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) 2017 OVH
9 * This is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License version 2.1, as published by the Free Software
12 * Foundation. See file COPYING.
23 #include <boost/scoped_ptr.hpp>
25 #ifdef HAVE_SYS_PARAM_H
26 #include <sys/param.h>
29 #ifdef HAVE_SYS_MOUNT_H
30 #include <sys/mount.h>
35 #include "include/types.h"
36 #include "include/compat.h"
41 #include "osdc/Objecter.h"
43 #include "common/errno.h"
44 #include "common/ceph_argparse.h"
45 #include "common/ceph_time.h"
46 #include "common/version.h"
47 #include "common/io_priority.h"
49 #include "os/ObjectStore.h"
51 #include "os/FuseStore.h"
54 #include "PrimaryLogPG.h"
57 #include "msg/Messenger.h"
58 #include "msg/Message.h"
60 #include "mon/MonClient.h"
62 #include "messages/MLog.h"
64 #include "messages/MGenericMessage.h"
65 #include "messages/MOSDPing.h"
66 #include "messages/MOSDFailure.h"
67 #include "messages/MOSDMarkMeDown.h"
68 #include "messages/MOSDFull.h"
69 #include "messages/MOSDOp.h"
70 #include "messages/MOSDOpReply.h"
71 #include "messages/MOSDBackoff.h"
72 #include "messages/MOSDBeacon.h"
73 #include "messages/MOSDRepOp.h"
74 #include "messages/MOSDRepOpReply.h"
75 #include "messages/MOSDBoot.h"
76 #include "messages/MOSDPGTemp.h"
78 #include "messages/MOSDMap.h"
79 #include "messages/MMonGetOSDMap.h"
80 #include "messages/MOSDPGNotify.h"
81 #include "messages/MOSDPGQuery.h"
82 #include "messages/MOSDPGLog.h"
83 #include "messages/MOSDPGRemove.h"
84 #include "messages/MOSDPGInfo.h"
85 #include "messages/MOSDPGCreate.h"
86 #include "messages/MOSDPGTrim.h"
87 #include "messages/MOSDPGScan.h"
88 #include "messages/MOSDPGBackfill.h"
89 #include "messages/MBackfillReserve.h"
90 #include "messages/MRecoveryReserve.h"
91 #include "messages/MOSDForceRecovery.h"
92 #include "messages/MOSDECSubOpWrite.h"
93 #include "messages/MOSDECSubOpWriteReply.h"
94 #include "messages/MOSDECSubOpRead.h"
95 #include "messages/MOSDECSubOpReadReply.h"
96 #include "messages/MOSDPGCreated.h"
97 #include "messages/MOSDPGUpdateLogMissing.h"
98 #include "messages/MOSDPGUpdateLogMissingReply.h"
100 #include "messages/MOSDAlive.h"
102 #include "messages/MOSDScrub.h"
103 #include "messages/MOSDScrubReserve.h"
104 #include "messages/MOSDRepScrub.h"
106 #include "messages/MMonCommand.h"
107 #include "messages/MCommand.h"
108 #include "messages/MCommandReply.h"
110 #include "messages/MPGStats.h"
111 #include "messages/MPGStatsAck.h"
113 #include "messages/MWatchNotify.h"
114 #include "messages/MOSDPGPush.h"
115 #include "messages/MOSDPGPushReply.h"
116 #include "messages/MOSDPGPull.h"
118 #include "common/perf_counters.h"
119 #include "common/Timer.h"
120 #include "common/LogClient.h"
121 #include "common/AsyncReserver.h"
122 #include "common/HeartbeatMap.h"
123 #include "common/admin_socket.h"
124 #include "common/ceph_context.h"
126 #include "global/signal_handler.h"
127 #include "global/pidfile.h"
129 #include "include/color.h"
130 #include "perfglue/cpu_profiler.h"
131 #include "perfglue/heap_profiler.h"
133 #include "osd/OpRequest.h"
135 #include "auth/AuthAuthorizeHandler.h"
136 #include "auth/RotatingKeyRing.h"
137 #include "common/errno.h"
139 #include "objclass/objclass.h"
141 #include "common/cmdparse.h"
142 #include "include/str_list.h"
143 #include "include/util.h"
145 #include "include/assert.h"
146 #include "common/config.h"
147 #include "common/EventTrace.h"
150 #define TRACEPOINT_DEFINE
151 #define TRACEPOINT_PROBE_DYNAMIC_LINKAGE
152 #include "tracing/osd.h"
153 #undef TRACEPOINT_PROBE_DYNAMIC_LINKAGE
154 #undef TRACEPOINT_DEFINE
156 #define tracepoint(...)
159 #define dout_context cct
160 #define dout_subsys ceph_subsys_osd
162 #define dout_prefix _prefix(_dout, whoami, get_osdmap_epoch())
165 const double OSD::OSD_TICK_INTERVAL
= 1.0;
167 static ostream
& _prefix(std::ostream
* _dout
, int whoami
, epoch_t epoch
) {
168 return *_dout
<< "osd." << whoami
<< " " << epoch
<< " ";
171 //Initial features in new superblock.
172 //Features here are also automatically upgraded
173 CompatSet
OSD::get_osd_initial_compat_set() {
174 CompatSet::FeatureSet ceph_osd_feature_compat
;
175 CompatSet::FeatureSet ceph_osd_feature_ro_compat
;
176 CompatSet::FeatureSet ceph_osd_feature_incompat
;
177 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE
);
178 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_PGINFO
);
179 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_OLOC
);
180 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_LEC
);
181 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_CATEGORIES
);
182 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_HOBJECTPOOL
);
183 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BIGINFO
);
184 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_LEVELDBINFO
);
185 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_LEVELDBLOG
);
186 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SNAPMAPPER
);
187 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_HINTS
);
188 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_PGMETA
);
189 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_MISSING
);
190 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_FASTINFO
);
191 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_RECOVERY_DELETES
);
192 return CompatSet(ceph_osd_feature_compat
, ceph_osd_feature_ro_compat
,
193 ceph_osd_feature_incompat
);
196 //Features are added here that this OSD supports.
197 CompatSet
OSD::get_osd_compat_set() {
198 CompatSet compat
= get_osd_initial_compat_set();
199 //Any features here can be set in code, but not in initial superblock
200 compat
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
);
204 OSDService::OSDService(OSD
*osd
) :
207 meta_osr(new ObjectStore::Sequencer("meta")),
208 whoami(osd
->whoami
), store(osd
->store
),
209 log_client(osd
->log_client
), clog(osd
->clog
),
210 pg_recovery_stats(osd
->pg_recovery_stats
),
211 cluster_messenger(osd
->cluster_messenger
),
212 client_messenger(osd
->client_messenger
),
214 recoverystate_perf(osd
->recoverystate_perf
),
216 peering_wq(osd
->peering_wq
),
217 recovery_gen_wq("recovery_gen_wq", cct
->_conf
->osd_recovery_thread_timeout
,
219 class_handler(osd
->class_handler
),
220 pg_epoch_lock("OSDService::pg_epoch_lock"),
221 publish_lock("OSDService::publish_lock"),
222 pre_publish_lock("OSDService::pre_publish_lock"),
224 peer_map_epoch_lock("OSDService::peer_map_epoch_lock"),
225 sched_scrub_lock("OSDService::sched_scrub_lock"), scrubs_pending(0),
227 agent_lock("OSDService::agent_lock"),
228 agent_valid_iterator(false),
230 flush_mode_high_count(0),
233 agent_stop_flag(false),
234 agent_timer_lock("OSDService::agent_timer_lock"),
235 agent_timer(osd
->client_messenger
->cct
, agent_timer_lock
),
236 last_recalibrate(ceph_clock_now()),
237 promote_max_objects(0),
238 promote_max_bytes(0),
239 objecter(new Objecter(osd
->client_messenger
->cct
, osd
->objecter_messenger
, osd
->monc
, NULL
, 0, 0)),
240 objecter_finisher(osd
->client_messenger
->cct
),
241 watch_lock("OSDService::watch_lock"),
242 watch_timer(osd
->client_messenger
->cct
, watch_lock
),
244 recovery_request_lock("OSDService::recovery_request_lock"),
245 recovery_request_timer(cct
, recovery_request_lock
, false),
246 recovery_sleep_lock("OSDService::recovery_sleep_lock"),
247 recovery_sleep_timer(cct
, recovery_sleep_lock
, false),
248 reserver_finisher(cct
),
249 local_reserver(&reserver_finisher
, cct
->_conf
->osd_max_backfills
,
250 cct
->_conf
->osd_min_recovery_priority
),
251 remote_reserver(&reserver_finisher
, cct
->_conf
->osd_max_backfills
,
252 cct
->_conf
->osd_min_recovery_priority
),
253 pg_temp_lock("OSDService::pg_temp_lock"),
254 snap_sleep_lock("OSDService::snap_sleep_lock"),
256 osd
->client_messenger
->cct
, snap_sleep_lock
, false /* relax locking */),
257 scrub_sleep_lock("OSDService::scrub_sleep_lock"),
259 osd
->client_messenger
->cct
, scrub_sleep_lock
, false /* relax locking */),
260 snap_reserver(&reserver_finisher
,
261 cct
->_conf
->osd_max_trimming_pgs
),
262 recovery_lock("OSDService::recovery_lock"),
263 recovery_ops_active(0),
264 recovery_ops_reserved(0),
265 recovery_paused(false),
266 map_cache_lock("OSDService::map_cache_lock"),
267 map_cache(cct
, cct
->_conf
->osd_map_cache_size
),
268 map_bl_cache(cct
->_conf
->osd_map_cache_size
),
269 map_bl_inc_cache(cct
->_conf
->osd_map_cache_size
),
270 in_progress_split_lock("OSDService::in_progress_split_lock"),
271 stat_lock("OSDService::stat_lock"),
272 full_status_lock("OSDService::full_status_lock"),
275 epoch_lock("OSDService::epoch_lock"),
276 boot_epoch(0), up_epoch(0), bind_epoch(0),
277 is_stopping_lock("OSDService::is_stopping_lock")
279 , pgid_lock("OSDService::pgid_lock")
285 OSDService::~OSDService()
293 void OSDService::add_pgid(spg_t pgid
, PG
*pg
){
294 Mutex::Locker
l(pgid_lock
);
295 if (!pgid_tracker
.count(pgid
)) {
298 pgid_tracker
[pgid
]++;
300 void OSDService::remove_pgid(spg_t pgid
, PG
*pg
)
302 Mutex::Locker
l(pgid_lock
);
303 assert(pgid_tracker
.count(pgid
));
304 assert(pgid_tracker
[pgid
] > 0);
305 pgid_tracker
[pgid
]--;
306 if (pgid_tracker
[pgid
] == 0) {
307 pgid_tracker
.erase(pgid
);
308 live_pgs
.erase(pgid
);
311 void OSDService::dump_live_pgids()
313 Mutex::Locker
l(pgid_lock
);
314 derr
<< "live pgids:" << dendl
;
315 for (map
<spg_t
, int>::const_iterator i
= pgid_tracker
.cbegin();
316 i
!= pgid_tracker
.cend();
318 derr
<< "\t" << *i
<< dendl
;
319 live_pgs
[i
->first
]->dump_live_ids();
325 void OSDService::_start_split(spg_t parent
, const set
<spg_t
> &children
)
327 for (set
<spg_t
>::const_iterator i
= children
.begin();
330 dout(10) << __func__
<< ": Starting split on pg " << *i
331 << ", parent=" << parent
<< dendl
;
332 assert(!pending_splits
.count(*i
));
333 assert(!in_progress_splits
.count(*i
));
334 pending_splits
.insert(make_pair(*i
, parent
));
336 assert(!rev_pending_splits
[parent
].count(*i
));
337 rev_pending_splits
[parent
].insert(*i
);
341 void OSDService::mark_split_in_progress(spg_t parent
, const set
<spg_t
> &children
)
343 Mutex::Locker
l(in_progress_split_lock
);
344 map
<spg_t
, set
<spg_t
> >::iterator piter
= rev_pending_splits
.find(parent
);
345 assert(piter
!= rev_pending_splits
.end());
346 for (set
<spg_t
>::const_iterator i
= children
.begin();
349 assert(piter
->second
.count(*i
));
350 assert(pending_splits
.count(*i
));
351 assert(!in_progress_splits
.count(*i
));
352 assert(pending_splits
[*i
] == parent
);
354 pending_splits
.erase(*i
);
355 piter
->second
.erase(*i
);
356 in_progress_splits
.insert(*i
);
358 if (piter
->second
.empty())
359 rev_pending_splits
.erase(piter
);
362 void OSDService::cancel_pending_splits_for_parent(spg_t parent
)
364 Mutex::Locker
l(in_progress_split_lock
);
365 _cancel_pending_splits_for_parent(parent
);
368 void OSDService::_cancel_pending_splits_for_parent(spg_t parent
)
370 map
<spg_t
, set
<spg_t
> >::iterator piter
= rev_pending_splits
.find(parent
);
371 if (piter
== rev_pending_splits
.end())
374 for (set
<spg_t
>::iterator i
= piter
->second
.begin();
375 i
!= piter
->second
.end();
377 assert(pending_splits
.count(*i
));
378 assert(!in_progress_splits
.count(*i
));
379 pending_splits
.erase(*i
);
380 dout(10) << __func__
<< ": Completing split on pg " << *i
381 << " for parent: " << parent
<< dendl
;
382 _cancel_pending_splits_for_parent(*i
);
384 rev_pending_splits
.erase(piter
);
387 void OSDService::_maybe_split_pgid(OSDMapRef old_map
,
391 assert(old_map
->have_pg_pool(pgid
.pool()));
392 int old_pgnum
= old_map
->get_pg_num(pgid
.pool());
393 if (pgid
.ps() < static_cast<unsigned>(old_pgnum
)) {
395 if (pgid
.is_split(old_pgnum
,
396 new_map
->get_pg_num(pgid
.pool()), &children
)) {
397 _start_split(pgid
, children
); }
399 assert(pgid
.ps() < static_cast<unsigned>(new_map
->get_pg_num(pgid
.pool())));
403 void OSDService::init_splits_between(spg_t pgid
,
407 // First, check whether we can avoid this potentially expensive check
408 if (tomap
->have_pg_pool(pgid
.pool()) &&
410 frommap
->get_pg_num(pgid
.pool()),
411 tomap
->get_pg_num(pgid
.pool()),
413 // Ok, a split happened, so we need to walk the osdmaps
414 set
<spg_t
> new_pgs
; // pgs to scan on each map
415 new_pgs
.insert(pgid
);
416 OSDMapRef
curmap(get_map(frommap
->get_epoch()));
417 for (epoch_t e
= frommap
->get_epoch() + 1;
418 e
<= tomap
->get_epoch();
420 OSDMapRef
nextmap(try_get_map(e
));
423 set
<spg_t
> even_newer_pgs
; // pgs added in this loop
424 for (set
<spg_t
>::iterator i
= new_pgs
.begin(); i
!= new_pgs
.end(); ++i
) {
425 set
<spg_t
> split_pgs
;
426 if (i
->is_split(curmap
->get_pg_num(i
->pool()),
427 nextmap
->get_pg_num(i
->pool()),
429 start_split(*i
, split_pgs
);
430 even_newer_pgs
.insert(split_pgs
.begin(), split_pgs
.end());
433 new_pgs
.insert(even_newer_pgs
.begin(), even_newer_pgs
.end());
436 assert(curmap
== tomap
); // we must have had both frommap and tomap
440 void OSDService::expand_pg_num(OSDMapRef old_map
,
443 Mutex::Locker
l(in_progress_split_lock
);
444 for (set
<spg_t
>::iterator i
= in_progress_splits
.begin();
445 i
!= in_progress_splits
.end();
447 if (!new_map
->have_pg_pool(i
->pool())) {
448 in_progress_splits
.erase(i
++);
450 _maybe_split_pgid(old_map
, new_map
, *i
);
454 for (map
<spg_t
, spg_t
>::iterator i
= pending_splits
.begin();
455 i
!= pending_splits
.end();
457 if (!new_map
->have_pg_pool(i
->first
.pool())) {
458 rev_pending_splits
.erase(i
->second
);
459 pending_splits
.erase(i
++);
461 _maybe_split_pgid(old_map
, new_map
, i
->first
);
467 bool OSDService::splitting(spg_t pgid
)
469 Mutex::Locker
l(in_progress_split_lock
);
470 return in_progress_splits
.count(pgid
) ||
471 pending_splits
.count(pgid
);
474 void OSDService::complete_split(const set
<spg_t
> &pgs
)
476 Mutex::Locker
l(in_progress_split_lock
);
477 for (set
<spg_t
>::const_iterator i
= pgs
.begin();
480 dout(10) << __func__
<< ": Completing split on pg " << *i
<< dendl
;
481 assert(!pending_splits
.count(*i
));
482 assert(in_progress_splits
.count(*i
));
483 in_progress_splits
.erase(*i
);
487 void OSDService::need_heartbeat_peer_update()
489 osd
->need_heartbeat_peer_update();
492 void OSDService::pg_stat_queue_enqueue(PG
*pg
)
494 osd
->pg_stat_queue_enqueue(pg
);
497 void OSDService::pg_stat_queue_dequeue(PG
*pg
)
499 osd
->pg_stat_queue_dequeue(pg
);
502 void OSDService::start_shutdown()
505 Mutex::Locker
l(agent_timer_lock
);
506 agent_timer
.shutdown();
510 Mutex::Locker
l(recovery_sleep_lock
);
511 recovery_sleep_timer
.shutdown();
515 void OSDService::shutdown_reserver()
517 reserver_finisher
.wait_for_empty();
518 reserver_finisher
.stop();
521 void OSDService::shutdown()
524 Mutex::Locker
l(watch_lock
);
525 watch_timer
.shutdown();
528 objecter
->shutdown();
529 objecter_finisher
.wait_for_empty();
530 objecter_finisher
.stop();
533 Mutex::Locker
l(recovery_request_lock
);
534 recovery_request_timer
.shutdown();
538 Mutex::Locker
l(snap_sleep_lock
);
539 snap_sleep_timer
.shutdown();
543 Mutex::Locker
l(scrub_sleep_lock
);
544 scrub_sleep_timer
.shutdown();
547 osdmap
= OSDMapRef();
548 next_osdmap
= OSDMapRef();
551 void OSDService::init()
553 reserver_finisher
.start();
554 objecter_finisher
.start();
555 objecter
->set_client_incarnation(0);
557 // deprioritize objecter in daemonperf output
558 objecter
->get_logger()->set_prio_adjust(-3);
562 snap_sleep_timer
.init();
563 scrub_sleep_timer
.init();
565 agent_thread
.create("osd_srv_agent");
567 if (cct
->_conf
->osd_recovery_delay_start
)
568 defer_recovery(cct
->_conf
->osd_recovery_delay_start
);
571 void OSDService::final_init()
573 objecter
->start(osdmap
.get());
576 void OSDService::activate_map()
578 // wake/unwake the tiering agent
581 !osdmap
->test_flag(CEPH_OSDMAP_NOTIERAGENT
) &&
587 class AgentTimeoutCB
: public Context
{
590 explicit AgentTimeoutCB(PGRef _pg
) : pg(_pg
) {}
591 void finish(int) override
{
592 pg
->agent_choose_mode_restart();
596 void OSDService::agent_entry()
598 dout(10) << __func__
<< " start" << dendl
;
601 while (!agent_stop_flag
) {
602 if (agent_queue
.empty()) {
603 dout(20) << __func__
<< " empty queue" << dendl
;
604 agent_cond
.Wait(agent_lock
);
607 uint64_t level
= agent_queue
.rbegin()->first
;
608 set
<PGRef
>& top
= agent_queue
.rbegin()->second
;
610 << " tiers " << agent_queue
.size()
611 << ", top is " << level
612 << " with pgs " << top
.size()
613 << ", ops " << agent_ops
<< "/"
614 << cct
->_conf
->osd_agent_max_ops
615 << (agent_active
? " active" : " NOT ACTIVE")
617 dout(20) << __func__
<< " oids " << agent_oids
<< dendl
;
618 int max
= cct
->_conf
->osd_agent_max_ops
- agent_ops
;
619 int agent_flush_quota
= max
;
620 if (!flush_mode_high_count
)
621 agent_flush_quota
= cct
->_conf
->osd_agent_max_low_ops
- agent_ops
;
622 if (agent_flush_quota
<= 0 || top
.empty() || !agent_active
) {
623 agent_cond
.Wait(agent_lock
);
627 if (!agent_valid_iterator
|| agent_queue_pos
== top
.end()) {
628 agent_queue_pos
= top
.begin();
629 agent_valid_iterator
= true;
631 PGRef pg
= *agent_queue_pos
;
632 dout(10) << "high_count " << flush_mode_high_count
633 << " agent_ops " << agent_ops
634 << " flush_quota " << agent_flush_quota
<< dendl
;
636 if (!pg
->agent_work(max
, agent_flush_quota
)) {
637 dout(10) << __func__
<< " " << pg
->get_pgid()
638 << " no agent_work, delay for " << cct
->_conf
->osd_agent_delay_time
639 << " seconds" << dendl
;
641 osd
->logger
->inc(l_osd_tier_delay
);
642 // Queue a timer to call agent_choose_mode for this pg in 5 seconds
643 agent_timer_lock
.Lock();
644 Context
*cb
= new AgentTimeoutCB(pg
);
645 agent_timer
.add_event_after(cct
->_conf
->osd_agent_delay_time
, cb
);
646 agent_timer_lock
.Unlock();
651 dout(10) << __func__
<< " finish" << dendl
;
654 void OSDService::agent_stop()
657 Mutex::Locker
l(agent_lock
);
659 // By this time all ops should be cancelled
660 assert(agent_ops
== 0);
661 // By this time all PGs are shutdown and dequeued
662 if (!agent_queue
.empty()) {
663 set
<PGRef
>& top
= agent_queue
.rbegin()->second
;
664 derr
<< "agent queue not empty, for example " << (*top
.begin())->info
.pgid
<< dendl
;
665 assert(0 == "agent queue not empty");
668 agent_stop_flag
= true;
674 // -------------------------------------
676 void OSDService::promote_throttle_recalibrate()
678 utime_t now
= ceph_clock_now();
679 double dur
= now
- last_recalibrate
;
680 last_recalibrate
= now
;
681 unsigned prob
= promote_probability_millis
;
683 uint64_t target_obj_sec
= cct
->_conf
->osd_tier_promote_max_objects_sec
;
684 uint64_t target_bytes_sec
= cct
->_conf
->osd_tier_promote_max_bytes_sec
;
686 unsigned min_prob
= 1;
688 uint64_t attempts
, obj
, bytes
;
689 promote_counter
.sample_and_attenuate(&attempts
, &obj
, &bytes
);
690 dout(10) << __func__
<< " " << attempts
<< " attempts, promoted "
691 << obj
<< " objects and " << pretty_si_t(bytes
) << " bytes; target "
692 << target_obj_sec
<< " obj/sec or "
693 << pretty_si_t(target_bytes_sec
) << " bytes/sec"
696 // calculate what the probability *should* be, given the targets
698 if (attempts
&& dur
> 0) {
699 uint64_t avg_size
= 1;
701 avg_size
= MAX(bytes
/ obj
, 1);
702 unsigned po
= (double)target_obj_sec
* dur
* 1000.0 / (double)attempts
;
703 unsigned pb
= (double)target_bytes_sec
/ (double)avg_size
* dur
* 1000.0
705 dout(20) << __func__
<< " po " << po
<< " pb " << pb
<< " avg_size "
706 << avg_size
<< dendl
;
707 if (target_obj_sec
&& target_bytes_sec
)
708 new_prob
= MIN(po
, pb
);
709 else if (target_obj_sec
)
711 else if (target_bytes_sec
)
718 dout(20) << __func__
<< " new_prob " << new_prob
<< dendl
;
720 // correct for persistent skew between target rate and actual rate, adjust
723 if (attempts
&& obj
) {
724 actual
= obj
* 1000 / attempts
;
725 ratio
= (double)actual
/ (double)prob
;
726 new_prob
= (double)new_prob
/ ratio
;
728 new_prob
= MAX(new_prob
, min_prob
);
729 new_prob
= MIN(new_prob
, 1000);
732 prob
= (prob
+ new_prob
) / 2;
733 prob
= MAX(prob
, min_prob
);
734 prob
= MIN(prob
, 1000);
735 dout(10) << __func__
<< " actual " << actual
736 << ", actual/prob ratio " << ratio
737 << ", adjusted new_prob " << new_prob
738 << ", prob " << promote_probability_millis
<< " -> " << prob
740 promote_probability_millis
= prob
;
742 // set hard limits for this interval to mitigate stampedes
743 promote_max_objects
= target_obj_sec
* OSD::OSD_TICK_INTERVAL
* 2;
744 promote_max_bytes
= target_bytes_sec
* OSD::OSD_TICK_INTERVAL
* 2;
747 // -------------------------------------
749 float OSDService::get_failsafe_full_ratio()
751 float full_ratio
= cct
->_conf
->osd_failsafe_full_ratio
;
752 if (full_ratio
> 1.0) full_ratio
/= 100.0;
756 void OSDService::check_full_status(float ratio
)
758 Mutex::Locker
l(full_status_lock
);
762 // The OSDMap ratios take precendence. So if the failsafe is .95 and
763 // the admin sets the cluster full to .96, the failsafe moves up to .96
764 // too. (Not that having failsafe == full is ideal, but it's better than
765 // dropping writes before the clusters appears full.)
766 OSDMapRef osdmap
= get_osdmap();
767 if (!osdmap
|| osdmap
->get_epoch() == 0) {
771 float nearfull_ratio
= osdmap
->get_nearfull_ratio();
772 float backfillfull_ratio
= std::max(osdmap
->get_backfillfull_ratio(), nearfull_ratio
);
773 float full_ratio
= std::max(osdmap
->get_full_ratio(), backfillfull_ratio
);
774 float failsafe_ratio
= std::max(get_failsafe_full_ratio(), full_ratio
);
776 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
777 // use the failsafe for nearfull and full; the mon isn't using the
778 // flags anyway because we're mid-upgrade.
779 full_ratio
= failsafe_ratio
;
780 backfillfull_ratio
= failsafe_ratio
;
781 nearfull_ratio
= failsafe_ratio
;
782 } else if (full_ratio
<= 0 ||
783 backfillfull_ratio
<= 0 ||
784 nearfull_ratio
<= 0) {
785 derr
<< __func__
<< " full_ratio, backfillfull_ratio or nearfull_ratio is <= 0" << dendl
;
786 // use failsafe flag. ick. the monitor did something wrong or the user
787 // did something stupid.
788 full_ratio
= failsafe_ratio
;
789 backfillfull_ratio
= failsafe_ratio
;
790 nearfull_ratio
= failsafe_ratio
;
795 if (injectfull_state
> NONE
&& injectfull
) {
796 new_state
= injectfull_state
;
797 inject
= "(Injected)";
798 } else if (ratio
> failsafe_ratio
) {
799 new_state
= FAILSAFE
;
800 } else if (ratio
> full_ratio
) {
802 } else if (ratio
> backfillfull_ratio
) {
803 new_state
= BACKFILLFULL
;
804 } else if (ratio
> nearfull_ratio
) {
805 new_state
= NEARFULL
;
809 dout(20) << __func__
<< " cur ratio " << ratio
810 << ". nearfull_ratio " << nearfull_ratio
811 << ". backfillfull_ratio " << backfillfull_ratio
812 << ", full_ratio " << full_ratio
813 << ", failsafe_ratio " << failsafe_ratio
814 << ", new state " << get_full_state_name(new_state
)
819 if (cur_state
!= new_state
) {
820 dout(10) << __func__
<< " " << get_full_state_name(cur_state
)
821 << " -> " << get_full_state_name(new_state
) << dendl
;
822 if (new_state
== FAILSAFE
) {
823 clog
->error() << "full status failsafe engaged, dropping updates, now "
824 << (int)roundf(ratio
* 100) << "% full";
825 } else if (cur_state
== FAILSAFE
) {
826 clog
->error() << "full status failsafe disengaged, no longer dropping "
827 << "updates, now " << (int)roundf(ratio
* 100) << "% full";
829 cur_state
= new_state
;
833 bool OSDService::need_fullness_update()
835 OSDMapRef osdmap
= get_osdmap();
837 if (osdmap
->exists(whoami
)) {
838 if (osdmap
->get_state(whoami
) & CEPH_OSD_FULL
) {
840 } else if (osdmap
->get_state(whoami
) & CEPH_OSD_BACKFILLFULL
) {
842 } else if (osdmap
->get_state(whoami
) & CEPH_OSD_NEARFULL
) {
849 else if (is_backfillfull())
851 else if (is_nearfull())
856 bool OSDService::_check_full(s_names type
, ostream
&ss
) const
858 Mutex::Locker
l(full_status_lock
);
860 if (injectfull
&& injectfull_state
>= type
) {
861 // injectfull is either a count of the number of times to return failsafe full
862 // or if -1 then always return full
865 ss
<< "Injected " << get_full_state_name(type
) << " OSD ("
866 << (injectfull
< 0 ? "set" : std::to_string(injectfull
)) << ")";
870 ss
<< "current usage is " << cur_ratio
;
871 return cur_state
>= type
;
874 bool OSDService::check_failsafe_full(ostream
&ss
) const
876 return _check_full(FAILSAFE
, ss
);
879 bool OSDService::check_full(ostream
&ss
) const
881 return _check_full(FULL
, ss
);
884 bool OSDService::check_backfill_full(ostream
&ss
) const
886 return _check_full(BACKFILLFULL
, ss
);
889 bool OSDService::check_nearfull(ostream
&ss
) const
891 return _check_full(NEARFULL
, ss
);
894 bool OSDService::is_failsafe_full() const
896 Mutex::Locker
l(full_status_lock
);
897 return cur_state
== FAILSAFE
;
900 bool OSDService::is_full() const
902 Mutex::Locker
l(full_status_lock
);
903 return cur_state
>= FULL
;
906 bool OSDService::is_backfillfull() const
908 Mutex::Locker
l(full_status_lock
);
909 return cur_state
>= BACKFILLFULL
;
912 bool OSDService::is_nearfull() const
914 Mutex::Locker
l(full_status_lock
);
915 return cur_state
>= NEARFULL
;
918 void OSDService::set_injectfull(s_names type
, int64_t count
)
920 Mutex::Locker
l(full_status_lock
);
921 injectfull_state
= type
;
925 osd_stat_t
OSDService::set_osd_stat(const struct store_statfs_t
&stbuf
,
926 vector
<int>& hb_peers
,
929 uint64_t bytes
= stbuf
.total
;
930 uint64_t used
= bytes
- stbuf
.available
;
931 uint64_t avail
= stbuf
.available
;
933 osd
->logger
->set(l_osd_stat_bytes
, bytes
);
934 osd
->logger
->set(l_osd_stat_bytes_used
, used
);
935 osd
->logger
->set(l_osd_stat_bytes_avail
, avail
);
938 Mutex::Locker
l(stat_lock
);
939 osd_stat
.hb_peers
.swap(hb_peers
);
940 osd
->op_tracker
.get_age_ms_histogram(&osd_stat
.op_queue_age_hist
);
941 osd_stat
.kb
= bytes
>> 10;
942 osd_stat
.kb_used
= used
>> 10;
943 osd_stat
.kb_avail
= avail
>> 10;
944 osd_stat
.num_pgs
= num_pgs
;
949 void OSDService::update_osd_stat(vector
<int>& hb_peers
)
951 // load osd stats first
952 struct store_statfs_t stbuf
;
953 int r
= osd
->store
->statfs(&stbuf
);
955 derr
<< "statfs() failed: " << cpp_strerror(r
) << dendl
;
959 auto new_stat
= set_osd_stat(stbuf
, hb_peers
, osd
->get_num_pgs());
960 dout(20) << "update_osd_stat " << new_stat
<< dendl
;
962 float ratio
= ((float)new_stat
.kb_used
) / ((float)new_stat
.kb
);
963 check_full_status(ratio
);
966 bool OSDService::check_osdmap_full(const set
<pg_shard_t
> &missing_on
)
968 OSDMapRef osdmap
= get_osdmap();
969 for (auto shard
: missing_on
) {
970 if (osdmap
->get_state(shard
.osd
) & CEPH_OSD_FULL
)
976 void OSDService::send_message_osd_cluster(int peer
, Message
*m
, epoch_t from_epoch
)
978 OSDMapRef next_map
= get_nextmap_reserved();
979 // service map is always newer/newest
980 assert(from_epoch
<= next_map
->get_epoch());
982 if (next_map
->is_down(peer
) ||
983 next_map
->get_info(peer
).up_from
> from_epoch
) {
985 release_map(next_map
);
988 const entity_inst_t
& peer_inst
= next_map
->get_cluster_inst(peer
);
989 ConnectionRef peer_con
= osd
->cluster_messenger
->get_connection(peer_inst
);
990 share_map_peer(peer
, peer_con
.get(), next_map
);
991 peer_con
->send_message(m
);
992 release_map(next_map
);
995 ConnectionRef
OSDService::get_con_osd_cluster(int peer
, epoch_t from_epoch
)
997 OSDMapRef next_map
= get_nextmap_reserved();
998 // service map is always newer/newest
999 assert(from_epoch
<= next_map
->get_epoch());
1001 if (next_map
->is_down(peer
) ||
1002 next_map
->get_info(peer
).up_from
> from_epoch
) {
1003 release_map(next_map
);
1006 ConnectionRef con
= osd
->cluster_messenger
->get_connection(next_map
->get_cluster_inst(peer
));
1007 release_map(next_map
);
1011 pair
<ConnectionRef
,ConnectionRef
> OSDService::get_con_osd_hb(int peer
, epoch_t from_epoch
)
1013 OSDMapRef next_map
= get_nextmap_reserved();
1014 // service map is always newer/newest
1015 assert(from_epoch
<= next_map
->get_epoch());
1017 pair
<ConnectionRef
,ConnectionRef
> ret
;
1018 if (next_map
->is_down(peer
) ||
1019 next_map
->get_info(peer
).up_from
> from_epoch
) {
1020 release_map(next_map
);
1023 ret
.first
= osd
->hb_back_client_messenger
->get_connection(next_map
->get_hb_back_inst(peer
));
1024 if (next_map
->get_hb_front_addr(peer
) != entity_addr_t())
1025 ret
.second
= osd
->hb_front_client_messenger
->get_connection(next_map
->get_hb_front_inst(peer
));
1026 release_map(next_map
);
1031 void OSDService::queue_want_pg_temp(pg_t pgid
, vector
<int>& want
)
1033 Mutex::Locker
l(pg_temp_lock
);
1034 map
<pg_t
,vector
<int> >::iterator p
= pg_temp_pending
.find(pgid
);
1035 if (p
== pg_temp_pending
.end() ||
1036 p
->second
!= want
) {
1037 pg_temp_wanted
[pgid
] = want
;
1041 void OSDService::remove_want_pg_temp(pg_t pgid
)
1043 Mutex::Locker
l(pg_temp_lock
);
1044 pg_temp_wanted
.erase(pgid
);
1045 pg_temp_pending
.erase(pgid
);
1048 void OSDService::_sent_pg_temp()
1050 for (map
<pg_t
,vector
<int> >::iterator p
= pg_temp_wanted
.begin();
1051 p
!= pg_temp_wanted
.end();
1053 pg_temp_pending
[p
->first
] = p
->second
;
1054 pg_temp_wanted
.clear();
1057 void OSDService::requeue_pg_temp()
1059 Mutex::Locker
l(pg_temp_lock
);
1060 // wanted overrides pending. note that remove_want_pg_temp
1061 // clears the item out of both.
1062 unsigned old_wanted
= pg_temp_wanted
.size();
1063 unsigned old_pending
= pg_temp_pending
.size();
1065 pg_temp_wanted
.swap(pg_temp_pending
);
1066 dout(10) << __func__
<< " " << old_wanted
<< " + " << old_pending
<< " -> "
1067 << pg_temp_wanted
.size() << dendl
;
1070 void OSDService::send_pg_temp()
1072 Mutex::Locker
l(pg_temp_lock
);
1073 if (pg_temp_wanted
.empty())
1075 dout(10) << "send_pg_temp " << pg_temp_wanted
<< dendl
;
1076 MOSDPGTemp
*m
= new MOSDPGTemp(osdmap
->get_epoch());
1077 m
->pg_temp
= pg_temp_wanted
;
1078 monc
->send_mon_message(m
);
1082 void OSDService::send_pg_created(pg_t pgid
)
1084 dout(20) << __func__
<< dendl
;
1085 if (osdmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
1086 monc
->send_mon_message(new MOSDPGCreated(pgid
));
1090 // --------------------------------------
1093 epoch_t
OSDService::get_peer_epoch(int peer
)
1095 Mutex::Locker
l(peer_map_epoch_lock
);
1096 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1097 if (p
== peer_map_epoch
.end())
1102 epoch_t
OSDService::note_peer_epoch(int peer
, epoch_t e
)
1104 Mutex::Locker
l(peer_map_epoch_lock
);
1105 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1106 if (p
!= peer_map_epoch
.end()) {
1107 if (p
->second
< e
) {
1108 dout(10) << "note_peer_epoch osd." << peer
<< " has " << e
<< dendl
;
1111 dout(30) << "note_peer_epoch osd." << peer
<< " has " << p
->second
<< " >= " << e
<< dendl
;
1115 dout(10) << "note_peer_epoch osd." << peer
<< " now has " << e
<< dendl
;
1116 peer_map_epoch
[peer
] = e
;
1121 void OSDService::forget_peer_epoch(int peer
, epoch_t as_of
)
1123 Mutex::Locker
l(peer_map_epoch_lock
);
1124 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1125 if (p
!= peer_map_epoch
.end()) {
1126 if (p
->second
<= as_of
) {
1127 dout(10) << "forget_peer_epoch osd." << peer
<< " as_of " << as_of
1128 << " had " << p
->second
<< dendl
;
1129 peer_map_epoch
.erase(p
);
1131 dout(10) << "forget_peer_epoch osd." << peer
<< " as_of " << as_of
1132 << " has " << p
->second
<< " - not forgetting" << dendl
;
1137 bool OSDService::should_share_map(entity_name_t name
, Connection
*con
,
1138 epoch_t epoch
, const OSDMapRef
& osdmap
,
1139 const epoch_t
*sent_epoch_p
)
1141 dout(20) << "should_share_map "
1142 << name
<< " " << con
->get_peer_addr()
1143 << " " << epoch
<< dendl
;
1145 // does client have old map?
1146 if (name
.is_client()) {
1147 bool message_sendmap
= epoch
< osdmap
->get_epoch();
1148 if (message_sendmap
&& sent_epoch_p
) {
1149 dout(20) << "client session last_sent_epoch: "
1151 << " versus osdmap epoch " << osdmap
->get_epoch() << dendl
;
1152 if (*sent_epoch_p
< osdmap
->get_epoch()) {
1154 } // else we don't need to send it out again
1158 if (con
->get_messenger() == osd
->cluster_messenger
&&
1159 con
!= osd
->cluster_messenger
->get_loopback_connection() &&
1160 osdmap
->is_up(name
.num()) &&
1161 (osdmap
->get_cluster_addr(name
.num()) == con
->get_peer_addr() ||
1162 osdmap
->get_hb_back_addr(name
.num()) == con
->get_peer_addr())) {
1164 epoch_t has
= MAX(get_peer_epoch(name
.num()), epoch
);
1167 if (has
< osdmap
->get_epoch()) {
1168 dout(10) << name
<< " " << con
->get_peer_addr()
1169 << " has old map " << epoch
<< " < "
1170 << osdmap
->get_epoch() << dendl
;
1178 void OSDService::share_map(
1183 epoch_t
*sent_epoch_p
)
1185 dout(20) << "share_map "
1186 << name
<< " " << con
->get_peer_addr()
1187 << " " << epoch
<< dendl
;
1189 if (!osd
->is_active()) {
1190 /*It is safe not to proceed as OSD is not in healthy state*/
1194 bool want_shared
= should_share_map(name
, con
, epoch
,
1195 osdmap
, sent_epoch_p
);
1198 if (name
.is_client()) {
1199 dout(10) << name
<< " has old map " << epoch
1200 << " < " << osdmap
->get_epoch() << dendl
;
1201 // we know the Session is valid or we wouldn't be sending
1203 *sent_epoch_p
= osdmap
->get_epoch();
1205 send_incremental_map(epoch
, con
, osdmap
);
1206 } else if (con
->get_messenger() == osd
->cluster_messenger
&&
1207 osdmap
->is_up(name
.num()) &&
1208 (osdmap
->get_cluster_addr(name
.num()) == con
->get_peer_addr() ||
1209 osdmap
->get_hb_back_addr(name
.num()) == con
->get_peer_addr())) {
1210 dout(10) << name
<< " " << con
->get_peer_addr()
1211 << " has old map " << epoch
<< " < "
1212 << osdmap
->get_epoch() << dendl
;
1213 note_peer_epoch(name
.num(), osdmap
->get_epoch());
1214 send_incremental_map(epoch
, con
, osdmap
);
1219 void OSDService::share_map_peer(int peer
, Connection
*con
, OSDMapRef map
)
1225 epoch_t pe
= get_peer_epoch(peer
);
1227 if (pe
< map
->get_epoch()) {
1228 send_incremental_map(pe
, con
, map
);
1229 note_peer_epoch(peer
, map
->get_epoch());
1231 dout(20) << "share_map_peer " << con
<< " already has epoch " << pe
<< dendl
;
1233 dout(20) << "share_map_peer " << con
<< " don't know epoch, doing nothing" << dendl
;
1234 // no idea about peer's epoch.
1235 // ??? send recent ???
1240 bool OSDService::can_inc_scrubs_pending()
1242 bool can_inc
= false;
1243 Mutex::Locker
l(sched_scrub_lock
);
1245 if (scrubs_pending
+ scrubs_active
< cct
->_conf
->osd_max_scrubs
) {
1246 dout(20) << __func__
<< " " << scrubs_pending
<< " -> " << (scrubs_pending
+1)
1247 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1250 dout(20) << __func__
<< scrubs_pending
<< " + " << scrubs_active
<< " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1256 bool OSDService::inc_scrubs_pending()
1258 bool result
= false;
1260 sched_scrub_lock
.Lock();
1261 if (scrubs_pending
+ scrubs_active
< cct
->_conf
->osd_max_scrubs
) {
1262 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
+1)
1263 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1267 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " + " << scrubs_active
<< " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1269 sched_scrub_lock
.Unlock();
1274 void OSDService::dec_scrubs_pending()
1276 sched_scrub_lock
.Lock();
1277 dout(20) << "dec_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
-1)
1278 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1280 assert(scrubs_pending
>= 0);
1281 sched_scrub_lock
.Unlock();
1284 void OSDService::inc_scrubs_active(bool reserved
)
1286 sched_scrub_lock
.Lock();
1290 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1291 << " (max " << cct
->_conf
->osd_max_scrubs
1292 << ", pending " << (scrubs_pending
+1) << " -> " << scrubs_pending
<< ")" << dendl
;
1293 assert(scrubs_pending
>= 0);
1295 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1296 << " (max " << cct
->_conf
->osd_max_scrubs
1297 << ", pending " << scrubs_pending
<< ")" << dendl
;
1299 sched_scrub_lock
.Unlock();
1302 void OSDService::dec_scrubs_active()
1304 sched_scrub_lock
.Lock();
1305 dout(20) << "dec_scrubs_active " << scrubs_active
<< " -> " << (scrubs_active
-1)
1306 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", pending " << scrubs_pending
<< ")" << dendl
;
1308 assert(scrubs_active
>= 0);
1309 sched_scrub_lock
.Unlock();
1312 void OSDService::retrieve_epochs(epoch_t
*_boot_epoch
, epoch_t
*_up_epoch
,
1313 epoch_t
*_bind_epoch
) const
1315 Mutex::Locker
l(epoch_lock
);
1317 *_boot_epoch
= boot_epoch
;
1319 *_up_epoch
= up_epoch
;
1321 *_bind_epoch
= bind_epoch
;
1324 void OSDService::set_epochs(const epoch_t
*_boot_epoch
, const epoch_t
*_up_epoch
,
1325 const epoch_t
*_bind_epoch
)
1327 Mutex::Locker
l(epoch_lock
);
1329 assert(*_boot_epoch
== 0 || *_boot_epoch
>= boot_epoch
);
1330 boot_epoch
= *_boot_epoch
;
1333 assert(*_up_epoch
== 0 || *_up_epoch
>= up_epoch
);
1334 up_epoch
= *_up_epoch
;
1337 assert(*_bind_epoch
== 0 || *_bind_epoch
>= bind_epoch
);
1338 bind_epoch
= *_bind_epoch
;
1342 bool OSDService::prepare_to_stop()
1344 Mutex::Locker
l(is_stopping_lock
);
1345 if (get_state() != NOT_STOPPING
)
1348 OSDMapRef osdmap
= get_osdmap();
1349 if (osdmap
&& osdmap
->is_up(whoami
)) {
1350 dout(0) << __func__
<< " telling mon we are shutting down" << dendl
;
1351 set_state(PREPARING_TO_STOP
);
1352 monc
->send_mon_message(new MOSDMarkMeDown(monc
->get_fsid(),
1353 osdmap
->get_inst(whoami
),
1354 osdmap
->get_epoch(),
1357 utime_t now
= ceph_clock_now();
1359 timeout
.set_from_double(now
+ cct
->_conf
->osd_mon_shutdown_timeout
);
1360 while ((ceph_clock_now() < timeout
) &&
1361 (get_state() != STOPPING
)) {
1362 is_stopping_cond
.WaitUntil(is_stopping_lock
, timeout
);
1365 dout(0) << __func__
<< " starting shutdown" << dendl
;
1366 set_state(STOPPING
);
1370 void OSDService::got_stop_ack()
1372 Mutex::Locker
l(is_stopping_lock
);
1373 if (get_state() == PREPARING_TO_STOP
) {
1374 dout(0) << __func__
<< " starting shutdown" << dendl
;
1375 set_state(STOPPING
);
1376 is_stopping_cond
.Signal();
1378 dout(10) << __func__
<< " ignoring msg" << dendl
;
1382 MOSDMap
*OSDService::build_incremental_map_msg(epoch_t since
, epoch_t to
,
1383 OSDSuperblock
& sblock
)
1385 MOSDMap
*m
= new MOSDMap(monc
->get_fsid());
1386 m
->oldest_map
= max_oldest_map
;
1387 m
->newest_map
= sblock
.newest_map
;
1389 for (epoch_t e
= to
; e
> since
; e
--) {
1391 if (e
> m
->oldest_map
&& get_inc_map_bl(e
, bl
)) {
1392 m
->incremental_maps
[e
].claim(bl
);
1393 } else if (get_map_bl(e
, bl
)) {
1394 m
->maps
[e
].claim(bl
);
1397 derr
<< "since " << since
<< " to " << to
1398 << " oldest " << m
->oldest_map
<< " newest " << m
->newest_map
1408 void OSDService::send_map(MOSDMap
*m
, Connection
*con
)
1410 con
->send_message(m
);
1413 void OSDService::send_incremental_map(epoch_t since
, Connection
*con
,
1416 epoch_t to
= osdmap
->get_epoch();
1417 dout(10) << "send_incremental_map " << since
<< " -> " << to
1418 << " to " << con
<< " " << con
->get_peer_addr() << dendl
;
1422 OSDSuperblock
sblock(get_superblock());
1423 if (since
< sblock
.oldest_map
) {
1424 // just send latest full map
1425 MOSDMap
*m
= new MOSDMap(monc
->get_fsid());
1426 m
->oldest_map
= max_oldest_map
;
1427 m
->newest_map
= sblock
.newest_map
;
1428 get_map_bl(to
, m
->maps
[to
]);
1433 if (to
> since
&& (int64_t)(to
- since
) > cct
->_conf
->osd_map_share_max_epochs
) {
1434 dout(10) << " " << (to
- since
) << " > max " << cct
->_conf
->osd_map_share_max_epochs
1435 << ", only sending most recent" << dendl
;
1436 since
= to
- cct
->_conf
->osd_map_share_max_epochs
;
1439 if (to
- since
> (epoch_t
)cct
->_conf
->osd_map_message_max
)
1440 to
= since
+ cct
->_conf
->osd_map_message_max
;
1441 m
= build_incremental_map_msg(since
, to
, sblock
);
1446 bool OSDService::_get_map_bl(epoch_t e
, bufferlist
& bl
)
1448 bool found
= map_bl_cache
.lookup(e
, &bl
);
1451 logger
->inc(l_osd_map_bl_cache_hit
);
1455 logger
->inc(l_osd_map_bl_cache_miss
);
1456 found
= store
->read(coll_t::meta(),
1457 OSD::get_osdmap_pobject_name(e
), 0, 0, bl
,
1458 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1465 bool OSDService::get_inc_map_bl(epoch_t e
, bufferlist
& bl
)
1467 Mutex::Locker
l(map_cache_lock
);
1468 bool found
= map_bl_inc_cache
.lookup(e
, &bl
);
1471 logger
->inc(l_osd_map_bl_cache_hit
);
1475 logger
->inc(l_osd_map_bl_cache_miss
);
1476 found
= store
->read(coll_t::meta(),
1477 OSD::get_inc_osdmap_pobject_name(e
), 0, 0, bl
,
1478 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1480 _add_map_inc_bl(e
, bl
);
1485 void OSDService::_add_map_bl(epoch_t e
, bufferlist
& bl
)
1487 dout(10) << "add_map_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1488 // cache a contiguous buffer
1489 if (bl
.get_num_buffers() > 1) {
1492 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1493 map_bl_cache
.add(e
, bl
);
1496 void OSDService::_add_map_inc_bl(epoch_t e
, bufferlist
& bl
)
1498 dout(10) << "add_map_inc_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1499 // cache a contiguous buffer
1500 if (bl
.get_num_buffers() > 1) {
1503 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1504 map_bl_inc_cache
.add(e
, bl
);
1507 void OSDService::pin_map_inc_bl(epoch_t e
, bufferlist
&bl
)
1509 Mutex::Locker
l(map_cache_lock
);
1510 // cache a contiguous buffer
1511 if (bl
.get_num_buffers() > 1) {
1514 map_bl_inc_cache
.pin(e
, bl
);
1517 void OSDService::pin_map_bl(epoch_t e
, bufferlist
&bl
)
1519 Mutex::Locker
l(map_cache_lock
);
1520 // cache a contiguous buffer
1521 if (bl
.get_num_buffers() > 1) {
1524 map_bl_cache
.pin(e
, bl
);
1527 void OSDService::clear_map_bl_cache_pins(epoch_t e
)
1529 Mutex::Locker
l(map_cache_lock
);
1530 map_bl_inc_cache
.clear_pinned(e
);
1531 map_bl_cache
.clear_pinned(e
);
1534 OSDMapRef
OSDService::_add_map(OSDMap
*o
)
1536 epoch_t e
= o
->get_epoch();
1538 if (cct
->_conf
->osd_map_dedup
) {
1539 // Dedup against an existing map at a nearby epoch
1540 OSDMapRef for_dedup
= map_cache
.lower_bound(e
);
1542 OSDMap::dedup(for_dedup
.get(), o
);
1546 OSDMapRef l
= map_cache
.add(e
, o
, &existed
);
1553 OSDMapRef
OSDService::try_get_map(epoch_t epoch
)
1555 Mutex::Locker
l(map_cache_lock
);
1556 OSDMapRef retval
= map_cache
.lookup(epoch
);
1558 dout(30) << "get_map " << epoch
<< " -cached" << dendl
;
1560 logger
->inc(l_osd_map_cache_hit
);
1565 logger
->inc(l_osd_map_cache_miss
);
1566 epoch_t lb
= map_cache
.cached_key_lower_bound();
1568 dout(30) << "get_map " << epoch
<< " - miss, below lower bound" << dendl
;
1569 logger
->inc(l_osd_map_cache_miss_low
);
1570 logger
->inc(l_osd_map_cache_miss_low_avg
, lb
- epoch
);
1574 OSDMap
*map
= new OSDMap
;
1576 dout(20) << "get_map " << epoch
<< " - loading and decoding " << map
<< dendl
;
1578 if (!_get_map_bl(epoch
, bl
) || bl
.length() == 0) {
1579 derr
<< "failed to load OSD map for epoch " << epoch
<< ", got " << bl
.length() << " bytes" << dendl
;
1585 dout(20) << "get_map " << epoch
<< " - return initial " << map
<< dendl
;
1587 return _add_map(map
);
1593 void OSDService::reply_op_error(OpRequestRef op
, int err
)
1595 reply_op_error(op
, err
, eversion_t(), 0);
1598 void OSDService::reply_op_error(OpRequestRef op
, int err
, eversion_t v
,
1601 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1602 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1604 flags
= m
->get_flags() & (CEPH_OSD_FLAG_ACK
|CEPH_OSD_FLAG_ONDISK
);
1606 MOSDOpReply
*reply
= new MOSDOpReply(m
, err
, osdmap
->get_epoch(), flags
,
1608 reply
->set_reply_versions(v
, uv
);
1609 m
->get_connection()->send_message(reply
);
1612 void OSDService::handle_misdirected_op(PG
*pg
, OpRequestRef op
)
1614 if (!cct
->_conf
->osd_debug_misdirected_ops
) {
1618 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1619 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1621 assert(m
->get_map_epoch() >= pg
->info
.history
.same_primary_since
);
1623 if (pg
->is_ec_pg()) {
1625 * OSD recomputes op target based on current OSDMap. With an EC pg, we
1626 * can get this result:
1627 * 1) client at map 512 sends an op to osd 3, pg_t 3.9 based on mapping
1628 * [CRUSH_ITEM_NONE, 2, 3]/3
1629 * 2) OSD 3 at map 513 remaps op to osd 3, spg_t 3.9s0 based on mapping
1631 * 3) PG 3.9s0 dequeues the op at epoch 512 and notices that it isn't primary
1633 * 4) client resends and this time PG 3.9s0 having caught up to 513 gets
1636 * We can't compute the op target based on the sending map epoch due to
1637 * splitting. The simplest thing is to detect such cases here and drop
1638 * them without an error (the client will resend anyway).
1640 assert(m
->get_map_epoch() <= superblock
.newest_map
);
1641 OSDMapRef opmap
= try_get_map(m
->get_map_epoch());
1643 dout(7) << __func__
<< ": " << *pg
<< " no longer have map for "
1644 << m
->get_map_epoch() << ", dropping" << dendl
;
1647 pg_t _pgid
= m
->get_raw_pg();
1649 if ((m
->get_flags() & CEPH_OSD_FLAG_PGOP
) == 0)
1650 _pgid
= opmap
->raw_pg_to_pg(_pgid
);
1651 if (opmap
->get_primary_shard(_pgid
, &pgid
) &&
1652 pgid
.shard
!= pg
->info
.pgid
.shard
) {
1653 dout(7) << __func__
<< ": " << *pg
<< " primary changed since "
1654 << m
->get_map_epoch() << ", dropping" << dendl
;
1659 dout(7) << *pg
<< " misdirected op in " << m
->get_map_epoch() << dendl
;
1660 clog
->warn() << m
->get_source_inst() << " misdirected " << m
->get_reqid()
1661 << " pg " << m
->get_raw_pg()
1662 << " to osd." << whoami
1663 << " not " << pg
->acting
1664 << " in e" << m
->get_map_epoch() << "/" << osdmap
->get_epoch();
1667 void OSDService::enqueue_back(spg_t pgid
, PGQueueable qi
)
1669 osd
->op_shardedwq
.queue(make_pair(pgid
, qi
));
1672 void OSDService::enqueue_front(spg_t pgid
, PGQueueable qi
)
1674 osd
->op_shardedwq
.queue_front(make_pair(pgid
, qi
));
1677 void OSDService::queue_for_peering(PG
*pg
)
1679 peering_wq
.queue(pg
);
1682 void OSDService::queue_for_snap_trim(PG
*pg
)
1684 dout(10) << "queueing " << *pg
<< " for snaptrim" << dendl
;
1685 osd
->op_shardedwq
.queue(
1689 PGSnapTrim(pg
->get_osdmap()->get_epoch()),
1690 cct
->_conf
->osd_snap_trim_cost
,
1691 cct
->_conf
->osd_snap_trim_priority
,
1694 pg
->get_osdmap()->get_epoch())));
1698 // ====================================================================
1702 #define dout_prefix *_dout
1704 // Commands shared between OSD's console and admin console:
1706 namespace osd_cmds
{
1708 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
);
1710 }} // namespace ceph::osd_cmds
1712 int OSD::mkfs(CephContext
*cct
, ObjectStore
*store
, const string
&dev
,
1713 uuid_d fsid
, int whoami
)
1717 ceph::shared_ptr
<ObjectStore::Sequencer
> osr(
1718 new ObjectStore::Sequencer("mkfs"));
1723 // if we are fed a uuid for this osd, use it.
1724 store
->set_fsid(cct
->_conf
->osd_uuid
);
1726 ret
= store
->mkfs();
1728 derr
<< "OSD::mkfs: ObjectStore::mkfs failed with error "
1729 << cpp_strerror(ret
) << dendl
;
1733 store
->set_cache_shards(1); // doesn't matter for mkfs!
1735 ret
= store
->mount();
1737 derr
<< "OSD::mkfs: couldn't mount ObjectStore: error "
1738 << cpp_strerror(ret
) << dendl
;
1742 ret
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, sbbl
);
1744 /* if we already have superblock, check content of superblock */
1745 dout(0) << " have superblock" << dendl
;
1746 bufferlist::iterator p
;
1749 if (whoami
!= sb
.whoami
) {
1750 derr
<< "provided osd id " << whoami
<< " != superblock's " << sb
.whoami
1755 if (fsid
!= sb
.cluster_fsid
) {
1756 derr
<< "provided cluster fsid " << fsid
1757 << " != superblock's " << sb
.cluster_fsid
<< dendl
;
1762 // create superblock
1763 sb
.cluster_fsid
= fsid
;
1764 sb
.osd_fsid
= store
->get_fsid();
1766 sb
.compat_features
= get_osd_initial_compat_set();
1771 ObjectStore::Transaction t
;
1772 t
.create_collection(coll_t::meta(), 0);
1773 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
1774 ret
= store
->apply_transaction(osr
.get(), std::move(t
));
1776 derr
<< "OSD::mkfs: error while writing OSD_SUPERBLOCK_GOBJECT: "
1777 << "apply_transaction returned " << cpp_strerror(ret
) << dendl
;
1782 if (!osr
->flush_commit(&waiter
)) {
1786 ret
= write_meta(store
, sb
.cluster_fsid
, sb
.osd_fsid
, whoami
);
1788 derr
<< "OSD::mkfs: failed to write fsid file: error "
1789 << cpp_strerror(ret
) << dendl
;
1800 int OSD::write_meta(ObjectStore
*store
, uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int whoami
)
1805 snprintf(val
, sizeof(val
), "%s", CEPH_OSD_ONDISK_MAGIC
);
1806 r
= store
->write_meta("magic", val
);
1810 snprintf(val
, sizeof(val
), "%d", whoami
);
1811 r
= store
->write_meta("whoami", val
);
1815 cluster_fsid
.print(val
);
1816 r
= store
->write_meta("ceph_fsid", val
);
1820 r
= store
->write_meta("ready", "ready");
1827 int OSD::peek_meta(ObjectStore
*store
, std::string
& magic
,
1828 uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int& whoami
)
1832 int r
= store
->read_meta("magic", &val
);
1837 r
= store
->read_meta("whoami", &val
);
1840 whoami
= atoi(val
.c_str());
1842 r
= store
->read_meta("ceph_fsid", &val
);
1845 r
= cluster_fsid
.parse(val
.c_str());
1849 r
= store
->read_meta("fsid", &val
);
1851 osd_fsid
= uuid_d();
1853 r
= osd_fsid
.parse(val
.c_str());
1863 #define dout_prefix _prefix(_dout, whoami, get_osdmap_epoch())
1867 OSD::OSD(CephContext
*cct_
, ObjectStore
*store_
,
1869 Messenger
*internal_messenger
,
1870 Messenger
*external_messenger
,
1871 Messenger
*hb_client_front
,
1872 Messenger
*hb_client_back
,
1873 Messenger
*hb_front_serverm
,
1874 Messenger
*hb_back_serverm
,
1875 Messenger
*osdc_messenger
,
1877 const std::string
&dev
, const std::string
&jdev
) :
1879 osd_lock("OSD::osd_lock"),
1880 tick_timer(cct
, osd_lock
),
1881 tick_timer_lock("OSD::tick_timer_lock"),
1882 tick_timer_without_osd_lock(cct
, tick_timer_lock
),
1883 authorize_handler_cluster_registry(new AuthAuthorizeHandlerRegistry(cct
,
1884 cct
->_conf
->auth_supported
.empty() ?
1885 cct
->_conf
->auth_cluster_required
:
1886 cct
->_conf
->auth_supported
)),
1887 authorize_handler_service_registry(new AuthAuthorizeHandlerRegistry(cct
,
1888 cct
->_conf
->auth_supported
.empty() ?
1889 cct
->_conf
->auth_service_required
:
1890 cct
->_conf
->auth_supported
)),
1891 cluster_messenger(internal_messenger
),
1892 client_messenger(external_messenger
),
1893 objecter_messenger(osdc_messenger
),
1895 mgrc(cct_
, client_messenger
),
1897 recoverystate_perf(NULL
),
1899 log_client(cct
, client_messenger
, &mc
->monmap
, LogClient::NO_FLAGS
),
1900 clog(log_client
.create_channel()),
1902 dev_path(dev
), journal_path(jdev
),
1903 store_is_rotational(store
->is_rotational()),
1904 trace_endpoint("0.0.0.0", 0, "osd"),
1906 osd_compat(get_osd_compat_set()),
1907 peering_tp(cct
, "OSD::peering_tp", "tp_peering",
1908 cct
->_conf
->osd_peering_wq_threads
,
1909 "osd_peering_tp_threads"),
1910 osd_op_tp(cct
, "OSD::osd_op_tp", "tp_osd_tp",
1911 get_num_op_threads()),
1912 disk_tp(cct
, "OSD::disk_tp", "tp_osd_disk", cct
->_conf
->osd_disk_threads
, "osd_disk_threads"),
1913 command_tp(cct
, "OSD::command_tp", "tp_osd_cmd", 1),
1914 session_waiting_lock("OSD::session_waiting_lock"),
1915 heartbeat_lock("OSD::heartbeat_lock"),
1916 heartbeat_stop(false),
1917 heartbeat_need_update(true),
1918 hb_front_client_messenger(hb_client_front
),
1919 hb_back_client_messenger(hb_client_back
),
1920 hb_front_server_messenger(hb_front_serverm
),
1921 hb_back_server_messenger(hb_back_serverm
),
1923 heartbeat_thread(this),
1924 heartbeat_dispatcher(this),
1925 op_tracker(cct
, cct
->_conf
->osd_enable_op_tracker
,
1926 cct
->_conf
->osd_num_op_tracker_shard
),
1927 test_ops_hook(NULL
),
1928 op_queue(get_io_queue()),
1929 op_prio_cutoff(get_io_prio_cut()),
1931 get_num_op_shards(),
1933 cct
->_conf
->osd_op_thread_timeout
,
1934 cct
->_conf
->osd_op_thread_suicide_timeout
,
1938 cct
->_conf
->osd_op_thread_timeout
,
1939 cct
->_conf
->osd_op_thread_suicide_timeout
,
1941 map_lock("OSD::map_lock"),
1942 pg_map_lock("OSD::pg_map_lock"),
1943 last_pg_create_epoch(0),
1944 mon_report_lock("OSD::mon_report_lock"),
1945 stats_ack_timeout(cct
->_conf
->osd_mon_ack_timeout
),
1947 requested_full_first(0),
1948 requested_full_last(0),
1949 pg_stat_queue_lock("OSD::pg_stat_queue_lock"),
1950 osd_stat_updated(false),
1951 pg_stat_tid(0), pg_stat_tid_flushed(0),
1954 cct
->_conf
->osd_command_thread_timeout
,
1955 cct
->_conf
->osd_command_thread_suicide_timeout
,
1960 cct
->_conf
->osd_remove_thread_timeout
,
1961 cct
->_conf
->osd_remove_thread_suicide_timeout
,
1965 monc
->set_messenger(client_messenger
);
1966 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
1967 cct
->_conf
->osd_op_log_threshold
);
1968 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
1969 cct
->_conf
->osd_op_history_duration
);
1970 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
1971 cct
->_conf
->osd_op_history_slow_op_threshold
);
1973 std::stringstream ss
;
1974 ss
<< "osd." << whoami
;
1975 trace_endpoint
.copy_name(ss
.str());
1981 delete authorize_handler_cluster_registry
;
1982 delete authorize_handler_service_registry
;
1983 delete class_handler
;
1984 cct
->get_perfcounters_collection()->remove(recoverystate_perf
);
1985 cct
->get_perfcounters_collection()->remove(logger
);
1986 delete recoverystate_perf
;
1991 void cls_initialize(ClassHandler
*ch
);
1993 void OSD::handle_signal(int signum
)
1995 assert(signum
== SIGINT
|| signum
== SIGTERM
);
1996 derr
<< "*** Got signal " << sig_str(signum
) << " ***" << dendl
;
2002 Mutex::Locker
lock(osd_lock
);
2006 if (store
->test_mount_in_use()) {
2007 derr
<< "OSD::pre_init: object store '" << dev_path
<< "' is "
2008 << "currently in use. (Is ceph-osd already running?)" << dendl
;
2012 cct
->_conf
->add_observer(this);
2018 class OSDSocketHook
: public AdminSocketHook
{
2021 explicit OSDSocketHook(OSD
*o
) : osd(o
) {}
2022 bool call(std::string admin_command
, cmdmap_t
& cmdmap
, std::string format
,
2023 bufferlist
& out
) override
{
2025 bool r
= osd
->asok_command(admin_command
, cmdmap
, format
, ss
);
2031 bool OSD::asok_command(string admin_command
, cmdmap_t
& cmdmap
, string format
,
2034 Formatter
*f
= Formatter::create(format
, "json-pretty", "json-pretty");
2035 if (admin_command
== "status") {
2036 f
->open_object_section("status");
2037 f
->dump_stream("cluster_fsid") << superblock
.cluster_fsid
;
2038 f
->dump_stream("osd_fsid") << superblock
.osd_fsid
;
2039 f
->dump_unsigned("whoami", superblock
.whoami
);
2040 f
->dump_string("state", get_state_name(get_state()));
2041 f
->dump_unsigned("oldest_map", superblock
.oldest_map
);
2042 f
->dump_unsigned("newest_map", superblock
.newest_map
);
2044 RWLock::RLocker
l(pg_map_lock
);
2045 f
->dump_unsigned("num_pgs", pg_map
.size());
2048 } else if (admin_command
== "flush_journal") {
2049 store
->flush_journal();
2050 } else if (admin_command
== "dump_ops_in_flight" ||
2051 admin_command
== "ops" ||
2052 admin_command
== "dump_blocked_ops" ||
2053 admin_command
== "dump_historic_ops" ||
2054 admin_command
== "dump_historic_ops_by_duration" ||
2055 admin_command
== "dump_historic_slow_ops") {
2057 const string error_str
= "op_tracker tracking is not enabled now, so no ops are tracked currently, \
2058 even those get stuck. Please enable \"osd_enable_op_tracker\", and the tracker \
2059 will start to track new ops received afterwards.";
2061 set
<string
> filters
;
2062 vector
<string
> filter_str
;
2063 if (cmd_getval(cct
, cmdmap
, "filterstr", filter_str
)) {
2064 copy(filter_str
.begin(), filter_str
.end(),
2065 inserter(filters
, filters
.end()));
2068 if (admin_command
== "dump_ops_in_flight" ||
2069 admin_command
== "ops") {
2070 if (!op_tracker
.dump_ops_in_flight(f
, false, filters
)) {
2074 if (admin_command
== "dump_blocked_ops") {
2075 if (!op_tracker
.dump_ops_in_flight(f
, true, filters
)) {
2079 if (admin_command
== "dump_historic_ops") {
2080 if (!op_tracker
.dump_historic_ops(f
, false, filters
)) {
2084 if (admin_command
== "dump_historic_ops_by_duration") {
2085 if (!op_tracker
.dump_historic_ops(f
, true, filters
)) {
2089 if (admin_command
== "dump_historic_slow_ops") {
2090 if (!op_tracker
.dump_historic_slow_ops(f
, filters
)) {
2094 } else if (admin_command
== "dump_op_pq_state") {
2095 f
->open_object_section("pq");
2096 op_shardedwq
.dump(f
);
2098 } else if (admin_command
== "dump_blacklist") {
2099 list
<pair
<entity_addr_t
,utime_t
> > bl
;
2100 OSDMapRef curmap
= service
.get_osdmap();
2102 f
->open_array_section("blacklist");
2103 curmap
->get_blacklist(&bl
);
2104 for (list
<pair
<entity_addr_t
,utime_t
> >::iterator it
= bl
.begin();
2105 it
!= bl
.end(); ++it
) {
2106 f
->open_object_section("entry");
2107 f
->open_object_section("entity_addr_t");
2109 f
->close_section(); //entity_addr_t
2110 it
->second
.localtime(f
->dump_stream("expire_time"));
2111 f
->close_section(); //entry
2113 f
->close_section(); //blacklist
2114 } else if (admin_command
== "dump_watchers") {
2115 list
<obj_watch_item_t
> watchers
;
2118 Mutex::Locker
l(osd_lock
);
2119 RWLock::RLocker
l2(pg_map_lock
);
2120 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2124 list
<obj_watch_item_t
> pg_watchers
;
2125 PG
*pg
= it
->second
;
2127 pg
->get_watchers(pg_watchers
);
2129 watchers
.splice(watchers
.end(), pg_watchers
);
2133 f
->open_array_section("watchers");
2134 for (list
<obj_watch_item_t
>::iterator it
= watchers
.begin();
2135 it
!= watchers
.end(); ++it
) {
2137 f
->open_object_section("watch");
2139 f
->dump_string("namespace", it
->obj
.nspace
);
2140 f
->dump_string("object", it
->obj
.oid
.name
);
2142 f
->open_object_section("entity_name");
2143 it
->wi
.name
.dump(f
);
2144 f
->close_section(); //entity_name_t
2146 f
->dump_unsigned("cookie", it
->wi
.cookie
);
2147 f
->dump_unsigned("timeout", it
->wi
.timeout_seconds
);
2149 f
->open_object_section("entity_addr_t");
2150 it
->wi
.addr
.dump(f
);
2151 f
->close_section(); //entity_addr_t
2153 f
->close_section(); //watch
2156 f
->close_section(); //watchers
2157 } else if (admin_command
== "dump_reservations") {
2158 f
->open_object_section("reservations");
2159 f
->open_object_section("local_reservations");
2160 service
.local_reserver
.dump(f
);
2162 f
->open_object_section("remote_reservations");
2163 service
.remote_reserver
.dump(f
);
2166 } else if (admin_command
== "get_latest_osdmap") {
2167 get_latest_osdmap();
2168 } else if (admin_command
== "heap") {
2169 auto result
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ss
);
2171 // Note: Failed heap profile commands won't necessarily trigger an error:
2172 f
->open_object_section("result");
2173 f
->dump_string("error", cpp_strerror(result
));
2174 f
->dump_bool("success", result
>= 0);
2176 } else if (admin_command
== "set_heap_property") {
2180 bool success
= false;
2181 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2182 error
= "unable to get property";
2184 } else if (!cmd_getval(cct
, cmdmap
, "value", value
)) {
2185 error
= "unable to get value";
2187 } else if (value
< 0) {
2188 error
= "negative value not allowed";
2190 } else if (!ceph_heap_set_numeric_property(property
.c_str(), (size_t)value
)) {
2191 error
= "invalid property";
2196 f
->open_object_section("result");
2197 f
->dump_string("error", error
);
2198 f
->dump_bool("success", success
);
2200 } else if (admin_command
== "get_heap_property") {
2204 bool success
= false;
2205 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2206 error
= "unable to get property";
2208 } else if (!ceph_heap_get_numeric_property(property
.c_str(), &value
)) {
2209 error
= "invalid property";
2214 f
->open_object_section("result");
2215 f
->dump_string("error", error
);
2216 f
->dump_bool("success", success
);
2217 f
->dump_int("value", value
);
2219 } else if (admin_command
== "dump_objectstore_kv_stats") {
2220 store
->get_db_statistics(f
);
2221 } else if (admin_command
== "dump_scrubs") {
2222 service
.dumps_scrub(f
);
2223 } else if (admin_command
== "calc_objectstore_db_histogram") {
2224 store
->generate_db_histogram(f
);
2225 } else if (admin_command
== "flush_store_cache") {
2226 store
->flush_cache();
2227 } else if (admin_command
== "dump_pgstate_history") {
2228 f
->open_object_section("pgstate_history");
2229 RWLock::RLocker
l2(pg_map_lock
);
2230 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2234 PG
*pg
= it
->second
;
2235 f
->dump_stream("pg") << pg
->get_pgid();
2237 pg
->pgstate_history
.dump(f
);
2241 } else if (admin_command
== "compact") {
2242 dout(1) << "triggering manual compaction" << dendl
;
2243 auto start
= ceph::coarse_mono_clock::now();
2245 auto end
= ceph::coarse_mono_clock::now();
2246 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
2247 dout(1) << "finished manual compaction in "
2248 << time_span
.count()
2249 << " seconds" << dendl
;
2250 f
->open_object_section("compact_result");
2251 f
->dump_float("elapsed_time", time_span
.count());
2254 assert(0 == "broken asok registration");
2261 class TestOpsSocketHook
: public AdminSocketHook
{
2262 OSDService
*service
;
2265 TestOpsSocketHook(OSDService
*s
, ObjectStore
*st
) : service(s
), store(st
) {}
2266 bool call(std::string command
, cmdmap_t
& cmdmap
, std::string format
,
2267 bufferlist
& out
) override
{
2269 test_ops(service
, store
, command
, cmdmap
, ss
);
2273 void test_ops(OSDService
*service
, ObjectStore
*store
,
2274 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
);
2278 class OSD::C_Tick
: public Context
{
2281 explicit C_Tick(OSD
*o
) : osd(o
) {}
2282 void finish(int r
) override
{
2287 class OSD::C_Tick_WithoutOSDLock
: public Context
{
2290 explicit C_Tick_WithoutOSDLock(OSD
*o
) : osd(o
) {}
2291 void finish(int r
) override
{
2292 osd
->tick_without_osd_lock();
2296 int OSD::enable_disable_fuse(bool stop
)
2300 string mntpath
= cct
->_conf
->osd_data
+ "/fuse";
2301 if (fuse_store
&& (stop
|| !cct
->_conf
->osd_objectstore_fuse
)) {
2302 dout(1) << __func__
<< " disabling" << dendl
;
2306 r
= ::rmdir(mntpath
.c_str());
2309 derr
<< __func__
<< " failed to rmdir " << mntpath
<< ": "
2310 << cpp_strerror(r
) << dendl
;
2315 if (!fuse_store
&& cct
->_conf
->osd_objectstore_fuse
) {
2316 dout(1) << __func__
<< " enabling" << dendl
;
2317 r
= ::mkdir(mntpath
.c_str(), 0700);
2320 if (r
< 0 && r
!= -EEXIST
) {
2321 derr
<< __func__
<< " unable to create " << mntpath
<< ": "
2322 << cpp_strerror(r
) << dendl
;
2325 fuse_store
= new FuseStore(store
, mntpath
);
2326 r
= fuse_store
->start();
2328 derr
<< __func__
<< " unable to start fuse: " << cpp_strerror(r
) << dendl
;
2334 #endif // HAVE_LIBFUSE
2338 int OSD::get_num_op_shards()
2340 if (cct
->_conf
->osd_op_num_shards
)
2341 return cct
->_conf
->osd_op_num_shards
;
2342 if (store_is_rotational
)
2343 return cct
->_conf
->osd_op_num_shards_hdd
;
2345 return cct
->_conf
->osd_op_num_shards_ssd
;
2348 int OSD::get_num_op_threads()
2350 if (cct
->_conf
->osd_op_num_threads_per_shard
)
2351 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard
;
2352 if (store_is_rotational
)
2353 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_hdd
;
2355 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_ssd
;
2358 float OSD::get_osd_recovery_sleep()
2360 if (cct
->_conf
->osd_recovery_sleep
)
2361 return cct
->_conf
->osd_recovery_sleep
;
2362 if (!store_is_rotational
&& !journal_is_rotational
)
2363 return cct
->_conf
->osd_recovery_sleep_ssd
;
2364 else if (store_is_rotational
&& !journal_is_rotational
)
2365 return cct
->_conf
->get_val
<double>("osd_recovery_sleep_hybrid");
2367 return cct
->_conf
->osd_recovery_sleep_hdd
;
2372 CompatSet initial
, diff
;
2373 Mutex::Locker
lock(osd_lock
);
2378 tick_timer_without_osd_lock
.init();
2379 service
.recovery_request_timer
.init();
2380 service
.recovery_sleep_timer
.init();
2383 dout(2) << "init " << dev_path
2384 << " (looks like " << (store_is_rotational
? "hdd" : "ssd") << ")"
2386 dout(2) << "journal " << journal_path
<< dendl
;
2387 assert(store
); // call pre_init() first!
2389 store
->set_cache_shards(get_num_op_shards());
2391 int r
= store
->mount();
2393 derr
<< "OSD:init: unable to mount object store" << dendl
;
2396 journal_is_rotational
= store
->is_journal_rotational();
2397 dout(2) << "journal looks like " << (journal_is_rotational
? "hdd" : "ssd")
2400 enable_disable_fuse(false);
2402 dout(2) << "boot" << dendl
;
2404 // initialize the daily loadavg with current 15min loadavg
2406 if (getloadavg(loadavgs
, 3) == 3) {
2407 daily_loadavg
= loadavgs
[2];
2409 derr
<< "OSD::init() : couldn't read loadavgs\n" << dendl
;
2410 daily_loadavg
= 1.0;
2413 int rotating_auth_attempts
= 0;
2415 // sanity check long object name handling
2418 l
.oid
.name
= string(cct
->_conf
->osd_max_object_name_len
, 'n');
2419 l
.set_key(string(cct
->_conf
->osd_max_object_name_len
, 'k'));
2420 l
.nspace
= string(cct
->_conf
->osd_max_object_namespace_len
, 's');
2421 r
= store
->validate_hobject_key(l
);
2423 derr
<< "backend (" << store
->get_type() << ") is unable to support max "
2424 << "object name[space] len" << dendl
;
2425 derr
<< " osd max object name len = "
2426 << cct
->_conf
->osd_max_object_name_len
<< dendl
;
2427 derr
<< " osd max object namespace len = "
2428 << cct
->_conf
->osd_max_object_namespace_len
<< dendl
;
2429 derr
<< cpp_strerror(r
) << dendl
;
2430 if (cct
->_conf
->osd_check_max_object_name_len_on_startup
) {
2433 derr
<< "osd_check_max_object_name_len_on_startup = false, starting anyway"
2436 dout(20) << "configured osd_max_object_name[space]_len looks ok" << dendl
;
2441 r
= read_superblock();
2443 derr
<< "OSD::init() : unable to read osd superblock" << dendl
;
2448 if (osd_compat
.compare(superblock
.compat_features
) < 0) {
2449 derr
<< "The disk uses features unsupported by the executable." << dendl
;
2450 derr
<< " ondisk features " << superblock
.compat_features
<< dendl
;
2451 derr
<< " daemon features " << osd_compat
<< dendl
;
2453 if (osd_compat
.writeable(superblock
.compat_features
)) {
2454 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2455 derr
<< "it is still writeable, though. Missing features: " << diff
<< dendl
;
2460 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2461 derr
<< "Cannot write to disk! Missing features: " << diff
<< dendl
;
2467 assert_warn(whoami
== superblock
.whoami
);
2468 if (whoami
!= superblock
.whoami
) {
2469 derr
<< "OSD::init: superblock says osd"
2470 << superblock
.whoami
<< " but I am osd." << whoami
<< dendl
;
2475 initial
= get_osd_initial_compat_set();
2476 diff
= superblock
.compat_features
.unsupported(initial
);
2477 if (superblock
.compat_features
.merge(initial
)) {
2478 // We need to persist the new compat_set before we
2480 dout(5) << "Upgrading superblock adding: " << diff
<< dendl
;
2481 ObjectStore::Transaction t
;
2482 write_superblock(t
);
2483 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2488 // make sure snap mapper object exists
2489 if (!store
->exists(coll_t::meta(), OSD::make_snapmapper_oid())) {
2490 dout(10) << "init creating/touching snapmapper object" << dendl
;
2491 ObjectStore::Transaction t
;
2492 t
.touch(coll_t::meta(), OSD::make_snapmapper_oid());
2493 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2498 class_handler
= new ClassHandler(cct
);
2499 cls_initialize(class_handler
);
2501 if (cct
->_conf
->osd_open_classes_on_start
) {
2502 int r
= class_handler
->open_all_classes();
2504 dout(1) << "warning: got an error loading one or more classes: " << cpp_strerror(r
) << dendl
;
2507 // load up "current" osdmap
2508 assert_warn(!osdmap
);
2510 derr
<< "OSD::init: unable to read current osdmap" << dendl
;
2514 osdmap
= get_map(superblock
.current_epoch
);
2515 check_osdmap_features(store
);
2517 create_recoverystate_perf();
2520 epoch_t bind_epoch
= osdmap
->get_epoch();
2521 service
.set_epochs(NULL
, NULL
, &bind_epoch
);
2524 clear_temp_objects();
2526 // initialize osdmap references in sharded wq
2527 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
2529 // load up pgs (as they previously existed)
2532 dout(2) << "superblock: I am osd." << superblock
.whoami
<< dendl
;
2533 dout(0) << "using " << op_queue
<< " op queue with priority op cut off at " <<
2534 op_prio_cutoff
<< "." << dendl
;
2539 client_messenger
->add_dispatcher_head(this);
2540 cluster_messenger
->add_dispatcher_head(this);
2542 hb_front_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2543 hb_back_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2544 hb_front_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2545 hb_back_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2547 objecter_messenger
->add_dispatcher_head(service
.objecter
);
2549 monc
->set_want_keys(CEPH_ENTITY_TYPE_MON
| CEPH_ENTITY_TYPE_OSD
2550 | CEPH_ENTITY_TYPE_MGR
);
2556 * FIXME: this is a placeholder implementation that unconditionally
2557 * sends every is_primary PG's stats every time we're called, unlike
2558 * the existing mon PGStats mechanism that uses pg_stat_queue and acks.
2559 * This has equivalent cost to the existing worst case where all
2560 * PGs are busy and their stats are always enqueued for sending.
2562 mgrc
.set_pgstats_cb([this](){
2563 RWLock::RLocker
l(map_lock
);
2565 utime_t had_for
= ceph_clock_now() - had_map_since
;
2566 osd_stat_t cur_stat
= service
.get_osd_stat();
2567 cur_stat
.os_perf_stat
= store
->get_cur_stats();
2569 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
2570 m
->osd_stat
= cur_stat
;
2572 Mutex::Locker lec
{min_last_epoch_clean_lock
};
2573 min_last_epoch_clean
= osdmap
->get_epoch();
2574 min_last_epoch_clean_pgs
.clear();
2575 RWLock::RLocker
lpg(pg_map_lock
);
2576 for (const auto &i
: pg_map
) {
2578 if (!pg
->is_primary()) {
2582 pg
->pg_stats_publish_lock
.Lock();
2583 if (pg
->pg_stats_publish_valid
) {
2584 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
2585 const auto lec
= pg
->pg_stats_publish
.get_effective_last_epoch_clean();
2586 min_last_epoch_clean
= min(min_last_epoch_clean
, lec
);
2587 min_last_epoch_clean_pgs
.push_back(pg
->info
.pgid
.pgid
);
2589 pg
->pg_stats_publish_lock
.Unlock();
2596 client_messenger
->add_dispatcher_head(&mgrc
);
2598 // tell monc about log_client so it will know about mon session resets
2599 monc
->set_log_client(&log_client
);
2600 update_log_config();
2607 set_disk_tp_priority();
2609 // start the heartbeat
2610 heartbeat_thread
.create("osd_srv_heartbt");
2613 tick_timer
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick(this));
2615 Mutex::Locker
l(tick_timer_lock
);
2616 tick_timer_without_osd_lock
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick_WithoutOSDLock(this));
2620 service
.publish_map(osdmap
);
2621 service
.publish_superblock(superblock
);
2622 service
.max_oldest_map
= superblock
.oldest_map
;
2626 r
= monc
->authenticate();
2628 derr
<< __func__
<< " authentication failed: " << cpp_strerror(r
)
2630 osd_lock
.Lock(); // locker is going to unlock this on function exit
2636 while (monc
->wait_auth_rotating(30.0) < 0) {
2637 derr
<< "unable to obtain rotating service keys; retrying" << dendl
;
2638 ++rotating_auth_attempts
;
2639 if (rotating_auth_attempts
> g_conf
->max_rotating_auth_attempts
) {
2640 derr
<< __func__
<< " wait_auth_rotating timed out" << dendl
;
2641 osd_lock
.Lock(); // make locker happy
2642 if (!is_stopping()) {
2649 r
= update_crush_device_class();
2651 derr
<< __func__
<< " unable to update_crush_device_class: "
2652 << cpp_strerror(r
) << dendl
;
2657 r
= update_crush_location();
2659 derr
<< __func__
<< " unable to update_crush_location: "
2660 << cpp_strerror(r
) << dendl
;
2669 // start objecter *after* we have authenticated, so that we don't ignore
2670 // the OSDMaps it requests.
2671 service
.final_init();
2675 dout(10) << "ensuring pgs have consumed prior maps" << dendl
;
2679 dout(0) << "done with init, starting boot process" << dendl
;
2681 // subscribe to any pg creations
2682 monc
->sub_want("osd_pg_creates", last_pg_create_epoch
, 0);
2684 // MgrClient needs this (it doesn't have MonClient reference itself)
2685 monc
->sub_want("mgrmap", 0, 0);
2687 // we don't need to ask for an osdmap here; objecter will
2688 //monc->sub_want("osdmap", osdmap->get_epoch(), CEPH_SUBSCRIBE_ONETIME);
2699 enable_disable_fuse(true);
2706 void OSD::final_init()
2708 AdminSocket
*admin_socket
= cct
->get_admin_socket();
2709 asok_hook
= new OSDSocketHook(this);
2710 int r
= admin_socket
->register_command("status", "status", asok_hook
,
2711 "high-level status of OSD");
2713 r
= admin_socket
->register_command("flush_journal", "flush_journal",
2715 "flush the journal to permanent store");
2717 r
= admin_socket
->register_command("dump_ops_in_flight",
2718 "dump_ops_in_flight " \
2719 "name=filterstr,type=CephString,n=N,req=false",
2721 "show the ops currently in flight");
2723 r
= admin_socket
->register_command("ops",
2725 "name=filterstr,type=CephString,n=N,req=false",
2727 "show the ops currently in flight");
2729 r
= admin_socket
->register_command("dump_blocked_ops",
2730 "dump_blocked_ops " \
2731 "name=filterstr,type=CephString,n=N,req=false",
2733 "show the blocked ops currently in flight");
2735 r
= admin_socket
->register_command("dump_historic_ops",
2736 "dump_historic_ops " \
2737 "name=filterstr,type=CephString,n=N,req=false",
2741 r
= admin_socket
->register_command("dump_historic_slow_ops",
2742 "dump_historic_slow_ops " \
2743 "name=filterstr,type=CephString,n=N,req=false",
2745 "show slowest recent ops");
2747 r
= admin_socket
->register_command("dump_historic_ops_by_duration",
2748 "dump_historic_ops_by_duration " \
2749 "name=filterstr,type=CephString,n=N,req=false",
2751 "show slowest recent ops, sorted by duration");
2753 r
= admin_socket
->register_command("dump_op_pq_state", "dump_op_pq_state",
2755 "dump op priority queue state");
2757 r
= admin_socket
->register_command("dump_blacklist", "dump_blacklist",
2759 "dump blacklisted clients and times");
2761 r
= admin_socket
->register_command("dump_watchers", "dump_watchers",
2763 "show clients which have active watches,"
2764 " and on which objects");
2766 r
= admin_socket
->register_command("dump_reservations", "dump_reservations",
2768 "show recovery reservations");
2770 r
= admin_socket
->register_command("get_latest_osdmap", "get_latest_osdmap",
2772 "force osd to update the latest map from "
2776 r
= admin_socket
->register_command( "heap",
2778 "name=heapcmd,type=CephString",
2780 "show heap usage info (available only if "
2781 "compiled with tcmalloc)");
2784 r
= admin_socket
->register_command("set_heap_property",
2785 "set_heap_property " \
2786 "name=property,type=CephString " \
2787 "name=value,type=CephInt",
2789 "update malloc extension heap property");
2792 r
= admin_socket
->register_command("get_heap_property",
2793 "get_heap_property " \
2794 "name=property,type=CephString",
2796 "get malloc extension heap property");
2799 r
= admin_socket
->register_command("dump_objectstore_kv_stats",
2800 "dump_objectstore_kv_stats",
2802 "print statistics of kvdb which used by bluestore");
2805 r
= admin_socket
->register_command("dump_scrubs",
2808 "print scheduled scrubs");
2811 r
= admin_socket
->register_command("calc_objectstore_db_histogram",
2812 "calc_objectstore_db_histogram",
2814 "Generate key value histogram of kvdb(rocksdb) which used by bluestore");
2817 r
= admin_socket
->register_command("flush_store_cache",
2818 "flush_store_cache",
2820 "Flush bluestore internal cache");
2822 r
= admin_socket
->register_command("dump_pgstate_history", "dump_pgstate_history",
2824 "show recent state history");
2827 r
= admin_socket
->register_command("compact", "compact",
2829 "Commpact object store's omap."
2830 " WARNING: Compaction probably slows your requests");
2833 test_ops_hook
= new TestOpsSocketHook(&(this->service
), this->store
);
2834 // Note: pools are CephString instead of CephPoolname because
2835 // these commands traditionally support both pool names and numbers
2836 r
= admin_socket
->register_command(
2839 "name=pool,type=CephString " \
2840 "name=objname,type=CephObjectname " \
2841 "name=key,type=CephString "\
2842 "name=val,type=CephString",
2846 r
= admin_socket
->register_command(
2849 "name=pool,type=CephString " \
2850 "name=objname,type=CephObjectname " \
2851 "name=key,type=CephString",
2855 r
= admin_socket
->register_command(
2858 "name=pool,type=CephString " \
2859 "name=objname,type=CephObjectname " \
2860 "name=header,type=CephString",
2865 r
= admin_socket
->register_command(
2868 "name=pool,type=CephString " \
2869 "name=objname,type=CephObjectname",
2871 "output entire object map");
2874 r
= admin_socket
->register_command(
2877 "name=pool,type=CephString " \
2878 "name=objname,type=CephObjectname " \
2879 "name=len,type=CephInt",
2881 "truncate object to length");
2884 r
= admin_socket
->register_command(
2887 "name=pool,type=CephString " \
2888 "name=objname,type=CephObjectname " \
2889 "name=shardid,type=CephInt,req=false,range=0|255",
2891 "inject data error to an object");
2894 r
= admin_socket
->register_command(
2897 "name=pool,type=CephString " \
2898 "name=objname,type=CephObjectname " \
2899 "name=shardid,type=CephInt,req=false,range=0|255",
2901 "inject metadata error to an object");
2903 r
= admin_socket
->register_command(
2904 "set_recovery_delay",
2905 "set_recovery_delay " \
2906 "name=utime,type=CephInt,req=false",
2908 "Delay osd recovery by specified seconds");
2910 r
= admin_socket
->register_command(
2913 "name=pgid,type=CephString ",
2915 "Trigger a scheduled scrub ");
2917 r
= admin_socket
->register_command(
2920 "name=type,type=CephString,req=false " \
2921 "name=count,type=CephInt,req=false ",
2923 "Inject a full disk (optional count times)");
2927 void OSD::create_logger()
2929 dout(10) << "create_logger" << dendl
;
2931 PerfCountersBuilder
osd_plb(cct
, "osd", l_osd_first
, l_osd_last
);
2933 // Latency axis configuration for op histograms, values are in nanoseconds
2934 PerfHistogramCommon::axis_config_d op_hist_x_axis_config
{
2936 PerfHistogramCommon::SCALE_LOG2
, ///< Latency in logarithmic scale
2938 100000, ///< Quantization unit is 100usec
2939 32, ///< Enough to cover much longer than slow requests
2942 // Op size axis configuration for op histograms, values are in bytes
2943 PerfHistogramCommon::axis_config_d op_hist_y_axis_config
{
2944 "Request size (bytes)",
2945 PerfHistogramCommon::SCALE_LOG2
, ///< Request size in logarithmic scale
2947 512, ///< Quantization unit is 512 bytes
2948 32, ///< Enough to cover requests larger than GB
2953 l_osd_op_wip
, "op_wip",
2954 "Replication operations currently being processed (primary)");
2955 osd_plb
.add_u64_counter(
2957 "Client operations",
2958 "ops", PerfCountersBuilder::PRIO_CRITICAL
);
2959 osd_plb
.add_u64_counter(
2960 l_osd_op_inb
, "op_in_bytes",
2961 "Client operations total write size",
2962 "wr", PerfCountersBuilder::PRIO_INTERESTING
);
2963 osd_plb
.add_u64_counter(
2964 l_osd_op_outb
, "op_out_bytes",
2965 "Client operations total read size",
2966 "rd", PerfCountersBuilder::PRIO_INTERESTING
);
2967 osd_plb
.add_time_avg(
2968 l_osd_op_lat
, "op_latency",
2969 "Latency of client operations (including queue time)",
2971 osd_plb
.add_time_avg(
2972 l_osd_op_process_lat
, "op_process_latency",
2973 "Latency of client operations (excluding queue time)");
2974 osd_plb
.add_time_avg(
2975 l_osd_op_prepare_lat
, "op_prepare_latency",
2976 "Latency of client operations (excluding queue time and wait for finished)");
2978 osd_plb
.add_u64_counter(
2979 l_osd_op_r
, "op_r", "Client read operations");
2980 osd_plb
.add_u64_counter(
2981 l_osd_op_r_outb
, "op_r_out_bytes", "Client data read");
2982 osd_plb
.add_time_avg(
2983 l_osd_op_r_lat
, "op_r_latency",
2984 "Latency of read operation (including queue time)");
2985 osd_plb
.add_u64_counter_histogram(
2986 l_osd_op_r_lat_outb_hist
, "op_r_latency_out_bytes_histogram",
2987 op_hist_x_axis_config
, op_hist_y_axis_config
,
2988 "Histogram of operation latency (including queue time) + data read");
2989 osd_plb
.add_time_avg(
2990 l_osd_op_r_process_lat
, "op_r_process_latency",
2991 "Latency of read operation (excluding queue time)");
2992 osd_plb
.add_time_avg(
2993 l_osd_op_r_prepare_lat
, "op_r_prepare_latency",
2994 "Latency of read operations (excluding queue time and wait for finished)");
2995 osd_plb
.add_u64_counter(
2996 l_osd_op_w
, "op_w", "Client write operations");
2997 osd_plb
.add_u64_counter(
2998 l_osd_op_w_inb
, "op_w_in_bytes", "Client data written");
2999 osd_plb
.add_time_avg(
3000 l_osd_op_w_lat
, "op_w_latency",
3001 "Latency of write operation (including queue time)");
3002 osd_plb
.add_u64_counter_histogram(
3003 l_osd_op_w_lat_inb_hist
, "op_w_latency_in_bytes_histogram",
3004 op_hist_x_axis_config
, op_hist_y_axis_config
,
3005 "Histogram of operation latency (including queue time) + data written");
3006 osd_plb
.add_time_avg(
3007 l_osd_op_w_process_lat
, "op_w_process_latency",
3008 "Latency of write operation (excluding queue time)");
3009 osd_plb
.add_time_avg(
3010 l_osd_op_w_prepare_lat
, "op_w_prepare_latency",
3011 "Latency of write operations (excluding queue time and wait for finished)");
3012 osd_plb
.add_u64_counter(
3013 l_osd_op_rw
, "op_rw",
3014 "Client read-modify-write operations");
3015 osd_plb
.add_u64_counter(
3016 l_osd_op_rw_inb
, "op_rw_in_bytes",
3017 "Client read-modify-write operations write in");
3018 osd_plb
.add_u64_counter(
3019 l_osd_op_rw_outb
,"op_rw_out_bytes",
3020 "Client read-modify-write operations read out ");
3021 osd_plb
.add_time_avg(
3022 l_osd_op_rw_lat
, "op_rw_latency",
3023 "Latency of read-modify-write operation (including queue time)");
3024 osd_plb
.add_u64_counter_histogram(
3025 l_osd_op_rw_lat_inb_hist
, "op_rw_latency_in_bytes_histogram",
3026 op_hist_x_axis_config
, op_hist_y_axis_config
,
3027 "Histogram of rw operation latency (including queue time) + data written");
3028 osd_plb
.add_u64_counter_histogram(
3029 l_osd_op_rw_lat_outb_hist
, "op_rw_latency_out_bytes_histogram",
3030 op_hist_x_axis_config
, op_hist_y_axis_config
,
3031 "Histogram of rw operation latency (including queue time) + data read");
3032 osd_plb
.add_time_avg(
3033 l_osd_op_rw_process_lat
, "op_rw_process_latency",
3034 "Latency of read-modify-write operation (excluding queue time)");
3035 osd_plb
.add_time_avg(
3036 l_osd_op_rw_prepare_lat
, "op_rw_prepare_latency",
3037 "Latency of read-modify-write operations (excluding queue time and wait for finished)");
3039 osd_plb
.add_time_avg(l_osd_op_before_queue_op_lat
, "op_before_queue_op_lat",
3040 "Latency of IO before calling queue(before really queue into ShardedOpWq)"); // client io before queue op_wq latency
3041 osd_plb
.add_time_avg(l_osd_op_before_dequeue_op_lat
, "op_before_dequeue_op_lat",
3042 "Latency of IO before calling dequeue_op(already dequeued and get PG lock)"); // client io before dequeue_op latency
3044 osd_plb
.add_u64_counter(
3045 l_osd_sop
, "subop", "Suboperations");
3046 osd_plb
.add_u64_counter(
3047 l_osd_sop_inb
, "subop_in_bytes", "Suboperations total size");
3048 osd_plb
.add_time_avg(l_osd_sop_lat
, "subop_latency", "Suboperations latency");
3050 osd_plb
.add_u64_counter(l_osd_sop_w
, "subop_w", "Replicated writes");
3051 osd_plb
.add_u64_counter(
3052 l_osd_sop_w_inb
, "subop_w_in_bytes", "Replicated written data size");
3053 osd_plb
.add_time_avg(
3054 l_osd_sop_w_lat
, "subop_w_latency", "Replicated writes latency");
3055 osd_plb
.add_u64_counter(
3056 l_osd_sop_pull
, "subop_pull", "Suboperations pull requests");
3057 osd_plb
.add_time_avg(
3058 l_osd_sop_pull_lat
, "subop_pull_latency", "Suboperations pull latency");
3059 osd_plb
.add_u64_counter(
3060 l_osd_sop_push
, "subop_push", "Suboperations push messages");
3061 osd_plb
.add_u64_counter(
3062 l_osd_sop_push_inb
, "subop_push_in_bytes", "Suboperations pushed size");
3063 osd_plb
.add_time_avg(
3064 l_osd_sop_push_lat
, "subop_push_latency", "Suboperations push latency");
3066 osd_plb
.add_u64_counter(l_osd_pull
, "pull", "Pull requests sent");
3067 osd_plb
.add_u64_counter(l_osd_push
, "push", "Push messages sent");
3068 osd_plb
.add_u64_counter(l_osd_push_outb
, "push_out_bytes", "Pushed size");
3070 osd_plb
.add_u64_counter(
3071 l_osd_rop
, "recovery_ops",
3072 "Started recovery operations",
3073 "rop", PerfCountersBuilder::PRIO_INTERESTING
);
3075 osd_plb
.add_u64(l_osd_loadavg
, "loadavg", "CPU load");
3076 osd_plb
.add_u64(l_osd_buf
, "buffer_bytes", "Total allocated buffer size");
3077 osd_plb
.add_u64(l_osd_history_alloc_bytes
, "history_alloc_Mbytes");
3078 osd_plb
.add_u64(l_osd_history_alloc_num
, "history_alloc_num");
3080 l_osd_cached_crc
, "cached_crc", "Total number getting crc from crc_cache");
3082 l_osd_cached_crc_adjusted
, "cached_crc_adjusted",
3083 "Total number getting crc from crc_cache with adjusting");
3084 osd_plb
.add_u64(l_osd_missed_crc
, "missed_crc",
3085 "Total number of crc cache misses");
3087 osd_plb
.add_u64(l_osd_pg
, "numpg", "Placement groups",
3088 "pgs", PerfCountersBuilder::PRIO_USEFUL
);
3090 l_osd_pg_primary
, "numpg_primary",
3091 "Placement groups for which this osd is primary");
3093 l_osd_pg_replica
, "numpg_replica",
3094 "Placement groups for which this osd is replica");
3096 l_osd_pg_stray
, "numpg_stray",
3097 "Placement groups ready to be deleted from this osd");
3099 l_osd_hb_to
, "heartbeat_to_peers", "Heartbeat (ping) peers we send to");
3100 osd_plb
.add_u64_counter(l_osd_map
, "map_messages", "OSD map messages");
3101 osd_plb
.add_u64_counter(l_osd_mape
, "map_message_epochs", "OSD map epochs");
3102 osd_plb
.add_u64_counter(
3103 l_osd_mape_dup
, "map_message_epoch_dups", "OSD map duplicates");
3104 osd_plb
.add_u64_counter(
3105 l_osd_waiting_for_map
, "messages_delayed_for_map",
3106 "Operations waiting for OSD map");
3108 osd_plb
.add_u64_counter(
3109 l_osd_map_cache_hit
, "osd_map_cache_hit", "osdmap cache hit");
3110 osd_plb
.add_u64_counter(
3111 l_osd_map_cache_miss
, "osd_map_cache_miss", "osdmap cache miss");
3112 osd_plb
.add_u64_counter(
3113 l_osd_map_cache_miss_low
, "osd_map_cache_miss_low",
3114 "osdmap cache miss below cache lower bound");
3115 osd_plb
.add_u64_avg(
3116 l_osd_map_cache_miss_low_avg
, "osd_map_cache_miss_low_avg",
3117 "osdmap cache miss, avg distance below cache lower bound");
3118 osd_plb
.add_u64_counter(
3119 l_osd_map_bl_cache_hit
, "osd_map_bl_cache_hit",
3120 "OSDMap buffer cache hits");
3121 osd_plb
.add_u64_counter(
3122 l_osd_map_bl_cache_miss
, "osd_map_bl_cache_miss",
3123 "OSDMap buffer cache misses");
3125 osd_plb
.add_u64(l_osd_stat_bytes
, "stat_bytes", "OSD size");
3126 osd_plb
.add_u64(l_osd_stat_bytes_used
, "stat_bytes_used", "Used space");
3127 osd_plb
.add_u64(l_osd_stat_bytes_avail
, "stat_bytes_avail", "Available space");
3129 osd_plb
.add_u64_counter(
3130 l_osd_copyfrom
, "copyfrom", "Rados \"copy-from\" operations");
3132 osd_plb
.add_u64_counter(l_osd_tier_promote
, "tier_promote", "Tier promotions");
3133 osd_plb
.add_u64_counter(l_osd_tier_flush
, "tier_flush", "Tier flushes");
3134 osd_plb
.add_u64_counter(
3135 l_osd_tier_flush_fail
, "tier_flush_fail", "Failed tier flushes");
3136 osd_plb
.add_u64_counter(
3137 l_osd_tier_try_flush
, "tier_try_flush", "Tier flush attempts");
3138 osd_plb
.add_u64_counter(
3139 l_osd_tier_try_flush_fail
, "tier_try_flush_fail",
3140 "Failed tier flush attempts");
3141 osd_plb
.add_u64_counter(
3142 l_osd_tier_evict
, "tier_evict", "Tier evictions");
3143 osd_plb
.add_u64_counter(
3144 l_osd_tier_whiteout
, "tier_whiteout", "Tier whiteouts");
3145 osd_plb
.add_u64_counter(
3146 l_osd_tier_dirty
, "tier_dirty", "Dirty tier flag set");
3147 osd_plb
.add_u64_counter(
3148 l_osd_tier_clean
, "tier_clean", "Dirty tier flag cleaned");
3149 osd_plb
.add_u64_counter(
3150 l_osd_tier_delay
, "tier_delay", "Tier delays (agent waiting)");
3151 osd_plb
.add_u64_counter(
3152 l_osd_tier_proxy_read
, "tier_proxy_read", "Tier proxy reads");
3153 osd_plb
.add_u64_counter(
3154 l_osd_tier_proxy_write
, "tier_proxy_write", "Tier proxy writes");
3156 osd_plb
.add_u64_counter(
3157 l_osd_agent_wake
, "agent_wake", "Tiering agent wake up");
3158 osd_plb
.add_u64_counter(
3159 l_osd_agent_skip
, "agent_skip", "Objects skipped by agent");
3160 osd_plb
.add_u64_counter(
3161 l_osd_agent_flush
, "agent_flush", "Tiering agent flushes");
3162 osd_plb
.add_u64_counter(
3163 l_osd_agent_evict
, "agent_evict", "Tiering agent evictions");
3165 osd_plb
.add_u64_counter(
3166 l_osd_object_ctx_cache_hit
, "object_ctx_cache_hit", "Object context cache hits");
3167 osd_plb
.add_u64_counter(
3168 l_osd_object_ctx_cache_total
, "object_ctx_cache_total", "Object context cache lookups");
3170 osd_plb
.add_u64_counter(l_osd_op_cache_hit
, "op_cache_hit");
3171 osd_plb
.add_time_avg(
3172 l_osd_tier_flush_lat
, "osd_tier_flush_lat", "Object flush latency");
3173 osd_plb
.add_time_avg(
3174 l_osd_tier_promote_lat
, "osd_tier_promote_lat", "Object promote latency");
3175 osd_plb
.add_time_avg(
3176 l_osd_tier_r_lat
, "osd_tier_r_lat", "Object proxy read latency");
3178 osd_plb
.add_u64_counter(
3179 l_osd_pg_info
, "osd_pg_info", "PG updated its info (using any method)");
3180 osd_plb
.add_u64_counter(
3181 l_osd_pg_fastinfo
, "osd_pg_fastinfo",
3182 "PG updated its info using fastinfo attr");
3183 osd_plb
.add_u64_counter(
3184 l_osd_pg_biginfo
, "osd_pg_biginfo", "PG updated its biginfo attr");
3186 logger
= osd_plb
.create_perf_counters();
3187 cct
->get_perfcounters_collection()->add(logger
);
3190 void OSD::create_recoverystate_perf()
3192 dout(10) << "create_recoverystate_perf" << dendl
;
3194 PerfCountersBuilder
rs_perf(cct
, "recoverystate_perf", rs_first
, rs_last
);
3196 rs_perf
.add_time_avg(rs_initial_latency
, "initial_latency", "Initial recovery state latency");
3197 rs_perf
.add_time_avg(rs_started_latency
, "started_latency", "Started recovery state latency");
3198 rs_perf
.add_time_avg(rs_reset_latency
, "reset_latency", "Reset recovery state latency");
3199 rs_perf
.add_time_avg(rs_start_latency
, "start_latency", "Start recovery state latency");
3200 rs_perf
.add_time_avg(rs_primary_latency
, "primary_latency", "Primary recovery state latency");
3201 rs_perf
.add_time_avg(rs_peering_latency
, "peering_latency", "Peering recovery state latency");
3202 rs_perf
.add_time_avg(rs_backfilling_latency
, "backfilling_latency", "Backfilling recovery state latency");
3203 rs_perf
.add_time_avg(rs_waitremotebackfillreserved_latency
, "waitremotebackfillreserved_latency", "Wait remote backfill reserved recovery state latency");
3204 rs_perf
.add_time_avg(rs_waitlocalbackfillreserved_latency
, "waitlocalbackfillreserved_latency", "Wait local backfill reserved recovery state latency");
3205 rs_perf
.add_time_avg(rs_notbackfilling_latency
, "notbackfilling_latency", "Notbackfilling recovery state latency");
3206 rs_perf
.add_time_avg(rs_repnotrecovering_latency
, "repnotrecovering_latency", "Repnotrecovering recovery state latency");
3207 rs_perf
.add_time_avg(rs_repwaitrecoveryreserved_latency
, "repwaitrecoveryreserved_latency", "Rep wait recovery reserved recovery state latency");
3208 rs_perf
.add_time_avg(rs_repwaitbackfillreserved_latency
, "repwaitbackfillreserved_latency", "Rep wait backfill reserved recovery state latency");
3209 rs_perf
.add_time_avg(rs_reprecovering_latency
, "reprecovering_latency", "RepRecovering recovery state latency");
3210 rs_perf
.add_time_avg(rs_activating_latency
, "activating_latency", "Activating recovery state latency");
3211 rs_perf
.add_time_avg(rs_waitlocalrecoveryreserved_latency
, "waitlocalrecoveryreserved_latency", "Wait local recovery reserved recovery state latency");
3212 rs_perf
.add_time_avg(rs_waitremoterecoveryreserved_latency
, "waitremoterecoveryreserved_latency", "Wait remote recovery reserved recovery state latency");
3213 rs_perf
.add_time_avg(rs_recovering_latency
, "recovering_latency", "Recovering recovery state latency");
3214 rs_perf
.add_time_avg(rs_recovered_latency
, "recovered_latency", "Recovered recovery state latency");
3215 rs_perf
.add_time_avg(rs_clean_latency
, "clean_latency", "Clean recovery state latency");
3216 rs_perf
.add_time_avg(rs_active_latency
, "active_latency", "Active recovery state latency");
3217 rs_perf
.add_time_avg(rs_replicaactive_latency
, "replicaactive_latency", "Replicaactive recovery state latency");
3218 rs_perf
.add_time_avg(rs_stray_latency
, "stray_latency", "Stray recovery state latency");
3219 rs_perf
.add_time_avg(rs_getinfo_latency
, "getinfo_latency", "Getinfo recovery state latency");
3220 rs_perf
.add_time_avg(rs_getlog_latency
, "getlog_latency", "Getlog recovery state latency");
3221 rs_perf
.add_time_avg(rs_waitactingchange_latency
, "waitactingchange_latency", "Waitactingchange recovery state latency");
3222 rs_perf
.add_time_avg(rs_incomplete_latency
, "incomplete_latency", "Incomplete recovery state latency");
3223 rs_perf
.add_time_avg(rs_down_latency
, "down_latency", "Down recovery state latency");
3224 rs_perf
.add_time_avg(rs_getmissing_latency
, "getmissing_latency", "Getmissing recovery state latency");
3225 rs_perf
.add_time_avg(rs_waitupthru_latency
, "waitupthru_latency", "Waitupthru recovery state latency");
3226 rs_perf
.add_time_avg(rs_notrecovering_latency
, "notrecovering_latency", "Notrecovering recovery state latency");
3228 recoverystate_perf
= rs_perf
.create_perf_counters();
3229 cct
->get_perfcounters_collection()->add(recoverystate_perf
);
3234 if (!service
.prepare_to_stop())
3235 return 0; // already shutting down
3237 if (is_stopping()) {
3241 derr
<< "shutdown" << dendl
;
3243 set_state(STATE_STOPPING
);
3246 cct
->_conf
->set_val("debug_osd", "100");
3247 cct
->_conf
->set_val("debug_journal", "100");
3248 cct
->_conf
->set_val("debug_filestore", "100");
3249 cct
->_conf
->set_val("debug_ms", "100");
3250 cct
->_conf
->apply_changes(NULL
);
3252 // stop MgrClient earlier as it's more like an internal consumer of OSD
3255 service
.start_shutdown();
3257 // stop sending work to pgs. this just prevents any new work in _process
3258 // from racing with on_shutdown and potentially entering the pg after.
3259 op_shardedwq
.drain();
3263 RWLock::RLocker
l(pg_map_lock
);
3264 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3267 dout(20) << " kicking pg " << p
->first
<< dendl
;
3269 p
->second
->on_shutdown();
3270 p
->second
->unlock();
3271 p
->second
->osr
->flush();
3274 clear_pg_stat_queue();
3276 // drain op queue again (in case PGs requeued something)
3277 op_shardedwq
.drain();
3279 finished
.clear(); // zap waiters (bleh, this is messy)
3282 op_shardedwq
.clear_pg_slots();
3284 // unregister commands
3285 cct
->get_admin_socket()->unregister_command("status");
3286 cct
->get_admin_socket()->unregister_command("flush_journal");
3287 cct
->get_admin_socket()->unregister_command("dump_ops_in_flight");
3288 cct
->get_admin_socket()->unregister_command("ops");
3289 cct
->get_admin_socket()->unregister_command("dump_blocked_ops");
3290 cct
->get_admin_socket()->unregister_command("dump_historic_ops");
3291 cct
->get_admin_socket()->unregister_command("dump_historic_ops_by_duration");
3292 cct
->get_admin_socket()->unregister_command("dump_historic_slow_ops");
3293 cct
->get_admin_socket()->unregister_command("dump_op_pq_state");
3294 cct
->get_admin_socket()->unregister_command("dump_blacklist");
3295 cct
->get_admin_socket()->unregister_command("dump_watchers");
3296 cct
->get_admin_socket()->unregister_command("dump_reservations");
3297 cct
->get_admin_socket()->unregister_command("get_latest_osdmap");
3298 cct
->get_admin_socket()->unregister_command("heap");
3299 cct
->get_admin_socket()->unregister_command("set_heap_property");
3300 cct
->get_admin_socket()->unregister_command("get_heap_property");
3301 cct
->get_admin_socket()->unregister_command("dump_objectstore_kv_stats");
3302 cct
->get_admin_socket()->unregister_command("dump_scrubs");
3303 cct
->get_admin_socket()->unregister_command("calc_objectstore_db_histogram");
3304 cct
->get_admin_socket()->unregister_command("flush_store_cache");
3305 cct
->get_admin_socket()->unregister_command("dump_pgstate_history");
3306 cct
->get_admin_socket()->unregister_command("compact");
3310 cct
->get_admin_socket()->unregister_command("setomapval");
3311 cct
->get_admin_socket()->unregister_command("rmomapkey");
3312 cct
->get_admin_socket()->unregister_command("setomapheader");
3313 cct
->get_admin_socket()->unregister_command("getomap");
3314 cct
->get_admin_socket()->unregister_command("truncobj");
3315 cct
->get_admin_socket()->unregister_command("injectdataerr");
3316 cct
->get_admin_socket()->unregister_command("injectmdataerr");
3317 cct
->get_admin_socket()->unregister_command("set_recovery_delay");
3318 cct
->get_admin_socket()->unregister_command("trigger_scrub");
3319 cct
->get_admin_socket()->unregister_command("injectfull");
3320 delete test_ops_hook
;
3321 test_ops_hook
= NULL
;
3325 heartbeat_lock
.Lock();
3326 heartbeat_stop
= true;
3327 heartbeat_cond
.Signal();
3328 heartbeat_lock
.Unlock();
3329 heartbeat_thread
.join();
3334 dout(10) << "osd tp stopped" << dendl
;
3338 dout(10) << "op sharded tp stopped" << dendl
;
3342 dout(10) << "command tp stopped" << dendl
;
3346 dout(10) << "disk tp paused (new)" << dendl
;
3348 dout(10) << "stopping agent" << dendl
;
3349 service
.agent_stop();
3353 reset_heartbeat_peers();
3355 tick_timer
.shutdown();
3358 Mutex::Locker
l(tick_timer_lock
);
3359 tick_timer_without_osd_lock
.shutdown();
3362 // note unmount epoch
3363 dout(10) << "noting clean unmount in epoch " << osdmap
->get_epoch() << dendl
;
3364 superblock
.mounted
= service
.get_boot_epoch();
3365 superblock
.clean_thru
= osdmap
->get_epoch();
3366 ObjectStore::Transaction t
;
3367 write_superblock(t
);
3368 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3370 derr
<< "OSD::shutdown: error writing superblock: "
3371 << cpp_strerror(r
) << dendl
;
3376 Mutex::Locker
l(pg_stat_queue_lock
);
3377 assert(pg_stat_queue
.empty());
3380 service
.shutdown_reserver();
3383 #ifdef PG_DEBUG_REFS
3384 service
.dump_live_pgids();
3387 RWLock::RLocker
l(pg_map_lock
);
3388 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3391 dout(20) << " kicking pg " << p
->first
<< dendl
;
3393 if (p
->second
->ref
!= 1) {
3394 derr
<< "pgid " << p
->first
<< " has ref count of "
3395 << p
->second
->ref
<< dendl
;
3396 #ifdef PG_DEBUG_REFS
3397 p
->second
->dump_live_ids();
3399 if (cct
->_conf
->osd_shutdown_pgref_assert
) {
3403 p
->second
->unlock();
3404 p
->second
->put("PGMap");
3408 #ifdef PG_DEBUG_REFS
3409 service
.dump_live_pgids();
3411 cct
->_conf
->remove_observer(this);
3413 dout(10) << "syncing store" << dendl
;
3414 enable_disable_fuse(true);
3416 if (cct
->_conf
->osd_journal_flush_on_shutdown
) {
3417 dout(10) << "flushing journal" << dendl
;
3418 store
->flush_journal();
3424 dout(10) << "Store synced" << dendl
;
3429 osdmap
= OSDMapRef();
3431 op_tracker
.on_shutdown();
3433 class_handler
->shutdown();
3434 client_messenger
->shutdown();
3435 cluster_messenger
->shutdown();
3436 hb_front_client_messenger
->shutdown();
3437 hb_back_client_messenger
->shutdown();
3438 objecter_messenger
->shutdown();
3439 hb_front_server_messenger
->shutdown();
3440 hb_back_server_messenger
->shutdown();
3447 int OSD::mon_cmd_maybe_osd_create(string
&cmd
)
3449 bool created
= false;
3451 dout(10) << __func__
<< " cmd: " << cmd
<< dendl
;
3452 vector
<string
> vcmd
{cmd
};
3456 monc
->start_mon_command(vcmd
, inbl
, NULL
, &outs
, &w
);
3459 if (r
== -ENOENT
&& !created
) {
3460 string newcmd
= "{\"prefix\": \"osd create\", \"id\": " + stringify(whoami
)
3461 + ", \"uuid\": \"" + stringify(superblock
.osd_fsid
) + "\"}";
3462 vector
<string
> vnewcmd
{newcmd
};
3466 monc
->start_mon_command(vnewcmd
, inbl
, NULL
, &outs
, &w
);
3469 derr
<< __func__
<< " fail: osd does not exist and created failed: "
3470 << cpp_strerror(r
) << dendl
;
3476 derr
<< __func__
<< " fail: '" << outs
<< "': " << cpp_strerror(r
) << dendl
;
3485 int OSD::update_crush_location()
3487 if (!cct
->_conf
->osd_crush_update_on_start
) {
3488 dout(10) << __func__
<< " osd_crush_update_on_start = false" << dendl
;
3493 if (cct
->_conf
->osd_crush_initial_weight
>= 0) {
3494 snprintf(weight
, sizeof(weight
), "%.4lf", cct
->_conf
->osd_crush_initial_weight
);
3496 struct store_statfs_t st
;
3497 int r
= store
->statfs(&st
);
3499 derr
<< "statfs: " << cpp_strerror(r
) << dendl
;
3502 snprintf(weight
, sizeof(weight
), "%.4lf",
3504 (double)(st
.total
) /
3505 (double)(1ull << 40 /* TB */)));
3508 std::multimap
<string
,string
> loc
= cct
->crush_location
.get_location();
3509 dout(10) << __func__
<< " crush location is " << loc
<< dendl
;
3512 string("{\"prefix\": \"osd crush create-or-move\", ") +
3513 string("\"id\": ") + stringify(whoami
) + string(", ") +
3514 string("\"weight\":") + weight
+ string(", ") +
3515 string("\"args\": [");
3516 for (multimap
<string
,string
>::iterator p
= loc
.begin(); p
!= loc
.end(); ++p
) {
3517 if (p
!= loc
.begin())
3519 cmd
+= "\"" + p
->first
+ "=" + p
->second
+ "\"";
3523 return mon_cmd_maybe_osd_create(cmd
);
3526 int OSD::update_crush_device_class()
3528 if (!cct
->_conf
->osd_class_update_on_start
) {
3529 dout(10) << __func__
<< " osd_class_update_on_start = false" << dendl
;
3533 string device_class
;
3534 int r
= store
->read_meta("crush_device_class", &device_class
);
3535 if (r
< 0 || device_class
.empty()) {
3536 device_class
= store
->get_default_device_class();
3539 if (device_class
.empty()) {
3540 dout(20) << __func__
<< " no device class stored locally" << dendl
;
3545 string("{\"prefix\": \"osd crush set-device-class\", ") +
3546 string("\"class\": \"") + device_class
+ string("\", ") +
3547 string("\"ids\": [\"") + stringify(whoami
) + string("\"]}");
3549 r
= mon_cmd_maybe_osd_create(cmd
);
3550 // the above cmd can fail for various reasons, e.g.:
3551 // (1) we are connecting to a pre-luminous monitor
3552 // (2) user manually specify a class other than
3553 // 'ceph-disk prepare --crush-device-class'
3554 // simply skip result-checking for now
3558 void OSD::write_superblock(ObjectStore::Transaction
& t
)
3560 dout(10) << "write_superblock " << superblock
<< dendl
;
3562 //hack: at minimum it's using the baseline feature set
3563 if (!superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_BASE
))
3564 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE
);
3567 ::encode(superblock
, bl
);
3568 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
3571 int OSD::read_superblock()
3574 int r
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, bl
);
3578 bufferlist::iterator p
= bl
.begin();
3579 ::decode(superblock
, p
);
3581 dout(10) << "read_superblock " << superblock
<< dendl
;
3586 void OSD::clear_temp_objects()
3588 dout(10) << __func__
<< dendl
;
3590 store
->list_collections(ls
);
3591 for (vector
<coll_t
>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
3593 if (!p
->is_pg(&pgid
))
3596 // list temp objects
3597 dout(20) << " clearing temps in " << *p
<< " pgid " << pgid
<< dendl
;
3599 vector
<ghobject_t
> temps
;
3602 vector
<ghobject_t
> objects
;
3603 store
->collection_list(*p
, next
, ghobject_t::get_max(),
3604 store
->get_ideal_list_max(),
3606 if (objects
.empty())
3608 vector
<ghobject_t
>::iterator q
;
3609 for (q
= objects
.begin(); q
!= objects
.end(); ++q
) {
3610 // Hammer set pool for temps to -1, so check for clean-up
3611 if (q
->hobj
.is_temp() || (q
->hobj
.pool
== -1)) {
3612 temps
.push_back(*q
);
3617 // If we saw a non-temp object and hit the break above we can
3618 // break out of the while loop too.
3619 if (q
!= objects
.end())
3622 if (!temps
.empty()) {
3623 ObjectStore::Transaction t
;
3625 for (vector
<ghobject_t
>::iterator q
= temps
.begin(); q
!= temps
.end(); ++q
) {
3626 dout(20) << " removing " << *p
<< " object " << *q
<< dendl
;
3628 if (++removed
> cct
->_conf
->osd_target_transaction_size
) {
3629 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3630 t
= ObjectStore::Transaction();
3635 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3641 void OSD::recursive_remove_collection(CephContext
* cct
,
3642 ObjectStore
*store
, spg_t pgid
,
3648 make_snapmapper_oid());
3650 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
3651 ObjectStore::Sequencer
>("rm"));
3652 ObjectStore::Transaction t
;
3653 SnapMapper
mapper(cct
, &driver
, 0, 0, 0, pgid
.shard
);
3655 vector
<ghobject_t
> objects
;
3656 store
->collection_list(tmp
, ghobject_t(), ghobject_t::get_max(),
3657 INT_MAX
, &objects
, 0);
3658 generic_dout(10) << __func__
<< " " << objects
<< dendl
;
3661 for (vector
<ghobject_t
>::iterator p
= objects
.begin();
3664 OSDriver::OSTransaction
_t(driver
.get_transaction(&t
));
3665 int r
= mapper
.remove_oid(p
->hobj
, &_t
);
3666 if (r
!= 0 && r
!= -ENOENT
)
3669 if (removed
> cct
->_conf
->osd_target_transaction_size
) {
3670 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3672 t
= ObjectStore::Transaction();
3676 t
.remove_collection(tmp
);
3677 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3681 if (!osr
->flush_commit(&waiter
)) {
3687 // ======================================================
3690 PGPool
OSD::_get_pool(int id
, OSDMapRef createmap
)
3692 if (!createmap
->have_pg_pool(id
)) {
3693 dout(5) << __func__
<< ": the OSDmap does not contain a PG pool with id = "
3698 PGPool p
= PGPool(cct
, createmap
, id
);
3700 dout(10) << "_get_pool " << p
.id
<< dendl
;
3704 PG
*OSD::_open_lock_pg(
3705 OSDMapRef createmap
,
3706 spg_t pgid
, bool no_lockdep_check
)
3708 assert(osd_lock
.is_locked());
3710 PG
* pg
= _make_pg(createmap
, pgid
);
3712 RWLock::WLocker
l(pg_map_lock
);
3713 pg
->lock(no_lockdep_check
);
3715 pg
->get("PGMap"); // because it's in pg_map
3716 service
.pg_add_epoch(pg
->info
.pgid
, createmap
->get_epoch());
3722 OSDMapRef createmap
,
3725 dout(10) << "_open_lock_pg " << pgid
<< dendl
;
3726 PGPool pool
= _get_pool(pgid
.pool(), createmap
);
3730 if (createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_REPLICATED
||
3731 createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_ERASURE
)
3732 pg
= new PrimaryLogPG(&service
, createmap
, pool
, pgid
);
3740 void OSD::add_newly_split_pg(PG
*pg
, PG::RecoveryCtx
*rctx
)
3742 epoch_t
e(service
.get_osdmap()->get_epoch());
3743 pg
->get("PGMap"); // For pg_map
3744 pg_map
[pg
->info
.pgid
] = pg
;
3745 service
.pg_add_epoch(pg
->info
.pgid
, pg
->get_osdmap()->get_epoch());
3747 dout(10) << "Adding newly split pg " << *pg
<< dendl
;
3748 pg
->handle_loaded(rctx
);
3749 pg
->write_if_dirty(*(rctx
->transaction
));
3750 pg
->queue_null(e
, e
);
3751 map
<spg_t
, list
<PG::CephPeeringEvtRef
> >::iterator to_wake
=
3752 peering_wait_for_split
.find(pg
->info
.pgid
);
3753 if (to_wake
!= peering_wait_for_split
.end()) {
3754 for (list
<PG::CephPeeringEvtRef
>::iterator i
=
3755 to_wake
->second
.begin();
3756 i
!= to_wake
->second
.end();
3758 pg
->queue_peering_event(*i
);
3760 peering_wait_for_split
.erase(to_wake
);
3762 if (!service
.get_osdmap()->have_pg_pool(pg
->info
.pgid
.pool()))
3766 OSD::res_result
OSD::_try_resurrect_pg(
3767 OSDMapRef curmap
, spg_t pgid
, spg_t
*resurrected
, PGRef
*old_pg_state
)
3769 assert(resurrected
);
3770 assert(old_pg_state
);
3771 // find nearest ancestor
3772 DeletingStateRef df
;
3775 df
= service
.deleting_pgs
.lookup(cur
);
3780 cur
= cur
.get_parent();
3783 return RES_NONE
; // good to go
3785 df
->old_pg_state
->lock();
3786 OSDMapRef create_map
= df
->old_pg_state
->get_osdmap();
3787 df
->old_pg_state
->unlock();
3789 set
<spg_t
> children
;
3791 if (df
->try_stop_deletion()) {
3792 dout(10) << __func__
<< ": halted deletion on pg " << pgid
<< dendl
;
3794 *old_pg_state
= df
->old_pg_state
;
3795 service
.deleting_pgs
.remove(pgid
); // PG is no longer being removed!
3798 // raced, ensure we don't see DeletingStateRef when we try to
3800 service
.deleting_pgs
.remove(pgid
);
3803 } else if (cur
.is_split(create_map
->get_pg_num(cur
.pool()),
3804 curmap
->get_pg_num(cur
.pool()),
3806 children
.count(pgid
)) {
3807 if (df
->try_stop_deletion()) {
3808 dout(10) << __func__
<< ": halted deletion on ancestor pg " << pgid
3811 *old_pg_state
= df
->old_pg_state
;
3812 service
.deleting_pgs
.remove(cur
); // PG is no longer being removed!
3815 /* this is not a problem, failing to cancel proves that all objects
3816 * have been removed, so no hobject_t overlap is possible
3824 PG
*OSD::_create_lock_pg(
3825 OSDMapRef createmap
,
3830 vector
<int>& up
, int up_primary
,
3831 vector
<int>& acting
, int acting_primary
,
3832 pg_history_t history
,
3833 const PastIntervals
& pi
,
3834 ObjectStore::Transaction
& t
)
3836 assert(osd_lock
.is_locked());
3837 dout(20) << "_create_lock_pg pgid " << pgid
<< dendl
;
3839 PG
*pg
= _open_lock_pg(createmap
, pgid
, true);
3841 service
.init_splits_between(pgid
, pg
->get_osdmap(), service
.get_osdmap());
3854 dout(7) << "_create_lock_pg " << *pg
<< dendl
;
3858 PG
*OSD::_lookup_lock_pg(spg_t pgid
)
3860 RWLock::RLocker
l(pg_map_lock
);
3862 auto pg_map_entry
= pg_map
.find(pgid
);
3863 if (pg_map_entry
== pg_map
.end())
3865 PG
*pg
= pg_map_entry
->second
;
3870 PG
*OSD::lookup_lock_pg(spg_t pgid
)
3872 return _lookup_lock_pg(pgid
);
3875 PG
*OSD::_lookup_lock_pg_with_map_lock_held(spg_t pgid
)
3877 assert(pg_map
.count(pgid
));
3878 PG
*pg
= pg_map
[pgid
];
3883 void OSD::load_pgs()
3885 assert(osd_lock
.is_locked());
3886 dout(0) << "load_pgs" << dendl
;
3888 RWLock::RLocker
l(pg_map_lock
);
3889 assert(pg_map
.empty());
3893 int r
= store
->list_collections(ls
);
3895 derr
<< "failed to list pgs: " << cpp_strerror(-r
) << dendl
;
3898 bool has_upgraded
= false;
3900 for (vector
<coll_t
>::iterator it
= ls
.begin();
3904 if (it
->is_temp(&pgid
) ||
3905 (it
->is_pg(&pgid
) && PG::_has_removal_flag(store
, pgid
))) {
3906 dout(10) << "load_pgs " << *it
<< " clearing temp" << dendl
;
3907 recursive_remove_collection(cct
, store
, pgid
, *it
);
3911 if (!it
->is_pg(&pgid
)) {
3912 dout(10) << "load_pgs ignoring unrecognized " << *it
<< dendl
;
3916 if (pgid
.preferred() >= 0) {
3917 dout(10) << __func__
<< ": skipping localized PG " << pgid
<< dendl
;
3918 // FIXME: delete it too, eventually
3922 dout(10) << "pgid " << pgid
<< " coll " << coll_t(pgid
) << dendl
;
3924 epoch_t map_epoch
= 0;
3925 int r
= PG::peek_map_epoch(store
, pgid
, &map_epoch
, &bl
);
3927 derr
<< __func__
<< " unable to peek at " << pgid
<< " metadata, skipping"
3933 if (map_epoch
> 0) {
3934 OSDMapRef pgosdmap
= service
.try_get_map(map_epoch
);
3936 if (!osdmap
->have_pg_pool(pgid
.pool())) {
3937 derr
<< __func__
<< ": could not find map for epoch " << map_epoch
3938 << " on pg " << pgid
<< ", but the pool is not present in the "
3939 << "current map, so this is probably a result of bug 10617. "
3940 << "Skipping the pg for now, you can use ceph-objectstore-tool "
3941 << "to clean it up later." << dendl
;
3944 derr
<< __func__
<< ": have pgid " << pgid
<< " at epoch "
3945 << map_epoch
<< ", but missing map. Crashing."
3947 assert(0 == "Missing map in load_pgs");
3950 pg
= _open_lock_pg(pgosdmap
, pgid
);
3952 pg
= _open_lock_pg(osdmap
, pgid
);
3954 // there can be no waiters here, so we don't call wake_pg_waiters
3956 pg
->ch
= store
->open_collection(pg
->coll
);
3958 // read pg state, log
3959 pg
->read_state(store
, bl
);
3961 if (pg
->must_upgrade()) {
3962 if (!pg
->can_upgrade()) {
3963 derr
<< "PG needs upgrade, but on-disk data is too old; upgrade to"
3964 << " an older version first." << dendl
;
3965 assert(0 == "PG too old to upgrade");
3967 if (!has_upgraded
) {
3968 derr
<< "PGs are upgrading" << dendl
;
3969 has_upgraded
= true;
3971 dout(10) << "PG " << pg
->info
.pgid
3972 << " must upgrade..." << dendl
;
3976 service
.init_splits_between(pg
->info
.pgid
, pg
->get_osdmap(), osdmap
);
3978 // generate state for PG's current mapping
3979 int primary
, up_primary
;
3980 vector
<int> acting
, up
;
3981 pg
->get_osdmap()->pg_to_up_acting_osds(
3982 pgid
.pgid
, &up
, &up_primary
, &acting
, &primary
);
3983 pg
->init_primary_up_acting(
3988 int role
= OSDMap::calc_pg_role(whoami
, pg
->acting
);
3989 if (pg
->pool
.info
.is_replicated() || role
== pg
->pg_whoami
.shard
)
3994 pg
->reg_next_scrub();
3996 PG::RecoveryCtx
rctx(0, 0, 0, 0, 0, 0);
3997 pg
->handle_loaded(&rctx
);
3999 dout(10) << "load_pgs loaded " << *pg
<< " " << pg
->pg_log
.get_log() << dendl
;
4000 if (pg
->pg_log
.is_dirty()) {
4001 ObjectStore::Transaction t
;
4002 pg
->write_if_dirty(t
);
4003 store
->apply_transaction(pg
->osr
.get(), std::move(t
));
4008 RWLock::RLocker
l(pg_map_lock
);
4009 dout(0) << "load_pgs opened " << pg_map
.size() << " pgs" << dendl
;
4012 // clean up old infos object?
4013 if (has_upgraded
&& store
->exists(coll_t::meta(), OSD::make_infos_oid())) {
4014 dout(1) << __func__
<< " removing legacy infos object" << dendl
;
4015 ObjectStore::Transaction t
;
4016 t
.remove(coll_t::meta(), OSD::make_infos_oid());
4017 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4019 derr
<< __func__
<< ": apply_transaction returned "
4020 << cpp_strerror(r
) << dendl
;
4025 build_past_intervals_parallel();
4030 * build past_intervals efficiently on old, degraded, and buried
4031 * clusters. this is important for efficiently catching up osds that
4032 * are way behind on maps to the current cluster state.
4034 * this is a parallel version of PG::generate_past_intervals().
4035 * follow the same logic, but do all pgs at the same time so that we
4036 * can make a single pass across the osdmap history.
4038 void OSD::build_past_intervals_parallel()
4042 vector
<int> old_acting
, old_up
;
4043 epoch_t same_interval_since
;
4047 map
<PG
*,pistate
> pis
;
4049 // calculate junction of map range
4050 epoch_t end_epoch
= superblock
.oldest_map
;
4051 epoch_t cur_epoch
= superblock
.newest_map
;
4053 RWLock::RLocker
l(pg_map_lock
);
4054 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4059 auto rpib
= pg
->get_required_past_interval_bounds(
4061 superblock
.oldest_map
);
4062 if (rpib
.first
>= rpib
.second
&& pg
->past_intervals
.empty()) {
4063 if (pg
->info
.history
.same_interval_since
== 0) {
4064 pg
->info
.history
.same_interval_since
= rpib
.second
;
4068 auto apib
= pg
->past_intervals
.get_bounds();
4069 if (apib
.second
>= rpib
.second
&&
4070 apib
.first
<= rpib
.first
) {
4071 if (pg
->info
.history
.same_interval_since
== 0) {
4072 pg
->info
.history
.same_interval_since
= rpib
.second
;
4078 dout(10) << pg
->info
.pgid
<< " needs " << rpib
.first
<< "-"
4079 << rpib
.second
<< dendl
;
4080 pistate
& p
= pis
[pg
];
4081 p
.start
= rpib
.first
;
4082 p
.end
= rpib
.second
;
4083 p
.same_interval_since
= 0;
4085 if (rpib
.first
< cur_epoch
)
4086 cur_epoch
= rpib
.first
;
4087 if (rpib
.second
> end_epoch
)
4088 end_epoch
= rpib
.second
;
4092 dout(10) << __func__
<< " nothing to build" << dendl
;
4096 dout(1) << __func__
<< " over " << cur_epoch
<< "-" << end_epoch
<< dendl
;
4097 assert(cur_epoch
<= end_epoch
);
4099 OSDMapRef cur_map
, last_map
;
4100 for ( ; cur_epoch
<= end_epoch
; cur_epoch
++) {
4101 dout(10) << __func__
<< " epoch " << cur_epoch
<< dendl
;
4103 cur_map
= get_map(cur_epoch
);
4105 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4107 pistate
& p
= i
->second
;
4109 if (cur_epoch
< p
.start
|| cur_epoch
> p
.end
)
4112 vector
<int> acting
, up
;
4115 pg_t pgid
= pg
->info
.pgid
.pgid
;
4116 if (p
.same_interval_since
&& last_map
->get_pools().count(pgid
.pool()))
4117 pgid
= pgid
.get_ancestor(last_map
->get_pg_num(pgid
.pool()));
4118 cur_map
->pg_to_up_acting_osds(
4119 pgid
, &up
, &up_primary
, &acting
, &primary
);
4121 if (p
.same_interval_since
== 0) {
4122 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4123 << " first map, acting " << acting
4124 << " up " << up
<< ", same_interval_since = " << cur_epoch
<< dendl
;
4125 p
.same_interval_since
= cur_epoch
;
4127 p
.old_acting
= acting
;
4128 p
.primary
= primary
;
4129 p
.up_primary
= up_primary
;
4134 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable(
4135 pg
->get_is_recoverable_predicate());
4136 std::stringstream debug
;
4137 bool new_interval
= PastIntervals::check_new_interval(
4140 p
.old_acting
, acting
,
4144 p
.same_interval_since
,
4145 pg
->info
.history
.last_epoch_clean
,
4149 &pg
->past_intervals
,
4152 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4153 << " " << debug
.str() << dendl
;
4155 p
.old_acting
= acting
;
4156 p
.primary
= primary
;
4157 p
.up_primary
= up_primary
;
4158 p
.same_interval_since
= cur_epoch
;
4163 // Now that past_intervals have been recomputed let's fix the same_interval_since
4164 // if it was cleared by import.
4165 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4167 pistate
& p
= i
->second
;
4169 if (pg
->info
.history
.same_interval_since
== 0) {
4170 assert(p
.same_interval_since
);
4171 dout(10) << __func__
<< " fix same_interval_since " << p
.same_interval_since
<< " pg " << *pg
<< dendl
;
4172 dout(10) << __func__
<< " past_intervals " << pg
->past_intervals
<< dendl
;
4174 pg
->info
.history
.same_interval_since
= p
.same_interval_since
;
4178 // write info only at the end. this is necessary because we check
4179 // whether the past_intervals go far enough back or forward in time,
4180 // but we don't check for holes. we could avoid it by discarding
4181 // the previous past_intervals and rebuilding from scratch, or we
4182 // can just do this and commit all our work at the end.
4183 ObjectStore::Transaction t
;
4185 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4188 pg
->dirty_big_info
= true;
4189 pg
->dirty_info
= true;
4190 pg
->write_if_dirty(t
);
4193 // don't let the transaction get too big
4194 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
4195 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4196 t
= ObjectStore::Transaction();
4201 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4205 * look up a pg. if we have it, great. if not, consider creating it IF the pg mapping
4206 * hasn't changed since the given epoch and we are the primary.
4208 int OSD::handle_pg_peering_evt(
4210 const pg_history_t
& orig_history
,
4211 const PastIntervals
& pi
,
4213 PG::CephPeeringEvtRef evt
)
4215 if (service
.splitting(pgid
)) {
4216 peering_wait_for_split
[pgid
].push_back(evt
);
4220 PG
*pg
= _lookup_lock_pg(pgid
);
4223 if (!osdmap
->have_pg_pool(pgid
.pool()))
4225 int up_primary
, acting_primary
;
4226 vector
<int> up
, acting
;
4227 osdmap
->pg_to_up_acting_osds(
4228 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4230 pg_history_t history
= orig_history
;
4231 bool valid_history
= project_pg_history(
4232 pgid
, history
, epoch
, up
, up_primary
, acting
, acting_primary
);
4234 if (!valid_history
|| epoch
< history
.same_interval_since
) {
4235 dout(10) << __func__
<< pgid
<< " acting changed in "
4236 << history
.same_interval_since
<< " (msg from " << epoch
<< ")"
4241 if (service
.splitting(pgid
)) {
4245 // do we need to resurrect a deleting pg?
4248 res_result result
= _try_resurrect_pg(
4249 service
.get_osdmap(),
4254 PG::RecoveryCtx rctx
= create_context();
4257 const pg_pool_t
* pp
= osdmap
->get_pg_pool(pgid
.pool());
4258 if (pp
->has_flag(pg_pool_t::FLAG_EC_OVERWRITES
) &&
4259 store
->get_type() != "bluestore") {
4260 clog
->warn() << "pg " << pgid
4261 << " is at risk of silent data corruption: "
4262 << "the pool allows ec overwrites but is not stored in "
4263 << "bluestore, so deep scrubbing will not detect bitrot";
4265 PG::_create(*rctx
.transaction
, pgid
, pgid
.get_split_bits(pp
->get_pg_num()));
4266 PG::_init(*rctx
.transaction
, pgid
, pp
);
4268 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
4269 if (!pp
->is_replicated() && role
!= pgid
.shard
)
4272 pg
= _create_lock_pg(
4277 acting
, acting_primary
,
4280 pg
->handle_create(&rctx
);
4281 pg
->write_if_dirty(*rctx
.transaction
);
4282 dispatch_context(rctx
, pg
, osdmap
);
4284 dout(10) << *pg
<< " is new" << dendl
;
4286 pg
->queue_peering_event(evt
);
4287 wake_pg_waiters(pg
);
4292 old_pg_state
->lock();
4293 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4294 int old_role
= old_pg_state
->role
;
4295 vector
<int> old_up
= old_pg_state
->up
;
4296 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4297 vector
<int> old_acting
= old_pg_state
->acting
;
4298 int old_primary
= old_pg_state
->primary
.osd
;
4299 pg_history_t old_history
= old_pg_state
->info
.history
;
4300 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4301 old_pg_state
->unlock();
4302 pg
= _create_lock_pg(
4315 pg
->handle_create(&rctx
);
4316 pg
->write_if_dirty(*rctx
.transaction
);
4317 dispatch_context(rctx
, pg
, osdmap
);
4319 dout(10) << *pg
<< " is new (resurrected)" << dendl
;
4321 pg
->queue_peering_event(evt
);
4322 wake_pg_waiters(pg
);
4327 assert(old_pg_state
);
4328 old_pg_state
->lock();
4329 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4330 int old_role
= old_pg_state
->role
;
4331 vector
<int> old_up
= old_pg_state
->up
;
4332 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4333 vector
<int> old_acting
= old_pg_state
->acting
;
4334 int old_primary
= old_pg_state
->primary
.osd
;
4335 pg_history_t old_history
= old_pg_state
->info
.history
;
4336 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4337 old_pg_state
->unlock();
4338 PG
*parent
= _create_lock_pg(
4352 parent
->handle_create(&rctx
);
4353 parent
->write_if_dirty(*rctx
.transaction
);
4354 dispatch_context(rctx
, parent
, osdmap
);
4356 dout(10) << *parent
<< " is new" << dendl
;
4358 assert(service
.splitting(pgid
));
4359 peering_wait_for_split
[pgid
].push_back(evt
);
4361 //parent->queue_peering_event(evt);
4362 parent
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
4363 wake_pg_waiters(parent
);
4372 // already had it. did the mapping change?
4373 if (epoch
< pg
->info
.history
.same_interval_since
) {
4374 dout(10) << *pg
<< __func__
<< " acting changed in "
4375 << pg
->info
.history
.same_interval_since
4376 << " (msg from " << epoch
<< ")" << dendl
;
4378 pg
->queue_peering_event(evt
);
4386 void OSD::build_initial_pg_history(
4389 utime_t created_stamp
,
4393 dout(10) << __func__
<< " " << pgid
<< " created " << created
<< dendl
;
4394 h
->epoch_created
= created
;
4395 h
->epoch_pool_created
= created
;
4396 h
->same_interval_since
= created
;
4397 h
->same_up_since
= created
;
4398 h
->same_primary_since
= created
;
4399 h
->last_scrub_stamp
= created_stamp
;
4400 h
->last_deep_scrub_stamp
= created_stamp
;
4401 h
->last_clean_scrub_stamp
= created_stamp
;
4403 OSDMapRef lastmap
= service
.get_map(created
);
4404 int up_primary
, acting_primary
;
4405 vector
<int> up
, acting
;
4406 lastmap
->pg_to_up_acting_osds(
4407 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4409 ostringstream debug
;
4410 for (epoch_t e
= created
+ 1; e
<= osdmap
->get_epoch(); ++e
) {
4411 OSDMapRef osdmap
= service
.get_map(e
);
4412 int new_up_primary
, new_acting_primary
;
4413 vector
<int> new_up
, new_acting
;
4414 osdmap
->pg_to_up_acting_osds(
4415 pgid
.pgid
, &new_up
, &new_up_primary
, &new_acting
, &new_acting_primary
);
4417 // this is a bit imprecise, but sufficient?
4418 struct min_size_predicate_t
: public IsPGRecoverablePredicate
{
4419 const pg_pool_t
*pi
;
4420 bool operator()(const set
<pg_shard_t
> &have
) const {
4421 return have
.size() >= pi
->min_size
;
4423 min_size_predicate_t(const pg_pool_t
*i
) : pi(i
) {}
4424 } min_size_predicate(osdmap
->get_pg_pool(pgid
.pgid
.pool()));
4426 bool new_interval
= PastIntervals::check_new_interval(
4433 h
->same_interval_since
,
4434 h
->last_epoch_clean
,
4438 &min_size_predicate
,
4442 h
->same_interval_since
= e
;
4445 h
->same_up_since
= e
;
4447 if (acting_primary
!= new_acting_primary
) {
4448 h
->same_primary_since
= e
;
4450 if (pgid
.pgid
.is_split(lastmap
->get_pg_num(pgid
.pgid
.pool()),
4451 osdmap
->get_pg_num(pgid
.pgid
.pool()),
4453 h
->last_epoch_split
= e
;
4457 dout(20) << __func__
<< " " << debug
.str() << dendl
;
4458 dout(10) << __func__
<< " " << *h
<< " " << *pi
4459 << " [" << (pi
->empty() ? pair
<epoch_t
,epoch_t
>(0,0) :
4460 pi
->get_bounds()) << ")"
4465 * Fill in the passed history so you know same_interval_since, same_up_since,
4466 * and same_primary_since.
4468 bool OSD::project_pg_history(spg_t pgid
, pg_history_t
& h
, epoch_t from
,
4469 const vector
<int>& currentup
,
4470 int currentupprimary
,
4471 const vector
<int>& currentacting
,
4472 int currentactingprimary
)
4474 dout(15) << "project_pg_history " << pgid
4475 << " from " << from
<< " to " << osdmap
->get_epoch()
4480 for (e
= osdmap
->get_epoch();
4483 // verify during intermediate epoch (e-1)
4484 OSDMapRef oldmap
= service
.try_get_map(e
-1);
4486 dout(15) << __func__
<< ": found map gap, returning false" << dendl
;
4489 assert(oldmap
->have_pg_pool(pgid
.pool()));
4491 int upprimary
, actingprimary
;
4492 vector
<int> up
, acting
;
4493 oldmap
->pg_to_up_acting_osds(
4500 // acting set change?
4501 if ((actingprimary
!= currentactingprimary
||
4502 upprimary
!= currentupprimary
||
4503 acting
!= currentacting
||
4504 up
!= currentup
) && e
> h
.same_interval_since
) {
4505 dout(15) << "project_pg_history " << pgid
<< " acting|up changed in " << e
4506 << " from " << acting
<< "/" << up
4507 << " " << actingprimary
<< "/" << upprimary
4508 << " -> " << currentacting
<< "/" << currentup
4509 << " " << currentactingprimary
<< "/" << currentupprimary
4511 h
.same_interval_since
= e
;
4514 if (pgid
.is_split(oldmap
->get_pg_num(pgid
.pool()),
4515 osdmap
->get_pg_num(pgid
.pool()),
4516 0) && e
> h
.same_interval_since
) {
4517 h
.same_interval_since
= e
;
4520 if ((up
!= currentup
|| upprimary
!= currentupprimary
)
4521 && e
> h
.same_up_since
) {
4522 dout(15) << "project_pg_history " << pgid
<< " up changed in " << e
4523 << " from " << up
<< " " << upprimary
4524 << " -> " << currentup
<< " " << currentupprimary
<< dendl
;
4525 h
.same_up_since
= e
;
4529 if (OSDMap::primary_changed(
4532 currentactingprimary
,
4534 e
> h
.same_primary_since
) {
4535 dout(15) << "project_pg_history " << pgid
<< " primary changed in " << e
<< dendl
;
4536 h
.same_primary_since
= e
;
4539 if (h
.same_interval_since
>= e
&& h
.same_up_since
>= e
&& h
.same_primary_since
>= e
)
4543 // base case: these floors should be the pg creation epoch if we didn't
4544 // find any changes.
4545 if (e
== h
.epoch_created
) {
4546 if (!h
.same_interval_since
)
4547 h
.same_interval_since
= e
;
4548 if (!h
.same_up_since
)
4549 h
.same_up_since
= e
;
4550 if (!h
.same_primary_since
)
4551 h
.same_primary_since
= e
;
4554 dout(15) << "project_pg_history end " << h
<< dendl
;
4560 void OSD::_add_heartbeat_peer(int p
)
4566 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(p
);
4567 if (i
== heartbeat_peers
.end()) {
4568 pair
<ConnectionRef
,ConnectionRef
> cons
= service
.get_con_osd_hb(p
, osdmap
->get_epoch());
4571 hi
= &heartbeat_peers
[p
];
4573 HeartbeatSession
*s
= new HeartbeatSession(p
);
4574 hi
->con_back
= cons
.first
.get();
4575 hi
->con_back
->set_priv(s
->get());
4577 hi
->con_front
= cons
.second
.get();
4578 hi
->con_front
->set_priv(s
->get());
4579 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4580 << " " << hi
->con_back
->get_peer_addr()
4581 << " " << hi
->con_front
->get_peer_addr()
4584 hi
->con_front
.reset(NULL
);
4585 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4586 << " " << hi
->con_back
->get_peer_addr()
4593 hi
->epoch
= osdmap
->get_epoch();
4596 void OSD::_remove_heartbeat_peer(int n
)
4598 map
<int,HeartbeatInfo
>::iterator q
= heartbeat_peers
.find(n
);
4599 assert(q
!= heartbeat_peers
.end());
4600 dout(20) << " removing heartbeat peer osd." << n
4601 << " " << q
->second
.con_back
->get_peer_addr()
4602 << " " << (q
->second
.con_front
? q
->second
.con_front
->get_peer_addr() : entity_addr_t())
4604 q
->second
.con_back
->mark_down();
4605 if (q
->second
.con_front
) {
4606 q
->second
.con_front
->mark_down();
4608 heartbeat_peers
.erase(q
);
4611 void OSD::need_heartbeat_peer_update()
4615 dout(20) << "need_heartbeat_peer_update" << dendl
;
4616 heartbeat_set_peers_need_update();
4619 void OSD::maybe_update_heartbeat_peers()
4621 assert(osd_lock
.is_locked());
4623 if (is_waiting_for_healthy()) {
4624 utime_t now
= ceph_clock_now();
4625 if (last_heartbeat_resample
== utime_t()) {
4626 last_heartbeat_resample
= now
;
4627 heartbeat_set_peers_need_update();
4628 } else if (!heartbeat_peers_need_update()) {
4629 utime_t dur
= now
- last_heartbeat_resample
;
4630 if (dur
> cct
->_conf
->osd_heartbeat_grace
) {
4631 dout(10) << "maybe_update_heartbeat_peers forcing update after " << dur
<< " seconds" << dendl
;
4632 heartbeat_set_peers_need_update();
4633 last_heartbeat_resample
= now
;
4634 reset_heartbeat_peers(); // we want *new* peers!
4639 if (!heartbeat_peers_need_update())
4641 heartbeat_clear_peers_need_update();
4643 Mutex::Locker
l(heartbeat_lock
);
4645 dout(10) << "maybe_update_heartbeat_peers updating" << dendl
;
4648 // build heartbeat from set
4650 RWLock::RLocker
l(pg_map_lock
);
4651 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4655 pg
->heartbeat_peer_lock
.Lock();
4656 dout(20) << i
->first
<< " heartbeat_peers " << pg
->heartbeat_peers
<< dendl
;
4657 for (set
<int>::iterator p
= pg
->heartbeat_peers
.begin();
4658 p
!= pg
->heartbeat_peers
.end();
4660 if (osdmap
->is_up(*p
))
4661 _add_heartbeat_peer(*p
);
4662 for (set
<int>::iterator p
= pg
->probe_targets
.begin();
4663 p
!= pg
->probe_targets
.end();
4665 if (osdmap
->is_up(*p
))
4666 _add_heartbeat_peer(*p
);
4667 pg
->heartbeat_peer_lock
.Unlock();
4671 // include next and previous up osds to ensure we have a fully-connected set
4672 set
<int> want
, extras
;
4673 int next
= osdmap
->get_next_up_osd_after(whoami
);
4676 int prev
= osdmap
->get_previous_up_osd_before(whoami
);
4677 if (prev
>= 0 && prev
!= next
)
4680 for (set
<int>::iterator p
= want
.begin(); p
!= want
.end(); ++p
) {
4681 dout(10) << " adding neighbor peer osd." << *p
<< dendl
;
4683 _add_heartbeat_peer(*p
);
4686 // remove down peers; enumerate extras
4687 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4688 while (p
!= heartbeat_peers
.end()) {
4689 if (!osdmap
->is_up(p
->first
)) {
4692 _remove_heartbeat_peer(o
);
4695 if (p
->second
.epoch
< osdmap
->get_epoch()) {
4696 extras
.insert(p
->first
);
4702 int start
= osdmap
->get_next_up_osd_after(whoami
);
4703 for (int n
= start
; n
>= 0; ) {
4704 if ((int)heartbeat_peers
.size() >= cct
->_conf
->osd_heartbeat_min_peers
)
4706 if (!extras
.count(n
) && !want
.count(n
) && n
!= whoami
) {
4707 dout(10) << " adding random peer osd." << n
<< dendl
;
4709 _add_heartbeat_peer(n
);
4711 n
= osdmap
->get_next_up_osd_after(n
);
4713 break; // came full circle; stop
4717 for (set
<int>::iterator p
= extras
.begin();
4718 (int)heartbeat_peers
.size() > cct
->_conf
->osd_heartbeat_min_peers
&& p
!= extras
.end();
4722 _remove_heartbeat_peer(*p
);
4725 dout(10) << "maybe_update_heartbeat_peers " << heartbeat_peers
.size() << " peers, extras " << extras
<< dendl
;
4728 void OSD::reset_heartbeat_peers()
4730 assert(osd_lock
.is_locked());
4731 dout(10) << "reset_heartbeat_peers" << dendl
;
4732 Mutex::Locker
l(heartbeat_lock
);
4733 while (!heartbeat_peers
.empty()) {
4734 HeartbeatInfo
& hi
= heartbeat_peers
.begin()->second
;
4735 hi
.con_back
->mark_down();
4737 hi
.con_front
->mark_down();
4739 heartbeat_peers
.erase(heartbeat_peers
.begin());
4741 failure_queue
.clear();
4744 void OSD::handle_osd_ping(MOSDPing
*m
)
4746 if (superblock
.cluster_fsid
!= m
->fsid
) {
4747 dout(20) << "handle_osd_ping from " << m
->get_source_inst()
4748 << " bad fsid " << m
->fsid
<< " != " << superblock
.cluster_fsid
<< dendl
;
4753 int from
= m
->get_source().num();
4755 heartbeat_lock
.Lock();
4756 if (is_stopping()) {
4757 heartbeat_lock
.Unlock();
4762 OSDMapRef curmap
= service
.get_osdmap();
4764 heartbeat_lock
.Unlock();
4771 case MOSDPing::PING
:
4773 if (cct
->_conf
->osd_debug_drop_ping_probability
> 0) {
4774 auto heartbeat_drop
= debug_heartbeat_drops_remaining
.find(from
);
4775 if (heartbeat_drop
!= debug_heartbeat_drops_remaining
.end()) {
4776 if (heartbeat_drop
->second
== 0) {
4777 debug_heartbeat_drops_remaining
.erase(heartbeat_drop
);
4779 --heartbeat_drop
->second
;
4780 dout(5) << "Dropping heartbeat from " << from
4781 << ", " << heartbeat_drop
->second
4782 << " remaining to drop" << dendl
;
4785 } else if (cct
->_conf
->osd_debug_drop_ping_probability
>
4786 ((((double)(rand()%100))/100.0))) {
4788 debug_heartbeat_drops_remaining
.insert(std::make_pair(from
,
4789 cct
->_conf
->osd_debug_drop_ping_duration
)).first
;
4790 dout(5) << "Dropping heartbeat from " << from
4791 << ", " << heartbeat_drop
->second
4792 << " remaining to drop" << dendl
;
4797 if (!cct
->get_heartbeat_map()->is_healthy()) {
4798 dout(10) << "internal heartbeat not healthy, dropping ping request" << dendl
;
4802 Message
*r
= new MOSDPing(monc
->get_fsid(),
4803 curmap
->get_epoch(),
4804 MOSDPing::PING_REPLY
, m
->stamp
,
4805 cct
->_conf
->osd_heartbeat_min_size
);
4806 m
->get_connection()->send_message(r
);
4808 if (curmap
->is_up(from
)) {
4809 service
.note_peer_epoch(from
, m
->map_epoch
);
4811 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
4813 service
.share_map_peer(from
, con
.get());
4816 } else if (!curmap
->exists(from
) ||
4817 curmap
->get_down_at(from
) > m
->map_epoch
) {
4818 // tell them they have died
4819 Message
*r
= new MOSDPing(monc
->get_fsid(),
4820 curmap
->get_epoch(),
4823 cct
->_conf
->osd_heartbeat_min_size
);
4824 m
->get_connection()->send_message(r
);
4829 case MOSDPing::PING_REPLY
:
4831 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(from
);
4832 if (i
!= heartbeat_peers
.end()) {
4833 if (m
->get_connection() == i
->second
.con_back
) {
4834 dout(25) << "handle_osd_ping got reply from osd." << from
4835 << " first_tx " << i
->second
.first_tx
4836 << " last_tx " << i
->second
.last_tx
4837 << " last_rx_back " << i
->second
.last_rx_back
<< " -> " << m
->stamp
4838 << " last_rx_front " << i
->second
.last_rx_front
4840 i
->second
.last_rx_back
= m
->stamp
;
4841 // if there is no front con, set both stamps.
4842 if (i
->second
.con_front
== NULL
)
4843 i
->second
.last_rx_front
= m
->stamp
;
4844 } else if (m
->get_connection() == i
->second
.con_front
) {
4845 dout(25) << "handle_osd_ping got reply from osd." << from
4846 << " first_tx " << i
->second
.first_tx
4847 << " last_tx " << i
->second
.last_tx
4848 << " last_rx_back " << i
->second
.last_rx_back
4849 << " last_rx_front " << i
->second
.last_rx_front
<< " -> " << m
->stamp
4851 i
->second
.last_rx_front
= m
->stamp
;
4854 utime_t cutoff
= ceph_clock_now();
4855 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
4856 if (i
->second
.is_healthy(cutoff
)) {
4857 // Cancel false reports
4858 auto failure_queue_entry
= failure_queue
.find(from
);
4859 if (failure_queue_entry
!= failure_queue
.end()) {
4860 dout(10) << "handle_osd_ping canceling queued "
4861 << "failure report for osd." << from
<< dendl
;
4862 failure_queue
.erase(failure_queue_entry
);
4865 auto failure_pending_entry
= failure_pending
.find(from
);
4866 if (failure_pending_entry
!= failure_pending
.end()) {
4867 dout(10) << "handle_osd_ping canceling in-flight "
4868 << "failure report for osd." << from
<< dendl
;
4869 send_still_alive(curmap
->get_epoch(),
4870 failure_pending_entry
->second
.second
);
4871 failure_pending
.erase(failure_pending_entry
);
4877 curmap
->is_up(from
)) {
4878 service
.note_peer_epoch(from
, m
->map_epoch
);
4880 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
4882 service
.share_map_peer(from
, con
.get());
4889 case MOSDPing::YOU_DIED
:
4890 dout(10) << "handle_osd_ping " << m
->get_source_inst()
4891 << " says i am down in " << m
->map_epoch
<< dendl
;
4892 osdmap_subscribe(curmap
->get_epoch()+1, false);
4896 heartbeat_lock
.Unlock();
4900 void OSD::heartbeat_entry()
4902 Mutex::Locker
l(heartbeat_lock
);
4905 while (!heartbeat_stop
) {
4908 double wait
= .5 + ((float)(rand() % 10)/10.0) * (float)cct
->_conf
->osd_heartbeat_interval
;
4910 w
.set_from_double(wait
);
4911 dout(30) << "heartbeat_entry sleeping for " << wait
<< dendl
;
4912 heartbeat_cond
.WaitInterval(heartbeat_lock
, w
);
4915 dout(30) << "heartbeat_entry woke up" << dendl
;
4919 void OSD::heartbeat_check()
4921 assert(heartbeat_lock
.is_locked());
4922 utime_t now
= ceph_clock_now();
4924 // check for heartbeat replies (move me elsewhere?)
4925 utime_t cutoff
= now
;
4926 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
4927 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4928 p
!= heartbeat_peers
.end();
4931 if (p
->second
.first_tx
== utime_t()) {
4932 dout(25) << "heartbeat_check we haven't sent ping to osd." << p
->first
4933 << "yet, skipping" << dendl
;
4937 dout(25) << "heartbeat_check osd." << p
->first
4938 << " first_tx " << p
->second
.first_tx
4939 << " last_tx " << p
->second
.last_tx
4940 << " last_rx_back " << p
->second
.last_rx_back
4941 << " last_rx_front " << p
->second
.last_rx_front
4943 if (p
->second
.is_unhealthy(cutoff
)) {
4944 if (p
->second
.last_rx_back
== utime_t() ||
4945 p
->second
.last_rx_front
== utime_t()) {
4946 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
4947 << " osd." << p
->first
<< " ever on either front or back, first ping sent "
4948 << p
->second
.first_tx
<< " (cutoff " << cutoff
<< ")" << dendl
;
4950 failure_queue
[p
->first
] = p
->second
.last_tx
;
4952 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
4953 << " osd." << p
->first
<< " since back " << p
->second
.last_rx_back
4954 << " front " << p
->second
.last_rx_front
4955 << " (cutoff " << cutoff
<< ")" << dendl
;
4957 failure_queue
[p
->first
] = MIN(p
->second
.last_rx_back
, p
->second
.last_rx_front
);
4963 void OSD::heartbeat()
4965 dout(30) << "heartbeat" << dendl
;
4969 int n_samples
= 86400 / cct
->_conf
->osd_heartbeat_interval
;
4970 if (getloadavg(loadavgs
, 1) == 1) {
4971 logger
->set(l_osd_loadavg
, 100 * loadavgs
[0]);
4972 daily_loadavg
= (daily_loadavg
* (n_samples
- 1) + loadavgs
[0]) / n_samples
;
4973 dout(30) << "heartbeat: daily_loadavg " << daily_loadavg
<< dendl
;
4976 dout(30) << "heartbeat checking stats" << dendl
;
4979 vector
<int> hb_peers
;
4980 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4981 p
!= heartbeat_peers
.end();
4983 hb_peers
.push_back(p
->first
);
4984 service
.update_osd_stat(hb_peers
);
4986 dout(5) << "heartbeat: " << service
.get_osd_stat() << dendl
;
4988 utime_t now
= ceph_clock_now();
4991 for (map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.begin();
4992 i
!= heartbeat_peers
.end();
4994 int peer
= i
->first
;
4995 i
->second
.last_tx
= now
;
4996 if (i
->second
.first_tx
== utime_t())
4997 i
->second
.first_tx
= now
;
4998 dout(30) << "heartbeat sending ping to osd." << peer
<< dendl
;
4999 i
->second
.con_back
->send_message(new MOSDPing(monc
->get_fsid(),
5000 service
.get_osdmap()->get_epoch(),
5001 MOSDPing::PING
, now
,
5002 cct
->_conf
->osd_heartbeat_min_size
));
5004 if (i
->second
.con_front
)
5005 i
->second
.con_front
->send_message(new MOSDPing(monc
->get_fsid(),
5006 service
.get_osdmap()->get_epoch(),
5007 MOSDPing::PING
, now
,
5008 cct
->_conf
->osd_heartbeat_min_size
));
5011 logger
->set(l_osd_hb_to
, heartbeat_peers
.size());
5013 // hmm.. am i all alone?
5014 dout(30) << "heartbeat lonely?" << dendl
;
5015 if (heartbeat_peers
.empty()) {
5016 if (now
- last_mon_heartbeat
> cct
->_conf
->osd_mon_heartbeat_interval
&& is_active()) {
5017 last_mon_heartbeat
= now
;
5018 dout(10) << "i have no heartbeat peers; checking mon for new map" << dendl
;
5019 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5023 dout(30) << "heartbeat done" << dendl
;
5026 bool OSD::heartbeat_reset(Connection
*con
)
5028 HeartbeatSession
*s
= static_cast<HeartbeatSession
*>(con
->get_priv());
5030 heartbeat_lock
.Lock();
5031 if (is_stopping()) {
5032 heartbeat_lock
.Unlock();
5036 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(s
->peer
);
5037 if (p
!= heartbeat_peers
.end() &&
5038 (p
->second
.con_back
== con
||
5039 p
->second
.con_front
== con
)) {
5040 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5041 << ", reopening" << dendl
;
5042 if (con
!= p
->second
.con_back
) {
5043 p
->second
.con_back
->mark_down();
5045 p
->second
.con_back
.reset(NULL
);
5046 if (p
->second
.con_front
&& con
!= p
->second
.con_front
) {
5047 p
->second
.con_front
->mark_down();
5049 p
->second
.con_front
.reset(NULL
);
5050 pair
<ConnectionRef
,ConnectionRef
> newcon
= service
.get_con_osd_hb(p
->second
.peer
, p
->second
.epoch
);
5052 p
->second
.con_back
= newcon
.first
.get();
5053 p
->second
.con_back
->set_priv(s
->get());
5054 if (newcon
.second
) {
5055 p
->second
.con_front
= newcon
.second
.get();
5056 p
->second
.con_front
->set_priv(s
->get());
5059 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5060 << ", raced with osdmap update, closing out peer" << dendl
;
5061 heartbeat_peers
.erase(p
);
5064 dout(10) << "heartbeat_reset closing (old) failed hb con " << con
<< dendl
;
5066 heartbeat_lock
.Unlock();
5074 // =========================================
5078 assert(osd_lock
.is_locked());
5079 dout(10) << "tick" << dendl
;
5081 if (is_active() || is_waiting_for_healthy()) {
5082 maybe_update_heartbeat_peers();
5085 if (is_waiting_for_healthy()) {
5087 } else if (is_preboot() &&
5088 waiting_for_luminous_mons
&&
5089 monc
->monmap
.get_required_features().contains_all(
5090 ceph::features::mon::FEATURE_LUMINOUS
)) {
5091 // mon upgrade finished!
5097 tick_timer
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick(this));
5100 void OSD::tick_without_osd_lock()
5102 assert(tick_timer_lock
.is_locked());
5103 dout(10) << "tick_without_osd_lock" << dendl
;
5105 logger
->set(l_osd_buf
, buffer::get_total_alloc());
5106 logger
->set(l_osd_history_alloc_bytes
, SHIFT_ROUND_UP(buffer::get_history_alloc_bytes(), 20));
5107 logger
->set(l_osd_history_alloc_num
, buffer::get_history_alloc_num());
5108 logger
->set(l_osd_cached_crc
, buffer::get_cached_crc());
5109 logger
->set(l_osd_cached_crc_adjusted
, buffer::get_cached_crc_adjusted());
5110 logger
->set(l_osd_missed_crc
, buffer::get_missed_crc());
5112 // osd_lock is not being held, which means the OSD state
5113 // might change when doing the monitor report
5114 if (is_active() || is_waiting_for_healthy()) {
5115 heartbeat_lock
.Lock();
5117 heartbeat_lock
.Unlock();
5119 map_lock
.get_read();
5120 Mutex::Locker
l(mon_report_lock
);
5124 bool report
= false;
5125 utime_t now
= ceph_clock_now();
5126 pg_stat_queue_lock
.Lock();
5127 double backoff
= stats_ack_timeout
/ cct
->_conf
->osd_mon_ack_timeout
;
5128 double adjusted_min
= cct
->_conf
->osd_mon_report_interval_min
* backoff
;
5129 // note: we shouldn't adjust max because it must remain < the
5130 // mon's mon_osd_report_timeout (which defaults to 1.5x our
5132 double max
= cct
->_conf
->osd_mon_report_interval_max
;
5133 if (!outstanding_pg_stats
.empty() &&
5134 (now
- stats_ack_timeout
) > last_pg_stats_ack
) {
5135 dout(1) << __func__
<< " mon hasn't acked PGStats in "
5136 << now
- last_pg_stats_ack
5137 << " seconds, reconnecting elsewhere" << dendl
;
5139 last_pg_stats_ack
= now
; // reset clock
5140 last_pg_stats_sent
= utime_t();
5142 MAX(cct
->_conf
->osd_mon_ack_timeout
,
5143 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_factor
);
5144 outstanding_pg_stats
.clear();
5146 if (now
- last_pg_stats_sent
> max
) {
5147 osd_stat_updated
= true;
5149 } else if (service
.need_fullness_update()) {
5151 } else if ((int)outstanding_pg_stats
.size() >=
5152 cct
->_conf
->osd_mon_report_max_in_flight
) {
5153 dout(20) << __func__
<< " have max " << outstanding_pg_stats
5154 << " stats updates in flight" << dendl
;
5156 if (now
- last_mon_report
> adjusted_min
) {
5157 dout(20) << __func__
<< " stats backoff " << backoff
5158 << " adjusted_min " << adjusted_min
<< " - sending report"
5160 osd_stat_updated
= true;
5164 pg_stat_queue_lock
.Unlock();
5167 monc
->reopen_session();
5168 } else if (report
) {
5169 last_mon_report
= now
;
5171 // do any pending reports
5174 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5178 map_lock
.put_read();
5182 if (!scrub_random_backoff()) {
5185 service
.promote_throttle_recalibrate();
5186 bool need_send_beacon
= false;
5187 const auto now
= ceph::coarse_mono_clock::now();
5189 // borrow lec lock to pretect last_sent_beacon from changing
5190 Mutex::Locker l
{min_last_epoch_clean_lock
};
5191 const auto elapsed
= now
- last_sent_beacon
;
5192 if (chrono::duration_cast
<chrono::seconds
>(elapsed
).count() >
5193 cct
->_conf
->osd_beacon_report_interval
) {
5194 need_send_beacon
= true;
5197 if (need_send_beacon
) {
5202 check_ops_in_flight();
5203 service
.kick_recovery_queue();
5204 tick_timer_without_osd_lock
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick_WithoutOSDLock(this));
5207 void OSD::check_ops_in_flight()
5209 vector
<string
> warnings
;
5210 if (op_tracker
.check_ops_in_flight(warnings
)) {
5211 for (vector
<string
>::iterator i
= warnings
.begin();
5212 i
!= warnings
.end();
5220 // setomapval <pool-id> [namespace/]<obj-name> <key> <val>
5221 // rmomapkey <pool-id> [namespace/]<obj-name> <key>
5222 // setomapheader <pool-id> [namespace/]<obj-name> <header>
5223 // getomap <pool> [namespace/]<obj-name>
5224 // truncobj <pool-id> [namespace/]<obj-name> <newlen>
5225 // injectmdataerr [namespace/]<obj-name> [shardid]
5226 // injectdataerr [namespace/]<obj-name> [shardid]
5228 // set_recovery_delay [utime]
5229 void TestOpsSocketHook::test_ops(OSDService
*service
, ObjectStore
*store
,
5230 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
)
5233 //Support changing the omap on a single osd by using the Admin Socket to
5234 //directly request the osd make a change.
5235 if (command
== "setomapval" || command
== "rmomapkey" ||
5236 command
== "setomapheader" || command
== "getomap" ||
5237 command
== "truncobj" || command
== "injectmdataerr" ||
5238 command
== "injectdataerr"
5242 OSDMapRef curmap
= service
->get_osdmap();
5247 cmd_getval(service
->cct
, cmdmap
, "pool", poolstr
);
5248 pool
= curmap
->lookup_pg_pool_name(poolstr
);
5249 //If we can't find it by name then maybe id specified
5250 if (pool
< 0 && isdigit(poolstr
[0]))
5251 pool
= atoll(poolstr
.c_str());
5253 ss
<< "Invalid pool" << poolstr
;
5257 string objname
, nspace
;
5258 cmd_getval(service
->cct
, cmdmap
, "objname", objname
);
5259 std::size_t found
= objname
.find_first_of('/');
5260 if (found
!= string::npos
) {
5261 nspace
= objname
.substr(0, found
);
5262 objname
= objname
.substr(found
+1);
5264 object_locator_t
oloc(pool
, nspace
);
5265 r
= curmap
->object_locator_to_pg(object_t(objname
), oloc
, rawpg
);
5268 ss
<< "Invalid namespace/objname";
5273 cmd_getval(service
->cct
, cmdmap
, "shardid", shardid
, int64_t(shard_id_t::NO_SHARD
));
5274 hobject_t
obj(object_t(objname
), string(""), CEPH_NOSNAP
, rawpg
.ps(), pool
, nspace
);
5275 ghobject_t
gobj(obj
, ghobject_t::NO_GEN
, shard_id_t(uint8_t(shardid
)));
5276 spg_t
pgid(curmap
->raw_pg_to_pg(rawpg
), shard_id_t(shardid
));
5277 if (curmap
->pg_is_ec(rawpg
)) {
5278 if ((command
!= "injectdataerr") && (command
!= "injectmdataerr")) {
5279 ss
<< "Must not call on ec pool, except injectdataerr or injectmdataerr";
5284 ObjectStore::Transaction t
;
5286 if (command
== "setomapval") {
5287 map
<string
, bufferlist
> newattrs
;
5290 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5291 cmd_getval(service
->cct
, cmdmap
, "val", valstr
);
5294 newattrs
[key
] = val
;
5295 t
.omap_setkeys(coll_t(pgid
), ghobject_t(obj
), newattrs
);
5296 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5298 ss
<< "error=" << r
;
5301 } else if (command
== "rmomapkey") {
5304 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5307 t
.omap_rmkeys(coll_t(pgid
), ghobject_t(obj
), keys
);
5308 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5310 ss
<< "error=" << r
;
5313 } else if (command
== "setomapheader") {
5314 bufferlist newheader
;
5317 cmd_getval(service
->cct
, cmdmap
, "header", headerstr
);
5318 newheader
.append(headerstr
);
5319 t
.omap_setheader(coll_t(pgid
), ghobject_t(obj
), newheader
);
5320 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5322 ss
<< "error=" << r
;
5325 } else if (command
== "getomap") {
5326 //Debug: Output entire omap
5328 map
<string
, bufferlist
> keyvals
;
5329 r
= store
->omap_get(coll_t(pgid
), ghobject_t(obj
), &hdrbl
, &keyvals
);
5331 ss
<< "header=" << string(hdrbl
.c_str(), hdrbl
.length());
5332 for (map
<string
, bufferlist
>::iterator it
= keyvals
.begin();
5333 it
!= keyvals
.end(); ++it
)
5334 ss
<< " key=" << (*it
).first
<< " val="
5335 << string((*it
).second
.c_str(), (*it
).second
.length());
5337 ss
<< "error=" << r
;
5339 } else if (command
== "truncobj") {
5341 cmd_getval(service
->cct
, cmdmap
, "len", trunclen
);
5342 t
.truncate(coll_t(pgid
), ghobject_t(obj
), trunclen
);
5343 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5345 ss
<< "error=" << r
;
5348 } else if (command
== "injectdataerr") {
5349 store
->inject_data_error(gobj
);
5351 } else if (command
== "injectmdataerr") {
5352 store
->inject_mdata_error(gobj
);
5357 if (command
== "set_recovery_delay") {
5359 cmd_getval(service
->cct
, cmdmap
, "utime", delay
, (int64_t)0);
5362 int r
= service
->cct
->_conf
->set_val("osd_recovery_delay_start",
5365 ss
<< "set_recovery_delay: error setting "
5366 << "osd_recovery_delay_start to '" << delay
<< "': error "
5370 service
->cct
->_conf
->apply_changes(NULL
);
5371 ss
<< "set_recovery_delay: set osd_recovery_delay_start "
5372 << "to " << service
->cct
->_conf
->osd_recovery_delay_start
;
5375 if (command
== "trigger_scrub") {
5377 OSDMapRef curmap
= service
->get_osdmap();
5381 cmd_getval(service
->cct
, cmdmap
, "pgid", pgidstr
);
5382 if (!pgid
.parse(pgidstr
.c_str())) {
5383 ss
<< "Invalid pgid specified";
5387 PG
*pg
= service
->osd
->_lookup_lock_pg(pgid
);
5388 if (pg
== nullptr) {
5389 ss
<< "Can't find pg " << pgid
;
5393 if (pg
->is_primary()) {
5394 pg
->unreg_next_scrub();
5395 const pg_pool_t
*p
= curmap
->get_pg_pool(pgid
.pool());
5396 double pool_scrub_max_interval
= 0;
5397 p
->opts
.get(pool_opts_t::SCRUB_MAX_INTERVAL
, &pool_scrub_max_interval
);
5398 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
5399 pool_scrub_max_interval
: g_conf
->osd_scrub_max_interval
;
5400 // Instead of marking must_scrub force a schedule scrub
5401 utime_t stamp
= ceph_clock_now();
5402 stamp
-= scrub_max_interval
;
5403 stamp
-= 100.0; // push back last scrub more for good measure
5404 pg
->info
.history
.last_scrub_stamp
= stamp
;
5405 pg
->reg_next_scrub();
5408 ss
<< "Not primary";
5413 if (command
== "injectfull") {
5416 OSDService::s_names state
;
5417 cmd_getval(service
->cct
, cmdmap
, "type", type
, string("full"));
5418 cmd_getval(service
->cct
, cmdmap
, "count", count
, (int64_t)-1);
5419 if (type
== "none" || count
== 0) {
5423 state
= service
->get_full_state(type
);
5424 if (state
== OSDService::s_names::INVALID
) {
5425 ss
<< "Invalid type use (none, nearfull, backfillfull, full, failsafe)";
5428 service
->set_injectfull(state
, count
);
5431 ss
<< "Internal error - command=" << command
;
5434 // =========================================
5437 ObjectStore
*store
, SnapMapper
*mapper
,
5439 ObjectStore::Sequencer
*osr
,
5440 coll_t coll
, DeletingStateRef dstate
,
5442 ThreadPool::TPHandle
&handle
)
5444 vector
<ghobject_t
> olist
;
5446 ObjectStore::Transaction t
;
5448 handle
.reset_tp_timeout();
5449 store
->collection_list(
5452 ghobject_t::get_max(),
5453 store
->get_ideal_list_max(),
5456 generic_dout(10) << __func__
<< " " << olist
<< dendl
;
5457 // default cont to true, this is safe because caller(OSD::RemoveWQ::_process())
5458 // will recheck the answer before it really goes on.
5460 for (vector
<ghobject_t
>::iterator i
= olist
.begin();
5465 OSDriver::OSTransaction
_t(osdriver
->get_transaction(&t
));
5466 int r
= mapper
->remove_oid(i
->hobj
, &_t
);
5467 if (r
!= 0 && r
!= -ENOENT
) {
5471 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
5473 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5474 cont
= dstate
->pause_clearing();
5475 handle
.suspend_tp_timeout();
5477 handle
.reset_tp_timeout();
5479 cont
= dstate
->resume_clearing();
5482 t
= ObjectStore::Transaction();
5488 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5489 cont
= dstate
->pause_clearing();
5490 handle
.suspend_tp_timeout();
5492 handle
.reset_tp_timeout();
5494 cont
= dstate
->resume_clearing();
5496 // whether there are more objects to remove in the collection
5497 *finished
= next
.is_max();
5501 void OSD::RemoveWQ::_process(
5502 pair
<PGRef
, DeletingStateRef
> item
,
5503 ThreadPool::TPHandle
&handle
)
5506 PGRef
pg(item
.first
);
5507 SnapMapper
&mapper
= pg
->snap_mapper
;
5508 OSDriver
&driver
= pg
->osdriver
;
5509 coll_t coll
= coll_t(pg
->info
.pgid
);
5511 bool finished
= false;
5513 if (!item
.second
->start_or_resume_clearing())
5516 bool cont
= remove_dir(
5517 pg
->cct
, store
, &mapper
, &driver
, pg
->osr
.get(), coll
, item
.second
,
5522 if (item
.second
->pause_clearing())
5527 if (!item
.second
->start_deleting())
5530 ObjectStore::Transaction t
;
5531 PGLog::clear_info_log(pg
->info
.pgid
, &t
);
5533 if (cct
->_conf
->osd_inject_failure_on_pg_removal
) {
5534 generic_derr
<< "osd_inject_failure_on_pg_removal" << dendl
;
5537 t
.remove_collection(coll
);
5539 // We need the sequencer to stick around until the op is complete
5540 store
->queue_transaction(
5545 0, // onreadable sync
5546 new ContainerContext
<PGRef
>(pg
),
5549 item
.second
->finish_deleting();
5551 // =========================================
5553 void OSD::ms_handle_connect(Connection
*con
)
5555 dout(10) << __func__
<< " con " << con
<< dendl
;
5556 if (con
->get_peer_type() == CEPH_ENTITY_TYPE_MON
) {
5557 Mutex::Locker
l(osd_lock
);
5560 dout(10) << __func__
<< " on mon" << dendl
;
5564 } else if (is_booting()) {
5565 _send_boot(); // resend boot message
5567 map_lock
.get_read();
5568 Mutex::Locker
l2(mon_report_lock
);
5570 utime_t now
= ceph_clock_now();
5571 last_mon_report
= now
;
5573 // resend everything, it's a new session
5576 service
.requeue_pg_temp();
5577 service
.send_pg_temp();
5580 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5584 map_lock
.put_read();
5586 send_beacon(ceph::coarse_mono_clock::now());
5590 // full map requests may happen while active or pre-boot
5591 if (requested_full_first
) {
5592 rerequest_full_maps();
5597 void OSD::ms_handle_fast_connect(Connection
*con
)
5599 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5600 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5601 Session
*s
= static_cast<Session
*>(con
->get_priv());
5603 s
= new Session(cct
);
5604 con
->set_priv(s
->get());
5606 dout(10) << " new session (outgoing) " << s
<< " con=" << s
->con
5607 << " addr=" << s
->con
->get_peer_addr() << dendl
;
5608 // we don't connect to clients
5609 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5610 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5616 void OSD::ms_handle_fast_accept(Connection
*con
)
5618 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5619 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5620 Session
*s
= static_cast<Session
*>(con
->get_priv());
5622 s
= new Session(cct
);
5623 con
->set_priv(s
->get());
5625 dout(10) << "new session (incoming)" << s
<< " con=" << con
5626 << " addr=" << con
->get_peer_addr()
5627 << " must have raced with connect" << dendl
;
5628 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5629 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5635 bool OSD::ms_handle_reset(Connection
*con
)
5637 Session
*session
= static_cast<Session
*>(con
->get_priv());
5638 dout(2) << "ms_handle_reset con " << con
<< " session " << session
<< dendl
;
5641 session
->wstate
.reset(con
);
5642 session
->con
.reset(NULL
); // break con <-> session ref cycle
5643 // note that we break session->con *before* the session_handle_reset
5644 // cleanup below. this avoids a race between us and
5645 // PG::add_backoff, Session::check_backoff, etc.
5646 session_handle_reset(session
);
5651 bool OSD::ms_handle_refused(Connection
*con
)
5653 if (!cct
->_conf
->osd_fast_fail_on_connection_refused
)
5656 Session
*session
= static_cast<Session
*>(con
->get_priv());
5657 dout(2) << "ms_handle_refused con " << con
<< " session " << session
<< dendl
;
5660 int type
= con
->get_peer_type();
5661 // handle only OSD failures here
5662 if (monc
&& (type
== CEPH_ENTITY_TYPE_OSD
)) {
5663 OSDMapRef osdmap
= get_osdmap();
5665 int id
= osdmap
->identify_osd_on_all_channels(con
->get_peer_addr());
5666 if (id
>= 0 && osdmap
->is_up(id
)) {
5667 // I'm cheating mon heartbeat grace logic, because we know it's not going
5668 // to respawn alone. +1 so we won't hit any boundary case.
5669 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(),
5670 osdmap
->get_inst(id
),
5671 cct
->_conf
->osd_heartbeat_grace
+ 1,
5672 osdmap
->get_epoch(),
5673 MOSDFailure::FLAG_IMMEDIATE
| MOSDFailure::FLAG_FAILED
5682 struct C_OSD_GetVersion
: public Context
{
5684 uint64_t oldest
, newest
;
5685 explicit C_OSD_GetVersion(OSD
*o
) : osd(o
), oldest(0), newest(0) {}
5686 void finish(int r
) override
{
5688 osd
->_got_mon_epochs(oldest
, newest
);
5692 void OSD::start_boot()
5694 if (!_is_healthy()) {
5695 // if we are not healthy, do not mark ourselves up (yet)
5696 dout(1) << "not healthy; waiting to boot" << dendl
;
5697 if (!is_waiting_for_healthy())
5698 start_waiting_for_healthy();
5699 // send pings sooner rather than later
5703 dout(1) << __func__
<< dendl
;
5704 set_state(STATE_PREBOOT
);
5705 waiting_for_luminous_mons
= false;
5706 dout(10) << "start_boot - have maps " << superblock
.oldest_map
5707 << ".." << superblock
.newest_map
<< dendl
;
5708 C_OSD_GetVersion
*c
= new C_OSD_GetVersion(this);
5709 monc
->get_version("osdmap", &c
->newest
, &c
->oldest
, c
);
5712 void OSD::_got_mon_epochs(epoch_t oldest
, epoch_t newest
)
5714 Mutex::Locker
l(osd_lock
);
5716 _preboot(oldest
, newest
);
5720 void OSD::_preboot(epoch_t oldest
, epoch_t newest
)
5722 assert(is_preboot());
5723 dout(10) << __func__
<< " _preboot mon has osdmaps "
5724 << oldest
<< ".." << newest
<< dendl
;
5726 // ensure our local fullness awareness is accurate
5729 // if our map within recent history, try to add ourselves to the osdmap.
5730 if (osdmap
->get_epoch() == 0) {
5731 derr
<< "waiting for initial osdmap" << dendl
;
5732 } else if (osdmap
->is_destroyed(whoami
)) {
5733 derr
<< "osdmap says I am destroyed, exiting" << dendl
;
5735 } else if (osdmap
->test_flag(CEPH_OSDMAP_NOUP
) || osdmap
->is_noup(whoami
)) {
5736 derr
<< "osdmap NOUP flag is set, waiting for it to clear" << dendl
;
5737 } else if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
)) {
5738 derr
<< "osdmap SORTBITWISE OSDMap flag is NOT set; please set it"
5740 } else if (osdmap
->require_osd_release
< CEPH_RELEASE_JEWEL
) {
5741 derr
<< "osdmap REQUIRE_JEWEL OSDMap flag is NOT set; please set it"
5743 } else if (!monc
->monmap
.get_required_features().contains_all(
5744 ceph::features::mon::FEATURE_LUMINOUS
)) {
5745 derr
<< "monmap REQUIRE_LUMINOUS is NOT set; must upgrade all monitors to "
5746 << "Luminous or later before Luminous OSDs will boot" << dendl
;
5747 waiting_for_luminous_mons
= true;
5748 } else if (service
.need_fullness_update()) {
5749 derr
<< "osdmap fullness state needs update" << dendl
;
5751 } else if (osdmap
->get_epoch() >= oldest
- 1 &&
5752 osdmap
->get_epoch() + cct
->_conf
->osd_map_message_max
> newest
) {
5757 // get all the latest maps
5758 if (osdmap
->get_epoch() + 1 >= oldest
)
5759 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5761 osdmap_subscribe(oldest
- 1, true);
5764 void OSD::send_full_update()
5766 if (!service
.need_fullness_update())
5769 if (service
.is_full()) {
5770 state
= CEPH_OSD_FULL
;
5771 } else if (service
.is_backfillfull()) {
5772 state
= CEPH_OSD_BACKFILLFULL
;
5773 } else if (service
.is_nearfull()) {
5774 state
= CEPH_OSD_NEARFULL
;
5777 OSDMap::calc_state_set(state
, s
);
5778 dout(10) << __func__
<< " want state " << s
<< dendl
;
5779 monc
->send_mon_message(new MOSDFull(osdmap
->get_epoch(), state
));
5782 void OSD::start_waiting_for_healthy()
5784 dout(1) << "start_waiting_for_healthy" << dendl
;
5785 set_state(STATE_WAITING_FOR_HEALTHY
);
5786 last_heartbeat_resample
= utime_t();
5789 bool OSD::_is_healthy()
5791 if (!cct
->get_heartbeat_map()->is_healthy()) {
5792 dout(1) << "is_healthy false -- internal heartbeat failed" << dendl
;
5796 if (is_waiting_for_healthy()) {
5797 Mutex::Locker
l(heartbeat_lock
);
5798 utime_t cutoff
= ceph_clock_now();
5799 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
5800 int num
= 0, up
= 0;
5801 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
5802 p
!= heartbeat_peers
.end();
5804 if (p
->second
.is_healthy(cutoff
))
5808 if ((float)up
< (float)num
* cct
->_conf
->osd_heartbeat_min_healthy_ratio
) {
5809 dout(1) << "is_healthy false -- only " << up
<< "/" << num
<< " up peers (less than "
5810 << int(cct
->_conf
->osd_heartbeat_min_healthy_ratio
* 100.0) << "%)" << dendl
;
5818 void OSD::_send_boot()
5820 dout(10) << "_send_boot" << dendl
;
5821 entity_addr_t cluster_addr
= cluster_messenger
->get_myaddr();
5822 Connection
*local_connection
= cluster_messenger
->get_loopback_connection().get();
5823 if (cluster_addr
.is_blank_ip()) {
5824 int port
= cluster_addr
.get_port();
5825 cluster_addr
= client_messenger
->get_myaddr();
5826 cluster_addr
.set_port(port
);
5827 cluster_messenger
->set_addr_unknowns(cluster_addr
);
5828 dout(10) << " assuming cluster_addr ip matches client_addr" << dendl
;
5830 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
5834 cluster_messenger
->ms_deliver_handle_fast_connect(local_connection
);
5837 entity_addr_t hb_back_addr
= hb_back_server_messenger
->get_myaddr();
5838 local_connection
= hb_back_server_messenger
->get_loopback_connection().get();
5839 if (hb_back_addr
.is_blank_ip()) {
5840 int port
= hb_back_addr
.get_port();
5841 hb_back_addr
= cluster_addr
;
5842 hb_back_addr
.set_port(port
);
5843 hb_back_server_messenger
->set_addr_unknowns(hb_back_addr
);
5844 dout(10) << " assuming hb_back_addr ip matches cluster_addr" << dendl
;
5846 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
5850 hb_back_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
5853 entity_addr_t hb_front_addr
= hb_front_server_messenger
->get_myaddr();
5854 local_connection
= hb_front_server_messenger
->get_loopback_connection().get();
5855 if (hb_front_addr
.is_blank_ip()) {
5856 int port
= hb_front_addr
.get_port();
5857 hb_front_addr
= client_messenger
->get_myaddr();
5858 hb_front_addr
.set_port(port
);
5859 hb_front_server_messenger
->set_addr_unknowns(hb_front_addr
);
5860 dout(10) << " assuming hb_front_addr ip matches client_addr" << dendl
;
5862 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
5866 hb_front_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
5869 MOSDBoot
*mboot
= new MOSDBoot(superblock
, get_osdmap_epoch(), service
.get_boot_epoch(),
5870 hb_back_addr
, hb_front_addr
, cluster_addr
,
5872 dout(10) << " client_addr " << client_messenger
->get_myaddr()
5873 << ", cluster_addr " << cluster_addr
5874 << ", hb_back_addr " << hb_back_addr
5875 << ", hb_front_addr " << hb_front_addr
5877 _collect_metadata(&mboot
->metadata
);
5878 monc
->send_mon_message(mboot
);
5879 set_state(STATE_BOOTING
);
5882 void OSD::_collect_metadata(map
<string
,string
> *pm
)
5885 (*pm
)["osd_data"] = dev_path
;
5886 if (store
->get_type() == "filestore") {
5887 // not applicable for bluestore
5888 (*pm
)["osd_journal"] = journal_path
;
5890 (*pm
)["front_addr"] = stringify(client_messenger
->get_myaddr());
5891 (*pm
)["back_addr"] = stringify(cluster_messenger
->get_myaddr());
5892 (*pm
)["hb_front_addr"] = stringify(hb_front_server_messenger
->get_myaddr());
5893 (*pm
)["hb_back_addr"] = stringify(hb_back_server_messenger
->get_myaddr());
5896 (*pm
)["osd_objectstore"] = store
->get_type();
5897 (*pm
)["rotational"] = store_is_rotational
? "1" : "0";
5898 (*pm
)["journal_rotational"] = journal_is_rotational
? "1" : "0";
5899 (*pm
)["default_device_class"] = store
->get_default_device_class();
5900 store
->collect_metadata(pm
);
5902 collect_sys_info(pm
, cct
);
5904 dout(10) << __func__
<< " " << *pm
<< dendl
;
5907 void OSD::queue_want_up_thru(epoch_t want
)
5909 map_lock
.get_read();
5910 epoch_t cur
= osdmap
->get_up_thru(whoami
);
5911 Mutex::Locker
l(mon_report_lock
);
5912 if (want
> up_thru_wanted
) {
5913 dout(10) << "queue_want_up_thru now " << want
<< " (was " << up_thru_wanted
<< ")"
5914 << ", currently " << cur
5916 up_thru_wanted
= want
;
5919 dout(10) << "queue_want_up_thru want " << want
<< " <= queued " << up_thru_wanted
5920 << ", currently " << cur
5923 map_lock
.put_read();
5926 void OSD::send_alive()
5928 assert(mon_report_lock
.is_locked());
5929 if (!osdmap
->exists(whoami
))
5931 epoch_t up_thru
= osdmap
->get_up_thru(whoami
);
5932 dout(10) << "send_alive up_thru currently " << up_thru
<< " want " << up_thru_wanted
<< dendl
;
5933 if (up_thru_wanted
> up_thru
) {
5934 dout(10) << "send_alive want " << up_thru_wanted
<< dendl
;
5935 monc
->send_mon_message(new MOSDAlive(osdmap
->get_epoch(), up_thru_wanted
));
5939 void OSD::request_full_map(epoch_t first
, epoch_t last
)
5941 dout(10) << __func__
<< " " << first
<< ".." << last
5942 << ", previously requested "
5943 << requested_full_first
<< ".." << requested_full_last
<< dendl
;
5944 assert(osd_lock
.is_locked());
5945 assert(first
> 0 && last
> 0);
5946 assert(first
<= last
);
5947 assert(first
>= requested_full_first
); // we shouldn't ever ask for older maps
5948 if (requested_full_first
== 0) {
5950 requested_full_first
= first
;
5951 requested_full_last
= last
;
5952 } else if (last
<= requested_full_last
) {
5956 // additional request
5957 first
= requested_full_last
+ 1;
5958 requested_full_last
= last
;
5960 MMonGetOSDMap
*req
= new MMonGetOSDMap
;
5961 req
->request_full(first
, last
);
5962 monc
->send_mon_message(req
);
5965 void OSD::got_full_map(epoch_t e
)
5967 assert(requested_full_first
<= requested_full_last
);
5968 assert(osd_lock
.is_locked());
5969 if (requested_full_first
== 0) {
5970 dout(20) << __func__
<< " " << e
<< ", nothing requested" << dendl
;
5973 if (e
< requested_full_first
) {
5974 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
5975 << ".." << requested_full_last
5976 << ", ignoring" << dendl
;
5979 if (e
>= requested_full_last
) {
5980 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
5981 << ".." << requested_full_last
<< ", resetting" << dendl
;
5982 requested_full_first
= requested_full_last
= 0;
5986 requested_full_first
= e
+ 1;
5988 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
5989 << ".." << requested_full_last
5990 << ", still need more" << dendl
;
5993 void OSD::requeue_failures()
5995 Mutex::Locker
l(heartbeat_lock
);
5996 unsigned old_queue
= failure_queue
.size();
5997 unsigned old_pending
= failure_pending
.size();
5998 for (map
<int,pair
<utime_t
,entity_inst_t
> >::iterator p
=
5999 failure_pending
.begin();
6000 p
!= failure_pending
.end(); ) {
6001 failure_queue
[p
->first
] = p
->second
.first
;
6002 failure_pending
.erase(p
++);
6004 dout(10) << __func__
<< " " << old_queue
<< " + " << old_pending
<< " -> "
6005 << failure_queue
.size() << dendl
;
6008 void OSD::send_failures()
6010 assert(map_lock
.is_locked());
6011 assert(mon_report_lock
.is_locked());
6012 Mutex::Locker
l(heartbeat_lock
);
6013 utime_t now
= ceph_clock_now();
6014 while (!failure_queue
.empty()) {
6015 int osd
= failure_queue
.begin()->first
;
6016 if (!failure_pending
.count(osd
)) {
6017 entity_inst_t i
= osdmap
->get_inst(osd
);
6018 int failed_for
= (int)(double)(now
- failure_queue
.begin()->second
);
6019 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(), i
, failed_for
,
6020 osdmap
->get_epoch()));
6021 failure_pending
[osd
] = make_pair(failure_queue
.begin()->second
, i
);
6023 failure_queue
.erase(osd
);
6027 void OSD::send_still_alive(epoch_t epoch
, const entity_inst_t
&i
)
6029 MOSDFailure
*m
= new MOSDFailure(monc
->get_fsid(), i
, 0, epoch
, MOSDFailure::FLAG_ALIVE
);
6030 monc
->send_mon_message(m
);
6033 void OSD::send_pg_stats(const utime_t
&now
)
6035 assert(map_lock
.is_locked());
6036 assert(osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
);
6037 dout(20) << "send_pg_stats" << dendl
;
6039 osd_stat_t cur_stat
= service
.get_osd_stat();
6041 cur_stat
.os_perf_stat
= store
->get_cur_stats();
6043 pg_stat_queue_lock
.Lock();
6045 if (osd_stat_updated
|| !pg_stat_queue
.empty()) {
6046 last_pg_stats_sent
= now
;
6047 osd_stat_updated
= false;
6049 dout(10) << "send_pg_stats - " << pg_stat_queue
.size() << " pgs updated" << dendl
;
6051 utime_t
had_for(now
);
6052 had_for
-= had_map_since
;
6054 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
6056 uint64_t tid
= ++pg_stat_tid
;
6058 m
->osd_stat
= cur_stat
;
6060 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6064 if (!pg
->is_primary()) { // we hold map_lock; role is stable.
6065 pg
->stat_queue_item
.remove_myself();
6066 pg
->put("pg_stat_queue");
6069 pg
->pg_stats_publish_lock
.Lock();
6070 if (pg
->pg_stats_publish_valid
) {
6071 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
6072 dout(25) << " sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6073 << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6075 dout(25) << " NOT sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6076 << pg
->pg_stats_publish
.reported_seq
<< ", not valid" << dendl
;
6078 pg
->pg_stats_publish_lock
.Unlock();
6081 if (last_pg_stats_ack
== utime_t() || !outstanding_pg_stats
.empty()) {
6082 last_pg_stats_ack
= ceph_clock_now();
6084 outstanding_pg_stats
.insert(tid
);
6085 dout(20) << __func__
<< " updates pending: " << outstanding_pg_stats
<< dendl
;
6087 monc
->send_mon_message(m
);
6090 pg_stat_queue_lock
.Unlock();
6093 void OSD::handle_pg_stats_ack(MPGStatsAck
*ack
)
6095 dout(10) << "handle_pg_stats_ack " << dendl
;
6097 if (!require_mon_peer(ack
)) {
6102 // NOTE: we may get replies from a previous mon even while
6103 // outstanding_pg_stats is empty if reconnecting races with replies
6106 pg_stat_queue_lock
.Lock();
6108 last_pg_stats_ack
= ceph_clock_now();
6110 // decay timeout slowly (analogous to TCP)
6112 MAX(cct
->_conf
->osd_mon_ack_timeout
,
6113 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_decay
);
6114 dout(20) << __func__
<< " timeout now " << stats_ack_timeout
<< dendl
;
6116 if (ack
->get_tid() > pg_stat_tid_flushed
) {
6117 pg_stat_tid_flushed
= ack
->get_tid();
6118 pg_stat_queue_cond
.Signal();
6121 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6127 auto acked
= ack
->pg_stat
.find(pg
->info
.pgid
.pgid
);
6128 if (acked
!= ack
->pg_stat
.end()) {
6129 pg
->pg_stats_publish_lock
.Lock();
6130 if (acked
->second
.first
== pg
->pg_stats_publish
.reported_seq
&&
6131 acked
->second
.second
== pg
->pg_stats_publish
.reported_epoch
) {
6132 dout(25) << " ack on " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6133 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6134 pg
->stat_queue_item
.remove_myself();
6135 pg
->put("pg_stat_queue");
6137 dout(25) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6138 << ":" << pg
->pg_stats_publish
.reported_seq
<< " > acked "
6139 << acked
->second
<< dendl
;
6141 pg
->pg_stats_publish_lock
.Unlock();
6143 dout(30) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6144 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6148 outstanding_pg_stats
.erase(ack
->get_tid());
6149 dout(20) << __func__
<< " still pending: " << outstanding_pg_stats
<< dendl
;
6151 pg_stat_queue_lock
.Unlock();
6156 void OSD::flush_pg_stats()
6158 dout(10) << "flush_pg_stats" << dendl
;
6160 utime_t now
= ceph_clock_now();
6161 map_lock
.get_read();
6162 mon_report_lock
.Lock();
6164 mon_report_lock
.Unlock();
6165 map_lock
.put_read();
6168 pg_stat_queue_lock
.Lock();
6169 uint64_t tid
= pg_stat_tid
;
6170 dout(10) << "flush_pg_stats waiting for stats tid " << tid
<< " to flush" << dendl
;
6171 while (tid
> pg_stat_tid_flushed
)
6172 pg_stat_queue_cond
.Wait(pg_stat_queue_lock
);
6173 dout(10) << "flush_pg_stats finished waiting for stats tid " << tid
<< " to flush" << dendl
;
6174 pg_stat_queue_lock
.Unlock();
6179 void OSD::send_beacon(const ceph::coarse_mono_clock::time_point
& now
)
6181 const auto& monmap
= monc
->monmap
;
6182 // send beacon to mon even if we are just connected, and the monmap is not
6183 // initialized yet by then.
6184 if (monmap
.epoch
> 0 &&
6185 monmap
.get_required_features().contains_all(
6186 ceph::features::mon::FEATURE_LUMINOUS
)) {
6187 dout(20) << __func__
<< " sending" << dendl
;
6188 MOSDBeacon
* beacon
= nullptr;
6190 Mutex::Locker l
{min_last_epoch_clean_lock
};
6191 beacon
= new MOSDBeacon(osdmap
->get_epoch(), min_last_epoch_clean
);
6192 std::swap(beacon
->pgs
, min_last_epoch_clean_pgs
);
6193 last_sent_beacon
= now
;
6195 monc
->send_mon_message(beacon
);
6197 dout(20) << __func__
<< " not sending" << dendl
;
6201 void OSD::handle_command(MMonCommand
*m
)
6203 if (!require_mon_peer(m
)) {
6208 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), NULL
);
6209 command_wq
.queue(c
);
6213 void OSD::handle_command(MCommand
*m
)
6215 ConnectionRef con
= m
->get_connection();
6216 Session
*session
= static_cast<Session
*>(con
->get_priv());
6218 con
->send_message(new MCommandReply(m
, -EPERM
));
6223 OSDCap
& caps
= session
->caps
;
6226 if (!caps
.allow_all() || m
->get_source().is_mon()) {
6227 con
->send_message(new MCommandReply(m
, -EPERM
));
6232 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), con
.get());
6233 command_wq
.queue(c
);
6243 string availability
;
6244 } osd_commands
[] = {
6246 #define COMMAND(parsesig, helptext, module, perm, availability) \
6247 {parsesig, helptext, module, perm, availability},
6249 // yes, these are really pg commands, but there's a limit to how
6250 // much work it's worth. The OSD returns all of them. Make this
6251 // form (pg <pgid> <cmd>) valid only for the cli.
6252 // Rest uses "tell <pgid> <cmd>"
6255 "name=pgid,type=CephPgid " \
6256 "name=cmd,type=CephChoices,strings=query", \
6257 "show details of a specific pg", "osd", "r", "cli")
6259 "name=pgid,type=CephPgid " \
6260 "name=cmd,type=CephChoices,strings=mark_unfound_lost " \
6261 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6262 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6265 "name=pgid,type=CephPgid " \
6266 "name=cmd,type=CephChoices,strings=list_missing " \
6267 "name=offset,type=CephString,req=false",
6268 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6271 // new form: tell <pgid> <cmd> for both cli and rest
6274 "show details of a specific pg", "osd", "r", "cli,rest")
6275 COMMAND("mark_unfound_lost " \
6276 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6277 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6278 "osd", "rw", "cli,rest")
6279 COMMAND("list_missing " \
6280 "name=offset,type=CephString,req=false",
6281 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6282 "osd", "r", "cli,rest")
6283 COMMAND("perf histogram dump "
6284 "name=logger,type=CephString,req=false "
6285 "name=counter,type=CephString,req=false",
6286 "Get histogram data",
6287 "osd", "r", "cli,rest")
6289 // tell <osd.n> commands. Validation of osd.n must be special-cased in client
6290 COMMAND("version", "report version of OSD", "osd", "r", "cli,rest")
6291 COMMAND("get_command_descriptions", "list commands descriptions", "osd", "r", "cli,rest")
6292 COMMAND("injectargs " \
6293 "name=injected_args,type=CephString,n=N",
6294 "inject configuration arguments into running OSD",
6295 "osd", "rw", "cli,rest")
6296 COMMAND("config set " \
6297 "name=key,type=CephString name=value,type=CephString",
6298 "Set a configuration option at runtime (not persistent)",
6299 "osd", "rw", "cli,rest")
6300 COMMAND("cluster_log " \
6301 "name=level,type=CephChoices,strings=error,warning,info,debug " \
6302 "name=message,type=CephString,n=N",
6303 "log a message to the cluster log",
6304 "osd", "rw", "cli,rest")
6306 "name=count,type=CephInt,req=false " \
6307 "name=size,type=CephInt,req=false " \
6308 "name=object_size,type=CephInt,req=false " \
6309 "name=object_num,type=CephInt,req=false ", \
6310 "OSD benchmark: write <count> <size>-byte objects, " \
6311 "(default 1G size 4MB). Results in log.",
6312 "osd", "rw", "cli,rest")
6313 COMMAND("flush_pg_stats", "flush pg stats", "osd", "rw", "cli,rest")
6315 "name=heapcmd,type=CephChoices,strings=dump|start_profiler|stop_profiler|release|stats", \
6316 "show heap usage info (available only if compiled with tcmalloc)", \
6317 "osd", "rw", "cli,rest")
6318 COMMAND("debug dump_missing " \
6319 "name=filename,type=CephFilepath",
6320 "dump missing objects to a named file", "osd", "r", "cli,rest")
6321 COMMAND("debug kick_recovery_wq " \
6322 "name=delay,type=CephInt,range=0",
6323 "set osd_recovery_delay_start to <val>", "osd", "rw", "cli,rest")
6324 COMMAND("cpu_profiler " \
6325 "name=arg,type=CephChoices,strings=status|flush",
6326 "run cpu profiling on daemon", "osd", "rw", "cli,rest")
6327 COMMAND("dump_pg_recovery_stats", "dump pg recovery statistics",
6328 "osd", "r", "cli,rest")
6329 COMMAND("reset_pg_recovery_stats", "reset pg recovery statistics",
6330 "osd", "rw", "cli,rest")
6332 "compact object store's omap. "
6333 "WARNING: Compaction probably slows your requests",
6334 "osd", "rw", "cli,rest")
6337 void OSD::do_command(Connection
*con
, ceph_tid_t tid
, vector
<string
>& cmd
, bufferlist
& data
)
6340 stringstream ss
, ds
;
6344 dout(20) << "do_command tid " << tid
<< " " << cmd
<< dendl
;
6346 map
<string
, cmd_vartype
> cmdmap
;
6350 boost::scoped_ptr
<Formatter
> f
;
6353 ss
<< "no command given";
6357 if (!cmdmap_from_json(cmd
, &cmdmap
, ss
)) {
6362 cmd_getval(cct
, cmdmap
, "prefix", prefix
);
6364 if (prefix
== "get_command_descriptions") {
6366 JSONFormatter
*f
= new JSONFormatter();
6367 f
->open_object_section("command_descriptions");
6368 for (OSDCommand
*cp
= osd_commands
;
6369 cp
< &osd_commands
[ARRAY_SIZE(osd_commands
)]; cp
++) {
6371 ostringstream secname
;
6372 secname
<< "cmd" << setfill('0') << std::setw(3) << cmdnum
;
6373 dump_cmddesc_to_json(f
, secname
.str(), cp
->cmdstring
, cp
->helpstring
,
6374 cp
->module
, cp
->perm
, cp
->availability
, 0);
6377 f
->close_section(); // command_descriptions
6384 cmd_getval(cct
, cmdmap
, "format", format
);
6385 f
.reset(Formatter::create(format
));
6387 if (prefix
== "version") {
6389 f
->open_object_section("version");
6390 f
->dump_string("version", pretty_version_to_str());
6394 ds
<< pretty_version_to_str();
6398 else if (prefix
== "injectargs") {
6399 vector
<string
> argsvec
;
6400 cmd_getval(cct
, cmdmap
, "injected_args", argsvec
);
6402 if (argsvec
.empty()) {
6404 ss
<< "ignoring empty injectargs";
6407 string args
= argsvec
.front();
6408 for (vector
<string
>::iterator a
= ++argsvec
.begin(); a
!= argsvec
.end(); ++a
)
6411 r
= cct
->_conf
->injectargs(args
, &ss
);
6414 else if (prefix
== "config set") {
6417 cmd_getval(cct
, cmdmap
, "key", key
);
6418 cmd_getval(cct
, cmdmap
, "value", val
);
6420 r
= cct
->_conf
->set_val(key
, val
, true, &ss
);
6422 cct
->_conf
->apply_changes(nullptr);
6426 else if (prefix
== "cluster_log") {
6428 cmd_getval(cct
, cmdmap
, "message", msg
);
6431 ss
<< "ignoring empty log message";
6434 string message
= msg
.front();
6435 for (vector
<string
>::iterator a
= ++msg
.begin(); a
!= msg
.end(); ++a
)
6436 message
+= " " + *a
;
6438 cmd_getval(cct
, cmdmap
, "level", lvl
);
6439 clog_type level
= string_to_clog_type(lvl
);
6442 ss
<< "unknown level '" << lvl
<< "'";
6445 clog
->do_log(level
, message
);
6448 // either 'pg <pgid> <command>' or
6449 // 'tell <pgid>' (which comes in without any of that prefix)?
6451 else if (prefix
== "pg" ||
6452 prefix
== "query" ||
6453 prefix
== "mark_unfound_lost" ||
6454 prefix
== "list_missing"
6458 if (!cmd_getval(cct
, cmdmap
, "pgid", pgidstr
)) {
6459 ss
<< "no pgid specified";
6461 } else if (!pgid
.parse(pgidstr
.c_str())) {
6462 ss
<< "couldn't parse pgid '" << pgidstr
<< "'";
6467 if (osdmap
->get_primary_shard(pgid
, &pcand
) &&
6468 (pg
= _lookup_lock_pg(pcand
))) {
6469 if (pg
->is_primary()) {
6470 // simulate pg <pgid> cmd= for pg->do-command
6472 cmd_putval(cct
, cmdmap
, "cmd", prefix
);
6473 r
= pg
->do_command(cmdmap
, ss
, data
, odata
, con
, tid
);
6476 // don't reply, pg will do so async
6480 ss
<< "not primary for pgid " << pgid
;
6482 // send them the latest diff to ensure they realize the mapping
6484 service
.send_incremental_map(osdmap
->get_epoch() - 1, con
, osdmap
);
6486 // do not reply; they will get newer maps and realize they
6493 ss
<< "i don't have pgid " << pgid
;
6499 else if (prefix
== "bench") {
6502 int64_t osize
, onum
;
6503 // default count 1G, size 4MB
6504 cmd_getval(cct
, cmdmap
, "count", count
, (int64_t)1 << 30);
6505 cmd_getval(cct
, cmdmap
, "size", bsize
, (int64_t)4 << 20);
6506 cmd_getval(cct
, cmdmap
, "object_size", osize
, (int64_t)0);
6507 cmd_getval(cct
, cmdmap
, "object_num", onum
, (int64_t)0);
6509 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
6510 ObjectStore::Sequencer
>("bench"));
6512 uint32_t duration
= cct
->_conf
->osd_bench_duration
;
6514 if (bsize
> (int64_t) cct
->_conf
->osd_bench_max_block_size
) {
6515 // let us limit the block size because the next checks rely on it
6516 // having a sane value. If we allow any block size to be set things
6517 // can still go sideways.
6518 ss
<< "block 'size' values are capped at "
6519 << prettybyte_t(cct
->_conf
->osd_bench_max_block_size
) << ". If you wish to use"
6520 << " a higher value, please adjust 'osd_bench_max_block_size'";
6523 } else if (bsize
< (int64_t) (1 << 20)) {
6524 // entering the realm of small block sizes.
6525 // limit the count to a sane value, assuming a configurable amount of
6526 // IOPS and duration, so that the OSD doesn't get hung up on this,
6527 // preventing timeouts from going off
6529 bsize
* duration
* cct
->_conf
->osd_bench_small_size_max_iops
;
6530 if (count
> max_count
) {
6531 ss
<< "'count' values greater than " << max_count
6532 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6533 << cct
->_conf
->osd_bench_small_size_max_iops
<< " IOPS,"
6534 << " for " << duration
<< " seconds,"
6535 << " can cause ill effects on osd. "
6536 << " Please adjust 'osd_bench_small_size_max_iops' with a higher"
6537 << " value if you wish to use a higher 'count'.";
6542 // 1MB block sizes are big enough so that we get more stuff done.
6543 // However, to avoid the osd from getting hung on this and having
6544 // timers being triggered, we are going to limit the count assuming
6545 // a configurable throughput and duration.
6546 // NOTE: max_count is the total amount of bytes that we believe we
6547 // will be able to write during 'duration' for the given
6548 // throughput. The block size hardly impacts this unless it's
6549 // way too big. Given we already check how big the block size
6550 // is, it's safe to assume everything will check out.
6552 cct
->_conf
->osd_bench_large_size_max_throughput
* duration
;
6553 if (count
> max_count
) {
6554 ss
<< "'count' values greater than " << max_count
6555 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6556 << prettybyte_t(cct
->_conf
->osd_bench_large_size_max_throughput
) << "/s,"
6557 << " for " << duration
<< " seconds,"
6558 << " can cause ill effects on osd. "
6559 << " Please adjust 'osd_bench_large_size_max_throughput'"
6560 << " with a higher value if you wish to use a higher 'count'.";
6566 if (osize
&& bsize
> osize
)
6569 dout(1) << " bench count " << count
6570 << " bsize " << prettybyte_t(bsize
) << dendl
;
6572 ObjectStore::Transaction cleanupt
;
6574 if (osize
&& onum
) {
6576 bufferptr
bp(osize
);
6578 bl
.push_back(std::move(bp
));
6579 bl
.rebuild_page_aligned();
6580 for (int i
=0; i
<onum
; ++i
) {
6582 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", i
);
6584 hobject_t
soid(sobject_t(oid
, 0));
6585 ObjectStore::Transaction t
;
6586 t
.write(coll_t(), ghobject_t(soid
), 0, osize
, bl
);
6587 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6588 cleanupt
.remove(coll_t(), ghobject_t(soid
));
6593 bufferptr
bp(bsize
);
6595 bl
.push_back(std::move(bp
));
6596 bl
.rebuild_page_aligned();
6600 if (!osr
->flush_commit(&waiter
)) {
6605 utime_t start
= ceph_clock_now();
6606 for (int64_t pos
= 0; pos
< count
; pos
+= bsize
) {
6608 unsigned offset
= 0;
6609 if (onum
&& osize
) {
6610 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", (int)(rand() % onum
));
6611 offset
= rand() % (osize
/ bsize
) * bsize
;
6613 snprintf(nm
, sizeof(nm
), "disk_bw_test_%lld", (long long)pos
);
6616 hobject_t
soid(sobject_t(oid
, 0));
6617 ObjectStore::Transaction t
;
6618 t
.write(coll_t::meta(), ghobject_t(soid
), offset
, bsize
, bl
);
6619 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6620 if (!onum
|| !osize
)
6621 cleanupt
.remove(coll_t::meta(), ghobject_t(soid
));
6626 if (!osr
->flush_commit(&waiter
)) {
6630 utime_t end
= ceph_clock_now();
6633 store
->queue_transaction(osr
.get(), std::move(cleanupt
), NULL
);
6636 if (!osr
->flush_commit(&waiter
)) {
6641 uint64_t rate
= (double)count
/ (end
- start
);
6643 f
->open_object_section("osd_bench_results");
6644 f
->dump_int("bytes_written", count
);
6645 f
->dump_int("blocksize", bsize
);
6646 f
->dump_unsigned("bytes_per_sec", rate
);
6650 ss
<< "bench: wrote " << prettybyte_t(count
)
6651 << " in blocks of " << prettybyte_t(bsize
) << " in "
6652 << (end
-start
) << " sec at " << prettybyte_t(rate
) << "/sec";
6656 else if (prefix
== "flush_pg_stats") {
6657 if (osdmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
6658 mgrc
.send_pgstats();
6659 ds
<< service
.get_osd_stat_seq() << "\n";
6665 else if (prefix
== "heap") {
6666 r
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ds
);
6669 else if (prefix
== "debug dump_missing") {
6671 cmd_getval(cct
, cmdmap
, "filename", file_name
);
6672 std::ofstream
fout(file_name
.c_str());
6673 if (!fout
.is_open()) {
6674 ss
<< "failed to open file '" << file_name
<< "'";
6679 fout
<< "*** osd " << whoami
<< ": dump_missing ***" << std::endl
;
6680 RWLock::RLocker
l(pg_map_lock
);
6681 for (ceph::unordered_map
<spg_t
, PG
*>::const_iterator pg_map_e
= pg_map
.begin();
6682 pg_map_e
!= pg_map
.end(); ++pg_map_e
) {
6683 PG
*pg
= pg_map_e
->second
;
6686 fout
<< *pg
<< std::endl
;
6687 std::map
<hobject_t
, pg_missing_item
>::const_iterator mend
=
6688 pg
->pg_log
.get_missing().get_items().end();
6689 std::map
<hobject_t
, pg_missing_item
>::const_iterator mi
=
6690 pg
->pg_log
.get_missing().get_items().begin();
6691 for (; mi
!= mend
; ++mi
) {
6692 fout
<< mi
->first
<< " -> " << mi
->second
<< std::endl
;
6693 if (!pg
->missing_loc
.needs_recovery(mi
->first
))
6695 if (pg
->missing_loc
.is_unfound(mi
->first
))
6696 fout
<< " unfound ";
6697 const set
<pg_shard_t
> &mls(pg
->missing_loc
.get_locations(mi
->first
));
6700 fout
<< "missing_loc: " << mls
<< std::endl
;
6708 else if (prefix
== "debug kick_recovery_wq") {
6710 cmd_getval(cct
, cmdmap
, "delay", delay
);
6713 r
= cct
->_conf
->set_val("osd_recovery_delay_start", oss
.str().c_str());
6715 ss
<< "kick_recovery_wq: error setting "
6716 << "osd_recovery_delay_start to '" << delay
<< "': error "
6720 cct
->_conf
->apply_changes(NULL
);
6721 ss
<< "kicking recovery queue. set osd_recovery_delay_start "
6722 << "to " << cct
->_conf
->osd_recovery_delay_start
;
6725 else if (prefix
== "cpu_profiler") {
6727 cmd_getval(cct
, cmdmap
, "arg", arg
);
6728 vector
<string
> argvec
;
6729 get_str_vec(arg
, argvec
);
6730 cpu_profiler_handle_command(argvec
, ds
);
6733 else if (prefix
== "dump_pg_recovery_stats") {
6736 pg_recovery_stats
.dump_formatted(f
.get());
6739 pg_recovery_stats
.dump(s
);
6740 ds
<< "dump pg recovery stats: " << s
.str();
6744 else if (prefix
== "reset_pg_recovery_stats") {
6745 ss
<< "reset pg recovery stats";
6746 pg_recovery_stats
.reset();
6749 else if (prefix
== "perf histogram dump") {
6751 std::string counter
;
6752 cmd_getval(cct
, cmdmap
, "logger", logger
);
6753 cmd_getval(cct
, cmdmap
, "counter", counter
);
6755 cct
->get_perfcounters_collection()->dump_formatted_histograms(
6756 f
.get(), false, logger
, counter
);
6761 else if (prefix
== "compact") {
6762 dout(1) << "triggering manual compaction" << dendl
;
6763 auto start
= ceph::coarse_mono_clock::now();
6765 auto end
= ceph::coarse_mono_clock::now();
6766 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
6767 dout(1) << "finished manual compaction in "
6768 << time_span
.count()
6769 << " seconds" << dendl
;
6770 ss
<< "compacted omap in " << time_span
.count() << " seconds";
6774 ss
<< "unrecognized command! " << cmd
;
6781 dout(0) << "do_command r=" << r
<< " " << rs
<< dendl
;
6784 MCommandReply
*reply
= new MCommandReply(r
, rs
);
6785 reply
->set_tid(tid
);
6786 reply
->set_data(odata
);
6787 con
->send_message(reply
);
6791 bool OSD::heartbeat_dispatch(Message
*m
)
6793 dout(30) << "heartbeat_dispatch " << m
<< dendl
;
6794 switch (m
->get_type()) {
6797 dout(10) << "ping from " << m
->get_source_inst() << dendl
;
6802 handle_osd_ping(static_cast<MOSDPing
*>(m
));
6806 dout(0) << "dropping unexpected message " << *m
<< " from " << m
->get_source_inst() << dendl
;
6813 bool OSD::ms_dispatch(Message
*m
)
6815 dout(20) << "OSD::ms_dispatch: " << *m
<< dendl
;
6816 if (m
->get_type() == MSG_OSD_MARK_ME_DOWN
) {
6817 service
.got_stop_ack();
6825 if (is_stopping()) {
6839 void OSD::maybe_share_map(
6844 if (!op
->check_send_map
) {
6847 epoch_t last_sent_epoch
= 0;
6849 session
->sent_epoch_lock
.lock();
6850 last_sent_epoch
= session
->last_sent_epoch
;
6851 session
->sent_epoch_lock
.unlock();
6853 const Message
*m
= op
->get_req();
6856 m
->get_connection().get(),
6859 session
? &last_sent_epoch
: NULL
);
6861 session
->sent_epoch_lock
.lock();
6862 if (session
->last_sent_epoch
< last_sent_epoch
) {
6863 session
->last_sent_epoch
= last_sent_epoch
;
6865 session
->sent_epoch_lock
.unlock();
6867 op
->check_send_map
= false;
6870 void OSD::dispatch_session_waiting(Session
*session
, OSDMapRef osdmap
)
6872 assert(session
->session_dispatch_lock
.is_locked());
6874 auto i
= session
->waiting_on_map
.begin();
6875 while (i
!= session
->waiting_on_map
.end()) {
6876 OpRequestRef op
= &(*i
);
6877 assert(ms_can_fast_dispatch(op
->get_req()));
6878 const MOSDFastDispatchOp
*m
= static_cast<const MOSDFastDispatchOp
*>(
6880 if (m
->get_min_epoch() > osdmap
->get_epoch()) {
6883 session
->waiting_on_map
.erase(i
++);
6887 if (m
->get_type() == CEPH_MSG_OSD_OP
) {
6888 pg_t actual_pgid
= osdmap
->raw_pg_to_pg(
6889 static_cast<const MOSDOp
*>(m
)->get_pg());
6890 if (!osdmap
->get_primary_shard(actual_pgid
, &pgid
)) {
6894 pgid
= m
->get_spg();
6896 enqueue_op(pgid
, op
, m
->get_map_epoch());
6899 if (session
->waiting_on_map
.empty()) {
6900 clear_session_waiting_on_map(session
);
6902 register_session_waiting_on_map(session
);
6906 void OSD::ms_fast_dispatch(Message
*m
)
6909 if (service
.is_stopping()) {
6913 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
6916 osd_reqid_t reqid
= op
->get_reqid();
6918 tracepoint(osd
, ms_fast_dispatch
, reqid
.name
._type
,
6919 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
6923 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
6925 // note sender epoch, min req'd epoch
6926 op
->sent_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch();
6927 op
->min_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_min_epoch();
6928 assert(op
->min_epoch
<= op
->sent_epoch
); // sanity check!
6930 service
.maybe_inject_dispatch_delay();
6932 if (m
->get_connection()->has_features(CEPH_FEATUREMASK_RESEND_ON_SPLIT
) ||
6933 m
->get_type() != CEPH_MSG_OSD_OP
) {
6934 // queue it directly
6936 static_cast<MOSDFastDispatchOp
*>(m
)->get_spg(),
6938 static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch());
6940 // legacy client, and this is an MOSDOp (the *only* fast dispatch
6941 // message that didn't have an explicit spg_t); we need to map
6942 // them to an spg_t while preserving delivery order.
6943 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
6946 Mutex::Locker
l(session
->session_dispatch_lock
);
6948 session
->waiting_on_map
.push_back(*op
);
6949 OSDMapRef nextmap
= service
.get_nextmap_reserved();
6950 dispatch_session_waiting(session
, nextmap
);
6951 service
.release_map(nextmap
);
6956 OID_EVENT_TRACE_WITH_MSG(m
, "MS_FAST_DISPATCH_END", false);
6959 void OSD::ms_fast_preprocess(Message
*m
)
6961 if (m
->get_connection()->get_peer_type() == CEPH_ENTITY_TYPE_OSD
) {
6962 if (m
->get_type() == CEPH_MSG_OSD_MAP
) {
6963 MOSDMap
*mm
= static_cast<MOSDMap
*>(m
);
6964 Session
*s
= static_cast<Session
*>(m
->get_connection()->get_priv());
6966 s
->received_map_lock
.lock();
6967 s
->received_map_epoch
= mm
->get_last();
6968 s
->received_map_lock
.unlock();
6975 bool OSD::ms_get_authorizer(int dest_type
, AuthAuthorizer
**authorizer
, bool force_new
)
6977 dout(10) << "OSD::ms_get_authorizer type=" << ceph_entity_type_name(dest_type
) << dendl
;
6979 if (is_stopping()) {
6980 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
6984 if (dest_type
== CEPH_ENTITY_TYPE_MON
)
6988 /* the MonClient checks keys every tick(), so we should just wait for that cycle
6990 if (monc
->wait_auth_rotating(10) < 0) {
6991 derr
<< "OSD::ms_get_authorizer wait_auth_rotating failed" << dendl
;
6996 *authorizer
= monc
->build_authorizer(dest_type
);
6997 return *authorizer
!= NULL
;
7001 bool OSD::ms_verify_authorizer(Connection
*con
, int peer_type
,
7002 int protocol
, bufferlist
& authorizer_data
, bufferlist
& authorizer_reply
,
7003 bool& isvalid
, CryptoKey
& session_key
)
7005 AuthAuthorizeHandler
*authorize_handler
= 0;
7006 switch (peer_type
) {
7007 case CEPH_ENTITY_TYPE_MDS
:
7009 * note: mds is technically a client from our perspective, but
7010 * this makes the 'cluster' consistent w/ monitor's usage.
7012 case CEPH_ENTITY_TYPE_OSD
:
7013 case CEPH_ENTITY_TYPE_MGR
:
7014 authorize_handler
= authorize_handler_cluster_registry
->get_handler(protocol
);
7017 authorize_handler
= authorize_handler_service_registry
->get_handler(protocol
);
7019 if (!authorize_handler
) {
7020 dout(0) << "No AuthAuthorizeHandler found for protocol " << protocol
<< dendl
;
7025 AuthCapsInfo caps_info
;
7028 uint64_t auid
= CEPH_AUTH_UID_DEFAULT
;
7030 RotatingKeyRing
*keys
= monc
->rotating_secrets
.get();
7032 isvalid
= authorize_handler
->verify_authorizer(
7034 authorizer_data
, authorizer_reply
, name
, global_id
, caps_info
, session_key
,
7037 dout(10) << __func__
<< " no rotating_keys (yet), denied" << dendl
;
7042 Session
*s
= static_cast<Session
*>(con
->get_priv());
7044 s
= new Session(cct
);
7045 con
->set_priv(s
->get());
7047 dout(10) << " new session " << s
<< " con=" << s
->con
<< " addr=" << s
->con
->get_peer_addr() << dendl
;
7050 s
->entity_name
= name
;
7051 if (caps_info
.allow_all
)
7052 s
->caps
.set_allow_all();
7055 if (caps_info
.caps
.length() > 0) {
7056 bufferlist::iterator p
= caps_info
.caps
.begin();
7061 catch (buffer::error
& e
) {
7063 bool success
= s
->caps
.parse(str
);
7065 dout(10) << " session " << s
<< " " << s
->entity_name
<< " has caps " << s
->caps
<< " '" << str
<< "'" << dendl
;
7067 dout(10) << " session " << s
<< " " << s
->entity_name
<< " failed to parse caps '" << str
<< "'" << dendl
;
7075 void OSD::do_waiters()
7077 assert(osd_lock
.is_locked());
7079 dout(10) << "do_waiters -- start" << dendl
;
7080 while (!finished
.empty()) {
7081 OpRequestRef next
= finished
.front();
7082 finished
.pop_front();
7085 dout(10) << "do_waiters -- finish" << dendl
;
7088 void OSD::dispatch_op(OpRequestRef op
)
7090 switch (op
->get_req()->get_type()) {
7092 case MSG_OSD_PG_CREATE
:
7093 handle_pg_create(op
);
7095 case MSG_OSD_PG_NOTIFY
:
7096 handle_pg_notify(op
);
7098 case MSG_OSD_PG_QUERY
:
7099 handle_pg_query(op
);
7101 case MSG_OSD_PG_LOG
:
7104 case MSG_OSD_PG_REMOVE
:
7105 handle_pg_remove(op
);
7107 case MSG_OSD_PG_INFO
:
7110 case MSG_OSD_PG_TRIM
:
7113 case MSG_OSD_BACKFILL_RESERVE
:
7114 handle_pg_backfill_reserve(op
);
7116 case MSG_OSD_RECOVERY_RESERVE
:
7117 handle_pg_recovery_reserve(op
);
7122 void OSD::_dispatch(Message
*m
)
7124 assert(osd_lock
.is_locked());
7125 dout(20) << "_dispatch " << m
<< " " << *m
<< dendl
;
7127 switch (m
->get_type()) {
7129 // -- don't need lock --
7131 dout(10) << "ping from " << m
->get_source() << dendl
;
7135 // -- don't need OSDMap --
7137 // map and replication
7138 case CEPH_MSG_OSD_MAP
:
7139 handle_osd_map(static_cast<MOSDMap
*>(m
));
7143 case MSG_PGSTATSACK
:
7144 handle_pg_stats_ack(static_cast<MPGStatsAck
*>(m
));
7147 case MSG_MON_COMMAND
:
7148 handle_command(static_cast<MMonCommand
*>(m
));
7151 handle_command(static_cast<MCommand
*>(m
));
7155 handle_scrub(static_cast<MOSDScrub
*>(m
));
7158 case MSG_OSD_FORCE_RECOVERY
:
7159 handle_force_recovery(m
);
7162 // -- need OSDMap --
7164 case MSG_OSD_PG_CREATE
:
7165 case MSG_OSD_PG_NOTIFY
:
7166 case MSG_OSD_PG_QUERY
:
7167 case MSG_OSD_PG_LOG
:
7168 case MSG_OSD_PG_REMOVE
:
7169 case MSG_OSD_PG_INFO
:
7170 case MSG_OSD_PG_TRIM
:
7171 case MSG_OSD_BACKFILL_RESERVE
:
7172 case MSG_OSD_RECOVERY_RESERVE
:
7174 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
7176 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
7177 // no map? starting up?
7179 dout(7) << "no OSDMap, not booted" << dendl
;
7180 logger
->inc(l_osd_waiting_for_map
);
7181 waiting_for_osdmap
.push_back(op
);
7182 op
->mark_delayed("no osdmap");
7192 void OSD::handle_pg_scrub(MOSDScrub
*m
, PG
*pg
)
7195 if (pg
->is_primary()) {
7196 pg
->unreg_next_scrub();
7197 pg
->scrubber
.must_scrub
= true;
7198 pg
->scrubber
.must_deep_scrub
= m
->deep
|| m
->repair
;
7199 pg
->scrubber
.must_repair
= m
->repair
;
7200 pg
->reg_next_scrub();
7201 dout(10) << "marking " << *pg
<< " for scrub" << dendl
;
7206 void OSD::handle_scrub(MOSDScrub
*m
)
7208 dout(10) << "handle_scrub " << *m
<< dendl
;
7209 if (!require_mon_or_mgr_peer(m
)) {
7213 if (m
->fsid
!= monc
->get_fsid()) {
7214 dout(0) << "handle_scrub fsid " << m
->fsid
<< " != " << monc
->get_fsid() << dendl
;
7219 RWLock::RLocker
l(pg_map_lock
);
7220 if (m
->scrub_pgs
.empty()) {
7221 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
7224 handle_pg_scrub(m
, p
->second
);
7226 for (vector
<pg_t
>::iterator p
= m
->scrub_pgs
.begin();
7227 p
!= m
->scrub_pgs
.end();
7230 if (osdmap
->get_primary_shard(*p
, &pcand
)) {
7231 auto pg_map_entry
= pg_map
.find(pcand
);
7232 if (pg_map_entry
!= pg_map
.end()) {
7233 handle_pg_scrub(m
, pg_map_entry
->second
);
7242 bool OSD::scrub_random_backoff()
7244 bool coin_flip
= (rand() / (double)RAND_MAX
>=
7245 cct
->_conf
->osd_scrub_backoff_ratio
);
7247 dout(20) << "scrub_random_backoff lost coin flip, randomly backing off" << dendl
;
7253 OSDService::ScrubJob::ScrubJob(CephContext
* cct
,
7254 const spg_t
& pg
, const utime_t
& timestamp
,
7255 double pool_scrub_min_interval
,
7256 double pool_scrub_max_interval
, bool must
)
7259 sched_time(timestamp
),
7262 // if not explicitly requested, postpone the scrub with a random delay
7264 double scrub_min_interval
= pool_scrub_min_interval
> 0 ?
7265 pool_scrub_min_interval
: cct
->_conf
->osd_scrub_min_interval
;
7266 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
7267 pool_scrub_max_interval
: cct
->_conf
->osd_scrub_max_interval
;
7269 sched_time
+= scrub_min_interval
;
7270 double r
= rand() / (double)RAND_MAX
;
7272 scrub_min_interval
* cct
->_conf
->osd_scrub_interval_randomize_ratio
* r
;
7273 deadline
+= scrub_max_interval
;
7277 bool OSDService::ScrubJob::ScrubJob::operator<(const OSDService::ScrubJob
& rhs
) const {
7278 if (sched_time
< rhs
.sched_time
)
7280 if (sched_time
> rhs
.sched_time
)
7282 return pgid
< rhs
.pgid
;
7285 bool OSD::scrub_time_permit(utime_t now
)
7288 time_t tt
= now
.sec();
7289 localtime_r(&tt
, &bdt
);
7290 bool time_permit
= false;
7291 if (cct
->_conf
->osd_scrub_begin_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7292 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
&& bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7296 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
|| bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7301 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7302 << " - " << cct
->_conf
->osd_scrub_end_hour
7303 << " now " << bdt
.tm_hour
<< " = no" << dendl
;
7305 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7306 << " - " << cct
->_conf
->osd_scrub_end_hour
7307 << " now " << bdt
.tm_hour
<< " = yes" << dendl
;
7312 bool OSD::scrub_load_below_threshold()
7315 if (getloadavg(loadavgs
, 3) != 3) {
7316 dout(10) << __func__
<< " couldn't read loadavgs\n" << dendl
;
7320 // allow scrub if below configured threshold
7321 if (loadavgs
[0] < cct
->_conf
->osd_scrub_load_threshold
) {
7322 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7323 << " < max " << cct
->_conf
->osd_scrub_load_threshold
7324 << " = yes" << dendl
;
7328 // allow scrub if below daily avg and currently decreasing
7329 if (loadavgs
[0] < daily_loadavg
&& loadavgs
[0] < loadavgs
[2]) {
7330 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7331 << " < daily_loadavg " << daily_loadavg
7332 << " and < 15m avg " << loadavgs
[2]
7333 << " = yes" << dendl
;
7337 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7338 << " >= max " << cct
->_conf
->osd_scrub_load_threshold
7339 << " and ( >= daily_loadavg " << daily_loadavg
7340 << " or >= 15m avg " << loadavgs
[2]
7341 << ") = no" << dendl
;
7345 void OSD::sched_scrub()
7347 // if not permitted, fail fast
7348 if (!service
.can_inc_scrubs_pending()) {
7352 utime_t now
= ceph_clock_now();
7353 bool time_permit
= scrub_time_permit(now
);
7354 bool load_is_low
= scrub_load_below_threshold();
7355 dout(20) << "sched_scrub load_is_low=" << (int)load_is_low
<< dendl
;
7357 OSDService::ScrubJob scrub
;
7358 if (service
.first_scrub_stamp(&scrub
)) {
7360 dout(30) << "sched_scrub examine " << scrub
.pgid
<< " at " << scrub
.sched_time
<< dendl
;
7362 if (scrub
.sched_time
> now
) {
7363 // save ourselves some effort
7364 dout(10) << "sched_scrub " << scrub
.pgid
<< " scheduled at " << scrub
.sched_time
7365 << " > " << now
<< dendl
;
7369 if (!cct
->_conf
->osd_scrub_during_recovery
&& service
.is_recovery_active()) {
7370 dout(10) << __func__
<< "not scheduling scrub of " << scrub
.pgid
<< " due to active recovery ops" << dendl
;
7374 if ((scrub
.deadline
>= now
) && !(time_permit
&& load_is_low
)) {
7375 dout(10) << __func__
<< " not scheduling scrub for " << scrub
.pgid
<< " due to "
7376 << (!time_permit
? "time not permit" : "high load") << dendl
;
7380 PG
*pg
= _lookup_lock_pg(scrub
.pgid
);
7383 if (pg
->get_pgbackend()->scrub_supported() && pg
->is_active()) {
7384 dout(10) << "sched_scrub scrubbing " << scrub
.pgid
<< " at " << scrub
.sched_time
7385 << (pg
->scrubber
.must_scrub
? ", explicitly requested" :
7386 (load_is_low
? ", load_is_low" : " deadline < now"))
7388 if (pg
->sched_scrub()) {
7394 } while (service
.next_scrub_stamp(scrub
, &scrub
));
7396 dout(20) << "sched_scrub done" << dendl
;
7401 // =====================================================
7404 void OSD::wait_for_new_map(OpRequestRef op
)
7407 if (waiting_for_osdmap
.empty()) {
7408 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
7411 logger
->inc(l_osd_waiting_for_map
);
7412 waiting_for_osdmap
.push_back(op
);
7413 op
->mark_delayed("wait for new map");
7418 * assimilate new OSDMap(s). scan pgs, etc.
7421 void OSD::note_down_osd(int peer
)
7423 assert(osd_lock
.is_locked());
7424 cluster_messenger
->mark_down(osdmap
->get_cluster_addr(peer
));
7426 heartbeat_lock
.Lock();
7427 failure_queue
.erase(peer
);
7428 failure_pending
.erase(peer
);
7429 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(peer
);
7430 if (p
!= heartbeat_peers
.end()) {
7431 p
->second
.con_back
->mark_down();
7432 if (p
->second
.con_front
) {
7433 p
->second
.con_front
->mark_down();
7435 heartbeat_peers
.erase(p
);
7437 heartbeat_lock
.Unlock();
7440 void OSD::note_up_osd(int peer
)
7442 service
.forget_peer_epoch(peer
, osdmap
->get_epoch() - 1);
7443 heartbeat_set_peers_need_update();
7446 struct C_OnMapCommit
: public Context
{
7448 epoch_t first
, last
;
7450 C_OnMapCommit(OSD
*o
, epoch_t f
, epoch_t l
, MOSDMap
*m
)
7451 : osd(o
), first(f
), last(l
), msg(m
) {}
7452 void finish(int r
) override
{
7453 osd
->_committed_osd_maps(first
, last
, msg
);
7458 struct C_OnMapApply
: public Context
{
7459 OSDService
*service
;
7460 list
<OSDMapRef
> pinned_maps
;
7462 C_OnMapApply(OSDService
*service
,
7463 const list
<OSDMapRef
> &pinned_maps
,
7465 : service(service
), pinned_maps(pinned_maps
), e(e
) {}
7466 void finish(int r
) override
{
7467 service
->clear_map_bl_cache_pins(e
);
7471 void OSD::osdmap_subscribe(version_t epoch
, bool force_request
)
7473 OSDMapRef osdmap
= service
.get_osdmap();
7474 if (osdmap
->get_epoch() >= epoch
)
7477 if (monc
->sub_want_increment("osdmap", epoch
, CEPH_SUBSCRIBE_ONETIME
) ||
7483 void OSD::trim_maps(epoch_t oldest
, int nreceived
, bool skip_maps
)
7485 epoch_t min
= std::min(oldest
, service
.map_cache
.cached_key_lower_bound());
7486 if (min
<= superblock
.oldest_map
)
7490 ObjectStore::Transaction t
;
7491 for (epoch_t e
= superblock
.oldest_map
; e
< min
; ++e
) {
7492 dout(20) << " removing old osdmap epoch " << e
<< dendl
;
7493 t
.remove(coll_t::meta(), get_osdmap_pobject_name(e
));
7494 t
.remove(coll_t::meta(), get_inc_osdmap_pobject_name(e
));
7495 superblock
.oldest_map
= e
+ 1;
7497 if (num
>= cct
->_conf
->osd_target_transaction_size
&& num
>= nreceived
) {
7498 service
.publish_superblock(superblock
);
7499 write_superblock(t
);
7500 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7504 // skip_maps leaves us with a range of old maps if we fail to remove all
7505 // of them before moving superblock.oldest_map forward to the first map
7506 // in the incoming MOSDMap msg. so we should continue removing them in
7507 // this case, even we could do huge series of delete transactions all at
7514 service
.publish_superblock(superblock
);
7515 write_superblock(t
);
7516 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7519 // we should not remove the cached maps
7520 assert(min
<= service
.map_cache
.cached_key_lower_bound());
7523 void OSD::handle_osd_map(MOSDMap
*m
)
7525 assert(osd_lock
.is_locked());
7526 // Keep a ref in the list until we get the newly received map written
7527 // onto disk. This is important because as long as the refs are alive,
7528 // the OSDMaps will be pinned in the cache and we won't try to read it
7529 // off of disk. Otherwise these maps will probably not stay in the cache,
7530 // and reading those OSDMaps before they are actually written can result
7532 list
<OSDMapRef
> pinned_maps
;
7533 if (m
->fsid
!= monc
->get_fsid()) {
7534 dout(0) << "handle_osd_map fsid " << m
->fsid
<< " != "
7535 << monc
->get_fsid() << dendl
;
7539 if (is_initializing()) {
7540 dout(0) << "ignoring osdmap until we have initialized" << dendl
;
7545 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
7546 if (session
&& !(session
->entity_name
.is_mon() ||
7547 session
->entity_name
.is_osd())) {
7549 dout(10) << "got osd map from Session " << session
7550 << " which we can't take maps from (not a mon or osd)" << dendl
;
7558 // share with the objecter
7560 service
.objecter
->handle_osd_map(m
);
7562 epoch_t first
= m
->get_first();
7563 epoch_t last
= m
->get_last();
7564 dout(3) << "handle_osd_map epochs [" << first
<< "," << last
<< "], i have "
7565 << superblock
.newest_map
7566 << ", src has [" << m
->oldest_map
<< "," << m
->newest_map
<< "]"
7569 logger
->inc(l_osd_map
);
7570 logger
->inc(l_osd_mape
, last
- first
+ 1);
7571 if (first
<= superblock
.newest_map
)
7572 logger
->inc(l_osd_mape_dup
, superblock
.newest_map
- first
+ 1);
7573 if (service
.max_oldest_map
< m
->oldest_map
) {
7574 service
.max_oldest_map
= m
->oldest_map
;
7575 assert(service
.max_oldest_map
>= superblock
.oldest_map
);
7578 // make sure there is something new, here, before we bother flushing
7579 // the queues and such
7580 if (last
<= superblock
.newest_map
) {
7581 dout(10) << " no new maps here, dropping" << dendl
;
7587 bool skip_maps
= false;
7588 if (first
> superblock
.newest_map
+ 1) {
7589 dout(10) << "handle_osd_map message skips epochs "
7590 << superblock
.newest_map
+ 1 << ".." << (first
-1) << dendl
;
7591 if (m
->oldest_map
<= superblock
.newest_map
+ 1) {
7592 osdmap_subscribe(superblock
.newest_map
+ 1, false);
7596 // always try to get the full range of maps--as many as we can. this
7597 // 1- is good to have
7598 // 2- is at present the only way to ensure that we get a *full* map as
7600 if (m
->oldest_map
< first
) {
7601 osdmap_subscribe(m
->oldest_map
- 1, true);
7608 ObjectStore::Transaction t
;
7609 uint64_t txn_size
= 0;
7611 // store new maps: queue for disk and put in the osdmap cache
7612 epoch_t start
= MAX(superblock
.newest_map
+ 1, first
);
7613 for (epoch_t e
= start
; e
<= last
; e
++) {
7614 if (txn_size
>= t
.get_num_bytes()) {
7615 derr
<< __func__
<< " transaction size overflowed" << dendl
;
7616 assert(txn_size
< t
.get_num_bytes());
7618 txn_size
= t
.get_num_bytes();
7619 map
<epoch_t
,bufferlist
>::iterator p
;
7620 p
= m
->maps
.find(e
);
7621 if (p
!= m
->maps
.end()) {
7622 dout(10) << "handle_osd_map got full map for epoch " << e
<< dendl
;
7623 OSDMap
*o
= new OSDMap
;
7624 bufferlist
& bl
= p
->second
;
7628 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7629 t
.write(coll_t::meta(), fulloid
, 0, bl
.length(), bl
);
7631 pinned_maps
.push_back(add_map(o
));
7637 p
= m
->incremental_maps
.find(e
);
7638 if (p
!= m
->incremental_maps
.end()) {
7639 dout(10) << "handle_osd_map got inc map for epoch " << e
<< dendl
;
7640 bufferlist
& bl
= p
->second
;
7641 ghobject_t oid
= get_inc_osdmap_pobject_name(e
);
7642 t
.write(coll_t::meta(), oid
, 0, bl
.length(), bl
);
7643 pin_map_inc_bl(e
, bl
);
7645 OSDMap
*o
= new OSDMap
;
7648 bool got
= get_map_bl(e
- 1, obl
);
7653 OSDMap::Incremental inc
;
7654 bufferlist::iterator p
= bl
.begin();
7656 if (o
->apply_incremental(inc
) < 0) {
7657 derr
<< "ERROR: bad fsid? i have " << osdmap
->get_fsid() << " and inc has " << inc
.fsid
<< dendl
;
7658 assert(0 == "bad fsid");
7662 o
->encode(fbl
, inc
.encode_features
| CEPH_FEATURE_RESERVED
);
7664 bool injected_failure
= false;
7665 if (cct
->_conf
->osd_inject_bad_map_crc_probability
> 0 &&
7666 (rand() % 10000) < cct
->_conf
->osd_inject_bad_map_crc_probability
*10000.0) {
7667 derr
<< __func__
<< " injecting map crc failure" << dendl
;
7668 injected_failure
= true;
7671 if ((inc
.have_crc
&& o
->get_crc() != inc
.full_crc
) || injected_failure
) {
7672 dout(2) << "got incremental " << e
7673 << " but failed to encode full with correct crc; requesting"
7675 clog
->warn() << "failed to encode map e" << e
<< " with expected crc";
7676 dout(20) << "my encoded map was:\n";
7677 fbl
.hexdump(*_dout
);
7680 request_full_map(e
, last
);
7686 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7687 t
.write(coll_t::meta(), fulloid
, 0, fbl
.length(), fbl
);
7689 pinned_maps
.push_back(add_map(o
));
7693 assert(0 == "MOSDMap lied about what maps it had?");
7696 // even if this map isn't from a mon, we may have satisfied our subscription
7697 monc
->sub_got("osdmap", last
);
7699 if (!m
->maps
.empty() && requested_full_first
) {
7700 dout(10) << __func__
<< " still missing full maps " << requested_full_first
7701 << ".." << requested_full_last
<< dendl
;
7702 rerequest_full_maps();
7705 if (superblock
.oldest_map
) {
7706 // make sure we at least keep pace with incoming maps
7707 trim_maps(m
->oldest_map
, last
- first
+ 1, skip_maps
);
7710 if (!superblock
.oldest_map
|| skip_maps
)
7711 superblock
.oldest_map
= first
;
7712 superblock
.newest_map
= last
;
7713 superblock
.current_epoch
= last
;
7715 // note in the superblock that we were clean thru the prior epoch
7716 epoch_t boot_epoch
= service
.get_boot_epoch();
7717 if (boot_epoch
&& boot_epoch
>= superblock
.mounted
) {
7718 superblock
.mounted
= boot_epoch
;
7719 superblock
.clean_thru
= last
;
7722 // superblock and commit
7723 write_superblock(t
);
7724 store
->queue_transaction(
7725 service
.meta_osr
.get(),
7727 new C_OnMapApply(&service
, pinned_maps
, last
),
7728 new C_OnMapCommit(this, start
, last
, m
), 0);
7729 service
.publish_superblock(superblock
);
7732 void OSD::_committed_osd_maps(epoch_t first
, epoch_t last
, MOSDMap
*m
)
7734 dout(10) << __func__
<< " " << first
<< ".." << last
<< dendl
;
7735 if (is_stopping()) {
7736 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7739 Mutex::Locker
l(osd_lock
);
7740 if (is_stopping()) {
7741 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7744 map_lock
.get_write();
7746 bool do_shutdown
= false;
7747 bool do_restart
= false;
7748 bool network_error
= false;
7750 // advance through the new maps
7751 for (epoch_t cur
= first
; cur
<= last
; cur
++) {
7752 dout(10) << " advance to epoch " << cur
7753 << " (<= last " << last
7754 << " <= newest_map " << superblock
.newest_map
7757 OSDMapRef newmap
= get_map(cur
);
7758 assert(newmap
); // we just cached it above!
7760 // start blacklisting messages sent to peers that go down.
7761 service
.pre_publish_map(newmap
);
7763 // kill connections to newly down osds
7764 bool waited_for_reservations
= false;
7766 osdmap
->get_all_osds(old
);
7767 for (set
<int>::iterator p
= old
.begin(); p
!= old
.end(); ++p
) {
7769 osdmap
->is_up(*p
) && // in old map
7770 newmap
->is_down(*p
)) { // but not the new one
7771 if (!waited_for_reservations
) {
7772 service
.await_reserved_maps();
7773 waited_for_reservations
= true;
7776 } else if (*p
!= whoami
&&
7777 osdmap
->is_down(*p
) &&
7778 newmap
->is_up(*p
)) {
7783 if ((osdmap
->test_flag(CEPH_OSDMAP_NOUP
) !=
7784 newmap
->test_flag(CEPH_OSDMAP_NOUP
)) ||
7785 (osdmap
->is_noup(whoami
) != newmap
->is_noup(whoami
))) {
7786 dout(10) << __func__
<< " NOUP flag changed in " << newmap
->get_epoch()
7789 // this captures the case where we sent the boot message while
7790 // NOUP was being set on the mon and our boot request was
7791 // dropped, and then later it is cleared. it imperfectly
7792 // handles the case where our original boot message was not
7793 // dropped and we restart even though we might have booted, but
7794 // that is harmless (boot will just take slightly longer).
7798 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
&&
7799 newmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
7800 dout(10) << __func__
<< " require_osd_release reached luminous in "
7801 << newmap
->get_epoch() << dendl
;
7802 clear_pg_stat_queue();
7803 clear_outstanding_pg_stats();
7809 service
.retrieve_epochs(&boot_epoch
, &up_epoch
, NULL
);
7811 osdmap
->is_up(whoami
) &&
7812 osdmap
->get_inst(whoami
) == client_messenger
->get_myinst()) {
7813 up_epoch
= osdmap
->get_epoch();
7814 dout(10) << "up_epoch is " << up_epoch
<< dendl
;
7816 boot_epoch
= osdmap
->get_epoch();
7817 dout(10) << "boot_epoch is " << boot_epoch
<< dendl
;
7819 service
.set_epochs(&boot_epoch
, &up_epoch
, NULL
);
7823 had_map_since
= ceph_clock_now();
7825 epoch_t _bind_epoch
= service
.get_bind_epoch();
7826 if (osdmap
->is_up(whoami
) &&
7827 osdmap
->get_addr(whoami
) == client_messenger
->get_myaddr() &&
7828 _bind_epoch
< osdmap
->get_up_from(whoami
)) {
7831 dout(1) << "state: booting -> active" << dendl
;
7832 set_state(STATE_ACTIVE
);
7834 // set incarnation so that osd_reqid_t's we generate for our
7835 // objecter requests are unique across restarts.
7836 service
.objecter
->set_client_incarnation(osdmap
->get_epoch());
7840 if (osdmap
->get_epoch() > 0 &&
7842 if (!osdmap
->exists(whoami
)) {
7843 dout(0) << "map says i do not exist. shutting down." << dendl
;
7844 do_shutdown
= true; // don't call shutdown() while we have
7845 // everything paused
7846 } else if (!osdmap
->is_up(whoami
) ||
7847 !osdmap
->get_addr(whoami
).probably_equals(
7848 client_messenger
->get_myaddr()) ||
7849 !osdmap
->get_cluster_addr(whoami
).probably_equals(
7850 cluster_messenger
->get_myaddr()) ||
7851 !osdmap
->get_hb_back_addr(whoami
).probably_equals(
7852 hb_back_server_messenger
->get_myaddr()) ||
7853 (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
7854 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
7855 hb_front_server_messenger
->get_myaddr()))) {
7856 if (!osdmap
->is_up(whoami
)) {
7857 if (service
.is_preparing_to_stop() || service
.is_stopping()) {
7858 service
.got_stop_ack();
7860 clog
->warn() << "Monitor daemon marked osd." << whoami
<< " down, "
7861 "but it is still running";
7862 clog
->debug() << "map e" << osdmap
->get_epoch()
7863 << " wrongly marked me down at e"
7864 << osdmap
->get_down_at(whoami
);
7866 } else if (!osdmap
->get_addr(whoami
).probably_equals(
7867 client_messenger
->get_myaddr())) {
7868 clog
->error() << "map e" << osdmap
->get_epoch()
7869 << " had wrong client addr (" << osdmap
->get_addr(whoami
)
7870 << " != my " << client_messenger
->get_myaddr() << ")";
7871 } else if (!osdmap
->get_cluster_addr(whoami
).probably_equals(
7872 cluster_messenger
->get_myaddr())) {
7873 clog
->error() << "map e" << osdmap
->get_epoch()
7874 << " had wrong cluster addr ("
7875 << osdmap
->get_cluster_addr(whoami
)
7876 << " != my " << cluster_messenger
->get_myaddr() << ")";
7877 } else if (!osdmap
->get_hb_back_addr(whoami
).probably_equals(
7878 hb_back_server_messenger
->get_myaddr())) {
7879 clog
->error() << "map e" << osdmap
->get_epoch()
7880 << " had wrong heartbeat back addr ("
7881 << osdmap
->get_hb_back_addr(whoami
)
7882 << " != my " << hb_back_server_messenger
->get_myaddr()
7884 } else if (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
7885 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
7886 hb_front_server_messenger
->get_myaddr())) {
7887 clog
->error() << "map e" << osdmap
->get_epoch()
7888 << " had wrong heartbeat front addr ("
7889 << osdmap
->get_hb_front_addr(whoami
)
7890 << " != my " << hb_front_server_messenger
->get_myaddr()
7894 if (!service
.is_stopping()) {
7895 epoch_t up_epoch
= 0;
7896 epoch_t bind_epoch
= osdmap
->get_epoch();
7897 service
.set_epochs(NULL
,&up_epoch
, &bind_epoch
);
7901 utime_t now
= ceph_clock_now();
7902 utime_t grace
= utime_t(cct
->_conf
->osd_max_markdown_period
, 0);
7903 osd_markdown_log
.push_back(now
);
7904 //clear all out-of-date log
7905 while (!osd_markdown_log
.empty() &&
7906 osd_markdown_log
.front() + grace
< now
)
7907 osd_markdown_log
.pop_front();
7908 if ((int)osd_markdown_log
.size() > cct
->_conf
->osd_max_markdown_count
) {
7909 dout(0) << __func__
<< " marked down "
7910 << osd_markdown_log
.size()
7911 << " > osd_max_markdown_count "
7912 << cct
->_conf
->osd_max_markdown_count
7913 << " in last " << grace
<< " seconds, shutting down"
7919 start_waiting_for_healthy();
7921 set
<int> avoid_ports
;
7922 #if defined(__FreeBSD__)
7923 // prevent FreeBSD from grabbing the client_messenger port during
7924 // rebinding. In which case a cluster_meesneger will connect also
7926 avoid_ports
.insert(client_messenger
->get_myaddr().get_port());
7928 avoid_ports
.insert(cluster_messenger
->get_myaddr().get_port());
7929 avoid_ports
.insert(hb_back_server_messenger
->get_myaddr().get_port());
7930 avoid_ports
.insert(hb_front_server_messenger
->get_myaddr().get_port());
7932 int r
= cluster_messenger
->rebind(avoid_ports
);
7934 do_shutdown
= true; // FIXME: do_restart?
7935 network_error
= true;
7936 dout(0) << __func__
<< " marked down:"
7937 << " rebind cluster_messenger failed" << dendl
;
7940 r
= hb_back_server_messenger
->rebind(avoid_ports
);
7942 do_shutdown
= true; // FIXME: do_restart?
7943 network_error
= true;
7944 dout(0) << __func__
<< " marked down:"
7945 << " rebind hb_back_server_messenger failed" << dendl
;
7948 r
= hb_front_server_messenger
->rebind(avoid_ports
);
7950 do_shutdown
= true; // FIXME: do_restart?
7951 network_error
= true;
7952 dout(0) << __func__
<< " marked down:"
7953 << " rebind hb_front_server_messenger failed" << dendl
;
7956 hb_front_client_messenger
->mark_down_all();
7957 hb_back_client_messenger
->mark_down_all();
7959 reset_heartbeat_peers();
7964 map_lock
.put_write();
7966 check_osdmap_features(store
);
7971 if (is_active() || is_waiting_for_healthy())
7972 maybe_update_heartbeat_peers();
7975 dout(10) << " not yet active; waiting for peering wq to drain" << dendl
;
7982 if (network_error
) {
7983 Mutex::Locker
l(heartbeat_lock
);
7984 map
<int,pair
<utime_t
,entity_inst_t
>>::iterator it
=
7985 failure_pending
.begin();
7986 while (it
!= failure_pending
.end()) {
7987 dout(10) << "handle_osd_ping canceling in-flight failure report for osd."
7988 << it
->first
<< dendl
;
7989 send_still_alive(osdmap
->get_epoch(), it
->second
.second
);
7990 failure_pending
.erase(it
++);
7993 // trigger shutdown in a different thread
7994 dout(0) << __func__
<< " shutdown OSD via async signal" << dendl
;
7995 queue_async_signal(SIGINT
);
7997 else if (m
->newest_map
&& m
->newest_map
> last
) {
7998 dout(10) << " msg say newest map is " << m
->newest_map
7999 << ", requesting more" << dendl
;
8000 osdmap_subscribe(osdmap
->get_epoch()+1, false);
8002 else if (is_preboot()) {
8003 if (m
->get_source().is_mon())
8004 _preboot(m
->oldest_map
, m
->newest_map
);
8008 else if (do_restart
)
8013 void OSD::check_osdmap_features(ObjectStore
*fs
)
8015 // adjust required feature bits?
8017 // we have to be a bit careful here, because we are accessing the
8018 // Policy structures without taking any lock. in particular, only
8019 // modify integer values that can safely be read by a racing CPU.
8020 // since we are only accessing existing Policy structures a their
8021 // current memory location, and setting or clearing bits in integer
8022 // fields, and we are the only writer, this is not a problem.
8025 Messenger::Policy p
= client_messenger
->get_default_policy();
8027 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_CLIENT
, &mask
);
8028 if ((p
.features_required
& mask
) != features
) {
8029 dout(0) << "crush map has features " << features
8030 << ", adjusting msgr requires for clients" << dendl
;
8031 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8032 client_messenger
->set_default_policy(p
);
8036 Messenger::Policy p
= client_messenger
->get_policy(entity_name_t::TYPE_MON
);
8038 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_MON
, &mask
);
8039 if ((p
.features_required
& mask
) != features
) {
8040 dout(0) << "crush map has features " << features
8041 << " was " << p
.features_required
8042 << ", adjusting msgr requires for mons" << dendl
;
8043 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8044 client_messenger
->set_policy(entity_name_t::TYPE_MON
, p
);
8048 Messenger::Policy p
= cluster_messenger
->get_policy(entity_name_t::TYPE_OSD
);
8050 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_OSD
, &mask
);
8052 if ((p
.features_required
& mask
) != features
) {
8053 dout(0) << "crush map has features " << features
8054 << ", adjusting msgr requires for osds" << dendl
;
8055 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8056 cluster_messenger
->set_policy(entity_name_t::TYPE_OSD
, p
);
8059 if ((features
& CEPH_FEATURE_OSD_ERASURE_CODES
) &&
8060 !superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
)) {
8061 dout(0) << __func__
<< " enabling on-disk ERASURE CODES compat feature" << dendl
;
8062 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
);
8063 ObjectStore::Transaction t
;
8064 write_superblock(t
);
8065 int err
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), NULL
);
8071 bool OSD::advance_pg(
8072 epoch_t osd_epoch
, PG
*pg
,
8073 ThreadPool::TPHandle
&handle
,
8074 PG::RecoveryCtx
*rctx
,
8075 set
<PGRef
> *new_pgs
)
8077 assert(pg
->is_locked());
8078 epoch_t next_epoch
= pg
->get_osdmap()->get_epoch() + 1;
8079 OSDMapRef lastmap
= pg
->get_osdmap();
8081 if (lastmap
->get_epoch() == osd_epoch
)
8083 assert(lastmap
->get_epoch() < osd_epoch
);
8085 epoch_t min_epoch
= service
.get_min_pg_epoch();
8088 max
= min_epoch
+ cct
->_conf
->osd_map_max_advance
;
8090 max
= next_epoch
+ cct
->_conf
->osd_map_max_advance
;
8094 next_epoch
<= osd_epoch
&& next_epoch
<= max
;
8096 OSDMapRef nextmap
= service
.try_get_map(next_epoch
);
8098 dout(20) << __func__
<< " missing map " << next_epoch
<< dendl
;
8099 // make sure max is bumped up so that we can get past any
8101 max
= MAX(max
, next_epoch
+ cct
->_conf
->osd_map_max_advance
);
8105 vector
<int> newup
, newacting
;
8106 int up_primary
, acting_primary
;
8107 nextmap
->pg_to_up_acting_osds(
8109 &newup
, &up_primary
,
8110 &newacting
, &acting_primary
);
8111 pg
->handle_advance_map(
8112 nextmap
, lastmap
, newup
, up_primary
,
8113 newacting
, acting_primary
, rctx
);
8116 set
<spg_t
> children
;
8117 spg_t
parent(pg
->info
.pgid
);
8118 if (parent
.is_split(
8119 lastmap
->get_pg_num(pg
->pool
.id
),
8120 nextmap
->get_pg_num(pg
->pool
.id
),
8122 service
.mark_split_in_progress(pg
->info
.pgid
, children
);
8124 pg
, children
, new_pgs
, lastmap
, nextmap
,
8129 handle
.reset_tp_timeout();
8131 service
.pg_update_epoch(pg
->info
.pgid
, lastmap
->get_epoch());
8132 pg
->handle_activate_map(rctx
);
8133 if (next_epoch
<= osd_epoch
) {
8134 dout(10) << __func__
<< " advanced to max " << max
8135 << " past min epoch " << min_epoch
8136 << " ... will requeue " << *pg
<< dendl
;
8142 void OSD::consume_map()
8144 assert(osd_lock
.is_locked());
8145 dout(7) << "consume_map version " << osdmap
->get_epoch() << dendl
;
8147 int num_pg_primary
= 0, num_pg_replica
= 0, num_pg_stray
= 0;
8148 list
<PGRef
> to_remove
;
8152 RWLock::RLocker
l(pg_map_lock
);
8153 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8156 PG
*pg
= it
->second
;
8158 if (pg
->is_primary())
8160 else if (pg
->is_replica())
8165 if (!osdmap
->have_pg_pool(pg
->info
.pgid
.pool())) {
8167 to_remove
.push_back(PGRef(pg
));
8169 service
.init_splits_between(it
->first
, service
.get_osdmap(), osdmap
);
8176 for (list
<PGRef
>::iterator i
= to_remove
.begin();
8177 i
!= to_remove
.end();
8178 to_remove
.erase(i
++)) {
8179 RWLock::WLocker
locker(pg_map_lock
);
8185 service
.expand_pg_num(service
.get_osdmap(), osdmap
);
8187 service
.pre_publish_map(osdmap
);
8188 service
.await_reserved_maps();
8189 service
.publish_map(osdmap
);
8191 service
.maybe_inject_dispatch_delay();
8193 dispatch_sessions_waiting_on_map();
8195 service
.maybe_inject_dispatch_delay();
8197 // remove any PGs which we no longer host from the session waiting_for_pg lists
8198 dout(20) << __func__
<< " checking waiting_for_pg" << dendl
;
8199 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
8201 service
.maybe_inject_dispatch_delay();
8205 RWLock::RLocker
l(pg_map_lock
);
8206 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8209 PG
*pg
= it
->second
;
8211 pg
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
8215 logger
->set(l_osd_pg
, pg_map
.size());
8217 logger
->set(l_osd_pg_primary
, num_pg_primary
);
8218 logger
->set(l_osd_pg_replica
, num_pg_replica
);
8219 logger
->set(l_osd_pg_stray
, num_pg_stray
);
8222 void OSD::activate_map()
8224 assert(osd_lock
.is_locked());
8226 dout(7) << "activate_map version " << osdmap
->get_epoch() << dendl
;
8228 if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
)) {
8229 derr
<< __func__
<< " SORTBITWISE flag is not set" << dendl
;
8233 if (osdmap
->test_flag(CEPH_OSDMAP_FULL
)) {
8234 dout(10) << " osdmap flagged full, doing onetime osdmap subscribe" << dendl
;
8235 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
8239 if (osdmap
->test_flag(CEPH_OSDMAP_NORECOVER
)) {
8240 if (!service
.recovery_is_paused()) {
8241 dout(1) << "pausing recovery (NORECOVER flag set)" << dendl
;
8242 service
.pause_recovery();
8245 if (service
.recovery_is_paused()) {
8246 dout(1) << "unpausing recovery (NORECOVER flag unset)" << dendl
;
8247 service
.unpause_recovery();
8251 service
.activate_map();
8254 take_waiters(waiting_for_osdmap
);
8257 bool OSD::require_mon_peer(const Message
*m
)
8259 if (!m
->get_connection()->peer_is_mon()) {
8260 dout(0) << "require_mon_peer received from non-mon "
8261 << m
->get_connection()->get_peer_addr()
8262 << " " << *m
<< dendl
;
8268 bool OSD::require_mon_or_mgr_peer(const Message
*m
)
8270 if (!m
->get_connection()->peer_is_mon() &&
8271 !m
->get_connection()->peer_is_mgr()) {
8272 dout(0) << "require_mon_or_mgr_peer received from non-mon, non-mgr "
8273 << m
->get_connection()->get_peer_addr()
8274 << " " << *m
<< dendl
;
8280 bool OSD::require_osd_peer(const Message
*m
)
8282 if (!m
->get_connection()->peer_is_osd()) {
8283 dout(0) << "require_osd_peer received from non-osd "
8284 << m
->get_connection()->get_peer_addr()
8285 << " " << *m
<< dendl
;
8291 bool OSD::require_self_aliveness(const Message
*m
, epoch_t epoch
)
8293 epoch_t up_epoch
= service
.get_up_epoch();
8294 if (epoch
< up_epoch
) {
8295 dout(7) << "from pre-up epoch " << epoch
<< " < " << up_epoch
<< dendl
;
8300 dout(7) << "still in boot state, dropping message " << *m
<< dendl
;
8307 bool OSD::require_same_peer_instance(const Message
*m
, OSDMapRef
& map
,
8308 bool is_fast_dispatch
)
8310 int from
= m
->get_source().num();
8312 if (map
->is_down(from
) ||
8313 (map
->get_cluster_addr(from
) != m
->get_source_inst().addr
)) {
8314 dout(5) << "from dead osd." << from
<< ", marking down, "
8315 << " msg was " << m
->get_source_inst().addr
8316 << " expected " << (map
->is_up(from
) ?
8317 map
->get_cluster_addr(from
) : entity_addr_t())
8319 ConnectionRef con
= m
->get_connection();
8321 Session
*s
= static_cast<Session
*>(con
->get_priv());
8323 if (!is_fast_dispatch
)
8324 s
->session_dispatch_lock
.Lock();
8325 clear_session_waiting_on_map(s
);
8326 con
->set_priv(NULL
); // break ref <-> session cycle, if any
8327 if (!is_fast_dispatch
)
8328 s
->session_dispatch_lock
.Unlock();
8338 * require that we have same (or newer) map, and that
8339 * the source is the pg primary.
8341 bool OSD::require_same_or_newer_map(OpRequestRef
& op
, epoch_t epoch
,
8342 bool is_fast_dispatch
)
8344 const Message
*m
= op
->get_req();
8345 dout(15) << "require_same_or_newer_map " << epoch
8346 << " (i am " << osdmap
->get_epoch() << ") " << m
<< dendl
;
8348 assert(osd_lock
.is_locked());
8350 // do they have a newer map?
8351 if (epoch
> osdmap
->get_epoch()) {
8352 dout(7) << "waiting for newer map epoch " << epoch
8353 << " > my " << osdmap
->get_epoch() << " with " << m
<< dendl
;
8354 wait_for_new_map(op
);
8358 if (!require_self_aliveness(op
->get_req(), epoch
)) {
8362 // ok, our map is same or newer.. do they still exist?
8363 if (m
->get_connection()->get_messenger() == cluster_messenger
&&
8364 !require_same_peer_instance(op
->get_req(), osdmap
, is_fast_dispatch
)) {
8375 // ----------------------------------------
8378 void OSD::split_pgs(
8380 const set
<spg_t
> &childpgids
, set
<PGRef
> *out_pgs
,
8383 PG::RecoveryCtx
*rctx
)
8385 unsigned pg_num
= nextmap
->get_pg_num(
8387 parent
->update_snap_mapper_bits(
8388 parent
->info
.pgid
.get_split_bits(pg_num
)
8391 vector
<object_stat_sum_t
> updated_stats(childpgids
.size() + 1);
8392 parent
->info
.stats
.stats
.sum
.split(updated_stats
);
8394 vector
<object_stat_sum_t
>::iterator stat_iter
= updated_stats
.begin();
8395 for (set
<spg_t
>::const_iterator i
= childpgids
.begin();
8396 i
!= childpgids
.end();
8398 assert(stat_iter
!= updated_stats
.end());
8399 dout(10) << "Splitting " << *parent
<< " into " << *i
<< dendl
;
8400 assert(service
.splitting(*i
));
8401 PG
* child
= _make_pg(nextmap
, *i
);
8403 out_pgs
->insert(child
);
8404 rctx
->created_pgs
.insert(child
);
8406 unsigned split_bits
= i
->get_split_bits(pg_num
);
8407 dout(10) << "pg_num is " << pg_num
<< dendl
;
8408 dout(10) << "m_seed " << i
->ps() << dendl
;
8409 dout(10) << "split_bits is " << split_bits
<< dendl
;
8411 parent
->split_colls(
8421 child
->info
.stats
.stats
.sum
= *stat_iter
;
8423 child
->write_if_dirty(*(rctx
->transaction
));
8426 assert(stat_iter
!= updated_stats
.end());
8427 parent
->info
.stats
.stats
.sum
= *stat_iter
;
8428 parent
->write_if_dirty(*(rctx
->transaction
));
8434 void OSD::handle_pg_create(OpRequestRef op
)
8436 const MOSDPGCreate
*m
= static_cast<const MOSDPGCreate
*>(op
->get_req());
8437 assert(m
->get_type() == MSG_OSD_PG_CREATE
);
8439 dout(10) << "handle_pg_create " << *m
<< dendl
;
8441 if (!require_mon_peer(op
->get_req())) {
8445 if (!require_same_or_newer_map(op
, m
->epoch
, false))
8450 map
<pg_t
,utime_t
>::const_iterator ci
= m
->ctimes
.begin();
8451 for (map
<pg_t
,pg_create_t
>::const_iterator p
= m
->mkpg
.begin();
8454 assert(ci
!= m
->ctimes
.end() && ci
->first
== p
->first
);
8455 epoch_t created
= p
->second
.created
;
8456 if (p
->second
.split_bits
) // Skip split pgs
8460 if (on
.preferred() >= 0) {
8461 dout(20) << "ignoring localized pg " << on
<< dendl
;
8465 if (!osdmap
->have_pg_pool(on
.pool())) {
8466 dout(20) << "ignoring pg on deleted pool " << on
<< dendl
;
8470 dout(20) << "mkpg " << on
<< " e" << created
<< "@" << ci
->second
<< dendl
;
8472 // is it still ours?
8473 vector
<int> up
, acting
;
8474 int up_primary
= -1;
8475 int acting_primary
= -1;
8476 osdmap
->pg_to_up_acting_osds(on
, &up
, &up_primary
, &acting
, &acting_primary
);
8477 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
8479 if (acting_primary
!= whoami
) {
8480 dout(10) << "mkpg " << on
<< " not acting_primary (" << acting_primary
8481 << "), my role=" << role
<< ", skipping" << dendl
;
8486 bool mapped
= osdmap
->get_primary_shard(on
, &pgid
);
8490 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
8492 pg_history_t history
;
8493 build_initial_pg_history(pgid
, created
, ci
->second
, &history
, &pi
);
8495 // The mon won't resend unless the primary changed, so
8496 // we ignore same_interval_since. We'll pass this history
8497 // to handle_pg_peering_evt with the current epoch as the
8498 // event -- the project_pg_history check in
8499 // handle_pg_peering_evt will be a noop.
8500 if (history
.same_primary_since
> m
->epoch
) {
8501 dout(10) << __func__
<< ": got obsolete pg create on pgid "
8502 << pgid
<< " from epoch " << m
->epoch
8503 << ", primary changed in " << history
.same_primary_since
8508 if (handle_pg_peering_evt(
8512 osdmap
->get_epoch(),
8513 PG::CephPeeringEvtRef(
8514 new PG::CephPeeringEvt(
8515 osdmap
->get_epoch(),
8516 osdmap
->get_epoch(),
8519 service
.send_pg_created(pgid
.pgid
);
8522 last_pg_create_epoch
= m
->epoch
;
8524 maybe_update_heartbeat_peers();
8528 // ----------------------------------------
8529 // peering and recovery
8531 PG::RecoveryCtx
OSD::create_context()
8533 ObjectStore::Transaction
*t
= new ObjectStore::Transaction
;
8534 C_Contexts
*on_applied
= new C_Contexts(cct
);
8535 C_Contexts
*on_safe
= new C_Contexts(cct
);
8536 map
<int, map
<spg_t
,pg_query_t
> > *query_map
=
8537 new map
<int, map
<spg_t
, pg_query_t
> >;
8538 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *notify_list
=
8539 new map
<int, vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8540 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *info_map
=
8541 new map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8542 PG::RecoveryCtx
rctx(query_map
, info_map
, notify_list
,
8543 on_applied
, on_safe
, t
);
8547 struct C_OpenPGs
: public Context
{
8551 C_OpenPGs(set
<PGRef
>& p
, ObjectStore
*s
, OSD
* o
) : store(s
), osd(o
) {
8554 void finish(int r
) override
{
8555 RWLock::RLocker
l(osd
->pg_map_lock
);
8556 for (auto p
: pgs
) {
8557 if (osd
->pg_map
.count(p
->info
.pgid
)) {
8558 p
->ch
= store
->open_collection(p
->coll
);
8565 void OSD::dispatch_context_transaction(PG::RecoveryCtx
&ctx
, PG
*pg
,
8566 ThreadPool::TPHandle
*handle
)
8568 if (!ctx
.transaction
->empty()) {
8569 if (!ctx
.created_pgs
.empty()) {
8570 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8572 int tr
= store
->queue_transaction(
8574 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
,
8575 TrackedOpRef(), handle
);
8576 delete (ctx
.transaction
);
8578 ctx
.transaction
= new ObjectStore::Transaction
;
8579 ctx
.on_applied
= new C_Contexts(cct
);
8580 ctx
.on_safe
= new C_Contexts(cct
);
8584 void OSD::dispatch_context(PG::RecoveryCtx
&ctx
, PG
*pg
, OSDMapRef curmap
,
8585 ThreadPool::TPHandle
*handle
)
8587 if (service
.get_osdmap()->is_up(whoami
) &&
8589 do_notifies(*ctx
.notify_list
, curmap
);
8590 do_queries(*ctx
.query_map
, curmap
);
8591 do_infos(*ctx
.info_map
, curmap
);
8593 delete ctx
.notify_list
;
8594 delete ctx
.query_map
;
8595 delete ctx
.info_map
;
8596 if ((ctx
.on_applied
->empty() &&
8597 ctx
.on_safe
->empty() &&
8598 ctx
.transaction
->empty() &&
8599 ctx
.created_pgs
.empty()) || !pg
) {
8600 delete ctx
.transaction
;
8601 delete ctx
.on_applied
;
8603 assert(ctx
.created_pgs
.empty());
8605 if (!ctx
.created_pgs
.empty()) {
8606 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8608 int tr
= store
->queue_transaction(
8610 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
, TrackedOpRef(),
8612 delete (ctx
.transaction
);
8618 * Send an MOSDPGNotify to a primary, with a list of PGs that I have
8619 * content for, and they are primary for.
8622 void OSD::do_notifies(
8623 map
<int,vector
<pair
<pg_notify_t
,PastIntervals
> > >& notify_list
,
8627 vector
<pair
<pg_notify_t
,PastIntervals
> > >::iterator it
=
8628 notify_list
.begin();
8629 it
!= notify_list
.end();
8631 if (!curmap
->is_up(it
->first
)) {
8632 dout(20) << __func__
<< " skipping down osd." << it
->first
<< dendl
;
8635 ConnectionRef con
= service
.get_con_osd_cluster(
8636 it
->first
, curmap
->get_epoch());
8638 dout(20) << __func__
<< " skipping osd." << it
->first
8639 << " (NULL con)" << dendl
;
8642 service
.share_map_peer(it
->first
, con
.get(), curmap
);
8643 dout(7) << __func__
<< " osd " << it
->first
8644 << " on " << it
->second
.size() << " PGs" << dendl
;
8645 MOSDPGNotify
*m
= new MOSDPGNotify(curmap
->get_epoch(),
8647 con
->send_message(m
);
8653 * send out pending queries for info | summaries
8655 void OSD::do_queries(map
<int, map
<spg_t
,pg_query_t
> >& query_map
,
8658 for (map
<int, map
<spg_t
,pg_query_t
> >::iterator pit
= query_map
.begin();
8659 pit
!= query_map
.end();
8661 if (!curmap
->is_up(pit
->first
)) {
8662 dout(20) << __func__
<< " skipping down osd." << pit
->first
<< dendl
;
8665 int who
= pit
->first
;
8666 ConnectionRef con
= service
.get_con_osd_cluster(who
, curmap
->get_epoch());
8668 dout(20) << __func__
<< " skipping osd." << who
8669 << " (NULL con)" << dendl
;
8672 service
.share_map_peer(who
, con
.get(), curmap
);
8673 dout(7) << __func__
<< " querying osd." << who
8674 << " on " << pit
->second
.size() << " PGs" << dendl
;
8675 MOSDPGQuery
*m
= new MOSDPGQuery(curmap
->get_epoch(), pit
->second
);
8676 con
->send_message(m
);
8681 void OSD::do_infos(map
<int,
8682 vector
<pair
<pg_notify_t
, PastIntervals
> > >& info_map
,
8686 vector
<pair
<pg_notify_t
, PastIntervals
> > >::iterator p
=
8688 p
!= info_map
.end();
8690 if (!curmap
->is_up(p
->first
)) {
8691 dout(20) << __func__
<< " skipping down osd." << p
->first
<< dendl
;
8694 for (vector
<pair
<pg_notify_t
,PastIntervals
> >::iterator i
= p
->second
.begin();
8695 i
!= p
->second
.end();
8697 dout(20) << __func__
<< " sending info " << i
->first
.info
8698 << " to shard " << p
->first
<< dendl
;
8700 ConnectionRef con
= service
.get_con_osd_cluster(
8701 p
->first
, curmap
->get_epoch());
8703 dout(20) << __func__
<< " skipping osd." << p
->first
8704 << " (NULL con)" << dendl
;
8707 service
.share_map_peer(p
->first
, con
.get(), curmap
);
8708 MOSDPGInfo
*m
= new MOSDPGInfo(curmap
->get_epoch());
8709 m
->pg_list
= p
->second
;
8710 con
->send_message(m
);
8717 * from non-primary to primary
8718 * includes pg_info_t.
8719 * NOTE: called with opqueue active.
8721 void OSD::handle_pg_notify(OpRequestRef op
)
8723 const MOSDPGNotify
*m
= static_cast<const MOSDPGNotify
*>(op
->get_req());
8724 assert(m
->get_type() == MSG_OSD_PG_NOTIFY
);
8726 dout(7) << "handle_pg_notify from " << m
->get_source() << dendl
;
8727 int from
= m
->get_source().num();
8729 if (!require_osd_peer(op
->get_req()))
8732 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
8737 for (auto it
= m
->get_pg_list().begin();
8738 it
!= m
->get_pg_list().end();
8740 if (it
->first
.info
.pgid
.preferred() >= 0) {
8741 dout(20) << "ignoring localized pg " << it
->first
.info
.pgid
<< dendl
;
8745 handle_pg_peering_evt(
8746 spg_t(it
->first
.info
.pgid
.pgid
, it
->first
.to
),
8747 it
->first
.info
.history
, it
->second
,
8748 it
->first
.query_epoch
,
8749 PG::CephPeeringEvtRef(
8750 new PG::CephPeeringEvt(
8751 it
->first
.epoch_sent
, it
->first
.query_epoch
,
8752 PG::MNotifyRec(pg_shard_t(from
, it
->first
.from
), it
->first
,
8753 op
->get_req()->get_connection()->get_features())))
8758 void OSD::handle_pg_log(OpRequestRef op
)
8760 MOSDPGLog
*m
= static_cast<MOSDPGLog
*>(op
->get_nonconst_req());
8761 assert(m
->get_type() == MSG_OSD_PG_LOG
);
8762 dout(7) << "handle_pg_log " << *m
<< " from " << m
->get_source() << dendl
;
8764 if (!require_osd_peer(op
->get_req()))
8767 int from
= m
->get_source().num();
8768 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
8771 if (m
->info
.pgid
.preferred() >= 0) {
8772 dout(10) << "ignoring localized pg " << m
->info
.pgid
<< dendl
;
8777 handle_pg_peering_evt(
8778 spg_t(m
->info
.pgid
.pgid
, m
->to
),
8779 m
->info
.history
, m
->past_intervals
, m
->get_epoch(),
8780 PG::CephPeeringEvtRef(
8781 new PG::CephPeeringEvt(
8782 m
->get_epoch(), m
->get_query_epoch(),
8783 PG::MLogRec(pg_shard_t(from
, m
->from
), m
)))
8787 void OSD::handle_pg_info(OpRequestRef op
)
8789 const MOSDPGInfo
*m
= static_cast<const MOSDPGInfo
*>(op
->get_req());
8790 assert(m
->get_type() == MSG_OSD_PG_INFO
);
8791 dout(7) << "handle_pg_info " << *m
<< " from " << m
->get_source() << dendl
;
8793 if (!require_osd_peer(op
->get_req()))
8796 int from
= m
->get_source().num();
8797 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
8802 for (auto p
= m
->pg_list
.begin();
8803 p
!= m
->pg_list
.end();
8805 if (p
->first
.info
.pgid
.preferred() >= 0) {
8806 dout(10) << "ignoring localized pg " << p
->first
.info
.pgid
<< dendl
;
8810 handle_pg_peering_evt(
8811 spg_t(p
->first
.info
.pgid
.pgid
, p
->first
.to
),
8812 p
->first
.info
.history
, p
->second
, p
->first
.epoch_sent
,
8813 PG::CephPeeringEvtRef(
8814 new PG::CephPeeringEvt(
8815 p
->first
.epoch_sent
, p
->first
.query_epoch
,
8818 from
, p
->first
.from
), p
->first
.info
, p
->first
.epoch_sent
)))
8823 void OSD::handle_pg_trim(OpRequestRef op
)
8825 const MOSDPGTrim
*m
= static_cast<const MOSDPGTrim
*>(op
->get_req());
8826 assert(m
->get_type() == MSG_OSD_PG_TRIM
);
8828 dout(7) << "handle_pg_trim " << *m
<< " from " << m
->get_source() << dendl
;
8830 if (!require_osd_peer(op
->get_req()))
8833 int from
= m
->get_source().num();
8834 if (!require_same_or_newer_map(op
, m
->epoch
, false))
8837 if (m
->pgid
.preferred() >= 0) {
8838 dout(10) << "ignoring localized pg " << m
->pgid
<< dendl
;
8844 PG
*pg
= _lookup_lock_pg(m
->pgid
);
8846 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
8850 if (m
->epoch
< pg
->info
.history
.same_interval_since
) {
8851 dout(10) << *pg
<< " got old trim to " << m
->trim_to
<< ", ignoring" << dendl
;
8856 if (pg
->is_primary()) {
8857 // peer is informing us of their last_complete_ondisk
8858 dout(10) << *pg
<< " replica osd." << from
<< " lcod " << m
->trim_to
<< dendl
;
8859 pg
->peer_last_complete_ondisk
[pg_shard_t(from
, m
->pgid
.shard
)] =
8861 // trim log when the pg is recovered
8862 pg
->calc_min_last_complete_ondisk();
8864 // primary is instructing us to trim
8865 ObjectStore::Transaction t
;
8866 pg
->pg_log
.trim(m
->trim_to
, pg
->info
);
8867 pg
->dirty_info
= true;
8868 pg
->write_if_dirty(t
);
8869 int tr
= store
->queue_transaction(pg
->osr
.get(), std::move(t
), NULL
);
8875 void OSD::handle_pg_backfill_reserve(OpRequestRef op
)
8877 const MBackfillReserve
*m
= static_cast<const MBackfillReserve
*>(op
->get_req());
8878 assert(m
->get_type() == MSG_OSD_BACKFILL_RESERVE
);
8880 if (!require_osd_peer(op
->get_req()))
8882 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
8885 PG::CephPeeringEvtRef evt
;
8886 if (m
->type
== MBackfillReserve::REQUEST
) {
8887 evt
= PG::CephPeeringEvtRef(
8888 new PG::CephPeeringEvt(
8891 PG::RequestBackfillPrio(m
->priority
)));
8892 } else if (m
->type
== MBackfillReserve::GRANT
) {
8893 evt
= PG::CephPeeringEvtRef(
8894 new PG::CephPeeringEvt(
8897 PG::RemoteBackfillReserved()));
8898 } else if (m
->type
== MBackfillReserve::REJECT
) {
8899 evt
= PG::CephPeeringEvtRef(
8900 new PG::CephPeeringEvt(
8903 PG::RemoteReservationRejected()));
8908 if (service
.splitting(m
->pgid
)) {
8909 peering_wait_for_split
[m
->pgid
].push_back(evt
);
8913 PG
*pg
= _lookup_lock_pg(m
->pgid
);
8915 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
8919 pg
->queue_peering_event(evt
);
8923 void OSD::handle_pg_recovery_reserve(OpRequestRef op
)
8925 const MRecoveryReserve
*m
= static_cast<const MRecoveryReserve
*>(op
->get_req());
8926 assert(m
->get_type() == MSG_OSD_RECOVERY_RESERVE
);
8928 if (!require_osd_peer(op
->get_req()))
8930 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
8933 PG::CephPeeringEvtRef evt
;
8934 if (m
->type
== MRecoveryReserve::REQUEST
) {
8935 evt
= PG::CephPeeringEvtRef(
8936 new PG::CephPeeringEvt(
8939 PG::RequestRecovery()));
8940 } else if (m
->type
== MRecoveryReserve::GRANT
) {
8941 evt
= PG::CephPeeringEvtRef(
8942 new PG::CephPeeringEvt(
8945 PG::RemoteRecoveryReserved()));
8946 } else if (m
->type
== MRecoveryReserve::RELEASE
) {
8947 evt
= PG::CephPeeringEvtRef(
8948 new PG::CephPeeringEvt(
8951 PG::RecoveryDone()));
8956 if (service
.splitting(m
->pgid
)) {
8957 peering_wait_for_split
[m
->pgid
].push_back(evt
);
8961 PG
*pg
= _lookup_lock_pg(m
->pgid
);
8963 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
8967 pg
->queue_peering_event(evt
);
8971 void OSD::handle_force_recovery(Message
*m
)
8973 MOSDForceRecovery
*msg
= static_cast<MOSDForceRecovery
*>(m
);
8974 assert(msg
->get_type() == MSG_OSD_FORCE_RECOVERY
);
8976 vector
<PGRef
> local_pgs
;
8977 local_pgs
.reserve(msg
->forced_pgs
.size());
8980 RWLock::RLocker
l(pg_map_lock
);
8981 for (auto& i
: msg
->forced_pgs
) {
8983 if (osdmap
->get_primary_shard(i
, &locpg
)) {
8984 auto pg_map_entry
= pg_map
.find(locpg
);
8985 if (pg_map_entry
!= pg_map
.end()) {
8986 local_pgs
.push_back(pg_map_entry
->second
);
8992 if (local_pgs
.size()) {
8993 service
.adjust_pg_priorities(local_pgs
, msg
->options
);
9000 * from primary to replica | stray
9001 * NOTE: called with opqueue active.
9003 void OSD::handle_pg_query(OpRequestRef op
)
9005 assert(osd_lock
.is_locked());
9007 const MOSDPGQuery
*m
= static_cast<const MOSDPGQuery
*>(op
->get_req());
9008 assert(m
->get_type() == MSG_OSD_PG_QUERY
);
9010 if (!require_osd_peer(op
->get_req()))
9013 dout(7) << "handle_pg_query from " << m
->get_source() << " epoch " << m
->get_epoch() << dendl
;
9014 int from
= m
->get_source().num();
9016 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9021 map
< int, vector
<pair
<pg_notify_t
, PastIntervals
> > > notify_list
;
9023 for (auto it
= m
->pg_list
.begin();
9024 it
!= m
->pg_list
.end();
9026 spg_t pgid
= it
->first
;
9028 if (pgid
.preferred() >= 0) {
9029 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9033 if (service
.splitting(pgid
)) {
9034 peering_wait_for_split
[pgid
].push_back(
9035 PG::CephPeeringEvtRef(
9036 new PG::CephPeeringEvt(
9037 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9038 PG::MQuery(pg_shard_t(from
, it
->second
.from
),
9039 it
->second
, it
->second
.epoch_sent
))));
9044 RWLock::RLocker
l(pg_map_lock
);
9045 if (pg_map
.count(pgid
)) {
9047 pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9049 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9050 pg_shard_t(from
, it
->second
.from
), it
->second
);
9056 if (!osdmap
->have_pg_pool(pgid
.pool()))
9059 // get active crush mapping
9060 int up_primary
, acting_primary
;
9061 vector
<int> up
, acting
;
9062 osdmap
->pg_to_up_acting_osds(
9063 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9066 pg_history_t history
= it
->second
.history
;
9067 bool valid_history
= project_pg_history(
9068 pgid
, history
, it
->second
.epoch_sent
,
9069 up
, up_primary
, acting
, acting_primary
);
9071 if (!valid_history
||
9072 it
->second
.epoch_sent
< history
.same_interval_since
) {
9073 dout(10) << " pg " << pgid
<< " dne, and pg has changed in "
9074 << history
.same_interval_since
9075 << " (msg from " << it
->second
.epoch_sent
<< ")" << dendl
;
9079 dout(10) << " pg " << pgid
<< " dne" << dendl
;
9080 pg_info_t
empty(spg_t(pgid
.pgid
, it
->second
.to
));
9081 /* This is racy, but that should be ok: if we complete the deletion
9082 * before the pg is recreated, we'll just start it off backfilling
9083 * instead of just empty */
9084 if (service
.deleting_pgs
.lookup(pgid
))
9085 empty
.set_last_backfill(hobject_t());
9086 if (it
->second
.type
== pg_query_t::LOG
||
9087 it
->second
.type
== pg_query_t::FULLLOG
) {
9088 ConnectionRef con
= service
.get_con_osd_cluster(from
, osdmap
->get_epoch());
9090 MOSDPGLog
*mlog
= new MOSDPGLog(
9091 it
->second
.from
, it
->second
.to
,
9092 osdmap
->get_epoch(), empty
,
9093 it
->second
.epoch_sent
);
9094 service
.share_map_peer(from
, con
.get(), osdmap
);
9095 con
->send_message(mlog
);
9098 notify_list
[from
].push_back(
9101 it
->second
.from
, it
->second
.to
,
9102 it
->second
.epoch_sent
,
9103 osdmap
->get_epoch(),
9106 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
9110 do_notifies(notify_list
, osdmap
);
9114 void OSD::handle_pg_remove(OpRequestRef op
)
9116 const MOSDPGRemove
*m
= static_cast<const MOSDPGRemove
*>(op
->get_req());
9117 assert(m
->get_type() == MSG_OSD_PG_REMOVE
);
9118 assert(osd_lock
.is_locked());
9120 if (!require_osd_peer(op
->get_req()))
9123 dout(7) << "handle_pg_remove from " << m
->get_source() << " on "
9124 << m
->pg_list
.size() << " pgs" << dendl
;
9126 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9131 for (auto it
= m
->pg_list
.begin();
9132 it
!= m
->pg_list
.end();
9135 if (pgid
.preferred() >= 0) {
9136 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9140 RWLock::WLocker
l(pg_map_lock
);
9141 if (pg_map
.count(pgid
) == 0) {
9142 dout(10) << " don't have pg " << pgid
<< dendl
;
9145 dout(5) << "queue_pg_for_deletion: " << pgid
<< dendl
;
9146 PG
*pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9147 pg_history_t history
= pg
->info
.history
;
9148 int up_primary
, acting_primary
;
9149 vector
<int> up
, acting
;
9150 osdmap
->pg_to_up_acting_osds(
9151 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9152 bool valid_history
= project_pg_history(
9153 pg
->info
.pgid
, history
, pg
->get_osdmap()->get_epoch(),
9154 up
, up_primary
, acting
, acting_primary
);
9155 if (valid_history
&&
9156 history
.same_interval_since
<= m
->get_epoch()) {
9157 assert(pg
->get_primary().osd
== m
->get_source().num());
9162 dout(10) << *pg
<< " ignoring remove request, pg changed in epoch "
9163 << history
.same_interval_since
9164 << " > " << m
->get_epoch() << dendl
;
9170 void OSD::_remove_pg(PG
*pg
)
9172 ObjectStore::Transaction rmt
;
9174 // on_removal, which calls remove_watchers_and_notifies, and the erasure from
9175 // the pg_map must be done together without unlocking the pg lock,
9176 // to avoid racing with watcher cleanup in ms_handle_reset
9177 // and handle_notify_timeout
9178 pg
->on_removal(&rmt
);
9180 service
.cancel_pending_splits_for_parent(pg
->info
.pgid
);
9181 int tr
= store
->queue_transaction(
9182 pg
->osr
.get(), std::move(rmt
), NULL
,
9183 new ContainerContext
<
9184 SequencerRef
>(pg
->osr
));
9187 DeletingStateRef deleting
= service
.deleting_pgs
.lookup_or_create(
9193 remove_wq
.queue(make_pair(PGRef(pg
), deleting
));
9195 service
.pg_remove_epoch(pg
->info
.pgid
);
9197 // dereference from op_wq
9198 op_shardedwq
.clear_pg_pointer(pg
->info
.pgid
);
9201 pg_map
.erase(pg
->info
.pgid
);
9202 pg
->put("PGMap"); // since we've taken it out of map
9206 // =========================================================
9209 void OSDService::_maybe_queue_recovery() {
9210 assert(recovery_lock
.is_locked_by_me());
9211 uint64_t available_pushes
;
9212 while (!awaiting_throttle
.empty() &&
9213 _recover_now(&available_pushes
)) {
9214 uint64_t to_start
= MIN(
9216 cct
->_conf
->osd_recovery_max_single_start
);
9217 _queue_for_recovery(awaiting_throttle
.front(), to_start
);
9218 awaiting_throttle
.pop_front();
9219 recovery_ops_reserved
+= to_start
;
9223 bool OSDService::_recover_now(uint64_t *available_pushes
)
9225 if (available_pushes
)
9226 *available_pushes
= 0;
9228 if (ceph_clock_now() < defer_recovery_until
) {
9229 dout(15) << __func__
<< " defer until " << defer_recovery_until
<< dendl
;
9233 if (recovery_paused
) {
9234 dout(15) << __func__
<< " paused" << dendl
;
9238 uint64_t max
= cct
->_conf
->osd_recovery_max_active
;
9239 if (max
<= recovery_ops_active
+ recovery_ops_reserved
) {
9240 dout(15) << __func__
<< " active " << recovery_ops_active
9241 << " + reserved " << recovery_ops_reserved
9242 << " >= max " << max
<< dendl
;
9246 if (available_pushes
)
9247 *available_pushes
= max
- recovery_ops_active
- recovery_ops_reserved
;
9253 void OSDService::adjust_pg_priorities(const vector
<PGRef
>& pgs
, int newflags
)
9255 if (!pgs
.size() || !(newflags
& (OFR_BACKFILL
| OFR_RECOVERY
)))
9259 if (newflags
& OFR_BACKFILL
) {
9260 newstate
= PG_STATE_FORCED_BACKFILL
;
9261 } else if (newflags
& OFR_RECOVERY
) {
9262 newstate
= PG_STATE_FORCED_RECOVERY
;
9265 // debug output here may get large, don't generate it if debug level is below
9266 // 10 and use abbreviated pg ids otherwise
9267 if ((cct
)->_conf
->subsys
.should_gather(ceph_subsys_osd
, 10)) {
9270 for (auto& i
: pgs
) {
9271 ss
<< i
->get_pgid() << " ";
9274 dout(10) << __func__
<< " working on " << ss
.str() << dendl
;
9277 if (newflags
& OFR_CANCEL
) {
9278 for (auto& i
: pgs
) {
9280 i
->_change_recovery_force_mode(newstate
, true);
9284 for (auto& i
: pgs
) {
9285 // make sure the PG is in correct state before forcing backfill or recovery, or
9286 // else we'll make PG keeping FORCE_* flag forever, requiring osds restart
9287 // or forcing somehow recovery/backfill.
9289 int pgstate
= i
->get_state();
9290 if ( ((newstate
== PG_STATE_FORCED_RECOVERY
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_RECOVERY_WAIT
| PG_STATE_RECOVERING
))) ||
9291 ((newstate
== PG_STATE_FORCED_BACKFILL
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_BACKFILL_WAIT
| PG_STATE_BACKFILL
))) )
9292 i
->_change_recovery_force_mode(newstate
, false);
9298 void OSD::do_recovery(
9299 PG
*pg
, epoch_t queued
, uint64_t reserved_pushes
,
9300 ThreadPool::TPHandle
&handle
)
9302 uint64_t started
= 0;
9305 * When the value of osd_recovery_sleep is set greater than zero, recovery
9306 * ops are scheduled after osd_recovery_sleep amount of time from the previous
9307 * recovery event's schedule time. This is done by adding a
9308 * recovery_requeue_callback event, which re-queues the recovery op using
9309 * queue_recovery_after_sleep.
9311 float recovery_sleep
= get_osd_recovery_sleep();
9312 if (recovery_sleep
> 0 && service
.recovery_needs_sleep
) {
9314 auto recovery_requeue_callback
= new FunctionContext([this, pgref
, queued
, reserved_pushes
](int r
) {
9315 dout(20) << "do_recovery wake up at "
9317 << ", re-queuing recovery" << dendl
;
9318 service
.recovery_needs_sleep
= false;
9319 service
.queue_recovery_after_sleep(pgref
.get(), queued
, reserved_pushes
);
9321 Mutex::Locker
l(service
.recovery_sleep_lock
);
9323 // This is true for the first recovery op and when the previous recovery op
9324 // has been scheduled in the past. The next recovery op is scheduled after
9325 // completing the sleep from now.
9326 if (service
.recovery_schedule_time
< ceph_clock_now()) {
9327 service
.recovery_schedule_time
= ceph_clock_now();
9329 service
.recovery_schedule_time
+= recovery_sleep
;
9330 service
.recovery_sleep_timer
.add_event_at(service
.recovery_schedule_time
,
9331 recovery_requeue_callback
);
9332 dout(20) << "Recovery event scheduled at "
9333 << service
.recovery_schedule_time
<< dendl
;
9338 service
.recovery_needs_sleep
= true;
9339 if (pg
->pg_has_reset_since(queued
)) {
9343 assert(!pg
->deleting
);
9344 assert(pg
->is_peered() && pg
->is_primary());
9346 assert(pg
->recovery_queued
);
9347 pg
->recovery_queued
= false;
9349 dout(10) << "do_recovery starting " << reserved_pushes
<< " " << *pg
<< dendl
;
9350 #ifdef DEBUG_RECOVERY_OIDS
9351 dout(20) << " active was " << service
.recovery_oids
[pg
->info
.pgid
] << dendl
;
9354 bool more
= pg
->start_recovery_ops(reserved_pushes
, handle
, &started
);
9355 dout(10) << "do_recovery started " << started
<< "/" << reserved_pushes
9356 << " on " << *pg
<< dendl
;
9358 // If no recovery op is started, don't bother to manipulate the RecoveryCtx
9359 if (!started
&& (more
|| !pg
->have_unfound())) {
9363 PG::RecoveryCtx rctx
= create_context();
9364 rctx
.handle
= &handle
;
9367 * if we couldn't start any recovery ops and things are still
9368 * unfound, see if we can discover more missing object locations.
9369 * It may be that our initial locations were bad and we errored
9370 * out while trying to pull.
9372 if (!more
&& pg
->have_unfound()) {
9373 pg
->discover_all_missing(*rctx
.query_map
);
9374 if (rctx
.query_map
->empty()) {
9376 if (pg
->state_test(PG_STATE_BACKFILL
)) {
9377 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9380 PG::CancelBackfill()));
9381 pg
->queue_peering_event(evt
);
9382 action
= "in backfill";
9383 } else if (pg
->state_test(PG_STATE_RECOVERING
)) {
9384 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9387 PG::CancelRecovery()));
9388 pg
->queue_peering_event(evt
);
9389 action
= "in recovery";
9391 action
= "already out of recovery/backfill";
9393 dout(10) << __func__
<< ": no luck, giving up on this pg for now (" << action
<< ")" << dendl
;
9395 dout(10) << __func__
<< ": no luck, giving up on this pg for now (queue_recovery)" << dendl
;
9396 pg
->queue_recovery();
9400 pg
->write_if_dirty(*rctx
.transaction
);
9401 OSDMapRef curmap
= pg
->get_osdmap();
9402 dispatch_context(rctx
, pg
, curmap
);
9406 assert(started
<= reserved_pushes
);
9407 service
.release_reserved_pushes(reserved_pushes
);
9410 void OSDService::start_recovery_op(PG
*pg
, const hobject_t
& soid
)
9412 Mutex::Locker
l(recovery_lock
);
9413 dout(10) << "start_recovery_op " << *pg
<< " " << soid
9414 << " (" << recovery_ops_active
<< "/"
9415 << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9417 recovery_ops_active
++;
9419 #ifdef DEBUG_RECOVERY_OIDS
9420 dout(20) << " active was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9421 assert(recovery_oids
[pg
->info
.pgid
].count(soid
) == 0);
9422 recovery_oids
[pg
->info
.pgid
].insert(soid
);
9426 void OSDService::finish_recovery_op(PG
*pg
, const hobject_t
& soid
, bool dequeue
)
9428 Mutex::Locker
l(recovery_lock
);
9429 dout(10) << "finish_recovery_op " << *pg
<< " " << soid
9430 << " dequeue=" << dequeue
9431 << " (" << recovery_ops_active
<< "/" << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9435 assert(recovery_ops_active
> 0);
9436 recovery_ops_active
--;
9438 #ifdef DEBUG_RECOVERY_OIDS
9439 dout(20) << " active oids was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9440 assert(recovery_oids
[pg
->info
.pgid
].count(soid
));
9441 recovery_oids
[pg
->info
.pgid
].erase(soid
);
9444 _maybe_queue_recovery();
9447 bool OSDService::is_recovery_active()
9449 Mutex::Locker
l(recovery_lock
);
9450 return recovery_ops_active
> 0;
9453 // =========================================================
9456 bool OSD::op_is_discardable(const MOSDOp
*op
)
9458 // drop client request if they are not connected and can't get the
9460 if (!op
->get_connection()->is_connected()) {
9466 void OSD::enqueue_op(spg_t pg
, OpRequestRef
& op
, epoch_t epoch
)
9468 utime_t latency
= ceph_clock_now() - op
->get_req()->get_recv_stamp();
9469 dout(15) << "enqueue_op " << op
<< " prio " << op
->get_req()->get_priority()
9470 << " cost " << op
->get_req()->get_cost()
9471 << " latency " << latency
9472 << " epoch " << epoch
9473 << " " << *(op
->get_req()) << dendl
;
9474 op
->osd_trace
.event("enqueue op");
9475 op
->osd_trace
.keyval("priority", op
->get_req()->get_priority());
9476 op
->osd_trace
.keyval("cost", op
->get_req()->get_cost());
9477 op
->mark_queued_for_pg();
9478 logger
->tinc(l_osd_op_before_queue_op_lat
, latency
);
9479 op_shardedwq
.queue(make_pair(pg
, PGQueueable(op
, epoch
)));
9485 * NOTE: dequeue called in worker thread, with pg lock
9487 void OSD::dequeue_op(
9488 PGRef pg
, OpRequestRef op
,
9489 ThreadPool::TPHandle
&handle
)
9492 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_BEGIN", false);
9494 utime_t now
= ceph_clock_now();
9495 op
->set_dequeued_time(now
);
9496 utime_t latency
= now
- op
->get_req()->get_recv_stamp();
9497 dout(10) << "dequeue_op " << op
<< " prio " << op
->get_req()->get_priority()
9498 << " cost " << op
->get_req()->get_cost()
9499 << " latency " << latency
9500 << " " << *(op
->get_req())
9501 << " pg " << *pg
<< dendl
;
9503 logger
->tinc(l_osd_op_before_dequeue_op_lat
, latency
);
9505 Session
*session
= static_cast<Session
*>(
9506 op
->get_req()->get_connection()->get_priv());
9508 maybe_share_map(session
, op
, pg
->get_osdmap());
9515 op
->mark_reached_pg();
9516 op
->osd_trace
.event("dequeue_op");
9518 pg
->do_request(op
, handle
);
9521 dout(10) << "dequeue_op " << op
<< " finish" << dendl
;
9522 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_END", false);
9526 struct C_CompleteSplits
: public Context
{
9529 C_CompleteSplits(OSD
*osd
, const set
<PGRef
> &in
)
9530 : osd(osd
), pgs(in
) {}
9531 void finish(int r
) override
{
9532 Mutex::Locker
l(osd
->osd_lock
);
9533 if (osd
->is_stopping())
9535 PG::RecoveryCtx rctx
= osd
->create_context();
9536 for (set
<PGRef
>::iterator i
= pgs
.begin();
9539 osd
->pg_map_lock
.get_write();
9542 osd
->add_newly_split_pg(pg
, &rctx
);
9543 if (!((*i
)->deleting
)) {
9544 set
<spg_t
> to_complete
;
9545 to_complete
.insert((*i
)->info
.pgid
);
9546 osd
->service
.complete_split(to_complete
);
9548 osd
->pg_map_lock
.put_write();
9549 osd
->dispatch_context_transaction(rctx
, pg
);
9550 osd
->wake_pg_waiters(*i
);
9554 osd
->dispatch_context(rctx
, 0, osd
->service
.get_osdmap());
9558 void OSD::process_peering_events(
9559 const list
<PG
*> &pgs
,
9560 ThreadPool::TPHandle
&handle
9563 bool need_up_thru
= false;
9564 epoch_t same_interval_since
= 0;
9566 PG::RecoveryCtx rctx
= create_context();
9567 rctx
.handle
= &handle
;
9568 for (list
<PG
*>::const_iterator i
= pgs
.begin();
9571 set
<PGRef
> split_pgs
;
9573 pg
->lock_suspend_timeout(handle
);
9574 curmap
= service
.get_osdmap();
9579 if (!advance_pg(curmap
->get_epoch(), pg
, handle
, &rctx
, &split_pgs
)) {
9580 // we need to requeue the PG explicitly since we didn't actually
9582 peering_wq
.queue(pg
);
9584 assert(!pg
->peering_queue
.empty());
9585 PG::CephPeeringEvtRef evt
= pg
->peering_queue
.front();
9586 pg
->peering_queue
.pop_front();
9587 pg
->handle_peering_event(evt
, &rctx
);
9589 need_up_thru
= pg
->need_up_thru
|| need_up_thru
;
9590 same_interval_since
= MAX(pg
->info
.history
.same_interval_since
,
9591 same_interval_since
);
9592 pg
->write_if_dirty(*rctx
.transaction
);
9593 if (!split_pgs
.empty()) {
9594 rctx
.on_applied
->add(new C_CompleteSplits(this, split_pgs
));
9597 dispatch_context_transaction(rctx
, pg
, &handle
);
9601 queue_want_up_thru(same_interval_since
);
9602 dispatch_context(rctx
, 0, curmap
, &handle
);
9604 service
.send_pg_temp();
9607 // --------------------------------
9609 const char** OSD::get_tracked_conf_keys() const
9611 static const char* KEYS
[] = {
9612 "osd_max_backfills",
9613 "osd_min_recovery_priority",
9614 "osd_max_trimming_pgs",
9615 "osd_op_complaint_time",
9616 "osd_op_log_threshold",
9617 "osd_op_history_size",
9618 "osd_op_history_duration",
9619 "osd_op_history_slow_op_size",
9620 "osd_op_history_slow_op_threshold",
9621 "osd_enable_op_tracker",
9622 "osd_map_cache_size",
9623 "osd_map_max_advance",
9624 "osd_pg_epoch_persisted_max_stale",
9625 "osd_disk_thread_ioprio_class",
9626 "osd_disk_thread_ioprio_priority",
9627 // clog & admin clog
9630 "clog_to_syslog_facility",
9631 "clog_to_syslog_level",
9632 "osd_objectstore_fuse",
9634 "clog_to_graylog_host",
9635 "clog_to_graylog_port",
9638 "osd_recovery_delay_start",
9639 "osd_client_message_size_cap",
9640 "osd_client_message_cap",
9641 "osd_heartbeat_min_size",
9642 "osd_heartbeat_interval",
9648 void OSD::handle_conf_change(const struct md_config_t
*conf
,
9649 const std::set
<std::string
> &changed
)
9651 if (changed
.count("osd_max_backfills")) {
9652 service
.local_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9653 service
.remote_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9655 if (changed
.count("osd_min_recovery_priority")) {
9656 service
.local_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9657 service
.remote_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9659 if (changed
.count("osd_max_trimming_pgs")) {
9660 service
.snap_reserver
.set_max(cct
->_conf
->osd_max_trimming_pgs
);
9662 if (changed
.count("osd_op_complaint_time") ||
9663 changed
.count("osd_op_log_threshold")) {
9664 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
9665 cct
->_conf
->osd_op_log_threshold
);
9667 if (changed
.count("osd_op_history_size") ||
9668 changed
.count("osd_op_history_duration")) {
9669 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
9670 cct
->_conf
->osd_op_history_duration
);
9672 if (changed
.count("osd_op_history_slow_op_size") ||
9673 changed
.count("osd_op_history_slow_op_threshold")) {
9674 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
9675 cct
->_conf
->osd_op_history_slow_op_threshold
);
9677 if (changed
.count("osd_enable_op_tracker")) {
9678 op_tracker
.set_tracking(cct
->_conf
->osd_enable_op_tracker
);
9680 if (changed
.count("osd_disk_thread_ioprio_class") ||
9681 changed
.count("osd_disk_thread_ioprio_priority")) {
9682 set_disk_tp_priority();
9684 if (changed
.count("osd_map_cache_size")) {
9685 service
.map_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9686 service
.map_bl_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9687 service
.map_bl_inc_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9689 if (changed
.count("clog_to_monitors") ||
9690 changed
.count("clog_to_syslog") ||
9691 changed
.count("clog_to_syslog_level") ||
9692 changed
.count("clog_to_syslog_facility") ||
9693 changed
.count("clog_to_graylog") ||
9694 changed
.count("clog_to_graylog_host") ||
9695 changed
.count("clog_to_graylog_port") ||
9696 changed
.count("host") ||
9697 changed
.count("fsid")) {
9698 update_log_config();
9702 if (changed
.count("osd_objectstore_fuse")) {
9704 enable_disable_fuse(false);
9709 if (changed
.count("osd_recovery_delay_start")) {
9710 service
.defer_recovery(cct
->_conf
->osd_recovery_delay_start
);
9711 service
.kick_recovery_queue();
9714 if (changed
.count("osd_client_message_cap")) {
9715 uint64_t newval
= cct
->_conf
->osd_client_message_cap
;
9716 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
9717 if (pol
.throttler_messages
&& newval
> 0) {
9718 pol
.throttler_messages
->reset_max(newval
);
9721 if (changed
.count("osd_client_message_size_cap")) {
9722 uint64_t newval
= cct
->_conf
->osd_client_message_size_cap
;
9723 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
9724 if (pol
.throttler_bytes
&& newval
> 0) {
9725 pol
.throttler_bytes
->reset_max(newval
);
9732 void OSD::update_log_config()
9734 map
<string
,string
> log_to_monitors
;
9735 map
<string
,string
> log_to_syslog
;
9736 map
<string
,string
> log_channel
;
9737 map
<string
,string
> log_prio
;
9738 map
<string
,string
> log_to_graylog
;
9739 map
<string
,string
> log_to_graylog_host
;
9740 map
<string
,string
> log_to_graylog_port
;
9744 if (parse_log_client_options(cct
, log_to_monitors
, log_to_syslog
,
9745 log_channel
, log_prio
, log_to_graylog
,
9746 log_to_graylog_host
, log_to_graylog_port
,
9748 clog
->update_config(log_to_monitors
, log_to_syslog
,
9749 log_channel
, log_prio
, log_to_graylog
,
9750 log_to_graylog_host
, log_to_graylog_port
,
9752 derr
<< "log_to_monitors " << log_to_monitors
<< dendl
;
9755 void OSD::check_config()
9757 // some sanity checks
9758 if (cct
->_conf
->osd_map_cache_size
<= cct
->_conf
->osd_map_max_advance
+ 2) {
9759 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
9760 << " is not > osd_map_max_advance ("
9761 << cct
->_conf
->osd_map_max_advance
<< ")";
9763 if (cct
->_conf
->osd_map_cache_size
<= (int)cct
->_conf
->osd_pg_epoch_persisted_max_stale
+ 2) {
9764 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
9765 << " is not > osd_pg_epoch_persisted_max_stale ("
9766 << cct
->_conf
->osd_pg_epoch_persisted_max_stale
<< ")";
9770 void OSD::set_disk_tp_priority()
9772 dout(10) << __func__
9773 << " class " << cct
->_conf
->osd_disk_thread_ioprio_class
9774 << " priority " << cct
->_conf
->osd_disk_thread_ioprio_priority
9776 if (cct
->_conf
->osd_disk_thread_ioprio_class
.empty() ||
9777 cct
->_conf
->osd_disk_thread_ioprio_priority
< 0)
9780 ceph_ioprio_string_to_class(cct
->_conf
->osd_disk_thread_ioprio_class
);
9782 derr
<< __func__
<< cpp_strerror(cls
) << ": "
9783 << "osd_disk_thread_ioprio_class is " << cct
->_conf
->osd_disk_thread_ioprio_class
9784 << " but only the following values are allowed: idle, be or rt" << dendl
;
9786 disk_tp
.set_ioprio(cls
, cct
->_conf
->osd_disk_thread_ioprio_priority
);
9789 // --------------------------------
9791 void OSD::get_latest_osdmap()
9793 dout(10) << __func__
<< " -- start" << dendl
;
9796 service
.objecter
->wait_for_latest_osdmap(&cond
);
9799 dout(10) << __func__
<< " -- finish" << dendl
;
9802 // --------------------------------
9804 int OSD::init_op_flags(OpRequestRef
& op
)
9806 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
9807 vector
<OSDOp
>::const_iterator iter
;
9809 // client flags have no bearing on whether an op is a read, write, etc.
9812 if (m
->has_flag(CEPH_OSD_FLAG_RWORDERED
)) {
9813 op
->set_force_rwordered();
9816 // set bits based on op codes, called methods.
9817 for (iter
= m
->ops
.begin(); iter
!= m
->ops
.end(); ++iter
) {
9818 if ((iter
->op
.op
== CEPH_OSD_OP_WATCH
&&
9819 iter
->op
.watch
.op
== CEPH_OSD_WATCH_OP_PING
)) {
9820 /* This a bit odd. PING isn't actually a write. It can't
9821 * result in an update to the object_info. PINGs also aren'ty
9822 * resent, so there's no reason to write out a log entry
9824 * However, we pipeline them behind writes, so let's force
9825 * the write_ordered flag.
9827 op
->set_force_rwordered();
9829 if (ceph_osd_op_mode_modify(iter
->op
.op
))
9832 if (ceph_osd_op_mode_read(iter
->op
.op
))
9835 // set READ flag if there are src_oids
9836 if (iter
->soid
.oid
.name
.length())
9839 // set PGOP flag if there are PG ops
9840 if (ceph_osd_op_type_pg(iter
->op
.op
))
9843 if (ceph_osd_op_mode_cache(iter
->op
.op
))
9846 // check for ec base pool
9847 int64_t poolid
= m
->get_pg().pool();
9848 const pg_pool_t
*pool
= osdmap
->get_pg_pool(poolid
);
9849 if (pool
&& pool
->is_tier()) {
9850 const pg_pool_t
*base_pool
= osdmap
->get_pg_pool(pool
->tier_of
);
9851 if (base_pool
&& base_pool
->require_rollback()) {
9852 if ((iter
->op
.op
!= CEPH_OSD_OP_READ
) &&
9853 (iter
->op
.op
!= CEPH_OSD_OP_CHECKSUM
) &&
9854 (iter
->op
.op
!= CEPH_OSD_OP_CMPEXT
) &&
9855 (iter
->op
.op
!= CEPH_OSD_OP_STAT
) &&
9856 (iter
->op
.op
!= CEPH_OSD_OP_ISDIRTY
) &&
9857 (iter
->op
.op
!= CEPH_OSD_OP_UNDIRTY
) &&
9858 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTR
) &&
9859 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTRS
) &&
9860 (iter
->op
.op
!= CEPH_OSD_OP_CMPXATTR
) &&
9861 (iter
->op
.op
!= CEPH_OSD_OP_ASSERT_VER
) &&
9862 (iter
->op
.op
!= CEPH_OSD_OP_LIST_WATCHERS
) &&
9863 (iter
->op
.op
!= CEPH_OSD_OP_LIST_SNAPS
) &&
9864 (iter
->op
.op
!= CEPH_OSD_OP_SETALLOCHINT
) &&
9865 (iter
->op
.op
!= CEPH_OSD_OP_WRITEFULL
) &&
9866 (iter
->op
.op
!= CEPH_OSD_OP_ROLLBACK
) &&
9867 (iter
->op
.op
!= CEPH_OSD_OP_CREATE
) &&
9868 (iter
->op
.op
!= CEPH_OSD_OP_DELETE
) &&
9869 (iter
->op
.op
!= CEPH_OSD_OP_SETXATTR
) &&
9870 (iter
->op
.op
!= CEPH_OSD_OP_RMXATTR
) &&
9871 (iter
->op
.op
!= CEPH_OSD_OP_STARTSYNC
) &&
9872 (iter
->op
.op
!= CEPH_OSD_OP_COPY_GET
) &&
9873 (iter
->op
.op
!= CEPH_OSD_OP_COPY_FROM
)) {
9879 switch (iter
->op
.op
) {
9880 case CEPH_OSD_OP_CALL
:
9882 bufferlist::iterator bp
= const_cast<bufferlist
&>(iter
->indata
).begin();
9883 int is_write
, is_read
;
9884 string cname
, mname
;
9885 bp
.copy(iter
->op
.cls
.class_len
, cname
);
9886 bp
.copy(iter
->op
.cls
.method_len
, mname
);
9888 ClassHandler::ClassData
*cls
;
9889 int r
= class_handler
->open_class(cname
, &cls
);
9891 derr
<< "class " << cname
<< " open got " << cpp_strerror(r
) << dendl
;
9894 else if (r
!= -EPERM
) // propagate permission errors
9898 int flags
= cls
->get_method_flags(mname
.c_str());
9900 if (flags
== -ENOENT
)
9906 is_read
= flags
& CLS_METHOD_RD
;
9907 is_write
= flags
& CLS_METHOD_WR
;
9908 bool is_promote
= flags
& CLS_METHOD_PROMOTE
;
9910 dout(10) << "class " << cname
<< " method " << mname
<< " "
9911 << "flags=" << (is_read
? "r" : "")
9912 << (is_write
? "w" : "")
9913 << (is_promote
? "p" : "")
9916 op
->set_class_read();
9918 op
->set_class_write();
9921 op
->add_class(cname
, is_read
, is_write
, cls
->whitelisted
);
9925 case CEPH_OSD_OP_WATCH
:
9926 // force the read bit for watch since it is depends on previous
9927 // watch state (and may return early if the watch exists) or, in
9928 // the case of ping, is simply a read op.
9931 case CEPH_OSD_OP_NOTIFY
:
9932 case CEPH_OSD_OP_NOTIFY_ACK
:
9938 case CEPH_OSD_OP_DELETE
:
9939 // if we get a delete with FAILOK we can skip handle cache. without
9940 // FAILOK we still need to promote (or do something smarter) to
9941 // determine whether to return ENOENT or 0.
9942 if (iter
== m
->ops
.begin() &&
9943 iter
->op
.flags
== CEPH_OSD_OP_FLAG_FAILOK
) {
9944 op
->set_skip_handle_cache();
9946 // skip promotion when proxying a delete op
9947 if (m
->ops
.size() == 1) {
9948 op
->set_skip_promote();
9952 case CEPH_OSD_OP_CACHE_TRY_FLUSH
:
9953 case CEPH_OSD_OP_CACHE_FLUSH
:
9954 case CEPH_OSD_OP_CACHE_EVICT
:
9955 // If try_flush/flush/evict is the only op, can skip handle cache.
9956 if (m
->ops
.size() == 1) {
9957 op
->set_skip_handle_cache();
9961 case CEPH_OSD_OP_READ
:
9962 case CEPH_OSD_OP_SYNC_READ
:
9963 case CEPH_OSD_OP_SPARSE_READ
:
9964 case CEPH_OSD_OP_CHECKSUM
:
9965 case CEPH_OSD_OP_WRITEFULL
:
9966 if (m
->ops
.size() == 1 &&
9967 (iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
||
9968 iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
)) {
9969 op
->set_skip_promote();
9973 // force promotion when pin an object in cache tier
9974 case CEPH_OSD_OP_CACHE_PIN
:
9983 if (op
->rmw_flags
== 0)
9989 void OSD::PeeringWQ::_dequeue(list
<PG
*> *out
) {
9990 for (list
<PG
*>::iterator i
= peering_queue
.begin();
9991 i
!= peering_queue
.end() &&
9992 out
->size() < osd
->cct
->_conf
->osd_peering_wq_batch_size
;
9994 if (in_use
.count(*i
)) {
9998 peering_queue
.erase(i
++);
10001 in_use
.insert(out
->begin(), out
->end());
10005 // =============================================================
10007 #undef dout_context
10008 #define dout_context osd->cct
10010 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq "
10012 void OSD::ShardedOpWQ::wake_pg_waiters(spg_t pgid
)
10014 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10015 auto sdata
= shard_list
[shard_index
];
10016 bool queued
= false;
10017 unsigned pushes_to_free
= 0;
10019 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10020 auto p
= sdata
->pg_slots
.find(pgid
);
10021 if (p
!= sdata
->pg_slots
.end()) {
10022 dout(20) << __func__
<< " " << pgid
10023 << " to_process " << p
->second
.to_process
10024 << " waiting_for_pg=" << (int)p
->second
.waiting_for_pg
<< dendl
;
10025 for (auto i
= p
->second
.to_process
.rbegin();
10026 i
!= p
->second
.to_process
.rend();
10028 sdata
->_enqueue_front(make_pair(pgid
, *i
), osd
->op_prio_cutoff
);
10030 for (auto& q
: p
->second
.to_process
) {
10031 pushes_to_free
+= q
.get_reserved_pushes();
10033 p
->second
.to_process
.clear();
10034 p
->second
.waiting_for_pg
= false;
10035 ++p
->second
.requeue_seq
;
10039 if (pushes_to_free
> 0) {
10040 osd
->service
.release_reserved_pushes(pushes_to_free
);
10043 sdata
->sdata_lock
.Lock();
10044 sdata
->sdata_cond
.SignalOne();
10045 sdata
->sdata_lock
.Unlock();
10049 void OSD::ShardedOpWQ::prune_pg_waiters(OSDMapRef osdmap
, int whoami
)
10051 unsigned pushes_to_free
= 0;
10052 for (auto sdata
: shard_list
) {
10053 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10054 sdata
->waiting_for_pg_osdmap
= osdmap
;
10055 auto p
= sdata
->pg_slots
.begin();
10056 while (p
!= sdata
->pg_slots
.end()) {
10057 ShardData::pg_slot
& slot
= p
->second
;
10058 if (!slot
.to_process
.empty() && slot
.num_running
== 0) {
10059 if (osdmap
->is_up_acting_osd_shard(p
->first
, whoami
)) {
10060 dout(20) << __func__
<< " " << p
->first
<< " maps to us, keeping"
10065 while (!slot
.to_process
.empty() &&
10066 slot
.to_process
.front().get_map_epoch() <= osdmap
->get_epoch()) {
10067 auto& qi
= slot
.to_process
.front();
10068 dout(20) << __func__
<< " " << p
->first
10070 << " epoch " << qi
.get_map_epoch()
10071 << " <= " << osdmap
->get_epoch()
10072 << ", stale, dropping" << dendl
;
10073 pushes_to_free
+= qi
.get_reserved_pushes();
10074 slot
.to_process
.pop_front();
10077 if (slot
.to_process
.empty() &&
10078 slot
.num_running
== 0 &&
10080 dout(20) << __func__
<< " " << p
->first
<< " empty, pruning" << dendl
;
10081 p
= sdata
->pg_slots
.erase(p
);
10087 if (pushes_to_free
> 0) {
10088 osd
->service
.release_reserved_pushes(pushes_to_free
);
10092 void OSD::ShardedOpWQ::clear_pg_pointer(spg_t pgid
)
10094 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10095 auto sdata
= shard_list
[shard_index
];
10096 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10097 auto p
= sdata
->pg_slots
.find(pgid
);
10098 if (p
!= sdata
->pg_slots
.end()) {
10099 auto& slot
= p
->second
;
10100 dout(20) << __func__
<< " " << pgid
<< " pg " << slot
.pg
<< dendl
;
10101 assert(!slot
.pg
|| slot
.pg
->deleting
);
10106 void OSD::ShardedOpWQ::clear_pg_slots()
10108 for (auto sdata
: shard_list
) {
10109 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10110 sdata
->pg_slots
.clear();
10111 sdata
->waiting_for_pg_osdmap
.reset();
10112 // don't bother with reserved pushes; we are shutting down
10117 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq(" << shard_index << ") "
10119 void OSD::ShardedOpWQ::_process(uint32_t thread_index
, heartbeat_handle_d
*hb
)
10121 uint32_t shard_index
= thread_index
% num_shards
;
10122 ShardData
*sdata
= shard_list
[shard_index
];
10123 assert(NULL
!= sdata
);
10126 sdata
->sdata_op_ordering_lock
.Lock();
10127 if (sdata
->pqueue
->empty()) {
10128 dout(20) << __func__
<< " empty q, waiting" << dendl
;
10129 // optimistically sleep a moment; maybe another work item will come along.
10130 osd
->cct
->get_heartbeat_map()->reset_timeout(hb
,
10131 osd
->cct
->_conf
->threadpool_default_timeout
, 0);
10132 sdata
->sdata_lock
.Lock();
10133 sdata
->sdata_op_ordering_lock
.Unlock();
10134 sdata
->sdata_cond
.WaitInterval(sdata
->sdata_lock
,
10135 utime_t(osd
->cct
->_conf
->threadpool_empty_queue_max_wait
, 0));
10136 sdata
->sdata_lock
.Unlock();
10137 sdata
->sdata_op_ordering_lock
.Lock();
10138 if (sdata
->pqueue
->empty()) {
10139 sdata
->sdata_op_ordering_lock
.Unlock();
10143 pair
<spg_t
, PGQueueable
> item
= sdata
->pqueue
->dequeue();
10144 if (osd
->is_stopping()) {
10145 sdata
->sdata_op_ordering_lock
.Unlock();
10146 return; // OSD shutdown, discard.
10149 uint64_t requeue_seq
;
10151 auto& slot
= sdata
->pg_slots
[item
.first
];
10152 dout(30) << __func__
<< " " << item
.first
10153 << " to_process " << slot
.to_process
10154 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10155 slot
.to_process
.push_back(item
.second
);
10156 // note the requeue seq now...
10157 requeue_seq
= slot
.requeue_seq
;
10158 if (slot
.waiting_for_pg
) {
10159 // save ourselves a bit of effort
10160 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10161 << " queued, waiting_for_pg" << dendl
;
10162 sdata
->sdata_op_ordering_lock
.Unlock();
10166 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10167 << " queued" << dendl
;
10168 ++slot
.num_running
;
10170 sdata
->sdata_op_ordering_lock
.Unlock();
10172 osd
->service
.maybe_inject_dispatch_delay();
10174 // [lookup +] lock pg (if we have it)
10176 pg
= osd
->_lookup_lock_pg(item
.first
);
10181 osd
->service
.maybe_inject_dispatch_delay();
10183 boost::optional
<PGQueueable
> qi
;
10185 // we don't use a Mutex::Locker here because of the
10186 // osd->service.release_reserved_pushes() call below
10187 sdata
->sdata_op_ordering_lock
.Lock();
10189 auto q
= sdata
->pg_slots
.find(item
.first
);
10190 assert(q
!= sdata
->pg_slots
.end());
10191 auto& slot
= q
->second
;
10192 --slot
.num_running
;
10194 if (slot
.to_process
.empty()) {
10195 // raced with wake_pg_waiters or prune_pg_waiters
10196 dout(20) << __func__
<< " " << item
.first
<< " nothing queued" << dendl
;
10200 sdata
->sdata_op_ordering_lock
.Unlock();
10203 if (requeue_seq
!= slot
.requeue_seq
) {
10204 dout(20) << __func__
<< " " << item
.first
10205 << " requeue_seq " << slot
.requeue_seq
<< " > our "
10206 << requeue_seq
<< ", we raced with wake_pg_waiters"
10211 sdata
->sdata_op_ordering_lock
.Unlock();
10214 if (pg
&& !slot
.pg
&& !pg
->deleting
) {
10215 dout(20) << __func__
<< " " << item
.first
<< " set pg to " << pg
<< dendl
;
10218 dout(30) << __func__
<< " " << item
.first
<< " to_process " << slot
.to_process
10219 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10221 // make sure we're not already waiting for this pg
10222 if (slot
.waiting_for_pg
) {
10223 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10224 << " slot is waiting_for_pg" << dendl
;
10228 sdata
->sdata_op_ordering_lock
.Unlock();
10233 qi
= slot
.to_process
.front();
10234 slot
.to_process
.pop_front();
10235 dout(20) << __func__
<< " " << item
.first
<< " item " << *qi
10236 << " pg " << pg
<< dendl
;
10239 // should this pg shard exist on this osd in this (or a later) epoch?
10240 OSDMapRef osdmap
= sdata
->waiting_for_pg_osdmap
;
10241 if (osdmap
->is_up_acting_osd_shard(item
.first
, osd
->whoami
)) {
10242 dout(20) << __func__
<< " " << item
.first
10243 << " no pg, should exist, will wait" << " on " << *qi
<< dendl
;
10244 slot
.to_process
.push_front(*qi
);
10245 slot
.waiting_for_pg
= true;
10246 } else if (qi
->get_map_epoch() > osdmap
->get_epoch()) {
10247 dout(20) << __func__
<< " " << item
.first
<< " no pg, item epoch is "
10248 << qi
->get_map_epoch() << " > " << osdmap
->get_epoch()
10249 << ", will wait on " << *qi
<< dendl
;
10250 slot
.to_process
.push_front(*qi
);
10251 slot
.waiting_for_pg
= true;
10253 dout(20) << __func__
<< " " << item
.first
<< " no pg, shouldn't exist,"
10254 << " dropping " << *qi
<< dendl
;
10255 // share map with client?
10256 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10257 Session
*session
= static_cast<Session
*>(
10258 (*_op
)->get_req()->get_connection()->get_priv());
10260 osd
->maybe_share_map(session
, *_op
, sdata
->waiting_for_pg_osdmap
);
10264 unsigned pushes_to_free
= qi
->get_reserved_pushes();
10265 if (pushes_to_free
> 0) {
10266 sdata
->sdata_op_ordering_lock
.Unlock();
10267 osd
->service
.release_reserved_pushes(pushes_to_free
);
10271 sdata
->sdata_op_ordering_lock
.Unlock();
10274 sdata
->sdata_op_ordering_lock
.Unlock();
10277 // osd_opwq_process marks the point at which an operation has been dequeued
10278 // and will begin to be handled by a worker thread.
10282 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10283 reqid
= (*_op
)->get_reqid();
10286 tracepoint(osd
, opwq_process_start
, reqid
.name
._type
,
10287 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10290 lgeneric_subdout(osd
->cct
, osd
, 30) << "dequeue status: ";
10291 Formatter
*f
= Formatter::create("json");
10292 f
->open_object_section("q");
10294 f
->close_section();
10299 ThreadPool::TPHandle
tp_handle(osd
->cct
, hb
, timeout_interval
,
10301 qi
->run(osd
, pg
, tp_handle
);
10306 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10307 reqid
= (*_op
)->get_reqid();
10310 tracepoint(osd
, opwq_process_finish
, reqid
.name
._type
,
10311 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10317 void OSD::ShardedOpWQ::_enqueue(pair
<spg_t
, PGQueueable
> item
) {
10318 uint32_t shard_index
=
10319 item
.first
.hash_to_shard(shard_list
.size());
10321 ShardData
* sdata
= shard_list
[shard_index
];
10322 assert (NULL
!= sdata
);
10323 unsigned priority
= item
.second
.get_priority();
10324 unsigned cost
= item
.second
.get_cost();
10325 sdata
->sdata_op_ordering_lock
.Lock();
10327 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10328 if (priority
>= osd
->op_prio_cutoff
)
10329 sdata
->pqueue
->enqueue_strict(
10330 item
.second
.get_owner(), priority
, item
);
10332 sdata
->pqueue
->enqueue(
10333 item
.second
.get_owner(),
10334 priority
, cost
, item
);
10335 sdata
->sdata_op_ordering_lock
.Unlock();
10337 sdata
->sdata_lock
.Lock();
10338 sdata
->sdata_cond
.SignalOne();
10339 sdata
->sdata_lock
.Unlock();
10343 void OSD::ShardedOpWQ::_enqueue_front(pair
<spg_t
, PGQueueable
> item
)
10345 uint32_t shard_index
= item
.first
.hash_to_shard(shard_list
.size());
10346 ShardData
* sdata
= shard_list
[shard_index
];
10347 assert (NULL
!= sdata
);
10348 sdata
->sdata_op_ordering_lock
.Lock();
10349 auto p
= sdata
->pg_slots
.find(item
.first
);
10350 if (p
!= sdata
->pg_slots
.end() && !p
->second
.to_process
.empty()) {
10351 // we may be racing with _process, which has dequeued a new item
10352 // from pqueue, put it on to_process, and is now busy taking the
10353 // pg lock. ensure this old requeued item is ordered before any
10354 // such newer item in to_process.
10355 p
->second
.to_process
.push_front(item
.second
);
10356 item
.second
= p
->second
.to_process
.back();
10357 p
->second
.to_process
.pop_back();
10358 dout(20) << __func__
<< " " << item
.first
10359 << " " << p
->second
.to_process
.front()
10360 << " shuffled w/ " << item
.second
<< dendl
;
10362 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10364 sdata
->_enqueue_front(item
, osd
->op_prio_cutoff
);
10365 sdata
->sdata_op_ordering_lock
.Unlock();
10366 sdata
->sdata_lock
.Lock();
10367 sdata
->sdata_cond
.SignalOne();
10368 sdata
->sdata_lock
.Unlock();
10372 namespace osd_cmds
{
10374 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
)
10376 if (!ceph_using_tcmalloc()) {
10377 os
<< "could not issue heap profiler command -- not using tcmalloc!";
10378 return -EOPNOTSUPP
;
10382 if (!cmd_getval(&cct
, cmdmap
, "heapcmd", cmd
)) {
10383 os
<< "unable to get value for command \"" << cmd
<< "\"";
10387 std::vector
<std::string
> cmd_vec
;
10388 get_str_vec(cmd
, cmd_vec
);
10390 ceph_heap_profiler_handle_command(cmd_vec
, os
);
10395 }} // namespace ceph::osd_cmds
10398 std::ostream
& operator<<(std::ostream
& out
, const OSD::io_queue
& q
) {
10400 case OSD::io_queue::prioritized
:
10401 out
<< "prioritized";
10403 case OSD::io_queue::weightedpriority
:
10404 out
<< "weightedpriority";
10406 case OSD::io_queue::mclock_opclass
:
10407 out
<< "mclock_opclass";
10409 case OSD::io_queue::mclock_client
:
10410 out
<< "mclock_client";