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"
48 #include "common/pick_address.h"
50 #include "os/ObjectStore.h"
52 #include "os/FuseStore.h"
55 #include "PrimaryLogPG.h"
58 #include "msg/Messenger.h"
59 #include "msg/Message.h"
61 #include "mon/MonClient.h"
63 #include "messages/MLog.h"
65 #include "messages/MGenericMessage.h"
66 #include "messages/MOSDPing.h"
67 #include "messages/MOSDFailure.h"
68 #include "messages/MOSDMarkMeDown.h"
69 #include "messages/MOSDFull.h"
70 #include "messages/MOSDOp.h"
71 #include "messages/MOSDOpReply.h"
72 #include "messages/MOSDBackoff.h"
73 #include "messages/MOSDBeacon.h"
74 #include "messages/MOSDRepOp.h"
75 #include "messages/MOSDRepOpReply.h"
76 #include "messages/MOSDBoot.h"
77 #include "messages/MOSDPGTemp.h"
79 #include "messages/MOSDMap.h"
80 #include "messages/MMonGetOSDMap.h"
81 #include "messages/MOSDPGNotify.h"
82 #include "messages/MOSDPGQuery.h"
83 #include "messages/MOSDPGLog.h"
84 #include "messages/MOSDPGRemove.h"
85 #include "messages/MOSDPGInfo.h"
86 #include "messages/MOSDPGCreate.h"
87 #include "messages/MOSDPGTrim.h"
88 #include "messages/MOSDPGScan.h"
89 #include "messages/MOSDPGBackfill.h"
90 #include "messages/MBackfillReserve.h"
91 #include "messages/MRecoveryReserve.h"
92 #include "messages/MOSDForceRecovery.h"
93 #include "messages/MOSDECSubOpWrite.h"
94 #include "messages/MOSDECSubOpWriteReply.h"
95 #include "messages/MOSDECSubOpRead.h"
96 #include "messages/MOSDECSubOpReadReply.h"
97 #include "messages/MOSDPGCreated.h"
98 #include "messages/MOSDPGUpdateLogMissing.h"
99 #include "messages/MOSDPGUpdateLogMissingReply.h"
101 #include "messages/MOSDAlive.h"
103 #include "messages/MOSDScrub.h"
104 #include "messages/MOSDScrubReserve.h"
105 #include "messages/MOSDRepScrub.h"
107 #include "messages/MMonCommand.h"
108 #include "messages/MCommand.h"
109 #include "messages/MCommandReply.h"
111 #include "messages/MPGStats.h"
112 #include "messages/MPGStatsAck.h"
114 #include "messages/MWatchNotify.h"
115 #include "messages/MOSDPGPush.h"
116 #include "messages/MOSDPGPushReply.h"
117 #include "messages/MOSDPGPull.h"
119 #include "common/perf_counters.h"
120 #include "common/Timer.h"
121 #include "common/LogClient.h"
122 #include "common/AsyncReserver.h"
123 #include "common/HeartbeatMap.h"
124 #include "common/admin_socket.h"
125 #include "common/ceph_context.h"
127 #include "global/signal_handler.h"
128 #include "global/pidfile.h"
130 #include "include/color.h"
131 #include "perfglue/cpu_profiler.h"
132 #include "perfglue/heap_profiler.h"
134 #include "osd/OpRequest.h"
136 #include "auth/AuthAuthorizeHandler.h"
137 #include "auth/RotatingKeyRing.h"
138 #include "common/errno.h"
140 #include "objclass/objclass.h"
142 #include "common/cmdparse.h"
143 #include "include/str_list.h"
144 #include "include/util.h"
146 #include "include/assert.h"
147 #include "common/config.h"
148 #include "common/EventTrace.h"
151 #define TRACEPOINT_DEFINE
152 #define TRACEPOINT_PROBE_DYNAMIC_LINKAGE
153 #include "tracing/osd.h"
154 #undef TRACEPOINT_PROBE_DYNAMIC_LINKAGE
155 #undef TRACEPOINT_DEFINE
157 #define tracepoint(...)
160 #define dout_context cct
161 #define dout_subsys ceph_subsys_osd
163 #define dout_prefix _prefix(_dout, whoami, get_osdmap_epoch())
166 const double OSD::OSD_TICK_INTERVAL
= 1.0;
168 static ostream
& _prefix(std::ostream
* _dout
, int whoami
, epoch_t epoch
) {
169 return *_dout
<< "osd." << whoami
<< " " << epoch
<< " ";
172 //Initial features in new superblock.
173 //Features here are also automatically upgraded
174 CompatSet
OSD::get_osd_initial_compat_set() {
175 CompatSet::FeatureSet ceph_osd_feature_compat
;
176 CompatSet::FeatureSet ceph_osd_feature_ro_compat
;
177 CompatSet::FeatureSet ceph_osd_feature_incompat
;
178 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE
);
179 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_PGINFO
);
180 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_OLOC
);
181 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_LEC
);
182 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_CATEGORIES
);
183 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_HOBJECTPOOL
);
184 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BIGINFO
);
185 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_LEVELDBINFO
);
186 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_LEVELDBLOG
);
187 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SNAPMAPPER
);
188 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_HINTS
);
189 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_PGMETA
);
190 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_MISSING
);
191 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_FASTINFO
);
192 ceph_osd_feature_incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_RECOVERY_DELETES
);
193 return CompatSet(ceph_osd_feature_compat
, ceph_osd_feature_ro_compat
,
194 ceph_osd_feature_incompat
);
197 //Features are added here that this OSD supports.
198 CompatSet
OSD::get_osd_compat_set() {
199 CompatSet compat
= get_osd_initial_compat_set();
200 //Any features here can be set in code, but not in initial superblock
201 compat
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
);
205 OSDService::OSDService(OSD
*osd
) :
208 meta_osr(new ObjectStore::Sequencer("meta")),
209 whoami(osd
->whoami
), store(osd
->store
),
210 log_client(osd
->log_client
), clog(osd
->clog
),
211 pg_recovery_stats(osd
->pg_recovery_stats
),
212 cluster_messenger(osd
->cluster_messenger
),
213 client_messenger(osd
->client_messenger
),
215 recoverystate_perf(osd
->recoverystate_perf
),
217 peering_wq(osd
->peering_wq
),
218 recovery_gen_wq("recovery_gen_wq", cct
->_conf
->osd_recovery_thread_timeout
,
220 class_handler(osd
->class_handler
),
221 pg_epoch_lock("OSDService::pg_epoch_lock"),
222 publish_lock("OSDService::publish_lock"),
223 pre_publish_lock("OSDService::pre_publish_lock"),
225 peer_map_epoch_lock("OSDService::peer_map_epoch_lock"),
226 sched_scrub_lock("OSDService::sched_scrub_lock"), scrubs_pending(0),
228 agent_lock("OSDService::agent_lock"),
229 agent_valid_iterator(false),
231 flush_mode_high_count(0),
234 agent_stop_flag(false),
235 agent_timer_lock("OSDService::agent_timer_lock"),
236 agent_timer(osd
->client_messenger
->cct
, agent_timer_lock
),
237 last_recalibrate(ceph_clock_now()),
238 promote_max_objects(0),
239 promote_max_bytes(0),
240 objecter(new Objecter(osd
->client_messenger
->cct
, osd
->objecter_messenger
, osd
->monc
, NULL
, 0, 0)),
241 objecter_finisher(osd
->client_messenger
->cct
),
242 watch_lock("OSDService::watch_lock"),
243 watch_timer(osd
->client_messenger
->cct
, watch_lock
),
245 recovery_request_lock("OSDService::recovery_request_lock"),
246 recovery_request_timer(cct
, recovery_request_lock
, false),
247 recovery_sleep_lock("OSDService::recovery_sleep_lock"),
248 recovery_sleep_timer(cct
, recovery_sleep_lock
, false),
249 reserver_finisher(cct
),
250 local_reserver(cct
, &reserver_finisher
, cct
->_conf
->osd_max_backfills
,
251 cct
->_conf
->osd_min_recovery_priority
),
252 remote_reserver(cct
, &reserver_finisher
, cct
->_conf
->osd_max_backfills
,
253 cct
->_conf
->osd_min_recovery_priority
),
254 pg_temp_lock("OSDService::pg_temp_lock"),
255 snap_sleep_lock("OSDService::snap_sleep_lock"),
257 osd
->client_messenger
->cct
, snap_sleep_lock
, false /* relax locking */),
258 scrub_sleep_lock("OSDService::scrub_sleep_lock"),
260 osd
->client_messenger
->cct
, scrub_sleep_lock
, false /* relax locking */),
261 snap_reserver(cct
, &reserver_finisher
,
262 cct
->_conf
->osd_max_trimming_pgs
),
263 recovery_lock("OSDService::recovery_lock"),
264 recovery_ops_active(0),
265 recovery_ops_reserved(0),
266 recovery_paused(false),
267 map_cache_lock("OSDService::map_cache_lock"),
268 map_cache(cct
, cct
->_conf
->osd_map_cache_size
),
269 map_bl_cache(cct
->_conf
->osd_map_cache_size
),
270 map_bl_inc_cache(cct
->_conf
->osd_map_cache_size
),
271 in_progress_split_lock("OSDService::in_progress_split_lock"),
272 stat_lock("OSDService::stat_lock"),
273 full_status_lock("OSDService::full_status_lock"),
276 epoch_lock("OSDService::epoch_lock"),
277 boot_epoch(0), up_epoch(0), bind_epoch(0),
278 is_stopping_lock("OSDService::is_stopping_lock")
280 , pgid_lock("OSDService::pgid_lock")
286 OSDService::~OSDService()
294 void OSDService::add_pgid(spg_t pgid
, PG
*pg
){
295 Mutex::Locker
l(pgid_lock
);
296 if (!pgid_tracker
.count(pgid
)) {
299 pgid_tracker
[pgid
]++;
301 void OSDService::remove_pgid(spg_t pgid
, PG
*pg
)
303 Mutex::Locker
l(pgid_lock
);
304 assert(pgid_tracker
.count(pgid
));
305 assert(pgid_tracker
[pgid
] > 0);
306 pgid_tracker
[pgid
]--;
307 if (pgid_tracker
[pgid
] == 0) {
308 pgid_tracker
.erase(pgid
);
309 live_pgs
.erase(pgid
);
312 void OSDService::dump_live_pgids()
314 Mutex::Locker
l(pgid_lock
);
315 derr
<< "live pgids:" << dendl
;
316 for (map
<spg_t
, int>::const_iterator i
= pgid_tracker
.cbegin();
317 i
!= pgid_tracker
.cend();
319 derr
<< "\t" << *i
<< dendl
;
320 live_pgs
[i
->first
]->dump_live_ids();
326 void OSDService::_start_split(spg_t parent
, const set
<spg_t
> &children
)
328 for (set
<spg_t
>::const_iterator i
= children
.begin();
331 dout(10) << __func__
<< ": Starting split on pg " << *i
332 << ", parent=" << parent
<< dendl
;
333 assert(!pending_splits
.count(*i
));
334 assert(!in_progress_splits
.count(*i
));
335 pending_splits
.insert(make_pair(*i
, parent
));
337 assert(!rev_pending_splits
[parent
].count(*i
));
338 rev_pending_splits
[parent
].insert(*i
);
342 void OSDService::mark_split_in_progress(spg_t parent
, const set
<spg_t
> &children
)
344 Mutex::Locker
l(in_progress_split_lock
);
345 map
<spg_t
, set
<spg_t
> >::iterator piter
= rev_pending_splits
.find(parent
);
346 assert(piter
!= rev_pending_splits
.end());
347 for (set
<spg_t
>::const_iterator i
= children
.begin();
350 assert(piter
->second
.count(*i
));
351 assert(pending_splits
.count(*i
));
352 assert(!in_progress_splits
.count(*i
));
353 assert(pending_splits
[*i
] == parent
);
355 pending_splits
.erase(*i
);
356 piter
->second
.erase(*i
);
357 in_progress_splits
.insert(*i
);
359 if (piter
->second
.empty())
360 rev_pending_splits
.erase(piter
);
363 void OSDService::cancel_pending_splits_for_parent(spg_t parent
)
365 Mutex::Locker
l(in_progress_split_lock
);
366 _cancel_pending_splits_for_parent(parent
);
369 void OSDService::_cancel_pending_splits_for_parent(spg_t parent
)
371 map
<spg_t
, set
<spg_t
> >::iterator piter
= rev_pending_splits
.find(parent
);
372 if (piter
== rev_pending_splits
.end())
375 for (set
<spg_t
>::iterator i
= piter
->second
.begin();
376 i
!= piter
->second
.end();
378 assert(pending_splits
.count(*i
));
379 assert(!in_progress_splits
.count(*i
));
380 pending_splits
.erase(*i
);
381 dout(10) << __func__
<< ": Completing split on pg " << *i
382 << " for parent: " << parent
<< dendl
;
383 _cancel_pending_splits_for_parent(*i
);
385 rev_pending_splits
.erase(piter
);
388 void OSDService::_maybe_split_pgid(OSDMapRef old_map
,
392 assert(old_map
->have_pg_pool(pgid
.pool()));
393 int old_pgnum
= old_map
->get_pg_num(pgid
.pool());
394 if (pgid
.ps() < static_cast<unsigned>(old_pgnum
)) {
396 if (pgid
.is_split(old_pgnum
,
397 new_map
->get_pg_num(pgid
.pool()), &children
)) {
398 _start_split(pgid
, children
); }
400 assert(pgid
.ps() < static_cast<unsigned>(new_map
->get_pg_num(pgid
.pool())));
404 void OSDService::init_splits_between(spg_t pgid
,
408 // First, check whether we can avoid this potentially expensive check
409 if (tomap
->have_pg_pool(pgid
.pool()) &&
411 frommap
->get_pg_num(pgid
.pool()),
412 tomap
->get_pg_num(pgid
.pool()),
414 // Ok, a split happened, so we need to walk the osdmaps
415 set
<spg_t
> new_pgs
; // pgs to scan on each map
416 new_pgs
.insert(pgid
);
417 OSDMapRef
curmap(get_map(frommap
->get_epoch()));
418 for (epoch_t e
= frommap
->get_epoch() + 1;
419 e
<= tomap
->get_epoch();
421 OSDMapRef
nextmap(try_get_map(e
));
424 set
<spg_t
> even_newer_pgs
; // pgs added in this loop
425 for (set
<spg_t
>::iterator i
= new_pgs
.begin(); i
!= new_pgs
.end(); ++i
) {
426 set
<spg_t
> split_pgs
;
427 if (i
->is_split(curmap
->get_pg_num(i
->pool()),
428 nextmap
->get_pg_num(i
->pool()),
430 start_split(*i
, split_pgs
);
431 even_newer_pgs
.insert(split_pgs
.begin(), split_pgs
.end());
434 new_pgs
.insert(even_newer_pgs
.begin(), even_newer_pgs
.end());
437 assert(curmap
== tomap
); // we must have had both frommap and tomap
441 void OSDService::expand_pg_num(OSDMapRef old_map
,
444 Mutex::Locker
l(in_progress_split_lock
);
445 for (set
<spg_t
>::iterator i
= in_progress_splits
.begin();
446 i
!= in_progress_splits
.end();
448 if (!new_map
->have_pg_pool(i
->pool())) {
449 in_progress_splits
.erase(i
++);
451 _maybe_split_pgid(old_map
, new_map
, *i
);
455 for (map
<spg_t
, spg_t
>::iterator i
= pending_splits
.begin();
456 i
!= pending_splits
.end();
458 if (!new_map
->have_pg_pool(i
->first
.pool())) {
459 rev_pending_splits
.erase(i
->second
);
460 pending_splits
.erase(i
++);
462 _maybe_split_pgid(old_map
, new_map
, i
->first
);
468 bool OSDService::splitting(spg_t pgid
)
470 Mutex::Locker
l(in_progress_split_lock
);
471 return in_progress_splits
.count(pgid
) ||
472 pending_splits
.count(pgid
);
475 void OSDService::complete_split(const set
<spg_t
> &pgs
)
477 Mutex::Locker
l(in_progress_split_lock
);
478 for (set
<spg_t
>::const_iterator i
= pgs
.begin();
481 dout(10) << __func__
<< ": Completing split on pg " << *i
<< dendl
;
482 assert(!pending_splits
.count(*i
));
483 assert(in_progress_splits
.count(*i
));
484 in_progress_splits
.erase(*i
);
488 void OSDService::need_heartbeat_peer_update()
490 osd
->need_heartbeat_peer_update();
493 void OSDService::pg_stat_queue_enqueue(PG
*pg
)
495 osd
->pg_stat_queue_enqueue(pg
);
498 void OSDService::pg_stat_queue_dequeue(PG
*pg
)
500 osd
->pg_stat_queue_dequeue(pg
);
503 void OSDService::start_shutdown()
506 Mutex::Locker
l(agent_timer_lock
);
507 agent_timer
.shutdown();
511 Mutex::Locker
l(recovery_sleep_lock
);
512 recovery_sleep_timer
.shutdown();
516 void OSDService::shutdown_reserver()
518 reserver_finisher
.wait_for_empty();
519 reserver_finisher
.stop();
522 void OSDService::shutdown()
525 Mutex::Locker
l(watch_lock
);
526 watch_timer
.shutdown();
529 objecter
->shutdown();
530 objecter_finisher
.wait_for_empty();
531 objecter_finisher
.stop();
534 Mutex::Locker
l(recovery_request_lock
);
535 recovery_request_timer
.shutdown();
539 Mutex::Locker
l(snap_sleep_lock
);
540 snap_sleep_timer
.shutdown();
544 Mutex::Locker
l(scrub_sleep_lock
);
545 scrub_sleep_timer
.shutdown();
548 osdmap
= OSDMapRef();
549 next_osdmap
= OSDMapRef();
552 void OSDService::init()
554 reserver_finisher
.start();
555 objecter_finisher
.start();
556 objecter
->set_client_incarnation(0);
558 // deprioritize objecter in daemonperf output
559 objecter
->get_logger()->set_prio_adjust(-3);
563 snap_sleep_timer
.init();
564 scrub_sleep_timer
.init();
566 agent_thread
.create("osd_srv_agent");
568 if (cct
->_conf
->osd_recovery_delay_start
)
569 defer_recovery(cct
->_conf
->osd_recovery_delay_start
);
572 void OSDService::final_init()
574 objecter
->start(osdmap
.get());
577 void OSDService::activate_map()
579 // wake/unwake the tiering agent
582 !osdmap
->test_flag(CEPH_OSDMAP_NOTIERAGENT
) &&
588 void OSDService::request_osdmap_update(epoch_t e
)
590 osd
->osdmap_subscribe(e
, false);
593 class AgentTimeoutCB
: public Context
{
596 explicit AgentTimeoutCB(PGRef _pg
) : pg(_pg
) {}
597 void finish(int) override
{
598 pg
->agent_choose_mode_restart();
602 void OSDService::agent_entry()
604 dout(10) << __func__
<< " start" << dendl
;
607 while (!agent_stop_flag
) {
608 if (agent_queue
.empty()) {
609 dout(20) << __func__
<< " empty queue" << dendl
;
610 agent_cond
.Wait(agent_lock
);
613 uint64_t level
= agent_queue
.rbegin()->first
;
614 set
<PGRef
>& top
= agent_queue
.rbegin()->second
;
616 << " tiers " << agent_queue
.size()
617 << ", top is " << level
618 << " with pgs " << top
.size()
619 << ", ops " << agent_ops
<< "/"
620 << cct
->_conf
->osd_agent_max_ops
621 << (agent_active
? " active" : " NOT ACTIVE")
623 dout(20) << __func__
<< " oids " << agent_oids
<< dendl
;
624 int max
= cct
->_conf
->osd_agent_max_ops
- agent_ops
;
625 int agent_flush_quota
= max
;
626 if (!flush_mode_high_count
)
627 agent_flush_quota
= cct
->_conf
->osd_agent_max_low_ops
- agent_ops
;
628 if (agent_flush_quota
<= 0 || top
.empty() || !agent_active
) {
629 agent_cond
.Wait(agent_lock
);
633 if (!agent_valid_iterator
|| agent_queue_pos
== top
.end()) {
634 agent_queue_pos
= top
.begin();
635 agent_valid_iterator
= true;
637 PGRef pg
= *agent_queue_pos
;
638 dout(10) << "high_count " << flush_mode_high_count
639 << " agent_ops " << agent_ops
640 << " flush_quota " << agent_flush_quota
<< dendl
;
642 if (!pg
->agent_work(max
, agent_flush_quota
)) {
643 dout(10) << __func__
<< " " << pg
->get_pgid()
644 << " no agent_work, delay for " << cct
->_conf
->osd_agent_delay_time
645 << " seconds" << dendl
;
647 osd
->logger
->inc(l_osd_tier_delay
);
648 // Queue a timer to call agent_choose_mode for this pg in 5 seconds
649 agent_timer_lock
.Lock();
650 Context
*cb
= new AgentTimeoutCB(pg
);
651 agent_timer
.add_event_after(cct
->_conf
->osd_agent_delay_time
, cb
);
652 agent_timer_lock
.Unlock();
657 dout(10) << __func__
<< " finish" << dendl
;
660 void OSDService::agent_stop()
663 Mutex::Locker
l(agent_lock
);
665 // By this time all ops should be cancelled
666 assert(agent_ops
== 0);
667 // By this time all PGs are shutdown and dequeued
668 if (!agent_queue
.empty()) {
669 set
<PGRef
>& top
= agent_queue
.rbegin()->second
;
670 derr
<< "agent queue not empty, for example " << (*top
.begin())->info
.pgid
<< dendl
;
671 assert(0 == "agent queue not empty");
674 agent_stop_flag
= true;
680 // -------------------------------------
682 void OSDService::promote_throttle_recalibrate()
684 utime_t now
= ceph_clock_now();
685 double dur
= now
- last_recalibrate
;
686 last_recalibrate
= now
;
687 unsigned prob
= promote_probability_millis
;
689 uint64_t target_obj_sec
= cct
->_conf
->osd_tier_promote_max_objects_sec
;
690 uint64_t target_bytes_sec
= cct
->_conf
->osd_tier_promote_max_bytes_sec
;
692 unsigned min_prob
= 1;
694 uint64_t attempts
, obj
, bytes
;
695 promote_counter
.sample_and_attenuate(&attempts
, &obj
, &bytes
);
696 dout(10) << __func__
<< " " << attempts
<< " attempts, promoted "
697 << obj
<< " objects and " << pretty_si_t(bytes
) << " bytes; target "
698 << target_obj_sec
<< " obj/sec or "
699 << pretty_si_t(target_bytes_sec
) << " bytes/sec"
702 // calculate what the probability *should* be, given the targets
704 if (attempts
&& dur
> 0) {
705 uint64_t avg_size
= 1;
707 avg_size
= MAX(bytes
/ obj
, 1);
708 unsigned po
= (double)target_obj_sec
* dur
* 1000.0 / (double)attempts
;
709 unsigned pb
= (double)target_bytes_sec
/ (double)avg_size
* dur
* 1000.0
711 dout(20) << __func__
<< " po " << po
<< " pb " << pb
<< " avg_size "
712 << avg_size
<< dendl
;
713 if (target_obj_sec
&& target_bytes_sec
)
714 new_prob
= MIN(po
, pb
);
715 else if (target_obj_sec
)
717 else if (target_bytes_sec
)
724 dout(20) << __func__
<< " new_prob " << new_prob
<< dendl
;
726 // correct for persistent skew between target rate and actual rate, adjust
729 if (attempts
&& obj
) {
730 actual
= obj
* 1000 / attempts
;
731 ratio
= (double)actual
/ (double)prob
;
732 new_prob
= (double)new_prob
/ ratio
;
734 new_prob
= MAX(new_prob
, min_prob
);
735 new_prob
= MIN(new_prob
, 1000);
738 prob
= (prob
+ new_prob
) / 2;
739 prob
= MAX(prob
, min_prob
);
740 prob
= MIN(prob
, 1000);
741 dout(10) << __func__
<< " actual " << actual
742 << ", actual/prob ratio " << ratio
743 << ", adjusted new_prob " << new_prob
744 << ", prob " << promote_probability_millis
<< " -> " << prob
746 promote_probability_millis
= prob
;
748 // set hard limits for this interval to mitigate stampedes
749 promote_max_objects
= target_obj_sec
* OSD::OSD_TICK_INTERVAL
* 2;
750 promote_max_bytes
= target_bytes_sec
* OSD::OSD_TICK_INTERVAL
* 2;
753 // -------------------------------------
755 float OSDService::get_failsafe_full_ratio()
757 float full_ratio
= cct
->_conf
->osd_failsafe_full_ratio
;
758 if (full_ratio
> 1.0) full_ratio
/= 100.0;
762 void OSDService::check_full_status(float ratio
)
764 Mutex::Locker
l(full_status_lock
);
768 // The OSDMap ratios take precendence. So if the failsafe is .95 and
769 // the admin sets the cluster full to .96, the failsafe moves up to .96
770 // too. (Not that having failsafe == full is ideal, but it's better than
771 // dropping writes before the clusters appears full.)
772 OSDMapRef osdmap
= get_osdmap();
773 if (!osdmap
|| osdmap
->get_epoch() == 0) {
777 float nearfull_ratio
= osdmap
->get_nearfull_ratio();
778 float backfillfull_ratio
= std::max(osdmap
->get_backfillfull_ratio(), nearfull_ratio
);
779 float full_ratio
= std::max(osdmap
->get_full_ratio(), backfillfull_ratio
);
780 float failsafe_ratio
= std::max(get_failsafe_full_ratio(), full_ratio
);
782 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
783 // use the failsafe for nearfull and full; the mon isn't using the
784 // flags anyway because we're mid-upgrade.
785 full_ratio
= failsafe_ratio
;
786 backfillfull_ratio
= failsafe_ratio
;
787 nearfull_ratio
= failsafe_ratio
;
788 } else if (full_ratio
<= 0 ||
789 backfillfull_ratio
<= 0 ||
790 nearfull_ratio
<= 0) {
791 derr
<< __func__
<< " full_ratio, backfillfull_ratio or nearfull_ratio is <= 0" << dendl
;
792 // use failsafe flag. ick. the monitor did something wrong or the user
793 // did something stupid.
794 full_ratio
= failsafe_ratio
;
795 backfillfull_ratio
= failsafe_ratio
;
796 nearfull_ratio
= failsafe_ratio
;
801 if (injectfull_state
> NONE
&& injectfull
) {
802 new_state
= injectfull_state
;
803 inject
= "(Injected)";
804 } else if (ratio
> failsafe_ratio
) {
805 new_state
= FAILSAFE
;
806 } else if (ratio
> full_ratio
) {
808 } else if (ratio
> backfillfull_ratio
) {
809 new_state
= BACKFILLFULL
;
810 } else if (ratio
> nearfull_ratio
) {
811 new_state
= NEARFULL
;
815 dout(20) << __func__
<< " cur ratio " << ratio
816 << ". nearfull_ratio " << nearfull_ratio
817 << ". backfillfull_ratio " << backfillfull_ratio
818 << ", full_ratio " << full_ratio
819 << ", failsafe_ratio " << failsafe_ratio
820 << ", new state " << get_full_state_name(new_state
)
825 if (cur_state
!= new_state
) {
826 dout(10) << __func__
<< " " << get_full_state_name(cur_state
)
827 << " -> " << get_full_state_name(new_state
) << dendl
;
828 if (new_state
== FAILSAFE
) {
829 clog
->error() << "full status failsafe engaged, dropping updates, now "
830 << (int)roundf(ratio
* 100) << "% full";
831 } else if (cur_state
== FAILSAFE
) {
832 clog
->error() << "full status failsafe disengaged, no longer dropping "
833 << "updates, now " << (int)roundf(ratio
* 100) << "% full";
835 cur_state
= new_state
;
839 bool OSDService::need_fullness_update()
841 OSDMapRef osdmap
= get_osdmap();
843 if (osdmap
->exists(whoami
)) {
844 if (osdmap
->get_state(whoami
) & CEPH_OSD_FULL
) {
846 } else if (osdmap
->get_state(whoami
) & CEPH_OSD_BACKFILLFULL
) {
848 } else if (osdmap
->get_state(whoami
) & CEPH_OSD_NEARFULL
) {
855 else if (is_backfillfull())
857 else if (is_nearfull())
862 bool OSDService::_check_full(s_names type
, ostream
&ss
) const
864 Mutex::Locker
l(full_status_lock
);
866 if (injectfull
&& injectfull_state
>= type
) {
867 // injectfull is either a count of the number of times to return failsafe full
868 // or if -1 then always return full
871 ss
<< "Injected " << get_full_state_name(type
) << " OSD ("
872 << (injectfull
< 0 ? "set" : std::to_string(injectfull
)) << ")";
876 ss
<< "current usage is " << cur_ratio
;
877 return cur_state
>= type
;
880 bool OSDService::check_failsafe_full(ostream
&ss
) const
882 return _check_full(FAILSAFE
, ss
);
885 bool OSDService::check_full(ostream
&ss
) const
887 return _check_full(FULL
, ss
);
890 bool OSDService::check_backfill_full(ostream
&ss
) const
892 return _check_full(BACKFILLFULL
, ss
);
895 bool OSDService::check_nearfull(ostream
&ss
) const
897 return _check_full(NEARFULL
, ss
);
900 bool OSDService::is_failsafe_full() const
902 Mutex::Locker
l(full_status_lock
);
903 return cur_state
== FAILSAFE
;
906 bool OSDService::is_full() const
908 Mutex::Locker
l(full_status_lock
);
909 return cur_state
>= FULL
;
912 bool OSDService::is_backfillfull() const
914 Mutex::Locker
l(full_status_lock
);
915 return cur_state
>= BACKFILLFULL
;
918 bool OSDService::is_nearfull() const
920 Mutex::Locker
l(full_status_lock
);
921 return cur_state
>= NEARFULL
;
924 void OSDService::set_injectfull(s_names type
, int64_t count
)
926 Mutex::Locker
l(full_status_lock
);
927 injectfull_state
= type
;
931 osd_stat_t
OSDService::set_osd_stat(const struct store_statfs_t
&stbuf
,
932 vector
<int>& hb_peers
,
935 uint64_t bytes
= stbuf
.total
;
936 uint64_t used
= bytes
- stbuf
.available
;
937 uint64_t avail
= stbuf
.available
;
939 osd
->logger
->set(l_osd_stat_bytes
, bytes
);
940 osd
->logger
->set(l_osd_stat_bytes_used
, used
);
941 osd
->logger
->set(l_osd_stat_bytes_avail
, avail
);
944 Mutex::Locker
l(stat_lock
);
945 osd_stat
.hb_peers
.swap(hb_peers
);
946 osd
->op_tracker
.get_age_ms_histogram(&osd_stat
.op_queue_age_hist
);
947 osd_stat
.kb
= bytes
>> 10;
948 osd_stat
.kb_used
= used
>> 10;
949 osd_stat
.kb_avail
= avail
>> 10;
950 osd_stat
.num_pgs
= num_pgs
;
955 void OSDService::update_osd_stat(vector
<int>& hb_peers
)
957 // load osd stats first
958 struct store_statfs_t stbuf
;
959 int r
= osd
->store
->statfs(&stbuf
);
961 derr
<< "statfs() failed: " << cpp_strerror(r
) << dendl
;
965 auto new_stat
= set_osd_stat(stbuf
, hb_peers
, osd
->get_num_pgs());
966 dout(20) << "update_osd_stat " << new_stat
<< dendl
;
968 float ratio
= ((float)new_stat
.kb_used
) / ((float)new_stat
.kb
);
969 check_full_status(ratio
);
972 bool OSDService::check_osdmap_full(const set
<pg_shard_t
> &missing_on
)
974 OSDMapRef osdmap
= get_osdmap();
975 for (auto shard
: missing_on
) {
976 if (osdmap
->get_state(shard
.osd
) & CEPH_OSD_FULL
)
982 void OSDService::send_message_osd_cluster(int peer
, Message
*m
, epoch_t from_epoch
)
984 OSDMapRef next_map
= get_nextmap_reserved();
985 // service map is always newer/newest
986 assert(from_epoch
<= next_map
->get_epoch());
988 if (next_map
->is_down(peer
) ||
989 next_map
->get_info(peer
).up_from
> from_epoch
) {
991 release_map(next_map
);
994 const entity_inst_t
& peer_inst
= next_map
->get_cluster_inst(peer
);
995 ConnectionRef peer_con
= osd
->cluster_messenger
->get_connection(peer_inst
);
996 share_map_peer(peer
, peer_con
.get(), next_map
);
997 peer_con
->send_message(m
);
998 release_map(next_map
);
1001 ConnectionRef
OSDService::get_con_osd_cluster(int peer
, epoch_t from_epoch
)
1003 OSDMapRef next_map
= get_nextmap_reserved();
1004 // service map is always newer/newest
1005 assert(from_epoch
<= next_map
->get_epoch());
1007 if (next_map
->is_down(peer
) ||
1008 next_map
->get_info(peer
).up_from
> from_epoch
) {
1009 release_map(next_map
);
1012 ConnectionRef con
= osd
->cluster_messenger
->get_connection(next_map
->get_cluster_inst(peer
));
1013 release_map(next_map
);
1017 pair
<ConnectionRef
,ConnectionRef
> OSDService::get_con_osd_hb(int peer
, epoch_t from_epoch
)
1019 OSDMapRef next_map
= get_nextmap_reserved();
1020 // service map is always newer/newest
1021 assert(from_epoch
<= next_map
->get_epoch());
1023 pair
<ConnectionRef
,ConnectionRef
> ret
;
1024 if (next_map
->is_down(peer
) ||
1025 next_map
->get_info(peer
).up_from
> from_epoch
) {
1026 release_map(next_map
);
1029 ret
.first
= osd
->hb_back_client_messenger
->get_connection(next_map
->get_hb_back_inst(peer
));
1030 if (next_map
->get_hb_front_addr(peer
) != entity_addr_t())
1031 ret
.second
= osd
->hb_front_client_messenger
->get_connection(next_map
->get_hb_front_inst(peer
));
1032 release_map(next_map
);
1037 void OSDService::queue_want_pg_temp(pg_t pgid
,
1038 const vector
<int>& want
,
1041 Mutex::Locker
l(pg_temp_lock
);
1042 auto p
= pg_temp_pending
.find(pgid
);
1043 if (p
== pg_temp_pending
.end() ||
1044 p
->second
.acting
!= want
||
1046 pg_temp_wanted
[pgid
] = pg_temp_t
{want
, forced
};
1050 void OSDService::remove_want_pg_temp(pg_t pgid
)
1052 Mutex::Locker
l(pg_temp_lock
);
1053 pg_temp_wanted
.erase(pgid
);
1054 pg_temp_pending
.erase(pgid
);
1057 void OSDService::_sent_pg_temp()
1059 pg_temp_pending
.insert(make_move_iterator(begin(pg_temp_wanted
)),
1060 make_move_iterator(end(pg_temp_wanted
)));
1061 pg_temp_wanted
.clear();
1064 void OSDService::requeue_pg_temp()
1066 Mutex::Locker
l(pg_temp_lock
);
1067 // wanted overrides pending. note that remove_want_pg_temp
1068 // clears the item out of both.
1069 unsigned old_wanted
= pg_temp_wanted
.size();
1070 unsigned old_pending
= pg_temp_pending
.size();
1072 pg_temp_wanted
.swap(pg_temp_pending
);
1073 dout(10) << __func__
<< " " << old_wanted
<< " + " << old_pending
<< " -> "
1074 << pg_temp_wanted
.size() << dendl
;
1077 std::ostream
& operator<<(std::ostream
& out
,
1078 const OSDService::pg_temp_t
& pg_temp
)
1080 out
<< pg_temp
.acting
;
1081 if (pg_temp
.forced
) {
1087 void OSDService::send_pg_temp()
1089 Mutex::Locker
l(pg_temp_lock
);
1090 if (pg_temp_wanted
.empty())
1092 dout(10) << "send_pg_temp " << pg_temp_wanted
<< dendl
;
1093 MOSDPGTemp
*ms
[2] = {nullptr, nullptr};
1094 for (auto& pg_temp
: pg_temp_wanted
) {
1095 auto& m
= ms
[pg_temp
.second
.forced
];
1097 m
= new MOSDPGTemp(osdmap
->get_epoch());
1098 m
->forced
= pg_temp
.second
.forced
;
1100 m
->pg_temp
.emplace(pg_temp
.first
,
1101 pg_temp
.second
.acting
);
1105 monc
->send_mon_message(m
);
1111 void OSDService::send_pg_created(pg_t pgid
)
1113 dout(20) << __func__
<< dendl
;
1114 if (osdmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
1115 monc
->send_mon_message(new MOSDPGCreated(pgid
));
1119 // --------------------------------------
1122 epoch_t
OSDService::get_peer_epoch(int peer
)
1124 Mutex::Locker
l(peer_map_epoch_lock
);
1125 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1126 if (p
== peer_map_epoch
.end())
1131 epoch_t
OSDService::note_peer_epoch(int peer
, epoch_t e
)
1133 Mutex::Locker
l(peer_map_epoch_lock
);
1134 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1135 if (p
!= peer_map_epoch
.end()) {
1136 if (p
->second
< e
) {
1137 dout(10) << "note_peer_epoch osd." << peer
<< " has " << e
<< dendl
;
1140 dout(30) << "note_peer_epoch osd." << peer
<< " has " << p
->second
<< " >= " << e
<< dendl
;
1144 dout(10) << "note_peer_epoch osd." << peer
<< " now has " << e
<< dendl
;
1145 peer_map_epoch
[peer
] = e
;
1150 void OSDService::forget_peer_epoch(int peer
, epoch_t as_of
)
1152 Mutex::Locker
l(peer_map_epoch_lock
);
1153 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1154 if (p
!= peer_map_epoch
.end()) {
1155 if (p
->second
<= as_of
) {
1156 dout(10) << "forget_peer_epoch osd." << peer
<< " as_of " << as_of
1157 << " had " << p
->second
<< dendl
;
1158 peer_map_epoch
.erase(p
);
1160 dout(10) << "forget_peer_epoch osd." << peer
<< " as_of " << as_of
1161 << " has " << p
->second
<< " - not forgetting" << dendl
;
1166 bool OSDService::should_share_map(entity_name_t name
, Connection
*con
,
1167 epoch_t epoch
, const OSDMapRef
& osdmap
,
1168 const epoch_t
*sent_epoch_p
)
1170 dout(20) << "should_share_map "
1171 << name
<< " " << con
->get_peer_addr()
1172 << " " << epoch
<< dendl
;
1174 // does client have old map?
1175 if (name
.is_client()) {
1176 bool message_sendmap
= epoch
< osdmap
->get_epoch();
1177 if (message_sendmap
&& sent_epoch_p
) {
1178 dout(20) << "client session last_sent_epoch: "
1180 << " versus osdmap epoch " << osdmap
->get_epoch() << dendl
;
1181 if (*sent_epoch_p
< osdmap
->get_epoch()) {
1183 } // else we don't need to send it out again
1187 if (con
->get_messenger() == osd
->cluster_messenger
&&
1188 con
!= osd
->cluster_messenger
->get_loopback_connection() &&
1189 osdmap
->is_up(name
.num()) &&
1190 (osdmap
->get_cluster_addr(name
.num()) == con
->get_peer_addr() ||
1191 osdmap
->get_hb_back_addr(name
.num()) == con
->get_peer_addr())) {
1193 epoch_t has
= MAX(get_peer_epoch(name
.num()), epoch
);
1196 if (has
< osdmap
->get_epoch()) {
1197 dout(10) << name
<< " " << con
->get_peer_addr()
1198 << " has old map " << epoch
<< " < "
1199 << osdmap
->get_epoch() << dendl
;
1207 void OSDService::share_map(
1212 epoch_t
*sent_epoch_p
)
1214 dout(20) << "share_map "
1215 << name
<< " " << con
->get_peer_addr()
1216 << " " << epoch
<< dendl
;
1218 if (!osd
->is_active()) {
1219 /*It is safe not to proceed as OSD is not in healthy state*/
1223 bool want_shared
= should_share_map(name
, con
, epoch
,
1224 osdmap
, sent_epoch_p
);
1227 if (name
.is_client()) {
1228 dout(10) << name
<< " has old map " << epoch
1229 << " < " << osdmap
->get_epoch() << dendl
;
1230 // we know the Session is valid or we wouldn't be sending
1232 *sent_epoch_p
= osdmap
->get_epoch();
1234 send_incremental_map(epoch
, con
, osdmap
);
1235 } else if (con
->get_messenger() == osd
->cluster_messenger
&&
1236 osdmap
->is_up(name
.num()) &&
1237 (osdmap
->get_cluster_addr(name
.num()) == con
->get_peer_addr() ||
1238 osdmap
->get_hb_back_addr(name
.num()) == con
->get_peer_addr())) {
1239 dout(10) << name
<< " " << con
->get_peer_addr()
1240 << " has old map " << epoch
<< " < "
1241 << osdmap
->get_epoch() << dendl
;
1242 note_peer_epoch(name
.num(), osdmap
->get_epoch());
1243 send_incremental_map(epoch
, con
, osdmap
);
1248 void OSDService::share_map_peer(int peer
, Connection
*con
, OSDMapRef map
)
1254 epoch_t pe
= get_peer_epoch(peer
);
1256 if (pe
< map
->get_epoch()) {
1257 send_incremental_map(pe
, con
, map
);
1258 note_peer_epoch(peer
, map
->get_epoch());
1260 dout(20) << "share_map_peer " << con
<< " already has epoch " << pe
<< dendl
;
1262 dout(20) << "share_map_peer " << con
<< " don't know epoch, doing nothing" << dendl
;
1263 // no idea about peer's epoch.
1264 // ??? send recent ???
1269 bool OSDService::can_inc_scrubs_pending()
1271 bool can_inc
= false;
1272 Mutex::Locker
l(sched_scrub_lock
);
1274 if (scrubs_pending
+ scrubs_active
< cct
->_conf
->osd_max_scrubs
) {
1275 dout(20) << __func__
<< " " << scrubs_pending
<< " -> " << (scrubs_pending
+1)
1276 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
1280 dout(20) << __func__
<< " " << scrubs_pending
<< " + " << scrubs_active
1281 << " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1287 bool OSDService::inc_scrubs_pending()
1289 bool result
= false;
1291 sched_scrub_lock
.Lock();
1292 if (scrubs_pending
+ scrubs_active
< cct
->_conf
->osd_max_scrubs
) {
1293 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
+1)
1294 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1298 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " + " << scrubs_active
<< " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1300 sched_scrub_lock
.Unlock();
1305 void OSDService::dec_scrubs_pending()
1307 sched_scrub_lock
.Lock();
1308 dout(20) << "dec_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
-1)
1309 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1311 assert(scrubs_pending
>= 0);
1312 sched_scrub_lock
.Unlock();
1315 void OSDService::inc_scrubs_active(bool reserved
)
1317 sched_scrub_lock
.Lock();
1321 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1322 << " (max " << cct
->_conf
->osd_max_scrubs
1323 << ", pending " << (scrubs_pending
+1) << " -> " << scrubs_pending
<< ")" << dendl
;
1324 assert(scrubs_pending
>= 0);
1326 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1327 << " (max " << cct
->_conf
->osd_max_scrubs
1328 << ", pending " << scrubs_pending
<< ")" << dendl
;
1330 sched_scrub_lock
.Unlock();
1333 void OSDService::dec_scrubs_active()
1335 sched_scrub_lock
.Lock();
1336 dout(20) << "dec_scrubs_active " << scrubs_active
<< " -> " << (scrubs_active
-1)
1337 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", pending " << scrubs_pending
<< ")" << dendl
;
1339 assert(scrubs_active
>= 0);
1340 sched_scrub_lock
.Unlock();
1343 void OSDService::retrieve_epochs(epoch_t
*_boot_epoch
, epoch_t
*_up_epoch
,
1344 epoch_t
*_bind_epoch
) const
1346 Mutex::Locker
l(epoch_lock
);
1348 *_boot_epoch
= boot_epoch
;
1350 *_up_epoch
= up_epoch
;
1352 *_bind_epoch
= bind_epoch
;
1355 void OSDService::set_epochs(const epoch_t
*_boot_epoch
, const epoch_t
*_up_epoch
,
1356 const epoch_t
*_bind_epoch
)
1358 Mutex::Locker
l(epoch_lock
);
1360 assert(*_boot_epoch
== 0 || *_boot_epoch
>= boot_epoch
);
1361 boot_epoch
= *_boot_epoch
;
1364 assert(*_up_epoch
== 0 || *_up_epoch
>= up_epoch
);
1365 up_epoch
= *_up_epoch
;
1368 assert(*_bind_epoch
== 0 || *_bind_epoch
>= bind_epoch
);
1369 bind_epoch
= *_bind_epoch
;
1373 bool OSDService::prepare_to_stop()
1375 Mutex::Locker
l(is_stopping_lock
);
1376 if (get_state() != NOT_STOPPING
)
1379 OSDMapRef osdmap
= get_osdmap();
1380 if (osdmap
&& osdmap
->is_up(whoami
)) {
1381 dout(0) << __func__
<< " telling mon we are shutting down" << dendl
;
1382 set_state(PREPARING_TO_STOP
);
1383 monc
->send_mon_message(new MOSDMarkMeDown(monc
->get_fsid(),
1384 osdmap
->get_inst(whoami
),
1385 osdmap
->get_epoch(),
1388 utime_t now
= ceph_clock_now();
1390 timeout
.set_from_double(now
+ cct
->_conf
->osd_mon_shutdown_timeout
);
1391 while ((ceph_clock_now() < timeout
) &&
1392 (get_state() != STOPPING
)) {
1393 is_stopping_cond
.WaitUntil(is_stopping_lock
, timeout
);
1396 dout(0) << __func__
<< " starting shutdown" << dendl
;
1397 set_state(STOPPING
);
1401 void OSDService::got_stop_ack()
1403 Mutex::Locker
l(is_stopping_lock
);
1404 if (get_state() == PREPARING_TO_STOP
) {
1405 dout(0) << __func__
<< " starting shutdown" << dendl
;
1406 set_state(STOPPING
);
1407 is_stopping_cond
.Signal();
1409 dout(10) << __func__
<< " ignoring msg" << dendl
;
1413 MOSDMap
*OSDService::build_incremental_map_msg(epoch_t since
, epoch_t to
,
1414 OSDSuperblock
& sblock
)
1416 MOSDMap
*m
= new MOSDMap(monc
->get_fsid(),
1417 osdmap
->get_encoding_features());
1418 m
->oldest_map
= max_oldest_map
;
1419 m
->newest_map
= sblock
.newest_map
;
1421 for (epoch_t e
= to
; e
> since
; e
--) {
1423 if (e
> m
->oldest_map
&& get_inc_map_bl(e
, bl
)) {
1424 m
->incremental_maps
[e
].claim(bl
);
1425 } else if (get_map_bl(e
, bl
)) {
1426 m
->maps
[e
].claim(bl
);
1429 derr
<< "since " << since
<< " to " << to
1430 << " oldest " << m
->oldest_map
<< " newest " << m
->newest_map
1440 void OSDService::send_map(MOSDMap
*m
, Connection
*con
)
1442 con
->send_message(m
);
1445 void OSDService::send_incremental_map(epoch_t since
, Connection
*con
,
1448 epoch_t to
= osdmap
->get_epoch();
1449 dout(10) << "send_incremental_map " << since
<< " -> " << to
1450 << " to " << con
<< " " << con
->get_peer_addr() << dendl
;
1454 OSDSuperblock
sblock(get_superblock());
1455 if (since
< sblock
.oldest_map
) {
1456 // just send latest full map
1457 MOSDMap
*m
= new MOSDMap(monc
->get_fsid(),
1458 osdmap
->get_encoding_features());
1459 m
->oldest_map
= max_oldest_map
;
1460 m
->newest_map
= sblock
.newest_map
;
1461 get_map_bl(to
, m
->maps
[to
]);
1466 if (to
> since
&& (int64_t)(to
- since
) > cct
->_conf
->osd_map_share_max_epochs
) {
1467 dout(10) << " " << (to
- since
) << " > max " << cct
->_conf
->osd_map_share_max_epochs
1468 << ", only sending most recent" << dendl
;
1469 since
= to
- cct
->_conf
->osd_map_share_max_epochs
;
1472 if (to
- since
> (epoch_t
)cct
->_conf
->osd_map_message_max
)
1473 to
= since
+ cct
->_conf
->osd_map_message_max
;
1474 m
= build_incremental_map_msg(since
, to
, sblock
);
1479 bool OSDService::_get_map_bl(epoch_t e
, bufferlist
& bl
)
1481 bool found
= map_bl_cache
.lookup(e
, &bl
);
1484 logger
->inc(l_osd_map_bl_cache_hit
);
1488 logger
->inc(l_osd_map_bl_cache_miss
);
1489 found
= store
->read(coll_t::meta(),
1490 OSD::get_osdmap_pobject_name(e
), 0, 0, bl
,
1491 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1498 bool OSDService::get_inc_map_bl(epoch_t e
, bufferlist
& bl
)
1500 Mutex::Locker
l(map_cache_lock
);
1501 bool found
= map_bl_inc_cache
.lookup(e
, &bl
);
1504 logger
->inc(l_osd_map_bl_cache_hit
);
1508 logger
->inc(l_osd_map_bl_cache_miss
);
1509 found
= store
->read(coll_t::meta(),
1510 OSD::get_inc_osdmap_pobject_name(e
), 0, 0, bl
,
1511 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1513 _add_map_inc_bl(e
, bl
);
1518 void OSDService::_add_map_bl(epoch_t e
, bufferlist
& bl
)
1520 dout(10) << "add_map_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1521 // cache a contiguous buffer
1522 if (bl
.get_num_buffers() > 1) {
1525 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1526 map_bl_cache
.add(e
, bl
);
1529 void OSDService::_add_map_inc_bl(epoch_t e
, bufferlist
& bl
)
1531 dout(10) << "add_map_inc_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1532 // cache a contiguous buffer
1533 if (bl
.get_num_buffers() > 1) {
1536 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1537 map_bl_inc_cache
.add(e
, bl
);
1540 void OSDService::pin_map_inc_bl(epoch_t e
, bufferlist
&bl
)
1542 Mutex::Locker
l(map_cache_lock
);
1543 // cache a contiguous buffer
1544 if (bl
.get_num_buffers() > 1) {
1547 map_bl_inc_cache
.pin(e
, bl
);
1550 void OSDService::pin_map_bl(epoch_t e
, bufferlist
&bl
)
1552 Mutex::Locker
l(map_cache_lock
);
1553 // cache a contiguous buffer
1554 if (bl
.get_num_buffers() > 1) {
1557 map_bl_cache
.pin(e
, bl
);
1560 void OSDService::clear_map_bl_cache_pins(epoch_t e
)
1562 Mutex::Locker
l(map_cache_lock
);
1563 map_bl_inc_cache
.clear_pinned(e
);
1564 map_bl_cache
.clear_pinned(e
);
1567 OSDMapRef
OSDService::_add_map(OSDMap
*o
)
1569 epoch_t e
= o
->get_epoch();
1571 if (cct
->_conf
->osd_map_dedup
) {
1572 // Dedup against an existing map at a nearby epoch
1573 OSDMapRef for_dedup
= map_cache
.lower_bound(e
);
1575 OSDMap::dedup(for_dedup
.get(), o
);
1579 OSDMapRef l
= map_cache
.add(e
, o
, &existed
);
1586 OSDMapRef
OSDService::try_get_map(epoch_t epoch
)
1588 Mutex::Locker
l(map_cache_lock
);
1589 OSDMapRef retval
= map_cache
.lookup(epoch
);
1591 dout(30) << "get_map " << epoch
<< " -cached" << dendl
;
1593 logger
->inc(l_osd_map_cache_hit
);
1598 logger
->inc(l_osd_map_cache_miss
);
1599 epoch_t lb
= map_cache
.cached_key_lower_bound();
1601 dout(30) << "get_map " << epoch
<< " - miss, below lower bound" << dendl
;
1602 logger
->inc(l_osd_map_cache_miss_low
);
1603 logger
->inc(l_osd_map_cache_miss_low_avg
, lb
- epoch
);
1607 OSDMap
*map
= new OSDMap
;
1609 dout(20) << "get_map " << epoch
<< " - loading and decoding " << map
<< dendl
;
1611 if (!_get_map_bl(epoch
, bl
) || bl
.length() == 0) {
1612 derr
<< "failed to load OSD map for epoch " << epoch
<< ", got " << bl
.length() << " bytes" << dendl
;
1618 dout(20) << "get_map " << epoch
<< " - return initial " << map
<< dendl
;
1620 return _add_map(map
);
1626 void OSDService::reply_op_error(OpRequestRef op
, int err
)
1628 reply_op_error(op
, err
, eversion_t(), 0);
1631 void OSDService::reply_op_error(OpRequestRef op
, int err
, eversion_t v
,
1634 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1635 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1637 flags
= m
->get_flags() & (CEPH_OSD_FLAG_ACK
|CEPH_OSD_FLAG_ONDISK
);
1639 MOSDOpReply
*reply
= new MOSDOpReply(m
, err
, osdmap
->get_epoch(), flags
,
1641 reply
->set_reply_versions(v
, uv
);
1642 m
->get_connection()->send_message(reply
);
1645 void OSDService::handle_misdirected_op(PG
*pg
, OpRequestRef op
)
1647 if (!cct
->_conf
->osd_debug_misdirected_ops
) {
1651 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1652 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1654 assert(m
->get_map_epoch() >= pg
->info
.history
.same_primary_since
);
1656 if (pg
->is_ec_pg()) {
1658 * OSD recomputes op target based on current OSDMap. With an EC pg, we
1659 * can get this result:
1660 * 1) client at map 512 sends an op to osd 3, pg_t 3.9 based on mapping
1661 * [CRUSH_ITEM_NONE, 2, 3]/3
1662 * 2) OSD 3 at map 513 remaps op to osd 3, spg_t 3.9s0 based on mapping
1664 * 3) PG 3.9s0 dequeues the op at epoch 512 and notices that it isn't primary
1666 * 4) client resends and this time PG 3.9s0 having caught up to 513 gets
1669 * We can't compute the op target based on the sending map epoch due to
1670 * splitting. The simplest thing is to detect such cases here and drop
1671 * them without an error (the client will resend anyway).
1673 assert(m
->get_map_epoch() <= superblock
.newest_map
);
1674 OSDMapRef opmap
= try_get_map(m
->get_map_epoch());
1676 dout(7) << __func__
<< ": " << *pg
<< " no longer have map for "
1677 << m
->get_map_epoch() << ", dropping" << dendl
;
1680 pg_t _pgid
= m
->get_raw_pg();
1682 if ((m
->get_flags() & CEPH_OSD_FLAG_PGOP
) == 0)
1683 _pgid
= opmap
->raw_pg_to_pg(_pgid
);
1684 if (opmap
->get_primary_shard(_pgid
, &pgid
) &&
1685 pgid
.shard
!= pg
->info
.pgid
.shard
) {
1686 dout(7) << __func__
<< ": " << *pg
<< " primary changed since "
1687 << m
->get_map_epoch() << ", dropping" << dendl
;
1692 dout(7) << *pg
<< " misdirected op in " << m
->get_map_epoch() << dendl
;
1693 clog
->warn() << m
->get_source_inst() << " misdirected " << m
->get_reqid()
1694 << " pg " << m
->get_raw_pg()
1695 << " to osd." << whoami
1696 << " not " << pg
->acting
1697 << " in e" << m
->get_map_epoch() << "/" << osdmap
->get_epoch();
1700 void OSDService::enqueue_back(spg_t pgid
, PGQueueable qi
)
1702 osd
->op_shardedwq
.queue(make_pair(pgid
, qi
));
1705 void OSDService::enqueue_front(spg_t pgid
, PGQueueable qi
)
1707 osd
->op_shardedwq
.queue_front(make_pair(pgid
, qi
));
1710 void OSDService::queue_for_peering(PG
*pg
)
1712 peering_wq
.queue(pg
);
1715 void OSDService::queue_for_snap_trim(PG
*pg
)
1717 dout(10) << "queueing " << *pg
<< " for snaptrim" << dendl
;
1718 osd
->op_shardedwq
.queue(
1722 PGSnapTrim(pg
->get_osdmap()->get_epoch()),
1723 cct
->_conf
->osd_snap_trim_cost
,
1724 cct
->_conf
->osd_snap_trim_priority
,
1727 pg
->get_osdmap()->get_epoch())));
1731 // ====================================================================
1735 #define dout_prefix *_dout
1737 // Commands shared between OSD's console and admin console:
1739 namespace osd_cmds
{
1741 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
);
1743 }} // namespace ceph::osd_cmds
1745 int OSD::mkfs(CephContext
*cct
, ObjectStore
*store
, const string
&dev
,
1746 uuid_d fsid
, int whoami
)
1750 ceph::shared_ptr
<ObjectStore::Sequencer
> osr(
1751 new ObjectStore::Sequencer("mkfs"));
1756 // if we are fed a uuid for this osd, use it.
1757 store
->set_fsid(cct
->_conf
->osd_uuid
);
1759 ret
= store
->mkfs();
1761 derr
<< "OSD::mkfs: ObjectStore::mkfs failed with error "
1762 << cpp_strerror(ret
) << dendl
;
1766 store
->set_cache_shards(1); // doesn't matter for mkfs!
1768 ret
= store
->mount();
1770 derr
<< "OSD::mkfs: couldn't mount ObjectStore: error "
1771 << cpp_strerror(ret
) << dendl
;
1775 ret
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, sbbl
);
1777 /* if we already have superblock, check content of superblock */
1778 dout(0) << " have superblock" << dendl
;
1779 bufferlist::iterator p
;
1782 if (whoami
!= sb
.whoami
) {
1783 derr
<< "provided osd id " << whoami
<< " != superblock's " << sb
.whoami
1788 if (fsid
!= sb
.cluster_fsid
) {
1789 derr
<< "provided cluster fsid " << fsid
1790 << " != superblock's " << sb
.cluster_fsid
<< dendl
;
1795 // create superblock
1796 sb
.cluster_fsid
= fsid
;
1797 sb
.osd_fsid
= store
->get_fsid();
1799 sb
.compat_features
= get_osd_initial_compat_set();
1804 ObjectStore::Transaction t
;
1805 t
.create_collection(coll_t::meta(), 0);
1806 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
1807 ret
= store
->apply_transaction(osr
.get(), std::move(t
));
1809 derr
<< "OSD::mkfs: error while writing OSD_SUPERBLOCK_GOBJECT: "
1810 << "apply_transaction returned " << cpp_strerror(ret
) << dendl
;
1815 if (!osr
->flush_commit(&waiter
)) {
1819 ret
= write_meta(cct
, store
, sb
.cluster_fsid
, sb
.osd_fsid
, whoami
);
1821 derr
<< "OSD::mkfs: failed to write fsid file: error "
1822 << cpp_strerror(ret
) << dendl
;
1833 int OSD::write_meta(CephContext
*cct
, ObjectStore
*store
, uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int whoami
)
1838 snprintf(val
, sizeof(val
), "%s", CEPH_OSD_ONDISK_MAGIC
);
1839 r
= store
->write_meta("magic", val
);
1843 snprintf(val
, sizeof(val
), "%d", whoami
);
1844 r
= store
->write_meta("whoami", val
);
1848 cluster_fsid
.print(val
);
1849 r
= store
->write_meta("ceph_fsid", val
);
1853 string key
= cct
->_conf
->get_val
<string
>("key");
1855 r
= store
->write_meta("osd_key", key
);
1859 string keyfile
= cct
->_conf
->get_val
<string
>("keyfile");
1860 if (!keyfile
.empty()) {
1863 if (keyfile
== "-") {
1864 static_assert(1024 * 1024 >
1865 (sizeof(CryptoKey
) - sizeof(bufferptr
) +
1866 sizeof(__u16
) + 16 /* AES_KEY_LEN */ + 3 - 1) / 3. * 4.,
1867 "1MB should be enough for a base64 encoded CryptoKey");
1868 r
= keybl
.read_fd(STDIN_FILENO
, 1024 * 1024);
1870 r
= keybl
.read_file(keyfile
.c_str(), &err
);
1873 derr
<< __func__
<< " failed to read keyfile " << keyfile
<< ": "
1874 << err
<< ": " << cpp_strerror(r
) << dendl
;
1877 r
= store
->write_meta("osd_key", keybl
.to_str());
1883 r
= store
->write_meta("ready", "ready");
1890 int OSD::peek_meta(ObjectStore
*store
, std::string
& magic
,
1891 uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int& whoami
)
1895 int r
= store
->read_meta("magic", &val
);
1900 r
= store
->read_meta("whoami", &val
);
1903 whoami
= atoi(val
.c_str());
1905 r
= store
->read_meta("ceph_fsid", &val
);
1908 r
= cluster_fsid
.parse(val
.c_str());
1912 r
= store
->read_meta("fsid", &val
);
1914 osd_fsid
= uuid_d();
1916 r
= osd_fsid
.parse(val
.c_str());
1926 #define dout_prefix _prefix(_dout, whoami, get_osdmap_epoch())
1930 OSD::OSD(CephContext
*cct_
, ObjectStore
*store_
,
1932 Messenger
*internal_messenger
,
1933 Messenger
*external_messenger
,
1934 Messenger
*hb_client_front
,
1935 Messenger
*hb_client_back
,
1936 Messenger
*hb_front_serverm
,
1937 Messenger
*hb_back_serverm
,
1938 Messenger
*osdc_messenger
,
1940 const std::string
&dev
, const std::string
&jdev
) :
1942 osd_lock("OSD::osd_lock"),
1943 tick_timer(cct
, osd_lock
),
1944 tick_timer_lock("OSD::tick_timer_lock"),
1945 tick_timer_without_osd_lock(cct
, tick_timer_lock
),
1946 authorize_handler_cluster_registry(new AuthAuthorizeHandlerRegistry(cct
,
1947 cct
->_conf
->auth_supported
.empty() ?
1948 cct
->_conf
->auth_cluster_required
:
1949 cct
->_conf
->auth_supported
)),
1950 authorize_handler_service_registry(new AuthAuthorizeHandlerRegistry(cct
,
1951 cct
->_conf
->auth_supported
.empty() ?
1952 cct
->_conf
->auth_service_required
:
1953 cct
->_conf
->auth_supported
)),
1954 cluster_messenger(internal_messenger
),
1955 client_messenger(external_messenger
),
1956 objecter_messenger(osdc_messenger
),
1958 mgrc(cct_
, client_messenger
),
1960 recoverystate_perf(NULL
),
1962 log_client(cct
, client_messenger
, &mc
->monmap
, LogClient::NO_FLAGS
),
1963 clog(log_client
.create_channel()),
1965 dev_path(dev
), journal_path(jdev
),
1966 store_is_rotational(store
->is_rotational()),
1967 trace_endpoint("0.0.0.0", 0, "osd"),
1969 osd_compat(get_osd_compat_set()),
1970 peering_tp(cct
, "OSD::peering_tp", "tp_peering",
1971 cct
->_conf
->osd_peering_wq_threads
,
1972 "osd_peering_tp_threads"),
1973 osd_op_tp(cct
, "OSD::osd_op_tp", "tp_osd_tp",
1974 get_num_op_threads()),
1975 disk_tp(cct
, "OSD::disk_tp", "tp_osd_disk", cct
->_conf
->osd_disk_threads
, "osd_disk_threads"),
1976 command_tp(cct
, "OSD::command_tp", "tp_osd_cmd", 1),
1977 session_waiting_lock("OSD::session_waiting_lock"),
1978 osdmap_subscribe_lock("OSD::osdmap_subscribe_lock"),
1979 heartbeat_lock("OSD::heartbeat_lock"),
1980 heartbeat_stop(false),
1981 heartbeat_need_update(true),
1982 hb_front_client_messenger(hb_client_front
),
1983 hb_back_client_messenger(hb_client_back
),
1984 hb_front_server_messenger(hb_front_serverm
),
1985 hb_back_server_messenger(hb_back_serverm
),
1987 heartbeat_thread(this),
1988 heartbeat_dispatcher(this),
1989 op_tracker(cct
, cct
->_conf
->osd_enable_op_tracker
,
1990 cct
->_conf
->osd_num_op_tracker_shard
),
1991 test_ops_hook(NULL
),
1992 op_queue(get_io_queue()),
1993 op_prio_cutoff(get_io_prio_cut()),
1995 get_num_op_shards(),
1997 cct
->_conf
->osd_op_thread_timeout
,
1998 cct
->_conf
->osd_op_thread_suicide_timeout
,
2002 cct
->_conf
->osd_op_thread_timeout
,
2003 cct
->_conf
->osd_op_thread_suicide_timeout
,
2005 map_lock("OSD::map_lock"),
2006 pg_map_lock("OSD::pg_map_lock"),
2007 last_pg_create_epoch(0),
2008 mon_report_lock("OSD::mon_report_lock"),
2009 stats_ack_timeout(cct
->_conf
->osd_mon_ack_timeout
),
2011 requested_full_first(0),
2012 requested_full_last(0),
2013 pg_stat_queue_lock("OSD::pg_stat_queue_lock"),
2014 osd_stat_updated(false),
2015 pg_stat_tid(0), pg_stat_tid_flushed(0),
2018 cct
->_conf
->osd_command_thread_timeout
,
2019 cct
->_conf
->osd_command_thread_suicide_timeout
,
2024 cct
->_conf
->osd_remove_thread_timeout
,
2025 cct
->_conf
->osd_remove_thread_suicide_timeout
,
2029 monc
->set_messenger(client_messenger
);
2030 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
2031 cct
->_conf
->osd_op_log_threshold
);
2032 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
2033 cct
->_conf
->osd_op_history_duration
);
2034 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
2035 cct
->_conf
->osd_op_history_slow_op_threshold
);
2037 std::stringstream ss
;
2038 ss
<< "osd." << whoami
;
2039 trace_endpoint
.copy_name(ss
.str());
2045 delete authorize_handler_cluster_registry
;
2046 delete authorize_handler_service_registry
;
2047 delete class_handler
;
2048 cct
->get_perfcounters_collection()->remove(recoverystate_perf
);
2049 cct
->get_perfcounters_collection()->remove(logger
);
2050 delete recoverystate_perf
;
2055 void cls_initialize(ClassHandler
*ch
);
2057 void OSD::handle_signal(int signum
)
2059 assert(signum
== SIGINT
|| signum
== SIGTERM
);
2060 derr
<< "*** Got signal " << sig_str(signum
) << " ***" << dendl
;
2066 Mutex::Locker
lock(osd_lock
);
2070 if (store
->test_mount_in_use()) {
2071 derr
<< "OSD::pre_init: object store '" << dev_path
<< "' is "
2072 << "currently in use. (Is ceph-osd already running?)" << dendl
;
2076 cct
->_conf
->add_observer(this);
2082 class OSDSocketHook
: public AdminSocketHook
{
2085 explicit OSDSocketHook(OSD
*o
) : osd(o
) {}
2086 bool call(std::string admin_command
, cmdmap_t
& cmdmap
, std::string format
,
2087 bufferlist
& out
) override
{
2089 bool r
= osd
->asok_command(admin_command
, cmdmap
, format
, ss
);
2095 bool OSD::asok_command(string admin_command
, cmdmap_t
& cmdmap
, string format
,
2098 Formatter
*f
= Formatter::create(format
, "json-pretty", "json-pretty");
2099 if (admin_command
== "status") {
2100 f
->open_object_section("status");
2101 f
->dump_stream("cluster_fsid") << superblock
.cluster_fsid
;
2102 f
->dump_stream("osd_fsid") << superblock
.osd_fsid
;
2103 f
->dump_unsigned("whoami", superblock
.whoami
);
2104 f
->dump_string("state", get_state_name(get_state()));
2105 f
->dump_unsigned("oldest_map", superblock
.oldest_map
);
2106 f
->dump_unsigned("newest_map", superblock
.newest_map
);
2108 RWLock::RLocker
l(pg_map_lock
);
2109 f
->dump_unsigned("num_pgs", pg_map
.size());
2112 } else if (admin_command
== "flush_journal") {
2113 store
->flush_journal();
2114 } else if (admin_command
== "dump_ops_in_flight" ||
2115 admin_command
== "ops" ||
2116 admin_command
== "dump_blocked_ops" ||
2117 admin_command
== "dump_historic_ops" ||
2118 admin_command
== "dump_historic_ops_by_duration" ||
2119 admin_command
== "dump_historic_slow_ops") {
2121 const string error_str
= "op_tracker tracking is not enabled now, so no ops are tracked currently, \
2122 even those get stuck. Please enable \"osd_enable_op_tracker\", and the tracker \
2123 will start to track new ops received afterwards.";
2125 set
<string
> filters
;
2126 vector
<string
> filter_str
;
2127 if (cmd_getval(cct
, cmdmap
, "filterstr", filter_str
)) {
2128 copy(filter_str
.begin(), filter_str
.end(),
2129 inserter(filters
, filters
.end()));
2132 if (admin_command
== "dump_ops_in_flight" ||
2133 admin_command
== "ops") {
2134 if (!op_tracker
.dump_ops_in_flight(f
, false, filters
)) {
2138 if (admin_command
== "dump_blocked_ops") {
2139 if (!op_tracker
.dump_ops_in_flight(f
, true, filters
)) {
2143 if (admin_command
== "dump_historic_ops") {
2144 if (!op_tracker
.dump_historic_ops(f
, false, filters
)) {
2148 if (admin_command
== "dump_historic_ops_by_duration") {
2149 if (!op_tracker
.dump_historic_ops(f
, true, filters
)) {
2153 if (admin_command
== "dump_historic_slow_ops") {
2154 if (!op_tracker
.dump_historic_slow_ops(f
, filters
)) {
2158 } else if (admin_command
== "dump_op_pq_state") {
2159 f
->open_object_section("pq");
2160 op_shardedwq
.dump(f
);
2162 } else if (admin_command
== "dump_blacklist") {
2163 list
<pair
<entity_addr_t
,utime_t
> > bl
;
2164 OSDMapRef curmap
= service
.get_osdmap();
2166 f
->open_array_section("blacklist");
2167 curmap
->get_blacklist(&bl
);
2168 for (list
<pair
<entity_addr_t
,utime_t
> >::iterator it
= bl
.begin();
2169 it
!= bl
.end(); ++it
) {
2170 f
->open_object_section("entry");
2171 f
->open_object_section("entity_addr_t");
2173 f
->close_section(); //entity_addr_t
2174 it
->second
.localtime(f
->dump_stream("expire_time"));
2175 f
->close_section(); //entry
2177 f
->close_section(); //blacklist
2178 } else if (admin_command
== "dump_watchers") {
2179 list
<obj_watch_item_t
> watchers
;
2182 Mutex::Locker
l(osd_lock
);
2183 RWLock::RLocker
l2(pg_map_lock
);
2184 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2188 list
<obj_watch_item_t
> pg_watchers
;
2189 PG
*pg
= it
->second
;
2191 pg
->get_watchers(pg_watchers
);
2193 watchers
.splice(watchers
.end(), pg_watchers
);
2197 f
->open_array_section("watchers");
2198 for (list
<obj_watch_item_t
>::iterator it
= watchers
.begin();
2199 it
!= watchers
.end(); ++it
) {
2201 f
->open_object_section("watch");
2203 f
->dump_string("namespace", it
->obj
.nspace
);
2204 f
->dump_string("object", it
->obj
.oid
.name
);
2206 f
->open_object_section("entity_name");
2207 it
->wi
.name
.dump(f
);
2208 f
->close_section(); //entity_name_t
2210 f
->dump_unsigned("cookie", it
->wi
.cookie
);
2211 f
->dump_unsigned("timeout", it
->wi
.timeout_seconds
);
2213 f
->open_object_section("entity_addr_t");
2214 it
->wi
.addr
.dump(f
);
2215 f
->close_section(); //entity_addr_t
2217 f
->close_section(); //watch
2220 f
->close_section(); //watchers
2221 } else if (admin_command
== "dump_reservations") {
2222 f
->open_object_section("reservations");
2223 f
->open_object_section("local_reservations");
2224 service
.local_reserver
.dump(f
);
2226 f
->open_object_section("remote_reservations");
2227 service
.remote_reserver
.dump(f
);
2230 } else if (admin_command
== "get_latest_osdmap") {
2231 get_latest_osdmap();
2232 } else if (admin_command
== "heap") {
2233 auto result
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ss
);
2235 // Note: Failed heap profile commands won't necessarily trigger an error:
2236 f
->open_object_section("result");
2237 f
->dump_string("error", cpp_strerror(result
));
2238 f
->dump_bool("success", result
>= 0);
2240 } else if (admin_command
== "set_heap_property") {
2244 bool success
= false;
2245 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2246 error
= "unable to get property";
2248 } else if (!cmd_getval(cct
, cmdmap
, "value", value
)) {
2249 error
= "unable to get value";
2251 } else if (value
< 0) {
2252 error
= "negative value not allowed";
2254 } else if (!ceph_heap_set_numeric_property(property
.c_str(), (size_t)value
)) {
2255 error
= "invalid property";
2260 f
->open_object_section("result");
2261 f
->dump_string("error", error
);
2262 f
->dump_bool("success", success
);
2264 } else if (admin_command
== "get_heap_property") {
2268 bool success
= false;
2269 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2270 error
= "unable to get property";
2272 } else if (!ceph_heap_get_numeric_property(property
.c_str(), &value
)) {
2273 error
= "invalid property";
2278 f
->open_object_section("result");
2279 f
->dump_string("error", error
);
2280 f
->dump_bool("success", success
);
2281 f
->dump_int("value", value
);
2283 } else if (admin_command
== "dump_objectstore_kv_stats") {
2284 store
->get_db_statistics(f
);
2285 } else if (admin_command
== "dump_scrubs") {
2286 service
.dumps_scrub(f
);
2287 } else if (admin_command
== "calc_objectstore_db_histogram") {
2288 store
->generate_db_histogram(f
);
2289 } else if (admin_command
== "flush_store_cache") {
2290 store
->flush_cache();
2291 } else if (admin_command
== "dump_pgstate_history") {
2292 f
->open_object_section("pgstate_history");
2293 RWLock::RLocker
l2(pg_map_lock
);
2294 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2298 PG
*pg
= it
->second
;
2299 f
->dump_stream("pg") << pg
->get_pgid();
2301 pg
->pgstate_history
.dump(f
);
2305 } else if (admin_command
== "compact") {
2306 dout(1) << "triggering manual compaction" << dendl
;
2307 auto start
= ceph::coarse_mono_clock::now();
2309 auto end
= ceph::coarse_mono_clock::now();
2310 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
2311 dout(1) << "finished manual compaction in "
2312 << time_span
.count()
2313 << " seconds" << dendl
;
2314 f
->open_object_section("compact_result");
2315 f
->dump_float("elapsed_time", time_span
.count());
2318 assert(0 == "broken asok registration");
2325 class TestOpsSocketHook
: public AdminSocketHook
{
2326 OSDService
*service
;
2329 TestOpsSocketHook(OSDService
*s
, ObjectStore
*st
) : service(s
), store(st
) {}
2330 bool call(std::string command
, cmdmap_t
& cmdmap
, std::string format
,
2331 bufferlist
& out
) override
{
2333 test_ops(service
, store
, command
, cmdmap
, ss
);
2337 void test_ops(OSDService
*service
, ObjectStore
*store
,
2338 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
);
2342 class OSD::C_Tick
: public Context
{
2345 explicit C_Tick(OSD
*o
) : osd(o
) {}
2346 void finish(int r
) override
{
2351 class OSD::C_Tick_WithoutOSDLock
: public Context
{
2354 explicit C_Tick_WithoutOSDLock(OSD
*o
) : osd(o
) {}
2355 void finish(int r
) override
{
2356 osd
->tick_without_osd_lock();
2360 int OSD::enable_disable_fuse(bool stop
)
2364 string mntpath
= cct
->_conf
->osd_data
+ "/fuse";
2365 if (fuse_store
&& (stop
|| !cct
->_conf
->osd_objectstore_fuse
)) {
2366 dout(1) << __func__
<< " disabling" << dendl
;
2370 r
= ::rmdir(mntpath
.c_str());
2373 derr
<< __func__
<< " failed to rmdir " << mntpath
<< ": "
2374 << cpp_strerror(r
) << dendl
;
2379 if (!fuse_store
&& cct
->_conf
->osd_objectstore_fuse
) {
2380 dout(1) << __func__
<< " enabling" << dendl
;
2381 r
= ::mkdir(mntpath
.c_str(), 0700);
2384 if (r
< 0 && r
!= -EEXIST
) {
2385 derr
<< __func__
<< " unable to create " << mntpath
<< ": "
2386 << cpp_strerror(r
) << dendl
;
2389 fuse_store
= new FuseStore(store
, mntpath
);
2390 r
= fuse_store
->start();
2392 derr
<< __func__
<< " unable to start fuse: " << cpp_strerror(r
) << dendl
;
2398 #endif // HAVE_LIBFUSE
2402 int OSD::get_num_op_shards()
2404 if (cct
->_conf
->osd_op_num_shards
)
2405 return cct
->_conf
->osd_op_num_shards
;
2406 if (store_is_rotational
)
2407 return cct
->_conf
->osd_op_num_shards_hdd
;
2409 return cct
->_conf
->osd_op_num_shards_ssd
;
2412 int OSD::get_num_op_threads()
2414 if (cct
->_conf
->osd_op_num_threads_per_shard
)
2415 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard
;
2416 if (store_is_rotational
)
2417 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_hdd
;
2419 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_ssd
;
2422 float OSD::get_osd_recovery_sleep()
2424 if (cct
->_conf
->osd_recovery_sleep
)
2425 return cct
->_conf
->osd_recovery_sleep
;
2426 if (!store_is_rotational
&& !journal_is_rotational
)
2427 return cct
->_conf
->osd_recovery_sleep_ssd
;
2428 else if (store_is_rotational
&& !journal_is_rotational
)
2429 return cct
->_conf
->get_val
<double>("osd_recovery_sleep_hybrid");
2431 return cct
->_conf
->osd_recovery_sleep_hdd
;
2436 CompatSet initial
, diff
;
2437 Mutex::Locker
lock(osd_lock
);
2442 tick_timer_without_osd_lock
.init();
2443 service
.recovery_request_timer
.init();
2444 service
.recovery_sleep_timer
.init();
2447 dout(2) << "init " << dev_path
2448 << " (looks like " << (store_is_rotational
? "hdd" : "ssd") << ")"
2450 dout(2) << "journal " << journal_path
<< dendl
;
2451 assert(store
); // call pre_init() first!
2453 store
->set_cache_shards(get_num_op_shards());
2455 int r
= store
->mount();
2457 derr
<< "OSD:init: unable to mount object store" << dendl
;
2460 journal_is_rotational
= store
->is_journal_rotational();
2461 dout(2) << "journal looks like " << (journal_is_rotational
? "hdd" : "ssd")
2464 enable_disable_fuse(false);
2466 dout(2) << "boot" << dendl
;
2468 // initialize the daily loadavg with current 15min loadavg
2470 if (getloadavg(loadavgs
, 3) == 3) {
2471 daily_loadavg
= loadavgs
[2];
2473 derr
<< "OSD::init() : couldn't read loadavgs\n" << dendl
;
2474 daily_loadavg
= 1.0;
2477 int rotating_auth_attempts
= 0;
2479 // sanity check long object name handling
2482 l
.oid
.name
= string(cct
->_conf
->osd_max_object_name_len
, 'n');
2483 l
.set_key(string(cct
->_conf
->osd_max_object_name_len
, 'k'));
2484 l
.nspace
= string(cct
->_conf
->osd_max_object_namespace_len
, 's');
2485 r
= store
->validate_hobject_key(l
);
2487 derr
<< "backend (" << store
->get_type() << ") is unable to support max "
2488 << "object name[space] len" << dendl
;
2489 derr
<< " osd max object name len = "
2490 << cct
->_conf
->osd_max_object_name_len
<< dendl
;
2491 derr
<< " osd max object namespace len = "
2492 << cct
->_conf
->osd_max_object_namespace_len
<< dendl
;
2493 derr
<< cpp_strerror(r
) << dendl
;
2494 if (cct
->_conf
->osd_check_max_object_name_len_on_startup
) {
2497 derr
<< "osd_check_max_object_name_len_on_startup = false, starting anyway"
2500 dout(20) << "configured osd_max_object_name[space]_len looks ok" << dendl
;
2505 r
= read_superblock();
2507 derr
<< "OSD::init() : unable to read osd superblock" << dendl
;
2512 if (osd_compat
.compare(superblock
.compat_features
) < 0) {
2513 derr
<< "The disk uses features unsupported by the executable." << dendl
;
2514 derr
<< " ondisk features " << superblock
.compat_features
<< dendl
;
2515 derr
<< " daemon features " << osd_compat
<< dendl
;
2517 if (osd_compat
.writeable(superblock
.compat_features
)) {
2518 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2519 derr
<< "it is still writeable, though. Missing features: " << diff
<< dendl
;
2524 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2525 derr
<< "Cannot write to disk! Missing features: " << diff
<< dendl
;
2531 assert_warn(whoami
== superblock
.whoami
);
2532 if (whoami
!= superblock
.whoami
) {
2533 derr
<< "OSD::init: superblock says osd"
2534 << superblock
.whoami
<< " but I am osd." << whoami
<< dendl
;
2539 initial
= get_osd_initial_compat_set();
2540 diff
= superblock
.compat_features
.unsupported(initial
);
2541 if (superblock
.compat_features
.merge(initial
)) {
2542 // We need to persist the new compat_set before we
2544 dout(5) << "Upgrading superblock adding: " << diff
<< dendl
;
2545 ObjectStore::Transaction t
;
2546 write_superblock(t
);
2547 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2552 // make sure snap mapper object exists
2553 if (!store
->exists(coll_t::meta(), OSD::make_snapmapper_oid())) {
2554 dout(10) << "init creating/touching snapmapper object" << dendl
;
2555 ObjectStore::Transaction t
;
2556 t
.touch(coll_t::meta(), OSD::make_snapmapper_oid());
2557 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2562 class_handler
= new ClassHandler(cct
);
2563 cls_initialize(class_handler
);
2565 if (cct
->_conf
->osd_open_classes_on_start
) {
2566 int r
= class_handler
->open_all_classes();
2568 dout(1) << "warning: got an error loading one or more classes: " << cpp_strerror(r
) << dendl
;
2571 // load up "current" osdmap
2572 assert_warn(!osdmap
);
2574 derr
<< "OSD::init: unable to read current osdmap" << dendl
;
2578 osdmap
= get_map(superblock
.current_epoch
);
2579 check_osdmap_features(store
);
2581 create_recoverystate_perf();
2584 epoch_t bind_epoch
= osdmap
->get_epoch();
2585 service
.set_epochs(NULL
, NULL
, &bind_epoch
);
2588 clear_temp_objects();
2590 // initialize osdmap references in sharded wq
2591 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
2593 // load up pgs (as they previously existed)
2596 dout(2) << "superblock: I am osd." << superblock
.whoami
<< dendl
;
2597 dout(0) << "using " << op_queue
<< " op queue with priority op cut off at " <<
2598 op_prio_cutoff
<< "." << dendl
;
2603 client_messenger
->add_dispatcher_head(this);
2604 cluster_messenger
->add_dispatcher_head(this);
2606 hb_front_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2607 hb_back_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2608 hb_front_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2609 hb_back_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2611 objecter_messenger
->add_dispatcher_head(service
.objecter
);
2613 monc
->set_want_keys(CEPH_ENTITY_TYPE_MON
| CEPH_ENTITY_TYPE_OSD
2614 | CEPH_ENTITY_TYPE_MGR
);
2620 * FIXME: this is a placeholder implementation that unconditionally
2621 * sends every is_primary PG's stats every time we're called, unlike
2622 * the existing mon PGStats mechanism that uses pg_stat_queue and acks.
2623 * This has equivalent cost to the existing worst case where all
2624 * PGs are busy and their stats are always enqueued for sending.
2626 mgrc
.set_pgstats_cb([this](){
2627 RWLock::RLocker
l(map_lock
);
2629 utime_t had_for
= ceph_clock_now() - had_map_since
;
2630 osd_stat_t cur_stat
= service
.get_osd_stat();
2631 cur_stat
.os_perf_stat
= store
->get_cur_stats();
2633 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
2634 m
->osd_stat
= cur_stat
;
2636 Mutex::Locker lec
{min_last_epoch_clean_lock
};
2637 min_last_epoch_clean
= osdmap
->get_epoch();
2638 min_last_epoch_clean_pgs
.clear();
2639 RWLock::RLocker
lpg(pg_map_lock
);
2640 for (const auto &i
: pg_map
) {
2642 if (!pg
->is_primary()) {
2646 pg
->pg_stats_publish_lock
.Lock();
2647 if (pg
->pg_stats_publish_valid
) {
2648 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
2649 const auto lec
= pg
->pg_stats_publish
.get_effective_last_epoch_clean();
2650 min_last_epoch_clean
= min(min_last_epoch_clean
, lec
);
2651 min_last_epoch_clean_pgs
.push_back(pg
->info
.pgid
.pgid
);
2653 pg
->pg_stats_publish_lock
.Unlock();
2660 client_messenger
->add_dispatcher_head(&mgrc
);
2662 // tell monc about log_client so it will know about mon session resets
2663 monc
->set_log_client(&log_client
);
2664 update_log_config();
2669 service
.publish_map(osdmap
);
2670 service
.publish_superblock(superblock
);
2671 service
.max_oldest_map
= superblock
.oldest_map
;
2677 set_disk_tp_priority();
2679 // start the heartbeat
2680 heartbeat_thread
.create("osd_srv_heartbt");
2683 tick_timer
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick(this));
2685 Mutex::Locker
l(tick_timer_lock
);
2686 tick_timer_without_osd_lock
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick_WithoutOSDLock(this));
2691 r
= monc
->authenticate();
2693 derr
<< __func__
<< " authentication failed: " << cpp_strerror(r
)
2695 osd_lock
.Lock(); // locker is going to unlock this on function exit
2701 while (monc
->wait_auth_rotating(30.0) < 0) {
2702 derr
<< "unable to obtain rotating service keys; retrying" << dendl
;
2703 ++rotating_auth_attempts
;
2704 if (rotating_auth_attempts
> g_conf
->max_rotating_auth_attempts
) {
2705 derr
<< __func__
<< " wait_auth_rotating timed out" << dendl
;
2706 osd_lock
.Lock(); // make locker happy
2707 if (!is_stopping()) {
2714 r
= update_crush_device_class();
2716 derr
<< __func__
<< " unable to update_crush_device_class: "
2717 << cpp_strerror(r
) << dendl
;
2722 r
= update_crush_location();
2724 derr
<< __func__
<< " unable to update_crush_location: "
2725 << cpp_strerror(r
) << dendl
;
2734 // start objecter *after* we have authenticated, so that we don't ignore
2735 // the OSDMaps it requests.
2736 service
.final_init();
2740 dout(10) << "ensuring pgs have consumed prior maps" << dendl
;
2744 dout(0) << "done with init, starting boot process" << dendl
;
2746 // subscribe to any pg creations
2747 monc
->sub_want("osd_pg_creates", last_pg_create_epoch
, 0);
2749 // MgrClient needs this (it doesn't have MonClient reference itself)
2750 monc
->sub_want("mgrmap", 0, 0);
2752 // we don't need to ask for an osdmap here; objecter will
2753 //monc->sub_want("osdmap", osdmap->get_epoch(), CEPH_SUBSCRIBE_ONETIME);
2764 enable_disable_fuse(true);
2771 void OSD::final_init()
2773 AdminSocket
*admin_socket
= cct
->get_admin_socket();
2774 asok_hook
= new OSDSocketHook(this);
2775 int r
= admin_socket
->register_command("status", "status", asok_hook
,
2776 "high-level status of OSD");
2778 r
= admin_socket
->register_command("flush_journal", "flush_journal",
2780 "flush the journal to permanent store");
2782 r
= admin_socket
->register_command("dump_ops_in_flight",
2783 "dump_ops_in_flight " \
2784 "name=filterstr,type=CephString,n=N,req=false",
2786 "show the ops currently in flight");
2788 r
= admin_socket
->register_command("ops",
2790 "name=filterstr,type=CephString,n=N,req=false",
2792 "show the ops currently in flight");
2794 r
= admin_socket
->register_command("dump_blocked_ops",
2795 "dump_blocked_ops " \
2796 "name=filterstr,type=CephString,n=N,req=false",
2798 "show the blocked ops currently in flight");
2800 r
= admin_socket
->register_command("dump_historic_ops",
2801 "dump_historic_ops " \
2802 "name=filterstr,type=CephString,n=N,req=false",
2806 r
= admin_socket
->register_command("dump_historic_slow_ops",
2807 "dump_historic_slow_ops " \
2808 "name=filterstr,type=CephString,n=N,req=false",
2810 "show slowest recent ops");
2812 r
= admin_socket
->register_command("dump_historic_ops_by_duration",
2813 "dump_historic_ops_by_duration " \
2814 "name=filterstr,type=CephString,n=N,req=false",
2816 "show slowest recent ops, sorted by duration");
2818 r
= admin_socket
->register_command("dump_op_pq_state", "dump_op_pq_state",
2820 "dump op priority queue state");
2822 r
= admin_socket
->register_command("dump_blacklist", "dump_blacklist",
2824 "dump blacklisted clients and times");
2826 r
= admin_socket
->register_command("dump_watchers", "dump_watchers",
2828 "show clients which have active watches,"
2829 " and on which objects");
2831 r
= admin_socket
->register_command("dump_reservations", "dump_reservations",
2833 "show recovery reservations");
2835 r
= admin_socket
->register_command("get_latest_osdmap", "get_latest_osdmap",
2837 "force osd to update the latest map from "
2841 r
= admin_socket
->register_command( "heap",
2843 "name=heapcmd,type=CephString",
2845 "show heap usage info (available only if "
2846 "compiled with tcmalloc)");
2849 r
= admin_socket
->register_command("set_heap_property",
2850 "set_heap_property " \
2851 "name=property,type=CephString " \
2852 "name=value,type=CephInt",
2854 "update malloc extension heap property");
2857 r
= admin_socket
->register_command("get_heap_property",
2858 "get_heap_property " \
2859 "name=property,type=CephString",
2861 "get malloc extension heap property");
2864 r
= admin_socket
->register_command("dump_objectstore_kv_stats",
2865 "dump_objectstore_kv_stats",
2867 "print statistics of kvdb which used by bluestore");
2870 r
= admin_socket
->register_command("dump_scrubs",
2873 "print scheduled scrubs");
2876 r
= admin_socket
->register_command("calc_objectstore_db_histogram",
2877 "calc_objectstore_db_histogram",
2879 "Generate key value histogram of kvdb(rocksdb) which used by bluestore");
2882 r
= admin_socket
->register_command("flush_store_cache",
2883 "flush_store_cache",
2885 "Flush bluestore internal cache");
2887 r
= admin_socket
->register_command("dump_pgstate_history", "dump_pgstate_history",
2889 "show recent state history");
2892 r
= admin_socket
->register_command("compact", "compact",
2894 "Commpact object store's omap."
2895 " WARNING: Compaction probably slows your requests");
2898 test_ops_hook
= new TestOpsSocketHook(&(this->service
), this->store
);
2899 // Note: pools are CephString instead of CephPoolname because
2900 // these commands traditionally support both pool names and numbers
2901 r
= admin_socket
->register_command(
2904 "name=pool,type=CephString " \
2905 "name=objname,type=CephObjectname " \
2906 "name=key,type=CephString "\
2907 "name=val,type=CephString",
2911 r
= admin_socket
->register_command(
2914 "name=pool,type=CephString " \
2915 "name=objname,type=CephObjectname " \
2916 "name=key,type=CephString",
2920 r
= admin_socket
->register_command(
2923 "name=pool,type=CephString " \
2924 "name=objname,type=CephObjectname " \
2925 "name=header,type=CephString",
2930 r
= admin_socket
->register_command(
2933 "name=pool,type=CephString " \
2934 "name=objname,type=CephObjectname",
2936 "output entire object map");
2939 r
= admin_socket
->register_command(
2942 "name=pool,type=CephString " \
2943 "name=objname,type=CephObjectname " \
2944 "name=len,type=CephInt",
2946 "truncate object to length");
2949 r
= admin_socket
->register_command(
2952 "name=pool,type=CephString " \
2953 "name=objname,type=CephObjectname " \
2954 "name=shardid,type=CephInt,req=false,range=0|255",
2956 "inject data error to an object");
2959 r
= admin_socket
->register_command(
2962 "name=pool,type=CephString " \
2963 "name=objname,type=CephObjectname " \
2964 "name=shardid,type=CephInt,req=false,range=0|255",
2966 "inject metadata error to an object");
2968 r
= admin_socket
->register_command(
2969 "set_recovery_delay",
2970 "set_recovery_delay " \
2971 "name=utime,type=CephInt,req=false",
2973 "Delay osd recovery by specified seconds");
2975 r
= admin_socket
->register_command(
2978 "name=pgid,type=CephString ",
2980 "Trigger a scheduled scrub ");
2982 r
= admin_socket
->register_command(
2985 "name=type,type=CephString,req=false " \
2986 "name=count,type=CephInt,req=false ",
2988 "Inject a full disk (optional count times)");
2992 void OSD::create_logger()
2994 dout(10) << "create_logger" << dendl
;
2996 PerfCountersBuilder
osd_plb(cct
, "osd", l_osd_first
, l_osd_last
);
2998 // Latency axis configuration for op histograms, values are in nanoseconds
2999 PerfHistogramCommon::axis_config_d op_hist_x_axis_config
{
3001 PerfHistogramCommon::SCALE_LOG2
, ///< Latency in logarithmic scale
3003 100000, ///< Quantization unit is 100usec
3004 32, ///< Enough to cover much longer than slow requests
3007 // Op size axis configuration for op histograms, values are in bytes
3008 PerfHistogramCommon::axis_config_d op_hist_y_axis_config
{
3009 "Request size (bytes)",
3010 PerfHistogramCommon::SCALE_LOG2
, ///< Request size in logarithmic scale
3012 512, ///< Quantization unit is 512 bytes
3013 32, ///< Enough to cover requests larger than GB
3017 // All the basic OSD operation stats are to be considered useful
3018 osd_plb
.set_prio_default(PerfCountersBuilder::PRIO_USEFUL
);
3021 l_osd_op_wip
, "op_wip",
3022 "Replication operations currently being processed (primary)");
3023 osd_plb
.add_u64_counter(
3025 "Client operations",
3026 "ops", PerfCountersBuilder::PRIO_CRITICAL
);
3027 osd_plb
.add_u64_counter(
3028 l_osd_op_inb
, "op_in_bytes",
3029 "Client operations total write size",
3030 "wr", PerfCountersBuilder::PRIO_INTERESTING
);
3031 osd_plb
.add_u64_counter(
3032 l_osd_op_outb
, "op_out_bytes",
3033 "Client operations total read size",
3034 "rd", PerfCountersBuilder::PRIO_INTERESTING
);
3035 osd_plb
.add_time_avg(
3036 l_osd_op_lat
, "op_latency",
3037 "Latency of client operations (including queue time)",
3039 osd_plb
.add_time_avg(
3040 l_osd_op_process_lat
, "op_process_latency",
3041 "Latency of client operations (excluding queue time)");
3042 osd_plb
.add_time_avg(
3043 l_osd_op_prepare_lat
, "op_prepare_latency",
3044 "Latency of client operations (excluding queue time and wait for finished)");
3046 osd_plb
.add_u64_counter(
3047 l_osd_op_r
, "op_r", "Client read operations");
3048 osd_plb
.add_u64_counter(
3049 l_osd_op_r_outb
, "op_r_out_bytes", "Client data read");
3050 osd_plb
.add_time_avg(
3051 l_osd_op_r_lat
, "op_r_latency",
3052 "Latency of read operation (including queue time)");
3053 osd_plb
.add_u64_counter_histogram(
3054 l_osd_op_r_lat_outb_hist
, "op_r_latency_out_bytes_histogram",
3055 op_hist_x_axis_config
, op_hist_y_axis_config
,
3056 "Histogram of operation latency (including queue time) + data read");
3057 osd_plb
.add_time_avg(
3058 l_osd_op_r_process_lat
, "op_r_process_latency",
3059 "Latency of read operation (excluding queue time)");
3060 osd_plb
.add_time_avg(
3061 l_osd_op_r_prepare_lat
, "op_r_prepare_latency",
3062 "Latency of read operations (excluding queue time and wait for finished)");
3063 osd_plb
.add_u64_counter(
3064 l_osd_op_w
, "op_w", "Client write operations");
3065 osd_plb
.add_u64_counter(
3066 l_osd_op_w_inb
, "op_w_in_bytes", "Client data written");
3067 osd_plb
.add_time_avg(
3068 l_osd_op_w_lat
, "op_w_latency",
3069 "Latency of write operation (including queue time)");
3070 osd_plb
.add_u64_counter_histogram(
3071 l_osd_op_w_lat_inb_hist
, "op_w_latency_in_bytes_histogram",
3072 op_hist_x_axis_config
, op_hist_y_axis_config
,
3073 "Histogram of operation latency (including queue time) + data written");
3074 osd_plb
.add_time_avg(
3075 l_osd_op_w_process_lat
, "op_w_process_latency",
3076 "Latency of write operation (excluding queue time)");
3077 osd_plb
.add_time_avg(
3078 l_osd_op_w_prepare_lat
, "op_w_prepare_latency",
3079 "Latency of write operations (excluding queue time and wait for finished)");
3080 osd_plb
.add_u64_counter(
3081 l_osd_op_rw
, "op_rw",
3082 "Client read-modify-write operations");
3083 osd_plb
.add_u64_counter(
3084 l_osd_op_rw_inb
, "op_rw_in_bytes",
3085 "Client read-modify-write operations write in");
3086 osd_plb
.add_u64_counter(
3087 l_osd_op_rw_outb
,"op_rw_out_bytes",
3088 "Client read-modify-write operations read out ");
3089 osd_plb
.add_time_avg(
3090 l_osd_op_rw_lat
, "op_rw_latency",
3091 "Latency of read-modify-write operation (including queue time)");
3092 osd_plb
.add_u64_counter_histogram(
3093 l_osd_op_rw_lat_inb_hist
, "op_rw_latency_in_bytes_histogram",
3094 op_hist_x_axis_config
, op_hist_y_axis_config
,
3095 "Histogram of rw operation latency (including queue time) + data written");
3096 osd_plb
.add_u64_counter_histogram(
3097 l_osd_op_rw_lat_outb_hist
, "op_rw_latency_out_bytes_histogram",
3098 op_hist_x_axis_config
, op_hist_y_axis_config
,
3099 "Histogram of rw operation latency (including queue time) + data read");
3100 osd_plb
.add_time_avg(
3101 l_osd_op_rw_process_lat
, "op_rw_process_latency",
3102 "Latency of read-modify-write operation (excluding queue time)");
3103 osd_plb
.add_time_avg(
3104 l_osd_op_rw_prepare_lat
, "op_rw_prepare_latency",
3105 "Latency of read-modify-write operations (excluding queue time and wait for finished)");
3107 // Now we move on to some more obscure stats, revert to assuming things
3108 // are low priority unless otherwise specified.
3109 osd_plb
.set_prio_default(PerfCountersBuilder::PRIO_DEBUGONLY
);
3111 osd_plb
.add_time_avg(l_osd_op_before_queue_op_lat
, "op_before_queue_op_lat",
3112 "Latency of IO before calling queue(before really queue into ShardedOpWq)"); // client io before queue op_wq latency
3113 osd_plb
.add_time_avg(l_osd_op_before_dequeue_op_lat
, "op_before_dequeue_op_lat",
3114 "Latency of IO before calling dequeue_op(already dequeued and get PG lock)"); // client io before dequeue_op latency
3116 osd_plb
.add_u64_counter(
3117 l_osd_sop
, "subop", "Suboperations");
3118 osd_plb
.add_u64_counter(
3119 l_osd_sop_inb
, "subop_in_bytes", "Suboperations total size");
3120 osd_plb
.add_time_avg(l_osd_sop_lat
, "subop_latency", "Suboperations latency");
3122 osd_plb
.add_u64_counter(l_osd_sop_w
, "subop_w", "Replicated writes");
3123 osd_plb
.add_u64_counter(
3124 l_osd_sop_w_inb
, "subop_w_in_bytes", "Replicated written data size");
3125 osd_plb
.add_time_avg(
3126 l_osd_sop_w_lat
, "subop_w_latency", "Replicated writes latency");
3127 osd_plb
.add_u64_counter(
3128 l_osd_sop_pull
, "subop_pull", "Suboperations pull requests");
3129 osd_plb
.add_time_avg(
3130 l_osd_sop_pull_lat
, "subop_pull_latency", "Suboperations pull latency");
3131 osd_plb
.add_u64_counter(
3132 l_osd_sop_push
, "subop_push", "Suboperations push messages");
3133 osd_plb
.add_u64_counter(
3134 l_osd_sop_push_inb
, "subop_push_in_bytes", "Suboperations pushed size");
3135 osd_plb
.add_time_avg(
3136 l_osd_sop_push_lat
, "subop_push_latency", "Suboperations push latency");
3138 osd_plb
.add_u64_counter(l_osd_pull
, "pull", "Pull requests sent");
3139 osd_plb
.add_u64_counter(l_osd_push
, "push", "Push messages sent");
3140 osd_plb
.add_u64_counter(l_osd_push_outb
, "push_out_bytes", "Pushed size");
3142 osd_plb
.add_u64_counter(
3143 l_osd_rop
, "recovery_ops",
3144 "Started recovery operations",
3145 "rop", PerfCountersBuilder::PRIO_INTERESTING
);
3147 osd_plb
.add_u64(l_osd_loadavg
, "loadavg", "CPU load");
3148 osd_plb
.add_u64(l_osd_buf
, "buffer_bytes", "Total allocated buffer size");
3149 osd_plb
.add_u64(l_osd_history_alloc_bytes
, "history_alloc_Mbytes");
3150 osd_plb
.add_u64(l_osd_history_alloc_num
, "history_alloc_num");
3152 l_osd_cached_crc
, "cached_crc", "Total number getting crc from crc_cache");
3154 l_osd_cached_crc_adjusted
, "cached_crc_adjusted",
3155 "Total number getting crc from crc_cache with adjusting");
3156 osd_plb
.add_u64(l_osd_missed_crc
, "missed_crc",
3157 "Total number of crc cache misses");
3159 osd_plb
.add_u64(l_osd_pg
, "numpg", "Placement groups",
3160 "pgs", PerfCountersBuilder::PRIO_USEFUL
);
3162 l_osd_pg_primary
, "numpg_primary",
3163 "Placement groups for which this osd is primary");
3165 l_osd_pg_replica
, "numpg_replica",
3166 "Placement groups for which this osd is replica");
3168 l_osd_pg_stray
, "numpg_stray",
3169 "Placement groups ready to be deleted from this osd");
3171 l_osd_pg_removing
, "numpg_removing",
3172 "Placement groups queued for local deletion", "pgsr",
3173 PerfCountersBuilder::PRIO_USEFUL
);
3175 l_osd_hb_to
, "heartbeat_to_peers", "Heartbeat (ping) peers we send to");
3176 osd_plb
.add_u64_counter(l_osd_map
, "map_messages", "OSD map messages");
3177 osd_plb
.add_u64_counter(l_osd_mape
, "map_message_epochs", "OSD map epochs");
3178 osd_plb
.add_u64_counter(
3179 l_osd_mape_dup
, "map_message_epoch_dups", "OSD map duplicates");
3180 osd_plb
.add_u64_counter(
3181 l_osd_waiting_for_map
, "messages_delayed_for_map",
3182 "Operations waiting for OSD map");
3184 osd_plb
.add_u64_counter(
3185 l_osd_map_cache_hit
, "osd_map_cache_hit", "osdmap cache hit");
3186 osd_plb
.add_u64_counter(
3187 l_osd_map_cache_miss
, "osd_map_cache_miss", "osdmap cache miss");
3188 osd_plb
.add_u64_counter(
3189 l_osd_map_cache_miss_low
, "osd_map_cache_miss_low",
3190 "osdmap cache miss below cache lower bound");
3191 osd_plb
.add_u64_avg(
3192 l_osd_map_cache_miss_low_avg
, "osd_map_cache_miss_low_avg",
3193 "osdmap cache miss, avg distance below cache lower bound");
3194 osd_plb
.add_u64_counter(
3195 l_osd_map_bl_cache_hit
, "osd_map_bl_cache_hit",
3196 "OSDMap buffer cache hits");
3197 osd_plb
.add_u64_counter(
3198 l_osd_map_bl_cache_miss
, "osd_map_bl_cache_miss",
3199 "OSDMap buffer cache misses");
3202 l_osd_stat_bytes
, "stat_bytes", "OSD size", "size",
3203 PerfCountersBuilder::PRIO_USEFUL
);
3205 l_osd_stat_bytes_used
, "stat_bytes_used", "Used space", "used",
3206 PerfCountersBuilder::PRIO_USEFUL
);
3207 osd_plb
.add_u64(l_osd_stat_bytes_avail
, "stat_bytes_avail", "Available space");
3209 osd_plb
.add_u64_counter(
3210 l_osd_copyfrom
, "copyfrom", "Rados \"copy-from\" operations");
3212 osd_plb
.add_u64_counter(l_osd_tier_promote
, "tier_promote", "Tier promotions");
3213 osd_plb
.add_u64_counter(l_osd_tier_flush
, "tier_flush", "Tier flushes");
3214 osd_plb
.add_u64_counter(
3215 l_osd_tier_flush_fail
, "tier_flush_fail", "Failed tier flushes");
3216 osd_plb
.add_u64_counter(
3217 l_osd_tier_try_flush
, "tier_try_flush", "Tier flush attempts");
3218 osd_plb
.add_u64_counter(
3219 l_osd_tier_try_flush_fail
, "tier_try_flush_fail",
3220 "Failed tier flush attempts");
3221 osd_plb
.add_u64_counter(
3222 l_osd_tier_evict
, "tier_evict", "Tier evictions");
3223 osd_plb
.add_u64_counter(
3224 l_osd_tier_whiteout
, "tier_whiteout", "Tier whiteouts");
3225 osd_plb
.add_u64_counter(
3226 l_osd_tier_dirty
, "tier_dirty", "Dirty tier flag set");
3227 osd_plb
.add_u64_counter(
3228 l_osd_tier_clean
, "tier_clean", "Dirty tier flag cleaned");
3229 osd_plb
.add_u64_counter(
3230 l_osd_tier_delay
, "tier_delay", "Tier delays (agent waiting)");
3231 osd_plb
.add_u64_counter(
3232 l_osd_tier_proxy_read
, "tier_proxy_read", "Tier proxy reads");
3233 osd_plb
.add_u64_counter(
3234 l_osd_tier_proxy_write
, "tier_proxy_write", "Tier proxy writes");
3236 osd_plb
.add_u64_counter(
3237 l_osd_agent_wake
, "agent_wake", "Tiering agent wake up");
3238 osd_plb
.add_u64_counter(
3239 l_osd_agent_skip
, "agent_skip", "Objects skipped by agent");
3240 osd_plb
.add_u64_counter(
3241 l_osd_agent_flush
, "agent_flush", "Tiering agent flushes");
3242 osd_plb
.add_u64_counter(
3243 l_osd_agent_evict
, "agent_evict", "Tiering agent evictions");
3245 osd_plb
.add_u64_counter(
3246 l_osd_object_ctx_cache_hit
, "object_ctx_cache_hit", "Object context cache hits");
3247 osd_plb
.add_u64_counter(
3248 l_osd_object_ctx_cache_total
, "object_ctx_cache_total", "Object context cache lookups");
3250 osd_plb
.add_u64_counter(l_osd_op_cache_hit
, "op_cache_hit");
3251 osd_plb
.add_time_avg(
3252 l_osd_tier_flush_lat
, "osd_tier_flush_lat", "Object flush latency");
3253 osd_plb
.add_time_avg(
3254 l_osd_tier_promote_lat
, "osd_tier_promote_lat", "Object promote latency");
3255 osd_plb
.add_time_avg(
3256 l_osd_tier_r_lat
, "osd_tier_r_lat", "Object proxy read latency");
3258 osd_plb
.add_u64_counter(
3259 l_osd_pg_info
, "osd_pg_info", "PG updated its info (using any method)");
3260 osd_plb
.add_u64_counter(
3261 l_osd_pg_fastinfo
, "osd_pg_fastinfo",
3262 "PG updated its info using fastinfo attr");
3263 osd_plb
.add_u64_counter(
3264 l_osd_pg_biginfo
, "osd_pg_biginfo", "PG updated its biginfo attr");
3266 logger
= osd_plb
.create_perf_counters();
3267 cct
->get_perfcounters_collection()->add(logger
);
3270 void OSD::create_recoverystate_perf()
3272 dout(10) << "create_recoverystate_perf" << dendl
;
3274 PerfCountersBuilder
rs_perf(cct
, "recoverystate_perf", rs_first
, rs_last
);
3276 rs_perf
.add_time_avg(rs_initial_latency
, "initial_latency", "Initial recovery state latency");
3277 rs_perf
.add_time_avg(rs_started_latency
, "started_latency", "Started recovery state latency");
3278 rs_perf
.add_time_avg(rs_reset_latency
, "reset_latency", "Reset recovery state latency");
3279 rs_perf
.add_time_avg(rs_start_latency
, "start_latency", "Start recovery state latency");
3280 rs_perf
.add_time_avg(rs_primary_latency
, "primary_latency", "Primary recovery state latency");
3281 rs_perf
.add_time_avg(rs_peering_latency
, "peering_latency", "Peering recovery state latency");
3282 rs_perf
.add_time_avg(rs_backfilling_latency
, "backfilling_latency", "Backfilling recovery state latency");
3283 rs_perf
.add_time_avg(rs_waitremotebackfillreserved_latency
, "waitremotebackfillreserved_latency", "Wait remote backfill reserved recovery state latency");
3284 rs_perf
.add_time_avg(rs_waitlocalbackfillreserved_latency
, "waitlocalbackfillreserved_latency", "Wait local backfill reserved recovery state latency");
3285 rs_perf
.add_time_avg(rs_notbackfilling_latency
, "notbackfilling_latency", "Notbackfilling recovery state latency");
3286 rs_perf
.add_time_avg(rs_repnotrecovering_latency
, "repnotrecovering_latency", "Repnotrecovering recovery state latency");
3287 rs_perf
.add_time_avg(rs_repwaitrecoveryreserved_latency
, "repwaitrecoveryreserved_latency", "Rep wait recovery reserved recovery state latency");
3288 rs_perf
.add_time_avg(rs_repwaitbackfillreserved_latency
, "repwaitbackfillreserved_latency", "Rep wait backfill reserved recovery state latency");
3289 rs_perf
.add_time_avg(rs_reprecovering_latency
, "reprecovering_latency", "RepRecovering recovery state latency");
3290 rs_perf
.add_time_avg(rs_activating_latency
, "activating_latency", "Activating recovery state latency");
3291 rs_perf
.add_time_avg(rs_waitlocalrecoveryreserved_latency
, "waitlocalrecoveryreserved_latency", "Wait local recovery reserved recovery state latency");
3292 rs_perf
.add_time_avg(rs_waitremoterecoveryreserved_latency
, "waitremoterecoveryreserved_latency", "Wait remote recovery reserved recovery state latency");
3293 rs_perf
.add_time_avg(rs_recovering_latency
, "recovering_latency", "Recovering recovery state latency");
3294 rs_perf
.add_time_avg(rs_recovered_latency
, "recovered_latency", "Recovered recovery state latency");
3295 rs_perf
.add_time_avg(rs_clean_latency
, "clean_latency", "Clean recovery state latency");
3296 rs_perf
.add_time_avg(rs_active_latency
, "active_latency", "Active recovery state latency");
3297 rs_perf
.add_time_avg(rs_replicaactive_latency
, "replicaactive_latency", "Replicaactive recovery state latency");
3298 rs_perf
.add_time_avg(rs_stray_latency
, "stray_latency", "Stray recovery state latency");
3299 rs_perf
.add_time_avg(rs_getinfo_latency
, "getinfo_latency", "Getinfo recovery state latency");
3300 rs_perf
.add_time_avg(rs_getlog_latency
, "getlog_latency", "Getlog recovery state latency");
3301 rs_perf
.add_time_avg(rs_waitactingchange_latency
, "waitactingchange_latency", "Waitactingchange recovery state latency");
3302 rs_perf
.add_time_avg(rs_incomplete_latency
, "incomplete_latency", "Incomplete recovery state latency");
3303 rs_perf
.add_time_avg(rs_down_latency
, "down_latency", "Down recovery state latency");
3304 rs_perf
.add_time_avg(rs_getmissing_latency
, "getmissing_latency", "Getmissing recovery state latency");
3305 rs_perf
.add_time_avg(rs_waitupthru_latency
, "waitupthru_latency", "Waitupthru recovery state latency");
3306 rs_perf
.add_time_avg(rs_notrecovering_latency
, "notrecovering_latency", "Notrecovering recovery state latency");
3308 recoverystate_perf
= rs_perf
.create_perf_counters();
3309 cct
->get_perfcounters_collection()->add(recoverystate_perf
);
3314 if (!service
.prepare_to_stop())
3315 return 0; // already shutting down
3317 if (is_stopping()) {
3321 derr
<< "shutdown" << dendl
;
3323 set_state(STATE_STOPPING
);
3326 if (cct
->_conf
->get_val
<bool>("osd_debug_shutdown")) {
3327 cct
->_conf
->set_val("debug_osd", "100");
3328 cct
->_conf
->set_val("debug_journal", "100");
3329 cct
->_conf
->set_val("debug_filestore", "100");
3330 cct
->_conf
->set_val("debug_bluestore", "100");
3331 cct
->_conf
->set_val("debug_ms", "100");
3332 cct
->_conf
->apply_changes(NULL
);
3335 // stop MgrClient earlier as it's more like an internal consumer of OSD
3338 service
.start_shutdown();
3340 // stop sending work to pgs. this just prevents any new work in _process
3341 // from racing with on_shutdown and potentially entering the pg after.
3342 op_shardedwq
.drain();
3346 RWLock::RLocker
l(pg_map_lock
);
3347 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3350 dout(20) << " kicking pg " << p
->first
<< dendl
;
3352 p
->second
->on_shutdown();
3353 p
->second
->unlock();
3354 p
->second
->osr
->flush();
3357 clear_pg_stat_queue();
3359 // drain op queue again (in case PGs requeued something)
3360 op_shardedwq
.drain();
3362 finished
.clear(); // zap waiters (bleh, this is messy)
3365 op_shardedwq
.clear_pg_slots();
3367 // unregister commands
3368 cct
->get_admin_socket()->unregister_command("status");
3369 cct
->get_admin_socket()->unregister_command("flush_journal");
3370 cct
->get_admin_socket()->unregister_command("dump_ops_in_flight");
3371 cct
->get_admin_socket()->unregister_command("ops");
3372 cct
->get_admin_socket()->unregister_command("dump_blocked_ops");
3373 cct
->get_admin_socket()->unregister_command("dump_historic_ops");
3374 cct
->get_admin_socket()->unregister_command("dump_historic_ops_by_duration");
3375 cct
->get_admin_socket()->unregister_command("dump_historic_slow_ops");
3376 cct
->get_admin_socket()->unregister_command("dump_op_pq_state");
3377 cct
->get_admin_socket()->unregister_command("dump_blacklist");
3378 cct
->get_admin_socket()->unregister_command("dump_watchers");
3379 cct
->get_admin_socket()->unregister_command("dump_reservations");
3380 cct
->get_admin_socket()->unregister_command("get_latest_osdmap");
3381 cct
->get_admin_socket()->unregister_command("heap");
3382 cct
->get_admin_socket()->unregister_command("set_heap_property");
3383 cct
->get_admin_socket()->unregister_command("get_heap_property");
3384 cct
->get_admin_socket()->unregister_command("dump_objectstore_kv_stats");
3385 cct
->get_admin_socket()->unregister_command("dump_scrubs");
3386 cct
->get_admin_socket()->unregister_command("calc_objectstore_db_histogram");
3387 cct
->get_admin_socket()->unregister_command("flush_store_cache");
3388 cct
->get_admin_socket()->unregister_command("dump_pgstate_history");
3389 cct
->get_admin_socket()->unregister_command("compact");
3393 cct
->get_admin_socket()->unregister_command("setomapval");
3394 cct
->get_admin_socket()->unregister_command("rmomapkey");
3395 cct
->get_admin_socket()->unregister_command("setomapheader");
3396 cct
->get_admin_socket()->unregister_command("getomap");
3397 cct
->get_admin_socket()->unregister_command("truncobj");
3398 cct
->get_admin_socket()->unregister_command("injectdataerr");
3399 cct
->get_admin_socket()->unregister_command("injectmdataerr");
3400 cct
->get_admin_socket()->unregister_command("set_recovery_delay");
3401 cct
->get_admin_socket()->unregister_command("trigger_scrub");
3402 cct
->get_admin_socket()->unregister_command("injectfull");
3403 delete test_ops_hook
;
3404 test_ops_hook
= NULL
;
3408 heartbeat_lock
.Lock();
3409 heartbeat_stop
= true;
3410 heartbeat_cond
.Signal();
3411 heartbeat_lock
.Unlock();
3412 heartbeat_thread
.join();
3417 dout(10) << "osd tp stopped" << dendl
;
3421 dout(10) << "op sharded tp stopped" << dendl
;
3425 dout(10) << "command tp stopped" << dendl
;
3429 dout(10) << "disk tp paused (new)" << dendl
;
3431 dout(10) << "stopping agent" << dendl
;
3432 service
.agent_stop();
3436 reset_heartbeat_peers();
3438 tick_timer
.shutdown();
3441 Mutex::Locker
l(tick_timer_lock
);
3442 tick_timer_without_osd_lock
.shutdown();
3445 // note unmount epoch
3446 dout(10) << "noting clean unmount in epoch " << osdmap
->get_epoch() << dendl
;
3447 superblock
.mounted
= service
.get_boot_epoch();
3448 superblock
.clean_thru
= osdmap
->get_epoch();
3449 ObjectStore::Transaction t
;
3450 write_superblock(t
);
3451 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3453 derr
<< "OSD::shutdown: error writing superblock: "
3454 << cpp_strerror(r
) << dendl
;
3459 Mutex::Locker
l(pg_stat_queue_lock
);
3460 assert(pg_stat_queue
.empty());
3463 service
.shutdown_reserver();
3466 #ifdef PG_DEBUG_REFS
3467 service
.dump_live_pgids();
3470 RWLock::RLocker
l(pg_map_lock
);
3471 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3474 dout(20) << " kicking pg " << p
->first
<< dendl
;
3476 if (p
->second
->ref
!= 1) {
3477 derr
<< "pgid " << p
->first
<< " has ref count of "
3478 << p
->second
->ref
<< dendl
;
3479 #ifdef PG_DEBUG_REFS
3480 p
->second
->dump_live_ids();
3482 if (cct
->_conf
->osd_shutdown_pgref_assert
) {
3486 p
->second
->unlock();
3487 p
->second
->put("PGMap");
3491 #ifdef PG_DEBUG_REFS
3492 service
.dump_live_pgids();
3494 cct
->_conf
->remove_observer(this);
3496 dout(10) << "syncing store" << dendl
;
3497 enable_disable_fuse(true);
3499 if (cct
->_conf
->osd_journal_flush_on_shutdown
) {
3500 dout(10) << "flushing journal" << dendl
;
3501 store
->flush_journal();
3507 dout(10) << "Store synced" << dendl
;
3512 osdmap
= OSDMapRef();
3514 op_tracker
.on_shutdown();
3516 class_handler
->shutdown();
3517 client_messenger
->shutdown();
3518 cluster_messenger
->shutdown();
3519 hb_front_client_messenger
->shutdown();
3520 hb_back_client_messenger
->shutdown();
3521 objecter_messenger
->shutdown();
3522 hb_front_server_messenger
->shutdown();
3523 hb_back_server_messenger
->shutdown();
3530 int OSD::mon_cmd_maybe_osd_create(string
&cmd
)
3532 bool created
= false;
3534 dout(10) << __func__
<< " cmd: " << cmd
<< dendl
;
3535 vector
<string
> vcmd
{cmd
};
3539 monc
->start_mon_command(vcmd
, inbl
, NULL
, &outs
, &w
);
3542 if (r
== -ENOENT
&& !created
) {
3543 string newcmd
= "{\"prefix\": \"osd create\", \"id\": " + stringify(whoami
)
3544 + ", \"uuid\": \"" + stringify(superblock
.osd_fsid
) + "\"}";
3545 vector
<string
> vnewcmd
{newcmd
};
3549 monc
->start_mon_command(vnewcmd
, inbl
, NULL
, &outs
, &w
);
3552 derr
<< __func__
<< " fail: osd does not exist and created failed: "
3553 << cpp_strerror(r
) << dendl
;
3559 derr
<< __func__
<< " fail: '" << outs
<< "': " << cpp_strerror(r
) << dendl
;
3568 int OSD::update_crush_location()
3570 if (!cct
->_conf
->osd_crush_update_on_start
) {
3571 dout(10) << __func__
<< " osd_crush_update_on_start = false" << dendl
;
3576 if (cct
->_conf
->osd_crush_initial_weight
>= 0) {
3577 snprintf(weight
, sizeof(weight
), "%.4lf", cct
->_conf
->osd_crush_initial_weight
);
3579 struct store_statfs_t st
;
3580 int r
= store
->statfs(&st
);
3582 derr
<< "statfs: " << cpp_strerror(r
) << dendl
;
3585 snprintf(weight
, sizeof(weight
), "%.4lf",
3587 (double)(st
.total
) /
3588 (double)(1ull << 40 /* TB */)));
3591 std::multimap
<string
,string
> loc
= cct
->crush_location
.get_location();
3592 dout(10) << __func__
<< " crush location is " << loc
<< dendl
;
3595 string("{\"prefix\": \"osd crush create-or-move\", ") +
3596 string("\"id\": ") + stringify(whoami
) + string(", ") +
3597 string("\"weight\":") + weight
+ string(", ") +
3598 string("\"args\": [");
3599 for (multimap
<string
,string
>::iterator p
= loc
.begin(); p
!= loc
.end(); ++p
) {
3600 if (p
!= loc
.begin())
3602 cmd
+= "\"" + p
->first
+ "=" + p
->second
+ "\"";
3606 return mon_cmd_maybe_osd_create(cmd
);
3609 int OSD::update_crush_device_class()
3611 if (!cct
->_conf
->osd_class_update_on_start
) {
3612 dout(10) << __func__
<< " osd_class_update_on_start = false" << dendl
;
3616 string device_class
;
3617 int r
= store
->read_meta("crush_device_class", &device_class
);
3618 if (r
< 0 || device_class
.empty()) {
3619 device_class
= store
->get_default_device_class();
3622 if (device_class
.empty()) {
3623 dout(20) << __func__
<< " no device class stored locally" << dendl
;
3628 string("{\"prefix\": \"osd crush set-device-class\", ") +
3629 string("\"class\": \"") + device_class
+ string("\", ") +
3630 string("\"ids\": [\"") + stringify(whoami
) + string("\"]}");
3632 r
= mon_cmd_maybe_osd_create(cmd
);
3633 // the above cmd can fail for various reasons, e.g.:
3634 // (1) we are connecting to a pre-luminous monitor
3635 // (2) user manually specify a class other than
3636 // 'ceph-disk prepare --crush-device-class'
3637 // simply skip result-checking for now
3641 void OSD::write_superblock(ObjectStore::Transaction
& t
)
3643 dout(10) << "write_superblock " << superblock
<< dendl
;
3645 //hack: at minimum it's using the baseline feature set
3646 if (!superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_BASE
))
3647 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE
);
3650 ::encode(superblock
, bl
);
3651 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
3654 int OSD::read_superblock()
3657 int r
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, bl
);
3661 bufferlist::iterator p
= bl
.begin();
3662 ::decode(superblock
, p
);
3664 dout(10) << "read_superblock " << superblock
<< dendl
;
3669 void OSD::clear_temp_objects()
3671 dout(10) << __func__
<< dendl
;
3673 store
->list_collections(ls
);
3674 for (vector
<coll_t
>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
3676 if (!p
->is_pg(&pgid
))
3679 // list temp objects
3680 dout(20) << " clearing temps in " << *p
<< " pgid " << pgid
<< dendl
;
3682 vector
<ghobject_t
> temps
;
3685 vector
<ghobject_t
> objects
;
3686 store
->collection_list(*p
, next
, ghobject_t::get_max(),
3687 store
->get_ideal_list_max(),
3689 if (objects
.empty())
3691 vector
<ghobject_t
>::iterator q
;
3692 for (q
= objects
.begin(); q
!= objects
.end(); ++q
) {
3693 // Hammer set pool for temps to -1, so check for clean-up
3694 if (q
->hobj
.is_temp() || (q
->hobj
.pool
== -1)) {
3695 temps
.push_back(*q
);
3700 // If we saw a non-temp object and hit the break above we can
3701 // break out of the while loop too.
3702 if (q
!= objects
.end())
3705 if (!temps
.empty()) {
3706 ObjectStore::Transaction t
;
3708 for (vector
<ghobject_t
>::iterator q
= temps
.begin(); q
!= temps
.end(); ++q
) {
3709 dout(20) << " removing " << *p
<< " object " << *q
<< dendl
;
3711 if (++removed
> cct
->_conf
->osd_target_transaction_size
) {
3712 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3713 t
= ObjectStore::Transaction();
3718 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3724 void OSD::recursive_remove_collection(CephContext
* cct
,
3725 ObjectStore
*store
, spg_t pgid
,
3731 make_snapmapper_oid());
3733 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
3734 ObjectStore::Sequencer
>("rm"));
3735 ObjectStore::Transaction t
;
3736 SnapMapper
mapper(cct
, &driver
, 0, 0, 0, pgid
.shard
);
3738 vector
<ghobject_t
> objects
;
3739 store
->collection_list(tmp
, ghobject_t(), ghobject_t::get_max(),
3740 INT_MAX
, &objects
, 0);
3741 generic_dout(10) << __func__
<< " " << objects
<< dendl
;
3744 for (vector
<ghobject_t
>::iterator p
= objects
.begin();
3747 OSDriver::OSTransaction
_t(driver
.get_transaction(&t
));
3748 int r
= mapper
.remove_oid(p
->hobj
, &_t
);
3749 if (r
!= 0 && r
!= -ENOENT
)
3752 if (removed
> cct
->_conf
->osd_target_transaction_size
) {
3753 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3755 t
= ObjectStore::Transaction();
3759 t
.remove_collection(tmp
);
3760 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3764 if (!osr
->flush_commit(&waiter
)) {
3770 // ======================================================
3773 PGPool
OSD::_get_pool(int id
, OSDMapRef createmap
)
3775 if (!createmap
->have_pg_pool(id
)) {
3776 dout(5) << __func__
<< ": the OSDmap does not contain a PG pool with id = "
3781 PGPool p
= PGPool(cct
, createmap
, id
);
3783 dout(10) << "_get_pool " << p
.id
<< dendl
;
3787 PG
*OSD::_open_lock_pg(
3788 OSDMapRef createmap
,
3789 spg_t pgid
, bool no_lockdep_check
)
3791 assert(osd_lock
.is_locked());
3793 PG
* pg
= _make_pg(createmap
, pgid
);
3795 RWLock::WLocker
l(pg_map_lock
);
3796 pg
->lock(no_lockdep_check
);
3798 pg
->get("PGMap"); // because it's in pg_map
3799 service
.pg_add_epoch(pg
->info
.pgid
, createmap
->get_epoch());
3805 OSDMapRef createmap
,
3808 dout(10) << "_open_lock_pg " << pgid
<< dendl
;
3809 PGPool pool
= _get_pool(pgid
.pool(), createmap
);
3813 if (createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_REPLICATED
||
3814 createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_ERASURE
)
3815 pg
= new PrimaryLogPG(&service
, createmap
, pool
, pgid
);
3823 void OSD::add_newly_split_pg(PG
*pg
, PG::RecoveryCtx
*rctx
)
3825 epoch_t
e(service
.get_osdmap()->get_epoch());
3826 pg
->get("PGMap"); // For pg_map
3827 pg_map
[pg
->info
.pgid
] = pg
;
3828 service
.pg_add_epoch(pg
->info
.pgid
, pg
->get_osdmap()->get_epoch());
3830 dout(10) << "Adding newly split pg " << *pg
<< dendl
;
3831 pg
->handle_loaded(rctx
);
3832 pg
->write_if_dirty(*(rctx
->transaction
));
3833 pg
->queue_null(e
, e
);
3834 map
<spg_t
, list
<PG::CephPeeringEvtRef
> >::iterator to_wake
=
3835 peering_wait_for_split
.find(pg
->info
.pgid
);
3836 if (to_wake
!= peering_wait_for_split
.end()) {
3837 for (list
<PG::CephPeeringEvtRef
>::iterator i
=
3838 to_wake
->second
.begin();
3839 i
!= to_wake
->second
.end();
3841 pg
->queue_peering_event(*i
);
3843 peering_wait_for_split
.erase(to_wake
);
3845 if (!service
.get_osdmap()->have_pg_pool(pg
->info
.pgid
.pool()))
3849 OSD::res_result
OSD::_try_resurrect_pg(
3850 OSDMapRef curmap
, spg_t pgid
, spg_t
*resurrected
, PGRef
*old_pg_state
)
3852 assert(resurrected
);
3853 assert(old_pg_state
);
3854 // find nearest ancestor
3855 DeletingStateRef df
;
3858 df
= service
.deleting_pgs
.lookup(cur
);
3863 cur
= cur
.get_parent();
3866 return RES_NONE
; // good to go
3868 df
->old_pg_state
->lock();
3869 OSDMapRef create_map
= df
->old_pg_state
->get_osdmap();
3870 df
->old_pg_state
->unlock();
3872 set
<spg_t
> children
;
3874 if (df
->try_stop_deletion()) {
3875 dout(10) << __func__
<< ": halted deletion on pg " << pgid
<< dendl
;
3877 *old_pg_state
= df
->old_pg_state
;
3878 service
.deleting_pgs
.remove(pgid
); // PG is no longer being removed!
3881 // raced, ensure we don't see DeletingStateRef when we try to
3883 service
.deleting_pgs
.remove(pgid
);
3886 } else if (cur
.is_split(create_map
->get_pg_num(cur
.pool()),
3887 curmap
->get_pg_num(cur
.pool()),
3889 children
.count(pgid
)) {
3890 if (df
->try_stop_deletion()) {
3891 dout(10) << __func__
<< ": halted deletion on ancestor pg " << pgid
3894 *old_pg_state
= df
->old_pg_state
;
3895 service
.deleting_pgs
.remove(cur
); // PG is no longer being removed!
3898 /* this is not a problem, failing to cancel proves that all objects
3899 * have been removed, so no hobject_t overlap is possible
3907 PG
*OSD::_create_lock_pg(
3908 OSDMapRef createmap
,
3913 vector
<int>& up
, int up_primary
,
3914 vector
<int>& acting
, int acting_primary
,
3915 pg_history_t history
,
3916 const PastIntervals
& pi
,
3917 ObjectStore::Transaction
& t
)
3919 assert(osd_lock
.is_locked());
3920 dout(20) << "_create_lock_pg pgid " << pgid
<< dendl
;
3922 PG
*pg
= _open_lock_pg(createmap
, pgid
, true);
3924 service
.init_splits_between(pgid
, pg
->get_osdmap(), service
.get_osdmap());
3937 dout(7) << "_create_lock_pg " << *pg
<< dendl
;
3941 PG
*OSD::_lookup_lock_pg(spg_t pgid
)
3943 RWLock::RLocker
l(pg_map_lock
);
3945 auto pg_map_entry
= pg_map
.find(pgid
);
3946 if (pg_map_entry
== pg_map
.end())
3948 PG
*pg
= pg_map_entry
->second
;
3953 PG
*OSD::lookup_lock_pg(spg_t pgid
)
3955 return _lookup_lock_pg(pgid
);
3958 PG
*OSD::_lookup_lock_pg_with_map_lock_held(spg_t pgid
)
3960 assert(pg_map
.count(pgid
));
3961 PG
*pg
= pg_map
[pgid
];
3966 void OSD::load_pgs()
3968 assert(osd_lock
.is_locked());
3969 dout(0) << "load_pgs" << dendl
;
3971 RWLock::RLocker
l(pg_map_lock
);
3972 assert(pg_map
.empty());
3976 int r
= store
->list_collections(ls
);
3978 derr
<< "failed to list pgs: " << cpp_strerror(-r
) << dendl
;
3981 bool has_upgraded
= false;
3983 for (vector
<coll_t
>::iterator it
= ls
.begin();
3987 if (it
->is_temp(&pgid
) ||
3988 (it
->is_pg(&pgid
) && PG::_has_removal_flag(store
, pgid
))) {
3989 dout(10) << "load_pgs " << *it
<< " clearing temp" << dendl
;
3990 recursive_remove_collection(cct
, store
, pgid
, *it
);
3994 if (!it
->is_pg(&pgid
)) {
3995 dout(10) << "load_pgs ignoring unrecognized " << *it
<< dendl
;
3999 if (pgid
.preferred() >= 0) {
4000 dout(10) << __func__
<< ": skipping localized PG " << pgid
<< dendl
;
4001 // FIXME: delete it too, eventually
4005 dout(10) << "pgid " << pgid
<< " coll " << coll_t(pgid
) << dendl
;
4007 epoch_t map_epoch
= 0;
4008 int r
= PG::peek_map_epoch(store
, pgid
, &map_epoch
, &bl
);
4010 derr
<< __func__
<< " unable to peek at " << pgid
<< " metadata, skipping"
4016 if (map_epoch
> 0) {
4017 OSDMapRef pgosdmap
= service
.try_get_map(map_epoch
);
4019 if (!osdmap
->have_pg_pool(pgid
.pool())) {
4020 derr
<< __func__
<< ": could not find map for epoch " << map_epoch
4021 << " on pg " << pgid
<< ", but the pool is not present in the "
4022 << "current map, so this is probably a result of bug 10617. "
4023 << "Skipping the pg for now, you can use ceph-objectstore-tool "
4024 << "to clean it up later." << dendl
;
4027 derr
<< __func__
<< ": have pgid " << pgid
<< " at epoch "
4028 << map_epoch
<< ", but missing map. Crashing."
4030 assert(0 == "Missing map in load_pgs");
4033 pg
= _open_lock_pg(pgosdmap
, pgid
);
4035 pg
= _open_lock_pg(osdmap
, pgid
);
4037 // there can be no waiters here, so we don't call wake_pg_waiters
4039 pg
->ch
= store
->open_collection(pg
->coll
);
4041 // read pg state, log
4042 pg
->read_state(store
, bl
);
4044 if (pg
->must_upgrade()) {
4045 if (!pg
->can_upgrade()) {
4046 derr
<< "PG needs upgrade, but on-disk data is too old; upgrade to"
4047 << " an older version first." << dendl
;
4048 assert(0 == "PG too old to upgrade");
4050 if (!has_upgraded
) {
4051 derr
<< "PGs are upgrading" << dendl
;
4052 has_upgraded
= true;
4054 dout(10) << "PG " << pg
->info
.pgid
4055 << " must upgrade..." << dendl
;
4060 dout(10) << "load_pgs " << *it
<< " deleting dne" << dendl
;
4062 service
.pg_remove_epoch(pg
->pg_id
);
4066 RWLock::WLocker
l(pg_map_lock
);
4067 auto p
= pg_map
.find(pg
->get_pgid());
4068 assert(p
!= pg_map
.end() && p
->second
== pg
);
4069 dout(20) << __func__
<< " removed pg " << pg
<< " from pg_map" << dendl
;
4073 recursive_remove_collection(cct
, store
, pgid
, *it
);
4077 service
.init_splits_between(pg
->info
.pgid
, pg
->get_osdmap(), osdmap
);
4079 // generate state for PG's current mapping
4080 int primary
, up_primary
;
4081 vector
<int> acting
, up
;
4082 pg
->get_osdmap()->pg_to_up_acting_osds(
4083 pgid
.pgid
, &up
, &up_primary
, &acting
, &primary
);
4084 pg
->init_primary_up_acting(
4089 int role
= OSDMap::calc_pg_role(whoami
, pg
->acting
);
4090 if (pg
->pool
.info
.is_replicated() || role
== pg
->pg_whoami
.shard
)
4095 pg
->reg_next_scrub();
4097 PG::RecoveryCtx
rctx(0, 0, 0, 0, 0, 0);
4098 pg
->handle_loaded(&rctx
);
4100 dout(10) << "load_pgs loaded " << *pg
<< " " << pg
->pg_log
.get_log() << dendl
;
4101 if (pg
->pg_log
.is_dirty()) {
4102 ObjectStore::Transaction t
;
4103 pg
->write_if_dirty(t
);
4104 store
->apply_transaction(pg
->osr
.get(), std::move(t
));
4109 RWLock::RLocker
l(pg_map_lock
);
4110 dout(0) << "load_pgs opened " << pg_map
.size() << " pgs" << dendl
;
4113 // clean up old infos object?
4114 if (has_upgraded
&& store
->exists(coll_t::meta(), OSD::make_infos_oid())) {
4115 dout(1) << __func__
<< " removing legacy infos object" << dendl
;
4116 ObjectStore::Transaction t
;
4117 t
.remove(coll_t::meta(), OSD::make_infos_oid());
4118 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4120 derr
<< __func__
<< ": apply_transaction returned "
4121 << cpp_strerror(r
) << dendl
;
4126 build_past_intervals_parallel();
4131 * build past_intervals efficiently on old, degraded, and buried
4132 * clusters. this is important for efficiently catching up osds that
4133 * are way behind on maps to the current cluster state.
4135 * this is a parallel version of PG::generate_past_intervals().
4136 * follow the same logic, but do all pgs at the same time so that we
4137 * can make a single pass across the osdmap history.
4139 void OSD::build_past_intervals_parallel()
4143 vector
<int> old_acting
, old_up
;
4144 epoch_t same_interval_since
;
4148 map
<PG
*,pistate
> pis
;
4150 // calculate junction of map range
4151 epoch_t end_epoch
= superblock
.oldest_map
;
4152 epoch_t cur_epoch
= superblock
.newest_map
;
4154 RWLock::RLocker
l(pg_map_lock
);
4155 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4160 // Ignore PGs only partially created (DNE)
4161 if (pg
->info
.dne()) {
4165 auto rpib
= pg
->get_required_past_interval_bounds(
4167 superblock
.oldest_map
);
4168 if (rpib
.first
>= rpib
.second
&& pg
->past_intervals
.empty()) {
4169 if (pg
->info
.history
.same_interval_since
== 0) {
4170 pg
->info
.history
.same_interval_since
= rpib
.second
;
4174 auto apib
= pg
->past_intervals
.get_bounds();
4175 if (apib
.second
>= rpib
.second
&&
4176 apib
.first
<= rpib
.first
) {
4177 if (pg
->info
.history
.same_interval_since
== 0) {
4178 pg
->info
.history
.same_interval_since
= rpib
.second
;
4184 dout(10) << pg
->info
.pgid
<< " needs " << rpib
.first
<< "-"
4185 << rpib
.second
<< dendl
;
4186 pistate
& p
= pis
[pg
];
4187 p
.start
= rpib
.first
;
4188 p
.end
= rpib
.second
;
4189 p
.same_interval_since
= 0;
4191 if (rpib
.first
< cur_epoch
)
4192 cur_epoch
= rpib
.first
;
4193 if (rpib
.second
> end_epoch
)
4194 end_epoch
= rpib
.second
;
4198 dout(10) << __func__
<< " nothing to build" << dendl
;
4202 dout(1) << __func__
<< " over " << cur_epoch
<< "-" << end_epoch
<< dendl
;
4203 assert(cur_epoch
<= end_epoch
);
4205 OSDMapRef cur_map
, last_map
;
4206 for ( ; cur_epoch
<= end_epoch
; cur_epoch
++) {
4207 dout(10) << __func__
<< " epoch " << cur_epoch
<< dendl
;
4209 cur_map
= get_map(cur_epoch
);
4211 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4213 pistate
& p
= i
->second
;
4215 if (cur_epoch
< p
.start
|| cur_epoch
> p
.end
)
4218 vector
<int> acting
, up
;
4221 pg_t pgid
= pg
->info
.pgid
.pgid
;
4222 if (p
.same_interval_since
&& last_map
->get_pools().count(pgid
.pool()))
4223 pgid
= pgid
.get_ancestor(last_map
->get_pg_num(pgid
.pool()));
4224 cur_map
->pg_to_up_acting_osds(
4225 pgid
, &up
, &up_primary
, &acting
, &primary
);
4227 if (p
.same_interval_since
== 0) {
4228 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4229 << " first map, acting " << acting
4230 << " up " << up
<< ", same_interval_since = " << cur_epoch
<< dendl
;
4231 p
.same_interval_since
= cur_epoch
;
4233 p
.old_acting
= acting
;
4234 p
.primary
= primary
;
4235 p
.up_primary
= up_primary
;
4240 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable(
4241 pg
->get_is_recoverable_predicate());
4242 std::stringstream debug
;
4243 bool new_interval
= PastIntervals::check_new_interval(
4246 p
.old_acting
, acting
,
4250 p
.same_interval_since
,
4251 pg
->info
.history
.last_epoch_clean
,
4255 &pg
->past_intervals
,
4258 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4259 << " " << debug
.str() << dendl
;
4261 p
.old_acting
= acting
;
4262 p
.primary
= primary
;
4263 p
.up_primary
= up_primary
;
4264 p
.same_interval_since
= cur_epoch
;
4269 // Now that past_intervals have been recomputed let's fix the same_interval_since
4270 // if it was cleared by import.
4271 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4273 pistate
& p
= i
->second
;
4275 if (pg
->info
.history
.same_interval_since
== 0) {
4276 assert(p
.same_interval_since
);
4277 dout(10) << __func__
<< " fix same_interval_since " << p
.same_interval_since
<< " pg " << *pg
<< dendl
;
4278 dout(10) << __func__
<< " past_intervals " << pg
->past_intervals
<< dendl
;
4280 pg
->info
.history
.same_interval_since
= p
.same_interval_since
;
4284 // write info only at the end. this is necessary because we check
4285 // whether the past_intervals go far enough back or forward in time,
4286 // but we don't check for holes. we could avoid it by discarding
4287 // the previous past_intervals and rebuilding from scratch, or we
4288 // can just do this and commit all our work at the end.
4289 ObjectStore::Transaction t
;
4291 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4294 pg
->dirty_big_info
= true;
4295 pg
->dirty_info
= true;
4296 pg
->write_if_dirty(t
);
4299 // don't let the transaction get too big
4300 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
4301 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4302 t
= ObjectStore::Transaction();
4307 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4311 * look up a pg. if we have it, great. if not, consider creating it IF the pg mapping
4312 * hasn't changed since the given epoch and we are the primary.
4314 int OSD::handle_pg_peering_evt(
4316 const pg_history_t
& orig_history
,
4317 const PastIntervals
& pi
,
4319 PG::CephPeeringEvtRef evt
)
4321 if (service
.splitting(pgid
)) {
4322 peering_wait_for_split
[pgid
].push_back(evt
);
4326 PG
*pg
= _lookup_lock_pg(pgid
);
4329 if (!osdmap
->have_pg_pool(pgid
.pool()))
4331 int up_primary
, acting_primary
;
4332 vector
<int> up
, acting
;
4333 osdmap
->pg_to_up_acting_osds(
4334 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4336 pg_history_t history
= orig_history
;
4337 bool valid_history
= project_pg_history(
4338 pgid
, history
, epoch
, up
, up_primary
, acting
, acting_primary
);
4340 if (!valid_history
|| epoch
< history
.same_interval_since
) {
4341 dout(10) << __func__
<< pgid
<< " acting changed in "
4342 << history
.same_interval_since
<< " (msg from " << epoch
<< ")"
4347 if (service
.splitting(pgid
)) {
4351 const bool is_mon_create
=
4352 evt
->get_event().dynamic_type() == PG::NullEvt::static_type();
4353 if (maybe_wait_for_max_pg(pgid
, is_mon_create
)) {
4356 // do we need to resurrect a deleting pg?
4359 res_result result
= _try_resurrect_pg(
4360 service
.get_osdmap(),
4365 PG::RecoveryCtx rctx
= create_context();
4368 const pg_pool_t
* pp
= osdmap
->get_pg_pool(pgid
.pool());
4369 if (pp
->has_flag(pg_pool_t::FLAG_EC_OVERWRITES
) &&
4370 store
->get_type() != "bluestore") {
4371 clog
->warn() << "pg " << pgid
4372 << " is at risk of silent data corruption: "
4373 << "the pool allows ec overwrites but is not stored in "
4374 << "bluestore, so deep scrubbing will not detect bitrot";
4376 PG::_create(*rctx
.transaction
, pgid
, pgid
.get_split_bits(pp
->get_pg_num()));
4377 PG::_init(*rctx
.transaction
, pgid
, pp
);
4379 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
4380 if (!pp
->is_replicated() && role
!= pgid
.shard
)
4383 pg
= _create_lock_pg(
4388 acting
, acting_primary
,
4391 pg
->handle_create(&rctx
);
4392 pg
->write_if_dirty(*rctx
.transaction
);
4393 dispatch_context(rctx
, pg
, osdmap
);
4395 dout(10) << *pg
<< " is new" << dendl
;
4397 pg
->queue_peering_event(evt
);
4398 wake_pg_waiters(pg
);
4403 old_pg_state
->lock();
4404 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4405 int old_role
= old_pg_state
->role
;
4406 vector
<int> old_up
= old_pg_state
->up
;
4407 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4408 vector
<int> old_acting
= old_pg_state
->acting
;
4409 int old_primary
= old_pg_state
->primary
.osd
;
4410 pg_history_t old_history
= old_pg_state
->info
.history
;
4411 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4412 old_pg_state
->unlock();
4413 pg
= _create_lock_pg(
4426 pg
->handle_create(&rctx
);
4427 pg
->write_if_dirty(*rctx
.transaction
);
4428 dispatch_context(rctx
, pg
, osdmap
);
4430 dout(10) << *pg
<< " is new (resurrected)" << dendl
;
4432 pg
->queue_peering_event(evt
);
4433 wake_pg_waiters(pg
);
4438 assert(old_pg_state
);
4439 old_pg_state
->lock();
4440 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4441 int old_role
= old_pg_state
->role
;
4442 vector
<int> old_up
= old_pg_state
->up
;
4443 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4444 vector
<int> old_acting
= old_pg_state
->acting
;
4445 int old_primary
= old_pg_state
->primary
.osd
;
4446 pg_history_t old_history
= old_pg_state
->info
.history
;
4447 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4448 old_pg_state
->unlock();
4449 PG
*parent
= _create_lock_pg(
4463 parent
->handle_create(&rctx
);
4464 parent
->write_if_dirty(*rctx
.transaction
);
4465 dispatch_context(rctx
, parent
, osdmap
);
4467 dout(10) << *parent
<< " is new" << dendl
;
4469 assert(service
.splitting(pgid
));
4470 peering_wait_for_split
[pgid
].push_back(evt
);
4472 //parent->queue_peering_event(evt);
4473 parent
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
4474 wake_pg_waiters(parent
);
4483 // already had it. did the mapping change?
4484 if (epoch
< pg
->info
.history
.same_interval_since
) {
4485 dout(10) << *pg
<< __func__
<< " acting changed in "
4486 << pg
->info
.history
.same_interval_since
4487 << " (msg from " << epoch
<< ")" << dendl
;
4489 pg
->queue_peering_event(evt
);
4496 bool OSD::maybe_wait_for_max_pg(spg_t pgid
, bool is_mon_create
)
4498 const auto max_pgs_per_osd
=
4499 (cct
->_conf
->get_val
<uint64_t>("mon_max_pg_per_osd") *
4500 cct
->_conf
->get_val
<double>("osd_max_pg_per_osd_hard_ratio"));
4502 RWLock::RLocker pg_map_locker
{pg_map_lock
};
4503 if (pg_map
.size() < max_pgs_per_osd
) {
4506 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
4507 if (is_mon_create
) {
4508 pending_creates_from_mon
++;
4510 bool is_primary
= osdmap
->get_pg_acting_rank(pgid
.pgid
, whoami
) == 0;
4511 pending_creates_from_osd
.emplace(pgid
.pgid
, is_primary
);
4513 dout(5) << __func__
<< " withhold creation of pg " << pgid
4514 << ": " << pg_map
.size() << " >= "<< max_pgs_per_osd
<< dendl
;
4518 // to re-trigger a peering, we have to twiddle the pg mapping a little bit,
4519 // see PG::should_restart_peering(). OSDMap::pg_to_up_acting_osds() will turn
4520 // to up set if pg_temp is empty. so an empty pg_temp won't work.
4521 static vector
<int32_t> twiddle(const vector
<int>& acting
) {
4522 if (acting
.size() > 1) {
4525 vector
<int32_t> twiddled(acting
.begin(), acting
.end());
4526 twiddled
.push_back(-1);
4531 void OSD::resume_creating_pg()
4533 bool do_sub_pg_creates
= false;
4534 bool have_pending_creates
= false;
4536 const auto max_pgs_per_osd
=
4537 (cct
->_conf
->get_val
<uint64_t>("mon_max_pg_per_osd") *
4538 cct
->_conf
->get_val
<double>("osd_max_pg_per_osd_hard_ratio"));
4539 RWLock::RLocker
l(pg_map_lock
);
4540 if (max_pgs_per_osd
<= pg_map
.size()) {
4541 // this could happen if admin decreases this setting before a PG is removed
4544 unsigned spare_pgs
= max_pgs_per_osd
- pg_map
.size();
4545 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
4546 if (pending_creates_from_mon
> 0) {
4547 do_sub_pg_creates
= true;
4548 if (pending_creates_from_mon
>= spare_pgs
) {
4549 spare_pgs
= pending_creates_from_mon
= 0;
4551 spare_pgs
-= pending_creates_from_mon
;
4552 pending_creates_from_mon
= 0;
4555 auto pg
= pending_creates_from_osd
.cbegin();
4556 while (spare_pgs
> 0 && pg
!= pending_creates_from_osd
.cend()) {
4557 dout(20) << __func__
<< " pg " << pg
->first
<< dendl
;
4559 osdmap
->pg_to_up_acting_osds(pg
->first
, nullptr, nullptr, &acting
, nullptr);
4560 service
.queue_want_pg_temp(pg
->first
, twiddle(acting
), true);
4561 pg
= pending_creates_from_osd
.erase(pg
);
4562 do_sub_pg_creates
= true;
4565 have_pending_creates
= (pending_creates_from_mon
> 0 ||
4566 !pending_creates_from_osd
.empty());
4569 bool do_renew_subs
= false;
4570 if (do_sub_pg_creates
) {
4571 if (monc
->sub_want("osd_pg_creates", last_pg_create_epoch
, 0)) {
4572 dout(4) << __func__
<< ": resolicit pg creates from mon since "
4573 << last_pg_create_epoch
<< dendl
;
4574 do_renew_subs
= true;
4577 version_t start
= osdmap
->get_epoch() + 1;
4578 if (have_pending_creates
) {
4579 // don't miss any new osdmap deleting PGs
4580 if (monc
->sub_want("osdmap", start
, 0)) {
4581 dout(4) << __func__
<< ": resolicit osdmap from mon since "
4583 do_renew_subs
= true;
4585 } else if (do_sub_pg_creates
) {
4586 // no need to subscribe the osdmap continuously anymore
4587 // once the pgtemp and/or mon_subscribe(pg_creates) is sent
4588 if (monc
->sub_want_increment("osdmap", start
, CEPH_SUBSCRIBE_ONETIME
)) {
4589 dout(4) << __func__
<< ": re-subscribe osdmap(onetime) since"
4591 do_renew_subs
= true;
4595 if (do_renew_subs
) {
4599 service
.send_pg_temp();
4602 void OSD::build_initial_pg_history(
4605 utime_t created_stamp
,
4609 dout(10) << __func__
<< " " << pgid
<< " created " << created
<< dendl
;
4610 h
->epoch_created
= created
;
4611 h
->epoch_pool_created
= created
;
4612 h
->same_interval_since
= created
;
4613 h
->same_up_since
= created
;
4614 h
->same_primary_since
= created
;
4615 h
->last_scrub_stamp
= created_stamp
;
4616 h
->last_deep_scrub_stamp
= created_stamp
;
4617 h
->last_clean_scrub_stamp
= created_stamp
;
4619 OSDMapRef lastmap
= service
.get_map(created
);
4620 int up_primary
, acting_primary
;
4621 vector
<int> up
, acting
;
4622 lastmap
->pg_to_up_acting_osds(
4623 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4625 ostringstream debug
;
4626 for (epoch_t e
= created
+ 1; e
<= osdmap
->get_epoch(); ++e
) {
4627 OSDMapRef osdmap
= service
.get_map(e
);
4628 int new_up_primary
, new_acting_primary
;
4629 vector
<int> new_up
, new_acting
;
4630 osdmap
->pg_to_up_acting_osds(
4631 pgid
.pgid
, &new_up
, &new_up_primary
, &new_acting
, &new_acting_primary
);
4633 // this is a bit imprecise, but sufficient?
4634 struct min_size_predicate_t
: public IsPGRecoverablePredicate
{
4635 const pg_pool_t
*pi
;
4636 bool operator()(const set
<pg_shard_t
> &have
) const {
4637 return have
.size() >= pi
->min_size
;
4639 min_size_predicate_t(const pg_pool_t
*i
) : pi(i
) {}
4640 } min_size_predicate(osdmap
->get_pg_pool(pgid
.pgid
.pool()));
4642 bool new_interval
= PastIntervals::check_new_interval(
4649 h
->same_interval_since
,
4650 h
->last_epoch_clean
,
4654 &min_size_predicate
,
4658 h
->same_interval_since
= e
;
4660 h
->same_up_since
= e
;
4662 if (acting_primary
!= new_acting_primary
) {
4663 h
->same_primary_since
= e
;
4665 if (pgid
.pgid
.is_split(lastmap
->get_pg_num(pgid
.pgid
.pool()),
4666 osdmap
->get_pg_num(pgid
.pgid
.pool()),
4668 h
->last_epoch_split
= e
;
4671 acting
= new_acting
;
4672 up_primary
= new_up_primary
;
4673 acting_primary
= new_acting_primary
;
4677 dout(20) << __func__
<< " " << debug
.str() << dendl
;
4678 dout(10) << __func__
<< " " << *h
<< " " << *pi
4679 << " [" << (pi
->empty() ? pair
<epoch_t
,epoch_t
>(0,0) :
4680 pi
->get_bounds()) << ")"
4685 * Fill in the passed history so you know same_interval_since, same_up_since,
4686 * and same_primary_since.
4688 bool OSD::project_pg_history(spg_t pgid
, pg_history_t
& h
, epoch_t from
,
4689 const vector
<int>& currentup
,
4690 int currentupprimary
,
4691 const vector
<int>& currentacting
,
4692 int currentactingprimary
)
4694 dout(15) << "project_pg_history " << pgid
4695 << " from " << from
<< " to " << osdmap
->get_epoch()
4700 for (e
= osdmap
->get_epoch();
4703 // verify during intermediate epoch (e-1)
4704 OSDMapRef oldmap
= service
.try_get_map(e
-1);
4706 dout(15) << __func__
<< ": found map gap, returning false" << dendl
;
4709 assert(oldmap
->have_pg_pool(pgid
.pool()));
4711 int upprimary
, actingprimary
;
4712 vector
<int> up
, acting
;
4713 oldmap
->pg_to_up_acting_osds(
4720 // acting set change?
4721 if ((actingprimary
!= currentactingprimary
||
4722 upprimary
!= currentupprimary
||
4723 acting
!= currentacting
||
4724 up
!= currentup
) && e
> h
.same_interval_since
) {
4725 dout(15) << "project_pg_history " << pgid
<< " acting|up changed in " << e
4726 << " from " << acting
<< "/" << up
4727 << " " << actingprimary
<< "/" << upprimary
4728 << " -> " << currentacting
<< "/" << currentup
4729 << " " << currentactingprimary
<< "/" << currentupprimary
4731 h
.same_interval_since
= e
;
4734 if (pgid
.is_split(oldmap
->get_pg_num(pgid
.pool()),
4735 osdmap
->get_pg_num(pgid
.pool()),
4736 0) && e
> h
.same_interval_since
) {
4737 h
.same_interval_since
= e
;
4740 if ((up
!= currentup
|| upprimary
!= currentupprimary
)
4741 && e
> h
.same_up_since
) {
4742 dout(15) << "project_pg_history " << pgid
<< " up changed in " << e
4743 << " from " << up
<< " " << upprimary
4744 << " -> " << currentup
<< " " << currentupprimary
<< dendl
;
4745 h
.same_up_since
= e
;
4749 if (OSDMap::primary_changed(
4752 currentactingprimary
,
4754 e
> h
.same_primary_since
) {
4755 dout(15) << "project_pg_history " << pgid
<< " primary changed in " << e
<< dendl
;
4756 h
.same_primary_since
= e
;
4759 if (h
.same_interval_since
>= e
&& h
.same_up_since
>= e
&& h
.same_primary_since
>= e
)
4763 // base case: these floors should be the pg creation epoch if we didn't
4764 // find any changes.
4765 if (e
== h
.epoch_created
) {
4766 if (!h
.same_interval_since
)
4767 h
.same_interval_since
= e
;
4768 if (!h
.same_up_since
)
4769 h
.same_up_since
= e
;
4770 if (!h
.same_primary_since
)
4771 h
.same_primary_since
= e
;
4774 dout(15) << "project_pg_history end " << h
<< dendl
;
4780 void OSD::_add_heartbeat_peer(int p
)
4786 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(p
);
4787 if (i
== heartbeat_peers
.end()) {
4788 pair
<ConnectionRef
,ConnectionRef
> cons
= service
.get_con_osd_hb(p
, osdmap
->get_epoch());
4791 hi
= &heartbeat_peers
[p
];
4793 HeartbeatSession
*s
= new HeartbeatSession(p
);
4794 hi
->con_back
= cons
.first
.get();
4795 hi
->con_back
->set_priv(s
->get());
4797 hi
->con_front
= cons
.second
.get();
4798 hi
->con_front
->set_priv(s
->get());
4799 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4800 << " " << hi
->con_back
->get_peer_addr()
4801 << " " << hi
->con_front
->get_peer_addr()
4804 hi
->con_front
.reset(NULL
);
4805 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4806 << " " << hi
->con_back
->get_peer_addr()
4813 hi
->epoch
= osdmap
->get_epoch();
4816 void OSD::_remove_heartbeat_peer(int n
)
4818 map
<int,HeartbeatInfo
>::iterator q
= heartbeat_peers
.find(n
);
4819 assert(q
!= heartbeat_peers
.end());
4820 dout(20) << " removing heartbeat peer osd." << n
4821 << " " << q
->second
.con_back
->get_peer_addr()
4822 << " " << (q
->second
.con_front
? q
->second
.con_front
->get_peer_addr() : entity_addr_t())
4824 q
->second
.con_back
->mark_down();
4825 if (q
->second
.con_front
) {
4826 q
->second
.con_front
->mark_down();
4828 heartbeat_peers
.erase(q
);
4831 void OSD::need_heartbeat_peer_update()
4835 dout(20) << "need_heartbeat_peer_update" << dendl
;
4836 heartbeat_set_peers_need_update();
4839 void OSD::maybe_update_heartbeat_peers()
4841 assert(osd_lock
.is_locked());
4843 if (is_waiting_for_healthy()) {
4844 utime_t now
= ceph_clock_now();
4845 if (last_heartbeat_resample
== utime_t()) {
4846 last_heartbeat_resample
= now
;
4847 heartbeat_set_peers_need_update();
4848 } else if (!heartbeat_peers_need_update()) {
4849 utime_t dur
= now
- last_heartbeat_resample
;
4850 if (dur
> cct
->_conf
->osd_heartbeat_grace
) {
4851 dout(10) << "maybe_update_heartbeat_peers forcing update after " << dur
<< " seconds" << dendl
;
4852 heartbeat_set_peers_need_update();
4853 last_heartbeat_resample
= now
;
4854 reset_heartbeat_peers(); // we want *new* peers!
4859 if (!heartbeat_peers_need_update())
4861 heartbeat_clear_peers_need_update();
4863 Mutex::Locker
l(heartbeat_lock
);
4865 dout(10) << "maybe_update_heartbeat_peers updating" << dendl
;
4868 // build heartbeat from set
4870 RWLock::RLocker
l(pg_map_lock
);
4871 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4875 pg
->heartbeat_peer_lock
.Lock();
4876 dout(20) << i
->first
<< " heartbeat_peers " << pg
->heartbeat_peers
<< dendl
;
4877 for (set
<int>::iterator p
= pg
->heartbeat_peers
.begin();
4878 p
!= pg
->heartbeat_peers
.end();
4880 if (osdmap
->is_up(*p
))
4881 _add_heartbeat_peer(*p
);
4882 for (set
<int>::iterator p
= pg
->probe_targets
.begin();
4883 p
!= pg
->probe_targets
.end();
4885 if (osdmap
->is_up(*p
))
4886 _add_heartbeat_peer(*p
);
4887 pg
->heartbeat_peer_lock
.Unlock();
4891 // include next and previous up osds to ensure we have a fully-connected set
4892 set
<int> want
, extras
;
4893 int next
= osdmap
->get_next_up_osd_after(whoami
);
4896 int prev
= osdmap
->get_previous_up_osd_before(whoami
);
4897 if (prev
>= 0 && prev
!= next
)
4900 for (set
<int>::iterator p
= want
.begin(); p
!= want
.end(); ++p
) {
4901 dout(10) << " adding neighbor peer osd." << *p
<< dendl
;
4903 _add_heartbeat_peer(*p
);
4906 // remove down peers; enumerate extras
4907 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4908 while (p
!= heartbeat_peers
.end()) {
4909 if (!osdmap
->is_up(p
->first
)) {
4912 _remove_heartbeat_peer(o
);
4915 if (p
->second
.epoch
< osdmap
->get_epoch()) {
4916 extras
.insert(p
->first
);
4922 int start
= osdmap
->get_next_up_osd_after(whoami
);
4923 for (int n
= start
; n
>= 0; ) {
4924 if ((int)heartbeat_peers
.size() >= cct
->_conf
->osd_heartbeat_min_peers
)
4926 if (!extras
.count(n
) && !want
.count(n
) && n
!= whoami
) {
4927 dout(10) << " adding random peer osd." << n
<< dendl
;
4929 _add_heartbeat_peer(n
);
4931 n
= osdmap
->get_next_up_osd_after(n
);
4933 break; // came full circle; stop
4937 for (set
<int>::iterator p
= extras
.begin();
4938 (int)heartbeat_peers
.size() > cct
->_conf
->osd_heartbeat_min_peers
&& p
!= extras
.end();
4942 _remove_heartbeat_peer(*p
);
4945 dout(10) << "maybe_update_heartbeat_peers " << heartbeat_peers
.size() << " peers, extras " << extras
<< dendl
;
4948 void OSD::reset_heartbeat_peers()
4950 assert(osd_lock
.is_locked());
4951 dout(10) << "reset_heartbeat_peers" << dendl
;
4952 Mutex::Locker
l(heartbeat_lock
);
4953 while (!heartbeat_peers
.empty()) {
4954 HeartbeatInfo
& hi
= heartbeat_peers
.begin()->second
;
4955 hi
.con_back
->mark_down();
4957 hi
.con_front
->mark_down();
4959 heartbeat_peers
.erase(heartbeat_peers
.begin());
4961 failure_queue
.clear();
4964 void OSD::handle_osd_ping(MOSDPing
*m
)
4966 if (superblock
.cluster_fsid
!= m
->fsid
) {
4967 dout(20) << "handle_osd_ping from " << m
->get_source_inst()
4968 << " bad fsid " << m
->fsid
<< " != " << superblock
.cluster_fsid
<< dendl
;
4973 int from
= m
->get_source().num();
4975 heartbeat_lock
.Lock();
4976 if (is_stopping()) {
4977 heartbeat_lock
.Unlock();
4982 OSDMapRef curmap
= service
.get_osdmap();
4984 heartbeat_lock
.Unlock();
4991 case MOSDPing::PING
:
4993 if (cct
->_conf
->osd_debug_drop_ping_probability
> 0) {
4994 auto heartbeat_drop
= debug_heartbeat_drops_remaining
.find(from
);
4995 if (heartbeat_drop
!= debug_heartbeat_drops_remaining
.end()) {
4996 if (heartbeat_drop
->second
== 0) {
4997 debug_heartbeat_drops_remaining
.erase(heartbeat_drop
);
4999 --heartbeat_drop
->second
;
5000 dout(5) << "Dropping heartbeat from " << from
5001 << ", " << heartbeat_drop
->second
5002 << " remaining to drop" << dendl
;
5005 } else if (cct
->_conf
->osd_debug_drop_ping_probability
>
5006 ((((double)(rand()%100))/100.0))) {
5008 debug_heartbeat_drops_remaining
.insert(std::make_pair(from
,
5009 cct
->_conf
->osd_debug_drop_ping_duration
)).first
;
5010 dout(5) << "Dropping heartbeat from " << from
5011 << ", " << heartbeat_drop
->second
5012 << " remaining to drop" << dendl
;
5017 if (!cct
->get_heartbeat_map()->is_healthy()) {
5018 dout(10) << "internal heartbeat not healthy, dropping ping request" << dendl
;
5022 Message
*r
= new MOSDPing(monc
->get_fsid(),
5023 curmap
->get_epoch(),
5024 MOSDPing::PING_REPLY
, m
->stamp
,
5025 cct
->_conf
->osd_heartbeat_min_size
);
5026 m
->get_connection()->send_message(r
);
5028 if (curmap
->is_up(from
)) {
5029 service
.note_peer_epoch(from
, m
->map_epoch
);
5031 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
5033 service
.share_map_peer(from
, con
.get());
5036 } else if (!curmap
->exists(from
) ||
5037 curmap
->get_down_at(from
) > m
->map_epoch
) {
5038 // tell them they have died
5039 Message
*r
= new MOSDPing(monc
->get_fsid(),
5040 curmap
->get_epoch(),
5043 cct
->_conf
->osd_heartbeat_min_size
);
5044 m
->get_connection()->send_message(r
);
5049 case MOSDPing::PING_REPLY
:
5051 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(from
);
5052 if (i
!= heartbeat_peers
.end()) {
5053 if (m
->get_connection() == i
->second
.con_back
) {
5054 dout(25) << "handle_osd_ping got reply from osd." << from
5055 << " first_tx " << i
->second
.first_tx
5056 << " last_tx " << i
->second
.last_tx
5057 << " last_rx_back " << i
->second
.last_rx_back
<< " -> " << m
->stamp
5058 << " last_rx_front " << i
->second
.last_rx_front
5060 i
->second
.last_rx_back
= m
->stamp
;
5061 // if there is no front con, set both stamps.
5062 if (i
->second
.con_front
== NULL
)
5063 i
->second
.last_rx_front
= m
->stamp
;
5064 } else if (m
->get_connection() == i
->second
.con_front
) {
5065 dout(25) << "handle_osd_ping got reply from osd." << from
5066 << " first_tx " << i
->second
.first_tx
5067 << " last_tx " << i
->second
.last_tx
5068 << " last_rx_back " << i
->second
.last_rx_back
5069 << " last_rx_front " << i
->second
.last_rx_front
<< " -> " << m
->stamp
5071 i
->second
.last_rx_front
= m
->stamp
;
5074 utime_t cutoff
= ceph_clock_now();
5075 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
5076 if (i
->second
.is_healthy(cutoff
)) {
5077 // Cancel false reports
5078 auto failure_queue_entry
= failure_queue
.find(from
);
5079 if (failure_queue_entry
!= failure_queue
.end()) {
5080 dout(10) << "handle_osd_ping canceling queued "
5081 << "failure report for osd." << from
<< dendl
;
5082 failure_queue
.erase(failure_queue_entry
);
5085 auto failure_pending_entry
= failure_pending
.find(from
);
5086 if (failure_pending_entry
!= failure_pending
.end()) {
5087 dout(10) << "handle_osd_ping canceling in-flight "
5088 << "failure report for osd." << from
<< dendl
;
5089 send_still_alive(curmap
->get_epoch(),
5090 failure_pending_entry
->second
.second
);
5091 failure_pending
.erase(failure_pending_entry
);
5097 curmap
->is_up(from
)) {
5098 service
.note_peer_epoch(from
, m
->map_epoch
);
5100 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
5102 service
.share_map_peer(from
, con
.get());
5109 case MOSDPing::YOU_DIED
:
5110 dout(10) << "handle_osd_ping " << m
->get_source_inst()
5111 << " says i am down in " << m
->map_epoch
<< dendl
;
5112 osdmap_subscribe(curmap
->get_epoch()+1, false);
5116 heartbeat_lock
.Unlock();
5120 void OSD::heartbeat_entry()
5122 Mutex::Locker
l(heartbeat_lock
);
5125 while (!heartbeat_stop
) {
5128 double wait
= .5 + ((float)(rand() % 10)/10.0) * (float)cct
->_conf
->osd_heartbeat_interval
;
5130 w
.set_from_double(wait
);
5131 dout(30) << "heartbeat_entry sleeping for " << wait
<< dendl
;
5132 heartbeat_cond
.WaitInterval(heartbeat_lock
, w
);
5135 dout(30) << "heartbeat_entry woke up" << dendl
;
5139 void OSD::heartbeat_check()
5141 assert(heartbeat_lock
.is_locked());
5142 utime_t now
= ceph_clock_now();
5144 // check for heartbeat replies (move me elsewhere?)
5145 utime_t cutoff
= now
;
5146 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
5147 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
5148 p
!= heartbeat_peers
.end();
5151 if (p
->second
.first_tx
== utime_t()) {
5152 dout(25) << "heartbeat_check we haven't sent ping to osd." << p
->first
5153 << "yet, skipping" << dendl
;
5157 dout(25) << "heartbeat_check osd." << p
->first
5158 << " first_tx " << p
->second
.first_tx
5159 << " last_tx " << p
->second
.last_tx
5160 << " last_rx_back " << p
->second
.last_rx_back
5161 << " last_rx_front " << p
->second
.last_rx_front
5163 if (p
->second
.is_unhealthy(cutoff
)) {
5164 if (p
->second
.last_rx_back
== utime_t() ||
5165 p
->second
.last_rx_front
== utime_t()) {
5166 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
5167 << " osd." << p
->first
<< " ever on either front or back, first ping sent "
5168 << p
->second
.first_tx
<< " (cutoff " << cutoff
<< ")" << dendl
;
5170 failure_queue
[p
->first
] = p
->second
.last_tx
;
5172 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
5173 << " osd." << p
->first
<< " since back " << p
->second
.last_rx_back
5174 << " front " << p
->second
.last_rx_front
5175 << " (cutoff " << cutoff
<< ")" << dendl
;
5177 failure_queue
[p
->first
] = MIN(p
->second
.last_rx_back
, p
->second
.last_rx_front
);
5183 void OSD::heartbeat()
5185 dout(30) << "heartbeat" << dendl
;
5189 int n_samples
= 86400 / cct
->_conf
->osd_heartbeat_interval
;
5190 if (getloadavg(loadavgs
, 1) == 1) {
5191 logger
->set(l_osd_loadavg
, 100 * loadavgs
[0]);
5192 daily_loadavg
= (daily_loadavg
* (n_samples
- 1) + loadavgs
[0]) / n_samples
;
5193 dout(30) << "heartbeat: daily_loadavg " << daily_loadavg
<< dendl
;
5196 dout(30) << "heartbeat checking stats" << dendl
;
5199 vector
<int> hb_peers
;
5200 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
5201 p
!= heartbeat_peers
.end();
5203 hb_peers
.push_back(p
->first
);
5204 service
.update_osd_stat(hb_peers
);
5206 dout(5) << "heartbeat: " << service
.get_osd_stat() << dendl
;
5208 utime_t now
= ceph_clock_now();
5211 for (map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.begin();
5212 i
!= heartbeat_peers
.end();
5214 int peer
= i
->first
;
5215 i
->second
.last_tx
= now
;
5216 if (i
->second
.first_tx
== utime_t())
5217 i
->second
.first_tx
= now
;
5218 dout(30) << "heartbeat sending ping to osd." << peer
<< dendl
;
5219 i
->second
.con_back
->send_message(new MOSDPing(monc
->get_fsid(),
5220 service
.get_osdmap()->get_epoch(),
5221 MOSDPing::PING
, now
,
5222 cct
->_conf
->osd_heartbeat_min_size
));
5224 if (i
->second
.con_front
)
5225 i
->second
.con_front
->send_message(new MOSDPing(monc
->get_fsid(),
5226 service
.get_osdmap()->get_epoch(),
5227 MOSDPing::PING
, now
,
5228 cct
->_conf
->osd_heartbeat_min_size
));
5231 logger
->set(l_osd_hb_to
, heartbeat_peers
.size());
5233 // hmm.. am i all alone?
5234 dout(30) << "heartbeat lonely?" << dendl
;
5235 if (heartbeat_peers
.empty()) {
5236 if (now
- last_mon_heartbeat
> cct
->_conf
->osd_mon_heartbeat_interval
&& is_active()) {
5237 last_mon_heartbeat
= now
;
5238 dout(10) << "i have no heartbeat peers; checking mon for new map" << dendl
;
5239 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5243 dout(30) << "heartbeat done" << dendl
;
5246 bool OSD::heartbeat_reset(Connection
*con
)
5248 HeartbeatSession
*s
= static_cast<HeartbeatSession
*>(con
->get_priv());
5250 heartbeat_lock
.Lock();
5251 if (is_stopping()) {
5252 heartbeat_lock
.Unlock();
5256 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(s
->peer
);
5257 if (p
!= heartbeat_peers
.end() &&
5258 (p
->second
.con_back
== con
||
5259 p
->second
.con_front
== con
)) {
5260 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5261 << ", reopening" << dendl
;
5262 if (con
!= p
->second
.con_back
) {
5263 p
->second
.con_back
->mark_down();
5265 p
->second
.con_back
.reset(NULL
);
5266 if (p
->second
.con_front
&& con
!= p
->second
.con_front
) {
5267 p
->second
.con_front
->mark_down();
5269 p
->second
.con_front
.reset(NULL
);
5270 pair
<ConnectionRef
,ConnectionRef
> newcon
= service
.get_con_osd_hb(p
->second
.peer
, p
->second
.epoch
);
5272 p
->second
.con_back
= newcon
.first
.get();
5273 p
->second
.con_back
->set_priv(s
->get());
5274 if (newcon
.second
) {
5275 p
->second
.con_front
= newcon
.second
.get();
5276 p
->second
.con_front
->set_priv(s
->get());
5279 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5280 << ", raced with osdmap update, closing out peer" << dendl
;
5281 heartbeat_peers
.erase(p
);
5284 dout(10) << "heartbeat_reset closing (old) failed hb con " << con
<< dendl
;
5286 heartbeat_lock
.Unlock();
5294 // =========================================
5298 assert(osd_lock
.is_locked());
5299 dout(10) << "tick" << dendl
;
5301 if (is_active() || is_waiting_for_healthy()) {
5302 maybe_update_heartbeat_peers();
5305 if (is_waiting_for_healthy()) {
5307 } else if (is_preboot() &&
5308 waiting_for_luminous_mons
&&
5309 monc
->monmap
.get_required_features().contains_all(
5310 ceph::features::mon::FEATURE_LUMINOUS
)) {
5311 // mon upgrade finished!
5317 tick_timer
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick(this));
5320 void OSD::tick_without_osd_lock()
5322 assert(tick_timer_lock
.is_locked());
5323 dout(10) << "tick_without_osd_lock" << dendl
;
5325 logger
->set(l_osd_buf
, buffer::get_total_alloc());
5326 logger
->set(l_osd_history_alloc_bytes
, SHIFT_ROUND_UP(buffer::get_history_alloc_bytes(), 20));
5327 logger
->set(l_osd_history_alloc_num
, buffer::get_history_alloc_num());
5328 logger
->set(l_osd_cached_crc
, buffer::get_cached_crc());
5329 logger
->set(l_osd_cached_crc_adjusted
, buffer::get_cached_crc_adjusted());
5330 logger
->set(l_osd_missed_crc
, buffer::get_missed_crc());
5331 logger
->set(l_osd_pg_removing
, remove_wq
.get_remove_queue_len());
5333 // osd_lock is not being held, which means the OSD state
5334 // might change when doing the monitor report
5335 if (is_active() || is_waiting_for_healthy()) {
5336 heartbeat_lock
.Lock();
5338 heartbeat_lock
.Unlock();
5340 map_lock
.get_read();
5341 Mutex::Locker
l(mon_report_lock
);
5345 bool report
= false;
5346 utime_t now
= ceph_clock_now();
5347 pg_stat_queue_lock
.Lock();
5348 double backoff
= stats_ack_timeout
/ cct
->_conf
->osd_mon_ack_timeout
;
5349 double adjusted_min
= cct
->_conf
->osd_mon_report_interval_min
* backoff
;
5350 // note: we shouldn't adjust max because it must remain < the
5351 // mon's mon_osd_report_timeout (which defaults to 1.5x our
5353 double max
= cct
->_conf
->osd_mon_report_interval_max
;
5354 if (!outstanding_pg_stats
.empty() &&
5355 (now
- stats_ack_timeout
) > last_pg_stats_ack
) {
5356 dout(1) << __func__
<< " mon hasn't acked PGStats in "
5357 << now
- last_pg_stats_ack
5358 << " seconds, reconnecting elsewhere" << dendl
;
5360 last_pg_stats_ack
= now
; // reset clock
5361 last_pg_stats_sent
= utime_t();
5363 MAX(cct
->_conf
->osd_mon_ack_timeout
,
5364 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_factor
);
5365 outstanding_pg_stats
.clear();
5367 if (now
- last_pg_stats_sent
> max
) {
5368 osd_stat_updated
= true;
5370 } else if (service
.need_fullness_update()) {
5372 } else if ((int)outstanding_pg_stats
.size() >=
5373 cct
->_conf
->osd_mon_report_max_in_flight
) {
5374 dout(20) << __func__
<< " have max " << outstanding_pg_stats
5375 << " stats updates in flight" << dendl
;
5377 if (now
- last_mon_report
> adjusted_min
) {
5378 dout(20) << __func__
<< " stats backoff " << backoff
5379 << " adjusted_min " << adjusted_min
<< " - sending report"
5381 osd_stat_updated
= true;
5385 pg_stat_queue_lock
.Unlock();
5388 monc
->reopen_session();
5389 } else if (report
) {
5390 last_mon_report
= now
;
5392 // do any pending reports
5395 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5399 map_lock
.put_read();
5403 if (!scrub_random_backoff()) {
5406 service
.promote_throttle_recalibrate();
5407 resume_creating_pg();
5408 bool need_send_beacon
= false;
5409 const auto now
= ceph::coarse_mono_clock::now();
5411 // borrow lec lock to pretect last_sent_beacon from changing
5412 Mutex::Locker l
{min_last_epoch_clean_lock
};
5413 const auto elapsed
= now
- last_sent_beacon
;
5414 if (chrono::duration_cast
<chrono::seconds
>(elapsed
).count() >
5415 cct
->_conf
->osd_beacon_report_interval
) {
5416 need_send_beacon
= true;
5419 if (need_send_beacon
) {
5424 mgrc
.update_osd_health(get_health_metrics());
5425 service
.kick_recovery_queue();
5426 tick_timer_without_osd_lock
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick_WithoutOSDLock(this));
5429 void OSD::check_ops_in_flight()
5431 vector
<string
> warnings
;
5432 if (op_tracker
.check_ops_in_flight(warnings
)) {
5433 for (vector
<string
>::iterator i
= warnings
.begin();
5434 i
!= warnings
.end();
5442 // setomapval <pool-id> [namespace/]<obj-name> <key> <val>
5443 // rmomapkey <pool-id> [namespace/]<obj-name> <key>
5444 // setomapheader <pool-id> [namespace/]<obj-name> <header>
5445 // getomap <pool> [namespace/]<obj-name>
5446 // truncobj <pool-id> [namespace/]<obj-name> <newlen>
5447 // injectmdataerr [namespace/]<obj-name> [shardid]
5448 // injectdataerr [namespace/]<obj-name> [shardid]
5450 // set_recovery_delay [utime]
5451 void TestOpsSocketHook::test_ops(OSDService
*service
, ObjectStore
*store
,
5452 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
)
5455 //Support changing the omap on a single osd by using the Admin Socket to
5456 //directly request the osd make a change.
5457 if (command
== "setomapval" || command
== "rmomapkey" ||
5458 command
== "setomapheader" || command
== "getomap" ||
5459 command
== "truncobj" || command
== "injectmdataerr" ||
5460 command
== "injectdataerr"
5464 OSDMapRef curmap
= service
->get_osdmap();
5469 cmd_getval(service
->cct
, cmdmap
, "pool", poolstr
);
5470 pool
= curmap
->lookup_pg_pool_name(poolstr
);
5471 //If we can't find it by name then maybe id specified
5472 if (pool
< 0 && isdigit(poolstr
[0]))
5473 pool
= atoll(poolstr
.c_str());
5475 ss
<< "Invalid pool '" << poolstr
<< "''";
5479 string objname
, nspace
;
5480 cmd_getval(service
->cct
, cmdmap
, "objname", objname
);
5481 std::size_t found
= objname
.find_first_of('/');
5482 if (found
!= string::npos
) {
5483 nspace
= objname
.substr(0, found
);
5484 objname
= objname
.substr(found
+1);
5486 object_locator_t
oloc(pool
, nspace
);
5487 r
= curmap
->object_locator_to_pg(object_t(objname
), oloc
, rawpg
);
5490 ss
<< "Invalid namespace/objname";
5495 cmd_getval(service
->cct
, cmdmap
, "shardid", shardid
, int64_t(shard_id_t::NO_SHARD
));
5496 hobject_t
obj(object_t(objname
), string(""), CEPH_NOSNAP
, rawpg
.ps(), pool
, nspace
);
5497 ghobject_t
gobj(obj
, ghobject_t::NO_GEN
, shard_id_t(uint8_t(shardid
)));
5498 spg_t
pgid(curmap
->raw_pg_to_pg(rawpg
), shard_id_t(shardid
));
5499 if (curmap
->pg_is_ec(rawpg
)) {
5500 if ((command
!= "injectdataerr") && (command
!= "injectmdataerr")) {
5501 ss
<< "Must not call on ec pool, except injectdataerr or injectmdataerr";
5506 ObjectStore::Transaction t
;
5508 if (command
== "setomapval") {
5509 map
<string
, bufferlist
> newattrs
;
5512 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5513 cmd_getval(service
->cct
, cmdmap
, "val", valstr
);
5516 newattrs
[key
] = val
;
5517 t
.omap_setkeys(coll_t(pgid
), ghobject_t(obj
), newattrs
);
5518 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5520 ss
<< "error=" << r
;
5523 } else if (command
== "rmomapkey") {
5526 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5529 t
.omap_rmkeys(coll_t(pgid
), ghobject_t(obj
), keys
);
5530 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5532 ss
<< "error=" << r
;
5535 } else if (command
== "setomapheader") {
5536 bufferlist newheader
;
5539 cmd_getval(service
->cct
, cmdmap
, "header", headerstr
);
5540 newheader
.append(headerstr
);
5541 t
.omap_setheader(coll_t(pgid
), ghobject_t(obj
), newheader
);
5542 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5544 ss
<< "error=" << r
;
5547 } else if (command
== "getomap") {
5548 //Debug: Output entire omap
5550 map
<string
, bufferlist
> keyvals
;
5551 r
= store
->omap_get(coll_t(pgid
), ghobject_t(obj
), &hdrbl
, &keyvals
);
5553 ss
<< "header=" << string(hdrbl
.c_str(), hdrbl
.length());
5554 for (map
<string
, bufferlist
>::iterator it
= keyvals
.begin();
5555 it
!= keyvals
.end(); ++it
)
5556 ss
<< " key=" << (*it
).first
<< " val="
5557 << string((*it
).second
.c_str(), (*it
).second
.length());
5559 ss
<< "error=" << r
;
5561 } else if (command
== "truncobj") {
5563 cmd_getval(service
->cct
, cmdmap
, "len", trunclen
);
5564 t
.truncate(coll_t(pgid
), ghobject_t(obj
), trunclen
);
5565 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5567 ss
<< "error=" << r
;
5570 } else if (command
== "injectdataerr") {
5571 store
->inject_data_error(gobj
);
5573 } else if (command
== "injectmdataerr") {
5574 store
->inject_mdata_error(gobj
);
5579 if (command
== "set_recovery_delay") {
5581 cmd_getval(service
->cct
, cmdmap
, "utime", delay
, (int64_t)0);
5584 int r
= service
->cct
->_conf
->set_val("osd_recovery_delay_start",
5587 ss
<< "set_recovery_delay: error setting "
5588 << "osd_recovery_delay_start to '" << delay
<< "': error "
5592 service
->cct
->_conf
->apply_changes(NULL
);
5593 ss
<< "set_recovery_delay: set osd_recovery_delay_start "
5594 << "to " << service
->cct
->_conf
->osd_recovery_delay_start
;
5597 if (command
== "trigger_scrub") {
5599 OSDMapRef curmap
= service
->get_osdmap();
5603 cmd_getval(service
->cct
, cmdmap
, "pgid", pgidstr
);
5604 if (!pgid
.parse(pgidstr
.c_str())) {
5605 ss
<< "Invalid pgid specified";
5609 PG
*pg
= service
->osd
->_lookup_lock_pg(pgid
);
5610 if (pg
== nullptr) {
5611 ss
<< "Can't find pg " << pgid
;
5615 if (pg
->is_primary()) {
5616 pg
->unreg_next_scrub();
5617 const pg_pool_t
*p
= curmap
->get_pg_pool(pgid
.pool());
5618 double pool_scrub_max_interval
= 0;
5619 p
->opts
.get(pool_opts_t::SCRUB_MAX_INTERVAL
, &pool_scrub_max_interval
);
5620 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
5621 pool_scrub_max_interval
: g_conf
->osd_scrub_max_interval
;
5622 // Instead of marking must_scrub force a schedule scrub
5623 utime_t stamp
= ceph_clock_now();
5624 stamp
-= scrub_max_interval
;
5625 stamp
-= 100.0; // push back last scrub more for good measure
5626 pg
->info
.history
.last_scrub_stamp
= stamp
;
5627 pg
->reg_next_scrub();
5630 ss
<< "Not primary";
5635 if (command
== "injectfull") {
5638 OSDService::s_names state
;
5639 cmd_getval(service
->cct
, cmdmap
, "type", type
, string("full"));
5640 cmd_getval(service
->cct
, cmdmap
, "count", count
, (int64_t)-1);
5641 if (type
== "none" || count
== 0) {
5645 state
= service
->get_full_state(type
);
5646 if (state
== OSDService::s_names::INVALID
) {
5647 ss
<< "Invalid type use (none, nearfull, backfillfull, full, failsafe)";
5650 service
->set_injectfull(state
, count
);
5653 ss
<< "Internal error - command=" << command
;
5656 // =========================================
5659 ObjectStore
*store
, SnapMapper
*mapper
,
5661 ObjectStore::Sequencer
*osr
,
5662 coll_t coll
, DeletingStateRef dstate
,
5664 ThreadPool::TPHandle
&handle
)
5666 vector
<ghobject_t
> olist
;
5668 ObjectStore::Transaction t
;
5670 handle
.reset_tp_timeout();
5671 store
->collection_list(
5674 ghobject_t::get_max(),
5675 store
->get_ideal_list_max(),
5678 generic_dout(10) << __func__
<< " " << olist
<< dendl
;
5679 // default cont to true, this is safe because caller(OSD::RemoveWQ::_process())
5680 // will recheck the answer before it really goes on.
5682 for (vector
<ghobject_t
>::iterator i
= olist
.begin();
5687 OSDriver::OSTransaction
_t(osdriver
->get_transaction(&t
));
5688 int r
= mapper
->remove_oid(i
->hobj
, &_t
);
5689 if (r
!= 0 && r
!= -ENOENT
) {
5693 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
5695 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5696 cont
= dstate
->pause_clearing();
5697 handle
.suspend_tp_timeout();
5699 handle
.reset_tp_timeout();
5701 cont
= dstate
->resume_clearing();
5704 t
= ObjectStore::Transaction();
5710 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5711 cont
= dstate
->pause_clearing();
5712 handle
.suspend_tp_timeout();
5714 handle
.reset_tp_timeout();
5716 cont
= dstate
->resume_clearing();
5718 // whether there are more objects to remove in the collection
5719 *finished
= next
.is_max();
5723 void OSD::RemoveWQ::_process(
5724 pair
<PGRef
, DeletingStateRef
> item
,
5725 ThreadPool::TPHandle
&handle
)
5728 PGRef
pg(item
.first
);
5729 SnapMapper
&mapper
= pg
->snap_mapper
;
5730 OSDriver
&driver
= pg
->osdriver
;
5731 coll_t coll
= coll_t(pg
->info
.pgid
);
5733 bool finished
= false;
5735 if (!item
.second
->start_or_resume_clearing())
5738 bool cont
= remove_dir(
5739 pg
->cct
, store
, &mapper
, &driver
, pg
->osr
.get(), coll
, item
.second
,
5744 if (item
.second
->pause_clearing())
5749 if (!item
.second
->start_deleting())
5752 ObjectStore::Transaction t
;
5753 PGLog::clear_info_log(pg
->info
.pgid
, &t
);
5755 if (cct
->_conf
->osd_inject_failure_on_pg_removal
) {
5756 generic_derr
<< "osd_inject_failure_on_pg_removal" << dendl
;
5759 t
.remove_collection(coll
);
5761 // We need the sequencer to stick around until the op is complete
5762 store
->queue_transaction(
5767 0, // onreadable sync
5768 new ContainerContext
<PGRef
>(pg
),
5771 item
.second
->finish_deleting();
5773 // =========================================
5775 void OSD::ms_handle_connect(Connection
*con
)
5777 dout(10) << __func__
<< " con " << con
<< dendl
;
5778 if (con
->get_peer_type() == CEPH_ENTITY_TYPE_MON
) {
5779 Mutex::Locker
l(osd_lock
);
5782 dout(10) << __func__
<< " on mon" << dendl
;
5786 } else if (is_booting()) {
5787 _send_boot(); // resend boot message
5789 map_lock
.get_read();
5790 Mutex::Locker
l2(mon_report_lock
);
5792 utime_t now
= ceph_clock_now();
5793 last_mon_report
= now
;
5795 // resend everything, it's a new session
5798 service
.requeue_pg_temp();
5799 service
.send_pg_temp();
5802 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5806 map_lock
.put_read();
5808 send_beacon(ceph::coarse_mono_clock::now());
5812 // full map requests may happen while active or pre-boot
5813 if (requested_full_first
) {
5814 rerequest_full_maps();
5819 void OSD::ms_handle_fast_connect(Connection
*con
)
5821 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5822 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5823 Session
*s
= static_cast<Session
*>(con
->get_priv());
5825 s
= new Session(cct
);
5826 con
->set_priv(s
->get());
5828 dout(10) << " new session (outgoing) " << s
<< " con=" << s
->con
5829 << " addr=" << s
->con
->get_peer_addr() << dendl
;
5830 // we don't connect to clients
5831 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5832 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5838 void OSD::ms_handle_fast_accept(Connection
*con
)
5840 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5841 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5842 Session
*s
= static_cast<Session
*>(con
->get_priv());
5844 s
= new Session(cct
);
5845 con
->set_priv(s
->get());
5847 dout(10) << "new session (incoming)" << s
<< " con=" << con
5848 << " addr=" << con
->get_peer_addr()
5849 << " must have raced with connect" << dendl
;
5850 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5851 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5857 bool OSD::ms_handle_reset(Connection
*con
)
5859 Session
*session
= static_cast<Session
*>(con
->get_priv());
5860 dout(2) << "ms_handle_reset con " << con
<< " session " << session
<< dendl
;
5863 session
->wstate
.reset(con
);
5864 session
->con
.reset(NULL
); // break con <-> session ref cycle
5865 // note that we break session->con *before* the session_handle_reset
5866 // cleanup below. this avoids a race between us and
5867 // PG::add_backoff, Session::check_backoff, etc.
5868 session_handle_reset(session
);
5873 bool OSD::ms_handle_refused(Connection
*con
)
5875 if (!cct
->_conf
->osd_fast_fail_on_connection_refused
)
5878 Session
*session
= static_cast<Session
*>(con
->get_priv());
5879 dout(2) << "ms_handle_refused con " << con
<< " session " << session
<< dendl
;
5882 int type
= con
->get_peer_type();
5883 // handle only OSD failures here
5884 if (monc
&& (type
== CEPH_ENTITY_TYPE_OSD
)) {
5885 OSDMapRef osdmap
= get_osdmap();
5887 int id
= osdmap
->identify_osd_on_all_channels(con
->get_peer_addr());
5888 if (id
>= 0 && osdmap
->is_up(id
)) {
5889 // I'm cheating mon heartbeat grace logic, because we know it's not going
5890 // to respawn alone. +1 so we won't hit any boundary case.
5891 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(),
5892 osdmap
->get_inst(id
),
5893 cct
->_conf
->osd_heartbeat_grace
+ 1,
5894 osdmap
->get_epoch(),
5895 MOSDFailure::FLAG_IMMEDIATE
| MOSDFailure::FLAG_FAILED
5904 struct C_OSD_GetVersion
: public Context
{
5906 uint64_t oldest
, newest
;
5907 explicit C_OSD_GetVersion(OSD
*o
) : osd(o
), oldest(0), newest(0) {}
5908 void finish(int r
) override
{
5910 osd
->_got_mon_epochs(oldest
, newest
);
5914 void OSD::start_boot()
5916 if (!_is_healthy()) {
5917 // if we are not healthy, do not mark ourselves up (yet)
5918 dout(1) << "not healthy; waiting to boot" << dendl
;
5919 if (!is_waiting_for_healthy())
5920 start_waiting_for_healthy();
5921 // send pings sooner rather than later
5925 dout(1) << __func__
<< dendl
;
5926 set_state(STATE_PREBOOT
);
5927 waiting_for_luminous_mons
= false;
5928 dout(10) << "start_boot - have maps " << superblock
.oldest_map
5929 << ".." << superblock
.newest_map
<< dendl
;
5930 C_OSD_GetVersion
*c
= new C_OSD_GetVersion(this);
5931 monc
->get_version("osdmap", &c
->newest
, &c
->oldest
, c
);
5934 void OSD::_got_mon_epochs(epoch_t oldest
, epoch_t newest
)
5936 Mutex::Locker
l(osd_lock
);
5938 _preboot(oldest
, newest
);
5942 void OSD::_preboot(epoch_t oldest
, epoch_t newest
)
5944 assert(is_preboot());
5945 dout(10) << __func__
<< " _preboot mon has osdmaps "
5946 << oldest
<< ".." << newest
<< dendl
;
5948 // ensure our local fullness awareness is accurate
5951 // if our map within recent history, try to add ourselves to the osdmap.
5952 if (osdmap
->get_epoch() == 0) {
5953 derr
<< "waiting for initial osdmap" << dendl
;
5954 } else if (osdmap
->is_destroyed(whoami
)) {
5955 derr
<< "osdmap says I am destroyed" << dendl
;
5956 // provide a small margin so we don't livelock seeing if we
5957 // un-destroyed ourselves.
5958 if (osdmap
->get_epoch() > newest
- 1) {
5961 } else if (osdmap
->test_flag(CEPH_OSDMAP_NOUP
) || osdmap
->is_noup(whoami
)) {
5962 derr
<< "osdmap NOUP flag is set, waiting for it to clear" << dendl
;
5963 } else if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
)) {
5964 derr
<< "osdmap SORTBITWISE OSDMap flag is NOT set; please set it"
5966 } else if (osdmap
->require_osd_release
< CEPH_RELEASE_JEWEL
) {
5967 derr
<< "osdmap REQUIRE_JEWEL OSDMap flag is NOT set; please set it"
5969 } else if (!monc
->monmap
.get_required_features().contains_all(
5970 ceph::features::mon::FEATURE_LUMINOUS
)) {
5971 derr
<< "monmap REQUIRE_LUMINOUS is NOT set; must upgrade all monitors to "
5972 << "Luminous or later before Luminous OSDs will boot" << dendl
;
5973 waiting_for_luminous_mons
= true;
5974 } else if (service
.need_fullness_update()) {
5975 derr
<< "osdmap fullness state needs update" << dendl
;
5977 } else if (osdmap
->get_epoch() >= oldest
- 1 &&
5978 osdmap
->get_epoch() + cct
->_conf
->osd_map_message_max
> newest
) {
5983 // get all the latest maps
5984 if (osdmap
->get_epoch() + 1 >= oldest
)
5985 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5987 osdmap_subscribe(oldest
- 1, true);
5990 void OSD::send_full_update()
5992 if (!service
.need_fullness_update())
5995 if (service
.is_full()) {
5996 state
= CEPH_OSD_FULL
;
5997 } else if (service
.is_backfillfull()) {
5998 state
= CEPH_OSD_BACKFILLFULL
;
5999 } else if (service
.is_nearfull()) {
6000 state
= CEPH_OSD_NEARFULL
;
6003 OSDMap::calc_state_set(state
, s
);
6004 dout(10) << __func__
<< " want state " << s
<< dendl
;
6005 monc
->send_mon_message(new MOSDFull(osdmap
->get_epoch(), state
));
6008 void OSD::start_waiting_for_healthy()
6010 dout(1) << "start_waiting_for_healthy" << dendl
;
6011 set_state(STATE_WAITING_FOR_HEALTHY
);
6012 last_heartbeat_resample
= utime_t();
6014 // subscribe to osdmap updates, in case our peers really are known to be dead
6015 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
6018 bool OSD::_is_healthy()
6020 if (!cct
->get_heartbeat_map()->is_healthy()) {
6021 dout(1) << "is_healthy false -- internal heartbeat failed" << dendl
;
6025 if (is_waiting_for_healthy()) {
6026 Mutex::Locker
l(heartbeat_lock
);
6027 utime_t cutoff
= ceph_clock_now();
6028 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
6029 int num
= 0, up
= 0;
6030 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
6031 p
!= heartbeat_peers
.end();
6033 if (p
->second
.is_healthy(cutoff
))
6037 if ((float)up
< (float)num
* cct
->_conf
->osd_heartbeat_min_healthy_ratio
) {
6038 dout(1) << "is_healthy false -- only " << up
<< "/" << num
<< " up peers (less than "
6039 << int(cct
->_conf
->osd_heartbeat_min_healthy_ratio
* 100.0) << "%)" << dendl
;
6047 void OSD::_send_boot()
6049 dout(10) << "_send_boot" << dendl
;
6050 entity_addr_t cluster_addr
= cluster_messenger
->get_myaddr();
6051 Connection
*local_connection
= cluster_messenger
->get_loopback_connection().get();
6052 if (cluster_addr
.is_blank_ip()) {
6053 int port
= cluster_addr
.get_port();
6054 cluster_addr
= client_messenger
->get_myaddr();
6055 cluster_addr
.set_port(port
);
6056 cluster_messenger
->set_addr_unknowns(cluster_addr
);
6057 dout(10) << " assuming cluster_addr ip matches client_addr" << dendl
;
6059 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
6063 cluster_messenger
->ms_deliver_handle_fast_connect(local_connection
);
6066 entity_addr_t hb_back_addr
= hb_back_server_messenger
->get_myaddr();
6067 local_connection
= hb_back_server_messenger
->get_loopback_connection().get();
6068 if (hb_back_addr
.is_blank_ip()) {
6069 int port
= hb_back_addr
.get_port();
6070 hb_back_addr
= cluster_addr
;
6071 hb_back_addr
.set_port(port
);
6072 hb_back_server_messenger
->set_addr_unknowns(hb_back_addr
);
6073 dout(10) << " assuming hb_back_addr ip matches cluster_addr" << dendl
;
6075 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
6079 hb_back_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
6082 entity_addr_t hb_front_addr
= hb_front_server_messenger
->get_myaddr();
6083 local_connection
= hb_front_server_messenger
->get_loopback_connection().get();
6084 if (hb_front_addr
.is_blank_ip()) {
6085 int port
= hb_front_addr
.get_port();
6086 hb_front_addr
= client_messenger
->get_myaddr();
6087 hb_front_addr
.set_port(port
);
6088 hb_front_server_messenger
->set_addr_unknowns(hb_front_addr
);
6089 dout(10) << " assuming hb_front_addr ip matches client_addr" << dendl
;
6091 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
6095 hb_front_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
6098 MOSDBoot
*mboot
= new MOSDBoot(superblock
, get_osdmap_epoch(), service
.get_boot_epoch(),
6099 hb_back_addr
, hb_front_addr
, cluster_addr
,
6101 dout(10) << " client_addr " << client_messenger
->get_myaddr()
6102 << ", cluster_addr " << cluster_addr
6103 << ", hb_back_addr " << hb_back_addr
6104 << ", hb_front_addr " << hb_front_addr
6106 _collect_metadata(&mboot
->metadata
);
6107 monc
->send_mon_message(mboot
);
6108 set_state(STATE_BOOTING
);
6111 void OSD::_collect_metadata(map
<string
,string
> *pm
)
6114 (*pm
)["osd_data"] = dev_path
;
6115 if (store
->get_type() == "filestore") {
6116 // not applicable for bluestore
6117 (*pm
)["osd_journal"] = journal_path
;
6119 (*pm
)["front_addr"] = stringify(client_messenger
->get_myaddr());
6120 (*pm
)["back_addr"] = stringify(cluster_messenger
->get_myaddr());
6121 (*pm
)["hb_front_addr"] = stringify(hb_front_server_messenger
->get_myaddr());
6122 (*pm
)["hb_back_addr"] = stringify(hb_back_server_messenger
->get_myaddr());
6125 (*pm
)["osd_objectstore"] = store
->get_type();
6126 (*pm
)["rotational"] = store_is_rotational
? "1" : "0";
6127 (*pm
)["journal_rotational"] = journal_is_rotational
? "1" : "0";
6128 (*pm
)["default_device_class"] = store
->get_default_device_class();
6129 store
->collect_metadata(pm
);
6131 collect_sys_info(pm
, cct
);
6133 std::string front_iface
, back_iface
;
6136 CEPH_PICK_ADDRESS_PUBLIC | CEPH_PICK_ADDRESS_CLUSTER,
6137 &front_iface, &back_iface);
6139 (*pm
)["front_iface"] = pick_iface(cct
,
6140 client_messenger
->get_myaddr().get_sockaddr_storage());
6141 (*pm
)["back_iface"] = pick_iface(cct
,
6142 cluster_messenger
->get_myaddr().get_sockaddr_storage());
6144 dout(10) << __func__
<< " " << *pm
<< dendl
;
6147 void OSD::queue_want_up_thru(epoch_t want
)
6149 map_lock
.get_read();
6150 epoch_t cur
= osdmap
->get_up_thru(whoami
);
6151 Mutex::Locker
l(mon_report_lock
);
6152 if (want
> up_thru_wanted
) {
6153 dout(10) << "queue_want_up_thru now " << want
<< " (was " << up_thru_wanted
<< ")"
6154 << ", currently " << cur
6156 up_thru_wanted
= want
;
6159 dout(10) << "queue_want_up_thru want " << want
<< " <= queued " << up_thru_wanted
6160 << ", currently " << cur
6163 map_lock
.put_read();
6166 void OSD::send_alive()
6168 assert(mon_report_lock
.is_locked());
6169 if (!osdmap
->exists(whoami
))
6171 epoch_t up_thru
= osdmap
->get_up_thru(whoami
);
6172 dout(10) << "send_alive up_thru currently " << up_thru
<< " want " << up_thru_wanted
<< dendl
;
6173 if (up_thru_wanted
> up_thru
) {
6174 dout(10) << "send_alive want " << up_thru_wanted
<< dendl
;
6175 monc
->send_mon_message(new MOSDAlive(osdmap
->get_epoch(), up_thru_wanted
));
6179 void OSD::request_full_map(epoch_t first
, epoch_t last
)
6181 dout(10) << __func__
<< " " << first
<< ".." << last
6182 << ", previously requested "
6183 << requested_full_first
<< ".." << requested_full_last
<< dendl
;
6184 assert(osd_lock
.is_locked());
6185 assert(first
> 0 && last
> 0);
6186 assert(first
<= last
);
6187 assert(first
>= requested_full_first
); // we shouldn't ever ask for older maps
6188 if (requested_full_first
== 0) {
6190 requested_full_first
= first
;
6191 requested_full_last
= last
;
6192 } else if (last
<= requested_full_last
) {
6196 // additional request
6197 first
= requested_full_last
+ 1;
6198 requested_full_last
= last
;
6200 MMonGetOSDMap
*req
= new MMonGetOSDMap
;
6201 req
->request_full(first
, last
);
6202 monc
->send_mon_message(req
);
6205 void OSD::got_full_map(epoch_t e
)
6207 assert(requested_full_first
<= requested_full_last
);
6208 assert(osd_lock
.is_locked());
6209 if (requested_full_first
== 0) {
6210 dout(20) << __func__
<< " " << e
<< ", nothing requested" << dendl
;
6213 if (e
< requested_full_first
) {
6214 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6215 << ".." << requested_full_last
6216 << ", ignoring" << dendl
;
6219 if (e
>= requested_full_last
) {
6220 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6221 << ".." << requested_full_last
<< ", resetting" << dendl
;
6222 requested_full_first
= requested_full_last
= 0;
6226 requested_full_first
= e
+ 1;
6228 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6229 << ".." << requested_full_last
6230 << ", still need more" << dendl
;
6233 void OSD::requeue_failures()
6235 Mutex::Locker
l(heartbeat_lock
);
6236 unsigned old_queue
= failure_queue
.size();
6237 unsigned old_pending
= failure_pending
.size();
6238 for (map
<int,pair
<utime_t
,entity_inst_t
> >::iterator p
=
6239 failure_pending
.begin();
6240 p
!= failure_pending
.end(); ) {
6241 failure_queue
[p
->first
] = p
->second
.first
;
6242 failure_pending
.erase(p
++);
6244 dout(10) << __func__
<< " " << old_queue
<< " + " << old_pending
<< " -> "
6245 << failure_queue
.size() << dendl
;
6248 void OSD::send_failures()
6250 assert(map_lock
.is_locked());
6251 assert(mon_report_lock
.is_locked());
6252 Mutex::Locker
l(heartbeat_lock
);
6253 utime_t now
= ceph_clock_now();
6254 while (!failure_queue
.empty()) {
6255 int osd
= failure_queue
.begin()->first
;
6256 if (!failure_pending
.count(osd
)) {
6257 entity_inst_t i
= osdmap
->get_inst(osd
);
6258 int failed_for
= (int)(double)(now
- failure_queue
.begin()->second
);
6259 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(), i
, failed_for
,
6260 osdmap
->get_epoch()));
6261 failure_pending
[osd
] = make_pair(failure_queue
.begin()->second
, i
);
6263 failure_queue
.erase(osd
);
6267 void OSD::send_still_alive(epoch_t epoch
, const entity_inst_t
&i
)
6269 MOSDFailure
*m
= new MOSDFailure(monc
->get_fsid(), i
, 0, epoch
, MOSDFailure::FLAG_ALIVE
);
6270 monc
->send_mon_message(m
);
6273 void OSD::send_pg_stats(const utime_t
&now
)
6275 assert(map_lock
.is_locked());
6276 assert(osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
);
6277 dout(20) << "send_pg_stats" << dendl
;
6279 osd_stat_t cur_stat
= service
.get_osd_stat();
6281 cur_stat
.os_perf_stat
= store
->get_cur_stats();
6283 pg_stat_queue_lock
.Lock();
6285 if (osd_stat_updated
|| !pg_stat_queue
.empty()) {
6286 last_pg_stats_sent
= now
;
6287 osd_stat_updated
= false;
6289 dout(10) << "send_pg_stats - " << pg_stat_queue
.size() << " pgs updated" << dendl
;
6291 utime_t
had_for(now
);
6292 had_for
-= had_map_since
;
6294 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
6296 uint64_t tid
= ++pg_stat_tid
;
6298 m
->osd_stat
= cur_stat
;
6300 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6304 if (!pg
->is_primary()) { // we hold map_lock; role is stable.
6305 pg
->stat_queue_item
.remove_myself();
6306 pg
->put("pg_stat_queue");
6309 pg
->pg_stats_publish_lock
.Lock();
6310 if (pg
->pg_stats_publish_valid
) {
6311 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
6312 dout(25) << " sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6313 << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6315 dout(25) << " NOT sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6316 << pg
->pg_stats_publish
.reported_seq
<< ", not valid" << dendl
;
6318 pg
->pg_stats_publish_lock
.Unlock();
6321 if (last_pg_stats_ack
== utime_t() || !outstanding_pg_stats
.empty()) {
6322 last_pg_stats_ack
= ceph_clock_now();
6324 outstanding_pg_stats
.insert(tid
);
6325 dout(20) << __func__
<< " updates pending: " << outstanding_pg_stats
<< dendl
;
6327 monc
->send_mon_message(m
);
6330 pg_stat_queue_lock
.Unlock();
6333 void OSD::handle_pg_stats_ack(MPGStatsAck
*ack
)
6335 dout(10) << "handle_pg_stats_ack " << dendl
;
6337 if (!require_mon_peer(ack
)) {
6342 // NOTE: we may get replies from a previous mon even while
6343 // outstanding_pg_stats is empty if reconnecting races with replies
6346 pg_stat_queue_lock
.Lock();
6348 last_pg_stats_ack
= ceph_clock_now();
6350 // decay timeout slowly (analogous to TCP)
6352 MAX(cct
->_conf
->osd_mon_ack_timeout
,
6353 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_decay
);
6354 dout(20) << __func__
<< " timeout now " << stats_ack_timeout
<< dendl
;
6356 if (ack
->get_tid() > pg_stat_tid_flushed
) {
6357 pg_stat_tid_flushed
= ack
->get_tid();
6358 pg_stat_queue_cond
.Signal();
6361 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6367 auto acked
= ack
->pg_stat
.find(pg
->info
.pgid
.pgid
);
6368 if (acked
!= ack
->pg_stat
.end()) {
6369 pg
->pg_stats_publish_lock
.Lock();
6370 if (acked
->second
.first
== pg
->pg_stats_publish
.reported_seq
&&
6371 acked
->second
.second
== pg
->pg_stats_publish
.reported_epoch
) {
6372 dout(25) << " ack on " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6373 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6374 pg
->stat_queue_item
.remove_myself();
6375 pg
->put("pg_stat_queue");
6377 dout(25) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6378 << ":" << pg
->pg_stats_publish
.reported_seq
<< " > acked "
6379 << acked
->second
<< dendl
;
6381 pg
->pg_stats_publish_lock
.Unlock();
6383 dout(30) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6384 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6388 outstanding_pg_stats
.erase(ack
->get_tid());
6389 dout(20) << __func__
<< " still pending: " << outstanding_pg_stats
<< dendl
;
6391 pg_stat_queue_lock
.Unlock();
6396 void OSD::flush_pg_stats()
6398 dout(10) << "flush_pg_stats" << dendl
;
6400 utime_t now
= ceph_clock_now();
6401 map_lock
.get_read();
6402 mon_report_lock
.Lock();
6404 mon_report_lock
.Unlock();
6405 map_lock
.put_read();
6408 pg_stat_queue_lock
.Lock();
6409 uint64_t tid
= pg_stat_tid
;
6410 dout(10) << "flush_pg_stats waiting for stats tid " << tid
<< " to flush" << dendl
;
6411 while (tid
> pg_stat_tid_flushed
)
6412 pg_stat_queue_cond
.Wait(pg_stat_queue_lock
);
6413 dout(10) << "flush_pg_stats finished waiting for stats tid " << tid
<< " to flush" << dendl
;
6414 pg_stat_queue_lock
.Unlock();
6419 void OSD::send_beacon(const ceph::coarse_mono_clock::time_point
& now
)
6421 const auto& monmap
= monc
->monmap
;
6422 // send beacon to mon even if we are just connected, and the monmap is not
6423 // initialized yet by then.
6424 if (monmap
.epoch
> 0 &&
6425 monmap
.get_required_features().contains_all(
6426 ceph::features::mon::FEATURE_LUMINOUS
)) {
6427 dout(20) << __func__
<< " sending" << dendl
;
6428 MOSDBeacon
* beacon
= nullptr;
6430 Mutex::Locker l
{min_last_epoch_clean_lock
};
6431 beacon
= new MOSDBeacon(osdmap
->get_epoch(), min_last_epoch_clean
);
6432 std::swap(beacon
->pgs
, min_last_epoch_clean_pgs
);
6433 last_sent_beacon
= now
;
6435 monc
->send_mon_message(beacon
);
6437 dout(20) << __func__
<< " not sending" << dendl
;
6441 void OSD::handle_command(MMonCommand
*m
)
6443 if (!require_mon_peer(m
)) {
6448 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), NULL
);
6449 command_wq
.queue(c
);
6453 void OSD::handle_command(MCommand
*m
)
6455 ConnectionRef con
= m
->get_connection();
6456 Session
*session
= static_cast<Session
*>(con
->get_priv());
6458 con
->send_message(new MCommandReply(m
, -EPERM
));
6463 OSDCap
& caps
= session
->caps
;
6466 if (!caps
.allow_all() || m
->get_source().is_mon()) {
6467 con
->send_message(new MCommandReply(m
, -EPERM
));
6472 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), con
.get());
6473 command_wq
.queue(c
);
6483 string availability
;
6484 } osd_commands
[] = {
6486 #define COMMAND(parsesig, helptext, module, perm, availability) \
6487 {parsesig, helptext, module, perm, availability},
6489 // yes, these are really pg commands, but there's a limit to how
6490 // much work it's worth. The OSD returns all of them. Make this
6491 // form (pg <pgid> <cmd>) valid only for the cli.
6492 // Rest uses "tell <pgid> <cmd>"
6495 "name=pgid,type=CephPgid " \
6496 "name=cmd,type=CephChoices,strings=query", \
6497 "show details of a specific pg", "osd", "r", "cli")
6499 "name=pgid,type=CephPgid " \
6500 "name=cmd,type=CephChoices,strings=mark_unfound_lost " \
6501 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6502 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6505 "name=pgid,type=CephPgid " \
6506 "name=cmd,type=CephChoices,strings=list_missing " \
6507 "name=offset,type=CephString,req=false",
6508 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6511 // new form: tell <pgid> <cmd> for both cli and rest
6514 "show details of a specific pg", "osd", "r", "cli,rest")
6515 COMMAND("mark_unfound_lost " \
6516 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6517 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6518 "osd", "rw", "cli,rest")
6519 COMMAND("list_missing " \
6520 "name=offset,type=CephString,req=false",
6521 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6522 "osd", "r", "cli,rest")
6523 COMMAND("perf histogram dump "
6524 "name=logger,type=CephString,req=false "
6525 "name=counter,type=CephString,req=false",
6526 "Get histogram data",
6527 "osd", "r", "cli,rest")
6529 // tell <osd.n> commands. Validation of osd.n must be special-cased in client
6530 COMMAND("version", "report version of OSD", "osd", "r", "cli,rest")
6531 COMMAND("get_command_descriptions", "list commands descriptions", "osd", "r", "cli,rest")
6532 COMMAND("injectargs " \
6533 "name=injected_args,type=CephString,n=N",
6534 "inject configuration arguments into running OSD",
6535 "osd", "rw", "cli,rest")
6536 COMMAND("config set " \
6537 "name=key,type=CephString name=value,type=CephString",
6538 "Set a configuration option at runtime (not persistent)",
6539 "osd", "rw", "cli,rest")
6540 COMMAND("cluster_log " \
6541 "name=level,type=CephChoices,strings=error,warning,info,debug " \
6542 "name=message,type=CephString,n=N",
6543 "log a message to the cluster log",
6544 "osd", "rw", "cli,rest")
6546 "name=count,type=CephInt,req=false " \
6547 "name=size,type=CephInt,req=false " \
6548 "name=object_size,type=CephInt,req=false " \
6549 "name=object_num,type=CephInt,req=false ", \
6550 "OSD benchmark: write <count> <size>-byte objects, " \
6551 "(default 1G size 4MB). Results in log.",
6552 "osd", "rw", "cli,rest")
6553 COMMAND("flush_pg_stats", "flush pg stats", "osd", "rw", "cli,rest")
6555 "name=heapcmd,type=CephChoices,strings=dump|start_profiler|stop_profiler|release|stats", \
6556 "show heap usage info (available only if compiled with tcmalloc)", \
6557 "osd", "rw", "cli,rest")
6558 COMMAND("debug dump_missing " \
6559 "name=filename,type=CephFilepath",
6560 "dump missing objects to a named file", "osd", "r", "cli,rest")
6561 COMMAND("debug kick_recovery_wq " \
6562 "name=delay,type=CephInt,range=0",
6563 "set osd_recovery_delay_start to <val>", "osd", "rw", "cli,rest")
6564 COMMAND("cpu_profiler " \
6565 "name=arg,type=CephChoices,strings=status|flush",
6566 "run cpu profiling on daemon", "osd", "rw", "cli,rest")
6567 COMMAND("dump_pg_recovery_stats", "dump pg recovery statistics",
6568 "osd", "r", "cli,rest")
6569 COMMAND("reset_pg_recovery_stats", "reset pg recovery statistics",
6570 "osd", "rw", "cli,rest")
6572 "compact object store's omap. "
6573 "WARNING: Compaction probably slows your requests",
6574 "osd", "rw", "cli,rest")
6577 void OSD::do_command(Connection
*con
, ceph_tid_t tid
, vector
<string
>& cmd
, bufferlist
& data
)
6580 stringstream ss
, ds
;
6584 dout(20) << "do_command tid " << tid
<< " " << cmd
<< dendl
;
6586 map
<string
, cmd_vartype
> cmdmap
;
6590 boost::scoped_ptr
<Formatter
> f
;
6593 ss
<< "no command given";
6597 if (!cmdmap_from_json(cmd
, &cmdmap
, ss
)) {
6602 cmd_getval(cct
, cmdmap
, "prefix", prefix
);
6604 if (prefix
== "get_command_descriptions") {
6606 JSONFormatter
*f
= new JSONFormatter();
6607 f
->open_object_section("command_descriptions");
6608 for (OSDCommand
*cp
= osd_commands
;
6609 cp
< &osd_commands
[ARRAY_SIZE(osd_commands
)]; cp
++) {
6611 ostringstream secname
;
6612 secname
<< "cmd" << setfill('0') << std::setw(3) << cmdnum
;
6613 dump_cmddesc_to_json(f
, secname
.str(), cp
->cmdstring
, cp
->helpstring
,
6614 cp
->module
, cp
->perm
, cp
->availability
, 0);
6617 f
->close_section(); // command_descriptions
6624 cmd_getval(cct
, cmdmap
, "format", format
);
6625 f
.reset(Formatter::create(format
));
6627 if (prefix
== "version") {
6629 f
->open_object_section("version");
6630 f
->dump_string("version", pretty_version_to_str());
6634 ds
<< pretty_version_to_str();
6638 else if (prefix
== "injectargs") {
6639 vector
<string
> argsvec
;
6640 cmd_getval(cct
, cmdmap
, "injected_args", argsvec
);
6642 if (argsvec
.empty()) {
6644 ss
<< "ignoring empty injectargs";
6647 string args
= argsvec
.front();
6648 for (vector
<string
>::iterator a
= ++argsvec
.begin(); a
!= argsvec
.end(); ++a
)
6651 r
= cct
->_conf
->injectargs(args
, &ss
);
6654 else if (prefix
== "config set") {
6657 cmd_getval(cct
, cmdmap
, "key", key
);
6658 cmd_getval(cct
, cmdmap
, "value", val
);
6660 r
= cct
->_conf
->set_val(key
, val
, true, &ss
);
6662 cct
->_conf
->apply_changes(nullptr);
6666 else if (prefix
== "cluster_log") {
6668 cmd_getval(cct
, cmdmap
, "message", msg
);
6671 ss
<< "ignoring empty log message";
6674 string message
= msg
.front();
6675 for (vector
<string
>::iterator a
= ++msg
.begin(); a
!= msg
.end(); ++a
)
6676 message
+= " " + *a
;
6678 cmd_getval(cct
, cmdmap
, "level", lvl
);
6679 clog_type level
= string_to_clog_type(lvl
);
6682 ss
<< "unknown level '" << lvl
<< "'";
6685 clog
->do_log(level
, message
);
6688 // either 'pg <pgid> <command>' or
6689 // 'tell <pgid>' (which comes in without any of that prefix)?
6691 else if (prefix
== "pg" ||
6692 prefix
== "query" ||
6693 prefix
== "mark_unfound_lost" ||
6694 prefix
== "list_missing"
6698 if (!cmd_getval(cct
, cmdmap
, "pgid", pgidstr
)) {
6699 ss
<< "no pgid specified";
6701 } else if (!pgid
.parse(pgidstr
.c_str())) {
6702 ss
<< "couldn't parse pgid '" << pgidstr
<< "'";
6707 if (osdmap
->get_primary_shard(pgid
, &pcand
) &&
6708 (pg
= _lookup_lock_pg(pcand
))) {
6709 if (pg
->is_primary()) {
6710 // simulate pg <pgid> cmd= for pg->do-command
6712 cmd_putval(cct
, cmdmap
, "cmd", prefix
);
6713 r
= pg
->do_command(cmdmap
, ss
, data
, odata
, con
, tid
);
6716 // don't reply, pg will do so async
6720 ss
<< "not primary for pgid " << pgid
;
6722 // send them the latest diff to ensure they realize the mapping
6724 service
.send_incremental_map(osdmap
->get_epoch() - 1, con
, osdmap
);
6726 // do not reply; they will get newer maps and realize they
6733 ss
<< "i don't have pgid " << pgid
;
6739 else if (prefix
== "bench") {
6742 int64_t osize
, onum
;
6743 // default count 1G, size 4MB
6744 cmd_getval(cct
, cmdmap
, "count", count
, (int64_t)1 << 30);
6745 cmd_getval(cct
, cmdmap
, "size", bsize
, (int64_t)4 << 20);
6746 cmd_getval(cct
, cmdmap
, "object_size", osize
, (int64_t)0);
6747 cmd_getval(cct
, cmdmap
, "object_num", onum
, (int64_t)0);
6749 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
6750 ObjectStore::Sequencer
>("bench"));
6752 uint32_t duration
= cct
->_conf
->osd_bench_duration
;
6754 if (bsize
> (int64_t) cct
->_conf
->osd_bench_max_block_size
) {
6755 // let us limit the block size because the next checks rely on it
6756 // having a sane value. If we allow any block size to be set things
6757 // can still go sideways.
6758 ss
<< "block 'size' values are capped at "
6759 << prettybyte_t(cct
->_conf
->osd_bench_max_block_size
) << ". If you wish to use"
6760 << " a higher value, please adjust 'osd_bench_max_block_size'";
6763 } else if (bsize
< (int64_t) (1 << 20)) {
6764 // entering the realm of small block sizes.
6765 // limit the count to a sane value, assuming a configurable amount of
6766 // IOPS and duration, so that the OSD doesn't get hung up on this,
6767 // preventing timeouts from going off
6769 bsize
* duration
* cct
->_conf
->osd_bench_small_size_max_iops
;
6770 if (count
> max_count
) {
6771 ss
<< "'count' values greater than " << max_count
6772 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6773 << cct
->_conf
->osd_bench_small_size_max_iops
<< " IOPS,"
6774 << " for " << duration
<< " seconds,"
6775 << " can cause ill effects on osd. "
6776 << " Please adjust 'osd_bench_small_size_max_iops' with a higher"
6777 << " value if you wish to use a higher 'count'.";
6782 // 1MB block sizes are big enough so that we get more stuff done.
6783 // However, to avoid the osd from getting hung on this and having
6784 // timers being triggered, we are going to limit the count assuming
6785 // a configurable throughput and duration.
6786 // NOTE: max_count is the total amount of bytes that we believe we
6787 // will be able to write during 'duration' for the given
6788 // throughput. The block size hardly impacts this unless it's
6789 // way too big. Given we already check how big the block size
6790 // is, it's safe to assume everything will check out.
6792 cct
->_conf
->osd_bench_large_size_max_throughput
* duration
;
6793 if (count
> max_count
) {
6794 ss
<< "'count' values greater than " << max_count
6795 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6796 << prettybyte_t(cct
->_conf
->osd_bench_large_size_max_throughput
) << "/s,"
6797 << " for " << duration
<< " seconds,"
6798 << " can cause ill effects on osd. "
6799 << " Please adjust 'osd_bench_large_size_max_throughput'"
6800 << " with a higher value if you wish to use a higher 'count'.";
6806 if (osize
&& bsize
> osize
)
6809 dout(1) << " bench count " << count
6810 << " bsize " << prettybyte_t(bsize
) << dendl
;
6812 ObjectStore::Transaction cleanupt
;
6814 if (osize
&& onum
) {
6816 bufferptr
bp(osize
);
6818 bl
.push_back(std::move(bp
));
6819 bl
.rebuild_page_aligned();
6820 for (int i
=0; i
<onum
; ++i
) {
6822 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", i
);
6824 hobject_t
soid(sobject_t(oid
, 0));
6825 ObjectStore::Transaction t
;
6826 t
.write(coll_t(), ghobject_t(soid
), 0, osize
, bl
);
6827 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6828 cleanupt
.remove(coll_t(), ghobject_t(soid
));
6833 bufferptr
bp(bsize
);
6835 bl
.push_back(std::move(bp
));
6836 bl
.rebuild_page_aligned();
6840 if (!osr
->flush_commit(&waiter
)) {
6845 utime_t start
= ceph_clock_now();
6846 for (int64_t pos
= 0; pos
< count
; pos
+= bsize
) {
6848 unsigned offset
= 0;
6849 if (onum
&& osize
) {
6850 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", (int)(rand() % onum
));
6851 offset
= rand() % (osize
/ bsize
) * bsize
;
6853 snprintf(nm
, sizeof(nm
), "disk_bw_test_%lld", (long long)pos
);
6856 hobject_t
soid(sobject_t(oid
, 0));
6857 ObjectStore::Transaction t
;
6858 t
.write(coll_t::meta(), ghobject_t(soid
), offset
, bsize
, bl
);
6859 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6860 if (!onum
|| !osize
)
6861 cleanupt
.remove(coll_t::meta(), ghobject_t(soid
));
6866 if (!osr
->flush_commit(&waiter
)) {
6870 utime_t end
= ceph_clock_now();
6873 store
->queue_transaction(osr
.get(), std::move(cleanupt
), NULL
);
6876 if (!osr
->flush_commit(&waiter
)) {
6881 uint64_t rate
= (double)count
/ (end
- start
);
6883 f
->open_object_section("osd_bench_results");
6884 f
->dump_int("bytes_written", count
);
6885 f
->dump_int("blocksize", bsize
);
6886 f
->dump_unsigned("bytes_per_sec", rate
);
6890 ss
<< "bench: wrote " << prettybyte_t(count
)
6891 << " in blocks of " << prettybyte_t(bsize
) << " in "
6892 << (end
-start
) << " sec at " << prettybyte_t(rate
) << "/sec";
6896 else if (prefix
== "flush_pg_stats") {
6897 if (osdmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
6898 mgrc
.send_pgstats();
6899 ds
<< service
.get_osd_stat_seq() << "\n";
6905 else if (prefix
== "heap") {
6906 r
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ds
);
6909 else if (prefix
== "debug dump_missing") {
6911 cmd_getval(cct
, cmdmap
, "filename", file_name
);
6912 std::ofstream
fout(file_name
.c_str());
6913 if (!fout
.is_open()) {
6914 ss
<< "failed to open file '" << file_name
<< "'";
6919 fout
<< "*** osd " << whoami
<< ": dump_missing ***" << std::endl
;
6920 RWLock::RLocker
l(pg_map_lock
);
6921 for (ceph::unordered_map
<spg_t
, PG
*>::const_iterator pg_map_e
= pg_map
.begin();
6922 pg_map_e
!= pg_map
.end(); ++pg_map_e
) {
6923 PG
*pg
= pg_map_e
->second
;
6926 fout
<< *pg
<< std::endl
;
6927 std::map
<hobject_t
, pg_missing_item
>::const_iterator mend
=
6928 pg
->pg_log
.get_missing().get_items().end();
6929 std::map
<hobject_t
, pg_missing_item
>::const_iterator mi
=
6930 pg
->pg_log
.get_missing().get_items().begin();
6931 for (; mi
!= mend
; ++mi
) {
6932 fout
<< mi
->first
<< " -> " << mi
->second
<< std::endl
;
6933 if (!pg
->missing_loc
.needs_recovery(mi
->first
))
6935 if (pg
->missing_loc
.is_unfound(mi
->first
))
6936 fout
<< " unfound ";
6937 const set
<pg_shard_t
> &mls(pg
->missing_loc
.get_locations(mi
->first
));
6940 fout
<< "missing_loc: " << mls
<< std::endl
;
6948 else if (prefix
== "debug kick_recovery_wq") {
6950 cmd_getval(cct
, cmdmap
, "delay", delay
);
6953 r
= cct
->_conf
->set_val("osd_recovery_delay_start", oss
.str().c_str());
6955 ss
<< "kick_recovery_wq: error setting "
6956 << "osd_recovery_delay_start to '" << delay
<< "': error "
6960 cct
->_conf
->apply_changes(NULL
);
6961 ss
<< "kicking recovery queue. set osd_recovery_delay_start "
6962 << "to " << cct
->_conf
->osd_recovery_delay_start
;
6965 else if (prefix
== "cpu_profiler") {
6967 cmd_getval(cct
, cmdmap
, "arg", arg
);
6968 vector
<string
> argvec
;
6969 get_str_vec(arg
, argvec
);
6970 cpu_profiler_handle_command(argvec
, ds
);
6973 else if (prefix
== "dump_pg_recovery_stats") {
6976 pg_recovery_stats
.dump_formatted(f
.get());
6979 pg_recovery_stats
.dump(s
);
6980 ds
<< "dump pg recovery stats: " << s
.str();
6984 else if (prefix
== "reset_pg_recovery_stats") {
6985 ss
<< "reset pg recovery stats";
6986 pg_recovery_stats
.reset();
6989 else if (prefix
== "perf histogram dump") {
6991 std::string counter
;
6992 cmd_getval(cct
, cmdmap
, "logger", logger
);
6993 cmd_getval(cct
, cmdmap
, "counter", counter
);
6995 cct
->get_perfcounters_collection()->dump_formatted_histograms(
6996 f
.get(), false, logger
, counter
);
7001 else if (prefix
== "compact") {
7002 dout(1) << "triggering manual compaction" << dendl
;
7003 auto start
= ceph::coarse_mono_clock::now();
7005 auto end
= ceph::coarse_mono_clock::now();
7006 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
7007 dout(1) << "finished manual compaction in "
7008 << time_span
.count()
7009 << " seconds" << dendl
;
7010 ss
<< "compacted omap in " << time_span
.count() << " seconds";
7014 ss
<< "unrecognized command! " << cmd
;
7021 dout(0) << "do_command r=" << r
<< " " << rs
<< dendl
;
7024 MCommandReply
*reply
= new MCommandReply(r
, rs
);
7025 reply
->set_tid(tid
);
7026 reply
->set_data(odata
);
7027 con
->send_message(reply
);
7031 bool OSD::heartbeat_dispatch(Message
*m
)
7033 dout(30) << "heartbeat_dispatch " << m
<< dendl
;
7034 switch (m
->get_type()) {
7037 dout(10) << "ping from " << m
->get_source_inst() << dendl
;
7042 handle_osd_ping(static_cast<MOSDPing
*>(m
));
7046 dout(0) << "dropping unexpected message " << *m
<< " from " << m
->get_source_inst() << dendl
;
7053 bool OSD::ms_dispatch(Message
*m
)
7055 dout(20) << "OSD::ms_dispatch: " << *m
<< dendl
;
7056 if (m
->get_type() == MSG_OSD_MARK_ME_DOWN
) {
7057 service
.got_stop_ack();
7065 if (is_stopping()) {
7079 void OSD::maybe_share_map(
7084 if (!op
->check_send_map
) {
7087 epoch_t last_sent_epoch
= 0;
7089 session
->sent_epoch_lock
.lock();
7090 last_sent_epoch
= session
->last_sent_epoch
;
7091 session
->sent_epoch_lock
.unlock();
7093 const Message
*m
= op
->get_req();
7096 m
->get_connection().get(),
7099 session
? &last_sent_epoch
: NULL
);
7101 session
->sent_epoch_lock
.lock();
7102 if (session
->last_sent_epoch
< last_sent_epoch
) {
7103 session
->last_sent_epoch
= last_sent_epoch
;
7105 session
->sent_epoch_lock
.unlock();
7107 op
->check_send_map
= false;
7110 void OSD::dispatch_session_waiting(Session
*session
, OSDMapRef osdmap
)
7112 assert(session
->session_dispatch_lock
.is_locked());
7114 auto i
= session
->waiting_on_map
.begin();
7115 while (i
!= session
->waiting_on_map
.end()) {
7116 OpRequestRef op
= &(*i
);
7117 assert(ms_can_fast_dispatch(op
->get_req()));
7118 const MOSDFastDispatchOp
*m
= static_cast<const MOSDFastDispatchOp
*>(
7120 if (m
->get_min_epoch() > osdmap
->get_epoch()) {
7123 session
->waiting_on_map
.erase(i
++);
7127 if (m
->get_type() == CEPH_MSG_OSD_OP
) {
7128 pg_t actual_pgid
= osdmap
->raw_pg_to_pg(
7129 static_cast<const MOSDOp
*>(m
)->get_pg());
7130 if (!osdmap
->get_primary_shard(actual_pgid
, &pgid
)) {
7134 pgid
= m
->get_spg();
7136 enqueue_op(pgid
, op
, m
->get_map_epoch());
7139 if (session
->waiting_on_map
.empty()) {
7140 clear_session_waiting_on_map(session
);
7142 register_session_waiting_on_map(session
);
7146 void OSD::ms_fast_dispatch(Message
*m
)
7149 if (service
.is_stopping()) {
7153 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
7156 osd_reqid_t reqid
= op
->get_reqid();
7158 tracepoint(osd
, ms_fast_dispatch
, reqid
.name
._type
,
7159 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
7163 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
7165 // note sender epoch, min req'd epoch
7166 op
->sent_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch();
7167 op
->min_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_min_epoch();
7168 assert(op
->min_epoch
<= op
->sent_epoch
); // sanity check!
7170 service
.maybe_inject_dispatch_delay();
7172 if (m
->get_connection()->has_features(CEPH_FEATUREMASK_RESEND_ON_SPLIT
) ||
7173 m
->get_type() != CEPH_MSG_OSD_OP
) {
7174 // queue it directly
7176 static_cast<MOSDFastDispatchOp
*>(m
)->get_spg(),
7178 static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch());
7180 // legacy client, and this is an MOSDOp (the *only* fast dispatch
7181 // message that didn't have an explicit spg_t); we need to map
7182 // them to an spg_t while preserving delivery order.
7183 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
7186 Mutex::Locker
l(session
->session_dispatch_lock
);
7188 session
->waiting_on_map
.push_back(*op
);
7189 OSDMapRef nextmap
= service
.get_nextmap_reserved();
7190 dispatch_session_waiting(session
, nextmap
);
7191 service
.release_map(nextmap
);
7196 OID_EVENT_TRACE_WITH_MSG(m
, "MS_FAST_DISPATCH_END", false);
7199 void OSD::ms_fast_preprocess(Message
*m
)
7201 if (m
->get_connection()->get_peer_type() == CEPH_ENTITY_TYPE_OSD
) {
7202 if (m
->get_type() == CEPH_MSG_OSD_MAP
) {
7203 MOSDMap
*mm
= static_cast<MOSDMap
*>(m
);
7204 Session
*s
= static_cast<Session
*>(m
->get_connection()->get_priv());
7206 s
->received_map_lock
.lock();
7207 s
->received_map_epoch
= mm
->get_last();
7208 s
->received_map_lock
.unlock();
7215 bool OSD::ms_get_authorizer(int dest_type
, AuthAuthorizer
**authorizer
, bool force_new
)
7217 dout(10) << "OSD::ms_get_authorizer type=" << ceph_entity_type_name(dest_type
) << dendl
;
7219 if (is_stopping()) {
7220 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7224 if (dest_type
== CEPH_ENTITY_TYPE_MON
)
7228 /* the MonClient checks keys every tick(), so we should just wait for that cycle
7230 if (monc
->wait_auth_rotating(10) < 0) {
7231 derr
<< "OSD::ms_get_authorizer wait_auth_rotating failed" << dendl
;
7236 *authorizer
= monc
->build_authorizer(dest_type
);
7237 return *authorizer
!= NULL
;
7241 bool OSD::ms_verify_authorizer(
7242 Connection
*con
, int peer_type
,
7243 int protocol
, bufferlist
& authorizer_data
, bufferlist
& authorizer_reply
,
7244 bool& isvalid
, CryptoKey
& session_key
,
7245 std::unique_ptr
<AuthAuthorizerChallenge
> *challenge
)
7247 AuthAuthorizeHandler
*authorize_handler
= 0;
7248 switch (peer_type
) {
7249 case CEPH_ENTITY_TYPE_MDS
:
7251 * note: mds is technically a client from our perspective, but
7252 * this makes the 'cluster' consistent w/ monitor's usage.
7254 case CEPH_ENTITY_TYPE_OSD
:
7255 case CEPH_ENTITY_TYPE_MGR
:
7256 authorize_handler
= authorize_handler_cluster_registry
->get_handler(protocol
);
7259 authorize_handler
= authorize_handler_service_registry
->get_handler(protocol
);
7261 if (!authorize_handler
) {
7262 dout(0) << "No AuthAuthorizeHandler found for protocol " << protocol
<< dendl
;
7267 AuthCapsInfo caps_info
;
7270 uint64_t auid
= CEPH_AUTH_UID_DEFAULT
;
7272 RotatingKeyRing
*keys
= monc
->rotating_secrets
.get();
7274 isvalid
= authorize_handler
->verify_authorizer(
7276 authorizer_data
, authorizer_reply
, name
, global_id
, caps_info
, session_key
,
7279 dout(10) << __func__
<< " no rotating_keys (yet), denied" << dendl
;
7284 Session
*s
= static_cast<Session
*>(con
->get_priv());
7286 s
= new Session(cct
);
7287 con
->set_priv(s
->get());
7289 dout(10) << " new session " << s
<< " con=" << s
->con
<< " addr=" << s
->con
->get_peer_addr() << dendl
;
7292 s
->entity_name
= name
;
7293 if (caps_info
.allow_all
)
7294 s
->caps
.set_allow_all();
7297 if (caps_info
.caps
.length() > 0) {
7298 bufferlist::iterator p
= caps_info
.caps
.begin();
7303 catch (buffer::error
& e
) {
7305 bool success
= s
->caps
.parse(str
);
7307 dout(10) << " session " << s
<< " " << s
->entity_name
<< " has caps " << s
->caps
<< " '" << str
<< "'" << dendl
;
7309 dout(10) << " session " << s
<< " " << s
->entity_name
<< " failed to parse caps '" << str
<< "'" << dendl
;
7317 void OSD::do_waiters()
7319 assert(osd_lock
.is_locked());
7321 dout(10) << "do_waiters -- start" << dendl
;
7322 while (!finished
.empty()) {
7323 OpRequestRef next
= finished
.front();
7324 finished
.pop_front();
7327 dout(10) << "do_waiters -- finish" << dendl
;
7330 void OSD::dispatch_op(OpRequestRef op
)
7332 switch (op
->get_req()->get_type()) {
7334 case MSG_OSD_PG_CREATE
:
7335 handle_pg_create(op
);
7337 case MSG_OSD_PG_NOTIFY
:
7338 handle_pg_notify(op
);
7340 case MSG_OSD_PG_QUERY
:
7341 handle_pg_query(op
);
7343 case MSG_OSD_PG_LOG
:
7346 case MSG_OSD_PG_REMOVE
:
7347 handle_pg_remove(op
);
7349 case MSG_OSD_PG_INFO
:
7352 case MSG_OSD_PG_TRIM
:
7355 case MSG_OSD_BACKFILL_RESERVE
:
7356 handle_pg_backfill_reserve(op
);
7358 case MSG_OSD_RECOVERY_RESERVE
:
7359 handle_pg_recovery_reserve(op
);
7364 void OSD::_dispatch(Message
*m
)
7366 assert(osd_lock
.is_locked());
7367 dout(20) << "_dispatch " << m
<< " " << *m
<< dendl
;
7369 switch (m
->get_type()) {
7371 // -- don't need lock --
7373 dout(10) << "ping from " << m
->get_source() << dendl
;
7377 // -- don't need OSDMap --
7379 // map and replication
7380 case CEPH_MSG_OSD_MAP
:
7381 handle_osd_map(static_cast<MOSDMap
*>(m
));
7385 case MSG_PGSTATSACK
:
7386 handle_pg_stats_ack(static_cast<MPGStatsAck
*>(m
));
7389 case MSG_MON_COMMAND
:
7390 handle_command(static_cast<MMonCommand
*>(m
));
7393 handle_command(static_cast<MCommand
*>(m
));
7397 handle_scrub(static_cast<MOSDScrub
*>(m
));
7400 case MSG_OSD_FORCE_RECOVERY
:
7401 handle_force_recovery(m
);
7404 // -- need OSDMap --
7406 case MSG_OSD_PG_CREATE
:
7407 case MSG_OSD_PG_NOTIFY
:
7408 case MSG_OSD_PG_QUERY
:
7409 case MSG_OSD_PG_LOG
:
7410 case MSG_OSD_PG_REMOVE
:
7411 case MSG_OSD_PG_INFO
:
7412 case MSG_OSD_PG_TRIM
:
7413 case MSG_OSD_BACKFILL_RESERVE
:
7414 case MSG_OSD_RECOVERY_RESERVE
:
7416 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
7418 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
7419 // no map? starting up?
7421 dout(7) << "no OSDMap, not booted" << dendl
;
7422 logger
->inc(l_osd_waiting_for_map
);
7423 waiting_for_osdmap
.push_back(op
);
7424 op
->mark_delayed("no osdmap");
7434 void OSD::handle_pg_scrub(MOSDScrub
*m
, PG
*pg
)
7437 if (pg
->is_primary()) {
7438 pg
->unreg_next_scrub();
7439 pg
->scrubber
.must_scrub
= true;
7440 pg
->scrubber
.must_deep_scrub
= m
->deep
|| m
->repair
;
7441 pg
->scrubber
.must_repair
= m
->repair
;
7442 pg
->reg_next_scrub();
7443 dout(10) << "marking " << *pg
<< " for scrub" << dendl
;
7448 void OSD::handle_scrub(MOSDScrub
*m
)
7450 dout(10) << "handle_scrub " << *m
<< dendl
;
7451 if (!require_mon_or_mgr_peer(m
)) {
7455 if (m
->fsid
!= monc
->get_fsid()) {
7456 dout(0) << "handle_scrub fsid " << m
->fsid
<< " != " << monc
->get_fsid() << dendl
;
7461 RWLock::RLocker
l(pg_map_lock
);
7462 if (m
->scrub_pgs
.empty()) {
7463 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
7466 handle_pg_scrub(m
, p
->second
);
7468 for (vector
<pg_t
>::iterator p
= m
->scrub_pgs
.begin();
7469 p
!= m
->scrub_pgs
.end();
7472 if (osdmap
->get_primary_shard(*p
, &pcand
)) {
7473 auto pg_map_entry
= pg_map
.find(pcand
);
7474 if (pg_map_entry
!= pg_map
.end()) {
7475 handle_pg_scrub(m
, pg_map_entry
->second
);
7484 bool OSD::scrub_random_backoff()
7486 bool coin_flip
= (rand() / (double)RAND_MAX
>=
7487 cct
->_conf
->osd_scrub_backoff_ratio
);
7489 dout(20) << "scrub_random_backoff lost coin flip, randomly backing off" << dendl
;
7495 OSDService::ScrubJob::ScrubJob(CephContext
* cct
,
7496 const spg_t
& pg
, const utime_t
& timestamp
,
7497 double pool_scrub_min_interval
,
7498 double pool_scrub_max_interval
, bool must
)
7501 sched_time(timestamp
),
7504 // if not explicitly requested, postpone the scrub with a random delay
7506 double scrub_min_interval
= pool_scrub_min_interval
> 0 ?
7507 pool_scrub_min_interval
: cct
->_conf
->osd_scrub_min_interval
;
7508 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
7509 pool_scrub_max_interval
: cct
->_conf
->osd_scrub_max_interval
;
7511 sched_time
+= scrub_min_interval
;
7512 double r
= rand() / (double)RAND_MAX
;
7514 scrub_min_interval
* cct
->_conf
->osd_scrub_interval_randomize_ratio
* r
;
7515 deadline
+= scrub_max_interval
;
7519 bool OSDService::ScrubJob::ScrubJob::operator<(const OSDService::ScrubJob
& rhs
) const {
7520 if (sched_time
< rhs
.sched_time
)
7522 if (sched_time
> rhs
.sched_time
)
7524 return pgid
< rhs
.pgid
;
7527 bool OSD::scrub_time_permit(utime_t now
)
7530 time_t tt
= now
.sec();
7531 localtime_r(&tt
, &bdt
);
7533 bool day_permit
= false;
7534 if (cct
->_conf
->osd_scrub_begin_week_day
< cct
->_conf
->osd_scrub_end_week_day
) {
7535 if (bdt
.tm_wday
>= cct
->_conf
->osd_scrub_begin_week_day
&& bdt
.tm_wday
< cct
->_conf
->osd_scrub_end_week_day
) {
7539 if (bdt
.tm_wday
>= cct
->_conf
->osd_scrub_begin_week_day
|| bdt
.tm_wday
< cct
->_conf
->osd_scrub_end_week_day
) {
7545 dout(20) << __func__
<< " should run between week day " << cct
->_conf
->osd_scrub_begin_week_day
7546 << " - " << cct
->_conf
->osd_scrub_end_week_day
7547 << " now " << bdt
.tm_wday
<< " = no" << dendl
;
7551 bool time_permit
= false;
7552 if (cct
->_conf
->osd_scrub_begin_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7553 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
&& bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7557 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
|| bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7562 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7563 << " - " << cct
->_conf
->osd_scrub_end_hour
7564 << " now " << bdt
.tm_hour
<< " = no" << dendl
;
7566 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7567 << " - " << cct
->_conf
->osd_scrub_end_hour
7568 << " now " << bdt
.tm_hour
<< " = yes" << dendl
;
7573 bool OSD::scrub_load_below_threshold()
7576 if (getloadavg(loadavgs
, 3) != 3) {
7577 dout(10) << __func__
<< " couldn't read loadavgs\n" << dendl
;
7581 // allow scrub if below configured threshold
7582 if (loadavgs
[0] < cct
->_conf
->osd_scrub_load_threshold
) {
7583 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7584 << " < max " << cct
->_conf
->osd_scrub_load_threshold
7585 << " = yes" << dendl
;
7589 // allow scrub if below daily avg and currently decreasing
7590 if (loadavgs
[0] < daily_loadavg
&& loadavgs
[0] < loadavgs
[2]) {
7591 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7592 << " < daily_loadavg " << daily_loadavg
7593 << " and < 15m avg " << loadavgs
[2]
7594 << " = yes" << dendl
;
7598 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7599 << " >= max " << cct
->_conf
->osd_scrub_load_threshold
7600 << " and ( >= daily_loadavg " << daily_loadavg
7601 << " or >= 15m avg " << loadavgs
[2]
7602 << ") = no" << dendl
;
7606 void OSD::sched_scrub()
7608 // if not permitted, fail fast
7609 if (!service
.can_inc_scrubs_pending()) {
7612 if (!cct
->_conf
->osd_scrub_during_recovery
&& service
.is_recovery_active()) {
7613 dout(20) << __func__
<< " not scheduling scrubs due to active recovery" << dendl
;
7618 utime_t now
= ceph_clock_now();
7619 bool time_permit
= scrub_time_permit(now
);
7620 bool load_is_low
= scrub_load_below_threshold();
7621 dout(20) << "sched_scrub load_is_low=" << (int)load_is_low
<< dendl
;
7623 OSDService::ScrubJob scrub
;
7624 if (service
.first_scrub_stamp(&scrub
)) {
7626 dout(30) << "sched_scrub examine " << scrub
.pgid
<< " at " << scrub
.sched_time
<< dendl
;
7628 if (scrub
.sched_time
> now
) {
7629 // save ourselves some effort
7630 dout(10) << "sched_scrub " << scrub
.pgid
<< " scheduled at " << scrub
.sched_time
7631 << " > " << now
<< dendl
;
7635 if ((scrub
.deadline
>= now
) && !(time_permit
&& load_is_low
)) {
7636 dout(10) << __func__
<< " not scheduling scrub for " << scrub
.pgid
<< " due to "
7637 << (!time_permit
? "time not permit" : "high load") << dendl
;
7641 PG
*pg
= _lookup_lock_pg(scrub
.pgid
);
7644 if (pg
->get_pgbackend()->scrub_supported() && pg
->is_active()) {
7645 dout(10) << "sched_scrub scrubbing " << scrub
.pgid
<< " at " << scrub
.sched_time
7646 << (pg
->scrubber
.must_scrub
? ", explicitly requested" :
7647 (load_is_low
? ", load_is_low" : " deadline < now"))
7649 if (pg
->sched_scrub()) {
7655 } while (service
.next_scrub_stamp(scrub
, &scrub
));
7657 dout(20) << "sched_scrub done" << dendl
;
7662 vector
<OSDHealthMetric
> OSD::get_health_metrics()
7664 vector
<OSDHealthMetric
> metrics
;
7665 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
7666 auto n_primaries
= pending_creates_from_mon
;
7667 for (const auto& create
: pending_creates_from_osd
) {
7668 if (create
.second
) {
7672 metrics
.emplace_back(osd_metric::PENDING_CREATING_PGS
, n_primaries
);
7676 // =====================================================
7679 void OSD::wait_for_new_map(OpRequestRef op
)
7682 if (waiting_for_osdmap
.empty()) {
7683 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
7686 logger
->inc(l_osd_waiting_for_map
);
7687 waiting_for_osdmap
.push_back(op
);
7688 op
->mark_delayed("wait for new map");
7693 * assimilate new OSDMap(s). scan pgs, etc.
7696 void OSD::note_down_osd(int peer
)
7698 assert(osd_lock
.is_locked());
7699 cluster_messenger
->mark_down(osdmap
->get_cluster_addr(peer
));
7701 heartbeat_lock
.Lock();
7702 failure_queue
.erase(peer
);
7703 failure_pending
.erase(peer
);
7704 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(peer
);
7705 if (p
!= heartbeat_peers
.end()) {
7706 p
->second
.con_back
->mark_down();
7707 if (p
->second
.con_front
) {
7708 p
->second
.con_front
->mark_down();
7710 heartbeat_peers
.erase(p
);
7712 heartbeat_lock
.Unlock();
7715 void OSD::note_up_osd(int peer
)
7717 service
.forget_peer_epoch(peer
, osdmap
->get_epoch() - 1);
7718 heartbeat_set_peers_need_update();
7721 struct C_OnMapCommit
: public Context
{
7723 epoch_t first
, last
;
7725 C_OnMapCommit(OSD
*o
, epoch_t f
, epoch_t l
, MOSDMap
*m
)
7726 : osd(o
), first(f
), last(l
), msg(m
) {}
7727 void finish(int r
) override
{
7728 osd
->_committed_osd_maps(first
, last
, msg
);
7733 struct C_OnMapApply
: public Context
{
7734 OSDService
*service
;
7735 list
<OSDMapRef
> pinned_maps
;
7737 C_OnMapApply(OSDService
*service
,
7738 const list
<OSDMapRef
> &pinned_maps
,
7740 : service(service
), pinned_maps(pinned_maps
), e(e
) {}
7741 void finish(int r
) override
{
7742 service
->clear_map_bl_cache_pins(e
);
7746 void OSD::osdmap_subscribe(version_t epoch
, bool force_request
)
7748 Mutex::Locker
l(osdmap_subscribe_lock
);
7749 if (latest_subscribed_epoch
>= epoch
&& !force_request
)
7752 latest_subscribed_epoch
= MAX(epoch
, latest_subscribed_epoch
);
7754 if (monc
->sub_want_increment("osdmap", epoch
, CEPH_SUBSCRIBE_ONETIME
) ||
7760 void OSD::trim_maps(epoch_t oldest
, int nreceived
, bool skip_maps
)
7762 epoch_t min
= std::min(oldest
, service
.map_cache
.cached_key_lower_bound());
7763 if (min
<= superblock
.oldest_map
)
7767 ObjectStore::Transaction t
;
7768 for (epoch_t e
= superblock
.oldest_map
; e
< min
; ++e
) {
7769 dout(20) << " removing old osdmap epoch " << e
<< dendl
;
7770 t
.remove(coll_t::meta(), get_osdmap_pobject_name(e
));
7771 t
.remove(coll_t::meta(), get_inc_osdmap_pobject_name(e
));
7772 superblock
.oldest_map
= e
+ 1;
7774 if (num
>= cct
->_conf
->osd_target_transaction_size
&& num
>= nreceived
) {
7775 service
.publish_superblock(superblock
);
7776 write_superblock(t
);
7777 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7781 // skip_maps leaves us with a range of old maps if we fail to remove all
7782 // of them before moving superblock.oldest_map forward to the first map
7783 // in the incoming MOSDMap msg. so we should continue removing them in
7784 // this case, even we could do huge series of delete transactions all at
7791 service
.publish_superblock(superblock
);
7792 write_superblock(t
);
7793 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7796 // we should not remove the cached maps
7797 assert(min
<= service
.map_cache
.cached_key_lower_bound());
7800 void OSD::handle_osd_map(MOSDMap
*m
)
7802 assert(osd_lock
.is_locked());
7803 // Keep a ref in the list until we get the newly received map written
7804 // onto disk. This is important because as long as the refs are alive,
7805 // the OSDMaps will be pinned in the cache and we won't try to read it
7806 // off of disk. Otherwise these maps will probably not stay in the cache,
7807 // and reading those OSDMaps before they are actually written can result
7809 list
<OSDMapRef
> pinned_maps
;
7810 if (m
->fsid
!= monc
->get_fsid()) {
7811 dout(0) << "handle_osd_map fsid " << m
->fsid
<< " != "
7812 << monc
->get_fsid() << dendl
;
7816 if (is_initializing()) {
7817 dout(0) << "ignoring osdmap until we have initialized" << dendl
;
7822 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
7823 if (session
&& !(session
->entity_name
.is_mon() ||
7824 session
->entity_name
.is_osd())) {
7826 dout(10) << "got osd map from Session " << session
7827 << " which we can't take maps from (not a mon or osd)" << dendl
;
7835 // share with the objecter
7837 service
.objecter
->handle_osd_map(m
);
7839 epoch_t first
= m
->get_first();
7840 epoch_t last
= m
->get_last();
7841 dout(3) << "handle_osd_map epochs [" << first
<< "," << last
<< "], i have "
7842 << superblock
.newest_map
7843 << ", src has [" << m
->oldest_map
<< "," << m
->newest_map
<< "]"
7846 logger
->inc(l_osd_map
);
7847 logger
->inc(l_osd_mape
, last
- first
+ 1);
7848 if (first
<= superblock
.newest_map
)
7849 logger
->inc(l_osd_mape_dup
, superblock
.newest_map
- first
+ 1);
7850 if (service
.max_oldest_map
< m
->oldest_map
) {
7851 service
.max_oldest_map
= m
->oldest_map
;
7852 assert(service
.max_oldest_map
>= superblock
.oldest_map
);
7855 // make sure there is something new, here, before we bother flushing
7856 // the queues and such
7857 if (last
<= superblock
.newest_map
) {
7858 dout(10) << " no new maps here, dropping" << dendl
;
7864 bool skip_maps
= false;
7865 if (first
> superblock
.newest_map
+ 1) {
7866 dout(10) << "handle_osd_map message skips epochs "
7867 << superblock
.newest_map
+ 1 << ".." << (first
-1) << dendl
;
7868 if (m
->oldest_map
<= superblock
.newest_map
+ 1) {
7869 osdmap_subscribe(superblock
.newest_map
+ 1, false);
7873 // always try to get the full range of maps--as many as we can. this
7874 // 1- is good to have
7875 // 2- is at present the only way to ensure that we get a *full* map as
7877 if (m
->oldest_map
< first
) {
7878 osdmap_subscribe(m
->oldest_map
- 1, true);
7885 ObjectStore::Transaction t
;
7886 uint64_t txn_size
= 0;
7888 // store new maps: queue for disk and put in the osdmap cache
7889 epoch_t start
= MAX(superblock
.newest_map
+ 1, first
);
7890 for (epoch_t e
= start
; e
<= last
; e
++) {
7891 if (txn_size
>= t
.get_num_bytes()) {
7892 derr
<< __func__
<< " transaction size overflowed" << dendl
;
7893 assert(txn_size
< t
.get_num_bytes());
7895 txn_size
= t
.get_num_bytes();
7896 map
<epoch_t
,bufferlist
>::iterator p
;
7897 p
= m
->maps
.find(e
);
7898 if (p
!= m
->maps
.end()) {
7899 dout(10) << "handle_osd_map got full map for epoch " << e
<< dendl
;
7900 OSDMap
*o
= new OSDMap
;
7901 bufferlist
& bl
= p
->second
;
7905 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7906 t
.write(coll_t::meta(), fulloid
, 0, bl
.length(), bl
);
7908 pinned_maps
.push_back(add_map(o
));
7914 p
= m
->incremental_maps
.find(e
);
7915 if (p
!= m
->incremental_maps
.end()) {
7916 dout(10) << "handle_osd_map got inc map for epoch " << e
<< dendl
;
7917 bufferlist
& bl
= p
->second
;
7918 ghobject_t oid
= get_inc_osdmap_pobject_name(e
);
7919 t
.write(coll_t::meta(), oid
, 0, bl
.length(), bl
);
7920 pin_map_inc_bl(e
, bl
);
7922 OSDMap
*o
= new OSDMap
;
7925 bool got
= get_map_bl(e
- 1, obl
);
7930 OSDMap::Incremental inc
;
7931 bufferlist::iterator p
= bl
.begin();
7933 if (o
->apply_incremental(inc
) < 0) {
7934 derr
<< "ERROR: bad fsid? i have " << osdmap
->get_fsid() << " and inc has " << inc
.fsid
<< dendl
;
7935 assert(0 == "bad fsid");
7939 o
->encode(fbl
, inc
.encode_features
| CEPH_FEATURE_RESERVED
);
7941 bool injected_failure
= false;
7942 if (cct
->_conf
->osd_inject_bad_map_crc_probability
> 0 &&
7943 (rand() % 10000) < cct
->_conf
->osd_inject_bad_map_crc_probability
*10000.0) {
7944 derr
<< __func__
<< " injecting map crc failure" << dendl
;
7945 injected_failure
= true;
7948 if ((inc
.have_crc
&& o
->get_crc() != inc
.full_crc
) || injected_failure
) {
7949 dout(2) << "got incremental " << e
7950 << " but failed to encode full with correct crc; requesting"
7952 clog
->warn() << "failed to encode map e" << e
<< " with expected crc";
7953 dout(20) << "my encoded map was:\n";
7954 fbl
.hexdump(*_dout
);
7957 request_full_map(e
, last
);
7963 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7964 t
.write(coll_t::meta(), fulloid
, 0, fbl
.length(), fbl
);
7966 pinned_maps
.push_back(add_map(o
));
7970 assert(0 == "MOSDMap lied about what maps it had?");
7973 // even if this map isn't from a mon, we may have satisfied our subscription
7974 monc
->sub_got("osdmap", last
);
7976 if (!m
->maps
.empty() && requested_full_first
) {
7977 dout(10) << __func__
<< " still missing full maps " << requested_full_first
7978 << ".." << requested_full_last
<< dendl
;
7979 rerequest_full_maps();
7982 if (superblock
.oldest_map
) {
7983 // make sure we at least keep pace with incoming maps
7984 trim_maps(m
->oldest_map
, last
- first
+ 1, skip_maps
);
7987 if (!superblock
.oldest_map
|| skip_maps
)
7988 superblock
.oldest_map
= first
;
7989 superblock
.newest_map
= last
;
7990 superblock
.current_epoch
= last
;
7992 // note in the superblock that we were clean thru the prior epoch
7993 epoch_t boot_epoch
= service
.get_boot_epoch();
7994 if (boot_epoch
&& boot_epoch
>= superblock
.mounted
) {
7995 superblock
.mounted
= boot_epoch
;
7996 superblock
.clean_thru
= last
;
7999 // superblock and commit
8000 write_superblock(t
);
8001 store
->queue_transaction(
8002 service
.meta_osr
.get(),
8004 new C_OnMapApply(&service
, pinned_maps
, last
),
8005 new C_OnMapCommit(this, start
, last
, m
), 0);
8006 service
.publish_superblock(superblock
);
8009 void OSD::_committed_osd_maps(epoch_t first
, epoch_t last
, MOSDMap
*m
)
8011 dout(10) << __func__
<< " " << first
<< ".." << last
<< dendl
;
8012 if (is_stopping()) {
8013 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
8016 Mutex::Locker
l(osd_lock
);
8017 if (is_stopping()) {
8018 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
8021 map_lock
.get_write();
8023 bool do_shutdown
= false;
8024 bool do_restart
= false;
8025 bool network_error
= false;
8027 // advance through the new maps
8028 for (epoch_t cur
= first
; cur
<= last
; cur
++) {
8029 dout(10) << " advance to epoch " << cur
8030 << " (<= last " << last
8031 << " <= newest_map " << superblock
.newest_map
8034 OSDMapRef newmap
= get_map(cur
);
8035 assert(newmap
); // we just cached it above!
8037 // start blacklisting messages sent to peers that go down.
8038 service
.pre_publish_map(newmap
);
8040 // kill connections to newly down osds
8041 bool waited_for_reservations
= false;
8043 osdmap
->get_all_osds(old
);
8044 for (set
<int>::iterator p
= old
.begin(); p
!= old
.end(); ++p
) {
8046 osdmap
->is_up(*p
) && // in old map
8047 newmap
->is_down(*p
)) { // but not the new one
8048 if (!waited_for_reservations
) {
8049 service
.await_reserved_maps();
8050 waited_for_reservations
= true;
8053 } else if (*p
!= whoami
&&
8054 osdmap
->is_down(*p
) &&
8055 newmap
->is_up(*p
)) {
8060 if ((osdmap
->test_flag(CEPH_OSDMAP_NOUP
) !=
8061 newmap
->test_flag(CEPH_OSDMAP_NOUP
)) ||
8062 (osdmap
->is_noup(whoami
) != newmap
->is_noup(whoami
))) {
8063 dout(10) << __func__
<< " NOUP flag changed in " << newmap
->get_epoch()
8066 // this captures the case where we sent the boot message while
8067 // NOUP was being set on the mon and our boot request was
8068 // dropped, and then later it is cleared. it imperfectly
8069 // handles the case where our original boot message was not
8070 // dropped and we restart even though we might have booted, but
8071 // that is harmless (boot will just take slightly longer).
8075 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
&&
8076 newmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
8077 dout(10) << __func__
<< " require_osd_release reached luminous in "
8078 << newmap
->get_epoch() << dendl
;
8079 clear_pg_stat_queue();
8080 clear_outstanding_pg_stats();
8086 service
.retrieve_epochs(&boot_epoch
, &up_epoch
, NULL
);
8088 osdmap
->is_up(whoami
) &&
8089 osdmap
->get_inst(whoami
) == client_messenger
->get_myinst()) {
8090 up_epoch
= osdmap
->get_epoch();
8091 dout(10) << "up_epoch is " << up_epoch
<< dendl
;
8093 boot_epoch
= osdmap
->get_epoch();
8094 dout(10) << "boot_epoch is " << boot_epoch
<< dendl
;
8096 service
.set_epochs(&boot_epoch
, &up_epoch
, NULL
);
8100 had_map_since
= ceph_clock_now();
8102 epoch_t _bind_epoch
= service
.get_bind_epoch();
8103 if (osdmap
->is_up(whoami
) &&
8104 osdmap
->get_addr(whoami
) == client_messenger
->get_myaddr() &&
8105 _bind_epoch
< osdmap
->get_up_from(whoami
)) {
8108 dout(1) << "state: booting -> active" << dendl
;
8109 set_state(STATE_ACTIVE
);
8111 // set incarnation so that osd_reqid_t's we generate for our
8112 // objecter requests are unique across restarts.
8113 service
.objecter
->set_client_incarnation(osdmap
->get_epoch());
8117 if (osdmap
->get_epoch() > 0 &&
8119 if (!osdmap
->exists(whoami
)) {
8120 dout(0) << "map says i do not exist. shutting down." << dendl
;
8121 do_shutdown
= true; // don't call shutdown() while we have
8122 // everything paused
8123 } else if (!osdmap
->is_up(whoami
) ||
8124 !osdmap
->get_addr(whoami
).probably_equals(
8125 client_messenger
->get_myaddr()) ||
8126 !osdmap
->get_cluster_addr(whoami
).probably_equals(
8127 cluster_messenger
->get_myaddr()) ||
8128 !osdmap
->get_hb_back_addr(whoami
).probably_equals(
8129 hb_back_server_messenger
->get_myaddr()) ||
8130 (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
8131 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
8132 hb_front_server_messenger
->get_myaddr()))) {
8133 if (!osdmap
->is_up(whoami
)) {
8134 if (service
.is_preparing_to_stop() || service
.is_stopping()) {
8135 service
.got_stop_ack();
8137 clog
->warn() << "Monitor daemon marked osd." << whoami
<< " down, "
8138 "but it is still running";
8139 clog
->debug() << "map e" << osdmap
->get_epoch()
8140 << " wrongly marked me down at e"
8141 << osdmap
->get_down_at(whoami
);
8143 } else if (!osdmap
->get_addr(whoami
).probably_equals(
8144 client_messenger
->get_myaddr())) {
8145 clog
->error() << "map e" << osdmap
->get_epoch()
8146 << " had wrong client addr (" << osdmap
->get_addr(whoami
)
8147 << " != my " << client_messenger
->get_myaddr() << ")";
8148 } else if (!osdmap
->get_cluster_addr(whoami
).probably_equals(
8149 cluster_messenger
->get_myaddr())) {
8150 clog
->error() << "map e" << osdmap
->get_epoch()
8151 << " had wrong cluster addr ("
8152 << osdmap
->get_cluster_addr(whoami
)
8153 << " != my " << cluster_messenger
->get_myaddr() << ")";
8154 } else if (!osdmap
->get_hb_back_addr(whoami
).probably_equals(
8155 hb_back_server_messenger
->get_myaddr())) {
8156 clog
->error() << "map e" << osdmap
->get_epoch()
8157 << " had wrong heartbeat back addr ("
8158 << osdmap
->get_hb_back_addr(whoami
)
8159 << " != my " << hb_back_server_messenger
->get_myaddr()
8161 } else if (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
8162 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
8163 hb_front_server_messenger
->get_myaddr())) {
8164 clog
->error() << "map e" << osdmap
->get_epoch()
8165 << " had wrong heartbeat front addr ("
8166 << osdmap
->get_hb_front_addr(whoami
)
8167 << " != my " << hb_front_server_messenger
->get_myaddr()
8171 if (!service
.is_stopping()) {
8172 epoch_t up_epoch
= 0;
8173 epoch_t bind_epoch
= osdmap
->get_epoch();
8174 service
.set_epochs(NULL
,&up_epoch
, &bind_epoch
);
8178 utime_t now
= ceph_clock_now();
8179 utime_t grace
= utime_t(cct
->_conf
->osd_max_markdown_period
, 0);
8180 osd_markdown_log
.push_back(now
);
8181 //clear all out-of-date log
8182 while (!osd_markdown_log
.empty() &&
8183 osd_markdown_log
.front() + grace
< now
)
8184 osd_markdown_log
.pop_front();
8185 if ((int)osd_markdown_log
.size() > cct
->_conf
->osd_max_markdown_count
) {
8186 dout(0) << __func__
<< " marked down "
8187 << osd_markdown_log
.size()
8188 << " > osd_max_markdown_count "
8189 << cct
->_conf
->osd_max_markdown_count
8190 << " in last " << grace
<< " seconds, shutting down"
8196 start_waiting_for_healthy();
8198 set
<int> avoid_ports
;
8199 #if defined(__FreeBSD__)
8200 // prevent FreeBSD from grabbing the client_messenger port during
8201 // rebinding. In which case a cluster_meesneger will connect also
8203 avoid_ports
.insert(client_messenger
->get_myaddr().get_port());
8205 avoid_ports
.insert(cluster_messenger
->get_myaddr().get_port());
8206 avoid_ports
.insert(hb_back_server_messenger
->get_myaddr().get_port());
8207 avoid_ports
.insert(hb_front_server_messenger
->get_myaddr().get_port());
8209 int r
= cluster_messenger
->rebind(avoid_ports
);
8211 do_shutdown
= true; // FIXME: do_restart?
8212 network_error
= true;
8213 dout(0) << __func__
<< " marked down:"
8214 << " rebind cluster_messenger failed" << dendl
;
8217 r
= hb_back_server_messenger
->rebind(avoid_ports
);
8219 do_shutdown
= true; // FIXME: do_restart?
8220 network_error
= true;
8221 dout(0) << __func__
<< " marked down:"
8222 << " rebind hb_back_server_messenger failed" << dendl
;
8225 r
= hb_front_server_messenger
->rebind(avoid_ports
);
8227 do_shutdown
= true; // FIXME: do_restart?
8228 network_error
= true;
8229 dout(0) << __func__
<< " marked down:"
8230 << " rebind hb_front_server_messenger failed" << dendl
;
8233 hb_front_client_messenger
->mark_down_all();
8234 hb_back_client_messenger
->mark_down_all();
8236 reset_heartbeat_peers();
8241 map_lock
.put_write();
8243 check_osdmap_features(store
);
8248 if (is_active() || is_waiting_for_healthy())
8249 maybe_update_heartbeat_peers();
8252 dout(10) << " not yet active; waiting for peering wq to drain" << dendl
;
8259 if (network_error
) {
8260 Mutex::Locker
l(heartbeat_lock
);
8261 map
<int,pair
<utime_t
,entity_inst_t
>>::iterator it
=
8262 failure_pending
.begin();
8263 while (it
!= failure_pending
.end()) {
8264 dout(10) << "handle_osd_ping canceling in-flight failure report for osd."
8265 << it
->first
<< dendl
;
8266 send_still_alive(osdmap
->get_epoch(), it
->second
.second
);
8267 failure_pending
.erase(it
++);
8270 // trigger shutdown in a different thread
8271 dout(0) << __func__
<< " shutdown OSD via async signal" << dendl
;
8272 queue_async_signal(SIGINT
);
8274 else if (m
->newest_map
&& m
->newest_map
> last
) {
8275 dout(10) << " msg say newest map is " << m
->newest_map
8276 << ", requesting more" << dendl
;
8277 osdmap_subscribe(osdmap
->get_epoch()+1, false);
8279 else if (is_preboot()) {
8280 if (m
->get_source().is_mon())
8281 _preboot(m
->oldest_map
, m
->newest_map
);
8285 else if (do_restart
)
8290 void OSD::check_osdmap_features(ObjectStore
*fs
)
8292 // adjust required feature bits?
8294 // we have to be a bit careful here, because we are accessing the
8295 // Policy structures without taking any lock. in particular, only
8296 // modify integer values that can safely be read by a racing CPU.
8297 // since we are only accessing existing Policy structures a their
8298 // current memory location, and setting or clearing bits in integer
8299 // fields, and we are the only writer, this is not a problem.
8302 Messenger::Policy p
= client_messenger
->get_default_policy();
8304 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_CLIENT
, &mask
);
8305 if ((p
.features_required
& mask
) != features
) {
8306 dout(0) << "crush map has features " << features
8307 << ", adjusting msgr requires for clients" << dendl
;
8308 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8309 client_messenger
->set_default_policy(p
);
8313 Messenger::Policy p
= client_messenger
->get_policy(entity_name_t::TYPE_MON
);
8315 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_MON
, &mask
);
8316 if ((p
.features_required
& mask
) != features
) {
8317 dout(0) << "crush map has features " << features
8318 << " was " << p
.features_required
8319 << ", adjusting msgr requires for mons" << dendl
;
8320 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8321 client_messenger
->set_policy(entity_name_t::TYPE_MON
, p
);
8325 Messenger::Policy p
= cluster_messenger
->get_policy(entity_name_t::TYPE_OSD
);
8327 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_OSD
, &mask
);
8329 if ((p
.features_required
& mask
) != features
) {
8330 dout(0) << "crush map has features " << features
8331 << ", adjusting msgr requires for osds" << dendl
;
8332 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8333 cluster_messenger
->set_policy(entity_name_t::TYPE_OSD
, p
);
8336 if ((features
& CEPH_FEATURE_OSD_ERASURE_CODES
) &&
8337 !superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
)) {
8338 dout(0) << __func__
<< " enabling on-disk ERASURE CODES compat feature" << dendl
;
8339 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
);
8340 ObjectStore::Transaction t
;
8341 write_superblock(t
);
8342 int err
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), NULL
);
8348 bool OSD::advance_pg(
8349 epoch_t osd_epoch
, PG
*pg
,
8350 ThreadPool::TPHandle
&handle
,
8351 PG::RecoveryCtx
*rctx
,
8352 set
<PGRef
> *new_pgs
)
8354 assert(pg
->is_locked());
8355 epoch_t next_epoch
= pg
->get_osdmap()->get_epoch() + 1;
8356 OSDMapRef lastmap
= pg
->get_osdmap();
8358 if (lastmap
->get_epoch() == osd_epoch
)
8360 assert(lastmap
->get_epoch() < osd_epoch
);
8362 epoch_t min_epoch
= service
.get_min_pg_epoch();
8365 max
= min_epoch
+ cct
->_conf
->osd_map_max_advance
;
8367 max
= next_epoch
+ cct
->_conf
->osd_map_max_advance
;
8371 next_epoch
<= osd_epoch
&& next_epoch
<= max
;
8373 OSDMapRef nextmap
= service
.try_get_map(next_epoch
);
8375 dout(20) << __func__
<< " missing map " << next_epoch
<< dendl
;
8376 // make sure max is bumped up so that we can get past any
8378 max
= MAX(max
, next_epoch
+ cct
->_conf
->osd_map_max_advance
);
8382 vector
<int> newup
, newacting
;
8383 int up_primary
, acting_primary
;
8384 nextmap
->pg_to_up_acting_osds(
8386 &newup
, &up_primary
,
8387 &newacting
, &acting_primary
);
8388 pg
->handle_advance_map(
8389 nextmap
, lastmap
, newup
, up_primary
,
8390 newacting
, acting_primary
, rctx
);
8393 set
<spg_t
> children
;
8394 spg_t
parent(pg
->info
.pgid
);
8395 if (parent
.is_split(
8396 lastmap
->get_pg_num(pg
->pool
.id
),
8397 nextmap
->get_pg_num(pg
->pool
.id
),
8399 service
.mark_split_in_progress(pg
->info
.pgid
, children
);
8401 pg
, children
, new_pgs
, lastmap
, nextmap
,
8406 handle
.reset_tp_timeout();
8408 service
.pg_update_epoch(pg
->info
.pgid
, lastmap
->get_epoch());
8409 pg
->handle_activate_map(rctx
);
8410 if (next_epoch
<= osd_epoch
) {
8411 dout(10) << __func__
<< " advanced to max " << max
8412 << " past min epoch " << min_epoch
8413 << " ... will requeue " << *pg
<< dendl
;
8419 void OSD::consume_map()
8421 assert(osd_lock
.is_locked());
8422 dout(7) << "consume_map version " << osdmap
->get_epoch() << dendl
;
8424 /** make sure the cluster is speaking in SORTBITWISE, because we don't
8425 * speak the older sorting version any more. Be careful not to force
8426 * a shutdown if we are merely processing old maps, though.
8428 if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
) && is_active()) {
8429 derr
<< __func__
<< " SORTBITWISE flag is not set" << dendl
;
8433 int num_pg_primary
= 0, num_pg_replica
= 0, num_pg_stray
= 0;
8434 list
<PGRef
> to_remove
;
8438 RWLock::RLocker
l(pg_map_lock
);
8439 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8442 PG
*pg
= it
->second
;
8444 if (pg
->is_primary())
8446 else if (pg
->is_replica())
8451 if (!osdmap
->have_pg_pool(pg
->info
.pgid
.pool())) {
8453 to_remove
.push_back(PGRef(pg
));
8455 service
.init_splits_between(it
->first
, service
.get_osdmap(), osdmap
);
8461 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
8462 for (auto pg
= pending_creates_from_osd
.cbegin();
8463 pg
!= pending_creates_from_osd
.cend();) {
8464 if (osdmap
->get_pg_acting_rank(pg
->first
, whoami
) < 0) {
8465 pg
= pending_creates_from_osd
.erase(pg
);
8472 for (list
<PGRef
>::iterator i
= to_remove
.begin();
8473 i
!= to_remove
.end();
8474 to_remove
.erase(i
++)) {
8475 RWLock::WLocker
locker(pg_map_lock
);
8481 service
.expand_pg_num(service
.get_osdmap(), osdmap
);
8483 service
.pre_publish_map(osdmap
);
8484 service
.await_reserved_maps();
8485 service
.publish_map(osdmap
);
8487 service
.maybe_inject_dispatch_delay();
8489 dispatch_sessions_waiting_on_map();
8491 service
.maybe_inject_dispatch_delay();
8493 // remove any PGs which we no longer host from the session waiting_for_pg lists
8494 dout(20) << __func__
<< " checking waiting_for_pg" << dendl
;
8495 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
8497 service
.maybe_inject_dispatch_delay();
8501 RWLock::RLocker
l(pg_map_lock
);
8502 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8505 PG
*pg
= it
->second
;
8507 pg
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
8511 logger
->set(l_osd_pg
, pg_map
.size());
8513 logger
->set(l_osd_pg_primary
, num_pg_primary
);
8514 logger
->set(l_osd_pg_replica
, num_pg_replica
);
8515 logger
->set(l_osd_pg_stray
, num_pg_stray
);
8516 logger
->set(l_osd_pg_removing
, remove_wq
.get_remove_queue_len());
8519 void OSD::activate_map()
8521 assert(osd_lock
.is_locked());
8523 dout(7) << "activate_map version " << osdmap
->get_epoch() << dendl
;
8525 if (osdmap
->test_flag(CEPH_OSDMAP_FULL
)) {
8526 dout(10) << " osdmap flagged full, doing onetime osdmap subscribe" << dendl
;
8527 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
8531 if (osdmap
->test_flag(CEPH_OSDMAP_NORECOVER
)) {
8532 if (!service
.recovery_is_paused()) {
8533 dout(1) << "pausing recovery (NORECOVER flag set)" << dendl
;
8534 service
.pause_recovery();
8537 if (service
.recovery_is_paused()) {
8538 dout(1) << "unpausing recovery (NORECOVER flag unset)" << dendl
;
8539 service
.unpause_recovery();
8543 service
.activate_map();
8546 take_waiters(waiting_for_osdmap
);
8549 bool OSD::require_mon_peer(const Message
*m
)
8551 if (!m
->get_connection()->peer_is_mon()) {
8552 dout(0) << "require_mon_peer received from non-mon "
8553 << m
->get_connection()->get_peer_addr()
8554 << " " << *m
<< dendl
;
8560 bool OSD::require_mon_or_mgr_peer(const Message
*m
)
8562 if (!m
->get_connection()->peer_is_mon() &&
8563 !m
->get_connection()->peer_is_mgr()) {
8564 dout(0) << "require_mon_or_mgr_peer received from non-mon, non-mgr "
8565 << m
->get_connection()->get_peer_addr()
8566 << " " << *m
<< dendl
;
8572 bool OSD::require_osd_peer(const Message
*m
)
8574 if (!m
->get_connection()->peer_is_osd()) {
8575 dout(0) << "require_osd_peer received from non-osd "
8576 << m
->get_connection()->get_peer_addr()
8577 << " " << *m
<< dendl
;
8583 bool OSD::require_self_aliveness(const Message
*m
, epoch_t epoch
)
8585 epoch_t up_epoch
= service
.get_up_epoch();
8586 if (epoch
< up_epoch
) {
8587 dout(7) << "from pre-up epoch " << epoch
<< " < " << up_epoch
<< dendl
;
8592 dout(7) << "still in boot state, dropping message " << *m
<< dendl
;
8599 bool OSD::require_same_peer_instance(const Message
*m
, OSDMapRef
& map
,
8600 bool is_fast_dispatch
)
8602 int from
= m
->get_source().num();
8604 if (map
->is_down(from
) ||
8605 (map
->get_cluster_addr(from
) != m
->get_source_inst().addr
)) {
8606 dout(5) << "from dead osd." << from
<< ", marking down, "
8607 << " msg was " << m
->get_source_inst().addr
8608 << " expected " << (map
->is_up(from
) ?
8609 map
->get_cluster_addr(from
) : entity_addr_t())
8611 ConnectionRef con
= m
->get_connection();
8613 Session
*s
= static_cast<Session
*>(con
->get_priv());
8615 if (!is_fast_dispatch
)
8616 s
->session_dispatch_lock
.Lock();
8617 clear_session_waiting_on_map(s
);
8618 con
->set_priv(NULL
); // break ref <-> session cycle, if any
8619 if (!is_fast_dispatch
)
8620 s
->session_dispatch_lock
.Unlock();
8630 * require that we have same (or newer) map, and that
8631 * the source is the pg primary.
8633 bool OSD::require_same_or_newer_map(OpRequestRef
& op
, epoch_t epoch
,
8634 bool is_fast_dispatch
)
8636 const Message
*m
= op
->get_req();
8637 dout(15) << "require_same_or_newer_map " << epoch
8638 << " (i am " << osdmap
->get_epoch() << ") " << m
<< dendl
;
8640 assert(osd_lock
.is_locked());
8642 // do they have a newer map?
8643 if (epoch
> osdmap
->get_epoch()) {
8644 dout(7) << "waiting for newer map epoch " << epoch
8645 << " > my " << osdmap
->get_epoch() << " with " << m
<< dendl
;
8646 wait_for_new_map(op
);
8650 if (!require_self_aliveness(op
->get_req(), epoch
)) {
8654 // ok, our map is same or newer.. do they still exist?
8655 if (m
->get_connection()->get_messenger() == cluster_messenger
&&
8656 !require_same_peer_instance(op
->get_req(), osdmap
, is_fast_dispatch
)) {
8667 // ----------------------------------------
8670 void OSD::split_pgs(
8672 const set
<spg_t
> &childpgids
, set
<PGRef
> *out_pgs
,
8675 PG::RecoveryCtx
*rctx
)
8677 unsigned pg_num
= nextmap
->get_pg_num(
8679 parent
->update_snap_mapper_bits(
8680 parent
->info
.pgid
.get_split_bits(pg_num
)
8683 vector
<object_stat_sum_t
> updated_stats(childpgids
.size() + 1);
8684 parent
->info
.stats
.stats
.sum
.split(updated_stats
);
8686 vector
<object_stat_sum_t
>::iterator stat_iter
= updated_stats
.begin();
8687 for (set
<spg_t
>::const_iterator i
= childpgids
.begin();
8688 i
!= childpgids
.end();
8690 assert(stat_iter
!= updated_stats
.end());
8691 dout(10) << "Splitting " << *parent
<< " into " << *i
<< dendl
;
8692 assert(service
.splitting(*i
));
8693 PG
* child
= _make_pg(nextmap
, *i
);
8695 out_pgs
->insert(child
);
8696 rctx
->created_pgs
.insert(child
);
8698 unsigned split_bits
= i
->get_split_bits(pg_num
);
8699 dout(10) << "pg_num is " << pg_num
<< dendl
;
8700 dout(10) << "m_seed " << i
->ps() << dendl
;
8701 dout(10) << "split_bits is " << split_bits
<< dendl
;
8703 parent
->split_colls(
8713 child
->info
.stats
.stats
.sum
= *stat_iter
;
8715 child
->write_if_dirty(*(rctx
->transaction
));
8718 assert(stat_iter
!= updated_stats
.end());
8719 parent
->info
.stats
.stats
.sum
= *stat_iter
;
8720 parent
->write_if_dirty(*(rctx
->transaction
));
8726 void OSD::handle_pg_create(OpRequestRef op
)
8728 const MOSDPGCreate
*m
= static_cast<const MOSDPGCreate
*>(op
->get_req());
8729 assert(m
->get_type() == MSG_OSD_PG_CREATE
);
8731 dout(10) << "handle_pg_create " << *m
<< dendl
;
8733 if (!require_mon_peer(op
->get_req())) {
8737 if (!require_same_or_newer_map(op
, m
->epoch
, false))
8742 map
<pg_t
,utime_t
>::const_iterator ci
= m
->ctimes
.begin();
8743 for (map
<pg_t
,pg_create_t
>::const_iterator p
= m
->mkpg
.begin();
8746 assert(ci
!= m
->ctimes
.end() && ci
->first
== p
->first
);
8747 epoch_t created
= p
->second
.created
;
8748 if (p
->second
.split_bits
) // Skip split pgs
8752 if (on
.preferred() >= 0) {
8753 dout(20) << "ignoring localized pg " << on
<< dendl
;
8757 if (!osdmap
->have_pg_pool(on
.pool())) {
8758 dout(20) << "ignoring pg on deleted pool " << on
<< dendl
;
8762 dout(20) << "mkpg " << on
<< " e" << created
<< "@" << ci
->second
<< dendl
;
8764 // is it still ours?
8765 vector
<int> up
, acting
;
8766 int up_primary
= -1;
8767 int acting_primary
= -1;
8768 osdmap
->pg_to_up_acting_osds(on
, &up
, &up_primary
, &acting
, &acting_primary
);
8769 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
8771 if (acting_primary
!= whoami
) {
8772 dout(10) << "mkpg " << on
<< " not acting_primary (" << acting_primary
8773 << "), my role=" << role
<< ", skipping" << dendl
;
8778 bool mapped
= osdmap
->get_primary_shard(on
, &pgid
);
8782 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
8784 pg_history_t history
;
8785 build_initial_pg_history(pgid
, created
, ci
->second
, &history
, &pi
);
8787 // The mon won't resend unless the primary changed, so
8788 // we ignore same_interval_since. We'll pass this history
8789 // to handle_pg_peering_evt with the current epoch as the
8790 // event -- the project_pg_history check in
8791 // handle_pg_peering_evt will be a noop.
8792 if (history
.same_primary_since
> m
->epoch
) {
8793 dout(10) << __func__
<< ": got obsolete pg create on pgid "
8794 << pgid
<< " from epoch " << m
->epoch
8795 << ", primary changed in " << history
.same_primary_since
8799 if (handle_pg_peering_evt(
8803 osdmap
->get_epoch(),
8804 PG::CephPeeringEvtRef(
8805 new PG::CephPeeringEvt(
8806 osdmap
->get_epoch(),
8807 osdmap
->get_epoch(),
8810 service
.send_pg_created(pgid
.pgid
);
8815 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
8816 if (pending_creates_from_mon
== 0) {
8817 last_pg_create_epoch
= m
->epoch
;
8820 maybe_update_heartbeat_peers();
8824 // ----------------------------------------
8825 // peering and recovery
8827 PG::RecoveryCtx
OSD::create_context()
8829 ObjectStore::Transaction
*t
= new ObjectStore::Transaction
;
8830 C_Contexts
*on_applied
= new C_Contexts(cct
);
8831 C_Contexts
*on_safe
= new C_Contexts(cct
);
8832 map
<int, map
<spg_t
,pg_query_t
> > *query_map
=
8833 new map
<int, map
<spg_t
, pg_query_t
> >;
8834 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *notify_list
=
8835 new map
<int, vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8836 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *info_map
=
8837 new map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8838 PG::RecoveryCtx
rctx(query_map
, info_map
, notify_list
,
8839 on_applied
, on_safe
, t
);
8843 struct C_OpenPGs
: public Context
{
8847 C_OpenPGs(set
<PGRef
>& p
, ObjectStore
*s
, OSD
* o
) : store(s
), osd(o
) {
8850 void finish(int r
) override
{
8851 RWLock::RLocker
l(osd
->pg_map_lock
);
8852 for (auto p
: pgs
) {
8853 if (osd
->pg_map
.count(p
->info
.pgid
)) {
8854 p
->ch
= store
->open_collection(p
->coll
);
8861 void OSD::dispatch_context_transaction(PG::RecoveryCtx
&ctx
, PG
*pg
,
8862 ThreadPool::TPHandle
*handle
)
8864 if (!ctx
.transaction
->empty()) {
8865 if (!ctx
.created_pgs
.empty()) {
8866 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8868 int tr
= store
->queue_transaction(
8870 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
,
8871 TrackedOpRef(), handle
);
8872 delete (ctx
.transaction
);
8874 ctx
.transaction
= new ObjectStore::Transaction
;
8875 ctx
.on_applied
= new C_Contexts(cct
);
8876 ctx
.on_safe
= new C_Contexts(cct
);
8880 void OSD::dispatch_context(PG::RecoveryCtx
&ctx
, PG
*pg
, OSDMapRef curmap
,
8881 ThreadPool::TPHandle
*handle
)
8883 if (service
.get_osdmap()->is_up(whoami
) &&
8885 do_notifies(*ctx
.notify_list
, curmap
);
8886 do_queries(*ctx
.query_map
, curmap
);
8887 do_infos(*ctx
.info_map
, curmap
);
8889 delete ctx
.notify_list
;
8890 delete ctx
.query_map
;
8891 delete ctx
.info_map
;
8892 if ((ctx
.on_applied
->empty() &&
8893 ctx
.on_safe
->empty() &&
8894 ctx
.transaction
->empty() &&
8895 ctx
.created_pgs
.empty()) || !pg
) {
8896 delete ctx
.transaction
;
8897 delete ctx
.on_applied
;
8899 assert(ctx
.created_pgs
.empty());
8901 if (!ctx
.created_pgs
.empty()) {
8902 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8904 int tr
= store
->queue_transaction(
8906 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
, TrackedOpRef(),
8908 delete (ctx
.transaction
);
8914 * Send an MOSDPGNotify to a primary, with a list of PGs that I have
8915 * content for, and they are primary for.
8918 void OSD::do_notifies(
8919 map
<int,vector
<pair
<pg_notify_t
,PastIntervals
> > >& notify_list
,
8923 vector
<pair
<pg_notify_t
,PastIntervals
> > >::iterator it
=
8924 notify_list
.begin();
8925 it
!= notify_list
.end();
8927 if (!curmap
->is_up(it
->first
)) {
8928 dout(20) << __func__
<< " skipping down osd." << it
->first
<< dendl
;
8931 ConnectionRef con
= service
.get_con_osd_cluster(
8932 it
->first
, curmap
->get_epoch());
8934 dout(20) << __func__
<< " skipping osd." << it
->first
8935 << " (NULL con)" << dendl
;
8938 service
.share_map_peer(it
->first
, con
.get(), curmap
);
8939 dout(7) << __func__
<< " osd." << it
->first
8940 << " on " << it
->second
.size() << " PGs" << dendl
;
8941 MOSDPGNotify
*m
= new MOSDPGNotify(curmap
->get_epoch(),
8943 con
->send_message(m
);
8949 * send out pending queries for info | summaries
8951 void OSD::do_queries(map
<int, map
<spg_t
,pg_query_t
> >& query_map
,
8954 for (map
<int, map
<spg_t
,pg_query_t
> >::iterator pit
= query_map
.begin();
8955 pit
!= query_map
.end();
8957 if (!curmap
->is_up(pit
->first
)) {
8958 dout(20) << __func__
<< " skipping down osd." << pit
->first
<< dendl
;
8961 int who
= pit
->first
;
8962 ConnectionRef con
= service
.get_con_osd_cluster(who
, curmap
->get_epoch());
8964 dout(20) << __func__
<< " skipping osd." << who
8965 << " (NULL con)" << dendl
;
8968 service
.share_map_peer(who
, con
.get(), curmap
);
8969 dout(7) << __func__
<< " querying osd." << who
8970 << " on " << pit
->second
.size() << " PGs" << dendl
;
8971 MOSDPGQuery
*m
= new MOSDPGQuery(curmap
->get_epoch(), pit
->second
);
8972 con
->send_message(m
);
8977 void OSD::do_infos(map
<int,
8978 vector
<pair
<pg_notify_t
, PastIntervals
> > >& info_map
,
8982 vector
<pair
<pg_notify_t
, PastIntervals
> > >::iterator p
=
8984 p
!= info_map
.end();
8986 if (!curmap
->is_up(p
->first
)) {
8987 dout(20) << __func__
<< " skipping down osd." << p
->first
<< dendl
;
8990 for (vector
<pair
<pg_notify_t
,PastIntervals
> >::iterator i
= p
->second
.begin();
8991 i
!= p
->second
.end();
8993 dout(20) << __func__
<< " sending info " << i
->first
.info
8994 << " to shard " << p
->first
<< dendl
;
8996 ConnectionRef con
= service
.get_con_osd_cluster(
8997 p
->first
, curmap
->get_epoch());
8999 dout(20) << __func__
<< " skipping osd." << p
->first
9000 << " (NULL con)" << dendl
;
9003 service
.share_map_peer(p
->first
, con
.get(), curmap
);
9004 MOSDPGInfo
*m
= new MOSDPGInfo(curmap
->get_epoch());
9005 m
->pg_list
= p
->second
;
9006 con
->send_message(m
);
9013 * from non-primary to primary
9014 * includes pg_info_t.
9015 * NOTE: called with opqueue active.
9017 void OSD::handle_pg_notify(OpRequestRef op
)
9019 const MOSDPGNotify
*m
= static_cast<const MOSDPGNotify
*>(op
->get_req());
9020 assert(m
->get_type() == MSG_OSD_PG_NOTIFY
);
9022 dout(7) << "handle_pg_notify from " << m
->get_source() << dendl
;
9023 int from
= m
->get_source().num();
9025 if (!require_osd_peer(op
->get_req()))
9028 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9033 for (auto it
= m
->get_pg_list().begin();
9034 it
!= m
->get_pg_list().end();
9036 if (it
->first
.info
.pgid
.preferred() >= 0) {
9037 dout(20) << "ignoring localized pg " << it
->first
.info
.pgid
<< dendl
;
9041 handle_pg_peering_evt(
9042 spg_t(it
->first
.info
.pgid
.pgid
, it
->first
.to
),
9043 it
->first
.info
.history
, it
->second
,
9044 it
->first
.query_epoch
,
9045 PG::CephPeeringEvtRef(
9046 new PG::CephPeeringEvt(
9047 it
->first
.epoch_sent
, it
->first
.query_epoch
,
9048 PG::MNotifyRec(pg_shard_t(from
, it
->first
.from
), it
->first
,
9049 op
->get_req()->get_connection()->get_features())))
9054 void OSD::handle_pg_log(OpRequestRef op
)
9056 MOSDPGLog
*m
= static_cast<MOSDPGLog
*>(op
->get_nonconst_req());
9057 assert(m
->get_type() == MSG_OSD_PG_LOG
);
9058 dout(7) << "handle_pg_log " << *m
<< " from " << m
->get_source() << dendl
;
9060 if (!require_osd_peer(op
->get_req()))
9063 int from
= m
->get_source().num();
9064 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9067 if (m
->info
.pgid
.preferred() >= 0) {
9068 dout(10) << "ignoring localized pg " << m
->info
.pgid
<< dendl
;
9073 handle_pg_peering_evt(
9074 spg_t(m
->info
.pgid
.pgid
, m
->to
),
9075 m
->info
.history
, m
->past_intervals
, m
->get_epoch(),
9076 PG::CephPeeringEvtRef(
9077 new PG::CephPeeringEvt(
9078 m
->get_epoch(), m
->get_query_epoch(),
9079 PG::MLogRec(pg_shard_t(from
, m
->from
), m
)))
9083 void OSD::handle_pg_info(OpRequestRef op
)
9085 const MOSDPGInfo
*m
= static_cast<const MOSDPGInfo
*>(op
->get_req());
9086 assert(m
->get_type() == MSG_OSD_PG_INFO
);
9087 dout(7) << "handle_pg_info " << *m
<< " from " << m
->get_source() << dendl
;
9089 if (!require_osd_peer(op
->get_req()))
9092 int from
= m
->get_source().num();
9093 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9098 for (auto p
= m
->pg_list
.begin();
9099 p
!= m
->pg_list
.end();
9101 if (p
->first
.info
.pgid
.preferred() >= 0) {
9102 dout(10) << "ignoring localized pg " << p
->first
.info
.pgid
<< dendl
;
9106 handle_pg_peering_evt(
9107 spg_t(p
->first
.info
.pgid
.pgid
, p
->first
.to
),
9108 p
->first
.info
.history
, p
->second
, p
->first
.epoch_sent
,
9109 PG::CephPeeringEvtRef(
9110 new PG::CephPeeringEvt(
9111 p
->first
.epoch_sent
, p
->first
.query_epoch
,
9114 from
, p
->first
.from
), p
->first
.info
, p
->first
.epoch_sent
)))
9119 void OSD::handle_pg_trim(OpRequestRef op
)
9121 const MOSDPGTrim
*m
= static_cast<const MOSDPGTrim
*>(op
->get_req());
9122 assert(m
->get_type() == MSG_OSD_PG_TRIM
);
9124 dout(7) << "handle_pg_trim " << *m
<< " from " << m
->get_source() << dendl
;
9126 if (!require_osd_peer(op
->get_req()))
9129 int from
= m
->get_source().num();
9130 if (!require_same_or_newer_map(op
, m
->epoch
, false))
9133 if (m
->pgid
.preferred() >= 0) {
9134 dout(10) << "ignoring localized pg " << m
->pgid
<< dendl
;
9140 PG
*pg
= _lookup_lock_pg(m
->pgid
);
9142 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
9146 if (m
->epoch
< pg
->info
.history
.same_interval_since
) {
9147 dout(10) << *pg
<< " got old trim to " << m
->trim_to
<< ", ignoring" << dendl
;
9152 if (pg
->is_primary()) {
9153 // peer is informing us of their last_complete_ondisk
9154 dout(10) << *pg
<< " replica osd." << from
<< " lcod " << m
->trim_to
<< dendl
;
9155 pg
->peer_last_complete_ondisk
[pg_shard_t(from
, m
->pgid
.shard
)] =
9157 // trim log when the pg is recovered
9158 pg
->calc_min_last_complete_ondisk();
9160 // primary is instructing us to trim
9161 ObjectStore::Transaction t
;
9162 pg
->pg_log
.trim(m
->trim_to
, pg
->info
);
9163 pg
->dirty_info
= true;
9164 pg
->write_if_dirty(t
);
9165 int tr
= store
->queue_transaction(pg
->osr
.get(), std::move(t
), NULL
);
9171 void OSD::handle_pg_backfill_reserve(OpRequestRef op
)
9173 const MBackfillReserve
*m
= static_cast<const MBackfillReserve
*>(op
->get_req());
9174 assert(m
->get_type() == MSG_OSD_BACKFILL_RESERVE
);
9176 if (!require_osd_peer(op
->get_req()))
9178 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
9181 PG::CephPeeringEvtRef evt
;
9182 if (m
->type
== MBackfillReserve::REQUEST
) {
9183 evt
= PG::CephPeeringEvtRef(
9184 new PG::CephPeeringEvt(
9187 PG::RequestBackfillPrio(m
->priority
)));
9188 } else if (m
->type
== MBackfillReserve::GRANT
) {
9189 evt
= PG::CephPeeringEvtRef(
9190 new PG::CephPeeringEvt(
9193 PG::RemoteBackfillReserved()));
9194 } else if (m
->type
== MBackfillReserve::REJECT
) {
9195 // NOTE: this is replica -> primary "i reject your request"
9196 // and also primary -> replica "cancel my previously-granted request"
9197 evt
= PG::CephPeeringEvtRef(
9198 new PG::CephPeeringEvt(
9201 PG::RemoteReservationRejected()));
9206 if (service
.splitting(m
->pgid
)) {
9207 peering_wait_for_split
[m
->pgid
].push_back(evt
);
9211 PG
*pg
= _lookup_lock_pg(m
->pgid
);
9213 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
9217 pg
->queue_peering_event(evt
);
9221 void OSD::handle_pg_recovery_reserve(OpRequestRef op
)
9223 const MRecoveryReserve
*m
= static_cast<const MRecoveryReserve
*>(op
->get_req());
9224 assert(m
->get_type() == MSG_OSD_RECOVERY_RESERVE
);
9226 if (!require_osd_peer(op
->get_req()))
9228 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
9231 PG::CephPeeringEvtRef evt
;
9232 if (m
->type
== MRecoveryReserve::REQUEST
) {
9233 evt
= PG::CephPeeringEvtRef(
9234 new PG::CephPeeringEvt(
9237 PG::RequestRecovery()));
9238 } else if (m
->type
== MRecoveryReserve::GRANT
) {
9239 evt
= PG::CephPeeringEvtRef(
9240 new PG::CephPeeringEvt(
9243 PG::RemoteRecoveryReserved()));
9244 } else if (m
->type
== MRecoveryReserve::RELEASE
) {
9245 evt
= PG::CephPeeringEvtRef(
9246 new PG::CephPeeringEvt(
9249 PG::RecoveryDone()));
9254 if (service
.splitting(m
->pgid
)) {
9255 peering_wait_for_split
[m
->pgid
].push_back(evt
);
9259 PG
*pg
= _lookup_lock_pg(m
->pgid
);
9261 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
9265 pg
->queue_peering_event(evt
);
9269 void OSD::handle_force_recovery(Message
*m
)
9271 MOSDForceRecovery
*msg
= static_cast<MOSDForceRecovery
*>(m
);
9272 assert(msg
->get_type() == MSG_OSD_FORCE_RECOVERY
);
9274 vector
<PGRef
> local_pgs
;
9275 local_pgs
.reserve(msg
->forced_pgs
.size());
9278 RWLock::RLocker
l(pg_map_lock
);
9279 for (auto& i
: msg
->forced_pgs
) {
9281 if (osdmap
->get_primary_shard(i
, &locpg
)) {
9282 auto pg_map_entry
= pg_map
.find(locpg
);
9283 if (pg_map_entry
!= pg_map
.end()) {
9284 local_pgs
.push_back(pg_map_entry
->second
);
9290 if (local_pgs
.size()) {
9291 service
.adjust_pg_priorities(local_pgs
, msg
->options
);
9298 * from primary to replica | stray
9299 * NOTE: called with opqueue active.
9301 void OSD::handle_pg_query(OpRequestRef op
)
9303 assert(osd_lock
.is_locked());
9305 const MOSDPGQuery
*m
= static_cast<const MOSDPGQuery
*>(op
->get_req());
9306 assert(m
->get_type() == MSG_OSD_PG_QUERY
);
9308 if (!require_osd_peer(op
->get_req()))
9311 dout(7) << "handle_pg_query from " << m
->get_source() << " epoch " << m
->get_epoch() << dendl
;
9312 int from
= m
->get_source().num();
9314 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9319 map
< int, vector
<pair
<pg_notify_t
, PastIntervals
> > > notify_list
;
9321 for (auto it
= m
->pg_list
.begin();
9322 it
!= m
->pg_list
.end();
9324 spg_t pgid
= it
->first
;
9326 if (pgid
.preferred() >= 0) {
9327 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9331 if (service
.splitting(pgid
)) {
9332 peering_wait_for_split
[pgid
].push_back(
9333 PG::CephPeeringEvtRef(
9334 new PG::CephPeeringEvt(
9335 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9336 PG::MQuery(pg_shard_t(from
, it
->second
.from
),
9337 it
->second
, it
->second
.epoch_sent
))));
9342 RWLock::RLocker
l(pg_map_lock
);
9343 if (pg_map
.count(pgid
)) {
9345 pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9347 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9348 pg_shard_t(from
, it
->second
.from
), it
->second
);
9354 if (!osdmap
->have_pg_pool(pgid
.pool()))
9357 // get active crush mapping
9358 int up_primary
, acting_primary
;
9359 vector
<int> up
, acting
;
9360 osdmap
->pg_to_up_acting_osds(
9361 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9364 pg_history_t history
= it
->second
.history
;
9365 bool valid_history
= project_pg_history(
9366 pgid
, history
, it
->second
.epoch_sent
,
9367 up
, up_primary
, acting
, acting_primary
);
9369 if (!valid_history
||
9370 it
->second
.epoch_sent
< history
.same_interval_since
) {
9371 dout(10) << " pg " << pgid
<< " dne, and pg has changed in "
9372 << history
.same_interval_since
9373 << " (msg from " << it
->second
.epoch_sent
<< ")" << dendl
;
9377 dout(10) << " pg " << pgid
<< " dne" << dendl
;
9378 pg_info_t
empty(spg_t(pgid
.pgid
, it
->second
.to
));
9379 /* This is racy, but that should be ok: if we complete the deletion
9380 * before the pg is recreated, we'll just start it off backfilling
9381 * instead of just empty */
9382 if (service
.deleting_pgs
.lookup(pgid
))
9383 empty
.set_last_backfill(hobject_t());
9384 if (it
->second
.type
== pg_query_t::LOG
||
9385 it
->second
.type
== pg_query_t::FULLLOG
) {
9386 ConnectionRef con
= service
.get_con_osd_cluster(from
, osdmap
->get_epoch());
9388 MOSDPGLog
*mlog
= new MOSDPGLog(
9389 it
->second
.from
, it
->second
.to
,
9390 osdmap
->get_epoch(), empty
,
9391 it
->second
.epoch_sent
);
9392 service
.share_map_peer(from
, con
.get(), osdmap
);
9393 con
->send_message(mlog
);
9396 notify_list
[from
].push_back(
9399 it
->second
.from
, it
->second
.to
,
9400 it
->second
.epoch_sent
,
9401 osdmap
->get_epoch(),
9404 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
9408 do_notifies(notify_list
, osdmap
);
9412 void OSD::handle_pg_remove(OpRequestRef op
)
9414 const MOSDPGRemove
*m
= static_cast<const MOSDPGRemove
*>(op
->get_req());
9415 assert(m
->get_type() == MSG_OSD_PG_REMOVE
);
9416 assert(osd_lock
.is_locked());
9418 if (!require_osd_peer(op
->get_req()))
9421 dout(7) << "handle_pg_remove from " << m
->get_source() << " on "
9422 << m
->pg_list
.size() << " pgs" << dendl
;
9424 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9429 for (auto it
= m
->pg_list
.begin();
9430 it
!= m
->pg_list
.end();
9433 if (pgid
.preferred() >= 0) {
9434 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9438 RWLock::WLocker
l(pg_map_lock
);
9439 if (pg_map
.count(pgid
) == 0) {
9440 dout(10) << " don't have pg " << pgid
<< dendl
;
9443 dout(5) << "queue_pg_for_deletion: " << pgid
<< dendl
;
9444 PG
*pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9445 pg_history_t history
= pg
->info
.history
;
9446 int up_primary
, acting_primary
;
9447 vector
<int> up
, acting
;
9448 osdmap
->pg_to_up_acting_osds(
9449 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9450 bool valid_history
= project_pg_history(
9451 pg
->info
.pgid
, history
, pg
->get_osdmap()->get_epoch(),
9452 up
, up_primary
, acting
, acting_primary
);
9453 if (valid_history
&&
9454 history
.same_interval_since
<= m
->get_epoch()) {
9455 assert(pg
->get_primary().osd
== m
->get_source().num());
9460 dout(10) << *pg
<< " ignoring remove request, pg changed in epoch "
9461 << history
.same_interval_since
9462 << " > " << m
->get_epoch() << dendl
;
9468 void OSD::_remove_pg(PG
*pg
)
9470 ObjectStore::Transaction rmt
;
9472 // on_removal, which calls remove_watchers_and_notifies, and the erasure from
9473 // the pg_map must be done together without unlocking the pg lock,
9474 // to avoid racing with watcher cleanup in ms_handle_reset
9475 // and handle_notify_timeout
9476 pg
->on_removal(&rmt
);
9478 service
.cancel_pending_splits_for_parent(pg
->info
.pgid
);
9479 int tr
= store
->queue_transaction(
9480 pg
->osr
.get(), std::move(rmt
), NULL
,
9481 new ContainerContext
<
9482 SequencerRef
>(pg
->osr
));
9485 DeletingStateRef deleting
= service
.deleting_pgs
.lookup_or_create(
9491 remove_wq
.queue(make_pair(PGRef(pg
), deleting
));
9493 service
.pg_remove_epoch(pg
->info
.pgid
);
9495 // dereference from op_wq
9496 op_shardedwq
.clear_pg_pointer(pg
->info
.pgid
);
9499 pg_map
.erase(pg
->info
.pgid
);
9500 pg
->put("PGMap"); // since we've taken it out of map
9503 // =========================================================
9506 void OSDService::_maybe_queue_recovery() {
9507 assert(recovery_lock
.is_locked_by_me());
9508 uint64_t available_pushes
;
9509 while (!awaiting_throttle
.empty() &&
9510 _recover_now(&available_pushes
)) {
9511 uint64_t to_start
= MIN(
9513 cct
->_conf
->osd_recovery_max_single_start
);
9514 _queue_for_recovery(awaiting_throttle
.front(), to_start
);
9515 awaiting_throttle
.pop_front();
9516 recovery_ops_reserved
+= to_start
;
9520 bool OSDService::_recover_now(uint64_t *available_pushes
)
9522 if (available_pushes
)
9523 *available_pushes
= 0;
9525 if (ceph_clock_now() < defer_recovery_until
) {
9526 dout(15) << __func__
<< " defer until " << defer_recovery_until
<< dendl
;
9530 if (recovery_paused
) {
9531 dout(15) << __func__
<< " paused" << dendl
;
9535 uint64_t max
= cct
->_conf
->osd_recovery_max_active
;
9536 if (max
<= recovery_ops_active
+ recovery_ops_reserved
) {
9537 dout(15) << __func__
<< " active " << recovery_ops_active
9538 << " + reserved " << recovery_ops_reserved
9539 << " >= max " << max
<< dendl
;
9543 if (available_pushes
)
9544 *available_pushes
= max
- recovery_ops_active
- recovery_ops_reserved
;
9550 void OSDService::adjust_pg_priorities(const vector
<PGRef
>& pgs
, int newflags
)
9552 if (!pgs
.size() || !(newflags
& (OFR_BACKFILL
| OFR_RECOVERY
)))
9556 if (newflags
& OFR_BACKFILL
) {
9557 newstate
= PG_STATE_FORCED_BACKFILL
;
9558 } else if (newflags
& OFR_RECOVERY
) {
9559 newstate
= PG_STATE_FORCED_RECOVERY
;
9562 // debug output here may get large, don't generate it if debug level is below
9563 // 10 and use abbreviated pg ids otherwise
9564 if ((cct
)->_conf
->subsys
.should_gather(ceph_subsys_osd
, 10)) {
9567 for (auto& i
: pgs
) {
9568 ss
<< i
->get_pgid() << " ";
9571 dout(10) << __func__
<< " working on " << ss
.str() << dendl
;
9574 if (newflags
& OFR_CANCEL
) {
9575 for (auto& i
: pgs
) {
9577 i
->_change_recovery_force_mode(newstate
, true);
9581 for (auto& i
: pgs
) {
9582 // make sure the PG is in correct state before forcing backfill or recovery, or
9583 // else we'll make PG keeping FORCE_* flag forever, requiring osds restart
9584 // or forcing somehow recovery/backfill.
9586 int pgstate
= i
->get_state();
9587 if ( ((newstate
== PG_STATE_FORCED_RECOVERY
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_RECOVERY_WAIT
| PG_STATE_RECOVERING
))) ||
9588 ((newstate
== PG_STATE_FORCED_BACKFILL
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_BACKFILL_WAIT
| PG_STATE_BACKFILLING
))) )
9589 i
->_change_recovery_force_mode(newstate
, false);
9595 void OSD::do_recovery(
9596 PG
*pg
, epoch_t queued
, uint64_t reserved_pushes
,
9597 ThreadPool::TPHandle
&handle
)
9599 uint64_t started
= 0;
9602 * When the value of osd_recovery_sleep is set greater than zero, recovery
9603 * ops are scheduled after osd_recovery_sleep amount of time from the previous
9604 * recovery event's schedule time. This is done by adding a
9605 * recovery_requeue_callback event, which re-queues the recovery op using
9606 * queue_recovery_after_sleep.
9608 float recovery_sleep
= get_osd_recovery_sleep();
9610 Mutex::Locker
l(service
.recovery_sleep_lock
);
9611 if (recovery_sleep
> 0 && service
.recovery_needs_sleep
) {
9613 auto recovery_requeue_callback
= new FunctionContext([this, pgref
, queued
, reserved_pushes
](int r
) {
9614 dout(20) << "do_recovery wake up at "
9616 << ", re-queuing recovery" << dendl
;
9617 Mutex::Locker
l(service
.recovery_sleep_lock
);
9618 service
.recovery_needs_sleep
= false;
9619 service
.queue_recovery_after_sleep(pgref
.get(), queued
, reserved_pushes
);
9622 // This is true for the first recovery op and when the previous recovery op
9623 // has been scheduled in the past. The next recovery op is scheduled after
9624 // completing the sleep from now.
9625 if (service
.recovery_schedule_time
< ceph_clock_now()) {
9626 service
.recovery_schedule_time
= ceph_clock_now();
9628 service
.recovery_schedule_time
+= recovery_sleep
;
9629 service
.recovery_sleep_timer
.add_event_at(service
.recovery_schedule_time
,
9630 recovery_requeue_callback
);
9631 dout(20) << "Recovery event scheduled at "
9632 << service
.recovery_schedule_time
<< dendl
;
9639 Mutex::Locker
l(service
.recovery_sleep_lock
);
9640 service
.recovery_needs_sleep
= true;
9643 if (pg
->pg_has_reset_since(queued
)) {
9647 assert(!pg
->deleting
);
9648 assert(pg
->is_peered() && pg
->is_primary());
9650 assert(pg
->recovery_queued
);
9651 pg
->recovery_queued
= false;
9653 dout(10) << "do_recovery starting " << reserved_pushes
<< " " << *pg
<< dendl
;
9654 #ifdef DEBUG_RECOVERY_OIDS
9655 dout(20) << " active was " << service
.recovery_oids
[pg
->info
.pgid
] << dendl
;
9658 bool more
= pg
->start_recovery_ops(reserved_pushes
, handle
, &started
);
9659 dout(10) << "do_recovery started " << started
<< "/" << reserved_pushes
9660 << " on " << *pg
<< dendl
;
9662 // If no recovery op is started, don't bother to manipulate the RecoveryCtx
9663 if (!started
&& (more
|| !pg
->have_unfound())) {
9667 PG::RecoveryCtx rctx
= create_context();
9668 rctx
.handle
= &handle
;
9671 * if we couldn't start any recovery ops and things are still
9672 * unfound, see if we can discover more missing object locations.
9673 * It may be that our initial locations were bad and we errored
9674 * out while trying to pull.
9676 if (!more
&& pg
->have_unfound()) {
9677 pg
->discover_all_missing(*rctx
.query_map
);
9678 if (rctx
.query_map
->empty()) {
9680 if (pg
->state_test(PG_STATE_BACKFILLING
)) {
9681 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9684 PG::DeferBackfill(cct
->_conf
->osd_recovery_retry_interval
)));
9685 pg
->queue_peering_event(evt
);
9686 action
= "in backfill";
9687 } else if (pg
->state_test(PG_STATE_RECOVERING
)) {
9688 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9691 PG::DeferRecovery(cct
->_conf
->osd_recovery_retry_interval
)));
9692 pg
->queue_peering_event(evt
);
9693 action
= "in recovery";
9695 action
= "already out of recovery/backfill";
9697 dout(10) << __func__
<< ": no luck, giving up on this pg for now (" << action
<< ")" << dendl
;
9699 dout(10) << __func__
<< ": no luck, giving up on this pg for now (queue_recovery)" << dendl
;
9700 pg
->queue_recovery();
9704 pg
->write_if_dirty(*rctx
.transaction
);
9705 OSDMapRef curmap
= pg
->get_osdmap();
9706 dispatch_context(rctx
, pg
, curmap
);
9710 assert(started
<= reserved_pushes
);
9711 service
.release_reserved_pushes(reserved_pushes
);
9714 void OSDService::start_recovery_op(PG
*pg
, const hobject_t
& soid
)
9716 Mutex::Locker
l(recovery_lock
);
9717 dout(10) << "start_recovery_op " << *pg
<< " " << soid
9718 << " (" << recovery_ops_active
<< "/"
9719 << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9721 recovery_ops_active
++;
9723 #ifdef DEBUG_RECOVERY_OIDS
9724 dout(20) << " active was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9725 assert(recovery_oids
[pg
->info
.pgid
].count(soid
) == 0);
9726 recovery_oids
[pg
->info
.pgid
].insert(soid
);
9730 void OSDService::finish_recovery_op(PG
*pg
, const hobject_t
& soid
, bool dequeue
)
9732 Mutex::Locker
l(recovery_lock
);
9733 dout(10) << "finish_recovery_op " << *pg
<< " " << soid
9734 << " dequeue=" << dequeue
9735 << " (" << recovery_ops_active
<< "/" << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9739 assert(recovery_ops_active
> 0);
9740 recovery_ops_active
--;
9742 #ifdef DEBUG_RECOVERY_OIDS
9743 dout(20) << " active oids was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9744 assert(recovery_oids
[pg
->info
.pgid
].count(soid
));
9745 recovery_oids
[pg
->info
.pgid
].erase(soid
);
9748 _maybe_queue_recovery();
9751 bool OSDService::is_recovery_active()
9753 return local_reserver
.has_reservation() || remote_reserver
.has_reservation();
9756 // =========================================================
9759 bool OSD::op_is_discardable(const MOSDOp
*op
)
9761 // drop client request if they are not connected and can't get the
9763 if (!op
->get_connection()->is_connected()) {
9769 void OSD::enqueue_op(spg_t pg
, OpRequestRef
& op
, epoch_t epoch
)
9771 utime_t latency
= ceph_clock_now() - op
->get_req()->get_recv_stamp();
9772 dout(15) << "enqueue_op " << op
<< " prio " << op
->get_req()->get_priority()
9773 << " cost " << op
->get_req()->get_cost()
9774 << " latency " << latency
9775 << " epoch " << epoch
9776 << " " << *(op
->get_req()) << dendl
;
9777 op
->osd_trace
.event("enqueue op");
9778 op
->osd_trace
.keyval("priority", op
->get_req()->get_priority());
9779 op
->osd_trace
.keyval("cost", op
->get_req()->get_cost());
9780 op
->mark_queued_for_pg();
9781 logger
->tinc(l_osd_op_before_queue_op_lat
, latency
);
9782 op_shardedwq
.queue(make_pair(pg
, PGQueueable(op
, epoch
)));
9788 * NOTE: dequeue called in worker thread, with pg lock
9790 void OSD::dequeue_op(
9791 PGRef pg
, OpRequestRef op
,
9792 ThreadPool::TPHandle
&handle
)
9795 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_BEGIN", false);
9797 utime_t now
= ceph_clock_now();
9798 op
->set_dequeued_time(now
);
9799 utime_t latency
= now
- op
->get_req()->get_recv_stamp();
9800 dout(10) << "dequeue_op " << op
<< " prio " << op
->get_req()->get_priority()
9801 << " cost " << op
->get_req()->get_cost()
9802 << " latency " << latency
9803 << " " << *(op
->get_req())
9804 << " pg " << *pg
<< dendl
;
9806 logger
->tinc(l_osd_op_before_dequeue_op_lat
, latency
);
9808 Session
*session
= static_cast<Session
*>(
9809 op
->get_req()->get_connection()->get_priv());
9811 maybe_share_map(session
, op
, pg
->get_osdmap());
9818 op
->mark_reached_pg();
9819 op
->osd_trace
.event("dequeue_op");
9821 pg
->do_request(op
, handle
);
9824 dout(10) << "dequeue_op " << op
<< " finish" << dendl
;
9825 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_END", false);
9829 struct C_CompleteSplits
: public Context
{
9832 C_CompleteSplits(OSD
*osd
, const set
<PGRef
> &in
)
9833 : osd(osd
), pgs(in
) {}
9834 void finish(int r
) override
{
9835 Mutex::Locker
l(osd
->osd_lock
);
9836 if (osd
->is_stopping())
9838 PG::RecoveryCtx rctx
= osd
->create_context();
9839 for (set
<PGRef
>::iterator i
= pgs
.begin();
9842 osd
->pg_map_lock
.get_write();
9845 osd
->add_newly_split_pg(pg
, &rctx
);
9846 if (!((*i
)->deleting
)) {
9847 set
<spg_t
> to_complete
;
9848 to_complete
.insert((*i
)->info
.pgid
);
9849 osd
->service
.complete_split(to_complete
);
9851 osd
->pg_map_lock
.put_write();
9852 osd
->dispatch_context_transaction(rctx
, pg
);
9853 osd
->wake_pg_waiters(*i
);
9857 osd
->dispatch_context(rctx
, 0, osd
->service
.get_osdmap());
9861 void OSD::process_peering_events(
9862 const list
<PG
*> &pgs
,
9863 ThreadPool::TPHandle
&handle
9866 bool need_up_thru
= false;
9867 epoch_t same_interval_since
= 0;
9869 PG::RecoveryCtx rctx
= create_context();
9870 rctx
.handle
= &handle
;
9871 for (list
<PG
*>::const_iterator i
= pgs
.begin();
9874 set
<PGRef
> split_pgs
;
9876 pg
->lock_suspend_timeout(handle
);
9877 curmap
= service
.get_osdmap();
9882 if (!advance_pg(curmap
->get_epoch(), pg
, handle
, &rctx
, &split_pgs
)) {
9883 // we need to requeue the PG explicitly since we didn't actually
9885 peering_wq
.queue(pg
);
9887 assert(!pg
->peering_queue
.empty());
9888 PG::CephPeeringEvtRef evt
= pg
->peering_queue
.front();
9889 pg
->peering_queue
.pop_front();
9890 pg
->handle_peering_event(evt
, &rctx
);
9892 need_up_thru
= pg
->need_up_thru
|| need_up_thru
;
9893 same_interval_since
= MAX(pg
->info
.history
.same_interval_since
,
9894 same_interval_since
);
9895 pg
->write_if_dirty(*rctx
.transaction
);
9896 if (!split_pgs
.empty()) {
9897 rctx
.on_applied
->add(new C_CompleteSplits(this, split_pgs
));
9900 dispatch_context_transaction(rctx
, pg
, &handle
);
9904 queue_want_up_thru(same_interval_since
);
9905 dispatch_context(rctx
, 0, curmap
, &handle
);
9907 service
.send_pg_temp();
9910 // --------------------------------
9912 const char** OSD::get_tracked_conf_keys() const
9914 static const char* KEYS
[] = {
9915 "osd_max_backfills",
9916 "osd_min_recovery_priority",
9917 "osd_max_trimming_pgs",
9918 "osd_op_complaint_time",
9919 "osd_op_log_threshold",
9920 "osd_op_history_size",
9921 "osd_op_history_duration",
9922 "osd_op_history_slow_op_size",
9923 "osd_op_history_slow_op_threshold",
9924 "osd_enable_op_tracker",
9925 "osd_map_cache_size",
9926 "osd_map_max_advance",
9927 "osd_pg_epoch_persisted_max_stale",
9928 "osd_disk_thread_ioprio_class",
9929 "osd_disk_thread_ioprio_priority",
9930 // clog & admin clog
9933 "clog_to_syslog_facility",
9934 "clog_to_syslog_level",
9935 "osd_objectstore_fuse",
9937 "clog_to_graylog_host",
9938 "clog_to_graylog_port",
9941 "osd_recovery_delay_start",
9942 "osd_client_message_size_cap",
9943 "osd_client_message_cap",
9944 "osd_heartbeat_min_size",
9945 "osd_heartbeat_interval",
9951 void OSD::handle_conf_change(const struct md_config_t
*conf
,
9952 const std::set
<std::string
> &changed
)
9954 if (changed
.count("osd_max_backfills")) {
9955 service
.local_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9956 service
.remote_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9958 if (changed
.count("osd_min_recovery_priority")) {
9959 service
.local_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9960 service
.remote_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9962 if (changed
.count("osd_max_trimming_pgs")) {
9963 service
.snap_reserver
.set_max(cct
->_conf
->osd_max_trimming_pgs
);
9965 if (changed
.count("osd_op_complaint_time") ||
9966 changed
.count("osd_op_log_threshold")) {
9967 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
9968 cct
->_conf
->osd_op_log_threshold
);
9970 if (changed
.count("osd_op_history_size") ||
9971 changed
.count("osd_op_history_duration")) {
9972 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
9973 cct
->_conf
->osd_op_history_duration
);
9975 if (changed
.count("osd_op_history_slow_op_size") ||
9976 changed
.count("osd_op_history_slow_op_threshold")) {
9977 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
9978 cct
->_conf
->osd_op_history_slow_op_threshold
);
9980 if (changed
.count("osd_enable_op_tracker")) {
9981 op_tracker
.set_tracking(cct
->_conf
->osd_enable_op_tracker
);
9983 if (changed
.count("osd_disk_thread_ioprio_class") ||
9984 changed
.count("osd_disk_thread_ioprio_priority")) {
9985 set_disk_tp_priority();
9987 if (changed
.count("osd_map_cache_size")) {
9988 service
.map_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9989 service
.map_bl_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9990 service
.map_bl_inc_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9992 if (changed
.count("clog_to_monitors") ||
9993 changed
.count("clog_to_syslog") ||
9994 changed
.count("clog_to_syslog_level") ||
9995 changed
.count("clog_to_syslog_facility") ||
9996 changed
.count("clog_to_graylog") ||
9997 changed
.count("clog_to_graylog_host") ||
9998 changed
.count("clog_to_graylog_port") ||
9999 changed
.count("host") ||
10000 changed
.count("fsid")) {
10001 update_log_config();
10004 #ifdef HAVE_LIBFUSE
10005 if (changed
.count("osd_objectstore_fuse")) {
10007 enable_disable_fuse(false);
10012 if (changed
.count("osd_recovery_delay_start")) {
10013 service
.defer_recovery(cct
->_conf
->osd_recovery_delay_start
);
10014 service
.kick_recovery_queue();
10017 if (changed
.count("osd_client_message_cap")) {
10018 uint64_t newval
= cct
->_conf
->osd_client_message_cap
;
10019 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
10020 if (pol
.throttler_messages
&& newval
> 0) {
10021 pol
.throttler_messages
->reset_max(newval
);
10024 if (changed
.count("osd_client_message_size_cap")) {
10025 uint64_t newval
= cct
->_conf
->osd_client_message_size_cap
;
10026 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
10027 if (pol
.throttler_bytes
&& newval
> 0) {
10028 pol
.throttler_bytes
->reset_max(newval
);
10035 void OSD::update_log_config()
10037 map
<string
,string
> log_to_monitors
;
10038 map
<string
,string
> log_to_syslog
;
10039 map
<string
,string
> log_channel
;
10040 map
<string
,string
> log_prio
;
10041 map
<string
,string
> log_to_graylog
;
10042 map
<string
,string
> log_to_graylog_host
;
10043 map
<string
,string
> log_to_graylog_port
;
10047 if (parse_log_client_options(cct
, log_to_monitors
, log_to_syslog
,
10048 log_channel
, log_prio
, log_to_graylog
,
10049 log_to_graylog_host
, log_to_graylog_port
,
10051 clog
->update_config(log_to_monitors
, log_to_syslog
,
10052 log_channel
, log_prio
, log_to_graylog
,
10053 log_to_graylog_host
, log_to_graylog_port
,
10055 derr
<< "log_to_monitors " << log_to_monitors
<< dendl
;
10058 void OSD::check_config()
10060 // some sanity checks
10061 if (cct
->_conf
->osd_map_cache_size
<= cct
->_conf
->osd_map_max_advance
+ 2) {
10062 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
10063 << " is not > osd_map_max_advance ("
10064 << cct
->_conf
->osd_map_max_advance
<< ")";
10066 if (cct
->_conf
->osd_map_cache_size
<= (int)cct
->_conf
->osd_pg_epoch_persisted_max_stale
+ 2) {
10067 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
10068 << " is not > osd_pg_epoch_persisted_max_stale ("
10069 << cct
->_conf
->osd_pg_epoch_persisted_max_stale
<< ")";
10073 void OSD::set_disk_tp_priority()
10075 dout(10) << __func__
10076 << " class " << cct
->_conf
->osd_disk_thread_ioprio_class
10077 << " priority " << cct
->_conf
->osd_disk_thread_ioprio_priority
10079 if (cct
->_conf
->osd_disk_thread_ioprio_class
.empty() ||
10080 cct
->_conf
->osd_disk_thread_ioprio_priority
< 0)
10083 ceph_ioprio_string_to_class(cct
->_conf
->osd_disk_thread_ioprio_class
);
10085 derr
<< __func__
<< cpp_strerror(cls
) << ": "
10086 << "osd_disk_thread_ioprio_class is " << cct
->_conf
->osd_disk_thread_ioprio_class
10087 << " but only the following values are allowed: idle, be or rt" << dendl
;
10089 disk_tp
.set_ioprio(cls
, cct
->_conf
->osd_disk_thread_ioprio_priority
);
10092 // --------------------------------
10094 void OSD::get_latest_osdmap()
10096 dout(10) << __func__
<< " -- start" << dendl
;
10099 service
.objecter
->wait_for_latest_osdmap(&cond
);
10102 dout(10) << __func__
<< " -- finish" << dendl
;
10105 // --------------------------------
10107 int OSD::init_op_flags(OpRequestRef
& op
)
10109 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
10110 vector
<OSDOp
>::const_iterator iter
;
10112 // client flags have no bearing on whether an op is a read, write, etc.
10115 if (m
->has_flag(CEPH_OSD_FLAG_RWORDERED
)) {
10116 op
->set_force_rwordered();
10119 // set bits based on op codes, called methods.
10120 for (iter
= m
->ops
.begin(); iter
!= m
->ops
.end(); ++iter
) {
10121 if ((iter
->op
.op
== CEPH_OSD_OP_WATCH
&&
10122 iter
->op
.watch
.op
== CEPH_OSD_WATCH_OP_PING
)) {
10123 /* This a bit odd. PING isn't actually a write. It can't
10124 * result in an update to the object_info. PINGs also aren'ty
10125 * resent, so there's no reason to write out a log entry
10127 * However, we pipeline them behind writes, so let's force
10128 * the write_ordered flag.
10130 op
->set_force_rwordered();
10132 if (ceph_osd_op_mode_modify(iter
->op
.op
))
10135 if (ceph_osd_op_mode_read(iter
->op
.op
))
10138 // set READ flag if there are src_oids
10139 if (iter
->soid
.oid
.name
.length())
10142 // set PGOP flag if there are PG ops
10143 if (ceph_osd_op_type_pg(iter
->op
.op
))
10146 if (ceph_osd_op_mode_cache(iter
->op
.op
))
10149 // check for ec base pool
10150 int64_t poolid
= m
->get_pg().pool();
10151 const pg_pool_t
*pool
= osdmap
->get_pg_pool(poolid
);
10152 if (pool
&& pool
->is_tier()) {
10153 const pg_pool_t
*base_pool
= osdmap
->get_pg_pool(pool
->tier_of
);
10154 if (base_pool
&& base_pool
->require_rollback()) {
10155 if ((iter
->op
.op
!= CEPH_OSD_OP_READ
) &&
10156 (iter
->op
.op
!= CEPH_OSD_OP_CHECKSUM
) &&
10157 (iter
->op
.op
!= CEPH_OSD_OP_CMPEXT
) &&
10158 (iter
->op
.op
!= CEPH_OSD_OP_STAT
) &&
10159 (iter
->op
.op
!= CEPH_OSD_OP_ISDIRTY
) &&
10160 (iter
->op
.op
!= CEPH_OSD_OP_UNDIRTY
) &&
10161 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTR
) &&
10162 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTRS
) &&
10163 (iter
->op
.op
!= CEPH_OSD_OP_CMPXATTR
) &&
10164 (iter
->op
.op
!= CEPH_OSD_OP_ASSERT_VER
) &&
10165 (iter
->op
.op
!= CEPH_OSD_OP_LIST_WATCHERS
) &&
10166 (iter
->op
.op
!= CEPH_OSD_OP_LIST_SNAPS
) &&
10167 (iter
->op
.op
!= CEPH_OSD_OP_SETALLOCHINT
) &&
10168 (iter
->op
.op
!= CEPH_OSD_OP_WRITEFULL
) &&
10169 (iter
->op
.op
!= CEPH_OSD_OP_ROLLBACK
) &&
10170 (iter
->op
.op
!= CEPH_OSD_OP_CREATE
) &&
10171 (iter
->op
.op
!= CEPH_OSD_OP_DELETE
) &&
10172 (iter
->op
.op
!= CEPH_OSD_OP_SETXATTR
) &&
10173 (iter
->op
.op
!= CEPH_OSD_OP_RMXATTR
) &&
10174 (iter
->op
.op
!= CEPH_OSD_OP_STARTSYNC
) &&
10175 (iter
->op
.op
!= CEPH_OSD_OP_COPY_GET
) &&
10176 (iter
->op
.op
!= CEPH_OSD_OP_COPY_FROM
)) {
10182 switch (iter
->op
.op
) {
10183 case CEPH_OSD_OP_CALL
:
10185 bufferlist::iterator bp
= const_cast<bufferlist
&>(iter
->indata
).begin();
10186 int is_write
, is_read
;
10187 string cname
, mname
;
10188 bp
.copy(iter
->op
.cls
.class_len
, cname
);
10189 bp
.copy(iter
->op
.cls
.method_len
, mname
);
10191 ClassHandler::ClassData
*cls
;
10192 int r
= class_handler
->open_class(cname
, &cls
);
10194 derr
<< "class " << cname
<< " open got " << cpp_strerror(r
) << dendl
;
10197 else if (r
!= -EPERM
) // propagate permission errors
10201 int flags
= cls
->get_method_flags(mname
.c_str());
10203 if (flags
== -ENOENT
)
10209 is_read
= flags
& CLS_METHOD_RD
;
10210 is_write
= flags
& CLS_METHOD_WR
;
10211 bool is_promote
= flags
& CLS_METHOD_PROMOTE
;
10213 dout(10) << "class " << cname
<< " method " << mname
<< " "
10214 << "flags=" << (is_read
? "r" : "")
10215 << (is_write
? "w" : "")
10216 << (is_promote
? "p" : "")
10219 op
->set_class_read();
10221 op
->set_class_write();
10224 op
->add_class(cname
, is_read
, is_write
, cls
->whitelisted
);
10228 case CEPH_OSD_OP_WATCH
:
10229 // force the read bit for watch since it is depends on previous
10230 // watch state (and may return early if the watch exists) or, in
10231 // the case of ping, is simply a read op.
10234 case CEPH_OSD_OP_NOTIFY
:
10235 case CEPH_OSD_OP_NOTIFY_ACK
:
10241 case CEPH_OSD_OP_DELETE
:
10242 // if we get a delete with FAILOK we can skip handle cache. without
10243 // FAILOK we still need to promote (or do something smarter) to
10244 // determine whether to return ENOENT or 0.
10245 if (iter
== m
->ops
.begin() &&
10246 iter
->op
.flags
== CEPH_OSD_OP_FLAG_FAILOK
) {
10247 op
->set_skip_handle_cache();
10249 // skip promotion when proxying a delete op
10250 if (m
->ops
.size() == 1) {
10251 op
->set_skip_promote();
10255 case CEPH_OSD_OP_CACHE_TRY_FLUSH
:
10256 case CEPH_OSD_OP_CACHE_FLUSH
:
10257 case CEPH_OSD_OP_CACHE_EVICT
:
10258 // If try_flush/flush/evict is the only op, can skip handle cache.
10259 if (m
->ops
.size() == 1) {
10260 op
->set_skip_handle_cache();
10264 case CEPH_OSD_OP_READ
:
10265 case CEPH_OSD_OP_SYNC_READ
:
10266 case CEPH_OSD_OP_SPARSE_READ
:
10267 case CEPH_OSD_OP_CHECKSUM
:
10268 case CEPH_OSD_OP_WRITEFULL
:
10269 if (m
->ops
.size() == 1 &&
10270 (iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
||
10271 iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
)) {
10272 op
->set_skip_promote();
10276 // force promotion when pin an object in cache tier
10277 case CEPH_OSD_OP_CACHE_PIN
:
10286 if (op
->rmw_flags
== 0)
10292 void OSD::PeeringWQ::_dequeue(list
<PG
*> *out
) {
10293 for (list
<PG
*>::iterator i
= peering_queue
.begin();
10294 i
!= peering_queue
.end() &&
10295 out
->size() < osd
->cct
->_conf
->osd_peering_wq_batch_size
;
10297 if (in_use
.count(*i
)) {
10300 out
->push_back(*i
);
10301 peering_queue
.erase(i
++);
10304 in_use
.insert(out
->begin(), out
->end());
10308 // =============================================================
10310 #undef dout_context
10311 #define dout_context osd->cct
10313 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq "
10315 void OSD::ShardedOpWQ::wake_pg_waiters(spg_t pgid
)
10317 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10318 auto sdata
= shard_list
[shard_index
];
10319 bool queued
= false;
10321 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10322 auto p
= sdata
->pg_slots
.find(pgid
);
10323 if (p
!= sdata
->pg_slots
.end()) {
10324 dout(20) << __func__
<< " " << pgid
10325 << " to_process " << p
->second
.to_process
10326 << " waiting_for_pg=" << (int)p
->second
.waiting_for_pg
<< dendl
;
10327 for (auto i
= p
->second
.to_process
.rbegin();
10328 i
!= p
->second
.to_process
.rend();
10330 sdata
->_enqueue_front(make_pair(pgid
, *i
), osd
->op_prio_cutoff
);
10332 p
->second
.to_process
.clear();
10333 p
->second
.waiting_for_pg
= false;
10334 ++p
->second
.requeue_seq
;
10339 sdata
->sdata_lock
.Lock();
10340 sdata
->sdata_cond
.SignalOne();
10341 sdata
->sdata_lock
.Unlock();
10345 void OSD::ShardedOpWQ::prune_pg_waiters(OSDMapRef osdmap
, int whoami
)
10347 unsigned pushes_to_free
= 0;
10348 for (auto sdata
: shard_list
) {
10349 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10350 sdata
->waiting_for_pg_osdmap
= osdmap
;
10351 auto p
= sdata
->pg_slots
.begin();
10352 while (p
!= sdata
->pg_slots
.end()) {
10353 ShardData::pg_slot
& slot
= p
->second
;
10354 if (!slot
.to_process
.empty() && slot
.num_running
== 0) {
10355 if (osdmap
->is_up_acting_osd_shard(p
->first
, whoami
)) {
10356 dout(20) << __func__
<< " " << p
->first
<< " maps to us, keeping"
10361 while (!slot
.to_process
.empty() &&
10362 slot
.to_process
.front().get_map_epoch() <= osdmap
->get_epoch()) {
10363 auto& qi
= slot
.to_process
.front();
10364 dout(20) << __func__
<< " " << p
->first
10366 << " epoch " << qi
.get_map_epoch()
10367 << " <= " << osdmap
->get_epoch()
10368 << ", stale, dropping" << dendl
;
10369 pushes_to_free
+= qi
.get_reserved_pushes();
10370 slot
.to_process
.pop_front();
10373 if (slot
.to_process
.empty() &&
10374 slot
.num_running
== 0 &&
10376 dout(20) << __func__
<< " " << p
->first
<< " empty, pruning" << dendl
;
10377 p
= sdata
->pg_slots
.erase(p
);
10383 if (pushes_to_free
> 0) {
10384 osd
->service
.release_reserved_pushes(pushes_to_free
);
10388 void OSD::ShardedOpWQ::clear_pg_pointer(spg_t pgid
)
10390 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10391 auto sdata
= shard_list
[shard_index
];
10392 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10393 auto p
= sdata
->pg_slots
.find(pgid
);
10394 if (p
!= sdata
->pg_slots
.end()) {
10395 auto& slot
= p
->second
;
10396 dout(20) << __func__
<< " " << pgid
<< " pg " << slot
.pg
<< dendl
;
10397 assert(!slot
.pg
|| slot
.pg
->deleting
);
10402 void OSD::ShardedOpWQ::clear_pg_slots()
10404 for (auto sdata
: shard_list
) {
10405 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10406 sdata
->pg_slots
.clear();
10407 sdata
->waiting_for_pg_osdmap
.reset();
10408 // don't bother with reserved pushes; we are shutting down
10413 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq(" << shard_index << ") "
10415 void OSD::ShardedOpWQ::_process(uint32_t thread_index
, heartbeat_handle_d
*hb
)
10417 uint32_t shard_index
= thread_index
% num_shards
;
10418 ShardData
*sdata
= shard_list
[shard_index
];
10419 assert(NULL
!= sdata
);
10422 sdata
->sdata_op_ordering_lock
.Lock();
10423 if (sdata
->pqueue
->empty()) {
10424 dout(20) << __func__
<< " empty q, waiting" << dendl
;
10425 // optimistically sleep a moment; maybe another work item will come along.
10426 osd
->cct
->get_heartbeat_map()->reset_timeout(hb
,
10427 osd
->cct
->_conf
->threadpool_default_timeout
, 0);
10428 sdata
->sdata_lock
.Lock();
10429 sdata
->sdata_op_ordering_lock
.Unlock();
10430 sdata
->sdata_cond
.WaitInterval(sdata
->sdata_lock
,
10431 utime_t(osd
->cct
->_conf
->threadpool_empty_queue_max_wait
, 0));
10432 sdata
->sdata_lock
.Unlock();
10433 sdata
->sdata_op_ordering_lock
.Lock();
10434 if (sdata
->pqueue
->empty()) {
10435 sdata
->sdata_op_ordering_lock
.Unlock();
10439 pair
<spg_t
, PGQueueable
> item
= sdata
->pqueue
->dequeue();
10440 if (osd
->is_stopping()) {
10441 sdata
->sdata_op_ordering_lock
.Unlock();
10442 return; // OSD shutdown, discard.
10445 uint64_t requeue_seq
;
10447 auto& slot
= sdata
->pg_slots
[item
.first
];
10448 dout(30) << __func__
<< " " << item
.first
10449 << " to_process " << slot
.to_process
10450 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10451 slot
.to_process
.push_back(item
.second
);
10452 // note the requeue seq now...
10453 requeue_seq
= slot
.requeue_seq
;
10454 if (slot
.waiting_for_pg
) {
10455 // save ourselves a bit of effort
10456 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10457 << " queued, waiting_for_pg" << dendl
;
10458 sdata
->sdata_op_ordering_lock
.Unlock();
10462 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10463 << " queued" << dendl
;
10464 ++slot
.num_running
;
10466 sdata
->sdata_op_ordering_lock
.Unlock();
10468 osd
->service
.maybe_inject_dispatch_delay();
10470 // [lookup +] lock pg (if we have it)
10472 pg
= osd
->_lookup_lock_pg(item
.first
);
10477 osd
->service
.maybe_inject_dispatch_delay();
10479 boost::optional
<PGQueueable
> qi
;
10481 // we don't use a Mutex::Locker here because of the
10482 // osd->service.release_reserved_pushes() call below
10483 sdata
->sdata_op_ordering_lock
.Lock();
10485 auto q
= sdata
->pg_slots
.find(item
.first
);
10486 assert(q
!= sdata
->pg_slots
.end());
10487 auto& slot
= q
->second
;
10488 --slot
.num_running
;
10490 if (slot
.to_process
.empty()) {
10491 // raced with wake_pg_waiters or prune_pg_waiters
10492 dout(20) << __func__
<< " " << item
.first
<< " nothing queued" << dendl
;
10496 sdata
->sdata_op_ordering_lock
.Unlock();
10499 if (requeue_seq
!= slot
.requeue_seq
) {
10500 dout(20) << __func__
<< " " << item
.first
10501 << " requeue_seq " << slot
.requeue_seq
<< " > our "
10502 << requeue_seq
<< ", we raced with wake_pg_waiters"
10507 sdata
->sdata_op_ordering_lock
.Unlock();
10510 if (pg
&& !slot
.pg
&& !pg
->deleting
) {
10511 dout(20) << __func__
<< " " << item
.first
<< " set pg to " << pg
<< dendl
;
10514 dout(30) << __func__
<< " " << item
.first
<< " to_process " << slot
.to_process
10515 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10517 // make sure we're not already waiting for this pg
10518 if (slot
.waiting_for_pg
) {
10519 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10520 << " slot is waiting_for_pg" << dendl
;
10524 sdata
->sdata_op_ordering_lock
.Unlock();
10529 qi
= slot
.to_process
.front();
10530 slot
.to_process
.pop_front();
10531 dout(20) << __func__
<< " " << item
.first
<< " item " << *qi
10532 << " pg " << pg
<< dendl
;
10535 // should this pg shard exist on this osd in this (or a later) epoch?
10536 OSDMapRef osdmap
= sdata
->waiting_for_pg_osdmap
;
10537 if (osdmap
->is_up_acting_osd_shard(item
.first
, osd
->whoami
)) {
10538 dout(20) << __func__
<< " " << item
.first
10539 << " no pg, should exist, will wait" << " on " << *qi
<< dendl
;
10540 slot
.to_process
.push_front(*qi
);
10541 slot
.waiting_for_pg
= true;
10542 } else if (qi
->get_map_epoch() > osdmap
->get_epoch()) {
10543 dout(20) << __func__
<< " " << item
.first
<< " no pg, item epoch is "
10544 << qi
->get_map_epoch() << " > " << osdmap
->get_epoch()
10545 << ", will wait on " << *qi
<< dendl
;
10546 slot
.to_process
.push_front(*qi
);
10547 slot
.waiting_for_pg
= true;
10549 dout(20) << __func__
<< " " << item
.first
<< " no pg, shouldn't exist,"
10550 << " dropping " << *qi
<< dendl
;
10551 // share map with client?
10552 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10553 Session
*session
= static_cast<Session
*>(
10554 (*_op
)->get_req()->get_connection()->get_priv());
10556 osd
->maybe_share_map(session
, *_op
, sdata
->waiting_for_pg_osdmap
);
10560 unsigned pushes_to_free
= qi
->get_reserved_pushes();
10561 if (pushes_to_free
> 0) {
10562 sdata
->sdata_op_ordering_lock
.Unlock();
10563 osd
->service
.release_reserved_pushes(pushes_to_free
);
10567 sdata
->sdata_op_ordering_lock
.Unlock();
10570 sdata
->sdata_op_ordering_lock
.Unlock();
10573 // osd_opwq_process marks the point at which an operation has been dequeued
10574 // and will begin to be handled by a worker thread.
10578 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10579 reqid
= (*_op
)->get_reqid();
10582 tracepoint(osd
, opwq_process_start
, reqid
.name
._type
,
10583 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10586 lgeneric_subdout(osd
->cct
, osd
, 30) << "dequeue status: ";
10587 Formatter
*f
= Formatter::create("json");
10588 f
->open_object_section("q");
10590 f
->close_section();
10595 ThreadPool::TPHandle
tp_handle(osd
->cct
, hb
, timeout_interval
,
10597 qi
->run(osd
, pg
, tp_handle
);
10602 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10603 reqid
= (*_op
)->get_reqid();
10606 tracepoint(osd
, opwq_process_finish
, reqid
.name
._type
,
10607 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10613 void OSD::ShardedOpWQ::_enqueue(pair
<spg_t
, PGQueueable
> item
) {
10614 uint32_t shard_index
=
10615 item
.first
.hash_to_shard(shard_list
.size());
10617 ShardData
* sdata
= shard_list
[shard_index
];
10618 assert (NULL
!= sdata
);
10619 unsigned priority
= item
.second
.get_priority();
10620 unsigned cost
= item
.second
.get_cost();
10621 sdata
->sdata_op_ordering_lock
.Lock();
10623 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10624 if (priority
>= osd
->op_prio_cutoff
)
10625 sdata
->pqueue
->enqueue_strict(
10626 item
.second
.get_owner(), priority
, item
);
10628 sdata
->pqueue
->enqueue(
10629 item
.second
.get_owner(),
10630 priority
, cost
, item
);
10631 sdata
->sdata_op_ordering_lock
.Unlock();
10633 sdata
->sdata_lock
.Lock();
10634 sdata
->sdata_cond
.SignalOne();
10635 sdata
->sdata_lock
.Unlock();
10639 void OSD::ShardedOpWQ::_enqueue_front(pair
<spg_t
, PGQueueable
> item
)
10641 uint32_t shard_index
= item
.first
.hash_to_shard(shard_list
.size());
10642 ShardData
* sdata
= shard_list
[shard_index
];
10643 assert (NULL
!= sdata
);
10644 sdata
->sdata_op_ordering_lock
.Lock();
10645 auto p
= sdata
->pg_slots
.find(item
.first
);
10646 if (p
!= sdata
->pg_slots
.end() && !p
->second
.to_process
.empty()) {
10647 // we may be racing with _process, which has dequeued a new item
10648 // from pqueue, put it on to_process, and is now busy taking the
10649 // pg lock. ensure this old requeued item is ordered before any
10650 // such newer item in to_process.
10651 p
->second
.to_process
.push_front(item
.second
);
10652 item
.second
= p
->second
.to_process
.back();
10653 p
->second
.to_process
.pop_back();
10654 dout(20) << __func__
<< " " << item
.first
10655 << " " << p
->second
.to_process
.front()
10656 << " shuffled w/ " << item
.second
<< dendl
;
10658 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10660 sdata
->_enqueue_front(item
, osd
->op_prio_cutoff
);
10661 sdata
->sdata_op_ordering_lock
.Unlock();
10662 sdata
->sdata_lock
.Lock();
10663 sdata
->sdata_cond
.SignalOne();
10664 sdata
->sdata_lock
.Unlock();
10668 namespace osd_cmds
{
10670 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
)
10672 if (!ceph_using_tcmalloc()) {
10673 os
<< "could not issue heap profiler command -- not using tcmalloc!";
10674 return -EOPNOTSUPP
;
10678 if (!cmd_getval(&cct
, cmdmap
, "heapcmd", cmd
)) {
10679 os
<< "unable to get value for command \"" << cmd
<< "\"";
10683 std::vector
<std::string
> cmd_vec
;
10684 get_str_vec(cmd
, cmd_vec
);
10686 ceph_heap_profiler_handle_command(cmd_vec
, os
);
10691 }} // namespace ceph::osd_cmds
10694 std::ostream
& operator<<(std::ostream
& out
, const OSD::io_queue
& q
) {
10696 case OSD::io_queue::prioritized
:
10697 out
<< "prioritized";
10699 case OSD::io_queue::weightedpriority
:
10700 out
<< "weightedpriority";
10702 case OSD::io_queue::mclock_opclass
:
10703 out
<< "mclock_opclass";
10705 case OSD::io_queue::mclock_client
:
10706 out
<< "mclock_client";