemoteReservationRejected());
7285 return discard_event();
7288 boost::statechart::result
7289 PG::RecoveryState::RepWaitBackfillReserved::react(
7290 const RemoteReservationRejected
&evt
)
7292 PG
*pg
= context
< RecoveryMachine
>().pg
;
7293 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
7294 return transit
<RepNotRecovering
>();
7297 boost::statechart::result
7298 PG::RecoveryState::RepWaitBackfillReserved::react(
7299 const RemoteReservationCanceled
&evt
)
7301 PG
*pg
= context
< RecoveryMachine
>().pg
;
7302 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
7303 return transit
<RepNotRecovering
>();
7306 /*---RepRecovering-------*/
7307 PG::RecoveryState::RepRecovering::RepRecovering(my_context ctx
)
7309 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepRecovering")
7311 context
< RecoveryMachine
>().log_enter(state_name
);
7314 boost::statechart::result
7315 PG::RecoveryState::RepRecovering::react(const BackfillTooFull
&)
7317 PG
*pg
= context
< RecoveryMachine
>().pg
;
7318 pg
->reject_reservation();
7319 return discard_event();
7322 void PG::RecoveryState::RepRecovering::exit()
7324 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7325 PG
*pg
= context
< RecoveryMachine
>().pg
;
7326 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
7327 utime_t dur
= ceph_clock_now() - enter_time
;
7328 pg
->osd
->recoverystate_perf
->tinc(rs_reprecovering_latency
, dur
);
7331 /*------Activating--------*/
7332 PG::RecoveryState::Activating::Activating(my_context ctx
)
7334 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Activating")
7336 context
< RecoveryMachine
>().log_enter(state_name
);
7339 void PG::RecoveryState::Activating::exit()
7341 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7342 PG
*pg
= context
< RecoveryMachine
>().pg
;
7343 utime_t dur
= ceph_clock_now() - enter_time
;
7344 pg
->osd
->recoverystate_perf
->tinc(rs_activating_latency
, dur
);
7347 PG::RecoveryState::WaitLocalRecoveryReserved::WaitLocalRecoveryReserved(my_context ctx
)
7349 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitLocalRecoveryReserved")
7351 context
< RecoveryMachine
>().log_enter(state_name
);
7352 PG
*pg
= context
< RecoveryMachine
>().pg
;
7354 // Make sure all nodes that part of the recovery aren't full
7355 if (!pg
->cct
->_conf
->osd_debug_skip_full_check_in_recovery
&&
7356 pg
->osd
->check_osdmap_full(pg
->actingbackfill
)) {
7357 post_event(RecoveryTooFull());
7361 pg
->state_clear(PG_STATE_RECOVERY_TOOFULL
);
7362 pg
->state_set(PG_STATE_RECOVERY_WAIT
);
7363 pg
->osd
->local_reserver
.request_reservation(
7365 new QueuePeeringEvt
<LocalRecoveryReserved
>(
7366 pg
, pg
->get_osdmap()->get_epoch(),
7367 LocalRecoveryReserved()),
7368 pg
->get_recovery_priority(),
7369 new QueuePeeringEvt
<DeferRecovery
>(
7370 pg
, pg
->get_osdmap()->get_epoch(),
7371 DeferRecovery(0.0)));
7372 pg
->publish_stats_to_osd();
7375 boost::statechart::result
7376 PG::RecoveryState::WaitLocalRecoveryReserved::react(const RecoveryTooFull
&evt
)
7378 PG
*pg
= context
< RecoveryMachine
>().pg
;
7379 pg
->state_set(PG_STATE_RECOVERY_TOOFULL
);
7380 pg
->schedule_recovery_retry(pg
->cct
->_conf
->osd_recovery_retry_interval
);
7381 return transit
<NotRecovering
>();
7384 void PG::RecoveryState::WaitLocalRecoveryReserved::exit()
7386 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7387 PG
*pg
= context
< RecoveryMachine
>().pg
;
7388 utime_t dur
= ceph_clock_now() - enter_time
;
7389 pg
->osd
->recoverystate_perf
->tinc(rs_waitlocalrecoveryreserved_latency
, dur
);
7392 PG::RecoveryState::WaitRemoteRecoveryReserved::WaitRemoteRecoveryReserved(my_context ctx
)
7394 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitRemoteRecoveryReserved"),
7395 remote_recovery_reservation_it(context
< Active
>().remote_shards_to_reserve_recovery
.begin())
7397 context
< RecoveryMachine
>().log_enter(state_name
);
7398 post_event(RemoteRecoveryReserved());
7401 boost::statechart::result
7402 PG::RecoveryState::WaitRemoteRecoveryReserved::react(const RemoteRecoveryReserved
&evt
) {
7403 PG
*pg
= context
< RecoveryMachine
>().pg
;
7405 if (remote_recovery_reservation_it
!= context
< Active
>().remote_shards_to_reserve_recovery
.end()) {
7406 assert(*remote_recovery_reservation_it
!= pg
->pg_whoami
);
7407 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
7408 remote_recovery_reservation_it
->osd
, pg
->get_osdmap()->get_epoch());
7410 pg
->osd
->send_message_osd_cluster(
7411 new MRecoveryReserve(
7412 MRecoveryReserve::REQUEST
,
7413 spg_t(pg
->info
.pgid
.pgid
, remote_recovery_reservation_it
->shard
),
7414 pg
->get_osdmap()->get_epoch()),
7417 ++remote_recovery_reservation_it
;
7419 post_event(AllRemotesReserved());
7421 return discard_event();
7424 void PG::RecoveryState::WaitRemoteRecoveryReserved::exit()
7426 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7427 PG
*pg
= context
< RecoveryMachine
>().pg
;
7428 utime_t dur
= ceph_clock_now() - enter_time
;
7429 pg
->osd
->recoverystate_perf
->tinc(rs_waitremoterecoveryreserved_latency
, dur
);
7432 PG::RecoveryState::Recovering::Recovering(my_context ctx
)
7434 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Recovering")
7436 context
< RecoveryMachine
>().log_enter(state_name
);
7438 PG
*pg
= context
< RecoveryMachine
>().pg
;
7439 pg
->state_clear(PG_STATE_RECOVERY_WAIT
);
7440 pg
->state_clear(PG_STATE_RECOVERY_TOOFULL
);
7441 pg
->state_set(PG_STATE_RECOVERING
);
7442 assert(!pg
->state_test(PG_STATE_ACTIVATING
));
7443 pg
->publish_stats_to_osd();
7444 pg
->queue_recovery();
7447 void PG::RecoveryState::Recovering::release_reservations(bool cancel
)
7449 PG
*pg
= context
< RecoveryMachine
>().pg
;
7450 assert(cancel
|| !pg
->pg_log
.get_missing().have_missing());
7452 // release remote reservations
7453 for (set
<pg_shard_t
>::const_iterator i
=
7454 context
< Active
>().remote_shards_to_reserve_recovery
.begin();
7455 i
!= context
< Active
>().remote_shards_to_reserve_recovery
.end();
7457 if (*i
== pg
->pg_whoami
) // skip myself
7459 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
7460 i
->osd
, pg
->get_osdmap()->get_epoch());
7462 pg
->osd
->send_message_osd_cluster(
7463 new MRecoveryReserve(
7464 MRecoveryReserve::RELEASE
,
7465 spg_t(pg
->info
.pgid
.pgid
, i
->shard
),
7466 pg
->get_osdmap()->get_epoch()),
7472 boost::statechart::result
7473 PG::RecoveryState::Recovering::react(const AllReplicasRecovered
&evt
)
7475 PG
*pg
= context
< RecoveryMachine
>().pg
;
7476 pg
->state_clear(PG_STATE_RECOVERING
);
7477 pg
->state_clear(PG_STATE_FORCED_RECOVERY
);
7478 release_reservations();
7479 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7480 return transit
<Recovered
>();
7483 boost::statechart::result
7484 PG::RecoveryState::Recovering::react(const RequestBackfill
&evt
)
7486 PG
*pg
= context
< RecoveryMachine
>().pg
;
7487 pg
->state_clear(PG_STATE_RECOVERING
);
7488 pg
->state_clear(PG_STATE_FORCED_RECOVERY
);
7489 release_reservations();
7490 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7491 // XXX: Is this needed?
7492 pg
->publish_stats_to_osd();
7493 return transit
<WaitLocalBackfillReserved
>();
7496 boost::statechart::result
7497 PG::RecoveryState::Recovering::react(const DeferRecovery
&evt
)
7499 PG
*pg
= context
< RecoveryMachine
>().pg
;
7500 if (!pg
->state_test(PG_STATE_RECOVERING
)) {
7501 // we may have finished recovery and have an AllReplicasRecovered
7502 // event queued to move us to the next state.
7503 ldout(pg
->cct
, 10) << "got defer recovery but not recovering" << dendl
;
7504 return discard_event();
7506 ldout(pg
->cct
, 10) << "defer recovery, retry delay " << evt
.delay
<< dendl
;
7507 pg
->state_clear(PG_STATE_RECOVERING
);
7508 pg
->state_set(PG_STATE_RECOVERY_WAIT
);
7509 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7510 release_reservations(true);
7511 pg
->schedule_recovery_retry(evt
.delay
);
7512 return transit
<NotRecovering
>();
7515 boost::statechart::result
7516 PG::RecoveryState::Recovering::react(const UnfoundRecovery
&evt
)
7518 PG
*pg
= context
< RecoveryMachine
>().pg
;
7519 ldout(pg
->cct
, 10) << "recovery has unfound, can't continue" << dendl
;
7520 pg
->state_set(PG_STATE_RECOVERY_UNFOUND
);
7521 pg
->state_clear(PG_STATE_RECOVERING
);
7522 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7523 release_reservations(true);
7524 return transit
<NotRecovering
>();
7527 void PG::RecoveryState::Recovering::exit()
7529 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7530 PG
*pg
= context
< RecoveryMachine
>().pg
;
7531 utime_t dur
= ceph_clock_now() - enter_time
;
7532 pg
->osd
->recoverystate_perf
->tinc(rs_recovering_latency
, dur
);
7535 PG::RecoveryState::Recovered::Recovered(my_context ctx
)
7537 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Recovered")
7539 pg_shard_t auth_log_shard
;
7541 context
< RecoveryMachine
>().log_enter(state_name
);
7543 PG
*pg
= context
< RecoveryMachine
>().pg
;
7545 assert(!pg
->needs_recovery());
7547 // if we finished backfill, all acting are active; recheck if
7548 // DEGRADED | UNDERSIZED is appropriate.
7549 assert(!pg
->actingbackfill
.empty());
7550 if (pg
->get_osdmap()->get_pg_size(pg
->info
.pgid
.pgid
) <=
7551 pg
->actingbackfill
.size()) {
7552 pg
->state_clear(PG_STATE_FORCED_BACKFILL
| PG_STATE_FORCED_RECOVERY
);
7553 pg
->publish_stats_to_osd();
7556 // trim pglog on recovered
7559 // adjust acting set? (e.g. because backfill completed...)
7560 bool history_les_bound
= false;
7561 if (pg
->acting
!= pg
->up
&& !pg
->choose_acting(auth_log_shard
,
7562 true, &history_les_bound
))
7563 assert(pg
->want_acting
.size());
7565 if (context
< Active
>().all_replicas_activated
)
7566 post_event(GoClean());
7569 void PG::RecoveryState::Recovered::exit()
7571 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7572 PG
*pg
= context
< RecoveryMachine
>().pg
;
7573 utime_t dur
= ceph_clock_now() - enter_time
;
7574 pg
->osd
->recoverystate_perf
->tinc(rs_recovered_latency
, dur
);
7577 PG::RecoveryState::Clean::Clean(my_context ctx
)
7579 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Clean")
7581 context
< RecoveryMachine
>().log_enter(state_name
);
7583 PG
*pg
= context
< RecoveryMachine
>().pg
;
7585 if (pg
->info
.last_complete
!= pg
->info
.last_update
) {
7588 pg
->finish_recovery(*context
< RecoveryMachine
>().get_on_safe_context_list());
7590 if (pg
->is_active()) {
7594 pg
->share_pg_info();
7595 pg
->publish_stats_to_osd();
7596 pg
->requeue_ops(pg
->waiting_for_clean_to_primary_repair
);
7599 void PG::RecoveryState::Clean::exit()
7601 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7602 PG
*pg
= context
< RecoveryMachine
>().pg
;
7603 pg
->state_clear(PG_STATE_CLEAN
);
7604 utime_t dur
= ceph_clock_now() - enter_time
;
7605 pg
->osd
->recoverystate_perf
->tinc(rs_clean_latency
, dur
);
7608 template <typename T
>
7609 set
<pg_shard_t
> unique_osd_shard_set(const pg_shard_t
& skip
, const T
&in
)
7611 set
<int> osds_found
;
7612 set
<pg_shard_t
> out
;
7613 for (typename
T::const_iterator i
= in
.begin();
7616 if (*i
!= skip
&& !osds_found
.count(i
->osd
)) {
7617 osds_found
.insert(i
->osd
);
7624 /*---------Active---------*/
7625 PG::RecoveryState::Active::Active(my_context ctx
)
7627 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active"),
7628 remote_shards_to_reserve_recovery(
7629 unique_osd_shard_set(
7630 context
< RecoveryMachine
>().pg
->pg_whoami
,
7631 context
< RecoveryMachine
>().pg
->actingbackfill
)),
7632 remote_shards_to_reserve_backfill(
7633 unique_osd_shard_set(
7634 context
< RecoveryMachine
>().pg
->pg_whoami
,
7635 context
< RecoveryMachine
>().pg
->backfill_targets
)),
7636 all_replicas_activated(false)
7638 context
< RecoveryMachine
>().log_enter(state_name
);
7640 PG
*pg
= context
< RecoveryMachine
>().pg
;
7642 assert(!pg
->backfill_reserving
);
7643 assert(!pg
->backfill_reserved
);
7644 assert(pg
->is_primary());
7645 ldout(pg
->cct
, 10) << "In Active, about to call activate" << dendl
;
7647 context
< RecoveryMachine
>().get_cur_transaction(),
7648 context
< RecoveryMachine
>().get_on_applied_context_list(),
7649 context
< RecoveryMachine
>().get_on_safe_context_list());
7650 pg
->activate(*context
< RecoveryMachine
>().get_cur_transaction(),
7651 pg
->get_osdmap()->get_epoch(),
7652 *context
< RecoveryMachine
>().get_on_safe_context_list(),
7653 *context
< RecoveryMachine
>().get_query_map(),
7654 context
< RecoveryMachine
>().get_info_map(),
7655 context
< RecoveryMachine
>().get_recovery_ctx());
7657 // everyone has to commit/ack before we are truly active
7658 pg
->blocked_by
.clear();
7659 for (set
<pg_shard_t
>::iterator p
= pg
->actingbackfill
.begin();
7660 p
!= pg
->actingbackfill
.end();
7662 if (p
->shard
!= pg
->pg_whoami
.shard
) {
7663 pg
->blocked_by
.insert(p
->shard
);
7666 pg
->publish_stats_to_osd();
7667 ldout(pg
->cct
, 10) << "Activate Finished" << dendl
;
7670 boost::statechart::result
PG::RecoveryState::Active::react(const AdvMap
& advmap
)
7672 PG
*pg
= context
< RecoveryMachine
>().pg
;
7673 ldout(pg
->cct
, 10) << "Active advmap" << dendl
;
7674 if (!pg
->pool
.newly_removed_snaps
.empty()) {
7675 pg
->snap_trimq
.union_of(pg
->pool
.newly_removed_snaps
);
7676 ldout(pg
->cct
, 10) << *pg
<< " snap_trimq now " << pg
->snap_trimq
<< dendl
;
7677 pg
->dirty_info
= true;
7678 pg
->dirty_big_info
= true;
7681 for (size_t i
= 0; i
< pg
->want_acting
.size(); i
++) {
7682 int osd
= pg
->want_acting
[i
];
7683 if (!advmap
.osdmap
->is_up(osd
)) {
7684 pg_shard_t
osd_with_shard(osd
, shard_id_t(i
));
7685 assert(pg
->is_acting(osd_with_shard
) || pg
->is_up(osd_with_shard
));
7689 bool need_publish
= false;
7690 /* Check for changes in pool size (if the acting set changed as a result,
7691 * this does not matter) */
7692 if (advmap
.lastmap
->get_pg_size(pg
->info
.pgid
.pgid
) !=
7693 pg
->get_osdmap()->get_pg_size(pg
->info
.pgid
.pgid
)) {
7694 if (pg
->get_osdmap()->get_pg_size(pg
->info
.pgid
.pgid
) <= pg
->actingset
.size()) {
7695 pg
->state_clear(PG_STATE_UNDERSIZED
);
7697 pg
->state_set(PG_STATE_UNDERSIZED
);
7699 // degraded changes will be detected by call from publish_stats_to_osd()
7700 need_publish
= true;
7703 // if we haven't reported our PG stats in a long time, do so now.
7704 if (pg
->info
.stats
.reported_epoch
+ pg
->cct
->_conf
->osd_pg_stat_report_interval_max
< advmap
.osdmap
->get_epoch()) {
7705 ldout(pg
->cct
, 20) << "reporting stats to osd after " << (advmap
.osdmap
->get_epoch() - pg
->info
.stats
.reported_epoch
)
7706 << " epochs" << dendl
;
7707 need_publish
= true;
7711 pg
->publish_stats_to_osd();
7713 return forward_event();
7716 boost::statechart::result
PG::RecoveryState::Active::react(const ActMap
&)
7718 PG
*pg
= context
< RecoveryMachine
>().pg
;
7719 ldout(pg
->cct
, 10) << "Active: handling ActMap" << dendl
;
7720 assert(pg
->is_primary());
7722 if (pg
->have_unfound()) {
7723 // object may have become unfound
7724 pg
->discover_all_missing(*context
< RecoveryMachine
>().get_query_map());
7727 if (pg
->cct
->_conf
->osd_check_for_log_corruption
)
7728 pg
->check_log_for_corruption(pg
->osd
->store
);
7730 uint64_t unfound
= pg
->missing_loc
.num_unfound();
7732 pg
->all_unfound_are_queried_or_lost(pg
->get_osdmap())) {
7733 if (pg
->cct
->_conf
->osd_auto_mark_unfound_lost
) {
7734 pg
->osd
->clog
->error() << pg
->info
.pgid
.pgid
<< " has " << unfound
7735 << " objects unfound and apparently lost, would automatically "
7736 << "mark these objects lost but this feature is not yet implemented "
7737 << "(osd_auto_mark_unfound_lost)";
7739 pg
->osd
->clog
->error() << pg
->info
.pgid
.pgid
<< " has "
7740 << unfound
<< " objects unfound and apparently lost";
7743 if (pg
->is_active()) {
7744 ldout(pg
->cct
, 10) << "Active: kicking snap trim" << dendl
;
7745 pg
->kick_snap_trim();
7748 if (pg
->is_peered() &&
7750 !pg
->get_osdmap()->test_flag(CEPH_OSDMAP_NOBACKFILL
) &&
7751 (!pg
->get_osdmap()->test_flag(CEPH_OSDMAP_NOREBALANCE
) || pg
->is_degraded())) {
7752 pg
->queue_recovery();
7754 return forward_event();
7757 boost::statechart::result
PG::RecoveryState::Active::react(const MNotifyRec
& notevt
)
7759 PG
*pg
= context
< RecoveryMachine
>().pg
;
7760 assert(pg
->is_primary());
7761 if (pg
->peer_info
.count(notevt
.from
)) {
7762 ldout(pg
->cct
, 10) << "Active: got notify from " << notevt
.from
7763 << ", already have info from that osd, ignoring"
7765 } else if (pg
->peer_purged
.count(notevt
.from
)) {
7766 ldout(pg
->cct
, 10) << "Active: got notify from " << notevt
.from
7767 << ", already purged that peer, ignoring"
7770 ldout(pg
->cct
, 10) << "Active: got notify from " << notevt
.from
7771 << ", calling proc_replica_info and discover_all_missing"
7773 pg
->proc_replica_info(
7774 notevt
.from
, notevt
.notify
.info
, notevt
.notify
.epoch_sent
);
7775 if (pg
->have_unfound()) {
7776 pg
->discover_all_missing(*context
< RecoveryMachine
>().get_query_map());
7779 return discard_event();
7782 boost::statechart::result
PG::RecoveryState::Active::react(const MInfoRec
& infoevt
)
7784 PG
*pg
= context
< RecoveryMachine
>().pg
;
7785 assert(pg
->is_primary());
7787 assert(!pg
->actingbackfill
.empty());
7788 // don't update history (yet) if we are active and primary; the replica
7789 // may be telling us they have activated (and committed) but we can't
7790 // share that until _everyone_ does the same.
7791 if (pg
->is_actingbackfill(infoevt
.from
)) {
7792 ldout(pg
->cct
, 10) << " peer osd." << infoevt
.from
7793 << " activated and committed" << dendl
;
7794 pg
->peer_activated
.insert(infoevt
.from
);
7795 pg
->blocked_by
.erase(infoevt
.from
.shard
);
7796 pg
->publish_stats_to_osd();
7797 if (pg
->peer_activated
.size() == pg
->actingbackfill
.size()) {
7798 pg
->all_activated_and_committed();
7801 return discard_event();
7804 boost::statechart::result
PG::RecoveryState::Active::react(const MLogRec
& logevt
)
7806 PG
*pg
= context
< RecoveryMachine
>().pg
;
7807 ldout(pg
->cct
, 10) << "searching osd." << logevt
.from
7808 << " log for unfound items" << dendl
;
7809 pg
->proc_replica_log(
7810 logevt
.msg
->info
, logevt
.msg
->log
, logevt
.msg
->missing
, logevt
.from
);
7811 bool got_missing
= pg
->search_for_missing(
7812 pg
->peer_info
[logevt
.from
],
7813 pg
->peer_missing
[logevt
.from
],
7815 context
< RecoveryMachine
>().get_recovery_ctx());
7816 // If there are missing AND we are "fully" active then start recovery now
7817 if (got_missing
&& pg
->state_test(PG_STATE_ACTIVE
)) {
7818 post_event(DoRecovery());
7820 return discard_event();
7823 boost::statechart::result
PG::RecoveryState::Active::react(const QueryState
& q
)
7825 PG
*pg
= context
< RecoveryMachine
>().pg
;
7827 q
.f
->open_object_section("state");
7828 q
.f
->dump_string("name", state_name
);
7829 q
.f
->dump_stream("enter_time") << enter_time
;
7832 q
.f
->open_array_section("might_have_unfound");
7833 for (set
<pg_shard_t
>::iterator p
= pg
->might_have_unfound
.begin();
7834 p
!= pg
->might_have_unfound
.end();
7836 q
.f
->open_object_section("osd");
7837 q
.f
->dump_stream("osd") << *p
;
7838 if (pg
->peer_missing
.count(*p
)) {
7839 q
.f
->dump_string("status", "already probed");
7840 } else if (pg
->peer_missing_requested
.count(*p
)) {
7841 q
.f
->dump_string("status", "querying");
7842 } else if (!pg
->get_osdmap()->is_up(p
->osd
)) {
7843 q
.f
->dump_string("status", "osd is down");
7845 q
.f
->dump_string("status", "not queried");
7847 q
.f
->close_section();
7849 q
.f
->close_section();
7852 q
.f
->open_object_section("recovery_progress");
7853 pg
->dump_recovery_info(q
.f
);
7854 q
.f
->close_section();
7858 q
.f
->open_object_section("scrub");
7859 q
.f
->dump_stream("scrubber.epoch_start") << pg
->scrubber
.epoch_start
;
7860 q
.f
->dump_bool("scrubber.active", pg
->scrubber
.active
);
7861 q
.f
->dump_string("scrubber.state", Scrubber::state_string(pg
->scrubber
.state
));
7862 q
.f
->dump_stream("scrubber.start") << pg
->scrubber
.start
;
7863 q
.f
->dump_stream("scrubber.end") << pg
->scrubber
.end
;
7864 q
.f
->dump_stream("scrubber.max_end") << pg
->scrubber
.max_end
;
7865 q
.f
->dump_stream("scrubber.subset_last_update") << pg
->scrubber
.subset_last_update
;
7866 q
.f
->dump_bool("scrubber.deep", pg
->scrubber
.deep
);
7868 q
.f
->open_array_section("scrubber.waiting_on_whom");
7869 for (set
<pg_shard_t
>::iterator p
= pg
->scrubber
.waiting_on_whom
.begin();
7870 p
!= pg
->scrubber
.waiting_on_whom
.end();
7872 q
.f
->dump_stream("shard") << *p
;
7874 q
.f
->close_section();
7876 q
.f
->close_section();
7879 q
.f
->close_section();
7880 return forward_event();
7883 boost::statechart::result
PG::RecoveryState::Active::react(const AllReplicasActivated
&evt
)
7885 PG
*pg
= context
< RecoveryMachine
>().pg
;
7886 all_replicas_activated
= true;
7888 pg
->state_clear(PG_STATE_ACTIVATING
);
7889 pg
->state_clear(PG_STATE_CREATING
);
7890 if (pg
->acting
.size() >= pg
->pool
.info
.min_size
) {
7891 pg
->state_set(PG_STATE_ACTIVE
);
7893 pg
->state_set(PG_STATE_PEERED
);
7896 // info.last_epoch_started is set during activate()
7897 pg
->info
.history
.last_epoch_started
= pg
->info
.last_epoch_started
;
7898 pg
->info
.history
.last_interval_started
= pg
->info
.last_interval_started
;
7899 pg
->dirty_info
= true;
7901 pg
->share_pg_info();
7902 pg
->publish_stats_to_osd();
7907 if (pg
->flushes_in_progress
== 0) {
7908 pg
->requeue_ops(pg
->waiting_for_peered
);
7909 } else if (!pg
->waiting_for_peered
.empty()) {
7910 ldout(pg
->cct
, 10) << __func__
<< " flushes in progress, moving "
7911 << pg
->waiting_for_peered
.size()
7912 << " items to waiting_for_flush"
7914 assert(pg
->waiting_for_flush
.empty());
7915 pg
->waiting_for_flush
.swap(pg
->waiting_for_peered
);
7920 return discard_event();
7923 void PG::RecoveryState::Active::exit()
7925 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7926 PG
*pg
= context
< RecoveryMachine
>().pg
;
7927 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7929 pg
->blocked_by
.clear();
7930 pg
->backfill_reserved
= false;
7931 pg
->backfill_reserving
= false;
7932 pg
->state_clear(PG_STATE_ACTIVATING
);
7933 pg
->state_clear(PG_STATE_DEGRADED
);
7934 pg
->state_clear(PG_STATE_UNDERSIZED
);
7935 pg
->state_clear(PG_STATE_BACKFILL_TOOFULL
);
7936 pg
->state_clear(PG_STATE_BACKFILL_WAIT
);
7937 pg
->state_clear(PG_STATE_RECOVERY_WAIT
);
7938 pg
->state_clear(PG_STATE_RECOVERY_TOOFULL
);
7939 utime_t dur
= ceph_clock_now() - enter_time
;
7940 pg
->osd
->recoverystate_perf
->tinc(rs_active_latency
, dur
);
7944 /*------ReplicaActive-----*/
7945 PG::RecoveryState::ReplicaActive::ReplicaActive(my_context ctx
)
7947 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive")
7949 context
< RecoveryMachine
>().log_enter(state_name
);
7951 PG
*pg
= context
< RecoveryMachine
>().pg
;
7953 context
< RecoveryMachine
>().get_cur_transaction(),
7954 context
< RecoveryMachine
>().get_on_applied_context_list(),
7955 context
< RecoveryMachine
>().get_on_safe_context_list());
7959 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(
7960 const Activate
& actevt
) {
7961 PG
*pg
= context
< RecoveryMachine
>().pg
;
7962 ldout(pg
->cct
, 10) << "In ReplicaActive, about to call activate" << dendl
;
7963 map
<int, map
<spg_t
, pg_query_t
> > query_map
;
7964 pg
->activate(*context
< RecoveryMachine
>().get_cur_transaction(),
7965 actevt
.activation_epoch
,
7966 *context
< RecoveryMachine
>().get_on_safe_context_list(),
7967 query_map
, NULL
, NULL
);
7968 ldout(pg
->cct
, 10) << "Activate Finished" << dendl
;
7969 return discard_event();
7972 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const MInfoRec
& infoevt
)
7974 PG
*pg
= context
< RecoveryMachine
>().pg
;
7975 pg
->proc_primary_info(*context
<RecoveryMachine
>().get_cur_transaction(),
7977 return discard_event();
7980 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const MLogRec
& logevt
)
7982 PG
*pg
= context
< RecoveryMachine
>().pg
;
7983 ldout(pg
->cct
, 10) << "received log from " << logevt
.from
<< dendl
;
7984 ObjectStore::Transaction
* t
= context
<RecoveryMachine
>().get_cur_transaction();
7985 pg
->merge_log(*t
, logevt
.msg
->info
, logevt
.msg
->log
, logevt
.from
);
7986 assert(pg
->pg_log
.get_head() == pg
->info
.last_update
);
7988 return discard_event();
7991 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const ActMap
&)
7993 PG
*pg
= context
< RecoveryMachine
>().pg
;
7994 if (pg
->should_send_notify() && pg
->get_primary().osd
>= 0) {
7995 context
< RecoveryMachine
>().send_notify(
7998 pg
->get_primary().shard
, pg
->pg_whoami
.shard
,
7999 pg
->get_osdmap()->get_epoch(),
8000 pg
->get_osdmap()->get_epoch(),
8002 pg
->past_intervals
);
8005 return discard_event();
8008 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(
8009 const MQuery
& query
)
8011 PG
*pg
= context
< RecoveryMachine
>().pg
;
8012 pg
->fulfill_query(query
, context
<RecoveryMachine
>().get_recovery_ctx());
8013 return discard_event();
8016 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const QueryState
& q
)
8018 q
.f
->open_object_section("state");
8019 q
.f
->dump_string("name", state_name
);
8020 q
.f
->dump_stream("enter_time") << enter_time
;
8021 q
.f
->close_section();
8022 return forward_event();
8025 void PG::RecoveryState::ReplicaActive::exit()
8027 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8028 PG
*pg
= context
< RecoveryMachine
>().pg
;
8029 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
8030 utime_t dur
= ceph_clock_now() - enter_time
;
8031 pg
->osd
->recoverystate_perf
->tinc(rs_replicaactive_latency
, dur
);
8035 PG::RecoveryState::Stray::Stray(my_context ctx
)
8037 NamedState(context
< RecoveryMachine
>().pg
, "Started/Stray")
8039 context
< RecoveryMachine
>().log_enter(state_name
);
8041 PG
*pg
= context
< RecoveryMachine
>().pg
;
8042 assert(!pg
->is_peered());
8043 assert(!pg
->is_peering());
8044 assert(!pg
->is_primary());
8046 context
< RecoveryMachine
>().get_cur_transaction(),
8047 context
< RecoveryMachine
>().get_on_applied_context_list(),
8048 context
< RecoveryMachine
>().get_on_safe_context_list());
8051 boost::statechart::result
PG::RecoveryState::Stray::react(const MLogRec
& logevt
)
8053 PG
*pg
= context
< RecoveryMachine
>().pg
;
8054 MOSDPGLog
*msg
= logevt
.msg
.get();
8055 ldout(pg
->cct
, 10) << "got info+log from osd." << logevt
.from
<< " " << msg
->info
<< " " << msg
->log
<< dendl
;
8057 ObjectStore::Transaction
* t
= context
<RecoveryMachine
>().get_cur_transaction();
8058 if (msg
->info
.last_backfill
== hobject_t()) {
8060 pg
->unreg_next_scrub();
8061 pg
->info
= msg
->info
;
8062 pg
->reg_next_scrub();
8063 pg
->dirty_info
= true;
8064 pg
->dirty_big_info
= true; // maybe.
8066 PGLogEntryHandler rollbacker
{pg
, t
};
8067 pg
->pg_log
.reset_backfill_claim_log(msg
->log
, &rollbacker
);
8069 pg
->pg_log
.reset_backfill();
8071 pg
->merge_log(*t
, msg
->info
, msg
->log
, logevt
.from
);
8074 assert(pg
->pg_log
.get_head() == pg
->info
.last_update
);
8076 post_event(Activate(logevt
.msg
->info
.last_epoch_started
));
8077 return transit
<ReplicaActive
>();
8080 boost::statechart::result
PG::RecoveryState::Stray::react(const MInfoRec
& infoevt
)
8082 PG
*pg
= context
< RecoveryMachine
>().pg
;
8083 ldout(pg
->cct
, 10) << "got info from osd." << infoevt
.from
<< " " << infoevt
.info
<< dendl
;
8085 if (pg
->info
.last_update
> infoevt
.info
.last_update
) {
8086 // rewind divergent log entries
8087 ObjectStore::Transaction
* t
= context
<RecoveryMachine
>().get_cur_transaction();
8088 pg
->rewind_divergent_log(*t
, infoevt
.info
.last_update
);
8089 pg
->info
.stats
= infoevt
.info
.stats
;
8090 pg
->info
.hit_set
= infoevt
.info
.hit_set
;
8093 assert(infoevt
.info
.last_update
== pg
->info
.last_update
);
8094 assert(pg
->pg_log
.get_head() == pg
->info
.last_update
);
8096 post_event(Activate(infoevt
.info
.last_epoch_started
));
8097 return transit
<ReplicaActive
>();
8100 boost::statechart::result
PG::RecoveryState::Stray::react(const MQuery
& query
)
8102 PG
*pg
= context
< RecoveryMachine
>().pg
;
8103 pg
->fulfill_query(query
, context
<RecoveryMachine
>().get_recovery_ctx());
8104 return discard_event();
8107 boost::statechart::result
PG::RecoveryState::Stray::react(const ActMap
&)
8109 PG
*pg
= context
< RecoveryMachine
>().pg
;
8110 if (pg
->should_send_notify() && pg
->get_primary().osd
>= 0) {
8111 context
< RecoveryMachine
>().send_notify(
8114 pg
->get_primary().shard
, pg
->pg_whoami
.shard
,
8115 pg
->get_osdmap()->get_epoch(),
8116 pg
->get_osdmap()->get_epoch(),
8118 pg
->past_intervals
);
8121 return discard_event();
8124 void PG::RecoveryState::Stray::exit()
8126 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8127 PG
*pg
= context
< RecoveryMachine
>().pg
;
8128 utime_t dur
= ceph_clock_now() - enter_time
;
8129 pg
->osd
->recoverystate_perf
->tinc(rs_stray_latency
, dur
);
8132 /*--------GetInfo---------*/
8133 PG::RecoveryState::GetInfo::GetInfo(my_context ctx
)
8135 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/GetInfo")
8137 context
< RecoveryMachine
>().log_enter(state_name
);
8139 PG
*pg
= context
< RecoveryMachine
>().pg
;
8140 pg
->check_past_interval_bounds();
8141 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
8143 assert(pg
->blocked_by
.empty());
8145 prior_set
= pg
->build_prior();
8147 pg
->reset_min_peer_features();
8149 if (prior_set
.pg_down
) {
8150 post_event(IsDown());
8151 } else if (peer_info_requested
.empty()) {
8152 post_event(GotInfo());
8156 void PG::RecoveryState::GetInfo::get_infos()
8158 PG
*pg
= context
< RecoveryMachine
>().pg
;
8159 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
8161 pg
->blocked_by
.clear();
8162 for (set
<pg_shard_t
>::const_iterator it
= prior_set
.probe
.begin();
8163 it
!= prior_set
.probe
.end();
8165 pg_shard_t peer
= *it
;
8166 if (peer
== pg
->pg_whoami
) {
8169 if (pg
->peer_info
.count(peer
)) {
8170 ldout(pg
->cct
, 10) << " have osd." << peer
<< " info " << pg
->peer_info
[peer
] << dendl
;
8173 if (peer_info_requested
.count(peer
)) {
8174 ldout(pg
->cct
, 10) << " already requested info from osd." << peer
<< dendl
;
8175 pg
->blocked_by
.insert(peer
.osd
);
8176 } else if (!pg
->get_osdmap()->is_up(peer
.osd
)) {
8177 ldout(pg
->cct
, 10) << " not querying info from down osd." << peer
<< dendl
;
8179 ldout(pg
->cct
, 10) << " querying info from osd." << peer
<< dendl
;
8180 context
< RecoveryMachine
>().send_query(
8181 peer
, pg_query_t(pg_query_t::INFO
,
8182 it
->shard
, pg
->pg_whoami
.shard
,
8184 pg
->get_osdmap()->get_epoch()));
8185 peer_info_requested
.insert(peer
);
8186 pg
->blocked_by
.insert(peer
.osd
);
8190 pg
->publish_stats_to_osd();
8193 boost::statechart::result
PG::RecoveryState::GetInfo::react(const MNotifyRec
& infoevt
)
8195 PG
*pg
= context
< RecoveryMachine
>().pg
;
8197 set
<pg_shard_t
>::iterator p
= peer_info_requested
.find(infoevt
.from
);
8198 if (p
!= peer_info_requested
.end()) {
8199 peer_info_requested
.erase(p
);
8200 pg
->blocked_by
.erase(infoevt
.from
.osd
);
8203 epoch_t old_start
= pg
->info
.history
.last_epoch_started
;
8204 if (pg
->proc_replica_info(
8205 infoevt
.from
, infoevt
.notify
.info
, infoevt
.notify
.epoch_sent
)) {
8206 // we got something new ...
8207 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
8208 if (old_start
< pg
->info
.history
.last_epoch_started
) {
8209 ldout(pg
->cct
, 10) << " last_epoch_started moved forward, rebuilding prior" << dendl
;
8210 prior_set
= pg
->build_prior();
8212 // filter out any osds that got dropped from the probe set from
8213 // peer_info_requested. this is less expensive than restarting
8214 // peering (which would re-probe everyone).
8215 set
<pg_shard_t
>::iterator p
= peer_info_requested
.begin();
8216 while (p
!= peer_info_requested
.end()) {
8217 if (prior_set
.probe
.count(*p
) == 0) {
8218 ldout(pg
->cct
, 20) << " dropping osd." << *p
<< " from info_requested, no longer in probe set" << dendl
;
8219 peer_info_requested
.erase(p
++);
8226 ldout(pg
->cct
, 20) << "Adding osd: " << infoevt
.from
.osd
<< " peer features: "
8227 << hex
<< infoevt
.features
<< dec
<< dendl
;
8228 pg
->apply_peer_features(infoevt
.features
);
8230 // are we done getting everything?
8231 if (peer_info_requested
.empty() && !prior_set
.pg_down
) {
8232 ldout(pg
->cct
, 20) << "Common peer features: " << hex
<< pg
->get_min_peer_features() << dec
<< dendl
;
8233 ldout(pg
->cct
, 20) << "Common acting features: " << hex
<< pg
->get_min_acting_features() << dec
<< dendl
;
8234 ldout(pg
->cct
, 20) << "Common upacting features: " << hex
<< pg
->get_min_upacting_features() << dec
<< dendl
;
8235 post_event(GotInfo());
8238 return discard_event();
8241 boost::statechart::result
PG::RecoveryState::GetInfo::react(const QueryState
& q
)
8243 PG
*pg
= context
< RecoveryMachine
>().pg
;
8244 q
.f
->open_object_section("state");
8245 q
.f
->dump_string("name", state_name
);
8246 q
.f
->dump_stream("enter_time") << enter_time
;
8248 q
.f
->open_array_section("requested_info_from");
8249 for (set
<pg_shard_t
>::iterator p
= peer_info_requested
.begin();
8250 p
!= peer_info_requested
.end();
8252 q
.f
->open_object_section("osd");
8253 q
.f
->dump_stream("osd") << *p
;
8254 if (pg
->peer_info
.count(*p
)) {
8255 q
.f
->open_object_section("got_info");
8256 pg
->peer_info
[*p
].dump(q
.f
);
8257 q
.f
->close_section();
8259 q
.f
->close_section();
8261 q
.f
->close_section();
8263 q
.f
->close_section();
8264 return forward_event();
8267 void PG::RecoveryState::GetInfo::exit()
8269 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8270 PG
*pg
= context
< RecoveryMachine
>().pg
;
8271 utime_t dur
= ceph_clock_now() - enter_time
;
8272 pg
->osd
->recoverystate_perf
->tinc(rs_getinfo_latency
, dur
);
8273 pg
->blocked_by
.clear();
8274 pg
->publish_stats_to_osd();
8277 /*------GetLog------------*/
8278 PG::RecoveryState::GetLog::GetLog(my_context ctx
)
8281 context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/GetLog"),
8284 context
< RecoveryMachine
>().log_enter(state_name
);
8286 PG
*pg
= context
< RecoveryMachine
>().pg
;
8289 if (!pg
->choose_acting(auth_log_shard
, false,
8290 &context
< Peering
>().history_les_bound
)) {
8291 if (!pg
->want_acting
.empty()) {
8292 post_event(NeedActingChange());
8294 post_event(IsIncomplete());
8300 if (auth_log_shard
== pg
->pg_whoami
) {
8301 post_event(GotLog());
8305 const pg_info_t
& best
= pg
->peer_info
[auth_log_shard
];
8308 if (pg
->info
.last_update
< best
.log_tail
) {
8309 ldout(pg
->cct
, 10) << " not contiguous with osd." << auth_log_shard
<< ", down" << dendl
;
8310 post_event(IsIncomplete());
8314 // how much log to request?
8315 eversion_t request_log_from
= pg
->info
.last_update
;
8316 assert(!pg
->actingbackfill
.empty());
8317 for (set
<pg_shard_t
>::iterator p
= pg
->actingbackfill
.begin();
8318 p
!= pg
->actingbackfill
.end();
8320 if (*p
== pg
->pg_whoami
) continue;
8321 pg_info_t
& ri
= pg
->peer_info
[*p
];
8322 if (ri
.last_update
< pg
->info
.log_tail
&& ri
.last_update
>= best
.log_tail
&&
8323 ri
.last_update
< request_log_from
)
8324 request_log_from
= ri
.last_update
;
8328 ldout(pg
->cct
, 10) << " requesting log from osd." << auth_log_shard
<< dendl
;
8329 context
<RecoveryMachine
>().send_query(
8333 auth_log_shard
.shard
, pg
->pg_whoami
.shard
,
8334 request_log_from
, pg
->info
.history
,
8335 pg
->get_osdmap()->get_epoch()));
8337 assert(pg
->blocked_by
.empty());
8338 pg
->blocked_by
.insert(auth_log_shard
.osd
);
8339 pg
->publish_stats_to_osd();
8342 boost::statechart::result
PG::RecoveryState::GetLog::react(const AdvMap
& advmap
)
8344 PG
*pg
= context
< RecoveryMachine
>().pg
;
8345 // make sure our log source didn't go down. we need to check
8346 // explicitly because it may not be part of the prior set, which
8347 // means the Peering state check won't catch it going down.
8348 if (!advmap
.osdmap
->is_up(auth_log_shard
.osd
)) {
8349 ldout(pg
->cct
, 10) << "GetLog: auth_log_shard osd."
8350 << auth_log_shard
.osd
<< " went down" << dendl
;
8352 return transit
< Reset
>();
8355 // let the Peering state do its checks.
8356 return forward_event();
8359 boost::statechart::result
PG::RecoveryState::GetLog::react(const MLogRec
& logevt
)
8361 PG
*pg
= context
< RecoveryMachine
>().pg
;
8363 if (logevt
.from
!= auth_log_shard
) {
8364 ldout(pg
->cct
, 10) << "GetLog: discarding log from "
8365 << "non-auth_log_shard osd." << logevt
.from
<< dendl
;
8366 return discard_event();
8368 ldout(pg
->cct
, 10) << "GetLog: received master log from osd"
8369 << logevt
.from
<< dendl
;
8371 post_event(GotLog());
8372 return discard_event();
8375 boost::statechart::result
PG::RecoveryState::GetLog::react(const GotLog
&)
8377 PG
*pg
= context
< RecoveryMachine
>().pg
;
8378 ldout(pg
->cct
, 10) << "leaving GetLog" << dendl
;
8380 ldout(pg
->cct
, 10) << "processing master log" << dendl
;
8381 pg
->proc_master_log(*context
<RecoveryMachine
>().get_cur_transaction(),
8382 msg
->info
, msg
->log
, msg
->missing
,
8386 context
< RecoveryMachine
>().get_cur_transaction(),
8387 context
< RecoveryMachine
>().get_on_applied_context_list(),
8388 context
< RecoveryMachine
>().get_on_safe_context_list());
8389 return transit
< GetMissing
>();
8392 boost::statechart::result
PG::RecoveryState::GetLog::react(const QueryState
& q
)
8394 q
.f
->open_object_section("state");
8395 q
.f
->dump_string("name", state_name
);
8396 q
.f
->dump_stream("enter_time") << enter_time
;
8397 q
.f
->dump_stream("auth_log_shard") << auth_log_shard
;
8398 q
.f
->close_section();
8399 return forward_event();
8402 void PG::RecoveryState::GetLog::exit()
8404 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8405 PG
*pg
= context
< RecoveryMachine
>().pg
;
8406 utime_t dur
= ceph_clock_now() - enter_time
;
8407 pg
->osd
->recoverystate_perf
->tinc(rs_getlog_latency
, dur
);
8408 pg
->blocked_by
.clear();
8409 pg
->publish_stats_to_osd();
8412 /*------WaitActingChange--------*/
8413 PG::RecoveryState::WaitActingChange::WaitActingChange(my_context ctx
)
8415 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/WaitActingChange")
8417 context
< RecoveryMachine
>().log_enter(state_name
);
8420 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const AdvMap
& advmap
)
8422 PG
*pg
= context
< RecoveryMachine
>().pg
;
8423 OSDMapRef osdmap
= advmap
.osdmap
;
8425 ldout(pg
->cct
, 10) << "verifying no want_acting " << pg
->want_acting
<< " targets didn't go down" << dendl
;
8426 for (vector
<int>::iterator p
= pg
->want_acting
.begin(); p
!= pg
->want_acting
.end(); ++p
) {
8427 if (!osdmap
->is_up(*p
)) {
8428 ldout(pg
->cct
, 10) << " want_acting target osd." << *p
<< " went down, resetting" << dendl
;
8430 return transit
< Reset
>();
8433 return forward_event();
8436 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const MLogRec
& logevt
)
8438 PG
*pg
= context
< RecoveryMachine
>().pg
;
8439 ldout(pg
->cct
, 10) << "In WaitActingChange, ignoring MLocRec" << dendl
;
8440 return discard_event();
8443 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const MInfoRec
& evt
)
8445 PG
*pg
= context
< RecoveryMachine
>().pg
;
8446 ldout(pg
->cct
, 10) << "In WaitActingChange, ignoring MInfoRec" << dendl
;
8447 return discard_event();
8450 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const MNotifyRec
& evt
)
8452 PG
*pg
= context
< RecoveryMachine
>().pg
;
8453 ldout(pg
->cct
, 10) << "In WaitActingChange, ignoring MNotifyRec" << dendl
;
8454 return discard_event();
8457 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const QueryState
& q
)
8459 q
.f
->open_object_section("state");
8460 q
.f
->dump_string("name", state_name
);
8461 q
.f
->dump_stream("enter_time") << enter_time
;
8462 q
.f
->dump_string("comment", "waiting for pg acting set to change");
8463 q
.f
->close_section();
8464 return forward_event();
8467 void PG::RecoveryState::WaitActingChange::exit()
8469 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8470 PG
*pg
= context
< RecoveryMachine
>().pg
;
8471 utime_t dur
= ceph_clock_now() - enter_time
;
8472 pg
->osd
->recoverystate_perf
->tinc(rs_waitactingchange_latency
, dur
);
8475 /*------Down--------*/
8476 PG::RecoveryState::Down::Down(my_context ctx
)
8478 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/Down")
8480 context
< RecoveryMachine
>().log_enter(state_name
);
8481 PG
*pg
= context
< RecoveryMachine
>().pg
;
8483 pg
->state_clear(PG_STATE_PEERING
);
8484 pg
->state_set(PG_STATE_DOWN
);
8486 auto &prior_set
= context
< Peering
>().prior_set
;
8487 assert(pg
->blocked_by
.empty());
8488 pg
->blocked_by
.insert(prior_set
.down
.begin(), prior_set
.down
.end());
8489 pg
->publish_stats_to_osd();
8492 void PG::RecoveryState::Down::exit()
8494 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8495 PG
*pg
= context
< RecoveryMachine
>().pg
;
8497 pg
->state_clear(PG_STATE_DOWN
);
8498 utime_t dur
= ceph_clock_now() - enter_time
;
8499 pg
->osd
->recoverystate_perf
->tinc(rs_down_latency
, dur
);
8501 pg
->blocked_by
.clear();
8502 pg
->publish_stats_to_osd();
8505 boost::statechart::result
PG::RecoveryState::Down::react(const QueryState
& q
)
8507 q
.f
->open_object_section("state");
8508 q
.f
->dump_string("name", state_name
);
8509 q
.f
->dump_stream("enter_time") << enter_time
;
8510 q
.f
->dump_string("comment",
8511 "not enough up instances of this PG to go active");
8512 q
.f
->close_section();
8513 return forward_event();
8516 /*------Incomplete--------*/
8517 PG::RecoveryState::Incomplete::Incomplete(my_context ctx
)
8519 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/Incomplete")
8521 context
< RecoveryMachine
>().log_enter(state_name
);
8522 PG
*pg
= context
< RecoveryMachine
>().pg
;
8524 pg
->state_clear(PG_STATE_PEERING
);
8525 pg
->state_set(PG_STATE_INCOMPLETE
);
8527 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
8528 assert(pg
->blocked_by
.empty());
8529 pg
->blocked_by
.insert(prior_set
.down
.begin(), prior_set
.down
.end());
8530 pg
->publish_stats_to_osd();
8533 boost::statechart::result
PG::RecoveryState::Incomplete::react(const AdvMap
&advmap
) {
8534 PG
*pg
= context
< RecoveryMachine
>().pg
;
8535 int64_t poolnum
= pg
->info
.pgid
.pool();
8537 // Reset if min_size turn smaller than previous value, pg might now be able to go active
8538 if (!advmap
.osdmap
->have_pg_pool(poolnum
) ||
8539 advmap
.lastmap
->get_pools().find(poolnum
)->second
.min_size
>
8540 advmap
.osdmap
->get_pools().find(poolnum
)->second
.min_size
) {
8542 return transit
< Reset
>();
8545 return forward_event();
8548 boost::statechart::result
PG::RecoveryState::Incomplete::react(const MNotifyRec
& notevt
) {
8549 PG
*pg
= context
< RecoveryMachine
>().pg
;
8550 ldout(pg
->cct
, 7) << "handle_pg_notify from osd." << notevt
.from
<< dendl
;
8551 if (pg
->proc_replica_info(
8552 notevt
.from
, notevt
.notify
.info
, notevt
.notify
.epoch_sent
)) {
8553 // We got something new, try again!
8554 return transit
< GetLog
>();
8556 return discard_event();
8560 boost::statechart::result
PG::RecoveryState::Incomplete::react(
8561 const QueryState
& q
)
8563 q
.f
->open_object_section("state");
8564 q
.f
->dump_string("name", state_name
);
8565 q
.f
->dump_stream("enter_time") << enter_time
;
8566 q
.f
->dump_string("comment", "not enough complete instances of this PG");
8567 q
.f
->close_section();
8568 return forward_event();
8571 void PG::RecoveryState::Incomplete::exit()
8573 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8574 PG
*pg
= context
< RecoveryMachine
>().pg
;
8576 pg
->state_clear(PG_STATE_INCOMPLETE
);
8577 utime_t dur
= ceph_clock_now() - enter_time
;
8578 pg
->osd
->recoverystate_perf
->tinc(rs_incomplete_latency
, dur
);
8580 pg
->blocked_by
.clear();
8581 pg
->publish_stats_to_osd();
8584 /*------GetMissing--------*/
8585 PG::RecoveryState::GetMissing::GetMissing(my_context ctx
)
8587 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/GetMissing")
8589 context
< RecoveryMachine
>().log_enter(state_name
);
8591 PG
*pg
= context
< RecoveryMachine
>().pg
;
8592 assert(!pg
->actingbackfill
.empty());
8594 for (set
<pg_shard_t
>::iterator i
= pg
->actingbackfill
.begin();
8595 i
!= pg
->actingbackfill
.end();
8597 if (*i
== pg
->get_primary()) continue;
8598 const pg_info_t
& pi
= pg
->peer_info
[*i
];
8599 // reset this so to make sure the pg_missing_t is initialized and
8600 // has the correct semantics even if we don't need to get a
8601 // missing set from a shard. This way later additions due to
8602 // lost+unfound delete work properly.
8603 pg
->peer_missing
[*i
].may_include_deletes
= !pg
->perform_deletes_during_peering();
8606 continue; // no pg data, nothing divergent
8608 if (pi
.last_update
< pg
->pg_log
.get_tail()) {
8609 ldout(pg
->cct
, 10) << " osd." << *i
<< " is not contiguous, will restart backfill" << dendl
;
8610 pg
->peer_missing
[*i
].clear();
8613 if (pi
.last_backfill
== hobject_t()) {
8614 ldout(pg
->cct
, 10) << " osd." << *i
<< " will fully backfill; can infer empty missing set" << dendl
;
8615 pg
->peer_missing
[*i
].clear();
8619 if (pi
.last_update
== pi
.last_complete
&& // peer has no missing
8620 pi
.last_update
== pg
->info
.last_update
) { // peer is up to date
8621 // replica has no missing and identical log as us. no need to
8623 // FIXME: we can do better here. if last_update==last_complete we
8624 // can infer the rest!
8625 ldout(pg
->cct
, 10) << " osd." << *i
<< " has no missing, identical log" << dendl
;
8626 pg
->peer_missing
[*i
].clear();
8630 // We pull the log from the peer's last_epoch_started to ensure we
8631 // get enough log to detect divergent updates.
8632 since
.epoch
= pi
.last_epoch_started
;
8633 assert(pi
.last_update
>= pg
->info
.log_tail
); // or else choose_acting() did a bad thing
8634 if (pi
.log_tail
<= since
) {
8635 ldout(pg
->cct
, 10) << " requesting log+missing since " << since
<< " from osd." << *i
<< dendl
;
8636 context
< RecoveryMachine
>().send_query(
8640 i
->shard
, pg
->pg_whoami
.shard
,
8641 since
, pg
->info
.history
,
8642 pg
->get_osdmap()->get_epoch()));
8644 ldout(pg
->cct
, 10) << " requesting fulllog+missing from osd." << *i
8645 << " (want since " << since
<< " < log.tail "
8646 << pi
.log_tail
<< ")" << dendl
;
8647 context
< RecoveryMachine
>().send_query(
8649 pg_query_t::FULLLOG
,
8650 i
->shard
, pg
->pg_whoami
.shard
,
8651 pg
->info
.history
, pg
->get_osdmap()->get_epoch()));
8653 peer_missing_requested
.insert(*i
);
8654 pg
->blocked_by
.insert(i
->osd
);
8657 if (peer_missing_requested
.empty()) {
8658 if (pg
->need_up_thru
) {
8659 ldout(pg
->cct
, 10) << " still need up_thru update before going active"
8661 post_event(NeedUpThru());
8666 post_event(Activate(pg
->get_osdmap()->get_epoch()));
8668 pg
->publish_stats_to_osd();
8672 boost::statechart::result
PG::RecoveryState::GetMissing::react(const MLogRec
& logevt
)
8674 PG
*pg
= context
< RecoveryMachine
>().pg
;
8676 peer_missing_requested
.erase(logevt
.from
);
8677 pg
->proc_replica_log(logevt
.msg
->info
, logevt
.msg
->log
, logevt
.msg
->missing
, logevt
.from
);
8679 if (peer_missing_requested
.empty()) {
8680 if (pg
->need_up_thru
) {
8681 ldout(pg
->cct
, 10) << " still need up_thru update before going active"
8683 post_event(NeedUpThru());
8685 ldout(pg
->cct
, 10) << "Got last missing, don't need missing "
8686 << "posting Activate" << dendl
;
8687 post_event(Activate(pg
->get_osdmap()->get_epoch()));
8690 return discard_event();
8693 boost::statechart::result
PG::RecoveryState::GetMissing::react(const QueryState
& q
)
8695 PG
*pg
= context
< RecoveryMachine
>().pg
;
8696 q
.f
->open_object_section("state");
8697 q
.f
->dump_string("name", state_name
);
8698 q
.f
->dump_stream("enter_time") << enter_time
;
8700 q
.f
->open_array_section("peer_missing_requested");
8701 for (set
<pg_shard_t
>::iterator p
= peer_missing_requested
.begin();
8702 p
!= peer_missing_requested
.end();
8704 q
.f
->open_object_section("osd");
8705 q
.f
->dump_stream("osd") << *p
;
8706 if (pg
->peer_missing
.count(*p
)) {
8707 q
.f
->open_object_section("got_missing");
8708 pg
->peer_missing
[*p
].dump(q
.f
);
8709 q
.f
->close_section();
8711 q
.f
->close_section();
8713 q
.f
->close_section();
8715 q
.f
->close_section();
8716 return forward_event();
8719 void PG::RecoveryState::GetMissing::exit()
8721 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8722 PG
*pg
= context
< RecoveryMachine
>().pg
;
8723 utime_t dur
= ceph_clock_now() - enter_time
;
8724 pg
->osd
->recoverystate_perf
->tinc(rs_getmissing_latency
, dur
);
8725 pg
->blocked_by
.clear();
8726 pg
->publish_stats_to_osd();
8729 /*------WaitUpThru--------*/
8730 PG::RecoveryState::WaitUpThru::WaitUpThru(my_context ctx
)
8732 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/WaitUpThru")
8734 context
< RecoveryMachine
>().log_enter(state_name
);
8737 boost::statechart::result
PG::RecoveryState::WaitUpThru::react(const ActMap
& am
)
8739 PG
*pg
= context
< RecoveryMachine
>().pg
;
8740 if (!pg
->need_up_thru
) {
8741 post_event(Activate(pg
->get_osdmap()->get_epoch()));
8743 return forward_event();
8746 boost::statechart::result
PG::RecoveryState::WaitUpThru::react(const MLogRec
& logevt
)
8748 PG
*pg
= context
< RecoveryMachine
>().pg
;
8749 ldout(pg
->cct
, 10) << "Noting missing from osd." << logevt
.from
<< dendl
;
8750 pg
->peer_missing
[logevt
.from
].claim(logevt
.msg
->missing
);
8751 pg
->peer_info
[logevt
.from
] = logevt
.msg
->info
;
8752 return discard_event();
8755 boost::statechart::result
PG::RecoveryState::WaitUpThru::react(const QueryState
& q
)
8757 q
.f
->open_object_section("state");
8758 q
.f
->dump_string("name", state_name
);
8759 q
.f
->dump_stream("enter_time") << enter_time
;
8760 q
.f
->dump_string("comment", "waiting for osdmap to reflect a new up_thru for this osd");
8761 q
.f
->close_section();
8762 return forward_event();
8765 void PG::RecoveryState::WaitUpThru::exit()
8767 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8768 PG
*pg
= context
< RecoveryMachine
>().pg
;
8769 utime_t dur
= ceph_clock_now() - enter_time
;
8770 pg
->osd
->recoverystate_perf
->tinc(rs_waitupthru_latency
, dur
);
8773 /*----RecoveryState::RecoveryMachine Methods-----*/
8775 #define dout_prefix *_dout << pg->gen_prefix()
8777 void PG::RecoveryState::RecoveryMachine::log_enter(const char *state_name
)
8779 PG
*pg
= context
< RecoveryMachine
>().pg
;
8780 ldout(pg
->cct
, 5) << "enter " << state_name
<< dendl
;
8781 pg
->osd
->pg_recovery_stats
.log_enter(state_name
);
8784 void PG::RecoveryState::RecoveryMachine::log_exit(const char *state_name
, utime_t enter_time
)
8786 utime_t dur
= ceph_clock_now() - enter_time
;
8787 PG
*pg
= context
< RecoveryMachine
>().pg
;
8788 ldout(pg
->cct
, 5) << "exit " << state_name
<< " " << dur
<< " " << event_count
<< " " << event_time
<< dendl
;
8789 pg
->osd
->pg_recovery_stats
.log_exit(state_name
, ceph_clock_now() - enter_time
,
8790 event_count
, event_time
);
8792 event_time
= utime_t();
8796 /*---------------------------------------------------*/
8798 #define dout_prefix (*_dout << (debug_pg ? debug_pg->gen_prefix() : string()) << " PriorSet: ")
8800 void PG::RecoveryState::start_handle(RecoveryCtx
*new_ctx
) {
8805 if (messages_pending_flush
) {
8806 rctx
= RecoveryCtx(*messages_pending_flush
, *new_ctx
);
8810 rctx
->start_time
= ceph_clock_now();
8814 void PG::RecoveryState::begin_block_outgoing() {
8815 assert(!messages_pending_flush
);
8818 messages_pending_flush
= BufferedRecoveryMessages();
8819 rctx
= RecoveryCtx(*messages_pending_flush
, *orig_ctx
);
8822 void PG::RecoveryState::clear_blocked_outgoing() {
8825 messages_pending_flush
= boost::optional
<BufferedRecoveryMessages
>();
8828 void PG::RecoveryState::end_block_outgoing() {
8829 assert(messages_pending_flush
);
8833 rctx
= RecoveryCtx(*orig_ctx
);
8834 rctx
->accept_buffered_messages(*messages_pending_flush
);
8835 messages_pending_flush
= boost::optional
<BufferedRecoveryMessages
>();
8838 void PG::RecoveryState::end_handle() {
8840 utime_t dur
= ceph_clock_now() - rctx
->start_time
;
8841 machine
.event_time
+= dur
;
8844 machine
.event_count
++;
8845 rctx
= boost::optional
<RecoveryCtx
>();
8849 ostream
& operator<<(ostream
& out
, const PG::BackfillInterval
& bi
)
8851 out
<< "BackfillInfo(" << bi
.begin
<< "-" << bi
.end
8852 << " " << bi
.objects
.size() << " objects";
8853 if (!bi
.objects
.empty())
8854 out
<< " " << bi
.objects
;
8859 void intrusive_ptr_add_ref(PG
*pg
) { pg
->get("intptr"); }
8860 void intrusive_ptr_release(PG
*pg
) { pg
->put("intptr"); }
8862 #ifdef PG_DEBUG_REFS
8863 uint64_t get_with_id(PG
*pg
) { return pg
->get_with_id(); }
8864 void put_with_id(PG
*pg
, uint64_t id
) { return pg
->put_with_id(id
); }