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>
7 * Copyright (C) 2013,2014 Cloudwatt <libre.licensing@cloudwatt.com>
9 * Author: Loic Dachary <loic@dachary.org>
11 * This is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public
13 * License version 2.1, as published by the Free Software
14 * Foundation. See file COPYING.
18 #include "boost/tuple/tuple.hpp"
19 #include "boost/intrusive_ptr.hpp"
21 #include "PrimaryLogPG.h"
23 #include "OpRequest.h"
24 #include "ScrubStore.h"
26 #include "objclass/objclass.h"
28 #include "common/errno.h"
29 #include "common/scrub_types.h"
30 #include "common/perf_counters.h"
32 #include "messages/MOSDOp.h"
33 #include "messages/MOSDBackoff.h"
34 #include "messages/MOSDSubOp.h"
35 #include "messages/MOSDSubOpReply.h"
36 #include "messages/MOSDPGTrim.h"
37 #include "messages/MOSDPGScan.h"
38 #include "messages/MOSDRepScrub.h"
39 #include "messages/MOSDPGBackfill.h"
40 #include "messages/MOSDPGBackfillRemove.h"
41 #include "messages/MOSDPGUpdateLogMissing.h"
42 #include "messages/MOSDPGUpdateLogMissingReply.h"
43 #include "messages/MCommandReply.h"
44 #include "messages/MOSDScrubReserve.h"
45 #include "mds/inode_backtrace.h" // Ugh
46 #include "common/EventTrace.h"
48 #include "common/config.h"
49 #include "include/compat.h"
50 #include "mon/MonClient.h"
51 #include "osdc/Objecter.h"
52 #include "json_spirit/json_spirit_value.h"
53 #include "json_spirit/json_spirit_reader.h"
54 #include "include/assert.h" // json_spirit clobbers it
55 #include "include/rados/rados_types.hpp"
58 #include "tracing/osd.h"
60 #define tracepoint(...)
63 #define dout_context cct
64 #define dout_subsys ceph_subsys_osd
65 #define DOUT_PREFIX_ARGS this, osd->whoami, get_osdmap()
67 #define dout_prefix _prefix(_dout, this)
69 static ostream
& _prefix(std::ostream
*_dout
, T
*pg
) {
70 return *_dout
<< pg
->gen_prefix();
79 MEMPOOL_DEFINE_OBJECT_FACTORY(PrimaryLogPG
, replicatedpg
, osd
);
81 PGLSFilter::PGLSFilter() : cct(nullptr)
85 PGLSFilter::~PGLSFilter()
89 struct PrimaryLogPG::C_OSD_OnApplied
: Context
{
97 : pg(pg
), epoch(epoch
), v(v
) {}
98 void finish(int) override
{
100 if (!pg
->pg_has_reset_since(epoch
))
107 * The CopyCallback class defines an interface for completions to the
108 * copy_start code. Users of the copy infrastructure must implement
109 * one and give an instance of the class to start_copy.
111 * The implementer is responsible for making sure that the CopyCallback
112 * can associate itself with the correct copy operation.
114 class PrimaryLogPG::CopyCallback
: public GenContext
<CopyCallbackResults
> {
118 * results.get<0>() is the return code: 0 for success; -ECANCELED if
119 * the operation was cancelled by the local OSD; -errno for other issues.
120 * results.get<1>() is a pointer to a CopyResults object, which you are
121 * responsible for deleting.
123 void finish(CopyCallbackResults results_
) override
= 0;
126 /// Provide the final size of the copied object to the CopyCallback
127 ~CopyCallback() override
{}
130 template <typename T
>
131 class PrimaryLogPG::BlessedGenContext
: public GenContext
<T
> {
133 unique_ptr
<GenContext
<T
>> c
;
136 BlessedGenContext(PrimaryLogPG
*pg
, GenContext
<T
> *c
, epoch_t e
)
137 : pg(pg
), c(c
), e(e
) {}
138 void finish(T t
) override
{
140 if (pg
->pg_has_reset_since(e
))
143 c
.release()->complete(t
);
148 GenContext
<ThreadPool::TPHandle
&> *PrimaryLogPG::bless_gencontext(
149 GenContext
<ThreadPool::TPHandle
&> *c
) {
150 return new BlessedGenContext
<ThreadPool::TPHandle
&>(
151 this, c
, get_osdmap()->get_epoch());
154 class PrimaryLogPG::BlessedContext
: public Context
{
156 unique_ptr
<Context
> c
;
159 BlessedContext(PrimaryLogPG
*pg
, Context
*c
, epoch_t e
)
160 : pg(pg
), c(c
), e(e
) {}
161 void finish(int r
) override
{
163 if (pg
->pg_has_reset_since(e
))
166 c
.release()->complete(r
);
172 Context
*PrimaryLogPG::bless_context(Context
*c
) {
173 return new BlessedContext(this, c
, get_osdmap()->get_epoch());
176 class PrimaryLogPG::C_PG_ObjectContext
: public Context
{
180 C_PG_ObjectContext(PrimaryLogPG
*p
, ObjectContext
*o
) :
182 void finish(int r
) override
{
183 pg
->object_context_destructor_callback(obc
);
187 class PrimaryLogPG::C_OSD_OndiskWriteUnlock
: public Context
{
188 ObjectContextRef obc
, obc2
, obc3
;
190 C_OSD_OndiskWriteUnlock(
192 ObjectContextRef o2
= ObjectContextRef(),
193 ObjectContextRef o3
= ObjectContextRef()) : obc(o
), obc2(o2
), obc3(o3
) {}
194 void finish(int r
) override
{
195 obc
->ondisk_write_unlock();
197 obc2
->ondisk_write_unlock();
199 obc3
->ondisk_write_unlock();
203 struct OnReadComplete
: public Context
{
205 PrimaryLogPG::OpContext
*opcontext
;
208 PrimaryLogPG::OpContext
*ctx
) : pg(pg
), opcontext(ctx
) {}
209 void finish(int r
) override
{
210 opcontext
->finish_read(pg
);
212 ~OnReadComplete() override
{}
215 class PrimaryLogPG::C_OSD_AppliedRecoveredObject
: public Context
{
217 ObjectContextRef obc
;
219 C_OSD_AppliedRecoveredObject(PrimaryLogPG
*p
, ObjectContextRef o
) :
221 void finish(int r
) override
{
222 pg
->_applied_recovered_object(obc
);
226 class PrimaryLogPG::C_OSD_CommittedPushedObject
: public Context
{
229 eversion_t last_complete
;
231 C_OSD_CommittedPushedObject(
232 PrimaryLogPG
*p
, epoch_t epoch
, eversion_t lc
) :
233 pg(p
), epoch(epoch
), last_complete(lc
) {
235 void finish(int r
) override
{
236 pg
->_committed_pushed_object(epoch
, last_complete
);
240 class PrimaryLogPG::C_OSD_AppliedRecoveredObjectReplica
: public Context
{
243 explicit C_OSD_AppliedRecoveredObjectReplica(PrimaryLogPG
*p
) :
245 void finish(int r
) override
{
246 pg
->_applied_recovered_object_replica();
251 void PrimaryLogPG::OpContext::start_async_reads(PrimaryLogPG
*pg
)
254 list
<pair
<boost::tuple
<uint64_t, uint64_t, unsigned>,
255 pair
<bufferlist
*, Context
*> > > in
;
256 in
.swap(pending_async_reads
);
257 pg
->pgbackend
->objects_read_async(
260 new OnReadComplete(pg
, this), pg
->get_pool().fast_read
);
262 void PrimaryLogPG::OpContext::finish_read(PrimaryLogPG
*pg
)
264 assert(inflightreads
> 0);
266 if (async_reads_complete()) {
267 assert(pg
->in_progress_async_reads
.size());
268 assert(pg
->in_progress_async_reads
.front().second
== this);
269 pg
->in_progress_async_reads
.pop_front();
271 // Restart the op context now that all reads have been
272 // completed. Read failures will be handled by the op finisher
273 pg
->execute_ctx(this);
277 class CopyFromCallback
: public PrimaryLogPG::CopyCallback
{
279 PrimaryLogPG::CopyResults
*results
= nullptr;
280 PrimaryLogPG::OpContext
*ctx
;
283 CopyFromCallback(PrimaryLogPG::OpContext
*ctx
, OSDOp
&osd_op
)
284 : ctx(ctx
), osd_op(osd_op
) {
286 ~CopyFromCallback() override
{}
288 void finish(PrimaryLogPG::CopyCallbackResults results_
) override
{
289 results
= results_
.get
<1>();
290 int r
= results_
.get
<0>();
292 // for finish_copyfrom
293 ctx
->user_at_version
= results
->user_version
;
296 ctx
->pg
->execute_ctx(ctx
);
298 if (r
!= -ECANCELED
) { // on cancel just toss it out; client resends
300 ctx
->pg
->osd
->reply_op_error(ctx
->op
, r
);
301 } else if (results
->should_requeue
) {
303 ctx
->pg
->requeue_op(ctx
->op
);
305 ctx
->pg
->close_op_ctx(ctx
);
309 bool is_temp_obj_used() {
310 return results
->started_temp_obj
;
312 uint64_t get_data_size() {
313 return results
->object_size
;
317 struct CopyFromFinisher
: public PrimaryLogPG::OpFinisher
{
318 CopyFromCallback
*copy_from_callback
;
320 CopyFromFinisher(CopyFromCallback
*copy_from_callback
)
321 : copy_from_callback(copy_from_callback
) {
324 int execute() override
{
325 // instance will be destructed after this method completes
326 copy_from_callback
->ctx
->pg
->finish_copyfrom(copy_from_callback
);
331 // ======================
332 // PGBackend::Listener
334 void PrimaryLogPG::on_local_recover(
335 const hobject_t
&hoid
,
336 const ObjectRecoveryInfo
&_recovery_info
,
337 ObjectContextRef obc
,
339 ObjectStore::Transaction
*t
342 dout(10) << __func__
<< ": " << hoid
<< dendl
;
344 ObjectRecoveryInfo
recovery_info(_recovery_info
);
345 clear_object_snap_mapping(t
, hoid
);
346 if (!is_delete
&& recovery_info
.soid
.is_snap()) {
347 OSDriver::OSTransaction
_t(osdriver
.get_transaction(t
));
349 dout(20) << " snapset " << recovery_info
.ss
350 << " legacy_snaps " << recovery_info
.oi
.legacy_snaps
<< dendl
;
351 if (recovery_info
.ss
.is_legacy() ||
352 recovery_info
.ss
.seq
== 0 /* jewel osd doesn't populate this */) {
353 assert(recovery_info
.oi
.legacy_snaps
.size());
354 snaps
.insert(recovery_info
.oi
.legacy_snaps
.begin(),
355 recovery_info
.oi
.legacy_snaps
.end());
357 auto p
= recovery_info
.ss
.clone_snaps
.find(hoid
.snap
);
358 assert(p
!= recovery_info
.ss
.clone_snaps
.end()); // hmm, should we warn?
359 snaps
.insert(p
->second
.begin(), p
->second
.end());
361 dout(20) << " snaps " << snaps
<< dendl
;
367 if (!is_delete
&& pg_log
.get_missing().is_missing(recovery_info
.soid
) &&
368 pg_log
.get_missing().get_items().find(recovery_info
.soid
)->second
.need
> recovery_info
.version
) {
369 assert(is_primary());
370 const pg_log_entry_t
*latest
= pg_log
.get_log().objects
.find(recovery_info
.soid
)->second
;
371 if (latest
->op
== pg_log_entry_t::LOST_REVERT
&&
372 latest
->reverting_to
== recovery_info
.version
) {
373 dout(10) << " got old revert version " << recovery_info
.version
374 << " for " << *latest
<< dendl
;
375 recovery_info
.version
= latest
->version
;
376 // update the attr to the revert event version
377 recovery_info
.oi
.prior_version
= recovery_info
.oi
.version
;
378 recovery_info
.oi
.version
= latest
->version
;
380 ::encode(recovery_info
.oi
, bl
,
381 get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
382 assert(!pool
.info
.require_rollback());
383 t
->setattr(coll
, ghobject_t(recovery_info
.soid
), OI_ATTR
, bl
);
385 obc
->attr_cache
[OI_ATTR
] = bl
;
389 // keep track of active pushes for scrub
392 if (recovery_info
.version
> pg_log
.get_can_rollback_to()) {
393 /* This can only happen during a repair, and even then, it would
394 * be one heck of a race. If we are repairing the object, the
395 * write in question must be fully committed, so it's not valid
396 * to roll it back anyway (and we'll be rolled forward shortly
398 PGLogEntryHandler h
{this, t
};
399 pg_log
.roll_forward_to(recovery_info
.version
, &h
);
401 recover_got(recovery_info
.soid
, recovery_info
.version
);
405 obc
->obs
.exists
= true;
406 obc
->ondisk_write_lock();
408 bool got
= obc
->get_recovery_read();
411 assert(recovering
.count(obc
->obs
.oi
.soid
));
412 recovering
[obc
->obs
.oi
.soid
] = obc
;
413 obc
->obs
.oi
= recovery_info
.oi
; // may have been updated above
414 t
->register_on_applied_sync(new C_OSD_OndiskWriteUnlock(obc
));
417 t
->register_on_applied(new C_OSD_AppliedRecoveredObject(this, obc
));
419 publish_stats_to_osd();
420 assert(missing_loc
.needs_recovery(hoid
));
422 missing_loc
.add_location(hoid
, pg_whoami
);
423 release_backoffs(hoid
);
424 if (!is_unreadable_object(hoid
)) {
425 auto unreadable_object_entry
= waiting_for_unreadable_object
.find(hoid
);
426 if (unreadable_object_entry
!= waiting_for_unreadable_object
.end()) {
427 dout(20) << " kicking unreadable waiters on " << hoid
<< dendl
;
428 requeue_ops(unreadable_object_entry
->second
);
429 waiting_for_unreadable_object
.erase(unreadable_object_entry
);
433 t
->register_on_applied(
434 new C_OSD_AppliedRecoveredObjectReplica(this));
438 t
->register_on_commit(
439 new C_OSD_CommittedPushedObject(
441 get_osdmap()->get_epoch(),
442 info
.last_complete
));
449 void PrimaryLogPG::on_global_recover(
450 const hobject_t
&soid
,
451 const object_stat_sum_t
&stat_diff
,
454 info
.stats
.stats
.sum
.add(stat_diff
);
455 missing_loc
.recovered(soid
);
456 publish_stats_to_osd();
457 dout(10) << "pushed " << soid
<< " to all replicas" << dendl
;
458 map
<hobject_t
, ObjectContextRef
>::iterator i
= recovering
.find(soid
);
459 assert(i
!= recovering
.end());
462 // recover missing won't have had an obc, but it gets filled in
463 // during on_local_recover
465 list
<OpRequestRef
> requeue_list
;
466 i
->second
->drop_recovery_read(&requeue_list
);
467 requeue_ops(requeue_list
);
470 backfills_in_flight
.erase(soid
);
473 finish_recovery_op(soid
);
474 release_backoffs(soid
);
475 auto degraded_object_entry
= waiting_for_degraded_object
.find(soid
);
476 if (degraded_object_entry
!= waiting_for_degraded_object
.end()) {
477 dout(20) << " kicking degraded waiters on " << soid
<< dendl
;
478 requeue_ops(degraded_object_entry
->second
);
479 waiting_for_degraded_object
.erase(degraded_object_entry
);
481 auto unreadable_object_entry
= waiting_for_unreadable_object
.find(soid
);
482 if (unreadable_object_entry
!= waiting_for_unreadable_object
.end()) {
483 dout(20) << " kicking unreadable waiters on " << soid
<< dendl
;
484 requeue_ops(unreadable_object_entry
->second
);
485 waiting_for_unreadable_object
.erase(unreadable_object_entry
);
487 finish_degraded_object(soid
);
490 void PrimaryLogPG::on_peer_recover(
492 const hobject_t
&soid
,
493 const ObjectRecoveryInfo
&recovery_info
)
495 publish_stats_to_osd();
497 peer_missing
[peer
].got(soid
, recovery_info
.version
);
500 void PrimaryLogPG::begin_peer_recover(
502 const hobject_t soid
)
504 peer_missing
[peer
].revise_have(soid
, eversion_t());
507 void PrimaryLogPG::schedule_recovery_work(
508 GenContext
<ThreadPool::TPHandle
&> *c
)
510 osd
->recovery_gen_wq
.queue(c
);
513 void PrimaryLogPG::send_message_osd_cluster(
514 int peer
, Message
*m
, epoch_t from_epoch
)
516 osd
->send_message_osd_cluster(peer
, m
, from_epoch
);
519 void PrimaryLogPG::send_message_osd_cluster(
520 Message
*m
, Connection
*con
)
522 osd
->send_message_osd_cluster(m
, con
);
525 void PrimaryLogPG::send_message_osd_cluster(
526 Message
*m
, const ConnectionRef
& con
)
528 osd
->send_message_osd_cluster(m
, con
);
531 void PrimaryLogPG::on_primary_error(
532 const hobject_t
&oid
,
535 dout(0) << __func__
<< ": oid " << oid
<< " version " << v
<< dendl
;
537 primary_error(oid
, v
);
538 backfill_add_missing(oid
, v
);
541 void PrimaryLogPG::backfill_add_missing(
542 const hobject_t
&oid
,
545 dout(0) << __func__
<< ": oid " << oid
<< " version " << v
<< dendl
;
546 backfills_in_flight
.erase(oid
);
547 missing_loc
.add_missing(oid
, v
, eversion_t());
550 ConnectionRef
PrimaryLogPG::get_con_osd_cluster(
551 int peer
, epoch_t from_epoch
)
553 return osd
->get_con_osd_cluster(peer
, from_epoch
);
556 PerfCounters
*PrimaryLogPG::get_logger()
562 // ====================
565 bool PrimaryLogPG::is_missing_object(const hobject_t
& soid
) const
567 return pg_log
.get_missing().get_items().count(soid
);
570 void PrimaryLogPG::maybe_kick_recovery(
571 const hobject_t
&soid
)
574 if (!missing_loc
.needs_recovery(soid
, &v
))
577 map
<hobject_t
, ObjectContextRef
>::const_iterator p
= recovering
.find(soid
);
578 if (p
!= recovering
.end()) {
579 dout(7) << "object " << soid
<< " v " << v
<< ", already recovering." << dendl
;
580 } else if (missing_loc
.is_unfound(soid
)) {
581 dout(7) << "object " << soid
<< " v " << v
<< ", is unfound." << dendl
;
583 dout(7) << "object " << soid
<< " v " << v
<< ", recovering." << dendl
;
584 PGBackend::RecoveryHandle
*h
= pgbackend
->open_recovery_op();
585 if (is_missing_object(soid
)) {
586 recover_missing(soid
, v
, cct
->_conf
->osd_client_op_priority
, h
);
587 } else if (missing_loc
.is_deleted(soid
)) {
588 prep_object_replica_deletes(soid
, v
, h
);
590 prep_object_replica_pushes(soid
, v
, h
);
592 pgbackend
->run_recovery_op(h
, cct
->_conf
->osd_client_op_priority
);
596 void PrimaryLogPG::wait_for_unreadable_object(
597 const hobject_t
& soid
, OpRequestRef op
)
599 assert(is_unreadable_object(soid
));
600 maybe_kick_recovery(soid
);
601 waiting_for_unreadable_object
[soid
].push_back(op
);
602 op
->mark_delayed("waiting for missing object");
605 bool PrimaryLogPG::is_degraded_or_backfilling_object(const hobject_t
& soid
)
607 /* The conditions below may clear (on_local_recover, before we queue
608 * the transaction) before we actually requeue the degraded waiters
609 * in on_global_recover after the transaction completes.
611 if (waiting_for_degraded_object
.count(soid
))
613 if (pg_log
.get_missing().get_items().count(soid
))
615 assert(!actingbackfill
.empty());
616 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
617 i
!= actingbackfill
.end();
619 if (*i
== get_primary()) continue;
620 pg_shard_t peer
= *i
;
621 auto peer_missing_entry
= peer_missing
.find(peer
);
622 if (peer_missing_entry
!= peer_missing
.end() &&
623 peer_missing_entry
->second
.get_items().count(soid
))
626 // Object is degraded if after last_backfill AND
627 // we are backfilling it
628 if (is_backfill_targets(peer
) &&
629 peer_info
[peer
].last_backfill
<= soid
&&
630 last_backfill_started
>= soid
&&
631 backfills_in_flight
.count(soid
))
637 void PrimaryLogPG::wait_for_degraded_object(const hobject_t
& soid
, OpRequestRef op
)
639 assert(is_degraded_or_backfilling_object(soid
));
641 maybe_kick_recovery(soid
);
642 waiting_for_degraded_object
[soid
].push_back(op
);
643 op
->mark_delayed("waiting for degraded object");
646 void PrimaryLogPG::block_write_on_full_cache(
647 const hobject_t
& _oid
, OpRequestRef op
)
649 const hobject_t oid
= _oid
.get_head();
650 dout(20) << __func__
<< ": blocking object " << oid
651 << " on full cache" << dendl
;
652 objects_blocked_on_cache_full
.insert(oid
);
653 waiting_for_cache_not_full
.push_back(op
);
654 op
->mark_delayed("waiting for cache not full");
657 void PrimaryLogPG::block_for_clean(
658 const hobject_t
& oid
, OpRequestRef op
)
660 dout(20) << __func__
<< ": blocking object " << oid
661 << " on primary repair" << dendl
;
662 waiting_for_clean_to_primary_repair
.push_back(op
);
663 op
->mark_delayed("waiting for clean to repair");
666 void PrimaryLogPG::block_write_on_snap_rollback(
667 const hobject_t
& oid
, ObjectContextRef obc
, OpRequestRef op
)
669 dout(20) << __func__
<< ": blocking object " << oid
.get_head()
670 << " on snap promotion " << obc
->obs
.oi
.soid
<< dendl
;
671 // otherwise, we'd have blocked in do_op
672 assert(oid
.is_head());
673 assert(objects_blocked_on_snap_promotion
.count(oid
) == 0);
674 objects_blocked_on_snap_promotion
[oid
] = obc
;
675 wait_for_blocked_object(obc
->obs
.oi
.soid
, op
);
678 void PrimaryLogPG::block_write_on_degraded_snap(
679 const hobject_t
& snap
, OpRequestRef op
)
681 dout(20) << __func__
<< ": blocking object " << snap
.get_head()
682 << " on degraded snap " << snap
<< dendl
;
683 // otherwise, we'd have blocked in do_op
684 assert(objects_blocked_on_degraded_snap
.count(snap
.get_head()) == 0);
685 objects_blocked_on_degraded_snap
[snap
.get_head()] = snap
.snap
;
686 wait_for_degraded_object(snap
, op
);
689 bool PrimaryLogPG::maybe_await_blocked_snapset(
690 const hobject_t
&hoid
,
693 ObjectContextRef obc
;
694 obc
= object_contexts
.lookup(hoid
.get_head());
696 if (obc
->is_blocked()) {
697 wait_for_blocked_object(obc
->obs
.oi
.soid
, op
);
703 obc
= object_contexts
.lookup(hoid
.get_snapdir());
705 if (obc
->is_blocked()) {
706 wait_for_blocked_object(obc
->obs
.oi
.soid
, op
);
715 void PrimaryLogPG::wait_for_blocked_object(const hobject_t
& soid
, OpRequestRef op
)
717 dout(10) << __func__
<< " " << soid
<< " " << op
<< dendl
;
718 waiting_for_blocked_object
[soid
].push_back(op
);
719 op
->mark_delayed("waiting for blocked object");
722 void PrimaryLogPG::maybe_force_recovery()
724 // no force if not in degraded/recovery/backfill states
725 if (!is_degraded() &&
726 !state_test(PG_STATE_RECOVERING
|
727 PG_STATE_RECOVERY_WAIT
|
728 PG_STATE_BACKFILLING
|
729 PG_STATE_BACKFILL_WAIT
|
730 PG_STATE_BACKFILL_TOOFULL
))
733 if (pg_log
.get_log().approx_size() <
734 cct
->_conf
->osd_max_pg_log_entries
*
735 cct
->_conf
->osd_force_recovery_pg_log_entries_factor
)
738 // find the oldest missing object
739 version_t min_version
= 0;
741 if (!pg_log
.get_missing().get_items().empty()) {
742 min_version
= pg_log
.get_missing().get_rmissing().begin()->first
;
743 soid
= pg_log
.get_missing().get_rmissing().begin()->second
;
745 assert(!actingbackfill
.empty());
746 for (set
<pg_shard_t
>::iterator it
= actingbackfill
.begin();
747 it
!= actingbackfill
.end();
749 if (*it
== get_primary()) continue;
750 pg_shard_t peer
= *it
;
751 if (peer_missing
.count(peer
) &&
752 !peer_missing
[peer
].get_items().empty() &&
753 min_version
> peer_missing
[peer
].get_rmissing().begin()->first
) {
754 min_version
= peer_missing
[peer
].get_rmissing().begin()->first
;
755 soid
= peer_missing
[peer
].get_rmissing().begin()->second
;
760 if (soid
!= hobject_t())
761 maybe_kick_recovery(soid
);
764 class PGLSPlainFilter
: public PGLSFilter
{
767 int init(bufferlist::iterator
¶ms
) override
770 ::decode(xattr
, params
);
771 ::decode(val
, params
);
772 } catch (buffer::error
&e
) {
778 ~PGLSPlainFilter() override
{}
779 bool filter(const hobject_t
&obj
, bufferlist
& xattr_data
,
780 bufferlist
& outdata
) override
;
783 class PGLSParentFilter
: public PGLSFilter
{
784 inodeno_t parent_ino
;
787 PGLSParentFilter(CephContext
* cct
) : cct(cct
) {
790 int init(bufferlist::iterator
¶ms
) override
793 ::decode(parent_ino
, params
);
794 } catch (buffer::error
&e
) {
797 generic_dout(0) << "parent_ino=" << parent_ino
<< dendl
;
801 ~PGLSParentFilter() override
{}
802 bool filter(const hobject_t
&obj
, bufferlist
& xattr_data
,
803 bufferlist
& outdata
) override
;
806 bool PGLSParentFilter::filter(const hobject_t
&obj
,
807 bufferlist
& xattr_data
, bufferlist
& outdata
)
809 bufferlist::iterator iter
= xattr_data
.begin();
810 inode_backtrace_t bt
;
812 generic_dout(0) << "PGLSParentFilter::filter" << dendl
;
816 vector
<inode_backpointer_t
>::iterator vi
;
817 for (vi
= bt
.ancestors
.begin(); vi
!= bt
.ancestors
.end(); ++vi
) {
818 generic_dout(0) << "vi->dirino=" << vi
->dirino
<< " parent_ino=" << parent_ino
<< dendl
;
819 if (vi
->dirino
== parent_ino
) {
820 ::encode(*vi
, outdata
);
828 bool PGLSPlainFilter::filter(const hobject_t
&obj
,
829 bufferlist
& xattr_data
, bufferlist
& outdata
)
831 if (val
.size() != xattr_data
.length())
834 if (memcmp(val
.c_str(), xattr_data
.c_str(), val
.size()))
840 bool PrimaryLogPG::pgls_filter(PGLSFilter
*filter
, hobject_t
& sobj
, bufferlist
& outdata
)
844 // If filter has expressed an interest in an xattr, load it.
845 if (!filter
->get_xattr().empty()) {
846 int ret
= pgbackend
->objects_get_attr(
850 dout(0) << "getattr (sobj=" << sobj
<< ", attr=" << filter
->get_xattr() << ") returned " << ret
<< dendl
;
852 if (ret
!= -ENODATA
|| filter
->reject_empty_xattr()) {
858 return filter
->filter(sobj
, bl
, outdata
);
861 int PrimaryLogPG::get_pgls_filter(bufferlist::iterator
& iter
, PGLSFilter
**pfilter
)
867 ::decode(type
, iter
);
869 catch (buffer::error
& e
) {
873 if (type
.compare("parent") == 0) {
874 filter
= new PGLSParentFilter(cct
);
875 } else if (type
.compare("plain") == 0) {
876 filter
= new PGLSPlainFilter();
878 std::size_t dot
= type
.find(".");
879 if (dot
== std::string::npos
|| dot
== 0 || dot
== type
.size() - 1) {
883 const std::string class_name
= type
.substr(0, dot
);
884 const std::string filter_name
= type
.substr(dot
+ 1);
885 ClassHandler::ClassData
*cls
= NULL
;
886 int r
= osd
->class_handler
->open_class(class_name
, &cls
);
888 derr
<< "Error opening class '" << class_name
<< "': "
889 << cpp_strerror(r
) << dendl
;
890 if (r
!= -EPERM
) // propogate permission error
897 ClassHandler::ClassFilter
*class_filter
= cls
->get_filter(filter_name
);
898 if (class_filter
== NULL
) {
899 derr
<< "Error finding filter '" << filter_name
<< "' in class "
900 << class_name
<< dendl
;
903 filter
= class_filter
->fn();
905 // Object classes are obliged to return us something, but let's
906 // give an error rather than asserting out.
907 derr
<< "Buggy class " << class_name
<< " failed to construct "
908 "filter " << filter_name
<< dendl
;
914 int r
= filter
->init(iter
);
916 derr
<< "Error initializing filter " << type
<< ": "
917 << cpp_strerror(r
) << dendl
;
921 // Successfully constructed and initialized, return it.
928 // ==========================================================
930 int PrimaryLogPG::do_command(
938 const auto &missing
= pg_log
.get_missing();
942 cmd_getval(cct
, cmdmap
, "format", format
);
943 boost::scoped_ptr
<Formatter
> f(Formatter::create(format
, "json-pretty", "json"));
946 cmd_getval(cct
, cmdmap
, "cmd", command
);
947 if (command
== "query") {
948 f
->open_object_section("pg");
949 f
->dump_string("state", pg_state_string(get_state()));
950 f
->dump_stream("snap_trimq") << snap_trimq
;
951 f
->dump_unsigned("snap_trimq_len", snap_trimq
.size());
952 f
->dump_unsigned("epoch", get_osdmap()->get_epoch());
953 f
->open_array_section("up");
954 for (vector
<int>::iterator p
= up
.begin(); p
!= up
.end(); ++p
)
955 f
->dump_unsigned("osd", *p
);
957 f
->open_array_section("acting");
958 for (vector
<int>::iterator p
= acting
.begin(); p
!= acting
.end(); ++p
)
959 f
->dump_unsigned("osd", *p
);
961 if (!backfill_targets
.empty()) {
962 f
->open_array_section("backfill_targets");
963 for (set
<pg_shard_t
>::iterator p
= backfill_targets
.begin();
964 p
!= backfill_targets
.end();
966 f
->dump_stream("shard") << *p
;
969 if (!actingbackfill
.empty()) {
970 f
->open_array_section("actingbackfill");
971 for (set
<pg_shard_t
>::iterator p
= actingbackfill
.begin();
972 p
!= actingbackfill
.end();
974 f
->dump_stream("shard") << *p
;
977 f
->open_object_section("info");
978 _update_calc_stats();
982 f
->open_array_section("peer_info");
983 for (map
<pg_shard_t
, pg_info_t
>::iterator p
= peer_info
.begin();
984 p
!= peer_info
.end();
986 f
->open_object_section("info");
987 f
->dump_stream("peer") << p
->first
;
988 p
->second
.dump(f
.get());
993 f
->open_array_section("recovery_state");
994 handle_query_state(f
.get());
997 f
->open_object_section("agent_state");
999 agent_state
->dump(f
.get());
1006 else if (command
== "mark_unfound_lost") {
1008 cmd_getval(cct
, cmdmap
, "mulcmd", mulcmd
);
1010 if (mulcmd
== "revert") {
1011 if (pool
.info
.ec_pool()) {
1012 ss
<< "mode must be 'delete' for ec pool";
1015 mode
= pg_log_entry_t::LOST_REVERT
;
1016 } else if (mulcmd
== "delete") {
1017 mode
= pg_log_entry_t::LOST_DELETE
;
1019 ss
<< "mode must be 'revert' or 'delete'; mark not yet implemented";
1022 assert(mode
== pg_log_entry_t::LOST_REVERT
||
1023 mode
== pg_log_entry_t::LOST_DELETE
);
1025 if (!is_primary()) {
1026 ss
<< "not primary";
1030 uint64_t unfound
= missing_loc
.num_unfound();
1032 ss
<< "pg has no unfound objects";
1033 return 0; // make command idempotent
1036 if (!all_unfound_are_queried_or_lost(get_osdmap())) {
1037 ss
<< "pg has " << unfound
1038 << " unfound objects but we haven't probed all sources, not marking lost";
1042 mark_all_unfound_lost(mode
, con
, tid
);
1045 else if (command
== "list_missing") {
1048 if (cmd_getval(cct
, cmdmap
, "offset", offset_json
)) {
1049 json_spirit::Value v
;
1051 if (!json_spirit::read(offset_json
, v
))
1052 throw std::runtime_error("bad json");
1054 } catch (std::runtime_error
& e
) {
1055 ss
<< "error parsing offset: " << e
.what();
1059 f
->open_object_section("missing");
1061 f
->open_object_section("offset");
1062 offset
.dump(f
.get());
1065 f
->dump_int("num_missing", missing
.num_missing());
1066 f
->dump_int("num_unfound", get_num_unfound());
1067 const map
<hobject_t
, pg_missing_item
> &needs_recovery_map
=
1068 missing_loc
.get_needs_recovery();
1069 map
<hobject_t
, pg_missing_item
>::const_iterator p
=
1070 needs_recovery_map
.upper_bound(offset
);
1072 f
->open_array_section("objects");
1074 for (; p
!= needs_recovery_map
.end() && num
< cct
->_conf
->osd_command_max_records
; ++p
) {
1075 if (missing_loc
.is_unfound(p
->first
)) {
1076 f
->open_object_section("object");
1078 f
->open_object_section("oid");
1079 p
->first
.dump(f
.get());
1082 p
->second
.dump(f
.get()); // have, need keys
1084 f
->open_array_section("locations");
1085 for (set
<pg_shard_t
>::iterator r
=
1086 missing_loc
.get_locations(p
->first
).begin();
1087 r
!= missing_loc
.get_locations(p
->first
).end();
1089 f
->dump_stream("shard") << *r
;
1098 f
->dump_bool("more", p
!= needs_recovery_map
.end());
1104 ss
<< "unknown pg command " << prefix
;
1108 // ==========================================================
1110 void PrimaryLogPG::do_pg_op(OpRequestRef op
)
1112 // NOTE: this is non-const because we modify the OSDOp.outdata in
1114 MOSDOp
*m
= static_cast<MOSDOp
*>(op
->get_nonconst_req());
1115 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1116 dout(10) << "do_pg_op " << *m
<< dendl
;
1121 string cname
, mname
;
1122 PGLSFilter
*filter
= NULL
;
1123 bufferlist filter_out
;
1125 snapid_t snapid
= m
->get_snapid();
1127 vector
<OSDOp
> ops
= m
->ops
;
1129 for (vector
<OSDOp
>::iterator p
= ops
.begin(); p
!= ops
.end(); ++p
) {
1131 bufferlist::iterator bp
= p
->indata
.begin();
1133 case CEPH_OSD_OP_PGNLS_FILTER
:
1135 ::decode(cname
, bp
);
1136 ::decode(mname
, bp
);
1138 catch (const buffer::error
& e
) {
1139 dout(0) << "unable to decode PGLS_FILTER description in " << *m
<< dendl
;
1147 result
= get_pgls_filter(bp
, &filter
);
1155 case CEPH_OSD_OP_PGNLS
:
1156 if (snapid
!= CEPH_NOSNAP
) {
1160 if (get_osdmap()->raw_pg_to_pg(m
->get_pg()) != info
.pgid
.pgid
) {
1161 dout(10) << " pgnls pg=" << m
->get_pg()
1162 << " " << get_osdmap()->raw_pg_to_pg(m
->get_pg())
1163 << " != " << info
.pgid
<< dendl
;
1166 unsigned list_size
= MIN(cct
->_conf
->osd_max_pgls
, p
->op
.pgls
.count
);
1168 dout(10) << " pgnls pg=" << m
->get_pg() << " count " << list_size
<< dendl
;
1169 // read into a buffer
1170 vector
<hobject_t
> sentries
;
1171 pg_nls_response_t response
;
1173 ::decode(response
.handle
, bp
);
1175 catch (const buffer::error
& e
) {
1176 dout(0) << "unable to decode PGNLS handle in " << *m
<< dendl
;
1182 hobject_t lower_bound
= response
.handle
;
1183 hobject_t pg_start
= info
.pgid
.pgid
.get_hobj_start();
1184 hobject_t pg_end
= info
.pgid
.pgid
.get_hobj_end(pool
.info
.get_pg_num());
1185 dout(10) << " pgnls lower_bound " << lower_bound
1186 << " pg_end " << pg_end
<< dendl
;
1187 if (((!lower_bound
.is_max() && lower_bound
>= pg_end
) ||
1188 (lower_bound
!= hobject_t() && lower_bound
< pg_start
))) {
1189 // this should only happen with a buggy client.
1190 dout(10) << "outside of PG bounds " << pg_start
<< " .. "
1196 hobject_t current
= lower_bound
;
1198 int r
= pgbackend
->objects_list_partial(
1209 map
<hobject_t
, pg_missing_item
>::const_iterator missing_iter
=
1210 pg_log
.get_missing().get_items().lower_bound(current
);
1211 vector
<hobject_t
>::iterator ls_iter
= sentries
.begin();
1212 hobject_t _max
= hobject_t::get_max();
1214 const hobject_t
&mcand
=
1215 missing_iter
== pg_log
.get_missing().get_items().end() ?
1217 missing_iter
->first
;
1218 const hobject_t
&lcand
=
1219 ls_iter
== sentries
.end() ?
1223 hobject_t candidate
;
1224 if (mcand
== lcand
) {
1226 if (!mcand
.is_max()) {
1230 } else if (mcand
< lcand
) {
1232 assert(!mcand
.is_max());
1236 assert(!lcand
.is_max());
1240 dout(10) << " pgnls candidate 0x" << std::hex
<< candidate
.get_hash()
1241 << " vs lower bound 0x" << lower_bound
.get_hash() << dendl
;
1243 if (candidate
>= next
) {
1247 if (response
.entries
.size() == list_size
) {
1252 // skip snapdir objects
1253 if (candidate
.snap
== CEPH_SNAPDIR
)
1256 if (candidate
.snap
!= CEPH_NOSNAP
)
1259 // skip internal namespace
1260 if (candidate
.get_namespace() == cct
->_conf
->osd_hit_set_namespace
)
1263 if (missing_loc
.is_deleted(candidate
))
1266 // skip wrong namespace
1267 if (m
->get_hobj().nspace
!= librados::all_nspaces
&&
1268 candidate
.get_namespace() != m
->get_hobj().nspace
)
1271 if (filter
&& !pgls_filter(filter
, candidate
, filter_out
))
1274 dout(20) << "pgnls item 0x" << std::hex
1275 << candidate
.get_hash()
1276 << ", rev 0x" << hobject_t::_reverse_bits(candidate
.get_hash())
1278 << candidate
.oid
.name
<< dendl
;
1280 librados::ListObjectImpl item
;
1281 item
.nspace
= candidate
.get_namespace();
1282 item
.oid
= candidate
.oid
.name
;
1283 item
.locator
= candidate
.get_key();
1284 response
.entries
.push_back(item
);
1287 if (next
.is_max() &&
1288 missing_iter
== pg_log
.get_missing().get_items().end() &&
1289 ls_iter
== sentries
.end()) {
1292 // Set response.handle to the start of the next PG according
1293 // to the object sort order.
1294 response
.handle
= info
.pgid
.pgid
.get_hobj_end(pool
.info
.get_pg_num());
1296 response
.handle
= next
;
1298 dout(10) << "pgnls handle=" << response
.handle
<< dendl
;
1299 ::encode(response
, osd_op
.outdata
);
1301 ::encode(filter_out
, osd_op
.outdata
);
1302 dout(10) << " pgnls result=" << result
<< " outdata.length()="
1303 << osd_op
.outdata
.length() << dendl
;
1307 case CEPH_OSD_OP_PGLS_FILTER
:
1309 ::decode(cname
, bp
);
1310 ::decode(mname
, bp
);
1312 catch (const buffer::error
& e
) {
1313 dout(0) << "unable to decode PGLS_FILTER description in " << *m
<< dendl
;
1321 result
= get_pgls_filter(bp
, &filter
);
1329 case CEPH_OSD_OP_PGLS
:
1330 if (snapid
!= CEPH_NOSNAP
) {
1334 if (get_osdmap()->raw_pg_to_pg(m
->get_pg()) != info
.pgid
.pgid
) {
1335 dout(10) << " pgls pg=" << m
->get_pg()
1336 << " " << get_osdmap()->raw_pg_to_pg(m
->get_pg())
1337 << " != " << info
.pgid
<< dendl
;
1340 unsigned list_size
= MIN(cct
->_conf
->osd_max_pgls
, p
->op
.pgls
.count
);
1342 dout(10) << " pgls pg=" << m
->get_pg() << " count " << list_size
<< dendl
;
1343 // read into a buffer
1344 vector
<hobject_t
> sentries
;
1345 pg_ls_response_t response
;
1347 ::decode(response
.handle
, bp
);
1349 catch (const buffer::error
& e
) {
1350 dout(0) << "unable to decode PGLS handle in " << *m
<< dendl
;
1356 hobject_t current
= response
.handle
;
1358 int r
= pgbackend
->objects_list_partial(
1369 assert(snapid
== CEPH_NOSNAP
|| pg_log
.get_missing().get_items().empty());
1371 map
<hobject_t
, pg_missing_item
>::const_iterator missing_iter
=
1372 pg_log
.get_missing().get_items().lower_bound(current
);
1373 vector
<hobject_t
>::iterator ls_iter
= sentries
.begin();
1374 hobject_t _max
= hobject_t::get_max();
1376 const hobject_t
&mcand
=
1377 missing_iter
== pg_log
.get_missing().get_items().end() ?
1379 missing_iter
->first
;
1380 const hobject_t
&lcand
=
1381 ls_iter
== sentries
.end() ?
1385 hobject_t candidate
;
1386 if (mcand
== lcand
) {
1388 if (!mcand
.is_max()) {
1392 } else if (mcand
< lcand
) {
1394 assert(!mcand
.is_max());
1398 assert(!lcand
.is_max());
1402 if (candidate
>= next
) {
1406 if (response
.entries
.size() == list_size
) {
1411 // skip snapdir objects
1412 if (candidate
.snap
== CEPH_SNAPDIR
)
1415 if (candidate
.snap
!= CEPH_NOSNAP
)
1418 // skip wrong namespace
1419 if (candidate
.get_namespace() != m
->get_hobj().nspace
)
1422 if (missing_loc
.is_deleted(candidate
))
1425 if (filter
&& !pgls_filter(filter
, candidate
, filter_out
))
1428 response
.entries
.push_back(make_pair(candidate
.oid
,
1429 candidate
.get_key()));
1431 if (next
.is_max() &&
1432 missing_iter
== pg_log
.get_missing().get_items().end() &&
1433 ls_iter
== sentries
.end()) {
1436 response
.handle
= next
;
1437 ::encode(response
, osd_op
.outdata
);
1439 ::encode(filter_out
, osd_op
.outdata
);
1440 dout(10) << " pgls result=" << result
<< " outdata.length()="
1441 << osd_op
.outdata
.length() << dendl
;
1445 case CEPH_OSD_OP_PG_HITSET_LS
:
1447 list
< pair
<utime_t
,utime_t
> > ls
;
1448 for (list
<pg_hit_set_info_t
>::const_iterator p
= info
.hit_set
.history
.begin();
1449 p
!= info
.hit_set
.history
.end();
1451 ls
.push_back(make_pair(p
->begin
, p
->end
));
1453 ls
.push_back(make_pair(hit_set_start_stamp
, utime_t()));
1454 ::encode(ls
, osd_op
.outdata
);
1458 case CEPH_OSD_OP_PG_HITSET_GET
:
1460 utime_t
stamp(osd_op
.op
.hit_set_get
.stamp
);
1461 if (hit_set_start_stamp
&& stamp
>= hit_set_start_stamp
) {
1462 // read the current in-memory HitSet, not the version we've
1468 ::encode(*hit_set
, osd_op
.outdata
);
1469 result
= osd_op
.outdata
.length();
1471 // read an archived HitSet.
1473 for (list
<pg_hit_set_info_t
>::const_iterator p
= info
.hit_set
.history
.begin();
1474 p
!= info
.hit_set
.history
.end();
1476 if (stamp
>= p
->begin
&& stamp
<= p
->end
) {
1477 oid
= get_hit_set_archive_object(p
->begin
, p
->end
, p
->using_gmt
);
1481 if (oid
== hobject_t()) {
1485 if (!pool
.info
.is_replicated()) {
1486 // FIXME: EC not supported yet
1487 result
= -EOPNOTSUPP
;
1490 if (is_unreadable_object(oid
)) {
1491 wait_for_unreadable_object(oid
, op
);
1495 result
= osd
->store
->read(ch
, ghobject_t(oid
), 0, 0, osd_op
.outdata
);
1500 case CEPH_OSD_OP_SCRUBLS
:
1501 result
= do_scrub_ls(m
, &osd_op
);
1514 MOSDOpReply
*reply
= new MOSDOpReply(m
, 0, get_osdmap()->get_epoch(),
1515 CEPH_OSD_FLAG_ACK
| CEPH_OSD_FLAG_ONDISK
,
1517 reply
->claim_op_out_data(ops
);
1518 reply
->set_result(result
);
1519 reply
->set_reply_versions(info
.last_update
, info
.last_user_version
);
1520 osd
->send_message_osd_client(reply
, m
->get_connection());
1524 int PrimaryLogPG::do_scrub_ls(MOSDOp
*m
, OSDOp
*osd_op
)
1526 if (m
->get_pg() != info
.pgid
.pgid
) {
1527 dout(10) << " scrubls pg=" << m
->get_pg() << " != " << info
.pgid
<< dendl
;
1528 return -EINVAL
; // hmm?
1530 auto bp
= osd_op
->indata
.begin();
1534 } catch (buffer::error
&) {
1535 dout(10) << " corrupted scrub_ls_arg_t" << dendl
;
1539 scrub_ls_result_t result
= {.interval
= info
.history
.same_interval_since
};
1540 if (arg
.interval
!= 0 && arg
.interval
!= info
.history
.same_interval_since
) {
1542 } else if (!scrubber
.store
) {
1544 } else if (arg
.get_snapsets
) {
1545 result
.vals
= scrubber
.store
->get_snap_errors(osd
->store
,
1550 result
.vals
= scrubber
.store
->get_object_errors(osd
->store
,
1555 ::encode(result
, osd_op
->outdata
);
1559 void PrimaryLogPG::calc_trim_to()
1561 size_t target
= cct
->_conf
->osd_min_pg_log_entries
;
1562 if (is_degraded() ||
1563 state_test(PG_STATE_RECOVERING
|
1564 PG_STATE_RECOVERY_WAIT
|
1565 PG_STATE_BACKFILLING
|
1566 PG_STATE_BACKFILL_WAIT
|
1567 PG_STATE_BACKFILL_TOOFULL
)) {
1568 target
= cct
->_conf
->osd_max_pg_log_entries
;
1571 eversion_t limit
= MIN(
1572 min_last_complete_ondisk
,
1573 pg_log
.get_can_rollback_to());
1574 if (limit
!= eversion_t() &&
1575 limit
!= pg_trim_to
&&
1576 pg_log
.get_log().approx_size() > target
) {
1577 size_t num_to_trim
= MIN(pg_log
.get_log().approx_size() - target
,
1578 cct
->_conf
->osd_pg_log_trim_max
);
1579 if (num_to_trim
< cct
->_conf
->osd_pg_log_trim_min
&&
1580 cct
->_conf
->osd_pg_log_trim_max
>= cct
->_conf
->osd_pg_log_trim_min
) {
1583 list
<pg_log_entry_t
>::const_iterator it
= pg_log
.get_log().log
.begin();
1584 eversion_t new_trim_to
;
1585 for (size_t i
= 0; i
< num_to_trim
; ++i
) {
1586 new_trim_to
= it
->version
;
1588 if (new_trim_to
> limit
) {
1589 new_trim_to
= limit
;
1590 dout(10) << "calc_trim_to trimming to min_last_complete_ondisk" << dendl
;
1594 dout(10) << "calc_trim_to " << pg_trim_to
<< " -> " << new_trim_to
<< dendl
;
1595 pg_trim_to
= new_trim_to
;
1596 assert(pg_trim_to
<= pg_log
.get_head());
1597 assert(pg_trim_to
<= min_last_complete_ondisk
);
1601 PrimaryLogPG::PrimaryLogPG(OSDService
*o
, OSDMapRef curmap
,
1602 const PGPool
&_pool
, spg_t p
) :
1603 PG(o
, curmap
, _pool
, p
),
1605 PGBackend::build_pg_backend(
1606 _pool
.info
, curmap
, this, coll_t(p
), ch
, o
->store
, cct
)),
1607 object_contexts(o
->cct
, o
->cct
->_conf
->osd_pg_object_context_cache_count
),
1608 snapset_contexts_lock("PrimaryLogPG::snapset_contexts_lock"),
1609 new_backfill(false),
1611 snap_trimmer_machine(this)
1613 missing_loc
.set_backend_predicates(
1614 pgbackend
->get_is_readable_predicate(),
1615 pgbackend
->get_is_recoverable_predicate());
1616 snap_trimmer_machine
.initiate();
1619 void PrimaryLogPG::get_src_oloc(const object_t
& oid
, const object_locator_t
& oloc
, object_locator_t
& src_oloc
)
1622 if (oloc
.key
.empty())
1623 src_oloc
.key
= oid
.name
;
1626 void PrimaryLogPG::handle_backoff(OpRequestRef
& op
)
1628 const MOSDBackoff
*m
= static_cast<const MOSDBackoff
*>(op
->get_req());
1629 SessionRef session
= static_cast<Session
*>(m
->get_connection()->get_priv());
1632 session
->put(); // get_priv takes a ref, and so does the SessionRef
1633 hobject_t begin
= info
.pgid
.pgid
.get_hobj_start();
1634 hobject_t end
= info
.pgid
.pgid
.get_hobj_end(pool
.info
.get_pg_num());
1635 if (begin
< m
->begin
) {
1641 dout(10) << __func__
<< " backoff ack id " << m
->id
1642 << " [" << begin
<< "," << end
<< ")" << dendl
;
1643 session
->ack_backoff(cct
, m
->pgid
, m
->id
, begin
, end
);
1646 void PrimaryLogPG::do_request(
1648 ThreadPool::TPHandle
&handle
)
1650 if (op
->osd_trace
) {
1651 op
->pg_trace
.init("pg op", &trace_endpoint
, &op
->osd_trace
);
1652 op
->pg_trace
.event("do request");
1654 // make sure we have a new enough map
1655 auto p
= waiting_for_map
.find(op
->get_source());
1656 if (p
!= waiting_for_map
.end()) {
1657 // preserve ordering
1658 dout(20) << __func__
<< " waiting_for_map "
1659 << p
->first
<< " not empty, queueing" << dendl
;
1660 p
->second
.push_back(op
);
1661 op
->mark_delayed("waiting_for_map not empty");
1664 if (!have_same_or_newer_map(op
->min_epoch
)) {
1665 dout(20) << __func__
<< " min " << op
->min_epoch
1666 << ", queue on waiting_for_map " << op
->get_source() << dendl
;
1667 waiting_for_map
[op
->get_source()].push_back(op
);
1668 op
->mark_delayed("op must wait for map");
1669 osd
->request_osdmap_update(op
->min_epoch
);
1673 if (can_discard_request(op
)) {
1678 const Message
*m
= op
->get_req();
1679 if (m
->get_connection()->has_feature(CEPH_FEATURE_RADOS_BACKOFF
)) {
1680 SessionRef session
= static_cast<Session
*>(m
->get_connection()->get_priv());
1683 session
->put(); // get_priv takes a ref, and so does the SessionRef
1685 if (op
->get_req()->get_type() == CEPH_MSG_OSD_OP
) {
1686 if (session
->check_backoff(cct
, info
.pgid
,
1687 info
.pgid
.pgid
.get_hobj_start(), m
)) {
1694 (!is_active() && is_peered());
1695 if (g_conf
->osd_backoff_on_peering
&& !backoff
) {
1701 add_pg_backoff(session
);
1705 // pg backoff acks at pg-level
1706 if (op
->get_req()->get_type() == CEPH_MSG_OSD_BACKOFF
) {
1707 const MOSDBackoff
*ba
= static_cast<const MOSDBackoff
*>(m
);
1708 if (ba
->begin
!= ba
->end
) {
1716 // Delay unless PGBackend says it's ok
1717 if (pgbackend
->can_handle_while_inactive(op
)) {
1718 bool handled
= pgbackend
->handle_message(op
);
1722 waiting_for_peered
.push_back(op
);
1723 op
->mark_delayed("waiting for peered");
1728 if (flushes_in_progress
> 0) {
1729 dout(20) << flushes_in_progress
1730 << " flushes_in_progress pending "
1731 << "waiting for flush on " << op
<< dendl
;
1732 waiting_for_flush
.push_back(op
);
1733 op
->mark_delayed("waiting for flush");
1737 assert(is_peered() && flushes_in_progress
== 0);
1738 if (pgbackend
->handle_message(op
))
1741 switch (op
->get_req()->get_type()) {
1742 case CEPH_MSG_OSD_OP
:
1743 case CEPH_MSG_OSD_BACKOFF
:
1745 dout(20) << " peered, not active, waiting for active on " << op
<< dendl
;
1746 waiting_for_active
.push_back(op
);
1747 op
->mark_delayed("waiting for active");
1750 switch (op
->get_req()->get_type()) {
1751 case CEPH_MSG_OSD_OP
:
1752 // verify client features
1753 if ((pool
.info
.has_tiers() || pool
.info
.is_tier()) &&
1754 !op
->has_feature(CEPH_FEATURE_OSD_CACHEPOOL
)) {
1755 osd
->reply_op_error(op
, -EOPNOTSUPP
);
1760 case CEPH_MSG_OSD_BACKOFF
:
1761 // object-level backoff acks handled in osdop context
1771 case MSG_OSD_SUBOPREPLY
:
1772 do_sub_op_reply(op
);
1775 case MSG_OSD_PG_SCAN
:
1776 do_scan(op
, handle
);
1779 case MSG_OSD_PG_BACKFILL
:
1783 case MSG_OSD_PG_BACKFILL_REMOVE
:
1784 do_backfill_remove(op
);
1787 case MSG_OSD_SCRUB_RESERVE
:
1789 const MOSDScrubReserve
*m
=
1790 static_cast<const MOSDScrubReserve
*>(op
->get_req());
1792 case MOSDScrubReserve::REQUEST
:
1793 handle_scrub_reserve_request(op
);
1795 case MOSDScrubReserve::GRANT
:
1796 handle_scrub_reserve_grant(op
, m
->from
);
1798 case MOSDScrubReserve::REJECT
:
1799 handle_scrub_reserve_reject(op
, m
->from
);
1801 case MOSDScrubReserve::RELEASE
:
1802 handle_scrub_reserve_release(op
);
1808 case MSG_OSD_REP_SCRUB
:
1809 replica_scrub(op
, handle
);
1812 case MSG_OSD_REP_SCRUBMAP
:
1813 do_replica_scrub_map(op
);
1816 case MSG_OSD_PG_UPDATE_LOG_MISSING
:
1817 do_update_log_missing(op
);
1820 case MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY
:
1821 do_update_log_missing_reply(op
);
1825 assert(0 == "bad message type in do_request");
1829 hobject_t
PrimaryLogPG::earliest_backfill() const
1831 hobject_t e
= hobject_t::get_max();
1832 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
1833 i
!= backfill_targets
.end();
1836 map
<pg_shard_t
, pg_info_t
>::const_iterator iter
= peer_info
.find(bt
);
1837 assert(iter
!= peer_info
.end());
1838 if (iter
->second
.last_backfill
< e
)
1839 e
= iter
->second
.last_backfill
;
1844 /** do_op - do an op
1845 * pg lock will be held (if multithreaded)
1846 * osd_lock NOT held.
1848 void PrimaryLogPG::do_op(OpRequestRef
& op
)
1851 // NOTE: take a non-const pointer here; we must be careful not to
1852 // change anything that will break other reads on m (operator<<).
1853 MOSDOp
*m
= static_cast<MOSDOp
*>(op
->get_nonconst_req());
1854 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1855 if (m
->finish_decode()) {
1856 op
->reset_desc(); // for TrackedOp
1860 dout(20) << __func__
<< ": op " << *m
<< dendl
;
1862 hobject_t head
= m
->get_hobj();
1863 head
.snap
= CEPH_NOSNAP
;
1865 if (!info
.pgid
.pgid
.contains(
1866 info
.pgid
.pgid
.get_split_bits(pool
.info
.get_pg_num()), head
)) {
1867 derr
<< __func__
<< " " << info
.pgid
.pgid
<< " does not contain "
1868 << head
<< " pg_num " << pool
.info
.get_pg_num() << " hash "
1869 << std::hex
<< head
.get_hash() << std::dec
<< dendl
;
1870 osd
->clog
->warn() << info
.pgid
.pgid
<< " does not contain " << head
1872 assert(!cct
->_conf
->osd_debug_misdirected_ops
);
1877 m
->get_connection()->has_feature(CEPH_FEATURE_RADOS_BACKOFF
);
1880 session
= static_cast<Session
*>(m
->get_connection()->get_priv());
1881 if (!session
.get()) {
1882 dout(10) << __func__
<< " no session" << dendl
;
1885 session
->put(); // get_priv() takes a ref, and so does the intrusive_ptr
1887 if (session
->check_backoff(cct
, info
.pgid
, head
, m
)) {
1892 if (m
->has_flag(CEPH_OSD_FLAG_PARALLELEXEC
)) {
1894 dout(20) << __func__
<< ": PARALLELEXEC not implemented " << *m
<< dendl
;
1895 osd
->reply_op_error(op
, -EINVAL
);
1899 if (op
->rmw_flags
== 0) {
1900 int r
= osd
->osd
->init_op_flags(op
);
1902 osd
->reply_op_error(op
, r
);
1907 if ((m
->get_flags() & (CEPH_OSD_FLAG_BALANCE_READS
|
1908 CEPH_OSD_FLAG_LOCALIZE_READS
)) &&
1910 !(op
->may_write() || op
->may_cache())) {
1911 // balanced reads; any replica will do
1912 if (!(is_primary() || is_replica())) {
1913 osd
->handle_misdirected_op(this, op
);
1917 // normal case; must be primary
1918 if (!is_primary()) {
1919 osd
->handle_misdirected_op(this, op
);
1924 if (!op_has_sufficient_caps(op
)) {
1925 osd
->reply_op_error(op
, -EPERM
);
1929 if (op
->includes_pg_op()) {
1930 return do_pg_op(op
);
1933 // object name too long?
1934 if (m
->get_oid().name
.size() > cct
->_conf
->osd_max_object_name_len
) {
1935 dout(4) << "do_op name is longer than "
1936 << cct
->_conf
->osd_max_object_name_len
1937 << " bytes" << dendl
;
1938 osd
->reply_op_error(op
, -ENAMETOOLONG
);
1941 if (m
->get_hobj().get_key().size() > cct
->_conf
->osd_max_object_name_len
) {
1942 dout(4) << "do_op locator is longer than "
1943 << cct
->_conf
->osd_max_object_name_len
1944 << " bytes" << dendl
;
1945 osd
->reply_op_error(op
, -ENAMETOOLONG
);
1948 if (m
->get_hobj().nspace
.size() > cct
->_conf
->osd_max_object_namespace_len
) {
1949 dout(4) << "do_op namespace is longer than "
1950 << cct
->_conf
->osd_max_object_namespace_len
1951 << " bytes" << dendl
;
1952 osd
->reply_op_error(op
, -ENAMETOOLONG
);
1956 if (int r
= osd
->store
->validate_hobject_key(head
)) {
1957 dout(4) << "do_op object " << head
<< " invalid for backing store: "
1959 osd
->reply_op_error(op
, r
);
1964 if (get_osdmap()->is_blacklisted(m
->get_source_addr())) {
1965 dout(10) << "do_op " << m
->get_source_addr() << " is blacklisted" << dendl
;
1966 osd
->reply_op_error(op
, -EBLACKLISTED
);
1970 // order this op as a write?
1971 bool write_ordered
= op
->rwordered();
1973 // discard due to cluster full transition? (we discard any op that
1974 // originates before the cluster or pool is marked full; the client
1975 // will resend after the full flag is removed or if they expect the
1976 // op to succeed despite being full). The except is FULL_FORCE and
1977 // FULL_TRY ops, which there is no reason to discard because they
1978 // bypass all full checks anyway. If this op isn't write or
1979 // read-ordered, we skip.
1980 // FIXME: we exclude mds writes for now.
1981 if (write_ordered
&& !(m
->get_source().is_mds() ||
1982 m
->has_flag(CEPH_OSD_FLAG_FULL_TRY
) ||
1983 m
->has_flag(CEPH_OSD_FLAG_FULL_FORCE
)) &&
1984 info
.history
.last_epoch_marked_full
> m
->get_map_epoch()) {
1985 dout(10) << __func__
<< " discarding op sent before full " << m
<< " "
1989 // mds should have stopped writing before this point.
1990 // We can't allow OSD to become non-startable even if mds
1991 // could be writing as part of file removals.
1993 if (write_ordered
&& osd
->check_failsafe_full(ss
)) {
1994 dout(10) << __func__
<< " fail-safe full check failed, dropping request"
1999 int64_t poolid
= get_pgid().pool();
2000 if (op
->may_write()) {
2002 const pg_pool_t
*pi
= get_osdmap()->get_pg_pool(poolid
);
2008 if (m
->get_snapid() != CEPH_NOSNAP
) {
2009 dout(20) << __func__
<< ": write to clone not valid " << *m
<< dendl
;
2010 osd
->reply_op_error(op
, -EINVAL
);
2015 if (cct
->_conf
->osd_max_write_size
&&
2016 m
->get_data_len() > cct
->_conf
->osd_max_write_size
<< 20) {
2017 // journal can't hold commit!
2018 derr
<< "do_op msg data len " << m
->get_data_len()
2019 << " > osd_max_write_size " << (cct
->_conf
->osd_max_write_size
<< 20)
2020 << " on " << *m
<< dendl
;
2021 osd
->reply_op_error(op
, -OSD_WRITETOOBIG
);
2026 dout(10) << "do_op " << *m
2027 << (op
->may_write() ? " may_write" : "")
2028 << (op
->may_read() ? " may_read" : "")
2029 << (op
->may_cache() ? " may_cache" : "")
2030 << " -> " << (write_ordered
? "write-ordered" : "read-ordered")
2031 << " flags " << ceph_osd_flag_string(m
->get_flags())
2035 if (is_unreadable_object(head
)) {
2036 if (!is_primary()) {
2037 osd
->reply_op_error(op
, -EAGAIN
);
2041 (g_conf
->osd_backoff_on_degraded
||
2042 (g_conf
->osd_backoff_on_unfound
&& missing_loc
.is_unfound(head
)))) {
2043 add_backoff(session
, head
, head
);
2044 maybe_kick_recovery(head
);
2046 wait_for_unreadable_object(head
, op
);
2052 if (write_ordered
&& is_degraded_or_backfilling_object(head
)) {
2053 if (can_backoff
&& g_conf
->osd_backoff_on_degraded
) {
2054 add_backoff(session
, head
, head
);
2055 maybe_kick_recovery(head
);
2057 wait_for_degraded_object(head
, op
);
2062 if (write_ordered
&& scrubber
.is_chunky_scrub_active() &&
2063 write_blocked_by_scrub(head
)) {
2064 dout(20) << __func__
<< ": waiting for scrub" << dendl
;
2065 waiting_for_scrub
.push_back(op
);
2066 op
->mark_delayed("waiting for scrub");
2071 map
<hobject_t
, snapid_t
>::iterator blocked_iter
=
2072 objects_blocked_on_degraded_snap
.find(head
);
2073 if (write_ordered
&& blocked_iter
!= objects_blocked_on_degraded_snap
.end()) {
2074 hobject_t
to_wait_on(head
);
2075 to_wait_on
.snap
= blocked_iter
->second
;
2076 wait_for_degraded_object(to_wait_on
, op
);
2079 map
<hobject_t
, ObjectContextRef
>::iterator blocked_snap_promote_iter
=
2080 objects_blocked_on_snap_promotion
.find(head
);
2081 if (write_ordered
&&
2082 blocked_snap_promote_iter
!= objects_blocked_on_snap_promotion
.end()) {
2083 wait_for_blocked_object(
2084 blocked_snap_promote_iter
->second
->obs
.oi
.soid
,
2088 if (write_ordered
&& objects_blocked_on_cache_full
.count(head
)) {
2089 block_write_on_full_cache(head
, op
);
2094 hobject_t snapdir
= head
.get_snapdir();
2096 if (is_unreadable_object(snapdir
)) {
2097 wait_for_unreadable_object(snapdir
, op
);
2102 if (write_ordered
&& is_degraded_or_backfilling_object(snapdir
)) {
2103 wait_for_degraded_object(snapdir
, op
);
2108 if (op
->may_write() || op
->may_cache()) {
2109 // warning: we will get back *a* request for this reqid, but not
2110 // necessarily the most recent. this happens with flush and
2111 // promote ops, but we can't possible have both in our log where
2112 // the original request is still not stable on disk, so for our
2113 // purposes here it doesn't matter which one we get.
2115 version_t user_version
;
2116 int return_code
= 0;
2117 bool got
= check_in_progress_op(
2118 m
->get_reqid(), &version
, &user_version
, &return_code
);
2120 dout(3) << __func__
<< " dup " << m
->get_reqid()
2121 << " version " << version
<< dendl
;
2122 if (already_complete(version
)) {
2123 osd
->reply_op_error(op
, return_code
, version
, user_version
);
2125 dout(10) << " waiting for " << version
<< " to commit" << dendl
;
2126 // always queue ondisk waiters, so that we can requeue if needed
2127 waiting_for_ondisk
[version
].push_back(make_pair(op
, user_version
));
2128 op
->mark_delayed("waiting for ondisk");
2134 ObjectContextRef obc
;
2135 bool can_create
= op
->may_write() || op
->may_cache();
2136 hobject_t missing_oid
;
2137 const hobject_t
& oid
= m
->get_hobj();
2139 // io blocked on obc?
2140 if (!m
->has_flag(CEPH_OSD_FLAG_FLUSH
) &&
2141 maybe_await_blocked_snapset(oid
, op
)) {
2145 int r
= find_object_context(
2146 oid
, &obc
, can_create
,
2147 m
->has_flag(CEPH_OSD_FLAG_MAP_SNAP_CLONE
),
2151 // If we're not the primary of this OSD, we just return -EAGAIN. Otherwise,
2152 // we have to wait for the object.
2154 // missing the specific snap we need; requeue and wait.
2155 assert(!op
->may_write()); // only happens on a read/cache
2156 wait_for_unreadable_object(missing_oid
, op
);
2159 } else if (r
== 0) {
2160 if (is_unreadable_object(obc
->obs
.oi
.soid
)) {
2161 dout(10) << __func__
<< ": clone " << obc
->obs
.oi
.soid
2162 << " is unreadable, waiting" << dendl
;
2163 wait_for_unreadable_object(obc
->obs
.oi
.soid
, op
);
2167 // degraded object? (the check above was for head; this could be a clone)
2168 if (write_ordered
&&
2169 obc
->obs
.oi
.soid
.snap
!= CEPH_NOSNAP
&&
2170 is_degraded_or_backfilling_object(obc
->obs
.oi
.soid
)) {
2171 dout(10) << __func__
<< ": clone " << obc
->obs
.oi
.soid
2172 << " is degraded, waiting" << dendl
;
2173 wait_for_degraded_object(obc
->obs
.oi
.soid
, op
);
2178 bool in_hit_set
= false;
2181 if (obc
->obs
.oi
.soid
!= hobject_t() && hit_set
->contains(obc
->obs
.oi
.soid
))
2184 if (missing_oid
!= hobject_t() && hit_set
->contains(missing_oid
))
2187 if (!op
->hitset_inserted
) {
2188 hit_set
->insert(oid
);
2189 op
->hitset_inserted
= true;
2190 if (hit_set
->is_full() ||
2191 hit_set_start_stamp
+ pool
.info
.hit_set_period
<= m
->get_recv_stamp()) {
2198 if (agent_choose_mode(false, op
))
2202 if (obc
.get() && obc
->obs
.exists
&& obc
->obs
.oi
.has_manifest()) {
2203 if (maybe_handle_manifest(op
,
2209 if (maybe_handle_cache(op
,
2218 if (r
&& (r
!= -ENOENT
|| !obc
)) {
2219 // copy the reqids for copy get on ENOENT
2221 (m
->ops
[0].op
.op
== CEPH_OSD_OP_COPY_GET
)) {
2222 fill_in_copy_get_noent(op
, oid
, m
->ops
[0]);
2225 dout(20) << __func__
<< ": find_object_context got error " << r
<< dendl
;
2226 if (op
->may_write() &&
2227 get_osdmap()->require_osd_release
>= CEPH_RELEASE_KRAKEN
) {
2228 record_write_error(op
, oid
, nullptr, r
);
2230 osd
->reply_op_error(op
, r
);
2235 // make sure locator is consistent
2236 object_locator_t
oloc(obc
->obs
.oi
.soid
);
2237 if (m
->get_object_locator() != oloc
) {
2238 dout(10) << " provided locator " << m
->get_object_locator()
2239 << " != object's " << obc
->obs
.oi
.soid
<< dendl
;
2240 osd
->clog
->warn() << "bad locator " << m
->get_object_locator()
2241 << " on object " << oloc
2245 // io blocked on obc?
2246 if (obc
->is_blocked() &&
2247 !m
->has_flag(CEPH_OSD_FLAG_FLUSH
)) {
2248 wait_for_blocked_object(obc
->obs
.oi
.soid
, op
);
2252 dout(25) << __func__
<< " oi " << obc
->obs
.oi
<< dendl
;
2254 for (vector
<OSDOp
>::iterator p
= m
->ops
.begin(); p
!= m
->ops
.end(); ++p
) {
2257 // make sure LIST_SNAPS is on CEPH_SNAPDIR and nothing else
2258 if (osd_op
.op
.op
== CEPH_OSD_OP_LIST_SNAPS
&&
2259 m
->get_snapid() != CEPH_SNAPDIR
) {
2260 dout(10) << "LIST_SNAPS with incorrect context" << dendl
;
2261 osd
->reply_op_error(op
, -EINVAL
);
2266 OpContext
*ctx
= new OpContext(op
, m
->get_reqid(), &m
->ops
, obc
, this);
2268 if (!obc
->obs
.exists
)
2269 ctx
->snapset_obc
= get_object_context(obc
->obs
.oi
.soid
.get_snapdir(), false);
2271 /* Due to obc caching, we might have a cached non-existent snapset_obc
2272 * for the snapdir. If so, we can ignore it. Subsequent parts of the
2273 * do_op pipeline make decisions based on whether snapset_obc is
2276 if (ctx
->snapset_obc
&& !ctx
->snapset_obc
->obs
.exists
)
2277 ctx
->snapset_obc
= ObjectContextRef();
2279 if (m
->has_flag(CEPH_OSD_FLAG_SKIPRWLOCKS
)) {
2280 dout(20) << __func__
<< ": skipping rw locks" << dendl
;
2281 } else if (m
->get_flags() & CEPH_OSD_FLAG_FLUSH
) {
2282 dout(20) << __func__
<< ": part of flush, will ignore write lock" << dendl
;
2284 // verify there is in fact a flush in progress
2285 // FIXME: we could make this a stronger test.
2286 map
<hobject_t
,FlushOpRef
>::iterator p
= flush_ops
.find(obc
->obs
.oi
.soid
);
2287 if (p
== flush_ops
.end()) {
2288 dout(10) << __func__
<< " no flush in progress, aborting" << dendl
;
2289 reply_ctx(ctx
, -EINVAL
);
2292 } else if (!get_rw_locks(write_ordered
, ctx
)) {
2293 dout(20) << __func__
<< " waiting for rw locks " << dendl
;
2294 op
->mark_delayed("waiting for rw locks");
2298 dout(20) << __func__
<< " obc " << *obc
<< dendl
;
2301 dout(20) << __func__
<< " returned an error: " << r
<< dendl
;
2303 if (op
->may_write() &&
2304 get_osdmap()->require_osd_release
>= CEPH_RELEASE_KRAKEN
) {
2305 record_write_error(op
, oid
, nullptr, r
);
2307 osd
->reply_op_error(op
, r
);
2312 if (m
->has_flag(CEPH_OSD_FLAG_IGNORE_CACHE
)) {
2313 ctx
->ignore_cache
= true;
2316 if ((op
->may_read()) && (obc
->obs
.oi
.is_lost())) {
2317 // This object is lost. Reading from it returns an error.
2318 dout(20) << __func__
<< ": object " << obc
->obs
.oi
.soid
2319 << " is lost" << dendl
;
2320 reply_ctx(ctx
, -ENFILE
);
2323 if (!op
->may_write() &&
2325 (!obc
->obs
.exists
||
2326 ((m
->get_snapid() != CEPH_SNAPDIR
) &&
2327 obc
->obs
.oi
.is_whiteout()))) {
2328 // copy the reqids for copy get on ENOENT
2329 if (m
->ops
[0].op
.op
== CEPH_OSD_OP_COPY_GET
) {
2330 fill_in_copy_get_noent(op
, oid
, m
->ops
[0]);
2334 reply_ctx(ctx
, -ENOENT
);
2341 utime_t prepare_latency
= ceph_clock_now();
2342 prepare_latency
-= op
->get_dequeued_time();
2343 osd
->logger
->tinc(l_osd_op_prepare_lat
, prepare_latency
);
2344 if (op
->may_read() && op
->may_write()) {
2345 osd
->logger
->tinc(l_osd_op_rw_prepare_lat
, prepare_latency
);
2346 } else if (op
->may_read()) {
2347 osd
->logger
->tinc(l_osd_op_r_prepare_lat
, prepare_latency
);
2348 } else if (op
->may_write() || op
->may_cache()) {
2349 osd
->logger
->tinc(l_osd_op_w_prepare_lat
, prepare_latency
);
2352 // force recovery of the oldest missing object if too many logs
2353 maybe_force_recovery();
2356 PrimaryLogPG::cache_result_t
PrimaryLogPG::maybe_handle_manifest_detail(
2359 ObjectContextRef obc
)
2361 if (static_cast<const MOSDOp
*>(op
->get_req())->get_flags() &
2362 CEPH_OSD_FLAG_IGNORE_REDIRECT
) {
2363 dout(20) << __func__
<< ": ignoring redirect due to flag" << dendl
;
2364 return cache_result_t::NOOP
;
2368 dout(10) << __func__
<< " " << obc
->obs
.oi
<< " "
2369 << (obc
->obs
.exists
? "exists" : "DNE")
2372 // if it is write-ordered and blocked, stop now
2373 if (obc
.get() && obc
->is_blocked() && write_ordered
) {
2374 // we're already doing something with this object
2375 dout(20) << __func__
<< " blocked on " << obc
->obs
.oi
.soid
<< dendl
;
2376 return cache_result_t::NOOP
;
2379 vector
<OSDOp
> ops
= static_cast<const MOSDOp
*>(op
->get_req())->ops
;
2380 for (vector
<OSDOp
>::iterator p
= ops
.begin(); p
!= ops
.end(); ++p
) {
2382 ceph_osd_op
& op
= osd_op
.op
;
2383 if (op
.op
== CEPH_OSD_OP_SET_REDIRECT
) {
2384 return cache_result_t::NOOP
;
2388 switch (obc
->obs
.oi
.manifest
.type
) {
2389 case object_manifest_t::TYPE_REDIRECT
:
2390 if (op
->may_write() || write_ordered
) {
2391 do_proxy_write(op
, obc
->obs
.oi
.soid
, obc
);
2393 do_proxy_read(op
, obc
);
2395 return cache_result_t::HANDLED_PROXY
;
2396 case object_manifest_t::TYPE_CHUNKED
:
2398 assert(0 == "unrecognized manifest type");
2401 return cache_result_t::NOOP
;
2404 void PrimaryLogPG::record_write_error(OpRequestRef op
, const hobject_t
&soid
,
2405 MOSDOpReply
*orig_reply
, int r
)
2407 dout(20) << __func__
<< " r=" << r
<< dendl
;
2408 assert(op
->may_write());
2409 const osd_reqid_t
&reqid
= static_cast<const MOSDOp
*>(op
->get_req())->get_reqid();
2410 mempool::osd_pglog::list
<pg_log_entry_t
> entries
;
2411 entries
.push_back(pg_log_entry_t(pg_log_entry_t::ERROR
, soid
,
2412 get_next_version(), eversion_t(), 0,
2413 reqid
, utime_t(), r
));
2418 boost::intrusive_ptr
<MOSDOpReply
> orig_reply
;
2423 MOSDOpReply
*orig_reply
,
2426 orig_reply(orig_reply
, false /* take over ref */), r(r
)
2429 ldpp_dout(pg
, 20) << "finished " << __func__
<< " r=" << r
<< dendl
;
2430 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
2431 int flags
= m
->get_flags() & (CEPH_OSD_FLAG_ACK
| CEPH_OSD_FLAG_ONDISK
);
2432 MOSDOpReply
*reply
= orig_reply
.detach();
2433 if (reply
== nullptr) {
2434 reply
= new MOSDOpReply(m
, r
, pg
->get_osdmap()->get_epoch(),
2437 ldpp_dout(pg
, 10) << " sending commit on " << *m
<< " " << reply
<< dendl
;
2438 pg
->osd
->send_message_osd_client(reply
, m
->get_connection());
2442 ObcLockManager lock_manager
;
2445 std::move(lock_manager
),
2446 boost::optional
<std::function
<void(void)> >(
2447 OnComplete(this, op
, orig_reply
, r
)),
2452 PrimaryLogPG::cache_result_t
PrimaryLogPG::maybe_handle_cache_detail(
2455 ObjectContextRef obc
,
2456 int r
, hobject_t missing_oid
,
2459 ObjectContextRef
*promote_obc
)
2461 // return quickly if caching is not enabled
2462 if (pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_NONE
)
2463 return cache_result_t::NOOP
;
2467 op
->get_req()->get_type() == CEPH_MSG_OSD_OP
&&
2468 (static_cast<const MOSDOp
*>(op
->get_req())->get_flags() &
2469 CEPH_OSD_FLAG_IGNORE_CACHE
)) {
2470 dout(20) << __func__
<< ": ignoring cache due to flag" << dendl
;
2471 return cache_result_t::NOOP
;
2474 must_promote
= must_promote
|| op
->need_promote();
2477 dout(25) << __func__
<< " " << obc
->obs
.oi
<< " "
2478 << (obc
->obs
.exists
? "exists" : "DNE")
2479 << " missing_oid " << missing_oid
2480 << " must_promote " << (int)must_promote
2481 << " in_hit_set " << (int)in_hit_set
2484 dout(25) << __func__
<< " (no obc)"
2485 << " missing_oid " << missing_oid
2486 << " must_promote " << (int)must_promote
2487 << " in_hit_set " << (int)in_hit_set
2490 // if it is write-ordered and blocked, stop now
2491 if (obc
.get() && obc
->is_blocked() && write_ordered
) {
2492 // we're already doing something with this object
2493 dout(20) << __func__
<< " blocked on " << obc
->obs
.oi
.soid
<< dendl
;
2494 return cache_result_t::NOOP
;
2497 if (r
== -ENOENT
&& missing_oid
== hobject_t()) {
2498 // we know this object is logically absent (e.g., an undefined clone)
2499 return cache_result_t::NOOP
;
2502 if (obc
.get() && obc
->obs
.exists
) {
2503 osd
->logger
->inc(l_osd_op_cache_hit
);
2504 return cache_result_t::NOOP
;
2506 if (!is_primary()) {
2507 dout(20) << __func__
<< " cache miss; ask the primary" << dendl
;
2508 osd
->reply_op_error(op
, -EAGAIN
);
2509 return cache_result_t::REPLIED_WITH_EAGAIN
;
2512 if (missing_oid
== hobject_t() && obc
.get()) {
2513 missing_oid
= obc
->obs
.oi
.soid
;
2516 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
2517 const object_locator_t oloc
= m
->get_object_locator();
2519 if (op
->need_skip_handle_cache()) {
2520 return cache_result_t::NOOP
;
2523 // older versions do not proxy the feature bits.
2524 bool can_proxy_write
= get_osdmap()->get_up_osd_features() &
2525 CEPH_FEATURE_OSD_PROXY_WRITE_FEATURES
;
2526 OpRequestRef promote_op
;
2528 switch (pool
.info
.cache_mode
) {
2529 case pg_pool_t::CACHEMODE_WRITEBACK
:
2531 agent_state
->evict_mode
== TierAgentState::EVICT_MODE_FULL
) {
2532 if (!op
->may_write() && !op
->may_cache() &&
2533 !write_ordered
&& !must_promote
) {
2534 dout(20) << __func__
<< " cache pool full, proxying read" << dendl
;
2536 return cache_result_t::HANDLED_PROXY
;
2538 dout(20) << __func__
<< " cache pool full, waiting" << dendl
;
2539 block_write_on_full_cache(missing_oid
, op
);
2540 return cache_result_t::BLOCKED_FULL
;
2543 if (must_promote
|| (!hit_set
&& !op
->need_skip_promote())) {
2544 promote_object(obc
, missing_oid
, oloc
, op
, promote_obc
);
2545 return cache_result_t::BLOCKED_PROMOTE
;
2548 if (op
->may_write() || op
->may_cache()) {
2549 if (can_proxy_write
) {
2550 do_proxy_write(op
, missing_oid
);
2552 // promote if can't proxy the write
2553 promote_object(obc
, missing_oid
, oloc
, op
, promote_obc
);
2554 return cache_result_t::BLOCKED_PROMOTE
;
2558 if (!op
->need_skip_promote() &&
2559 maybe_promote(obc
, missing_oid
, oloc
, in_hit_set
,
2560 pool
.info
.min_write_recency_for_promote
,
2563 return cache_result_t::BLOCKED_PROMOTE
;
2565 return cache_result_t::HANDLED_PROXY
;
2569 // Avoid duplicate promotion
2570 if (obc
.get() && obc
->is_blocked()) {
2573 return cache_result_t::BLOCKED_PROMOTE
;
2577 if (!op
->need_skip_promote()) {
2578 (void)maybe_promote(obc
, missing_oid
, oloc
, in_hit_set
,
2579 pool
.info
.min_read_recency_for_promote
,
2580 promote_op
, promote_obc
);
2583 return cache_result_t::HANDLED_PROXY
;
2585 assert(0 == "unreachable");
2586 return cache_result_t::NOOP
;
2588 case pg_pool_t::CACHEMODE_FORWARD
:
2589 // FIXME: this mode allows requests to be reordered.
2590 do_cache_redirect(op
);
2591 return cache_result_t::HANDLED_REDIRECT
;
2593 case pg_pool_t::CACHEMODE_READONLY
:
2594 // TODO: clean this case up
2595 if (!obc
.get() && r
== -ENOENT
) {
2596 // we don't have the object and op's a read
2597 promote_object(obc
, missing_oid
, oloc
, op
, promote_obc
);
2598 return cache_result_t::BLOCKED_PROMOTE
;
2600 if (!r
) { // it must be a write
2601 do_cache_redirect(op
);
2602 return cache_result_t::HANDLED_REDIRECT
;
2604 // crap, there was a failure of some kind
2605 return cache_result_t::NOOP
;
2607 case pg_pool_t::CACHEMODE_READFORWARD
:
2608 // Do writeback to the cache tier for writes
2609 if (op
->may_write() || write_ordered
|| must_promote
) {
2611 agent_state
->evict_mode
== TierAgentState::EVICT_MODE_FULL
) {
2612 dout(20) << __func__
<< " cache pool full, waiting" << dendl
;
2613 block_write_on_full_cache(missing_oid
, op
);
2614 return cache_result_t::BLOCKED_FULL
;
2616 promote_object(obc
, missing_oid
, oloc
, op
, promote_obc
);
2617 return cache_result_t::BLOCKED_PROMOTE
;
2620 // If it is a read, we can read, we need to forward it
2621 do_cache_redirect(op
);
2622 return cache_result_t::HANDLED_REDIRECT
;
2624 case pg_pool_t::CACHEMODE_PROXY
:
2625 if (!must_promote
) {
2626 if (op
->may_write() || op
->may_cache() || write_ordered
) {
2627 if (can_proxy_write
) {
2628 do_proxy_write(op
, missing_oid
);
2629 return cache_result_t::HANDLED_PROXY
;
2633 return cache_result_t::HANDLED_PROXY
;
2636 // ugh, we're forced to promote.
2638 agent_state
->evict_mode
== TierAgentState::EVICT_MODE_FULL
) {
2639 dout(20) << __func__
<< " cache pool full, waiting" << dendl
;
2640 block_write_on_full_cache(missing_oid
, op
);
2641 return cache_result_t::BLOCKED_FULL
;
2643 promote_object(obc
, missing_oid
, oloc
, op
, promote_obc
);
2644 return cache_result_t::BLOCKED_PROMOTE
;
2646 case pg_pool_t::CACHEMODE_READPROXY
:
2647 // Do writeback to the cache tier for writes
2648 if (op
->may_write() || write_ordered
|| must_promote
) {
2650 agent_state
->evict_mode
== TierAgentState::EVICT_MODE_FULL
) {
2651 dout(20) << __func__
<< " cache pool full, waiting" << dendl
;
2652 block_write_on_full_cache(missing_oid
, op
);
2653 return cache_result_t::BLOCKED_FULL
;
2655 promote_object(obc
, missing_oid
, oloc
, op
, promote_obc
);
2656 return cache_result_t::BLOCKED_PROMOTE
;
2659 // If it is a read, we can read, we need to proxy it
2661 return cache_result_t::HANDLED_PROXY
;
2664 assert(0 == "unrecognized cache_mode");
2666 return cache_result_t::NOOP
;
2669 bool PrimaryLogPG::maybe_promote(ObjectContextRef obc
,
2670 const hobject_t
& missing_oid
,
2671 const object_locator_t
& oloc
,
2674 OpRequestRef promote_op
,
2675 ObjectContextRef
*promote_obc
)
2677 dout(20) << __func__
<< " missing_oid " << missing_oid
2678 << " in_hit_set " << in_hit_set
<< dendl
;
2684 // Check if in the current hit set
2694 unsigned count
= (int)in_hit_set
;
2696 // Check if in other hit sets
2697 const hobject_t
& oid
= obc
.get() ? obc
->obs
.oi
.soid
: missing_oid
;
2698 for (map
<time_t,HitSetRef
>::reverse_iterator itor
=
2699 agent_state
->hit_set_map
.rbegin();
2700 itor
!= agent_state
->hit_set_map
.rend();
2702 if (!itor
->second
->contains(oid
)) {
2706 if (count
>= recency
) {
2711 if (count
>= recency
) {
2714 return false; // not promoting
2719 if (osd
->promote_throttle()) {
2720 dout(10) << __func__
<< " promote throttled" << dendl
;
2723 promote_object(obc
, missing_oid
, oloc
, promote_op
, promote_obc
);
2727 void PrimaryLogPG::do_cache_redirect(OpRequestRef op
)
2729 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
2730 int flags
= m
->get_flags() & (CEPH_OSD_FLAG_ACK
|CEPH_OSD_FLAG_ONDISK
);
2731 MOSDOpReply
*reply
= new MOSDOpReply(m
, -ENOENT
,
2732 get_osdmap()->get_epoch(), flags
, false);
2733 request_redirect_t
redir(m
->get_object_locator(), pool
.info
.tier_of
);
2734 reply
->set_redirect(redir
);
2735 dout(10) << "sending redirect to pool " << pool
.info
.tier_of
<< " for op "
2737 m
->get_connection()->send_message(reply
);
2741 struct C_ProxyRead
: public Context
{
2744 epoch_t last_peering_reset
;
2746 PrimaryLogPG::ProxyReadOpRef prdop
;
2748 C_ProxyRead(PrimaryLogPG
*p
, hobject_t o
, epoch_t lpr
,
2749 const PrimaryLogPG::ProxyReadOpRef
& prd
)
2750 : pg(p
), oid(o
), last_peering_reset(lpr
),
2751 tid(0), prdop(prd
), start(ceph_clock_now())
2753 void finish(int r
) override
{
2754 if (prdop
->canceled
)
2757 if (prdop
->canceled
) {
2761 if (last_peering_reset
== pg
->get_last_peering_reset()) {
2762 pg
->finish_proxy_read(oid
, tid
, r
);
2763 pg
->osd
->logger
->tinc(l_osd_tier_r_lat
, ceph_clock_now() - start
);
2769 void PrimaryLogPG::do_proxy_read(OpRequestRef op
, ObjectContextRef obc
)
2771 // NOTE: non-const here because the ProxyReadOp needs mutable refs to
2772 // stash the result in the request's OSDOp vector
2773 MOSDOp
*m
= static_cast<MOSDOp
*>(op
->get_nonconst_req());
2774 object_locator_t oloc
;
2776 /* extensible tier */
2777 if (obc
&& obc
->obs
.exists
&& obc
->obs
.oi
.has_manifest()) {
2778 switch (obc
->obs
.oi
.manifest
.type
) {
2779 case object_manifest_t::TYPE_REDIRECT
:
2780 oloc
= object_locator_t(obc
->obs
.oi
.manifest
.redirect_target
);
2781 soid
= obc
->obs
.oi
.manifest
.redirect_target
;
2783 case object_manifest_t::TYPE_CHUNKED
:
2785 assert(0 == "unrecognized manifest type");
2789 soid
= m
->get_hobj();
2790 oloc
= object_locator_t(m
->get_object_locator());
2791 oloc
.pool
= pool
.info
.tier_of
;
2793 unsigned flags
= CEPH_OSD_FLAG_IGNORE_CACHE
| CEPH_OSD_FLAG_IGNORE_OVERLAY
;
2795 // pass through some original flags that make sense.
2796 // - leave out redirection and balancing flags since we are
2797 // already proxying through the primary
2798 // - leave off read/write/exec flags that are derived from the op
2799 flags
|= m
->get_flags() & (CEPH_OSD_FLAG_RWORDERED
|
2800 CEPH_OSD_FLAG_ORDERSNAP
|
2801 CEPH_OSD_FLAG_ENFORCE_SNAPC
|
2802 CEPH_OSD_FLAG_MAP_SNAP_CLONE
);
2804 dout(10) << __func__
<< " Start proxy read for " << *m
<< dendl
;
2806 ProxyReadOpRef
prdop(std::make_shared
<ProxyReadOp
>(op
, soid
, m
->ops
));
2808 ObjectOperation obj_op
;
2809 obj_op
.dup(prdop
->ops
);
2811 if (pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_WRITEBACK
&&
2812 (agent_state
&& agent_state
->evict_mode
!= TierAgentState::EVICT_MODE_FULL
)) {
2813 for (unsigned i
= 0; i
< obj_op
.ops
.size(); i
++) {
2814 ceph_osd_op op
= obj_op
.ops
[i
].op
;
2816 case CEPH_OSD_OP_READ
:
2817 case CEPH_OSD_OP_SYNC_READ
:
2818 case CEPH_OSD_OP_SPARSE_READ
:
2819 case CEPH_OSD_OP_CHECKSUM
:
2820 case CEPH_OSD_OP_CMPEXT
:
2821 op
.flags
= (op
.flags
| CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL
) &
2822 ~(CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
| CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
);
2827 C_ProxyRead
*fin
= new C_ProxyRead(this, soid
, get_last_peering_reset(),
2829 ceph_tid_t tid
= osd
->objecter
->read(
2830 soid
.oid
, oloc
, obj_op
,
2831 m
->get_snapid(), NULL
,
2832 flags
, new C_OnFinisher(fin
, &osd
->objecter_finisher
),
2833 &prdop
->user_version
,
2834 &prdop
->data_offset
,
2837 prdop
->objecter_tid
= tid
;
2838 proxyread_ops
[tid
] = prdop
;
2839 in_progress_proxy_ops
[soid
].push_back(op
);
2842 void PrimaryLogPG::finish_proxy_read(hobject_t oid
, ceph_tid_t tid
, int r
)
2844 dout(10) << __func__
<< " " << oid
<< " tid " << tid
2845 << " " << cpp_strerror(r
) << dendl
;
2847 map
<ceph_tid_t
, ProxyReadOpRef
>::iterator p
= proxyread_ops
.find(tid
);
2848 if (p
== proxyread_ops
.end()) {
2849 dout(10) << __func__
<< " no proxyread_op found" << dendl
;
2852 ProxyReadOpRef prdop
= p
->second
;
2853 if (tid
!= prdop
->objecter_tid
) {
2854 dout(10) << __func__
<< " tid " << tid
<< " != prdop " << prdop
2855 << " tid " << prdop
->objecter_tid
<< dendl
;
2858 if (oid
!= prdop
->soid
) {
2859 dout(10) << __func__
<< " oid " << oid
<< " != prdop " << prdop
2860 << " soid " << prdop
->soid
<< dendl
;
2863 proxyread_ops
.erase(tid
);
2865 map
<hobject_t
, list
<OpRequestRef
>>::iterator q
= in_progress_proxy_ops
.find(oid
);
2866 if (q
== in_progress_proxy_ops
.end()) {
2867 dout(10) << __func__
<< " no in_progress_proxy_ops found" << dendl
;
2870 assert(q
->second
.size());
2871 list
<OpRequestRef
>::iterator it
= std::find(q
->second
.begin(),
2874 assert(it
!= q
->second
.end());
2875 OpRequestRef op
= *it
;
2876 q
->second
.erase(it
);
2877 if (q
->second
.size() == 0) {
2878 in_progress_proxy_ops
.erase(oid
);
2881 osd
->logger
->inc(l_osd_tier_proxy_read
);
2883 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
2884 OpContext
*ctx
= new OpContext(op
, m
->get_reqid(), &prdop
->ops
, this);
2885 ctx
->reply
= new MOSDOpReply(m
, 0, get_osdmap()->get_epoch(), 0, false);
2886 ctx
->user_at_version
= prdop
->user_version
;
2887 ctx
->data_off
= prdop
->data_offset
;
2888 ctx
->ignore_log_op_stats
= true;
2889 complete_read_ctx(r
, ctx
);
2892 void PrimaryLogPG::kick_proxy_ops_blocked(hobject_t
& soid
)
2894 map
<hobject_t
, list
<OpRequestRef
>>::iterator p
= in_progress_proxy_ops
.find(soid
);
2895 if (p
== in_progress_proxy_ops
.end())
2898 list
<OpRequestRef
>& ls
= p
->second
;
2899 dout(10) << __func__
<< " " << soid
<< " requeuing " << ls
.size() << " requests" << dendl
;
2901 in_progress_proxy_ops
.erase(p
);
2904 void PrimaryLogPG::cancel_proxy_read(ProxyReadOpRef prdop
,
2905 vector
<ceph_tid_t
> *tids
)
2907 dout(10) << __func__
<< " " << prdop
->soid
<< dendl
;
2908 prdop
->canceled
= true;
2910 // cancel objecter op, if we can
2911 if (prdop
->objecter_tid
) {
2912 tids
->push_back(prdop
->objecter_tid
);
2913 for (uint32_t i
= 0; i
< prdop
->ops
.size(); i
++) {
2914 prdop
->ops
[i
].outdata
.clear();
2916 proxyread_ops
.erase(prdop
->objecter_tid
);
2917 prdop
->objecter_tid
= 0;
2921 void PrimaryLogPG::cancel_proxy_ops(bool requeue
, vector
<ceph_tid_t
> *tids
)
2923 dout(10) << __func__
<< dendl
;
2925 // cancel proxy reads
2926 map
<ceph_tid_t
, ProxyReadOpRef
>::iterator p
= proxyread_ops
.begin();
2927 while (p
!= proxyread_ops
.end()) {
2928 cancel_proxy_read((p
++)->second
, tids
);
2931 // cancel proxy writes
2932 map
<ceph_tid_t
, ProxyWriteOpRef
>::iterator q
= proxywrite_ops
.begin();
2933 while (q
!= proxywrite_ops
.end()) {
2934 cancel_proxy_write((q
++)->second
, tids
);
2938 map
<hobject_t
, list
<OpRequestRef
>>::iterator p
=
2939 in_progress_proxy_ops
.begin();
2940 while (p
!= in_progress_proxy_ops
.end()) {
2941 list
<OpRequestRef
>& ls
= p
->second
;
2942 dout(10) << __func__
<< " " << p
->first
<< " requeuing " << ls
.size()
2943 << " requests" << dendl
;
2945 in_progress_proxy_ops
.erase(p
++);
2948 in_progress_proxy_ops
.clear();
2952 struct C_ProxyWrite_Commit
: public Context
{
2955 epoch_t last_peering_reset
;
2957 PrimaryLogPG::ProxyWriteOpRef pwop
;
2958 C_ProxyWrite_Commit(PrimaryLogPG
*p
, hobject_t o
, epoch_t lpr
,
2959 const PrimaryLogPG::ProxyWriteOpRef
& pw
)
2960 : pg(p
), oid(o
), last_peering_reset(lpr
),
2963 void finish(int r
) override
{
2967 if (pwop
->canceled
) {
2971 if (last_peering_reset
== pg
->get_last_peering_reset()) {
2972 pg
->finish_proxy_write(oid
, tid
, r
);
2978 void PrimaryLogPG::do_proxy_write(OpRequestRef op
, const hobject_t
& missing_oid
, ObjectContextRef obc
)
2980 // NOTE: non-const because ProxyWriteOp takes a mutable ref
2981 MOSDOp
*m
= static_cast<MOSDOp
*>(op
->get_nonconst_req());
2982 object_locator_t oloc
;
2983 SnapContext
snapc(m
->get_snap_seq(), m
->get_snaps());
2985 /* extensible tier */
2986 if (obc
&& obc
->obs
.exists
&& obc
->obs
.oi
.has_manifest()) {
2987 switch (obc
->obs
.oi
.manifest
.type
) {
2988 case object_manifest_t::TYPE_REDIRECT
:
2989 oloc
= object_locator_t(obc
->obs
.oi
.manifest
.redirect_target
);
2990 soid
= obc
->obs
.oi
.manifest
.redirect_target
;
2992 case object_manifest_t::TYPE_CHUNKED
:
2994 assert(0 == "unrecognized manifest type");
2998 soid
= m
->get_hobj();
2999 oloc
= object_locator_t(m
->get_object_locator());
3000 oloc
.pool
= pool
.info
.tier_of
;
3003 unsigned flags
= CEPH_OSD_FLAG_IGNORE_CACHE
| CEPH_OSD_FLAG_IGNORE_OVERLAY
;
3004 if (!(op
->may_write() || op
->may_cache())) {
3005 flags
|= CEPH_OSD_FLAG_RWORDERED
;
3007 dout(10) << __func__
<< " Start proxy write for " << *m
<< dendl
;
3009 ProxyWriteOpRef
pwop(std::make_shared
<ProxyWriteOp
>(op
, soid
, m
->ops
, m
->get_reqid()));
3010 pwop
->ctx
= new OpContext(op
, m
->get_reqid(), &pwop
->ops
, this);
3011 pwop
->mtime
= m
->get_mtime();
3013 ObjectOperation obj_op
;
3014 obj_op
.dup(pwop
->ops
);
3016 C_ProxyWrite_Commit
*fin
= new C_ProxyWrite_Commit(
3017 this, soid
, get_last_peering_reset(), pwop
);
3018 ceph_tid_t tid
= osd
->objecter
->mutate(
3019 soid
.oid
, oloc
, obj_op
, snapc
,
3020 ceph::real_clock::from_ceph_timespec(pwop
->mtime
),
3021 flags
, new C_OnFinisher(fin
, &osd
->objecter_finisher
),
3022 &pwop
->user_version
, pwop
->reqid
);
3024 pwop
->objecter_tid
= tid
;
3025 proxywrite_ops
[tid
] = pwop
;
3026 in_progress_proxy_ops
[soid
].push_back(op
);
3029 void PrimaryLogPG::finish_proxy_write(hobject_t oid
, ceph_tid_t tid
, int r
)
3031 dout(10) << __func__
<< " " << oid
<< " tid " << tid
3032 << " " << cpp_strerror(r
) << dendl
;
3034 map
<ceph_tid_t
, ProxyWriteOpRef
>::iterator p
= proxywrite_ops
.find(tid
);
3035 if (p
== proxywrite_ops
.end()) {
3036 dout(10) << __func__
<< " no proxywrite_op found" << dendl
;
3039 ProxyWriteOpRef pwop
= p
->second
;
3040 assert(tid
== pwop
->objecter_tid
);
3041 assert(oid
== pwop
->soid
);
3043 proxywrite_ops
.erase(tid
);
3045 map
<hobject_t
, list
<OpRequestRef
> >::iterator q
= in_progress_proxy_ops
.find(oid
);
3046 if (q
== in_progress_proxy_ops
.end()) {
3047 dout(10) << __func__
<< " no in_progress_proxy_ops found" << dendl
;
3052 list
<OpRequestRef
>& in_progress_op
= q
->second
;
3053 assert(in_progress_op
.size());
3054 list
<OpRequestRef
>::iterator it
= std::find(in_progress_op
.begin(),
3055 in_progress_op
.end(),
3057 assert(it
!= in_progress_op
.end());
3058 in_progress_op
.erase(it
);
3059 if (in_progress_op
.size() == 0) {
3060 in_progress_proxy_ops
.erase(oid
);
3063 osd
->logger
->inc(l_osd_tier_proxy_write
);
3065 const MOSDOp
*m
= static_cast<const MOSDOp
*>(pwop
->op
->get_req());
3068 if (!pwop
->sent_reply
) {
3070 MOSDOpReply
*reply
= pwop
->ctx
->reply
;
3072 pwop
->ctx
->reply
= NULL
;
3074 reply
= new MOSDOpReply(m
, r
, get_osdmap()->get_epoch(), 0, true);
3075 reply
->set_reply_versions(eversion_t(), pwop
->user_version
);
3077 reply
->add_flags(CEPH_OSD_FLAG_ACK
| CEPH_OSD_FLAG_ONDISK
);
3078 dout(10) << " sending commit on " << pwop
<< " " << reply
<< dendl
;
3079 osd
->send_message_osd_client(reply
, m
->get_connection());
3080 pwop
->sent_reply
= true;
3081 pwop
->ctx
->op
->mark_commit_sent();
3088 void PrimaryLogPG::cancel_proxy_write(ProxyWriteOpRef pwop
,
3089 vector
<ceph_tid_t
> *tids
)
3091 dout(10) << __func__
<< " " << pwop
->soid
<< dendl
;
3092 pwop
->canceled
= true;
3094 // cancel objecter op, if we can
3095 if (pwop
->objecter_tid
) {
3096 tids
->push_back(pwop
->objecter_tid
);
3099 proxywrite_ops
.erase(pwop
->objecter_tid
);
3100 pwop
->objecter_tid
= 0;
3104 class PromoteCallback
: public PrimaryLogPG::CopyCallback
{
3105 ObjectContextRef obc
;
3109 PromoteCallback(ObjectContextRef obc_
, PrimaryLogPG
*pg_
)
3112 start(ceph_clock_now()) {}
3114 void finish(PrimaryLogPG::CopyCallbackResults results
) override
{
3115 PrimaryLogPG::CopyResults
*results_data
= results
.get
<1>();
3116 int r
= results
.get
<0>();
3117 pg
->finish_promote(r
, results_data
, obc
);
3118 pg
->osd
->logger
->tinc(l_osd_tier_promote_lat
, ceph_clock_now() - start
);
3122 void PrimaryLogPG::promote_object(ObjectContextRef obc
,
3123 const hobject_t
& missing_oid
,
3124 const object_locator_t
& oloc
,
3126 ObjectContextRef
*promote_obc
)
3128 hobject_t hoid
= obc
? obc
->obs
.oi
.soid
: missing_oid
;
3129 assert(hoid
!= hobject_t());
3130 if (write_blocked_by_scrub(hoid
)) {
3131 dout(10) << __func__
<< " " << hoid
3132 << " blocked by scrub" << dendl
;
3134 waiting_for_scrub
.push_back(op
);
3135 op
->mark_delayed("waiting for scrub");
3136 dout(10) << __func__
<< " " << hoid
3137 << " placing op in waiting_for_scrub" << dendl
;
3139 dout(10) << __func__
<< " " << hoid
3140 << " no op, dropping on the floor" << dendl
;
3144 if (!obc
) { // we need to create an ObjectContext
3145 assert(missing_oid
!= hobject_t());
3146 obc
= get_object_context(missing_oid
, true);
3152 * Before promote complete, if there are proxy-reads for the object,
3153 * for this case we don't use DONTNEED.
3155 unsigned src_fadvise_flags
= LIBRADOS_OP_FLAG_FADVISE_SEQUENTIAL
;
3156 map
<hobject_t
, list
<OpRequestRef
>>::iterator q
= in_progress_proxy_ops
.find(obc
->obs
.oi
.soid
);
3157 if (q
== in_progress_proxy_ops
.end()) {
3158 src_fadvise_flags
|= LIBRADOS_OP_FLAG_FADVISE_DONTNEED
;
3161 PromoteCallback
*cb
= new PromoteCallback(obc
, this);
3162 object_locator_t my_oloc
= oloc
;
3163 my_oloc
.pool
= pool
.info
.tier_of
;
3165 unsigned flags
= CEPH_OSD_COPY_FROM_FLAG_IGNORE_OVERLAY
|
3166 CEPH_OSD_COPY_FROM_FLAG_IGNORE_CACHE
|
3167 CEPH_OSD_COPY_FROM_FLAG_MAP_SNAP_CLONE
|
3168 CEPH_OSD_COPY_FROM_FLAG_RWORDERED
;
3169 start_copy(cb
, obc
, obc
->obs
.oi
.soid
, my_oloc
, 0, flags
,
3170 obc
->obs
.oi
.soid
.snap
== CEPH_NOSNAP
,
3171 src_fadvise_flags
, 0);
3173 assert(obc
->is_blocked());
3176 wait_for_blocked_object(obc
->obs
.oi
.soid
, op
);
3177 info
.stats
.stats
.sum
.num_promote
++;
3180 void PrimaryLogPG::execute_ctx(OpContext
*ctx
)
3183 dout(10) << __func__
<< " " << ctx
<< dendl
;
3184 ctx
->reset_obs(ctx
->obc
);
3185 ctx
->update_log_only
= false; // reset in case finish_copyfrom() is re-running execute_ctx
3186 OpRequestRef op
= ctx
->op
;
3187 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
3188 ObjectContextRef obc
= ctx
->obc
;
3189 const hobject_t
& soid
= obc
->obs
.oi
.soid
;
3191 // this method must be idempotent since we may call it several times
3192 // before we finally apply the resulting transaction.
3193 ctx
->op_t
.reset(new PGTransaction
);
3195 if (op
->may_write() || op
->may_cache()) {
3197 if (!(m
->has_flag(CEPH_OSD_FLAG_ENFORCE_SNAPC
)) &&
3198 pool
.info
.is_pool_snaps_mode()) {
3200 ctx
->snapc
= pool
.snapc
;
3202 // client specified snapc
3203 ctx
->snapc
.seq
= m
->get_snap_seq();
3204 ctx
->snapc
.snaps
= m
->get_snaps();
3205 filter_snapc(ctx
->snapc
.snaps
);
3207 if ((m
->has_flag(CEPH_OSD_FLAG_ORDERSNAP
)) &&
3208 ctx
->snapc
.seq
< obc
->ssc
->snapset
.seq
) {
3209 dout(10) << " ORDERSNAP flag set and snapc seq " << ctx
->snapc
.seq
3210 << " < snapset seq " << obc
->ssc
->snapset
.seq
3211 << " on " << obc
->obs
.oi
.soid
<< dendl
;
3212 reply_ctx(ctx
, -EOLDSNAPC
);
3217 ctx
->at_version
= get_next_version();
3218 ctx
->mtime
= m
->get_mtime();
3220 dout(10) << __func__
<< " " << soid
<< " " << *ctx
->ops
3221 << " ov " << obc
->obs
.oi
.version
<< " av " << ctx
->at_version
3222 << " snapc " << ctx
->snapc
3223 << " snapset " << obc
->ssc
->snapset
3226 dout(10) << __func__
<< " " << soid
<< " " << *ctx
->ops
3227 << " ov " << obc
->obs
.oi
.version
3231 if (!ctx
->user_at_version
)
3232 ctx
->user_at_version
= obc
->obs
.oi
.user_version
;
3233 dout(30) << __func__
<< " user_at_version " << ctx
->user_at_version
<< dendl
;
3235 if (op
->may_read()) {
3236 dout(10) << " taking ondisk_read_lock" << dendl
;
3237 obc
->ondisk_read_lock();
3242 osd_reqid_t reqid
= ctx
->op
->get_reqid();
3244 tracepoint(osd
, prepare_tx_enter
, reqid
.name
._type
,
3245 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
3248 int result
= prepare_transaction(ctx
);
3252 osd_reqid_t reqid
= ctx
->op
->get_reqid();
3254 tracepoint(osd
, prepare_tx_exit
, reqid
.name
._type
,
3255 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
3258 if (op
->may_read()) {
3259 dout(10) << " dropping ondisk_read_lock" << dendl
;
3260 obc
->ondisk_read_unlock();
3263 bool pending_async_reads
= !ctx
->pending_async_reads
.empty();
3264 if (result
== -EINPROGRESS
|| pending_async_reads
) {
3266 if (pending_async_reads
) {
3267 in_progress_async_reads
.push_back(make_pair(op
, ctx
));
3268 ctx
->start_async_reads(this);
3273 if (result
== -EAGAIN
) {
3274 // clean up after the ctx
3279 bool successful_write
= !ctx
->op_t
->empty() && op
->may_write() && result
>= 0;
3280 // prepare the reply
3281 ctx
->reply
= new MOSDOpReply(m
, 0, get_osdmap()->get_epoch(), 0,
3284 // Write operations aren't allowed to return a data payload because
3285 // we can't do so reliably. If the client has to resend the request
3286 // and it has already been applied, we will return 0 with no
3287 // payload. Non-deterministic behavior is no good. However, it is
3288 // possible to construct an operation that does a read, does a guard
3289 // check (e.g., CMPXATTR), and then a write. Then we either succeed
3290 // with the write, or return a CMPXATTR and the read value.
3291 if (successful_write
) {
3292 // write. normalize the result code.
3293 dout(20) << " zeroing write result code " << result
<< dendl
;
3296 ctx
->reply
->set_result(result
);
3299 if ((ctx
->op_t
->empty() || result
< 0) && !ctx
->update_log_only
) {
3300 // finish side-effects
3302 do_osd_op_effects(ctx
, m
->get_connection());
3304 complete_read_ctx(result
, ctx
);
3308 ctx
->reply
->set_reply_versions(ctx
->at_version
, ctx
->user_at_version
);
3310 assert(op
->may_write() || op
->may_cache());
3315 // verify that we are doing this in order?
3316 if (cct
->_conf
->osd_debug_op_order
&& m
->get_source().is_client() &&
3317 !pool
.info
.is_tier() && !pool
.info
.has_tiers()) {
3318 map
<client_t
,ceph_tid_t
>& cm
= debug_op_order
[obc
->obs
.oi
.soid
];
3319 ceph_tid_t t
= m
->get_tid();
3320 client_t n
= m
->get_source().num();
3321 map
<client_t
,ceph_tid_t
>::iterator p
= cm
.find(n
);
3322 if (p
== cm
.end()) {
3323 dout(20) << " op order client." << n
<< " tid " << t
<< " (first)" << dendl
;
3326 dout(20) << " op order client." << n
<< " tid " << t
<< " last was " << p
->second
<< dendl
;
3327 if (p
->second
> t
) {
3328 derr
<< "bad op order, already applied " << p
->second
<< " > this " << t
<< dendl
;
3329 assert(0 == "out of order op");
3335 if (ctx
->update_log_only
) {
3337 do_osd_op_effects(ctx
, m
->get_connection());
3339 dout(20) << __func__
<< " update_log_only -- result=" << result
<< dendl
;
3340 // save just what we need from ctx
3341 MOSDOpReply
*reply
= ctx
->reply
;
3342 ctx
->reply
= nullptr;
3343 reply
->claim_op_out_data(*ctx
->ops
);
3344 reply
->get_header().data_off
= (ctx
->data_off
? *ctx
->data_off
: 0);
3347 if (result
== -ENOENT
) {
3348 reply
->set_enoent_reply_versions(info
.last_update
,
3349 info
.last_user_version
);
3351 reply
->add_flags(CEPH_OSD_FLAG_ACK
| CEPH_OSD_FLAG_ONDISK
);
3352 // append to pg log for dup detection - don't save buffers for now
3353 record_write_error(op
, soid
, reply
, result
);
3357 // no need to capture PG ref, repop cancel will handle that
3358 // Can capture the ctx by pointer, it's owned by the repop
3359 ctx
->register_on_commit(
3365 if (m
&& !ctx
->sent_reply
) {
3366 MOSDOpReply
*reply
= ctx
->reply
;
3368 ctx
->reply
= nullptr;
3370 reply
= new MOSDOpReply(m
, 0, get_osdmap()->get_epoch(), 0, true);
3371 reply
->set_reply_versions(ctx
->at_version
,
3372 ctx
->user_at_version
);
3374 reply
->add_flags(CEPH_OSD_FLAG_ACK
| CEPH_OSD_FLAG_ONDISK
);
3375 dout(10) << " sending reply on " << *m
<< " " << reply
<< dendl
;
3376 osd
->send_message_osd_client(reply
, m
->get_connection());
3377 ctx
->sent_reply
= true;
3378 ctx
->op
->mark_commit_sent();
3381 ctx
->register_on_success(
3385 ctx
->op
? ctx
->op
->get_req()->get_connection() :
3388 ctx
->register_on_finish(
3393 // issue replica writes
3394 ceph_tid_t rep_tid
= osd
->get_tid();
3396 RepGather
*repop
= new_repop(ctx
, obc
, rep_tid
);
3398 issue_repop(repop
, ctx
);
3403 void PrimaryLogPG::close_op_ctx(OpContext
*ctx
) {
3404 release_object_locks(ctx
->lock_manager
);
3408 for (auto p
= ctx
->on_finish
.begin(); p
!= ctx
->on_finish
.end();
3409 ctx
->on_finish
.erase(p
++)) {
3415 void PrimaryLogPG::reply_ctx(OpContext
*ctx
, int r
)
3418 osd
->reply_op_error(ctx
->op
, r
);
3422 void PrimaryLogPG::reply_ctx(OpContext
*ctx
, int r
, eversion_t v
, version_t uv
)
3425 osd
->reply_op_error(ctx
->op
, r
, v
, uv
);
3429 void PrimaryLogPG::log_op_stats(OpContext
*ctx
)
3431 OpRequestRef op
= ctx
->op
;
3432 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
3434 utime_t now
= ceph_clock_now();
3435 utime_t latency
= now
;
3436 latency
-= ctx
->op
->get_req()->get_recv_stamp();
3437 utime_t process_latency
= now
;
3438 process_latency
-= ctx
->op
->get_dequeued_time();
3440 uint64_t inb
= ctx
->bytes_written
;
3441 uint64_t outb
= ctx
->bytes_read
;
3443 osd
->logger
->inc(l_osd_op
);
3445 osd
->logger
->inc(l_osd_op_outb
, outb
);
3446 osd
->logger
->inc(l_osd_op_inb
, inb
);
3447 osd
->logger
->tinc(l_osd_op_lat
, latency
);
3448 osd
->logger
->tinc(l_osd_op_process_lat
, process_latency
);
3450 if (op
->may_read() && op
->may_write()) {
3451 osd
->logger
->inc(l_osd_op_rw
);
3452 osd
->logger
->inc(l_osd_op_rw_inb
, inb
);
3453 osd
->logger
->inc(l_osd_op_rw_outb
, outb
);
3454 osd
->logger
->tinc(l_osd_op_rw_lat
, latency
);
3455 osd
->logger
->hinc(l_osd_op_rw_lat_inb_hist
, latency
.to_nsec(), inb
);
3456 osd
->logger
->hinc(l_osd_op_rw_lat_outb_hist
, latency
.to_nsec(), outb
);
3457 osd
->logger
->tinc(l_osd_op_rw_process_lat
, process_latency
);
3458 } else if (op
->may_read()) {
3459 osd
->logger
->inc(l_osd_op_r
);
3460 osd
->logger
->inc(l_osd_op_r_outb
, outb
);
3461 osd
->logger
->tinc(l_osd_op_r_lat
, latency
);
3462 osd
->logger
->hinc(l_osd_op_r_lat_outb_hist
, latency
.to_nsec(), outb
);
3463 osd
->logger
->tinc(l_osd_op_r_process_lat
, process_latency
);
3464 } else if (op
->may_write() || op
->may_cache()) {
3465 osd
->logger
->inc(l_osd_op_w
);
3466 osd
->logger
->inc(l_osd_op_w_inb
, inb
);
3467 osd
->logger
->tinc(l_osd_op_w_lat
, latency
);
3468 osd
->logger
->hinc(l_osd_op_w_lat_inb_hist
, latency
.to_nsec(), inb
);
3469 osd
->logger
->tinc(l_osd_op_w_process_lat
, process_latency
);
3473 dout(15) << "log_op_stats " << *m
3476 << " lat " << latency
<< dendl
;
3479 void PrimaryLogPG::do_sub_op(OpRequestRef op
)
3481 const MOSDSubOp
*m
= static_cast<const MOSDSubOp
*>(op
->get_req());
3482 assert(have_same_or_newer_map(m
->map_epoch
));
3483 assert(m
->get_type() == MSG_OSD_SUBOP
);
3484 dout(15) << "do_sub_op " << *op
->get_req() << dendl
;
3487 waiting_for_peered
.push_back(op
);
3488 op
->mark_delayed("waiting for active");
3492 const OSDOp
*first
= NULL
;
3493 if (m
->ops
.size() >= 1) {
3498 switch (first
->op
.op
) {
3499 case CEPH_OSD_OP_DELETE
:
3502 case CEPH_OSD_OP_SCRUB_RESERVE
:
3503 handle_scrub_reserve_request(op
);
3505 case CEPH_OSD_OP_SCRUB_UNRESERVE
:
3506 handle_scrub_reserve_release(op
);
3508 case CEPH_OSD_OP_SCRUB_MAP
:
3509 sub_op_scrub_map(op
);
3515 void PrimaryLogPG::do_sub_op_reply(OpRequestRef op
)
3517 const MOSDSubOpReply
*r
= static_cast<const MOSDSubOpReply
*>(op
->get_req());
3518 assert(r
->get_type() == MSG_OSD_SUBOPREPLY
);
3519 if (r
->ops
.size() >= 1) {
3520 const OSDOp
& first
= r
->ops
[0];
3521 switch (first
.op
.op
) {
3522 case CEPH_OSD_OP_SCRUB_RESERVE
:
3524 pg_shard_t from
= r
->from
;
3525 bufferlist::iterator p
= const_cast<bufferlist
&>(r
->get_data()).begin();
3527 ::decode(reserved
, p
);
3529 handle_scrub_reserve_grant(op
, from
);
3531 handle_scrub_reserve_reject(op
, from
);
3539 void PrimaryLogPG::do_scan(
3541 ThreadPool::TPHandle
&handle
)
3543 const MOSDPGScan
*m
= static_cast<const MOSDPGScan
*>(op
->get_req());
3544 assert(m
->get_type() == MSG_OSD_PG_SCAN
);
3545 dout(10) << "do_scan " << *m
<< dendl
;
3550 case MOSDPGScan::OP_SCAN_GET_DIGEST
:
3553 if (osd
->check_backfill_full(ss
)) {
3554 dout(1) << __func__
<< ": Canceling backfill, " << ss
.str() << dendl
;
3555 queue_peering_event(
3557 std::make_shared
<CephPeeringEvt
>(
3558 get_osdmap()->get_epoch(),
3559 get_osdmap()->get_epoch(),
3560 BackfillTooFull())));
3564 BackfillInterval bi
;
3565 bi
.begin
= m
->begin
;
3566 // No need to flush, there won't be any in progress writes occuring
3569 cct
->_conf
->osd_backfill_scan_min
,
3570 cct
->_conf
->osd_backfill_scan_max
,
3573 MOSDPGScan
*reply
= new MOSDPGScan(
3574 MOSDPGScan::OP_SCAN_DIGEST
,
3576 get_osdmap()->get_epoch(), m
->query_epoch
,
3577 spg_t(info
.pgid
.pgid
, get_primary().shard
), bi
.begin
, bi
.end
);
3578 ::encode(bi
.objects
, reply
->get_data());
3579 osd
->send_message_osd_cluster(reply
, m
->get_connection());
3583 case MOSDPGScan::OP_SCAN_DIGEST
:
3585 pg_shard_t from
= m
->from
;
3587 // Check that from is in backfill_targets vector
3588 assert(is_backfill_targets(from
));
3590 BackfillInterval
& bi
= peer_backfill_info
[from
];
3591 bi
.begin
= m
->begin
;
3593 bufferlist::iterator p
= const_cast<bufferlist
&>(m
->get_data()).begin();
3595 // take care to preserve ordering!
3597 ::decode_noclear(bi
.objects
, p
);
3599 if (waiting_on_backfill
.erase(from
)) {
3600 if (waiting_on_backfill
.empty()) {
3601 assert(peer_backfill_info
.size() == backfill_targets
.size());
3602 finish_recovery_op(hobject_t::get_max());
3605 // we canceled backfill for a while due to a too full, and this
3606 // is an extra response from a non-too-full peer
3613 void PrimaryLogPG::do_backfill(OpRequestRef op
)
3615 const MOSDPGBackfill
*m
= static_cast<const MOSDPGBackfill
*>(op
->get_req());
3616 assert(m
->get_type() == MSG_OSD_PG_BACKFILL
);
3617 dout(10) << "do_backfill " << *m
<< dendl
;
3622 case MOSDPGBackfill::OP_BACKFILL_FINISH
:
3624 assert(cct
->_conf
->osd_kill_backfill_at
!= 1);
3626 MOSDPGBackfill
*reply
= new MOSDPGBackfill(
3627 MOSDPGBackfill::OP_BACKFILL_FINISH_ACK
,
3628 get_osdmap()->get_epoch(),
3630 spg_t(info
.pgid
.pgid
, get_primary().shard
));
3631 reply
->set_priority(get_recovery_op_priority());
3632 osd
->send_message_osd_cluster(reply
, m
->get_connection());
3633 queue_peering_event(
3635 std::make_shared
<CephPeeringEvt
>(
3636 get_osdmap()->get_epoch(),
3637 get_osdmap()->get_epoch(),
3642 case MOSDPGBackfill::OP_BACKFILL_PROGRESS
:
3644 assert(cct
->_conf
->osd_kill_backfill_at
!= 2);
3646 info
.set_last_backfill(m
->last_backfill
);
3647 info
.stats
= m
->stats
;
3649 ObjectStore::Transaction t
;
3652 int tr
= osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
3657 case MOSDPGBackfill::OP_BACKFILL_FINISH_ACK
:
3659 assert(is_primary());
3660 assert(cct
->_conf
->osd_kill_backfill_at
!= 3);
3661 finish_recovery_op(hobject_t::get_max());
3667 void PrimaryLogPG::do_backfill_remove(OpRequestRef op
)
3669 const MOSDPGBackfillRemove
*m
= static_cast<const MOSDPGBackfillRemove
*>(
3671 assert(m
->get_type() == MSG_OSD_PG_BACKFILL_REMOVE
);
3672 dout(7) << __func__
<< " " << m
->ls
<< dendl
;
3676 ObjectStore::Transaction t
;
3677 for (auto& p
: m
->ls
) {
3678 remove_snap_mapped_object(t
, p
.first
);
3680 int r
= osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
3684 int PrimaryLogPG::trim_object(
3685 bool first
, const hobject_t
&coid
, PrimaryLogPG::OpContextUPtr
*ctxp
)
3690 ObjectContextRef obc
= get_object_context(coid
, false, NULL
);
3691 if (!obc
|| !obc
->ssc
|| !obc
->ssc
->exists
) {
3692 osd
->clog
->error() << __func__
<< ": Can not trim " << coid
3693 << " repair needed " << (obc
? "(no obc->ssc or !exists)" : "(no obc)");
3698 coid
.oid
, coid
.get_key(),
3699 obc
->ssc
->snapset
.head_exists
? CEPH_NOSNAP
:CEPH_SNAPDIR
, coid
.get_hash(),
3700 info
.pgid
.pool(), coid
.get_namespace());
3701 ObjectContextRef snapset_obc
= get_object_context(snapoid
, false);
3703 osd
->clog
->error() << __func__
<< ": Can not trim " << coid
3704 << " repair needed, no snapset obc for " << snapoid
;
3708 SnapSet
& snapset
= obc
->ssc
->snapset
;
3710 bool legacy
= snapset
.is_legacy() ||
3711 get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
;
3713 object_info_t
&coi
= obc
->obs
.oi
;
3714 set
<snapid_t
> old_snaps
;
3716 old_snaps
.insert(coi
.legacy_snaps
.begin(), coi
.legacy_snaps
.end());
3718 auto p
= snapset
.clone_snaps
.find(coid
.snap
);
3719 if (p
== snapset
.clone_snaps
.end()) {
3720 osd
->clog
->error() << "No clone_snaps in snapset " << snapset
3721 << " for object " << coid
<< "\n";
3724 old_snaps
.insert(snapset
.clone_snaps
[coid
.snap
].begin(),
3725 snapset
.clone_snaps
[coid
.snap
].end());
3727 if (old_snaps
.empty()) {
3728 osd
->clog
->error() << "No object info snaps for object " << coid
;
3732 dout(10) << coid
<< " old_snaps " << old_snaps
3733 << " old snapset " << snapset
<< dendl
;
3734 if (snapset
.seq
== 0) {
3735 osd
->clog
->error() << "No snapset.seq for object " << coid
;
3739 set
<snapid_t
> new_snaps
;
3740 for (set
<snapid_t
>::iterator i
= old_snaps
.begin();
3741 i
!= old_snaps
.end();
3743 if (!pool
.info
.is_removed_snap(*i
))
3744 new_snaps
.insert(*i
);
3747 vector
<snapid_t
>::iterator p
= snapset
.clones
.end();
3749 if (new_snaps
.empty()) {
3750 p
= std::find(snapset
.clones
.begin(), snapset
.clones
.end(), coid
.snap
);
3751 if (p
== snapset
.clones
.end()) {
3752 osd
->clog
->error() << "Snap " << coid
.snap
<< " not in clones";
3757 OpContextUPtr ctx
= simple_opc_create(obc
);
3758 ctx
->snapset_obc
= snapset_obc
;
3760 if (!ctx
->lock_manager
.get_snaptrimmer_write(
3764 close_op_ctx(ctx
.release());
3765 dout(10) << __func__
<< ": Unable to get a wlock on " << coid
<< dendl
;
3769 if (!ctx
->lock_manager
.get_snaptrimmer_write(
3773 close_op_ctx(ctx
.release());
3774 dout(10) << __func__
<< ": Unable to get a wlock on " << snapoid
<< dendl
;
3778 ctx
->at_version
= get_next_version();
3780 PGTransaction
*t
= ctx
->op_t
.get();
3782 if (new_snaps
.empty()) {
3784 dout(10) << coid
<< " snaps " << old_snaps
<< " -> "
3785 << new_snaps
<< " ... deleting" << dendl
;
3788 assert(p
!= snapset
.clones
.end());
3790 snapid_t last
= coid
.snap
;
3791 ctx
->delta_stats
.num_bytes
-= snapset
.get_clone_bytes(last
);
3793 if (p
!= snapset
.clones
.begin()) {
3794 // not the oldest... merge overlap into next older clone
3795 vector
<snapid_t
>::iterator n
= p
- 1;
3796 hobject_t prev_coid
= coid
;
3797 prev_coid
.snap
= *n
;
3798 bool adjust_prev_bytes
= is_present_clone(prev_coid
);
3800 if (adjust_prev_bytes
)
3801 ctx
->delta_stats
.num_bytes
-= snapset
.get_clone_bytes(*n
);
3803 snapset
.clone_overlap
[*n
].intersection_of(
3804 snapset
.clone_overlap
[*p
]);
3806 if (adjust_prev_bytes
)
3807 ctx
->delta_stats
.num_bytes
+= snapset
.get_clone_bytes(*n
);
3809 ctx
->delta_stats
.num_objects
--;
3811 ctx
->delta_stats
.num_objects_dirty
--;
3813 ctx
->delta_stats
.num_objects_omap
--;
3814 if (coi
.is_whiteout()) {
3815 dout(20) << __func__
<< " trimming whiteout on " << coid
<< dendl
;
3816 ctx
->delta_stats
.num_whiteouts
--;
3818 ctx
->delta_stats
.num_object_clones
--;
3819 if (coi
.is_cache_pinned())
3820 ctx
->delta_stats
.num_objects_pinned
--;
3821 obc
->obs
.exists
= false;
3823 snapset
.clones
.erase(p
);
3824 snapset
.clone_overlap
.erase(last
);
3825 snapset
.clone_size
.erase(last
);
3826 snapset
.clone_snaps
.erase(last
);
3830 pg_log_entry_t::DELETE
,
3833 ctx
->obs
->oi
.version
,
3845 coi
= object_info_t(coid
);
3847 ctx
->at_version
.version
++;
3849 // save adjusted snaps for this object
3850 dout(10) << coid
<< " snaps " << old_snaps
<< " -> " << new_snaps
<< dendl
;
3852 coi
.legacy_snaps
= vector
<snapid_t
>(new_snaps
.rbegin(), new_snaps
.rend());
3854 snapset
.clone_snaps
[coid
.snap
] = vector
<snapid_t
>(new_snaps
.rbegin(),
3856 // we still do a 'modify' event on this object just to trigger a
3857 // snapmapper.update ... :(
3860 coi
.prior_version
= coi
.version
;
3861 coi
.version
= ctx
->at_version
;
3863 ::encode(coi
, bl
, get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
3864 t
->setattr(coid
, OI_ATTR
, bl
);
3868 pg_log_entry_t::MODIFY
,
3877 ctx
->at_version
.version
++;
3885 // save head snapset
3886 dout(10) << coid
<< " new snapset " << snapset
<< " on "
3887 << snapset_obc
->obs
.oi
<< dendl
;
3888 if (snapset
.clones
.empty() &&
3889 (!snapset
.head_exists
||
3890 (snapset_obc
->obs
.oi
.is_whiteout() &&
3891 !(snapset_obc
->obs
.oi
.is_dirty() && pool
.info
.is_tier()) &&
3892 !snapset_obc
->obs
.oi
.is_cache_pinned()))) {
3893 // NOTE: this arguably constitutes minor interference with the
3894 // tiering agent if this is a cache tier since a snap trim event
3895 // is effectively evicting a whiteout we might otherwise want to
3897 dout(10) << coid
<< " removing " << snapoid
<< dendl
;
3900 pg_log_entry_t::DELETE
,
3903 ctx
->snapset_obc
->obs
.oi
.version
,
3909 if (snapoid
.is_head()) {
3910 derr
<< "removing snap head" << dendl
;
3911 object_info_t
& oi
= ctx
->snapset_obc
->obs
.oi
;
3912 ctx
->delta_stats
.num_objects
--;
3913 if (oi
.is_dirty()) {
3914 ctx
->delta_stats
.num_objects_dirty
--;
3917 ctx
->delta_stats
.num_objects_omap
--;
3918 if (oi
.is_whiteout()) {
3919 dout(20) << __func__
<< " trimming whiteout on " << oi
.soid
<< dendl
;
3920 ctx
->delta_stats
.num_whiteouts
--;
3922 if (oi
.is_cache_pinned()) {
3923 ctx
->delta_stats
.num_objects_pinned
--;
3926 ctx
->snapset_obc
->obs
.exists
= false;
3927 ctx
->snapset_obc
->obs
.oi
= object_info_t(snapoid
);
3930 dout(10) << coid
<< " filtering snapset on " << snapoid
<< dendl
;
3931 snapset
.filter(pool
.info
);
3932 dout(10) << coid
<< " writing updated snapset on " << snapoid
3933 << ", snapset is " << snapset
<< dendl
;
3936 pg_log_entry_t::MODIFY
,
3939 ctx
->snapset_obc
->obs
.oi
.version
,
3946 ctx
->snapset_obc
->obs
.oi
.prior_version
=
3947 ctx
->snapset_obc
->obs
.oi
.version
;
3948 ctx
->snapset_obc
->obs
.oi
.version
= ctx
->at_version
;
3950 map
<string
, bufferlist
> attrs
;
3952 ::encode(snapset
, bl
);
3953 attrs
[SS_ATTR
].claim(bl
);
3956 ::encode(ctx
->snapset_obc
->obs
.oi
, bl
,
3957 get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
3958 attrs
[OI_ATTR
].claim(bl
);
3959 t
->setattrs(snapoid
, attrs
);
3962 *ctxp
= std::move(ctx
);
3966 void PrimaryLogPG::kick_snap_trim()
3968 assert(is_active());
3969 assert(is_primary());
3970 if (is_clean() && !snap_trimq
.empty()) {
3971 dout(10) << __func__
<< ": clean and snaps to trim, kicking" << dendl
;
3972 snap_trimmer_machine
.process_event(KickTrim());
3976 void PrimaryLogPG::snap_trimmer_scrub_complete()
3978 if (is_primary() && is_active() && is_clean()) {
3979 assert(!snap_trimq
.empty());
3980 snap_trimmer_machine
.process_event(ScrubComplete());
3984 void PrimaryLogPG::snap_trimmer(epoch_t queued
)
3986 if (deleting
|| pg_has_reset_since(queued
)) {
3990 assert(is_primary());
3992 dout(10) << "snap_trimmer posting" << dendl
;
3993 snap_trimmer_machine
.process_event(DoSnapWork());
3994 dout(10) << "snap_trimmer complete" << dendl
;
3998 int PrimaryLogPG::do_xattr_cmp_u64(int op
, __u64 v1
, bufferlist
& xattr
)
4002 string
v2s(xattr
.c_str(), xattr
.length());
4004 v2
= strtoull(v2s
.c_str(), NULL
, 10);
4008 dout(20) << "do_xattr_cmp_u64 '" << v1
<< "' vs '" << v2
<< "' op " << op
<< dendl
;
4011 case CEPH_OSD_CMPXATTR_OP_EQ
:
4013 case CEPH_OSD_CMPXATTR_OP_NE
:
4015 case CEPH_OSD_CMPXATTR_OP_GT
:
4017 case CEPH_OSD_CMPXATTR_OP_GTE
:
4019 case CEPH_OSD_CMPXATTR_OP_LT
:
4021 case CEPH_OSD_CMPXATTR_OP_LTE
:
4028 int PrimaryLogPG::do_xattr_cmp_str(int op
, string
& v1s
, bufferlist
& xattr
)
4030 string
v2s(xattr
.c_str(), xattr
.length());
4032 dout(20) << "do_xattr_cmp_str '" << v1s
<< "' vs '" << v2s
<< "' op " << op
<< dendl
;
4035 case CEPH_OSD_CMPXATTR_OP_EQ
:
4036 return (v1s
.compare(v2s
) == 0);
4037 case CEPH_OSD_CMPXATTR_OP_NE
:
4038 return (v1s
.compare(v2s
) != 0);
4039 case CEPH_OSD_CMPXATTR_OP_GT
:
4040 return (v1s
.compare(v2s
) > 0);
4041 case CEPH_OSD_CMPXATTR_OP_GTE
:
4042 return (v1s
.compare(v2s
) >= 0);
4043 case CEPH_OSD_CMPXATTR_OP_LT
:
4044 return (v1s
.compare(v2s
) < 0);
4045 case CEPH_OSD_CMPXATTR_OP_LTE
:
4046 return (v1s
.compare(v2s
) <= 0);
4052 int PrimaryLogPG::do_writesame(OpContext
*ctx
, OSDOp
& osd_op
)
4054 ceph_osd_op
& op
= osd_op
.op
;
4055 vector
<OSDOp
> write_ops(1);
4056 OSDOp
& write_op
= write_ops
[0];
4057 uint64_t write_length
= op
.writesame
.length
;
4063 if (!op
.writesame
.data_length
|| write_length
% op
.writesame
.data_length
)
4066 if (op
.writesame
.data_length
!= osd_op
.indata
.length()) {
4067 derr
<< "invalid length ws data length " << op
.writesame
.data_length
<< " actual len " << osd_op
.indata
.length() << dendl
;
4071 while (write_length
) {
4072 write_op
.indata
.append(osd_op
.indata
);
4073 write_length
-= op
.writesame
.data_length
;
4076 write_op
.op
.op
= CEPH_OSD_OP_WRITE
;
4077 write_op
.op
.extent
.offset
= op
.writesame
.offset
;
4078 write_op
.op
.extent
.length
= op
.writesame
.length
;
4079 result
= do_osd_ops(ctx
, write_ops
);
4081 derr
<< "do_writesame do_osd_ops failed " << result
<< dendl
;
4086 // ========================================================================
4087 // low level osd ops
4089 int PrimaryLogPG::do_tmap2omap(OpContext
*ctx
, unsigned flags
)
4091 dout(20) << " convert tmap to omap for " << ctx
->new_obs
.oi
.soid
<< dendl
;
4092 bufferlist header
, vals
;
4093 int r
= _get_tmap(ctx
, &header
, &vals
);
4095 if (r
== -ENODATA
&& (flags
& CEPH_OSD_TMAP2OMAP_NULLOK
))
4100 vector
<OSDOp
> ops(3);
4102 ops
[0].op
.op
= CEPH_OSD_OP_TRUNCATE
;
4103 ops
[0].op
.extent
.offset
= 0;
4104 ops
[0].op
.extent
.length
= 0;
4106 ops
[1].op
.op
= CEPH_OSD_OP_OMAPSETHEADER
;
4107 ops
[1].indata
.claim(header
);
4109 ops
[2].op
.op
= CEPH_OSD_OP_OMAPSETVALS
;
4110 ops
[2].indata
.claim(vals
);
4112 return do_osd_ops(ctx
, ops
);
4115 int PrimaryLogPG::do_tmapup_slow(OpContext
*ctx
, bufferlist::iterator
& bp
, OSDOp
& osd_op
,
4120 map
<string
, bufferlist
> m
;
4122 bufferlist::iterator p
= bl
.begin();
4123 ::decode(header
, p
);
4135 case CEPH_OSD_TMAP_SET
: // insert key
4143 case CEPH_OSD_TMAP_RM
: // remove key
4145 if (!m
.count(key
)) {
4150 case CEPH_OSD_TMAP_RMSLOPPY
: // remove key
4154 case CEPH_OSD_TMAP_HDR
: // update header
4156 ::decode(header
, bp
);
4166 ::encode(header
, obl
);
4170 vector
<OSDOp
> nops(1);
4171 OSDOp
& newop
= nops
[0];
4172 newop
.op
.op
= CEPH_OSD_OP_WRITEFULL
;
4173 newop
.op
.extent
.offset
= 0;
4174 newop
.op
.extent
.length
= obl
.length();
4176 do_osd_ops(ctx
, nops
);
4177 osd_op
.outdata
.claim(newop
.outdata
);
4181 int PrimaryLogPG::do_tmapup(OpContext
*ctx
, bufferlist::iterator
& bp
, OSDOp
& osd_op
)
4183 bufferlist::iterator orig_bp
= bp
;
4186 dout(10) << "tmapup is a no-op" << dendl
;
4188 // read the whole object
4189 vector
<OSDOp
> nops(1);
4190 OSDOp
& newop
= nops
[0];
4191 newop
.op
.op
= CEPH_OSD_OP_READ
;
4192 newop
.op
.extent
.offset
= 0;
4193 newop
.op
.extent
.length
= 0;
4194 result
= do_osd_ops(ctx
, nops
);
4196 dout(10) << "tmapup read " << newop
.outdata
.length() << dendl
;
4198 dout(30) << " starting is \n";
4199 newop
.outdata
.hexdump(*_dout
);
4202 bufferlist::iterator ip
= newop
.outdata
.begin();
4205 dout(30) << "the update command is: \n";
4206 osd_op
.indata
.hexdump(*_dout
);
4212 if (newop
.outdata
.length()) {
4213 ::decode(header
, ip
);
4214 ::decode(nkeys
, ip
);
4216 dout(10) << "tmapup header " << header
.length() << dendl
;
4218 if (!bp
.end() && *bp
== CEPH_OSD_TMAP_HDR
) {
4220 ::decode(header
, bp
);
4221 dout(10) << "tmapup new header " << header
.length() << dendl
;
4224 ::encode(header
, obl
);
4226 dout(20) << "tmapup initial nkeys " << nkeys
<< dendl
;
4229 bufferlist newkeydata
;
4230 string nextkey
, last_in_key
;
4232 bool have_next
= false;
4235 ::decode(nextkey
, ip
);
4236 ::decode(nextval
, ip
);
4238 while (!bp
.end() && !result
) {
4245 catch (buffer::error
& e
) {
4248 if (key
< last_in_key
) {
4249 dout(5) << "tmapup warning: key '" << key
<< "' < previous key '" << last_in_key
4250 << "', falling back to an inefficient (unsorted) update" << dendl
;
4252 return do_tmapup_slow(ctx
, bp
, osd_op
, newop
.outdata
);
4256 dout(10) << "tmapup op " << (int)op
<< " key " << key
<< dendl
;
4258 // skip existing intervening keys
4259 bool key_exists
= false;
4260 while (have_next
&& !key_exists
) {
4261 dout(20) << " (have_next=" << have_next
<< " nextkey=" << nextkey
<< ")" << dendl
;
4264 if (nextkey
< key
) {
4266 ::encode(nextkey
, newkeydata
);
4267 ::encode(nextval
, newkeydata
);
4268 dout(20) << " keep " << nextkey
<< " " << nextval
.length() << dendl
;
4270 // don't copy; discard old value. and stop.
4271 dout(20) << " drop " << nextkey
<< " " << nextval
.length() << dendl
;
4276 ::decode(nextkey
, ip
);
4277 ::decode(nextval
, ip
);
4283 if (op
== CEPH_OSD_TMAP_SET
) {
4288 catch (buffer::error
& e
) {
4291 ::encode(key
, newkeydata
);
4292 ::encode(val
, newkeydata
);
4293 dout(20) << " set " << key
<< " " << val
.length() << dendl
;
4295 } else if (op
== CEPH_OSD_TMAP_CREATE
) {
4303 catch (buffer::error
& e
) {
4306 ::encode(key
, newkeydata
);
4307 ::encode(val
, newkeydata
);
4308 dout(20) << " create " << key
<< " " << val
.length() << dendl
;
4310 } else if (op
== CEPH_OSD_TMAP_RM
) {
4315 } else if (op
== CEPH_OSD_TMAP_RMSLOPPY
) {
4318 dout(10) << " invalid tmap op " << (int)op
<< dendl
;
4325 ::encode(nextkey
, newkeydata
);
4326 ::encode(nextval
, newkeydata
);
4327 dout(20) << " keep " << nextkey
<< " " << nextval
.length() << dendl
;
4331 rest
.substr_of(newop
.outdata
, ip
.get_off(), newop
.outdata
.length() - ip
.get_off());
4332 dout(20) << " keep trailing " << rest
.length()
4333 << " at " << newkeydata
.length() << dendl
;
4334 newkeydata
.claim_append(rest
);
4337 // encode final key count + key data
4338 dout(20) << "tmapup final nkeys " << nkeys
<< dendl
;
4339 ::encode(nkeys
, obl
);
4340 obl
.claim_append(newkeydata
);
4343 dout(30) << " final is \n";
4344 obl
.hexdump(*_dout
);
4348 bufferlist::iterator tp
= obl
.begin();
4351 map
<string
,bufferlist
> d
;
4354 dout(0) << " **** debug sanity check, looks ok ****" << dendl
;
4359 dout(20) << "tmapput write " << obl
.length() << dendl
;
4360 newop
.op
.op
= CEPH_OSD_OP_WRITEFULL
;
4361 newop
.op
.extent
.offset
= 0;
4362 newop
.op
.extent
.length
= obl
.length();
4364 do_osd_ops(ctx
, nops
);
4365 osd_op
.outdata
.claim(newop
.outdata
);
4371 static int check_offset_and_length(uint64_t offset
, uint64_t length
, uint64_t max
)
4373 if (offset
>= max
||
4375 offset
+ length
> max
)
4381 struct FillInVerifyExtent
: public Context
{
4384 bufferlist
*outdatap
;
4385 boost::optional
<uint32_t> maybe_crc
;
4390 FillInVerifyExtent(ceph_le64
*r
, int32_t *rv
, bufferlist
*blp
,
4391 boost::optional
<uint32_t> mc
, uint64_t size
,
4392 OSDService
*osd
, hobject_t soid
, __le32 flags
) :
4393 r(r
), rval(rv
), outdatap(blp
), maybe_crc(mc
),
4394 size(size
), osd(osd
), soid(soid
), flags(flags
) {}
4395 void finish(int len
) override
{
4403 // whole object? can we verify the checksum?
4404 if (maybe_crc
&& *r
== size
) {
4405 uint32_t crc
= outdatap
->crc32c(-1);
4406 if (maybe_crc
!= crc
) {
4407 osd
->clog
->error() << std::hex
<< " full-object read crc 0x" << crc
4408 << " != expected 0x" << *maybe_crc
4409 << std::dec
<< " on " << soid
;
4410 if (!(flags
& CEPH_OSD_OP_FLAG_FAILOK
)) {
4419 struct ToSparseReadResult
: public Context
{
4421 bufferlist
* data_bl
;
4422 uint64_t data_offset
;
4424 ToSparseReadResult(int* result
, bufferlist
* bl
, uint64_t offset
,
4426 : result(result
), data_bl(bl
), data_offset(offset
),len(len
) {}
4427 void finish(int r
) override
{
4435 map
<uint64_t, uint64_t> extents
= {{data_offset
, r
}};
4436 ::encode(extents
, outdata
);
4437 ::encode_destructively(*data_bl
, outdata
);
4438 data_bl
->swap(outdata
);
4442 template<typename V
>
4443 static string
list_keys(const map
<string
, V
>& m
) {
4445 for (typename map
<string
, V
>::const_iterator itr
= m
.begin(); itr
!= m
.end(); ++itr
) {
4449 s
.append(itr
->first
);
4454 template<typename T
>
4455 static string
list_entries(const T
& m
) {
4457 for (typename
T::const_iterator itr
= m
.begin(); itr
!= m
.end(); ++itr
) {
4466 void PrimaryLogPG::maybe_create_new_object(
4468 bool ignore_transaction
)
4470 ObjectState
& obs
= ctx
->new_obs
;
4472 ctx
->delta_stats
.num_objects
++;
4474 assert(!obs
.oi
.is_whiteout());
4475 obs
.oi
.new_object();
4476 if (!ignore_transaction
)
4477 ctx
->op_t
->create(obs
.oi
.soid
);
4478 } else if (obs
.oi
.is_whiteout()) {
4479 dout(10) << __func__
<< " clearing whiteout on " << obs
.oi
.soid
<< dendl
;
4480 ctx
->new_obs
.oi
.clear_flag(object_info_t::FLAG_WHITEOUT
);
4481 --ctx
->delta_stats
.num_whiteouts
;
4485 struct ReadFinisher
: public PrimaryLogPG::OpFinisher
{
4488 ReadFinisher(OSDOp
& osd_op
) : osd_op(osd_op
) {
4491 int execute() override
{
4496 struct C_ChecksumRead
: public Context
{
4497 PrimaryLogPG
*primary_log_pg
;
4499 Checksummer::CSumType csum_type
;
4500 bufferlist init_value_bl
;
4501 ceph_le64 read_length
;
4503 Context
*fill_extent_ctx
;
4505 C_ChecksumRead(PrimaryLogPG
*primary_log_pg
, OSDOp
&osd_op
,
4506 Checksummer::CSumType csum_type
, bufferlist
&&init_value_bl
,
4507 boost::optional
<uint32_t> maybe_crc
, uint64_t size
,
4508 OSDService
*osd
, hobject_t soid
, __le32 flags
)
4509 : primary_log_pg(primary_log_pg
), osd_op(osd_op
),
4510 csum_type(csum_type
), init_value_bl(std::move(init_value_bl
)),
4511 fill_extent_ctx(new FillInVerifyExtent(&read_length
, &osd_op
.rval
,
4512 &read_bl
, maybe_crc
, size
,
4513 osd
, soid
, flags
)) {
4515 ~C_ChecksumRead() override
{
4516 delete fill_extent_ctx
;
4519 void finish(int r
) override
{
4520 fill_extent_ctx
->complete(r
);
4521 fill_extent_ctx
= nullptr;
4523 if (osd_op
.rval
>= 0) {
4524 bufferlist::iterator init_value_bl_it
= init_value_bl
.begin();
4525 osd_op
.rval
= primary_log_pg
->finish_checksum(osd_op
, csum_type
,
4526 &init_value_bl_it
, read_bl
);
4531 int PrimaryLogPG::do_checksum(OpContext
*ctx
, OSDOp
& osd_op
,
4532 bufferlist::iterator
*bl_it
)
4534 dout(20) << __func__
<< dendl
;
4535 bool skip_data_digest
=
4536 (osd
->store
->has_builtin_csum() && g_conf
->osd_skip_data_digest
) ||
4537 g_conf
->osd_distrust_data_digest
;
4539 auto& op
= osd_op
.op
;
4540 if (op
.checksum
.chunk_size
> 0) {
4541 if (op
.checksum
.length
== 0) {
4542 dout(10) << __func__
<< ": length required when chunk size provided"
4546 if (op
.checksum
.length
% op
.checksum
.chunk_size
!= 0) {
4547 dout(10) << __func__
<< ": length not aligned to chunk size" << dendl
;
4552 auto& oi
= ctx
->new_obs
.oi
;
4553 if (op
.checksum
.offset
== 0 && op
.checksum
.length
== 0) {
4554 // zeroed offset+length implies checksum whole object
4555 op
.checksum
.length
= oi
.size
;
4556 } else if (op
.checksum
.offset
+ op
.checksum
.length
> oi
.size
) {
4560 Checksummer::CSumType csum_type
;
4561 switch (op
.checksum
.type
) {
4562 case CEPH_OSD_CHECKSUM_OP_TYPE_XXHASH32
:
4563 csum_type
= Checksummer::CSUM_XXHASH32
;
4565 case CEPH_OSD_CHECKSUM_OP_TYPE_XXHASH64
:
4566 csum_type
= Checksummer::CSUM_XXHASH64
;
4568 case CEPH_OSD_CHECKSUM_OP_TYPE_CRC32C
:
4569 csum_type
= Checksummer::CSUM_CRC32C
;
4572 dout(10) << __func__
<< ": unknown crc type ("
4573 << static_cast<uint32_t>(op
.checksum
.type
) << ")" << dendl
;
4577 size_t csum_init_value_size
= Checksummer::get_csum_init_value_size(csum_type
);
4578 if (bl_it
->get_remaining() < csum_init_value_size
) {
4579 dout(10) << __func__
<< ": init value not provided" << dendl
;
4583 bufferlist init_value_bl
;
4584 init_value_bl
.substr_of(bl_it
->get_bl(), bl_it
->get_off(),
4585 csum_init_value_size
);
4586 bl_it
->advance(csum_init_value_size
);
4588 if (pool
.info
.require_rollback() && op
.checksum
.length
> 0) {
4589 // If there is a data digest and it is possible we are reading
4590 // entire object, pass the digest.
4591 boost::optional
<uint32_t> maybe_crc
;
4592 if (!skip_data_digest
&&
4593 oi
.is_data_digest() && op
.checksum
.offset
== 0 &&
4594 op
.checksum
.length
>= oi
.size
) {
4595 maybe_crc
= oi
.data_digest
;
4599 auto& soid
= oi
.soid
;
4600 auto checksum_ctx
= new C_ChecksumRead(this, osd_op
, csum_type
,
4601 std::move(init_value_bl
), maybe_crc
,
4602 oi
.size
, osd
, soid
, op
.flags
);
4604 ctx
->pending_async_reads
.push_back({
4605 {op
.checksum
.offset
, op
.checksum
.length
, op
.flags
},
4606 {&checksum_ctx
->read_bl
, checksum_ctx
}});
4608 dout(10) << __func__
<< ": async_read noted for " << soid
<< dendl
;
4609 ctx
->op_finishers
[ctx
->current_osd_subop_num
].reset(
4610 new ReadFinisher(osd_op
));
4611 return -EINPROGRESS
;
4615 std::vector
<OSDOp
> read_ops(1);
4616 auto& read_op
= read_ops
[0];
4617 if (op
.checksum
.length
> 0) {
4618 read_op
.op
.op
= CEPH_OSD_OP_READ
;
4619 read_op
.op
.flags
= op
.flags
;
4620 read_op
.op
.extent
.offset
= op
.checksum
.offset
;
4621 read_op
.op
.extent
.length
= op
.checksum
.length
;
4622 read_op
.op
.extent
.truncate_size
= 0;
4623 read_op
.op
.extent
.truncate_seq
= 0;
4625 int r
= do_osd_ops(ctx
, read_ops
);
4627 derr
<< __func__
<< ": do_osd_ops failed: " << cpp_strerror(r
) << dendl
;
4632 bufferlist::iterator init_value_bl_it
= init_value_bl
.begin();
4633 return finish_checksum(osd_op
, csum_type
, &init_value_bl_it
,
4637 int PrimaryLogPG::finish_checksum(OSDOp
& osd_op
,
4638 Checksummer::CSumType csum_type
,
4639 bufferlist::iterator
*init_value_bl_it
,
4640 const bufferlist
&read_bl
) {
4641 dout(20) << __func__
<< dendl
;
4643 auto& op
= osd_op
.op
;
4645 if (op
.checksum
.length
> 0 && read_bl
.length() != op
.checksum
.length
) {
4646 derr
<< __func__
<< ": bytes read " << read_bl
.length() << " != "
4647 << op
.checksum
.length
<< dendl
;
4651 size_t csum_chunk_size
= (op
.checksum
.chunk_size
!= 0 ?
4652 op
.checksum
.chunk_size
: read_bl
.length());
4653 uint32_t csum_count
= (csum_chunk_size
> 0 ?
4654 read_bl
.length() / csum_chunk_size
: 0);
4657 bufferptr csum_data
;
4658 if (csum_count
> 0) {
4659 size_t csum_value_size
= Checksummer::get_csum_value_size(csum_type
);
4660 csum_data
= buffer::create(csum_value_size
* csum_count
);
4662 csum
.append(csum_data
);
4664 switch (csum_type
) {
4665 case Checksummer::CSUM_XXHASH32
:
4667 Checksummer::xxhash32::init_value_t init_value
;
4668 ::decode(init_value
, *init_value_bl_it
);
4669 Checksummer::calculate
<Checksummer::xxhash32
>(
4670 init_value
, csum_chunk_size
, 0, read_bl
.length(), read_bl
,
4674 case Checksummer::CSUM_XXHASH64
:
4676 Checksummer::xxhash64::init_value_t init_value
;
4677 ::decode(init_value
, *init_value_bl_it
);
4678 Checksummer::calculate
<Checksummer::xxhash64
>(
4679 init_value
, csum_chunk_size
, 0, read_bl
.length(), read_bl
,
4683 case Checksummer::CSUM_CRC32C
:
4685 Checksummer::crc32c::init_value_t init_value
;
4686 ::decode(init_value
, *init_value_bl_it
);
4687 Checksummer::calculate
<Checksummer::crc32c
>(
4688 init_value
, csum_chunk_size
, 0, read_bl
.length(), read_bl
,
4697 ::encode(csum_count
, osd_op
.outdata
);
4698 osd_op
.outdata
.claim_append(csum
);
4702 struct C_ExtentCmpRead
: public Context
{
4703 PrimaryLogPG
*primary_log_pg
;
4705 ceph_le64 read_length
;
4707 Context
*fill_extent_ctx
;
4709 C_ExtentCmpRead(PrimaryLogPG
*primary_log_pg
, OSDOp
&osd_op
,
4710 boost::optional
<uint32_t> maybe_crc
, uint64_t size
,
4711 OSDService
*osd
, hobject_t soid
, __le32 flags
)
4712 : primary_log_pg(primary_log_pg
), osd_op(osd_op
),
4713 fill_extent_ctx(new FillInVerifyExtent(&read_length
, &osd_op
.rval
,
4714 &read_bl
, maybe_crc
, size
,
4715 osd
, soid
, flags
)) {
4717 ~C_ExtentCmpRead() override
{
4718 delete fill_extent_ctx
;
4721 void finish(int r
) override
{
4725 delete fill_extent_ctx
;
4727 fill_extent_ctx
->complete(r
);
4729 fill_extent_ctx
= nullptr;
4731 if (osd_op
.rval
>= 0) {
4732 osd_op
.rval
= primary_log_pg
->finish_extent_cmp(osd_op
, read_bl
);
4737 int PrimaryLogPG::do_extent_cmp(OpContext
*ctx
, OSDOp
& osd_op
)
4739 dout(20) << __func__
<< dendl
;
4740 ceph_osd_op
& op
= osd_op
.op
;
4741 bool skip_data_digest
=
4742 (osd
->store
->has_builtin_csum() && g_conf
->osd_skip_data_digest
) ||
4743 g_conf
->osd_distrust_data_digest
;
4745 auto& oi
= ctx
->new_obs
.oi
;
4746 uint64_t size
= oi
.size
;
4747 if ((oi
.truncate_seq
< op
.extent
.truncate_seq
) &&
4748 (op
.extent
.offset
+ op
.extent
.length
> op
.extent
.truncate_size
)) {
4749 size
= op
.extent
.truncate_size
;
4752 if (op
.extent
.offset
>= size
) {
4753 op
.extent
.length
= 0;
4754 } else if (op
.extent
.offset
+ op
.extent
.length
> size
) {
4755 op
.extent
.length
= size
- op
.extent
.offset
;
4758 if (op
.extent
.length
== 0) {
4759 dout(20) << __func__
<< " zero length extent" << dendl
;
4760 return finish_extent_cmp(osd_op
, bufferlist
{});
4761 } else if (!ctx
->obs
->exists
|| ctx
->obs
->oi
.is_whiteout()) {
4762 dout(20) << __func__
<< " object DNE" << dendl
;
4763 return finish_extent_cmp(osd_op
, {});
4764 } else if (pool
.info
.require_rollback()) {
4765 // If there is a data digest and it is possible we are reading
4766 // entire object, pass the digest.
4767 boost::optional
<uint32_t> maybe_crc
;
4768 if (!skip_data_digest
&&
4769 oi
.is_data_digest() && op
.checksum
.offset
== 0 &&
4770 op
.checksum
.length
>= oi
.size
) {
4771 maybe_crc
= oi
.data_digest
;
4775 auto& soid
= oi
.soid
;
4776 auto extent_cmp_ctx
= new C_ExtentCmpRead(this, osd_op
, maybe_crc
, oi
.size
,
4777 osd
, soid
, op
.flags
);
4778 ctx
->pending_async_reads
.push_back({
4779 {op
.extent
.offset
, op
.extent
.length
, op
.flags
},
4780 {&extent_cmp_ctx
->read_bl
, extent_cmp_ctx
}});
4782 dout(10) << __func__
<< ": async_read noted for " << soid
<< dendl
;
4784 ctx
->op_finishers
[ctx
->current_osd_subop_num
].reset(
4785 new ReadFinisher(osd_op
));
4786 return -EINPROGRESS
;
4790 vector
<OSDOp
> read_ops(1);
4791 OSDOp
& read_op
= read_ops
[0];
4793 read_op
.op
.op
= CEPH_OSD_OP_SYNC_READ
;
4794 read_op
.op
.extent
.offset
= op
.extent
.offset
;
4795 read_op
.op
.extent
.length
= op
.extent
.length
;
4796 read_op
.op
.extent
.truncate_seq
= op
.extent
.truncate_seq
;
4797 read_op
.op
.extent
.truncate_size
= op
.extent
.truncate_size
;
4799 int result
= do_osd_ops(ctx
, read_ops
);
4801 derr
<< __func__
<< " failed " << result
<< dendl
;
4804 return finish_extent_cmp(osd_op
, read_op
.outdata
);
4807 int PrimaryLogPG::finish_extent_cmp(OSDOp
& osd_op
, const bufferlist
&read_bl
)
4809 for (uint64_t idx
= 0; idx
< osd_op
.indata
.length(); ++idx
) {
4810 char read_byte
= (idx
< read_bl
.length() ? read_bl
[idx
] : 0);
4811 if (osd_op
.indata
[idx
] != read_byte
) {
4812 return (-MAX_ERRNO
- idx
);
4819 int PrimaryLogPG::do_read(OpContext
*ctx
, OSDOp
& osd_op
) {
4820 dout(20) << __func__
<< dendl
;
4821 auto& op
= osd_op
.op
;
4822 auto& oi
= ctx
->new_obs
.oi
;
4823 auto& soid
= oi
.soid
;
4824 __u32 seq
= oi
.truncate_seq
;
4825 uint64_t size
= oi
.size
;
4826 bool trimmed_read
= false;
4827 bool skip_data_digest
=
4828 (osd
->store
->has_builtin_csum() && g_conf
->osd_skip_data_digest
) ||
4829 g_conf
->osd_distrust_data_digest
;
4831 // are we beyond truncate_size?
4832 if ( (seq
< op
.extent
.truncate_seq
) &&
4833 (op
.extent
.offset
+ op
.extent
.length
> op
.extent
.truncate_size
) )
4834 size
= op
.extent
.truncate_size
;
4836 if (op
.extent
.length
== 0) //length is zero mean read the whole object
4837 op
.extent
.length
= size
;
4839 if (op
.extent
.offset
>= size
) {
4840 op
.extent
.length
= 0;
4841 trimmed_read
= true;
4842 } else if (op
.extent
.offset
+ op
.extent
.length
> size
) {
4843 op
.extent
.length
= size
- op
.extent
.offset
;
4844 trimmed_read
= true;
4847 // read into a buffer
4849 if (trimmed_read
&& op
.extent
.length
== 0) {
4850 // read size was trimmed to zero and it is expected to do nothing
4851 // a read operation of 0 bytes does *not* do nothing, this is why
4852 // the trimmed_read boolean is needed
4853 } else if (pool
.info
.require_rollback()) {
4854 boost::optional
<uint32_t> maybe_crc
;
4855 // If there is a data digest and it is possible we are reading
4856 // entire object, pass the digest. FillInVerifyExtent will
4857 // will check the oi.size again.
4858 if (!skip_data_digest
&&
4859 oi
.is_data_digest() && op
.extent
.offset
== 0 &&
4860 op
.extent
.length
>= oi
.size
)
4861 maybe_crc
= oi
.data_digest
;
4862 ctx
->pending_async_reads
.push_back(
4864 boost::make_tuple(op
.extent
.offset
, op
.extent
.length
, op
.flags
),
4865 make_pair(&osd_op
.outdata
,
4866 new FillInVerifyExtent(&op
.extent
.length
, &osd_op
.rval
,
4867 &osd_op
.outdata
, maybe_crc
, oi
.size
,
4868 osd
, soid
, op
.flags
))));
4869 dout(10) << " async_read noted for " << soid
<< dendl
;
4871 ctx
->op_finishers
[ctx
->current_osd_subop_num
].reset(
4872 new ReadFinisher(osd_op
));
4874 int r
= pgbackend
->objects_read_sync(
4875 soid
, op
.extent
.offset
, op
.extent
.length
, op
.flags
, &osd_op
.outdata
);
4877 r
= rep_repair_primary_object(soid
, ctx
->op
);
4880 op
.extent
.length
= r
;
4883 op
.extent
.length
= 0;
4885 dout(10) << " read got " << r
<< " / " << op
.extent
.length
4886 << " bytes from obj " << soid
<< dendl
;
4888 // whole object? can we verify the checksum?
4889 if (!skip_data_digest
&&
4890 op
.extent
.length
== oi
.size
&& oi
.is_data_digest()) {
4891 uint32_t crc
= osd_op
.outdata
.crc32c(-1);
4892 if (oi
.data_digest
!= crc
) {
4893 osd
->clog
->error() << info
.pgid
<< std::hex
4894 << " full-object read crc 0x" << crc
4895 << " != expected 0x" << oi
.data_digest
4896 << std::dec
<< " on " << soid
;
4897 // FIXME fall back to replica or something?
4903 // XXX the op.extent.length is the requested length for async read
4904 // On error this length is changed to 0 after the error comes back.
4905 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(op
.extent
.length
, 10);
4906 ctx
->delta_stats
.num_rd
++;
4910 int PrimaryLogPG::do_sparse_read(OpContext
*ctx
, OSDOp
& osd_op
) {
4911 dout(20) << __func__
<< dendl
;
4912 auto& op
= osd_op
.op
;
4913 auto& oi
= ctx
->new_obs
.oi
;
4914 auto& soid
= oi
.soid
;
4915 bool skip_data_digest
=
4916 (osd
->store
->has_builtin_csum() && g_conf
->osd_skip_data_digest
) ||
4917 g_conf
->osd_distrust_data_digest
;
4919 if (op
.extent
.truncate_seq
) {
4920 dout(0) << "sparse_read does not support truncation sequence " << dendl
;
4925 if (pool
.info
.ec_pool()) {
4926 // translate sparse read to a normal one if not supported
4927 uint64_t offset
= op
.extent
.offset
;
4928 uint64_t length
= op
.extent
.length
;
4929 if (offset
> oi
.size
) {
4931 } else if (offset
+ length
> oi
.size
) {
4932 length
= oi
.size
- offset
;
4936 ctx
->pending_async_reads
.push_back(
4938 boost::make_tuple(offset
, length
, op
.flags
),
4941 new ToSparseReadResult(&osd_op
.rval
, &osd_op
.outdata
, offset
,
4942 &op
.extent
.length
))));
4943 dout(10) << " async_read (was sparse_read) noted for " << soid
<< dendl
;
4945 ctx
->op_finishers
[ctx
->current_osd_subop_num
].reset(
4946 new ReadFinisher(osd_op
));
4948 dout(10) << " sparse read ended up empty for " << soid
<< dendl
;
4949 map
<uint64_t, uint64_t> extents
;
4950 ::encode(extents
, osd_op
.outdata
);
4953 // read into a buffer
4954 map
<uint64_t, uint64_t> m
;
4955 uint32_t total_read
= 0;
4956 int r
= osd
->store
->fiemap(ch
, ghobject_t(soid
, ghobject_t::NO_GEN
,
4958 op
.extent
.offset
, op
.extent
.length
, m
);
4963 map
<uint64_t, uint64_t>::iterator miter
;
4965 uint64_t last
= op
.extent
.offset
;
4966 for (miter
= m
.begin(); miter
!= m
.end(); ++miter
) {
4968 if (cct
->_conf
->osd_verify_sparse_read_holes
&&
4969 last
< miter
->first
) {
4971 uint64_t len
= miter
->first
- last
;
4972 r
= pgbackend
->objects_read_sync(soid
, last
, len
, op
.flags
, &t
);
4974 osd
->clog
->error() << coll
<< " " << soid
4975 << " sparse-read failed to read: "
4977 } else if (!t
.is_zero()) {
4978 osd
->clog
->error() << coll
<< " " << soid
4979 << " sparse-read found data in hole "
4980 << last
<< "~" << len
;
4985 r
= pgbackend
->objects_read_sync(soid
, miter
->first
, miter
->second
,
4988 r
= rep_repair_primary_object(soid
, ctx
->op
);
4994 // this is usually happen when we get extent that exceeds the actual file
4996 if (r
< (int)miter
->second
)
4999 dout(10) << "sparse-read " << miter
->first
<< "@" << miter
->second
5001 data_bl
.claim_append(tmpbl
);
5002 last
= miter
->first
+ r
;
5009 // verify trailing hole?
5010 if (cct
->_conf
->osd_verify_sparse_read_holes
) {
5011 uint64_t end
= MIN(op
.extent
.offset
+ op
.extent
.length
, oi
.size
);
5014 uint64_t len
= end
- last
;
5015 r
= pgbackend
->objects_read_sync(soid
, last
, len
, op
.flags
, &t
);
5017 osd
->clog
->error() << coll
<< " " << soid
5018 << " sparse-read failed to read: " << r
;
5019 } else if (!t
.is_zero()) {
5020 osd
->clog
->error() << coll
<< " " << soid
5021 << " sparse-read found data in hole "
5022 << last
<< "~" << len
;
5027 // Why SPARSE_READ need checksum? In fact, librbd always use sparse-read.
5028 // Maybe at first, there is no much whole objects. With continued use, more
5029 // and more whole object exist. So from this point, for spare-read add
5030 // checksum make sense.
5031 if (!skip_data_digest
&&
5032 total_read
== oi
.size
&& oi
.is_data_digest()) {
5033 uint32_t crc
= data_bl
.crc32c(-1);
5034 if (oi
.data_digest
!= crc
) {
5035 osd
->clog
->error() << info
.pgid
<< std::hex
5036 << " full-object read crc 0x" << crc
5037 << " != expected 0x" << oi
.data_digest
5038 << std::dec
<< " on " << soid
;
5039 // FIXME fall back to replica or something?
5044 op
.extent
.length
= total_read
;
5046 ::encode(m
, osd_op
.outdata
); // re-encode since it might be modified
5047 ::encode_destructively(data_bl
, osd_op
.outdata
);
5049 dout(10) << " sparse_read got " << total_read
<< " bytes from object "
5053 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(op
.extent
.length
, 10);
5054 ctx
->delta_stats
.num_rd
++;
5058 int PrimaryLogPG::do_osd_ops(OpContext
*ctx
, vector
<OSDOp
>& ops
)
5061 SnapSetContext
*ssc
= ctx
->obc
->ssc
;
5062 ObjectState
& obs
= ctx
->new_obs
;
5063 object_info_t
& oi
= obs
.oi
;
5064 const hobject_t
& soid
= oi
.soid
;
5065 bool skip_data_digest
=
5066 (osd
->store
->has_builtin_csum() && g_conf
->osd_skip_data_digest
) ||
5067 g_conf
->osd_distrust_data_digest
;
5069 PGTransaction
* t
= ctx
->op_t
.get();
5071 dout(10) << "do_osd_op " << soid
<< " " << ops
<< dendl
;
5073 ctx
->current_osd_subop_num
= 0;
5074 for (auto p
= ops
.begin(); p
!= ops
.end(); ++p
, ctx
->current_osd_subop_num
++, ctx
->processed_subop_count
++) {
5076 ceph_osd_op
& op
= osd_op
.op
;
5078 OpFinisher
* op_finisher
= nullptr;
5080 auto op_finisher_it
= ctx
->op_finishers
.find(ctx
->current_osd_subop_num
);
5081 if (op_finisher_it
!= ctx
->op_finishers
.end()) {
5082 op_finisher
= op_finisher_it
->second
.get();
5086 // TODO: check endianness (__le32 vs uint32_t, etc.)
5087 // The fields in ceph_osd_op are little-endian (according to the definition in rados.h),
5088 // but the code in this function seems to treat them as native-endian. What should the
5090 tracepoint(osd
, do_osd_op_pre
, soid
.oid
.name
.c_str(), soid
.snap
.val
, op
.op
, ceph_osd_op_name(op
.op
), op
.flags
);
5092 dout(10) << "do_osd_op " << osd_op
<< dendl
;
5094 bufferlist::iterator bp
= osd_op
.indata
.begin();
5096 // user-visible modifcation?
5098 // non user-visible modifications
5099 case CEPH_OSD_OP_WATCH
:
5100 case CEPH_OSD_OP_CACHE_EVICT
:
5101 case CEPH_OSD_OP_CACHE_FLUSH
:
5102 case CEPH_OSD_OP_CACHE_TRY_FLUSH
:
5103 case CEPH_OSD_OP_UNDIRTY
:
5104 case CEPH_OSD_OP_COPY_FROM
: // we handle user_version update explicitly
5105 case CEPH_OSD_OP_CACHE_PIN
:
5106 case CEPH_OSD_OP_CACHE_UNPIN
:
5107 case CEPH_OSD_OP_SET_REDIRECT
:
5110 if (op
.op
& CEPH_OSD_OP_MODE_WR
)
5111 ctx
->user_modify
= true;
5114 // munge -1 truncate to 0 truncate
5115 if (ceph_osd_op_uses_extent(op
.op
) &&
5116 op
.extent
.truncate_seq
== 1 &&
5117 op
.extent
.truncate_size
== (-1ULL)) {
5118 op
.extent
.truncate_size
= 0;
5119 op
.extent
.truncate_seq
= 0;
5122 // munge ZERO -> TRUNCATE? (don't munge to DELETE or we risk hosing attributes)
5123 if (op
.op
== CEPH_OSD_OP_ZERO
&&
5125 op
.extent
.offset
< cct
->_conf
->osd_max_object_size
&&
5126 op
.extent
.length
>= 1 &&
5127 op
.extent
.length
<= cct
->_conf
->osd_max_object_size
&&
5128 op
.extent
.offset
+ op
.extent
.length
>= oi
.size
) {
5129 if (op
.extent
.offset
>= oi
.size
) {
5133 dout(10) << " munging ZERO " << op
.extent
.offset
<< "~" << op
.extent
.length
5134 << " -> TRUNCATE " << op
.extent
.offset
<< " (old size is " << oi
.size
<< ")" << dendl
;
5135 op
.op
= CEPH_OSD_OP_TRUNCATE
;
5142 case CEPH_OSD_OP_CMPEXT
:
5144 tracepoint(osd
, do_osd_op_pre_extent_cmp
, soid
.oid
.name
.c_str(),
5145 soid
.snap
.val
, oi
.size
, oi
.truncate_seq
, op
.extent
.offset
,
5146 op
.extent
.length
, op
.extent
.truncate_size
,
5147 op
.extent
.truncate_seq
);
5149 if (op_finisher
== nullptr) {
5150 result
= do_extent_cmp(ctx
, osd_op
);
5152 result
= op_finisher
->execute();
5156 case CEPH_OSD_OP_SYNC_READ
:
5157 if (pool
.info
.require_rollback()) {
5158 result
= -EOPNOTSUPP
;
5162 case CEPH_OSD_OP_READ
:
5164 tracepoint(osd
, do_osd_op_pre_read
, soid
.oid
.name
.c_str(),
5165 soid
.snap
.val
, oi
.size
, oi
.truncate_seq
, op
.extent
.offset
,
5166 op
.extent
.length
, op
.extent
.truncate_size
,
5167 op
.extent
.truncate_seq
);
5168 if (op_finisher
== nullptr) {
5169 if (!ctx
->data_off
) {
5170 ctx
->data_off
= op
.extent
.offset
;
5172 result
= do_read(ctx
, osd_op
);
5174 result
= op_finisher
->execute();
5178 case CEPH_OSD_OP_CHECKSUM
:
5181 tracepoint(osd
, do_osd_op_pre_checksum
, soid
.oid
.name
.c_str(),
5182 soid
.snap
.val
, oi
.size
, oi
.truncate_seq
, op
.checksum
.type
,
5183 op
.checksum
.offset
, op
.checksum
.length
,
5184 op
.checksum
.chunk_size
);
5186 if (op_finisher
== nullptr) {
5187 result
= do_checksum(ctx
, osd_op
, &bp
);
5189 result
= op_finisher
->execute();
5195 case CEPH_OSD_OP_MAPEXT
:
5196 tracepoint(osd
, do_osd_op_pre_mapext
, soid
.oid
.name
.c_str(), soid
.snap
.val
, op
.extent
.offset
, op
.extent
.length
);
5197 if (pool
.info
.require_rollback()) {
5198 result
= -EOPNOTSUPP
;
5203 // read into a buffer
5205 int r
= osd
->store
->fiemap(ch
, ghobject_t(soid
, ghobject_t::NO_GEN
,
5207 op
.extent
.offset
, op
.extent
.length
, bl
);
5208 osd_op
.outdata
.claim(bl
);
5212 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(bl
.length(), 10);
5213 ctx
->delta_stats
.num_rd
++;
5214 dout(10) << " map_extents done on object " << soid
<< dendl
;
5219 case CEPH_OSD_OP_SPARSE_READ
:
5220 tracepoint(osd
, do_osd_op_pre_sparse_read
, soid
.oid
.name
.c_str(),
5221 soid
.snap
.val
, oi
.size
, oi
.truncate_seq
, op
.extent
.offset
,
5222 op
.extent
.length
, op
.extent
.truncate_size
,
5223 op
.extent
.truncate_seq
);
5224 if (op_finisher
== nullptr) {
5225 result
= do_sparse_read(ctx
, osd_op
);
5227 result
= op_finisher
->execute();
5231 case CEPH_OSD_OP_CALL
:
5233 string cname
, mname
;
5236 bp
.copy(op
.cls
.class_len
, cname
);
5237 bp
.copy(op
.cls
.method_len
, mname
);
5238 bp
.copy(op
.cls
.indata_len
, indata
);
5239 } catch (buffer::error
& e
) {
5240 dout(10) << "call unable to decode class + method + indata" << dendl
;
5241 dout(30) << "in dump: ";
5242 osd_op
.indata
.hexdump(*_dout
);
5245 tracepoint(osd
, do_osd_op_pre_call
, soid
.oid
.name
.c_str(), soid
.snap
.val
, "???", "???");
5248 tracepoint(osd
, do_osd_op_pre_call
, soid
.oid
.name
.c_str(), soid
.snap
.val
, cname
.c_str(), mname
.c_str());
5250 ClassHandler::ClassData
*cls
;
5251 result
= osd
->class_handler
->open_class(cname
, &cls
);
5252 assert(result
== 0); // init_op_flags() already verified this works.
5254 ClassHandler::ClassMethod
*method
= cls
->get_method(mname
.c_str());
5256 dout(10) << "call method " << cname
<< "." << mname
<< " does not exist" << dendl
;
5257 result
= -EOPNOTSUPP
;
5261 int flags
= method
->get_flags();
5262 if (flags
& CLS_METHOD_WR
)
5263 ctx
->user_modify
= true;
5266 dout(10) << "call method " << cname
<< "." << mname
<< dendl
;
5267 int prev_rd
= ctx
->num_read
;
5268 int prev_wr
= ctx
->num_write
;
5269 result
= method
->exec((cls_method_context_t
)&ctx
, indata
, outdata
);
5271 if (ctx
->num_read
> prev_rd
&& !(flags
& CLS_METHOD_RD
)) {
5272 derr
<< "method " << cname
<< "." << mname
<< " tried to read object but is not marked RD" << dendl
;
5276 if (ctx
->num_write
> prev_wr
&& !(flags
& CLS_METHOD_WR
)) {
5277 derr
<< "method " << cname
<< "." << mname
<< " tried to update object but is not marked WR" << dendl
;
5282 dout(10) << "method called response length=" << outdata
.length() << dendl
;
5283 op
.extent
.length
= outdata
.length();
5284 osd_op
.outdata
.claim_append(outdata
);
5285 dout(30) << "out dump: ";
5286 osd_op
.outdata
.hexdump(*_dout
);
5291 case CEPH_OSD_OP_STAT
:
5292 // note: stat does not require RD
5294 tracepoint(osd
, do_osd_op_pre_stat
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5296 if (obs
.exists
&& !oi
.is_whiteout()) {
5297 ::encode(oi
.size
, osd_op
.outdata
);
5298 ::encode(oi
.mtime
, osd_op
.outdata
);
5299 dout(10) << "stat oi has " << oi
.size
<< " " << oi
.mtime
<< dendl
;
5302 dout(10) << "stat oi object does not exist" << dendl
;
5305 ctx
->delta_stats
.num_rd
++;
5309 case CEPH_OSD_OP_ISDIRTY
:
5312 tracepoint(osd
, do_osd_op_pre_isdirty
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5313 bool is_dirty
= obs
.oi
.is_dirty();
5314 ::encode(is_dirty
, osd_op
.outdata
);
5315 ctx
->delta_stats
.num_rd
++;
5320 case CEPH_OSD_OP_UNDIRTY
:
5323 tracepoint(osd
, do_osd_op_pre_undirty
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5324 if (oi
.is_dirty()) {
5325 ctx
->undirty
= true; // see make_writeable()
5327 ctx
->delta_stats
.num_wr
++;
5333 case CEPH_OSD_OP_CACHE_TRY_FLUSH
:
5336 tracepoint(osd
, do_osd_op_pre_try_flush
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5337 if (ctx
->lock_type
!= ObjectContext::RWState::RWNONE
) {
5338 dout(10) << "cache-try-flush without SKIPRWLOCKS flag set" << dendl
;
5342 if (pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_NONE
) {
5350 if (oi
.is_cache_pinned()) {
5351 dout(10) << "cache-try-flush on a pinned object, consider unpin this object first" << dendl
;
5355 if (oi
.is_dirty()) {
5356 result
= start_flush(ctx
->op
, ctx
->obc
, false, NULL
, boost::none
);
5357 if (result
== -EINPROGRESS
)
5365 case CEPH_OSD_OP_CACHE_FLUSH
:
5368 tracepoint(osd
, do_osd_op_pre_cache_flush
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5369 if (ctx
->lock_type
== ObjectContext::RWState::RWNONE
) {
5370 dout(10) << "cache-flush with SKIPRWLOCKS flag set" << dendl
;
5374 if (pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_NONE
) {
5382 if (oi
.is_cache_pinned()) {
5383 dout(10) << "cache-flush on a pinned object, consider unpin this object first" << dendl
;
5388 if (oi
.is_dirty()) {
5389 result
= start_flush(ctx
->op
, ctx
->obc
, true, &missing
, boost::none
);
5390 if (result
== -EINPROGRESS
)
5395 // Check special return value which has set missing_return
5396 if (result
== -ENOENT
) {
5397 dout(10) << __func__
<< " CEPH_OSD_OP_CACHE_FLUSH got ENOENT" << dendl
;
5398 assert(!missing
.is_min());
5399 wait_for_unreadable_object(missing
, ctx
->op
);
5400 // Error code which is used elsewhere when wait_for_unreadable_object() is used
5406 case CEPH_OSD_OP_CACHE_EVICT
:
5409 tracepoint(osd
, do_osd_op_pre_cache_evict
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5410 if (pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_NONE
) {
5418 if (oi
.is_cache_pinned()) {
5419 dout(10) << "cache-evict on a pinned object, consider unpin this object first" << dendl
;
5423 if (oi
.is_dirty()) {
5427 if (!oi
.watchers
.empty()) {
5431 if (soid
.snap
== CEPH_NOSNAP
) {
5432 result
= _verify_no_head_clones(soid
, ssc
->snapset
);
5436 result
= _delete_oid(ctx
, true, false);
5438 // mark that this is a cache eviction to avoid triggering normal
5439 // make_writeable() clone or snapdir object creation in finish_ctx()
5440 ctx
->cache_evict
= true;
5442 osd
->logger
->inc(l_osd_tier_evict
);
5446 case CEPH_OSD_OP_GETXATTR
:
5450 bp
.copy(op
.xattr
.name_len
, aname
);
5451 tracepoint(osd
, do_osd_op_pre_getxattr
, soid
.oid
.name
.c_str(), soid
.snap
.val
, aname
.c_str());
5452 string name
= "_" + aname
;
5453 int r
= getattr_maybe_cache(
5458 op
.xattr
.value_len
= osd_op
.outdata
.length();
5460 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(osd_op
.outdata
.length(), 10);
5464 ctx
->delta_stats
.num_rd
++;
5468 case CEPH_OSD_OP_GETXATTRS
:
5471 tracepoint(osd
, do_osd_op_pre_getxattrs
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5472 map
<string
, bufferlist
> out
;
5473 result
= getattrs_maybe_cache(
5479 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(bl
.length(), 10);
5480 ctx
->delta_stats
.num_rd
++;
5481 osd_op
.outdata
.claim_append(bl
);
5485 case CEPH_OSD_OP_CMPXATTR
:
5489 bp
.copy(op
.xattr
.name_len
, aname
);
5490 tracepoint(osd
, do_osd_op_pre_cmpxattr
, soid
.oid
.name
.c_str(), soid
.snap
.val
, aname
.c_str());
5491 string name
= "_" + aname
;
5492 name
[op
.xattr
.name_len
+ 1] = 0;
5495 result
= getattr_maybe_cache(
5499 if (result
< 0 && result
!= -EEXIST
&& result
!= -ENODATA
)
5502 ctx
->delta_stats
.num_rd
++;
5503 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(xattr
.length(), 10);
5505 switch (op
.xattr
.cmp_mode
) {
5506 case CEPH_OSD_CMPXATTR_MODE_STRING
:
5509 bp
.copy(op
.xattr
.value_len
, val
);
5510 val
[op
.xattr
.value_len
] = 0;
5511 dout(10) << "CEPH_OSD_OP_CMPXATTR name=" << name
<< " val=" << val
5512 << " op=" << (int)op
.xattr
.cmp_op
<< " mode=" << (int)op
.xattr
.cmp_mode
<< dendl
;
5513 result
= do_xattr_cmp_str(op
.xattr
.cmp_op
, val
, xattr
);
5517 case CEPH_OSD_CMPXATTR_MODE_U64
:
5521 ::decode(u64val
, bp
);
5523 catch (buffer::error
& e
) {
5527 dout(10) << "CEPH_OSD_OP_CMPXATTR name=" << name
<< " val=" << u64val
5528 << " op=" << (int)op
.xattr
.cmp_op
<< " mode=" << (int)op
.xattr
.cmp_mode
<< dendl
;
5529 result
= do_xattr_cmp_u64(op
.xattr
.cmp_op
, u64val
, xattr
);
5534 dout(10) << "bad cmp mode " << (int)op
.xattr
.cmp_mode
<< dendl
;
5539 dout(10) << "comparison returned false" << dendl
;
5540 result
= -ECANCELED
;
5544 dout(10) << "comparison returned " << result
<< " " << cpp_strerror(-result
) << dendl
;
5548 dout(10) << "comparison returned true" << dendl
;
5552 case CEPH_OSD_OP_ASSERT_VER
:
5555 uint64_t ver
= op
.assert_ver
.ver
;
5556 tracepoint(osd
, do_osd_op_pre_assert_ver
, soid
.oid
.name
.c_str(), soid
.snap
.val
, ver
);
5559 else if (ver
< oi
.user_version
)
5561 else if (ver
> oi
.user_version
)
5562 result
= -EOVERFLOW
;
5566 case CEPH_OSD_OP_LIST_WATCHERS
:
5569 tracepoint(osd
, do_osd_op_pre_list_watchers
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5570 obj_list_watch_response_t resp
;
5572 map
<pair
<uint64_t, entity_name_t
>, watch_info_t
>::const_iterator oi_iter
;
5573 for (oi_iter
= oi
.watchers
.begin(); oi_iter
!= oi
.watchers
.end();
5575 dout(20) << "key cookie=" << oi_iter
->first
.first
5576 << " entity=" << oi_iter
->first
.second
<< " "
5577 << oi_iter
->second
<< dendl
;
5578 assert(oi_iter
->first
.first
== oi_iter
->second
.cookie
);
5579 assert(oi_iter
->first
.second
.is_client());
5581 watch_item_t
wi(oi_iter
->first
.second
, oi_iter
->second
.cookie
,
5582 oi_iter
->second
.timeout_seconds
, oi_iter
->second
.addr
);
5583 resp
.entries
.push_back(wi
);
5586 resp
.encode(osd_op
.outdata
, ctx
->get_features());
5589 ctx
->delta_stats
.num_rd
++;
5593 case CEPH_OSD_OP_LIST_SNAPS
:
5596 tracepoint(osd
, do_osd_op_pre_list_snaps
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5597 obj_list_snap_response_t resp
;
5600 ssc
= ctx
->obc
->ssc
= get_snapset_context(soid
, false);
5604 int clonecount
= ssc
->snapset
.clones
.size();
5605 if (ssc
->snapset
.head_exists
)
5607 resp
.clones
.reserve(clonecount
);
5608 for (auto clone_iter
= ssc
->snapset
.clones
.begin();
5609 clone_iter
!= ssc
->snapset
.clones
.end(); ++clone_iter
) {
5611 ci
.cloneid
= *clone_iter
;
5613 hobject_t clone_oid
= soid
;
5614 clone_oid
.snap
= *clone_iter
;
5616 if (!ssc
->snapset
.is_legacy()) {
5617 auto p
= ssc
->snapset
.clone_snaps
.find(*clone_iter
);
5618 if (p
== ssc
->snapset
.clone_snaps
.end()) {
5619 osd
->clog
->error() << "osd." << osd
->whoami
5620 << ": inconsistent clone_snaps found for oid "
5621 << soid
<< " clone " << *clone_iter
5622 << " snapset " << ssc
->snapset
;
5626 for (auto q
= p
->second
.rbegin(); q
!= p
->second
.rend(); ++q
) {
5627 ci
.snaps
.push_back(*q
);
5630 /* No need to take a lock here. We are only inspecting state cached on
5631 * in the ObjectContext, so we aren't performing an actual read unless
5632 * the clone obc is not already loaded (in which case, it cannot have
5633 * an in progress write). We also do not risk exposing uncommitted
5634 * state since we do have a read lock on the head object or snapdir,
5635 * which we would have to write lock in order to make user visible
5636 * modifications to the snapshot state (snap trim related mutations
5637 * are not user visible).
5639 if (is_missing_object(clone_oid
)) {
5640 dout(20) << "LIST_SNAPS " << clone_oid
<< " missing" << dendl
;
5641 wait_for_unreadable_object(clone_oid
, ctx
->op
);
5646 ObjectContextRef clone_obc
= get_object_context(clone_oid
, false);
5648 if (maybe_handle_cache(
5649 ctx
->op
, true, clone_obc
, -ENOENT
, clone_oid
, true)) {
5650 // promoting the clone
5653 osd
->clog
->error() << "osd." << osd
->whoami
5654 << ": missing clone " << clone_oid
5657 // should not happen
5662 for (vector
<snapid_t
>::reverse_iterator p
=
5663 clone_obc
->obs
.oi
.legacy_snaps
.rbegin();
5664 p
!= clone_obc
->obs
.oi
.legacy_snaps
.rend();
5666 ci
.snaps
.push_back(*p
);
5670 dout(20) << " clone " << *clone_iter
<< " snaps " << ci
.snaps
<< dendl
;
5672 map
<snapid_t
, interval_set
<uint64_t> >::const_iterator coi
;
5673 coi
= ssc
->snapset
.clone_overlap
.find(ci
.cloneid
);
5674 if (coi
== ssc
->snapset
.clone_overlap
.end()) {
5675 osd
->clog
->error() << "osd." << osd
->whoami
5676 << ": inconsistent clone_overlap found for oid "
5677 << soid
<< " clone " << *clone_iter
;
5681 const interval_set
<uint64_t> &o
= coi
->second
;
5682 ci
.overlap
.reserve(o
.num_intervals());
5683 for (interval_set
<uint64_t>::const_iterator r
= o
.begin();
5684 r
!= o
.end(); ++r
) {
5685 ci
.overlap
.push_back(pair
<uint64_t,uint64_t>(r
.get_start(),
5689 map
<snapid_t
, uint64_t>::const_iterator si
;
5690 si
= ssc
->snapset
.clone_size
.find(ci
.cloneid
);
5691 if (si
== ssc
->snapset
.clone_size
.end()) {
5692 osd
->clog
->error() << "osd." << osd
->whoami
5693 << ": inconsistent clone_size found for oid "
5694 << soid
<< " clone " << *clone_iter
;
5698 ci
.size
= si
->second
;
5700 resp
.clones
.push_back(ci
);
5705 if (ssc
->snapset
.head_exists
&&
5706 !ctx
->obc
->obs
.oi
.is_whiteout()) {
5709 ci
.cloneid
= CEPH_NOSNAP
;
5711 //Size for HEAD is oi.size
5714 resp
.clones
.push_back(ci
);
5716 resp
.seq
= ssc
->snapset
.seq
;
5718 resp
.encode(osd_op
.outdata
);
5721 ctx
->delta_stats
.num_rd
++;
5725 case CEPH_OSD_OP_NOTIFY
:
5732 uint32_t ver
; // obsolete
5734 ::decode(timeout
, bp
);
5736 } catch (const buffer::error
&e
) {
5739 tracepoint(osd
, do_osd_op_pre_notify
, soid
.oid
.name
.c_str(), soid
.snap
.val
, timeout
);
5741 timeout
= cct
->_conf
->osd_default_notify_timeout
;
5744 n
.timeout
= timeout
;
5745 n
.notify_id
= osd
->get_next_id(get_osdmap()->get_epoch());
5746 n
.cookie
= op
.watch
.cookie
;
5748 ctx
->notifies
.push_back(n
);
5750 // return our unique notify id to the client
5751 ::encode(n
.notify_id
, osd_op
.outdata
);
5755 case CEPH_OSD_OP_NOTIFY_ACK
:
5759 uint64_t notify_id
= 0;
5760 uint64_t watch_cookie
= 0;
5761 ::decode(notify_id
, bp
);
5762 ::decode(watch_cookie
, bp
);
5763 bufferlist reply_bl
;
5765 ::decode(reply_bl
, bp
);
5767 tracepoint(osd
, do_osd_op_pre_notify_ack
, soid
.oid
.name
.c_str(), soid
.snap
.val
, notify_id
, watch_cookie
, "Y");
5768 OpContext::NotifyAck
ack(notify_id
, watch_cookie
, reply_bl
);
5769 ctx
->notify_acks
.push_back(ack
);
5770 } catch (const buffer::error
&e
) {
5771 tracepoint(osd
, do_osd_op_pre_notify_ack
, soid
.oid
.name
.c_str(), soid
.snap
.val
, op
.watch
.cookie
, 0, "N");
5772 OpContext::NotifyAck
ack(
5773 // op.watch.cookie is actually the notify_id for historical reasons
5776 ctx
->notify_acks
.push_back(ack
);
5781 case CEPH_OSD_OP_SETALLOCHINT
:
5784 tracepoint(osd
, do_osd_op_pre_setallochint
, soid
.oid
.name
.c_str(), soid
.snap
.val
, op
.alloc_hint
.expected_object_size
, op
.alloc_hint
.expected_write_size
);
5785 maybe_create_new_object(ctx
);
5786 oi
.expected_object_size
= op
.alloc_hint
.expected_object_size
;
5787 oi
.expected_write_size
= op
.alloc_hint
.expected_write_size
;
5788 oi
.alloc_hint_flags
= op
.alloc_hint
.flags
;
5789 t
->set_alloc_hint(soid
, op
.alloc_hint
.expected_object_size
,
5790 op
.alloc_hint
.expected_write_size
,
5791 op
.alloc_hint
.flags
);
5792 ctx
->delta_stats
.num_wr
++;
5800 // -- object data --
5802 case CEPH_OSD_OP_WRITE
:
5805 __u32 seq
= oi
.truncate_seq
;
5806 tracepoint(osd
, do_osd_op_pre_write
, soid
.oid
.name
.c_str(), soid
.snap
.val
, oi
.size
, seq
, op
.extent
.offset
, op
.extent
.length
, op
.extent
.truncate_size
, op
.extent
.truncate_seq
);
5807 if (op
.extent
.length
!= osd_op
.indata
.length()) {
5812 if (pool
.info
.has_flag(pg_pool_t::FLAG_WRITE_FADVISE_DONTNEED
))
5813 op
.flags
= op
.flags
| CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
;
5815 if (pool
.info
.requires_aligned_append() &&
5816 (op
.extent
.offset
% pool
.info
.required_alignment() != 0)) {
5817 result
= -EOPNOTSUPP
;
5822 if (pool
.info
.requires_aligned_append() && op
.extent
.offset
) {
5823 result
= -EOPNOTSUPP
;
5826 } else if (op
.extent
.offset
!= oi
.size
&&
5827 pool
.info
.requires_aligned_append()) {
5828 result
= -EOPNOTSUPP
;
5832 if (seq
&& (seq
> op
.extent
.truncate_seq
) &&
5833 (op
.extent
.offset
+ op
.extent
.length
> oi
.size
)) {
5834 // old write, arrived after trimtrunc
5835 op
.extent
.length
= (op
.extent
.offset
> oi
.size
? 0 : oi
.size
- op
.extent
.offset
);
5836 dout(10) << " old truncate_seq " << op
.extent
.truncate_seq
<< " < current " << seq
5837 << ", adjusting write length to " << op
.extent
.length
<< dendl
;
5839 t
.substr_of(osd_op
.indata
, 0, op
.extent
.length
);
5840 osd_op
.indata
.swap(t
);
5842 if (op
.extent
.truncate_seq
> seq
) {
5843 // write arrives before trimtrunc
5844 if (obs
.exists
&& !oi
.is_whiteout()) {
5845 dout(10) << " truncate_seq " << op
.extent
.truncate_seq
<< " > current " << seq
5846 << ", truncating to " << op
.extent
.truncate_size
<< dendl
;
5847 t
->truncate(soid
, op
.extent
.truncate_size
);
5848 oi
.truncate_seq
= op
.extent
.truncate_seq
;
5849 oi
.truncate_size
= op
.extent
.truncate_size
;
5850 if (op
.extent
.truncate_size
!= oi
.size
) {
5851 ctx
->delta_stats
.num_bytes
-= oi
.size
;
5852 ctx
->delta_stats
.num_bytes
+= op
.extent
.truncate_size
;
5853 oi
.size
= op
.extent
.truncate_size
;
5856 dout(10) << " truncate_seq " << op
.extent
.truncate_seq
<< " > current " << seq
5857 << ", but object is new" << dendl
;
5858 oi
.truncate_seq
= op
.extent
.truncate_seq
;
5859 oi
.truncate_size
= op
.extent
.truncate_size
;
5862 result
= check_offset_and_length(op
.extent
.offset
, op
.extent
.length
, cct
->_conf
->osd_max_object_size
);
5866 maybe_create_new_object(ctx
);
5868 if (op
.extent
.length
== 0) {
5869 if (op
.extent
.offset
> oi
.size
) {
5871 soid
, op
.extent
.offset
);
5877 soid
, op
.extent
.offset
, op
.extent
.length
, osd_op
.indata
, op
.flags
);
5880 if (op
.extent
.offset
== 0 && op
.extent
.length
>= oi
.size
5881 && !skip_data_digest
) {
5882 obs
.oi
.set_data_digest(osd_op
.indata
.crc32c(-1));
5883 } else if (op
.extent
.offset
== oi
.size
&& obs
.oi
.is_data_digest()) {
5884 if (skip_data_digest
) {
5885 obs
.oi
.clear_data_digest();
5887 obs
.oi
.set_data_digest(osd_op
.indata
.crc32c(obs
.oi
.data_digest
));
5890 obs
.oi
.clear_data_digest();
5892 write_update_size_and_usage(ctx
->delta_stats
, oi
, ctx
->modified_ranges
,
5893 op
.extent
.offset
, op
.extent
.length
);
5898 case CEPH_OSD_OP_WRITEFULL
:
5900 { // write full object
5901 tracepoint(osd
, do_osd_op_pre_writefull
, soid
.oid
.name
.c_str(), soid
.snap
.val
, oi
.size
, 0, op
.extent
.length
);
5903 if (op
.extent
.length
!= osd_op
.indata
.length()) {
5907 result
= check_offset_and_length(0, op
.extent
.length
, cct
->_conf
->osd_max_object_size
);
5911 if (pool
.info
.has_flag(pg_pool_t::FLAG_WRITE_FADVISE_DONTNEED
))
5912 op
.flags
= op
.flags
| CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
;
5914 maybe_create_new_object(ctx
);
5915 if (pool
.info
.require_rollback()) {
5916 t
->truncate(soid
, 0);
5917 } else if (obs
.exists
&& op
.extent
.length
< oi
.size
) {
5918 t
->truncate(soid
, op
.extent
.length
);
5920 if (op
.extent
.length
) {
5921 t
->write(soid
, 0, op
.extent
.length
, osd_op
.indata
, op
.flags
);
5923 if (!skip_data_digest
) {
5924 obs
.oi
.set_data_digest(osd_op
.indata
.crc32c(-1));
5926 obs
.oi
.clear_data_digest();
5929 write_update_size_and_usage(ctx
->delta_stats
, oi
, ctx
->modified_ranges
,
5930 0, op
.extent
.length
, true);
5934 case CEPH_OSD_OP_WRITESAME
:
5936 tracepoint(osd
, do_osd_op_pre_writesame
, soid
.oid
.name
.c_str(), soid
.snap
.val
, oi
.size
, op
.writesame
.offset
, op
.writesame
.length
, op
.writesame
.data_length
);
5937 result
= do_writesame(ctx
, osd_op
);
5940 case CEPH_OSD_OP_ROLLBACK
:
5942 tracepoint(osd
, do_osd_op_pre_rollback
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5943 result
= _rollback_to(ctx
, op
);
5946 case CEPH_OSD_OP_ZERO
:
5947 tracepoint(osd
, do_osd_op_pre_zero
, soid
.oid
.name
.c_str(), soid
.snap
.val
, op
.extent
.offset
, op
.extent
.length
);
5948 if (pool
.info
.requires_aligned_append()) {
5949 result
= -EOPNOTSUPP
;
5954 result
= check_offset_and_length(op
.extent
.offset
, op
.extent
.length
, cct
->_conf
->osd_max_object_size
);
5957 assert(op
.extent
.length
);
5958 if (obs
.exists
&& !oi
.is_whiteout()) {
5959 t
->zero(soid
, op
.extent
.offset
, op
.extent
.length
);
5960 interval_set
<uint64_t> ch
;
5961 ch
.insert(op
.extent
.offset
, op
.extent
.length
);
5962 ctx
->modified_ranges
.union_of(ch
);
5963 ctx
->delta_stats
.num_wr
++;
5964 oi
.clear_data_digest();
5970 case CEPH_OSD_OP_CREATE
:
5973 tracepoint(osd
, do_osd_op_pre_create
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
5974 int flags
= le32_to_cpu(op
.flags
);
5975 if (obs
.exists
&& !oi
.is_whiteout() &&
5976 (flags
& CEPH_OSD_OP_FLAG_EXCL
)) {
5977 result
= -EEXIST
; /* this is an exclusive create */
5979 if (osd_op
.indata
.length()) {
5980 bufferlist::iterator p
= osd_op
.indata
.begin();
5983 ::decode(category
, p
);
5985 catch (buffer::error
& e
) {
5989 // category is no longer implemented.
5992 maybe_create_new_object(ctx
);
5999 case CEPH_OSD_OP_TRIMTRUNC
:
6000 op
.extent
.offset
= op
.extent
.truncate_size
;
6003 case CEPH_OSD_OP_TRUNCATE
:
6004 tracepoint(osd
, do_osd_op_pre_truncate
, soid
.oid
.name
.c_str(), soid
.snap
.val
, oi
.size
, oi
.truncate_seq
, op
.extent
.offset
, op
.extent
.length
, op
.extent
.truncate_size
, op
.extent
.truncate_seq
);
6005 if (pool
.info
.requires_aligned_append()) {
6006 result
= -EOPNOTSUPP
;
6012 if (!obs
.exists
|| oi
.is_whiteout()) {
6013 dout(10) << " object dne, truncate is a no-op" << dendl
;
6017 if (op
.extent
.offset
> cct
->_conf
->osd_max_object_size
) {
6022 if (op
.extent
.truncate_seq
) {
6023 assert(op
.extent
.offset
== op
.extent
.truncate_size
);
6024 if (op
.extent
.truncate_seq
<= oi
.truncate_seq
) {
6025 dout(10) << " truncate seq " << op
.extent
.truncate_seq
<< " <= current " << oi
.truncate_seq
6026 << ", no-op" << dendl
;
6029 dout(10) << " truncate seq " << op
.extent
.truncate_seq
<< " > current " << oi
.truncate_seq
6030 << ", truncating" << dendl
;
6031 oi
.truncate_seq
= op
.extent
.truncate_seq
;
6032 oi
.truncate_size
= op
.extent
.truncate_size
;
6035 maybe_create_new_object(ctx
);
6036 t
->truncate(soid
, op
.extent
.offset
);
6037 if (oi
.size
> op
.extent
.offset
) {
6038 interval_set
<uint64_t> trim
;
6039 trim
.insert(op
.extent
.offset
, oi
.size
-op
.extent
.offset
);
6040 ctx
->modified_ranges
.union_of(trim
);
6042 if (op
.extent
.offset
!= oi
.size
) {
6043 ctx
->delta_stats
.num_bytes
-= oi
.size
;
6044 ctx
->delta_stats
.num_bytes
+= op
.extent
.offset
;
6045 oi
.size
= op
.extent
.offset
;
6047 ctx
->delta_stats
.num_wr
++;
6048 // do no set exists, or we will break above DELETE -> TRUNCATE munging.
6050 oi
.clear_data_digest();
6054 case CEPH_OSD_OP_DELETE
:
6056 tracepoint(osd
, do_osd_op_pre_delete
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6058 result
= _delete_oid(ctx
, false, ctx
->ignore_cache
);
6062 case CEPH_OSD_OP_WATCH
:
6065 tracepoint(osd
, do_osd_op_pre_watch
, soid
.oid
.name
.c_str(), soid
.snap
.val
,
6066 op
.watch
.cookie
, op
.watch
.op
);
6071 uint64_t cookie
= op
.watch
.cookie
;
6072 entity_name_t entity
= ctx
->reqid
.name
;
6073 ObjectContextRef obc
= ctx
->obc
;
6075 dout(10) << "watch " << ceph_osd_watch_op_name(op
.watch
.op
)
6076 << ": ctx->obc=" << (void *)obc
.get() << " cookie=" << cookie
6077 << " oi.version=" << oi
.version
.version
<< " ctx->at_version=" << ctx
->at_version
<< dendl
;
6078 dout(10) << "watch: oi.user_version=" << oi
.user_version
<< dendl
;
6079 dout(10) << "watch: peer_addr="
6080 << ctx
->op
->get_req()->get_connection()->get_peer_addr() << dendl
;
6082 uint32_t timeout
= cct
->_conf
->osd_client_watch_timeout
;
6083 if (op
.watch
.timeout
!= 0) {
6084 timeout
= op
.watch
.timeout
;
6087 watch_info_t
w(cookie
, timeout
,
6088 ctx
->op
->get_req()->get_connection()->get_peer_addr());
6089 if (op
.watch
.op
== CEPH_OSD_WATCH_OP_WATCH
||
6090 op
.watch
.op
== CEPH_OSD_WATCH_OP_LEGACY_WATCH
) {
6091 if (oi
.watchers
.count(make_pair(cookie
, entity
))) {
6092 dout(10) << " found existing watch " << w
<< " by " << entity
<< dendl
;
6094 dout(10) << " registered new watch " << w
<< " by " << entity
<< dendl
;
6095 oi
.watchers
[make_pair(cookie
, entity
)] = w
;
6096 t
->nop(soid
); // make sure update the object_info on disk!
6098 bool will_ping
= (op
.watch
.op
== CEPH_OSD_WATCH_OP_WATCH
);
6099 ctx
->watch_connects
.push_back(make_pair(w
, will_ping
));
6100 } else if (op
.watch
.op
== CEPH_OSD_WATCH_OP_RECONNECT
) {
6101 if (!oi
.watchers
.count(make_pair(cookie
, entity
))) {
6105 dout(10) << " found existing watch " << w
<< " by " << entity
<< dendl
;
6106 ctx
->watch_connects
.push_back(make_pair(w
, true));
6107 } else if (op
.watch
.op
== CEPH_OSD_WATCH_OP_PING
) {
6108 /* Note: WATCH with PING doesn't cause may_write() to return true,
6109 * so if there is nothing else in the transaction, this is going
6110 * to run do_osd_op_effects, but not write out a log entry */
6111 if (!oi
.watchers
.count(make_pair(cookie
, entity
))) {
6115 map
<pair
<uint64_t,entity_name_t
>,WatchRef
>::iterator p
=
6116 obc
->watchers
.find(make_pair(cookie
, entity
));
6117 if (p
== obc
->watchers
.end() ||
6118 !p
->second
->is_connected()) {
6119 // client needs to reconnect
6120 result
= -ETIMEDOUT
;
6123 dout(10) << " found existing watch " << w
<< " by " << entity
<< dendl
;
6124 p
->second
->got_ping(ceph_clock_now());
6126 } else if (op
.watch
.op
== CEPH_OSD_WATCH_OP_UNWATCH
) {
6127 map
<pair
<uint64_t, entity_name_t
>, watch_info_t
>::iterator oi_iter
=
6128 oi
.watchers
.find(make_pair(cookie
, entity
));
6129 if (oi_iter
!= oi
.watchers
.end()) {
6130 dout(10) << " removed watch " << oi_iter
->second
<< " by "
6132 oi
.watchers
.erase(oi_iter
);
6133 t
->nop(soid
); // update oi on disk
6134 ctx
->watch_disconnects
.push_back(
6135 watch_disconnect_t(cookie
, entity
, false));
6137 dout(10) << " can't remove: no watch by " << entity
<< dendl
;
6143 case CEPH_OSD_OP_CACHE_PIN
:
6144 tracepoint(osd
, do_osd_op_pre_cache_pin
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6145 if ((!pool
.info
.is_tier() ||
6146 pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_NONE
)) {
6148 dout(10) << " pin object is only allowed on the cache tier " << dendl
;
6153 if (!obs
.exists
|| oi
.is_whiteout()) {
6158 if (!oi
.is_cache_pinned()) {
6159 oi
.set_flag(object_info_t::FLAG_CACHE_PIN
);
6161 ctx
->delta_stats
.num_objects_pinned
++;
6162 ctx
->delta_stats
.num_wr
++;
6168 case CEPH_OSD_OP_CACHE_UNPIN
:
6169 tracepoint(osd
, do_osd_op_pre_cache_unpin
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6170 if ((!pool
.info
.is_tier() ||
6171 pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_NONE
)) {
6173 dout(10) << " pin object is only allowed on the cache tier " << dendl
;
6178 if (!obs
.exists
|| oi
.is_whiteout()) {
6183 if (oi
.is_cache_pinned()) {
6184 oi
.clear_flag(object_info_t::FLAG_CACHE_PIN
);
6186 ctx
->delta_stats
.num_objects_pinned
--;
6187 ctx
->delta_stats
.num_wr
++;
6193 case CEPH_OSD_OP_SET_REDIRECT
:
6196 if (pool
.info
.is_tier()) {
6204 if (get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
6205 result
= -EOPNOTSUPP
;
6209 object_t target_name
;
6210 object_locator_t target_oloc
;
6211 snapid_t target_snapid
= (uint64_t)op
.copy_from
.snapid
;
6212 version_t target_version
= op
.copy_from
.src_version
;
6214 ::decode(target_name
, bp
);
6215 ::decode(target_oloc
, bp
);
6217 catch (buffer::error
& e
) {
6222 get_osdmap()->object_locator_to_pg(target_name
, target_oloc
, raw_pg
);
6223 hobject_t
target(target_name
, target_oloc
.key
, target_snapid
,
6224 raw_pg
.ps(), raw_pg
.pool(),
6225 target_oloc
.nspace
);
6226 if (target
== soid
) {
6227 dout(20) << " set-redirect self is invalid" << dendl
;
6231 oi
.set_flag(object_info_t::FLAG_MANIFEST
);
6232 oi
.manifest
.redirect_target
= target
;
6233 oi
.manifest
.type
= object_manifest_t::TYPE_REDIRECT
;
6234 t
->truncate(soid
, 0);
6235 if (oi
.is_omap() && pool
.info
.supports_omap()) {
6236 t
->omap_clear(soid
);
6237 obs
.oi
.clear_omap_digest();
6238 obs
.oi
.clear_flag(object_info_t::FLAG_OMAP
);
6240 ctx
->delta_stats
.num_bytes
-= oi
.size
;
6243 oi
.user_version
= target_version
;
6244 ctx
->user_at_version
= target_version
;
6246 map
<string
,bufferlist
> rmattrs
;
6247 result
= getattrs_maybe_cache(ctx
->obc
,
6252 map
<string
, bufferlist
>::iterator iter
;
6253 for (iter
= rmattrs
.begin(); iter
!= rmattrs
.end(); ++iter
) {
6254 const string
& name
= iter
->first
;
6255 t
->rmattr(soid
, name
);
6257 dout(10) << "set-redirect oid:" << oi
.soid
<< " user_version: " << oi
.user_version
<< dendl
;
6262 // -- object attrs --
6264 case CEPH_OSD_OP_SETXATTR
:
6267 if (cct
->_conf
->osd_max_attr_size
> 0 &&
6268 op
.xattr
.value_len
> cct
->_conf
->osd_max_attr_size
) {
6269 tracepoint(osd
, do_osd_op_pre_setxattr
, soid
.oid
.name
.c_str(), soid
.snap
.val
, "???");
6273 unsigned max_name_len
= MIN(osd
->store
->get_max_attr_name_length(),
6274 cct
->_conf
->osd_max_attr_name_len
);
6275 if (op
.xattr
.name_len
> max_name_len
) {
6276 result
= -ENAMETOOLONG
;
6279 maybe_create_new_object(ctx
);
6281 bp
.copy(op
.xattr
.name_len
, aname
);
6282 tracepoint(osd
, do_osd_op_pre_setxattr
, soid
.oid
.name
.c_str(), soid
.snap
.val
, aname
.c_str());
6283 string name
= "_" + aname
;
6285 bp
.copy(op
.xattr
.value_len
, bl
);
6286 t
->setattr(soid
, name
, bl
);
6287 ctx
->delta_stats
.num_wr
++;
6291 case CEPH_OSD_OP_RMXATTR
:
6295 bp
.copy(op
.xattr
.name_len
, aname
);
6296 tracepoint(osd
, do_osd_op_pre_rmxattr
, soid
.oid
.name
.c_str(), soid
.snap
.val
, aname
.c_str());
6297 if (!obs
.exists
|| oi
.is_whiteout()) {
6301 string name
= "_" + aname
;
6302 t
->rmattr(soid
, name
);
6303 ctx
->delta_stats
.num_wr
++;
6308 // -- fancy writers --
6309 case CEPH_OSD_OP_APPEND
:
6311 tracepoint(osd
, do_osd_op_pre_append
, soid
.oid
.name
.c_str(), soid
.snap
.val
, oi
.size
, oi
.truncate_seq
, op
.extent
.offset
, op
.extent
.length
, op
.extent
.truncate_size
, op
.extent
.truncate_seq
);
6312 // just do it inline; this works because we are happy to execute
6313 // fancy op on replicas as well.
6314 vector
<OSDOp
> nops(1);
6315 OSDOp
& newop
= nops
[0];
6316 newop
.op
.op
= CEPH_OSD_OP_WRITE
;
6317 newop
.op
.extent
.offset
= oi
.size
;
6318 newop
.op
.extent
.length
= op
.extent
.length
;
6319 newop
.op
.extent
.truncate_seq
= oi
.truncate_seq
;
6320 newop
.indata
= osd_op
.indata
;
6321 result
= do_osd_ops(ctx
, nops
);
6322 osd_op
.outdata
.claim(newop
.outdata
);
6326 case CEPH_OSD_OP_STARTSYNC
:
6327 tracepoint(osd
, do_osd_op_pre_startsync
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6332 // -- trivial map --
6333 case CEPH_OSD_OP_TMAPGET
:
6334 tracepoint(osd
, do_osd_op_pre_tmapget
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6335 if (pool
.info
.require_rollback()) {
6336 result
= -EOPNOTSUPP
;
6340 vector
<OSDOp
> nops(1);
6341 OSDOp
& newop
= nops
[0];
6342 newop
.op
.op
= CEPH_OSD_OP_SYNC_READ
;
6343 newop
.op
.extent
.offset
= 0;
6344 newop
.op
.extent
.length
= 0;
6345 do_osd_ops(ctx
, nops
);
6346 osd_op
.outdata
.claim(newop
.outdata
);
6350 case CEPH_OSD_OP_TMAPPUT
:
6351 tracepoint(osd
, do_osd_op_pre_tmapput
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6352 if (pool
.info
.require_rollback()) {
6353 result
= -EOPNOTSUPP
;
6357 //_dout_lock.Lock();
6358 //osd_op.data.hexdump(*_dout);
6359 //_dout_lock.Unlock();
6361 // verify sort order
6362 bool unsorted
= false;
6365 ::decode(header
, bp
);
6372 dout(10) << "tmapput key " << key
<< dendl
;
6375 if (key
< last_key
) {
6376 dout(10) << "TMAPPUT is unordered; resorting" << dendl
;
6385 vector
<OSDOp
> nops(1);
6386 OSDOp
& newop
= nops
[0];
6387 newop
.op
.op
= CEPH_OSD_OP_WRITEFULL
;
6388 newop
.op
.extent
.offset
= 0;
6389 newop
.op
.extent
.length
= osd_op
.indata
.length();
6390 newop
.indata
= osd_op
.indata
;
6393 bp
= osd_op
.indata
.begin();
6395 map
<string
, bufferlist
> m
;
6396 ::decode(header
, bp
);
6400 ::encode(header
, newbl
);
6402 newop
.indata
= newbl
;
6404 result
= do_osd_ops(ctx
, nops
);
6405 assert(result
== 0);
6409 case CEPH_OSD_OP_TMAPUP
:
6410 tracepoint(osd
, do_osd_op_pre_tmapup
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6411 if (pool
.info
.require_rollback()) {
6412 result
= -EOPNOTSUPP
;
6416 result
= do_tmapup(ctx
, bp
, osd_op
);
6419 case CEPH_OSD_OP_TMAP2OMAP
:
6421 tracepoint(osd
, do_osd_op_pre_tmap2omap
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6422 result
= do_tmap2omap(ctx
, op
.tmap2omap
.flags
);
6426 case CEPH_OSD_OP_OMAPGETKEYS
:
6430 uint64_t max_return
;
6432 ::decode(start_after
, bp
);
6433 ::decode(max_return
, bp
);
6435 catch (buffer::error
& e
) {
6437 tracepoint(osd
, do_osd_op_pre_omapgetkeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
, "???", 0);
6440 if (max_return
> cct
->_conf
->osd_max_omap_entries_per_request
) {
6441 max_return
= cct
->_conf
->osd_max_omap_entries_per_request
;
6443 tracepoint(osd
, do_osd_op_pre_omapgetkeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
, start_after
.c_str(), max_return
);
6447 bool truncated
= false;
6449 ObjectMap::ObjectMapIterator iter
= osd
->store
->get_omap_iterator(
6450 coll
, ghobject_t(soid
)
6453 iter
->upper_bound(start_after
);
6454 for (num
= 0; iter
->valid(); ++num
, iter
->next(false)) {
6455 if (num
>= max_return
||
6456 bl
.length() >= cct
->_conf
->osd_max_omap_bytes_per_request
) {
6460 ::encode(iter
->key(), bl
);
6462 } // else return empty out_set
6463 ::encode(num
, osd_op
.outdata
);
6464 osd_op
.outdata
.claim_append(bl
);
6465 ::encode(truncated
, osd_op
.outdata
);
6466 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(osd_op
.outdata
.length(), 10);
6467 ctx
->delta_stats
.num_rd
++;
6471 case CEPH_OSD_OP_OMAPGETVALS
:
6475 uint64_t max_return
;
6476 string filter_prefix
;
6478 ::decode(start_after
, bp
);
6479 ::decode(max_return
, bp
);
6480 ::decode(filter_prefix
, bp
);
6482 catch (buffer::error
& e
) {
6484 tracepoint(osd
, do_osd_op_pre_omapgetvals
, soid
.oid
.name
.c_str(), soid
.snap
.val
, "???", 0, "???");
6487 if (max_return
> cct
->_conf
->osd_max_omap_entries_per_request
) {
6488 max_return
= cct
->_conf
->osd_max_omap_entries_per_request
;
6490 tracepoint(osd
, do_osd_op_pre_omapgetvals
, soid
.oid
.name
.c_str(), soid
.snap
.val
, start_after
.c_str(), max_return
, filter_prefix
.c_str());
6493 bool truncated
= false;
6496 ObjectMap::ObjectMapIterator iter
= osd
->store
->get_omap_iterator(
6497 coll
, ghobject_t(soid
)
6503 iter
->upper_bound(start_after
);
6504 if (filter_prefix
> start_after
) iter
->lower_bound(filter_prefix
);
6507 iter
->key().substr(0, filter_prefix
.size()) == filter_prefix
;
6508 ++num
, iter
->next(false)) {
6509 dout(20) << "Found key " << iter
->key() << dendl
;
6510 if (num
>= max_return
||
6511 bl
.length() >= cct
->_conf
->osd_max_omap_bytes_per_request
) {
6515 ::encode(iter
->key(), bl
);
6516 ::encode(iter
->value(), bl
);
6518 } // else return empty out_set
6519 ::encode(num
, osd_op
.outdata
);
6520 osd_op
.outdata
.claim_append(bl
);
6521 ::encode(truncated
, osd_op
.outdata
);
6522 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(osd_op
.outdata
.length(), 10);
6523 ctx
->delta_stats
.num_rd
++;
6527 case CEPH_OSD_OP_OMAPGETHEADER
:
6528 tracepoint(osd
, do_osd_op_pre_omapgetheader
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6529 if (!oi
.is_omap()) {
6530 // return empty header
6535 osd
->store
->omap_get_header(ch
, ghobject_t(soid
), &osd_op
.outdata
);
6536 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(osd_op
.outdata
.length(), 10);
6537 ctx
->delta_stats
.num_rd
++;
6541 case CEPH_OSD_OP_OMAPGETVALSBYKEYS
:
6544 set
<string
> keys_to_get
;
6546 ::decode(keys_to_get
, bp
);
6548 catch (buffer::error
& e
) {
6550 tracepoint(osd
, do_osd_op_pre_omapgetvalsbykeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
, "???");
6553 tracepoint(osd
, do_osd_op_pre_omapgetvalsbykeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
, list_entries(keys_to_get
).c_str());
6554 map
<string
, bufferlist
> out
;
6556 osd
->store
->omap_get_values(ch
, ghobject_t(soid
), keys_to_get
, &out
);
6557 } // else return empty omap entries
6558 ::encode(out
, osd_op
.outdata
);
6559 ctx
->delta_stats
.num_rd_kb
+= SHIFT_ROUND_UP(osd_op
.outdata
.length(), 10);
6560 ctx
->delta_stats
.num_rd
++;
6564 case CEPH_OSD_OP_OMAP_CMP
:
6567 if (!obs
.exists
|| oi
.is_whiteout()) {
6569 tracepoint(osd
, do_osd_op_pre_omap_cmp
, soid
.oid
.name
.c_str(), soid
.snap
.val
, "???");
6572 map
<string
, pair
<bufferlist
, int> > assertions
;
6574 ::decode(assertions
, bp
);
6576 catch (buffer::error
& e
) {
6578 tracepoint(osd
, do_osd_op_pre_omap_cmp
, soid
.oid
.name
.c_str(), soid
.snap
.val
, "???");
6581 tracepoint(osd
, do_osd_op_pre_omap_cmp
, soid
.oid
.name
.c_str(), soid
.snap
.val
, list_keys(assertions
).c_str());
6583 map
<string
, bufferlist
> out
;
6587 for (map
<string
, pair
<bufferlist
, int> >::iterator i
= assertions
.begin();
6588 i
!= assertions
.end();
6590 to_get
.insert(i
->first
);
6591 int r
= osd
->store
->omap_get_values(ch
, ghobject_t(soid
),
6597 } // else leave out empty
6599 //Should set num_rd_kb based on encode length of map
6600 ctx
->delta_stats
.num_rd
++;
6604 for (map
<string
, pair
<bufferlist
, int> >::iterator i
= assertions
.begin();
6605 i
!= assertions
.end();
6607 auto out_entry
= out
.find(i
->first
);
6608 bufferlist
&bl
= (out_entry
!= out
.end()) ?
6609 out_entry
->second
: empty
;
6610 switch (i
->second
.second
) {
6611 case CEPH_OSD_CMPXATTR_OP_EQ
:
6612 if (!(bl
== i
->second
.first
)) {
6616 case CEPH_OSD_CMPXATTR_OP_LT
:
6617 if (!(bl
< i
->second
.first
)) {
6621 case CEPH_OSD_CMPXATTR_OP_GT
:
6622 if (!(bl
> i
->second
.first
)) {
6640 case CEPH_OSD_OP_OMAPSETVALS
:
6641 if (!pool
.info
.supports_omap()) {
6642 result
= -EOPNOTSUPP
;
6643 tracepoint(osd
, do_osd_op_pre_omapsetvals
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6648 maybe_create_new_object(ctx
);
6649 bufferlist to_set_bl
;
6651 decode_str_str_map_to_bl(bp
, &to_set_bl
);
6653 catch (buffer::error
& e
) {
6655 tracepoint(osd
, do_osd_op_pre_omapsetvals
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6658 tracepoint(osd
, do_osd_op_pre_omapsetvals
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6659 if (cct
->_conf
->subsys
.should_gather(dout_subsys
, 20)) {
6660 dout(20) << "setting vals: " << dendl
;
6661 map
<string
,bufferlist
> to_set
;
6662 bufferlist::iterator pt
= to_set_bl
.begin();
6663 ::decode(to_set
, pt
);
6664 for (map
<string
, bufferlist
>::iterator i
= to_set
.begin();
6667 dout(20) << "\t" << i
->first
<< dendl
;
6670 t
->omap_setkeys(soid
, to_set_bl
);
6671 ctx
->delta_stats
.num_wr
++;
6673 obs
.oi
.set_flag(object_info_t::FLAG_OMAP
);
6674 obs
.oi
.clear_omap_digest();
6677 case CEPH_OSD_OP_OMAPSETHEADER
:
6678 tracepoint(osd
, do_osd_op_pre_omapsetheader
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6679 if (!pool
.info
.supports_omap()) {
6680 result
= -EOPNOTSUPP
;
6685 maybe_create_new_object(ctx
);
6686 t
->omap_setheader(soid
, osd_op
.indata
);
6687 ctx
->delta_stats
.num_wr
++;
6689 obs
.oi
.set_flag(object_info_t::FLAG_OMAP
);
6690 obs
.oi
.clear_omap_digest();
6693 case CEPH_OSD_OP_OMAPCLEAR
:
6694 tracepoint(osd
, do_osd_op_pre_omapclear
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6695 if (!pool
.info
.supports_omap()) {
6696 result
= -EOPNOTSUPP
;
6701 if (!obs
.exists
|| oi
.is_whiteout()) {
6706 t
->omap_clear(soid
);
6707 ctx
->delta_stats
.num_wr
++;
6708 obs
.oi
.clear_omap_digest();
6709 obs
.oi
.clear_flag(object_info_t::FLAG_OMAP
);
6714 case CEPH_OSD_OP_OMAPRMKEYS
:
6715 if (!pool
.info
.supports_omap()) {
6716 result
= -EOPNOTSUPP
;
6717 tracepoint(osd
, do_osd_op_pre_omaprmkeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6722 if (!obs
.exists
|| oi
.is_whiteout()) {
6724 tracepoint(osd
, do_osd_op_pre_omaprmkeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6727 bufferlist to_rm_bl
;
6729 decode_str_set_to_bl(bp
, &to_rm_bl
);
6731 catch (buffer::error
& e
) {
6733 tracepoint(osd
, do_osd_op_pre_omaprmkeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6736 tracepoint(osd
, do_osd_op_pre_omaprmkeys
, soid
.oid
.name
.c_str(), soid
.snap
.val
);
6737 t
->omap_rmkeys(soid
, to_rm_bl
);
6738 ctx
->delta_stats
.num_wr
++;
6740 obs
.oi
.clear_omap_digest();
6743 case CEPH_OSD_OP_COPY_GET
:
6745 tracepoint(osd
, do_osd_op_pre_copy_get
, soid
.oid
.name
.c_str(),
6747 if (op_finisher
== nullptr) {
6748 result
= do_copy_get(ctx
, bp
, osd_op
, ctx
->obc
);
6750 result
= op_finisher
->execute();
6754 case CEPH_OSD_OP_COPY_FROM
:
6758 object_locator_t src_oloc
;
6759 snapid_t src_snapid
= (uint64_t)op
.copy_from
.snapid
;
6760 version_t src_version
= op
.copy_from
.src_version
;
6762 ::decode(src_name
, bp
);
6763 ::decode(src_oloc
, bp
);
6765 catch (buffer::error
& e
) {
6768 do_osd_op_pre_copy_from
,
6769 soid
.oid
.name
.c_str(),
6781 do_osd_op_pre_copy_from
,
6782 soid
.oid
.name
.c_str(),
6784 src_name
.name
.c_str(),
6786 src_oloc
.key
.c_str(),
6787 src_oloc
.nspace
.c_str(),
6791 if (op_finisher
== nullptr) {
6794 get_osdmap()->object_locator_to_pg(src_name
, src_oloc
, raw_pg
);
6795 hobject_t
src(src_name
, src_oloc
.key
, src_snapid
,
6796 raw_pg
.ps(), raw_pg
.pool(),
6799 dout(20) << " copy from self is invalid" << dendl
;
6803 CopyFromCallback
*cb
= new CopyFromCallback(ctx
, osd_op
);
6804 ctx
->op_finishers
[ctx
->current_osd_subop_num
].reset(
6805 new CopyFromFinisher(cb
));
6806 start_copy(cb
, ctx
->obc
, src
, src_oloc
, src_version
,
6809 op
.copy_from
.src_fadvise_flags
,
6811 result
= -EINPROGRESS
;
6814 result
= op_finisher
->execute();
6815 assert(result
== 0);
6817 // COPY_FROM cannot be executed multiple times -- it must restart
6818 ctx
->op_finishers
.erase(ctx
->current_osd_subop_num
);
6824 tracepoint(osd
, do_osd_op_pre_unknown
, soid
.oid
.name
.c_str(), soid
.snap
.val
, op
.op
, ceph_osd_op_name(op
.op
));
6825 dout(1) << "unrecognized osd op " << op
.op
6826 << " " << ceph_osd_op_name(op
.op
)
6828 result
= -EOPNOTSUPP
;
6832 osd_op
.rval
= result
;
6833 tracepoint(osd
, do_osd_op_post
, soid
.oid
.name
.c_str(), soid
.snap
.val
, op
.op
, ceph_osd_op_name(op
.op
), op
.flags
, result
);
6834 if (result
< 0 && (op
.flags
& CEPH_OSD_OP_FLAG_FAILOK
))
6843 int PrimaryLogPG::_get_tmap(OpContext
*ctx
, bufferlist
*header
, bufferlist
*vals
)
6845 if (ctx
->new_obs
.oi
.size
== 0) {
6846 dout(20) << "unable to get tmap for zero sized " << ctx
->new_obs
.oi
.soid
<< dendl
;
6849 vector
<OSDOp
> nops(1);
6850 OSDOp
&newop
= nops
[0];
6851 newop
.op
.op
= CEPH_OSD_OP_TMAPGET
;
6852 do_osd_ops(ctx
, nops
);
6854 bufferlist::iterator i
= newop
.outdata
.begin();
6855 ::decode(*header
, i
);
6856 (*vals
).substr_of(newop
.outdata
, i
.get_off(), i
.get_remaining());
6858 dout(20) << "unsuccessful at decoding tmap for " << ctx
->new_obs
.oi
.soid
6862 dout(20) << "successful at decoding tmap for " << ctx
->new_obs
.oi
.soid
6867 int PrimaryLogPG::_verify_no_head_clones(const hobject_t
& soid
,
6870 // verify that all clones have been evicted
6871 dout(20) << __func__
<< " verifying clones are absent "
6873 for (vector
<snapid_t
>::const_iterator p
= ss
.clones
.begin();
6874 p
!= ss
.clones
.end();
6876 hobject_t clone_oid
= soid
;
6877 clone_oid
.snap
= *p
;
6878 if (is_missing_object(clone_oid
))
6880 ObjectContextRef clone_obc
= get_object_context(clone_oid
, false);
6881 if (clone_obc
&& clone_obc
->obs
.exists
) {
6882 dout(10) << __func__
<< " cannot evict head before clone "
6883 << clone_oid
<< dendl
;
6886 if (copy_ops
.count(clone_oid
)) {
6887 dout(10) << __func__
<< " cannot evict head, pending promote on clone "
6888 << clone_oid
<< dendl
;
6895 inline int PrimaryLogPG::_delete_oid(
6897 bool no_whiteout
, // no whiteouts, no matter what.
6898 bool try_no_whiteout
) // try not to whiteout
6900 SnapSet
& snapset
= ctx
->new_snapset
;
6901 ObjectState
& obs
= ctx
->new_obs
;
6902 object_info_t
& oi
= obs
.oi
;
6903 const hobject_t
& soid
= oi
.soid
;
6904 PGTransaction
* t
= ctx
->op_t
.get();
6906 // cache: cache: set whiteout on delete?
6907 bool whiteout
= false;
6908 if (pool
.info
.cache_mode
!= pg_pool_t::CACHEMODE_NONE
6910 && !try_no_whiteout
) {
6914 if (get_osdmap()->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
6916 // in luminous or later, we can't delete the head if there are
6917 // clones. we trust the caller passing no_whiteout has already
6918 // verified they don't exist.
6919 if (!snapset
.clones
.empty() ||
6920 (!ctx
->snapc
.snaps
.empty() && ctx
->snapc
.snaps
[0] > snapset
.seq
)) {
6922 dout(20) << __func__
<< " has or will have clones but no_whiteout=1"
6925 dout(20) << __func__
<< " has or will have clones; will whiteout"
6933 dout(20) << __func__
<< " " << soid
<< " whiteout=" << (int)whiteout
6934 << " no_whiteout=" << (int)no_whiteout
6935 << " try_no_whiteout=" << (int)try_no_whiteout
6937 if (!obs
.exists
|| (obs
.oi
.is_whiteout() && whiteout
))
6943 interval_set
<uint64_t> ch
;
6944 ch
.insert(0, oi
.size
);
6945 ctx
->modified_ranges
.union_of(ch
);
6948 ctx
->delta_stats
.num_wr
++;
6949 if (soid
.is_snap()) {
6950 assert(ctx
->obc
->ssc
->snapset
.clone_overlap
.count(soid
.snap
));
6951 ctx
->delta_stats
.num_bytes
-= ctx
->obc
->ssc
->snapset
.get_clone_bytes(soid
.snap
);
6953 ctx
->delta_stats
.num_bytes
-= oi
.size
;
6958 // disconnect all watchers
6959 for (map
<pair
<uint64_t, entity_name_t
>, watch_info_t
>::iterator p
=
6960 oi
.watchers
.begin();
6961 p
!= oi
.watchers
.end();
6963 dout(20) << __func__
<< " will disconnect watcher " << p
->first
<< dendl
;
6964 ctx
->watch_disconnects
.push_back(
6965 watch_disconnect_t(p
->first
.first
, p
->first
.second
, true));
6967 oi
.watchers
.clear();
6970 dout(20) << __func__
<< " setting whiteout on " << soid
<< dendl
;
6971 oi
.set_flag(object_info_t::FLAG_WHITEOUT
);
6972 ctx
->delta_stats
.num_whiteouts
++;
6974 osd
->logger
->inc(l_osd_tier_whiteout
);
6979 ctx
->delta_stats
.num_objects
--;
6981 ctx
->delta_stats
.num_object_clones
--;
6982 if (oi
.is_whiteout()) {
6983 dout(20) << __func__
<< " deleting whiteout on " << soid
<< dendl
;
6984 ctx
->delta_stats
.num_whiteouts
--;
6985 oi
.clear_flag(object_info_t::FLAG_WHITEOUT
);
6987 if (oi
.is_cache_pinned()) {
6988 ctx
->delta_stats
.num_objects_pinned
--;
6990 if ((legacy
|| snapset
.is_legacy()) && soid
.is_head()) {
6991 snapset
.head_exists
= false;
6997 int PrimaryLogPG::_rollback_to(OpContext
*ctx
, ceph_osd_op
& op
)
6999 SnapSet
& snapset
= ctx
->new_snapset
;
7000 ObjectState
& obs
= ctx
->new_obs
;
7001 object_info_t
& oi
= obs
.oi
;
7002 const hobject_t
& soid
= oi
.soid
;
7003 PGTransaction
* t
= ctx
->op_t
.get();
7004 snapid_t snapid
= (uint64_t)op
.snap
.snapid
;
7005 hobject_t missing_oid
;
7007 dout(10) << "_rollback_to " << soid
<< " snapid " << snapid
<< dendl
;
7009 ObjectContextRef rollback_to
;
7010 int ret
= find_object_context(
7011 hobject_t(soid
.oid
, soid
.get_key(), snapid
, soid
.get_hash(), info
.pgid
.pool(),
7012 soid
.get_namespace()),
7013 &rollback_to
, false, false, &missing_oid
);
7014 if (ret
== -EAGAIN
) {
7015 /* clone must be missing */
7016 assert(is_degraded_or_backfilling_object(missing_oid
));
7017 dout(20) << "_rollback_to attempted to roll back to a missing or backfilling clone "
7018 << missing_oid
<< " (requested snapid: ) " << snapid
<< dendl
;
7019 block_write_on_degraded_snap(missing_oid
, ctx
->op
);
7023 ObjectContextRef promote_obc
;
7024 cache_result_t tier_mode_result
;
7025 if (obs
.exists
&& obs
.oi
.has_manifest()) {
7027 maybe_handle_manifest_detail(
7033 maybe_handle_cache_detail(
7043 switch (tier_mode_result
) {
7044 case cache_result_t::NOOP
:
7046 case cache_result_t::BLOCKED_PROMOTE
:
7047 assert(promote_obc
);
7048 block_write_on_snap_rollback(soid
, promote_obc
, ctx
->op
);
7050 case cache_result_t::BLOCKED_FULL
:
7051 block_write_on_full_cache(soid
, ctx
->op
);
7053 case cache_result_t::REPLIED_WITH_EAGAIN
:
7054 assert(0 == "this can't happen, no rollback on replica");
7056 assert(0 == "must promote was set, other values are not valid");
7061 if (ret
== -ENOENT
|| (rollback_to
&& rollback_to
->obs
.oi
.is_whiteout())) {
7062 // there's no snapshot here, or there's no object.
7063 // if there's no snapshot, we delete the object; otherwise, do nothing.
7064 dout(20) << "_rollback_to deleting head on " << soid
.oid
7065 << " because got ENOENT|whiteout on find_object_context" << dendl
;
7066 if (ctx
->obc
->obs
.oi
.watchers
.size()) {
7067 // Cannot delete an object with watchers
7070 _delete_oid(ctx
, false, false);
7074 // ummm....huh? It *can't* return anything else at time of writing.
7075 assert(0 == "unexpected error code in _rollback_to");
7076 } else { //we got our context, let's use it to do the rollback!
7077 hobject_t
& rollback_to_sobject
= rollback_to
->obs
.oi
.soid
;
7078 if (is_degraded_or_backfilling_object(rollback_to_sobject
)) {
7079 dout(20) << "_rollback_to attempted to roll back to a degraded object "
7080 << rollback_to_sobject
<< " (requested snapid: ) " << snapid
<< dendl
;
7081 block_write_on_degraded_snap(rollback_to_sobject
, ctx
->op
);
7083 } else if (rollback_to
->obs
.oi
.soid
.snap
== CEPH_NOSNAP
) {
7084 // rolling back to the head; we just need to clone it.
7087 /* 1) Delete current head
7088 * 2) Clone correct snapshot into head
7089 * 3) Calculate clone_overlaps by following overlaps
7090 * forward from rollback snapshot */
7091 dout(10) << "_rollback_to deleting " << soid
.oid
7092 << " and rolling back to old snap" << dendl
;
7097 t
->clone(soid
, rollback_to_sobject
);
7098 snapset
.head_exists
= true;
7099 t
->add_obc(rollback_to
);
7101 map
<snapid_t
, interval_set
<uint64_t> >::iterator iter
=
7102 snapset
.clone_overlap
.lower_bound(snapid
);
7103 interval_set
<uint64_t> overlaps
= iter
->second
;
7104 assert(iter
!= snapset
.clone_overlap
.end());
7106 iter
!= snapset
.clone_overlap
.end();
7108 overlaps
.intersection_of(iter
->second
);
7110 if (obs
.oi
.size
> 0) {
7111 interval_set
<uint64_t> modified
;
7112 modified
.insert(0, obs
.oi
.size
);
7113 overlaps
.intersection_of(modified
);
7114 modified
.subtract(overlaps
);
7115 ctx
->modified_ranges
.union_of(modified
);
7118 // Adjust the cached objectcontext
7119 maybe_create_new_object(ctx
, true);
7120 ctx
->delta_stats
.num_bytes
-= obs
.oi
.size
;
7121 ctx
->delta_stats
.num_bytes
+= rollback_to
->obs
.oi
.size
;
7122 obs
.oi
.size
= rollback_to
->obs
.oi
.size
;
7123 if (rollback_to
->obs
.oi
.is_data_digest())
7124 obs
.oi
.set_data_digest(rollback_to
->obs
.oi
.data_digest
);
7126 obs
.oi
.clear_data_digest();
7127 if (rollback_to
->obs
.oi
.is_omap_digest())
7128 obs
.oi
.set_omap_digest(rollback_to
->obs
.oi
.omap_digest
);
7130 obs
.oi
.clear_omap_digest();
7132 if (rollback_to
->obs
.oi
.is_omap()) {
7133 dout(10) << __func__
<< " setting omap flag on " << obs
.oi
.soid
<< dendl
;
7134 obs
.oi
.set_flag(object_info_t::FLAG_OMAP
);
7136 dout(10) << __func__
<< " clearing omap flag on " << obs
.oi
.soid
<< dendl
;
7137 obs
.oi
.clear_flag(object_info_t::FLAG_OMAP
);
7140 snapset
.head_exists
= true;
7146 void PrimaryLogPG::_make_clone(
7149 ObjectContextRef obc
,
7150 const hobject_t
& head
, const hobject_t
& coid
,
7154 ::encode(*poi
, bv
, get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
7156 t
->clone(coid
, head
);
7157 setattr_maybe_cache(obc
, ctx
, t
, OI_ATTR
, bv
);
7158 rmattr_maybe_cache(obc
, ctx
, t
, SS_ATTR
);
7161 void PrimaryLogPG::make_writeable(OpContext
*ctx
)
7163 const hobject_t
& soid
= ctx
->obs
->oi
.soid
;
7164 SnapContext
& snapc
= ctx
->snapc
;
7167 assert(soid
.snap
== CEPH_NOSNAP
);
7168 dout(20) << "make_writeable " << soid
<< " snapset=" << ctx
->new_snapset
7169 << " snapc=" << snapc
<< dendl
;
7171 bool was_dirty
= ctx
->obc
->obs
.oi
.is_dirty();
7172 if (ctx
->new_obs
.exists
) {
7173 // we will mark the object dirty
7174 if (ctx
->undirty
&& was_dirty
) {
7175 dout(20) << " clearing DIRTY flag" << dendl
;
7176 assert(ctx
->new_obs
.oi
.is_dirty());
7177 ctx
->new_obs
.oi
.clear_flag(object_info_t::FLAG_DIRTY
);
7178 --ctx
->delta_stats
.num_objects_dirty
;
7179 osd
->logger
->inc(l_osd_tier_clean
);
7180 } else if (!was_dirty
&& !ctx
->undirty
) {
7181 dout(20) << " setting DIRTY flag" << dendl
;
7182 ctx
->new_obs
.oi
.set_flag(object_info_t::FLAG_DIRTY
);
7183 ++ctx
->delta_stats
.num_objects_dirty
;
7184 osd
->logger
->inc(l_osd_tier_dirty
);
7188 dout(20) << " deletion, decrementing num_dirty and clearing flag" << dendl
;
7189 ctx
->new_obs
.oi
.clear_flag(object_info_t::FLAG_DIRTY
);
7190 --ctx
->delta_stats
.num_objects_dirty
;
7194 if ((ctx
->new_obs
.exists
&&
7195 ctx
->new_obs
.oi
.is_omap()) &&
7196 (!ctx
->obc
->obs
.exists
||
7197 !ctx
->obc
->obs
.oi
.is_omap())) {
7198 ++ctx
->delta_stats
.num_objects_omap
;
7200 if ((!ctx
->new_obs
.exists
||
7201 !ctx
->new_obs
.oi
.is_omap()) &&
7202 (ctx
->obc
->obs
.exists
&&
7203 ctx
->obc
->obs
.oi
.is_omap())) {
7204 --ctx
->delta_stats
.num_objects_omap
;
7208 if (ctx
->new_snapset
.seq
> snapc
.seq
) {
7209 snapc
.seq
= ctx
->new_snapset
.seq
;
7210 snapc
.snaps
= ctx
->new_snapset
.snaps
;
7211 filter_snapc(snapc
.snaps
);
7212 dout(10) << " using newer snapc " << snapc
<< dendl
;
7215 if ((ctx
->obs
->exists
&& !ctx
->obs
->oi
.is_whiteout()) && // head exist(ed)
7216 snapc
.snaps
.size() && // there are snaps
7217 !ctx
->cache_evict
&&
7218 snapc
.snaps
[0] > ctx
->new_snapset
.seq
) { // existing object is old
7220 hobject_t coid
= soid
;
7221 coid
.snap
= snapc
.seq
;
7224 for (l
=1; l
<snapc
.snaps
.size() && snapc
.snaps
[l
] > ctx
->new_snapset
.seq
; l
++) ;
7226 vector
<snapid_t
> snaps(l
);
7227 for (unsigned i
=0; i
<l
; i
++)
7228 snaps
[i
] = snapc
.snaps
[i
];
7231 object_info_t
static_snap_oi(coid
);
7232 object_info_t
*snap_oi
;
7234 ctx
->clone_obc
= object_contexts
.lookup_or_create(static_snap_oi
.soid
);
7235 ctx
->clone_obc
->destructor_callback
= new C_PG_ObjectContext(this, ctx
->clone_obc
.get());
7236 ctx
->clone_obc
->obs
.oi
= static_snap_oi
;
7237 ctx
->clone_obc
->obs
.exists
= true;
7238 ctx
->clone_obc
->ssc
= ctx
->obc
->ssc
;
7239 ctx
->clone_obc
->ssc
->ref
++;
7240 if (pool
.info
.require_rollback())
7241 ctx
->clone_obc
->attr_cache
= ctx
->obc
->attr_cache
;
7242 snap_oi
= &ctx
->clone_obc
->obs
.oi
;
7243 bool got
= ctx
->lock_manager
.get_write_greedy(
7248 dout(20) << " got greedy write on clone_obc " << *ctx
->clone_obc
<< dendl
;
7250 snap_oi
= &static_snap_oi
;
7252 snap_oi
->version
= ctx
->at_version
;
7253 snap_oi
->prior_version
= ctx
->obs
->oi
.version
;
7254 snap_oi
->copy_user_bits(ctx
->obs
->oi
);
7256 bool legacy
= ctx
->new_snapset
.is_legacy() ||
7257 get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
;
7259 snap_oi
->legacy_snaps
= snaps
;
7262 _make_clone(ctx
, ctx
->op_t
.get(), ctx
->clone_obc
, soid
, coid
, snap_oi
);
7264 ctx
->delta_stats
.num_objects
++;
7265 if (snap_oi
->is_dirty()) {
7266 ctx
->delta_stats
.num_objects_dirty
++;
7267 osd
->logger
->inc(l_osd_tier_dirty
);
7269 if (snap_oi
->is_omap())
7270 ctx
->delta_stats
.num_objects_omap
++;
7271 if (snap_oi
->is_cache_pinned())
7272 ctx
->delta_stats
.num_objects_pinned
++;
7273 ctx
->delta_stats
.num_object_clones
++;
7274 ctx
->new_snapset
.clones
.push_back(coid
.snap
);
7275 ctx
->new_snapset
.clone_size
[coid
.snap
] = ctx
->obs
->oi
.size
;
7277 ctx
->new_snapset
.clone_snaps
[coid
.snap
] = snaps
;
7280 // clone_overlap should contain an entry for each clone
7281 // (an empty interval_set if there is no overlap)
7282 ctx
->new_snapset
.clone_overlap
[coid
.snap
];
7283 if (ctx
->obs
->oi
.size
)
7284 ctx
->new_snapset
.clone_overlap
[coid
.snap
].insert(0, ctx
->obs
->oi
.size
);
7287 dout(10) << " cloning v " << ctx
->obs
->oi
.version
7288 << " to " << coid
<< " v " << ctx
->at_version
7289 << " snaps=" << snaps
7290 << " snapset=" << ctx
->new_snapset
<< dendl
;
7291 ctx
->log
.push_back(pg_log_entry_t(pg_log_entry_t::CLONE
, coid
, ctx
->at_version
,
7292 ctx
->obs
->oi
.version
,
7293 ctx
->obs
->oi
.user_version
,
7294 osd_reqid_t(), ctx
->new_obs
.oi
.mtime
, 0));
7295 ::encode(snaps
, ctx
->log
.back().snaps
);
7297 ctx
->at_version
.version
++;
7300 // update most recent clone_overlap and usage stats
7301 if (ctx
->new_snapset
.clones
.size() > 0) {
7302 /* we need to check whether the most recent clone exists, if it's been evicted,
7303 * it's not included in the stats */
7304 hobject_t last_clone_oid
= soid
;
7305 last_clone_oid
.snap
= ctx
->new_snapset
.clone_overlap
.rbegin()->first
;
7306 if (is_present_clone(last_clone_oid
)) {
7307 interval_set
<uint64_t> &newest_overlap
= ctx
->new_snapset
.clone_overlap
.rbegin()->second
;
7308 ctx
->modified_ranges
.intersection_of(newest_overlap
);
7309 // modified_ranges is still in use by the clone
7310 add_interval_usage(ctx
->modified_ranges
, ctx
->delta_stats
);
7311 newest_overlap
.subtract(ctx
->modified_ranges
);
7315 // update snapset with latest snap context
7316 ctx
->new_snapset
.seq
= snapc
.seq
;
7317 ctx
->new_snapset
.snaps
= snapc
.snaps
;
7318 if (get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
7319 // pessimistic assumption that this is a net-new legacy SnapSet
7320 ctx
->delta_stats
.num_legacy_snapsets
++;
7321 ctx
->new_snapset
.head_exists
= ctx
->new_obs
.exists
;
7322 } else if (ctx
->new_snapset
.is_legacy()) {
7323 ctx
->new_snapset
.head_exists
= ctx
->new_obs
.exists
;
7325 dout(20) << "make_writeable " << soid
7326 << " done, snapset=" << ctx
->new_snapset
<< dendl
;
7330 void PrimaryLogPG::write_update_size_and_usage(object_stat_sum_t
& delta_stats
, object_info_t
& oi
,
7331 interval_set
<uint64_t>& modified
, uint64_t offset
,
7332 uint64_t length
, bool write_full
)
7334 interval_set
<uint64_t> ch
;
7337 ch
.insert(0, oi
.size
);
7339 ch
.insert(offset
, length
);
7340 modified
.union_of(ch
);
7341 if (write_full
|| offset
+ length
> oi
.size
) {
7342 uint64_t new_size
= offset
+ length
;
7343 delta_stats
.num_bytes
-= oi
.size
;
7344 delta_stats
.num_bytes
+= new_size
;
7347 delta_stats
.num_wr
++;
7348 delta_stats
.num_wr_kb
+= SHIFT_ROUND_UP(length
, 10);
7351 void PrimaryLogPG::add_interval_usage(interval_set
<uint64_t>& s
, object_stat_sum_t
& delta_stats
)
7353 for (interval_set
<uint64_t>::const_iterator p
= s
.begin(); p
!= s
.end(); ++p
) {
7354 delta_stats
.num_bytes
+= p
.get_len();
7358 void PrimaryLogPG::complete_disconnect_watches(
7359 ObjectContextRef obc
,
7360 const list
<watch_disconnect_t
> &to_disconnect
)
7362 for (list
<watch_disconnect_t
>::const_iterator i
=
7363 to_disconnect
.begin();
7364 i
!= to_disconnect
.end();
7366 pair
<uint64_t, entity_name_t
> watcher(i
->cookie
, i
->name
);
7367 auto watchers_entry
= obc
->watchers
.find(watcher
);
7368 if (watchers_entry
!= obc
->watchers
.end()) {
7369 WatchRef watch
= watchers_entry
->second
;
7370 dout(10) << "do_osd_op_effects disconnect watcher " << watcher
<< dendl
;
7371 obc
->watchers
.erase(watcher
);
7372 watch
->remove(i
->send_disconnect
);
7374 dout(10) << "do_osd_op_effects disconnect failed to find watcher "
7375 << watcher
<< dendl
;
7380 void PrimaryLogPG::do_osd_op_effects(OpContext
*ctx
, const ConnectionRef
& conn
)
7382 entity_name_t entity
= ctx
->reqid
.name
;
7383 dout(15) << "do_osd_op_effects " << entity
<< " con " << conn
.get() << dendl
;
7385 // disconnects first
7386 complete_disconnect_watches(ctx
->obc
, ctx
->watch_disconnects
);
7390 boost::intrusive_ptr
<Session
> session((Session
*)conn
->get_priv());
7393 session
->put(); // get_priv() takes a ref, and so does the intrusive_ptr
7395 for (list
<pair
<watch_info_t
,bool> >::iterator i
= ctx
->watch_connects
.begin();
7396 i
!= ctx
->watch_connects
.end();
7398 pair
<uint64_t, entity_name_t
> watcher(i
->first
.cookie
, entity
);
7399 dout(15) << "do_osd_op_effects applying watch connect on session "
7400 << session
.get() << " watcher " << watcher
<< dendl
;
7402 if (ctx
->obc
->watchers
.count(watcher
)) {
7403 dout(15) << "do_osd_op_effects found existing watch watcher " << watcher
7405 watch
= ctx
->obc
->watchers
[watcher
];
7407 dout(15) << "do_osd_op_effects new watcher " << watcher
7409 watch
= Watch::makeWatchRef(
7410 this, osd
, ctx
->obc
, i
->first
.timeout_seconds
,
7411 i
->first
.cookie
, entity
, conn
->get_peer_addr());
7412 ctx
->obc
->watchers
.insert(
7417 watch
->connect(conn
, i
->second
);
7420 for (list
<notify_info_t
>::iterator p
= ctx
->notifies
.begin();
7421 p
!= ctx
->notifies
.end();
7423 dout(10) << "do_osd_op_effects, notify " << *p
<< dendl
;
7424 ConnectionRef
conn(ctx
->op
->get_req()->get_connection());
7426 Notify::makeNotifyRef(
7428 ctx
->reqid
.name
.num(),
7433 ctx
->obc
->obs
.oi
.user_version
,
7435 for (map
<pair
<uint64_t, entity_name_t
>, WatchRef
>::iterator i
=
7436 ctx
->obc
->watchers
.begin();
7437 i
!= ctx
->obc
->watchers
.end();
7439 dout(10) << "starting notify on watch " << i
->first
<< dendl
;
7440 i
->second
->start_notify(notif
);
7445 for (list
<OpContext::NotifyAck
>::iterator p
= ctx
->notify_acks
.begin();
7446 p
!= ctx
->notify_acks
.end();
7448 if (p
->watch_cookie
)
7449 dout(10) << "notify_ack " << make_pair(p
->watch_cookie
.get(), p
->notify_id
) << dendl
;
7451 dout(10) << "notify_ack " << make_pair("NULL", p
->notify_id
) << dendl
;
7452 for (map
<pair
<uint64_t, entity_name_t
>, WatchRef
>::iterator i
=
7453 ctx
->obc
->watchers
.begin();
7454 i
!= ctx
->obc
->watchers
.end();
7456 if (i
->first
.second
!= entity
) continue;
7457 if (p
->watch_cookie
&&
7458 p
->watch_cookie
.get() != i
->first
.first
) continue;
7459 dout(10) << "acking notify on watch " << i
->first
<< dendl
;
7460 i
->second
->notify_ack(p
->notify_id
, p
->reply_bl
);
7465 hobject_t
PrimaryLogPG::generate_temp_object(const hobject_t
& target
)
7468 ss
<< "temp_" << info
.pgid
<< "_" << get_role()
7469 << "_" << osd
->monc
->get_global_id() << "_" << (++temp_seq
);
7470 hobject_t hoid
= target
.make_temp_hobject(ss
.str());
7471 dout(20) << __func__
<< " " << hoid
<< dendl
;
7475 hobject_t
PrimaryLogPG::get_temp_recovery_object(
7476 const hobject_t
& target
,
7480 ss
<< "temp_recovering_" << info
.pgid
// (note this includes the shardid)
7482 << "_" << info
.history
.same_interval_since
7483 << "_" << target
.snap
;
7484 // pgid + version + interval + snapid is unique, and short
7485 hobject_t hoid
= target
.make_temp_hobject(ss
.str());
7486 dout(20) << __func__
<< " " << hoid
<< dendl
;
7490 int PrimaryLogPG::prepare_transaction(OpContext
*ctx
)
7492 assert(!ctx
->ops
->empty());
7494 const hobject_t
& soid
= ctx
->obs
->oi
.soid
;
7496 // valid snap context?
7497 if (!ctx
->snapc
.is_valid()) {
7498 dout(10) << " invalid snapc " << ctx
->snapc
<< dendl
;
7502 // prepare the actual mutation
7503 int result
= do_osd_ops(ctx
, *ctx
->ops
);
7505 if (ctx
->op
->may_write() &&
7506 get_osdmap()->require_osd_release
>= CEPH_RELEASE_KRAKEN
) {
7507 // need to save the error code in the pg log, to detect dup ops,
7508 // but do nothing else
7509 ctx
->update_log_only
= true;
7514 // read-op? write-op noop? done?
7515 if (ctx
->op_t
->empty() && !ctx
->modify
) {
7516 unstable_stats
.add(ctx
->delta_stats
);
7517 if (ctx
->op
->may_write() &&
7518 get_osdmap()->require_osd_release
>= CEPH_RELEASE_KRAKEN
) {
7519 ctx
->update_log_only
= true;
7525 if ((ctx
->delta_stats
.num_bytes
> 0 ||
7526 ctx
->delta_stats
.num_objects
> 0) && // FIXME: keys?
7527 (pool
.info
.has_flag(pg_pool_t::FLAG_FULL
) ||
7528 get_osdmap()->test_flag(CEPH_OSDMAP_FULL
))) {
7529 const MOSDOp
*m
= static_cast<const MOSDOp
*>(ctx
->op
->get_req());
7530 if (ctx
->reqid
.name
.is_mds() || // FIXME: ignore MDS for now
7531 m
->has_flag(CEPH_OSD_FLAG_FULL_FORCE
)) {
7532 dout(20) << __func__
<< " full, but proceeding due to FULL_FORCE or MDS"
7534 } else if (m
->has_flag(CEPH_OSD_FLAG_FULL_TRY
)) {
7535 // they tried, they failed.
7536 dout(20) << __func__
<< " full, replying to FULL_TRY op" << dendl
;
7537 return pool
.info
.has_flag(pg_pool_t::FLAG_FULL
) ? -EDQUOT
: -ENOSPC
;
7540 dout(20) << __func__
<< " full, dropping request (bad client)" << dendl
;
7545 // clone, if necessary
7546 if (soid
.snap
== CEPH_NOSNAP
)
7547 make_writeable(ctx
);
7550 ctx
->new_obs
.exists
? pg_log_entry_t::MODIFY
:
7551 pg_log_entry_t::DELETE
);
7556 void PrimaryLogPG::finish_ctx(OpContext
*ctx
, int log_op_type
, bool maintain_ssc
)
7558 const hobject_t
& soid
= ctx
->obs
->oi
.soid
;
7559 dout(20) << __func__
<< " " << soid
<< " " << ctx
7560 << " op " << pg_log_entry_t::get_op_name(log_op_type
)
7562 utime_t now
= ceph_clock_now();
7567 if (soid
.snap
== CEPH_NOSNAP
&& maintain_ssc
) {
7568 ::encode(ctx
->new_snapset
, bss
);
7569 assert(ctx
->new_obs
.exists
== ctx
->new_snapset
.head_exists
||
7570 !ctx
->new_snapset
.is_legacy());
7572 if (ctx
->new_obs
.exists
) {
7573 if (!ctx
->obs
->exists
) {
7574 if (ctx
->snapset_obc
&& ctx
->snapset_obc
->obs
.exists
) {
7575 hobject_t snapoid
= soid
.get_snapdir();
7576 dout(10) << " removing unneeded snapdir " << snapoid
<< dendl
;
7577 ctx
->log
.push_back(pg_log_entry_t(pg_log_entry_t::DELETE
, snapoid
,
7579 ctx
->snapset_obc
->obs
.oi
.version
,
7580 0, osd_reqid_t(), ctx
->mtime
, 0));
7581 ctx
->op_t
->remove(snapoid
);
7583 ctx
->at_version
.version
++;
7585 ctx
->snapset_obc
->obs
.exists
= false;
7588 } else if (!ctx
->new_snapset
.clones
.empty() &&
7589 !ctx
->cache_evict
&&
7590 !ctx
->new_snapset
.head_exists
&&
7591 (!ctx
->snapset_obc
|| !ctx
->snapset_obc
->obs
.exists
)) {
7592 // save snapset on _snap
7593 hobject_t
snapoid(soid
.oid
, soid
.get_key(), CEPH_SNAPDIR
, soid
.get_hash(),
7594 info
.pgid
.pool(), soid
.get_namespace());
7595 dout(10) << " final snapset " << ctx
->new_snapset
7596 << " in " << snapoid
<< dendl
;
7597 assert(get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
);
7598 ctx
->log
.push_back(pg_log_entry_t(pg_log_entry_t::MODIFY
, snapoid
,
7601 0, osd_reqid_t(), ctx
->mtime
, 0));
7603 if (!ctx
->snapset_obc
)
7604 ctx
->snapset_obc
= get_object_context(snapoid
, true);
7606 if (ctx
->lock_type
== ObjectContext::RWState::RWWRITE
) {
7607 got
= ctx
->lock_manager
.get_write_greedy(
7612 assert(ctx
->lock_type
== ObjectContext::RWState::RWEXCL
);
7613 got
= ctx
->lock_manager
.get_lock_type(
7614 ObjectContext::RWState::RWEXCL
,
7620 dout(20) << " got greedy write on snapset_obc " << *ctx
->snapset_obc
<< dendl
;
7621 ctx
->snapset_obc
->obs
.exists
= true;
7622 ctx
->snapset_obc
->obs
.oi
.version
= ctx
->at_version
;
7623 ctx
->snapset_obc
->obs
.oi
.last_reqid
= ctx
->reqid
;
7624 ctx
->snapset_obc
->obs
.oi
.mtime
= ctx
->mtime
;
7625 ctx
->snapset_obc
->obs
.oi
.local_mtime
= now
;
7627 map
<string
, bufferlist
> attrs
;
7628 bufferlist
bv(sizeof(ctx
->new_obs
.oi
));
7629 ::encode(ctx
->snapset_obc
->obs
.oi
, bv
,
7630 get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
7631 ctx
->op_t
->create(snapoid
);
7632 attrs
[OI_ATTR
].claim(bv
);
7633 attrs
[SS_ATTR
].claim(bss
);
7634 setattrs_maybe_cache(ctx
->snapset_obc
, ctx
, ctx
->op_t
.get(), attrs
);
7635 ctx
->at_version
.version
++;
7639 // finish and log the op.
7640 if (ctx
->user_modify
) {
7641 // update the user_version for any modify ops, except for the watch op
7642 ctx
->user_at_version
= MAX(info
.last_user_version
, ctx
->new_obs
.oi
.user_version
) + 1;
7643 /* In order for new clients and old clients to interoperate properly
7644 * when exchanging versions, we need to lower bound the user_version
7645 * (which our new clients pay proper attention to)
7646 * by the at_version (which is all the old clients can ever see). */
7647 if (ctx
->at_version
.version
> ctx
->user_at_version
)
7648 ctx
->user_at_version
= ctx
->at_version
.version
;
7649 ctx
->new_obs
.oi
.user_version
= ctx
->user_at_version
;
7651 ctx
->bytes_written
= ctx
->op_t
->get_bytes_written();
7653 if (ctx
->new_obs
.exists
) {
7654 // on the head object
7655 ctx
->new_obs
.oi
.version
= ctx
->at_version
;
7656 ctx
->new_obs
.oi
.prior_version
= ctx
->obs
->oi
.version
;
7657 ctx
->new_obs
.oi
.last_reqid
= ctx
->reqid
;
7658 if (ctx
->mtime
!= utime_t()) {
7659 ctx
->new_obs
.oi
.mtime
= ctx
->mtime
;
7660 dout(10) << " set mtime to " << ctx
->new_obs
.oi
.mtime
<< dendl
;
7661 ctx
->new_obs
.oi
.local_mtime
= now
;
7663 dout(10) << " mtime unchanged at " << ctx
->new_obs
.oi
.mtime
<< dendl
;
7666 map
<string
, bufferlist
> attrs
;
7667 bufferlist
bv(sizeof(ctx
->new_obs
.oi
));
7668 ::encode(ctx
->new_obs
.oi
, bv
,
7669 get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
7670 attrs
[OI_ATTR
].claim(bv
);
7672 if (soid
.snap
== CEPH_NOSNAP
) {
7673 dout(10) << " final snapset " << ctx
->new_snapset
7674 << " in " << soid
<< dendl
;
7675 attrs
[SS_ATTR
].claim(bss
);
7677 dout(10) << " no snapset (this is a clone)" << dendl
;
7679 ctx
->op_t
->setattrs(soid
, attrs
);
7681 ctx
->new_obs
.oi
= object_info_t(ctx
->obc
->obs
.oi
.soid
);
7684 bool legacy_snapset
= ctx
->new_snapset
.is_legacy() ||
7685 get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
;
7688 ctx
->log
.push_back(pg_log_entry_t(log_op_type
, soid
, ctx
->at_version
,
7689 ctx
->obs
->oi
.version
,
7690 ctx
->user_at_version
, ctx
->reqid
,
7692 if (soid
.snap
< CEPH_NOSNAP
) {
7693 switch (log_op_type
) {
7694 case pg_log_entry_t::MODIFY
:
7695 case pg_log_entry_t::PROMOTE
:
7696 case pg_log_entry_t::CLEAN
:
7697 if (legacy_snapset
) {
7698 dout(20) << __func__
<< " encoding legacy_snaps "
7699 << ctx
->new_obs
.oi
.legacy_snaps
7701 ::encode(ctx
->new_obs
.oi
.legacy_snaps
, ctx
->log
.back().snaps
);
7703 dout(20) << __func__
<< " encoding snaps from " << ctx
->new_snapset
7705 ::encode(ctx
->new_snapset
.clone_snaps
[soid
.snap
], ctx
->log
.back().snaps
);
7713 if (!ctx
->extra_reqids
.empty()) {
7714 dout(20) << __func__
<< " extra_reqids " << ctx
->extra_reqids
<< dendl
;
7715 ctx
->log
.back().extra_reqids
.swap(ctx
->extra_reqids
);
7718 // apply new object state.
7719 ctx
->obc
->obs
= ctx
->new_obs
;
7721 if (soid
.is_head() && !ctx
->obc
->obs
.exists
&&
7722 (!maintain_ssc
|| ctx
->cache_evict
)) {
7723 ctx
->obc
->ssc
->exists
= false;
7724 ctx
->obc
->ssc
->snapset
= SnapSet();
7726 ctx
->obc
->ssc
->exists
= true;
7727 ctx
->obc
->ssc
->snapset
= ctx
->new_snapset
;
7731 void PrimaryLogPG::apply_stats(
7732 const hobject_t
&soid
,
7733 const object_stat_sum_t
&delta_stats
) {
7735 info
.stats
.stats
.add(delta_stats
);
7737 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
7738 i
!= backfill_targets
.end();
7741 pg_info_t
& pinfo
= peer_info
[bt
];
7742 if (soid
<= pinfo
.last_backfill
)
7743 pinfo
.stats
.stats
.add(delta_stats
);
7744 else if (soid
<= last_backfill_started
)
7745 pending_backfill_updates
[soid
].stats
.add(delta_stats
);
7748 if (is_primary() && scrubber
.active
) {
7749 if (soid
< scrubber
.start
) {
7750 dout(20) << __func__
<< " " << soid
<< " < [" << scrubber
.start
7751 << "," << scrubber
.end
<< ")" << dendl
;
7752 scrub_cstat
.add(delta_stats
);
7754 dout(20) << __func__
<< " " << soid
<< " >= [" << scrubber
.start
7755 << "," << scrubber
.end
<< ")" << dendl
;
7760 void PrimaryLogPG::complete_read_ctx(int result
, OpContext
*ctx
)
7762 const MOSDOp
*m
= static_cast<const MOSDOp
*>(ctx
->op
->get_req());
7763 assert(ctx
->async_reads_complete());
7765 for (vector
<OSDOp
>::iterator p
= ctx
->ops
->begin();
7766 p
!= ctx
->ops
->end() && result
>= 0; ++p
) {
7767 if (p
->rval
< 0 && !(p
->op
.flags
& CEPH_OSD_OP_FLAG_FAILOK
)) {
7771 ctx
->bytes_read
+= p
->outdata
.length();
7773 ctx
->reply
->claim_op_out_data(*ctx
->ops
);
7774 ctx
->reply
->get_header().data_off
= (ctx
->data_off
? *ctx
->data_off
: 0);
7776 MOSDOpReply
*reply
= ctx
->reply
;
7777 ctx
->reply
= nullptr;
7780 if (!ctx
->ignore_log_op_stats
) {
7782 publish_stats_to_osd();
7785 // on read, return the current object version
7787 reply
->set_reply_versions(eversion_t(), ctx
->obs
->oi
.user_version
);
7789 reply
->set_reply_versions(eversion_t(), ctx
->user_at_version
);
7791 } else if (result
== -ENOENT
) {
7792 // on ENOENT, set a floor for what the next user version will be.
7793 reply
->set_enoent_reply_versions(info
.last_update
, info
.last_user_version
);
7796 reply
->set_result(result
);
7797 reply
->add_flags(CEPH_OSD_FLAG_ACK
| CEPH_OSD_FLAG_ONDISK
);
7798 osd
->send_message_osd_client(reply
, m
->get_connection());
7802 // ========================================================================
7805 struct C_Copyfrom
: public Context
{
7808 epoch_t last_peering_reset
;
7810 PrimaryLogPG::CopyOpRef cop
;
7811 C_Copyfrom(PrimaryLogPG
*p
, hobject_t o
, epoch_t lpr
,
7812 const PrimaryLogPG::CopyOpRef
& c
)
7813 : pg(p
), oid(o
), last_peering_reset(lpr
),
7816 void finish(int r
) override
{
7817 if (r
== -ECANCELED
)
7820 if (last_peering_reset
== pg
->get_last_peering_reset()) {
7821 pg
->process_copy_chunk(oid
, tid
, r
);
7827 struct C_CopyFrom_AsyncReadCb
: public Context
{
7829 object_copy_data_t reply_obj
;
7832 C_CopyFrom_AsyncReadCb(OSDOp
*osd_op
, uint64_t features
) :
7833 osd_op(osd_op
), features(features
), len(0) {}
7834 void finish(int r
) override
{
7841 assert(len
<= reply_obj
.data
.length());
7843 bl
.substr_of(reply_obj
.data
, 0, len
);
7844 reply_obj
.data
.swap(bl
);
7845 ::encode(reply_obj
, osd_op
->outdata
, features
);
7849 int PrimaryLogPG::do_copy_get(OpContext
*ctx
, bufferlist::iterator
& bp
,
7850 OSDOp
& osd_op
, ObjectContextRef
&obc
)
7852 object_info_t
& oi
= obc
->obs
.oi
;
7853 hobject_t
& soid
= oi
.soid
;
7855 object_copy_cursor_t cursor
;
7857 bool skip_data_digest
=
7858 (osd
->store
->has_builtin_csum() && g_conf
->osd_skip_data_digest
) ||
7859 g_conf
->osd_distrust_data_digest
;
7862 ::decode(cursor
, bp
);
7863 ::decode(out_max
, bp
);
7865 catch (buffer::error
& e
) {
7870 const MOSDOp
*op
= reinterpret_cast<const MOSDOp
*>(ctx
->op
->get_req());
7871 uint64_t features
= op
->get_features();
7873 bool async_read_started
= false;
7874 object_copy_data_t _reply_obj
;
7875 C_CopyFrom_AsyncReadCb
*cb
= NULL
;
7876 if (pool
.info
.require_rollback()) {
7877 cb
= new C_CopyFrom_AsyncReadCb(&osd_op
, features
);
7879 object_copy_data_t
&reply_obj
= cb
? cb
->reply_obj
: _reply_obj
;
7881 reply_obj
.size
= oi
.size
;
7882 reply_obj
.mtime
= oi
.mtime
;
7884 if (soid
.snap
< CEPH_NOSNAP
) {
7885 if (obc
->ssc
->snapset
.is_legacy()) {
7886 reply_obj
.snaps
= oi
.legacy_snaps
;
7888 auto p
= obc
->ssc
->snapset
.clone_snaps
.find(soid
.snap
);
7889 assert(p
!= obc
->ssc
->snapset
.clone_snaps
.end()); // warn?
7890 reply_obj
.snaps
= p
->second
;
7893 reply_obj
.snap_seq
= obc
->ssc
->snapset
.seq
;
7895 if (!skip_data_digest
&& oi
.is_data_digest()) {
7896 reply_obj
.flags
|= object_copy_data_t::FLAG_DATA_DIGEST
;
7897 reply_obj
.data_digest
= oi
.data_digest
;
7899 if (oi
.is_omap_digest()) {
7900 reply_obj
.flags
|= object_copy_data_t::FLAG_OMAP_DIGEST
;
7901 reply_obj
.omap_digest
= oi
.omap_digest
;
7903 reply_obj
.truncate_seq
= oi
.truncate_seq
;
7904 reply_obj
.truncate_size
= oi
.truncate_size
;
7907 map
<string
,bufferlist
>& out_attrs
= reply_obj
.attrs
;
7908 if (!cursor
.attr_complete
) {
7909 result
= getattrs_maybe_cache(
7918 cursor
.attr_complete
= true;
7919 dout(20) << " got attrs" << dendl
;
7922 int64_t left
= out_max
- osd_op
.outdata
.length();
7925 bufferlist
& bl
= reply_obj
.data
;
7926 if (left
> 0 && !cursor
.data_complete
) {
7927 if (cursor
.data_offset
< oi
.size
) {
7928 uint64_t max_read
= MIN(oi
.size
- cursor
.data_offset
, (uint64_t)left
);
7930 async_read_started
= true;
7931 ctx
->pending_async_reads
.push_back(
7933 boost::make_tuple(cursor
.data_offset
, max_read
, osd_op
.op
.flags
),
7934 make_pair(&bl
, cb
)));
7937 ctx
->op_finishers
[ctx
->current_osd_subop_num
].reset(
7938 new ReadFinisher(osd_op
));
7939 result
= -EINPROGRESS
;
7941 dout(10) << __func__
<< ": async_read noted for " << soid
<< dendl
;
7943 result
= pgbackend
->objects_read_sync(
7944 oi
.soid
, cursor
.data_offset
, max_read
, osd_op
.op
.flags
, &bl
);
7949 cursor
.data_offset
+= max_read
;
7951 if (cursor
.data_offset
== oi
.size
) {
7952 cursor
.data_complete
= true;
7953 dout(20) << " got data" << dendl
;
7955 assert(cursor
.data_offset
<= oi
.size
);
7959 uint32_t omap_keys
= 0;
7960 if (!pool
.info
.supports_omap() || !oi
.is_omap()) {
7961 cursor
.omap_complete
= true;
7963 if (left
> 0 && !cursor
.omap_complete
) {
7964 assert(cursor
.data_complete
);
7965 if (cursor
.omap_offset
.empty()) {
7966 osd
->store
->omap_get_header(ch
, ghobject_t(oi
.soid
),
7967 &reply_obj
.omap_header
);
7969 bufferlist omap_data
;
7970 ObjectMap::ObjectMapIterator iter
=
7971 osd
->store
->get_omap_iterator(coll
, ghobject_t(oi
.soid
));
7973 iter
->upper_bound(cursor
.omap_offset
);
7974 for (; iter
->valid(); iter
->next(false)) {
7976 ::encode(iter
->key(), omap_data
);
7977 ::encode(iter
->value(), omap_data
);
7978 left
-= iter
->key().length() + 4 + iter
->value().length() + 4;
7983 ::encode(omap_keys
, reply_obj
.omap_data
);
7984 reply_obj
.omap_data
.claim_append(omap_data
);
7986 if (iter
->valid()) {
7987 cursor
.omap_offset
= iter
->key();
7989 cursor
.omap_complete
= true;
7990 dout(20) << " got omap" << dendl
;
7995 if (cursor
.is_complete()) {
7996 // include reqids only in the final step. this is a bit fragile
7998 pg_log
.get_log().get_object_reqids(ctx
->obc
->obs
.oi
.soid
, 10, &reply_obj
.reqids
);
7999 dout(20) << " got reqids" << dendl
;
8002 dout(20) << " cursor.is_complete=" << cursor
.is_complete()
8003 << " " << out_attrs
.size() << " attrs"
8004 << " " << bl
.length() << " bytes"
8005 << " " << reply_obj
.omap_header
.length() << " omap header bytes"
8006 << " " << reply_obj
.omap_data
.length() << " omap data bytes in "
8007 << omap_keys
<< " keys"
8008 << " " << reply_obj
.reqids
.size() << " reqids"
8010 reply_obj
.cursor
= cursor
;
8011 if (!async_read_started
) {
8012 ::encode(reply_obj
, osd_op
.outdata
, features
);
8014 if (cb
&& !async_read_started
) {
8024 void PrimaryLogPG::fill_in_copy_get_noent(OpRequestRef
& op
, hobject_t oid
,
8027 // NOTE: we take non-const ref here for claim_op_out_data below; we must
8028 // be careful not to modify anything else that will upset a racing
8030 MOSDOp
*m
= static_cast<MOSDOp
*>(op
->get_nonconst_req());
8031 uint64_t features
= m
->get_features();
8032 object_copy_data_t reply_obj
;
8034 pg_log
.get_log().get_object_reqids(oid
, 10, &reply_obj
.reqids
);
8035 dout(20) << __func__
<< " got reqids " << reply_obj
.reqids
<< dendl
;
8036 ::encode(reply_obj
, osd_op
.outdata
, features
);
8037 osd_op
.rval
= -ENOENT
;
8038 MOSDOpReply
*reply
= new MOSDOpReply(m
, 0, get_osdmap()->get_epoch(), 0, false);
8039 reply
->claim_op_out_data(m
->ops
);
8040 reply
->set_result(-ENOENT
);
8041 reply
->add_flags(CEPH_OSD_FLAG_ACK
| CEPH_OSD_FLAG_ONDISK
);
8042 osd
->send_message_osd_client(reply
, m
->get_connection());
8045 void PrimaryLogPG::start_copy(CopyCallback
*cb
, ObjectContextRef obc
,
8046 hobject_t src
, object_locator_t oloc
,
8047 version_t version
, unsigned flags
,
8048 bool mirror_snapset
,
8049 unsigned src_obj_fadvise_flags
,
8050 unsigned dest_obj_fadvise_flags
)
8052 const hobject_t
& dest
= obc
->obs
.oi
.soid
;
8053 dout(10) << __func__
<< " " << dest
8054 << " from " << src
<< " " << oloc
<< " v" << version
8055 << " flags " << flags
8056 << (mirror_snapset
? " mirror_snapset" : "")
8059 assert(!mirror_snapset
|| (src
.snap
== CEPH_NOSNAP
||
8060 src
.snap
== CEPH_SNAPDIR
));
8062 // cancel a previous in-progress copy?
8063 if (copy_ops
.count(dest
)) {
8064 // FIXME: if the src etc match, we could avoid restarting from the
8066 CopyOpRef cop
= copy_ops
[dest
];
8067 vector
<ceph_tid_t
> tids
;
8068 cancel_copy(cop
, false, &tids
);
8069 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
8072 CopyOpRef
cop(std::make_shared
<CopyOp
>(cb
, obc
, src
, oloc
, version
, flags
,
8073 mirror_snapset
, src_obj_fadvise_flags
,
8074 dest_obj_fadvise_flags
));
8075 copy_ops
[dest
] = cop
;
8078 _copy_some(obc
, cop
);
8081 void PrimaryLogPG::_copy_some(ObjectContextRef obc
, CopyOpRef cop
)
8083 dout(10) << __func__
<< " " << obc
<< " " << cop
<< dendl
;
8086 if (cop
->flags
& CEPH_OSD_COPY_FROM_FLAG_FLUSH
)
8087 flags
|= CEPH_OSD_FLAG_FLUSH
;
8088 if (cop
->flags
& CEPH_OSD_COPY_FROM_FLAG_IGNORE_CACHE
)
8089 flags
|= CEPH_OSD_FLAG_IGNORE_CACHE
;
8090 if (cop
->flags
& CEPH_OSD_COPY_FROM_FLAG_IGNORE_OVERLAY
)
8091 flags
|= CEPH_OSD_FLAG_IGNORE_OVERLAY
;
8092 if (cop
->flags
& CEPH_OSD_COPY_FROM_FLAG_MAP_SNAP_CLONE
)
8093 flags
|= CEPH_OSD_FLAG_MAP_SNAP_CLONE
;
8094 if (cop
->flags
& CEPH_OSD_COPY_FROM_FLAG_RWORDERED
)
8095 flags
|= CEPH_OSD_FLAG_RWORDERED
;
8097 C_GatherBuilder
gather(cct
);
8099 if (cop
->cursor
.is_initial() && cop
->mirror_snapset
) {
8101 assert(cop
->src
.snap
== CEPH_NOSNAP
);
8103 op
.list_snaps(&cop
->results
.snapset
, NULL
);
8104 ceph_tid_t tid
= osd
->objecter
->read(cop
->src
.oid
, cop
->oloc
, op
,
8106 flags
, gather
.new_sub(), NULL
);
8107 cop
->objecter_tid2
= tid
;
8111 if (cop
->results
.user_version
) {
8112 op
.assert_version(cop
->results
.user_version
);
8114 // we should learn the version after the first chunk, if we didn't know
8116 assert(cop
->cursor
.is_initial());
8118 op
.copy_get(&cop
->cursor
, get_copy_chunk_size(),
8119 &cop
->results
.object_size
, &cop
->results
.mtime
,
8120 &cop
->attrs
, &cop
->data
, &cop
->omap_header
, &cop
->omap_data
,
8121 &cop
->results
.snaps
, &cop
->results
.snap_seq
,
8122 &cop
->results
.flags
,
8123 &cop
->results
.source_data_digest
,
8124 &cop
->results
.source_omap_digest
,
8125 &cop
->results
.reqids
,
8126 &cop
->results
.truncate_seq
,
8127 &cop
->results
.truncate_size
,
8129 op
.set_last_op_flags(cop
->src_obj_fadvise_flags
);
8131 C_Copyfrom
*fin
= new C_Copyfrom(this, obc
->obs
.oi
.soid
,
8132 get_last_peering_reset(), cop
);
8133 gather
.set_finisher(new C_OnFinisher(fin
,
8134 &osd
->objecter_finisher
));
8136 ceph_tid_t tid
= osd
->objecter
->read(cop
->src
.oid
, cop
->oloc
, op
,
8137 cop
->src
.snap
, NULL
,
8140 // discover the object version if we don't know it yet
8141 cop
->results
.user_version
? NULL
: &cop
->results
.user_version
);
8143 cop
->objecter_tid
= tid
;
8147 void PrimaryLogPG::process_copy_chunk(hobject_t oid
, ceph_tid_t tid
, int r
)
8149 vector
<ceph_tid_t
> tids
;
8150 dout(10) << __func__
<< " " << oid
<< " tid " << tid
8151 << " " << cpp_strerror(r
) << dendl
;
8152 map
<hobject_t
,CopyOpRef
>::iterator p
= copy_ops
.find(oid
);
8153 if (p
== copy_ops
.end()) {
8154 dout(10) << __func__
<< " no copy_op found" << dendl
;
8157 CopyOpRef cop
= p
->second
;
8158 if (tid
!= cop
->objecter_tid
) {
8159 dout(10) << __func__
<< " tid " << tid
<< " != cop " << cop
8160 << " tid " << cop
->objecter_tid
<< dendl
;
8164 if (cop
->omap_data
.length() || cop
->omap_header
.length())
8165 cop
->results
.has_omap
= true;
8167 if (r
>= 0 && !pool
.info
.supports_omap() &&
8168 (cop
->omap_data
.length() || cop
->omap_header
.length())) {
8171 cop
->objecter_tid
= 0;
8172 cop
->objecter_tid2
= 0; // assume this ordered before us (if it happened)
8173 ObjectContextRef
& cobc
= cop
->obc
;
8178 assert(cop
->rval
>= 0);
8180 if (oid
.snap
< CEPH_NOSNAP
&& !cop
->results
.snaps
.empty()) {
8181 // verify snap hasn't been deleted
8182 vector
<snapid_t
>::iterator p
= cop
->results
.snaps
.begin();
8183 while (p
!= cop
->results
.snaps
.end()) {
8184 if (pool
.info
.is_removed_snap(*p
)) {
8185 dout(10) << __func__
<< " clone snap " << *p
<< " has been deleted"
8187 for (vector
<snapid_t
>::iterator q
= p
+ 1;
8188 q
!= cop
->results
.snaps
.end();
8191 cop
->results
.snaps
.resize(cop
->results
.snaps
.size() - 1);
8196 if (cop
->results
.snaps
.empty()) {
8197 dout(10) << __func__
<< " no more snaps for " << oid
<< dendl
;
8203 assert(cop
->rval
>= 0);
8205 if (!cop
->temp_cursor
.data_complete
) {
8206 cop
->results
.data_digest
= cop
->data
.crc32c(cop
->results
.data_digest
);
8208 if (pool
.info
.supports_omap() && !cop
->temp_cursor
.omap_complete
) {
8209 if (cop
->omap_header
.length()) {
8210 cop
->results
.omap_digest
=
8211 cop
->omap_header
.crc32c(cop
->results
.omap_digest
);
8213 if (cop
->omap_data
.length()) {
8215 keys
.substr_of(cop
->omap_data
, 4, cop
->omap_data
.length() - 4);
8216 cop
->results
.omap_digest
= keys
.crc32c(cop
->results
.omap_digest
);
8220 if (!cop
->temp_cursor
.attr_complete
) {
8221 for (map
<string
,bufferlist
>::iterator p
= cop
->attrs
.begin();
8222 p
!= cop
->attrs
.end();
8224 cop
->results
.attrs
[string("_") + p
->first
] = p
->second
;
8229 if (!cop
->cursor
.is_complete()) {
8230 // write out what we have so far
8231 if (cop
->temp_cursor
.is_initial()) {
8232 assert(!cop
->results
.started_temp_obj
);
8233 cop
->results
.started_temp_obj
= true;
8234 cop
->results
.temp_oid
= generate_temp_object(oid
);
8235 dout(20) << __func__
<< " using temp " << cop
->results
.temp_oid
<< dendl
;
8237 ObjectContextRef tempobc
= get_object_context(cop
->results
.temp_oid
, true);
8238 OpContextUPtr ctx
= simple_opc_create(tempobc
);
8239 if (cop
->temp_cursor
.is_initial()) {
8240 ctx
->new_temp_oid
= cop
->results
.temp_oid
;
8242 _write_copy_chunk(cop
, ctx
->op_t
.get());
8243 simple_opc_submit(std::move(ctx
));
8244 dout(10) << __func__
<< " fetching more" << dendl
;
8245 _copy_some(cobc
, cop
);
8250 if (cop
->results
.is_data_digest() || cop
->results
.is_omap_digest()) {
8251 dout(20) << __func__
<< std::hex
8252 << " got digest: rx data 0x" << cop
->results
.data_digest
8253 << " omap 0x" << cop
->results
.omap_digest
8254 << ", source: data 0x" << cop
->results
.source_data_digest
8255 << " omap 0x" << cop
->results
.source_omap_digest
8257 << " flags " << cop
->results
.flags
8260 if (cop
->results
.is_data_digest() &&
8261 cop
->results
.data_digest
!= cop
->results
.source_data_digest
) {
8262 derr
<< __func__
<< std::hex
<< " data digest 0x" << cop
->results
.data_digest
8263 << " != source 0x" << cop
->results
.source_data_digest
<< std::dec
8265 osd
->clog
->error() << info
.pgid
<< " copy from " << cop
->src
8266 << " to " << cop
->obc
->obs
.oi
.soid
<< std::hex
8267 << " data digest 0x" << cop
->results
.data_digest
8268 << " != source 0x" << cop
->results
.source_data_digest
8273 if (cop
->results
.is_omap_digest() &&
8274 cop
->results
.omap_digest
!= cop
->results
.source_omap_digest
) {
8275 derr
<< __func__
<< std::hex
8276 << " omap digest 0x" << cop
->results
.omap_digest
8277 << " != source 0x" << cop
->results
.source_omap_digest
8278 << std::dec
<< dendl
;
8279 osd
->clog
->error() << info
.pgid
<< " copy from " << cop
->src
8280 << " to " << cop
->obc
->obs
.oi
.soid
<< std::hex
8281 << " omap digest 0x" << cop
->results
.omap_digest
8282 << " != source 0x" << cop
->results
.source_omap_digest
8287 if (cct
->_conf
->osd_debug_inject_copyfrom_error
) {
8288 derr
<< __func__
<< " injecting copyfrom failure" << dendl
;
8293 cop
->results
.fill_in_final_tx
= std::function
<void(PGTransaction
*)>(
8294 [this, &cop
/* avoid ref cycle */](PGTransaction
*t
) {
8295 ObjectState
& obs
= cop
->obc
->obs
;
8296 if (cop
->temp_cursor
.is_initial()) {
8297 dout(20) << "fill_in_final_tx: writing "
8298 << "directly to final object" << dendl
;
8299 // write directly to final object
8300 cop
->results
.temp_oid
= obs
.oi
.soid
;
8301 _write_copy_chunk(cop
, t
);
8303 // finish writing to temp object, then move into place
8304 dout(20) << "fill_in_final_tx: writing to temp object" << dendl
;
8305 _write_copy_chunk(cop
, t
);
8306 t
->rename(obs
.oi
.soid
, cop
->results
.temp_oid
);
8308 t
->setattrs(obs
.oi
.soid
, cop
->results
.attrs
);
8311 dout(20) << __func__
<< " success; committing" << dendl
;
8314 dout(20) << __func__
<< " complete r = " << cpp_strerror(r
) << dendl
;
8315 CopyCallbackResults
results(r
, &cop
->results
);
8316 cop
->cb
->complete(results
);
8318 copy_ops
.erase(cobc
->obs
.oi
.soid
);
8321 if (r
< 0 && cop
->results
.started_temp_obj
) {
8322 dout(10) << __func__
<< " deleting partial temp object "
8323 << cop
->results
.temp_oid
<< dendl
;
8324 ObjectContextRef tempobc
= get_object_context(cop
->results
.temp_oid
, true);
8325 OpContextUPtr ctx
= simple_opc_create(tempobc
);
8326 ctx
->op_t
->remove(cop
->results
.temp_oid
);
8327 ctx
->discard_temp_oid
= cop
->results
.temp_oid
;
8328 simple_opc_submit(std::move(ctx
));
8331 // cancel and requeue proxy ops on this object
8333 for (map
<ceph_tid_t
, ProxyReadOpRef
>::iterator it
= proxyread_ops
.begin();
8334 it
!= proxyread_ops
.end();) {
8335 if (it
->second
->soid
== cobc
->obs
.oi
.soid
) {
8336 cancel_proxy_read((it
++)->second
, &tids
);
8341 for (map
<ceph_tid_t
, ProxyWriteOpRef
>::iterator it
= proxywrite_ops
.begin();
8342 it
!= proxywrite_ops
.end();) {
8343 if (it
->second
->soid
== cobc
->obs
.oi
.soid
) {
8344 cancel_proxy_write((it
++)->second
, &tids
);
8349 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
8350 kick_proxy_ops_blocked(cobc
->obs
.oi
.soid
);
8353 kick_object_context_blocked(cobc
);
8356 void PrimaryLogPG::cancel_and_requeue_proxy_ops(hobject_t oid
) {
8357 vector
<ceph_tid_t
> tids
;
8358 for (map
<ceph_tid_t
, ProxyReadOpRef
>::iterator it
= proxyread_ops
.begin();
8359 it
!= proxyread_ops
.end();) {
8360 if (it
->second
->soid
== oid
) {
8361 cancel_proxy_read((it
++)->second
, &tids
);
8366 for (map
<ceph_tid_t
, ProxyWriteOpRef
>::iterator it
= proxywrite_ops
.begin();
8367 it
!= proxywrite_ops
.end();) {
8368 if (it
->second
->soid
== oid
) {
8369 cancel_proxy_write((it
++)->second
, &tids
);
8374 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
8375 kick_proxy_ops_blocked(oid
);
8378 void PrimaryLogPG::_write_copy_chunk(CopyOpRef cop
, PGTransaction
*t
)
8380 dout(20) << __func__
<< " " << cop
8381 << " " << cop
->attrs
.size() << " attrs"
8382 << " " << cop
->data
.length() << " bytes"
8383 << " " << cop
->omap_header
.length() << " omap header bytes"
8384 << " " << cop
->omap_data
.length() << " omap data bytes"
8386 if (!cop
->temp_cursor
.attr_complete
) {
8387 t
->create(cop
->results
.temp_oid
);
8389 if (!cop
->temp_cursor
.data_complete
) {
8390 assert(cop
->data
.length() + cop
->temp_cursor
.data_offset
==
8391 cop
->cursor
.data_offset
);
8392 if (pool
.info
.requires_aligned_append() &&
8393 !cop
->cursor
.data_complete
) {
8395 * Trim off the unaligned bit at the end, we'll adjust cursor.data_offset
8396 * to pick it up on the next pass.
8398 assert(cop
->temp_cursor
.data_offset
%
8399 pool
.info
.required_alignment() == 0);
8400 if (cop
->data
.length() % pool
.info
.required_alignment() != 0) {
8402 cop
->data
.length() % pool
.info
.required_alignment();
8404 bl
.substr_of(cop
->data
, 0, cop
->data
.length() - to_trim
);
8406 cop
->cursor
.data_offset
-= to_trim
;
8407 assert(cop
->data
.length() + cop
->temp_cursor
.data_offset
==
8408 cop
->cursor
.data_offset
);
8411 if (cop
->data
.length()) {
8413 cop
->results
.temp_oid
,
8414 cop
->temp_cursor
.data_offset
,
8417 cop
->dest_obj_fadvise_flags
);
8421 if (pool
.info
.supports_omap()) {
8422 if (!cop
->temp_cursor
.omap_complete
) {
8423 if (cop
->omap_header
.length()) {
8425 cop
->results
.temp_oid
,
8427 cop
->omap_header
.clear();
8429 if (cop
->omap_data
.length()) {
8430 map
<string
,bufferlist
> omap
;
8431 bufferlist::iterator p
= cop
->omap_data
.begin();
8433 t
->omap_setkeys(cop
->results
.temp_oid
, omap
);
8434 cop
->omap_data
.clear();
8438 assert(cop
->omap_header
.length() == 0);
8439 assert(cop
->omap_data
.length() == 0);
8441 cop
->temp_cursor
= cop
->cursor
;
8444 void PrimaryLogPG::finish_copyfrom(CopyFromCallback
*cb
)
8446 OpContext
*ctx
= cb
->ctx
;
8447 dout(20) << "finish_copyfrom on " << ctx
->obs
->oi
.soid
<< dendl
;
8449 ObjectState
& obs
= ctx
->new_obs
;
8451 dout(20) << __func__
<< ": exists, removing" << dendl
;
8452 ctx
->op_t
->remove(obs
.oi
.soid
);
8454 ctx
->delta_stats
.num_objects
++;
8457 if (cb
->is_temp_obj_used()) {
8458 ctx
->discard_temp_oid
= cb
->results
->temp_oid
;
8460 cb
->results
->fill_in_final_tx(ctx
->op_t
.get());
8462 // CopyFromCallback fills this in for us
8463 obs
.oi
.user_version
= ctx
->user_at_version
;
8465 if (cb
->results
->is_data_digest()) {
8466 obs
.oi
.set_data_digest(cb
->results
->data_digest
);
8468 obs
.oi
.clear_data_digest();
8470 if (cb
->results
->is_omap_digest()) {
8471 obs
.oi
.set_omap_digest(cb
->results
->omap_digest
);
8473 obs
.oi
.clear_omap_digest();
8476 obs
.oi
.truncate_seq
= cb
->results
->truncate_seq
;
8477 obs
.oi
.truncate_size
= cb
->results
->truncate_size
;
8479 ctx
->extra_reqids
= cb
->results
->reqids
;
8481 // cache: clear whiteout?
8482 if (obs
.oi
.is_whiteout()) {
8483 dout(10) << __func__
<< " clearing whiteout on " << obs
.oi
.soid
<< dendl
;
8484 obs
.oi
.clear_flag(object_info_t::FLAG_WHITEOUT
);
8485 --ctx
->delta_stats
.num_whiteouts
;
8488 if (cb
->results
->has_omap
) {
8489 dout(10) << __func__
<< " setting omap flag on " << obs
.oi
.soid
<< dendl
;
8490 obs
.oi
.set_flag(object_info_t::FLAG_OMAP
);
8492 dout(10) << __func__
<< " clearing omap flag on " << obs
.oi
.soid
<< dendl
;
8493 obs
.oi
.clear_flag(object_info_t::FLAG_OMAP
);
8496 interval_set
<uint64_t> ch
;
8497 if (obs
.oi
.size
> 0)
8498 ch
.insert(0, obs
.oi
.size
);
8499 ctx
->modified_ranges
.union_of(ch
);
8501 if (cb
->get_data_size() != obs
.oi
.size
) {
8502 ctx
->delta_stats
.num_bytes
-= obs
.oi
.size
;
8503 obs
.oi
.size
= cb
->get_data_size();
8504 ctx
->delta_stats
.num_bytes
+= obs
.oi
.size
;
8506 ctx
->delta_stats
.num_wr
++;
8507 ctx
->delta_stats
.num_wr_kb
+= SHIFT_ROUND_UP(obs
.oi
.size
, 10);
8509 osd
->logger
->inc(l_osd_copyfrom
);
8512 void PrimaryLogPG::finish_promote(int r
, CopyResults
*results
,
8513 ObjectContextRef obc
)
8515 const hobject_t
& soid
= obc
->obs
.oi
.soid
;
8516 dout(10) << __func__
<< " " << soid
<< " r=" << r
8517 << " uv" << results
->user_version
<< dendl
;
8519 if (r
== -ECANCELED
) {
8523 if (r
!= -ENOENT
&& soid
.is_snap()) {
8524 if (results
->snaps
.empty()) {
8525 // we must have read "snap" content from the head object in
8526 // the base pool. use snap_seq to construct what snaps should
8527 // be for this clone (what is was before we evicted the clean
8528 // clone from this pool, and what it will be when we flush and
8529 // the clone eventually happens in the base pool).
8530 SnapSet
& snapset
= obc
->ssc
->snapset
;
8531 vector
<snapid_t
>::iterator p
= snapset
.snaps
.begin();
8532 while (p
!= snapset
.snaps
.end() && *p
> soid
.snap
)
8534 while (p
!= snapset
.snaps
.end() && *p
> results
->snap_seq
) {
8535 results
->snaps
.push_back(*p
);
8540 dout(20) << __func__
<< " snaps " << results
->snaps
<< dendl
;
8541 filter_snapc(results
->snaps
);
8543 dout(20) << __func__
<< " filtered snaps " << results
->snaps
<< dendl
;
8544 if (results
->snaps
.empty()) {
8545 dout(20) << __func__
8546 << " snaps are empty, clone is invalid,"
8547 << " setting r to ENOENT" << dendl
;
8552 if (r
< 0 && results
->started_temp_obj
) {
8553 dout(10) << __func__
<< " abort; will clean up partial work" << dendl
;
8554 ObjectContextRef tempobc
= get_object_context(results
->temp_oid
, false);
8556 OpContextUPtr ctx
= simple_opc_create(tempobc
);
8557 ctx
->op_t
->remove(results
->temp_oid
);
8558 simple_opc_submit(std::move(ctx
));
8559 results
->started_temp_obj
= false;
8562 if (r
== -ENOENT
&& soid
.is_snap()) {
8563 dout(10) << __func__
8564 << ": enoent while trying to promote clone, " << soid
8565 << " must have been trimmed, removing from snapset"
8567 hobject_t
head(soid
.get_head());
8568 ObjectContextRef obc
= get_object_context(head
, false);
8571 OpContextUPtr tctx
= simple_opc_create(obc
);
8572 tctx
->at_version
= get_next_version();
8573 filter_snapc(tctx
->new_snapset
.snaps
);
8574 vector
<snapid_t
> new_clones
;
8575 map
<snapid_t
, vector
<snapid_t
>> new_clone_snaps
;
8576 for (vector
<snapid_t
>::iterator i
= tctx
->new_snapset
.clones
.begin();
8577 i
!= tctx
->new_snapset
.clones
.end();
8579 if (*i
!= soid
.snap
) {
8580 new_clones
.push_back(*i
);
8581 auto p
= tctx
->new_snapset
.clone_snaps
.find(*i
);
8582 if (p
!= tctx
->new_snapset
.clone_snaps
.end()) {
8583 new_clone_snaps
[*i
] = p
->second
;
8587 tctx
->new_snapset
.clones
.swap(new_clones
);
8588 tctx
->new_snapset
.clone_overlap
.erase(soid
.snap
);
8589 tctx
->new_snapset
.clone_size
.erase(soid
.snap
);
8590 tctx
->new_snapset
.clone_snaps
.swap(new_clone_snaps
);
8592 // take RWWRITE lock for duration of our local write. ignore starvation.
8593 if (!tctx
->lock_manager
.take_write_lock(
8596 assert(0 == "problem!");
8598 dout(20) << __func__
<< " took lock on obc, " << obc
->rwstate
<< dendl
;
8600 finish_ctx(tctx
.get(), pg_log_entry_t::PROMOTE
);
8602 simple_opc_submit(std::move(tctx
));
8606 bool whiteout
= false;
8608 assert(soid
.snap
== CEPH_NOSNAP
); // snap case is above
8609 dout(10) << __func__
<< " whiteout " << soid
<< dendl
;
8613 if (r
< 0 && !whiteout
) {
8614 derr
<< __func__
<< " unexpected promote error " << cpp_strerror(r
) << dendl
;
8615 // pass error to everyone blocked on this object
8616 // FIXME: this is pretty sloppy, but at this point we got
8617 // something unexpected and don't have many other options.
8618 map
<hobject_t
,list
<OpRequestRef
>>::iterator blocked_iter
=
8619 waiting_for_blocked_object
.find(soid
);
8620 if (blocked_iter
!= waiting_for_blocked_object
.end()) {
8621 while (!blocked_iter
->second
.empty()) {
8622 osd
->reply_op_error(blocked_iter
->second
.front(), r
);
8623 blocked_iter
->second
.pop_front();
8625 waiting_for_blocked_object
.erase(blocked_iter
);
8630 osd
->promote_finish(results
->object_size
);
8632 OpContextUPtr tctx
= simple_opc_create(obc
);
8633 tctx
->at_version
= get_next_version();
8635 ++tctx
->delta_stats
.num_objects
;
8636 if (soid
.snap
< CEPH_NOSNAP
)
8637 ++tctx
->delta_stats
.num_object_clones
;
8638 tctx
->new_obs
.exists
= true;
8640 tctx
->extra_reqids
= results
->reqids
;
8642 bool legacy_snapset
= tctx
->new_snapset
.is_legacy() ||
8643 get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
;
8646 // create a whiteout
8647 tctx
->op_t
->create(soid
);
8648 tctx
->new_obs
.oi
.set_flag(object_info_t::FLAG_WHITEOUT
);
8649 ++tctx
->delta_stats
.num_whiteouts
;
8650 dout(20) << __func__
<< " creating whiteout on " << soid
<< dendl
;
8651 osd
->logger
->inc(l_osd_tier_whiteout
);
8653 if (results
->has_omap
) {
8654 dout(10) << __func__
<< " setting omap flag on " << soid
<< dendl
;
8655 tctx
->new_obs
.oi
.set_flag(object_info_t::FLAG_OMAP
);
8656 ++tctx
->delta_stats
.num_objects_omap
;
8659 results
->fill_in_final_tx(tctx
->op_t
.get());
8660 if (results
->started_temp_obj
) {
8661 tctx
->discard_temp_oid
= results
->temp_oid
;
8663 tctx
->new_obs
.oi
.size
= results
->object_size
;
8664 tctx
->new_obs
.oi
.user_version
= results
->user_version
;
8665 if (results
->is_data_digest()) {
8666 tctx
->new_obs
.oi
.set_data_digest(results
->data_digest
);
8668 tctx
->new_obs
.oi
.clear_data_digest();
8670 if (results
->is_omap_digest()) {
8671 tctx
->new_obs
.oi
.set_omap_digest(results
->omap_digest
);
8673 tctx
->new_obs
.oi
.clear_omap_digest();
8675 tctx
->new_obs
.oi
.truncate_seq
= results
->truncate_seq
;
8676 tctx
->new_obs
.oi
.truncate_size
= results
->truncate_size
;
8678 if (soid
.snap
!= CEPH_NOSNAP
) {
8679 if (legacy_snapset
) {
8680 tctx
->new_obs
.oi
.legacy_snaps
= results
->snaps
;
8681 assert(!tctx
->new_obs
.oi
.legacy_snaps
.empty());
8683 // it's already in the snapset
8684 assert(obc
->ssc
->snapset
.clone_snaps
.count(soid
.snap
));
8686 assert(obc
->ssc
->snapset
.clone_size
.count(soid
.snap
));
8687 assert(obc
->ssc
->snapset
.clone_size
[soid
.snap
] ==
8688 results
->object_size
);
8689 assert(obc
->ssc
->snapset
.clone_overlap
.count(soid
.snap
));
8691 tctx
->delta_stats
.num_bytes
+= obc
->ssc
->snapset
.get_clone_bytes(soid
.snap
);
8693 tctx
->delta_stats
.num_bytes
+= results
->object_size
;
8697 if (results
->mirror_snapset
) {
8698 assert(tctx
->new_obs
.oi
.soid
.snap
== CEPH_NOSNAP
);
8699 tctx
->new_snapset
.from_snap_set(
8701 get_osdmap()->require_osd_release
< CEPH_RELEASE_LUMINOUS
);
8703 tctx
->new_snapset
.head_exists
= true;
8704 dout(20) << __func__
<< " new_snapset " << tctx
->new_snapset
<< dendl
;
8706 // take RWWRITE lock for duration of our local write. ignore starvation.
8707 if (!tctx
->lock_manager
.take_write_lock(
8710 assert(0 == "problem!");
8712 dout(20) << __func__
<< " took lock on obc, " << obc
->rwstate
<< dendl
;
8714 finish_ctx(tctx
.get(), pg_log_entry_t::PROMOTE
);
8716 simple_opc_submit(std::move(tctx
));
8718 osd
->logger
->inc(l_osd_tier_promote
);
8721 agent_state
->is_idle())
8722 agent_choose_mode();
8725 void PrimaryLogPG::cancel_copy(CopyOpRef cop
, bool requeue
,
8726 vector
<ceph_tid_t
> *tids
)
8728 dout(10) << __func__
<< " " << cop
->obc
->obs
.oi
.soid
8729 << " from " << cop
->src
<< " " << cop
->oloc
8730 << " v" << cop
->results
.user_version
<< dendl
;
8732 // cancel objecter op, if we can
8733 if (cop
->objecter_tid
) {
8734 tids
->push_back(cop
->objecter_tid
);
8735 cop
->objecter_tid
= 0;
8736 if (cop
->objecter_tid2
) {
8737 tids
->push_back(cop
->objecter_tid2
);
8738 cop
->objecter_tid2
= 0;
8742 copy_ops
.erase(cop
->obc
->obs
.oi
.soid
);
8743 cop
->obc
->stop_block();
8745 kick_object_context_blocked(cop
->obc
);
8746 cop
->results
.should_requeue
= requeue
;
8747 CopyCallbackResults
result(-ECANCELED
, &cop
->results
);
8748 cop
->cb
->complete(result
);
8750 // There may still be an objecter callback referencing this copy op.
8751 // That callback will not need the obc since it's been canceled, and
8752 // we need the obc reference to go away prior to flush.
8753 cop
->obc
= ObjectContextRef();
8756 void PrimaryLogPG::cancel_copy_ops(bool requeue
, vector
<ceph_tid_t
> *tids
)
8758 dout(10) << __func__
<< dendl
;
8759 map
<hobject_t
,CopyOpRef
>::iterator p
= copy_ops
.begin();
8760 while (p
!= copy_ops
.end()) {
8761 // requeue this op? can I queue up all of them?
8762 cancel_copy((p
++)->second
, requeue
, tids
);
8767 // ========================================================================
8770 // Flush a dirty object in the cache tier by writing it back to the
8771 // base tier. The sequence looks like:
8773 // * send a copy-from operation to the base tier to copy the current
8774 // version of the object
8775 // * base tier will pull the object via (perhaps multiple) copy-get(s)
8776 // * on completion, we check if the object has been modified. if so,
8777 // just reply with -EAGAIN.
8778 // * try to take a write lock so we can clear the dirty flag. if this
8779 // fails, wait and retry
8780 // * start a repop that clears the bit.
8782 // If we have to wait, we will retry by coming back through the
8783 // start_flush method. We check if a flush is already in progress
8784 // and, if so, try to finish it by rechecking the version and trying
8785 // to clear the dirty bit.
8787 // In order for the cache-flush (a write op) to not block the copy-get
8788 // from reading the object, the client *must* set the SKIPRWLOCKS
8791 // NOTE: normally writes are strictly ordered for the client, but
8792 // flushes are special in that they can be reordered with respect to
8793 // other writes. In particular, we can't have a flush request block
8794 // an update to the cache pool object!
8796 struct C_Flush
: public Context
{
8799 epoch_t last_peering_reset
;
8802 C_Flush(PrimaryLogPG
*p
, hobject_t o
, epoch_t lpr
)
8803 : pg(p
), oid(o
), last_peering_reset(lpr
),
8804 tid(0), start(ceph_clock_now())
8806 void finish(int r
) override
{
8807 if (r
== -ECANCELED
)
8810 if (last_peering_reset
== pg
->get_last_peering_reset()) {
8811 pg
->finish_flush(oid
, tid
, r
);
8812 pg
->osd
->logger
->tinc(l_osd_tier_flush_lat
, ceph_clock_now() - start
);
8818 int PrimaryLogPG::start_flush(
8819 OpRequestRef op
, ObjectContextRef obc
,
8820 bool blocking
, hobject_t
*pmissing
,
8821 boost::optional
<std::function
<void()>> &&on_flush
)
8823 const object_info_t
& oi
= obc
->obs
.oi
;
8824 const hobject_t
& soid
= oi
.soid
;
8825 dout(10) << __func__
<< " " << soid
8826 << " v" << oi
.version
8827 << " uv" << oi
.user_version
8828 << " " << (blocking
? "blocking" : "non-blocking/best-effort")
8831 // get a filtered snapset, need to remove removed snaps
8832 SnapSet snapset
= obc
->ssc
->snapset
.get_filtered(pool
.info
);
8834 // verify there are no (older) check for dirty clones
8836 dout(20) << " snapset " << snapset
<< dendl
;
8837 vector
<snapid_t
>::reverse_iterator p
= snapset
.clones
.rbegin();
8838 while (p
!= snapset
.clones
.rend() && *p
>= soid
.snap
)
8840 if (p
!= snapset
.clones
.rend()) {
8841 hobject_t next
= soid
;
8843 assert(next
.snap
< soid
.snap
);
8844 if (pg_log
.get_missing().is_missing(next
)) {
8845 dout(10) << __func__
<< " missing clone is " << next
<< dendl
;
8850 ObjectContextRef older_obc
= get_object_context(next
, false);
8852 dout(20) << __func__
<< " next oldest clone is " << older_obc
->obs
.oi
8854 if (older_obc
->obs
.oi
.is_dirty()) {
8855 dout(10) << __func__
<< " next oldest clone is dirty: "
8856 << older_obc
->obs
.oi
<< dendl
;
8860 dout(20) << __func__
<< " next oldest clone " << next
8861 << " is not present; implicitly clean" << dendl
;
8864 dout(20) << __func__
<< " no older clones" << dendl
;
8871 map
<hobject_t
,FlushOpRef
>::iterator p
= flush_ops
.find(soid
);
8872 if (p
!= flush_ops
.end()) {
8873 FlushOpRef fop
= p
->second
;
8874 if (fop
->op
== op
) {
8875 // we couldn't take the write lock on a cache-try-flush before;
8876 // now we are trying again for the lock.
8877 return try_flush_mark_clean(fop
);
8879 if (fop
->flushed_version
== obc
->obs
.oi
.user_version
&&
8880 (fop
->blocking
|| !blocking
)) {
8881 // nonblocking can join anything
8882 // blocking can only join a blocking flush
8883 dout(20) << __func__
<< " piggybacking on existing flush " << dendl
;
8885 fop
->dup_ops
.push_back(op
);
8886 return -EAGAIN
; // clean up this ctx; op will retry later
8889 // cancel current flush since it will fail anyway, or because we
8890 // are blocking and the existing flush is nonblocking.
8891 dout(20) << __func__
<< " canceling previous flush; it will fail" << dendl
;
8893 osd
->reply_op_error(fop
->op
, -EBUSY
);
8894 while (!fop
->dup_ops
.empty()) {
8895 osd
->reply_op_error(fop
->dup_ops
.front(), -EBUSY
);
8896 fop
->dup_ops
.pop_front();
8898 vector
<ceph_tid_t
> tids
;
8899 cancel_flush(fop
, false, &tids
);
8900 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
8904 * In general, we need to send a delete and a copyfrom.
8905 * Consider snapc 10:[10, 9, 8, 4, 3, 2]:[10(10, 9), 4(4,3,2)]
8906 * where 4 is marked as clean. To flush 10, we have to:
8907 * 1) delete 4:[4,3,2] -- Logically, the object does not exist after 4
8908 * 2) copyfrom 8:[8,4,3,2] -- flush object after snap 8
8910 * There is a complicating case. Supposed there had been a clone 7
8911 * for snaps [7, 6] which has been trimmed since they no longer exist.
8912 * In the base pool, we'd have 5:[4,3,2]:[4(4,3,2)]+head. When we submit
8913 * the delete, the snap will be promoted to 5, and the head will become
8914 * a snapdir. When the copy-from goes through, we'll end up with
8915 * 8:[8,4,3,2]:[4(4,3,2)]+head.
8917 * Another complication is the case where there is an interval change
8918 * after doing the delete and the flush but before marking the object
8919 * clean. We'll happily delete head and then recreate it at the same
8920 * sequence number, which works out ok.
8923 SnapContext snapc
, dsnapc
;
8924 if (snapset
.seq
!= 0) {
8925 if (soid
.snap
== CEPH_NOSNAP
) {
8926 snapc
.seq
= snapset
.seq
;
8927 snapc
.snaps
= snapset
.snaps
;
8929 snapid_t min_included_snap
;
8930 if (snapset
.is_legacy()) {
8931 min_included_snap
= oi
.legacy_snaps
.back();
8933 auto p
= snapset
.clone_snaps
.find(soid
.snap
);
8934 assert(p
!= snapset
.clone_snaps
.end());
8935 min_included_snap
= p
->second
.back();
8937 snapc
= snapset
.get_ssc_as_of(min_included_snap
- 1);
8940 snapid_t prev_snapc
= 0;
8941 for (vector
<snapid_t
>::reverse_iterator citer
= snapset
.clones
.rbegin();
8942 citer
!= snapset
.clones
.rend();
8944 if (*citer
< soid
.snap
) {
8945 prev_snapc
= *citer
;
8950 dsnapc
= snapset
.get_ssc_as_of(prev_snapc
);
8953 object_locator_t
base_oloc(soid
);
8954 base_oloc
.pool
= pool
.info
.tier_of
;
8956 if (dsnapc
.seq
< snapc
.seq
) {
8959 osd
->objecter
->mutate(
8964 ceph::real_clock::from_ceph_timespec(oi
.mtime
),
8965 (CEPH_OSD_FLAG_IGNORE_OVERLAY
|
8966 CEPH_OSD_FLAG_ENFORCE_SNAPC
),
8967 NULL
/* no callback, we'll rely on the ordering w.r.t the next op */);
8970 FlushOpRef
fop(std::make_shared
<FlushOp
>());
8972 fop
->flushed_version
= oi
.user_version
;
8973 fop
->blocking
= blocking
;
8974 fop
->on_flush
= std::move(on_flush
);
8978 if (oi
.is_whiteout()) {
8979 fop
->removal
= true;
8982 object_locator_t
oloc(soid
);
8983 o
.copy_from(soid
.oid
.name
, soid
.snap
, oloc
, oi
.user_version
,
8984 CEPH_OSD_COPY_FROM_FLAG_FLUSH
|
8985 CEPH_OSD_COPY_FROM_FLAG_IGNORE_OVERLAY
|
8986 CEPH_OSD_COPY_FROM_FLAG_IGNORE_CACHE
|
8987 CEPH_OSD_COPY_FROM_FLAG_MAP_SNAP_CLONE
,
8988 LIBRADOS_OP_FLAG_FADVISE_SEQUENTIAL
|LIBRADOS_OP_FLAG_FADVISE_NOCACHE
);
8990 //mean the base tier don't cache data after this
8991 if (agent_state
&& agent_state
->evict_mode
!= TierAgentState::EVICT_MODE_FULL
)
8992 o
.set_last_op_flags(LIBRADOS_OP_FLAG_FADVISE_DONTNEED
);
8994 C_Flush
*fin
= new C_Flush(this, soid
, get_last_peering_reset());
8996 ceph_tid_t tid
= osd
->objecter
->mutate(
8997 soid
.oid
, base_oloc
, o
, snapc
,
8998 ceph::real_clock::from_ceph_timespec(oi
.mtime
),
8999 CEPH_OSD_FLAG_IGNORE_OVERLAY
| CEPH_OSD_FLAG_ENFORCE_SNAPC
,
9000 new C_OnFinisher(fin
,
9001 &osd
->objecter_finisher
));
9002 /* we're under the pg lock and fin->finish() is grabbing that */
9004 fop
->objecter_tid
= tid
;
9006 flush_ops
[soid
] = fop
;
9007 info
.stats
.stats
.sum
.num_flush
++;
9008 info
.stats
.stats
.sum
.num_flush_kb
+= SHIFT_ROUND_UP(oi
.size
, 10);
9009 return -EINPROGRESS
;
9012 void PrimaryLogPG::finish_flush(hobject_t oid
, ceph_tid_t tid
, int r
)
9014 dout(10) << __func__
<< " " << oid
<< " tid " << tid
9015 << " " << cpp_strerror(r
) << dendl
;
9016 map
<hobject_t
,FlushOpRef
>::iterator p
= flush_ops
.find(oid
);
9017 if (p
== flush_ops
.end()) {
9018 dout(10) << __func__
<< " no flush_op found" << dendl
;
9021 FlushOpRef fop
= p
->second
;
9022 if (tid
!= fop
->objecter_tid
) {
9023 dout(10) << __func__
<< " tid " << tid
<< " != fop " << fop
9024 << " tid " << fop
->objecter_tid
<< dendl
;
9027 ObjectContextRef obc
= fop
->obc
;
9028 fop
->objecter_tid
= 0;
9030 if (r
< 0 && !(r
== -ENOENT
&& fop
->removal
)) {
9032 osd
->reply_op_error(fop
->op
, -EBUSY
);
9033 if (fop
->blocking
) {
9035 kick_object_context_blocked(obc
);
9038 if (!fop
->dup_ops
.empty()) {
9039 dout(20) << __func__
<< " requeueing dups" << dendl
;
9040 requeue_ops(fop
->dup_ops
);
9042 if (fop
->on_flush
) {
9043 (*(fop
->on_flush
))();
9044 fop
->on_flush
= boost::none
;
9046 flush_ops
.erase(oid
);
9050 r
= try_flush_mark_clean(fop
);
9051 if (r
== -EBUSY
&& fop
->op
) {
9052 osd
->reply_op_error(fop
->op
, r
);
9056 int PrimaryLogPG::try_flush_mark_clean(FlushOpRef fop
)
9058 ObjectContextRef obc
= fop
->obc
;
9059 const hobject_t
& oid
= obc
->obs
.oi
.soid
;
9061 if (fop
->blocking
) {
9063 kick_object_context_blocked(obc
);
9066 if (fop
->flushed_version
!= obc
->obs
.oi
.user_version
||
9068 if (obc
->obs
.exists
)
9069 dout(10) << __func__
<< " flushed_version " << fop
->flushed_version
9070 << " != current " << obc
->obs
.oi
.user_version
9073 dout(10) << __func__
<< " object no longer exists" << dendl
;
9075 if (!fop
->dup_ops
.empty()) {
9076 dout(20) << __func__
<< " requeueing dups" << dendl
;
9077 requeue_ops(fop
->dup_ops
);
9079 if (fop
->on_flush
) {
9080 (*(fop
->on_flush
))();
9081 fop
->on_flush
= boost::none
;
9083 flush_ops
.erase(oid
);
9085 osd
->logger
->inc(l_osd_tier_flush_fail
);
9087 osd
->logger
->inc(l_osd_tier_try_flush_fail
);
9091 if (!fop
->blocking
&&
9092 write_blocked_by_scrub(oid
)) {
9094 dout(10) << __func__
<< " blocked by scrub" << dendl
;
9095 requeue_op(fop
->op
);
9096 requeue_ops(fop
->dup_ops
);
9097 return -EAGAIN
; // will retry
9099 osd
->logger
->inc(l_osd_tier_try_flush_fail
);
9100 vector
<ceph_tid_t
> tids
;
9101 cancel_flush(fop
, false, &tids
);
9102 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
9107 // successfully flushed, can we evict this object?
9108 if (!fop
->op
&& agent_state
->evict_mode
!= TierAgentState::EVICT_MODE_IDLE
&&
9109 agent_maybe_evict(obc
, true)) {
9110 osd
->logger
->inc(l_osd_tier_clean
);
9111 if (fop
->on_flush
) {
9112 (*(fop
->on_flush
))();
9113 fop
->on_flush
= boost::none
;
9115 flush_ops
.erase(oid
);
9119 dout(10) << __func__
<< " clearing DIRTY flag for " << oid
<< dendl
;
9120 OpContextUPtr ctx
= simple_opc_create(fop
->obc
);
9122 // successfully flushed; can we clear the dirty bit?
9123 // try to take the lock manually, since we don't
9125 if (ctx
->lock_manager
.get_lock_type(
9126 ObjectContext::RWState::RWWRITE
,
9130 dout(20) << __func__
<< " took write lock" << dendl
;
9131 } else if (fop
->op
) {
9132 dout(10) << __func__
<< " waiting on write lock " << fop
->op
<< " "
9133 << fop
->dup_ops
<< dendl
;
9134 close_op_ctx(ctx
.release());
9135 // fop->op is now waiting on the lock; get fop->dup_ops to wait too.
9136 for (auto op
: fop
->dup_ops
) {
9137 bool locked
= ctx
->lock_manager
.get_lock_type(
9138 ObjectContext::RWState::RWWRITE
,
9144 return -EAGAIN
; // will retry
9146 dout(10) << __func__
<< " failed write lock, no op; failing" << dendl
;
9147 close_op_ctx(ctx
.release());
9148 osd
->logger
->inc(l_osd_tier_try_flush_fail
);
9149 vector
<ceph_tid_t
> tids
;
9150 cancel_flush(fop
, false, &tids
);
9151 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
9155 if (fop
->on_flush
) {
9156 ctx
->register_on_finish(*(fop
->on_flush
));
9157 fop
->on_flush
= boost::none
;
9160 ctx
->at_version
= get_next_version();
9162 ctx
->new_obs
= obc
->obs
;
9163 ctx
->new_obs
.oi
.clear_flag(object_info_t::FLAG_DIRTY
);
9164 --ctx
->delta_stats
.num_objects_dirty
;
9166 finish_ctx(ctx
.get(), pg_log_entry_t::CLEAN
);
9168 osd
->logger
->inc(l_osd_tier_clean
);
9170 if (!fop
->dup_ops
.empty() || fop
->op
) {
9171 dout(20) << __func__
<< " requeueing for " << ctx
->at_version
<< dendl
;
9172 list
<OpRequestRef
> ls
;
9174 ls
.push_back(fop
->op
);
9175 ls
.splice(ls
.end(), fop
->dup_ops
);
9179 simple_opc_submit(std::move(ctx
));
9181 flush_ops
.erase(oid
);
9184 osd
->logger
->inc(l_osd_tier_flush
);
9186 osd
->logger
->inc(l_osd_tier_try_flush
);
9188 return -EINPROGRESS
;
9191 void PrimaryLogPG::cancel_flush(FlushOpRef fop
, bool requeue
,
9192 vector
<ceph_tid_t
> *tids
)
9194 dout(10) << __func__
<< " " << fop
->obc
->obs
.oi
.soid
<< " tid "
9195 << fop
->objecter_tid
<< dendl
;
9196 if (fop
->objecter_tid
) {
9197 tids
->push_back(fop
->objecter_tid
);
9198 fop
->objecter_tid
= 0;
9200 if (fop
->io_tids
.size()) {
9201 for (auto &p
: fop
->io_tids
) {
9202 tids
->push_back(p
.second
);
9206 if (fop
->blocking
&& fop
->obc
->is_blocked()) {
9207 fop
->obc
->stop_block();
9208 kick_object_context_blocked(fop
->obc
);
9212 requeue_op(fop
->op
);
9213 requeue_ops(fop
->dup_ops
);
9215 if (fop
->on_flush
) {
9216 (*(fop
->on_flush
))();
9217 fop
->on_flush
= boost::none
;
9219 flush_ops
.erase(fop
->obc
->obs
.oi
.soid
);
9222 void PrimaryLogPG::cancel_flush_ops(bool requeue
, vector
<ceph_tid_t
> *tids
)
9224 dout(10) << __func__
<< dendl
;
9225 map
<hobject_t
,FlushOpRef
>::iterator p
= flush_ops
.begin();
9226 while (p
!= flush_ops
.end()) {
9227 cancel_flush((p
++)->second
, requeue
, tids
);
9231 bool PrimaryLogPG::is_present_clone(hobject_t coid
)
9233 if (!pool
.info
.allow_incomplete_clones())
9235 if (is_missing_object(coid
))
9237 ObjectContextRef obc
= get_object_context(coid
, false);
9238 return obc
&& obc
->obs
.exists
;
9241 // ========================================================================
9244 class C_OSD_RepopApplied
: public Context
{
9246 boost::intrusive_ptr
<PrimaryLogPG::RepGather
> repop
;
9248 C_OSD_RepopApplied(PrimaryLogPG
*pg
, PrimaryLogPG::RepGather
*repop
)
9249 : pg(pg
), repop(repop
) {}
9250 void finish(int) override
{
9251 pg
->repop_all_applied(repop
.get());
9256 void PrimaryLogPG::repop_all_applied(RepGather
*repop
)
9258 dout(10) << __func__
<< ": repop tid " << repop
->rep_tid
<< " all applied "
9260 assert(!repop
->applies_with_commit
);
9261 repop
->all_applied
= true;
9262 if (!repop
->rep_aborted
) {
9267 class C_OSD_RepopCommit
: public Context
{
9269 boost::intrusive_ptr
<PrimaryLogPG::RepGather
> repop
;
9271 C_OSD_RepopCommit(PrimaryLogPG
*pg
, PrimaryLogPG::RepGather
*repop
)
9272 : pg(pg
), repop(repop
) {}
9273 void finish(int) override
{
9274 pg
->repop_all_committed(repop
.get());
9278 void PrimaryLogPG::repop_all_committed(RepGather
*repop
)
9280 dout(10) << __func__
<< ": repop tid " << repop
->rep_tid
<< " all committed "
9282 repop
->all_committed
= true;
9283 if (repop
->applies_with_commit
) {
9284 assert(!repop
->all_applied
);
9285 repop
->all_applied
= true;
9288 if (!repop
->rep_aborted
) {
9289 if (repop
->v
!= eversion_t()) {
9290 last_update_ondisk
= repop
->v
;
9291 last_complete_ondisk
= repop
->pg_local_last_complete
;
9297 void PrimaryLogPG::op_applied(const eversion_t
&applied_version
)
9299 dout(10) << "op_applied version " << applied_version
<< dendl
;
9300 if (applied_version
== eversion_t())
9302 assert(applied_version
> last_update_applied
);
9303 assert(applied_version
<= info
.last_update
);
9304 last_update_applied
= applied_version
;
9306 if (scrubber
.active
) {
9307 if (last_update_applied
>= scrubber
.subset_last_update
) {
9308 if (ops_blocked_by_scrub()) {
9309 requeue_scrub(true);
9311 requeue_scrub(false);
9316 assert(scrubber
.start
== scrubber
.end
);
9319 if (scrubber
.active_rep_scrub
) {
9320 if (last_update_applied
>= static_cast<const MOSDRepScrub
*>(
9321 scrubber
.active_rep_scrub
->get_req())->scrub_to
) {
9324 PGQueueable(scrubber
.active_rep_scrub
, get_osdmap()->get_epoch()));
9325 scrubber
.active_rep_scrub
= OpRequestRef();
9331 void PrimaryLogPG::eval_repop(RepGather
*repop
)
9333 const MOSDOp
*m
= NULL
;
9335 m
= static_cast<const MOSDOp
*>(repop
->op
->get_req());
9338 dout(10) << "eval_repop " << *repop
9339 << (repop
->rep_done
? " DONE" : "")
9342 dout(10) << "eval_repop " << *repop
<< " (no op)"
9343 << (repop
->rep_done
? " DONE" : "")
9346 if (repop
->rep_done
)
9350 if (repop
->all_committed
) {
9351 dout(10) << " commit: " << *repop
<< dendl
;
9352 for (auto p
= repop
->on_committed
.begin();
9353 p
!= repop
->on_committed
.end();
9354 repop
->on_committed
.erase(p
++)) {
9357 // send dup commits, in order
9358 if (waiting_for_ondisk
.count(repop
->v
)) {
9359 assert(waiting_for_ondisk
.begin()->first
== repop
->v
);
9360 for (list
<pair
<OpRequestRef
, version_t
> >::iterator i
=
9361 waiting_for_ondisk
[repop
->v
].begin();
9362 i
!= waiting_for_ondisk
[repop
->v
].end();
9364 osd
->reply_op_error(i
->first
, repop
->r
, repop
->v
,
9367 waiting_for_ondisk
.erase(repop
->v
);
9372 if (repop
->all_applied
) {
9373 if (repop
->applies_with_commit
) {
9374 assert(repop
->on_applied
.empty());
9376 dout(10) << " applied: " << *repop
<< " " << dendl
;
9377 for (auto p
= repop
->on_applied
.begin();
9378 p
!= repop
->on_applied
.end();
9379 repop
->on_applied
.erase(p
++)) {
9385 if (repop
->all_applied
&& repop
->all_committed
) {
9386 repop
->rep_done
= true;
9388 publish_stats_to_osd();
9389 calc_min_last_complete_ondisk();
9391 dout(10) << " removing " << *repop
<< dendl
;
9392 assert(!repop_queue
.empty());
9393 dout(20) << " q front is " << *repop_queue
.front() << dendl
;
9394 if (repop_queue
.front() != repop
) {
9395 if (!repop
->applies_with_commit
) {
9396 dout(0) << " removing " << *repop
<< dendl
;
9397 dout(0) << " q front is " << *repop_queue
.front() << dendl
;
9398 assert(repop_queue
.front() == repop
);
9401 RepGather
*to_remove
= nullptr;
9402 while (!repop_queue
.empty() &&
9403 (to_remove
= repop_queue
.front())->rep_done
) {
9404 repop_queue
.pop_front();
9405 for (auto p
= to_remove
->on_success
.begin();
9406 p
!= to_remove
->on_success
.end();
9407 to_remove
->on_success
.erase(p
++)) {
9410 remove_repop(to_remove
);
9416 void PrimaryLogPG::issue_repop(RepGather
*repop
, OpContext
*ctx
)
9419 const hobject_t
& soid
= ctx
->obs
->oi
.soid
;
9420 dout(7) << "issue_repop rep_tid " << repop
->rep_tid
9424 repop
->v
= ctx
->at_version
;
9425 if (ctx
->at_version
> eversion_t()) {
9426 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
9427 i
!= actingbackfill
.end();
9429 if (*i
== get_primary()) continue;
9430 pg_info_t
&pinfo
= peer_info
[*i
];
9431 // keep peer_info up to date
9432 if (pinfo
.last_complete
== pinfo
.last_update
)
9433 pinfo
.last_complete
= ctx
->at_version
;
9434 pinfo
.last_update
= ctx
->at_version
;
9438 ctx
->obc
->ondisk_write_lock();
9440 bool unlock_snapset_obc
= false;
9441 ctx
->op_t
->add_obc(ctx
->obc
);
9442 if (ctx
->clone_obc
) {
9443 ctx
->clone_obc
->ondisk_write_lock();
9444 ctx
->op_t
->add_obc(ctx
->clone_obc
);
9446 if (ctx
->snapset_obc
&& ctx
->snapset_obc
->obs
.oi
.soid
!=
9447 ctx
->obc
->obs
.oi
.soid
) {
9448 ctx
->snapset_obc
->ondisk_write_lock();
9449 unlock_snapset_obc
= true;
9450 ctx
->op_t
->add_obc(ctx
->snapset_obc
);
9453 Context
*on_all_commit
= new C_OSD_RepopCommit(this, repop
);
9454 Context
*on_all_applied
= new C_OSD_RepopApplied(this, repop
);
9455 Context
*onapplied_sync
= new C_OSD_OndiskWriteUnlock(
9458 unlock_snapset_obc
? ctx
->snapset_obc
: ObjectContextRef());
9459 if (!(ctx
->log
.empty())) {
9460 assert(ctx
->at_version
>= projected_last_update
);
9461 projected_last_update
= ctx
->at_version
;
9463 for (auto &&entry
: ctx
->log
) {
9464 projected_log
.add(entry
);
9466 pgbackend
->submit_transaction(
9470 std::move(ctx
->op_t
),
9472 min_last_complete_ondisk
,
9474 ctx
->updated_hset_history
,
9483 PrimaryLogPG::RepGather
*PrimaryLogPG::new_repop(
9484 OpContext
*ctx
, ObjectContextRef obc
,
9488 dout(10) << "new_repop rep_tid " << rep_tid
<< " on " << *ctx
->op
->get_req() << dendl
;
9490 dout(10) << "new_repop rep_tid " << rep_tid
<< " (no op)" << dendl
;
9492 RepGather
*repop
= new RepGather(
9493 ctx
, rep_tid
, info
.last_complete
, false);
9495 repop
->start
= ceph_clock_now();
9497 repop_queue
.push_back(&repop
->queue_item
);
9500 osd
->logger
->inc(l_osd_op_wip
);
9502 dout(10) << __func__
<< ": " << *repop
<< dendl
;
9506 boost::intrusive_ptr
<PrimaryLogPG::RepGather
> PrimaryLogPG::new_repop(
9509 ObcLockManager
&&manager
,
9511 boost::optional
<std::function
<void(void)> > &&on_complete
)
9513 RepGather
*repop
= new RepGather(
9516 std::move(on_complete
),
9523 repop
->start
= ceph_clock_now();
9525 repop_queue
.push_back(&repop
->queue_item
);
9527 osd
->logger
->inc(l_osd_op_wip
);
9529 dout(10) << __func__
<< ": " << *repop
<< dendl
;
9530 return boost::intrusive_ptr
<RepGather
>(repop
);
9533 void PrimaryLogPG::remove_repop(RepGather
*repop
)
9535 dout(20) << __func__
<< " " << *repop
<< dendl
;
9537 for (auto p
= repop
->on_finish
.begin();
9538 p
!= repop
->on_finish
.end();
9539 repop
->on_finish
.erase(p
++)) {
9543 release_object_locks(
9544 repop
->lock_manager
);
9547 osd
->logger
->dec(l_osd_op_wip
);
9550 PrimaryLogPG::OpContextUPtr
PrimaryLogPG::simple_opc_create(ObjectContextRef obc
)
9552 dout(20) << __func__
<< " " << obc
->obs
.oi
.soid
<< dendl
;
9553 ceph_tid_t rep_tid
= osd
->get_tid();
9554 osd_reqid_t
reqid(osd
->get_cluster_msgr_name(), 0, rep_tid
);
9555 OpContextUPtr
ctx(new OpContext(OpRequestRef(), reqid
, nullptr, obc
, this));
9556 ctx
->op_t
.reset(new PGTransaction());
9557 ctx
->mtime
= ceph_clock_now();
9561 void PrimaryLogPG::simple_opc_submit(OpContextUPtr ctx
)
9563 RepGather
*repop
= new_repop(ctx
.get(), ctx
->obc
, ctx
->reqid
.tid
);
9564 dout(20) << __func__
<< " " << repop
<< dendl
;
9565 issue_repop(repop
, ctx
.get());
9572 void PrimaryLogPG::submit_log_entries(
9573 const mempool::osd_pglog::list
<pg_log_entry_t
> &entries
,
9574 ObcLockManager
&&manager
,
9575 boost::optional
<std::function
<void(void)> > &&_on_complete
,
9579 dout(10) << __func__
<< " " << entries
<< dendl
;
9580 assert(is_primary());
9583 if (!entries
.empty()) {
9584 assert(entries
.rbegin()->version
>= projected_last_update
);
9585 version
= projected_last_update
= entries
.rbegin()->version
;
9588 boost::intrusive_ptr
<RepGather
> repop
;
9589 boost::optional
<std::function
<void(void)> > on_complete
;
9590 if (get_osdmap()->require_osd_release
>= CEPH_RELEASE_JEWEL
) {
9596 std::move(_on_complete
));
9598 on_complete
= std::move(_on_complete
);
9601 pgbackend
->call_write_ordered(
9602 [this, entries
, repop
, on_complete
]() {
9603 ObjectStore::Transaction t
;
9604 eversion_t old_last_update
= info
.last_update
;
9605 merge_new_log_entries(entries
, t
, pg_trim_to
, min_last_complete_ondisk
);
9608 set
<pg_shard_t
> waiting_on
;
9609 for (set
<pg_shard_t
>::const_iterator i
= actingbackfill
.begin();
9610 i
!= actingbackfill
.end();
9612 pg_shard_t
peer(*i
);
9613 if (peer
== pg_whoami
) continue;
9614 assert(peer_missing
.count(peer
));
9615 assert(peer_info
.count(peer
));
9616 if (get_osdmap()->require_osd_release
>= CEPH_RELEASE_JEWEL
) {
9618 MOSDPGUpdateLogMissing
*m
= new MOSDPGUpdateLogMissing(
9620 spg_t(info
.pgid
.pgid
, i
->shard
),
9622 get_osdmap()->get_epoch(),
9626 min_last_complete_ondisk
);
9627 osd
->send_message_osd_cluster(
9628 peer
.osd
, m
, get_osdmap()->get_epoch());
9629 waiting_on
.insert(peer
);
9631 MOSDPGLog
*m
= new MOSDPGLog(
9632 peer
.shard
, pg_whoami
.shard
,
9633 info
.last_update
.epoch
,
9635 m
->log
.log
= entries
;
9636 m
->log
.tail
= old_last_update
;
9637 m
->log
.head
= info
.last_update
;
9638 osd
->send_message_osd_cluster(
9639 peer
.osd
, m
, get_osdmap()->get_epoch());
9642 if (get_osdmap()->require_osd_release
>= CEPH_RELEASE_JEWEL
) {
9643 ceph_tid_t rep_tid
= repop
->rep_tid
;
9644 waiting_on
.insert(pg_whoami
);
9645 log_entry_update_waiting_on
.insert(
9648 LogUpdateCtx
{std::move(repop
), std::move(waiting_on
)}
9650 struct OnComplete
: public Context
{
9658 : pg(pg
), rep_tid(rep_tid
), epoch(epoch
) {}
9659 void finish(int) override
{
9661 if (!pg
->pg_has_reset_since(epoch
)) {
9662 auto it
= pg
->log_entry_update_waiting_on
.find(rep_tid
);
9663 assert(it
!= pg
->log_entry_update_waiting_on
.end());
9664 auto it2
= it
->second
.waiting_on
.find(pg
->pg_whoami
);
9665 assert(it2
!= it
->second
.waiting_on
.end());
9666 it
->second
.waiting_on
.erase(it2
);
9667 if (it
->second
.waiting_on
.empty()) {
9668 pg
->repop_all_committed(it
->second
.repop
.get());
9669 pg
->log_entry_update_waiting_on
.erase(it
);
9675 t
.register_on_commit(
9676 new OnComplete
{this, rep_tid
, get_osdmap()->get_epoch()});
9679 struct OnComplete
: public Context
{
9681 std::function
<void(void)> on_complete
;
9685 const std::function
<void(void)> &on_complete
,
9688 on_complete(std::move(on_complete
)),
9690 void finish(int) override
{
9692 if (!pg
->pg_has_reset_since(epoch
))
9697 t
.register_on_complete(
9699 this, *on_complete
, get_osdmap()->get_epoch()
9703 t
.register_on_applied(
9704 new C_OSD_OnApplied
{this, get_osdmap()->get_epoch(), info
.last_update
});
9705 int r
= osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
9712 void PrimaryLogPG::cancel_log_updates()
9714 // get rid of all the LogUpdateCtx so their references to repops are
9716 log_entry_update_waiting_on
.clear();
9719 // -------------------------------------------------------
9721 void PrimaryLogPG::get_watchers(list
<obj_watch_item_t
> &pg_watchers
)
9723 pair
<hobject_t
, ObjectContextRef
> i
;
9724 while (object_contexts
.get_next(i
.first
, &i
)) {
9725 ObjectContextRef
obc(i
.second
);
9726 get_obc_watchers(obc
, pg_watchers
);
9730 void PrimaryLogPG::get_obc_watchers(ObjectContextRef obc
, list
<obj_watch_item_t
> &pg_watchers
)
9732 for (map
<pair
<uint64_t, entity_name_t
>, WatchRef
>::iterator j
=
9733 obc
->watchers
.begin();
9734 j
!= obc
->watchers
.end();
9736 obj_watch_item_t owi
;
9738 owi
.obj
= obc
->obs
.oi
.soid
;
9739 owi
.wi
.addr
= j
->second
->get_peer_addr();
9740 owi
.wi
.name
= j
->second
->get_entity();
9741 owi
.wi
.cookie
= j
->second
->get_cookie();
9742 owi
.wi
.timeout_seconds
= j
->second
->get_timeout();
9744 dout(30) << "watch: Found oid=" << owi
.obj
<< " addr=" << owi
.wi
.addr
9745 << " name=" << owi
.wi
.name
<< " cookie=" << owi
.wi
.cookie
<< dendl
;
9747 pg_watchers
.push_back(owi
);
9751 void PrimaryLogPG::check_blacklisted_watchers()
9753 dout(20) << "PrimaryLogPG::check_blacklisted_watchers for pg " << get_pgid() << dendl
;
9754 pair
<hobject_t
, ObjectContextRef
> i
;
9755 while (object_contexts
.get_next(i
.first
, &i
))
9756 check_blacklisted_obc_watchers(i
.second
);
9759 void PrimaryLogPG::check_blacklisted_obc_watchers(ObjectContextRef obc
)
9761 dout(20) << "PrimaryLogPG::check_blacklisted_obc_watchers for obc " << obc
->obs
.oi
.soid
<< dendl
;
9762 for (map
<pair
<uint64_t, entity_name_t
>, WatchRef
>::iterator k
=
9763 obc
->watchers
.begin();
9764 k
!= obc
->watchers
.end();
9766 //Advance iterator now so handle_watch_timeout() can erase element
9767 map
<pair
<uint64_t, entity_name_t
>, WatchRef
>::iterator j
= k
++;
9768 dout(30) << "watch: Found " << j
->second
->get_entity() << " cookie " << j
->second
->get_cookie() << dendl
;
9769 entity_addr_t ea
= j
->second
->get_peer_addr();
9770 dout(30) << "watch: Check entity_addr_t " << ea
<< dendl
;
9771 if (get_osdmap()->is_blacklisted(ea
)) {
9772 dout(10) << "watch: Found blacklisted watcher for " << ea
<< dendl
;
9773 assert(j
->second
->get_pg() == this);
9774 j
->second
->unregister_cb();
9775 handle_watch_timeout(j
->second
);
9780 void PrimaryLogPG::populate_obc_watchers(ObjectContextRef obc
)
9782 assert(is_active());
9783 assert((recovering
.count(obc
->obs
.oi
.soid
) ||
9784 !is_missing_object(obc
->obs
.oi
.soid
)) ||
9785 (pg_log
.get_log().objects
.count(obc
->obs
.oi
.soid
) && // or this is a revert... see recover_primary()
9786 pg_log
.get_log().objects
.find(obc
->obs
.oi
.soid
)->second
->op
==
9787 pg_log_entry_t::LOST_REVERT
&&
9788 pg_log
.get_log().objects
.find(obc
->obs
.oi
.soid
)->second
->reverting_to
==
9789 obc
->obs
.oi
.version
));
9791 dout(10) << "populate_obc_watchers " << obc
->obs
.oi
.soid
<< dendl
;
9792 assert(obc
->watchers
.empty());
9793 // populate unconnected_watchers
9794 for (map
<pair
<uint64_t, entity_name_t
>, watch_info_t
>::iterator p
=
9795 obc
->obs
.oi
.watchers
.begin();
9796 p
!= obc
->obs
.oi
.watchers
.end();
9798 utime_t expire
= info
.stats
.last_became_active
;
9799 expire
+= p
->second
.timeout_seconds
;
9800 dout(10) << " unconnected watcher " << p
->first
<< " will expire " << expire
<< dendl
;
9802 Watch::makeWatchRef(
9803 this, osd
, obc
, p
->second
.timeout_seconds
, p
->first
.first
,
9804 p
->first
.second
, p
->second
.addr
));
9805 watch
->disconnect();
9806 obc
->watchers
.insert(
9808 make_pair(p
->first
.first
, p
->first
.second
),
9811 // Look for watchers from blacklisted clients and drop
9812 check_blacklisted_obc_watchers(obc
);
9815 void PrimaryLogPG::handle_watch_timeout(WatchRef watch
)
9817 ObjectContextRef obc
= watch
->get_obc(); // handle_watch_timeout owns this ref
9818 dout(10) << "handle_watch_timeout obc " << obc
<< dendl
;
9821 dout(10) << "handle_watch_timeout not active, no-op" << dendl
;
9824 if (is_degraded_or_backfilling_object(obc
->obs
.oi
.soid
)) {
9825 callbacks_for_degraded_object
[obc
->obs
.oi
.soid
].push_back(
9826 watch
->get_delayed_cb()
9828 dout(10) << "handle_watch_timeout waiting for degraded on obj "
9834 if (write_blocked_by_scrub(obc
->obs
.oi
.soid
)) {
9835 dout(10) << "handle_watch_timeout waiting for scrub on obj "
9838 scrubber
.add_callback(
9839 watch
->get_delayed_cb() // This callback!
9844 OpContextUPtr ctx
= simple_opc_create(obc
);
9845 ctx
->at_version
= get_next_version();
9847 object_info_t
& oi
= ctx
->new_obs
.oi
;
9848 oi
.watchers
.erase(make_pair(watch
->get_cookie(),
9849 watch
->get_entity()));
9851 list
<watch_disconnect_t
> watch_disconnects
= {
9852 watch_disconnect_t(watch
->get_cookie(), watch
->get_entity(), true)
9854 ctx
->register_on_success(
9855 [this, obc
, watch_disconnects
]() {
9856 complete_disconnect_watches(obc
, watch_disconnects
);
9860 PGTransaction
*t
= ctx
->op_t
.get();
9861 ctx
->log
.push_back(pg_log_entry_t(pg_log_entry_t::MODIFY
, obc
->obs
.oi
.soid
,
9865 osd_reqid_t(), ctx
->mtime
, 0));
9867 oi
.prior_version
= obc
->obs
.oi
.version
;
9868 oi
.version
= ctx
->at_version
;
9870 ::encode(oi
, bl
, get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
9871 t
->setattr(obc
->obs
.oi
.soid
, OI_ATTR
, bl
);
9873 // apply new object state.
9874 ctx
->obc
->obs
= ctx
->new_obs
;
9876 // no ctx->delta_stats
9877 simple_opc_submit(std::move(ctx
));
9880 ObjectContextRef
PrimaryLogPG::create_object_context(const object_info_t
& oi
,
9881 SnapSetContext
*ssc
)
9883 ObjectContextRef
obc(object_contexts
.lookup_or_create(oi
.soid
));
9884 assert(obc
->destructor_callback
== NULL
);
9885 obc
->destructor_callback
= new C_PG_ObjectContext(this, obc
.get());
9887 obc
->obs
.exists
= false;
9890 register_snapset_context(ssc
);
9891 dout(10) << "create_object_context " << (void*)obc
.get() << " " << oi
.soid
<< " " << dendl
;
9893 populate_obc_watchers(obc
);
9897 ObjectContextRef
PrimaryLogPG::get_object_context(
9898 const hobject_t
& soid
,
9900 const map
<string
, bufferlist
> *attrs
)
9903 attrs
|| !pg_log
.get_missing().is_missing(soid
) ||
9904 // or this is a revert... see recover_primary()
9905 (pg_log
.get_log().objects
.count(soid
) &&
9906 pg_log
.get_log().objects
.find(soid
)->second
->op
==
9907 pg_log_entry_t::LOST_REVERT
));
9908 ObjectContextRef obc
= object_contexts
.lookup(soid
);
9909 osd
->logger
->inc(l_osd_object_ctx_cache_total
);
9911 osd
->logger
->inc(l_osd_object_ctx_cache_hit
);
9912 dout(10) << __func__
<< ": found obc in cache: " << obc
9915 dout(10) << __func__
<< ": obc NOT found in cache: " << soid
<< dendl
;
9919 assert(attrs
->count(OI_ATTR
));
9920 bv
= attrs
->find(OI_ATTR
)->second
;
9922 int r
= pgbackend
->objects_get_attr(soid
, OI_ATTR
, &bv
);
9925 dout(10) << __func__
<< ": no obc for soid "
9926 << soid
<< " and !can_create"
9928 return ObjectContextRef(); // -ENOENT!
9931 dout(10) << __func__
<< ": no obc for soid "
9932 << soid
<< " but can_create"
9935 object_info_t
oi(soid
);
9936 SnapSetContext
*ssc
= get_snapset_context(
9937 soid
, true, 0, false);
9939 obc
= create_object_context(oi
, ssc
);
9940 dout(10) << __func__
<< ": " << obc
<< " " << soid
9941 << " " << obc
->rwstate
9942 << " oi: " << obc
->obs
.oi
9943 << " ssc: " << obc
->ssc
9944 << " snapset: " << obc
->ssc
->snapset
<< dendl
;
9951 bufferlist::iterator bliter
= bv
.begin();
9952 ::decode(oi
, bliter
);
9954 dout(0) << __func__
<< ": obc corrupt: " << soid
<< dendl
;
9955 return ObjectContextRef(); // -ENOENT!
9958 assert(oi
.soid
.pool
== (int64_t)info
.pgid
.pool());
9960 obc
= object_contexts
.lookup_or_create(oi
.soid
);
9961 obc
->destructor_callback
= new C_PG_ObjectContext(this, obc
.get());
9963 obc
->obs
.exists
= true;
9965 obc
->ssc
= get_snapset_context(
9967 soid
.has_snapset() ? attrs
: 0);
9970 populate_obc_watchers(obc
);
9972 if (pool
.info
.require_rollback()) {
9974 obc
->attr_cache
= *attrs
;
9976 int r
= pgbackend
->objects_get_attrs(
9983 dout(10) << __func__
<< ": creating obc from disk: " << obc
9987 // XXX: Caller doesn't expect this
9988 if (obc
->ssc
== NULL
) {
9989 derr
<< __func__
<< ": obc->ssc not available, not returning context" << dendl
;
9990 return ObjectContextRef(); // -ENOENT!
9993 dout(10) << __func__
<< ": " << obc
<< " " << soid
9994 << " " << obc
->rwstate
9995 << " oi: " << obc
->obs
.oi
9996 << " exists: " << (int)obc
->obs
.exists
9997 << " ssc: " << obc
->ssc
9998 << " snapset: " << obc
->ssc
->snapset
<< dendl
;
10002 void PrimaryLogPG::context_registry_on_change()
10004 pair
<hobject_t
, ObjectContextRef
> i
;
10005 while (object_contexts
.get_next(i
.first
, &i
)) {
10006 ObjectContextRef
obc(i
.second
);
10008 for (map
<pair
<uint64_t, entity_name_t
>, WatchRef
>::iterator j
=
10009 obc
->watchers
.begin();
10010 j
!= obc
->watchers
.end();
10011 obc
->watchers
.erase(j
++)) {
10012 j
->second
->discard();
10020 * If we return an error, and set *pmissing, then promoting that
10023 * If we return -EAGAIN, we will always set *pmissing to the missing
10024 * object to wait for.
10026 * If we return an error but do not set *pmissing, then we know the
10027 * object does not exist.
10029 int PrimaryLogPG::find_object_context(const hobject_t
& oid
,
10030 ObjectContextRef
*pobc
,
10032 bool map_snapid_to_clone
,
10033 hobject_t
*pmissing
)
10036 assert(oid
.pool
== static_cast<int64_t>(info
.pgid
.pool()));
10038 if (oid
.snap
== CEPH_NOSNAP
) {
10039 ObjectContextRef obc
= get_object_context(oid
, can_create
);
10045 dout(10) << "find_object_context " << oid
10046 << " @" << oid
.snap
10047 << " oi=" << obc
->obs
.oi
10054 hobject_t head
= oid
.get_head();
10056 // want the snapdir?
10057 if (oid
.snap
== CEPH_SNAPDIR
) {
10058 // return head or snapdir, whichever exists.
10059 ObjectContextRef headobc
= get_object_context(head
, can_create
);
10060 ObjectContextRef obc
= headobc
;
10061 if (!obc
|| !obc
->obs
.exists
)
10062 obc
= get_object_context(oid
, can_create
);
10063 if (!obc
|| !obc
->obs
.exists
) {
10064 // if we have neither, we would want to promote the head.
10068 *pobc
= headobc
; // may be null
10071 dout(10) << "find_object_context " << oid
10072 << " @" << oid
.snap
10073 << " oi=" << obc
->obs
.oi
10077 // always populate ssc for SNAPDIR...
10079 obc
->ssc
= get_snapset_context(
10085 if (!map_snapid_to_clone
&& pool
.info
.is_removed_snap(oid
.snap
)) {
10086 dout(10) << __func__
<< " snap " << oid
.snap
<< " is removed" << dendl
;
10090 SnapSetContext
*ssc
= get_snapset_context(oid
, can_create
);
10091 if (!ssc
|| !(ssc
->exists
|| can_create
)) {
10092 dout(20) << __func__
<< " " << oid
<< " no snapset" << dendl
;
10094 *pmissing
= head
; // start by getting the head
10096 put_snapset_context(ssc
);
10100 if (map_snapid_to_clone
) {
10101 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10102 << " snapset " << ssc
->snapset
10103 << " map_snapid_to_clone=true" << dendl
;
10104 if (oid
.snap
> ssc
->snapset
.seq
) {
10105 // already must be readable
10106 ObjectContextRef obc
= get_object_context(head
, false);
10107 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10108 << " snapset " << ssc
->snapset
10109 << " maps to head" << dendl
;
10111 put_snapset_context(ssc
);
10112 return (obc
&& obc
->obs
.exists
) ? 0 : -ENOENT
;
10114 vector
<snapid_t
>::const_iterator citer
= std::find(
10115 ssc
->snapset
.clones
.begin(),
10116 ssc
->snapset
.clones
.end(),
10118 if (citer
== ssc
->snapset
.clones
.end()) {
10119 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10120 << " snapset " << ssc
->snapset
10121 << " maps to nothing" << dendl
;
10122 put_snapset_context(ssc
);
10126 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10127 << " snapset " << ssc
->snapset
10128 << " maps to " << oid
<< dendl
;
10130 if (pg_log
.get_missing().is_missing(oid
)) {
10131 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10132 << " snapset " << ssc
->snapset
10133 << " " << oid
<< " is missing" << dendl
;
10136 put_snapset_context(ssc
);
10140 ObjectContextRef obc
= get_object_context(oid
, false);
10141 if (!obc
|| !obc
->obs
.exists
) {
10142 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10143 << " snapset " << ssc
->snapset
10144 << " " << oid
<< " is not present" << dendl
;
10147 put_snapset_context(ssc
);
10150 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10151 << " snapset " << ssc
->snapset
10152 << " " << oid
<< " HIT" << dendl
;
10154 put_snapset_context(ssc
);
10157 ceph_abort(); //unreachable
10160 dout(10) << "find_object_context " << oid
<< " @" << oid
.snap
10161 << " snapset " << ssc
->snapset
<< dendl
;
10164 if (oid
.snap
> ssc
->snapset
.seq
) {
10165 if (ssc
->snapset
.head_exists
) {
10166 ObjectContextRef obc
= get_object_context(head
, false);
10167 dout(10) << "find_object_context " << head
10168 << " want " << oid
.snap
<< " > snapset seq " << ssc
->snapset
.seq
10169 << " -- HIT " << obc
->obs
10174 assert(ssc
== obc
->ssc
);
10175 put_snapset_context(ssc
);
10180 dout(10) << "find_object_context " << head
10181 << " want " << oid
.snap
<< " > snapset seq " << ssc
->snapset
.seq
10182 << " but head dne -- DNE"
10184 put_snapset_context(ssc
);
10188 // which clone would it be?
10190 while (k
< ssc
->snapset
.clones
.size() &&
10191 ssc
->snapset
.clones
[k
] < oid
.snap
)
10193 if (k
== ssc
->snapset
.clones
.size()) {
10194 dout(10) << "find_object_context no clones with last >= oid.snap "
10195 << oid
.snap
<< " -- DNE" << dendl
;
10196 put_snapset_context(ssc
);
10199 hobject_t
soid(oid
.oid
, oid
.get_key(), ssc
->snapset
.clones
[k
], oid
.get_hash(),
10200 info
.pgid
.pool(), oid
.get_namespace());
10202 if (pg_log
.get_missing().is_missing(soid
)) {
10203 dout(20) << "find_object_context " << soid
<< " missing, try again later"
10207 put_snapset_context(ssc
);
10211 ObjectContextRef obc
= get_object_context(soid
, false);
10212 if (!obc
|| !obc
->obs
.exists
) {
10215 put_snapset_context(ssc
);
10216 if (is_degraded_or_backfilling_object(soid
)) {
10217 dout(20) << __func__
<< " clone is degraded or backfilling " << soid
<< dendl
;
10220 dout(20) << __func__
<< " missing clone " << soid
<< dendl
;
10228 assert(obc
->ssc
== ssc
);
10229 put_snapset_context(ssc
);
10234 dout(20) << "find_object_context " << soid
10235 << " snapset " << obc
->ssc
->snapset
10236 << " legacy_snaps " << obc
->obs
.oi
.legacy_snaps
10238 snapid_t first
, last
;
10239 if (obc
->ssc
->snapset
.is_legacy()) {
10240 first
= obc
->obs
.oi
.legacy_snaps
.back();
10241 last
= obc
->obs
.oi
.legacy_snaps
.front();
10243 auto p
= obc
->ssc
->snapset
.clone_snaps
.find(soid
.snap
);
10244 assert(p
!= obc
->ssc
->snapset
.clone_snaps
.end());
10245 if (p
->second
.empty()) {
10246 dout(1) << __func__
<< " " << soid
<< " empty snapset -- DNE" << dendl
;
10247 assert(!cct
->_conf
->osd_debug_verify_snaps
);
10250 first
= p
->second
.back();
10251 last
= p
->second
.front();
10253 if (first
<= oid
.snap
) {
10254 dout(20) << "find_object_context " << soid
<< " [" << first
<< "," << last
10255 << "] contains " << oid
.snap
<< " -- HIT " << obc
->obs
<< dendl
;
10259 dout(20) << "find_object_context " << soid
<< " [" << first
<< "," << last
10260 << "] does not contain " << oid
.snap
<< " -- DNE" << dendl
;
10265 void PrimaryLogPG::object_context_destructor_callback(ObjectContext
*obc
)
10268 put_snapset_context(obc
->ssc
);
10271 void PrimaryLogPG::add_object_context_to_pg_stat(ObjectContextRef obc
, pg_stat_t
*pgstat
)
10273 object_info_t
& oi
= obc
->obs
.oi
;
10275 dout(10) << "add_object_context_to_pg_stat " << oi
.soid
<< dendl
;
10276 object_stat_sum_t stat
;
10278 stat
.num_bytes
+= oi
.size
;
10280 if (oi
.soid
.snap
!= CEPH_SNAPDIR
)
10281 stat
.num_objects
++;
10283 stat
.num_objects_dirty
++;
10284 if (oi
.is_whiteout())
10285 stat
.num_whiteouts
++;
10287 stat
.num_objects_omap
++;
10288 if (oi
.is_cache_pinned())
10289 stat
.num_objects_pinned
++;
10291 if (oi
.soid
.snap
&& oi
.soid
.snap
!= CEPH_NOSNAP
&& oi
.soid
.snap
!= CEPH_SNAPDIR
) {
10292 stat
.num_object_clones
++;
10295 obc
->ssc
= get_snapset_context(oi
.soid
, false);
10298 // subtract off clone overlap
10299 if (obc
->ssc
->snapset
.clone_overlap
.count(oi
.soid
.snap
)) {
10300 interval_set
<uint64_t>& o
= obc
->ssc
->snapset
.clone_overlap
[oi
.soid
.snap
];
10301 for (interval_set
<uint64_t>::const_iterator r
= o
.begin();
10304 stat
.num_bytes
-= r
.get_len();
10310 pgstat
->stats
.sum
.add(stat
);
10313 void PrimaryLogPG::kick_object_context_blocked(ObjectContextRef obc
)
10315 const hobject_t
& soid
= obc
->obs
.oi
.soid
;
10316 if (obc
->is_blocked()) {
10317 dout(10) << __func__
<< " " << soid
<< " still blocked" << dendl
;
10321 map
<hobject_t
, list
<OpRequestRef
>>::iterator p
= waiting_for_blocked_object
.find(soid
);
10322 if (p
!= waiting_for_blocked_object
.end()) {
10323 list
<OpRequestRef
>& ls
= p
->second
;
10324 dout(10) << __func__
<< " " << soid
<< " requeuing " << ls
.size() << " requests" << dendl
;
10326 waiting_for_blocked_object
.erase(p
);
10329 map
<hobject_t
, ObjectContextRef
>::iterator i
=
10330 objects_blocked_on_snap_promotion
.find(obc
->obs
.oi
.soid
.get_head());
10331 if (i
!= objects_blocked_on_snap_promotion
.end()) {
10332 assert(i
->second
== obc
);
10333 objects_blocked_on_snap_promotion
.erase(i
);
10336 if (obc
->requeue_scrub_on_unblock
) {
10337 obc
->requeue_scrub_on_unblock
= false;
10342 SnapSetContext
*PrimaryLogPG::get_snapset_context(
10343 const hobject_t
& oid
,
10345 const map
<string
, bufferlist
> *attrs
,
10348 Mutex::Locker
l(snapset_contexts_lock
);
10349 SnapSetContext
*ssc
;
10350 map
<hobject_t
, SnapSetContext
*>::iterator p
= snapset_contexts
.find(
10351 oid
.get_snapdir());
10352 if (p
!= snapset_contexts
.end()) {
10353 if (can_create
|| p
->second
->exists
) {
10362 if (!(oid
.is_head() && !oid_existed
))
10363 r
= pgbackend
->objects_get_attr(oid
.get_head(), SS_ATTR
, &bv
);
10366 if (!(oid
.is_snapdir() && !oid_existed
))
10367 r
= pgbackend
->objects_get_attr(oid
.get_snapdir(), SS_ATTR
, &bv
);
10368 if (r
< 0 && !can_create
)
10372 assert(attrs
->count(SS_ATTR
));
10373 bv
= attrs
->find(SS_ATTR
)->second
;
10375 ssc
= new SnapSetContext(oid
.get_snapdir());
10376 _register_snapset_context(ssc
);
10378 bufferlist::iterator bvp
= bv
.begin();
10380 ssc
->snapset
.decode(bvp
);
10381 } catch (buffer::error
& e
) {
10382 dout(0) << __func__
<< " Can't decode snapset: " << e
<< dendl
;
10385 ssc
->exists
= true;
10387 ssc
->exists
= false;
10395 void PrimaryLogPG::put_snapset_context(SnapSetContext
*ssc
)
10397 Mutex::Locker
l(snapset_contexts_lock
);
10399 if (ssc
->ref
== 0) {
10400 if (ssc
->registered
)
10401 snapset_contexts
.erase(ssc
->oid
);
10406 /** pull - request object from a peer
10411 * NONE - didn't pull anything
10412 * YES - pulled what the caller wanted
10413 * OTHER - needed to pull something else first (_head or _snapdir)
10415 enum { PULL_NONE
, PULL_OTHER
, PULL_YES
};
10417 int PrimaryLogPG::recover_missing(
10418 const hobject_t
&soid
, eversion_t v
,
10420 PGBackend::RecoveryHandle
*h
)
10422 if (missing_loc
.is_unfound(soid
)) {
10423 dout(7) << "pull " << soid
10425 << " but it is unfound" << dendl
;
10429 if (missing_loc
.is_deleted(soid
)) {
10430 start_recovery_op(soid
);
10431 assert(!recovering
.count(soid
));
10432 recovering
.insert(make_pair(soid
, ObjectContextRef()));
10433 epoch_t cur_epoch
= get_osdmap()->get_epoch();
10434 remove_missing_object(soid
, v
, new FunctionContext(
10437 if (!pg_has_reset_since(cur_epoch
)) {
10438 bool object_missing
= false;
10439 for (const auto& shard
: actingbackfill
) {
10440 if (shard
== pg_whoami
)
10442 if (peer_missing
[shard
].is_missing(soid
)) {
10443 dout(20) << __func__
<< ": soid " << soid
<< " needs to be deleted from replica " << shard
<< dendl
;
10444 object_missing
= true;
10448 if (!object_missing
) {
10449 object_stat_sum_t stat_diff
;
10450 stat_diff
.num_objects_recovered
= 1;
10451 on_global_recover(soid
, stat_diff
, true);
10453 auto recovery_handle
= pgbackend
->open_recovery_op();
10454 pgbackend
->recover_delete_object(soid
, v
, recovery_handle
);
10455 pgbackend
->run_recovery_op(recovery_handle
, priority
);
10463 // is this a snapped object? if so, consult the snapset.. we may not need the entire object!
10464 ObjectContextRef obc
;
10465 ObjectContextRef head_obc
;
10466 if (soid
.snap
&& soid
.snap
< CEPH_NOSNAP
) {
10467 // do we have the head and/or snapdir?
10468 hobject_t head
= soid
.get_head();
10469 if (pg_log
.get_missing().is_missing(head
)) {
10470 if (recovering
.count(head
)) {
10471 dout(10) << " missing but already recovering head " << head
<< dendl
;
10474 int r
= recover_missing(
10475 head
, pg_log
.get_missing().get_items().find(head
)->second
.need
, priority
,
10477 if (r
!= PULL_NONE
)
10482 head
= soid
.get_snapdir();
10483 if (pg_log
.get_missing().is_missing(head
)) {
10484 if (recovering
.count(head
)) {
10485 dout(10) << " missing but already recovering snapdir " << head
<< dendl
;
10488 int r
= recover_missing(
10489 head
, pg_log
.get_missing().get_items().find(head
)->second
.need
, priority
,
10491 if (r
!= PULL_NONE
)
10497 // we must have one or the other
10498 head_obc
= get_object_context(
10503 head_obc
= get_object_context(
10504 soid
.get_snapdir(),
10509 start_recovery_op(soid
);
10510 assert(!recovering
.count(soid
));
10511 recovering
.insert(make_pair(soid
, obc
));
10512 int r
= pgbackend
->recover_object(
10518 // This is only a pull which shouldn't return an error
10523 void PrimaryLogPG::send_remove_op(
10524 const hobject_t
& oid
, eversion_t v
, pg_shard_t peer
)
10526 ceph_tid_t tid
= osd
->get_tid();
10527 osd_reqid_t
rid(osd
->get_cluster_msgr_name(), 0, tid
);
10529 dout(10) << "send_remove_op " << oid
<< " from osd." << peer
10530 << " tid " << tid
<< dendl
;
10532 MOSDSubOp
*subop
= new MOSDSubOp(
10533 rid
, pg_whoami
, spg_t(info
.pgid
.pgid
, peer
.shard
),
10534 oid
, CEPH_OSD_FLAG_ACK
,
10535 get_osdmap()->get_epoch(), tid
, v
);
10536 subop
->ops
= vector
<OSDOp
>(1);
10537 subop
->ops
[0].op
.op
= CEPH_OSD_OP_DELETE
;
10539 osd
->send_message_osd_cluster(peer
.osd
, subop
, get_osdmap()->get_epoch());
10542 void PrimaryLogPG::remove_missing_object(const hobject_t
&soid
,
10543 eversion_t v
, Context
*on_complete
)
10545 dout(20) << __func__
<< " " << soid
<< " " << v
<< dendl
;
10546 assert(on_complete
!= nullptr);
10548 ObjectStore::Transaction t
;
10549 remove_snap_mapped_object(t
, soid
);
10551 ObjectRecoveryInfo recovery_info
;
10552 recovery_info
.soid
= soid
;
10553 recovery_info
.version
= v
;
10555 epoch_t cur_epoch
= get_osdmap()->get_epoch();
10556 t
.register_on_complete(new FunctionContext(
10559 if (!pg_has_reset_since(cur_epoch
)) {
10560 ObjectStore::Transaction t2
;
10561 on_local_recover(soid
, recovery_info
, ObjectContextRef(), true, &t2
);
10562 t2
.register_on_complete(on_complete
);
10563 int r
= osd
->store
->queue_transaction(osr
.get(), std::move(t2
), nullptr);
10568 on_complete
->complete(-EAGAIN
);
10571 int r
= osd
->store
->queue_transaction(osr
.get(), std::move(t
), nullptr);
10575 void PrimaryLogPG::finish_degraded_object(const hobject_t
& oid
)
10577 dout(10) << "finish_degraded_object " << oid
<< dendl
;
10578 if (callbacks_for_degraded_object
.count(oid
)) {
10579 list
<Context
*> contexts
;
10580 contexts
.swap(callbacks_for_degraded_object
[oid
]);
10581 callbacks_for_degraded_object
.erase(oid
);
10582 for (list
<Context
*>::iterator i
= contexts
.begin();
10583 i
!= contexts
.end();
10588 map
<hobject_t
, snapid_t
>::iterator i
= objects_blocked_on_degraded_snap
.find(
10590 if (i
!= objects_blocked_on_degraded_snap
.end() &&
10591 i
->second
== oid
.snap
)
10592 objects_blocked_on_degraded_snap
.erase(i
);
10595 void PrimaryLogPG::_committed_pushed_object(
10596 epoch_t epoch
, eversion_t last_complete
)
10599 if (!pg_has_reset_since(epoch
)) {
10600 dout(10) << "_committed_pushed_object last_complete " << last_complete
<< " now ondisk" << dendl
;
10601 last_complete_ondisk
= last_complete
;
10603 if (last_complete_ondisk
== info
.last_update
) {
10604 if (!is_primary()) {
10605 // Either we are a replica or backfill target.
10606 // we are fully up to date. tell the primary!
10607 osd
->send_message_osd_cluster(
10610 get_osdmap()->get_epoch(),
10611 spg_t(info
.pgid
.pgid
, get_primary().shard
),
10612 last_complete_ondisk
),
10613 get_osdmap()->get_epoch());
10615 calc_min_last_complete_ondisk();
10620 dout(10) << "_committed_pushed_object pg has changed, not touching last_complete_ondisk" << dendl
;
10626 void PrimaryLogPG::_applied_recovered_object(ObjectContextRef obc
)
10629 dout(20) << __func__
<< dendl
;
10631 dout(20) << "obc = " << *obc
<< dendl
;
10633 assert(active_pushes
>= 1);
10636 // requeue an active chunky scrub waiting on recovery ops
10637 if (!deleting
&& active_pushes
== 0
10638 && scrubber
.is_chunky_scrub_active()) {
10639 if (ops_blocked_by_scrub()) {
10640 requeue_scrub(true);
10642 requeue_scrub(false);
10648 void PrimaryLogPG::_applied_recovered_object_replica()
10651 dout(20) << __func__
<< dendl
;
10652 assert(active_pushes
>= 1);
10655 // requeue an active chunky scrub waiting on recovery ops
10656 if (!deleting
&& active_pushes
== 0 &&
10657 scrubber
.active_rep_scrub
&& static_cast<const MOSDRepScrub
*>(
10658 scrubber
.active_rep_scrub
->get_req())->chunky
) {
10661 PGQueueable(scrubber
.active_rep_scrub
, get_osdmap()->get_epoch()));
10662 scrubber
.active_rep_scrub
= OpRequestRef();
10667 void PrimaryLogPG::recover_got(hobject_t oid
, eversion_t v
)
10669 dout(10) << "got missing " << oid
<< " v " << v
<< dendl
;
10670 pg_log
.recover_got(oid
, v
, info
);
10671 if (pg_log
.get_log().complete_to
!= pg_log
.get_log().log
.end()) {
10672 dout(10) << "last_complete now " << info
.last_complete
10673 << " log.complete_to " << pg_log
.get_log().complete_to
->version
10676 dout(10) << "last_complete now " << info
.last_complete
10677 << " log.complete_to at end" << dendl
;
10678 //below is not true in the repair case.
10679 //assert(missing.num_missing() == 0); // otherwise, complete_to was wrong.
10680 assert(info
.last_complete
== info
.last_update
);
10684 void PrimaryLogPG::primary_failed(const hobject_t
&soid
)
10686 list
<pg_shard_t
> fl
= { pg_whoami
};
10687 failed_push(fl
, soid
);
10690 void PrimaryLogPG::failed_push(const list
<pg_shard_t
> &from
, const hobject_t
&soid
)
10692 dout(20) << __func__
<< ": " << soid
<< dendl
;
10693 assert(recovering
.count(soid
));
10694 auto obc
= recovering
[soid
];
10696 list
<OpRequestRef
> blocked_ops
;
10697 obc
->drop_recovery_read(&blocked_ops
);
10698 requeue_ops(blocked_ops
);
10700 recovering
.erase(soid
);
10701 for (auto&& i
: from
)
10702 missing_loc
.remove_location(soid
, i
);
10703 dout(0) << __func__
<< " " << soid
<< " from shard " << from
10704 << ", reps on " << missing_loc
.get_locations(soid
)
10705 << " unfound? " << missing_loc
.is_unfound(soid
) << dendl
;
10706 finish_recovery_op(soid
); // close out this attempt,
10709 void PrimaryLogPG::sub_op_remove(OpRequestRef op
)
10711 const MOSDSubOp
*m
= static_cast<const MOSDSubOp
*>(op
->get_req());
10712 assert(m
->get_type() == MSG_OSD_SUBOP
);
10713 dout(7) << "sub_op_remove " << m
->poid
<< dendl
;
10715 op
->mark_started();
10717 ObjectStore::Transaction t
;
10718 remove_snap_mapped_object(t
, m
->poid
);
10719 int r
= osd
->store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
10723 eversion_t
PrimaryLogPG::pick_newest_available(const hobject_t
& oid
)
10726 pg_missing_item pmi
;
10727 bool is_missing
= pg_log
.get_missing().is_missing(oid
, &pmi
);
10728 assert(is_missing
);
10730 dout(10) << "pick_newest_available " << oid
<< " " << v
<< " on osd." << osd
->whoami
<< " (local)" << dendl
;
10732 assert(!actingbackfill
.empty());
10733 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
10734 i
!= actingbackfill
.end();
10736 if (*i
== get_primary()) continue;
10737 pg_shard_t peer
= *i
;
10738 if (!peer_missing
[peer
].is_missing(oid
)) {
10741 eversion_t h
= peer_missing
[peer
].get_items().at(oid
).have
;
10742 dout(10) << "pick_newest_available " << oid
<< " " << h
<< " on osd." << peer
<< dendl
;
10747 dout(10) << "pick_newest_available " << oid
<< " " << v
<< " (newest)" << dendl
;
10751 void PrimaryLogPG::do_update_log_missing(OpRequestRef
&op
)
10753 const MOSDPGUpdateLogMissing
*m
= static_cast<const MOSDPGUpdateLogMissing
*>(
10755 assert(m
->get_type() == MSG_OSD_PG_UPDATE_LOG_MISSING
);
10756 ObjectStore::Transaction t
;
10757 boost::optional
<eversion_t
> op_trim_to
, op_roll_forward_to
;
10758 if (m
->pg_trim_to
!= eversion_t())
10759 op_trim_to
= m
->pg_trim_to
;
10760 if (m
->pg_roll_forward_to
!= eversion_t())
10761 op_roll_forward_to
= m
->pg_roll_forward_to
;
10763 dout(20) << __func__
<< " op_trim_to = " << op_trim_to
<< " op_roll_forward_to = " << op_roll_forward_to
<< dendl
;
10765 append_log_entries_update_missing(m
->entries
, t
, op_trim_to
, op_roll_forward_to
);
10766 eversion_t new_lcod
= info
.last_complete
;
10768 Context
*complete
= new FunctionContext(
10770 const MOSDPGUpdateLogMissing
*msg
= static_cast<const MOSDPGUpdateLogMissing
*>(
10773 if (!pg_has_reset_since(msg
->get_epoch())) {
10774 update_last_complete_ondisk(new_lcod
);
10775 MOSDPGUpdateLogMissingReply
*reply
=
10776 new MOSDPGUpdateLogMissingReply(
10777 spg_t(info
.pgid
.pgid
, primary_shard().shard
),
10783 reply
->set_priority(CEPH_MSG_PRIO_HIGH
);
10784 msg
->get_connection()->send_message(reply
);
10789 if (get_osdmap()->require_osd_release
>= CEPH_RELEASE_KRAKEN
) {
10790 t
.register_on_commit(complete
);
10792 /* Hack to work around the fact that ReplicatedBackend sends
10793 * ack+commit if commit happens first
10795 * This behavior is no longer necessary, but we preserve it so old
10796 * primaries can keep their repops in order */
10797 if (pool
.info
.ec_pool()) {
10798 t
.register_on_complete(complete
);
10800 t
.register_on_commit(complete
);
10803 t
.register_on_applied(
10804 new C_OSD_OnApplied
{this, get_osdmap()->get_epoch(), info
.last_update
});
10805 int tr
= osd
->store
->queue_transaction(
10812 void PrimaryLogPG::do_update_log_missing_reply(OpRequestRef
&op
)
10814 const MOSDPGUpdateLogMissingReply
*m
=
10815 static_cast<const MOSDPGUpdateLogMissingReply
*>(
10817 dout(20) << __func__
<< " got reply from "
10818 << m
->get_from() << dendl
;
10820 auto it
= log_entry_update_waiting_on
.find(m
->get_tid());
10821 if (it
!= log_entry_update_waiting_on
.end()) {
10822 if (it
->second
.waiting_on
.count(m
->get_from())) {
10823 it
->second
.waiting_on
.erase(m
->get_from());
10824 if (m
->last_complete_ondisk
!= eversion_t()) {
10825 update_peer_last_complete_ondisk(m
->get_from(), m
->last_complete_ondisk
);
10829 << info
.pgid
<< " got reply "
10830 << *m
<< " from shard we are not waiting for "
10834 if (it
->second
.waiting_on
.empty()) {
10835 repop_all_committed(it
->second
.repop
.get());
10836 log_entry_update_waiting_on
.erase(it
);
10840 << info
.pgid
<< " got reply "
10841 << *m
<< " on unknown tid " << m
->get_tid();
10845 /* Mark all unfound objects as lost.
10847 void PrimaryLogPG::mark_all_unfound_lost(
10852 dout(3) << __func__
<< " " << pg_log_entry_t::get_op_name(what
) << dendl
;
10853 list
<hobject_t
> oids
;
10855 dout(30) << __func__
<< ": log before:\n";
10856 pg_log
.get_log().print(*_dout
);
10859 mempool::osd_pglog::list
<pg_log_entry_t
> log_entries
;
10861 utime_t mtime
= ceph_clock_now();
10862 map
<hobject_t
, pg_missing_item
>::const_iterator m
=
10863 missing_loc
.get_needs_recovery().begin();
10864 map
<hobject_t
, pg_missing_item
>::const_iterator mend
=
10865 missing_loc
.get_needs_recovery().end();
10867 ObcLockManager manager
;
10868 eversion_t v
= get_next_version();
10869 v
.epoch
= get_osdmap()->get_epoch();
10870 uint64_t num_unfound
= missing_loc
.num_unfound();
10871 while (m
!= mend
) {
10872 const hobject_t
&oid(m
->first
);
10873 if (!missing_loc
.is_unfound(oid
)) {
10874 // We only care about unfound objects
10879 ObjectContextRef obc
;
10883 case pg_log_entry_t::LOST_MARK
:
10884 assert(0 == "actually, not implemented yet!");
10887 case pg_log_entry_t::LOST_REVERT
:
10888 prev
= pick_newest_available(oid
);
10889 if (prev
> eversion_t()) {
10892 pg_log_entry_t::LOST_REVERT
, oid
, v
,
10893 m
->second
.need
, 0, osd_reqid_t(), mtime
, 0);
10894 e
.reverting_to
= prev
;
10895 e
.mark_unrollbackable();
10896 log_entries
.push_back(e
);
10897 dout(10) << e
<< dendl
;
10899 // we are now missing the new version; recovery code will sort it out.
10905 case pg_log_entry_t::LOST_DELETE
:
10907 pg_log_entry_t
e(pg_log_entry_t::LOST_DELETE
, oid
, v
, m
->second
.need
,
10908 0, osd_reqid_t(), mtime
, 0);
10909 if (get_osdmap()->require_osd_release
>= CEPH_RELEASE_JEWEL
) {
10910 if (pool
.info
.require_rollback()) {
10911 e
.mod_desc
.try_rmobject(v
.version
);
10913 e
.mark_unrollbackable();
10915 } // otherwise, just do what we used to do
10916 dout(10) << e
<< dendl
;
10917 log_entries
.push_back(e
);
10918 oids
.push_back(oid
);
10920 // If context found mark object as deleted in case
10921 // of racing with new creation. This can happen if
10922 // object lost and EIO at primary.
10923 obc
= object_contexts
.lookup(oid
);
10925 obc
->obs
.exists
= false;
10937 info
.stats
.stats_invalid
= true;
10939 submit_log_entries(
10941 std::move(manager
),
10942 boost::optional
<std::function
<void(void)> >(
10943 [this, oids
, con
, num_unfound
, tid
]() {
10944 if (perform_deletes_during_peering()) {
10945 for (auto oid
: oids
) {
10946 // clear old locations - merge_new_log_entries will have
10947 // handled rebuilding missing_loc for each of these
10948 // objects if we have the RECOVERY_DELETES flag
10949 missing_loc
.recovered(oid
);
10953 if (is_recovery_unfound()) {
10954 queue_peering_event(
10956 std::make_shared
<CephPeeringEvt
>(
10957 get_osdmap()->get_epoch(),
10958 get_osdmap()->get_epoch(),
10960 } else if (is_backfill_unfound()) {
10961 queue_peering_event(
10963 std::make_shared
<CephPeeringEvt
>(
10964 get_osdmap()->get_epoch(),
10965 get_osdmap()->get_epoch(),
10966 RequestBackfill())));
10972 ss
<< "pg has " << num_unfound
10973 << " objects unfound and apparently lost marking";
10974 string rs
= ss
.str();
10975 dout(0) << "do_command r=" << 0 << " " << rs
<< dendl
;
10976 osd
->clog
->info() << rs
;
10978 MCommandReply
*reply
= new MCommandReply(0, rs
);
10979 reply
->set_tid(tid
);
10980 con
->send_message(reply
);
10986 void PrimaryLogPG::_split_into(pg_t child_pgid
, PG
*child
, unsigned split_bits
)
10988 assert(repop_queue
.empty());
10992 * pg status change notification
10995 void PrimaryLogPG::apply_and_flush_repops(bool requeue
)
10997 list
<OpRequestRef
> rq
;
10999 // apply all repops
11000 while (!repop_queue
.empty()) {
11001 RepGather
*repop
= repop_queue
.front();
11002 repop_queue
.pop_front();
11003 dout(10) << " canceling repop tid " << repop
->rep_tid
<< dendl
;
11004 repop
->rep_aborted
= true;
11005 repop
->on_applied
.clear();
11006 repop
->on_committed
.clear();
11007 repop
->on_success
.clear();
11011 dout(10) << " requeuing " << *repop
->op
->get_req() << dendl
;
11012 rq
.push_back(repop
->op
);
11013 repop
->op
= OpRequestRef();
11016 // also requeue any dups, interleaved into position
11017 map
<eversion_t
, list
<pair
<OpRequestRef
, version_t
> > >::iterator p
=
11018 waiting_for_ondisk
.find(repop
->v
);
11019 if (p
!= waiting_for_ondisk
.end()) {
11020 dout(10) << " also requeuing ondisk waiters " << p
->second
<< dendl
;
11021 for (list
<pair
<OpRequestRef
, version_t
> >::iterator i
=
11023 i
!= p
->second
.end();
11025 rq
.push_back(i
->first
);
11027 waiting_for_ondisk
.erase(p
);
11031 remove_repop(repop
);
11034 assert(repop_queue
.empty());
11038 if (!waiting_for_ondisk
.empty()) {
11039 for (map
<eversion_t
, list
<pair
<OpRequestRef
, version_t
> > >::iterator i
=
11040 waiting_for_ondisk
.begin();
11041 i
!= waiting_for_ondisk
.end();
11043 for (list
<pair
<OpRequestRef
, version_t
> >::iterator j
=
11045 j
!= i
->second
.end();
11047 derr
<< __func__
<< ": op " << *(j
->first
->get_req()) << " waiting on "
11048 << i
->first
<< dendl
;
11051 assert(waiting_for_ondisk
.empty());
11055 waiting_for_ondisk
.clear();
11058 void PrimaryLogPG::on_flushed()
11060 assert(flushes_in_progress
> 0);
11061 flushes_in_progress
--;
11062 if (flushes_in_progress
== 0) {
11063 requeue_ops(waiting_for_flush
);
11065 if (!is_peered() || !is_primary()) {
11066 pair
<hobject_t
, ObjectContextRef
> i
;
11067 while (object_contexts
.get_next(i
.first
, &i
)) {
11068 derr
<< "on_flushed: object " << i
.first
<< " obc still alive" << dendl
;
11070 assert(object_contexts
.empty());
11072 pgbackend
->on_flushed();
11075 void PrimaryLogPG::on_removal(ObjectStore::Transaction
*t
)
11077 dout(10) << "on_removal" << dendl
;
11079 // adjust info to backfill
11080 info
.set_last_backfill(hobject_t());
11081 pg_log
.reset_backfill();
11086 PGLogEntryHandler rollbacker
{this, t
};
11087 pg_log
.roll_forward(&rollbacker
);
11089 write_if_dirty(*t
);
11095 void PrimaryLogPG::clear_async_reads()
11097 dout(10) << __func__
<< dendl
;
11098 for(auto& i
: in_progress_async_reads
) {
11099 dout(10) << "clear ctx: "
11100 << "OpRequestRef " << i
.first
11101 << " OpContext " << i
.second
11103 close_op_ctx(i
.second
);
11107 void PrimaryLogPG::on_shutdown()
11109 dout(10) << "on_shutdown" << dendl
;
11111 // remove from queues
11112 osd
->pg_stat_queue_dequeue(this);
11113 osd
->peering_wq
.dequeue(this);
11115 // handles queue races
11118 if (recovery_queued
) {
11119 recovery_queued
= false;
11120 osd
->clear_queued_recovery(this);
11123 clear_scrub_reserved();
11124 scrub_clear_state();
11126 unreg_next_scrub();
11128 vector
<ceph_tid_t
> tids
;
11129 cancel_copy_ops(false, &tids
);
11130 cancel_flush_ops(false, &tids
);
11131 cancel_proxy_ops(false, &tids
);
11132 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
11134 apply_and_flush_repops(false);
11135 cancel_log_updates();
11136 // we must remove PGRefs, so do this this prior to release_backoffs() callers
11138 // clean up snap trim references
11139 snap_trimmer_machine
.process_event(Reset());
11141 pgbackend
->on_change();
11143 context_registry_on_change();
11144 object_contexts
.clear();
11146 clear_async_reads();
11148 osd
->remote_reserver
.cancel_reservation(info
.pgid
);
11149 osd
->local_reserver
.cancel_reservation(info
.pgid
);
11151 clear_primary_state();
11155 void PrimaryLogPG::on_activate()
11158 if (needs_recovery()) {
11159 dout(10) << "activate not all replicas are up-to-date, queueing recovery" << dendl
;
11160 queue_peering_event(
11162 std::make_shared
<CephPeeringEvt
>(
11163 get_osdmap()->get_epoch(),
11164 get_osdmap()->get_epoch(),
11166 } else if (needs_backfill()) {
11167 dout(10) << "activate queueing backfill" << dendl
;
11168 queue_peering_event(
11170 std::make_shared
<CephPeeringEvt
>(
11171 get_osdmap()->get_epoch(),
11172 get_osdmap()->get_epoch(),
11173 RequestBackfill())));
11175 dout(10) << "activate all replicas clean, no recovery" << dendl
;
11176 eio_errors_to_process
= false;
11177 queue_peering_event(
11179 std::make_shared
<CephPeeringEvt
>(
11180 get_osdmap()->get_epoch(),
11181 get_osdmap()->get_epoch(),
11182 AllReplicasRecovered())));
11185 publish_stats_to_osd();
11187 if (!backfill_targets
.empty()) {
11188 last_backfill_started
= earliest_backfill();
11189 new_backfill
= true;
11190 assert(!last_backfill_started
.is_max());
11191 dout(5) << "on activate: bft=" << backfill_targets
11192 << " from " << last_backfill_started
<< dendl
;
11193 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
11194 i
!= backfill_targets
.end();
11196 dout(5) << "target shard " << *i
11197 << " from " << peer_info
[*i
].last_backfill
11206 void PrimaryLogPG::_on_new_interval()
11208 dout(20) << __func__
<< " checking missing set deletes flag. missing = " << pg_log
.get_missing() << dendl
;
11209 if (!pg_log
.get_missing().may_include_deletes
&&
11210 get_osdmap()->test_flag(CEPH_OSDMAP_RECOVERY_DELETES
)) {
11211 pg_log
.rebuild_missing_set_with_deletes(osd
->store
, coll
, info
);
11213 assert(pg_log
.get_missing().may_include_deletes
== get_osdmap()->test_flag(CEPH_OSDMAP_RECOVERY_DELETES
));
11216 void PrimaryLogPG::on_change(ObjectStore::Transaction
*t
)
11218 dout(10) << "on_change" << dendl
;
11220 if (hit_set
&& hit_set
->insert_count() == 0) {
11221 dout(20) << " discarding empty hit_set" << dendl
;
11225 if (recovery_queued
) {
11226 recovery_queued
= false;
11227 osd
->clear_queued_recovery(this);
11230 // requeue everything in the reverse order they should be
11232 requeue_ops(waiting_for_peered
);
11233 requeue_ops(waiting_for_flush
);
11234 requeue_ops(waiting_for_active
);
11236 clear_scrub_reserved();
11238 vector
<ceph_tid_t
> tids
;
11239 cancel_copy_ops(is_primary(), &tids
);
11240 cancel_flush_ops(is_primary(), &tids
);
11241 cancel_proxy_ops(is_primary(), &tids
);
11242 osd
->objecter
->op_cancel(tids
, -ECANCELED
);
11244 // requeue object waiters
11245 for (auto& p
: waiting_for_unreadable_object
) {
11246 release_backoffs(p
.first
);
11248 if (is_primary()) {
11249 requeue_object_waiters(waiting_for_unreadable_object
);
11251 waiting_for_unreadable_object
.clear();
11253 for (map
<hobject_t
,list
<OpRequestRef
>>::iterator p
= waiting_for_degraded_object
.begin();
11254 p
!= waiting_for_degraded_object
.end();
11255 waiting_for_degraded_object
.erase(p
++)) {
11256 release_backoffs(p
->first
);
11258 requeue_ops(p
->second
);
11261 finish_degraded_object(p
->first
);
11264 // requeues waiting_for_scrub
11265 scrub_clear_state();
11267 for (auto p
= waiting_for_blocked_object
.begin();
11268 p
!= waiting_for_blocked_object
.end();
11269 waiting_for_blocked_object
.erase(p
++)) {
11271 requeue_ops(p
->second
);
11275 for (auto i
= callbacks_for_degraded_object
.begin();
11276 i
!= callbacks_for_degraded_object
.end();
11278 finish_degraded_object((i
++)->first
);
11280 assert(callbacks_for_degraded_object
.empty());
11282 if (is_primary()) {
11283 requeue_ops(waiting_for_cache_not_full
);
11285 waiting_for_cache_not_full
.clear();
11287 objects_blocked_on_cache_full
.clear();
11289 for (list
<pair
<OpRequestRef
, OpContext
*> >::iterator i
=
11290 in_progress_async_reads
.begin();
11291 i
!= in_progress_async_reads
.end();
11292 in_progress_async_reads
.erase(i
++)) {
11293 close_op_ctx(i
->second
);
11295 requeue_op(i
->first
);
11298 // this will requeue ops we were working on but didn't finish, and
11300 apply_and_flush_repops(is_primary());
11301 cancel_log_updates();
11303 // do this *after* apply_and_flush_repops so that we catch any newly
11304 // registered watches.
11305 context_registry_on_change();
11307 pgbackend
->on_change_cleanup(t
);
11308 scrubber
.cleanup_store(t
);
11309 pgbackend
->on_change();
11311 // clear snap_trimmer state
11312 snap_trimmer_machine
.process_event(Reset());
11314 debug_op_order
.clear();
11315 unstable_stats
.clear();
11317 // we don't want to cache object_contexts through the interval change
11318 // NOTE: we actually assert that all currently live references are dead
11319 // by the time the flush for the next interval completes.
11320 object_contexts
.clear();
11322 // should have been cleared above by finishing all of the degraded objects
11323 assert(objects_blocked_on_degraded_snap
.empty());
11326 void PrimaryLogPG::on_role_change()
11328 dout(10) << "on_role_change" << dendl
;
11329 if (get_role() != 0 && hit_set
) {
11330 dout(10) << " clearing hit set" << dendl
;
11335 void PrimaryLogPG::on_pool_change()
11337 dout(10) << __func__
<< dendl
;
11338 // requeue cache full waiters just in case the cache_mode is
11339 // changing away from writeback mode. note that if we are not
11340 // active the normal requeuing machinery is sufficient (and properly
11343 pool
.info
.cache_mode
!= pg_pool_t::CACHEMODE_WRITEBACK
&&
11344 !waiting_for_cache_not_full
.empty()) {
11345 dout(10) << __func__
<< " requeuing full waiters (not in writeback) "
11347 requeue_ops(waiting_for_cache_not_full
);
11348 objects_blocked_on_cache_full
.clear();
11354 // clear state. called on recovery completion AND cancellation.
11355 void PrimaryLogPG::_clear_recovery_state()
11357 missing_loc
.clear();
11358 #ifdef DEBUG_RECOVERY_OIDS
11359 recovering_oids
.clear();
11361 last_backfill_started
= hobject_t();
11362 set
<hobject_t
>::iterator i
= backfills_in_flight
.begin();
11363 while (i
!= backfills_in_flight
.end()) {
11364 assert(recovering
.count(*i
));
11365 backfills_in_flight
.erase(i
++);
11368 list
<OpRequestRef
> blocked_ops
;
11369 for (map
<hobject_t
, ObjectContextRef
>::iterator i
= recovering
.begin();
11370 i
!= recovering
.end();
11371 recovering
.erase(i
++)) {
11373 i
->second
->drop_recovery_read(&blocked_ops
);
11374 requeue_ops(blocked_ops
);
11377 assert(backfills_in_flight
.empty());
11378 pending_backfill_updates
.clear();
11379 assert(recovering
.empty());
11380 pgbackend
->clear_recovery_state();
11383 void PrimaryLogPG::cancel_pull(const hobject_t
&soid
)
11385 dout(20) << __func__
<< ": " << soid
<< dendl
;
11386 assert(recovering
.count(soid
));
11387 ObjectContextRef obc
= recovering
[soid
];
11389 list
<OpRequestRef
> blocked_ops
;
11390 obc
->drop_recovery_read(&blocked_ops
);
11391 requeue_ops(blocked_ops
);
11393 recovering
.erase(soid
);
11394 finish_recovery_op(soid
);
11395 release_backoffs(soid
);
11396 if (waiting_for_degraded_object
.count(soid
)) {
11397 dout(20) << " kicking degraded waiters on " << soid
<< dendl
;
11398 requeue_ops(waiting_for_degraded_object
[soid
]);
11399 waiting_for_degraded_object
.erase(soid
);
11401 if (waiting_for_unreadable_object
.count(soid
)) {
11402 dout(20) << " kicking unreadable waiters on " << soid
<< dendl
;
11403 requeue_ops(waiting_for_unreadable_object
[soid
]);
11404 waiting_for_unreadable_object
.erase(soid
);
11406 if (is_missing_object(soid
))
11407 pg_log
.set_last_requested(0); // get recover_primary to start over
11408 finish_degraded_object(soid
);
11411 void PrimaryLogPG::check_recovery_sources(const OSDMapRef
& osdmap
)
11414 * check that any peers we are planning to (or currently) pulling
11415 * objects from are dealt with.
11417 missing_loc
.check_recovery_sources(osdmap
);
11418 pgbackend
->check_recovery_sources(osdmap
);
11420 for (set
<pg_shard_t
>::iterator i
= peer_log_requested
.begin();
11421 i
!= peer_log_requested
.end();
11423 if (!osdmap
->is_up(i
->osd
)) {
11424 dout(10) << "peer_log_requested removing " << *i
<< dendl
;
11425 peer_log_requested
.erase(i
++);
11431 for (set
<pg_shard_t
>::iterator i
= peer_missing_requested
.begin();
11432 i
!= peer_missing_requested
.end();
11434 if (!osdmap
->is_up(i
->osd
)) {
11435 dout(10) << "peer_missing_requested removing " << *i
<< dendl
;
11436 peer_missing_requested
.erase(i
++);
11443 void PG::MissingLoc::check_recovery_sources(const OSDMapRef
& osdmap
)
11445 set
<pg_shard_t
> now_down
;
11446 for (set
<pg_shard_t
>::iterator p
= missing_loc_sources
.begin();
11447 p
!= missing_loc_sources
.end();
11449 if (osdmap
->is_up(p
->osd
)) {
11453 ldout(pg
->cct
, 10) << "check_recovery_sources source osd." << *p
<< " now down" << dendl
;
11454 now_down
.insert(*p
);
11455 missing_loc_sources
.erase(p
++);
11458 if (now_down
.empty()) {
11459 ldout(pg
->cct
, 10) << "check_recovery_sources no source osds (" << missing_loc_sources
<< ") went down" << dendl
;
11461 ldout(pg
->cct
, 10) << "check_recovery_sources sources osds " << now_down
<< " now down, remaining sources are "
11462 << missing_loc_sources
<< dendl
;
11464 // filter missing_loc
11465 map
<hobject_t
, set
<pg_shard_t
>>::iterator p
= missing_loc
.begin();
11466 while (p
!= missing_loc
.end()) {
11467 set
<pg_shard_t
>::iterator q
= p
->second
.begin();
11468 while (q
!= p
->second
.end())
11469 if (now_down
.count(*q
)) {
11470 p
->second
.erase(q
++);
11474 if (p
->second
.empty())
11475 missing_loc
.erase(p
++);
11483 bool PrimaryLogPG::start_recovery_ops(
11485 ThreadPool::TPHandle
&handle
,
11486 uint64_t *ops_started
)
11488 uint64_t& started
= *ops_started
;
11490 bool work_in_progress
= false;
11491 assert(is_primary());
11493 if (!state_test(PG_STATE_RECOVERING
) &&
11494 !state_test(PG_STATE_BACKFILLING
)) {
11495 /* TODO: I think this case is broken and will make do_recovery()
11496 * unhappy since we're returning false */
11497 dout(10) << "recovery raced and were queued twice, ignoring!" << dendl
;
11501 const auto &missing
= pg_log
.get_missing();
11503 unsigned int num_missing
= missing
.num_missing();
11504 uint64_t num_unfound
= get_num_unfound();
11506 if (num_missing
== 0) {
11507 info
.last_complete
= info
.last_update
;
11510 if (num_missing
== num_unfound
) {
11511 // All of the missing objects we have are unfound.
11512 // Recover the replicas.
11513 started
= recover_replicas(max
, handle
);
11516 // We still have missing objects that we should grab from replicas.
11517 started
+= recover_primary(max
, handle
);
11519 if (!started
&& num_unfound
!= get_num_unfound()) {
11520 // second chance to recovery replicas
11521 started
= recover_replicas(max
, handle
);
11525 work_in_progress
= true;
11527 bool deferred_backfill
= false;
11528 if (recovering
.empty() &&
11529 state_test(PG_STATE_BACKFILLING
) &&
11530 !backfill_targets
.empty() && started
< max
&&
11531 missing
.num_missing() == 0 &&
11532 waiting_on_backfill
.empty()) {
11533 if (get_osdmap()->test_flag(CEPH_OSDMAP_NOBACKFILL
)) {
11534 dout(10) << "deferring backfill due to NOBACKFILL" << dendl
;
11535 deferred_backfill
= true;
11536 } else if (get_osdmap()->test_flag(CEPH_OSDMAP_NOREBALANCE
) &&
11538 dout(10) << "deferring backfill due to NOREBALANCE" << dendl
;
11539 deferred_backfill
= true;
11540 } else if (!backfill_reserved
) {
11541 dout(10) << "deferring backfill due to !backfill_reserved" << dendl
;
11542 if (!backfill_reserving
) {
11543 dout(10) << "queueing RequestBackfill" << dendl
;
11544 backfill_reserving
= true;
11545 queue_peering_event(
11547 std::make_shared
<CephPeeringEvt
>(
11548 get_osdmap()->get_epoch(),
11549 get_osdmap()->get_epoch(),
11550 RequestBackfill())));
11552 deferred_backfill
= true;
11554 started
+= recover_backfill(max
- started
, handle
, &work_in_progress
);
11558 dout(10) << " started " << started
<< dendl
;
11559 osd
->logger
->inc(l_osd_rop
, started
);
11561 if (!recovering
.empty() ||
11562 work_in_progress
|| recovery_ops_active
> 0 || deferred_backfill
)
11563 return work_in_progress
;
11565 assert(recovering
.empty());
11566 assert(recovery_ops_active
== 0);
11568 dout(10) << __func__
<< " needs_recovery: "
11569 << missing_loc
.get_needs_recovery()
11571 dout(10) << __func__
<< " missing_loc: "
11572 << missing_loc
.get_missing_locs()
11574 int unfound
= get_num_unfound();
11576 dout(10) << " still have " << unfound
<< " unfound" << dendl
;
11577 return work_in_progress
;
11580 if (missing
.num_missing() > 0) {
11581 // this shouldn't happen!
11582 osd
->clog
->error() << info
.pgid
<< " Unexpected Error: recovery ending with "
11583 << missing
.num_missing() << ": " << missing
.get_items();
11584 return work_in_progress
;
11587 if (needs_recovery()) {
11588 // this shouldn't happen!
11589 // We already checked num_missing() so we must have missing replicas
11590 osd
->clog
->error() << info
.pgid
11591 << " Unexpected Error: recovery ending with missing replicas";
11592 return work_in_progress
;
11595 if (state_test(PG_STATE_RECOVERING
)) {
11596 state_clear(PG_STATE_RECOVERING
);
11597 state_clear(PG_STATE_FORCED_RECOVERY
);
11598 if (needs_backfill()) {
11599 dout(10) << "recovery done, queuing backfill" << dendl
;
11600 queue_peering_event(
11602 std::make_shared
<CephPeeringEvt
>(
11603 get_osdmap()->get_epoch(),
11604 get_osdmap()->get_epoch(),
11605 RequestBackfill())));
11607 dout(10) << "recovery done, no backfill" << dendl
;
11608 eio_errors_to_process
= false;
11609 state_clear(PG_STATE_FORCED_BACKFILL
);
11610 queue_peering_event(
11612 std::make_shared
<CephPeeringEvt
>(
11613 get_osdmap()->get_epoch(),
11614 get_osdmap()->get_epoch(),
11615 AllReplicasRecovered())));
11617 } else { // backfilling
11618 state_clear(PG_STATE_BACKFILLING
);
11619 state_clear(PG_STATE_FORCED_BACKFILL
);
11620 state_clear(PG_STATE_FORCED_RECOVERY
);
11621 dout(10) << "recovery done, backfill done" << dendl
;
11622 eio_errors_to_process
= false;
11623 queue_peering_event(
11625 std::make_shared
<CephPeeringEvt
>(
11626 get_osdmap()->get_epoch(),
11627 get_osdmap()->get_epoch(),
11635 * do one recovery op.
11636 * return true if done, false if nothing left to do.
11638 uint64_t PrimaryLogPG::recover_primary(uint64_t max
, ThreadPool::TPHandle
&handle
)
11640 assert(is_primary());
11642 const auto &missing
= pg_log
.get_missing();
11644 dout(10) << "recover_primary recovering " << recovering
.size()
11645 << " in pg" << dendl
;
11646 dout(10) << "recover_primary " << missing
<< dendl
;
11647 dout(25) << "recover_primary " << missing
.get_items() << dendl
;
11650 pg_log_entry_t
*latest
= 0;
11651 unsigned started
= 0;
11654 PGBackend::RecoveryHandle
*h
= pgbackend
->open_recovery_op();
11655 map
<version_t
, hobject_t
>::const_iterator p
=
11656 missing
.get_rmissing().lower_bound(pg_log
.get_log().last_requested
);
11657 while (p
!= missing
.get_rmissing().end()) {
11658 handle
.reset_tp_timeout();
11660 version_t v
= p
->first
;
11662 if (pg_log
.get_log().objects
.count(p
->second
)) {
11663 latest
= pg_log
.get_log().objects
.find(p
->second
)->second
;
11664 assert(latest
->is_update() || latest
->is_delete());
11665 soid
= latest
->soid
;
11670 const pg_missing_item
& item
= missing
.get_items().find(p
->second
)->second
;
11673 hobject_t head
= soid
.get_head();
11675 eversion_t need
= item
.need
;
11677 dout(10) << "recover_primary "
11678 << soid
<< " " << item
.need
11679 << (missing
.is_missing(soid
) ? " (missing)":"")
11680 << (missing
.is_missing(head
) ? " (missing head)":"")
11681 << (recovering
.count(soid
) ? " (recovering)":"")
11682 << (recovering
.count(head
) ? " (recovering head)":"")
11686 switch (latest
->op
) {
11687 case pg_log_entry_t::CLONE
:
11689 * Handling for this special case removed for now, until we
11690 * can correctly construct an accurate SnapSet from the old
11695 case pg_log_entry_t::LOST_REVERT
:
11697 if (item
.have
== latest
->reverting_to
) {
11698 ObjectContextRef obc
= get_object_context(soid
, true);
11700 if (obc
->obs
.oi
.version
== latest
->version
) {
11701 // I'm already reverting
11702 dout(10) << " already reverting " << soid
<< dendl
;
11704 dout(10) << " reverting " << soid
<< " to " << latest
->prior_version
<< dendl
;
11705 obc
->ondisk_write_lock();
11706 obc
->obs
.oi
.version
= latest
->version
;
11708 ObjectStore::Transaction t
;
11710 obc
->obs
.oi
.encode(
11712 get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
11713 assert(!pool
.info
.require_rollback());
11714 t
.setattr(coll
, ghobject_t(soid
), OI_ATTR
, b2
);
11716 recover_got(soid
, latest
->version
);
11717 missing_loc
.add_location(soid
, pg_whoami
);
11721 osd
->store
->queue_transaction(osr
.get(), std::move(t
),
11722 new C_OSD_AppliedRecoveredObject(this, obc
),
11723 new C_OSD_CommittedPushedObject(
11725 get_osdmap()->get_epoch(),
11726 info
.last_complete
),
11727 new C_OSD_OndiskWriteUnlock(obc
));
11732 * Pull the old version of the object. Update missing_loc here to have the location
11733 * of the version we want.
11735 * This doesn't use the usual missing_loc paths, but that's okay:
11736 * - if we have it locally, we hit the case above, and go from there.
11737 * - if we don't, we always pass through this case during recovery and set up the location
11739 * - this way we don't need to mangle the missing code to be general about needing an old
11742 eversion_t alternate_need
= latest
->reverting_to
;
11743 dout(10) << " need to pull prior_version " << alternate_need
<< " for revert " << item
<< dendl
;
11745 for (map
<pg_shard_t
, pg_missing_t
>::iterator p
= peer_missing
.begin();
11746 p
!= peer_missing
.end();
11748 if (p
->second
.is_missing(soid
, need
) &&
11749 p
->second
.get_items().at(soid
).have
== alternate_need
) {
11750 missing_loc
.add_location(soid
, p
->first
);
11752 dout(10) << " will pull " << alternate_need
<< " or " << need
11753 << " from one of " << missing_loc
.get_locations(soid
)
11761 if (!recovering
.count(soid
)) {
11762 if (recovering
.count(head
)) {
11765 int r
= recover_missing(
11766 soid
, need
, get_recovery_op_priority(), h
);
11779 if (started
>= max
)
11784 // only advance last_requested if we haven't skipped anything
11786 pg_log
.set_last_requested(v
);
11789 pgbackend
->run_recovery_op(h
, get_recovery_op_priority());
11793 bool PrimaryLogPG::primary_error(
11794 const hobject_t
& soid
, eversion_t v
)
11796 pg_log
.missing_add(soid
, v
, eversion_t());
11797 pg_log
.set_last_requested(0);
11798 missing_loc
.remove_location(soid
, pg_whoami
);
11800 assert(!actingbackfill
.empty());
11801 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
11802 i
!= actingbackfill
.end();
11804 if (*i
== get_primary()) continue;
11805 pg_shard_t peer
= *i
;
11806 if (!peer_missing
[peer
].is_missing(soid
, v
)) {
11807 missing_loc
.add_location(soid
, peer
);
11808 dout(10) << info
.pgid
<< " unexpectedly missing " << soid
<< " v" << v
11809 << ", there should be a copy on shard " << peer
<< dendl
;
11814 osd
->clog
->error() << info
.pgid
<< " missing primary copy of " << soid
<< ", unfound";
11816 osd
->clog
->error() << info
.pgid
<< " missing primary copy of " << soid
11817 << ", will try copies on " << missing_loc
.get_locations(soid
);
11821 int PrimaryLogPG::prep_object_replica_deletes(
11822 const hobject_t
& soid
, eversion_t v
,
11823 PGBackend::RecoveryHandle
*h
)
11825 assert(is_primary());
11826 dout(10) << __func__
<< ": on " << soid
<< dendl
;
11828 start_recovery_op(soid
);
11829 assert(!recovering
.count(soid
));
11830 recovering
.insert(make_pair(soid
, ObjectContextRef()));
11832 pgbackend
->recover_delete_object(soid
, v
, h
);
11836 int PrimaryLogPG::prep_object_replica_pushes(
11837 const hobject_t
& soid
, eversion_t v
,
11838 PGBackend::RecoveryHandle
*h
)
11840 assert(is_primary());
11841 dout(10) << __func__
<< ": on " << soid
<< dendl
;
11843 // NOTE: we know we will get a valid oloc off of disk here.
11844 ObjectContextRef obc
= get_object_context(soid
, false);
11846 primary_error(soid
, v
);
11850 if (!obc
->get_recovery_read()) {
11851 dout(20) << "recovery delayed on " << soid
11852 << "; could not get rw_manager lock" << dendl
;
11855 dout(20) << "recovery got recovery read lock on " << soid
11859 start_recovery_op(soid
);
11860 assert(!recovering
.count(soid
));
11861 recovering
.insert(make_pair(soid
, obc
));
11863 /* We need this in case there is an in progress write on the object. In fact,
11864 * the only possible write is an update to the xattr due to a lost_revert --
11865 * a client write would be blocked since the object is degraded.
11866 * In almost all cases, therefore, this lock should be uncontended.
11868 obc
->ondisk_read_lock();
11869 int r
= pgbackend
->recover_object(
11872 ObjectContextRef(),
11873 obc
, // has snapset context
11875 obc
->ondisk_read_unlock();
11877 dout(0) << __func__
<< " Error " << r
<< " on oid " << soid
<< dendl
;
11878 primary_failed(soid
);
11879 primary_error(soid
, v
);
11885 uint64_t PrimaryLogPG::recover_replicas(uint64_t max
, ThreadPool::TPHandle
&handle
)
11887 dout(10) << __func__
<< "(" << max
<< ")" << dendl
;
11888 uint64_t started
= 0;
11890 PGBackend::RecoveryHandle
*h
= pgbackend
->open_recovery_op();
11892 // this is FAR from an optimal recovery order. pretty lame, really.
11893 assert(!actingbackfill
.empty());
11894 for (set
<pg_shard_t
>::iterator i
= actingbackfill
.begin();
11895 i
!= actingbackfill
.end();
11897 if (*i
== get_primary()) continue;
11898 pg_shard_t peer
= *i
;
11899 map
<pg_shard_t
, pg_missing_t
>::const_iterator pm
= peer_missing
.find(peer
);
11900 assert(pm
!= peer_missing
.end());
11901 map
<pg_shard_t
, pg_info_t
>::const_iterator pi
= peer_info
.find(peer
);
11902 assert(pi
!= peer_info
.end());
11903 size_t m_sz
= pm
->second
.num_missing();
11905 dout(10) << " peer osd." << peer
<< " missing " << m_sz
<< " objects." << dendl
;
11906 dout(20) << " peer osd." << peer
<< " missing " << pm
->second
.get_items() << dendl
;
11909 const pg_missing_t
&m(pm
->second
);
11910 for (map
<version_t
, hobject_t
>::const_iterator p
= m
.get_rmissing().begin();
11911 p
!= m
.get_rmissing().end() && started
< max
;
11913 handle
.reset_tp_timeout();
11914 const hobject_t
soid(p
->second
);
11916 if (missing_loc
.is_unfound(soid
)) {
11917 dout(10) << __func__
<< ": " << soid
<< " still unfound" << dendl
;
11921 if (soid
> pi
->second
.last_backfill
) {
11922 if (!recovering
.count(soid
)) {
11923 derr
<< __func__
<< ": object " << soid
<< " last_backfill " << pi
->second
.last_backfill
<< dendl
;
11924 derr
<< __func__
<< ": object added to missing set for backfill, but "
11925 << "is not in recovering, error!" << dendl
;
11931 if (recovering
.count(soid
)) {
11932 dout(10) << __func__
<< ": already recovering " << soid
<< dendl
;
11936 if (missing_loc
.is_deleted(soid
)) {
11937 dout(10) << __func__
<< ": " << soid
<< " is a delete, removing" << dendl
;
11938 map
<hobject_t
,pg_missing_item
>::const_iterator r
= m
.get_items().find(soid
);
11939 started
+= prep_object_replica_deletes(soid
, r
->second
.need
, h
);
11943 if (soid
.is_snap() && pg_log
.get_missing().is_missing(soid
.get_head())) {
11944 dout(10) << __func__
<< ": " << soid
.get_head()
11945 << " still missing on primary" << dendl
;
11949 if (soid
.is_snap() && pg_log
.get_missing().is_missing(soid
.get_snapdir())) {
11950 dout(10) << __func__
<< ": " << soid
.get_snapdir()
11951 << " still missing on primary" << dendl
;
11955 if (pg_log
.get_missing().is_missing(soid
)) {
11956 dout(10) << __func__
<< ": " << soid
<< " still missing on primary" << dendl
;
11960 dout(10) << __func__
<< ": recover_object_replicas(" << soid
<< ")" << dendl
;
11961 map
<hobject_t
,pg_missing_item
>::const_iterator r
= m
.get_items().find(soid
);
11962 started
+= prep_object_replica_pushes(soid
, r
->second
.need
,
11967 pgbackend
->run_recovery_op(h
, get_recovery_op_priority());
11971 hobject_t
PrimaryLogPG::earliest_peer_backfill() const
11973 hobject_t e
= hobject_t::get_max();
11974 for (set
<pg_shard_t
>::const_iterator i
= backfill_targets
.begin();
11975 i
!= backfill_targets
.end();
11977 pg_shard_t peer
= *i
;
11978 map
<pg_shard_t
, BackfillInterval
>::const_iterator iter
=
11979 peer_backfill_info
.find(peer
);
11980 assert(iter
!= peer_backfill_info
.end());
11981 if (iter
->second
.begin
< e
)
11982 e
= iter
->second
.begin
;
11987 bool PrimaryLogPG::all_peer_done() const
11989 // Primary hasn't got any more objects
11990 assert(backfill_info
.empty());
11992 for (set
<pg_shard_t
>::const_iterator i
= backfill_targets
.begin();
11993 i
!= backfill_targets
.end();
11995 pg_shard_t bt
= *i
;
11996 map
<pg_shard_t
, BackfillInterval
>::const_iterator piter
=
11997 peer_backfill_info
.find(bt
);
11998 assert(piter
!= peer_backfill_info
.end());
11999 const BackfillInterval
& pbi
= piter
->second
;
12000 // See if peer has more to process
12001 if (!pbi
.extends_to_end() || !pbi
.empty())
12012 * backfilled: fully pushed to replica or present in replica's missing set (both
12013 * our copy and theirs).
12015 * All objects on a backfill_target in
12016 * [MIN,peer_backfill_info[backfill_target].begin) are valid; logically-removed
12017 * objects have been actually deleted and all logically-valid objects are replicated.
12018 * There may be PG objects in this interval yet to be backfilled.
12020 * All objects in PG in [MIN,backfill_info.begin) have been backfilled to all
12021 * backfill_targets. There may be objects on backfill_target(s) yet to be deleted.
12023 * For a backfill target, all objects < MIN(peer_backfill_info[target].begin,
12024 * backfill_info.begin) in PG are backfilled. No deleted objects in this
12025 * interval remain on the backfill target.
12027 * For a backfill target, all objects <= peer_info[target].last_backfill
12028 * have been backfilled to target
12030 * There *MAY* be missing/outdated objects between last_backfill_started and
12031 * MIN(peer_backfill_info[*].begin, backfill_info.begin) in the event that client
12032 * io created objects since the last scan. For this reason, we call
12033 * update_range() again before continuing backfill.
12035 uint64_t PrimaryLogPG::recover_backfill(
12037 ThreadPool::TPHandle
&handle
, bool *work_started
)
12039 dout(10) << "recover_backfill (" << max
<< ")"
12040 << " bft=" << backfill_targets
12041 << " last_backfill_started " << last_backfill_started
12042 << (new_backfill
? " new_backfill":"")
12044 assert(!backfill_targets
.empty());
12046 // Initialize from prior backfill state
12047 if (new_backfill
) {
12048 // on_activate() was called prior to getting here
12049 assert(last_backfill_started
== earliest_backfill());
12050 new_backfill
= false;
12052 // initialize BackfillIntervals
12053 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
12054 i
!= backfill_targets
.end();
12056 peer_backfill_info
[*i
].reset(peer_info
[*i
].last_backfill
);
12058 backfill_info
.reset(last_backfill_started
);
12060 backfills_in_flight
.clear();
12061 pending_backfill_updates
.clear();
12064 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
12065 i
!= backfill_targets
.end();
12067 dout(10) << "peer osd." << *i
12068 << " info " << peer_info
[*i
]
12069 << " interval " << peer_backfill_info
[*i
].begin
12070 << "-" << peer_backfill_info
[*i
].end
12071 << " " << peer_backfill_info
[*i
].objects
.size() << " objects"
12075 // update our local interval to cope with recent changes
12076 backfill_info
.begin
= last_backfill_started
;
12077 update_range(&backfill_info
, handle
);
12080 vector
<boost::tuple
<hobject_t
, eversion_t
, pg_shard_t
> > to_remove
;
12081 set
<hobject_t
> add_to_stat
;
12083 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
12084 i
!= backfill_targets
.end();
12086 peer_backfill_info
[*i
].trim_to(
12087 std::max(peer_info
[*i
].last_backfill
, last_backfill_started
));
12089 backfill_info
.trim_to(last_backfill_started
);
12091 PGBackend::RecoveryHandle
*h
= pgbackend
->open_recovery_op();
12092 while (ops
< max
) {
12093 if (backfill_info
.begin
<= earliest_peer_backfill() &&
12094 !backfill_info
.extends_to_end() && backfill_info
.empty()) {
12095 hobject_t next
= backfill_info
.end
;
12096 backfill_info
.reset(next
);
12097 backfill_info
.end
= hobject_t::get_max();
12098 update_range(&backfill_info
, handle
);
12099 backfill_info
.trim();
12102 dout(20) << " my backfill interval " << backfill_info
<< dendl
;
12104 bool sent_scan
= false;
12105 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
12106 i
!= backfill_targets
.end();
12108 pg_shard_t bt
= *i
;
12109 BackfillInterval
& pbi
= peer_backfill_info
[bt
];
12111 dout(20) << " peer shard " << bt
<< " backfill " << pbi
<< dendl
;
12112 if (pbi
.begin
<= backfill_info
.begin
&&
12113 !pbi
.extends_to_end() && pbi
.empty()) {
12114 dout(10) << " scanning peer osd." << bt
<< " from " << pbi
.end
<< dendl
;
12115 epoch_t e
= get_osdmap()->get_epoch();
12116 MOSDPGScan
*m
= new MOSDPGScan(
12117 MOSDPGScan::OP_SCAN_GET_DIGEST
, pg_whoami
, e
, last_peering_reset
,
12118 spg_t(info
.pgid
.pgid
, bt
.shard
),
12119 pbi
.end
, hobject_t());
12120 osd
->send_message_osd_cluster(bt
.osd
, m
, get_osdmap()->get_epoch());
12121 assert(waiting_on_backfill
.find(bt
) == waiting_on_backfill
.end());
12122 waiting_on_backfill
.insert(bt
);
12127 // Count simultaneous scans as a single op and let those complete
12130 start_recovery_op(hobject_t::get_max()); // XXX: was pbi.end
12134 if (backfill_info
.empty() && all_peer_done()) {
12135 dout(10) << " reached end for both local and all peers" << dendl
;
12139 // Get object within set of peers to operate on and
12140 // the set of targets for which that object applies.
12141 hobject_t check
= earliest_peer_backfill();
12143 if (check
< backfill_info
.begin
) {
12145 set
<pg_shard_t
> check_targets
;
12146 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
12147 i
!= backfill_targets
.end();
12149 pg_shard_t bt
= *i
;
12150 BackfillInterval
& pbi
= peer_backfill_info
[bt
];
12151 if (pbi
.begin
== check
)
12152 check_targets
.insert(bt
);
12154 assert(!check_targets
.empty());
12156 dout(20) << " BACKFILL removing " << check
12157 << " from peers " << check_targets
<< dendl
;
12158 for (set
<pg_shard_t
>::iterator i
= check_targets
.begin();
12159 i
!= check_targets
.end();
12161 pg_shard_t bt
= *i
;
12162 BackfillInterval
& pbi
= peer_backfill_info
[bt
];
12163 assert(pbi
.begin
== check
);
12165 to_remove
.push_back(boost::make_tuple(check
, pbi
.objects
.begin()->second
, bt
));
12169 /* This requires a bit of explanation. We compare head against
12170 * last_backfill to determine whether to send an operation
12171 * to the replica. A single write operation can touch up to three
12172 * objects: head, the snapdir, and a new clone which sorts closer to
12173 * head than any existing clone. If last_backfill points at a clone,
12174 * the transaction won't be sent and all 3 must lie on the right side
12175 * of the line (i.e., we'll backfill them later). If last_backfill
12176 * points at snapdir, it sorts greater than head, so we send the
12177 * transaction which is correct because all three must lie to the left
12180 * If it points at head, we have a bit of an issue. If head actually
12181 * exists, no problem, because any transaction which touches snapdir
12182 * must end up creating it (and deleting head), so sending the
12183 * operation won't pose a problem -- we'll end up having to scan it,
12184 * but it'll end up being the right version so we won't bother to
12185 * rebackfill it. However, if head doesn't exist, any write on head
12186 * will remove snapdir. For a replicated pool, this isn't a problem,
12187 * ENOENT on remove isn't an issue and it's in backfill future anyway.
12188 * It only poses a problem for EC pools, because we never just delete
12189 * an object, we rename it into a rollback object. That operation
12190 * will end up crashing the osd with ENOENT. Tolerating the failure
12191 * wouldn't work either, even if snapdir exists, we'd be creating a
12192 * rollback object past the last_backfill line which wouldn't get
12193 * cleaned up (no rollback objects past the last_backfill line is an
12194 * existing important invariant). Thus, let's avoid the whole issue
12195 * by just not updating last_backfill_started here if head doesn't
12196 * exist and snapdir does. We aren't using up a recovery count here,
12197 * so we're going to recover snapdir immediately anyway. We'll only
12198 * fail "backward" if we fail to get the rw lock and that just means
12199 * we'll re-process this section of the hash space again.
12201 * I'm choosing this hack here because the really "correct" answer is
12202 * going to be to unify snapdir and head into a single object (a
12203 * snapdir is really just a confusing way to talk about head existing
12204 * as a whiteout), but doing that is going to be a somewhat larger
12207 * @see http://tracker.ceph.com/issues/17668
12209 if (!(check
.is_head() &&
12210 backfill_info
.begin
.is_snapdir() &&
12211 check
== backfill_info
.begin
.get_head()))
12212 last_backfill_started
= check
;
12214 // Don't increment ops here because deletions
12215 // are cheap and not replied to unlike real recovery_ops,
12216 // and we can't increment ops without requeueing ourself
12219 eversion_t
& obj_v
= backfill_info
.objects
.begin()->second
;
12221 vector
<pg_shard_t
> need_ver_targs
, missing_targs
, keep_ver_targs
, skip_targs
;
12222 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
12223 i
!= backfill_targets
.end();
12225 pg_shard_t bt
= *i
;
12226 BackfillInterval
& pbi
= peer_backfill_info
[bt
];
12227 // Find all check peers that have the wrong version
12228 if (check
== backfill_info
.begin
&& check
== pbi
.begin
) {
12229 if (pbi
.objects
.begin()->second
!= obj_v
) {
12230 need_ver_targs
.push_back(bt
);
12232 keep_ver_targs
.push_back(bt
);
12235 pg_info_t
& pinfo
= peer_info
[bt
];
12237 // Only include peers that we've caught up to their backfill line
12238 // otherwise, they only appear to be missing this object
12239 // because their pbi.begin > backfill_info.begin.
12240 if (backfill_info
.begin
> pinfo
.last_backfill
)
12241 missing_targs
.push_back(bt
);
12243 skip_targs
.push_back(bt
);
12247 if (!keep_ver_targs
.empty()) {
12248 // These peers have version obj_v
12249 dout(20) << " BACKFILL keeping " << check
12250 << " with ver " << obj_v
12251 << " on peers " << keep_ver_targs
<< dendl
;
12252 //assert(!waiting_for_degraded_object.count(check));
12254 if (!need_ver_targs
.empty() || !missing_targs
.empty()) {
12255 ObjectContextRef obc
= get_object_context(backfill_info
.begin
, false);
12257 if (obc
->get_recovery_read()) {
12258 if (!need_ver_targs
.empty()) {
12259 dout(20) << " BACKFILL replacing " << check
12260 << " with ver " << obj_v
12261 << " to peers " << need_ver_targs
<< dendl
;
12263 if (!missing_targs
.empty()) {
12264 dout(20) << " BACKFILL pushing " << backfill_info
.begin
12265 << " with ver " << obj_v
12266 << " to peers " << missing_targs
<< dendl
;
12268 vector
<pg_shard_t
> all_push
= need_ver_targs
;
12269 all_push
.insert(all_push
.end(), missing_targs
.begin(), missing_targs
.end());
12271 handle
.reset_tp_timeout();
12272 int r
= prep_backfill_object_push(backfill_info
.begin
, obj_v
, obc
, all_push
, h
);
12274 *work_started
= true;
12275 dout(0) << __func__
<< " Error " << r
<< " trying to backfill " << backfill_info
.begin
<< dendl
;
12280 *work_started
= true;
12281 dout(20) << "backfill blocking on " << backfill_info
.begin
12282 << "; could not get rw_manager lock" << dendl
;
12286 dout(20) << "need_ver_targs=" << need_ver_targs
12287 << " keep_ver_targs=" << keep_ver_targs
<< dendl
;
12288 dout(20) << "backfill_targets=" << backfill_targets
12289 << " missing_targs=" << missing_targs
12290 << " skip_targs=" << skip_targs
<< dendl
;
12292 last_backfill_started
= backfill_info
.begin
;
12293 add_to_stat
.insert(backfill_info
.begin
); // XXX: Only one for all pushes?
12294 backfill_info
.pop_front();
12295 vector
<pg_shard_t
> check_targets
= need_ver_targs
;
12296 check_targets
.insert(check_targets
.end(), keep_ver_targs
.begin(), keep_ver_targs
.end());
12297 for (vector
<pg_shard_t
>::iterator i
= check_targets
.begin();
12298 i
!= check_targets
.end();
12300 pg_shard_t bt
= *i
;
12301 BackfillInterval
& pbi
= peer_backfill_info
[bt
];
12307 hobject_t backfill_pos
=
12308 std::min(backfill_info
.begin
, earliest_peer_backfill());
12310 for (set
<hobject_t
>::iterator i
= add_to_stat
.begin();
12311 i
!= add_to_stat
.end();
12313 ObjectContextRef obc
= get_object_context(*i
, false);
12316 add_object_context_to_pg_stat(obc
, &stat
);
12317 pending_backfill_updates
[*i
] = stat
;
12319 if (HAVE_FEATURE(get_min_upacting_features(), SERVER_LUMINOUS
)) {
12320 map
<pg_shard_t
,MOSDPGBackfillRemove
*> reqs
;
12321 for (unsigned i
= 0; i
< to_remove
.size(); ++i
) {
12322 handle
.reset_tp_timeout();
12323 const hobject_t
& oid
= to_remove
[i
].get
<0>();
12324 eversion_t v
= to_remove
[i
].get
<1>();
12325 pg_shard_t peer
= to_remove
[i
].get
<2>();
12326 MOSDPGBackfillRemove
*m
;
12327 auto it
= reqs
.find(peer
);
12328 if (it
!= reqs
.end()) {
12331 m
= reqs
[peer
] = new MOSDPGBackfillRemove(
12332 spg_t(info
.pgid
.pgid
, peer
.shard
),
12333 get_osdmap()->get_epoch());
12335 m
->ls
.push_back(make_pair(oid
, v
));
12337 if (oid
<= last_backfill_started
)
12338 pending_backfill_updates
[oid
]; // add empty stat!
12340 for (auto p
: reqs
) {
12341 osd
->send_message_osd_cluster(p
.first
.osd
, p
.second
,
12342 get_osdmap()->get_epoch());
12345 // for jewel targets
12346 for (unsigned i
= 0; i
< to_remove
.size(); ++i
) {
12347 handle
.reset_tp_timeout();
12349 // ordered before any subsequent updates
12350 send_remove_op(to_remove
[i
].get
<0>(), to_remove
[i
].get
<1>(),
12351 to_remove
[i
].get
<2>());
12353 if (to_remove
[i
].get
<0>() <= last_backfill_started
)
12354 pending_backfill_updates
[to_remove
[i
].get
<0>()]; // add empty stat!
12358 pgbackend
->run_recovery_op(h
, get_recovery_op_priority());
12360 dout(5) << "backfill_pos is " << backfill_pos
<< dendl
;
12361 for (set
<hobject_t
>::iterator i
= backfills_in_flight
.begin();
12362 i
!= backfills_in_flight
.end();
12364 dout(20) << *i
<< " is still in flight" << dendl
;
12367 hobject_t next_backfill_to_complete
= backfills_in_flight
.empty() ?
12368 backfill_pos
: *(backfills_in_flight
.begin());
12369 hobject_t new_last_backfill
= earliest_backfill();
12370 dout(10) << "starting new_last_backfill at " << new_last_backfill
<< dendl
;
12371 for (map
<hobject_t
, pg_stat_t
>::iterator i
=
12372 pending_backfill_updates
.begin();
12373 i
!= pending_backfill_updates
.end() &&
12374 i
->first
< next_backfill_to_complete
;
12375 pending_backfill_updates
.erase(i
++)) {
12376 dout(20) << " pending_backfill_update " << i
->first
<< dendl
;
12377 assert(i
->first
> new_last_backfill
);
12378 for (set
<pg_shard_t
>::iterator j
= backfill_targets
.begin();
12379 j
!= backfill_targets
.end();
12381 pg_shard_t bt
= *j
;
12382 pg_info_t
& pinfo
= peer_info
[bt
];
12383 //Add stats to all peers that were missing object
12384 if (i
->first
> pinfo
.last_backfill
)
12385 pinfo
.stats
.add(i
->second
);
12387 new_last_backfill
= i
->first
;
12389 dout(10) << "possible new_last_backfill at " << new_last_backfill
<< dendl
;
12391 assert(!pending_backfill_updates
.empty() ||
12392 new_last_backfill
== last_backfill_started
);
12393 if (pending_backfill_updates
.empty() &&
12394 backfill_pos
.is_max()) {
12395 assert(backfills_in_flight
.empty());
12396 new_last_backfill
= backfill_pos
;
12397 last_backfill_started
= backfill_pos
;
12399 dout(10) << "final new_last_backfill at " << new_last_backfill
<< dendl
;
12401 // If new_last_backfill == MAX, then we will send OP_BACKFILL_FINISH to
12402 // all the backfill targets. Otherwise, we will move last_backfill up on
12403 // those targets need it and send OP_BACKFILL_PROGRESS to them.
12404 for (set
<pg_shard_t
>::iterator i
= backfill_targets
.begin();
12405 i
!= backfill_targets
.end();
12407 pg_shard_t bt
= *i
;
12408 pg_info_t
& pinfo
= peer_info
[bt
];
12410 if (new_last_backfill
> pinfo
.last_backfill
) {
12411 pinfo
.set_last_backfill(new_last_backfill
);
12412 epoch_t e
= get_osdmap()->get_epoch();
12413 MOSDPGBackfill
*m
= NULL
;
12414 if (pinfo
.last_backfill
.is_max()) {
12415 m
= new MOSDPGBackfill(
12416 MOSDPGBackfill::OP_BACKFILL_FINISH
,
12418 last_peering_reset
,
12419 spg_t(info
.pgid
.pgid
, bt
.shard
));
12420 // Use default priority here, must match sub_op priority
12421 /* pinfo.stats might be wrong if we did log-based recovery on the
12422 * backfilled portion in addition to continuing backfill.
12424 pinfo
.stats
= info
.stats
;
12425 start_recovery_op(hobject_t::get_max());
12427 m
= new MOSDPGBackfill(
12428 MOSDPGBackfill::OP_BACKFILL_PROGRESS
,
12430 last_peering_reset
,
12431 spg_t(info
.pgid
.pgid
, bt
.shard
));
12432 // Use default priority here, must match sub_op priority
12434 m
->last_backfill
= pinfo
.last_backfill
;
12435 m
->stats
= pinfo
.stats
;
12436 osd
->send_message_osd_cluster(bt
.osd
, m
, get_osdmap()->get_epoch());
12437 dout(10) << " peer " << bt
12438 << " num_objects now " << pinfo
.stats
.stats
.sum
.num_objects
12439 << " / " << info
.stats
.stats
.sum
.num_objects
<< dendl
;
12444 *work_started
= true;
12448 int PrimaryLogPG::prep_backfill_object_push(
12449 hobject_t oid
, eversion_t v
,
12450 ObjectContextRef obc
,
12451 vector
<pg_shard_t
> peers
,
12452 PGBackend::RecoveryHandle
*h
)
12454 dout(10) << __func__
<< " " << oid
<< " v " << v
<< " to peers " << peers
<< dendl
;
12455 assert(!peers
.empty());
12457 backfills_in_flight
.insert(oid
);
12458 for (unsigned int i
= 0 ; i
< peers
.size(); ++i
) {
12459 map
<pg_shard_t
, pg_missing_t
>::iterator bpm
= peer_missing
.find(peers
[i
]);
12460 assert(bpm
!= peer_missing
.end());
12461 bpm
->second
.add(oid
, eversion_t(), eversion_t(), false);
12464 assert(!recovering
.count(oid
));
12466 start_recovery_op(oid
);
12467 recovering
.insert(make_pair(oid
, obc
));
12469 // We need to take the read_lock here in order to flush in-progress writes
12470 obc
->ondisk_read_lock();
12471 int r
= pgbackend
->recover_object(
12474 ObjectContextRef(),
12477 obc
->ondisk_read_unlock();
12479 dout(0) << __func__
<< " Error " << r
<< " on oid " << oid
<< dendl
;
12480 primary_failed(oid
);
12481 primary_error(oid
, v
);
12482 backfills_in_flight
.erase(oid
);
12483 missing_loc
.add_missing(oid
, v
, eversion_t());
12488 void PrimaryLogPG::update_range(
12489 BackfillInterval
*bi
,
12490 ThreadPool::TPHandle
&handle
)
12492 int local_min
= cct
->_conf
->osd_backfill_scan_min
;
12493 int local_max
= cct
->_conf
->osd_backfill_scan_max
;
12495 if (bi
->version
< info
.log_tail
) {
12496 dout(10) << __func__
<< ": bi is old, rescanning local backfill_info"
12499 if (last_update_applied
>= info
.log_tail
) {
12500 bi
->version
= last_update_applied
;
12502 bi
->version
= info
.last_update
;
12504 scan_range(local_min
, local_max
, bi
, handle
);
12507 if (bi
->version
>= projected_last_update
) {
12508 dout(10) << __func__
<< ": bi is current " << dendl
;
12509 assert(bi
->version
== projected_last_update
);
12510 } else if (bi
->version
>= info
.log_tail
) {
12511 if (pg_log
.get_log().empty() && projected_log
.empty()) {
12512 /* Because we don't move log_tail on split, the log might be
12513 * empty even if log_tail != last_update. However, the only
12514 * way to get here with an empty log is if log_tail is actually
12515 * eversion_t(), because otherwise the entry which changed
12516 * last_update since the last scan would have to be present.
12518 assert(bi
->version
== eversion_t());
12522 dout(10) << __func__
<< ": bi is old, (" << bi
->version
12523 << ") can be updated with log to projected_last_update "
12524 << projected_last_update
<< dendl
;
12526 auto func
= [&](const pg_log_entry_t
&e
) {
12527 dout(10) << __func__
<< ": updating from version " << e
.version
12529 const hobject_t
&soid
= e
.soid
;
12530 if (soid
>= bi
->begin
&&
12532 if (e
.is_update()) {
12533 dout(10) << __func__
<< ": " << e
.soid
<< " updated to version "
12534 << e
.version
<< dendl
;
12535 bi
->objects
.erase(e
.soid
);
12536 bi
->objects
.insert(
12540 } else if (e
.is_delete()) {
12541 dout(10) << __func__
<< ": " << e
.soid
<< " removed" << dendl
;
12542 bi
->objects
.erase(e
.soid
);
12546 dout(10) << "scanning pg log first" << dendl
;
12547 pg_log
.get_log().scan_log_after(bi
->version
, func
);
12548 dout(10) << "scanning projected log" << dendl
;
12549 projected_log
.scan_log_after(bi
->version
, func
);
12550 bi
->version
= projected_last_update
;
12552 assert(0 == "scan_range should have raised bi->version past log_tail");
12556 void PrimaryLogPG::scan_range(
12557 int min
, int max
, BackfillInterval
*bi
,
12558 ThreadPool::TPHandle
&handle
)
12560 assert(is_locked());
12561 dout(10) << "scan_range from " << bi
->begin
<< dendl
;
12562 bi
->clear_objects();
12564 vector
<hobject_t
> ls
;
12566 int r
= pgbackend
->objects_list_partial(bi
->begin
, min
, max
, &ls
, &bi
->end
);
12568 dout(10) << " got " << ls
.size() << " items, next " << bi
->end
<< dendl
;
12569 dout(20) << ls
<< dendl
;
12571 for (vector
<hobject_t
>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
12572 handle
.reset_tp_timeout();
12573 ObjectContextRef obc
;
12575 obc
= object_contexts
.lookup(*p
);
12577 bi
->objects
[*p
] = obc
->obs
.oi
.version
;
12578 dout(20) << " " << *p
<< " " << obc
->obs
.oi
.version
<< dendl
;
12581 int r
= pgbackend
->objects_get_attr(*p
, OI_ATTR
, &bl
);
12583 /* If the object does not exist here, it must have been removed
12584 * between the collection_list_partial and here. This can happen
12585 * for the first item in the range, which is usually last_backfill.
12591 object_info_t
oi(bl
);
12592 bi
->objects
[*p
] = oi
.version
;
12593 dout(20) << " " << *p
<< " " << oi
.version
<< dendl
;
12601 * verifies that stray objects have been deleted
12603 void PrimaryLogPG::check_local()
12605 dout(10) << __func__
<< dendl
;
12607 assert(info
.last_update
>= pg_log
.get_tail()); // otherwise we need some help!
12609 if (!cct
->_conf
->osd_debug_verify_stray_on_activate
)
12612 // just scan the log.
12613 set
<hobject_t
> did
;
12614 for (list
<pg_log_entry_t
>::const_reverse_iterator p
= pg_log
.get_log().log
.rbegin();
12615 p
!= pg_log
.get_log().log
.rend();
12617 if (did
.count(p
->soid
))
12619 did
.insert(p
->soid
);
12621 if (p
->is_delete() && !is_missing_object(p
->soid
)) {
12622 dout(10) << " checking " << p
->soid
12623 << " at " << p
->version
<< dendl
;
12625 int r
= osd
->store
->stat(
12627 ghobject_t(p
->soid
, ghobject_t::NO_GEN
, pg_whoami
.shard
),
12629 if (r
!= -ENOENT
) {
12630 derr
<< __func__
<< " " << p
->soid
<< " exists, but should have been "
12631 << "deleted" << dendl
;
12632 assert(0 == "erroneously present object");
12635 // ignore old(+missing) objects
12642 // ===========================
12645 hobject_t
PrimaryLogPG::get_hit_set_current_object(utime_t stamp
)
12648 ss
<< "hit_set_" << info
.pgid
.pgid
<< "_current_" << stamp
;
12649 hobject_t
hoid(sobject_t(ss
.str(), CEPH_NOSNAP
), "",
12650 info
.pgid
.ps(), info
.pgid
.pool(),
12651 cct
->_conf
->osd_hit_set_namespace
);
12652 dout(20) << __func__
<< " " << hoid
<< dendl
;
12656 hobject_t
PrimaryLogPG::get_hit_set_archive_object(utime_t start
,
12661 ss
<< "hit_set_" << info
.pgid
.pgid
<< "_archive_";
12663 start
.gmtime(ss
) << "_";
12666 start
.localtime(ss
) << "_";
12669 hobject_t
hoid(sobject_t(ss
.str(), CEPH_NOSNAP
), "",
12670 info
.pgid
.ps(), info
.pgid
.pool(),
12671 cct
->_conf
->osd_hit_set_namespace
);
12672 dout(20) << __func__
<< " " << hoid
<< dendl
;
12676 void PrimaryLogPG::hit_set_clear()
12678 dout(20) << __func__
<< dendl
;
12680 hit_set_start_stamp
= utime_t();
12683 void PrimaryLogPG::hit_set_setup()
12685 if (!is_active() ||
12691 if (is_active() && is_primary() &&
12692 (!pool
.info
.hit_set_count
||
12693 !pool
.info
.hit_set_period
||
12694 pool
.info
.hit_set_params
.get_type() == HitSet::TYPE_NONE
)) {
12697 // only primary is allowed to remove all the hit set objects
12698 hit_set_remove_all();
12702 // FIXME: discard any previous data for now
12705 // include any writes we know about from the pg log. this doesn't
12706 // capture reads, but it is better than nothing!
12707 hit_set_apply_log();
12710 void PrimaryLogPG::hit_set_remove_all()
12712 // If any archives are degraded we skip this
12713 for (list
<pg_hit_set_info_t
>::iterator p
= info
.hit_set
.history
.begin();
12714 p
!= info
.hit_set
.history
.end();
12716 hobject_t aoid
= get_hit_set_archive_object(p
->begin
, p
->end
, p
->using_gmt
);
12718 // Once we hit a degraded object just skip
12719 if (is_degraded_or_backfilling_object(aoid
))
12721 if (write_blocked_by_scrub(aoid
))
12725 if (!info
.hit_set
.history
.empty()) {
12726 list
<pg_hit_set_info_t
>::reverse_iterator p
= info
.hit_set
.history
.rbegin();
12727 assert(p
!= info
.hit_set
.history
.rend());
12728 hobject_t oid
= get_hit_set_archive_object(p
->begin
, p
->end
, p
->using_gmt
);
12729 assert(!is_degraded_or_backfilling_object(oid
));
12730 ObjectContextRef obc
= get_object_context(oid
, false);
12733 OpContextUPtr ctx
= simple_opc_create(obc
);
12734 ctx
->at_version
= get_next_version();
12735 ctx
->updated_hset_history
= info
.hit_set
;
12736 utime_t now
= ceph_clock_now();
12738 hit_set_trim(ctx
, 0);
12739 simple_opc_submit(std::move(ctx
));
12742 info
.hit_set
= pg_hit_set_history_t();
12744 agent_state
->discard_hit_sets();
12748 void PrimaryLogPG::hit_set_create()
12750 utime_t now
= ceph_clock_now();
12751 // make a copy of the params to modify
12752 HitSet::Params
params(pool
.info
.hit_set_params
);
12754 dout(20) << __func__
<< " " << params
<< dendl
;
12755 if (pool
.info
.hit_set_params
.get_type() == HitSet::TYPE_BLOOM
) {
12756 BloomHitSet::Params
*p
=
12757 static_cast<BloomHitSet::Params
*>(params
.impl
.get());
12759 // convert false positive rate so it holds up across the full period
12760 p
->set_fpp(p
->get_fpp() / pool
.info
.hit_set_count
);
12761 if (p
->get_fpp() <= 0.0)
12762 p
->set_fpp(.01); // fpp cannot be zero!
12764 // if we don't have specified size, estimate target size based on the
12766 if (p
->target_size
== 0 && hit_set
) {
12767 utime_t dur
= now
- hit_set_start_stamp
;
12768 unsigned unique
= hit_set
->approx_unique_insert_count();
12769 dout(20) << __func__
<< " previous set had approx " << unique
12770 << " unique items over " << dur
<< " seconds" << dendl
;
12771 p
->target_size
= (double)unique
* (double)pool
.info
.hit_set_period
12774 if (p
->target_size
<
12775 static_cast<uint64_t>(cct
->_conf
->osd_hit_set_min_size
))
12776 p
->target_size
= cct
->_conf
->osd_hit_set_min_size
;
12779 > static_cast<uint64_t>(cct
->_conf
->osd_hit_set_max_size
))
12780 p
->target_size
= cct
->_conf
->osd_hit_set_max_size
;
12782 p
->seed
= now
.sec();
12784 dout(10) << __func__
<< " target_size " << p
->target_size
12785 << " fpp " << p
->get_fpp() << dendl
;
12787 hit_set
.reset(new HitSet(params
));
12788 hit_set_start_stamp
= now
;
12792 * apply log entries to set
12794 * this would only happen after peering, to at least capture writes
12795 * during an interval that was potentially lost.
12797 bool PrimaryLogPG::hit_set_apply_log()
12802 eversion_t to
= info
.last_update
;
12803 eversion_t from
= info
.hit_set
.current_last_update
;
12805 dout(20) << __func__
<< " no update" << dendl
;
12809 dout(20) << __func__
<< " " << to
<< " .. " << info
.last_update
<< dendl
;
12810 list
<pg_log_entry_t
>::const_reverse_iterator p
= pg_log
.get_log().log
.rbegin();
12811 while (p
!= pg_log
.get_log().log
.rend() && p
->version
> to
)
12813 while (p
!= pg_log
.get_log().log
.rend() && p
->version
> from
) {
12814 hit_set
->insert(p
->soid
);
12821 void PrimaryLogPG::hit_set_persist()
12823 dout(10) << __func__
<< dendl
;
12825 unsigned max
= pool
.info
.hit_set_count
;
12827 utime_t now
= ceph_clock_now();
12830 // If any archives are degraded we skip this persist request
12831 // account for the additional entry being added below
12832 for (list
<pg_hit_set_info_t
>::iterator p
= info
.hit_set
.history
.begin();
12833 p
!= info
.hit_set
.history
.end();
12835 hobject_t aoid
= get_hit_set_archive_object(p
->begin
, p
->end
, p
->using_gmt
);
12837 // Once we hit a degraded object just skip further trim
12838 if (is_degraded_or_backfilling_object(aoid
))
12840 if (write_blocked_by_scrub(aoid
))
12844 // If backfill is in progress and we could possibly overlap with the
12845 // hit_set_* objects, back off. Since these all have
12846 // hobject_t::hash set to pgid.ps(), and those sort first, we can
12847 // look just at that. This is necessary because our transactions
12848 // may include a modify of the new hit_set *and* a delete of the
12849 // old one, and this may span the backfill boundary.
12850 for (set
<pg_shard_t
>::iterator p
= backfill_targets
.begin();
12851 p
!= backfill_targets
.end();
12853 assert(peer_info
.count(*p
));
12854 const pg_info_t
& pi
= peer_info
[*p
];
12855 if (pi
.last_backfill
== hobject_t() ||
12856 pi
.last_backfill
.get_hash() == info
.pgid
.ps()) {
12857 dout(10) << __func__
<< " backfill target osd." << *p
12858 << " last_backfill has not progressed past pgid ps"
12865 pg_hit_set_info_t new_hset
= pg_hit_set_info_t(pool
.info
.use_gmt_hitset
);
12866 new_hset
.begin
= hit_set_start_stamp
;
12867 new_hset
.end
= now
;
12868 oid
= get_hit_set_archive_object(
12871 new_hset
.using_gmt
);
12873 // If the current object is degraded we skip this persist request
12874 if (write_blocked_by_scrub(oid
))
12878 ::encode(*hit_set
, bl
);
12879 dout(20) << __func__
<< " archive " << oid
<< dendl
;
12882 agent_state
->add_hit_set(new_hset
.begin
, hit_set
);
12883 uint32_t size
= agent_state
->hit_set_map
.size();
12884 if (size
>= pool
.info
.hit_set_count
) {
12885 size
= pool
.info
.hit_set_count
> 0 ? pool
.info
.hit_set_count
- 1: 0;
12887 hit_set_in_memory_trim(size
);
12890 ObjectContextRef obc
= get_object_context(oid
, true);
12891 OpContextUPtr ctx
= simple_opc_create(obc
);
12893 ctx
->at_version
= get_next_version();
12894 ctx
->updated_hset_history
= info
.hit_set
;
12895 pg_hit_set_history_t
&updated_hit_set_hist
= *(ctx
->updated_hset_history
);
12897 updated_hit_set_hist
.current_last_update
= info
.last_update
;
12898 new_hset
.version
= ctx
->at_version
;
12900 updated_hit_set_hist
.history
.push_back(new_hset
);
12903 // fabricate an object_info_t and SnapSet
12904 obc
->obs
.oi
.version
= ctx
->at_version
;
12905 obc
->obs
.oi
.mtime
= now
;
12906 obc
->obs
.oi
.size
= bl
.length();
12907 obc
->obs
.exists
= true;
12908 obc
->obs
.oi
.set_data_digest(bl
.crc32c(-1));
12910 ctx
->new_obs
= obc
->obs
;
12912 obc
->ssc
->snapset
.head_exists
= true;
12913 ctx
->new_snapset
= obc
->ssc
->snapset
;
12915 ctx
->delta_stats
.num_objects
++;
12916 ctx
->delta_stats
.num_objects_hit_set_archive
++;
12917 ctx
->delta_stats
.num_bytes
+= bl
.length();
12918 ctx
->delta_stats
.num_bytes_hit_set_archive
+= bl
.length();
12921 ::encode(ctx
->new_snapset
, bss
);
12922 bufferlist
boi(sizeof(ctx
->new_obs
.oi
));
12923 ::encode(ctx
->new_obs
.oi
, boi
,
12924 get_osdmap()->get_features(CEPH_ENTITY_TYPE_OSD
, nullptr));
12926 ctx
->op_t
->create(oid
);
12928 ctx
->op_t
->write(oid
, 0, bl
.length(), bl
, 0);
12930 map
<string
, bufferlist
> attrs
;
12931 attrs
[OI_ATTR
].claim(boi
);
12932 attrs
[SS_ATTR
].claim(bss
);
12933 setattrs_maybe_cache(ctx
->obc
, ctx
.get(), ctx
->op_t
.get(), attrs
);
12934 ctx
->log
.push_back(
12936 pg_log_entry_t::MODIFY
,
12946 hit_set_trim(ctx
, max
);
12948 simple_opc_submit(std::move(ctx
));
12951 void PrimaryLogPG::hit_set_trim(OpContextUPtr
&ctx
, unsigned max
)
12953 assert(ctx
->updated_hset_history
);
12954 pg_hit_set_history_t
&updated_hit_set_hist
=
12955 *(ctx
->updated_hset_history
);
12956 for (unsigned num
= updated_hit_set_hist
.history
.size(); num
> max
; --num
) {
12957 list
<pg_hit_set_info_t
>::iterator p
= updated_hit_set_hist
.history
.begin();
12958 assert(p
!= updated_hit_set_hist
.history
.end());
12959 hobject_t oid
= get_hit_set_archive_object(p
->begin
, p
->end
, p
->using_gmt
);
12961 assert(!is_degraded_or_backfilling_object(oid
));
12963 dout(20) << __func__
<< " removing " << oid
<< dendl
;
12964 ++ctx
->at_version
.version
;
12965 ctx
->log
.push_back(
12966 pg_log_entry_t(pg_log_entry_t::DELETE
,
12975 ctx
->op_t
->remove(oid
);
12976 updated_hit_set_hist
.history
.pop_front();
12978 ObjectContextRef obc
= get_object_context(oid
, false);
12980 --ctx
->delta_stats
.num_objects
;
12981 --ctx
->delta_stats
.num_objects_hit_set_archive
;
12982 ctx
->delta_stats
.num_bytes
-= obc
->obs
.oi
.size
;
12983 ctx
->delta_stats
.num_bytes_hit_set_archive
-= obc
->obs
.oi
.size
;
12987 void PrimaryLogPG::hit_set_in_memory_trim(uint32_t max_in_memory
)
12989 while (agent_state
->hit_set_map
.size() > max_in_memory
) {
12990 agent_state
->remove_oldest_hit_set();
12995 // =======================================
12998 void PrimaryLogPG::agent_setup()
13000 assert(is_locked());
13001 if (!is_active() ||
13003 pool
.info
.cache_mode
== pg_pool_t::CACHEMODE_NONE
||
13004 pool
.info
.tier_of
< 0 ||
13005 !get_osdmap()->have_pg_pool(pool
.info
.tier_of
)) {
13009 if (!agent_state
) {
13010 agent_state
.reset(new TierAgentState
);
13012 // choose random starting position
13013 agent_state
->position
= hobject_t();
13014 agent_state
->position
.pool
= info
.pgid
.pool();
13015 agent_state
->position
.set_hash(pool
.info
.get_random_pg_position(
13018 agent_state
->start
= agent_state
->position
;
13020 dout(10) << __func__
<< " allocated new state, position "
13021 << agent_state
->position
<< dendl
;
13023 dout(10) << __func__
<< " keeping existing state" << dendl
;
13026 if (info
.stats
.stats_invalid
) {
13027 osd
->clog
->warn() << "pg " << info
.pgid
<< " has invalid (post-split) stats; must scrub before tier agent can activate";
13030 agent_choose_mode();
13033 void PrimaryLogPG::agent_clear()
13036 agent_state
.reset(NULL
);
13039 // Return false if no objects operated on since start of object hash space
13040 bool PrimaryLogPG::agent_work(int start_max
, int agent_flush_quota
)
13043 if (!agent_state
) {
13044 dout(10) << __func__
<< " no agent state, stopping" << dendl
;
13051 if (agent_state
->is_idle()) {
13052 dout(10) << __func__
<< " idle, stopping" << dendl
;
13057 osd
->logger
->inc(l_osd_agent_wake
);
13059 dout(10) << __func__
13060 << " max " << start_max
13061 << ", flush " << agent_state
->get_flush_mode_name()
13062 << ", evict " << agent_state
->get_evict_mode_name()
13063 << ", pos " << agent_state
->position
13065 assert(is_primary());
13066 assert(is_active());
13068 agent_load_hit_sets();
13070 const pg_pool_t
*base_pool
= get_osdmap()->get_pg_pool(pool
.info
.tier_of
);
13074 int ls_max
= cct
->_conf
->osd_pool_default_cache_max_evict_check_size
;
13076 // list some objects. this conveniently lists clones (oldest to
13077 // newest) before heads... the same order we want to flush in.
13079 // NOTE: do not flush the Sequencer. we will assume that the
13080 // listing we get back is imprecise.
13081 vector
<hobject_t
> ls
;
13083 int r
= pgbackend
->objects_list_partial(agent_state
->position
, ls_min
, ls_max
,
13086 dout(20) << __func__
<< " got " << ls
.size() << " objects" << dendl
;
13088 for (vector
<hobject_t
>::iterator p
= ls
.begin();
13091 if (p
->nspace
== cct
->_conf
->osd_hit_set_namespace
) {
13092 dout(20) << __func__
<< " skip (hit set) " << *p
<< dendl
;
13093 osd
->logger
->inc(l_osd_agent_skip
);
13096 if (is_degraded_or_backfilling_object(*p
)) {
13097 dout(20) << __func__
<< " skip (degraded) " << *p
<< dendl
;
13098 osd
->logger
->inc(l_osd_agent_skip
);
13101 if (is_missing_object(p
->get_head())) {
13102 dout(20) << __func__
<< " skip (missing head) " << *p
<< dendl
;
13103 osd
->logger
->inc(l_osd_agent_skip
);
13106 ObjectContextRef obc
= get_object_context(*p
, false, NULL
);
13108 // we didn't flush; we may miss something here.
13109 dout(20) << __func__
<< " skip (no obc) " << *p
<< dendl
;
13110 osd
->logger
->inc(l_osd_agent_skip
);
13113 if (!obc
->obs
.exists
) {
13114 dout(20) << __func__
<< " skip (dne) " << obc
->obs
.oi
.soid
<< dendl
;
13115 osd
->logger
->inc(l_osd_agent_skip
);
13118 if (range_intersects_scrub(obc
->obs
.oi
.soid
,
13119 obc
->obs
.oi
.soid
.get_head())) {
13120 dout(20) << __func__
<< " skip (scrubbing) " << obc
->obs
.oi
<< dendl
;
13121 osd
->logger
->inc(l_osd_agent_skip
);
13124 if (obc
->is_blocked()) {
13125 dout(20) << __func__
<< " skip (blocked) " << obc
->obs
.oi
<< dendl
;
13126 osd
->logger
->inc(l_osd_agent_skip
);
13129 if (obc
->is_request_pending()) {
13130 dout(20) << __func__
<< " skip (request pending) " << obc
->obs
.oi
<< dendl
;
13131 osd
->logger
->inc(l_osd_agent_skip
);
13135 // be careful flushing omap to an EC pool.
13136 if (!base_pool
->supports_omap() &&
13137 obc
->obs
.oi
.is_omap()) {
13138 dout(20) << __func__
<< " skip (omap to EC) " << obc
->obs
.oi
<< dendl
;
13139 osd
->logger
->inc(l_osd_agent_skip
);
13143 if (agent_state
->evict_mode
!= TierAgentState::EVICT_MODE_IDLE
&&
13144 agent_maybe_evict(obc
, false))
13146 else if (agent_state
->flush_mode
!= TierAgentState::FLUSH_MODE_IDLE
&&
13147 agent_flush_quota
> 0 && agent_maybe_flush(obc
)) {
13149 --agent_flush_quota
;
13151 if (started
>= start_max
) {
13152 // If finishing early, set "next" to the next object
13153 if (++p
!= ls
.end())
13159 if (++agent_state
->hist_age
> cct
->_conf
->osd_agent_hist_halflife
) {
13160 dout(20) << __func__
<< " resetting atime and temp histograms" << dendl
;
13161 agent_state
->hist_age
= 0;
13162 agent_state
->temp_hist
.decay();
13165 // Total objects operated on so far
13166 int total_started
= agent_state
->started
+ started
;
13167 bool need_delay
= false;
13169 dout(20) << __func__
<< " start pos " << agent_state
->position
13170 << " next start pos " << next
13171 << " started " << total_started
<< dendl
;
13173 // See if we've made a full pass over the object hash space
13174 // This might check at most ls_max objects a second time to notice that
13175 // we've checked every objects at least once.
13176 if (agent_state
->position
< agent_state
->start
&&
13177 next
>= agent_state
->start
) {
13178 dout(20) << __func__
<< " wrap around " << agent_state
->start
<< dendl
;
13179 if (total_started
== 0)
13183 agent_state
->start
= next
;
13185 agent_state
->started
= total_started
;
13187 // See if we are starting from beginning
13189 agent_state
->position
= hobject_t();
13191 agent_state
->position
= next
;
13193 // Discard old in memory HitSets
13194 hit_set_in_memory_trim(pool
.info
.hit_set_count
);
13197 assert(agent_state
->delaying
== false);
13202 agent_choose_mode();
13207 void PrimaryLogPG::agent_load_hit_sets()
13209 if (agent_state
->evict_mode
== TierAgentState::EVICT_MODE_IDLE
) {
13213 if (agent_state
->hit_set_map
.size() < info
.hit_set
.history
.size()) {
13214 dout(10) << __func__
<< dendl
;
13215 for (list
<pg_hit_set_info_t
>::iterator p
= info
.hit_set
.history
.begin();
13216 p
!= info
.hit_set
.history
.end(); ++p
) {
13217 if (agent_state
->hit_set_map
.count(p
->begin
.sec()) == 0) {
13218 dout(10) << __func__
<< " loading " << p
->begin
<< "-"
13219 << p
->end
<< dendl
;
13220 if (!pool
.info
.is_replicated()) {
13221 // FIXME: EC not supported here yet
13222 derr
<< __func__
<< " on non-replicated pool" << dendl
;
13226 hobject_t oid
= get_hit_set_archive_object(p
->begin
, p
->end
, p
->using_gmt
);
13227 if (is_unreadable_object(oid
)) {
13228 dout(10) << __func__
<< " unreadable " << oid
<< ", waiting" << dendl
;
13232 ObjectContextRef obc
= get_object_context(oid
, false);
13234 derr
<< __func__
<< ": could not load hitset " << oid
<< dendl
;
13240 obc
->ondisk_read_lock();
13241 int r
= osd
->store
->read(ch
, ghobject_t(oid
), 0, 0, bl
);
13243 obc
->ondisk_read_unlock();
13245 HitSetRef
hs(new HitSet
);
13246 bufferlist::iterator pbl
= bl
.begin();
13247 ::decode(*hs
, pbl
);
13248 agent_state
->add_hit_set(p
->begin
.sec(), hs
);
13254 bool PrimaryLogPG::agent_maybe_flush(ObjectContextRef
& obc
)
13256 if (!obc
->obs
.oi
.is_dirty()) {
13257 dout(20) << __func__
<< " skip (clean) " << obc
->obs
.oi
<< dendl
;
13258 osd
->logger
->inc(l_osd_agent_skip
);
13261 if (obc
->obs
.oi
.is_cache_pinned()) {
13262 dout(20) << __func__
<< " skip (cache_pinned) " << obc
->obs
.oi
<< dendl
;
13263 osd
->logger
->inc(l_osd_agent_skip
);
13267 utime_t now
= ceph_clock_now();
13268 utime_t ob_local_mtime
;
13269 if (obc
->obs
.oi
.local_mtime
!= utime_t()) {
13270 ob_local_mtime
= obc
->obs
.oi
.local_mtime
;
13272 ob_local_mtime
= obc
->obs
.oi
.mtime
;
13274 bool evict_mode_full
=
13275 (agent_state
->evict_mode
== TierAgentState::EVICT_MODE_FULL
);
13276 if (!evict_mode_full
&&
13277 obc
->obs
.oi
.soid
.snap
== CEPH_NOSNAP
&& // snaps immutable; don't delay
13278 (ob_local_mtime
+ utime_t(pool
.info
.cache_min_flush_age
, 0) > now
)) {
13279 dout(20) << __func__
<< " skip (too young) " << obc
->obs
.oi
<< dendl
;
13280 osd
->logger
->inc(l_osd_agent_skip
);
13284 if (osd
->agent_is_active_oid(obc
->obs
.oi
.soid
)) {
13285 dout(20) << __func__
<< " skip (flushing) " << obc
->obs
.oi
<< dendl
;
13286 osd
->logger
->inc(l_osd_agent_skip
);
13290 dout(10) << __func__
<< " flushing " << obc
->obs
.oi
<< dendl
;
13292 // FIXME: flush anything dirty, regardless of what distribution of
13295 hobject_t oid
= obc
->obs
.oi
.soid
;
13296 osd
->agent_start_op(oid
);
13297 // no need to capture a pg ref, can't outlive fop or ctx
13298 std::function
<void()> on_flush
= [this, oid
]() {
13299 osd
->agent_finish_op(oid
);
13302 int result
= start_flush(
13303 OpRequestRef(), obc
, false, NULL
,
13305 if (result
!= -EINPROGRESS
) {
13307 dout(10) << __func__
<< " start_flush() failed " << obc
->obs
.oi
13308 << " with " << result
<< dendl
;
13309 osd
->logger
->inc(l_osd_agent_skip
);
13313 osd
->logger
->inc(l_osd_agent_flush
);
13317 bool PrimaryLogPG::agent_maybe_evict(ObjectContextRef
& obc
, bool after_flush
)
13319 const hobject_t
& soid
= obc
->obs
.oi
.soid
;
13320 if (!after_flush
&& obc
->obs
.oi
.is_dirty()) {
13321 dout(20) << __func__
<< " skip (dirty) " << obc
->obs
.oi
<< dendl
;
13324 if (!obc
->obs
.oi
.watchers
.empty()) {
13325 dout(20) << __func__
<< " skip (watchers) " << obc
->obs
.oi
<< dendl
;
13328 if (obc
->is_blocked()) {
13329 dout(20) << __func__
<< " skip (blocked) " << obc
->obs
.oi
<< dendl
;
13332 if (obc
->obs
.oi
.is_cache_pinned()) {
13333 dout(20) << __func__
<< " skip (cache_pinned) " << obc
->obs
.oi
<< dendl
;
13337 if (soid
.snap
== CEPH_NOSNAP
) {
13338 int result
= _verify_no_head_clones(soid
, obc
->ssc
->snapset
);
13340 dout(20) << __func__
<< " skip (clones) " << obc
->obs
.oi
<< dendl
;
13345 if (agent_state
->evict_mode
!= TierAgentState::EVICT_MODE_FULL
) {
13346 // is this object old than cache_min_evict_age?
13347 utime_t now
= ceph_clock_now();
13348 utime_t ob_local_mtime
;
13349 if (obc
->obs
.oi
.local_mtime
!= utime_t()) {
13350 ob_local_mtime
= obc
->obs
.oi
.local_mtime
;
13352 ob_local_mtime
= obc
->obs
.oi
.mtime
;
13354 if (ob_local_mtime
+ utime_t(pool
.info
.cache_min_evict_age
, 0) > now
) {
13355 dout(20) << __func__
<< " skip (too young) " << obc
->obs
.oi
<< dendl
;
13356 osd
->logger
->inc(l_osd_agent_skip
);
13359 // is this object old and/or cold enough?
13361 uint64_t temp_upper
= 0, temp_lower
= 0;
13363 agent_estimate_temp(soid
, &temp
);
13364 agent_state
->temp_hist
.add(temp
);
13365 agent_state
->temp_hist
.get_position_micro(temp
, &temp_lower
, &temp_upper
);
13367 dout(20) << __func__
13368 << " temp " << temp
13369 << " pos " << temp_lower
<< "-" << temp_upper
13370 << ", evict_effort " << agent_state
->evict_effort
13372 dout(30) << "agent_state:\n";
13373 Formatter
*f
= Formatter::create("");
13374 f
->open_object_section("agent_state");
13375 agent_state
->dump(f
);
13376 f
->close_section();
13381 if (1000000 - temp_upper
>= agent_state
->evict_effort
)
13385 dout(10) << __func__
<< " evicting " << obc
->obs
.oi
<< dendl
;
13386 OpContextUPtr ctx
= simple_opc_create(obc
);
13388 if (!ctx
->lock_manager
.get_lock_type(
13389 ObjectContext::RWState::RWWRITE
,
13393 close_op_ctx(ctx
.release());
13394 dout(20) << __func__
<< " skip (cannot get lock) " << obc
->obs
.oi
<< dendl
;
13398 osd
->agent_start_evict_op();
13399 ctx
->register_on_finish(
13401 osd
->agent_finish_evict_op();
13404 ctx
->at_version
= get_next_version();
13405 assert(ctx
->new_obs
.exists
);
13406 int r
= _delete_oid(ctx
.get(), true, false);
13407 if (obc
->obs
.oi
.is_omap())
13408 ctx
->delta_stats
.num_objects_omap
--;
13409 ctx
->delta_stats
.num_evict
++;
13410 ctx
->delta_stats
.num_evict_kb
+= SHIFT_ROUND_UP(obc
->obs
.oi
.size
, 10);
13411 if (obc
->obs
.oi
.is_dirty())
13412 --ctx
->delta_stats
.num_objects_dirty
;
13414 finish_ctx(ctx
.get(), pg_log_entry_t::DELETE
, false);
13415 simple_opc_submit(std::move(ctx
));
13416 osd
->logger
->inc(l_osd_tier_evict
);
13417 osd
->logger
->inc(l_osd_agent_evict
);
13421 void PrimaryLogPG::agent_stop()
13423 dout(20) << __func__
<< dendl
;
13424 if (agent_state
&& !agent_state
->is_idle()) {
13425 agent_state
->evict_mode
= TierAgentState::EVICT_MODE_IDLE
;
13426 agent_state
->flush_mode
= TierAgentState::FLUSH_MODE_IDLE
;
13427 osd
->agent_disable_pg(this, agent_state
->evict_effort
);
13431 void PrimaryLogPG::agent_delay()
13433 dout(20) << __func__
<< dendl
;
13434 if (agent_state
&& !agent_state
->is_idle()) {
13435 assert(agent_state
->delaying
== false);
13436 agent_state
->delaying
= true;
13437 osd
->agent_disable_pg(this, agent_state
->evict_effort
);
13441 void PrimaryLogPG::agent_choose_mode_restart()
13443 dout(20) << __func__
<< dendl
;
13445 if (agent_state
&& agent_state
->delaying
) {
13446 agent_state
->delaying
= false;
13447 agent_choose_mode(true);
13452 bool PrimaryLogPG::agent_choose_mode(bool restart
, OpRequestRef op
)
13454 bool requeued
= false;
13455 // Let delay play out
13456 if (agent_state
->delaying
) {
13457 dout(20) << __func__
<< this << " delaying, ignored" << dendl
;
13461 TierAgentState::flush_mode_t flush_mode
= TierAgentState::FLUSH_MODE_IDLE
;
13462 TierAgentState::evict_mode_t evict_mode
= TierAgentState::EVICT_MODE_IDLE
;
13463 unsigned evict_effort
= 0;
13465 if (info
.stats
.stats_invalid
) {
13466 // idle; stats can't be trusted until we scrub.
13467 dout(20) << __func__
<< " stats invalid (post-split), idle" << dendl
;
13472 uint64_t divisor
= pool
.info
.get_pg_num_divisor(info
.pgid
.pgid
);
13473 assert(divisor
> 0);
13475 // adjust (effective) user objects down based on the number
13476 // of HitSet objects, which should not count toward our total since
13477 // they cannot be flushed.
13478 uint64_t unflushable
= info
.stats
.stats
.sum
.num_objects_hit_set_archive
;
13480 // also exclude omap objects if ec backing pool
13481 const pg_pool_t
*base_pool
= get_osdmap()->get_pg_pool(pool
.info
.tier_of
);
13483 if (!base_pool
->supports_omap())
13484 unflushable
+= info
.stats
.stats
.sum
.num_objects_omap
;
13486 uint64_t num_user_objects
= info
.stats
.stats
.sum
.num_objects
;
13487 if (num_user_objects
> unflushable
)
13488 num_user_objects
-= unflushable
;
13490 num_user_objects
= 0;
13492 uint64_t num_user_bytes
= info
.stats
.stats
.sum
.num_bytes
;
13493 uint64_t unflushable_bytes
= info
.stats
.stats
.sum
.num_bytes_hit_set_archive
;
13494 num_user_bytes
-= unflushable_bytes
;
13495 uint64_t num_overhead_bytes
= osd
->store
->estimate_objects_overhead(num_user_objects
);
13496 num_user_bytes
+= num_overhead_bytes
;
13498 // also reduce the num_dirty by num_objects_omap
13499 int64_t num_dirty
= info
.stats
.stats
.sum
.num_objects_dirty
;
13500 if (!base_pool
->supports_omap()) {
13501 if (num_dirty
> info
.stats
.stats
.sum
.num_objects_omap
)
13502 num_dirty
-= info
.stats
.stats
.sum
.num_objects_omap
;
13507 dout(10) << __func__
13509 << TierAgentState::get_flush_mode_name(agent_state
->flush_mode
)
13511 << TierAgentState::get_evict_mode_name(agent_state
->evict_mode
)
13512 << " num_objects: " << info
.stats
.stats
.sum
.num_objects
13513 << " num_bytes: " << info
.stats
.stats
.sum
.num_bytes
13514 << " num_objects_dirty: " << info
.stats
.stats
.sum
.num_objects_dirty
13515 << " num_objects_omap: " << info
.stats
.stats
.sum
.num_objects_omap
13516 << " num_dirty: " << num_dirty
13517 << " num_user_objects: " << num_user_objects
13518 << " num_user_bytes: " << num_user_bytes
13519 << " num_overhead_bytes: " << num_overhead_bytes
13520 << " pool.info.target_max_bytes: " << pool
.info
.target_max_bytes
13521 << " pool.info.target_max_objects: " << pool
.info
.target_max_objects
13524 // get dirty, full ratios
13525 uint64_t dirty_micro
= 0;
13526 uint64_t full_micro
= 0;
13527 if (pool
.info
.target_max_bytes
&& num_user_objects
> 0) {
13528 uint64_t avg_size
= num_user_bytes
/ num_user_objects
;
13530 num_dirty
* avg_size
* 1000000 /
13531 MAX(pool
.info
.target_max_bytes
/ divisor
, 1);
13533 num_user_objects
* avg_size
* 1000000 /
13534 MAX(pool
.info
.target_max_bytes
/ divisor
, 1);
13536 if (pool
.info
.target_max_objects
> 0) {
13537 uint64_t dirty_objects_micro
=
13538 num_dirty
* 1000000 /
13539 MAX(pool
.info
.target_max_objects
/ divisor
, 1);
13540 if (dirty_objects_micro
> dirty_micro
)
13541 dirty_micro
= dirty_objects_micro
;
13542 uint64_t full_objects_micro
=
13543 num_user_objects
* 1000000 /
13544 MAX(pool
.info
.target_max_objects
/ divisor
, 1);
13545 if (full_objects_micro
> full_micro
)
13546 full_micro
= full_objects_micro
;
13548 dout(20) << __func__
<< " dirty " << ((float)dirty_micro
/ 1000000.0)
13549 << " full " << ((float)full_micro
/ 1000000.0)
13553 uint64_t flush_target
= pool
.info
.cache_target_dirty_ratio_micro
;
13554 uint64_t flush_high_target
= pool
.info
.cache_target_dirty_high_ratio_micro
;
13555 uint64_t flush_slop
= (float)flush_target
* cct
->_conf
->osd_agent_slop
;
13556 if (restart
|| agent_state
->flush_mode
== TierAgentState::FLUSH_MODE_IDLE
) {
13557 flush_target
+= flush_slop
;
13558 flush_high_target
+= flush_slop
;
13560 flush_target
-= MIN(flush_target
, flush_slop
);
13561 flush_high_target
-= MIN(flush_high_target
, flush_slop
);
13564 if (dirty_micro
> flush_high_target
) {
13565 flush_mode
= TierAgentState::FLUSH_MODE_HIGH
;
13566 } else if (dirty_micro
> flush_target
) {
13567 flush_mode
= TierAgentState::FLUSH_MODE_LOW
;
13571 uint64_t evict_target
= pool
.info
.cache_target_full_ratio_micro
;
13572 uint64_t evict_slop
= (float)evict_target
* cct
->_conf
->osd_agent_slop
;
13573 if (restart
|| agent_state
->evict_mode
== TierAgentState::EVICT_MODE_IDLE
)
13574 evict_target
+= evict_slop
;
13576 evict_target
-= MIN(evict_target
, evict_slop
);
13578 if (full_micro
> 1000000) {
13579 // evict anything clean
13580 evict_mode
= TierAgentState::EVICT_MODE_FULL
;
13581 evict_effort
= 1000000;
13582 } else if (full_micro
> evict_target
) {
13583 // set effort in [0..1] range based on where we are between
13584 evict_mode
= TierAgentState::EVICT_MODE_SOME
;
13585 uint64_t over
= full_micro
- evict_target
;
13586 uint64_t span
= 1000000 - evict_target
;
13587 evict_effort
= MAX(over
* 1000000 / span
,
13588 (unsigned)(1000000.0 * cct
->_conf
->osd_agent_min_evict_effort
));
13590 // quantize effort to avoid too much reordering in the agent_queue.
13591 uint64_t inc
= cct
->_conf
->osd_agent_quantize_effort
* 1000000;
13593 uint64_t was
= evict_effort
;
13594 evict_effort
-= evict_effort
% inc
;
13595 if (evict_effort
< inc
)
13596 evict_effort
= inc
;
13597 assert(evict_effort
>= inc
&& evict_effort
<= 1000000);
13598 dout(30) << __func__
<< " evict_effort " << was
<< " quantized by " << inc
<< " to " << evict_effort
<< dendl
;
13603 bool old_idle
= agent_state
->is_idle();
13604 if (flush_mode
!= agent_state
->flush_mode
) {
13605 dout(5) << __func__
<< " flush_mode "
13606 << TierAgentState::get_flush_mode_name(agent_state
->flush_mode
)
13608 << TierAgentState::get_flush_mode_name(flush_mode
)
13610 if (flush_mode
== TierAgentState::FLUSH_MODE_HIGH
) {
13611 osd
->agent_inc_high_count();
13612 info
.stats
.stats
.sum
.num_flush_mode_high
= 1;
13613 } else if (flush_mode
== TierAgentState::FLUSH_MODE_LOW
) {
13614 info
.stats
.stats
.sum
.num_flush_mode_low
= 1;
13616 if (agent_state
->flush_mode
== TierAgentState::FLUSH_MODE_HIGH
) {
13617 osd
->agent_dec_high_count();
13618 info
.stats
.stats
.sum
.num_flush_mode_high
= 0;
13619 } else if (agent_state
->flush_mode
== TierAgentState::FLUSH_MODE_LOW
) {
13620 info
.stats
.stats
.sum
.num_flush_mode_low
= 0;
13622 agent_state
->flush_mode
= flush_mode
;
13624 if (evict_mode
!= agent_state
->evict_mode
) {
13625 dout(5) << __func__
<< " evict_mode "
13626 << TierAgentState::get_evict_mode_name(agent_state
->evict_mode
)
13628 << TierAgentState::get_evict_mode_name(evict_mode
)
13630 if (agent_state
->evict_mode
== TierAgentState::EVICT_MODE_FULL
&&
13634 requeue_ops(waiting_for_flush
);
13635 requeue_ops(waiting_for_active
);
13636 requeue_ops(waiting_for_scrub
);
13637 requeue_ops(waiting_for_cache_not_full
);
13638 objects_blocked_on_cache_full
.clear();
13641 if (evict_mode
== TierAgentState::EVICT_MODE_SOME
) {
13642 info
.stats
.stats
.sum
.num_evict_mode_some
= 1;
13643 } else if (evict_mode
== TierAgentState::EVICT_MODE_FULL
) {
13644 info
.stats
.stats
.sum
.num_evict_mode_full
= 1;
13646 if (agent_state
->evict_mode
== TierAgentState::EVICT_MODE_SOME
) {
13647 info
.stats
.stats
.sum
.num_evict_mode_some
= 0;
13648 } else if (agent_state
->evict_mode
== TierAgentState::EVICT_MODE_FULL
) {
13649 info
.stats
.stats
.sum
.num_evict_mode_full
= 0;
13651 agent_state
->evict_mode
= evict_mode
;
13653 uint64_t old_effort
= agent_state
->evict_effort
;
13654 if (evict_effort
!= agent_state
->evict_effort
) {
13655 dout(5) << __func__
<< " evict_effort "
13656 << ((float)agent_state
->evict_effort
/ 1000000.0)
13658 << ((float)evict_effort
/ 1000000.0)
13660 agent_state
->evict_effort
= evict_effort
;
13663 // NOTE: we are using evict_effort as a proxy for *all* agent effort
13664 // (including flush). This is probably fine (they should be
13665 // correlated) but it is not precisely correct.
13666 if (agent_state
->is_idle()) {
13667 if (!restart
&& !old_idle
) {
13668 osd
->agent_disable_pg(this, old_effort
);
13671 if (restart
|| old_idle
) {
13672 osd
->agent_enable_pg(this, agent_state
->evict_effort
);
13673 } else if (old_effort
!= agent_state
->evict_effort
) {
13674 osd
->agent_adjust_pg(this, old_effort
, agent_state
->evict_effort
);
13680 void PrimaryLogPG::agent_estimate_temp(const hobject_t
& oid
, int *temp
)
13685 if (hit_set
->contains(oid
))
13688 int last_n
= pool
.info
.hit_set_search_last_n
;
13689 for (map
<time_t,HitSetRef
>::reverse_iterator p
=
13690 agent_state
->hit_set_map
.rbegin(); last_n
> 0 &&
13691 p
!= agent_state
->hit_set_map
.rend(); ++p
, ++i
) {
13692 if (p
->second
->contains(oid
)) {
13693 *temp
+= pool
.info
.get_grade(i
);
13699 // Dup op detection
13701 bool PrimaryLogPG::already_complete(eversion_t v
)
13703 dout(20) << __func__
<< ": " << v
<< dendl
;
13704 for (xlist
<RepGather
*>::iterator i
= repop_queue
.begin();
13707 dout(20) << __func__
<< ": " << **i
<< dendl
;
13708 // skip copy from temp object ops
13709 if ((*i
)->v
== eversion_t()) {
13710 dout(20) << __func__
<< ": " << **i
13711 << " version is empty" << dendl
;
13715 dout(20) << __func__
<< ": " << **i
13716 << " (*i)->v past v" << dendl
;
13719 if (!(*i
)->all_committed
) {
13720 dout(20) << __func__
<< ": " << **i
13721 << " not committed, returning false"
13726 dout(20) << __func__
<< ": returning true" << dendl
;
13730 bool PrimaryLogPG::already_ack(eversion_t v
)
13732 dout(20) << __func__
<< ": " << v
<< dendl
;
13733 for (xlist
<RepGather
*>::iterator i
= repop_queue
.begin();
13736 // skip copy from temp object ops
13737 if ((*i
)->v
== eversion_t()) {
13738 dout(20) << __func__
<< ": " << **i
13739 << " version is empty" << dendl
;
13743 dout(20) << __func__
<< ": " << **i
13744 << " (*i)->v past v" << dendl
;
13747 if (!(*i
)->all_applied
) {
13748 dout(20) << __func__
<< ": " << **i
13749 << " not applied, returning false"
13754 dout(20) << __func__
<< ": returning true" << dendl
;
13759 // ==========================================================================================
13763 bool PrimaryLogPG::_range_available_for_scrub(
13764 const hobject_t
&begin
, const hobject_t
&end
)
13766 pair
<hobject_t
, ObjectContextRef
> next
;
13767 next
.second
= object_contexts
.lookup(begin
);
13768 next
.first
= begin
;
13770 while (more
&& next
.first
< end
) {
13771 if (next
.second
&& next
.second
->is_blocked()) {
13772 next
.second
->requeue_scrub_on_unblock
= true;
13773 dout(10) << __func__
<< ": scrub delayed, "
13774 << next
.first
<< " is blocked"
13778 more
= object_contexts
.get_next(next
.first
, &next
);
13783 static bool doing_clones(const boost::optional
<SnapSet
> &snapset
,
13784 const vector
<snapid_t
>::reverse_iterator
&curclone
) {
13785 return snapset
&& curclone
!= snapset
.get().clones
.rend();
13788 void PrimaryLogPG::log_missing(unsigned missing
,
13789 const boost::optional
<hobject_t
> &head
,
13790 LogChannelRef clog
,
13794 bool allow_incomplete_clones
)
13797 if (allow_incomplete_clones
) {
13798 dout(20) << func
<< " " << mode
<< " " << pgid
<< " " << head
.get()
13799 << " skipped " << missing
<< " clone(s) in cache tier" << dendl
;
13801 clog
->info() << mode
<< " " << pgid
<< " " << head
.get()
13802 << " " << missing
<< " missing clone(s)";
13806 unsigned PrimaryLogPG::process_clones_to(const boost::optional
<hobject_t
> &head
,
13807 const boost::optional
<SnapSet
> &snapset
,
13808 LogChannelRef clog
,
13811 bool allow_incomplete_clones
,
13812 boost::optional
<snapid_t
> target
,
13813 vector
<snapid_t
>::reverse_iterator
*curclone
,
13814 inconsistent_snapset_wrapper
&e
)
13818 unsigned missing
= 0;
13820 // NOTE: clones are in descending order, thus **curclone > target test here
13821 hobject_t
next_clone(head
.get());
13822 while(doing_clones(snapset
, *curclone
) && (!target
|| **curclone
> *target
)) {
13824 // it is okay to be missing one or more clones in a cache tier.
13825 // skip higher-numbered clones in the list.
13826 if (!allow_incomplete_clones
) {
13827 next_clone
.snap
= **curclone
;
13828 clog
->error() << mode
<< " " << pgid
<< " " << head
.get()
13829 << " expected clone " << next_clone
<< " " << missing
13831 ++scrubber
.shallow_errors
;
13832 e
.set_clone_missing(next_clone
.snap
);
13834 // Clones are descending
13841 * Validate consistency of the object info and snap sets.
13843 * We are sort of comparing 2 lists. The main loop is on objmap.objects. But
13844 * the comparison of the objects is against multiple snapset.clones. There are
13845 * multiple clone lists and in between lists we expect head or snapdir.
13851 * obj1 snap 1 head/snapdir, unexpected obj1 snap 1
13852 * obj2 head head/snapdir, head ok
13853 * [SnapSet clones 6 4 2 1]
13854 * obj2 snap 7 obj2 snap 6, unexpected obj2 snap 7
13855 * obj2 snap 6 obj2 snap 6, match
13856 * obj2 snap 4 obj2 snap 4, match
13857 * obj3 head obj2 snap 2 (expected), obj2 snap 1 (expected), head ok
13858 * [Snapset clones 3 1]
13859 * obj3 snap 3 obj3 snap 3 match
13860 * obj3 snap 1 obj3 snap 1 match
13861 * obj4 snapdir head/snapdir, snapdir ok
13862 * [Snapset clones 4]
13863 * EOL obj4 snap 4, (expected)
13865 void PrimaryLogPG::scrub_snapshot_metadata(
13866 ScrubMap
&scrubmap
,
13867 const map
<hobject_t
,
13868 pair
<boost::optional
<uint32_t>,
13869 boost::optional
<uint32_t>>> &missing_digest
)
13871 dout(10) << __func__
<< dendl
;
13873 coll_t
c(info
.pgid
);
13874 bool repair
= state_test(PG_STATE_REPAIR
);
13875 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
13876 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
13877 boost::optional
<snapid_t
> all_clones
; // Unspecified snapid_t or boost::none
13879 /// snapsets to repair
13880 map
<hobject_t
,SnapSet
> snapset_to_repair
;
13882 // traverse in reverse order.
13883 boost::optional
<hobject_t
> head
;
13884 boost::optional
<SnapSet
> snapset
; // If initialized so will head (above)
13885 vector
<snapid_t
>::reverse_iterator curclone
; // Defined only if snapset initialized
13886 unsigned missing
= 0;
13887 inconsistent_snapset_wrapper soid_error
, head_error
;
13888 unsigned soid_error_count
= 0;
13890 bufferlist last_data
;
13892 for (map
<hobject_t
,ScrubMap::object
>::reverse_iterator
13893 p
= scrubmap
.objects
.rbegin(); p
!= scrubmap
.objects
.rend(); ++p
) {
13894 const hobject_t
& soid
= p
->first
;
13895 soid_error
= inconsistent_snapset_wrapper
{soid
};
13896 object_stat_sum_t stat
;
13897 boost::optional
<object_info_t
> oi
;
13899 if (!soid
.is_snapdir())
13900 stat
.num_objects
++;
13902 if (soid
.nspace
== cct
->_conf
->osd_hit_set_namespace
)
13903 stat
.num_objects_hit_set_archive
++;
13905 if (soid
.is_snap()) {
13907 stat
.num_object_clones
++;
13911 if (p
->second
.attrs
.count(OI_ATTR
) == 0) {
13913 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
13914 << " no '" << OI_ATTR
<< "' attr";
13915 ++scrubber
.shallow_errors
;
13916 soid_error
.set_info_missing();
13919 bv
.push_back(p
->second
.attrs
[OI_ATTR
]);
13921 oi
= object_info_t(); // Initialize optional<> before decode into it
13922 oi
.get().decode(bv
);
13923 } catch (buffer::error
& e
) {
13925 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
13926 << " can't decode '" << OI_ATTR
<< "' attr " << e
.what();
13927 ++scrubber
.shallow_errors
;
13928 soid_error
.set_info_corrupted();
13929 soid_error
.set_info_missing(); // Not available too
13934 if (pgbackend
->be_get_ondisk_size(oi
->size
) != p
->second
.size
) {
13935 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
13936 << " on disk size (" << p
->second
.size
13937 << ") does not match object info size ("
13938 << oi
->size
<< ") adjusted for ondisk to ("
13939 << pgbackend
->be_get_ondisk_size(oi
->size
)
13941 soid_error
.set_size_mismatch();
13942 ++scrubber
.shallow_errors
;
13945 dout(20) << mode
<< " " << soid
<< " " << oi
.get() << dendl
;
13947 // A clone num_bytes will be added later when we have snapset
13948 if (!soid
.is_snap()) {
13949 stat
.num_bytes
+= oi
->size
;
13951 if (soid
.nspace
== cct
->_conf
->osd_hit_set_namespace
)
13952 stat
.num_bytes_hit_set_archive
+= oi
->size
;
13954 if (!soid
.is_snapdir()) {
13955 if (oi
->is_dirty())
13956 ++stat
.num_objects_dirty
;
13957 if (oi
->is_whiteout())
13958 ++stat
.num_whiteouts
;
13960 ++stat
.num_objects_omap
;
13961 if (oi
->is_cache_pinned())
13962 ++stat
.num_objects_pinned
;
13965 // pessimistic assumption that this object might contain a
13967 stat
.num_legacy_snapsets
++;
13970 // Check for any problems while processing clones
13971 if (doing_clones(snapset
, curclone
)) {
13972 boost::optional
<snapid_t
> target
;
13973 // Expecting an object with snap for current head
13974 if (soid
.has_snapset() || soid
.get_head() != head
->get_head()) {
13976 dout(10) << __func__
<< " " << mode
<< " " << info
.pgid
<< " new object "
13977 << soid
<< " while processing " << head
.get() << dendl
;
13979 target
= all_clones
;
13981 assert(soid
.is_snap());
13982 target
= soid
.snap
;
13985 // Log any clones we were expecting to be there up to target
13986 // This will set missing, but will be a no-op if snap.soid == *curclone.
13987 missing
+= process_clones_to(head
, snapset
, osd
->clog
, info
.pgid
, mode
,
13988 pool
.info
.allow_incomplete_clones(), target
, &curclone
,
13992 // Check doing_clones() again in case we ran process_clones_to()
13993 if (doing_clones(snapset
, curclone
)) {
13994 // A head/snapdir would have processed all clones above
13995 // or all greater than *curclone.
13996 assert(soid
.is_snap() && *curclone
<= soid
.snap
);
13998 // After processing above clone snap should match the expected curclone
13999 expected
= (*curclone
== soid
.snap
);
14001 // If we aren't doing clones any longer, then expecting head/snapdir
14002 expected
= soid
.has_snapset();
14005 // If we couldn't read the head's snapset, just ignore clones
14006 if (head
&& !snapset
) {
14007 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14008 << " clone ignored due to missing snapset";
14010 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14011 << " is an unexpected clone";
14013 ++scrubber
.shallow_errors
;
14014 soid_error
.set_headless();
14015 scrubber
.store
->add_snap_error(pool
.id
, soid_error
);
14016 ++soid_error_count
;
14017 if (head
&& soid
.get_head() == head
->get_head())
14018 head_error
.set_clone(soid
.snap
);
14023 if (soid
.has_snapset()) {
14026 log_missing(missing
, head
, osd
->clog
, info
.pgid
, __func__
, mode
,
14027 pool
.info
.allow_incomplete_clones());
14030 // Save previous head error information
14031 if (head
&& (head_error
.errors
|| soid_error_count
))
14032 scrubber
.store
->add_snap_error(pool
.id
, head_error
);
14033 // Set this as a new head object
14036 head_error
= soid_error
;
14037 soid_error_count
= 0;
14039 dout(20) << __func__
<< " " << mode
<< " new head " << head
<< dendl
;
14041 if (p
->second
.attrs
.count(SS_ATTR
) == 0) {
14042 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14043 << " no '" << SS_ATTR
<< "' attr";
14044 ++scrubber
.shallow_errors
;
14045 snapset
= boost::none
;
14046 head_error
.set_snapset_missing();
14049 bl
.push_back(p
->second
.attrs
[SS_ATTR
]);
14050 bufferlist::iterator blp
= bl
.begin();
14052 snapset
= SnapSet(); // Initialize optional<> before decoding into it
14053 ::decode(snapset
.get(), blp
);
14054 head_error
.ss_bl
.push_back(p
->second
.attrs
[SS_ATTR
]);
14055 } catch (buffer::error
& e
) {
14056 snapset
= boost::none
;
14057 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14058 << " can't decode '" << SS_ATTR
<< "' attr " << e
.what();
14059 ++scrubber
.shallow_errors
;
14060 head_error
.set_snapset_corrupted();
14065 // what will be next?
14066 curclone
= snapset
->clones
.rbegin();
14068 if (!snapset
->clones
.empty()) {
14069 dout(20) << " snapset " << snapset
.get() << dendl
;
14070 if (snapset
->seq
== 0) {
14071 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14072 << " snaps.seq not set";
14073 ++scrubber
.shallow_errors
;
14074 head_error
.set_snapset_error();
14078 if (soid
.is_head() && !snapset
->head_exists
) {
14079 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14080 << " snapset.head_exists=false, but head exists";
14081 ++scrubber
.shallow_errors
;
14082 head_error
.set_head_mismatch();
14083 // Fix head_exists locally so is_legacy() returns correctly
14084 snapset
->head_exists
= true;
14086 if (soid
.is_snapdir() && snapset
->head_exists
) {
14087 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14088 << " snapset.head_exists=true, but snapdir exists";
14089 ++scrubber
.shallow_errors
;
14090 head_error
.set_head_mismatch();
14091 // For symmetry fix this too, but probably doesn't matter
14092 snapset
->head_exists
= false;
14095 if (get_osdmap()->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
14096 if (soid
.is_snapdir()) {
14097 dout(10) << " will move snapset to head from " << soid
<< dendl
;
14098 snapset_to_repair
[soid
.get_head()] = *snapset
;
14099 } else if (snapset
->is_legacy()) {
14100 dout(10) << " will convert legacy snapset on " << soid
<< " " << *snapset
14102 snapset_to_repair
[soid
.get_head()] = *snapset
;
14105 stat
.num_legacy_snapsets
++;
14108 // pessimistic assumption that this object might contain a
14110 stat
.num_legacy_snapsets
++;
14113 assert(soid
.is_snap());
14116 assert(soid
.snap
== *curclone
);
14118 dout(20) << __func__
<< " " << mode
<< " matched clone " << soid
<< dendl
;
14120 if (snapset
->clone_size
.count(soid
.snap
) == 0) {
14121 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14122 << " is missing in clone_size";
14123 ++scrubber
.shallow_errors
;
14124 soid_error
.set_size_mismatch();
14126 if (oi
&& oi
->size
!= snapset
->clone_size
[soid
.snap
]) {
14127 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14128 << " size " << oi
->size
<< " != clone_size "
14129 << snapset
->clone_size
[*curclone
];
14130 ++scrubber
.shallow_errors
;
14131 soid_error
.set_size_mismatch();
14134 if (snapset
->clone_overlap
.count(soid
.snap
) == 0) {
14135 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14136 << " is missing in clone_overlap";
14137 ++scrubber
.shallow_errors
;
14138 soid_error
.set_size_mismatch();
14140 // This checking is based on get_clone_bytes(). The first 2 asserts
14141 // can't happen because we know we have a clone_size and
14142 // a clone_overlap. Now we check that the interval_set won't
14143 // cause the last assert.
14144 uint64_t size
= snapset
->clone_size
.find(soid
.snap
)->second
;
14145 const interval_set
<uint64_t> &overlap
=
14146 snapset
->clone_overlap
.find(soid
.snap
)->second
;
14147 bool bad_interval_set
= false;
14148 for (interval_set
<uint64_t>::const_iterator i
= overlap
.begin();
14149 i
!= overlap
.end(); ++i
) {
14150 if (size
< i
.get_len()) {
14151 bad_interval_set
= true;
14154 size
-= i
.get_len();
14157 if (bad_interval_set
) {
14158 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14159 << " bad interval_set in clone_overlap";
14160 ++scrubber
.shallow_errors
;
14161 soid_error
.set_size_mismatch();
14163 stat
.num_bytes
+= snapset
->get_clone_bytes(soid
.snap
);
14168 // migrate legacy_snaps to snapset?
14169 auto p
= snapset_to_repair
.find(soid
.get_head());
14170 if (p
!= snapset_to_repair
.end()) {
14171 if (!oi
|| oi
->legacy_snaps
.empty()) {
14172 osd
->clog
->error() << mode
<< " " << info
.pgid
<< " " << soid
14173 << " has no oi or legacy_snaps; cannot convert "
14175 ++scrubber
.shallow_errors
;
14177 dout(20) << __func__
<< " copying legacy_snaps " << oi
->legacy_snaps
14178 << " to snapset " << p
->second
<< dendl
;
14179 p
->second
.clone_snaps
[soid
.snap
] = oi
->legacy_snaps
;
14185 if (soid_error
.errors
) {
14186 scrubber
.store
->add_snap_error(pool
.id
, soid_error
);
14187 ++soid_error_count
;
14191 scrub_cstat
.add(stat
);
14194 if (doing_clones(snapset
, curclone
)) {
14195 dout(10) << __func__
<< " " << mode
<< " " << info
.pgid
14196 << " No more objects while processing " << head
.get() << dendl
;
14198 missing
+= process_clones_to(head
, snapset
, osd
->clog
, info
.pgid
, mode
,
14199 pool
.info
.allow_incomplete_clones(), all_clones
, &curclone
,
14202 // There could be missing found by the test above or even
14203 // before dropping out of the loop for the last head.
14205 log_missing(missing
, head
, osd
->clog
, info
.pgid
, __func__
,
14206 mode
, pool
.info
.allow_incomplete_clones());
14208 if (head
&& (head_error
.errors
|| soid_error_count
))
14209 scrubber
.store
->add_snap_error(pool
.id
, head_error
);
14211 for (auto p
= missing_digest
.begin(); p
!= missing_digest
.end(); ++p
) {
14212 if (p
->first
.is_snapdir())
14214 dout(10) << __func__
<< " recording digests for " << p
->first
<< dendl
;
14215 ObjectContextRef obc
= get_object_context(p
->first
, false);
14217 osd
->clog
->error() << info
.pgid
<< " " << mode
14218 << " cannot get object context for object "
14221 } else if (obc
->obs
.oi
.soid
!= p
->first
) {
14222 osd
->clog
->error() << info
.pgid
<< " " << mode
14223 << " object " << p
->first
14224 << " has a valid oi attr with a mismatched name, "
14225 << " obc->obs.oi.soid: " << obc
->obs
.oi
.soid
;
14228 OpContextUPtr ctx
= simple_opc_create(obc
);
14229 ctx
->at_version
= get_next_version();
14230 ctx
->mtime
= utime_t(); // do not update mtime
14231 if (p
->second
.first
) {
14232 ctx
->new_obs
.oi
.set_data_digest(*p
->second
.first
);
14234 ctx
->new_obs
.oi
.clear_data_digest();
14236 if (p
->second
.second
) {
14237 ctx
->new_obs
.oi
.set_omap_digest(*p
->second
.second
);
14239 ctx
->new_obs
.oi
.clear_omap_digest();
14241 finish_ctx(ctx
.get(), pg_log_entry_t::MODIFY
);
14243 ctx
->register_on_success(
14245 dout(20) << "updating scrub digest" << dendl
;
14246 if (--scrubber
.num_digest_updates_pending
== 0) {
14251 simple_opc_submit(std::move(ctx
));
14252 ++scrubber
.num_digest_updates_pending
;
14254 for (auto& p
: snapset_to_repair
) {
14255 // cache pools may not have the clones, which means we won't know
14256 // what snaps they have. fake out the clone_snaps entries anyway (with
14257 // blank snap lists).
14258 p
.second
.head_exists
= true;
14259 if (pool
.info
.allow_incomplete_clones()) {
14260 for (auto s
: p
.second
.clones
) {
14261 if (p
.second
.clone_snaps
.count(s
) == 0) {
14262 dout(10) << __func__
<< " " << p
.first
<< " faking clone_snaps for "
14264 p
.second
.clone_snaps
[s
];
14268 if (p
.second
.clones
.size() != p
.second
.clone_snaps
.size() ||
14269 p
.second
.is_legacy()) {
14270 // this happens if we encounter other errors above, like a missing
14272 dout(10) << __func__
<< " not writing snapset to " << p
.first
14273 << " snapset " << p
.second
<< " clones " << p
.second
.clones
14274 << "; didn't convert fully" << dendl
;
14275 scrub_cstat
.sum
.num_legacy_snapsets
++;
14278 dout(10) << __func__
<< " writing snapset to " << p
.first
14279 << " " << p
.second
<< dendl
;
14280 ObjectContextRef obc
= get_object_context(p
.first
, true);
14282 osd
->clog
->error() << info
.pgid
<< " " << mode
14283 << " cannot get object context for object "
14286 } else if (obc
->obs
.oi
.soid
!= p
.first
) {
14287 osd
->clog
->error() << info
.pgid
<< " " << mode
14288 << " object " << p
.first
14289 << " has a valid oi attr with a mismatched name, "
14290 << " obc->obs.oi.soid: " << obc
->obs
.oi
.soid
;
14293 ObjectContextRef snapset_obc
;
14294 if (!obc
->obs
.exists
) {
14295 snapset_obc
= get_object_context(p
.first
.get_snapdir(), false);
14296 if (!snapset_obc
) {
14297 osd
->clog
->error() << info
.pgid
<< " " << mode
14298 << " cannot get object context for "
14299 << p
.first
.get_snapdir();
14303 OpContextUPtr ctx
= simple_opc_create(obc
);
14304 PGTransaction
*t
= ctx
->op_t
.get();
14305 ctx
->snapset_obc
= snapset_obc
;
14306 ctx
->at_version
= get_next_version();
14307 ctx
->mtime
= utime_t(); // do not update mtime
14308 ctx
->new_snapset
= p
.second
;
14309 if (!ctx
->new_obs
.exists
) {
14310 dout(20) << __func__
<< " making " << p
.first
<< " a whiteout" << dendl
;
14311 ctx
->new_obs
.exists
= true;
14312 ctx
->new_snapset
.head_exists
= true;
14313 ctx
->new_obs
.oi
.set_flag(object_info_t::FLAG_WHITEOUT
);
14314 ++ctx
->delta_stats
.num_whiteouts
;
14315 ++ctx
->delta_stats
.num_objects
;
14316 t
->create(p
.first
);
14317 if (p
.first
< scrubber
.start
) {
14318 dout(20) << __func__
<< " kludging around update outside of scrub range"
14321 scrub_cstat
.add(ctx
->delta_stats
);
14324 dout(20) << __func__
<< " final snapset " << ctx
->new_snapset
<< dendl
;
14325 assert(!ctx
->new_snapset
.is_legacy());
14326 finish_ctx(ctx
.get(), pg_log_entry_t::MODIFY
);
14327 ctx
->register_on_success(
14329 dout(20) << "updating snapset" << dendl
;
14330 if (--scrubber
.num_digest_updates_pending
== 0) {
14335 simple_opc_submit(std::move(ctx
));
14336 ++scrubber
.num_digest_updates_pending
;
14339 dout(10) << __func__
<< " (" << mode
<< ") finish" << dendl
;
14342 void PrimaryLogPG::_scrub_clear_state()
14344 scrub_cstat
= object_stat_collection_t();
14347 void PrimaryLogPG::_scrub_finish()
14349 bool repair
= state_test(PG_STATE_REPAIR
);
14350 bool deep_scrub
= state_test(PG_STATE_DEEP_SCRUB
);
14351 const char *mode
= (repair
? "repair": (deep_scrub
? "deep-scrub" : "scrub"));
14353 if (info
.stats
.stats_invalid
) {
14354 info
.stats
.stats
= scrub_cstat
;
14355 info
.stats
.stats_invalid
= false;
14358 agent_choose_mode();
14361 dout(10) << mode
<< " got "
14362 << scrub_cstat
.sum
.num_objects
<< "/" << info
.stats
.stats
.sum
.num_objects
<< " objects, "
14363 << scrub_cstat
.sum
.num_object_clones
<< "/" << info
.stats
.stats
.sum
.num_object_clones
<< " clones, "
14364 << scrub_cstat
.sum
.num_objects_dirty
<< "/" << info
.stats
.stats
.sum
.num_objects_dirty
<< " dirty, "
14365 << scrub_cstat
.sum
.num_objects_omap
<< "/" << info
.stats
.stats
.sum
.num_objects_omap
<< " omap, "
14366 << scrub_cstat
.sum
.num_objects_pinned
<< "/" << info
.stats
.stats
.sum
.num_objects_pinned
<< " pinned, "
14367 << scrub_cstat
.sum
.num_objects_hit_set_archive
<< "/" << info
.stats
.stats
.sum
.num_objects_hit_set_archive
<< " hit_set_archive, "
14368 << scrub_cstat
.sum
.num_bytes
<< "/" << info
.stats
.stats
.sum
.num_bytes
<< " bytes, "
14369 << scrub_cstat
.sum
.num_bytes_hit_set_archive
<< "/" << info
.stats
.stats
.sum
.num_bytes_hit_set_archive
<< " hit_set_archive bytes."
14372 if (scrub_cstat
.sum
.num_objects
!= info
.stats
.stats
.sum
.num_objects
||
14373 scrub_cstat
.sum
.num_object_clones
!= info
.stats
.stats
.sum
.num_object_clones
||
14374 (scrub_cstat
.sum
.num_objects_dirty
!= info
.stats
.stats
.sum
.num_objects_dirty
&&
14375 !info
.stats
.dirty_stats_invalid
) ||
14376 (scrub_cstat
.sum
.num_objects_omap
!= info
.stats
.stats
.sum
.num_objects_omap
&&
14377 !info
.stats
.omap_stats_invalid
) ||
14378 (scrub_cstat
.sum
.num_objects_pinned
!= info
.stats
.stats
.sum
.num_objects_pinned
&&
14379 !info
.stats
.pin_stats_invalid
) ||
14380 (scrub_cstat
.sum
.num_objects_hit_set_archive
!= info
.stats
.stats
.sum
.num_objects_hit_set_archive
&&
14381 !info
.stats
.hitset_stats_invalid
) ||
14382 (scrub_cstat
.sum
.num_bytes_hit_set_archive
!= info
.stats
.stats
.sum
.num_bytes_hit_set_archive
&&
14383 !info
.stats
.hitset_bytes_stats_invalid
) ||
14384 scrub_cstat
.sum
.num_whiteouts
!= info
.stats
.stats
.sum
.num_whiteouts
||
14385 scrub_cstat
.sum
.num_bytes
!= info
.stats
.stats
.sum
.num_bytes
) {
14386 osd
->clog
->error() << info
.pgid
<< " " << mode
14387 << " stat mismatch, got "
14388 << scrub_cstat
.sum
.num_objects
<< "/" << info
.stats
.stats
.sum
.num_objects
<< " objects, "
14389 << scrub_cstat
.sum
.num_object_clones
<< "/" << info
.stats
.stats
.sum
.num_object_clones
<< " clones, "
14390 << scrub_cstat
.sum
.num_objects_dirty
<< "/" << info
.stats
.stats
.sum
.num_objects_dirty
<< " dirty, "
14391 << scrub_cstat
.sum
.num_objects_omap
<< "/" << info
.stats
.stats
.sum
.num_objects_omap
<< " omap, "
14392 << scrub_cstat
.sum
.num_objects_pinned
<< "/" << info
.stats
.stats
.sum
.num_objects_pinned
<< " pinned, "
14393 << scrub_cstat
.sum
.num_objects_hit_set_archive
<< "/" << info
.stats
.stats
.sum
.num_objects_hit_set_archive
<< " hit_set_archive, "
14394 << scrub_cstat
.sum
.num_whiteouts
<< "/" << info
.stats
.stats
.sum
.num_whiteouts
<< " whiteouts, "
14395 << scrub_cstat
.sum
.num_bytes
<< "/" << info
.stats
.stats
.sum
.num_bytes
<< " bytes, "
14396 << scrub_cstat
.sum
.num_bytes_hit_set_archive
<< "/" << info
.stats
.stats
.sum
.num_bytes_hit_set_archive
<< " hit_set_archive bytes.";
14397 ++scrubber
.shallow_errors
;
14401 info
.stats
.stats
= scrub_cstat
;
14402 info
.stats
.dirty_stats_invalid
= false;
14403 info
.stats
.omap_stats_invalid
= false;
14404 info
.stats
.hitset_stats_invalid
= false;
14405 info
.stats
.hitset_bytes_stats_invalid
= false;
14406 publish_stats_to_osd();
14409 } else if (scrub_cstat
.sum
.num_legacy_snapsets
!=
14410 info
.stats
.stats
.sum
.num_legacy_snapsets
) {
14411 osd
->clog
->info() << info
.pgid
<< " " << mode
<< " updated num_legacy_snapsets"
14412 << " from " << info
.stats
.stats
.sum
.num_legacy_snapsets
14413 << " -> " << scrub_cstat
.sum
.num_legacy_snapsets
<< "\n";
14414 info
.stats
.stats
.sum
.num_legacy_snapsets
= scrub_cstat
.sum
.num_legacy_snapsets
;
14415 publish_stats_to_osd();
14418 // Clear object context cache to get repair information
14420 object_contexts
.clear();
14423 bool PrimaryLogPG::check_osdmap_full(const set
<pg_shard_t
> &missing_on
)
14425 return osd
->check_osdmap_full(missing_on
);
14428 int PrimaryLogPG::rep_repair_primary_object(const hobject_t
& soid
, OpRequestRef op
)
14430 // Only supports replicated pools
14431 assert(!pool
.info
.require_rollback());
14432 assert(is_primary());
14434 dout(10) << __func__
<< " " << soid
14435 << " peers osd.{" << actingbackfill
<< "}" << dendl
;
14438 block_for_clean(soid
, op
);
14442 assert(!pg_log
.get_missing().is_missing(soid
));
14446 int r
= get_pgbackend()->objects_get_attr(soid
, OI_ATTR
, &bv
);
14448 // Leave v and try to repair without a version, getting attr failed
14449 dout(0) << __func__
<< ": Need version of replica, objects_get_attr failed: "
14450 << soid
<< " error=" << r
<< dendl
;
14452 bufferlist::iterator bliter
= bv
.begin();
14453 ::decode(oi
, bliter
);
14456 // Leave v as default constructed. This will fail when sent to older OSDs, but
14457 // not much worse than failing here.
14458 dout(0) << __func__
<< ": Need version of replica, bad object_info_t: " << soid
<< dendl
;
14461 missing_loc
.add_missing(soid
, v
, eversion_t());
14462 if (primary_error(soid
, v
)) {
14463 dout(0) << __func__
<< " No other replicas available for " << soid
<< dendl
;
14464 // XXX: If we knew that there is no down osd which could include this
14465 // object, it would be nice if we could return EIO here.
14466 // If a "never fail" flag was available, that could be used
14467 // for rbd to NOT return EIO until object marked lost.
14469 // Drop through to save this op in case an osd comes up with the object.
14472 // Restart the op after object becomes readable again
14473 waiting_for_unreadable_object
[soid
].push_back(op
);
14474 op
->mark_delayed("waiting for missing object");
14476 if (!eio_errors_to_process
) {
14477 eio_errors_to_process
= true;
14478 assert(is_clean());
14479 queue_peering_event(
14481 std::make_shared
<CephPeeringEvt
>(
14482 get_osdmap()->get_epoch(),
14483 get_osdmap()->get_epoch(),
14486 // A prior error must have already cleared clean state and queued recovery
14487 // or a map change has triggered re-peering.
14488 // Not inlining the recovery by calling maybe_kick_recovery(soid);
14489 dout(5) << __func__
<< ": Read error on " << soid
<< ", but already seen errors" << dendl
;
14495 /*---SnapTrimmer Logging---*/
14497 #define dout_prefix *_dout << pg->gen_prefix()
14499 void PrimaryLogPG::SnapTrimmer::log_enter(const char *state_name
)
14501 ldout(pg
->cct
, 20) << "enter " << state_name
<< dendl
;
14504 void PrimaryLogPG::SnapTrimmer::log_exit(const char *state_name
, utime_t enter_time
)
14506 ldout(pg
->cct
, 20) << "exit " << state_name
<< dendl
;
14509 /*---SnapTrimmer states---*/
14511 #define dout_prefix (*_dout << context< SnapTrimmer >().pg->gen_prefix() \
14512 << "SnapTrimmer state<" << get_state_name() << ">: ")
14515 PrimaryLogPG::NotTrimming::NotTrimming(my_context ctx
)
14517 NamedState(context
< SnapTrimmer
>().pg
, "NotTrimming")
14519 context
< SnapTrimmer
>().log_enter(state_name
);
14522 void PrimaryLogPG::NotTrimming::exit()
14524 context
< SnapTrimmer
>().log_exit(state_name
, enter_time
);
14527 boost::statechart::result
PrimaryLogPG::NotTrimming::react(const KickTrim
&)
14529 PrimaryLogPG
*pg
= context
< SnapTrimmer
>().pg
;
14530 ldout(pg
->cct
, 10) << "NotTrimming react KickTrim" << dendl
;
14532 if (!(pg
->is_primary() && pg
->is_active())) {
14533 ldout(pg
->cct
, 10) << "NotTrimming not primary or active" << dendl
;
14534 return discard_event();
14536 if (!pg
->is_clean() ||
14537 pg
->snap_trimq
.empty()) {
14538 ldout(pg
->cct
, 10) << "NotTrimming not clean or nothing to trim" << dendl
;
14539 return discard_event();
14541 if (pg
->scrubber
.active
) {
14542 ldout(pg
->cct
, 10) << " scrubbing, will requeue snap_trimmer after" << dendl
;
14543 return transit
< WaitScrub
>();
14545 return transit
< Trimming
>();
14549 boost::statechart::result
PrimaryLogPG::WaitReservation::react(const SnapTrimReserved
&)
14551 PrimaryLogPG
*pg
= context
< SnapTrimmer
>().pg
;
14552 ldout(pg
->cct
, 10) << "WaitReservation react SnapTrimReserved" << dendl
;
14555 if (!context
< SnapTrimmer
>().can_trim()) {
14556 post_event(KickTrim());
14557 return transit
< NotTrimming
>();
14560 context
<Trimming
>().snap_to_trim
= pg
->snap_trimq
.range_start();
14561 ldout(pg
->cct
, 10) << "NotTrimming: trimming "
14562 << pg
->snap_trimq
.range_start()
14564 return transit
< AwaitAsyncWork
>();
14567 /* AwaitAsyncWork */
14568 PrimaryLogPG::AwaitAsyncWork::AwaitAsyncWork(my_context ctx
)
14570 NamedState(context
< SnapTrimmer
>().pg
, "Trimming/AwaitAsyncWork")
14572 auto *pg
= context
< SnapTrimmer
>().pg
;
14573 context
< SnapTrimmer
>().log_enter(state_name
);
14574 context
< SnapTrimmer
>().pg
->osd
->queue_for_snap_trim(pg
);
14575 pg
->state_set(PG_STATE_SNAPTRIM
);
14576 pg
->state_clear(PG_STATE_SNAPTRIM_ERROR
);
14577 pg
->publish_stats_to_osd();
14580 boost::statechart::result
PrimaryLogPG::AwaitAsyncWork::react(const DoSnapWork
&)
14582 PrimaryLogPGRef pg
= context
< SnapTrimmer
>().pg
;
14583 snapid_t snap_to_trim
= context
<Trimming
>().snap_to_trim
;
14584 auto &in_flight
= context
<Trimming
>().in_flight
;
14585 assert(in_flight
.empty());
14587 assert(pg
->is_primary() && pg
->is_active());
14588 if (!context
< SnapTrimmer
>().can_trim()) {
14589 ldout(pg
->cct
, 10) << "something changed, reverting to NotTrimming" << dendl
;
14590 post_event(KickTrim());
14591 return transit
< NotTrimming
>();
14594 ldout(pg
->cct
, 10) << "AwaitAsyncWork: trimming snap " << snap_to_trim
<< dendl
;
14596 vector
<hobject_t
> to_trim
;
14597 unsigned max
= pg
->cct
->_conf
->osd_pg_max_concurrent_snap_trims
;
14598 to_trim
.reserve(max
);
14599 int r
= pg
->snap_mapper
.get_next_objects_to_trim(
14603 if (r
!= 0 && r
!= -ENOENT
) {
14604 lderr(pg
->cct
) << "get_next_objects_to_trim returned "
14605 << cpp_strerror(r
) << dendl
;
14606 assert(0 == "get_next_objects_to_trim returned an invalid code");
14607 } else if (r
== -ENOENT
) {
14609 ldout(pg
->cct
, 10) << "got ENOENT" << dendl
;
14611 ldout(pg
->cct
, 10) << "adding snap " << snap_to_trim
14612 << " to purged_snaps"
14614 pg
->info
.purged_snaps
.insert(snap_to_trim
);
14615 pg
->snap_trimq
.erase(snap_to_trim
);
14616 ldout(pg
->cct
, 10) << "purged_snaps now "
14617 << pg
->info
.purged_snaps
<< ", snap_trimq now "
14618 << pg
->snap_trimq
<< dendl
;
14620 ObjectStore::Transaction t
;
14621 pg
->dirty_big_info
= true;
14622 pg
->write_if_dirty(t
);
14623 int tr
= pg
->osd
->store
->queue_transaction(pg
->osr
.get(), std::move(t
), NULL
);
14626 pg
->share_pg_info();
14627 post_event(KickTrim());
14628 return transit
< NotTrimming
>();
14630 assert(!to_trim
.empty());
14632 for (auto &&object
: to_trim
) {
14634 ldout(pg
->cct
, 10) << "AwaitAsyncWork react trimming " << object
<< dendl
;
14636 int error
= pg
->trim_object(in_flight
.empty(), object
, &ctx
);
14638 if (error
== -ENOLCK
) {
14639 ldout(pg
->cct
, 10) << "could not get write lock on obj "
14640 << object
<< dendl
;
14642 pg
->state_set(PG_STATE_SNAPTRIM_ERROR
);
14643 ldout(pg
->cct
, 10) << "Snaptrim error=" << error
<< dendl
;
14645 if (!in_flight
.empty()) {
14646 ldout(pg
->cct
, 10) << "letting the ones we already started finish" << dendl
;
14647 return transit
< WaitRepops
>();
14649 if (error
== -ENOLCK
) {
14650 ldout(pg
->cct
, 10) << "waiting for it to clear"
14652 return transit
< WaitRWLock
>();
14654 return transit
< NotTrimming
>();
14658 in_flight
.insert(object
);
14659 ctx
->register_on_success(
14660 [pg
, object
, &in_flight
]() {
14661 assert(in_flight
.find(object
) != in_flight
.end());
14662 in_flight
.erase(object
);
14663 if (in_flight
.empty()) {
14664 if (pg
->state_test(PG_STATE_SNAPTRIM_ERROR
)) {
14665 pg
->snap_trimmer_machine
.process_event(Reset());
14667 pg
->snap_trimmer_machine
.process_event(RepopsComplete());
14672 pg
->simple_opc_submit(std::move(ctx
));
14675 return transit
< WaitRepops
>();
14678 void PrimaryLogPG::setattr_maybe_cache(
14679 ObjectContextRef obc
,
14685 t
->setattr(obc
->obs
.oi
.soid
, key
, val
);
14688 void PrimaryLogPG::setattrs_maybe_cache(
14689 ObjectContextRef obc
,
14692 map
<string
, bufferlist
> &attrs
)
14694 t
->setattrs(obc
->obs
.oi
.soid
, attrs
);
14697 void PrimaryLogPG::rmattr_maybe_cache(
14698 ObjectContextRef obc
,
14703 t
->rmattr(obc
->obs
.oi
.soid
, key
);
14706 int PrimaryLogPG::getattr_maybe_cache(
14707 ObjectContextRef obc
,
14711 if (pool
.info
.require_rollback()) {
14712 map
<string
, bufferlist
>::iterator i
= obc
->attr_cache
.find(key
);
14713 if (i
!= obc
->attr_cache
.end()) {
14721 return pgbackend
->objects_get_attr(obc
->obs
.oi
.soid
, key
, val
);
14724 int PrimaryLogPG::getattrs_maybe_cache(
14725 ObjectContextRef obc
,
14726 map
<string
, bufferlist
> *out
)
14730 if (pool
.info
.require_rollback()) {
14731 *out
= obc
->attr_cache
;
14733 r
= pgbackend
->objects_get_attrs(obc
->obs
.oi
.soid
, out
);
14735 map
<string
, bufferlist
> tmp
;
14736 for (map
<string
, bufferlist
>::iterator i
= out
->begin();
14739 if (i
->first
.size() > 1 && i
->first
[0] == '_')
14740 tmp
[i
->first
.substr(1, i
->first
.size())].claim(i
->second
);
14746 bool PrimaryLogPG::check_failsafe_full(ostream
&ss
) {
14747 return osd
->check_failsafe_full(ss
);
14750 void intrusive_ptr_add_ref(PrimaryLogPG
*pg
) { pg
->get("intptr"); }
14751 void intrusive_ptr_release(PrimaryLogPG
*pg
) { pg
->put("intptr"); }
14753 #ifdef PG_DEBUG_REFS
14754 uint64_t get_with_id(PrimaryLogPG
*pg
) { return pg
->get_with_id(); }
14755 void put_with_id(PrimaryLogPG
*pg
, uint64_t id
) { return pg
->put_with_id(id
); }
14758 void intrusive_ptr_add_ref(PrimaryLogPG::RepGather
*repop
) { repop
->get(); }
14759 void intrusive_ptr_release(PrimaryLogPG::RepGather
*repop
) { repop
->put(); }