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 pi
->build_removed_snaps(newly_removed_snaps
);
241 interval_set
<snapid_t
> intersection
;
242 intersection
.intersection_of(newly_removed_snaps
, cached_removed_snaps
);
243 if (intersection
== cached_removed_snaps
) {
244 newly_removed_snaps
.subtract(cached_removed_snaps
);
245 cached_removed_snaps
.union_of(newly_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
= newly_removed_snaps
;
251 newly_removed_snaps
.clear();
253 snapc
= pi
->get_snap_context();
255 /* 1) map->get_epoch() == cached_epoch + 1 &&
256 * 2) pi->get_snap_epoch() != map->get_epoch()
258 * From the if branch, 1 && 2 must be true. From 2, we know that
259 * this map didn't change the set of removed snaps. From 1, we
260 * know that our cached_removed_snaps matches the previous map.
261 * Thus, from 1 && 2, cached_removed snaps matches the current
262 * set of removed snaps and all we have to do is clear
263 * newly_removed_snaps.
265 newly_removed_snaps
.clear();
267 cached_epoch
= map
->get_epoch();
268 lgeneric_subdout(cct
, osd
, 20)
269 << "PGPool::update cached_removed_snaps "
270 << cached_removed_snaps
271 << " newly_removed_snaps "
272 << newly_removed_snaps
273 << " snapc " << snapc
274 << (updated
? " (updated)":" (no change)")
278 PG::PG(OSDService
*o
, OSDMapRef curmap
,
279 const PGPool
&_pool
, spg_t p
) :
282 osdriver(osd
->store
, coll_t(), OSD::make_snapmapper_oid()),
287 p
.get_split_bits(curmap
->get_pg_num(_pool
.id
)),
290 osdmap_ref(curmap
), last_persisted_osdmap_ref(curmap
), pool(_pool
),
293 _ref_id_lock("PG::_ref_id_lock"), _ref_id(0),
296 trace_endpoint("0.0.0.0", 0, "PG"),
297 dirty_info(false), dirty_big_info(false),
302 pgmeta_oid(p
.make_pgmeta_oid()),
305 curmap
->get_pools().at(p
.pgid
.pool()).ec_pool(),
307 stat_queue_item(this),
309 recovery_queued(false),
310 recovery_ops_active(0),
314 pg_whoami(osd
->whoami
, p
.shard
),
316 last_peering_reset(0),
317 heartbeat_peer_lock("PG::heartbeat_peer_lock"),
318 backfill_reserved(false),
319 backfill_reserving(false),
320 flushes_in_progress(0),
321 pg_stats_publish_lock("PG::pg_stats_publish_lock"),
322 pg_stats_publish_valid(false),
323 osr(osd
->osr_registry
.lookup_or_create(p
, (stringify(p
)))),
324 finish_sync_event(NULL
),
325 backoff_lock("PG::backoff_lock"),
326 scrub_after_recovery(false),
328 recovery_state(this),
330 peer_features(CEPH_FEATURES_SUPPORTED_DEFAULT
),
331 acting_features(CEPH_FEATURES_SUPPORTED_DEFAULT
),
332 upacting_features(CEPH_FEATURES_SUPPORTED_DEFAULT
),
336 osd
->add_pgid(p
, this);
339 std::stringstream ss
;
340 ss
<< "PG " << info
.pgid
;
341 trace_endpoint
.copy_name(ss
.str());
348 pgstate_history
.set_pg_in_destructor();
350 osd
->remove_pgid(info
.pgid
, this);
354 void PG::lock_suspend_timeout(ThreadPool::TPHandle
&handle
)
356 handle
.suspend_tp_timeout();
358 handle
.reset_tp_timeout();
361 void PG::lock(bool no_lockdep
) const
363 _lock
.Lock(no_lockdep
);
364 // if we have unrecorded dirty state with the lock dropped, there is a bug
366 assert(!dirty_big_info
);
368 dout(30) << "lock" << dendl
;
371 std::string
PG::gen_prefix() const
374 OSDMapRef mapref
= osdmap_ref
;
375 if (_lock
.is_locked_by_me()) {
376 out
<< "osd." << osd
->whoami
377 << " pg_epoch: " << (mapref
? mapref
->get_epoch():0)
378 << " " << *this << " ";
380 out
<< "osd." << osd
->whoami
381 << " pg_epoch: " << (mapref
? mapref
->get_epoch():0)
382 << " pg[" << info
.pgid
<< "(unlocked)] ";
387 /********* PG **********/
389 void PG::proc_master_log(
390 ObjectStore::Transaction
& t
, pg_info_t
&oinfo
,
391 pg_log_t
&olog
, pg_missing_t
& omissing
, pg_shard_t from
)
393 dout(10) << "proc_master_log for osd." << from
<< ": "
394 << olog
<< " " << omissing
<< dendl
;
395 assert(!is_peered() && is_primary());
397 // merge log into our own log to build master log. no need to
398 // make any adjustments to their missing map; we are taking their
399 // log to be authoritative (i.e., their entries are by definitely
401 merge_log(t
, oinfo
, olog
, from
);
402 peer_info
[from
] = oinfo
;
403 dout(10) << " peer osd." << from
<< " now " << oinfo
<< " " << omissing
<< dendl
;
404 might_have_unfound
.insert(from
);
406 // See doc/dev/osd_internals/last_epoch_started
407 if (oinfo
.last_epoch_started
> info
.last_epoch_started
) {
408 info
.last_epoch_started
= oinfo
.last_epoch_started
;
411 if (oinfo
.last_interval_started
> info
.last_interval_started
) {
412 info
.last_interval_started
= oinfo
.last_interval_started
;
415 update_history(oinfo
.history
);
416 assert(cct
->_conf
->osd_find_best_info_ignore_history_les
||
417 info
.last_epoch_started
>= info
.history
.last_epoch_started
);
419 peer_missing
[from
].claim(omissing
);
422 void PG::proc_replica_log(
424 const pg_log_t
&olog
,
425 pg_missing_t
& omissing
,
428 dout(10) << "proc_replica_log for osd." << from
<< ": "
429 << oinfo
<< " " << olog
<< " " << omissing
<< dendl
;
431 pg_log
.proc_replica_log(oinfo
, olog
, omissing
, from
);
433 peer_info
[from
] = oinfo
;
434 dout(10) << " peer osd." << from
<< " now " << oinfo
<< " " << omissing
<< dendl
;
435 might_have_unfound
.insert(from
);
437 for (map
<hobject_t
, pg_missing_item
>::const_iterator i
=
438 omissing
.get_items().begin();
439 i
!= omissing
.get_items().end();
441 dout(20) << " after missing " << i
->first
<< " need " << i
->second
.need
442 << " have " << i
->second
.have
<< dendl
;
444 peer_missing
[from
].claim(omissing
);
447 bool PG::proc_replica_info(
448 pg_shard_t from
, const pg_info_t
&oinfo
, epoch_t send_epoch
)
450 map
<pg_shard_t
, pg_info_t
>::iterator p
= peer_info
.find(from
);
451 if (p
!= peer_info
.end() && p
->second
.last_update
== oinfo
.last_update
) {
452 dout(10) << " got dup osd." << from
<< " info " << oinfo
<< ", identical to ours" << dendl
;
456 if (!get_osdmap()->has_been_up_since(from
.osd
, send_epoch
)) {
457 dout(10) << " got info " << oinfo
<< " from down osd." << from
458 << " discarding" << dendl
;
462 dout(10) << " got osd." << from
<< " " << oinfo
<< dendl
;
463 assert(is_primary());
464 peer_info
[from
] = oinfo
;
465 might_have_unfound
.insert(from
);
467 update_history(oinfo
.history
);
470 if (!is_up(from
) && !is_acting(from
)) {
471 dout(10) << " osd." << from
<< " has stray content: " << oinfo
<< dendl
;
472 stray_set
.insert(from
);
478 // was this a new info? if so, update peers!
479 if (p
== peer_info
.end())
480 update_heartbeat_peers();
485 void PG::remove_snap_mapped_object(
486 ObjectStore::Transaction
&t
, const hobject_t
&soid
)
490 ghobject_t(soid
, ghobject_t::NO_GEN
, pg_whoami
.shard
));
491 clear_object_snap_mapping(&t
, soid
);
494 void PG::clear_object_snap_mapping(
495 ObjectStore::Transaction
*t
, const hobject_t
&soid
)
497 OSDriver::OSTransaction
_t(osdriver
.get_transaction(t
));
498 if (soid
.snap
< CEPH_MAXSNAP
) {
499 int r
= snap_mapper
.remove_oid(
502 if (!(r
== 0 || r
== -ENOENT
)) {
503 derr
<< __func__
<< ": remove_oid returned " << cpp_strerror(r
) << dendl
;
509 void PG::update_object_snap_mapping(
510 ObjectStore::Transaction
*t
, const hobject_t
&soid
, const set
<snapid_t
> &snaps
)
512 OSDriver::OSTransaction
_t(osdriver
.get_transaction(t
));
513 assert(soid
.snap
< CEPH_MAXSNAP
);
514 int r
= snap_mapper
.remove_oid(
517 if (!(r
== 0 || r
== -ENOENT
)) {
518 derr
<< __func__
<< ": remove_oid returned " << cpp_strerror(r
) << dendl
;
528 ObjectStore::Transaction
& t
, pg_info_t
&oinfo
, pg_log_t
&olog
, pg_shard_t from
)
530 PGLogEntryHandler rollbacker
{this, &t
};
532 oinfo
, olog
, from
, info
, &rollbacker
, dirty_info
, dirty_big_info
);
535 void PG::rewind_divergent_log(ObjectStore::Transaction
& t
, eversion_t newhead
)
537 PGLogEntryHandler rollbacker
{this, &t
};
538 pg_log
.rewind_divergent_log(
539 newhead
, info
, &rollbacker
, dirty_info
, dirty_big_info
);
543 * Process information from a replica to determine if it could have any
544 * objects that i need.
546 * TODO: if the missing set becomes very large, this could get expensive.
547 * Instead, we probably want to just iterate over our unfound set.
549 bool PG::search_for_missing(
550 const pg_info_t
&oinfo
, const pg_missing_t
&omissing
,
554 uint64_t num_unfound_before
= missing_loc
.num_unfound();
555 bool found_missing
= missing_loc
.add_source_info(
556 from
, oinfo
, omissing
, ctx
->handle
);
557 if (found_missing
&& num_unfound_before
!= missing_loc
.num_unfound())
558 publish_stats_to_osd();
560 (get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, NULL
) &
561 CEPH_FEATURE_OSD_ERASURE_CODES
)) {
562 pg_info_t
tinfo(oinfo
);
563 tinfo
.pgid
.shard
= pg_whoami
.shard
;
564 (*(ctx
->info_map
))[from
.osd
].push_back(
567 from
.shard
, pg_whoami
.shard
,
568 get_osdmap()->get_epoch(),
569 get_osdmap()->get_epoch(),
573 return found_missing
;
576 bool PG::MissingLoc::readable_with_acting(
577 const hobject_t
&hoid
,
578 const set
<pg_shard_t
> &acting
) const {
579 if (!needs_recovery(hoid
))
581 if (is_deleted(hoid
))
583 auto missing_loc_entry
= missing_loc
.find(hoid
);
584 if (missing_loc_entry
== missing_loc
.end())
586 const set
<pg_shard_t
> &locs
= missing_loc_entry
->second
;
587 ldout(pg
->cct
, 10) << __func__
<< ": locs:" << locs
<< dendl
;
588 set
<pg_shard_t
> have_acting
;
589 for (set
<pg_shard_t
>::const_iterator i
= locs
.begin();
592 if (acting
.count(*i
))
593 have_acting
.insert(*i
);
595 return (*is_readable
)(have_acting
);
598 void PG::MissingLoc::add_batch_sources_info(
599 const set
<pg_shard_t
> &sources
, ThreadPool::TPHandle
* handle
)
601 ldout(pg
->cct
, 10) << __func__
<< ": adding sources in batch "
602 << sources
.size() << dendl
;
604 for (map
<hobject_t
, pg_missing_item
>::const_iterator i
= needs_recovery_map
.begin();
605 i
!= needs_recovery_map
.end();
607 if (handle
&& ++loop
>= pg
->cct
->_conf
->osd_loop_before_reset_tphandle
) {
608 handle
->reset_tp_timeout();
611 if (i
->second
.is_delete())
613 missing_loc
[i
->first
].insert(sources
.begin(), sources
.end());
614 missing_loc_sources
.insert(sources
.begin(), sources
.end());
618 bool PG::MissingLoc::add_source_info(
620 const pg_info_t
&oinfo
,
621 const pg_missing_t
&omissing
,
622 ThreadPool::TPHandle
* handle
)
624 bool found_missing
= false;
627 for (map
<hobject_t
,pg_missing_item
>::const_iterator p
= needs_recovery_map
.begin();
628 p
!= needs_recovery_map
.end();
630 const hobject_t
&soid(p
->first
);
631 eversion_t need
= p
->second
.need
;
632 if (handle
&& ++loop
>= pg
->cct
->_conf
->osd_loop_before_reset_tphandle
) {
633 handle
->reset_tp_timeout();
636 if (p
->second
.is_delete()) {
637 ldout(pg
->cct
, 10) << __func__
<< " " << soid
638 << " delete, ignoring source" << dendl
;
639 found_missing
= true;
642 if (oinfo
.last_update
< need
) {
643 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
644 << " also missing on osd." << fromosd
645 << " (last_update " << oinfo
.last_update
646 << " < needed " << need
<< ")" << dendl
;
649 if (!oinfo
.last_backfill
.is_max() &&
650 !oinfo
.last_backfill_bitwise
) {
651 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
652 << " also missing on osd." << fromosd
653 << " (last_backfill " << oinfo
.last_backfill
654 << " but with wrong sort order)"
658 if (p
->first
>= oinfo
.last_backfill
) {
659 // FIXME: this is _probably_ true, although it could conceivably
660 // be in the undefined region! Hmm!
661 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
662 << " also missing on osd." << fromosd
663 << " (past last_backfill " << oinfo
.last_backfill
667 if (oinfo
.last_complete
< need
) {
668 if (omissing
.is_missing(soid
)) {
669 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
670 << " also missing on osd." << fromosd
<< dendl
;
675 ldout(pg
->cct
, 10) << "search_for_missing " << soid
<< " " << need
676 << " is on osd." << fromosd
<< dendl
;
678 missing_loc
[soid
].insert(fromosd
);
679 missing_loc_sources
.insert(fromosd
);
680 found_missing
= true;
683 ldout(pg
->cct
, 20) << "needs_recovery_map missing " << needs_recovery_map
685 return found_missing
;
688 void PG::discover_all_missing(map
<int, map
<spg_t
,pg_query_t
> > &query_map
)
690 auto &missing
= pg_log
.get_missing();
691 uint64_t unfound
= get_num_unfound();
694 dout(10) << __func__
<< " "
695 << missing
.num_missing() << " missing, "
696 << unfound
<< " unfound"
699 std::set
<pg_shard_t
>::const_iterator m
= might_have_unfound
.begin();
700 std::set
<pg_shard_t
>::const_iterator mend
= might_have_unfound
.end();
701 for (; m
!= mend
; ++m
) {
704 if (!get_osdmap()->is_up(peer
.osd
)) {
705 dout(20) << __func__
<< " skipping down osd." << peer
<< dendl
;
709 map
<pg_shard_t
, pg_info_t
>::const_iterator iter
= peer_info
.find(peer
);
710 if (iter
!= peer_info
.end() &&
711 (iter
->second
.is_empty() || iter
->second
.dne())) {
712 // ignore empty peers
716 // If we've requested any of this stuff, the pg_missing_t information
717 // should be on its way.
718 // TODO: coalsce requested_* into a single data structure
719 if (peer_missing
.find(peer
) != peer_missing
.end()) {
720 dout(20) << __func__
<< ": osd." << peer
721 << ": we already have pg_missing_t" << dendl
;
724 if (peer_log_requested
.find(peer
) != peer_log_requested
.end()) {
725 dout(20) << __func__
<< ": osd." << peer
726 << ": in peer_log_requested" << dendl
;
729 if (peer_missing_requested
.find(peer
) != peer_missing_requested
.end()) {
730 dout(20) << __func__
<< ": osd." << peer
731 << ": in peer_missing_requested" << dendl
;
736 dout(10) << __func__
<< ": osd." << peer
<< ": requesting pg_missing_t"
738 peer_missing_requested
.insert(peer
);
739 query_map
[peer
.osd
][spg_t(info
.pgid
.pgid
, peer
.shard
)] =
742 peer
.shard
, pg_whoami
.shard
,
743 info
.history
, get_osdmap()->get_epoch());
747 /******* PG ***********/
748 bool PG::needs_recovery() const
750 assert(is_primary());
752 auto &missing
= pg_log
.get_missing();
754 if (missing
.num_missing()) {
755 dout(10) << __func__
<< " primary has " << missing
.num_missing()
756 << " missing" << dendl
;
760 assert(!actingbackfill
.empty());
761 set
<pg_shard_t
>::const_iterator end
= actingbackfill
.end();
762 set
<pg_shard_t
>::const_iterator a
= actingbackfill
.begin();
763 for (; a
!= end
; ++a
) {
764 if (*a
== get_primary()) continue;
765 pg_shard_t peer
= *a
;
766 map
<pg_shard_t
, pg_missing_t
>::const_iterator pm
= peer_missing
.find(peer
);
767 if (pm
== peer_missing
.end()) {
768 dout(10) << __func__
<< " osd." << peer
<< " doesn't have missing set"
772 if (pm
->second
.num_missing()) {
773 dout(10) << __func__
<< " osd." << peer
<< " has "
774 << pm
->second
.num_missing() << " missing" << dendl
;
779 dout(10) << __func__
<< " is recovered" << dendl
;
783 bool PG::needs_backfill() const
785 assert(is_primary());
787 // We can assume that only possible osds that need backfill
788 // are on the backfill_targets vector nodes.
789 set
<pg_shard_t
>::const_iterator end
= backfill_targets
.end();
790 set
<pg_shard_t
>::const_iterator a
= backfill_targets
.begin();
791 for (; a
!= end
; ++a
) {
792 pg_shard_t peer
= *a
;
793 map
<pg_shard_t
, pg_info_t
>::const_iterator pi
= peer_info
.find(peer
);
794 if (!pi
->second
.last_backfill
.is_max()) {
795 dout(10) << __func__
<< " osd." << peer
<< " has last_backfill " << pi
->second
.last_backfill
<< dendl
;
800 dout(10) << __func__
<< " does not need backfill" << dendl
;
805 void PG::check_past_interval_bounds() const
807 auto rpib
= get_required_past_interval_bounds(
809 osd
->get_superblock().oldest_map
);
810 if (rpib
.first
>= rpib
.second
) {
811 if (!past_intervals
.empty()) {
812 osd
->clog
->error() << info
.pgid
<< " required past_interval bounds are"
813 << " empty [" << rpib
<< ") but past_intervals is not: "
815 derr
<< info
.pgid
<< " required past_interval bounds are"
816 << " empty [" << rpib
<< ") but past_intervals is not: "
817 << past_intervals
<< dendl
;
820 if (past_intervals
.empty()) {
821 osd
->clog
->error() << info
.pgid
<< " required past_interval bounds are"
822 << " not empty [" << rpib
<< ") but past_intervals "
823 << past_intervals
<< " is empty";
824 derr
<< info
.pgid
<< " required past_interval bounds are"
825 << " not empty [" << rpib
<< ") but past_intervals "
826 << past_intervals
<< " is empty" << dendl
;
827 assert(!past_intervals
.empty());
830 auto apib
= past_intervals
.get_bounds();
831 if (apib
.first
> rpib
.first
) {
832 osd
->clog
->error() << info
.pgid
<< " past_intervals [" << apib
833 << ") start interval does not contain the required"
834 << " bound [" << rpib
<< ") start";
835 derr
<< info
.pgid
<< " past_intervals [" << apib
836 << ") start interval does not contain the required"
837 << " bound [" << rpib
<< ") start" << dendl
;
838 assert(0 == "past_interval start interval mismatch");
840 if (apib
.second
!= rpib
.second
) {
841 osd
->clog
->error() << info
.pgid
<< " past_interal bound [" << apib
842 << ") end does not match required [" << rpib
844 derr
<< info
.pgid
<< " past_interal bound [" << apib
845 << ") end does not match required [" << rpib
847 assert(0 == "past_interval end mismatch");
852 bool PG::adjust_need_up_thru(const OSDMapRef osdmap
)
854 epoch_t up_thru
= osdmap
->get_up_thru(osd
->whoami
);
856 up_thru
>= info
.history
.same_interval_since
) {
857 dout(10) << "adjust_need_up_thru now " << up_thru
<< ", need_up_thru now false" << dendl
;
858 need_up_thru
= false;
864 void PG::remove_down_peer_info(const OSDMapRef osdmap
)
866 // Remove any downed osds from peer_info
867 bool removed
= false;
868 map
<pg_shard_t
, pg_info_t
>::iterator p
= peer_info
.begin();
869 while (p
!= peer_info
.end()) {
870 if (!osdmap
->is_up(p
->first
.osd
)) {
871 dout(10) << " dropping down osd." << p
->first
<< " info " << p
->second
<< dendl
;
872 peer_missing
.erase(p
->first
);
873 peer_log_requested
.erase(p
->first
);
874 peer_missing_requested
.erase(p
->first
);
875 peer_info
.erase(p
++);
881 // if we removed anyone, update peers (which include peer_info)
883 update_heartbeat_peers();
884 check_recovery_sources(osdmap
);
888 * Returns true unless there is a non-lost OSD in might_have_unfound.
890 bool PG::all_unfound_are_queried_or_lost(const OSDMapRef osdmap
) const
892 assert(is_primary());
894 set
<pg_shard_t
>::const_iterator peer
= might_have_unfound
.begin();
895 set
<pg_shard_t
>::const_iterator mend
= might_have_unfound
.end();
896 for (; peer
!= mend
; ++peer
) {
897 if (peer_missing
.count(*peer
))
899 map
<pg_shard_t
, pg_info_t
>::const_iterator iter
= peer_info
.find(*peer
);
900 if (iter
!= peer_info
.end() &&
901 (iter
->second
.is_empty() || iter
->second
.dne()))
903 if (!osdmap
->exists(peer
->osd
))
905 const osd_info_t
&osd_info(osdmap
->get_info(peer
->osd
));
906 if (osd_info
.lost_at
<= osd_info
.up_from
) {
907 // If there is even one OSD in might_have_unfound that isn't lost, we
908 // still might retrieve our unfound.
912 dout(10) << "all_unfound_are_queried_or_lost all of might_have_unfound " << might_have_unfound
913 << " have been queried or are marked lost" << dendl
;
917 PastIntervals::PriorSet
PG::build_prior()
921 for (map
<pg_shard_t
,pg_info_t
>::iterator it
= peer_info
.begin();
922 it
!= peer_info
.end();
924 assert(info
.history
.last_epoch_started
>= it
->second
.history
.last_epoch_started
);
928 const OSDMap
&osdmap
= *get_osdmap();
929 PastIntervals::PriorSet prior
= past_intervals
.get_prior_set(
931 info
.history
.last_epoch_started
,
932 get_pgbackend()->get_is_recoverable_predicate(),
933 [&](epoch_t start
, int osd
, epoch_t
*lost_at
) {
934 const osd_info_t
*pinfo
= 0;
935 if (osdmap
.exists(osd
)) {
936 pinfo
= &osdmap
.get_info(osd
);
938 *lost_at
= pinfo
->lost_at
;
941 if (osdmap
.is_up(osd
)) {
942 return PastIntervals::UP
;
944 return PastIntervals::DNE
;
945 } else if (pinfo
->lost_at
> start
) {
946 return PastIntervals::LOST
;
948 return PastIntervals::DOWN
;
956 state_set(PG_STATE_DOWN
);
959 if (get_osdmap()->get_up_thru(osd
->whoami
) < info
.history
.same_interval_since
) {
960 dout(10) << "up_thru " << get_osdmap()->get_up_thru(osd
->whoami
)
961 << " < same_since " << info
.history
.same_interval_since
962 << ", must notify monitor" << dendl
;
965 dout(10) << "up_thru " << get_osdmap()->get_up_thru(osd
->whoami
)
966 << " >= same_since " << info
.history
.same_interval_since
967 << ", all is well" << dendl
;
968 need_up_thru
= false;
970 set_probe_targets(prior
.probe
);
974 void PG::clear_primary_state()
976 dout(10) << "clear_primary_state" << dendl
;
978 // clear peering state
980 peer_log_requested
.clear();
981 peer_missing_requested
.clear();
983 peer_missing
.clear();
984 need_up_thru
= false;
985 peer_last_complete_ondisk
.clear();
986 peer_activated
.clear();
987 min_last_complete_ondisk
= eversion_t();
988 pg_trim_to
= eversion_t();
989 might_have_unfound
.clear();
990 projected_log
= PGLog::IndexedLog();
992 last_update_ondisk
= eversion_t();
996 finish_sync_event
= 0; // so that _finish_recovery doesn't go off in another thread
1000 release_pg_backoffs();
1002 pg_log
.reset_recovery_pointers();
1004 scrubber
.reserved_peers
.clear();
1005 scrub_after_recovery
= false;
1010 PG::Scrubber::Scrubber()
1011 : reserved(false), reserve_failed(false),
1014 waiting_on(0), shallow_errors(0), deep_errors(0), fixed(0),
1015 must_scrub(false), must_deep_scrub(false), must_repair(false),
1017 num_digest_updates_pending(0),
1023 PG::Scrubber::~Scrubber() {}
1028 * Returns an iterator to the best info in infos sorted by:
1029 * 1) Prefer newer last_update
1030 * 2) Prefer longer tail if it brings another info into contiguity
1031 * 3) Prefer current primary
1033 map
<pg_shard_t
, pg_info_t
>::const_iterator
PG::find_best_info(
1034 const map
<pg_shard_t
, pg_info_t
> &infos
,
1035 bool restrict_to_up_acting
,
1036 bool *history_les_bound
) const
1038 assert(history_les_bound
);
1039 /* See doc/dev/osd_internals/last_epoch_started.rst before attempting
1040 * to make changes to this process. Also, make sure to update it
1041 * when you find bugs! */
1042 eversion_t min_last_update_acceptable
= eversion_t::max();
1043 epoch_t max_last_epoch_started_found
= 0;
1044 for (map
<pg_shard_t
, pg_info_t
>::const_iterator i
= infos
.begin();
1047 if (!cct
->_conf
->osd_find_best_info_ignore_history_les
&&
1048 max_last_epoch_started_found
< i
->second
.history
.last_epoch_started
) {
1049 *history_les_bound
= true;
1050 max_last_epoch_started_found
= i
->second
.history
.last_epoch_started
;
1052 if (!i
->second
.is_incomplete() &&
1053 max_last_epoch_started_found
< i
->second
.last_epoch_started
) {
1054 max_last_epoch_started_found
= i
->second
.last_epoch_started
;
1057 for (map
<pg_shard_t
, pg_info_t
>::const_iterator i
= infos
.begin();
1060 if (max_last_epoch_started_found
<= i
->second
.last_epoch_started
) {
1061 if (min_last_update_acceptable
> i
->second
.last_update
)
1062 min_last_update_acceptable
= i
->second
.last_update
;
1065 if (min_last_update_acceptable
== eversion_t::max())
1068 map
<pg_shard_t
, pg_info_t
>::const_iterator best
= infos
.end();
1069 // find osd with newest last_update (oldest for ec_pool).
1070 // if there are multiples, prefer
1071 // - a longer tail, if it brings another peer into log contiguity
1072 // - the current primary
1073 for (map
<pg_shard_t
, pg_info_t
>::const_iterator p
= infos
.begin();
1076 if (restrict_to_up_acting
&& !is_up(p
->first
) &&
1077 !is_acting(p
->first
))
1079 // Only consider peers with last_update >= min_last_update_acceptable
1080 if (p
->second
.last_update
< min_last_update_acceptable
)
1082 // Disqualify anyone with a too old last_epoch_started
1083 if (p
->second
.last_epoch_started
< max_last_epoch_started_found
)
1085 // Disqualify anyone who is incomplete (not fully backfilled)
1086 if (p
->second
.is_incomplete())
1088 if (best
== infos
.end()) {
1092 // Prefer newer last_update
1093 if (pool
.info
.require_rollback()) {
1094 if (p
->second
.last_update
> best
->second
.last_update
)
1096 if (p
->second
.last_update
< best
->second
.last_update
) {
1101 if (p
->second
.last_update
< best
->second
.last_update
)
1103 if (p
->second
.last_update
> best
->second
.last_update
) {
1109 // Prefer longer tail
1110 if (p
->second
.log_tail
> best
->second
.log_tail
) {
1112 } else if (p
->second
.log_tail
< best
->second
.log_tail
) {
1117 // prefer current primary (usually the caller), all things being equal
1118 if (p
->first
== pg_whoami
) {
1119 dout(10) << "calc_acting prefer osd." << p
->first
1120 << " because it is current primary" << dendl
;
1128 void PG::calc_ec_acting(
1129 map
<pg_shard_t
, pg_info_t
>::const_iterator auth_log_shard
,
1131 const vector
<int> &acting
,
1132 pg_shard_t acting_primary
,
1133 const vector
<int> &up
,
1134 pg_shard_t up_primary
,
1135 const map
<pg_shard_t
, pg_info_t
> &all_info
,
1136 bool restrict_to_up_acting
,
1138 set
<pg_shard_t
> *backfill
,
1139 set
<pg_shard_t
> *acting_backfill
,
1140 pg_shard_t
*want_primary
,
1143 vector
<int> want(size
, CRUSH_ITEM_NONE
);
1144 map
<shard_id_t
, set
<pg_shard_t
> > all_info_by_shard
;
1145 unsigned usable
= 0;
1146 for (map
<pg_shard_t
, pg_info_t
>::const_iterator i
= all_info
.begin();
1147 i
!= all_info
.end();
1149 all_info_by_shard
[i
->first
.shard
].insert(i
->first
);
1151 for (uint8_t i
= 0; i
< want
.size(); ++i
) {
1152 ss
<< "For position " << (unsigned)i
<< ": ";
1153 if (up
.size() > (unsigned)i
&& up
[i
] != CRUSH_ITEM_NONE
&&
1154 !all_info
.find(pg_shard_t(up
[i
], shard_id_t(i
)))->second
.is_incomplete() &&
1155 all_info
.find(pg_shard_t(up
[i
], shard_id_t(i
)))->second
.last_update
>=
1156 auth_log_shard
->second
.log_tail
) {
1157 ss
<< " selecting up[i]: " << pg_shard_t(up
[i
], shard_id_t(i
)) << std::endl
;
1162 if (up
.size() > (unsigned)i
&& up
[i
] != CRUSH_ITEM_NONE
) {
1163 ss
<< " backfilling up[i]: " << pg_shard_t(up
[i
], shard_id_t(i
))
1165 backfill
->insert(pg_shard_t(up
[i
], shard_id_t(i
)));
1168 if (acting
.size() > (unsigned)i
&& acting
[i
] != CRUSH_ITEM_NONE
&&
1169 !all_info
.find(pg_shard_t(acting
[i
], shard_id_t(i
)))->second
.is_incomplete() &&
1170 all_info
.find(pg_shard_t(acting
[i
], shard_id_t(i
)))->second
.last_update
>=
1171 auth_log_shard
->second
.log_tail
) {
1172 ss
<< " selecting acting[i]: " << pg_shard_t(acting
[i
], shard_id_t(i
)) << std::endl
;
1173 want
[i
] = acting
[i
];
1175 } else if (!restrict_to_up_acting
) {
1176 for (set
<pg_shard_t
>::iterator j
= all_info_by_shard
[shard_id_t(i
)].begin();
1177 j
!= all_info_by_shard
[shard_id_t(i
)].end();
1179 assert(j
->shard
== i
);
1180 if (!all_info
.find(*j
)->second
.is_incomplete() &&
1181 all_info
.find(*j
)->second
.last_update
>=
1182 auth_log_shard
->second
.log_tail
) {
1183 ss
<< " selecting stray: " << *j
<< std::endl
;
1189 if (want
[i
] == CRUSH_ITEM_NONE
)
1190 ss
<< " failed to fill position " << (int)i
<< std::endl
;
1194 bool found_primary
= false;
1195 for (uint8_t i
= 0; i
< want
.size(); ++i
) {
1196 if (want
[i
] != CRUSH_ITEM_NONE
) {
1197 acting_backfill
->insert(pg_shard_t(want
[i
], shard_id_t(i
)));
1198 if (!found_primary
) {
1199 *want_primary
= pg_shard_t(want
[i
], shard_id_t(i
));
1200 found_primary
= true;
1204 acting_backfill
->insert(backfill
->begin(), backfill
->end());
1209 * calculate the desired acting set.
1211 * Choose an appropriate acting set. Prefer up[0], unless it is
1212 * incomplete, or another osd has a longer tail that allows us to
1213 * bring other up nodes up to date.
1215 void PG::calc_replicated_acting(
1216 map
<pg_shard_t
, pg_info_t
>::const_iterator auth_log_shard
,
1218 const vector
<int> &acting
,
1219 pg_shard_t acting_primary
,
1220 const vector
<int> &up
,
1221 pg_shard_t up_primary
,
1222 const map
<pg_shard_t
, pg_info_t
> &all_info
,
1223 bool restrict_to_up_acting
,
1225 set
<pg_shard_t
> *backfill
,
1226 set
<pg_shard_t
> *acting_backfill
,
1227 pg_shard_t
*want_primary
,
1230 ss
<< "calc_acting newest update on osd." << auth_log_shard
->first
1231 << " with " << auth_log_shard
->second
1232 << (restrict_to_up_acting
? " restrict_to_up_acting" : "") << std::endl
;
1233 pg_shard_t auth_log_shard_id
= auth_log_shard
->first
;
1236 map
<pg_shard_t
,pg_info_t
>::const_iterator primary
;
1238 !all_info
.find(up_primary
)->second
.is_incomplete() &&
1239 all_info
.find(up_primary
)->second
.last_update
>=
1240 auth_log_shard
->second
.log_tail
) {
1241 ss
<< "up_primary: " << up_primary
<< ") selected as primary" << std::endl
;
1242 primary
= all_info
.find(up_primary
); // prefer up[0], all thing being equal
1244 assert(!auth_log_shard
->second
.is_incomplete());
1245 ss
<< "up[0] needs backfill, osd." << auth_log_shard_id
1246 << " selected as primary instead" << std::endl
;
1247 primary
= auth_log_shard
;
1250 ss
<< "calc_acting primary is osd." << primary
->first
1251 << " with " << primary
->second
<< std::endl
;
1252 *want_primary
= primary
->first
;
1253 want
->push_back(primary
->first
.osd
);
1254 acting_backfill
->insert(primary
->first
);
1255 unsigned usable
= 1;
1257 // select replicas that have log contiguity with primary.
1258 // prefer up, then acting, then any peer_info osds
1259 for (vector
<int>::const_iterator i
= up
.begin();
1262 pg_shard_t up_cand
= pg_shard_t(*i
, shard_id_t::NO_SHARD
);
1263 if (up_cand
== primary
->first
)
1265 const pg_info_t
&cur_info
= all_info
.find(up_cand
)->second
;
1266 if (cur_info
.is_incomplete() ||
1267 cur_info
.last_update
< MIN(
1268 primary
->second
.log_tail
,
1269 auth_log_shard
->second
.log_tail
)) {
1270 /* We include auth_log_shard->second.log_tail because in GetLog,
1271 * we will request logs back to the min last_update over our
1272 * acting_backfill set, which will result in our log being extended
1273 * as far backwards as necessary to pick up any peers which can
1274 * be log recovered by auth_log_shard's log */
1275 ss
<< " shard " << up_cand
<< " (up) backfill " << cur_info
<< std::endl
;
1276 backfill
->insert(up_cand
);
1277 acting_backfill
->insert(up_cand
);
1279 want
->push_back(*i
);
1280 acting_backfill
->insert(up_cand
);
1282 ss
<< " osd." << *i
<< " (up) accepted " << cur_info
<< std::endl
;
1286 // This no longer has backfill OSDs, but they are covered above.
1287 for (vector
<int>::const_iterator i
= acting
.begin();
1290 pg_shard_t
acting_cand(*i
, shard_id_t::NO_SHARD
);
1294 // skip up osds we already considered above
1295 if (acting_cand
== primary
->first
)
1297 vector
<int>::const_iterator up_it
= find(up
.begin(), up
.end(), acting_cand
.osd
);
1298 if (up_it
!= up
.end())
1301 const pg_info_t
&cur_info
= all_info
.find(acting_cand
)->second
;
1302 if (cur_info
.is_incomplete() ||
1303 cur_info
.last_update
< primary
->second
.log_tail
) {
1304 ss
<< " shard " << acting_cand
<< " (stray) REJECTED "
1305 << cur_info
<< std::endl
;
1307 want
->push_back(*i
);
1308 acting_backfill
->insert(acting_cand
);
1309 ss
<< " shard " << acting_cand
<< " (stray) accepted "
1310 << cur_info
<< std::endl
;
1315 if (restrict_to_up_acting
) {
1318 for (map
<pg_shard_t
,pg_info_t
>::const_iterator i
= all_info
.begin();
1319 i
!= all_info
.end();
1324 // skip up osds we already considered above
1325 if (i
->first
== primary
->first
)
1327 vector
<int>::const_iterator up_it
= find(up
.begin(), up
.end(), i
->first
.osd
);
1328 if (up_it
!= up
.end())
1330 vector
<int>::const_iterator acting_it
= find(
1331 acting
.begin(), acting
.end(), i
->first
.osd
);
1332 if (acting_it
!= acting
.end())
1335 if (i
->second
.is_incomplete() ||
1336 i
->second
.last_update
< primary
->second
.log_tail
) {
1337 ss
<< " shard " << i
->first
<< " (stray) REJECTED "
1338 << i
->second
<< std::endl
;
1340 want
->push_back(i
->first
.osd
);
1341 acting_backfill
->insert(i
->first
);
1342 ss
<< " shard " << i
->first
<< " (stray) accepted "
1343 << i
->second
<< std::endl
;
1352 * calculate the desired acting, and request a change with the monitor
1353 * if it differs from the current acting.
1355 * if restrict_to_up_acting=true, we filter out anything that's not in
1356 * up/acting. in order to lift this restriction, we need to
1357 * 1) check whether it's worth switching the acting set any time we get
1358 * a new pg info (not just here, when recovery finishes)
1359 * 2) check whether anything in want_acting went down on each new map
1360 * (and, if so, calculate a new want_acting)
1361 * 3) remove the assertion in PG::RecoveryState::Active::react(const AdvMap)
1364 bool PG::choose_acting(pg_shard_t
&auth_log_shard_id
,
1365 bool restrict_to_up_acting
,
1366 bool *history_les_bound
)
1368 map
<pg_shard_t
, pg_info_t
> all_info(peer_info
.begin(), peer_info
.end());
1369 all_info
[pg_whoami
] = info
;
1371 for (map
<pg_shard_t
, pg_info_t
>::iterator p
= all_info
.begin();
1372 p
!= all_info
.end();
1374 dout(10) << "calc_acting osd." << p
->first
<< " " << p
->second
<< dendl
;
1377 map
<pg_shard_t
, pg_info_t
>::const_iterator auth_log_shard
=
1378 find_best_info(all_info
, restrict_to_up_acting
, history_les_bound
);
1380 if (auth_log_shard
== all_info
.end()) {
1382 dout(10) << "choose_acting no suitable info found (incomplete backfills?),"
1383 << " reverting to up" << dendl
;
1386 osd
->queue_want_pg_temp(info
.pgid
.pgid
, empty
);
1388 dout(10) << "choose_acting failed" << dendl
;
1389 assert(want_acting
.empty());
1394 assert(!auth_log_shard
->second
.is_incomplete());
1395 auth_log_shard_id
= auth_log_shard
->first
;
1397 set
<pg_shard_t
> want_backfill
, want_acting_backfill
;
1399 pg_shard_t want_primary
;
1401 if (!pool
.info
.ec_pool())
1402 calc_replicated_acting(
1404 get_osdmap()->get_pg_size(info
.pgid
.pgid
),
1410 restrict_to_up_acting
,
1413 &want_acting_backfill
,
1419 get_osdmap()->get_pg_size(info
.pgid
.pgid
),
1425 restrict_to_up_acting
,
1428 &want_acting_backfill
,
1431 dout(10) << ss
.str() << dendl
;
1433 unsigned num_want_acting
= 0;
1434 set
<pg_shard_t
> have
;
1435 for (int i
= 0; i
< (int)want
.size(); ++i
) {
1436 if (want
[i
] != CRUSH_ITEM_NONE
) {
1441 pool
.info
.ec_pool() ? shard_id_t(i
) : shard_id_t::NO_SHARD
));
1445 // We go incomplete if below min_size for ec_pools since backfill
1446 // does not currently maintain rollbackability
1447 // Otherwise, we will go "peered", but not "active"
1448 if (num_want_acting
< pool
.info
.min_size
&&
1449 (pool
.info
.ec_pool() ||
1450 !cct
->_conf
->osd_allow_recovery_below_min_size
)) {
1451 want_acting
.clear();
1452 dout(10) << "choose_acting failed, below min size" << dendl
;
1456 /* Check whether we have enough acting shards to later perform recovery */
1457 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable_predicate(
1458 get_pgbackend()->get_is_recoverable_predicate());
1459 if (!(*recoverable_predicate
)(have
)) {
1460 want_acting
.clear();
1461 dout(10) << "choose_acting failed, not recoverable" << dendl
;
1465 if (want
!= acting
) {
1466 dout(10) << "choose_acting want " << want
<< " != acting " << acting
1467 << ", requesting pg_temp change" << dendl
;
1470 if (want_acting
== up
) {
1471 // There can't be any pending backfill if
1472 // want is the same as crush map up OSDs.
1473 assert(want_backfill
.empty());
1475 osd
->queue_want_pg_temp(info
.pgid
.pgid
, empty
);
1477 osd
->queue_want_pg_temp(info
.pgid
.pgid
, want
);
1480 want_acting
.clear();
1481 actingbackfill
= want_acting_backfill
;
1482 dout(10) << "actingbackfill is " << actingbackfill
<< dendl
;
1483 assert(backfill_targets
.empty() || backfill_targets
== want_backfill
);
1484 if (backfill_targets
.empty()) {
1485 // Caller is GetInfo
1486 backfill_targets
= want_backfill
;
1488 // Will not change if already set because up would have had to change
1489 // Verify that nothing in backfill is in stray_set
1490 for (set
<pg_shard_t
>::iterator i
= want_backfill
.begin();
1491 i
!= want_backfill
.end();
1493 assert(stray_set
.find(*i
) == stray_set
.end());
1495 dout(10) << "choose_acting want " << want
<< " (== acting) backfill_targets "
1496 << want_backfill
<< dendl
;
1500 /* Build the might_have_unfound set.
1502 * This is used by the primary OSD during recovery.
1504 * This set tracks the OSDs which might have unfound objects that the primary
1505 * OSD needs. As we receive pg_missing_t from each OSD in might_have_unfound, we
1506 * will remove the OSD from the set.
1508 void PG::build_might_have_unfound()
1510 assert(might_have_unfound
.empty());
1511 assert(is_primary());
1513 dout(10) << __func__
<< dendl
;
1515 check_past_interval_bounds();
1517 might_have_unfound
= past_intervals
.get_might_have_unfound(
1519 pool
.info
.ec_pool());
1521 // include any (stray) peers
1522 for (map
<pg_shard_t
, pg_info_t
>::iterator p
= peer_info
.begin();
1523 p
!= peer_info
.end();
1525 might_have_unfound
.insert(p
->first
);
1527 dout(15) << __func__
<< ": built " << might_have_unfound
<< dendl
;
1530 struct C_PG_ActivateCommitted
: public Context
{
1533 epoch_t activation_epoch
;
1534 C_PG_ActivateCommitted(PG
*p
, epoch_t e
, epoch_t ae
)
1535 : pg(p
), epoch(e
), activation_epoch(ae
) {}
1536 void finish(int r
) override
{
1537 pg
->_activate_committed(epoch
, activation_epoch
);
1541 void PG::activate(ObjectStore::Transaction
& t
,
1542 epoch_t activation_epoch
,
1543 list
<Context
*>& tfin
,
1544 map
<int, map
<spg_t
,pg_query_t
> >& query_map
,
1548 PastIntervals
> > > *activator_map
,
1551 assert(!is_peered());
1552 assert(scrubber
.callbacks
.empty());
1553 assert(callbacks_for_degraded_object
.empty());
1556 state_clear(PG_STATE_DOWN
);
1558 send_notify
= false;
1561 // only update primary last_epoch_started if we will go active
1562 if (acting
.size() >= pool
.info
.min_size
) {
1563 assert(cct
->_conf
->osd_find_best_info_ignore_history_les
||
1564 info
.last_epoch_started
<= activation_epoch
);
1565 info
.last_epoch_started
= activation_epoch
;
1566 info
.last_interval_started
= info
.history
.same_interval_since
;
1568 } else if (is_acting(pg_whoami
)) {
1569 /* update last_epoch_started on acting replica to whatever the primary sent
1570 * unless it's smaller (could happen if we are going peered rather than
1571 * active, see doc/dev/osd_internals/last_epoch_started.rst) */
1572 if (info
.last_epoch_started
< activation_epoch
) {
1573 info
.last_epoch_started
= activation_epoch
;
1574 info
.last_interval_started
= info
.history
.same_interval_since
;
1578 auto &missing
= pg_log
.get_missing();
1581 last_update_ondisk
= info
.last_update
;
1582 min_last_complete_ondisk
= eversion_t(0,0); // we don't know (yet)!
1584 last_update_applied
= info
.last_update
;
1585 last_rollback_info_trimmed_to_applied
= pg_log
.get_can_rollback_to();
1587 need_up_thru
= false;
1589 // write pg info, log
1591 dirty_big_info
= true; // maybe
1593 // find out when we commit
1594 t
.register_on_complete(
1595 new C_PG_ActivateCommitted(
1597 get_osdmap()->get_epoch(),
1600 // initialize snap_trimq
1602 dout(20) << "activate - purged_snaps " << info
.purged_snaps
1603 << " cached_removed_snaps " << pool
.cached_removed_snaps
<< dendl
;
1604 snap_trimq
= pool
.cached_removed_snaps
;
1605 interval_set
<snapid_t
> intersection
;
1606 intersection
.intersection_of(snap_trimq
, info
.purged_snaps
);
1607 if (intersection
== info
.purged_snaps
) {
1608 snap_trimq
.subtract(info
.purged_snaps
);
1610 dout(0) << "warning: info.purged_snaps (" << info
.purged_snaps
1611 << ") is not a subset of pool.cached_removed_snaps ("
1612 << pool
.cached_removed_snaps
<< ")" << dendl
;
1613 snap_trimq
.subtract(intersection
);
1617 // init complete pointer
1618 if (missing
.num_missing() == 0) {
1619 dout(10) << "activate - no missing, moving last_complete " << info
.last_complete
1620 << " -> " << info
.last_update
<< dendl
;
1621 info
.last_complete
= info
.last_update
;
1622 pg_log
.reset_recovery_pointers();
1624 dout(10) << "activate - not complete, " << missing
<< dendl
;
1625 pg_log
.activate_not_complete(info
);
1633 // start up replicas
1635 assert(!actingbackfill
.empty());
1636 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
1637 i
!= actingbackfill
.end();
1639 if (*i
== pg_whoami
) continue;
1640 pg_shard_t peer
= *i
;
1641 assert(peer_info
.count(peer
));
1642 pg_info_t
& pi
= peer_info
[peer
];
1644 dout(10) << "activate peer osd." << peer
<< " " << pi
<< dendl
;
1647 assert(peer_missing
.count(peer
));
1648 pg_missing_t
& pm
= peer_missing
[peer
];
1650 bool needs_past_intervals
= pi
.dne();
1653 * cover case where peer sort order was different and
1654 * last_backfill cannot be interpreted
1656 bool force_restart_backfill
=
1657 !pi
.last_backfill
.is_max() &&
1658 !pi
.last_backfill_bitwise
;
1660 if (pi
.last_update
== info
.last_update
&& !force_restart_backfill
) {
1662 if (!pi
.last_backfill
.is_max())
1663 osd
->clog
->info() << info
.pgid
<< " continuing backfill to osd."
1665 << " from (" << pi
.log_tail
<< "," << pi
.last_update
1666 << "] " << pi
.last_backfill
1667 << " to " << info
.last_update
;
1668 if (!pi
.is_empty() && activator_map
) {
1669 dout(10) << "activate peer osd." << peer
<< " is up to date, queueing in pending_activators" << dendl
;
1670 (*activator_map
)[peer
.osd
].push_back(
1673 peer
.shard
, pg_whoami
.shard
,
1674 get_osdmap()->get_epoch(),
1675 get_osdmap()->get_epoch(),
1679 dout(10) << "activate peer osd." << peer
<< " is up to date, but sending pg_log anyway" << dendl
;
1681 i
->shard
, pg_whoami
.shard
,
1682 get_osdmap()->get_epoch(), info
);
1685 pg_log
.get_tail() > pi
.last_update
||
1686 pi
.last_backfill
== hobject_t() ||
1687 force_restart_backfill
||
1688 (backfill_targets
.count(*i
) && pi
.last_backfill
.is_max())) {
1689 /* ^ This last case covers a situation where a replica is not contiguous
1690 * with the auth_log, but is contiguous with this replica. Reshuffling
1691 * the active set to handle this would be tricky, so instead we just go
1692 * ahead and backfill it anyway. This is probably preferrable in any
1693 * case since the replica in question would have to be significantly
1697 osd
->clog
->debug() << info
.pgid
<< " starting backfill to osd." << peer
1698 << " from (" << pi
.log_tail
<< "," << pi
.last_update
1699 << "] " << pi
.last_backfill
1700 << " to " << info
.last_update
;
1702 pi
.last_update
= info
.last_update
;
1703 pi
.last_complete
= info
.last_update
;
1704 pi
.set_last_backfill(hobject_t());
1705 pi
.last_epoch_started
= info
.last_epoch_started
;
1706 pi
.last_interval_started
= info
.last_interval_started
;
1707 pi
.history
= info
.history
;
1708 pi
.hit_set
= info
.hit_set
;
1709 pi
.stats
.stats
.clear();
1711 // initialize peer with our purged_snaps.
1712 pi
.purged_snaps
= info
.purged_snaps
;
1715 i
->shard
, pg_whoami
.shard
,
1716 get_osdmap()->get_epoch(), pi
);
1718 // send some recent log, so that op dup detection works well.
1719 m
->log
.copy_up_to(pg_log
.get_log(), cct
->_conf
->osd_min_pg_log_entries
);
1720 m
->info
.log_tail
= m
->log
.tail
;
1721 pi
.log_tail
= m
->log
.tail
; // sigh...
1726 assert(pg_log
.get_tail() <= pi
.last_update
);
1728 i
->shard
, pg_whoami
.shard
,
1729 get_osdmap()->get_epoch(), info
);
1730 // send new stuff to append to replicas log
1731 m
->log
.copy_after(pg_log
.get_log(), pi
.last_update
);
1734 // share past_intervals if we are creating the pg on the replica
1735 // based on whether our info for that peer was dne() *before*
1736 // updating pi.history in the backfill block above.
1737 if (m
&& needs_past_intervals
)
1738 m
->past_intervals
= past_intervals
;
1740 // update local version of peer's missing list!
1741 if (m
&& pi
.last_backfill
!= hobject_t()) {
1742 for (list
<pg_log_entry_t
>::iterator p
= m
->log
.log
.begin();
1743 p
!= m
->log
.log
.end();
1745 if (p
->soid
<= pi
.last_backfill
&&
1747 if (perform_deletes_during_peering() && p
->is_delete()) {
1748 pm
.rm(p
->soid
, p
->version
);
1750 pm
.add_next_event(*p
);
1757 dout(10) << "activate peer osd." << peer
<< " sending " << m
->log
<< dendl
;
1758 //m->log.print(cout);
1759 osd
->send_message_osd_cluster(peer
.osd
, m
, get_osdmap()->get_epoch());
1763 pi
.last_update
= info
.last_update
;
1765 // update our missing
1766 if (pm
.num_missing() == 0) {
1767 pi
.last_complete
= pi
.last_update
;
1768 dout(10) << "activate peer osd." << peer
<< " " << pi
<< " uptodate" << dendl
;
1770 dout(10) << "activate peer osd." << peer
<< " " << pi
<< " missing " << pm
<< dendl
;
1774 // Set up missing_loc
1775 set
<pg_shard_t
> complete_shards
;
1776 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
1777 i
!= actingbackfill
.end();
1779 dout(20) << __func__
<< " setting up missing_loc from shard " << *i
<< " " << dendl
;
1780 if (*i
== get_primary()) {
1781 missing_loc
.add_active_missing(missing
);
1782 if (!missing
.have_missing())
1783 complete_shards
.insert(*i
);
1785 auto peer_missing_entry
= peer_missing
.find(*i
);
1786 assert(peer_missing_entry
!= peer_missing
.end());
1787 missing_loc
.add_active_missing(peer_missing_entry
->second
);
1788 if (!peer_missing_entry
->second
.have_missing() &&
1789 peer_info
[*i
].last_backfill
.is_max())
1790 complete_shards
.insert(*i
);
1793 // If necessary, create might_have_unfound to help us find our unfound objects.
1794 // NOTE: It's important that we build might_have_unfound before trimming the
1796 might_have_unfound
.clear();
1797 if (needs_recovery()) {
1798 // If only one shard has missing, we do a trick to add all others as recovery
1799 // source, this is considered safe since the PGLogs have been merged locally,
1800 // and covers vast majority of the use cases, like one OSD/host is down for
1801 // a while for hardware repairing
1802 if (complete_shards
.size() + 1 == actingbackfill
.size()) {
1803 missing_loc
.add_batch_sources_info(complete_shards
, ctx
->handle
);
1805 missing_loc
.add_source_info(pg_whoami
, info
, pg_log
.get_missing(),
1807 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
1808 i
!= actingbackfill
.end();
1810 if (*i
== pg_whoami
) continue;
1811 dout(10) << __func__
<< ": adding " << *i
<< " as a source" << dendl
;
1812 assert(peer_missing
.count(*i
));
1813 assert(peer_info
.count(*i
));
1814 missing_loc
.add_source_info(
1821 for (map
<pg_shard_t
, pg_missing_t
>::iterator i
= peer_missing
.begin();
1822 i
!= peer_missing
.end();
1824 if (is_actingbackfill(i
->first
))
1826 assert(peer_info
.count(i
->first
));
1828 peer_info
[i
->first
],
1834 build_might_have_unfound();
1837 discover_all_missing(query_map
);
1840 // num_objects_degraded if calculated should reflect this too, unless no
1841 // missing and we are about to go clean.
1842 if (get_osdmap()->get_pg_size(info
.pgid
.pgid
) > actingset
.size()) {
1843 state_set(PG_STATE_UNDERSIZED
);
1846 state_set(PG_STATE_ACTIVATING
);
1847 release_pg_backoffs();
1848 projected_last_update
= info
.last_update
;
1850 if (acting
.size() >= pool
.info
.min_size
) {
1851 PGLogEntryHandler handler
{this, &t
};
1852 pg_log
.roll_forward(&handler
);
1856 bool PG::op_has_sufficient_caps(OpRequestRef
& op
)
1858 // only check MOSDOp
1859 if (op
->get_req()->get_type() != CEPH_MSG_OSD_OP
)
1862 const MOSDOp
*req
= static_cast<const MOSDOp
*>(op
->get_req());
1864 Session
*session
= static_cast<Session
*>(req
->get_connection()->get_priv());
1866 dout(0) << "op_has_sufficient_caps: no session for op " << *req
<< dendl
;
1869 OSDCap
& caps
= session
->caps
;
1872 const string
&key
= req
->get_hobj().get_key().empty() ?
1873 req
->get_oid().name
:
1874 req
->get_hobj().get_key();
1876 bool cap
= caps
.is_capable(pool
.name
, req
->get_hobj().nspace
,
1878 op
->need_read_cap(),
1879 op
->need_write_cap(),
1882 dout(20) << "op_has_sufficient_caps "
1883 << "session=" << session
1884 << " pool=" << pool
.id
<< " (" << pool
.name
1885 << " " << req
->get_hobj().nspace
1886 << ") owner=" << pool
.auid
1887 << " need_read_cap=" << op
->need_read_cap()
1888 << " need_write_cap=" << op
->need_write_cap()
1889 << " classes=" << op
->classes()
1890 << " -> " << (cap
? "yes" : "NO")
1895 void PG::_activate_committed(epoch_t epoch
, epoch_t activation_epoch
)
1898 if (pg_has_reset_since(epoch
)) {
1899 dout(10) << "_activate_committed " << epoch
1900 << ", that was an old interval" << dendl
;
1901 } else if (is_primary()) {
1902 peer_activated
.insert(pg_whoami
);
1903 dout(10) << "_activate_committed " << epoch
1904 << " peer_activated now " << peer_activated
1905 << " last_interval_started " << info
.history
.last_interval_started
1906 << " last_epoch_started " << info
.history
.last_epoch_started
1907 << " same_interval_since " << info
.history
.same_interval_since
<< dendl
;
1908 assert(!actingbackfill
.empty());
1909 if (peer_activated
.size() == actingbackfill
.size())
1910 all_activated_and_committed();
1912 dout(10) << "_activate_committed " << epoch
<< " telling primary" << dendl
;
1913 MOSDPGInfo
*m
= new MOSDPGInfo(epoch
);
1914 pg_notify_t i
= pg_notify_t(
1915 get_primary().shard
, pg_whoami
.shard
,
1916 get_osdmap()->get_epoch(),
1917 get_osdmap()->get_epoch(),
1920 i
.info
.history
.last_epoch_started
= activation_epoch
;
1921 i
.info
.history
.last_interval_started
= i
.info
.history
.same_interval_since
;
1922 if (acting
.size() >= pool
.info
.min_size
) {
1923 state_set(PG_STATE_ACTIVE
);
1925 state_set(PG_STATE_PEERED
);
1928 m
->pg_list
.push_back(make_pair(i
, PastIntervals()));
1929 osd
->send_message_osd_cluster(get_primary().osd
, m
, get_osdmap()->get_epoch());
1932 if (flushes_in_progress
== 0) {
1933 requeue_ops(waiting_for_peered
);
1934 } else if (!waiting_for_peered
.empty()) {
1935 dout(10) << __func__
<< " flushes in progress, moving "
1936 << waiting_for_peered
.size() << " items to waiting_for_flush"
1938 assert(waiting_for_flush
.empty());
1939 waiting_for_flush
.swap(waiting_for_peered
);
1943 assert(!dirty_info
);
1949 * update info.history.last_epoch_started ONLY after we and all
1950 * replicas have activated AND committed the activate transaction
1951 * (i.e. the peering results are stable on disk).
1953 void PG::all_activated_and_committed()
1955 dout(10) << "all_activated_and_committed" << dendl
;
1956 assert(is_primary());
1957 assert(peer_activated
.size() == actingbackfill
.size());
1958 assert(!actingbackfill
.empty());
1959 assert(blocked_by
.empty());
1962 _update_calc_stats();
1963 if (info
.stats
.stats
.sum
.num_objects_degraded
) {
1964 state_set(PG_STATE_DEGRADED
);
1966 state_clear(PG_STATE_DEGRADED
);
1969 queue_peering_event(
1971 std::make_shared
<CephPeeringEvt
>(
1972 get_osdmap()->get_epoch(),
1973 get_osdmap()->get_epoch(),
1974 AllReplicasActivated())));
1977 bool PG::requeue_scrub(bool high_priority
)
1979 assert(is_locked());
1981 dout(10) << __func__
<< ": already queued" << dendl
;
1984 dout(10) << __func__
<< ": queueing" << dendl
;
1985 scrub_queued
= true;
1986 osd
->queue_for_scrub(this, high_priority
);
1991 void PG::queue_recovery()
1993 if (!is_primary() || !is_peered()) {
1994 dout(10) << "queue_recovery -- not primary or not peered " << dendl
;
1995 assert(!recovery_queued
);
1996 } else if (recovery_queued
) {
1997 dout(10) << "queue_recovery -- already queued" << dendl
;
1999 dout(10) << "queue_recovery -- queuing" << dendl
;
2000 recovery_queued
= true;
2001 osd
->queue_for_recovery(this);
2005 bool PG::queue_scrub()
2007 assert(is_locked());
2008 if (is_scrubbing()) {
2011 scrubber
.priority
= scrubber
.must_scrub
?
2012 cct
->_conf
->osd_requested_scrub_priority
: get_scrub_priority();
2013 scrubber
.must_scrub
= false;
2014 state_set(PG_STATE_SCRUBBING
);
2015 if (scrubber
.must_deep_scrub
) {
2016 state_set(PG_STATE_DEEP_SCRUB
);
2017 scrubber
.must_deep_scrub
= false;
2019 if (scrubber
.must_repair
|| scrubber
.auto_repair
) {
2020 state_set(PG_STATE_REPAIR
);
2021 scrubber
.must_repair
= false;
2027 unsigned PG::get_scrub_priority()
2029 // a higher value -> a higher priority
2030 int pool_scrub_priority
= 0;
2031 pool
.info
.opts
.get(pool_opts_t::SCRUB_PRIORITY
, &pool_scrub_priority
);
2032 return pool_scrub_priority
> 0 ? pool_scrub_priority
: cct
->_conf
->osd_scrub_priority
;
2035 struct C_PG_FinishRecovery
: public Context
{
2037 explicit C_PG_FinishRecovery(PG
*p
) : pg(p
) {}
2038 void finish(int r
) override
{
2039 pg
->_finish_recovery(this);
2043 void PG::mark_clean()
2045 if (actingset
.size() == get_osdmap()->get_pg_size(info
.pgid
.pgid
)) {
2046 state_clear(PG_STATE_FORCED_BACKFILL
| PG_STATE_FORCED_RECOVERY
);
2047 state_set(PG_STATE_CLEAN
);
2048 info
.history
.last_epoch_clean
= get_osdmap()->get_epoch();
2049 info
.history
.last_interval_clean
= info
.history
.same_interval_since
;
2050 past_intervals
.clear();
2051 dirty_big_info
= true;
2058 void PG::_change_recovery_force_mode(int new_mode
, bool clear
)
2061 // we can't and shouldn't do anything if the PG is being deleted locally
2063 state_clear(new_mode
);
2065 state_set(new_mode
);
2067 publish_stats_to_osd();
2071 inline int PG::clamp_recovery_priority(int priority
)
2073 static_assert(OSD_RECOVERY_PRIORITY_MIN
< OSD_RECOVERY_PRIORITY_MAX
, "Invalid priority range");
2074 static_assert(OSD_RECOVERY_PRIORITY_MIN
>= 0, "Priority range must match unsigned type");
2076 // Clamp to valid range
2077 if (priority
> OSD_RECOVERY_PRIORITY_MAX
) {
2078 return OSD_RECOVERY_PRIORITY_MAX
;
2079 } else if (priority
< OSD_RECOVERY_PRIORITY_MIN
) {
2080 return OSD_RECOVERY_PRIORITY_MIN
;
2086 unsigned PG::get_recovery_priority()
2088 // a higher value -> a higher priority
2091 if (state
& PG_STATE_FORCED_RECOVERY
) {
2092 ret
= OSD_RECOVERY_PRIORITY_FORCED
;
2094 pool
.info
.opts
.get(pool_opts_t::RECOVERY_PRIORITY
, &ret
);
2095 ret
= clamp_recovery_priority(OSD_RECOVERY_PRIORITY_BASE
+ ret
);
2097 dout(20) << __func__
<< " recovery priority for " << *this << " is " << ret
<< ", state is " << state
<< dendl
;
2098 return static_cast<unsigned>(ret
);
2101 unsigned PG::get_backfill_priority()
2103 // a higher value -> a higher priority
2104 int ret
= OSD_BACKFILL_PRIORITY_BASE
;
2105 if (state
& PG_STATE_FORCED_BACKFILL
) {
2106 ret
= OSD_RECOVERY_PRIORITY_FORCED
;
2108 if (acting
.size() < pool
.info
.min_size
) {
2109 // inactive: no. of replicas < min_size, highest priority since it blocks IO
2110 ret
= OSD_BACKFILL_INACTIVE_PRIORITY_BASE
+ (pool
.info
.min_size
- acting
.size());
2112 } else if (is_undersized()) {
2113 // undersized: OSD_BACKFILL_DEGRADED_PRIORITY_BASE + num missing replicas
2114 assert(pool
.info
.size
> actingset
.size());
2115 ret
= OSD_BACKFILL_DEGRADED_PRIORITY_BASE
+ (pool
.info
.size
- actingset
.size());
2117 } else if (is_degraded()) {
2118 // degraded: baseline degraded
2119 ret
= OSD_BACKFILL_DEGRADED_PRIORITY_BASE
;
2122 // Adjust with pool's recovery priority
2123 int pool_recovery_priority
= 0;
2124 pool
.info
.opts
.get(pool_opts_t::RECOVERY_PRIORITY
, &pool_recovery_priority
);
2126 ret
= clamp_recovery_priority(pool_recovery_priority
+ ret
);
2129 return static_cast<unsigned>(ret
);
2132 void PG::finish_recovery(list
<Context
*>& tfin
)
2134 dout(10) << "finish_recovery" << dendl
;
2135 assert(info
.last_complete
== info
.last_update
);
2137 clear_recovery_state();
2140 * sync all this before purging strays. but don't block!
2142 finish_sync_event
= new C_PG_FinishRecovery(this);
2143 tfin
.push_back(finish_sync_event
);
2146 void PG::_finish_recovery(Context
*c
)
2153 if (c
== finish_sync_event
) {
2154 dout(10) << "_finish_recovery" << dendl
;
2155 finish_sync_event
= 0;
2158 publish_stats_to_osd();
2160 if (scrub_after_recovery
) {
2161 dout(10) << "_finish_recovery requeueing for scrub" << dendl
;
2162 scrub_after_recovery
= false;
2163 scrubber
.must_deep_scrub
= true;
2167 dout(10) << "_finish_recovery -- stale" << dendl
;
2172 void PG::start_recovery_op(const hobject_t
& soid
)
2174 dout(10) << "start_recovery_op " << soid
2175 #ifdef DEBUG_RECOVERY_OIDS
2176 << " (" << recovering_oids
<< ")"
2179 assert(recovery_ops_active
>= 0);
2180 recovery_ops_active
++;
2181 #ifdef DEBUG_RECOVERY_OIDS
2182 assert(recovering_oids
.count(soid
) == 0);
2183 recovering_oids
.insert(soid
);
2185 osd
->start_recovery_op(this, soid
);
2188 void PG::finish_recovery_op(const hobject_t
& soid
, bool dequeue
)
2190 dout(10) << "finish_recovery_op " << soid
2191 #ifdef DEBUG_RECOVERY_OIDS
2192 << " (" << recovering_oids
<< ")"
2195 assert(recovery_ops_active
> 0);
2196 recovery_ops_active
--;
2197 #ifdef DEBUG_RECOVERY_OIDS
2198 assert(recovering_oids
.count(soid
));
2199 recovering_oids
.erase(soid
);
2201 osd
->finish_recovery_op(this, soid
, dequeue
);
2208 void PG::split_into(pg_t child_pgid
, PG
*child
, unsigned split_bits
)
2210 child
->update_snap_mapper_bits(split_bits
);
2211 child
->update_osdmap_ref(get_osdmap());
2216 pg_log
.split_into(child_pgid
, split_bits
, &(child
->pg_log
));
2217 child
->info
.last_complete
= info
.last_complete
;
2219 info
.last_update
= pg_log
.get_head();
2220 child
->info
.last_update
= child
->pg_log
.get_head();
2222 child
->info
.last_user_version
= info
.last_user_version
;
2224 info
.log_tail
= pg_log
.get_tail();
2225 child
->info
.log_tail
= child
->pg_log
.get_tail();
2227 if (info
.last_complete
< pg_log
.get_tail())
2228 info
.last_complete
= pg_log
.get_tail();
2229 if (child
->info
.last_complete
< child
->pg_log
.get_tail())
2230 child
->info
.last_complete
= child
->pg_log
.get_tail();
2233 child
->info
.history
= info
.history
;
2234 child
->info
.history
.epoch_created
= get_osdmap()->get_epoch();
2235 child
->info
.purged_snaps
= info
.purged_snaps
;
2237 if (info
.last_backfill
.is_max()) {
2238 child
->info
.set_last_backfill(hobject_t::get_max());
2240 // restart backfill on parent and child to be safe. we could
2241 // probably do better in the bitwise sort case, but it's more
2242 // fragile (there may be special work to do on backfill completion
2244 info
.set_last_backfill(hobject_t());
2245 child
->info
.set_last_backfill(hobject_t());
2246 // restarting backfill implies that the missing set is empty,
2247 // since it is only used for objects prior to last_backfill
2248 pg_log
.reset_backfill();
2249 child
->pg_log
.reset_backfill();
2252 child
->info
.stats
= info
.stats
;
2253 child
->info
.stats
.parent_split_bits
= split_bits
;
2254 info
.stats
.stats_invalid
= true;
2255 child
->info
.stats
.stats_invalid
= true;
2256 child
->info
.last_epoch_started
= info
.last_epoch_started
;
2257 child
->info
.last_interval_started
= info
.last_interval_started
;
2259 child
->snap_trimq
= snap_trimq
;
2261 // There can't be recovery/backfill going on now
2262 int primary
, up_primary
;
2263 vector
<int> newup
, newacting
;
2264 get_osdmap()->pg_to_up_acting_osds(
2265 child
->info
.pgid
.pgid
, &newup
, &up_primary
, &newacting
, &primary
);
2266 child
->init_primary_up_acting(
2271 child
->role
= OSDMap::calc_pg_role(osd
->whoami
, child
->acting
);
2273 // this comparison includes primary rank via pg_shard_t
2274 if (get_primary() != child
->get_primary())
2275 child
->info
.history
.same_primary_since
= get_osdmap()->get_epoch();
2277 child
->info
.stats
.up
= up
;
2278 child
->info
.stats
.up_primary
= up_primary
;
2279 child
->info
.stats
.acting
= acting
;
2280 child
->info
.stats
.acting_primary
= primary
;
2281 child
->info
.stats
.mapping_epoch
= get_osdmap()->get_epoch();
2284 child
->past_intervals
= past_intervals
;
2286 _split_into(child_pgid
, child
, split_bits
);
2288 // release all backoffs for simplicity
2289 release_backoffs(hobject_t(), hobject_t::get_max());
2291 child
->on_new_interval();
2293 child
->dirty_info
= true;
2294 child
->dirty_big_info
= true;
2296 dirty_big_info
= true;
2299 void PG::add_backoff(SessionRef s
, const hobject_t
& begin
, const hobject_t
& end
)
2301 ConnectionRef con
= s
->con
;
2302 if (!con
) // OSD::ms_handle_reset clears s->con without a lock
2304 BackoffRef
b(s
->have_backoff(info
.pgid
, begin
));
2306 derr
<< __func__
<< " already have backoff for " << s
<< " begin " << begin
2307 << " " << *b
<< dendl
;
2310 Mutex::Locker
l(backoff_lock
);
2312 b
= new Backoff(info
.pgid
, this, s
, ++s
->backoff_seq
, begin
, end
);
2313 backoffs
[begin
].insert(b
);
2315 dout(10) << __func__
<< " session " << s
<< " added " << *b
<< dendl
;
2320 get_osdmap()->get_epoch(),
2321 CEPH_OSD_BACKOFF_OP_BLOCK
,
2327 void PG::release_backoffs(const hobject_t
& begin
, const hobject_t
& end
)
2329 dout(10) << __func__
<< " [" << begin
<< "," << end
<< ")" << dendl
;
2330 vector
<BackoffRef
> bv
;
2332 Mutex::Locker
l(backoff_lock
);
2333 auto p
= backoffs
.lower_bound(begin
);
2334 while (p
!= backoffs
.end()) {
2335 int r
= cmp(p
->first
, end
);
2336 dout(20) << __func__
<< " ? " << r
<< " " << p
->first
2337 << " " << p
->second
<< dendl
;
2338 // note: must still examine begin=end=p->first case
2339 if (r
> 0 || (r
== 0 && begin
< end
)) {
2342 dout(20) << __func__
<< " checking " << p
->first
2343 << " " << p
->second
<< dendl
;
2344 auto q
= p
->second
.begin();
2345 while (q
!= p
->second
.end()) {
2346 dout(20) << __func__
<< " checking " << *q
<< dendl
;
2347 int r
= cmp((*q
)->begin
, begin
);
2348 if (r
== 0 || (r
> 0 && (*q
)->end
< end
)) {
2350 q
= p
->second
.erase(q
);
2355 if (p
->second
.empty()) {
2356 p
= backoffs
.erase(p
);
2363 Mutex::Locker
l(b
->lock
);
2364 dout(10) << __func__
<< " " << *b
<< dendl
;
2366 assert(b
->pg
== this);
2367 ConnectionRef con
= b
->session
->con
;
2368 if (con
) { // OSD::ms_handle_reset clears s->con without a lock
2372 get_osdmap()->get_epoch(),
2373 CEPH_OSD_BACKOFF_OP_UNBLOCK
,
2379 b
->state
= Backoff::STATE_DELETING
;
2381 b
->session
->rm_backoff(b
);
2389 void PG::clear_backoffs()
2391 dout(10) << __func__
<< " " << dendl
;
2392 map
<hobject_t
,set
<BackoffRef
>> ls
;
2394 Mutex::Locker
l(backoff_lock
);
2397 for (auto& p
: ls
) {
2398 for (auto& b
: p
.second
) {
2399 Mutex::Locker
l(b
->lock
);
2400 dout(10) << __func__
<< " " << *b
<< dendl
;
2402 assert(b
->pg
== this);
2404 b
->state
= Backoff::STATE_DELETING
;
2406 b
->session
->rm_backoff(b
);
2415 // called by Session::clear_backoffs()
2416 void PG::rm_backoff(BackoffRef b
)
2418 dout(10) << __func__
<< " " << *b
<< dendl
;
2419 Mutex::Locker
l(backoff_lock
);
2420 assert(b
->lock
.is_locked_by_me());
2421 assert(b
->pg
== this);
2422 auto p
= backoffs
.find(b
->begin
);
2423 // may race with release_backoffs()
2424 if (p
!= backoffs
.end()) {
2425 auto q
= p
->second
.find(b
);
2426 if (q
!= p
->second
.end()) {
2428 if (p
->second
.empty()) {
2435 void PG::clear_recovery_state()
2437 dout(10) << "clear_recovery_state" << dendl
;
2439 pg_log
.reset_recovery_pointers();
2440 finish_sync_event
= 0;
2443 while (recovery_ops_active
> 0) {
2444 #ifdef DEBUG_RECOVERY_OIDS
2445 soid
= *recovering_oids
.begin();
2447 finish_recovery_op(soid
, true);
2450 backfill_targets
.clear();
2451 backfill_info
.clear();
2452 peer_backfill_info
.clear();
2453 waiting_on_backfill
.clear();
2454 _clear_recovery_state(); // pg impl specific hook
2457 void PG::cancel_recovery()
2459 dout(10) << "cancel_recovery" << dendl
;
2460 clear_recovery_state();
2464 void PG::purge_strays()
2466 dout(10) << "purge_strays " << stray_set
<< dendl
;
2468 bool removed
= false;
2469 for (set
<pg_shard_t
>::iterator p
= stray_set
.begin();
2470 p
!= stray_set
.end();
2472 assert(!is_actingbackfill(*p
));
2473 if (get_osdmap()->is_up(p
->osd
)) {
2474 dout(10) << "sending PGRemove to osd." << *p
<< dendl
;
2475 vector
<spg_t
> to_remove
;
2476 to_remove
.push_back(spg_t(info
.pgid
.pgid
, p
->shard
));
2477 MOSDPGRemove
*m
= new MOSDPGRemove(
2478 get_osdmap()->get_epoch(),
2480 osd
->send_message_osd_cluster(p
->osd
, m
, get_osdmap()->get_epoch());
2482 dout(10) << "not sending PGRemove to down osd." << *p
<< dendl
;
2484 peer_missing
.erase(*p
);
2485 peer_info
.erase(*p
);
2486 peer_purged
.insert(*p
);
2490 // if we removed anyone, update peers (which include peer_info)
2492 update_heartbeat_peers();
2496 // clear _requested maps; we may have to peer() again if we discover
2497 // (more) stray content
2498 peer_log_requested
.clear();
2499 peer_missing_requested
.clear();
2502 void PG::set_probe_targets(const set
<pg_shard_t
> &probe_set
)
2504 Mutex::Locker
l(heartbeat_peer_lock
);
2505 probe_targets
.clear();
2506 for (set
<pg_shard_t
>::iterator i
= probe_set
.begin();
2507 i
!= probe_set
.end();
2509 probe_targets
.insert(i
->osd
);
2513 void PG::clear_probe_targets()
2515 Mutex::Locker
l(heartbeat_peer_lock
);
2516 probe_targets
.clear();
2519 void PG::update_heartbeat_peers()
2521 assert(is_locked());
2527 for (unsigned i
=0; i
<acting
.size(); i
++) {
2528 if (acting
[i
] != CRUSH_ITEM_NONE
)
2529 new_peers
.insert(acting
[i
]);
2531 for (unsigned i
=0; i
<up
.size(); i
++) {
2532 if (up
[i
] != CRUSH_ITEM_NONE
)
2533 new_peers
.insert(up
[i
]);
2535 for (map
<pg_shard_t
,pg_info_t
>::iterator p
= peer_info
.begin();
2536 p
!= peer_info
.end();
2538 new_peers
.insert(p
->first
.osd
);
2540 bool need_update
= false;
2541 heartbeat_peer_lock
.Lock();
2542 if (new_peers
== heartbeat_peers
) {
2543 dout(10) << "update_heartbeat_peers " << heartbeat_peers
<< " unchanged" << dendl
;
2545 dout(10) << "update_heartbeat_peers " << heartbeat_peers
<< " -> " << new_peers
<< dendl
;
2546 heartbeat_peers
.swap(new_peers
);
2549 heartbeat_peer_lock
.Unlock();
2552 osd
->need_heartbeat_peer_update();
2556 bool PG::check_in_progress_op(
2557 const osd_reqid_t
&r
,
2558 eversion_t
*version
,
2559 version_t
*user_version
,
2560 int *return_code
) const
2563 projected_log
.get_request(r
, version
, user_version
, return_code
) ||
2564 pg_log
.get_log().get_request(r
, version
, user_version
, return_code
));
2567 void PG::_update_calc_stats()
2569 info
.stats
.version
= info
.last_update
;
2570 info
.stats
.created
= info
.history
.epoch_created
;
2571 info
.stats
.last_scrub
= info
.history
.last_scrub
;
2572 info
.stats
.last_scrub_stamp
= info
.history
.last_scrub_stamp
;
2573 info
.stats
.last_deep_scrub
= info
.history
.last_deep_scrub
;
2574 info
.stats
.last_deep_scrub_stamp
= info
.history
.last_deep_scrub_stamp
;
2575 info
.stats
.last_clean_scrub_stamp
= info
.history
.last_clean_scrub_stamp
;
2576 info
.stats
.last_epoch_clean
= info
.history
.last_epoch_clean
;
2578 info
.stats
.log_size
= pg_log
.get_head().version
- pg_log
.get_tail().version
;
2579 info
.stats
.ondisk_log_size
= info
.stats
.log_size
;
2580 info
.stats
.log_start
= pg_log
.get_tail();
2581 info
.stats
.ondisk_log_start
= pg_log
.get_tail();
2582 info
.stats
.snaptrimq_len
= snap_trimq
.size();
2584 unsigned num_shards
= get_osdmap()->get_pg_size(info
.pgid
.pgid
);
2586 // In rare case that upset is too large (usually transient), use as target
2587 // for calculations below.
2588 unsigned target
= std::max(num_shards
, (unsigned)upset
.size());
2589 // Not sure this could ever happen, that actingset > upset
2590 // which only matters if actingset > num_shards.
2591 unsigned nrep
= std::max(actingset
.size(), upset
.size());
2592 // calc num_object_copies
2593 info
.stats
.stats
.calc_copies(MAX(target
, nrep
));
2594 info
.stats
.stats
.sum
.num_objects_degraded
= 0;
2595 info
.stats
.stats
.sum
.num_objects_unfound
= 0;
2596 info
.stats
.stats
.sum
.num_objects_misplaced
= 0;
2597 if ((is_remapped() || is_undersized() || !is_clean()) && (is_peered() || is_activating())) {
2598 dout(20) << __func__
<< " actingset " << actingset
<< " upset "
2599 << upset
<< " actingbackfill " << actingbackfill
<< dendl
;
2600 dout(20) << __func__
<< " acting " << acting
<< " up " << up
<< dendl
;
2602 assert(!actingbackfill
.empty());
2604 // NOTE: we only generate degraded, misplaced and unfound
2605 // values for the summation, not individual stat categories.
2606 int64_t num_objects
= info
.stats
.stats
.sum
.num_objects
;
2608 // Objects missing from up nodes, sorted by # objects.
2609 boost::container::flat_set
<pair
<int64_t,pg_shard_t
>> missing_target_objects
;
2610 // Objects missing from nodes not in up, sort by # objects
2611 boost::container::flat_set
<pair
<int64_t,pg_shard_t
>> acting_source_objects
;
2616 missing
= pg_log
.get_missing().num_missing();
2617 assert(actingbackfill
.count(pg_whoami
));
2618 if (upset
.count(pg_whoami
)) {
2619 missing_target_objects
.insert(make_pair(missing
, pg_whoami
));
2621 acting_source_objects
.insert(make_pair(missing
, pg_whoami
));
2623 info
.stats
.stats
.sum
.num_objects_missing_on_primary
= missing
;
2626 for (auto& peer
: peer_info
) {
2627 // Ignore other peers until we add code to look at detailed missing
2628 // information. (recovery)
2629 if (!actingbackfill
.count(peer
.first
)) {
2633 // Backfill targets always track num_objects accurately
2634 // all other peers track missing accurately.
2635 if (is_backfill_targets(peer
.first
)) {
2636 missing
= std::max((int64_t)0, num_objects
- peer
.second
.stats
.stats
.sum
.num_objects
);
2638 if (peer_missing
.count(peer
.first
)) {
2639 missing
= peer_missing
[peer
.first
].num_missing();
2641 dout(20) << __func__
<< " no peer_missing found for " << peer
.first
<< dendl
;
2644 if (upset
.count(peer
.first
)) {
2645 missing_target_objects
.insert(make_pair(missing
, peer
.first
));
2647 acting_source_objects
.insert(make_pair(missing
, peer
.first
));
2649 peer
.second
.stats
.stats
.sum
.num_objects_missing
= missing
;
2652 if (pool
.info
.is_replicated()) {
2653 // Add to missing_target_objects up to target elements (num_objects missing)
2654 assert(target
>= missing_target_objects
.size());
2655 unsigned needed
= target
- missing_target_objects
.size();
2656 for (; needed
; --needed
)
2657 missing_target_objects
.insert(make_pair(num_objects
, pg_shard_t(pg_shard_t::NO_OSD
)));
2659 for (unsigned i
= 0 ; i
< num_shards
; ++i
) {
2660 shard_id_t
shard(i
);
2662 for (const auto& t
: missing_target_objects
) {
2663 if (std::get
<1>(t
).shard
== shard
) {
2669 missing_target_objects
.insert(make_pair(num_objects
, pg_shard_t(pg_shard_t::NO_OSD
,shard
)));
2673 for (const auto& item
: missing_target_objects
)
2674 dout(20) << __func__
<< " missing shard " << std::get
<1>(item
) << " missing= " << std::get
<0>(item
) << dendl
;
2675 for (const auto& item
: acting_source_objects
)
2676 dout(20) << __func__
<< " acting shard " << std::get
<1>(item
) << " missing= " << std::get
<0>(item
) << dendl
;
2678 // A misplaced object is not stored on the correct OSD
2679 int64_t misplaced
= 0;
2680 // a degraded objects has fewer replicas or EC shards than the pool specifies.
2681 int64_t degraded
= 0;
2683 for (auto m
= missing_target_objects
.rbegin();
2684 m
!= missing_target_objects
.rend(); ++m
) {
2686 int64_t extra_missing
= -1;
2688 if (pool
.info
.is_replicated()) {
2689 if (!acting_source_objects
.empty()) {
2690 auto extra_copy
= acting_source_objects
.begin();
2691 extra_missing
= std::get
<0>(*extra_copy
);
2692 acting_source_objects
.erase(extra_copy
);
2694 } else { // Erasure coded
2695 // Use corresponding shard
2696 for (const auto& a
: acting_source_objects
) {
2697 if (std::get
<1>(a
).shard
== std::get
<1>(*m
).shard
) {
2698 extra_missing
= std::get
<0>(a
);
2699 acting_source_objects
.erase(a
);
2705 if (extra_missing
>= 0 && std::get
<0>(*m
) >= extra_missing
) {
2706 // We don't know which of the objects on the target
2707 // are part of extra_missing so assume are all degraded.
2708 misplaced
+= std::get
<0>(*m
) - extra_missing
;
2709 degraded
+= extra_missing
;
2711 // 1. extra_missing == -1, more targets than sources so degraded
2712 // 2. extra_missing > std::get<0>(m), so that we know that some extra_missing
2713 // previously degraded are now present on the target.
2714 degraded
+= std::get
<0>(*m
);
2717 // If there are still acting that haven't been accounted for
2718 // then they are misplaced
2719 for (const auto& a
: acting_source_objects
) {
2720 int64_t extra_misplaced
= std::max((int64_t)0, num_objects
- std::get
<0>(a
));
2721 dout(20) << __func__
<< " extra acting misplaced " << extra_misplaced
<< dendl
;
2722 misplaced
+= extra_misplaced
;
2724 dout(20) << __func__
<< " degraded " << degraded
<< dendl
;
2725 dout(20) << __func__
<< " misplaced " << misplaced
<< dendl
;
2727 info
.stats
.stats
.sum
.num_objects_degraded
= degraded
;
2728 info
.stats
.stats
.sum
.num_objects_unfound
= get_num_unfound();
2729 info
.stats
.stats
.sum
.num_objects_misplaced
= misplaced
;
2733 void PG::_update_blocked_by()
2735 // set a max on the number of blocking peers we report. if we go
2736 // over, report a random subset. keep the result sorted.
2737 unsigned keep
= MIN(blocked_by
.size(), cct
->_conf
->osd_max_pg_blocked_by
);
2738 unsigned skip
= blocked_by
.size() - keep
;
2739 info
.stats
.blocked_by
.clear();
2740 info
.stats
.blocked_by
.resize(keep
);
2742 for (set
<int>::iterator p
= blocked_by
.begin();
2743 p
!= blocked_by
.end() && keep
> 0;
2745 if (skip
> 0 && (rand() % (skip
+ keep
) < skip
)) {
2748 info
.stats
.blocked_by
[pos
++] = *p
;
2754 void PG::publish_stats_to_osd()
2759 pg_stats_publish_lock
.Lock();
2761 if (info
.stats
.stats
.sum
.num_scrub_errors
)
2762 state_set(PG_STATE_INCONSISTENT
);
2764 state_clear(PG_STATE_INCONSISTENT
);
2766 utime_t now
= ceph_clock_now();
2767 if (info
.stats
.state
!= state
) {
2768 info
.stats
.last_change
= now
;
2769 // Optimistic estimation, if we just find out an inactive PG,
2770 // assumt it is active till now.
2771 if (!(state
& PG_STATE_ACTIVE
) &&
2772 (info
.stats
.state
& PG_STATE_ACTIVE
))
2773 info
.stats
.last_active
= now
;
2775 if ((state
& PG_STATE_ACTIVE
) &&
2776 !(info
.stats
.state
& PG_STATE_ACTIVE
))
2777 info
.stats
.last_became_active
= now
;
2778 if ((state
& (PG_STATE_ACTIVE
|PG_STATE_PEERED
)) &&
2779 !(info
.stats
.state
& (PG_STATE_ACTIVE
|PG_STATE_PEERED
)))
2780 info
.stats
.last_became_peered
= now
;
2781 if (!(state
& PG_STATE_CREATING
) &&
2782 (info
.stats
.state
& PG_STATE_CREATING
)) {
2783 osd
->send_pg_created(get_pgid().pgid
);
2785 info
.stats
.state
= state
;
2788 _update_calc_stats();
2789 if (info
.stats
.stats
.sum
.num_objects_degraded
) {
2790 state_set(PG_STATE_DEGRADED
);
2792 state_clear(PG_STATE_DEGRADED
);
2794 _update_blocked_by();
2796 bool publish
= false;
2797 pg_stat_t pre_publish
= info
.stats
;
2798 pre_publish
.stats
.add(unstable_stats
);
2799 utime_t cutoff
= now
;
2800 cutoff
-= cct
->_conf
->osd_pg_stat_report_interval_max
;
2801 if (pg_stats_publish_valid
&& pre_publish
== pg_stats_publish
&&
2802 info
.stats
.last_fresh
> cutoff
) {
2803 dout(15) << "publish_stats_to_osd " << pg_stats_publish
.reported_epoch
2804 << ": no change since " << info
.stats
.last_fresh
<< dendl
;
2806 // update our stat summary and timestamps
2807 info
.stats
.reported_epoch
= get_osdmap()->get_epoch();
2808 ++info
.stats
.reported_seq
;
2810 info
.stats
.last_fresh
= now
;
2812 if (info
.stats
.state
& PG_STATE_CLEAN
)
2813 info
.stats
.last_clean
= now
;
2814 if (info
.stats
.state
& PG_STATE_ACTIVE
)
2815 info
.stats
.last_active
= now
;
2816 if (info
.stats
.state
& (PG_STATE_ACTIVE
|PG_STATE_PEERED
))
2817 info
.stats
.last_peered
= now
;
2818 info
.stats
.last_unstale
= now
;
2819 if ((info
.stats
.state
& PG_STATE_DEGRADED
) == 0)
2820 info
.stats
.last_undegraded
= now
;
2821 if ((info
.stats
.state
& PG_STATE_UNDERSIZED
) == 0)
2822 info
.stats
.last_fullsized
= now
;
2824 // do not send pgstat to mon anymore once we are luminous, since mgr takes
2825 // care of this by sending MMonMgrReport to mon.
2827 osd
->osd
->get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
;
2828 pg_stats_publish_valid
= true;
2829 pg_stats_publish
= pre_publish
;
2831 dout(15) << "publish_stats_to_osd " << pg_stats_publish
.reported_epoch
2832 << ":" << pg_stats_publish
.reported_seq
<< dendl
;
2834 pg_stats_publish_lock
.Unlock();
2837 osd
->pg_stat_queue_enqueue(this);
2840 void PG::clear_publish_stats()
2842 dout(15) << "clear_stats" << dendl
;
2843 pg_stats_publish_lock
.Lock();
2844 pg_stats_publish_valid
= false;
2845 pg_stats_publish_lock
.Unlock();
2847 osd
->pg_stat_queue_dequeue(this);
2851 * initialize a newly instantiated pg
2853 * Initialize PG state, as when a PG is initially created, or when it
2854 * is first instantiated on the current node.
2856 * @param role our role/rank
2857 * @param newup up set
2858 * @param newacting acting set
2859 * @param history pg history
2860 * @param pi past_intervals
2861 * @param backfill true if info should be marked as backfill
2862 * @param t transaction to write out our new state in
2866 const vector
<int>& newup
, int new_up_primary
,
2867 const vector
<int>& newacting
, int new_acting_primary
,
2868 const pg_history_t
& history
,
2869 const PastIntervals
& pi
,
2871 ObjectStore::Transaction
*t
)
2873 dout(10) << "init role " << role
<< " up " << newup
<< " acting " << newacting
2874 << " history " << history
2875 << " past_intervals " << pi
2881 init_primary_up_acting(
2885 new_acting_primary
);
2887 info
.history
= history
;
2888 past_intervals
= pi
;
2891 info
.stats
.up_primary
= new_up_primary
;
2892 info
.stats
.acting
= acting
;
2893 info
.stats
.acting_primary
= new_acting_primary
;
2894 info
.stats
.mapping_epoch
= info
.history
.same_interval_since
;
2897 dout(10) << __func__
<< ": Setting backfill" << dendl
;
2898 info
.set_last_backfill(hobject_t());
2899 info
.last_complete
= info
.last_update
;
2900 pg_log
.mark_log_for_rewrite();
2906 dirty_big_info
= true;
2910 #pragma GCC diagnostic ignored "-Wpragmas"
2911 #pragma GCC diagnostic push
2912 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
2914 void PG::upgrade(ObjectStore
*store
)
2916 assert(info_struct_v
<= 10);
2917 ObjectStore::Transaction t
;
2919 assert(info_struct_v
>= 7);
2922 if (info_struct_v
<= 7) {
2923 pg_log
.mark_log_for_rewrite();
2924 ghobject_t
log_oid(OSD::make_pg_log_oid(pg_id
));
2925 ghobject_t
biginfo_oid(OSD::make_pg_biginfo_oid(pg_id
));
2926 t
.remove(coll_t::meta(), log_oid
);
2927 t
.remove(coll_t::meta(), biginfo_oid
);
2928 t
.touch(coll
, pgmeta_oid
);
2932 if (info_struct_v
<= 8) {
2933 // no special action needed.
2937 if (info_struct_v
<= 9) {
2938 // previous versions weren't (as) aggressively clearing past_intervals
2939 if (info
.history
.last_epoch_clean
>= info
.history
.same_interval_since
) {
2940 dout(20) << __func__
<< " clearing past_intervals" << dendl
;
2941 past_intervals
.clear();
2945 // update infover_key
2946 if (info_struct_v
< cur_struct_v
) {
2947 map
<string
,bufferlist
> v
;
2948 __u8 ver
= cur_struct_v
;
2949 ::encode(ver
, v
[infover_key
]);
2950 t
.omap_setkeys(coll
, pgmeta_oid
, v
);
2954 dirty_big_info
= true;
2957 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
2958 ObjectStore::Sequencer
>("upgrade"));
2959 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
2961 derr
<< __func__
<< ": apply_transaction returned "
2962 << cpp_strerror(r
) << dendl
;
2968 if (!osr
->flush_commit(&waiter
)) {
2973 #pragma GCC diagnostic pop
2974 #pragma GCC diagnostic warning "-Wpragmas"
2976 int PG::_prepare_write_info(CephContext
* cct
,
2977 map
<string
,bufferlist
> *km
,
2979 pg_info_t
&info
, pg_info_t
&last_written_info
,
2980 PastIntervals
&past_intervals
,
2981 bool dirty_big_info
,
2984 PerfCounters
*logger
)
2987 ::encode(epoch
, (*km
)[epoch_key
]);
2991 logger
->inc(l_osd_pg_info
);
2993 // try to do info efficiently?
2994 if (!dirty_big_info
&& try_fast_info
&&
2995 info
.last_update
> last_written_info
.last_update
) {
2996 pg_fast_info_t fast
;
2997 fast
.populate_from(info
);
2998 bool did
= fast
.try_apply_to(&last_written_info
);
2999 assert(did
); // we verified last_update increased above
3000 if (info
== last_written_info
) {
3001 ::encode(fast
, (*km
)[fastinfo_key
]);
3003 logger
->inc(l_osd_pg_fastinfo
);
3006 generic_dout(30) << __func__
<< " fastinfo failed, info:\n";
3008 JSONFormatter
jf(true);
3009 jf
.dump_object("info", info
);
3013 *_dout
<< "\nlast_written_info:\n";
3014 JSONFormatter
jf(true);
3015 jf
.dump_object("last_written_info", last_written_info
);
3020 last_written_info
= info
;
3022 // info. store purged_snaps separately.
3023 interval_set
<snapid_t
> purged_snaps
;
3024 purged_snaps
.swap(info
.purged_snaps
);
3025 ::encode(info
, (*km
)[info_key
]);
3026 purged_snaps
.swap(info
.purged_snaps
);
3028 if (dirty_big_info
) {
3029 // potentially big stuff
3030 bufferlist
& bigbl
= (*km
)[biginfo_key
];
3031 ::encode(past_intervals
, bigbl
);
3032 ::encode(info
.purged_snaps
, bigbl
);
3033 //dout(20) << "write_info bigbl " << bigbl.length() << dendl;
3035 logger
->inc(l_osd_pg_biginfo
);
3041 void PG::_create(ObjectStore::Transaction
& t
, spg_t pgid
, int bits
)
3044 t
.create_collection(coll
, bits
);
3047 void PG::_init(ObjectStore::Transaction
& t
, spg_t pgid
, const pg_pool_t
*pool
)
3052 // Give a hint to the PG collection
3054 uint32_t pg_num
= pool
->get_pg_num();
3055 uint64_t expected_num_objects_pg
= pool
->expected_num_objects
/ pg_num
;
3056 ::encode(pg_num
, hint
);
3057 ::encode(expected_num_objects_pg
, hint
);
3058 uint32_t hint_type
= ObjectStore::Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS
;
3059 t
.collection_hint(coll
, hint_type
, hint
);
3062 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3063 t
.touch(coll
, pgmeta_oid
);
3064 map
<string
,bufferlist
> values
;
3065 __u8 struct_v
= cur_struct_v
;
3066 ::encode(struct_v
, values
[infover_key
]);
3067 t
.omap_setkeys(coll
, pgmeta_oid
, values
);
3070 void PG::prepare_write_info(map
<string
,bufferlist
> *km
)
3072 info
.stats
.stats
.add(unstable_stats
);
3073 unstable_stats
.clear();
3075 bool need_update_epoch
= last_epoch
< get_osdmap()->get_epoch();
3076 int ret
= _prepare_write_info(cct
, km
, get_osdmap()->get_epoch(),
3080 dirty_big_info
, need_update_epoch
,
3081 cct
->_conf
->osd_fast_info
,
3084 if (need_update_epoch
)
3085 last_epoch
= get_osdmap()->get_epoch();
3086 last_persisted_osdmap_ref
= osdmap_ref
;
3089 dirty_big_info
= false;
3092 #pragma GCC diagnostic ignored "-Wpragmas"
3093 #pragma GCC diagnostic push
3094 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
3096 bool PG::_has_removal_flag(ObjectStore
*store
,
3100 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3102 // first try new way
3104 keys
.insert("_remove");
3105 map
<string
,bufferlist
> values
;
3106 if (store
->omap_get_values(coll
, pgmeta_oid
, keys
, &values
) == 0 &&
3113 int PG::peek_map_epoch(ObjectStore
*store
,
3119 ghobject_t
legacy_infos_oid(OSD::make_infos_oid());
3120 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3121 epoch_t cur_epoch
= 0;
3125 // validate collection name
3126 assert(coll
.is_pg());
3131 keys
.insert(infover_key
);
3132 keys
.insert(epoch_key
);
3133 map
<string
,bufferlist
> values
;
3134 int r
= store
->omap_get_values(coll
, pgmeta_oid
, keys
, &values
);
3136 assert(values
.size() == 2);
3138 // sanity check version
3139 bufferlist::iterator bp
= values
[infover_key
].begin();
3141 ::decode(struct_v
, bp
);
3142 assert(struct_v
>= 8);
3145 bp
= values
[epoch_key
].begin();
3146 ::decode(cur_epoch
, bp
);
3148 // probably bug 10617; see OSD::load_pgs()
3152 *pepoch
= cur_epoch
;
3156 #pragma GCC diagnostic pop
3157 #pragma GCC diagnostic warning "-Wpragmas"
3159 void PG::write_if_dirty(ObjectStore::Transaction
& t
)
3161 map
<string
,bufferlist
> km
;
3162 if (dirty_big_info
|| dirty_info
)
3163 prepare_write_info(&km
);
3164 pg_log
.write_log_and_missing(t
, &km
, coll
, pgmeta_oid
, pool
.info
.require_rollback());
3166 t
.omap_setkeys(coll
, pgmeta_oid
, km
);
3171 assert(is_primary());
3173 dout(10) << __func__
<< " to " << pg_trim_to
<< dendl
;
3174 if (pg_trim_to
!= eversion_t()) {
3175 // inform peers to trim log
3176 assert(!actingbackfill
.empty());
3177 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
3178 i
!= actingbackfill
.end();
3180 if (*i
== pg_whoami
) continue;
3181 osd
->send_message_osd_cluster(
3184 get_osdmap()->get_epoch(),
3185 spg_t(info
.pgid
.pgid
, i
->shard
),
3187 get_osdmap()->get_epoch());
3190 // trim primary as well
3191 pg_log
.trim(pg_trim_to
, info
);
3196 void PG::add_log_entry(const pg_log_entry_t
& e
, bool applied
)
3198 // raise last_complete only if we were previously up to date
3199 if (info
.last_complete
== info
.last_update
)
3200 info
.last_complete
= e
.version
;
3202 // raise last_update.
3203 assert(e
.version
> info
.last_update
);
3204 info
.last_update
= e
.version
;
3206 // raise user_version, if it increased (it may have not get bumped
3207 // by all logged updates)
3208 if (e
.user_version
> info
.last_user_version
)
3209 info
.last_user_version
= e
.user_version
;
3212 pg_log
.add(e
, applied
);
3213 dout(10) << "add_log_entry " << e
<< dendl
;
3217 void PG::append_log(
3218 const vector
<pg_log_entry_t
>& logv
,
3220 eversion_t roll_forward_to
,
3221 ObjectStore::Transaction
&t
,
3222 bool transaction_applied
)
3224 if (transaction_applied
)
3225 update_snap_map(logv
, t
);
3227 /* The primary has sent an info updating the history, but it may not
3228 * have arrived yet. We want to make sure that we cannot remember this
3229 * write without remembering that it happened in an interval which went
3230 * active in epoch history.last_epoch_started.
3232 if (info
.last_epoch_started
!= info
.history
.last_epoch_started
) {
3233 info
.history
.last_epoch_started
= info
.last_epoch_started
;
3235 if (info
.last_interval_started
!= info
.history
.last_interval_started
) {
3236 info
.history
.last_interval_started
= info
.last_interval_started
;
3238 dout(10) << "append_log " << pg_log
.get_log() << " " << logv
<< dendl
;
3240 PGLogEntryHandler handler
{this, &t
};
3241 if (!transaction_applied
) {
3242 /* We must be a backfill peer, so it's ok if we apply
3243 * out-of-turn since we won't be considered when
3244 * determining a min possible last_update.
3246 pg_log
.roll_forward(&handler
);
3249 for (vector
<pg_log_entry_t
>::const_iterator p
= logv
.begin();
3252 add_log_entry(*p
, transaction_applied
);
3254 /* We don't want to leave the rollforward artifacts around
3255 * here past last_backfill. It's ok for the same reason as
3257 if (transaction_applied
&&
3258 p
->soid
> info
.last_backfill
) {
3259 pg_log
.roll_forward(&handler
);
3262 auto last
= logv
.rbegin();
3263 if (is_primary() && last
!= logv
.rend()) {
3264 projected_log
.skip_can_rollback_to_to_head();
3265 projected_log
.trim(cct
, last
->version
, nullptr, nullptr, nullptr);
3268 if (transaction_applied
&& roll_forward_to
> pg_log
.get_can_rollback_to()) {
3269 pg_log
.roll_forward_to(
3272 t
.register_on_applied(
3273 new C_UpdateLastRollbackInfoTrimmedToApplied(
3275 get_osdmap()->get_epoch(),
3279 pg_log
.trim(trim_to
, info
);
3281 // update the local pg, pg log
3286 bool PG::check_log_for_corruption(ObjectStore
*store
)
3288 /// TODO: this method needs to work with the omap log
3292 //! Get the name we're going to save our corrupt page log as
3293 std::string
PG::get_corrupt_pg_log_name() const
3295 const int MAX_BUF
= 512;
3298 time_t my_time(time(NULL
));
3299 const struct tm
*t
= localtime_r(&my_time
, &tm_buf
);
3300 int ret
= strftime(buf
, sizeof(buf
), "corrupt_log_%Y-%m-%d_%k:%M_", t
);
3302 dout(0) << "strftime failed" << dendl
;
3303 return "corrupt_log_unknown_time";
3306 out
+= stringify(info
.pgid
);
3311 ObjectStore
*store
, spg_t pgid
, const coll_t
&coll
, bufferlist
&bl
,
3312 pg_info_t
&info
, PastIntervals
&past_intervals
,
3315 // try for v8 or later
3317 keys
.insert(infover_key
);
3318 keys
.insert(info_key
);
3319 keys
.insert(biginfo_key
);
3320 keys
.insert(fastinfo_key
);
3321 ghobject_t
pgmeta_oid(pgid
.make_pgmeta_oid());
3322 map
<string
,bufferlist
> values
;
3323 int r
= store
->omap_get_values(coll
, pgmeta_oid
, keys
, &values
);
3325 assert(values
.size() == 3 ||
3326 values
.size() == 4);
3328 bufferlist::iterator p
= values
[infover_key
].begin();
3329 ::decode(struct_v
, p
);
3330 assert(struct_v
>= 8);
3332 p
= values
[info_key
].begin();
3335 p
= values
[biginfo_key
].begin();
3336 if (struct_v
>= 10) {
3337 ::decode(past_intervals
, p
);
3339 past_intervals
.decode_classic(p
);
3341 ::decode(info
.purged_snaps
, p
);
3343 p
= values
[fastinfo_key
].begin();
3345 pg_fast_info_t fast
;
3347 fast
.try_apply_to(&info
);
3353 ghobject_t
infos_oid(OSD::make_infos_oid());
3354 bufferlist::iterator p
= bl
.begin();
3355 ::decode(struct_v
, p
);
3356 assert(struct_v
== 7);
3358 // get info out of leveldb
3359 string k
= get_info_key(info
.pgid
);
3360 string bk
= get_biginfo_key(info
.pgid
);
3365 store
->omap_get_values(coll_t::meta(), ghobject_t(infos_oid
), keys
, &values
);
3366 assert(values
.size() == 2);
3368 p
= values
[k
].begin();
3371 p
= values
[bk
].begin();
3372 ::decode(past_intervals
, p
);
3373 interval_set
<snapid_t
> snap_collections
; // obsolete
3374 ::decode(snap_collections
, p
);
3375 ::decode(info
.purged_snaps
, p
);
3379 void PG::read_state(ObjectStore
*store
, bufferlist
&bl
)
3381 int r
= read_info(store
, pg_id
, coll
, bl
, info
, past_intervals
,
3385 last_written_info
= info
;
3387 // if we are upgrading from jewel, we need to force rebuild of
3388 // missing set. v9 was fastinfo, added v11.0.2-331-g1d5dc29a13
3389 // (before kraken). persisted missing set was circa
3390 // v11.0.0-866-gb0e239da95 (a bit earlier, also before kraken).
3391 // v8 was pre-jewel (per-pg meta object).
3392 bool force_rebuild_missing
= info_struct_v
< 9;
3393 if (force_rebuild_missing
) {
3394 dout(10) << __func__
<< " detected upgrade from jewel, force_rebuild_missing"
3399 pg_log
.read_log_and_missing(
3402 info_struct_v
< 8 ? coll_t::meta() : coll
,
3403 ghobject_t(info_struct_v
< 8 ? OSD::make_pg_log_oid(pg_id
) : pgmeta_oid
),
3405 force_rebuild_missing
,
3407 cct
->_conf
->osd_ignore_stale_divergent_priors
,
3408 cct
->_conf
->osd_debug_verify_missing_on_start
);
3410 osd
->clog
->error() << oss
.str();
3412 if (force_rebuild_missing
) {
3413 dout(10) << __func__
<< " forced rebuild of missing got "
3414 << pg_log
.get_missing()
3418 // log any weirdness
3422 void PG::log_weirdness()
3424 if (pg_log
.get_tail() != info
.log_tail
)
3425 osd
->clog
->error() << info
.pgid
3426 << " info mismatch, log.tail " << pg_log
.get_tail()
3427 << " != info.log_tail " << info
.log_tail
;
3428 if (pg_log
.get_head() != info
.last_update
)
3429 osd
->clog
->error() << info
.pgid
3430 << " info mismatch, log.head " << pg_log
.get_head()
3431 << " != info.last_update " << info
.last_update
;
3433 if (!pg_log
.get_log().empty()) {
3435 if ((pg_log
.get_log().log
.begin()->version
<= pg_log
.get_tail()))
3436 osd
->clog
->error() << info
.pgid
3437 << " log bound mismatch, info (tail,head] ("
3438 << pg_log
.get_tail() << "," << pg_log
.get_head() << "]"
3440 << pg_log
.get_log().log
.begin()->version
<< ","
3441 << pg_log
.get_log().log
.rbegin()->version
<< "]";
3444 if (pg_log
.get_log().caller_ops
.size() > pg_log
.get_log().log
.size()) {
3445 osd
->clog
->error() << info
.pgid
3446 << " caller_ops.size " << pg_log
.get_log().caller_ops
.size()
3447 << " > log size " << pg_log
.get_log().log
.size();
3451 void PG::update_snap_map(
3452 const vector
<pg_log_entry_t
> &log_entries
,
3453 ObjectStore::Transaction
&t
)
3455 for (vector
<pg_log_entry_t
>::const_iterator i
= log_entries
.begin();
3456 i
!= log_entries
.end();
3458 OSDriver::OSTransaction
_t(osdriver
.get_transaction(&t
));
3459 if (i
->soid
.snap
< CEPH_MAXSNAP
) {
3460 if (i
->is_delete()) {
3461 int r
= snap_mapper
.remove_oid(
3465 } else if (i
->is_update()) {
3466 assert(i
->snaps
.length() > 0);
3467 vector
<snapid_t
> snaps
;
3468 bufferlist snapbl
= i
->snaps
;
3469 bufferlist::iterator p
= snapbl
.begin();
3475 set
<snapid_t
> _snaps(snaps
.begin(), snaps
.end());
3477 if (i
->is_clone() || i
->is_promote()) {
3478 snap_mapper
.add_oid(
3482 } else if (i
->is_modify()) {
3483 assert(i
->is_modify());
3484 int r
= snap_mapper
.update_snaps(
3491 assert(i
->is_clean());
3499 * filter trimming|trimmed snaps out of snapcontext
3501 void PG::filter_snapc(vector
<snapid_t
> &snaps
)
3503 //nothing needs to trim, we can return immediately
3504 if(snap_trimq
.empty() && info
.purged_snaps
.empty())
3507 bool filtering
= false;
3508 vector
<snapid_t
> newsnaps
;
3509 for (vector
<snapid_t
>::iterator p
= snaps
.begin();
3512 if (snap_trimq
.contains(*p
) || info
.purged_snaps
.contains(*p
)) {
3514 // start building a new vector with what we've seen so far
3515 dout(10) << "filter_snapc filtering " << snaps
<< dendl
;
3516 newsnaps
.insert(newsnaps
.begin(), snaps
.begin(), p
);
3519 dout(20) << "filter_snapc removing trimq|purged snap " << *p
<< dendl
;
3522 newsnaps
.push_back(*p
); // continue building new vector
3526 snaps
.swap(newsnaps
);
3527 dout(10) << "filter_snapc result " << snaps
<< dendl
;
3531 void PG::requeue_object_waiters(map
<hobject_t
, list
<OpRequestRef
>>& m
)
3533 for (map
<hobject_t
, list
<OpRequestRef
>>::iterator it
= m
.begin();
3536 requeue_ops(it
->second
);
3540 void PG::requeue_op(OpRequestRef op
)
3542 auto p
= waiting_for_map
.find(op
->get_source());
3543 if (p
!= waiting_for_map
.end()) {
3544 dout(20) << __func__
<< " " << op
<< " (waiting_for_map " << p
->first
<< ")"
3546 p
->second
.push_front(op
);
3548 dout(20) << __func__
<< " " << op
<< dendl
;
3549 osd
->enqueue_front(info
.pgid
, PGQueueable(op
, get_osdmap()->get_epoch()));
3553 void PG::requeue_ops(list
<OpRequestRef
> &ls
)
3555 for (list
<OpRequestRef
>::reverse_iterator i
= ls
.rbegin();
3558 auto p
= waiting_for_map
.find((*i
)->get_source());
3559 if (p
!= waiting_for_map
.end()) {
3560 dout(20) << __func__
<< " " << *i
<< " (waiting_for_map " << p
->first
3562 p
->second
.push_front(*i
);
3564 dout(20) << __func__
<< " " << *i
<< dendl
;
3565 osd
->enqueue_front(info
.pgid
, PGQueueable(*i
, get_osdmap()->get_epoch()));
3571 void PG::requeue_map_waiters()
3573 epoch_t epoch
= get_osdmap()->get_epoch();
3574 auto p
= waiting_for_map
.begin();
3575 while (p
!= waiting_for_map
.end()) {
3576 if (epoch
< p
->second
.front()->min_epoch
) {
3577 dout(20) << __func__
<< " " << p
->first
<< " front op "
3578 << p
->second
.front() << " must still wait, doing nothing"
3582 dout(20) << __func__
<< " " << p
->first
<< " " << p
->second
<< dendl
;
3583 for (auto q
= p
->second
.rbegin(); q
!= p
->second
.rend(); ++q
) {
3584 osd
->enqueue_front(info
.pgid
, PGQueueable(*q
, epoch
));
3586 p
= waiting_for_map
.erase(p
);
3592 // ==========================================================================================
3596 * when holding pg and sched_scrub_lock, then the states are:
3598 * scrubber.reserved = true
3599 * scrub_rserved_peers includes whoami
3600 * osd->scrub_pending++
3601 * scheduling, replica declined:
3602 * scrubber.reserved = true
3603 * scrubber.reserved_peers includes -1
3604 * osd->scrub_pending++
3606 * scrubber.reserved = true
3607 * scrubber.reserved_peers.size() == acting.size();
3609 * osd->scrub_pending++
3611 * scrubber.reserved = false;
3612 * scrubber.reserved_peers empty
3613 * osd->scrubber.active++
3616 // returns true if a scrub has been newly kicked off
3617 bool PG::sched_scrub()
3619 bool nodeep_scrub
= false;
3620 assert(is_locked());
3621 if (!(is_primary() && is_active() && is_clean() && !is_scrubbing())) {
3625 double deep_scrub_interval
= 0;
3626 pool
.info
.opts
.get(pool_opts_t::DEEP_SCRUB_INTERVAL
, &deep_scrub_interval
);
3627 if (deep_scrub_interval
<= 0) {
3628 deep_scrub_interval
= cct
->_conf
->osd_deep_scrub_interval
;
3630 bool time_for_deep
= ceph_clock_now() >=
3631 info
.history
.last_deep_scrub_stamp
+ deep_scrub_interval
;
3633 bool deep_coin_flip
= false;
3634 // Only add random deep scrubs when NOT user initiated scrub
3635 if (!scrubber
.must_scrub
)
3636 deep_coin_flip
= (rand() % 100) < cct
->_conf
->osd_deep_scrub_randomize_ratio
* 100;
3637 dout(20) << __func__
<< ": time_for_deep=" << time_for_deep
<< " deep_coin_flip=" << deep_coin_flip
<< dendl
;
3639 time_for_deep
= (time_for_deep
|| deep_coin_flip
);
3641 //NODEEP_SCRUB so ignore time initiated deep-scrub
3642 if (osd
->osd
->get_osdmap()->test_flag(CEPH_OSDMAP_NODEEP_SCRUB
) ||
3643 pool
.info
.has_flag(pg_pool_t::FLAG_NODEEP_SCRUB
)) {
3644 time_for_deep
= false;
3645 nodeep_scrub
= true;
3648 if (!scrubber
.must_scrub
) {
3649 assert(!scrubber
.must_deep_scrub
);
3651 //NOSCRUB so skip regular scrubs
3652 if ((osd
->osd
->get_osdmap()->test_flag(CEPH_OSDMAP_NOSCRUB
) ||
3653 pool
.info
.has_flag(pg_pool_t::FLAG_NOSCRUB
)) && !time_for_deep
) {
3654 if (scrubber
.reserved
) {
3655 // cancel scrub if it is still in scheduling,
3656 // so pgs from other pools where scrub are still legal
3657 // have a chance to go ahead with scrubbing.
3658 clear_scrub_reserved();
3659 scrub_unreserve_replicas();
3665 if (cct
->_conf
->osd_scrub_auto_repair
3666 && get_pgbackend()->auto_repair_supported()
3668 // respect the command from user, and not do auto-repair
3669 && !scrubber
.must_repair
3670 && !scrubber
.must_scrub
3671 && !scrubber
.must_deep_scrub
) {
3672 dout(20) << __func__
<< ": auto repair with deep scrubbing" << dendl
;
3673 scrubber
.auto_repair
= true;
3675 // this happens when user issue the scrub/repair command during
3676 // the scheduling of the scrub/repair (e.g. request reservation)
3677 scrubber
.auto_repair
= false;
3681 if (!scrubber
.reserved
) {
3682 assert(scrubber
.reserved_peers
.empty());
3683 if ((cct
->_conf
->osd_scrub_during_recovery
|| !osd
->is_recovery_active()) &&
3684 osd
->inc_scrubs_pending()) {
3685 dout(20) << __func__
<< ": reserved locally, reserving replicas" << dendl
;
3686 scrubber
.reserved
= true;
3687 scrubber
.reserved_peers
.insert(pg_whoami
);
3688 scrub_reserve_replicas();
3690 dout(20) << __func__
<< ": failed to reserve locally" << dendl
;
3694 if (scrubber
.reserved
) {
3695 if (scrubber
.reserve_failed
) {
3696 dout(20) << "sched_scrub: failed, a peer declined" << dendl
;
3697 clear_scrub_reserved();
3698 scrub_unreserve_replicas();
3700 } else if (scrubber
.reserved_peers
.size() == acting
.size()) {
3701 dout(20) << "sched_scrub: success, reserved self and replicas" << dendl
;
3702 if (time_for_deep
) {
3703 dout(10) << "sched_scrub: scrub will be deep" << dendl
;
3704 state_set(PG_STATE_DEEP_SCRUB
);
3705 } else if (!scrubber
.must_deep_scrub
&& info
.stats
.stats
.sum
.num_deep_scrub_errors
) {
3706 if (!nodeep_scrub
) {
3707 osd
->clog
->info() << "osd." << osd
->whoami
3708 << " pg " << info
.pgid
3709 << " Deep scrub errors, upgrading scrub to deep-scrub";
3710 state_set(PG_STATE_DEEP_SCRUB
);
3711 } else if (!scrubber
.must_scrub
) {
3712 osd
->clog
->error() << "osd." << osd
->whoami
3713 << " pg " << info
.pgid
3714 << " Regular scrub skipped due to deep-scrub errors and nodeep-scrub set";
3715 clear_scrub_reserved();
3716 scrub_unreserve_replicas();
3719 osd
->clog
->error() << "osd." << osd
->whoami
3720 << " pg " << info
.pgid
3721 << " Regular scrub request, deep-scrub details will be lost";
3726 // none declined, since scrubber.reserved is set
3727 dout(20) << "sched_scrub: reserved " << scrubber
.reserved_peers
<< ", waiting for replicas" << dendl
;
3734 void PG::reg_next_scrub()
3740 if (scrubber
.must_scrub
||
3741 (info
.stats
.stats_invalid
&& cct
->_conf
->osd_scrub_invalid_stats
)) {
3742 reg_stamp
= ceph_clock_now();
3744 reg_stamp
= info
.history
.last_scrub_stamp
;
3746 // note down the sched_time, so we can locate this scrub, and remove it
3748 double scrub_min_interval
= 0, scrub_max_interval
= 0;
3749 pool
.info
.opts
.get(pool_opts_t::SCRUB_MIN_INTERVAL
, &scrub_min_interval
);
3750 pool
.info
.opts
.get(pool_opts_t::SCRUB_MAX_INTERVAL
, &scrub_max_interval
);
3751 assert(scrubber
.scrub_reg_stamp
== utime_t());
3752 scrubber
.scrub_reg_stamp
= osd
->reg_pg_scrub(info
.pgid
,
3756 scrubber
.must_scrub
);
3759 void PG::unreg_next_scrub()
3762 osd
->unreg_pg_scrub(info
.pgid
, scrubber
.scrub_reg_stamp
);
3763 scrubber
.scrub_reg_stamp
= utime_t();
3767 void PG::do_replica_scrub_map(OpRequestRef op
)
3769 const MOSDRepScrubMap
*m
= static_cast<const MOSDRepScrubMap
*>(op
->get_req());
3770 dout(7) << __func__
<< " " << *m
<< dendl
;
3771 if (m
->map_epoch
< info
.history
.same_interval_since
) {
3772 dout(10) << __func__
<< " discarding old from "
3773 << m
->map_epoch
<< " < " << info
.history
.same_interval_since
3777 if (!scrubber
.is_chunky_scrub_active()) {
3778 dout(10) << __func__
<< " scrub isn't active" << dendl
;
3784 bufferlist::iterator p
= const_cast<bufferlist
&>(m
->get_data()).begin();
3785 scrubber
.received_maps
[m
->from
].decode(p
, info
.pgid
.pool());
3786 dout(10) << "map version is "
3787 << scrubber
.received_maps
[m
->from
].valid_through
3790 --scrubber
.waiting_on
;
3791 scrubber
.waiting_on_whom
.erase(m
->from
);
3792 if (scrubber
.waiting_on
== 0) {
3793 if (ops_blocked_by_scrub()) {
3794 requeue_scrub(true);
3796 requeue_scrub(false);
3801 void PG::sub_op_scrub_map(OpRequestRef op
)
3803 // for legacy jewel compatibility only
3804 const MOSDSubOp
*m
= static_cast<const MOSDSubOp
*>(op
->get_req());
3805 assert(m
->get_type() == MSG_OSD_SUBOP
);
3806 dout(7) << "sub_op_scrub_map" << dendl
;
3808 if (m
->map_epoch
< info
.history
.same_interval_since
) {
3809 dout(10) << "sub_op_scrub discarding old sub_op from "
3810 << m
->map_epoch
<< " < " << info
.history
.same_interval_since
<< dendl
;
3814 if (!scrubber
.is_chunky_scrub_active()) {
3815 dout(10) << "sub_op_scrub_map scrub isn't active" << dendl
;
3821 dout(10) << " got " << m
->from
<< " scrub map" << dendl
;
3822 bufferlist::iterator p
= const_cast<bufferlist
&>(m
->get_data()).begin();
3824 scrubber
.received_maps
[m
->from
].decode(p
, info
.pgid
.pool());
3825 dout(10) << "map version is "
3826 << scrubber
.received_maps
[m
->from
].valid_through
3829 --scrubber
.waiting_on
;
3830 scrubber
.waiting_on_whom
.erase(m
->from
);
3832 if (scrubber
.waiting_on
== 0) {
3833 if (ops_blocked_by_scrub()) {
3834 requeue_scrub(true);
3836 requeue_scrub(false);
3841 // send scrub v3 messages (chunky scrub)
3842 void PG::_request_scrub_map(
3843 pg_shard_t replica
, eversion_t version
,
3844 hobject_t start
, hobject_t end
,
3845 bool deep
, uint32_t seed
)
3847 assert(replica
!= pg_whoami
);
3848 dout(10) << "scrub requesting scrubmap from osd." << replica
3849 << " deep " << (int)deep
<< " seed " << seed
<< dendl
;
3850 MOSDRepScrub
*repscrubop
= new MOSDRepScrub(
3851 spg_t(info
.pgid
.pgid
, replica
.shard
), version
,
3852 get_osdmap()->get_epoch(),
3853 get_last_peering_reset(),
3854 start
, end
, deep
, seed
);
3855 // default priority, we want the rep scrub processed prior to any recovery
3856 // or client io messages (we are holding a lock!)
3857 osd
->send_message_osd_cluster(
3858 replica
.osd
, repscrubop
, get_osdmap()->get_epoch());
3861 void PG::handle_scrub_reserve_request(OpRequestRef op
)
3863 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
3865 if (scrubber
.reserved
) {
3866 dout(10) << __func__
<< " ignoring reserve request: Already reserved"
3870 if ((cct
->_conf
->osd_scrub_during_recovery
|| !osd
->is_recovery_active()) &&
3871 osd
->inc_scrubs_pending()) {
3872 scrubber
.reserved
= true;
3874 dout(20) << __func__
<< ": failed to reserve remotely" << dendl
;
3875 scrubber
.reserved
= false;
3877 if (op
->get_req()->get_type() == MSG_OSD_SCRUB_RESERVE
) {
3878 const MOSDScrubReserve
*m
=
3879 static_cast<const MOSDScrubReserve
*>(op
->get_req());
3880 Message
*reply
= new MOSDScrubReserve(
3881 spg_t(info
.pgid
.pgid
, primary
.shard
),
3883 scrubber
.reserved
? MOSDScrubReserve::GRANT
: MOSDScrubReserve::REJECT
,
3885 osd
->send_message_osd_cluster(reply
, op
->get_req()->get_connection());
3887 // for jewel compat only
3888 const MOSDSubOp
*req
= static_cast<const MOSDSubOp
*>(op
->get_req());
3889 assert(req
->get_type() == MSG_OSD_SUBOP
);
3890 MOSDSubOpReply
*reply
= new MOSDSubOpReply(
3891 req
, pg_whoami
, 0, get_osdmap()->get_epoch(), CEPH_OSD_FLAG_ACK
);
3892 ::encode(scrubber
.reserved
, reply
->get_data());
3893 osd
->send_message_osd_cluster(reply
, op
->get_req()->get_connection());
3897 void PG::handle_scrub_reserve_grant(OpRequestRef op
, pg_shard_t from
)
3899 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
3901 if (!scrubber
.reserved
) {
3902 dout(10) << "ignoring obsolete scrub reserve reply" << dendl
;
3905 if (scrubber
.reserved_peers
.find(from
) != scrubber
.reserved_peers
.end()) {
3906 dout(10) << " already had osd." << from
<< " reserved" << dendl
;
3908 dout(10) << " osd." << from
<< " scrub reserve = success" << dendl
;
3909 scrubber
.reserved_peers
.insert(from
);
3914 void PG::handle_scrub_reserve_reject(OpRequestRef op
, pg_shard_t from
)
3916 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
3918 if (!scrubber
.reserved
) {
3919 dout(10) << "ignoring obsolete scrub reserve reply" << dendl
;
3922 if (scrubber
.reserved_peers
.find(from
) != scrubber
.reserved_peers
.end()) {
3923 dout(10) << " already had osd." << from
<< " reserved" << dendl
;
3925 /* One decline stops this pg from being scheduled for scrubbing. */
3926 dout(10) << " osd." << from
<< " scrub reserve = fail" << dendl
;
3927 scrubber
.reserve_failed
= true;
3932 void PG::handle_scrub_reserve_release(OpRequestRef op
)
3934 dout(7) << __func__
<< " " << *op
->get_req() << dendl
;
3936 clear_scrub_reserved();
3939 void PG::reject_reservation()
3941 osd
->send_message_osd_cluster(
3943 new MBackfillReserve(
3944 MBackfillReserve::REJECT
,
3945 spg_t(info
.pgid
.pgid
, primary
.shard
),
3946 get_osdmap()->get_epoch()),
3947 get_osdmap()->get_epoch());
3950 void PG::schedule_backfill_retry(float delay
)
3952 Mutex::Locker
lock(osd
->recovery_request_lock
);
3953 osd
->recovery_request_timer
.add_event_after(
3955 new QueuePeeringEvt
<RequestBackfill
>(
3956 this, get_osdmap()->get_epoch(),
3957 RequestBackfill()));
3960 void PG::schedule_recovery_retry(float delay
)
3962 Mutex::Locker
lock(osd
->recovery_request_lock
);
3963 osd
->recovery_request_timer
.add_event_after(
3965 new QueuePeeringEvt
<DoRecovery
>(
3966 this, get_osdmap()->get_epoch(),
3970 void PG::clear_scrub_reserved()
3972 scrubber
.reserved_peers
.clear();
3973 scrubber
.reserve_failed
= false;
3975 if (scrubber
.reserved
) {
3976 scrubber
.reserved
= false;
3977 osd
->dec_scrubs_pending();
3981 void PG::scrub_reserve_replicas()
3983 assert(backfill_targets
.empty());
3984 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
3985 i
!= actingbackfill
.end();
3987 if (*i
== pg_whoami
) continue;
3988 dout(10) << "scrub requesting reserve from osd." << *i
<< dendl
;
3989 if (HAVE_FEATURE(get_min_acting_features(), SERVER_LUMINOUS
)) {
3990 osd
->send_message_osd_cluster(
3992 new MOSDScrubReserve(spg_t(info
.pgid
.pgid
, i
->shard
),
3993 get_osdmap()->get_epoch(),
3994 MOSDScrubReserve::REQUEST
, pg_whoami
),
3995 get_osdmap()->get_epoch());
3997 // for jewel compat only
3998 vector
<OSDOp
> scrub(1);
3999 scrub
[0].op
.op
= CEPH_OSD_OP_SCRUB_RESERVE
;
4003 MOSDSubOp
*subop
= new MOSDSubOp(
4004 reqid
, pg_whoami
, spg_t(info
.pgid
.pgid
, i
->shard
), poid
, 0,
4005 get_osdmap()->get_epoch(), osd
->get_tid(), v
);
4007 osd
->send_message_osd_cluster(
4008 i
->osd
, subop
, get_osdmap()->get_epoch());
4013 void PG::scrub_unreserve_replicas()
4015 assert(backfill_targets
.empty());
4016 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
4017 i
!= actingbackfill
.end();
4019 if (*i
== pg_whoami
) continue;
4020 dout(10) << "scrub requesting unreserve from osd." << *i
<< dendl
;
4021 if (HAVE_FEATURE(get_min_acting_features(), SERVER_LUMINOUS
)) {
4022 osd
->send_message_osd_cluster(
4024 new MOSDScrubReserve(spg_t(info
.pgid
.pgid
, i
->shard
),
4025 get_osdmap()->get_epoch(),
4026 MOSDScrubReserve::RELEASE
, pg_whoami
),
4027 get_osdmap()->get_epoch());
4029 // for jewel compat only
4030 vector
<OSDOp
> scrub(1);
4031 scrub
[0].op
.op
= CEPH_OSD_OP_SCRUB_UNRESERVE
;
4035 MOSDSubOp
*subop
= new MOSDSubOp(
4036 reqid
, pg_whoami
, spg_t(info
.pgid
.pgid
, i
->shard
), poid
, 0,
4037 get_osdmap()->get_epoch(), osd
->get_tid(), v
);
4039 osd
->send_message_osd_cluster(i
->osd
, subop
, get_osdmap()->get_epoch());
4044 void PG::_scan_rollback_obs(
4045 const vector
<ghobject_t
> &rollback_obs
,
4046 ThreadPool::TPHandle
&handle
)
4048 ObjectStore::Transaction t
;
4049 eversion_t trimmed_to
= last_rollback_info_trimmed_to_applied
;
4050 for (vector
<ghobject_t
>::const_iterator i
= rollback_obs
.begin();
4051 i
!= rollback_obs
.end();
4053 if (i
->generation
< trimmed_to
.version
) {
4054 osd
->clog
->error() << "osd." << osd
->whoami
4055 << " pg " << info
.pgid
4056 << " found obsolete rollback obj "
4057 << *i
<< " generation < trimmed_to "
4064 derr
<< __func__
<< ": queueing trans to clean up obsolete rollback objs"
4066 osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
4070 void PG::_scan_snaps(ScrubMap
&smap
)
4074 for (map
<hobject_t
, ScrubMap::object
>::reverse_iterator i
= smap
.objects
.rbegin();
4075 i
!= smap
.objects
.rend();
4077 const hobject_t
&hoid
= i
->first
;
4078 ScrubMap::object
&o
= i
->second
;
4080 if (hoid
.is_head() || hoid
.is_snapdir()) {
4081 // parse the SnapSet
4083 if (o
.attrs
.find(SS_ATTR
) == o
.attrs
.end()) {
4086 bl
.push_back(o
.attrs
[SS_ATTR
]);
4087 auto p
= bl
.begin();
4089 ::decode(snapset
, p
);
4093 head
= hoid
.get_head();
4094 // Make sure head_exists is correct for is_legacy() check
4096 snapset
.head_exists
= true;
4099 if (hoid
.snap
< CEPH_MAXSNAP
) {
4100 // check and if necessary fix snap_mapper
4101 if (hoid
.get_head() != head
) {
4102 derr
<< __func__
<< " no head for " << hoid
<< " (have " << head
<< ")"
4106 set
<snapid_t
> obj_snaps
;
4107 if (!snapset
.is_legacy()) {
4108 auto p
= snapset
.clone_snaps
.find(hoid
.snap
);
4109 if (p
== snapset
.clone_snaps
.end()) {
4110 derr
<< __func__
<< " no clone_snaps for " << hoid
<< " in " << snapset
4114 obj_snaps
.insert(p
->second
.begin(), p
->second
.end());
4117 if (o
.attrs
.find(OI_ATTR
) == o
.attrs
.end()) {
4120 bl
.push_back(o
.attrs
[OI_ATTR
]);
4127 obj_snaps
.insert(oi
.legacy_snaps
.begin(), oi
.legacy_snaps
.end());
4129 set
<snapid_t
> cur_snaps
;
4130 int r
= snap_mapper
.get_snaps(hoid
, &cur_snaps
);
4131 if (r
!= 0 && r
!= -ENOENT
) {
4132 derr
<< __func__
<< ": get_snaps returned " << cpp_strerror(r
) << dendl
;
4135 if (r
== -ENOENT
|| cur_snaps
!= obj_snaps
) {
4136 ObjectStore::Transaction t
;
4137 OSDriver::OSTransaction
_t(osdriver
.get_transaction(&t
));
4139 r
= snap_mapper
.remove_oid(hoid
, &_t
);
4141 derr
<< __func__
<< ": remove_oid returned " << cpp_strerror(r
)
4145 osd
->clog
->error() << "osd." << osd
->whoami
4146 << " found snap mapper error on pg "
4148 << " oid " << hoid
<< " snaps in mapper: "
4149 << cur_snaps
<< ", oi: "
4153 osd
->clog
->error() << "osd." << osd
->whoami
4154 << " found snap mapper error on pg "
4156 << " oid " << hoid
<< " snaps missing in mapper"
4161 snap_mapper
.add_oid(hoid
, obj_snaps
, &_t
);
4162 r
= osd
->store
->apply_transaction(osr
.get(), std::move(t
));
4164 derr
<< __func__
<< ": apply_transaction got " << cpp_strerror(r
)
4172 void PG::_repair_oinfo_oid(ScrubMap
&smap
)
4174 for (map
<hobject_t
, ScrubMap::object
>::reverse_iterator i
= smap
.objects
.rbegin();
4175 i
!= smap
.objects
.rend();
4177 const hobject_t
&hoid
= i
->first
;
4178 ScrubMap::object
&o
= i
->second
;
4181 if (o
.attrs
.find(OI_ATTR
) == o
.attrs
.end()) {
4184 bl
.push_back(o
.attrs
[OI_ATTR
]);
4191 if (oi
.soid
!= hoid
) {
4192 ObjectStore::Transaction t
;
4193 OSDriver::OSTransaction
_t(osdriver
.get_transaction(&t
));
4194 osd
->clog
->error() << "osd." << osd
->whoami
4195 << " found object info error on pg "
4197 << " oid " << hoid
<< " oid in object info: "
4203 ::encode(oi
, bl
, get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
4205 bufferptr
bp(bl
.c_str(), bl
.length());
4206 o
.attrs
[OI_ATTR
] = bp
;
4208 t
.setattr(coll
, ghobject_t(hoid
), OI_ATTR
, bl
);
4209 int r
= osd
->store
->apply_transaction(osr
.get(), std::move(t
));
4211 derr
<< __func__
<< ": apply_transaction got " << cpp_strerror(r
)
4219 * build a scrub map over a chunk without releasing the lock
4220 * only used by chunky scrub
4222 int PG::build_scrub_map_chunk(
4224 hobject_t start
, hobject_t end
, bool deep
, uint32_t seed
,
4225 ThreadPool::TPHandle
&handle
)
4227 dout(10) << __func__
<< " [" << start
<< "," << end
<< ") "
4228 << " seed " << seed
<< dendl
;
4230 map
.valid_through
= info
.last_update
;
4233 vector
<hobject_t
> ls
;
4234 vector
<ghobject_t
> rollback_obs
;
4235 int ret
= get_pgbackend()->objects_list_range(
4242 dout(5) << "objects_list_range error: " << ret
<< dendl
;
4247 get_pgbackend()->be_scan_list(map
, ls
, deep
, seed
, handle
);
4248 _scan_rollback_obs(rollback_obs
, handle
);
4250 _repair_oinfo_oid(map
);
4252 dout(20) << __func__
<< " done" << dendl
;
4256 void PG::Scrubber::cleanup_store(ObjectStore::Transaction
*t
) {
4259 struct OnComplete
: Context
{
4260 std::unique_ptr
<Scrub::Store
> store
;
4262 std::unique_ptr
<Scrub::Store
> &&store
)
4263 : store(std::move(store
)) {}
4264 void finish(int) override
{}
4267 t
->register_on_complete(new OnComplete(std::move(store
)));
4271 void PG::repair_object(
4272 const hobject_t
& soid
, list
<pair
<ScrubMap::object
, pg_shard_t
> > *ok_peers
,
4273 pg_shard_t bad_peer
)
4275 list
<pg_shard_t
> op_shards
;
4276 for (auto i
: *ok_peers
) {
4277 op_shards
.push_back(i
.second
);
4279 dout(10) << "repair_object " << soid
<< " bad_peer osd."
4280 << bad_peer
<< " ok_peers osd.{" << op_shards
<< "}" << dendl
;
4281 ScrubMap::object
&po
= ok_peers
->back().first
;
4284 bv
.push_back(po
.attrs
[OI_ATTR
]);
4287 bufferlist::iterator bliter
= bv
.begin();
4288 ::decode(oi
, bliter
);
4290 dout(0) << __func__
<< ": Need version of replica, bad object_info_t: " << soid
<< dendl
;
4293 if (bad_peer
!= primary
) {
4294 peer_missing
[bad_peer
].add(soid
, oi
.version
, eversion_t(), false);
4296 // We should only be scrubbing if the PG is clean.
4297 assert(waiting_for_unreadable_object
.empty());
4299 pg_log
.missing_add(soid
, oi
.version
, eversion_t());
4301 pg_log
.set_last_requested(0);
4302 dout(10) << __func__
<< ": primary = " << primary
<< dendl
;
4305 if (is_ec_pg() || bad_peer
== primary
) {
4306 // we'd better collect all shard for EC pg, and prepare good peers as the
4307 // source of pull in the case of replicated pg.
4308 missing_loc
.add_missing(soid
, oi
.version
, eversion_t());
4309 list
<pair
<ScrubMap::object
, pg_shard_t
> >::iterator i
;
4310 for (i
= ok_peers
->begin();
4311 i
!= ok_peers
->end();
4313 missing_loc
.add_location(soid
, i
->second
);
4319 * Wait for last_update_applied to match msg->scrub_to as above. Wait
4320 * for pushes to complete in case of recent recovery. Build a single
4321 * scrubmap of objects that are in the range [msg->start, msg->end).
4323 void PG::replica_scrub(
4325 ThreadPool::TPHandle
&handle
)
4327 const MOSDRepScrub
*msg
= static_cast<const MOSDRepScrub
*>(op
->get_req());
4328 assert(!scrubber
.active_rep_scrub
);
4329 dout(7) << "replica_scrub" << dendl
;
4331 if (msg
->map_epoch
< info
.history
.same_interval_since
) {
4332 dout(10) << "replica_scrub discarding old replica_scrub from "
4333 << msg
->map_epoch
<< " < " << info
.history
.same_interval_since
4340 assert(msg
->chunky
);
4341 if (last_update_applied
< msg
->scrub_to
) {
4342 dout(10) << "waiting for last_update_applied to catch up" << dendl
;
4343 scrubber
.active_rep_scrub
= op
;
4347 if (active_pushes
> 0) {
4348 dout(10) << "waiting for active pushes to finish" << dendl
;
4349 scrubber
.active_rep_scrub
= op
;
4353 // compensate for hobject_t's with wrong pool from sloppy hammer OSDs
4354 hobject_t start
= msg
->start
;
4355 hobject_t end
= msg
->end
;
4356 if (!start
.is_max())
4357 start
.pool
= info
.pgid
.pool();
4359 end
.pool
= info
.pgid
.pool();
4361 build_scrub_map_chunk(
4362 map
, start
, end
, msg
->deep
, msg
->seed
,
4365 if (HAVE_FEATURE(acting_features
, SERVER_LUMINOUS
)) {
4366 MOSDRepScrubMap
*reply
= new MOSDRepScrubMap(
4367 spg_t(info
.pgid
.pgid
, get_primary().shard
),
4370 ::encode(map
, reply
->get_data());
4371 osd
->send_message_osd_cluster(reply
, msg
->get_connection());
4373 // for jewel compatibility
4374 vector
<OSDOp
> scrub(1);
4375 scrub
[0].op
.op
= CEPH_OSD_OP_SCRUB_MAP
;
4379 MOSDSubOp
*subop
= new MOSDSubOp(
4382 spg_t(info
.pgid
.pgid
, get_primary().shard
),
4388 ::encode(map
, subop
->get_data());
4390 osd
->send_message_osd_cluster(subop
, msg
->get_connection());
4395 * PG_STATE_SCRUBBING is set when the scrub is queued
4397 * scrub will be chunky if all OSDs in PG support chunky scrub
4398 * scrub will fail if OSDs are too old.
4400 void PG::scrub(epoch_t queued
, ThreadPool::TPHandle
&handle
)
4402 if (cct
->_conf
->osd_scrub_sleep
> 0 &&
4403 (scrubber
.state
== PG::Scrubber::NEW_CHUNK
||
4404 scrubber
.state
== PG::Scrubber::INACTIVE
) &&
4405 scrubber
.needs_sleep
) {
4406 ceph_assert(!scrubber
.sleeping
);
4407 dout(20) << __func__
<< " state is INACTIVE|NEW_CHUNK, sleeping" << dendl
;
4409 // Do an async sleep so we don't block the op queue
4410 OSDService
*osds
= osd
;
4411 spg_t pgid
= get_pgid();
4412 int state
= scrubber
.state
;
4413 auto scrub_requeue_callback
=
4414 new FunctionContext([osds
, pgid
, state
](int r
) {
4415 PG
*pg
= osds
->osd
->lookup_lock_pg(pgid
);
4416 if (pg
== nullptr) {
4417 lgeneric_dout(osds
->osd
->cct
, 20)
4418 << "scrub_requeue_callback: Could not find "
4419 << "PG " << pgid
<< " can't complete scrub requeue after sleep"
4423 pg
->scrubber
.sleeping
= false;
4424 pg
->scrubber
.needs_sleep
= false;
4425 lgeneric_dout(pg
->cct
, 20)
4426 << "scrub_requeue_callback: slept for "
4427 << ceph_clock_now() - pg
->scrubber
.sleep_start
4428 << ", re-queuing scrub with state " << state
<< dendl
;
4429 pg
->scrub_queued
= false;
4430 pg
->requeue_scrub();
4431 pg
->scrubber
.sleep_start
= utime_t();
4434 Mutex::Locker
l(osd
->scrub_sleep_lock
);
4435 osd
->scrub_sleep_timer
.add_event_after(cct
->_conf
->osd_scrub_sleep
,
4436 scrub_requeue_callback
);
4437 scrubber
.sleeping
= true;
4438 scrubber
.sleep_start
= ceph_clock_now();
4441 if (pg_has_reset_since(queued
)) {
4444 assert(scrub_queued
);
4445 scrub_queued
= false;
4446 scrubber
.needs_sleep
= true;
4448 if (!is_primary() || !is_active() || !is_clean() || !is_scrubbing()) {
4449 dout(10) << "scrub -- not primary or active or not clean" << dendl
;
4450 state_clear(PG_STATE_SCRUBBING
);
4451 state_clear(PG_STATE_REPAIR
);
4452 state_clear(PG_STATE_DEEP_SCRUB
);
4453 publish_stats_to_osd();
4457 if (!scrubber
.active
) {
4458 assert(backfill_targets
.empty());
4460 scrubber
.deep
= state_test(PG_STATE_DEEP_SCRUB
);
4462 dout(10) << "starting a new chunky scrub" << dendl
;
4465 chunky_scrub(handle
);
4469 * Chunky scrub scrubs objects one chunk at a time with writes blocked for that
4472 * The object store is partitioned into chunks which end on hash boundaries. For
4473 * each chunk, the following logic is performed:
4475 * (1) Block writes on the chunk
4476 * (2) Request maps from replicas
4477 * (3) Wait for pushes to be applied (after recovery)
4478 * (4) Wait for writes to flush on the chunk
4479 * (5) Wait for maps from replicas
4480 * (6) Compare / repair all scrub maps
4481 * (7) Wait for digest updates to apply
4483 * This logic is encoded in the mostly linear state machine:
4485 * +------------------+
4486 * _________v__________ |
4489 * |____________________| |
4492 * _________v___v______ | |
4495 * |____________________| | |
4497 * _________v__________ | |
4499 * | WAIT_PUSHES | | |
4500 * |____________________| | |
4502 * _________v__________ | |
4504 * | WAIT_LAST_UPDATE | | |
4505 * |____________________| | |
4507 * _________v__________ | |
4510 * |____________________| | |
4512 * _________v__________ | |
4514 * | WAIT_REPLICAS | | |
4515 * |____________________| | |
4517 * _________v__________ | |
4519 * | COMPARE_MAPS | | |
4520 * |____________________| | |
4523 * _________v__________ | |
4525 * |WAIT_DIGEST_UPDATES | | |
4526 * |____________________| | |
4529 * _________v__________ |
4532 * |____________________| |
4534 * +------------------+
4536 * The primary determines the last update from the subset by walking the log. If
4537 * it sees a log entry pertaining to a file in the chunk, it tells the replicas
4538 * to wait until that update is applied before building a scrub map. Both the
4539 * primary and replicas will wait for any active pushes to be applied.
4541 * In contrast to classic_scrub, chunky_scrub is entirely handled by scrub_wq.
4543 * scrubber.state encodes the current state of the scrub (refer to state diagram
4546 void PG::chunky_scrub(ThreadPool::TPHandle
&handle
)
4548 // check for map changes
4549 if (scrubber
.is_chunky_scrub_active()) {
4550 if (scrubber
.epoch_start
!= info
.history
.same_interval_since
) {
4551 dout(10) << "scrub pg changed, aborting" << dendl
;
4552 scrub_clear_state();
4553 scrub_unreserve_replicas();
4562 dout(20) << "scrub state " << Scrubber::state_string(scrubber
.state
)
4563 << " [" << scrubber
.start
<< "," << scrubber
.end
<< ")" << dendl
;
4565 switch (scrubber
.state
) {
4566 case PG::Scrubber::INACTIVE
:
4567 dout(10) << "scrub start" << dendl
;
4569 publish_stats_to_osd();
4570 scrubber
.epoch_start
= info
.history
.same_interval_since
;
4571 scrubber
.active
= true;
4573 osd
->inc_scrubs_active(scrubber
.reserved
);
4574 if (scrubber
.reserved
) {
4575 scrubber
.reserved
= false;
4576 scrubber
.reserved_peers
.clear();
4580 ObjectStore::Transaction t
;
4581 scrubber
.cleanup_store(&t
);
4582 scrubber
.store
.reset(Scrub::Store::create(osd
->store
, &t
,
4584 osd
->store
->queue_transaction(osr
.get(), std::move(t
), nullptr);
4587 // Don't include temporary objects when scrubbing
4588 scrubber
.start
= info
.pgid
.pgid
.get_hobj_start();
4589 scrubber
.state
= PG::Scrubber::NEW_CHUNK
;
4592 bool repair
= state_test(PG_STATE_REPAIR
);
4593 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
4594 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
4596 oss
<< info
.pgid
.pgid
<< " " << mode
<< " starts" << std::endl
;
4597 osd
->clog
->debug(oss
);
4604 case PG::Scrubber::NEW_CHUNK
:
4605 scrubber
.primary_scrubmap
= ScrubMap();
4606 scrubber
.received_maps
.clear();
4609 /* get the start and end of our scrub chunk
4611 * Our scrub chunk has an important restriction we're going to need to
4612 * respect. We can't let head or snapdir be start or end.
4613 * Using a half-open interval means that if end == head|snapdir,
4614 * we'd scrub/lock head and the clone right next to head in different
4615 * chunks which would allow us to miss clones created between
4616 * scrubbing that chunk and scrubbing the chunk including head.
4617 * This isn't true for any of the other clones since clones can
4618 * only be created "just to the left of" head. There is one exception
4619 * to this: promotion of clones which always happens to the left of the
4620 * left-most clone, but promote_object checks the scrubber in that
4621 * case, so it should be ok. Also, it's ok to "miss" clones at the
4622 * left end of the range if we are a tier because they may legitimately
4623 * not exist (see _scrub).
4625 int min
= MAX(3, cct
->_conf
->osd_scrub_chunk_min
);
4626 hobject_t start
= scrubber
.start
;
4627 hobject_t candidate_end
;
4628 vector
<hobject_t
> objects
;
4629 ret
= get_pgbackend()->objects_list_partial(
4632 MAX(min
, cct
->_conf
->osd_scrub_chunk_max
),
4637 if (!objects
.empty()) {
4638 hobject_t back
= objects
.back();
4639 while (candidate_end
.has_snapset() &&
4640 candidate_end
.get_head() == back
.get_head()) {
4641 candidate_end
= back
;
4643 if (objects
.empty()) {
4645 "Somehow we got more than 2 objects which"
4646 "have the same head but are not clones");
4648 back
= objects
.back();
4650 if (candidate_end
.has_snapset()) {
4651 assert(candidate_end
.get_head() != back
.get_head());
4652 candidate_end
= candidate_end
.get_object_boundary();
4655 assert(candidate_end
.is_max());
4658 if (!_range_available_for_scrub(scrubber
.start
, candidate_end
)) {
4659 // we'll be requeued by whatever made us unavailable for scrub
4660 dout(10) << __func__
<< ": scrub blocked somewhere in range "
4661 << "[" << scrubber
.start
<< ", " << candidate_end
<< ")"
4666 scrubber
.end
= candidate_end
;
4669 // walk the log to find the latest update that affects our chunk
4670 scrubber
.subset_last_update
= eversion_t();
4671 for (auto p
= projected_log
.log
.rbegin();
4672 p
!= projected_log
.log
.rend();
4674 if (p
->soid
>= scrubber
.start
&&
4675 p
->soid
< scrubber
.end
) {
4676 scrubber
.subset_last_update
= p
->version
;
4680 if (scrubber
.subset_last_update
== eversion_t()) {
4681 for (list
<pg_log_entry_t
>::const_reverse_iterator p
=
4682 pg_log
.get_log().log
.rbegin();
4683 p
!= pg_log
.get_log().log
.rend();
4685 if (p
->soid
>= scrubber
.start
&&
4686 p
->soid
< scrubber
.end
) {
4687 scrubber
.subset_last_update
= p
->version
;
4693 // ask replicas to wait until
4694 // last_update_applied >= scrubber.subset_last_update and then scan
4695 scrubber
.waiting_on_whom
.insert(pg_whoami
);
4696 ++scrubber
.waiting_on
;
4698 // request maps from replicas
4699 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
4700 i
!= actingbackfill
.end();
4702 if (*i
== pg_whoami
) continue;
4703 _request_scrub_map(*i
, scrubber
.subset_last_update
,
4704 scrubber
.start
, scrubber
.end
, scrubber
.deep
,
4706 scrubber
.waiting_on_whom
.insert(*i
);
4707 ++scrubber
.waiting_on
;
4710 scrubber
.state
= PG::Scrubber::WAIT_PUSHES
;
4714 case PG::Scrubber::WAIT_PUSHES
:
4715 if (active_pushes
== 0) {
4716 scrubber
.state
= PG::Scrubber::WAIT_LAST_UPDATE
;
4718 dout(15) << "wait for pushes to apply" << dendl
;
4723 case PG::Scrubber::WAIT_LAST_UPDATE
:
4724 if (last_update_applied
>= scrubber
.subset_last_update
) {
4725 scrubber
.state
= PG::Scrubber::BUILD_MAP
;
4727 // will be requeued by op_applied
4728 dout(15) << "wait for writes to flush" << dendl
;
4733 case PG::Scrubber::BUILD_MAP
:
4734 assert(last_update_applied
>= scrubber
.subset_last_update
);
4736 // build my own scrub map
4737 ret
= build_scrub_map_chunk(scrubber
.primary_scrubmap
,
4738 scrubber
.start
, scrubber
.end
,
4739 scrubber
.deep
, scrubber
.seed
,
4742 dout(5) << "error building scrub map: " << ret
<< ", aborting" << dendl
;
4743 scrub_clear_state();
4744 scrub_unreserve_replicas();
4748 --scrubber
.waiting_on
;
4749 scrubber
.waiting_on_whom
.erase(pg_whoami
);
4751 scrubber
.state
= PG::Scrubber::WAIT_REPLICAS
;
4754 case PG::Scrubber::WAIT_REPLICAS
:
4755 if (scrubber
.waiting_on
> 0) {
4756 // will be requeued by sub_op_scrub_map
4757 dout(10) << "wait for replicas to build scrub map" << dendl
;
4760 scrubber
.state
= PG::Scrubber::COMPARE_MAPS
;
4764 case PG::Scrubber::COMPARE_MAPS
:
4765 assert(last_update_applied
>= scrubber
.subset_last_update
);
4766 assert(scrubber
.waiting_on
== 0);
4768 scrub_compare_maps();
4769 scrubber
.start
= scrubber
.end
;
4770 scrubber
.run_callbacks();
4772 // requeue the writes from the chunk that just finished
4773 requeue_ops(waiting_for_scrub
);
4775 scrubber
.state
= PG::Scrubber::WAIT_DIGEST_UPDATES
;
4779 case PG::Scrubber::WAIT_DIGEST_UPDATES
:
4780 if (scrubber
.num_digest_updates_pending
) {
4781 dout(10) << __func__
<< " waiting on "
4782 << scrubber
.num_digest_updates_pending
4783 << " digest updates" << dendl
;
4788 if (!(scrubber
.end
.is_max())) {
4789 scrubber
.state
= PG::Scrubber::NEW_CHUNK
;
4793 scrubber
.state
= PG::Scrubber::FINISH
;
4798 case PG::Scrubber::FINISH
:
4800 scrubber
.state
= PG::Scrubber::INACTIVE
;
4803 if (!snap_trimq
.empty()) {
4804 dout(10) << "scrub finished, requeuing snap_trimmer" << dendl
;
4805 snap_trimmer_scrub_complete();
4814 dout(20) << "scrub final state " << Scrubber::state_string(scrubber
.state
)
4815 << " [" << scrubber
.start
<< "," << scrubber
.end
<< ")" << dendl
;
4818 void PG::scrub_clear_state()
4820 assert(is_locked());
4821 state_clear(PG_STATE_SCRUBBING
);
4822 state_clear(PG_STATE_REPAIR
);
4823 state_clear(PG_STATE_DEEP_SCRUB
);
4824 publish_stats_to_osd();
4826 // active -> nothing.
4827 if (scrubber
.active
)
4828 osd
->dec_scrubs_active();
4830 requeue_ops(waiting_for_scrub
);
4834 // type-specific state clear
4835 _scrub_clear_state();
4838 void PG::scrub_compare_maps()
4840 dout(10) << __func__
<< " has maps, analyzing" << dendl
;
4842 // construct authoritative scrub map for type specific scrubbing
4843 scrubber
.cleaned_meta_map
.insert(scrubber
.primary_scrubmap
);
4844 map
<hobject_t
, pair
<uint32_t, uint32_t>> missing_digest
;
4846 if (acting
.size() > 1) {
4847 dout(10) << __func__
<< " comparing replica scrub maps" << dendl
;
4851 // Map from object with errors to good peer
4852 map
<hobject_t
, list
<pg_shard_t
>> authoritative
;
4853 map
<pg_shard_t
, ScrubMap
*> maps
;
4855 dout(2) << __func__
<< " osd." << acting
[0] << " has "
4856 << scrubber
.primary_scrubmap
.objects
.size() << " items" << dendl
;
4857 maps
[pg_whoami
] = &scrubber
.primary_scrubmap
;
4859 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
4860 i
!= actingbackfill
.end();
4862 if (*i
== pg_whoami
) continue;
4863 dout(2) << __func__
<< " replica " << *i
<< " has "
4864 << scrubber
.received_maps
[*i
].objects
.size()
4865 << " items" << dendl
;
4866 maps
[*i
] = &scrubber
.received_maps
[*i
];
4869 get_pgbackend()->be_compare_scrubmaps(
4871 state_test(PG_STATE_REPAIR
),
4873 scrubber
.inconsistent
,
4876 scrubber
.shallow_errors
,
4877 scrubber
.deep_errors
,
4878 scrubber
.store
.get(),
4881 dout(2) << ss
.str() << dendl
;
4883 if (!ss
.str().empty()) {
4884 osd
->clog
->error(ss
);
4887 for (map
<hobject_t
, list
<pg_shard_t
>>::iterator i
= authoritative
.begin();
4888 i
!= authoritative
.end();
4890 list
<pair
<ScrubMap::object
, pg_shard_t
> > good_peers
;
4891 for (list
<pg_shard_t
>::const_iterator j
= i
->second
.begin();
4892 j
!= i
->second
.end();
4894 good_peers
.push_back(make_pair(maps
[*j
]->objects
[i
->first
], *j
));
4896 scrubber
.authoritative
.insert(
4902 for (map
<hobject_t
, list
<pg_shard_t
>>::iterator i
= authoritative
.begin();
4903 i
!= authoritative
.end();
4905 scrubber
.cleaned_meta_map
.objects
.erase(i
->first
);
4906 scrubber
.cleaned_meta_map
.objects
.insert(
4907 *(maps
[i
->second
.back()]->objects
.find(i
->first
))
4912 ScrubMap for_meta_scrub
;
4913 if (scrubber
.end
.is_max() ||
4914 scrubber
.cleaned_meta_map
.objects
.empty()) {
4915 scrubber
.cleaned_meta_map
.swap(for_meta_scrub
);
4917 auto iter
= scrubber
.cleaned_meta_map
.objects
.end();
4918 --iter
; // not empty, see if clause
4919 auto begin
= scrubber
.cleaned_meta_map
.objects
.begin();
4920 while (iter
!= begin
) {
4922 if (next
->first
.get_head() != iter
->first
.get_head()) {
4927 for_meta_scrub
.objects
.insert(begin
, iter
);
4928 scrubber
.cleaned_meta_map
.objects
.erase(begin
, iter
);
4931 // ok, do the pg-type specific scrubbing
4932 scrub_snapshot_metadata(for_meta_scrub
, missing_digest
);
4933 if (!scrubber
.store
->empty()) {
4934 if (state_test(PG_STATE_REPAIR
)) {
4935 dout(10) << __func__
<< ": discarding scrub results" << dendl
;
4936 scrubber
.store
->flush(nullptr);
4938 dout(10) << __func__
<< ": updating scrub object" << dendl
;
4939 ObjectStore::Transaction t
;
4940 scrubber
.store
->flush(&t
);
4941 osd
->store
->queue_transaction(osr
.get(), std::move(t
), nullptr);
4946 bool PG::scrub_process_inconsistent()
4948 dout(10) << __func__
<< ": checking authoritative" << dendl
;
4949 bool repair
= state_test(PG_STATE_REPAIR
);
4950 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
4951 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
4953 // authoriative only store objects which missing or inconsistent.
4954 if (!scrubber
.authoritative
.empty()) {
4956 ss
<< info
.pgid
<< " " << mode
<< " "
4957 << scrubber
.missing
.size() << " missing, "
4958 << scrubber
.inconsistent
.size() << " inconsistent objects";
4959 dout(2) << ss
.str() << dendl
;
4960 osd
->clog
->error(ss
);
4962 state_clear(PG_STATE_CLEAN
);
4963 for (map
<hobject_t
, list
<pair
<ScrubMap::object
, pg_shard_t
> >>::iterator i
=
4964 scrubber
.authoritative
.begin();
4965 i
!= scrubber
.authoritative
.end();
4967 set
<pg_shard_t
>::iterator j
;
4969 auto missing_entry
= scrubber
.missing
.find(i
->first
);
4970 if (missing_entry
!= scrubber
.missing
.end()) {
4971 for (j
= missing_entry
->second
.begin();
4972 j
!= missing_entry
->second
.end();
4981 if (scrubber
.inconsistent
.count(i
->first
)) {
4982 for (j
= scrubber
.inconsistent
[i
->first
].begin();
4983 j
!= scrubber
.inconsistent
[i
->first
].end();
4985 repair_object(i
->first
,
4994 return (!scrubber
.authoritative
.empty() && repair
);
4997 bool PG::ops_blocked_by_scrub() const {
4998 return (waiting_for_scrub
.size() != 0);
5001 // the part that actually finalizes a scrub
5002 void PG::scrub_finish()
5004 bool repair
= state_test(PG_STATE_REPAIR
);
5005 // if the repair request comes from auto-repair and large number of errors,
5006 // we would like to cancel auto-repair
5007 if (repair
&& scrubber
.auto_repair
5008 && scrubber
.authoritative
.size() > cct
->_conf
->osd_scrub_auto_repair_num_errors
) {
5009 state_clear(PG_STATE_REPAIR
);
5012 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
5013 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
5015 // type-specific finish (can tally more errors)
5018 bool has_error
= scrub_process_inconsistent();
5022 oss
<< info
.pgid
.pgid
<< " " << mode
<< " ";
5023 int total_errors
= scrubber
.shallow_errors
+ scrubber
.deep_errors
;
5025 oss
<< total_errors
<< " errors";
5028 if (!deep_scrub
&& info
.stats
.stats
.sum
.num_deep_scrub_errors
)
5029 oss
<< " ( " << info
.stats
.stats
.sum
.num_deep_scrub_errors
5030 << " remaining deep scrub error details lost)";
5032 oss
<< ", " << scrubber
.fixed
<< " fixed";
5034 osd
->clog
->error(oss
);
5036 osd
->clog
->debug(oss
);
5041 utime_t now
= ceph_clock_now();
5042 info
.history
.last_scrub
= info
.last_update
;
5043 info
.history
.last_scrub_stamp
= now
;
5044 if (scrubber
.deep
) {
5045 info
.history
.last_deep_scrub
= info
.last_update
;
5046 info
.history
.last_deep_scrub_stamp
= now
;
5048 // Since we don't know which errors were fixed, we can only clear them
5049 // when every one has been fixed.
5051 if (scrubber
.fixed
== scrubber
.shallow_errors
+ scrubber
.deep_errors
) {
5053 scrubber
.shallow_errors
= scrubber
.deep_errors
= 0;
5055 // Deep scrub in order to get corrected error counts
5056 scrub_after_recovery
= true;
5060 if ((scrubber
.shallow_errors
== 0) && (scrubber
.deep_errors
== 0))
5061 info
.history
.last_clean_scrub_stamp
= now
;
5062 info
.stats
.stats
.sum
.num_shallow_scrub_errors
= scrubber
.shallow_errors
;
5063 info
.stats
.stats
.sum
.num_deep_scrub_errors
= scrubber
.deep_errors
;
5065 info
.stats
.stats
.sum
.num_shallow_scrub_errors
= scrubber
.shallow_errors
;
5066 // XXX: last_clean_scrub_stamp doesn't mean the pg is not inconsistent
5067 // because of deep-scrub errors
5068 if (scrubber
.shallow_errors
== 0)
5069 info
.history
.last_clean_scrub_stamp
= now
;
5071 info
.stats
.stats
.sum
.num_scrub_errors
=
5072 info
.stats
.stats
.sum
.num_shallow_scrub_errors
+
5073 info
.stats
.stats
.sum
.num_deep_scrub_errors
;
5077 ObjectStore::Transaction t
;
5080 int tr
= osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
5086 queue_peering_event(
5088 std::make_shared
<CephPeeringEvt
>(
5089 get_osdmap()->get_epoch(),
5090 get_osdmap()->get_epoch(),
5094 scrub_clear_state();
5095 scrub_unreserve_replicas();
5097 if (is_active() && is_primary()) {
5102 void PG::share_pg_info()
5104 dout(10) << "share_pg_info" << dendl
;
5106 // share new pg_info_t with replicas
5107 assert(!actingbackfill
.empty());
5108 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
5109 i
!= actingbackfill
.end();
5111 if (*i
== pg_whoami
) continue;
5112 pg_shard_t peer
= *i
;
5113 if (peer_info
.count(peer
)) {
5114 peer_info
[peer
].last_epoch_started
= info
.last_epoch_started
;
5115 peer_info
[peer
].last_interval_started
= info
.last_interval_started
;
5116 peer_info
[peer
].history
.merge(info
.history
);
5118 MOSDPGInfo
*m
= new MOSDPGInfo(get_osdmap()->get_epoch());
5119 m
->pg_list
.push_back(
5122 peer
.shard
, pg_whoami
.shard
,
5123 get_osdmap()->get_epoch(),
5124 get_osdmap()->get_epoch(),
5127 osd
->send_message_osd_cluster(peer
.osd
, m
, get_osdmap()->get_epoch());
5131 bool PG::append_log_entries_update_missing(
5132 const mempool::osd_pglog::list
<pg_log_entry_t
> &entries
,
5133 ObjectStore::Transaction
&t
)
5135 assert(!entries
.empty());
5136 assert(entries
.begin()->version
> info
.last_update
);
5138 PGLogEntryHandler rollbacker
{this, &t
};
5139 bool invalidate_stats
=
5140 pg_log
.append_new_log_entries(info
.last_backfill
,
5141 info
.last_backfill_bitwise
,
5144 info
.last_update
= pg_log
.get_head();
5146 if (pg_log
.get_missing().num_missing() == 0) {
5147 // advance last_complete since nothing else is missing!
5148 info
.last_complete
= info
.last_update
;
5151 info
.stats
.stats_invalid
= info
.stats
.stats_invalid
|| invalidate_stats
;
5154 return invalidate_stats
;
5158 void PG::merge_new_log_entries(
5159 const mempool::osd_pglog::list
<pg_log_entry_t
> &entries
,
5160 ObjectStore::Transaction
&t
)
5162 dout(10) << __func__
<< " " << entries
<< dendl
;
5163 assert(is_primary());
5165 bool rebuild_missing
= append_log_entries_update_missing(entries
, t
);
5166 for (set
<pg_shard_t
>::const_iterator i
= actingbackfill
.begin();
5167 i
!= actingbackfill
.end();
5169 pg_shard_t
peer(*i
);
5170 if (peer
== pg_whoami
) continue;
5171 assert(peer_missing
.count(peer
));
5172 assert(peer_info
.count(peer
));
5173 pg_missing_t
& pmissing(peer_missing
[peer
]);
5174 dout(20) << __func__
<< " peer_missing for " << peer
<< " = " << pmissing
<< dendl
;
5175 pg_info_t
& pinfo(peer_info
[peer
]);
5176 bool invalidate_stats
= PGLog::append_log_entries_update_missing(
5177 pinfo
.last_backfill
,
5178 info
.last_backfill_bitwise
,
5185 pinfo
.last_update
= info
.last_update
;
5186 pinfo
.stats
.stats_invalid
= pinfo
.stats
.stats_invalid
|| invalidate_stats
;
5187 rebuild_missing
= rebuild_missing
|| invalidate_stats
;
5190 if (!rebuild_missing
) {
5194 for (auto &&i
: entries
) {
5195 missing_loc
.rebuild(
5200 pg_log
.get_missing(),
5206 void PG::update_history(const pg_history_t
& new_history
)
5209 if (info
.history
.merge(new_history
)) {
5210 dout(20) << __func__
<< " advanced history from " << new_history
<< dendl
;
5212 if (info
.history
.last_epoch_clean
>= info
.history
.same_interval_since
) {
5213 dout(20) << __func__
<< " clearing past_intervals" << dendl
;
5214 past_intervals
.clear();
5215 dirty_big_info
= true;
5221 void PG::fulfill_info(
5222 pg_shard_t from
, const pg_query_t
&query
,
5223 pair
<pg_shard_t
, pg_info_t
> ¬ify_info
)
5225 assert(from
== primary
);
5226 assert(query
.type
== pg_query_t::INFO
);
5229 dout(10) << "sending info" << dendl
;
5230 notify_info
= make_pair(from
, info
);
5233 void PG::fulfill_log(
5234 pg_shard_t from
, const pg_query_t
&query
, epoch_t query_epoch
)
5236 dout(10) << "log request from " << from
<< dendl
;
5237 assert(from
== primary
);
5238 assert(query
.type
!= pg_query_t::INFO
);
5239 ConnectionRef con
= osd
->get_con_osd_cluster(
5240 from
.osd
, get_osdmap()->get_epoch());
5243 MOSDPGLog
*mlog
= new MOSDPGLog(
5244 from
.shard
, pg_whoami
.shard
,
5245 get_osdmap()->get_epoch(),
5247 mlog
->missing
= pg_log
.get_missing();
5249 // primary -> other, when building master log
5250 if (query
.type
== pg_query_t::LOG
) {
5251 dout(10) << " sending info+missing+log since " << query
.since
5253 if (query
.since
!= eversion_t() && query
.since
< pg_log
.get_tail()) {
5254 osd
->clog
->error() << info
.pgid
<< " got broken pg_query_t::LOG since " << query
.since
5255 << " when my log.tail is " << pg_log
.get_tail()
5256 << ", sending full log instead";
5257 mlog
->log
= pg_log
.get_log(); // primary should not have requested this!!
5259 mlog
->log
.copy_after(pg_log
.get_log(), query
.since
);
5261 else if (query
.type
== pg_query_t::FULLLOG
) {
5262 dout(10) << " sending info+missing+full log" << dendl
;
5263 mlog
->log
= pg_log
.get_log();
5266 dout(10) << " sending " << mlog
->log
<< " " << mlog
->missing
<< dendl
;
5268 osd
->share_map_peer(from
.osd
, con
.get(), get_osdmap());
5269 osd
->send_message_osd_cluster(mlog
, con
.get());
5272 void PG::check_full_transition(OSDMapRef lastmap
, OSDMapRef osdmap
)
5274 bool changed
= false;
5275 if (osdmap
->test_flag(CEPH_OSDMAP_FULL
) &&
5276 !lastmap
->test_flag(CEPH_OSDMAP_FULL
)) {
5277 dout(10) << " cluster was marked full in " << osdmap
->get_epoch() << dendl
;
5280 const pg_pool_t
*pi
= osdmap
->get_pg_pool(info
.pgid
.pool());
5282 if (pi
->has_flag(pg_pool_t::FLAG_FULL
)) {
5283 const pg_pool_t
*opi
= lastmap
->get_pg_pool(info
.pgid
.pool());
5284 if (!opi
|| !opi
->has_flag(pg_pool_t::FLAG_FULL
)) {
5285 dout(10) << " pool was marked full in " << osdmap
->get_epoch() << dendl
;
5290 info
.history
.last_epoch_marked_full
= osdmap
->get_epoch();
5295 bool PG::should_restart_peering(
5297 int newactingprimary
,
5298 const vector
<int>& newup
,
5299 const vector
<int>& newacting
,
5303 if (PastIntervals::is_new_interval(
5315 dout(20) << "new interval newup " << newup
5316 << " newacting " << newacting
<< dendl
;
5323 bool PG::old_peering_msg(epoch_t reply_epoch
, epoch_t query_epoch
)
5325 if (last_peering_reset
> reply_epoch
||
5326 last_peering_reset
> query_epoch
) {
5327 dout(10) << "old_peering_msg reply_epoch " << reply_epoch
<< " query_epoch " << query_epoch
5328 << " last_peering_reset " << last_peering_reset
5335 void PG::set_last_peering_reset()
5337 dout(20) << "set_last_peering_reset " << get_osdmap()->get_epoch() << dendl
;
5338 if (last_peering_reset
!= get_osdmap()->get_epoch()) {
5339 last_peering_reset
= get_osdmap()->get_epoch();
5340 reset_interval_flush();
5347 FlushState(PG
*pg
, epoch_t epoch
) : pg(pg
), epoch(epoch
) {}
5350 if (!pg
->pg_has_reset_since(epoch
))
5351 pg
->queue_flushed(epoch
);
5355 typedef ceph::shared_ptr
<FlushState
> FlushStateRef
;
5357 void PG::start_flush(ObjectStore::Transaction
*t
,
5358 list
<Context
*> *on_applied
,
5359 list
<Context
*> *on_safe
)
5361 // flush in progress ops
5362 FlushStateRef
flush_trigger (std::make_shared
<FlushState
>(
5363 this, get_osdmap()->get_epoch()));
5365 flushes_in_progress
++;
5366 on_applied
->push_back(new ContainerContext
<FlushStateRef
>(flush_trigger
));
5367 on_safe
->push_back(new ContainerContext
<FlushStateRef
>(flush_trigger
));
5370 void PG::reset_interval_flush()
5372 dout(10) << "Clearing blocked outgoing recovery messages" << dendl
;
5373 recovery_state
.clear_blocked_outgoing();
5375 Context
*c
= new QueuePeeringEvt
<IntervalFlush
>(
5376 this, get_osdmap()->get_epoch(), IntervalFlush());
5377 if (!osr
->flush_commit(c
)) {
5378 dout(10) << "Beginning to block outgoing recovery messages" << dendl
;
5379 recovery_state
.begin_block_outgoing();
5381 dout(10) << "Not blocking outgoing recovery messages" << dendl
;
5386 /* Called before initializing peering during advance_map */
5387 void PG::start_peering_interval(
5388 const OSDMapRef lastmap
,
5389 const vector
<int>& newup
, int new_up_primary
,
5390 const vector
<int>& newacting
, int new_acting_primary
,
5391 ObjectStore::Transaction
*t
)
5393 const OSDMapRef osdmap
= get_osdmap();
5395 set_last_peering_reset();
5397 vector
<int> oldacting
, oldup
;
5398 int oldrole
= get_role();
5402 pg_shard_t old_acting_primary
= get_primary();
5403 pg_shard_t old_up_primary
= up_primary
;
5404 bool was_old_primary
= is_primary();
5406 acting
.swap(oldacting
);
5408 init_primary_up_acting(
5412 new_acting_primary
);
5414 if (info
.stats
.up
!= up
||
5415 info
.stats
.acting
!= acting
||
5416 info
.stats
.up_primary
!= new_up_primary
||
5417 info
.stats
.acting_primary
!= new_acting_primary
) {
5419 info
.stats
.up_primary
= new_up_primary
;
5420 info
.stats
.acting
= acting
;
5421 info
.stats
.acting_primary
= new_acting_primary
;
5422 info
.stats
.mapping_epoch
= osdmap
->get_epoch();
5425 pg_stats_publish_lock
.Lock();
5426 pg_stats_publish_valid
= false;
5427 pg_stats_publish_lock
.Unlock();
5429 // This will now be remapped during a backfill in cases
5430 // that it would not have been before.
5432 state_set(PG_STATE_REMAPPED
);
5434 state_clear(PG_STATE_REMAPPED
);
5436 int role
= osdmap
->calc_pg_role(osd
->whoami
, acting
, acting
.size());
5437 if (pool
.info
.is_replicated() || role
== pg_whoami
.shard
)
5442 // did acting, up, primary|acker change?
5444 dout(10) << " no lastmap" << dendl
;
5446 dirty_big_info
= true;
5447 info
.history
.same_interval_since
= osdmap
->get_epoch();
5449 std::stringstream debug
;
5450 assert(info
.history
.same_interval_since
!= 0);
5451 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable(
5452 get_is_recoverable_predicate());
5453 bool new_interval
= PastIntervals::check_new_interval(
5454 old_acting_primary
.osd
,
5456 oldacting
, newacting
,
5460 info
.history
.same_interval_since
,
5461 info
.history
.last_epoch_clean
,
5468 dout(10) << __func__
<< ": check_new_interval output: "
5469 << debug
.str() << dendl
;
5471 if (osdmap
->get_epoch() == osd
->get_superblock().oldest_map
&&
5472 info
.history
.last_epoch_clean
< osdmap
->get_epoch()) {
5473 dout(10) << " map gap, clearing past_intervals and faking" << dendl
;
5474 // our information is incomplete and useless; someone else was clean
5475 // after everything we know if osdmaps were trimmed.
5476 past_intervals
.clear();
5478 dout(10) << " noting past " << past_intervals
<< dendl
;
5481 dirty_big_info
= true;
5482 info
.history
.same_interval_since
= osdmap
->get_epoch();
5483 if (info
.pgid
.pgid
.is_split(lastmap
->get_pg_num(info
.pgid
.pgid
.pool()),
5484 osdmap
->get_pg_num(info
.pgid
.pgid
.pool()),
5486 info
.history
.last_epoch_split
= osdmap
->get_epoch();
5491 if (old_up_primary
!= up_primary
||
5493 info
.history
.same_up_since
= osdmap
->get_epoch();
5495 // this comparison includes primary rank via pg_shard_t
5496 if (old_acting_primary
!= get_primary()) {
5497 info
.history
.same_primary_since
= osdmap
->get_epoch();
5502 dout(1) << __func__
<< " up " << oldup
<< " -> " << up
5503 << ", acting " << oldacting
<< " -> " << acting
5504 << ", acting_primary " << old_acting_primary
<< " -> " << new_acting_primary
5505 << ", up_primary " << old_up_primary
<< " -> " << new_up_primary
5506 << ", role " << oldrole
<< " -> " << role
5507 << ", features acting " << acting_features
5508 << " upacting " << upacting_features
5512 state_clear(PG_STATE_ACTIVE
);
5513 state_clear(PG_STATE_PEERED
);
5514 state_clear(PG_STATE_DOWN
);
5515 state_clear(PG_STATE_RECOVERY_WAIT
);
5516 state_clear(PG_STATE_RECOVERY_TOOFULL
);
5517 state_clear(PG_STATE_RECOVERING
);
5519 peer_purged
.clear();
5520 actingbackfill
.clear();
5521 scrub_queued
= false;
5523 // reset primary state?
5524 if (was_old_primary
|| is_primary()) {
5525 osd
->remove_want_pg_temp(info
.pgid
.pgid
);
5527 clear_primary_state();
5533 projected_last_update
= eversion_t();
5537 // should we tell the primary we are here?
5538 send_notify
= !is_primary();
5540 if (role
!= oldrole
||
5541 was_old_primary
!= is_primary()) {
5542 // did primary change?
5543 if (was_old_primary
!= is_primary()) {
5544 state_clear(PG_STATE_CLEAN
);
5545 clear_publish_stats();
5550 // take active waiters
5551 requeue_ops(waiting_for_peered
);
5555 // did primary change?
5556 if (get_primary() != old_acting_primary
) {
5557 dout(10) << *this << " " << oldacting
<< " -> " << acting
5558 << ", acting primary "
5559 << old_acting_primary
<< " -> " << get_primary()
5562 // primary is the same.
5564 // i am (still) primary. but my replica set changed.
5565 state_clear(PG_STATE_CLEAN
);
5567 dout(10) << oldacting
<< " -> " << acting
5568 << ", replicas changed" << dendl
;
5574 if (acting
.empty() && !up
.empty() && up_primary
== pg_whoami
) {
5575 dout(10) << " acting empty, but i am up[0], clearing pg_temp" << dendl
;
5576 osd
->queue_want_pg_temp(info
.pgid
.pgid
, acting
);
5580 void PG::on_new_interval()
5582 const OSDMapRef osdmap
= get_osdmap();
5586 // initialize features
5587 acting_features
= CEPH_FEATURES_SUPPORTED_DEFAULT
;
5588 upacting_features
= CEPH_FEATURES_SUPPORTED_DEFAULT
;
5589 for (vector
<int>::iterator p
= acting
.begin(); p
!= acting
.end(); ++p
) {
5590 if (*p
== CRUSH_ITEM_NONE
)
5592 uint64_t f
= osdmap
->get_xinfo(*p
).features
;
5593 acting_features
&= f
;
5594 upacting_features
&= f
;
5596 for (vector
<int>::iterator p
= up
.begin(); p
!= up
.end(); ++p
) {
5597 if (*p
== CRUSH_ITEM_NONE
)
5599 upacting_features
&= osdmap
->get_xinfo(*p
).features
;
5605 void PG::proc_primary_info(ObjectStore::Transaction
&t
, const pg_info_t
&oinfo
)
5607 assert(!is_primary());
5609 update_history(oinfo
.history
);
5611 if (last_complete_ondisk
.epoch
>= info
.history
.last_epoch_started
) {
5612 // DEBUG: verify that the snaps are empty in snap_mapper
5613 if (cct
->_conf
->osd_debug_verify_snaps_on_info
) {
5614 interval_set
<snapid_t
> p
;
5615 p
.union_of(oinfo
.purged_snaps
, info
.purged_snaps
);
5616 p
.subtract(info
.purged_snaps
);
5618 for (interval_set
<snapid_t
>::iterator i
= p
.begin();
5621 for (snapid_t snap
= i
.get_start();
5622 snap
!= i
.get_len() + i
.get_start();
5624 vector
<hobject_t
> hoids
;
5625 int r
= snap_mapper
.get_next_objects_to_trim(snap
, 1, &hoids
);
5626 if (r
!= 0 && r
!= -ENOENT
) {
5627 derr
<< __func__
<< ": snap_mapper get_next_object_to_trim returned "
5628 << cpp_strerror(r
) << dendl
;
5630 } else if (r
!= -ENOENT
) {
5631 assert(!hoids
.empty());
5632 derr
<< __func__
<< ": snap_mapper get_next_object_to_trim returned "
5633 << cpp_strerror(r
) << " for object "
5634 << hoids
[0] << " on snap " << snap
5635 << " which should have been fully trimmed " << dendl
;
5642 info
.purged_snaps
= oinfo
.purged_snaps
;
5644 dirty_big_info
= true;
5648 ostream
& operator<<(ostream
& out
, const PG
& pg
)
5650 out
<< "pg[" << pg
.info
5652 if (pg
.acting
!= pg
.up
)
5653 out
<< "/" << pg
.acting
;
5655 out
<< "p" << pg
.get_primary();
5656 out
<< " r=" << pg
.get_role();
5657 out
<< " lpr=" << pg
.get_last_peering_reset();
5659 if (!pg
.past_intervals
.empty()) {
5660 out
<< " pi=[" << pg
.past_intervals
.get_bounds()
5661 << ")/" << pg
.past_intervals
.size();
5664 if (pg
.is_peered()) {
5665 if (pg
.last_update_ondisk
!= pg
.info
.last_update
)
5666 out
<< " luod=" << pg
.last_update_ondisk
;
5667 if (pg
.last_update_applied
!= pg
.info
.last_update
)
5668 out
<< " lua=" << pg
.last_update_applied
;
5671 if (pg
.recovery_ops_active
)
5672 out
<< " rops=" << pg
.recovery_ops_active
;
5674 if (pg
.pg_log
.get_tail() != pg
.info
.log_tail
||
5675 pg
.pg_log
.get_head() != pg
.info
.last_update
)
5676 out
<< " (info mismatch, " << pg
.pg_log
.get_log() << ")";
5678 if (!pg
.pg_log
.get_log().empty()) {
5679 if ((pg
.pg_log
.get_log().log
.begin()->version
<= pg
.pg_log
.get_tail())) {
5680 out
<< " (log bound mismatch, actual=["
5681 << pg
.pg_log
.get_log().log
.begin()->version
<< ","
5682 << pg
.pg_log
.get_log().log
.rbegin()->version
<< "]";
5687 if (!pg
.backfill_targets
.empty())
5688 out
<< " bft=" << pg
.backfill_targets
;
5689 out
<< " crt=" << pg
.pg_log
.get_can_rollback_to();
5691 if (pg
.last_complete_ondisk
!= pg
.info
.last_complete
)
5692 out
<< " lcod " << pg
.last_complete_ondisk
;
5694 if (pg
.is_primary()) {
5695 out
<< " mlcod " << pg
.min_last_complete_ondisk
;
5698 out
<< " " << pg_state_string(pg
.get_state());
5699 if (pg
.should_send_notify())
5702 if (pg
.scrubber
.must_repair
)
5703 out
<< " MUST_REPAIR";
5704 if (pg
.scrubber
.auto_repair
)
5705 out
<< " AUTO_REPAIR";
5706 if (pg
.scrubber
.must_deep_scrub
)
5707 out
<< " MUST_DEEP_SCRUB";
5708 if (pg
.scrubber
.must_scrub
)
5709 out
<< " MUST_SCRUB";
5711 //out << " (" << pg.pg_log.get_tail() << "," << pg.pg_log.get_head() << "]";
5712 if (pg
.pg_log
.get_missing().num_missing()) {
5713 out
<< " m=" << pg
.pg_log
.get_missing().num_missing();
5714 if (pg
.is_primary()) {
5715 uint64_t unfound
= pg
.get_num_unfound();
5717 out
<< " u=" << unfound
;
5720 if (pg
.snap_trimq
.size())
5721 out
<< " snaptrimq=" << pg
.snap_trimq
;
5729 bool PG::can_discard_op(OpRequestRef
& op
)
5731 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
5732 if (cct
->_conf
->osd_discard_disconnected_ops
&& OSD::op_is_discardable(m
)) {
5733 dout(20) << " discard " << *m
<< dendl
;
5737 if (m
->get_map_epoch() < info
.history
.same_primary_since
) {
5738 dout(7) << " changed after " << m
->get_map_epoch()
5739 << ", dropping " << *m
<< dendl
;
5743 if (m
->get_connection()->has_feature(CEPH_FEATURE_RESEND_ON_SPLIT
)) {
5744 if (m
->get_map_epoch() < pool
.info
.get_last_force_op_resend()) {
5745 dout(7) << __func__
<< " sent before last_force_op_resend "
5746 << pool
.info
.last_force_op_resend
<< ", dropping" << *m
<< dendl
;
5749 if (m
->get_map_epoch() < info
.history
.last_epoch_split
) {
5750 dout(7) << __func__
<< " pg split in "
5751 << info
.history
.last_epoch_split
<< ", dropping" << dendl
;
5754 } else if (m
->get_connection()->has_feature(CEPH_FEATURE_OSD_POOLRESEND
)) {
5755 if (m
->get_map_epoch() < pool
.info
.get_last_force_op_resend_preluminous()) {
5756 dout(7) << __func__
<< " sent before last_force_op_resend_preluminous "
5757 << pool
.info
.last_force_op_resend_preluminous
5758 << ", dropping" << *m
<< dendl
;
5766 template<typename T
, int MSGTYPE
>
5767 bool PG::can_discard_replica_op(OpRequestRef
& op
)
5769 const T
*m
= static_cast<const T
*>(op
->get_req());
5770 assert(m
->get_type() == MSGTYPE
);
5772 int from
= m
->get_source().num();
5774 // if a repop is replied after a replica goes down in a new osdmap, and
5775 // before the pg advances to this new osdmap, the repop replies before this
5776 // repop can be discarded by that replica OSD, because the primary resets the
5777 // connection to it when handling the new osdmap marking it down, and also
5778 // resets the messenger sesssion when the replica reconnects. to avoid the
5779 // out-of-order replies, the messages from that replica should be discarded.
5780 if (osd
->get_osdmap()->is_down(from
))
5782 /* Mostly, this overlaps with the old_peering_msg
5783 * condition. An important exception is pushes
5784 * sent by replicas not in the acting set, since
5785 * if such a replica goes down it does not cause
5786 * a new interval. */
5787 if (get_osdmap()->get_down_at(from
) >= m
->map_epoch
)
5791 // if pg changes _at all_, we reset and repeer!
5792 if (old_peering_msg(m
->map_epoch
, m
->map_epoch
)) {
5793 dout(10) << "can_discard_replica_op pg changed " << info
.history
5794 << " after " << m
->map_epoch
5795 << ", dropping" << dendl
;
5801 bool PG::can_discard_scan(OpRequestRef op
)
5803 const MOSDPGScan
*m
= static_cast<const MOSDPGScan
*>(op
->get_req());
5804 assert(m
->get_type() == MSG_OSD_PG_SCAN
);
5806 if (old_peering_msg(m
->map_epoch
, m
->query_epoch
)) {
5807 dout(10) << " got old scan, ignoring" << dendl
;
5813 bool PG::can_discard_backfill(OpRequestRef op
)
5815 const MOSDPGBackfill
*m
= static_cast<const MOSDPGBackfill
*>(op
->get_req());
5816 assert(m
->get_type() == MSG_OSD_PG_BACKFILL
);
5818 if (old_peering_msg(m
->map_epoch
, m
->query_epoch
)) {
5819 dout(10) << " got old backfill, ignoring" << dendl
;
5827 bool PG::can_discard_request(OpRequestRef
& op
)
5829 switch (op
->get_req()->get_type()) {
5830 case CEPH_MSG_OSD_OP
:
5831 return can_discard_op(op
);
5832 case CEPH_MSG_OSD_BACKOFF
:
5833 return false; // never discard
5835 return can_discard_replica_op
<MOSDSubOp
, MSG_OSD_SUBOP
>(op
);
5837 return can_discard_replica_op
<MOSDRepOp
, MSG_OSD_REPOP
>(op
);
5838 case MSG_OSD_PG_PUSH
:
5839 return can_discard_replica_op
<MOSDPGPush
, MSG_OSD_PG_PUSH
>(op
);
5840 case MSG_OSD_PG_PULL
:
5841 return can_discard_replica_op
<MOSDPGPull
, MSG_OSD_PG_PULL
>(op
);
5842 case MSG_OSD_PG_PUSH_REPLY
:
5843 return can_discard_replica_op
<MOSDPGPushReply
, MSG_OSD_PG_PUSH_REPLY
>(op
);
5844 case MSG_OSD_SUBOPREPLY
:
5845 return can_discard_replica_op
<MOSDSubOpReply
, MSG_OSD_SUBOPREPLY
>(op
);
5846 case MSG_OSD_REPOPREPLY
:
5847 return can_discard_replica_op
<MOSDRepOpReply
, MSG_OSD_REPOPREPLY
>(op
);
5848 case MSG_OSD_PG_RECOVERY_DELETE
:
5849 return can_discard_replica_op
<MOSDPGRecoveryDelete
, MSG_OSD_PG_RECOVERY_DELETE
>(op
);
5851 case MSG_OSD_PG_RECOVERY_DELETE_REPLY
:
5852 return can_discard_replica_op
<MOSDPGRecoveryDeleteReply
, MSG_OSD_PG_RECOVERY_DELETE_REPLY
>(op
);
5854 case MSG_OSD_EC_WRITE
:
5855 return can_discard_replica_op
<MOSDECSubOpWrite
, MSG_OSD_EC_WRITE
>(op
);
5856 case MSG_OSD_EC_WRITE_REPLY
:
5857 return can_discard_replica_op
<MOSDECSubOpWriteReply
, MSG_OSD_EC_WRITE_REPLY
>(op
);
5858 case MSG_OSD_EC_READ
:
5859 return can_discard_replica_op
<MOSDECSubOpRead
, MSG_OSD_EC_READ
>(op
);
5860 case MSG_OSD_EC_READ_REPLY
:
5861 return can_discard_replica_op
<MOSDECSubOpReadReply
, MSG_OSD_EC_READ_REPLY
>(op
);
5862 case MSG_OSD_REP_SCRUB
:
5863 return can_discard_replica_op
<MOSDRepScrub
, MSG_OSD_REP_SCRUB
>(op
);
5864 case MSG_OSD_SCRUB_RESERVE
:
5865 return can_discard_replica_op
<MOSDScrubReserve
, MSG_OSD_SCRUB_RESERVE
>(op
);
5866 case MSG_OSD_REP_SCRUBMAP
:
5867 return can_discard_replica_op
<MOSDRepScrubMap
, MSG_OSD_REP_SCRUBMAP
>(op
);
5868 case MSG_OSD_PG_UPDATE_LOG_MISSING
:
5869 return can_discard_replica_op
<
5870 MOSDPGUpdateLogMissing
, MSG_OSD_PG_UPDATE_LOG_MISSING
>(op
);
5871 case MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY
:
5872 return can_discard_replica_op
<
5873 MOSDPGUpdateLogMissingReply
, MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY
>(op
);
5875 case MSG_OSD_PG_SCAN
:
5876 return can_discard_scan(op
);
5877 case MSG_OSD_PG_BACKFILL
:
5878 return can_discard_backfill(op
);
5879 case MSG_OSD_PG_BACKFILL_REMOVE
:
5880 return can_discard_replica_op
<MOSDPGBackfillRemove
,
5881 MSG_OSD_PG_BACKFILL_REMOVE
>(op
);
5886 void PG::take_waiters()
5888 dout(10) << "take_waiters" << dendl
;
5889 requeue_map_waiters();
5890 for (list
<CephPeeringEvtRef
>::iterator i
= peering_waiters
.begin();
5891 i
!= peering_waiters
.end();
5892 ++i
) osd
->queue_for_peering(this);
5893 peering_queue
.splice(peering_queue
.begin(), peering_waiters
,
5894 peering_waiters
.begin(), peering_waiters
.end());
5897 void PG::handle_peering_event(CephPeeringEvtRef evt
, RecoveryCtx
*rctx
)
5899 dout(10) << "handle_peering_event: " << evt
->get_desc() << dendl
;
5900 if (!have_same_or_newer_map(evt
->get_epoch_sent())) {
5901 dout(10) << "deferring event " << evt
->get_desc() << dendl
;
5902 peering_waiters
.push_back(evt
);
5905 if (old_peering_evt(evt
))
5907 recovery_state
.handle_event(evt
, rctx
);
5910 void PG::queue_peering_event(CephPeeringEvtRef evt
)
5912 if (old_peering_evt(evt
))
5914 peering_queue
.push_back(evt
);
5915 osd
->queue_for_peering(this);
5918 void PG::queue_null(epoch_t msg_epoch
,
5919 epoch_t query_epoch
)
5921 dout(10) << "null" << dendl
;
5922 queue_peering_event(
5923 CephPeeringEvtRef(std::make_shared
<CephPeeringEvt
>(msg_epoch
, query_epoch
,
5927 void PG::queue_flushed(epoch_t e
)
5929 dout(10) << "flushed" << dendl
;
5930 queue_peering_event(
5931 CephPeeringEvtRef(std::make_shared
<CephPeeringEvt
>(e
, e
,
5935 void PG::queue_query(epoch_t msg_epoch
,
5936 epoch_t query_epoch
,
5937 pg_shard_t from
, const pg_query_t
& q
)
5939 dout(10) << "handle_query " << q
<< " from replica " << from
<< dendl
;
5940 queue_peering_event(
5941 CephPeeringEvtRef(std::make_shared
<CephPeeringEvt
>(msg_epoch
, query_epoch
,
5942 MQuery(from
, q
, query_epoch
))));
5945 void PG::handle_advance_map(
5946 OSDMapRef osdmap
, OSDMapRef lastmap
,
5947 vector
<int>& newup
, int up_primary
,
5948 vector
<int>& newacting
, int acting_primary
,
5951 assert(lastmap
->get_epoch() == osdmap_ref
->get_epoch());
5952 assert(lastmap
== osdmap_ref
);
5953 dout(10) << "handle_advance_map "
5954 << newup
<< "/" << newacting
5955 << " -- " << up_primary
<< "/" << acting_primary
5957 update_osdmap_ref(osdmap
);
5958 pool
.update(osdmap
);
5959 past_intervals
.update_type_from_map(pool
.info
.ec_pool(), *osdmap
);
5960 if (cct
->_conf
->osd_debug_verify_cached_snaps
) {
5961 interval_set
<snapid_t
> actual_removed_snaps
;
5962 const pg_pool_t
*pi
= osdmap
->get_pg_pool(info
.pgid
.pool());
5964 pi
->build_removed_snaps(actual_removed_snaps
);
5965 if (!(actual_removed_snaps
== pool
.cached_removed_snaps
)) {
5966 derr
<< __func__
<< ": mismatch between the actual removed snaps "
5967 << actual_removed_snaps
<< " and pool.cached_removed_snaps "
5968 << " pool.cached_removed_snaps " << pool
.cached_removed_snaps
5971 assert(actual_removed_snaps
== pool
.cached_removed_snaps
);
5974 osdmap
, lastmap
, newup
, up_primary
,
5975 newacting
, acting_primary
);
5976 recovery_state
.handle_event(evt
, rctx
);
5977 if (pool
.info
.last_change
== osdmap_ref
->get_epoch()) {
5979 update_store_with_options();
5983 void PG::handle_activate_map(RecoveryCtx
*rctx
)
5985 dout(10) << "handle_activate_map " << dendl
;
5987 recovery_state
.handle_event(evt
, rctx
);
5988 if (osdmap_ref
->get_epoch() - last_persisted_osdmap_ref
->get_epoch() >
5989 cct
->_conf
->osd_pg_epoch_persisted_max_stale
) {
5990 dout(20) << __func__
<< ": Dirtying info: last_persisted is "
5991 << last_persisted_osdmap_ref
->get_epoch()
5992 << " while current is " << osdmap_ref
->get_epoch() << dendl
;
5995 dout(20) << __func__
<< ": Not dirtying info: last_persisted is "
5996 << last_persisted_osdmap_ref
->get_epoch()
5997 << " while current is " << osdmap_ref
->get_epoch() << dendl
;
5999 if (osdmap_ref
->check_new_blacklist_entries()) check_blacklisted_watchers();
6002 void PG::handle_loaded(RecoveryCtx
*rctx
)
6004 dout(10) << "handle_loaded" << dendl
;
6006 recovery_state
.handle_event(evt
, rctx
);
6009 void PG::handle_create(RecoveryCtx
*rctx
)
6011 dout(10) << "handle_create" << dendl
;
6012 rctx
->created_pgs
.insert(this);
6014 recovery_state
.handle_event(evt
, rctx
);
6016 recovery_state
.handle_event(evt2
, rctx
);
6018 rctx
->on_applied
->add(make_lambda_context([this]() {
6019 update_store_with_options();
6023 void PG::handle_query_state(Formatter
*f
)
6025 dout(10) << "handle_query_state" << dendl
;
6027 recovery_state
.handle_event(q
, 0);
6030 void PG::update_store_with_options()
6032 auto r
= osd
->store
->set_collection_opts(coll
, pool
.info
.opts
);
6033 if(r
< 0 && r
!= -EOPNOTSUPP
) {
6034 derr
<< __func__
<< " set_collection_opts returns error:" << r
<< dendl
;
6038 void PG::update_store_on_load()
6040 if (osd
->store
->get_type() == "filestore") {
6041 // legacy filestore didn't store collection bit width; fix.
6042 int bits
= osd
->store
->collection_bits(coll
);
6044 assert(!coll
.is_meta()); // otherwise OSD::load_pgs() did a bad thing
6045 bits
= info
.pgid
.get_split_bits(pool
.info
.get_pg_num());
6046 lderr(cct
) << __func__
<< " setting bit width to " << bits
<< dendl
;
6047 ObjectStore::Transaction t
;
6048 t
.collection_set_bits(coll
, bits
);
6049 osd
->store
->apply_transaction(osr
.get(), std::move(t
));
6054 /*------------ Recovery State Machine----------------*/
6056 #define dout_prefix (*_dout << context< RecoveryMachine >().pg->gen_prefix() \
6057 << "state<" << get_state_name() << ">: ")
6059 /*------Crashed-------*/
6060 PG::RecoveryState::Crashed::Crashed(my_context ctx
)
6062 NamedState(context
< RecoveryMachine
>().pg
, "Crashed")
6064 context
< RecoveryMachine
>().log_enter(state_name
);
6065 assert(0 == "we got a bad state machine event");
6069 /*------Initial-------*/
6070 PG::RecoveryState::Initial::Initial(my_context ctx
)
6072 NamedState(context
< RecoveryMachine
>().pg
, "Initial")
6074 context
< RecoveryMachine
>().log_enter(state_name
);
6077 boost::statechart::result
PG::RecoveryState::Initial::react(const Load
& l
)
6079 PG
*pg
= context
< RecoveryMachine
>().pg
;
6081 // do we tell someone we're here?
6082 pg
->send_notify
= (!pg
->is_primary());
6083 pg
->update_store_with_options();
6085 pg
->update_store_on_load();
6087 return transit
< Reset
>();
6090 boost::statechart::result
PG::RecoveryState::Initial::react(const MNotifyRec
& notify
)
6092 PG
*pg
= context
< RecoveryMachine
>().pg
;
6093 pg
->proc_replica_info(
6094 notify
.from
, notify
.notify
.info
, notify
.notify
.epoch_sent
);
6095 pg
->set_last_peering_reset();
6096 return transit
< Primary
>();
6099 boost::statechart::result
PG::RecoveryState::Initial::react(const MInfoRec
& i
)
6101 PG
*pg
= context
< RecoveryMachine
>().pg
;
6102 assert(!pg
->is_primary());
6104 return transit
< Stray
>();
6107 boost::statechart::result
PG::RecoveryState::Initial::react(const MLogRec
& i
)
6109 PG
*pg
= context
< RecoveryMachine
>().pg
;
6110 assert(!pg
->is_primary());
6112 return transit
< Stray
>();
6115 void PG::RecoveryState::Initial::exit()
6117 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6118 PG
*pg
= context
< RecoveryMachine
>().pg
;
6119 utime_t dur
= ceph_clock_now() - enter_time
;
6120 pg
->osd
->recoverystate_perf
->tinc(rs_initial_latency
, dur
);
6123 /*------Started-------*/
6124 PG::RecoveryState::Started::Started(my_context ctx
)
6126 NamedState(context
< RecoveryMachine
>().pg
, "Started")
6128 context
< RecoveryMachine
>().log_enter(state_name
);
6131 boost::statechart::result
6132 PG::RecoveryState::Started::react(const IntervalFlush
&)
6134 PG
*pg
= context
< RecoveryMachine
>().pg
;
6135 ldout(pg
->cct
, 10) << "Ending blocked outgoing recovery messages" << dendl
;
6136 context
< RecoveryMachine
>().pg
->recovery_state
.end_block_outgoing();
6137 return discard_event();
6141 boost::statechart::result
6142 PG::RecoveryState::Started::react(const FlushedEvt
&)
6144 PG
*pg
= context
< RecoveryMachine
>().pg
;
6146 return discard_event();
6150 boost::statechart::result
PG::RecoveryState::Started::react(const AdvMap
& advmap
)
6152 PG
*pg
= context
< RecoveryMachine
>().pg
;
6153 ldout(pg
->cct
, 10) << "Started advmap" << dendl
;
6154 pg
->check_full_transition(advmap
.lastmap
, advmap
.osdmap
);
6155 if (pg
->should_restart_peering(
6157 advmap
.acting_primary
,
6162 ldout(pg
->cct
, 10) << "should_restart_peering, transitioning to Reset"
6165 return transit
< Reset
>();
6167 pg
->remove_down_peer_info(advmap
.osdmap
);
6168 return discard_event();
6171 boost::statechart::result
PG::RecoveryState::Started::react(const QueryState
& q
)
6173 q
.f
->open_object_section("state");
6174 q
.f
->dump_string("name", state_name
);
6175 q
.f
->dump_stream("enter_time") << enter_time
;
6176 q
.f
->close_section();
6177 return discard_event();
6180 void PG::RecoveryState::Started::exit()
6182 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6183 PG
*pg
= context
< RecoveryMachine
>().pg
;
6184 utime_t dur
= ceph_clock_now() - enter_time
;
6185 pg
->osd
->recoverystate_perf
->tinc(rs_started_latency
, dur
);
6188 /*--------Reset---------*/
6189 PG::RecoveryState::Reset::Reset(my_context ctx
)
6191 NamedState(context
< RecoveryMachine
>().pg
, "Reset")
6193 context
< RecoveryMachine
>().log_enter(state_name
);
6194 PG
*pg
= context
< RecoveryMachine
>().pg
;
6196 pg
->flushes_in_progress
= 0;
6197 pg
->set_last_peering_reset();
6200 boost::statechart::result
6201 PG::RecoveryState::Reset::react(const FlushedEvt
&)
6203 PG
*pg
= context
< RecoveryMachine
>().pg
;
6205 return discard_event();
6208 boost::statechart::result
6209 PG::RecoveryState::Reset::react(const IntervalFlush
&)
6211 PG
*pg
= context
< RecoveryMachine
>().pg
;
6212 ldout(pg
->cct
, 10) << "Ending blocked outgoing recovery messages" << dendl
;
6213 context
< RecoveryMachine
>().pg
->recovery_state
.end_block_outgoing();
6214 return discard_event();
6217 boost::statechart::result
PG::RecoveryState::Reset::react(const AdvMap
& advmap
)
6219 PG
*pg
= context
< RecoveryMachine
>().pg
;
6220 ldout(pg
->cct
, 10) << "Reset advmap" << dendl
;
6222 pg
->check_full_transition(advmap
.lastmap
, advmap
.osdmap
);
6224 if (pg
->should_restart_peering(
6226 advmap
.acting_primary
,
6231 ldout(pg
->cct
, 10) << "should restart peering, calling start_peering_interval again"
6233 pg
->start_peering_interval(
6235 advmap
.newup
, advmap
.up_primary
,
6236 advmap
.newacting
, advmap
.acting_primary
,
6237 context
< RecoveryMachine
>().get_cur_transaction());
6239 pg
->remove_down_peer_info(advmap
.osdmap
);
6240 pg
->check_past_interval_bounds();
6241 return discard_event();
6244 boost::statechart::result
PG::RecoveryState::Reset::react(const ActMap
&)
6246 PG
*pg
= context
< RecoveryMachine
>().pg
;
6247 if (pg
->should_send_notify() && pg
->get_primary().osd
>= 0) {
6248 context
< RecoveryMachine
>().send_notify(
6251 pg
->get_primary().shard
, pg
->pg_whoami
.shard
,
6252 pg
->get_osdmap()->get_epoch(),
6253 pg
->get_osdmap()->get_epoch(),
6255 pg
->past_intervals
);
6258 pg
->update_heartbeat_peers();
6261 return transit
< Started
>();
6264 boost::statechart::result
PG::RecoveryState::Reset::react(const QueryState
& q
)
6266 q
.f
->open_object_section("state");
6267 q
.f
->dump_string("name", state_name
);
6268 q
.f
->dump_stream("enter_time") << enter_time
;
6269 q
.f
->close_section();
6270 return discard_event();
6273 void PG::RecoveryState::Reset::exit()
6275 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6276 PG
*pg
= context
< RecoveryMachine
>().pg
;
6277 utime_t dur
= ceph_clock_now() - enter_time
;
6278 pg
->osd
->recoverystate_perf
->tinc(rs_reset_latency
, dur
);
6281 /*-------Start---------*/
6282 PG::RecoveryState::Start::Start(my_context ctx
)
6284 NamedState(context
< RecoveryMachine
>().pg
, "Start")
6286 context
< RecoveryMachine
>().log_enter(state_name
);
6288 PG
*pg
= context
< RecoveryMachine
>().pg
;
6289 if (pg
->is_primary()) {
6290 ldout(pg
->cct
, 1) << "transitioning to Primary" << dendl
;
6291 post_event(MakePrimary());
6293 ldout(pg
->cct
, 1) << "transitioning to Stray" << dendl
;
6294 post_event(MakeStray());
6298 void PG::RecoveryState::Start::exit()
6300 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6301 PG
*pg
= context
< RecoveryMachine
>().pg
;
6302 utime_t dur
= ceph_clock_now() - enter_time
;
6303 pg
->osd
->recoverystate_perf
->tinc(rs_start_latency
, dur
);
6306 /*---------Primary--------*/
6307 PG::RecoveryState::Primary::Primary(my_context ctx
)
6309 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary")
6311 context
< RecoveryMachine
>().log_enter(state_name
);
6312 PG
*pg
= context
< RecoveryMachine
>().pg
;
6313 assert(pg
->want_acting
.empty());
6315 // set CREATING bit until we have peered for the first time.
6316 if (pg
->info
.history
.last_epoch_started
== 0) {
6317 pg
->state_set(PG_STATE_CREATING
);
6318 // use the history timestamp, which ultimately comes from the
6319 // monitor in the create case.
6320 utime_t t
= pg
->info
.history
.last_scrub_stamp
;
6321 pg
->info
.stats
.last_fresh
= t
;
6322 pg
->info
.stats
.last_active
= t
;
6323 pg
->info
.stats
.last_change
= t
;
6324 pg
->info
.stats
.last_peered
= t
;
6325 pg
->info
.stats
.last_clean
= t
;
6326 pg
->info
.stats
.last_unstale
= t
;
6327 pg
->info
.stats
.last_undegraded
= t
;
6328 pg
->info
.stats
.last_fullsized
= t
;
6329 pg
->info
.stats
.last_scrub_stamp
= t
;
6330 pg
->info
.stats
.last_deep_scrub_stamp
= t
;
6331 pg
->info
.stats
.last_clean_scrub_stamp
= t
;
6335 boost::statechart::result
PG::RecoveryState::Primary::react(const MNotifyRec
& notevt
)
6337 PG
*pg
= context
< RecoveryMachine
>().pg
;
6338 ldout(pg
->cct
, 7) << "handle_pg_notify from osd." << notevt
.from
<< dendl
;
6339 pg
->proc_replica_info(
6340 notevt
.from
, notevt
.notify
.info
, notevt
.notify
.epoch_sent
);
6341 return discard_event();
6344 boost::statechart::result
PG::RecoveryState::Primary::react(const ActMap
&)
6346 PG
*pg
= context
< RecoveryMachine
>().pg
;
6347 ldout(pg
->cct
, 7) << "handle ActMap primary" << dendl
;
6348 pg
->publish_stats_to_osd();
6350 return discard_event();
6353 void PG::RecoveryState::Primary::exit()
6355 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6356 PG
*pg
= context
< RecoveryMachine
>().pg
;
6357 pg
->want_acting
.clear();
6358 utime_t dur
= ceph_clock_now() - enter_time
;
6359 pg
->osd
->recoverystate_perf
->tinc(rs_primary_latency
, dur
);
6360 pg
->clear_primary_state();
6361 pg
->state_clear(PG_STATE_CREATING
);
6364 /*---------Peering--------*/
6365 PG::RecoveryState::Peering::Peering(my_context ctx
)
6367 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering"),
6368 history_les_bound(false)
6370 context
< RecoveryMachine
>().log_enter(state_name
);
6372 PG
*pg
= context
< RecoveryMachine
>().pg
;
6373 assert(!pg
->is_peered());
6374 assert(!pg
->is_peering());
6375 assert(pg
->is_primary());
6376 pg
->state_set(PG_STATE_PEERING
);
6379 boost::statechart::result
PG::RecoveryState::Peering::react(const AdvMap
& advmap
)
6381 PG
*pg
= context
< RecoveryMachine
>().pg
;
6382 ldout(pg
->cct
, 10) << "Peering advmap" << dendl
;
6383 if (prior_set
.affected_by_map(*(advmap
.osdmap
), pg
)) {
6384 ldout(pg
->cct
, 1) << "Peering, affected_by_map, going to Reset" << dendl
;
6386 return transit
< Reset
>();
6389 pg
->adjust_need_up_thru(advmap
.osdmap
);
6391 return forward_event();
6394 boost::statechart::result
PG::RecoveryState::Peering::react(const QueryState
& q
)
6396 PG
*pg
= context
< RecoveryMachine
>().pg
;
6398 q
.f
->open_object_section("state");
6399 q
.f
->dump_string("name", state_name
);
6400 q
.f
->dump_stream("enter_time") << enter_time
;
6402 q
.f
->open_array_section("past_intervals");
6403 pg
->past_intervals
.dump(q
.f
);
6404 q
.f
->close_section();
6406 q
.f
->open_array_section("probing_osds");
6407 for (set
<pg_shard_t
>::iterator p
= prior_set
.probe
.begin();
6408 p
!= prior_set
.probe
.end();
6410 q
.f
->dump_stream("osd") << *p
;
6411 q
.f
->close_section();
6413 if (prior_set
.pg_down
)
6414 q
.f
->dump_string("blocked", "peering is blocked due to down osds");
6416 q
.f
->open_array_section("down_osds_we_would_probe");
6417 for (set
<int>::iterator p
= prior_set
.down
.begin();
6418 p
!= prior_set
.down
.end();
6420 q
.f
->dump_int("osd", *p
);
6421 q
.f
->close_section();
6423 q
.f
->open_array_section("peering_blocked_by");
6424 for (map
<int,epoch_t
>::iterator p
= prior_set
.blocked_by
.begin();
6425 p
!= prior_set
.blocked_by
.end();
6427 q
.f
->open_object_section("osd");
6428 q
.f
->dump_int("osd", p
->first
);
6429 q
.f
->dump_int("current_lost_at", p
->second
);
6430 q
.f
->dump_string("comment", "starting or marking this osd lost may let us proceed");
6431 q
.f
->close_section();
6433 q
.f
->close_section();
6435 if (history_les_bound
) {
6436 q
.f
->open_array_section("peering_blocked_by_detail");
6437 q
.f
->open_object_section("item");
6438 q
.f
->dump_string("detail","peering_blocked_by_history_les_bound");
6439 q
.f
->close_section();
6440 q
.f
->close_section();
6443 q
.f
->close_section();
6444 return forward_event();
6447 void PG::RecoveryState::Peering::exit()
6449 PG
*pg
= context
< RecoveryMachine
>().pg
;
6450 ldout(pg
->cct
, 10) << "Leaving Peering" << dendl
;
6451 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6452 pg
->state_clear(PG_STATE_PEERING
);
6453 pg
->clear_probe_targets();
6455 utime_t dur
= ceph_clock_now() - enter_time
;
6456 pg
->osd
->recoverystate_perf
->tinc(rs_peering_latency
, dur
);
6460 /*------Backfilling-------*/
6461 PG::RecoveryState::Backfilling::Backfilling(my_context ctx
)
6463 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Backfilling")
6465 context
< RecoveryMachine
>().log_enter(state_name
);
6466 PG
*pg
= context
< RecoveryMachine
>().pg
;
6467 pg
->backfill_reserved
= true;
6468 pg
->queue_recovery();
6469 pg
->state_clear(PG_STATE_BACKFILL_TOOFULL
);
6470 pg
->state_clear(PG_STATE_BACKFILL_WAIT
);
6471 pg
->state_set(PG_STATE_BACKFILLING
);
6472 pg
->publish_stats_to_osd();
6475 boost::statechart::result
6476 PG::RecoveryState::Backfilling::react(const DeferBackfill
&c
)
6478 PG
*pg
= context
< RecoveryMachine
>().pg
;
6479 ldout(pg
->cct
, 10) << "defer backfill, retry delay " << c
.delay
<< dendl
;
6480 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
6482 pg
->state_set(PG_STATE_BACKFILL_WAIT
);
6483 pg
->state_clear(PG_STATE_BACKFILLING
);
6485 for (set
<pg_shard_t
>::iterator it
= pg
->backfill_targets
.begin();
6486 it
!= pg
->backfill_targets
.end();
6488 assert(*it
!= pg
->pg_whoami
);
6489 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6490 it
->osd
, pg
->get_osdmap()->get_epoch());
6492 pg
->osd
->send_message_osd_cluster(
6493 new MBackfillReserve(
6494 MBackfillReserve::REJECT
,
6495 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
6496 pg
->get_osdmap()->get_epoch()),
6502 if (!pg
->waiting_on_backfill
.empty()) {
6503 pg
->waiting_on_backfill
.clear();
6504 pg
->finish_recovery_op(hobject_t::get_max());
6507 pg
->schedule_backfill_retry(c
.delay
);
6508 return transit
<NotBackfilling
>();
6511 boost::statechart::result
6512 PG::RecoveryState::Backfilling::react(const UnfoundBackfill
&c
)
6514 PG
*pg
= context
< RecoveryMachine
>().pg
;
6515 ldout(pg
->cct
, 10) << "backfill has unfound, can't continue" << dendl
;
6516 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
6518 pg
->state_set(PG_STATE_BACKFILL_UNFOUND
);
6519 pg
->state_clear(PG_STATE_BACKFILLING
);
6521 for (set
<pg_shard_t
>::iterator it
= pg
->backfill_targets
.begin();
6522 it
!= pg
->backfill_targets
.end();
6524 assert(*it
!= pg
->pg_whoami
);
6525 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6526 it
->osd
, pg
->get_osdmap()->get_epoch());
6528 pg
->osd
->send_message_osd_cluster(
6529 new MBackfillReserve(
6530 MBackfillReserve::REJECT
,
6531 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
6532 pg
->get_osdmap()->get_epoch()),
6537 pg
->waiting_on_backfill
.clear();
6539 return transit
<NotBackfilling
>();
6542 boost::statechart::result
6543 PG::RecoveryState::Backfilling::react(const RemoteReservationRejected
&)
6545 PG
*pg
= context
< RecoveryMachine
>().pg
;
6546 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
6547 pg
->state_set(PG_STATE_BACKFILL_TOOFULL
);
6549 for (set
<pg_shard_t
>::iterator it
= pg
->backfill_targets
.begin();
6550 it
!= pg
->backfill_targets
.end();
6552 assert(*it
!= pg
->pg_whoami
);
6553 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6554 it
->osd
, pg
->get_osdmap()->get_epoch());
6556 pg
->osd
->send_message_osd_cluster(
6557 new MBackfillReserve(
6558 MBackfillReserve::REJECT
,
6559 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
6560 pg
->get_osdmap()->get_epoch()),
6565 if (!pg
->waiting_on_backfill
.empty()) {
6566 pg
->waiting_on_backfill
.clear();
6567 pg
->finish_recovery_op(hobject_t::get_max());
6570 pg
->schedule_backfill_retry(pg
->cct
->_conf
->osd_recovery_retry_interval
);
6571 return transit
<NotBackfilling
>();
6574 void PG::RecoveryState::Backfilling::exit()
6576 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6577 PG
*pg
= context
< RecoveryMachine
>().pg
;
6578 pg
->backfill_reserved
= false;
6579 pg
->backfill_reserving
= false;
6580 pg
->state_clear(PG_STATE_BACKFILLING
);
6581 pg
->state_clear(PG_STATE_FORCED_BACKFILL
| PG_STATE_FORCED_RECOVERY
);
6582 utime_t dur
= ceph_clock_now() - enter_time
;
6583 pg
->osd
->recoverystate_perf
->tinc(rs_backfilling_latency
, dur
);
6586 /*--WaitRemoteBackfillReserved--*/
6588 PG::RecoveryState::WaitRemoteBackfillReserved::WaitRemoteBackfillReserved(my_context ctx
)
6590 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitRemoteBackfillReserved"),
6591 backfill_osd_it(context
< Active
>().remote_shards_to_reserve_backfill
.begin())
6593 context
< RecoveryMachine
>().log_enter(state_name
);
6594 PG
*pg
= context
< RecoveryMachine
>().pg
;
6595 pg
->state_set(PG_STATE_BACKFILL_WAIT
);
6596 pg
->publish_stats_to_osd();
6597 post_event(RemoteBackfillReserved());
6600 boost::statechart::result
6601 PG::RecoveryState::WaitRemoteBackfillReserved::react(const RemoteBackfillReserved
&evt
)
6603 PG
*pg
= context
< RecoveryMachine
>().pg
;
6605 if (backfill_osd_it
!= context
< Active
>().remote_shards_to_reserve_backfill
.end()) {
6606 //The primary never backfills itself
6607 assert(*backfill_osd_it
!= pg
->pg_whoami
);
6608 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6609 backfill_osd_it
->osd
, pg
->get_osdmap()->get_epoch());
6611 pg
->osd
->send_message_osd_cluster(
6612 new MBackfillReserve(
6613 MBackfillReserve::REQUEST
,
6614 spg_t(pg
->info
.pgid
.pgid
, backfill_osd_it
->shard
),
6615 pg
->get_osdmap()->get_epoch(),
6616 pg
->get_backfill_priority()),
6621 post_event(AllBackfillsReserved());
6623 return discard_event();
6626 void PG::RecoveryState::WaitRemoteBackfillReserved::exit()
6628 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6629 PG
*pg
= context
< RecoveryMachine
>().pg
;
6630 utime_t dur
= ceph_clock_now() - enter_time
;
6631 pg
->osd
->recoverystate_perf
->tinc(rs_waitremotebackfillreserved_latency
, dur
);
6634 boost::statechart::result
6635 PG::RecoveryState::WaitRemoteBackfillReserved::react(const RemoteReservationRejected
&evt
)
6637 PG
*pg
= context
< RecoveryMachine
>().pg
;
6638 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
6640 // Send REJECT to all previously acquired reservations
6641 set
<pg_shard_t
>::const_iterator it
, begin
, end
, next
;
6642 begin
= context
< Active
>().remote_shards_to_reserve_backfill
.begin();
6643 end
= context
< Active
>().remote_shards_to_reserve_backfill
.end();
6644 assert(begin
!= end
);
6645 for (next
= it
= begin
, ++next
; next
!= backfill_osd_it
; ++it
, ++next
) {
6646 //The primary never backfills itself
6647 assert(*it
!= pg
->pg_whoami
);
6648 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
6649 it
->osd
, pg
->get_osdmap()->get_epoch());
6651 pg
->osd
->send_message_osd_cluster(
6652 new MBackfillReserve(
6653 MBackfillReserve::REJECT
,
6654 spg_t(pg
->info
.pgid
.pgid
, it
->shard
),
6655 pg
->get_osdmap()->get_epoch()),
6660 pg
->state_clear(PG_STATE_BACKFILL_WAIT
);
6661 pg
->state_set(PG_STATE_BACKFILL_TOOFULL
);
6662 pg
->publish_stats_to_osd();
6664 pg
->schedule_backfill_retry(pg
->cct
->_conf
->osd_recovery_retry_interval
);
6666 return transit
<NotBackfilling
>();
6669 /*--WaitLocalBackfillReserved--*/
6670 PG::RecoveryState::WaitLocalBackfillReserved::WaitLocalBackfillReserved(my_context ctx
)
6672 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitLocalBackfillReserved")
6674 context
< RecoveryMachine
>().log_enter(state_name
);
6675 PG
*pg
= context
< RecoveryMachine
>().pg
;
6676 pg
->state_set(PG_STATE_BACKFILL_WAIT
);
6677 pg
->osd
->local_reserver
.request_reservation(
6679 new QueuePeeringEvt
<LocalBackfillReserved
>(
6680 pg
, pg
->get_osdmap()->get_epoch(),
6681 LocalBackfillReserved()),
6682 pg
->get_backfill_priority(),
6683 new QueuePeeringEvt
<DeferBackfill
>(
6684 pg
, pg
->get_osdmap()->get_epoch(),
6685 DeferBackfill(0.0)));
6686 pg
->publish_stats_to_osd();
6689 void PG::RecoveryState::WaitLocalBackfillReserved::exit()
6691 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6692 PG
*pg
= context
< RecoveryMachine
>().pg
;
6693 utime_t dur
= ceph_clock_now() - enter_time
;
6694 pg
->osd
->recoverystate_perf
->tinc(rs_waitlocalbackfillreserved_latency
, dur
);
6697 /*----NotBackfilling------*/
6698 PG::RecoveryState::NotBackfilling::NotBackfilling(my_context ctx
)
6700 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/NotBackfilling")
6702 context
< RecoveryMachine
>().log_enter(state_name
);
6703 PG
*pg
= context
< RecoveryMachine
>().pg
;
6704 pg
->publish_stats_to_osd();
6707 boost::statechart::result
6708 PG::RecoveryState::NotBackfilling::react(const RemoteBackfillReserved
&evt
)
6710 return discard_event();
6713 boost::statechart::result
6714 PG::RecoveryState::NotBackfilling::react(const RemoteReservationRejected
&evt
)
6716 return discard_event();
6719 void PG::RecoveryState::NotBackfilling::exit()
6721 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6722 PG
*pg
= context
< RecoveryMachine
>().pg
;
6723 pg
->state_clear(PG_STATE_BACKFILL_UNFOUND
);
6724 utime_t dur
= ceph_clock_now() - enter_time
;
6725 pg
->osd
->recoverystate_perf
->tinc(rs_notbackfilling_latency
, dur
);
6728 /*----NotRecovering------*/
6729 PG::RecoveryState::NotRecovering::NotRecovering(my_context ctx
)
6731 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/NotRecovering")
6733 context
< RecoveryMachine
>().log_enter(state_name
);
6734 PG
*pg
= context
< RecoveryMachine
>().pg
;
6735 pg
->publish_stats_to_osd();
6738 void PG::RecoveryState::NotRecovering::exit()
6740 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6741 PG
*pg
= context
< RecoveryMachine
>().pg
;
6742 pg
->state_clear(PG_STATE_RECOVERY_UNFOUND
);
6743 utime_t dur
= ceph_clock_now() - enter_time
;
6744 pg
->osd
->recoverystate_perf
->tinc(rs_notrecovering_latency
, dur
);
6747 /*---RepNotRecovering----*/
6748 PG::RecoveryState::RepNotRecovering::RepNotRecovering(my_context ctx
)
6750 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepNotRecovering")
6752 context
< RecoveryMachine
>().log_enter(state_name
);
6755 boost::statechart::result
6756 PG::RecoveryState::RepNotRecovering::react(const RejectRemoteReservation
&evt
)
6758 PG
*pg
= context
< RecoveryMachine
>().pg
;
6759 pg
->reject_reservation();
6760 post_event(RemoteReservationRejected());
6761 return discard_event();
6764 void PG::RecoveryState::RepNotRecovering::exit()
6766 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6767 PG
*pg
= context
< RecoveryMachine
>().pg
;
6768 utime_t dur
= ceph_clock_now() - enter_time
;
6769 pg
->osd
->recoverystate_perf
->tinc(rs_repnotrecovering_latency
, dur
);
6772 /*---RepWaitRecoveryReserved--*/
6773 PG::RecoveryState::RepWaitRecoveryReserved::RepWaitRecoveryReserved(my_context ctx
)
6775 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepWaitRecoveryReserved")
6777 context
< RecoveryMachine
>().log_enter(state_name
);
6778 PG
*pg
= context
< RecoveryMachine
>().pg
;
6780 pg
->osd
->remote_reserver
.request_reservation(
6782 new QueuePeeringEvt
<RemoteRecoveryReserved
>(
6783 pg
, pg
->get_osdmap()->get_epoch(),
6784 RemoteRecoveryReserved()),
6785 pg
->get_recovery_priority());
6788 boost::statechart::result
6789 PG::RecoveryState::RepWaitRecoveryReserved::react(const RemoteRecoveryReserved
&evt
)
6791 PG
*pg
= context
< RecoveryMachine
>().pg
;
6792 pg
->osd
->send_message_osd_cluster(
6794 new MRecoveryReserve(
6795 MRecoveryReserve::GRANT
,
6796 spg_t(pg
->info
.pgid
.pgid
, pg
->primary
.shard
),
6797 pg
->get_osdmap()->get_epoch()),
6798 pg
->get_osdmap()->get_epoch());
6799 return transit
<RepRecovering
>();
6802 boost::statechart::result
6803 PG::RecoveryState::RepWaitRecoveryReserved::react(
6804 const RemoteReservationCanceled
&evt
)
6806 PG
*pg
= context
< RecoveryMachine
>().pg
;
6807 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
6808 return transit
<RepNotRecovering
>();
6811 void PG::RecoveryState::RepWaitRecoveryReserved::exit()
6813 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6814 PG
*pg
= context
< RecoveryMachine
>().pg
;
6815 utime_t dur
= ceph_clock_now() - enter_time
;
6816 pg
->osd
->recoverystate_perf
->tinc(rs_repwaitrecoveryreserved_latency
, dur
);
6819 /*-RepWaitBackfillReserved*/
6820 PG::RecoveryState::RepWaitBackfillReserved::RepWaitBackfillReserved(my_context ctx
)
6822 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepWaitBackfillReserved")
6824 context
< RecoveryMachine
>().log_enter(state_name
);
6827 boost::statechart::result
6828 PG::RecoveryState::RepNotRecovering::react(const RequestBackfillPrio
&evt
)
6830 PG
*pg
= context
< RecoveryMachine
>().pg
;
6833 if (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
> 0 &&
6834 (rand()%1000 < (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
*1000.0))) {
6835 ldout(pg
->cct
, 10) << "backfill reservation rejected: failure injection"
6837 post_event(RejectRemoteReservation());
6838 } else if (!pg
->cct
->_conf
->osd_debug_skip_full_check_in_backfill_reservation
&&
6839 pg
->osd
->check_backfill_full(ss
)) {
6840 ldout(pg
->cct
, 10) << "backfill reservation rejected: "
6841 << ss
.str() << dendl
;
6842 post_event(RejectRemoteReservation());
6844 pg
->osd
->remote_reserver
.request_reservation(
6846 new QueuePeeringEvt
<RemoteBackfillReserved
>(
6847 pg
, pg
->get_osdmap()->get_epoch(),
6848 RemoteBackfillReserved()), evt
.priority
);
6850 return transit
<RepWaitBackfillReserved
>();
6853 void PG::RecoveryState::RepWaitBackfillReserved::exit()
6855 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6856 PG
*pg
= context
< RecoveryMachine
>().pg
;
6857 utime_t dur
= ceph_clock_now() - enter_time
;
6858 pg
->osd
->recoverystate_perf
->tinc(rs_repwaitbackfillreserved_latency
, dur
);
6861 boost::statechart::result
6862 PG::RecoveryState::RepWaitBackfillReserved::react(const RemoteBackfillReserved
&evt
)
6864 PG
*pg
= context
< RecoveryMachine
>().pg
;
6867 if (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
> 0 &&
6868 (rand()%1000 < (pg
->cct
->_conf
->osd_debug_reject_backfill_probability
*1000.0))) {
6869 ldout(pg
->cct
, 10) << "backfill reservation rejected after reservation: "
6870 << "failure injection" << dendl
;
6871 post_event(RejectRemoteReservation());
6872 return discard_event();
6873 } else if (!pg
->cct
->_conf
->osd_debug_skip_full_check_in_backfill_reservation
&&
6874 pg
->osd
->check_backfill_full(ss
)) {
6875 ldout(pg
->cct
, 10) << "backfill reservation rejected after reservation: "
6876 << ss
.str() << dendl
;
6877 post_event(RejectRemoteReservation());
6878 return discard_event();
6880 pg
->osd
->send_message_osd_cluster(
6882 new MBackfillReserve(
6883 MBackfillReserve::GRANT
,
6884 spg_t(pg
->info
.pgid
.pgid
, pg
->primary
.shard
),
6885 pg
->get_osdmap()->get_epoch()),
6886 pg
->get_osdmap()->get_epoch());
6887 return transit
<RepRecovering
>();
6891 boost::statechart::result
6892 PG::RecoveryState::RepWaitBackfillReserved::react(
6893 const RejectRemoteReservation
&evt
)
6895 PG
*pg
= context
< RecoveryMachine
>().pg
;
6896 pg
->reject_reservation();
6897 post_event(RemoteReservationRejected());
6898 return discard_event();
6901 boost::statechart::result
6902 PG::RecoveryState::RepWaitBackfillReserved::react(
6903 const RemoteReservationRejected
&evt
)
6905 PG
*pg
= context
< RecoveryMachine
>().pg
;
6906 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
6907 return transit
<RepNotRecovering
>();
6910 boost::statechart::result
6911 PG::RecoveryState::RepWaitBackfillReserved::react(
6912 const RemoteReservationCanceled
&evt
)
6914 PG
*pg
= context
< RecoveryMachine
>().pg
;
6915 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
6916 return transit
<RepNotRecovering
>();
6919 /*---RepRecovering-------*/
6920 PG::RecoveryState::RepRecovering::RepRecovering(my_context ctx
)
6922 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive/RepRecovering")
6924 context
< RecoveryMachine
>().log_enter(state_name
);
6927 boost::statechart::result
6928 PG::RecoveryState::RepRecovering::react(const BackfillTooFull
&)
6930 PG
*pg
= context
< RecoveryMachine
>().pg
;
6931 pg
->reject_reservation();
6932 return discard_event();
6935 void PG::RecoveryState::RepRecovering::exit()
6937 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6938 PG
*pg
= context
< RecoveryMachine
>().pg
;
6939 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
6940 utime_t dur
= ceph_clock_now() - enter_time
;
6941 pg
->osd
->recoverystate_perf
->tinc(rs_reprecovering_latency
, dur
);
6944 /*------Activating--------*/
6945 PG::RecoveryState::Activating::Activating(my_context ctx
)
6947 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Activating")
6949 context
< RecoveryMachine
>().log_enter(state_name
);
6952 void PG::RecoveryState::Activating::exit()
6954 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
6955 PG
*pg
= context
< RecoveryMachine
>().pg
;
6956 utime_t dur
= ceph_clock_now() - enter_time
;
6957 pg
->osd
->recoverystate_perf
->tinc(rs_activating_latency
, dur
);
6960 PG::RecoveryState::WaitLocalRecoveryReserved::WaitLocalRecoveryReserved(my_context ctx
)
6962 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitLocalRecoveryReserved")
6964 context
< RecoveryMachine
>().log_enter(state_name
);
6965 PG
*pg
= context
< RecoveryMachine
>().pg
;
6967 // Make sure all nodes that part of the recovery aren't full
6968 if (!pg
->cct
->_conf
->osd_debug_skip_full_check_in_recovery
&&
6969 pg
->osd
->check_osdmap_full(pg
->actingbackfill
)) {
6970 post_event(RecoveryTooFull());
6974 pg
->state_clear(PG_STATE_RECOVERY_TOOFULL
);
6975 pg
->state_set(PG_STATE_RECOVERY_WAIT
);
6976 pg
->osd
->local_reserver
.request_reservation(
6978 new QueuePeeringEvt
<LocalRecoveryReserved
>(
6979 pg
, pg
->get_osdmap()->get_epoch(),
6980 LocalRecoveryReserved()),
6981 pg
->get_recovery_priority(),
6982 new QueuePeeringEvt
<DeferRecovery
>(
6983 pg
, pg
->get_osdmap()->get_epoch(),
6984 DeferRecovery(0.0)));
6985 pg
->publish_stats_to_osd();
6988 boost::statechart::result
6989 PG::RecoveryState::WaitLocalRecoveryReserved::react(const RecoveryTooFull
&evt
)
6991 PG
*pg
= context
< RecoveryMachine
>().pg
;
6992 pg
->state_set(PG_STATE_RECOVERY_TOOFULL
);
6993 pg
->schedule_recovery_retry(pg
->cct
->_conf
->osd_recovery_retry_interval
);
6994 return transit
<NotRecovering
>();
6997 void PG::RecoveryState::WaitLocalRecoveryReserved::exit()
6999 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7000 PG
*pg
= context
< RecoveryMachine
>().pg
;
7001 utime_t dur
= ceph_clock_now() - enter_time
;
7002 pg
->osd
->recoverystate_perf
->tinc(rs_waitlocalrecoveryreserved_latency
, dur
);
7005 PG::RecoveryState::WaitRemoteRecoveryReserved::WaitRemoteRecoveryReserved(my_context ctx
)
7007 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/WaitRemoteRecoveryReserved"),
7008 remote_recovery_reservation_it(context
< Active
>().remote_shards_to_reserve_recovery
.begin())
7010 context
< RecoveryMachine
>().log_enter(state_name
);
7011 post_event(RemoteRecoveryReserved());
7014 boost::statechart::result
7015 PG::RecoveryState::WaitRemoteRecoveryReserved::react(const RemoteRecoveryReserved
&evt
) {
7016 PG
*pg
= context
< RecoveryMachine
>().pg
;
7018 if (remote_recovery_reservation_it
!= context
< Active
>().remote_shards_to_reserve_recovery
.end()) {
7019 assert(*remote_recovery_reservation_it
!= pg
->pg_whoami
);
7020 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
7021 remote_recovery_reservation_it
->osd
, pg
->get_osdmap()->get_epoch());
7023 pg
->osd
->send_message_osd_cluster(
7024 new MRecoveryReserve(
7025 MRecoveryReserve::REQUEST
,
7026 spg_t(pg
->info
.pgid
.pgid
, remote_recovery_reservation_it
->shard
),
7027 pg
->get_osdmap()->get_epoch()),
7030 ++remote_recovery_reservation_it
;
7032 post_event(AllRemotesReserved());
7034 return discard_event();
7037 void PG::RecoveryState::WaitRemoteRecoveryReserved::exit()
7039 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7040 PG
*pg
= context
< RecoveryMachine
>().pg
;
7041 utime_t dur
= ceph_clock_now() - enter_time
;
7042 pg
->osd
->recoverystate_perf
->tinc(rs_waitremoterecoveryreserved_latency
, dur
);
7045 PG::RecoveryState::Recovering::Recovering(my_context ctx
)
7047 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Recovering")
7049 context
< RecoveryMachine
>().log_enter(state_name
);
7051 PG
*pg
= context
< RecoveryMachine
>().pg
;
7052 pg
->state_clear(PG_STATE_RECOVERY_WAIT
);
7053 pg
->state_clear(PG_STATE_RECOVERY_TOOFULL
);
7054 pg
->state_set(PG_STATE_RECOVERING
);
7055 assert(!pg
->state_test(PG_STATE_ACTIVATING
));
7056 pg
->publish_stats_to_osd();
7057 pg
->queue_recovery();
7060 void PG::RecoveryState::Recovering::release_reservations(bool cancel
)
7062 PG
*pg
= context
< RecoveryMachine
>().pg
;
7063 assert(cancel
|| !pg
->pg_log
.get_missing().have_missing());
7065 // release remote reservations
7066 for (set
<pg_shard_t
>::const_iterator i
=
7067 context
< Active
>().remote_shards_to_reserve_recovery
.begin();
7068 i
!= context
< Active
>().remote_shards_to_reserve_recovery
.end();
7070 if (*i
== pg
->pg_whoami
) // skip myself
7072 ConnectionRef con
= pg
->osd
->get_con_osd_cluster(
7073 i
->osd
, pg
->get_osdmap()->get_epoch());
7075 pg
->osd
->send_message_osd_cluster(
7076 new MRecoveryReserve(
7077 MRecoveryReserve::RELEASE
,
7078 spg_t(pg
->info
.pgid
.pgid
, i
->shard
),
7079 pg
->get_osdmap()->get_epoch()),
7085 boost::statechart::result
7086 PG::RecoveryState::Recovering::react(const AllReplicasRecovered
&evt
)
7088 PG
*pg
= context
< RecoveryMachine
>().pg
;
7089 pg
->state_clear(PG_STATE_RECOVERING
);
7090 pg
->state_clear(PG_STATE_FORCED_RECOVERY
);
7091 release_reservations();
7092 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7093 return transit
<Recovered
>();
7096 boost::statechart::result
7097 PG::RecoveryState::Recovering::react(const RequestBackfill
&evt
)
7099 PG
*pg
= context
< RecoveryMachine
>().pg
;
7100 pg
->state_clear(PG_STATE_RECOVERING
);
7101 pg
->state_clear(PG_STATE_FORCED_RECOVERY
);
7102 release_reservations();
7103 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7104 // XXX: Is this needed?
7105 pg
->publish_stats_to_osd();
7106 return transit
<WaitLocalBackfillReserved
>();
7109 boost::statechart::result
7110 PG::RecoveryState::Recovering::react(const DeferRecovery
&evt
)
7112 PG
*pg
= context
< RecoveryMachine
>().pg
;
7113 ldout(pg
->cct
, 10) << "defer recovery, retry delay " << evt
.delay
<< dendl
;
7114 pg
->state_clear(PG_STATE_RECOVERING
);
7115 pg
->state_set(PG_STATE_RECOVERY_WAIT
);
7116 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7117 release_reservations(true);
7118 pg
->schedule_recovery_retry(evt
.delay
);
7119 return transit
<NotRecovering
>();
7122 boost::statechart::result
7123 PG::RecoveryState::Recovering::react(const UnfoundRecovery
&evt
)
7125 PG
*pg
= context
< RecoveryMachine
>().pg
;
7126 ldout(pg
->cct
, 10) << "recovery has unfound, can't continue" << dendl
;
7127 pg
->state_set(PG_STATE_RECOVERY_UNFOUND
);
7128 pg
->state_clear(PG_STATE_RECOVERING
);
7129 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7130 release_reservations(true);
7131 return transit
<NotRecovering
>();
7134 void PG::RecoveryState::Recovering::exit()
7136 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7137 PG
*pg
= context
< RecoveryMachine
>().pg
;
7138 utime_t dur
= ceph_clock_now() - enter_time
;
7139 pg
->osd
->recoverystate_perf
->tinc(rs_recovering_latency
, dur
);
7142 PG::RecoveryState::Recovered::Recovered(my_context ctx
)
7144 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Recovered")
7146 pg_shard_t auth_log_shard
;
7148 context
< RecoveryMachine
>().log_enter(state_name
);
7150 PG
*pg
= context
< RecoveryMachine
>().pg
;
7152 assert(!pg
->needs_recovery());
7154 // if we finished backfill, all acting are active; recheck if
7155 // DEGRADED | UNDERSIZED is appropriate.
7156 assert(!pg
->actingbackfill
.empty());
7157 if (pg
->get_osdmap()->get_pg_size(pg
->info
.pgid
.pgid
) <=
7158 pg
->actingbackfill
.size()) {
7159 pg
->state_clear(PG_STATE_FORCED_BACKFILL
| PG_STATE_FORCED_RECOVERY
);
7160 pg
->publish_stats_to_osd();
7163 // trim pglog on recovered
7166 // adjust acting set? (e.g. because backfill completed...)
7167 bool history_les_bound
= false;
7168 if (pg
->acting
!= pg
->up
&& !pg
->choose_acting(auth_log_shard
,
7169 true, &history_les_bound
))
7170 assert(pg
->want_acting
.size());
7172 if (context
< Active
>().all_replicas_activated
)
7173 post_event(GoClean());
7176 void PG::RecoveryState::Recovered::exit()
7178 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7179 PG
*pg
= context
< RecoveryMachine
>().pg
;
7180 utime_t dur
= ceph_clock_now() - enter_time
;
7181 pg
->osd
->recoverystate_perf
->tinc(rs_recovered_latency
, dur
);
7184 PG::RecoveryState::Clean::Clean(my_context ctx
)
7186 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active/Clean")
7188 context
< RecoveryMachine
>().log_enter(state_name
);
7190 PG
*pg
= context
< RecoveryMachine
>().pg
;
7192 if (pg
->info
.last_complete
!= pg
->info
.last_update
) {
7195 pg
->finish_recovery(*context
< RecoveryMachine
>().get_on_safe_context_list());
7197 if (pg
->is_active()) {
7201 pg
->share_pg_info();
7202 pg
->publish_stats_to_osd();
7203 pg
->requeue_ops(pg
->waiting_for_clean_to_primary_repair
);
7206 void PG::RecoveryState::Clean::exit()
7208 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7209 PG
*pg
= context
< RecoveryMachine
>().pg
;
7210 pg
->state_clear(PG_STATE_CLEAN
);
7211 utime_t dur
= ceph_clock_now() - enter_time
;
7212 pg
->osd
->recoverystate_perf
->tinc(rs_clean_latency
, dur
);
7215 template <typename T
>
7216 set
<pg_shard_t
> unique_osd_shard_set(const pg_shard_t
& skip
, const T
&in
)
7218 set
<int> osds_found
;
7219 set
<pg_shard_t
> out
;
7220 for (typename
T::const_iterator i
= in
.begin();
7223 if (*i
!= skip
&& !osds_found
.count(i
->osd
)) {
7224 osds_found
.insert(i
->osd
);
7231 /*---------Active---------*/
7232 PG::RecoveryState::Active::Active(my_context ctx
)
7234 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Active"),
7235 remote_shards_to_reserve_recovery(
7236 unique_osd_shard_set(
7237 context
< RecoveryMachine
>().pg
->pg_whoami
,
7238 context
< RecoveryMachine
>().pg
->actingbackfill
)),
7239 remote_shards_to_reserve_backfill(
7240 unique_osd_shard_set(
7241 context
< RecoveryMachine
>().pg
->pg_whoami
,
7242 context
< RecoveryMachine
>().pg
->backfill_targets
)),
7243 all_replicas_activated(false)
7245 context
< RecoveryMachine
>().log_enter(state_name
);
7247 PG
*pg
= context
< RecoveryMachine
>().pg
;
7249 assert(!pg
->backfill_reserving
);
7250 assert(!pg
->backfill_reserved
);
7251 assert(pg
->is_primary());
7252 ldout(pg
->cct
, 10) << "In Active, about to call activate" << dendl
;
7254 context
< RecoveryMachine
>().get_cur_transaction(),
7255 context
< RecoveryMachine
>().get_on_applied_context_list(),
7256 context
< RecoveryMachine
>().get_on_safe_context_list());
7257 pg
->activate(*context
< RecoveryMachine
>().get_cur_transaction(),
7258 pg
->get_osdmap()->get_epoch(),
7259 *context
< RecoveryMachine
>().get_on_safe_context_list(),
7260 *context
< RecoveryMachine
>().get_query_map(),
7261 context
< RecoveryMachine
>().get_info_map(),
7262 context
< RecoveryMachine
>().get_recovery_ctx());
7264 // everyone has to commit/ack before we are truly active
7265 pg
->blocked_by
.clear();
7266 for (set
<pg_shard_t
>::iterator p
= pg
->actingbackfill
.begin();
7267 p
!= pg
->actingbackfill
.end();
7269 if (p
->shard
!= pg
->pg_whoami
.shard
) {
7270 pg
->blocked_by
.insert(p
->shard
);
7273 pg
->publish_stats_to_osd();
7274 ldout(pg
->cct
, 10) << "Activate Finished" << dendl
;
7277 boost::statechart::result
PG::RecoveryState::Active::react(const AdvMap
& advmap
)
7279 PG
*pg
= context
< RecoveryMachine
>().pg
;
7280 ldout(pg
->cct
, 10) << "Active advmap" << dendl
;
7281 if (!pg
->pool
.newly_removed_snaps
.empty()) {
7282 pg
->snap_trimq
.union_of(pg
->pool
.newly_removed_snaps
);
7283 ldout(pg
->cct
, 10) << *pg
<< " snap_trimq now " << pg
->snap_trimq
<< dendl
;
7284 pg
->dirty_info
= true;
7285 pg
->dirty_big_info
= true;
7288 for (size_t i
= 0; i
< pg
->want_acting
.size(); i
++) {
7289 int osd
= pg
->want_acting
[i
];
7290 if (!advmap
.osdmap
->is_up(osd
)) {
7291 pg_shard_t
osd_with_shard(osd
, shard_id_t(i
));
7292 assert(pg
->is_acting(osd_with_shard
) || pg
->is_up(osd_with_shard
));
7296 bool need_publish
= false;
7297 /* Check for changes in pool size (if the acting set changed as a result,
7298 * this does not matter) */
7299 if (advmap
.lastmap
->get_pg_size(pg
->info
.pgid
.pgid
) !=
7300 pg
->get_osdmap()->get_pg_size(pg
->info
.pgid
.pgid
)) {
7301 if (pg
->get_osdmap()->get_pg_size(pg
->info
.pgid
.pgid
) <= pg
->actingset
.size()) {
7302 pg
->state_clear(PG_STATE_UNDERSIZED
);
7304 pg
->state_set(PG_STATE_UNDERSIZED
);
7306 // degraded changes will be detected by call from publish_stats_to_osd()
7307 need_publish
= true;
7310 // if we haven't reported our PG stats in a long time, do so now.
7311 if (pg
->info
.stats
.reported_epoch
+ pg
->cct
->_conf
->osd_pg_stat_report_interval_max
< advmap
.osdmap
->get_epoch()) {
7312 ldout(pg
->cct
, 20) << "reporting stats to osd after " << (advmap
.osdmap
->get_epoch() - pg
->info
.stats
.reported_epoch
)
7313 << " epochs" << dendl
;
7314 need_publish
= true;
7318 pg
->publish_stats_to_osd();
7320 return forward_event();
7323 boost::statechart::result
PG::RecoveryState::Active::react(const ActMap
&)
7325 PG
*pg
= context
< RecoveryMachine
>().pg
;
7326 ldout(pg
->cct
, 10) << "Active: handling ActMap" << dendl
;
7327 assert(pg
->is_primary());
7329 if (pg
->have_unfound()) {
7330 // object may have become unfound
7331 pg
->discover_all_missing(*context
< RecoveryMachine
>().get_query_map());
7334 if (pg
->cct
->_conf
->osd_check_for_log_corruption
)
7335 pg
->check_log_for_corruption(pg
->osd
->store
);
7337 uint64_t unfound
= pg
->missing_loc
.num_unfound();
7339 pg
->all_unfound_are_queried_or_lost(pg
->get_osdmap())) {
7340 if (pg
->cct
->_conf
->osd_auto_mark_unfound_lost
) {
7341 pg
->osd
->clog
->error() << pg
->info
.pgid
.pgid
<< " has " << unfound
7342 << " objects unfound and apparently lost, would automatically "
7343 << "mark these objects lost but this feature is not yet implemented "
7344 << "(osd_auto_mark_unfound_lost)";
7346 pg
->osd
->clog
->error() << pg
->info
.pgid
.pgid
<< " has "
7347 << unfound
<< " objects unfound and apparently lost";
7350 if (pg
->is_active()) {
7351 ldout(pg
->cct
, 10) << "Active: kicking snap trim" << dendl
;
7352 pg
->kick_snap_trim();
7355 if (pg
->is_peered() &&
7357 !pg
->get_osdmap()->test_flag(CEPH_OSDMAP_NOBACKFILL
) &&
7358 (!pg
->get_osdmap()->test_flag(CEPH_OSDMAP_NOREBALANCE
) || pg
->is_degraded())) {
7359 pg
->queue_recovery();
7361 return forward_event();
7364 boost::statechart::result
PG::RecoveryState::Active::react(const MNotifyRec
& notevt
)
7366 PG
*pg
= context
< RecoveryMachine
>().pg
;
7367 assert(pg
->is_primary());
7368 if (pg
->peer_info
.count(notevt
.from
)) {
7369 ldout(pg
->cct
, 10) << "Active: got notify from " << notevt
.from
7370 << ", already have info from that osd, ignoring"
7372 } else if (pg
->peer_purged
.count(notevt
.from
)) {
7373 ldout(pg
->cct
, 10) << "Active: got notify from " << notevt
.from
7374 << ", already purged that peer, ignoring"
7377 ldout(pg
->cct
, 10) << "Active: got notify from " << notevt
.from
7378 << ", calling proc_replica_info and discover_all_missing"
7380 pg
->proc_replica_info(
7381 notevt
.from
, notevt
.notify
.info
, notevt
.notify
.epoch_sent
);
7382 if (pg
->have_unfound()) {
7383 pg
->discover_all_missing(*context
< RecoveryMachine
>().get_query_map());
7386 return discard_event();
7389 boost::statechart::result
PG::RecoveryState::Active::react(const MInfoRec
& infoevt
)
7391 PG
*pg
= context
< RecoveryMachine
>().pg
;
7392 assert(pg
->is_primary());
7394 assert(!pg
->actingbackfill
.empty());
7395 // don't update history (yet) if we are active and primary; the replica
7396 // may be telling us they have activated (and committed) but we can't
7397 // share that until _everyone_ does the same.
7398 if (pg
->is_actingbackfill(infoevt
.from
)) {
7399 ldout(pg
->cct
, 10) << " peer osd." << infoevt
.from
7400 << " activated and committed" << dendl
;
7401 pg
->peer_activated
.insert(infoevt
.from
);
7402 pg
->blocked_by
.erase(infoevt
.from
.shard
);
7403 pg
->publish_stats_to_osd();
7404 if (pg
->peer_activated
.size() == pg
->actingbackfill
.size()) {
7405 pg
->all_activated_and_committed();
7408 return discard_event();
7411 boost::statechart::result
PG::RecoveryState::Active::react(const MLogRec
& logevt
)
7413 PG
*pg
= context
< RecoveryMachine
>().pg
;
7414 ldout(pg
->cct
, 10) << "searching osd." << logevt
.from
7415 << " log for unfound items" << dendl
;
7416 pg
->proc_replica_log(
7417 logevt
.msg
->info
, logevt
.msg
->log
, logevt
.msg
->missing
, logevt
.from
);
7418 bool got_missing
= pg
->search_for_missing(
7419 pg
->peer_info
[logevt
.from
],
7420 pg
->peer_missing
[logevt
.from
],
7422 context
< RecoveryMachine
>().get_recovery_ctx());
7423 // If there are missing AND we are "fully" active then start recovery now
7424 if (got_missing
&& pg
->state_test(PG_STATE_ACTIVE
)) {
7425 post_event(DoRecovery());
7427 return discard_event();
7430 boost::statechart::result
PG::RecoveryState::Active::react(const QueryState
& q
)
7432 PG
*pg
= context
< RecoveryMachine
>().pg
;
7434 q
.f
->open_object_section("state");
7435 q
.f
->dump_string("name", state_name
);
7436 q
.f
->dump_stream("enter_time") << enter_time
;
7439 q
.f
->open_array_section("might_have_unfound");
7440 for (set
<pg_shard_t
>::iterator p
= pg
->might_have_unfound
.begin();
7441 p
!= pg
->might_have_unfound
.end();
7443 q
.f
->open_object_section("osd");
7444 q
.f
->dump_stream("osd") << *p
;
7445 if (pg
->peer_missing
.count(*p
)) {
7446 q
.f
->dump_string("status", "already probed");
7447 } else if (pg
->peer_missing_requested
.count(*p
)) {
7448 q
.f
->dump_string("status", "querying");
7449 } else if (!pg
->get_osdmap()->is_up(p
->osd
)) {
7450 q
.f
->dump_string("status", "osd is down");
7452 q
.f
->dump_string("status", "not queried");
7454 q
.f
->close_section();
7456 q
.f
->close_section();
7459 q
.f
->open_object_section("recovery_progress");
7460 pg
->dump_recovery_info(q
.f
);
7461 q
.f
->close_section();
7465 q
.f
->open_object_section("scrub");
7466 q
.f
->dump_stream("scrubber.epoch_start") << pg
->scrubber
.epoch_start
;
7467 q
.f
->dump_bool("scrubber.active", pg
->scrubber
.active
);
7468 q
.f
->dump_string("scrubber.state", Scrubber::state_string(pg
->scrubber
.state
));
7469 q
.f
->dump_stream("scrubber.start") << pg
->scrubber
.start
;
7470 q
.f
->dump_stream("scrubber.end") << pg
->scrubber
.end
;
7471 q
.f
->dump_stream("scrubber.subset_last_update") << pg
->scrubber
.subset_last_update
;
7472 q
.f
->dump_bool("scrubber.deep", pg
->scrubber
.deep
);
7473 q
.f
->dump_unsigned("scrubber.seed", pg
->scrubber
.seed
);
7474 q
.f
->dump_int("scrubber.waiting_on", pg
->scrubber
.waiting_on
);
7476 q
.f
->open_array_section("scrubber.waiting_on_whom");
7477 for (set
<pg_shard_t
>::iterator p
= pg
->scrubber
.waiting_on_whom
.begin();
7478 p
!= pg
->scrubber
.waiting_on_whom
.end();
7480 q
.f
->dump_stream("shard") << *p
;
7482 q
.f
->close_section();
7484 q
.f
->close_section();
7487 q
.f
->close_section();
7488 return forward_event();
7491 boost::statechart::result
PG::RecoveryState::Active::react(const AllReplicasActivated
&evt
)
7493 PG
*pg
= context
< RecoveryMachine
>().pg
;
7494 all_replicas_activated
= true;
7496 pg
->state_clear(PG_STATE_ACTIVATING
);
7497 pg
->state_clear(PG_STATE_CREATING
);
7498 if (pg
->acting
.size() >= pg
->pool
.info
.min_size
) {
7499 pg
->state_set(PG_STATE_ACTIVE
);
7501 pg
->state_set(PG_STATE_PEERED
);
7504 // info.last_epoch_started is set during activate()
7505 pg
->info
.history
.last_epoch_started
= pg
->info
.last_epoch_started
;
7506 pg
->info
.history
.last_interval_started
= pg
->info
.last_interval_started
;
7507 pg
->dirty_info
= true;
7509 pg
->share_pg_info();
7510 pg
->publish_stats_to_osd();
7515 if (pg
->flushes_in_progress
== 0) {
7516 pg
->requeue_ops(pg
->waiting_for_peered
);
7517 } else if (!pg
->waiting_for_peered
.empty()) {
7518 ldout(pg
->cct
, 10) << __func__
<< " flushes in progress, moving "
7519 << pg
->waiting_for_peered
.size()
7520 << " items to waiting_for_flush"
7522 assert(pg
->waiting_for_flush
.empty());
7523 pg
->waiting_for_flush
.swap(pg
->waiting_for_peered
);
7528 return discard_event();
7531 void PG::RecoveryState::Active::exit()
7533 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7534 PG
*pg
= context
< RecoveryMachine
>().pg
;
7535 pg
->osd
->local_reserver
.cancel_reservation(pg
->info
.pgid
);
7537 pg
->blocked_by
.clear();
7538 pg
->backfill_reserved
= false;
7539 pg
->backfill_reserving
= false;
7540 pg
->state_clear(PG_STATE_ACTIVATING
);
7541 pg
->state_clear(PG_STATE_DEGRADED
);
7542 pg
->state_clear(PG_STATE_UNDERSIZED
);
7543 pg
->state_clear(PG_STATE_BACKFILL_TOOFULL
);
7544 pg
->state_clear(PG_STATE_BACKFILL_WAIT
);
7545 pg
->state_clear(PG_STATE_RECOVERY_WAIT
);
7546 pg
->state_clear(PG_STATE_RECOVERY_TOOFULL
);
7547 utime_t dur
= ceph_clock_now() - enter_time
;
7548 pg
->osd
->recoverystate_perf
->tinc(rs_active_latency
, dur
);
7552 /*------ReplicaActive-----*/
7553 PG::RecoveryState::ReplicaActive::ReplicaActive(my_context ctx
)
7555 NamedState(context
< RecoveryMachine
>().pg
, "Started/ReplicaActive")
7557 context
< RecoveryMachine
>().log_enter(state_name
);
7559 PG
*pg
= context
< RecoveryMachine
>().pg
;
7561 context
< RecoveryMachine
>().get_cur_transaction(),
7562 context
< RecoveryMachine
>().get_on_applied_context_list(),
7563 context
< RecoveryMachine
>().get_on_safe_context_list());
7567 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(
7568 const Activate
& actevt
) {
7569 PG
*pg
= context
< RecoveryMachine
>().pg
;
7570 ldout(pg
->cct
, 10) << "In ReplicaActive, about to call activate" << dendl
;
7571 map
<int, map
<spg_t
, pg_query_t
> > query_map
;
7572 pg
->activate(*context
< RecoveryMachine
>().get_cur_transaction(),
7573 actevt
.activation_epoch
,
7574 *context
< RecoveryMachine
>().get_on_safe_context_list(),
7575 query_map
, NULL
, NULL
);
7576 ldout(pg
->cct
, 10) << "Activate Finished" << dendl
;
7577 return discard_event();
7580 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const MInfoRec
& infoevt
)
7582 PG
*pg
= context
< RecoveryMachine
>().pg
;
7583 pg
->proc_primary_info(*context
<RecoveryMachine
>().get_cur_transaction(),
7585 return discard_event();
7588 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const MLogRec
& logevt
)
7590 PG
*pg
= context
< RecoveryMachine
>().pg
;
7591 ldout(pg
->cct
, 10) << "received log from " << logevt
.from
<< dendl
;
7592 ObjectStore::Transaction
* t
= context
<RecoveryMachine
>().get_cur_transaction();
7593 pg
->merge_log(*t
, logevt
.msg
->info
, logevt
.msg
->log
, logevt
.from
);
7594 assert(pg
->pg_log
.get_head() == pg
->info
.last_update
);
7596 return discard_event();
7599 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const ActMap
&)
7601 PG
*pg
= context
< RecoveryMachine
>().pg
;
7602 if (pg
->should_send_notify() && pg
->get_primary().osd
>= 0) {
7603 context
< RecoveryMachine
>().send_notify(
7606 pg
->get_primary().shard
, pg
->pg_whoami
.shard
,
7607 pg
->get_osdmap()->get_epoch(),
7608 pg
->get_osdmap()->get_epoch(),
7610 pg
->past_intervals
);
7613 return discard_event();
7616 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const MQuery
& query
)
7618 PG
*pg
= context
< RecoveryMachine
>().pg
;
7619 if (query
.query
.type
== pg_query_t::MISSING
) {
7620 pg
->update_history(query
.query
.history
);
7621 pg
->fulfill_log(query
.from
, query
.query
, query
.query_epoch
);
7622 } // else: from prior to activation, safe to ignore
7623 return discard_event();
7626 boost::statechart::result
PG::RecoveryState::ReplicaActive::react(const QueryState
& q
)
7628 q
.f
->open_object_section("state");
7629 q
.f
->dump_string("name", state_name
);
7630 q
.f
->dump_stream("enter_time") << enter_time
;
7631 q
.f
->close_section();
7632 return forward_event();
7635 void PG::RecoveryState::ReplicaActive::exit()
7637 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7638 PG
*pg
= context
< RecoveryMachine
>().pg
;
7639 pg
->osd
->remote_reserver
.cancel_reservation(pg
->info
.pgid
);
7640 utime_t dur
= ceph_clock_now() - enter_time
;
7641 pg
->osd
->recoverystate_perf
->tinc(rs_replicaactive_latency
, dur
);
7645 PG::RecoveryState::Stray::Stray(my_context ctx
)
7647 NamedState(context
< RecoveryMachine
>().pg
, "Started/Stray")
7649 context
< RecoveryMachine
>().log_enter(state_name
);
7651 PG
*pg
= context
< RecoveryMachine
>().pg
;
7652 assert(!pg
->is_peered());
7653 assert(!pg
->is_peering());
7654 assert(!pg
->is_primary());
7656 context
< RecoveryMachine
>().get_cur_transaction(),
7657 context
< RecoveryMachine
>().get_on_applied_context_list(),
7658 context
< RecoveryMachine
>().get_on_safe_context_list());
7661 boost::statechart::result
PG::RecoveryState::Stray::react(const MLogRec
& logevt
)
7663 PG
*pg
= context
< RecoveryMachine
>().pg
;
7664 MOSDPGLog
*msg
= logevt
.msg
.get();
7665 ldout(pg
->cct
, 10) << "got info+log from osd." << logevt
.from
<< " " << msg
->info
<< " " << msg
->log
<< dendl
;
7667 ObjectStore::Transaction
* t
= context
<RecoveryMachine
>().get_cur_transaction();
7668 if (msg
->info
.last_backfill
== hobject_t()) {
7670 pg
->unreg_next_scrub();
7671 pg
->info
= msg
->info
;
7672 pg
->reg_next_scrub();
7673 pg
->dirty_info
= true;
7674 pg
->dirty_big_info
= true; // maybe.
7676 PGLogEntryHandler rollbacker
{pg
, t
};
7677 pg
->pg_log
.reset_backfill_claim_log(msg
->log
, &rollbacker
);
7679 pg
->pg_log
.reset_backfill();
7681 pg
->merge_log(*t
, msg
->info
, msg
->log
, logevt
.from
);
7684 assert(pg
->pg_log
.get_head() == pg
->info
.last_update
);
7686 post_event(Activate(logevt
.msg
->info
.last_epoch_started
));
7687 return transit
<ReplicaActive
>();
7690 boost::statechart::result
PG::RecoveryState::Stray::react(const MInfoRec
& infoevt
)
7692 PG
*pg
= context
< RecoveryMachine
>().pg
;
7693 ldout(pg
->cct
, 10) << "got info from osd." << infoevt
.from
<< " " << infoevt
.info
<< dendl
;
7695 if (pg
->info
.last_update
> infoevt
.info
.last_update
) {
7696 // rewind divergent log entries
7697 ObjectStore::Transaction
* t
= context
<RecoveryMachine
>().get_cur_transaction();
7698 pg
->rewind_divergent_log(*t
, infoevt
.info
.last_update
);
7699 pg
->info
.stats
= infoevt
.info
.stats
;
7700 pg
->info
.hit_set
= infoevt
.info
.hit_set
;
7703 assert(infoevt
.info
.last_update
== pg
->info
.last_update
);
7704 assert(pg
->pg_log
.get_head() == pg
->info
.last_update
);
7706 post_event(Activate(infoevt
.info
.last_epoch_started
));
7707 return transit
<ReplicaActive
>();
7710 boost::statechart::result
PG::RecoveryState::Stray::react(const MQuery
& query
)
7712 PG
*pg
= context
< RecoveryMachine
>().pg
;
7713 if (query
.query
.type
== pg_query_t::INFO
) {
7714 pair
<pg_shard_t
, pg_info_t
> notify_info
;
7715 pg
->update_history(query
.query
.history
);
7716 pg
->fulfill_info(query
.from
, query
.query
, notify_info
);
7717 context
< RecoveryMachine
>().send_notify(
7720 notify_info
.first
.shard
, pg
->pg_whoami
.shard
,
7722 pg
->get_osdmap()->get_epoch(),
7723 notify_info
.second
),
7724 pg
->past_intervals
);
7726 pg
->fulfill_log(query
.from
, query
.query
, query
.query_epoch
);
7728 return discard_event();
7731 boost::statechart::result
PG::RecoveryState::Stray::react(const ActMap
&)
7733 PG
*pg
= context
< RecoveryMachine
>().pg
;
7734 if (pg
->should_send_notify() && pg
->get_primary().osd
>= 0) {
7735 context
< RecoveryMachine
>().send_notify(
7738 pg
->get_primary().shard
, pg
->pg_whoami
.shard
,
7739 pg
->get_osdmap()->get_epoch(),
7740 pg
->get_osdmap()->get_epoch(),
7742 pg
->past_intervals
);
7745 return discard_event();
7748 void PG::RecoveryState::Stray::exit()
7750 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7751 PG
*pg
= context
< RecoveryMachine
>().pg
;
7752 utime_t dur
= ceph_clock_now() - enter_time
;
7753 pg
->osd
->recoverystate_perf
->tinc(rs_stray_latency
, dur
);
7756 /*--------GetInfo---------*/
7757 PG::RecoveryState::GetInfo::GetInfo(my_context ctx
)
7759 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/GetInfo")
7761 context
< RecoveryMachine
>().log_enter(state_name
);
7763 PG
*pg
= context
< RecoveryMachine
>().pg
;
7764 pg
->check_past_interval_bounds();
7765 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
7767 assert(pg
->blocked_by
.empty());
7769 prior_set
= pg
->build_prior();
7771 pg
->reset_min_peer_features();
7773 if (prior_set
.pg_down
) {
7774 post_event(IsDown());
7775 } else if (peer_info_requested
.empty()) {
7776 post_event(GotInfo());
7780 void PG::RecoveryState::GetInfo::get_infos()
7782 PG
*pg
= context
< RecoveryMachine
>().pg
;
7783 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
7785 pg
->blocked_by
.clear();
7786 for (set
<pg_shard_t
>::const_iterator it
= prior_set
.probe
.begin();
7787 it
!= prior_set
.probe
.end();
7789 pg_shard_t peer
= *it
;
7790 if (peer
== pg
->pg_whoami
) {
7793 if (pg
->peer_info
.count(peer
)) {
7794 ldout(pg
->cct
, 10) << " have osd." << peer
<< " info " << pg
->peer_info
[peer
] << dendl
;
7797 if (peer_info_requested
.count(peer
)) {
7798 ldout(pg
->cct
, 10) << " already requested info from osd." << peer
<< dendl
;
7799 pg
->blocked_by
.insert(peer
.osd
);
7800 } else if (!pg
->get_osdmap()->is_up(peer
.osd
)) {
7801 ldout(pg
->cct
, 10) << " not querying info from down osd." << peer
<< dendl
;
7803 ldout(pg
->cct
, 10) << " querying info from osd." << peer
<< dendl
;
7804 context
< RecoveryMachine
>().send_query(
7805 peer
, pg_query_t(pg_query_t::INFO
,
7806 it
->shard
, pg
->pg_whoami
.shard
,
7808 pg
->get_osdmap()->get_epoch()));
7809 peer_info_requested
.insert(peer
);
7810 pg
->blocked_by
.insert(peer
.osd
);
7814 pg
->publish_stats_to_osd();
7817 boost::statechart::result
PG::RecoveryState::GetInfo::react(const MNotifyRec
& infoevt
)
7819 PG
*pg
= context
< RecoveryMachine
>().pg
;
7821 set
<pg_shard_t
>::iterator p
= peer_info_requested
.find(infoevt
.from
);
7822 if (p
!= peer_info_requested
.end()) {
7823 peer_info_requested
.erase(p
);
7824 pg
->blocked_by
.erase(infoevt
.from
.osd
);
7827 epoch_t old_start
= pg
->info
.history
.last_epoch_started
;
7828 if (pg
->proc_replica_info(
7829 infoevt
.from
, infoevt
.notify
.info
, infoevt
.notify
.epoch_sent
)) {
7830 // we got something new ...
7831 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
7832 if (old_start
< pg
->info
.history
.last_epoch_started
) {
7833 ldout(pg
->cct
, 10) << " last_epoch_started moved forward, rebuilding prior" << dendl
;
7834 prior_set
= pg
->build_prior();
7836 // filter out any osds that got dropped from the probe set from
7837 // peer_info_requested. this is less expensive than restarting
7838 // peering (which would re-probe everyone).
7839 set
<pg_shard_t
>::iterator p
= peer_info_requested
.begin();
7840 while (p
!= peer_info_requested
.end()) {
7841 if (prior_set
.probe
.count(*p
) == 0) {
7842 ldout(pg
->cct
, 20) << " dropping osd." << *p
<< " from info_requested, no longer in probe set" << dendl
;
7843 peer_info_requested
.erase(p
++);
7850 ldout(pg
->cct
, 20) << "Adding osd: " << infoevt
.from
.osd
<< " peer features: "
7851 << hex
<< infoevt
.features
<< dec
<< dendl
;
7852 pg
->apply_peer_features(infoevt
.features
);
7854 // are we done getting everything?
7855 if (peer_info_requested
.empty() && !prior_set
.pg_down
) {
7856 ldout(pg
->cct
, 20) << "Common peer features: " << hex
<< pg
->get_min_peer_features() << dec
<< dendl
;
7857 ldout(pg
->cct
, 20) << "Common acting features: " << hex
<< pg
->get_min_acting_features() << dec
<< dendl
;
7858 ldout(pg
->cct
, 20) << "Common upacting features: " << hex
<< pg
->get_min_upacting_features() << dec
<< dendl
;
7859 post_event(GotInfo());
7862 return discard_event();
7865 boost::statechart::result
PG::RecoveryState::GetInfo::react(const QueryState
& q
)
7867 PG
*pg
= context
< RecoveryMachine
>().pg
;
7868 q
.f
->open_object_section("state");
7869 q
.f
->dump_string("name", state_name
);
7870 q
.f
->dump_stream("enter_time") << enter_time
;
7872 q
.f
->open_array_section("requested_info_from");
7873 for (set
<pg_shard_t
>::iterator p
= peer_info_requested
.begin();
7874 p
!= peer_info_requested
.end();
7876 q
.f
->open_object_section("osd");
7877 q
.f
->dump_stream("osd") << *p
;
7878 if (pg
->peer_info
.count(*p
)) {
7879 q
.f
->open_object_section("got_info");
7880 pg
->peer_info
[*p
].dump(q
.f
);
7881 q
.f
->close_section();
7883 q
.f
->close_section();
7885 q
.f
->close_section();
7887 q
.f
->close_section();
7888 return forward_event();
7891 void PG::RecoveryState::GetInfo::exit()
7893 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
7894 PG
*pg
= context
< RecoveryMachine
>().pg
;
7895 utime_t dur
= ceph_clock_now() - enter_time
;
7896 pg
->osd
->recoverystate_perf
->tinc(rs_getinfo_latency
, dur
);
7897 pg
->blocked_by
.clear();
7898 pg
->publish_stats_to_osd();
7901 /*------GetLog------------*/
7902 PG::RecoveryState::GetLog::GetLog(my_context ctx
)
7905 context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/GetLog"),
7908 context
< RecoveryMachine
>().log_enter(state_name
);
7910 PG
*pg
= context
< RecoveryMachine
>().pg
;
7913 if (!pg
->choose_acting(auth_log_shard
, false,
7914 &context
< Peering
>().history_les_bound
)) {
7915 if (!pg
->want_acting
.empty()) {
7916 post_event(NeedActingChange());
7918 post_event(IsIncomplete());
7924 if (auth_log_shard
== pg
->pg_whoami
) {
7925 post_event(GotLog());
7929 const pg_info_t
& best
= pg
->peer_info
[auth_log_shard
];
7932 if (pg
->info
.last_update
< best
.log_tail
) {
7933 ldout(pg
->cct
, 10) << " not contiguous with osd." << auth_log_shard
<< ", down" << dendl
;
7934 post_event(IsIncomplete());
7938 // how much log to request?
7939 eversion_t request_log_from
= pg
->info
.last_update
;
7940 assert(!pg
->actingbackfill
.empty());
7941 for (set
<pg_shard_t
>::iterator p
= pg
->actingbackfill
.begin();
7942 p
!= pg
->actingbackfill
.end();
7944 if (*p
== pg
->pg_whoami
) continue;
7945 pg_info_t
& ri
= pg
->peer_info
[*p
];
7946 if (ri
.last_update
< pg
->info
.log_tail
&& ri
.last_update
>= best
.log_tail
&&
7947 ri
.last_update
< request_log_from
)
7948 request_log_from
= ri
.last_update
;
7952 ldout(pg
->cct
, 10) << " requesting log from osd." << auth_log_shard
<< dendl
;
7953 context
<RecoveryMachine
>().send_query(
7957 auth_log_shard
.shard
, pg
->pg_whoami
.shard
,
7958 request_log_from
, pg
->info
.history
,
7959 pg
->get_osdmap()->get_epoch()));
7961 assert(pg
->blocked_by
.empty());
7962 pg
->blocked_by
.insert(auth_log_shard
.osd
);
7963 pg
->publish_stats_to_osd();
7966 boost::statechart::result
PG::RecoveryState::GetLog::react(const AdvMap
& advmap
)
7968 PG
*pg
= context
< RecoveryMachine
>().pg
;
7969 // make sure our log source didn't go down. we need to check
7970 // explicitly because it may not be part of the prior set, which
7971 // means the Peering state check won't catch it going down.
7972 if (!advmap
.osdmap
->is_up(auth_log_shard
.osd
)) {
7973 ldout(pg
->cct
, 10) << "GetLog: auth_log_shard osd."
7974 << auth_log_shard
.osd
<< " went down" << dendl
;
7976 return transit
< Reset
>();
7979 // let the Peering state do its checks.
7980 return forward_event();
7983 boost::statechart::result
PG::RecoveryState::GetLog::react(const MLogRec
& logevt
)
7985 PG
*pg
= context
< RecoveryMachine
>().pg
;
7987 if (logevt
.from
!= auth_log_shard
) {
7988 ldout(pg
->cct
, 10) << "GetLog: discarding log from "
7989 << "non-auth_log_shard osd." << logevt
.from
<< dendl
;
7990 return discard_event();
7992 ldout(pg
->cct
, 10) << "GetLog: received master log from osd"
7993 << logevt
.from
<< dendl
;
7995 post_event(GotLog());
7996 return discard_event();
7999 boost::statechart::result
PG::RecoveryState::GetLog::react(const GotLog
&)
8001 PG
*pg
= context
< RecoveryMachine
>().pg
;
8002 ldout(pg
->cct
, 10) << "leaving GetLog" << dendl
;
8004 ldout(pg
->cct
, 10) << "processing master log" << dendl
;
8005 pg
->proc_master_log(*context
<RecoveryMachine
>().get_cur_transaction(),
8006 msg
->info
, msg
->log
, msg
->missing
,
8010 context
< RecoveryMachine
>().get_cur_transaction(),
8011 context
< RecoveryMachine
>().get_on_applied_context_list(),
8012 context
< RecoveryMachine
>().get_on_safe_context_list());
8013 return transit
< GetMissing
>();
8016 boost::statechart::result
PG::RecoveryState::GetLog::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
->dump_stream("auth_log_shard") << auth_log_shard
;
8022 q
.f
->close_section();
8023 return forward_event();
8026 void PG::RecoveryState::GetLog::exit()
8028 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8029 PG
*pg
= context
< RecoveryMachine
>().pg
;
8030 utime_t dur
= ceph_clock_now() - enter_time
;
8031 pg
->osd
->recoverystate_perf
->tinc(rs_getlog_latency
, dur
);
8032 pg
->blocked_by
.clear();
8033 pg
->publish_stats_to_osd();
8036 /*------WaitActingChange--------*/
8037 PG::RecoveryState::WaitActingChange::WaitActingChange(my_context ctx
)
8039 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/WaitActingChange")
8041 context
< RecoveryMachine
>().log_enter(state_name
);
8044 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const AdvMap
& advmap
)
8046 PG
*pg
= context
< RecoveryMachine
>().pg
;
8047 OSDMapRef osdmap
= advmap
.osdmap
;
8049 ldout(pg
->cct
, 10) << "verifying no want_acting " << pg
->want_acting
<< " targets didn't go down" << dendl
;
8050 for (vector
<int>::iterator p
= pg
->want_acting
.begin(); p
!= pg
->want_acting
.end(); ++p
) {
8051 if (!osdmap
->is_up(*p
)) {
8052 ldout(pg
->cct
, 10) << " want_acting target osd." << *p
<< " went down, resetting" << dendl
;
8054 return transit
< Reset
>();
8057 return forward_event();
8060 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const MLogRec
& logevt
)
8062 PG
*pg
= context
< RecoveryMachine
>().pg
;
8063 ldout(pg
->cct
, 10) << "In WaitActingChange, ignoring MLocRec" << dendl
;
8064 return discard_event();
8067 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const MInfoRec
& evt
)
8069 PG
*pg
= context
< RecoveryMachine
>().pg
;
8070 ldout(pg
->cct
, 10) << "In WaitActingChange, ignoring MInfoRec" << dendl
;
8071 return discard_event();
8074 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const MNotifyRec
& evt
)
8076 PG
*pg
= context
< RecoveryMachine
>().pg
;
8077 ldout(pg
->cct
, 10) << "In WaitActingChange, ignoring MNotifyRec" << dendl
;
8078 return discard_event();
8081 boost::statechart::result
PG::RecoveryState::WaitActingChange::react(const QueryState
& q
)
8083 q
.f
->open_object_section("state");
8084 q
.f
->dump_string("name", state_name
);
8085 q
.f
->dump_stream("enter_time") << enter_time
;
8086 q
.f
->dump_string("comment", "waiting for pg acting set to change");
8087 q
.f
->close_section();
8088 return forward_event();
8091 void PG::RecoveryState::WaitActingChange::exit()
8093 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8094 PG
*pg
= context
< RecoveryMachine
>().pg
;
8095 utime_t dur
= ceph_clock_now() - enter_time
;
8096 pg
->osd
->recoverystate_perf
->tinc(rs_waitactingchange_latency
, dur
);
8099 /*------Down--------*/
8100 PG::RecoveryState::Down::Down(my_context ctx
)
8102 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/Down")
8104 context
< RecoveryMachine
>().log_enter(state_name
);
8105 PG
*pg
= context
< RecoveryMachine
>().pg
;
8107 pg
->state_clear(PG_STATE_PEERING
);
8108 pg
->state_set(PG_STATE_DOWN
);
8110 auto &prior_set
= context
< Peering
>().prior_set
;
8111 assert(pg
->blocked_by
.empty());
8112 pg
->blocked_by
.insert(prior_set
.down
.begin(), prior_set
.down
.end());
8113 pg
->publish_stats_to_osd();
8116 void PG::RecoveryState::Down::exit()
8118 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8119 PG
*pg
= context
< RecoveryMachine
>().pg
;
8121 pg
->state_clear(PG_STATE_DOWN
);
8122 utime_t dur
= ceph_clock_now() - enter_time
;
8123 pg
->osd
->recoverystate_perf
->tinc(rs_down_latency
, dur
);
8125 pg
->blocked_by
.clear();
8126 pg
->publish_stats_to_osd();
8129 boost::statechart::result
PG::RecoveryState::Down::react(const QueryState
& q
)
8131 q
.f
->open_object_section("state");
8132 q
.f
->dump_string("name", state_name
);
8133 q
.f
->dump_stream("enter_time") << enter_time
;
8134 q
.f
->dump_string("comment",
8135 "not enough up instances of this PG to go active");
8136 q
.f
->close_section();
8137 return forward_event();
8140 /*------Incomplete--------*/
8141 PG::RecoveryState::Incomplete::Incomplete(my_context ctx
)
8143 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/Incomplete")
8145 context
< RecoveryMachine
>().log_enter(state_name
);
8146 PG
*pg
= context
< RecoveryMachine
>().pg
;
8148 pg
->state_clear(PG_STATE_PEERING
);
8149 pg
->state_set(PG_STATE_INCOMPLETE
);
8151 PastIntervals::PriorSet
&prior_set
= context
< Peering
>().prior_set
;
8152 assert(pg
->blocked_by
.empty());
8153 pg
->blocked_by
.insert(prior_set
.down
.begin(), prior_set
.down
.end());
8154 pg
->publish_stats_to_osd();
8157 boost::statechart::result
PG::RecoveryState::Incomplete::react(const AdvMap
&advmap
) {
8158 PG
*pg
= context
< RecoveryMachine
>().pg
;
8159 int64_t poolnum
= pg
->info
.pgid
.pool();
8161 // Reset if min_size turn smaller than previous value, pg might now be able to go active
8162 if (advmap
.lastmap
->get_pools().find(poolnum
)->second
.min_size
>
8163 advmap
.osdmap
->get_pools().find(poolnum
)->second
.min_size
) {
8165 return transit
< Reset
>();
8168 return forward_event();
8171 boost::statechart::result
PG::RecoveryState::Incomplete::react(const MNotifyRec
& notevt
) {
8172 PG
*pg
= context
< RecoveryMachine
>().pg
;
8173 ldout(pg
->cct
, 7) << "handle_pg_notify from osd." << notevt
.from
<< dendl
;
8174 if (pg
->proc_replica_info(
8175 notevt
.from
, notevt
.notify
.info
, notevt
.notify
.epoch_sent
)) {
8176 // We got something new, try again!
8177 return transit
< GetLog
>();
8179 return discard_event();
8183 boost::statechart::result
PG::RecoveryState::Incomplete::react(
8184 const QueryState
& q
)
8186 q
.f
->open_object_section("state");
8187 q
.f
->dump_string("name", state_name
);
8188 q
.f
->dump_stream("enter_time") << enter_time
;
8189 q
.f
->dump_string("comment", "not enough complete instances of this PG");
8190 q
.f
->close_section();
8191 return forward_event();
8194 void PG::RecoveryState::Incomplete::exit()
8196 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8197 PG
*pg
= context
< RecoveryMachine
>().pg
;
8199 pg
->state_clear(PG_STATE_INCOMPLETE
);
8200 utime_t dur
= ceph_clock_now() - enter_time
;
8201 pg
->osd
->recoverystate_perf
->tinc(rs_incomplete_latency
, dur
);
8203 pg
->blocked_by
.clear();
8204 pg
->publish_stats_to_osd();
8207 /*------GetMissing--------*/
8208 PG::RecoveryState::GetMissing::GetMissing(my_context ctx
)
8210 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/GetMissing")
8212 context
< RecoveryMachine
>().log_enter(state_name
);
8214 PG
*pg
= context
< RecoveryMachine
>().pg
;
8215 assert(!pg
->actingbackfill
.empty());
8217 for (set
<pg_shard_t
>::iterator i
= pg
->actingbackfill
.begin();
8218 i
!= pg
->actingbackfill
.end();
8220 if (*i
== pg
->get_primary()) continue;
8221 const pg_info_t
& pi
= pg
->peer_info
[*i
];
8222 // reset this so to make sure the pg_missing_t is initialized and
8223 // has the correct semantics even if we don't need to get a
8224 // missing set from a shard. This way later additions due to
8225 // lost+unfound delete work properly.
8226 pg
->peer_missing
[*i
].may_include_deletes
= !pg
->perform_deletes_during_peering();
8229 continue; // no pg data, nothing divergent
8231 if (pi
.last_update
< pg
->pg_log
.get_tail()) {
8232 ldout(pg
->cct
, 10) << " osd." << *i
<< " is not contiguous, will restart backfill" << dendl
;
8233 pg
->peer_missing
[*i
].clear();
8236 if (pi
.last_backfill
== hobject_t()) {
8237 ldout(pg
->cct
, 10) << " osd." << *i
<< " will fully backfill; can infer empty missing set" << dendl
;
8238 pg
->peer_missing
[*i
].clear();
8242 if (pi
.last_update
== pi
.last_complete
&& // peer has no missing
8243 pi
.last_update
== pg
->info
.last_update
) { // peer is up to date
8244 // replica has no missing and identical log as us. no need to
8246 // FIXME: we can do better here. if last_update==last_complete we
8247 // can infer the rest!
8248 ldout(pg
->cct
, 10) << " osd." << *i
<< " has no missing, identical log" << dendl
;
8249 pg
->peer_missing
[*i
].clear();
8253 // We pull the log from the peer's last_epoch_started to ensure we
8254 // get enough log to detect divergent updates.
8255 since
.epoch
= pi
.last_epoch_started
;
8256 assert(pi
.last_update
>= pg
->info
.log_tail
); // or else choose_acting() did a bad thing
8257 if (pi
.log_tail
<= since
) {
8258 ldout(pg
->cct
, 10) << " requesting log+missing since " << since
<< " from osd." << *i
<< dendl
;
8259 context
< RecoveryMachine
>().send_query(
8263 i
->shard
, pg
->pg_whoami
.shard
,
8264 since
, pg
->info
.history
,
8265 pg
->get_osdmap()->get_epoch()));
8267 ldout(pg
->cct
, 10) << " requesting fulllog+missing from osd." << *i
8268 << " (want since " << since
<< " < log.tail "
8269 << pi
.log_tail
<< ")" << dendl
;
8270 context
< RecoveryMachine
>().send_query(
8272 pg_query_t::FULLLOG
,
8273 i
->shard
, pg
->pg_whoami
.shard
,
8274 pg
->info
.history
, pg
->get_osdmap()->get_epoch()));
8276 peer_missing_requested
.insert(*i
);
8277 pg
->blocked_by
.insert(i
->osd
);
8280 if (peer_missing_requested
.empty()) {
8281 if (pg
->need_up_thru
) {
8282 ldout(pg
->cct
, 10) << " still need up_thru update before going active"
8284 post_event(NeedUpThru());
8289 post_event(Activate(pg
->get_osdmap()->get_epoch()));
8291 pg
->publish_stats_to_osd();
8295 boost::statechart::result
PG::RecoveryState::GetMissing::react(const MLogRec
& logevt
)
8297 PG
*pg
= context
< RecoveryMachine
>().pg
;
8299 peer_missing_requested
.erase(logevt
.from
);
8300 pg
->proc_replica_log(logevt
.msg
->info
, logevt
.msg
->log
, logevt
.msg
->missing
, logevt
.from
);
8302 if (peer_missing_requested
.empty()) {
8303 if (pg
->need_up_thru
) {
8304 ldout(pg
->cct
, 10) << " still need up_thru update before going active"
8306 post_event(NeedUpThru());
8308 ldout(pg
->cct
, 10) << "Got last missing, don't need missing "
8309 << "posting Activate" << dendl
;
8310 post_event(Activate(pg
->get_osdmap()->get_epoch()));
8313 return discard_event();
8316 boost::statechart::result
PG::RecoveryState::GetMissing::react(const QueryState
& q
)
8318 PG
*pg
= context
< RecoveryMachine
>().pg
;
8319 q
.f
->open_object_section("state");
8320 q
.f
->dump_string("name", state_name
);
8321 q
.f
->dump_stream("enter_time") << enter_time
;
8323 q
.f
->open_array_section("peer_missing_requested");
8324 for (set
<pg_shard_t
>::iterator p
= peer_missing_requested
.begin();
8325 p
!= peer_missing_requested
.end();
8327 q
.f
->open_object_section("osd");
8328 q
.f
->dump_stream("osd") << *p
;
8329 if (pg
->peer_missing
.count(*p
)) {
8330 q
.f
->open_object_section("got_missing");
8331 pg
->peer_missing
[*p
].dump(q
.f
);
8332 q
.f
->close_section();
8334 q
.f
->close_section();
8336 q
.f
->close_section();
8338 q
.f
->close_section();
8339 return forward_event();
8342 void PG::RecoveryState::GetMissing::exit()
8344 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8345 PG
*pg
= context
< RecoveryMachine
>().pg
;
8346 utime_t dur
= ceph_clock_now() - enter_time
;
8347 pg
->osd
->recoverystate_perf
->tinc(rs_getmissing_latency
, dur
);
8348 pg
->blocked_by
.clear();
8349 pg
->publish_stats_to_osd();
8352 /*------WaitUpThru--------*/
8353 PG::RecoveryState::WaitUpThru::WaitUpThru(my_context ctx
)
8355 NamedState(context
< RecoveryMachine
>().pg
, "Started/Primary/Peering/WaitUpThru")
8357 context
< RecoveryMachine
>().log_enter(state_name
);
8360 boost::statechart::result
PG::RecoveryState::WaitUpThru::react(const ActMap
& am
)
8362 PG
*pg
= context
< RecoveryMachine
>().pg
;
8363 if (!pg
->need_up_thru
) {
8364 post_event(Activate(pg
->get_osdmap()->get_epoch()));
8366 return forward_event();
8369 boost::statechart::result
PG::RecoveryState::WaitUpThru::react(const MLogRec
& logevt
)
8371 PG
*pg
= context
< RecoveryMachine
>().pg
;
8372 ldout(pg
->cct
, 10) << "Noting missing from osd." << logevt
.from
<< dendl
;
8373 pg
->peer_missing
[logevt
.from
].claim(logevt
.msg
->missing
);
8374 pg
->peer_info
[logevt
.from
] = logevt
.msg
->info
;
8375 return discard_event();
8378 boost::statechart::result
PG::RecoveryState::WaitUpThru::react(const QueryState
& q
)
8380 q
.f
->open_object_section("state");
8381 q
.f
->dump_string("name", state_name
);
8382 q
.f
->dump_stream("enter_time") << enter_time
;
8383 q
.f
->dump_string("comment", "waiting for osdmap to reflect a new up_thru for this osd");
8384 q
.f
->close_section();
8385 return forward_event();
8388 void PG::RecoveryState::WaitUpThru::exit()
8390 context
< RecoveryMachine
>().log_exit(state_name
, enter_time
);
8391 PG
*pg
= context
< RecoveryMachine
>().pg
;
8392 utime_t dur
= ceph_clock_now() - enter_time
;
8393 pg
->osd
->recoverystate_perf
->tinc(rs_waitupthru_latency
, dur
);
8396 /*----RecoveryState::RecoveryMachine Methods-----*/
8398 #define dout_prefix *_dout << pg->gen_prefix()
8400 void PG::RecoveryState::RecoveryMachine::log_enter(const char *state_name
)
8402 PG
*pg
= context
< RecoveryMachine
>().pg
;
8403 ldout(pg
->cct
, 5) << "enter " << state_name
<< dendl
;
8404 pg
->osd
->pg_recovery_stats
.log_enter(state_name
);
8407 void PG::RecoveryState::RecoveryMachine::log_exit(const char *state_name
, utime_t enter_time
)
8409 utime_t dur
= ceph_clock_now() - enter_time
;
8410 PG
*pg
= context
< RecoveryMachine
>().pg
;
8411 ldout(pg
->cct
, 5) << "exit " << state_name
<< " " << dur
<< " " << event_count
<< " " << event_time
<< dendl
;
8412 pg
->osd
->pg_recovery_stats
.log_exit(state_name
, ceph_clock_now() - enter_time
,
8413 event_count
, event_time
);
8415 event_time
= utime_t();
8419 /*---------------------------------------------------*/
8421 #define dout_prefix (*_dout << (debug_pg ? debug_pg->gen_prefix() : string()) << " PriorSet: ")
8423 void PG::RecoveryState::start_handle(RecoveryCtx
*new_ctx
) {
8428 if (messages_pending_flush
) {
8429 rctx
= RecoveryCtx(*messages_pending_flush
, *new_ctx
);
8433 rctx
->start_time
= ceph_clock_now();
8437 void PG::RecoveryState::begin_block_outgoing() {
8438 assert(!messages_pending_flush
);
8441 messages_pending_flush
= BufferedRecoveryMessages();
8442 rctx
= RecoveryCtx(*messages_pending_flush
, *orig_ctx
);
8445 void PG::RecoveryState::clear_blocked_outgoing() {
8448 messages_pending_flush
= boost::optional
<BufferedRecoveryMessages
>();
8451 void PG::RecoveryState::end_block_outgoing() {
8452 assert(messages_pending_flush
);
8456 rctx
= RecoveryCtx(*orig_ctx
);
8457 rctx
->accept_buffered_messages(*messages_pending_flush
);
8458 messages_pending_flush
= boost::optional
<BufferedRecoveryMessages
>();
8461 void PG::RecoveryState::end_handle() {
8463 utime_t dur
= ceph_clock_now() - rctx
->start_time
;
8464 machine
.event_time
+= dur
;
8467 machine
.event_count
++;
8468 rctx
= boost::optional
<RecoveryCtx
>();
8472 ostream
& operator<<(ostream
& out
, const PG::BackfillInterval
& bi
)
8474 out
<< "BackfillInfo(" << bi
.begin
<< "-" << bi
.end
8475 << " " << bi
.objects
.size() << " objects";
8476 if (!bi
.objects
.empty())
8477 out
<< " " << bi
.objects
;
8482 void intrusive_ptr_add_ref(PG
*pg
) { pg
->get("intptr"); }
8483 void intrusive_ptr_release(PG
*pg
) { pg
->put("intptr"); }
8485 #ifdef PG_DEBUG_REFS
8486 uint64_t get_with_id(PG
*pg
) { return pg
->get_with_id(); }
8487 void put_with_id(PG
*pg
, uint64_t id
) { return pg
->put_with_id(id
); }