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
<< ")" << dendl
;
1279 dout(20) << __func__
<< scrubs_pending
<< " + " << scrubs_active
<< " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1285 bool OSDService::inc_scrubs_pending()
1287 bool result
= false;
1289 sched_scrub_lock
.Lock();
1290 if (scrubs_pending
+ scrubs_active
< cct
->_conf
->osd_max_scrubs
) {
1291 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
+1)
1292 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1296 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " + " << scrubs_active
<< " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1298 sched_scrub_lock
.Unlock();
1303 void OSDService::dec_scrubs_pending()
1305 sched_scrub_lock
.Lock();
1306 dout(20) << "dec_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
-1)
1307 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1309 assert(scrubs_pending
>= 0);
1310 sched_scrub_lock
.Unlock();
1313 void OSDService::inc_scrubs_active(bool reserved
)
1315 sched_scrub_lock
.Lock();
1319 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1320 << " (max " << cct
->_conf
->osd_max_scrubs
1321 << ", pending " << (scrubs_pending
+1) << " -> " << scrubs_pending
<< ")" << dendl
;
1322 assert(scrubs_pending
>= 0);
1324 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1325 << " (max " << cct
->_conf
->osd_max_scrubs
1326 << ", pending " << scrubs_pending
<< ")" << dendl
;
1328 sched_scrub_lock
.Unlock();
1331 void OSDService::dec_scrubs_active()
1333 sched_scrub_lock
.Lock();
1334 dout(20) << "dec_scrubs_active " << scrubs_active
<< " -> " << (scrubs_active
-1)
1335 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", pending " << scrubs_pending
<< ")" << dendl
;
1337 assert(scrubs_active
>= 0);
1338 sched_scrub_lock
.Unlock();
1341 void OSDService::retrieve_epochs(epoch_t
*_boot_epoch
, epoch_t
*_up_epoch
,
1342 epoch_t
*_bind_epoch
) const
1344 Mutex::Locker
l(epoch_lock
);
1346 *_boot_epoch
= boot_epoch
;
1348 *_up_epoch
= up_epoch
;
1350 *_bind_epoch
= bind_epoch
;
1353 void OSDService::set_epochs(const epoch_t
*_boot_epoch
, const epoch_t
*_up_epoch
,
1354 const epoch_t
*_bind_epoch
)
1356 Mutex::Locker
l(epoch_lock
);
1358 assert(*_boot_epoch
== 0 || *_boot_epoch
>= boot_epoch
);
1359 boot_epoch
= *_boot_epoch
;
1362 assert(*_up_epoch
== 0 || *_up_epoch
>= up_epoch
);
1363 up_epoch
= *_up_epoch
;
1366 assert(*_bind_epoch
== 0 || *_bind_epoch
>= bind_epoch
);
1367 bind_epoch
= *_bind_epoch
;
1371 bool OSDService::prepare_to_stop()
1373 Mutex::Locker
l(is_stopping_lock
);
1374 if (get_state() != NOT_STOPPING
)
1377 OSDMapRef osdmap
= get_osdmap();
1378 if (osdmap
&& osdmap
->is_up(whoami
)) {
1379 dout(0) << __func__
<< " telling mon we are shutting down" << dendl
;
1380 set_state(PREPARING_TO_STOP
);
1381 monc
->send_mon_message(new MOSDMarkMeDown(monc
->get_fsid(),
1382 osdmap
->get_inst(whoami
),
1383 osdmap
->get_epoch(),
1386 utime_t now
= ceph_clock_now();
1388 timeout
.set_from_double(now
+ cct
->_conf
->osd_mon_shutdown_timeout
);
1389 while ((ceph_clock_now() < timeout
) &&
1390 (get_state() != STOPPING
)) {
1391 is_stopping_cond
.WaitUntil(is_stopping_lock
, timeout
);
1394 dout(0) << __func__
<< " starting shutdown" << dendl
;
1395 set_state(STOPPING
);
1399 void OSDService::got_stop_ack()
1401 Mutex::Locker
l(is_stopping_lock
);
1402 if (get_state() == PREPARING_TO_STOP
) {
1403 dout(0) << __func__
<< " starting shutdown" << dendl
;
1404 set_state(STOPPING
);
1405 is_stopping_cond
.Signal();
1407 dout(10) << __func__
<< " ignoring msg" << dendl
;
1411 MOSDMap
*OSDService::build_incremental_map_msg(epoch_t since
, epoch_t to
,
1412 OSDSuperblock
& sblock
)
1414 MOSDMap
*m
= new MOSDMap(monc
->get_fsid());
1415 m
->oldest_map
= max_oldest_map
;
1416 m
->newest_map
= sblock
.newest_map
;
1418 for (epoch_t e
= to
; e
> since
; e
--) {
1420 if (e
> m
->oldest_map
&& get_inc_map_bl(e
, bl
)) {
1421 m
->incremental_maps
[e
].claim(bl
);
1422 } else if (get_map_bl(e
, bl
)) {
1423 m
->maps
[e
].claim(bl
);
1426 derr
<< "since " << since
<< " to " << to
1427 << " oldest " << m
->oldest_map
<< " newest " << m
->newest_map
1437 void OSDService::send_map(MOSDMap
*m
, Connection
*con
)
1439 con
->send_message(m
);
1442 void OSDService::send_incremental_map(epoch_t since
, Connection
*con
,
1445 epoch_t to
= osdmap
->get_epoch();
1446 dout(10) << "send_incremental_map " << since
<< " -> " << to
1447 << " to " << con
<< " " << con
->get_peer_addr() << dendl
;
1451 OSDSuperblock
sblock(get_superblock());
1452 if (since
< sblock
.oldest_map
) {
1453 // just send latest full map
1454 MOSDMap
*m
= new MOSDMap(monc
->get_fsid());
1455 m
->oldest_map
= max_oldest_map
;
1456 m
->newest_map
= sblock
.newest_map
;
1457 get_map_bl(to
, m
->maps
[to
]);
1462 if (to
> since
&& (int64_t)(to
- since
) > cct
->_conf
->osd_map_share_max_epochs
) {
1463 dout(10) << " " << (to
- since
) << " > max " << cct
->_conf
->osd_map_share_max_epochs
1464 << ", only sending most recent" << dendl
;
1465 since
= to
- cct
->_conf
->osd_map_share_max_epochs
;
1468 if (to
- since
> (epoch_t
)cct
->_conf
->osd_map_message_max
)
1469 to
= since
+ cct
->_conf
->osd_map_message_max
;
1470 m
= build_incremental_map_msg(since
, to
, sblock
);
1475 bool OSDService::_get_map_bl(epoch_t e
, bufferlist
& bl
)
1477 bool found
= map_bl_cache
.lookup(e
, &bl
);
1480 logger
->inc(l_osd_map_bl_cache_hit
);
1484 logger
->inc(l_osd_map_bl_cache_miss
);
1485 found
= store
->read(coll_t::meta(),
1486 OSD::get_osdmap_pobject_name(e
), 0, 0, bl
,
1487 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1494 bool OSDService::get_inc_map_bl(epoch_t e
, bufferlist
& bl
)
1496 Mutex::Locker
l(map_cache_lock
);
1497 bool found
= map_bl_inc_cache
.lookup(e
, &bl
);
1500 logger
->inc(l_osd_map_bl_cache_hit
);
1504 logger
->inc(l_osd_map_bl_cache_miss
);
1505 found
= store
->read(coll_t::meta(),
1506 OSD::get_inc_osdmap_pobject_name(e
), 0, 0, bl
,
1507 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1509 _add_map_inc_bl(e
, bl
);
1514 void OSDService::_add_map_bl(epoch_t e
, bufferlist
& bl
)
1516 dout(10) << "add_map_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1517 // cache a contiguous buffer
1518 if (bl
.get_num_buffers() > 1) {
1521 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1522 map_bl_cache
.add(e
, bl
);
1525 void OSDService::_add_map_inc_bl(epoch_t e
, bufferlist
& bl
)
1527 dout(10) << "add_map_inc_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1528 // cache a contiguous buffer
1529 if (bl
.get_num_buffers() > 1) {
1532 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1533 map_bl_inc_cache
.add(e
, bl
);
1536 void OSDService::pin_map_inc_bl(epoch_t e
, bufferlist
&bl
)
1538 Mutex::Locker
l(map_cache_lock
);
1539 // cache a contiguous buffer
1540 if (bl
.get_num_buffers() > 1) {
1543 map_bl_inc_cache
.pin(e
, bl
);
1546 void OSDService::pin_map_bl(epoch_t e
, bufferlist
&bl
)
1548 Mutex::Locker
l(map_cache_lock
);
1549 // cache a contiguous buffer
1550 if (bl
.get_num_buffers() > 1) {
1553 map_bl_cache
.pin(e
, bl
);
1556 void OSDService::clear_map_bl_cache_pins(epoch_t e
)
1558 Mutex::Locker
l(map_cache_lock
);
1559 map_bl_inc_cache
.clear_pinned(e
);
1560 map_bl_cache
.clear_pinned(e
);
1563 OSDMapRef
OSDService::_add_map(OSDMap
*o
)
1565 epoch_t e
= o
->get_epoch();
1567 if (cct
->_conf
->osd_map_dedup
) {
1568 // Dedup against an existing map at a nearby epoch
1569 OSDMapRef for_dedup
= map_cache
.lower_bound(e
);
1571 OSDMap::dedup(for_dedup
.get(), o
);
1575 OSDMapRef l
= map_cache
.add(e
, o
, &existed
);
1582 OSDMapRef
OSDService::try_get_map(epoch_t epoch
)
1584 Mutex::Locker
l(map_cache_lock
);
1585 OSDMapRef retval
= map_cache
.lookup(epoch
);
1587 dout(30) << "get_map " << epoch
<< " -cached" << dendl
;
1589 logger
->inc(l_osd_map_cache_hit
);
1594 logger
->inc(l_osd_map_cache_miss
);
1595 epoch_t lb
= map_cache
.cached_key_lower_bound();
1597 dout(30) << "get_map " << epoch
<< " - miss, below lower bound" << dendl
;
1598 logger
->inc(l_osd_map_cache_miss_low
);
1599 logger
->inc(l_osd_map_cache_miss_low_avg
, lb
- epoch
);
1603 OSDMap
*map
= new OSDMap
;
1605 dout(20) << "get_map " << epoch
<< " - loading and decoding " << map
<< dendl
;
1607 if (!_get_map_bl(epoch
, bl
) || bl
.length() == 0) {
1608 derr
<< "failed to load OSD map for epoch " << epoch
<< ", got " << bl
.length() << " bytes" << dendl
;
1614 dout(20) << "get_map " << epoch
<< " - return initial " << map
<< dendl
;
1616 return _add_map(map
);
1622 void OSDService::reply_op_error(OpRequestRef op
, int err
)
1624 reply_op_error(op
, err
, eversion_t(), 0);
1627 void OSDService::reply_op_error(OpRequestRef op
, int err
, eversion_t v
,
1630 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1631 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1633 flags
= m
->get_flags() & (CEPH_OSD_FLAG_ACK
|CEPH_OSD_FLAG_ONDISK
);
1635 MOSDOpReply
*reply
= new MOSDOpReply(m
, err
, osdmap
->get_epoch(), flags
,
1637 reply
->set_reply_versions(v
, uv
);
1638 m
->get_connection()->send_message(reply
);
1641 void OSDService::handle_misdirected_op(PG
*pg
, OpRequestRef op
)
1643 if (!cct
->_conf
->osd_debug_misdirected_ops
) {
1647 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1648 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1650 assert(m
->get_map_epoch() >= pg
->info
.history
.same_primary_since
);
1652 if (pg
->is_ec_pg()) {
1654 * OSD recomputes op target based on current OSDMap. With an EC pg, we
1655 * can get this result:
1656 * 1) client at map 512 sends an op to osd 3, pg_t 3.9 based on mapping
1657 * [CRUSH_ITEM_NONE, 2, 3]/3
1658 * 2) OSD 3 at map 513 remaps op to osd 3, spg_t 3.9s0 based on mapping
1660 * 3) PG 3.9s0 dequeues the op at epoch 512 and notices that it isn't primary
1662 * 4) client resends and this time PG 3.9s0 having caught up to 513 gets
1665 * We can't compute the op target based on the sending map epoch due to
1666 * splitting. The simplest thing is to detect such cases here and drop
1667 * them without an error (the client will resend anyway).
1669 assert(m
->get_map_epoch() <= superblock
.newest_map
);
1670 OSDMapRef opmap
= try_get_map(m
->get_map_epoch());
1672 dout(7) << __func__
<< ": " << *pg
<< " no longer have map for "
1673 << m
->get_map_epoch() << ", dropping" << dendl
;
1676 pg_t _pgid
= m
->get_raw_pg();
1678 if ((m
->get_flags() & CEPH_OSD_FLAG_PGOP
) == 0)
1679 _pgid
= opmap
->raw_pg_to_pg(_pgid
);
1680 if (opmap
->get_primary_shard(_pgid
, &pgid
) &&
1681 pgid
.shard
!= pg
->info
.pgid
.shard
) {
1682 dout(7) << __func__
<< ": " << *pg
<< " primary changed since "
1683 << m
->get_map_epoch() << ", dropping" << dendl
;
1688 dout(7) << *pg
<< " misdirected op in " << m
->get_map_epoch() << dendl
;
1689 clog
->warn() << m
->get_source_inst() << " misdirected " << m
->get_reqid()
1690 << " pg " << m
->get_raw_pg()
1691 << " to osd." << whoami
1692 << " not " << pg
->acting
1693 << " in e" << m
->get_map_epoch() << "/" << osdmap
->get_epoch();
1696 void OSDService::enqueue_back(spg_t pgid
, PGQueueable qi
)
1698 osd
->op_shardedwq
.queue(make_pair(pgid
, qi
));
1701 void OSDService::enqueue_front(spg_t pgid
, PGQueueable qi
)
1703 osd
->op_shardedwq
.queue_front(make_pair(pgid
, qi
));
1706 void OSDService::queue_for_peering(PG
*pg
)
1708 peering_wq
.queue(pg
);
1711 void OSDService::queue_for_snap_trim(PG
*pg
)
1713 dout(10) << "queueing " << *pg
<< " for snaptrim" << dendl
;
1714 osd
->op_shardedwq
.queue(
1718 PGSnapTrim(pg
->get_osdmap()->get_epoch()),
1719 cct
->_conf
->osd_snap_trim_cost
,
1720 cct
->_conf
->osd_snap_trim_priority
,
1723 pg
->get_osdmap()->get_epoch())));
1727 // ====================================================================
1731 #define dout_prefix *_dout
1733 // Commands shared between OSD's console and admin console:
1735 namespace osd_cmds
{
1737 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
);
1739 }} // namespace ceph::osd_cmds
1741 int OSD::mkfs(CephContext
*cct
, ObjectStore
*store
, const string
&dev
,
1742 uuid_d fsid
, int whoami
)
1746 ceph::shared_ptr
<ObjectStore::Sequencer
> osr(
1747 new ObjectStore::Sequencer("mkfs"));
1752 // if we are fed a uuid for this osd, use it.
1753 store
->set_fsid(cct
->_conf
->osd_uuid
);
1755 ret
= store
->mkfs();
1757 derr
<< "OSD::mkfs: ObjectStore::mkfs failed with error "
1758 << cpp_strerror(ret
) << dendl
;
1762 store
->set_cache_shards(1); // doesn't matter for mkfs!
1764 ret
= store
->mount();
1766 derr
<< "OSD::mkfs: couldn't mount ObjectStore: error "
1767 << cpp_strerror(ret
) << dendl
;
1771 ret
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, sbbl
);
1773 /* if we already have superblock, check content of superblock */
1774 dout(0) << " have superblock" << dendl
;
1775 bufferlist::iterator p
;
1778 if (whoami
!= sb
.whoami
) {
1779 derr
<< "provided osd id " << whoami
<< " != superblock's " << sb
.whoami
1784 if (fsid
!= sb
.cluster_fsid
) {
1785 derr
<< "provided cluster fsid " << fsid
1786 << " != superblock's " << sb
.cluster_fsid
<< dendl
;
1791 // create superblock
1792 sb
.cluster_fsid
= fsid
;
1793 sb
.osd_fsid
= store
->get_fsid();
1795 sb
.compat_features
= get_osd_initial_compat_set();
1800 ObjectStore::Transaction t
;
1801 t
.create_collection(coll_t::meta(), 0);
1802 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
1803 ret
= store
->apply_transaction(osr
.get(), std::move(t
));
1805 derr
<< "OSD::mkfs: error while writing OSD_SUPERBLOCK_GOBJECT: "
1806 << "apply_transaction returned " << cpp_strerror(ret
) << dendl
;
1811 if (!osr
->flush_commit(&waiter
)) {
1815 ret
= write_meta(cct
, store
, sb
.cluster_fsid
, sb
.osd_fsid
, whoami
);
1817 derr
<< "OSD::mkfs: failed to write fsid file: error "
1818 << cpp_strerror(ret
) << dendl
;
1829 int OSD::write_meta(CephContext
*cct
, ObjectStore
*store
, uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int whoami
)
1834 snprintf(val
, sizeof(val
), "%s", CEPH_OSD_ONDISK_MAGIC
);
1835 r
= store
->write_meta("magic", val
);
1839 snprintf(val
, sizeof(val
), "%d", whoami
);
1840 r
= store
->write_meta("whoami", val
);
1844 cluster_fsid
.print(val
);
1845 r
= store
->write_meta("ceph_fsid", val
);
1849 string key
= cct
->_conf
->get_val
<string
>("key");
1851 r
= store
->write_meta("osd_key", key
);
1855 string keyfile
= cct
->_conf
->get_val
<string
>("keyfile");
1856 if (!keyfile
.empty()) {
1859 if (keyfile
== "-") {
1860 static_assert(1024 * 1024 >
1861 (sizeof(CryptoKey
) - sizeof(bufferptr
) +
1862 sizeof(__u16
) + 16 /* AES_KEY_LEN */ + 3 - 1) / 3. * 4.,
1863 "1MB should be enough for a base64 encoded CryptoKey");
1864 r
= keybl
.read_fd(STDIN_FILENO
, 1024 * 1024);
1866 r
= keybl
.read_file(keyfile
.c_str(), &err
);
1869 derr
<< __func__
<< " failed to read keyfile " << keyfile
<< ": "
1870 << err
<< ": " << cpp_strerror(r
) << dendl
;
1873 r
= store
->write_meta("osd_key", keybl
.to_str());
1879 r
= store
->write_meta("ready", "ready");
1886 int OSD::peek_meta(ObjectStore
*store
, std::string
& magic
,
1887 uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int& whoami
)
1891 int r
= store
->read_meta("magic", &val
);
1896 r
= store
->read_meta("whoami", &val
);
1899 whoami
= atoi(val
.c_str());
1901 r
= store
->read_meta("ceph_fsid", &val
);
1904 r
= cluster_fsid
.parse(val
.c_str());
1908 r
= store
->read_meta("fsid", &val
);
1910 osd_fsid
= uuid_d();
1912 r
= osd_fsid
.parse(val
.c_str());
1922 #define dout_prefix _prefix(_dout, whoami, get_osdmap_epoch())
1926 OSD::OSD(CephContext
*cct_
, ObjectStore
*store_
,
1928 Messenger
*internal_messenger
,
1929 Messenger
*external_messenger
,
1930 Messenger
*hb_client_front
,
1931 Messenger
*hb_client_back
,
1932 Messenger
*hb_front_serverm
,
1933 Messenger
*hb_back_serverm
,
1934 Messenger
*osdc_messenger
,
1936 const std::string
&dev
, const std::string
&jdev
) :
1938 osd_lock("OSD::osd_lock"),
1939 tick_timer(cct
, osd_lock
),
1940 tick_timer_lock("OSD::tick_timer_lock"),
1941 tick_timer_without_osd_lock(cct
, tick_timer_lock
),
1942 authorize_handler_cluster_registry(new AuthAuthorizeHandlerRegistry(cct
,
1943 cct
->_conf
->auth_supported
.empty() ?
1944 cct
->_conf
->auth_cluster_required
:
1945 cct
->_conf
->auth_supported
)),
1946 authorize_handler_service_registry(new AuthAuthorizeHandlerRegistry(cct
,
1947 cct
->_conf
->auth_supported
.empty() ?
1948 cct
->_conf
->auth_service_required
:
1949 cct
->_conf
->auth_supported
)),
1950 cluster_messenger(internal_messenger
),
1951 client_messenger(external_messenger
),
1952 objecter_messenger(osdc_messenger
),
1954 mgrc(cct_
, client_messenger
),
1956 recoverystate_perf(NULL
),
1958 log_client(cct
, client_messenger
, &mc
->monmap
, LogClient::NO_FLAGS
),
1959 clog(log_client
.create_channel()),
1961 dev_path(dev
), journal_path(jdev
),
1962 store_is_rotational(store
->is_rotational()),
1963 trace_endpoint("0.0.0.0", 0, "osd"),
1965 osd_compat(get_osd_compat_set()),
1966 peering_tp(cct
, "OSD::peering_tp", "tp_peering",
1967 cct
->_conf
->osd_peering_wq_threads
,
1968 "osd_peering_tp_threads"),
1969 osd_op_tp(cct
, "OSD::osd_op_tp", "tp_osd_tp",
1970 get_num_op_threads()),
1971 disk_tp(cct
, "OSD::disk_tp", "tp_osd_disk", cct
->_conf
->osd_disk_threads
, "osd_disk_threads"),
1972 command_tp(cct
, "OSD::command_tp", "tp_osd_cmd", 1),
1973 session_waiting_lock("OSD::session_waiting_lock"),
1974 osdmap_subscribe_lock("OSD::osdmap_subscribe_lock"),
1975 heartbeat_lock("OSD::heartbeat_lock"),
1976 heartbeat_stop(false),
1977 heartbeat_need_update(true),
1978 hb_front_client_messenger(hb_client_front
),
1979 hb_back_client_messenger(hb_client_back
),
1980 hb_front_server_messenger(hb_front_serverm
),
1981 hb_back_server_messenger(hb_back_serverm
),
1983 heartbeat_thread(this),
1984 heartbeat_dispatcher(this),
1985 op_tracker(cct
, cct
->_conf
->osd_enable_op_tracker
,
1986 cct
->_conf
->osd_num_op_tracker_shard
),
1987 test_ops_hook(NULL
),
1988 op_queue(get_io_queue()),
1989 op_prio_cutoff(get_io_prio_cut()),
1991 get_num_op_shards(),
1993 cct
->_conf
->osd_op_thread_timeout
,
1994 cct
->_conf
->osd_op_thread_suicide_timeout
,
1998 cct
->_conf
->osd_op_thread_timeout
,
1999 cct
->_conf
->osd_op_thread_suicide_timeout
,
2001 map_lock("OSD::map_lock"),
2002 pg_map_lock("OSD::pg_map_lock"),
2003 last_pg_create_epoch(0),
2004 mon_report_lock("OSD::mon_report_lock"),
2005 stats_ack_timeout(cct
->_conf
->osd_mon_ack_timeout
),
2007 requested_full_first(0),
2008 requested_full_last(0),
2009 pg_stat_queue_lock("OSD::pg_stat_queue_lock"),
2010 osd_stat_updated(false),
2011 pg_stat_tid(0), pg_stat_tid_flushed(0),
2014 cct
->_conf
->osd_command_thread_timeout
,
2015 cct
->_conf
->osd_command_thread_suicide_timeout
,
2020 cct
->_conf
->osd_remove_thread_timeout
,
2021 cct
->_conf
->osd_remove_thread_suicide_timeout
,
2025 monc
->set_messenger(client_messenger
);
2026 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
2027 cct
->_conf
->osd_op_log_threshold
);
2028 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
2029 cct
->_conf
->osd_op_history_duration
);
2030 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
2031 cct
->_conf
->osd_op_history_slow_op_threshold
);
2033 std::stringstream ss
;
2034 ss
<< "osd." << whoami
;
2035 trace_endpoint
.copy_name(ss
.str());
2041 delete authorize_handler_cluster_registry
;
2042 delete authorize_handler_service_registry
;
2043 delete class_handler
;
2044 cct
->get_perfcounters_collection()->remove(recoverystate_perf
);
2045 cct
->get_perfcounters_collection()->remove(logger
);
2046 delete recoverystate_perf
;
2051 void cls_initialize(ClassHandler
*ch
);
2053 void OSD::handle_signal(int signum
)
2055 assert(signum
== SIGINT
|| signum
== SIGTERM
);
2056 derr
<< "*** Got signal " << sig_str(signum
) << " ***" << dendl
;
2062 Mutex::Locker
lock(osd_lock
);
2066 if (store
->test_mount_in_use()) {
2067 derr
<< "OSD::pre_init: object store '" << dev_path
<< "' is "
2068 << "currently in use. (Is ceph-osd already running?)" << dendl
;
2072 cct
->_conf
->add_observer(this);
2078 class OSDSocketHook
: public AdminSocketHook
{
2081 explicit OSDSocketHook(OSD
*o
) : osd(o
) {}
2082 bool call(std::string admin_command
, cmdmap_t
& cmdmap
, std::string format
,
2083 bufferlist
& out
) override
{
2085 bool r
= osd
->asok_command(admin_command
, cmdmap
, format
, ss
);
2091 bool OSD::asok_command(string admin_command
, cmdmap_t
& cmdmap
, string format
,
2094 Formatter
*f
= Formatter::create(format
, "json-pretty", "json-pretty");
2095 if (admin_command
== "status") {
2096 f
->open_object_section("status");
2097 f
->dump_stream("cluster_fsid") << superblock
.cluster_fsid
;
2098 f
->dump_stream("osd_fsid") << superblock
.osd_fsid
;
2099 f
->dump_unsigned("whoami", superblock
.whoami
);
2100 f
->dump_string("state", get_state_name(get_state()));
2101 f
->dump_unsigned("oldest_map", superblock
.oldest_map
);
2102 f
->dump_unsigned("newest_map", superblock
.newest_map
);
2104 RWLock::RLocker
l(pg_map_lock
);
2105 f
->dump_unsigned("num_pgs", pg_map
.size());
2108 } else if (admin_command
== "flush_journal") {
2109 store
->flush_journal();
2110 } else if (admin_command
== "dump_ops_in_flight" ||
2111 admin_command
== "ops" ||
2112 admin_command
== "dump_blocked_ops" ||
2113 admin_command
== "dump_historic_ops" ||
2114 admin_command
== "dump_historic_ops_by_duration" ||
2115 admin_command
== "dump_historic_slow_ops") {
2117 const string error_str
= "op_tracker tracking is not enabled now, so no ops are tracked currently, \
2118 even those get stuck. Please enable \"osd_enable_op_tracker\", and the tracker \
2119 will start to track new ops received afterwards.";
2121 set
<string
> filters
;
2122 vector
<string
> filter_str
;
2123 if (cmd_getval(cct
, cmdmap
, "filterstr", filter_str
)) {
2124 copy(filter_str
.begin(), filter_str
.end(),
2125 inserter(filters
, filters
.end()));
2128 if (admin_command
== "dump_ops_in_flight" ||
2129 admin_command
== "ops") {
2130 if (!op_tracker
.dump_ops_in_flight(f
, false, filters
)) {
2134 if (admin_command
== "dump_blocked_ops") {
2135 if (!op_tracker
.dump_ops_in_flight(f
, true, filters
)) {
2139 if (admin_command
== "dump_historic_ops") {
2140 if (!op_tracker
.dump_historic_ops(f
, false, filters
)) {
2144 if (admin_command
== "dump_historic_ops_by_duration") {
2145 if (!op_tracker
.dump_historic_ops(f
, true, filters
)) {
2149 if (admin_command
== "dump_historic_slow_ops") {
2150 if (!op_tracker
.dump_historic_slow_ops(f
, filters
)) {
2154 } else if (admin_command
== "dump_op_pq_state") {
2155 f
->open_object_section("pq");
2156 op_shardedwq
.dump(f
);
2158 } else if (admin_command
== "dump_blacklist") {
2159 list
<pair
<entity_addr_t
,utime_t
> > bl
;
2160 OSDMapRef curmap
= service
.get_osdmap();
2162 f
->open_array_section("blacklist");
2163 curmap
->get_blacklist(&bl
);
2164 for (list
<pair
<entity_addr_t
,utime_t
> >::iterator it
= bl
.begin();
2165 it
!= bl
.end(); ++it
) {
2166 f
->open_object_section("entry");
2167 f
->open_object_section("entity_addr_t");
2169 f
->close_section(); //entity_addr_t
2170 it
->second
.localtime(f
->dump_stream("expire_time"));
2171 f
->close_section(); //entry
2173 f
->close_section(); //blacklist
2174 } else if (admin_command
== "dump_watchers") {
2175 list
<obj_watch_item_t
> watchers
;
2178 Mutex::Locker
l(osd_lock
);
2179 RWLock::RLocker
l2(pg_map_lock
);
2180 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2184 list
<obj_watch_item_t
> pg_watchers
;
2185 PG
*pg
= it
->second
;
2187 pg
->get_watchers(pg_watchers
);
2189 watchers
.splice(watchers
.end(), pg_watchers
);
2193 f
->open_array_section("watchers");
2194 for (list
<obj_watch_item_t
>::iterator it
= watchers
.begin();
2195 it
!= watchers
.end(); ++it
) {
2197 f
->open_object_section("watch");
2199 f
->dump_string("namespace", it
->obj
.nspace
);
2200 f
->dump_string("object", it
->obj
.oid
.name
);
2202 f
->open_object_section("entity_name");
2203 it
->wi
.name
.dump(f
);
2204 f
->close_section(); //entity_name_t
2206 f
->dump_unsigned("cookie", it
->wi
.cookie
);
2207 f
->dump_unsigned("timeout", it
->wi
.timeout_seconds
);
2209 f
->open_object_section("entity_addr_t");
2210 it
->wi
.addr
.dump(f
);
2211 f
->close_section(); //entity_addr_t
2213 f
->close_section(); //watch
2216 f
->close_section(); //watchers
2217 } else if (admin_command
== "dump_reservations") {
2218 f
->open_object_section("reservations");
2219 f
->open_object_section("local_reservations");
2220 service
.local_reserver
.dump(f
);
2222 f
->open_object_section("remote_reservations");
2223 service
.remote_reserver
.dump(f
);
2226 } else if (admin_command
== "get_latest_osdmap") {
2227 get_latest_osdmap();
2228 } else if (admin_command
== "heap") {
2229 auto result
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ss
);
2231 // Note: Failed heap profile commands won't necessarily trigger an error:
2232 f
->open_object_section("result");
2233 f
->dump_string("error", cpp_strerror(result
));
2234 f
->dump_bool("success", result
>= 0);
2236 } else if (admin_command
== "set_heap_property") {
2240 bool success
= false;
2241 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2242 error
= "unable to get property";
2244 } else if (!cmd_getval(cct
, cmdmap
, "value", value
)) {
2245 error
= "unable to get value";
2247 } else if (value
< 0) {
2248 error
= "negative value not allowed";
2250 } else if (!ceph_heap_set_numeric_property(property
.c_str(), (size_t)value
)) {
2251 error
= "invalid property";
2256 f
->open_object_section("result");
2257 f
->dump_string("error", error
);
2258 f
->dump_bool("success", success
);
2260 } else if (admin_command
== "get_heap_property") {
2264 bool success
= false;
2265 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2266 error
= "unable to get property";
2268 } else if (!ceph_heap_get_numeric_property(property
.c_str(), &value
)) {
2269 error
= "invalid property";
2274 f
->open_object_section("result");
2275 f
->dump_string("error", error
);
2276 f
->dump_bool("success", success
);
2277 f
->dump_int("value", value
);
2279 } else if (admin_command
== "dump_objectstore_kv_stats") {
2280 store
->get_db_statistics(f
);
2281 } else if (admin_command
== "dump_scrubs") {
2282 service
.dumps_scrub(f
);
2283 } else if (admin_command
== "calc_objectstore_db_histogram") {
2284 store
->generate_db_histogram(f
);
2285 } else if (admin_command
== "flush_store_cache") {
2286 store
->flush_cache();
2287 } else if (admin_command
== "dump_pgstate_history") {
2288 f
->open_object_section("pgstate_history");
2289 RWLock::RLocker
l2(pg_map_lock
);
2290 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2294 PG
*pg
= it
->second
;
2295 f
->dump_stream("pg") << pg
->get_pgid();
2297 pg
->pgstate_history
.dump(f
);
2301 } else if (admin_command
== "compact") {
2302 dout(1) << "triggering manual compaction" << dendl
;
2303 auto start
= ceph::coarse_mono_clock::now();
2305 auto end
= ceph::coarse_mono_clock::now();
2306 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
2307 dout(1) << "finished manual compaction in "
2308 << time_span
.count()
2309 << " seconds" << dendl
;
2310 f
->open_object_section("compact_result");
2311 f
->dump_float("elapsed_time", time_span
.count());
2314 assert(0 == "broken asok registration");
2321 class TestOpsSocketHook
: public AdminSocketHook
{
2322 OSDService
*service
;
2325 TestOpsSocketHook(OSDService
*s
, ObjectStore
*st
) : service(s
), store(st
) {}
2326 bool call(std::string command
, cmdmap_t
& cmdmap
, std::string format
,
2327 bufferlist
& out
) override
{
2329 test_ops(service
, store
, command
, cmdmap
, ss
);
2333 void test_ops(OSDService
*service
, ObjectStore
*store
,
2334 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
);
2338 class OSD::C_Tick
: public Context
{
2341 explicit C_Tick(OSD
*o
) : osd(o
) {}
2342 void finish(int r
) override
{
2347 class OSD::C_Tick_WithoutOSDLock
: public Context
{
2350 explicit C_Tick_WithoutOSDLock(OSD
*o
) : osd(o
) {}
2351 void finish(int r
) override
{
2352 osd
->tick_without_osd_lock();
2356 int OSD::enable_disable_fuse(bool stop
)
2360 string mntpath
= cct
->_conf
->osd_data
+ "/fuse";
2361 if (fuse_store
&& (stop
|| !cct
->_conf
->osd_objectstore_fuse
)) {
2362 dout(1) << __func__
<< " disabling" << dendl
;
2366 r
= ::rmdir(mntpath
.c_str());
2369 derr
<< __func__
<< " failed to rmdir " << mntpath
<< ": "
2370 << cpp_strerror(r
) << dendl
;
2375 if (!fuse_store
&& cct
->_conf
->osd_objectstore_fuse
) {
2376 dout(1) << __func__
<< " enabling" << dendl
;
2377 r
= ::mkdir(mntpath
.c_str(), 0700);
2380 if (r
< 0 && r
!= -EEXIST
) {
2381 derr
<< __func__
<< " unable to create " << mntpath
<< ": "
2382 << cpp_strerror(r
) << dendl
;
2385 fuse_store
= new FuseStore(store
, mntpath
);
2386 r
= fuse_store
->start();
2388 derr
<< __func__
<< " unable to start fuse: " << cpp_strerror(r
) << dendl
;
2394 #endif // HAVE_LIBFUSE
2398 int OSD::get_num_op_shards()
2400 if (cct
->_conf
->osd_op_num_shards
)
2401 return cct
->_conf
->osd_op_num_shards
;
2402 if (store_is_rotational
)
2403 return cct
->_conf
->osd_op_num_shards_hdd
;
2405 return cct
->_conf
->osd_op_num_shards_ssd
;
2408 int OSD::get_num_op_threads()
2410 if (cct
->_conf
->osd_op_num_threads_per_shard
)
2411 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard
;
2412 if (store_is_rotational
)
2413 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_hdd
;
2415 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_ssd
;
2418 float OSD::get_osd_recovery_sleep()
2420 if (cct
->_conf
->osd_recovery_sleep
)
2421 return cct
->_conf
->osd_recovery_sleep
;
2422 if (!store_is_rotational
&& !journal_is_rotational
)
2423 return cct
->_conf
->osd_recovery_sleep_ssd
;
2424 else if (store_is_rotational
&& !journal_is_rotational
)
2425 return cct
->_conf
->get_val
<double>("osd_recovery_sleep_hybrid");
2427 return cct
->_conf
->osd_recovery_sleep_hdd
;
2432 CompatSet initial
, diff
;
2433 Mutex::Locker
lock(osd_lock
);
2438 tick_timer_without_osd_lock
.init();
2439 service
.recovery_request_timer
.init();
2440 service
.recovery_sleep_timer
.init();
2443 dout(2) << "init " << dev_path
2444 << " (looks like " << (store_is_rotational
? "hdd" : "ssd") << ")"
2446 dout(2) << "journal " << journal_path
<< dendl
;
2447 assert(store
); // call pre_init() first!
2449 store
->set_cache_shards(get_num_op_shards());
2451 int r
= store
->mount();
2453 derr
<< "OSD:init: unable to mount object store" << dendl
;
2456 journal_is_rotational
= store
->is_journal_rotational();
2457 dout(2) << "journal looks like " << (journal_is_rotational
? "hdd" : "ssd")
2460 enable_disable_fuse(false);
2462 dout(2) << "boot" << dendl
;
2464 // initialize the daily loadavg with current 15min loadavg
2466 if (getloadavg(loadavgs
, 3) == 3) {
2467 daily_loadavg
= loadavgs
[2];
2469 derr
<< "OSD::init() : couldn't read loadavgs\n" << dendl
;
2470 daily_loadavg
= 1.0;
2473 int rotating_auth_attempts
= 0;
2475 // sanity check long object name handling
2478 l
.oid
.name
= string(cct
->_conf
->osd_max_object_name_len
, 'n');
2479 l
.set_key(string(cct
->_conf
->osd_max_object_name_len
, 'k'));
2480 l
.nspace
= string(cct
->_conf
->osd_max_object_namespace_len
, 's');
2481 r
= store
->validate_hobject_key(l
);
2483 derr
<< "backend (" << store
->get_type() << ") is unable to support max "
2484 << "object name[space] len" << dendl
;
2485 derr
<< " osd max object name len = "
2486 << cct
->_conf
->osd_max_object_name_len
<< dendl
;
2487 derr
<< " osd max object namespace len = "
2488 << cct
->_conf
->osd_max_object_namespace_len
<< dendl
;
2489 derr
<< cpp_strerror(r
) << dendl
;
2490 if (cct
->_conf
->osd_check_max_object_name_len_on_startup
) {
2493 derr
<< "osd_check_max_object_name_len_on_startup = false, starting anyway"
2496 dout(20) << "configured osd_max_object_name[space]_len looks ok" << dendl
;
2501 r
= read_superblock();
2503 derr
<< "OSD::init() : unable to read osd superblock" << dendl
;
2508 if (osd_compat
.compare(superblock
.compat_features
) < 0) {
2509 derr
<< "The disk uses features unsupported by the executable." << dendl
;
2510 derr
<< " ondisk features " << superblock
.compat_features
<< dendl
;
2511 derr
<< " daemon features " << osd_compat
<< dendl
;
2513 if (osd_compat
.writeable(superblock
.compat_features
)) {
2514 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2515 derr
<< "it is still writeable, though. Missing features: " << diff
<< dendl
;
2520 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2521 derr
<< "Cannot write to disk! Missing features: " << diff
<< dendl
;
2527 assert_warn(whoami
== superblock
.whoami
);
2528 if (whoami
!= superblock
.whoami
) {
2529 derr
<< "OSD::init: superblock says osd"
2530 << superblock
.whoami
<< " but I am osd." << whoami
<< dendl
;
2535 initial
= get_osd_initial_compat_set();
2536 diff
= superblock
.compat_features
.unsupported(initial
);
2537 if (superblock
.compat_features
.merge(initial
)) {
2538 // We need to persist the new compat_set before we
2540 dout(5) << "Upgrading superblock adding: " << diff
<< dendl
;
2541 ObjectStore::Transaction t
;
2542 write_superblock(t
);
2543 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2548 // make sure snap mapper object exists
2549 if (!store
->exists(coll_t::meta(), OSD::make_snapmapper_oid())) {
2550 dout(10) << "init creating/touching snapmapper object" << dendl
;
2551 ObjectStore::Transaction t
;
2552 t
.touch(coll_t::meta(), OSD::make_snapmapper_oid());
2553 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2558 class_handler
= new ClassHandler(cct
);
2559 cls_initialize(class_handler
);
2561 if (cct
->_conf
->osd_open_classes_on_start
) {
2562 int r
= class_handler
->open_all_classes();
2564 dout(1) << "warning: got an error loading one or more classes: " << cpp_strerror(r
) << dendl
;
2567 // load up "current" osdmap
2568 assert_warn(!osdmap
);
2570 derr
<< "OSD::init: unable to read current osdmap" << dendl
;
2574 osdmap
= get_map(superblock
.current_epoch
);
2575 check_osdmap_features(store
);
2577 create_recoverystate_perf();
2580 epoch_t bind_epoch
= osdmap
->get_epoch();
2581 service
.set_epochs(NULL
, NULL
, &bind_epoch
);
2584 clear_temp_objects();
2586 // initialize osdmap references in sharded wq
2587 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
2589 // load up pgs (as they previously existed)
2592 dout(2) << "superblock: I am osd." << superblock
.whoami
<< dendl
;
2593 dout(0) << "using " << op_queue
<< " op queue with priority op cut off at " <<
2594 op_prio_cutoff
<< "." << dendl
;
2599 client_messenger
->add_dispatcher_head(this);
2600 cluster_messenger
->add_dispatcher_head(this);
2602 hb_front_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2603 hb_back_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2604 hb_front_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2605 hb_back_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2607 objecter_messenger
->add_dispatcher_head(service
.objecter
);
2609 monc
->set_want_keys(CEPH_ENTITY_TYPE_MON
| CEPH_ENTITY_TYPE_OSD
2610 | CEPH_ENTITY_TYPE_MGR
);
2616 * FIXME: this is a placeholder implementation that unconditionally
2617 * sends every is_primary PG's stats every time we're called, unlike
2618 * the existing mon PGStats mechanism that uses pg_stat_queue and acks.
2619 * This has equivalent cost to the existing worst case where all
2620 * PGs are busy and their stats are always enqueued for sending.
2622 mgrc
.set_pgstats_cb([this](){
2623 RWLock::RLocker
l(map_lock
);
2625 utime_t had_for
= ceph_clock_now() - had_map_since
;
2626 osd_stat_t cur_stat
= service
.get_osd_stat();
2627 cur_stat
.os_perf_stat
= store
->get_cur_stats();
2629 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
2630 m
->osd_stat
= cur_stat
;
2632 Mutex::Locker lec
{min_last_epoch_clean_lock
};
2633 min_last_epoch_clean
= osdmap
->get_epoch();
2634 min_last_epoch_clean_pgs
.clear();
2635 RWLock::RLocker
lpg(pg_map_lock
);
2636 for (const auto &i
: pg_map
) {
2638 if (!pg
->is_primary()) {
2642 pg
->pg_stats_publish_lock
.Lock();
2643 if (pg
->pg_stats_publish_valid
) {
2644 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
2645 const auto lec
= pg
->pg_stats_publish
.get_effective_last_epoch_clean();
2646 min_last_epoch_clean
= min(min_last_epoch_clean
, lec
);
2647 min_last_epoch_clean_pgs
.push_back(pg
->info
.pgid
.pgid
);
2649 pg
->pg_stats_publish_lock
.Unlock();
2656 client_messenger
->add_dispatcher_head(&mgrc
);
2658 // tell monc about log_client so it will know about mon session resets
2659 monc
->set_log_client(&log_client
);
2660 update_log_config();
2667 set_disk_tp_priority();
2669 // start the heartbeat
2670 heartbeat_thread
.create("osd_srv_heartbt");
2673 tick_timer
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick(this));
2675 Mutex::Locker
l(tick_timer_lock
);
2676 tick_timer_without_osd_lock
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick_WithoutOSDLock(this));
2680 service
.publish_map(osdmap
);
2681 service
.publish_superblock(superblock
);
2682 service
.max_oldest_map
= superblock
.oldest_map
;
2686 r
= monc
->authenticate();
2688 derr
<< __func__
<< " authentication failed: " << cpp_strerror(r
)
2690 osd_lock
.Lock(); // locker is going to unlock this on function exit
2696 while (monc
->wait_auth_rotating(30.0) < 0) {
2697 derr
<< "unable to obtain rotating service keys; retrying" << dendl
;
2698 ++rotating_auth_attempts
;
2699 if (rotating_auth_attempts
> g_conf
->max_rotating_auth_attempts
) {
2700 derr
<< __func__
<< " wait_auth_rotating timed out" << dendl
;
2701 osd_lock
.Lock(); // make locker happy
2702 if (!is_stopping()) {
2709 r
= update_crush_device_class();
2711 derr
<< __func__
<< " unable to update_crush_device_class: "
2712 << cpp_strerror(r
) << dendl
;
2717 r
= update_crush_location();
2719 derr
<< __func__
<< " unable to update_crush_location: "
2720 << cpp_strerror(r
) << dendl
;
2729 // start objecter *after* we have authenticated, so that we don't ignore
2730 // the OSDMaps it requests.
2731 service
.final_init();
2735 dout(10) << "ensuring pgs have consumed prior maps" << dendl
;
2739 dout(0) << "done with init, starting boot process" << dendl
;
2741 // subscribe to any pg creations
2742 monc
->sub_want("osd_pg_creates", last_pg_create_epoch
, 0);
2744 // MgrClient needs this (it doesn't have MonClient reference itself)
2745 monc
->sub_want("mgrmap", 0, 0);
2747 // we don't need to ask for an osdmap here; objecter will
2748 //monc->sub_want("osdmap", osdmap->get_epoch(), CEPH_SUBSCRIBE_ONETIME);
2759 enable_disable_fuse(true);
2766 void OSD::final_init()
2768 AdminSocket
*admin_socket
= cct
->get_admin_socket();
2769 asok_hook
= new OSDSocketHook(this);
2770 int r
= admin_socket
->register_command("status", "status", asok_hook
,
2771 "high-level status of OSD");
2773 r
= admin_socket
->register_command("flush_journal", "flush_journal",
2775 "flush the journal to permanent store");
2777 r
= admin_socket
->register_command("dump_ops_in_flight",
2778 "dump_ops_in_flight " \
2779 "name=filterstr,type=CephString,n=N,req=false",
2781 "show the ops currently in flight");
2783 r
= admin_socket
->register_command("ops",
2785 "name=filterstr,type=CephString,n=N,req=false",
2787 "show the ops currently in flight");
2789 r
= admin_socket
->register_command("dump_blocked_ops",
2790 "dump_blocked_ops " \
2791 "name=filterstr,type=CephString,n=N,req=false",
2793 "show the blocked ops currently in flight");
2795 r
= admin_socket
->register_command("dump_historic_ops",
2796 "dump_historic_ops " \
2797 "name=filterstr,type=CephString,n=N,req=false",
2801 r
= admin_socket
->register_command("dump_historic_slow_ops",
2802 "dump_historic_slow_ops " \
2803 "name=filterstr,type=CephString,n=N,req=false",
2805 "show slowest recent ops");
2807 r
= admin_socket
->register_command("dump_historic_ops_by_duration",
2808 "dump_historic_ops_by_duration " \
2809 "name=filterstr,type=CephString,n=N,req=false",
2811 "show slowest recent ops, sorted by duration");
2813 r
= admin_socket
->register_command("dump_op_pq_state", "dump_op_pq_state",
2815 "dump op priority queue state");
2817 r
= admin_socket
->register_command("dump_blacklist", "dump_blacklist",
2819 "dump blacklisted clients and times");
2821 r
= admin_socket
->register_command("dump_watchers", "dump_watchers",
2823 "show clients which have active watches,"
2824 " and on which objects");
2826 r
= admin_socket
->register_command("dump_reservations", "dump_reservations",
2828 "show recovery reservations");
2830 r
= admin_socket
->register_command("get_latest_osdmap", "get_latest_osdmap",
2832 "force osd to update the latest map from "
2836 r
= admin_socket
->register_command( "heap",
2838 "name=heapcmd,type=CephString",
2840 "show heap usage info (available only if "
2841 "compiled with tcmalloc)");
2844 r
= admin_socket
->register_command("set_heap_property",
2845 "set_heap_property " \
2846 "name=property,type=CephString " \
2847 "name=value,type=CephInt",
2849 "update malloc extension heap property");
2852 r
= admin_socket
->register_command("get_heap_property",
2853 "get_heap_property " \
2854 "name=property,type=CephString",
2856 "get malloc extension heap property");
2859 r
= admin_socket
->register_command("dump_objectstore_kv_stats",
2860 "dump_objectstore_kv_stats",
2862 "print statistics of kvdb which used by bluestore");
2865 r
= admin_socket
->register_command("dump_scrubs",
2868 "print scheduled scrubs");
2871 r
= admin_socket
->register_command("calc_objectstore_db_histogram",
2872 "calc_objectstore_db_histogram",
2874 "Generate key value histogram of kvdb(rocksdb) which used by bluestore");
2877 r
= admin_socket
->register_command("flush_store_cache",
2878 "flush_store_cache",
2880 "Flush bluestore internal cache");
2882 r
= admin_socket
->register_command("dump_pgstate_history", "dump_pgstate_history",
2884 "show recent state history");
2887 r
= admin_socket
->register_command("compact", "compact",
2889 "Commpact object store's omap."
2890 " WARNING: Compaction probably slows your requests");
2893 test_ops_hook
= new TestOpsSocketHook(&(this->service
), this->store
);
2894 // Note: pools are CephString instead of CephPoolname because
2895 // these commands traditionally support both pool names and numbers
2896 r
= admin_socket
->register_command(
2899 "name=pool,type=CephString " \
2900 "name=objname,type=CephObjectname " \
2901 "name=key,type=CephString "\
2902 "name=val,type=CephString",
2906 r
= admin_socket
->register_command(
2909 "name=pool,type=CephString " \
2910 "name=objname,type=CephObjectname " \
2911 "name=key,type=CephString",
2915 r
= admin_socket
->register_command(
2918 "name=pool,type=CephString " \
2919 "name=objname,type=CephObjectname " \
2920 "name=header,type=CephString",
2925 r
= admin_socket
->register_command(
2928 "name=pool,type=CephString " \
2929 "name=objname,type=CephObjectname",
2931 "output entire object map");
2934 r
= admin_socket
->register_command(
2937 "name=pool,type=CephString " \
2938 "name=objname,type=CephObjectname " \
2939 "name=len,type=CephInt",
2941 "truncate object to length");
2944 r
= admin_socket
->register_command(
2947 "name=pool,type=CephString " \
2948 "name=objname,type=CephObjectname " \
2949 "name=shardid,type=CephInt,req=false,range=0|255",
2951 "inject data error to an object");
2954 r
= admin_socket
->register_command(
2957 "name=pool,type=CephString " \
2958 "name=objname,type=CephObjectname " \
2959 "name=shardid,type=CephInt,req=false,range=0|255",
2961 "inject metadata error to an object");
2963 r
= admin_socket
->register_command(
2964 "set_recovery_delay",
2965 "set_recovery_delay " \
2966 "name=utime,type=CephInt,req=false",
2968 "Delay osd recovery by specified seconds");
2970 r
= admin_socket
->register_command(
2973 "name=pgid,type=CephString ",
2975 "Trigger a scheduled scrub ");
2977 r
= admin_socket
->register_command(
2980 "name=type,type=CephString,req=false " \
2981 "name=count,type=CephInt,req=false ",
2983 "Inject a full disk (optional count times)");
2987 void OSD::create_logger()
2989 dout(10) << "create_logger" << dendl
;
2991 PerfCountersBuilder
osd_plb(cct
, "osd", l_osd_first
, l_osd_last
);
2993 // Latency axis configuration for op histograms, values are in nanoseconds
2994 PerfHistogramCommon::axis_config_d op_hist_x_axis_config
{
2996 PerfHistogramCommon::SCALE_LOG2
, ///< Latency in logarithmic scale
2998 100000, ///< Quantization unit is 100usec
2999 32, ///< Enough to cover much longer than slow requests
3002 // Op size axis configuration for op histograms, values are in bytes
3003 PerfHistogramCommon::axis_config_d op_hist_y_axis_config
{
3004 "Request size (bytes)",
3005 PerfHistogramCommon::SCALE_LOG2
, ///< Request size in logarithmic scale
3007 512, ///< Quantization unit is 512 bytes
3008 32, ///< Enough to cover requests larger than GB
3012 // All the basic OSD operation stats are to be considered useful
3013 osd_plb
.set_prio_default(PerfCountersBuilder::PRIO_USEFUL
);
3016 l_osd_op_wip
, "op_wip",
3017 "Replication operations currently being processed (primary)");
3018 osd_plb
.add_u64_counter(
3020 "Client operations",
3021 "ops", PerfCountersBuilder::PRIO_CRITICAL
);
3022 osd_plb
.add_u64_counter(
3023 l_osd_op_inb
, "op_in_bytes",
3024 "Client operations total write size",
3025 "wr", PerfCountersBuilder::PRIO_INTERESTING
);
3026 osd_plb
.add_u64_counter(
3027 l_osd_op_outb
, "op_out_bytes",
3028 "Client operations total read size",
3029 "rd", PerfCountersBuilder::PRIO_INTERESTING
);
3030 osd_plb
.add_time_avg(
3031 l_osd_op_lat
, "op_latency",
3032 "Latency of client operations (including queue time)",
3034 osd_plb
.add_time_avg(
3035 l_osd_op_process_lat
, "op_process_latency",
3036 "Latency of client operations (excluding queue time)");
3037 osd_plb
.add_time_avg(
3038 l_osd_op_prepare_lat
, "op_prepare_latency",
3039 "Latency of client operations (excluding queue time and wait for finished)");
3041 osd_plb
.add_u64_counter(
3042 l_osd_op_r
, "op_r", "Client read operations");
3043 osd_plb
.add_u64_counter(
3044 l_osd_op_r_outb
, "op_r_out_bytes", "Client data read");
3045 osd_plb
.add_time_avg(
3046 l_osd_op_r_lat
, "op_r_latency",
3047 "Latency of read operation (including queue time)");
3048 osd_plb
.add_u64_counter_histogram(
3049 l_osd_op_r_lat_outb_hist
, "op_r_latency_out_bytes_histogram",
3050 op_hist_x_axis_config
, op_hist_y_axis_config
,
3051 "Histogram of operation latency (including queue time) + data read");
3052 osd_plb
.add_time_avg(
3053 l_osd_op_r_process_lat
, "op_r_process_latency",
3054 "Latency of read operation (excluding queue time)");
3055 osd_plb
.add_time_avg(
3056 l_osd_op_r_prepare_lat
, "op_r_prepare_latency",
3057 "Latency of read operations (excluding queue time and wait for finished)");
3058 osd_plb
.add_u64_counter(
3059 l_osd_op_w
, "op_w", "Client write operations");
3060 osd_plb
.add_u64_counter(
3061 l_osd_op_w_inb
, "op_w_in_bytes", "Client data written");
3062 osd_plb
.add_time_avg(
3063 l_osd_op_w_lat
, "op_w_latency",
3064 "Latency of write operation (including queue time)");
3065 osd_plb
.add_u64_counter_histogram(
3066 l_osd_op_w_lat_inb_hist
, "op_w_latency_in_bytes_histogram",
3067 op_hist_x_axis_config
, op_hist_y_axis_config
,
3068 "Histogram of operation latency (including queue time) + data written");
3069 osd_plb
.add_time_avg(
3070 l_osd_op_w_process_lat
, "op_w_process_latency",
3071 "Latency of write operation (excluding queue time)");
3072 osd_plb
.add_time_avg(
3073 l_osd_op_w_prepare_lat
, "op_w_prepare_latency",
3074 "Latency of write operations (excluding queue time and wait for finished)");
3075 osd_plb
.add_u64_counter(
3076 l_osd_op_rw
, "op_rw",
3077 "Client read-modify-write operations");
3078 osd_plb
.add_u64_counter(
3079 l_osd_op_rw_inb
, "op_rw_in_bytes",
3080 "Client read-modify-write operations write in");
3081 osd_plb
.add_u64_counter(
3082 l_osd_op_rw_outb
,"op_rw_out_bytes",
3083 "Client read-modify-write operations read out ");
3084 osd_plb
.add_time_avg(
3085 l_osd_op_rw_lat
, "op_rw_latency",
3086 "Latency of read-modify-write operation (including queue time)");
3087 osd_plb
.add_u64_counter_histogram(
3088 l_osd_op_rw_lat_inb_hist
, "op_rw_latency_in_bytes_histogram",
3089 op_hist_x_axis_config
, op_hist_y_axis_config
,
3090 "Histogram of rw operation latency (including queue time) + data written");
3091 osd_plb
.add_u64_counter_histogram(
3092 l_osd_op_rw_lat_outb_hist
, "op_rw_latency_out_bytes_histogram",
3093 op_hist_x_axis_config
, op_hist_y_axis_config
,
3094 "Histogram of rw operation latency (including queue time) + data read");
3095 osd_plb
.add_time_avg(
3096 l_osd_op_rw_process_lat
, "op_rw_process_latency",
3097 "Latency of read-modify-write operation (excluding queue time)");
3098 osd_plb
.add_time_avg(
3099 l_osd_op_rw_prepare_lat
, "op_rw_prepare_latency",
3100 "Latency of read-modify-write operations (excluding queue time and wait for finished)");
3102 // Now we move on to some more obscure stats, revert to assuming things
3103 // are low priority unless otherwise specified.
3104 osd_plb
.set_prio_default(PerfCountersBuilder::PRIO_DEBUGONLY
);
3106 osd_plb
.add_time_avg(l_osd_op_before_queue_op_lat
, "op_before_queue_op_lat",
3107 "Latency of IO before calling queue(before really queue into ShardedOpWq)"); // client io before queue op_wq latency
3108 osd_plb
.add_time_avg(l_osd_op_before_dequeue_op_lat
, "op_before_dequeue_op_lat",
3109 "Latency of IO before calling dequeue_op(already dequeued and get PG lock)"); // client io before dequeue_op latency
3111 osd_plb
.add_u64_counter(
3112 l_osd_sop
, "subop", "Suboperations");
3113 osd_plb
.add_u64_counter(
3114 l_osd_sop_inb
, "subop_in_bytes", "Suboperations total size");
3115 osd_plb
.add_time_avg(l_osd_sop_lat
, "subop_latency", "Suboperations latency");
3117 osd_plb
.add_u64_counter(l_osd_sop_w
, "subop_w", "Replicated writes");
3118 osd_plb
.add_u64_counter(
3119 l_osd_sop_w_inb
, "subop_w_in_bytes", "Replicated written data size");
3120 osd_plb
.add_time_avg(
3121 l_osd_sop_w_lat
, "subop_w_latency", "Replicated writes latency");
3122 osd_plb
.add_u64_counter(
3123 l_osd_sop_pull
, "subop_pull", "Suboperations pull requests");
3124 osd_plb
.add_time_avg(
3125 l_osd_sop_pull_lat
, "subop_pull_latency", "Suboperations pull latency");
3126 osd_plb
.add_u64_counter(
3127 l_osd_sop_push
, "subop_push", "Suboperations push messages");
3128 osd_plb
.add_u64_counter(
3129 l_osd_sop_push_inb
, "subop_push_in_bytes", "Suboperations pushed size");
3130 osd_plb
.add_time_avg(
3131 l_osd_sop_push_lat
, "subop_push_latency", "Suboperations push latency");
3133 osd_plb
.add_u64_counter(l_osd_pull
, "pull", "Pull requests sent");
3134 osd_plb
.add_u64_counter(l_osd_push
, "push", "Push messages sent");
3135 osd_plb
.add_u64_counter(l_osd_push_outb
, "push_out_bytes", "Pushed size");
3137 osd_plb
.add_u64_counter(
3138 l_osd_rop
, "recovery_ops",
3139 "Started recovery operations",
3140 "rop", PerfCountersBuilder::PRIO_INTERESTING
);
3142 osd_plb
.add_u64(l_osd_loadavg
, "loadavg", "CPU load");
3143 osd_plb
.add_u64(l_osd_buf
, "buffer_bytes", "Total allocated buffer size");
3144 osd_plb
.add_u64(l_osd_history_alloc_bytes
, "history_alloc_Mbytes");
3145 osd_plb
.add_u64(l_osd_history_alloc_num
, "history_alloc_num");
3147 l_osd_cached_crc
, "cached_crc", "Total number getting crc from crc_cache");
3149 l_osd_cached_crc_adjusted
, "cached_crc_adjusted",
3150 "Total number getting crc from crc_cache with adjusting");
3151 osd_plb
.add_u64(l_osd_missed_crc
, "missed_crc",
3152 "Total number of crc cache misses");
3154 osd_plb
.add_u64(l_osd_pg
, "numpg", "Placement groups",
3155 "pgs", PerfCountersBuilder::PRIO_USEFUL
);
3157 l_osd_pg_primary
, "numpg_primary",
3158 "Placement groups for which this osd is primary");
3160 l_osd_pg_replica
, "numpg_replica",
3161 "Placement groups for which this osd is replica");
3163 l_osd_pg_stray
, "numpg_stray",
3164 "Placement groups ready to be deleted from this osd");
3166 l_osd_pg_removing
, "numpg_removing",
3167 "Placement groups queued for local deletion", "pgsr",
3168 PerfCountersBuilder::PRIO_USEFUL
);
3170 l_osd_hb_to
, "heartbeat_to_peers", "Heartbeat (ping) peers we send to");
3171 osd_plb
.add_u64_counter(l_osd_map
, "map_messages", "OSD map messages");
3172 osd_plb
.add_u64_counter(l_osd_mape
, "map_message_epochs", "OSD map epochs");
3173 osd_plb
.add_u64_counter(
3174 l_osd_mape_dup
, "map_message_epoch_dups", "OSD map duplicates");
3175 osd_plb
.add_u64_counter(
3176 l_osd_waiting_for_map
, "messages_delayed_for_map",
3177 "Operations waiting for OSD map");
3179 osd_plb
.add_u64_counter(
3180 l_osd_map_cache_hit
, "osd_map_cache_hit", "osdmap cache hit");
3181 osd_plb
.add_u64_counter(
3182 l_osd_map_cache_miss
, "osd_map_cache_miss", "osdmap cache miss");
3183 osd_plb
.add_u64_counter(
3184 l_osd_map_cache_miss_low
, "osd_map_cache_miss_low",
3185 "osdmap cache miss below cache lower bound");
3186 osd_plb
.add_u64_avg(
3187 l_osd_map_cache_miss_low_avg
, "osd_map_cache_miss_low_avg",
3188 "osdmap cache miss, avg distance below cache lower bound");
3189 osd_plb
.add_u64_counter(
3190 l_osd_map_bl_cache_hit
, "osd_map_bl_cache_hit",
3191 "OSDMap buffer cache hits");
3192 osd_plb
.add_u64_counter(
3193 l_osd_map_bl_cache_miss
, "osd_map_bl_cache_miss",
3194 "OSDMap buffer cache misses");
3197 l_osd_stat_bytes
, "stat_bytes", "OSD size", "size",
3198 PerfCountersBuilder::PRIO_USEFUL
);
3200 l_osd_stat_bytes_used
, "stat_bytes_used", "Used space", "used",
3201 PerfCountersBuilder::PRIO_USEFUL
);
3202 osd_plb
.add_u64(l_osd_stat_bytes_avail
, "stat_bytes_avail", "Available space");
3204 osd_plb
.add_u64_counter(
3205 l_osd_copyfrom
, "copyfrom", "Rados \"copy-from\" operations");
3207 osd_plb
.add_u64_counter(l_osd_tier_promote
, "tier_promote", "Tier promotions");
3208 osd_plb
.add_u64_counter(l_osd_tier_flush
, "tier_flush", "Tier flushes");
3209 osd_plb
.add_u64_counter(
3210 l_osd_tier_flush_fail
, "tier_flush_fail", "Failed tier flushes");
3211 osd_plb
.add_u64_counter(
3212 l_osd_tier_try_flush
, "tier_try_flush", "Tier flush attempts");
3213 osd_plb
.add_u64_counter(
3214 l_osd_tier_try_flush_fail
, "tier_try_flush_fail",
3215 "Failed tier flush attempts");
3216 osd_plb
.add_u64_counter(
3217 l_osd_tier_evict
, "tier_evict", "Tier evictions");
3218 osd_plb
.add_u64_counter(
3219 l_osd_tier_whiteout
, "tier_whiteout", "Tier whiteouts");
3220 osd_plb
.add_u64_counter(
3221 l_osd_tier_dirty
, "tier_dirty", "Dirty tier flag set");
3222 osd_plb
.add_u64_counter(
3223 l_osd_tier_clean
, "tier_clean", "Dirty tier flag cleaned");
3224 osd_plb
.add_u64_counter(
3225 l_osd_tier_delay
, "tier_delay", "Tier delays (agent waiting)");
3226 osd_plb
.add_u64_counter(
3227 l_osd_tier_proxy_read
, "tier_proxy_read", "Tier proxy reads");
3228 osd_plb
.add_u64_counter(
3229 l_osd_tier_proxy_write
, "tier_proxy_write", "Tier proxy writes");
3231 osd_plb
.add_u64_counter(
3232 l_osd_agent_wake
, "agent_wake", "Tiering agent wake up");
3233 osd_plb
.add_u64_counter(
3234 l_osd_agent_skip
, "agent_skip", "Objects skipped by agent");
3235 osd_plb
.add_u64_counter(
3236 l_osd_agent_flush
, "agent_flush", "Tiering agent flushes");
3237 osd_plb
.add_u64_counter(
3238 l_osd_agent_evict
, "agent_evict", "Tiering agent evictions");
3240 osd_plb
.add_u64_counter(
3241 l_osd_object_ctx_cache_hit
, "object_ctx_cache_hit", "Object context cache hits");
3242 osd_plb
.add_u64_counter(
3243 l_osd_object_ctx_cache_total
, "object_ctx_cache_total", "Object context cache lookups");
3245 osd_plb
.add_u64_counter(l_osd_op_cache_hit
, "op_cache_hit");
3246 osd_plb
.add_time_avg(
3247 l_osd_tier_flush_lat
, "osd_tier_flush_lat", "Object flush latency");
3248 osd_plb
.add_time_avg(
3249 l_osd_tier_promote_lat
, "osd_tier_promote_lat", "Object promote latency");
3250 osd_plb
.add_time_avg(
3251 l_osd_tier_r_lat
, "osd_tier_r_lat", "Object proxy read latency");
3253 osd_plb
.add_u64_counter(
3254 l_osd_pg_info
, "osd_pg_info", "PG updated its info (using any method)");
3255 osd_plb
.add_u64_counter(
3256 l_osd_pg_fastinfo
, "osd_pg_fastinfo",
3257 "PG updated its info using fastinfo attr");
3258 osd_plb
.add_u64_counter(
3259 l_osd_pg_biginfo
, "osd_pg_biginfo", "PG updated its biginfo attr");
3261 logger
= osd_plb
.create_perf_counters();
3262 cct
->get_perfcounters_collection()->add(logger
);
3265 void OSD::create_recoverystate_perf()
3267 dout(10) << "create_recoverystate_perf" << dendl
;
3269 PerfCountersBuilder
rs_perf(cct
, "recoverystate_perf", rs_first
, rs_last
);
3271 rs_perf
.add_time_avg(rs_initial_latency
, "initial_latency", "Initial recovery state latency");
3272 rs_perf
.add_time_avg(rs_started_latency
, "started_latency", "Started recovery state latency");
3273 rs_perf
.add_time_avg(rs_reset_latency
, "reset_latency", "Reset recovery state latency");
3274 rs_perf
.add_time_avg(rs_start_latency
, "start_latency", "Start recovery state latency");
3275 rs_perf
.add_time_avg(rs_primary_latency
, "primary_latency", "Primary recovery state latency");
3276 rs_perf
.add_time_avg(rs_peering_latency
, "peering_latency", "Peering recovery state latency");
3277 rs_perf
.add_time_avg(rs_backfilling_latency
, "backfilling_latency", "Backfilling recovery state latency");
3278 rs_perf
.add_time_avg(rs_waitremotebackfillreserved_latency
, "waitremotebackfillreserved_latency", "Wait remote backfill reserved recovery state latency");
3279 rs_perf
.add_time_avg(rs_waitlocalbackfillreserved_latency
, "waitlocalbackfillreserved_latency", "Wait local backfill reserved recovery state latency");
3280 rs_perf
.add_time_avg(rs_notbackfilling_latency
, "notbackfilling_latency", "Notbackfilling recovery state latency");
3281 rs_perf
.add_time_avg(rs_repnotrecovering_latency
, "repnotrecovering_latency", "Repnotrecovering recovery state latency");
3282 rs_perf
.add_time_avg(rs_repwaitrecoveryreserved_latency
, "repwaitrecoveryreserved_latency", "Rep wait recovery reserved recovery state latency");
3283 rs_perf
.add_time_avg(rs_repwaitbackfillreserved_latency
, "repwaitbackfillreserved_latency", "Rep wait backfill reserved recovery state latency");
3284 rs_perf
.add_time_avg(rs_reprecovering_latency
, "reprecovering_latency", "RepRecovering recovery state latency");
3285 rs_perf
.add_time_avg(rs_activating_latency
, "activating_latency", "Activating recovery state latency");
3286 rs_perf
.add_time_avg(rs_waitlocalrecoveryreserved_latency
, "waitlocalrecoveryreserved_latency", "Wait local recovery reserved recovery state latency");
3287 rs_perf
.add_time_avg(rs_waitremoterecoveryreserved_latency
, "waitremoterecoveryreserved_latency", "Wait remote recovery reserved recovery state latency");
3288 rs_perf
.add_time_avg(rs_recovering_latency
, "recovering_latency", "Recovering recovery state latency");
3289 rs_perf
.add_time_avg(rs_recovered_latency
, "recovered_latency", "Recovered recovery state latency");
3290 rs_perf
.add_time_avg(rs_clean_latency
, "clean_latency", "Clean recovery state latency");
3291 rs_perf
.add_time_avg(rs_active_latency
, "active_latency", "Active recovery state latency");
3292 rs_perf
.add_time_avg(rs_replicaactive_latency
, "replicaactive_latency", "Replicaactive recovery state latency");
3293 rs_perf
.add_time_avg(rs_stray_latency
, "stray_latency", "Stray recovery state latency");
3294 rs_perf
.add_time_avg(rs_getinfo_latency
, "getinfo_latency", "Getinfo recovery state latency");
3295 rs_perf
.add_time_avg(rs_getlog_latency
, "getlog_latency", "Getlog recovery state latency");
3296 rs_perf
.add_time_avg(rs_waitactingchange_latency
, "waitactingchange_latency", "Waitactingchange recovery state latency");
3297 rs_perf
.add_time_avg(rs_incomplete_latency
, "incomplete_latency", "Incomplete recovery state latency");
3298 rs_perf
.add_time_avg(rs_down_latency
, "down_latency", "Down recovery state latency");
3299 rs_perf
.add_time_avg(rs_getmissing_latency
, "getmissing_latency", "Getmissing recovery state latency");
3300 rs_perf
.add_time_avg(rs_waitupthru_latency
, "waitupthru_latency", "Waitupthru recovery state latency");
3301 rs_perf
.add_time_avg(rs_notrecovering_latency
, "notrecovering_latency", "Notrecovering recovery state latency");
3303 recoverystate_perf
= rs_perf
.create_perf_counters();
3304 cct
->get_perfcounters_collection()->add(recoverystate_perf
);
3309 if (!service
.prepare_to_stop())
3310 return 0; // already shutting down
3312 if (is_stopping()) {
3316 derr
<< "shutdown" << dendl
;
3318 set_state(STATE_STOPPING
);
3321 if (cct
->_conf
->get_val
<bool>("osd_debug_shutdown")) {
3322 cct
->_conf
->set_val("debug_osd", "100");
3323 cct
->_conf
->set_val("debug_journal", "100");
3324 cct
->_conf
->set_val("debug_filestore", "100");
3325 cct
->_conf
->set_val("debug_bluestore", "100");
3326 cct
->_conf
->set_val("debug_ms", "100");
3327 cct
->_conf
->apply_changes(NULL
);
3330 // stop MgrClient earlier as it's more like an internal consumer of OSD
3333 service
.start_shutdown();
3335 // stop sending work to pgs. this just prevents any new work in _process
3336 // from racing with on_shutdown and potentially entering the pg after.
3337 op_shardedwq
.drain();
3341 RWLock::RLocker
l(pg_map_lock
);
3342 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3345 dout(20) << " kicking pg " << p
->first
<< dendl
;
3347 p
->second
->on_shutdown();
3348 p
->second
->unlock();
3349 p
->second
->osr
->flush();
3352 clear_pg_stat_queue();
3354 // drain op queue again (in case PGs requeued something)
3355 op_shardedwq
.drain();
3357 finished
.clear(); // zap waiters (bleh, this is messy)
3360 op_shardedwq
.clear_pg_slots();
3362 // unregister commands
3363 cct
->get_admin_socket()->unregister_command("status");
3364 cct
->get_admin_socket()->unregister_command("flush_journal");
3365 cct
->get_admin_socket()->unregister_command("dump_ops_in_flight");
3366 cct
->get_admin_socket()->unregister_command("ops");
3367 cct
->get_admin_socket()->unregister_command("dump_blocked_ops");
3368 cct
->get_admin_socket()->unregister_command("dump_historic_ops");
3369 cct
->get_admin_socket()->unregister_command("dump_historic_ops_by_duration");
3370 cct
->get_admin_socket()->unregister_command("dump_historic_slow_ops");
3371 cct
->get_admin_socket()->unregister_command("dump_op_pq_state");
3372 cct
->get_admin_socket()->unregister_command("dump_blacklist");
3373 cct
->get_admin_socket()->unregister_command("dump_watchers");
3374 cct
->get_admin_socket()->unregister_command("dump_reservations");
3375 cct
->get_admin_socket()->unregister_command("get_latest_osdmap");
3376 cct
->get_admin_socket()->unregister_command("heap");
3377 cct
->get_admin_socket()->unregister_command("set_heap_property");
3378 cct
->get_admin_socket()->unregister_command("get_heap_property");
3379 cct
->get_admin_socket()->unregister_command("dump_objectstore_kv_stats");
3380 cct
->get_admin_socket()->unregister_command("dump_scrubs");
3381 cct
->get_admin_socket()->unregister_command("calc_objectstore_db_histogram");
3382 cct
->get_admin_socket()->unregister_command("flush_store_cache");
3383 cct
->get_admin_socket()->unregister_command("dump_pgstate_history");
3384 cct
->get_admin_socket()->unregister_command("compact");
3388 cct
->get_admin_socket()->unregister_command("setomapval");
3389 cct
->get_admin_socket()->unregister_command("rmomapkey");
3390 cct
->get_admin_socket()->unregister_command("setomapheader");
3391 cct
->get_admin_socket()->unregister_command("getomap");
3392 cct
->get_admin_socket()->unregister_command("truncobj");
3393 cct
->get_admin_socket()->unregister_command("injectdataerr");
3394 cct
->get_admin_socket()->unregister_command("injectmdataerr");
3395 cct
->get_admin_socket()->unregister_command("set_recovery_delay");
3396 cct
->get_admin_socket()->unregister_command("trigger_scrub");
3397 cct
->get_admin_socket()->unregister_command("injectfull");
3398 delete test_ops_hook
;
3399 test_ops_hook
= NULL
;
3403 heartbeat_lock
.Lock();
3404 heartbeat_stop
= true;
3405 heartbeat_cond
.Signal();
3406 heartbeat_lock
.Unlock();
3407 heartbeat_thread
.join();
3412 dout(10) << "osd tp stopped" << dendl
;
3416 dout(10) << "op sharded tp stopped" << dendl
;
3420 dout(10) << "command tp stopped" << dendl
;
3424 dout(10) << "disk tp paused (new)" << dendl
;
3426 dout(10) << "stopping agent" << dendl
;
3427 service
.agent_stop();
3431 reset_heartbeat_peers();
3433 tick_timer
.shutdown();
3436 Mutex::Locker
l(tick_timer_lock
);
3437 tick_timer_without_osd_lock
.shutdown();
3440 // note unmount epoch
3441 dout(10) << "noting clean unmount in epoch " << osdmap
->get_epoch() << dendl
;
3442 superblock
.mounted
= service
.get_boot_epoch();
3443 superblock
.clean_thru
= osdmap
->get_epoch();
3444 ObjectStore::Transaction t
;
3445 write_superblock(t
);
3446 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3448 derr
<< "OSD::shutdown: error writing superblock: "
3449 << cpp_strerror(r
) << dendl
;
3454 Mutex::Locker
l(pg_stat_queue_lock
);
3455 assert(pg_stat_queue
.empty());
3458 service
.shutdown_reserver();
3461 #ifdef PG_DEBUG_REFS
3462 service
.dump_live_pgids();
3465 RWLock::RLocker
l(pg_map_lock
);
3466 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3469 dout(20) << " kicking pg " << p
->first
<< dendl
;
3471 if (p
->second
->ref
!= 1) {
3472 derr
<< "pgid " << p
->first
<< " has ref count of "
3473 << p
->second
->ref
<< dendl
;
3474 #ifdef PG_DEBUG_REFS
3475 p
->second
->dump_live_ids();
3477 if (cct
->_conf
->osd_shutdown_pgref_assert
) {
3481 p
->second
->unlock();
3482 p
->second
->put("PGMap");
3486 #ifdef PG_DEBUG_REFS
3487 service
.dump_live_pgids();
3489 cct
->_conf
->remove_observer(this);
3491 dout(10) << "syncing store" << dendl
;
3492 enable_disable_fuse(true);
3494 if (cct
->_conf
->osd_journal_flush_on_shutdown
) {
3495 dout(10) << "flushing journal" << dendl
;
3496 store
->flush_journal();
3502 dout(10) << "Store synced" << dendl
;
3507 osdmap
= OSDMapRef();
3509 op_tracker
.on_shutdown();
3511 class_handler
->shutdown();
3512 client_messenger
->shutdown();
3513 cluster_messenger
->shutdown();
3514 hb_front_client_messenger
->shutdown();
3515 hb_back_client_messenger
->shutdown();
3516 objecter_messenger
->shutdown();
3517 hb_front_server_messenger
->shutdown();
3518 hb_back_server_messenger
->shutdown();
3525 int OSD::mon_cmd_maybe_osd_create(string
&cmd
)
3527 bool created
= false;
3529 dout(10) << __func__
<< " cmd: " << cmd
<< dendl
;
3530 vector
<string
> vcmd
{cmd
};
3534 monc
->start_mon_command(vcmd
, inbl
, NULL
, &outs
, &w
);
3537 if (r
== -ENOENT
&& !created
) {
3538 string newcmd
= "{\"prefix\": \"osd create\", \"id\": " + stringify(whoami
)
3539 + ", \"uuid\": \"" + stringify(superblock
.osd_fsid
) + "\"}";
3540 vector
<string
> vnewcmd
{newcmd
};
3544 monc
->start_mon_command(vnewcmd
, inbl
, NULL
, &outs
, &w
);
3547 derr
<< __func__
<< " fail: osd does not exist and created failed: "
3548 << cpp_strerror(r
) << dendl
;
3554 derr
<< __func__
<< " fail: '" << outs
<< "': " << cpp_strerror(r
) << dendl
;
3563 int OSD::update_crush_location()
3565 if (!cct
->_conf
->osd_crush_update_on_start
) {
3566 dout(10) << __func__
<< " osd_crush_update_on_start = false" << dendl
;
3571 if (cct
->_conf
->osd_crush_initial_weight
>= 0) {
3572 snprintf(weight
, sizeof(weight
), "%.4lf", cct
->_conf
->osd_crush_initial_weight
);
3574 struct store_statfs_t st
;
3575 int r
= store
->statfs(&st
);
3577 derr
<< "statfs: " << cpp_strerror(r
) << dendl
;
3580 snprintf(weight
, sizeof(weight
), "%.4lf",
3582 (double)(st
.total
) /
3583 (double)(1ull << 40 /* TB */)));
3586 std::multimap
<string
,string
> loc
= cct
->crush_location
.get_location();
3587 dout(10) << __func__
<< " crush location is " << loc
<< dendl
;
3590 string("{\"prefix\": \"osd crush create-or-move\", ") +
3591 string("\"id\": ") + stringify(whoami
) + string(", ") +
3592 string("\"weight\":") + weight
+ string(", ") +
3593 string("\"args\": [");
3594 for (multimap
<string
,string
>::iterator p
= loc
.begin(); p
!= loc
.end(); ++p
) {
3595 if (p
!= loc
.begin())
3597 cmd
+= "\"" + p
->first
+ "=" + p
->second
+ "\"";
3601 return mon_cmd_maybe_osd_create(cmd
);
3604 int OSD::update_crush_device_class()
3606 if (!cct
->_conf
->osd_class_update_on_start
) {
3607 dout(10) << __func__
<< " osd_class_update_on_start = false" << dendl
;
3611 string device_class
;
3612 int r
= store
->read_meta("crush_device_class", &device_class
);
3613 if (r
< 0 || device_class
.empty()) {
3614 device_class
= store
->get_default_device_class();
3617 if (device_class
.empty()) {
3618 dout(20) << __func__
<< " no device class stored locally" << dendl
;
3623 string("{\"prefix\": \"osd crush set-device-class\", ") +
3624 string("\"class\": \"") + device_class
+ string("\", ") +
3625 string("\"ids\": [\"") + stringify(whoami
) + string("\"]}");
3627 r
= mon_cmd_maybe_osd_create(cmd
);
3628 // the above cmd can fail for various reasons, e.g.:
3629 // (1) we are connecting to a pre-luminous monitor
3630 // (2) user manually specify a class other than
3631 // 'ceph-disk prepare --crush-device-class'
3632 // simply skip result-checking for now
3636 void OSD::write_superblock(ObjectStore::Transaction
& t
)
3638 dout(10) << "write_superblock " << superblock
<< dendl
;
3640 //hack: at minimum it's using the baseline feature set
3641 if (!superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_BASE
))
3642 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE
);
3645 ::encode(superblock
, bl
);
3646 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
3649 int OSD::read_superblock()
3652 int r
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, bl
);
3656 bufferlist::iterator p
= bl
.begin();
3657 ::decode(superblock
, p
);
3659 dout(10) << "read_superblock " << superblock
<< dendl
;
3664 void OSD::clear_temp_objects()
3666 dout(10) << __func__
<< dendl
;
3668 store
->list_collections(ls
);
3669 for (vector
<coll_t
>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
3671 if (!p
->is_pg(&pgid
))
3674 // list temp objects
3675 dout(20) << " clearing temps in " << *p
<< " pgid " << pgid
<< dendl
;
3677 vector
<ghobject_t
> temps
;
3680 vector
<ghobject_t
> objects
;
3681 store
->collection_list(*p
, next
, ghobject_t::get_max(),
3682 store
->get_ideal_list_max(),
3684 if (objects
.empty())
3686 vector
<ghobject_t
>::iterator q
;
3687 for (q
= objects
.begin(); q
!= objects
.end(); ++q
) {
3688 // Hammer set pool for temps to -1, so check for clean-up
3689 if (q
->hobj
.is_temp() || (q
->hobj
.pool
== -1)) {
3690 temps
.push_back(*q
);
3695 // If we saw a non-temp object and hit the break above we can
3696 // break out of the while loop too.
3697 if (q
!= objects
.end())
3700 if (!temps
.empty()) {
3701 ObjectStore::Transaction t
;
3703 for (vector
<ghobject_t
>::iterator q
= temps
.begin(); q
!= temps
.end(); ++q
) {
3704 dout(20) << " removing " << *p
<< " object " << *q
<< dendl
;
3706 if (++removed
> cct
->_conf
->osd_target_transaction_size
) {
3707 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3708 t
= ObjectStore::Transaction();
3713 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3719 void OSD::recursive_remove_collection(CephContext
* cct
,
3720 ObjectStore
*store
, spg_t pgid
,
3726 make_snapmapper_oid());
3728 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
3729 ObjectStore::Sequencer
>("rm"));
3730 ObjectStore::Transaction t
;
3731 SnapMapper
mapper(cct
, &driver
, 0, 0, 0, pgid
.shard
);
3733 vector
<ghobject_t
> objects
;
3734 store
->collection_list(tmp
, ghobject_t(), ghobject_t::get_max(),
3735 INT_MAX
, &objects
, 0);
3736 generic_dout(10) << __func__
<< " " << objects
<< dendl
;
3739 for (vector
<ghobject_t
>::iterator p
= objects
.begin();
3742 OSDriver::OSTransaction
_t(driver
.get_transaction(&t
));
3743 int r
= mapper
.remove_oid(p
->hobj
, &_t
);
3744 if (r
!= 0 && r
!= -ENOENT
)
3747 if (removed
> cct
->_conf
->osd_target_transaction_size
) {
3748 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3750 t
= ObjectStore::Transaction();
3754 t
.remove_collection(tmp
);
3755 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3759 if (!osr
->flush_commit(&waiter
)) {
3765 // ======================================================
3768 PGPool
OSD::_get_pool(int id
, OSDMapRef createmap
)
3770 if (!createmap
->have_pg_pool(id
)) {
3771 dout(5) << __func__
<< ": the OSDmap does not contain a PG pool with id = "
3776 PGPool p
= PGPool(cct
, createmap
, id
);
3778 dout(10) << "_get_pool " << p
.id
<< dendl
;
3782 PG
*OSD::_open_lock_pg(
3783 OSDMapRef createmap
,
3784 spg_t pgid
, bool no_lockdep_check
)
3786 assert(osd_lock
.is_locked());
3788 PG
* pg
= _make_pg(createmap
, pgid
);
3790 RWLock::WLocker
l(pg_map_lock
);
3791 pg
->lock(no_lockdep_check
);
3793 pg
->get("PGMap"); // because it's in pg_map
3794 service
.pg_add_epoch(pg
->info
.pgid
, createmap
->get_epoch());
3800 OSDMapRef createmap
,
3803 dout(10) << "_open_lock_pg " << pgid
<< dendl
;
3804 PGPool pool
= _get_pool(pgid
.pool(), createmap
);
3808 if (createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_REPLICATED
||
3809 createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_ERASURE
)
3810 pg
= new PrimaryLogPG(&service
, createmap
, pool
, pgid
);
3818 void OSD::add_newly_split_pg(PG
*pg
, PG::RecoveryCtx
*rctx
)
3820 epoch_t
e(service
.get_osdmap()->get_epoch());
3821 pg
->get("PGMap"); // For pg_map
3822 pg_map
[pg
->info
.pgid
] = pg
;
3823 service
.pg_add_epoch(pg
->info
.pgid
, pg
->get_osdmap()->get_epoch());
3825 dout(10) << "Adding newly split pg " << *pg
<< dendl
;
3826 pg
->handle_loaded(rctx
);
3827 pg
->write_if_dirty(*(rctx
->transaction
));
3828 pg
->queue_null(e
, e
);
3829 map
<spg_t
, list
<PG::CephPeeringEvtRef
> >::iterator to_wake
=
3830 peering_wait_for_split
.find(pg
->info
.pgid
);
3831 if (to_wake
!= peering_wait_for_split
.end()) {
3832 for (list
<PG::CephPeeringEvtRef
>::iterator i
=
3833 to_wake
->second
.begin();
3834 i
!= to_wake
->second
.end();
3836 pg
->queue_peering_event(*i
);
3838 peering_wait_for_split
.erase(to_wake
);
3840 if (!service
.get_osdmap()->have_pg_pool(pg
->info
.pgid
.pool()))
3844 OSD::res_result
OSD::_try_resurrect_pg(
3845 OSDMapRef curmap
, spg_t pgid
, spg_t
*resurrected
, PGRef
*old_pg_state
)
3847 assert(resurrected
);
3848 assert(old_pg_state
);
3849 // find nearest ancestor
3850 DeletingStateRef df
;
3853 df
= service
.deleting_pgs
.lookup(cur
);
3858 cur
= cur
.get_parent();
3861 return RES_NONE
; // good to go
3863 df
->old_pg_state
->lock();
3864 OSDMapRef create_map
= df
->old_pg_state
->get_osdmap();
3865 df
->old_pg_state
->unlock();
3867 set
<spg_t
> children
;
3869 if (df
->try_stop_deletion()) {
3870 dout(10) << __func__
<< ": halted deletion on pg " << pgid
<< dendl
;
3872 *old_pg_state
= df
->old_pg_state
;
3873 service
.deleting_pgs
.remove(pgid
); // PG is no longer being removed!
3876 // raced, ensure we don't see DeletingStateRef when we try to
3878 service
.deleting_pgs
.remove(pgid
);
3881 } else if (cur
.is_split(create_map
->get_pg_num(cur
.pool()),
3882 curmap
->get_pg_num(cur
.pool()),
3884 children
.count(pgid
)) {
3885 if (df
->try_stop_deletion()) {
3886 dout(10) << __func__
<< ": halted deletion on ancestor pg " << pgid
3889 *old_pg_state
= df
->old_pg_state
;
3890 service
.deleting_pgs
.remove(cur
); // PG is no longer being removed!
3893 /* this is not a problem, failing to cancel proves that all objects
3894 * have been removed, so no hobject_t overlap is possible
3902 PG
*OSD::_create_lock_pg(
3903 OSDMapRef createmap
,
3908 vector
<int>& up
, int up_primary
,
3909 vector
<int>& acting
, int acting_primary
,
3910 pg_history_t history
,
3911 const PastIntervals
& pi
,
3912 ObjectStore::Transaction
& t
)
3914 assert(osd_lock
.is_locked());
3915 dout(20) << "_create_lock_pg pgid " << pgid
<< dendl
;
3917 PG
*pg
= _open_lock_pg(createmap
, pgid
, true);
3919 service
.init_splits_between(pgid
, pg
->get_osdmap(), service
.get_osdmap());
3932 dout(7) << "_create_lock_pg " << *pg
<< dendl
;
3936 PG
*OSD::_lookup_lock_pg(spg_t pgid
)
3938 RWLock::RLocker
l(pg_map_lock
);
3940 auto pg_map_entry
= pg_map
.find(pgid
);
3941 if (pg_map_entry
== pg_map
.end())
3943 PG
*pg
= pg_map_entry
->second
;
3948 PG
*OSD::lookup_lock_pg(spg_t pgid
)
3950 return _lookup_lock_pg(pgid
);
3953 PG
*OSD::_lookup_lock_pg_with_map_lock_held(spg_t pgid
)
3955 assert(pg_map
.count(pgid
));
3956 PG
*pg
= pg_map
[pgid
];
3961 void OSD::load_pgs()
3963 assert(osd_lock
.is_locked());
3964 dout(0) << "load_pgs" << dendl
;
3966 RWLock::RLocker
l(pg_map_lock
);
3967 assert(pg_map
.empty());
3971 int r
= store
->list_collections(ls
);
3973 derr
<< "failed to list pgs: " << cpp_strerror(-r
) << dendl
;
3976 bool has_upgraded
= false;
3978 for (vector
<coll_t
>::iterator it
= ls
.begin();
3982 if (it
->is_temp(&pgid
) ||
3983 (it
->is_pg(&pgid
) && PG::_has_removal_flag(store
, pgid
))) {
3984 dout(10) << "load_pgs " << *it
<< " clearing temp" << dendl
;
3985 recursive_remove_collection(cct
, store
, pgid
, *it
);
3989 if (!it
->is_pg(&pgid
)) {
3990 dout(10) << "load_pgs ignoring unrecognized " << *it
<< dendl
;
3994 if (pgid
.preferred() >= 0) {
3995 dout(10) << __func__
<< ": skipping localized PG " << pgid
<< dendl
;
3996 // FIXME: delete it too, eventually
4000 dout(10) << "pgid " << pgid
<< " coll " << coll_t(pgid
) << dendl
;
4002 epoch_t map_epoch
= 0;
4003 int r
= PG::peek_map_epoch(store
, pgid
, &map_epoch
, &bl
);
4005 derr
<< __func__
<< " unable to peek at " << pgid
<< " metadata, skipping"
4011 if (map_epoch
> 0) {
4012 OSDMapRef pgosdmap
= service
.try_get_map(map_epoch
);
4014 if (!osdmap
->have_pg_pool(pgid
.pool())) {
4015 derr
<< __func__
<< ": could not find map for epoch " << map_epoch
4016 << " on pg " << pgid
<< ", but the pool is not present in the "
4017 << "current map, so this is probably a result of bug 10617. "
4018 << "Skipping the pg for now, you can use ceph-objectstore-tool "
4019 << "to clean it up later." << dendl
;
4022 derr
<< __func__
<< ": have pgid " << pgid
<< " at epoch "
4023 << map_epoch
<< ", but missing map. Crashing."
4025 assert(0 == "Missing map in load_pgs");
4028 pg
= _open_lock_pg(pgosdmap
, pgid
);
4030 pg
= _open_lock_pg(osdmap
, pgid
);
4032 // there can be no waiters here, so we don't call wake_pg_waiters
4034 pg
->ch
= store
->open_collection(pg
->coll
);
4036 // read pg state, log
4037 pg
->read_state(store
, bl
);
4039 if (pg
->must_upgrade()) {
4040 if (!pg
->can_upgrade()) {
4041 derr
<< "PG needs upgrade, but on-disk data is too old; upgrade to"
4042 << " an older version first." << dendl
;
4043 assert(0 == "PG too old to upgrade");
4045 if (!has_upgraded
) {
4046 derr
<< "PGs are upgrading" << dendl
;
4047 has_upgraded
= true;
4049 dout(10) << "PG " << pg
->info
.pgid
4050 << " must upgrade..." << dendl
;
4055 dout(10) << "load_pgs " << *it
<< " deleting dne" << dendl
;
4057 service
.pg_remove_epoch(pg
->pg_id
);
4061 RWLock::WLocker
l(pg_map_lock
);
4062 auto p
= pg_map
.find(pg
->get_pgid());
4063 assert(p
!= pg_map
.end() && p
->second
== pg
);
4064 dout(20) << __func__
<< " removed pg " << pg
<< " from pg_map" << dendl
;
4068 recursive_remove_collection(cct
, store
, pgid
, *it
);
4072 service
.init_splits_between(pg
->info
.pgid
, pg
->get_osdmap(), osdmap
);
4074 // generate state for PG's current mapping
4075 int primary
, up_primary
;
4076 vector
<int> acting
, up
;
4077 pg
->get_osdmap()->pg_to_up_acting_osds(
4078 pgid
.pgid
, &up
, &up_primary
, &acting
, &primary
);
4079 pg
->init_primary_up_acting(
4084 int role
= OSDMap::calc_pg_role(whoami
, pg
->acting
);
4085 if (pg
->pool
.info
.is_replicated() || role
== pg
->pg_whoami
.shard
)
4090 pg
->reg_next_scrub();
4092 PG::RecoveryCtx
rctx(0, 0, 0, 0, 0, 0);
4093 pg
->handle_loaded(&rctx
);
4095 dout(10) << "load_pgs loaded " << *pg
<< " " << pg
->pg_log
.get_log() << dendl
;
4096 if (pg
->pg_log
.is_dirty()) {
4097 ObjectStore::Transaction t
;
4098 pg
->write_if_dirty(t
);
4099 store
->apply_transaction(pg
->osr
.get(), std::move(t
));
4104 RWLock::RLocker
l(pg_map_lock
);
4105 dout(0) << "load_pgs opened " << pg_map
.size() << " pgs" << dendl
;
4108 // clean up old infos object?
4109 if (has_upgraded
&& store
->exists(coll_t::meta(), OSD::make_infos_oid())) {
4110 dout(1) << __func__
<< " removing legacy infos object" << dendl
;
4111 ObjectStore::Transaction t
;
4112 t
.remove(coll_t::meta(), OSD::make_infos_oid());
4113 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4115 derr
<< __func__
<< ": apply_transaction returned "
4116 << cpp_strerror(r
) << dendl
;
4121 build_past_intervals_parallel();
4126 * build past_intervals efficiently on old, degraded, and buried
4127 * clusters. this is important for efficiently catching up osds that
4128 * are way behind on maps to the current cluster state.
4130 * this is a parallel version of PG::generate_past_intervals().
4131 * follow the same logic, but do all pgs at the same time so that we
4132 * can make a single pass across the osdmap history.
4134 void OSD::build_past_intervals_parallel()
4138 vector
<int> old_acting
, old_up
;
4139 epoch_t same_interval_since
;
4143 map
<PG
*,pistate
> pis
;
4145 // calculate junction of map range
4146 epoch_t end_epoch
= superblock
.oldest_map
;
4147 epoch_t cur_epoch
= superblock
.newest_map
;
4149 RWLock::RLocker
l(pg_map_lock
);
4150 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4155 // Ignore PGs only partially created (DNE)
4156 if (pg
->info
.dne()) {
4160 auto rpib
= pg
->get_required_past_interval_bounds(
4162 superblock
.oldest_map
);
4163 if (rpib
.first
>= rpib
.second
&& pg
->past_intervals
.empty()) {
4164 if (pg
->info
.history
.same_interval_since
== 0) {
4165 pg
->info
.history
.same_interval_since
= rpib
.second
;
4169 auto apib
= pg
->past_intervals
.get_bounds();
4170 if (apib
.second
>= rpib
.second
&&
4171 apib
.first
<= rpib
.first
) {
4172 if (pg
->info
.history
.same_interval_since
== 0) {
4173 pg
->info
.history
.same_interval_since
= rpib
.second
;
4179 dout(10) << pg
->info
.pgid
<< " needs " << rpib
.first
<< "-"
4180 << rpib
.second
<< dendl
;
4181 pistate
& p
= pis
[pg
];
4182 p
.start
= rpib
.first
;
4183 p
.end
= rpib
.second
;
4184 p
.same_interval_since
= 0;
4186 if (rpib
.first
< cur_epoch
)
4187 cur_epoch
= rpib
.first
;
4188 if (rpib
.second
> end_epoch
)
4189 end_epoch
= rpib
.second
;
4193 dout(10) << __func__
<< " nothing to build" << dendl
;
4197 dout(1) << __func__
<< " over " << cur_epoch
<< "-" << end_epoch
<< dendl
;
4198 assert(cur_epoch
<= end_epoch
);
4200 OSDMapRef cur_map
, last_map
;
4201 for ( ; cur_epoch
<= end_epoch
; cur_epoch
++) {
4202 dout(10) << __func__
<< " epoch " << cur_epoch
<< dendl
;
4204 cur_map
= get_map(cur_epoch
);
4206 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4208 pistate
& p
= i
->second
;
4210 if (cur_epoch
< p
.start
|| cur_epoch
> p
.end
)
4213 vector
<int> acting
, up
;
4216 pg_t pgid
= pg
->info
.pgid
.pgid
;
4217 if (p
.same_interval_since
&& last_map
->get_pools().count(pgid
.pool()))
4218 pgid
= pgid
.get_ancestor(last_map
->get_pg_num(pgid
.pool()));
4219 cur_map
->pg_to_up_acting_osds(
4220 pgid
, &up
, &up_primary
, &acting
, &primary
);
4222 if (p
.same_interval_since
== 0) {
4223 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4224 << " first map, acting " << acting
4225 << " up " << up
<< ", same_interval_since = " << cur_epoch
<< dendl
;
4226 p
.same_interval_since
= cur_epoch
;
4228 p
.old_acting
= acting
;
4229 p
.primary
= primary
;
4230 p
.up_primary
= up_primary
;
4235 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable(
4236 pg
->get_is_recoverable_predicate());
4237 std::stringstream debug
;
4238 bool new_interval
= PastIntervals::check_new_interval(
4241 p
.old_acting
, acting
,
4245 p
.same_interval_since
,
4246 pg
->info
.history
.last_epoch_clean
,
4250 &pg
->past_intervals
,
4253 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4254 << " " << debug
.str() << dendl
;
4256 p
.old_acting
= acting
;
4257 p
.primary
= primary
;
4258 p
.up_primary
= up_primary
;
4259 p
.same_interval_since
= cur_epoch
;
4264 // Now that past_intervals have been recomputed let's fix the same_interval_since
4265 // if it was cleared by import.
4266 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4268 pistate
& p
= i
->second
;
4270 if (pg
->info
.history
.same_interval_since
== 0) {
4271 assert(p
.same_interval_since
);
4272 dout(10) << __func__
<< " fix same_interval_since " << p
.same_interval_since
<< " pg " << *pg
<< dendl
;
4273 dout(10) << __func__
<< " past_intervals " << pg
->past_intervals
<< dendl
;
4275 pg
->info
.history
.same_interval_since
= p
.same_interval_since
;
4279 // write info only at the end. this is necessary because we check
4280 // whether the past_intervals go far enough back or forward in time,
4281 // but we don't check for holes. we could avoid it by discarding
4282 // the previous past_intervals and rebuilding from scratch, or we
4283 // can just do this and commit all our work at the end.
4284 ObjectStore::Transaction t
;
4286 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4289 pg
->dirty_big_info
= true;
4290 pg
->dirty_info
= true;
4291 pg
->write_if_dirty(t
);
4294 // don't let the transaction get too big
4295 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
4296 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4297 t
= ObjectStore::Transaction();
4302 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4306 * look up a pg. if we have it, great. if not, consider creating it IF the pg mapping
4307 * hasn't changed since the given epoch and we are the primary.
4309 int OSD::handle_pg_peering_evt(
4311 const pg_history_t
& orig_history
,
4312 const PastIntervals
& pi
,
4314 PG::CephPeeringEvtRef evt
)
4316 if (service
.splitting(pgid
)) {
4317 peering_wait_for_split
[pgid
].push_back(evt
);
4321 PG
*pg
= _lookup_lock_pg(pgid
);
4324 if (!osdmap
->have_pg_pool(pgid
.pool()))
4326 int up_primary
, acting_primary
;
4327 vector
<int> up
, acting
;
4328 osdmap
->pg_to_up_acting_osds(
4329 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4331 pg_history_t history
= orig_history
;
4332 bool valid_history
= project_pg_history(
4333 pgid
, history
, epoch
, up
, up_primary
, acting
, acting_primary
);
4335 if (!valid_history
|| epoch
< history
.same_interval_since
) {
4336 dout(10) << __func__
<< pgid
<< " acting changed in "
4337 << history
.same_interval_since
<< " (msg from " << epoch
<< ")"
4342 if (service
.splitting(pgid
)) {
4346 const bool is_mon_create
=
4347 evt
->get_event().dynamic_type() == PG::NullEvt::static_type();
4348 if (maybe_wait_for_max_pg(pgid
, is_mon_create
)) {
4351 // do we need to resurrect a deleting pg?
4354 res_result result
= _try_resurrect_pg(
4355 service
.get_osdmap(),
4360 PG::RecoveryCtx rctx
= create_context();
4363 const pg_pool_t
* pp
= osdmap
->get_pg_pool(pgid
.pool());
4364 if (pp
->has_flag(pg_pool_t::FLAG_EC_OVERWRITES
) &&
4365 store
->get_type() != "bluestore") {
4366 clog
->warn() << "pg " << pgid
4367 << " is at risk of silent data corruption: "
4368 << "the pool allows ec overwrites but is not stored in "
4369 << "bluestore, so deep scrubbing will not detect bitrot";
4371 PG::_create(*rctx
.transaction
, pgid
, pgid
.get_split_bits(pp
->get_pg_num()));
4372 PG::_init(*rctx
.transaction
, pgid
, pp
);
4374 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
4375 if (!pp
->is_replicated() && role
!= pgid
.shard
)
4378 pg
= _create_lock_pg(
4383 acting
, acting_primary
,
4386 pg
->handle_create(&rctx
);
4387 pg
->write_if_dirty(*rctx
.transaction
);
4388 dispatch_context(rctx
, pg
, osdmap
);
4390 dout(10) << *pg
<< " is new" << dendl
;
4392 pg
->queue_peering_event(evt
);
4393 wake_pg_waiters(pg
);
4398 old_pg_state
->lock();
4399 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4400 int old_role
= old_pg_state
->role
;
4401 vector
<int> old_up
= old_pg_state
->up
;
4402 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4403 vector
<int> old_acting
= old_pg_state
->acting
;
4404 int old_primary
= old_pg_state
->primary
.osd
;
4405 pg_history_t old_history
= old_pg_state
->info
.history
;
4406 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4407 old_pg_state
->unlock();
4408 pg
= _create_lock_pg(
4421 pg
->handle_create(&rctx
);
4422 pg
->write_if_dirty(*rctx
.transaction
);
4423 dispatch_context(rctx
, pg
, osdmap
);
4425 dout(10) << *pg
<< " is new (resurrected)" << dendl
;
4427 pg
->queue_peering_event(evt
);
4428 wake_pg_waiters(pg
);
4433 assert(old_pg_state
);
4434 old_pg_state
->lock();
4435 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4436 int old_role
= old_pg_state
->role
;
4437 vector
<int> old_up
= old_pg_state
->up
;
4438 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4439 vector
<int> old_acting
= old_pg_state
->acting
;
4440 int old_primary
= old_pg_state
->primary
.osd
;
4441 pg_history_t old_history
= old_pg_state
->info
.history
;
4442 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4443 old_pg_state
->unlock();
4444 PG
*parent
= _create_lock_pg(
4458 parent
->handle_create(&rctx
);
4459 parent
->write_if_dirty(*rctx
.transaction
);
4460 dispatch_context(rctx
, parent
, osdmap
);
4462 dout(10) << *parent
<< " is new" << dendl
;
4464 assert(service
.splitting(pgid
));
4465 peering_wait_for_split
[pgid
].push_back(evt
);
4467 //parent->queue_peering_event(evt);
4468 parent
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
4469 wake_pg_waiters(parent
);
4478 // already had it. did the mapping change?
4479 if (epoch
< pg
->info
.history
.same_interval_since
) {
4480 dout(10) << *pg
<< __func__
<< " acting changed in "
4481 << pg
->info
.history
.same_interval_since
4482 << " (msg from " << epoch
<< ")" << dendl
;
4484 pg
->queue_peering_event(evt
);
4491 bool OSD::maybe_wait_for_max_pg(spg_t pgid
, bool is_mon_create
)
4493 const auto max_pgs_per_osd
=
4494 (cct
->_conf
->get_val
<uint64_t>("mon_max_pg_per_osd") *
4495 cct
->_conf
->get_val
<double>("osd_max_pg_per_osd_hard_ratio"));
4497 RWLock::RLocker pg_map_locker
{pg_map_lock
};
4498 if (pg_map
.size() < max_pgs_per_osd
) {
4501 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
4502 if (is_mon_create
) {
4503 pending_creates_from_mon
++;
4505 bool is_primary
= osdmap
->get_pg_acting_rank(pgid
.pgid
, whoami
) == 0;
4506 pending_creates_from_osd
.emplace(pgid
.pgid
, is_primary
);
4508 dout(5) << __func__
<< " withhold creation of pg " << pgid
4509 << ": " << pg_map
.size() << " >= "<< max_pgs_per_osd
<< dendl
;
4513 // to re-trigger a peering, we have to twiddle the pg mapping a little bit,
4514 // see PG::should_restart_peering(). OSDMap::pg_to_up_acting_osds() will turn
4515 // to up set if pg_temp is empty. so an empty pg_temp won't work.
4516 static vector
<int32_t> twiddle(const vector
<int>& acting
) {
4517 if (acting
.size() > 1) {
4520 vector
<int32_t> twiddled(acting
.begin(), acting
.end());
4521 twiddled
.push_back(-1);
4526 void OSD::resume_creating_pg()
4528 bool do_sub_pg_creates
= false;
4529 bool have_pending_creates
= false;
4531 const auto max_pgs_per_osd
=
4532 (cct
->_conf
->get_val
<uint64_t>("mon_max_pg_per_osd") *
4533 cct
->_conf
->get_val
<double>("osd_max_pg_per_osd_hard_ratio"));
4534 RWLock::RLocker
l(pg_map_lock
);
4535 if (max_pgs_per_osd
<= pg_map
.size()) {
4536 // this could happen if admin decreases this setting before a PG is removed
4539 unsigned spare_pgs
= max_pgs_per_osd
- pg_map
.size();
4540 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
4541 if (pending_creates_from_mon
> 0) {
4542 do_sub_pg_creates
= true;
4543 if (pending_creates_from_mon
>= spare_pgs
) {
4544 spare_pgs
= pending_creates_from_mon
= 0;
4546 spare_pgs
-= pending_creates_from_mon
;
4547 pending_creates_from_mon
= 0;
4550 auto pg
= pending_creates_from_osd
.cbegin();
4551 while (spare_pgs
> 0 && pg
!= pending_creates_from_osd
.cend()) {
4552 dout(20) << __func__
<< " pg " << pg
->first
<< dendl
;
4554 osdmap
->pg_to_up_acting_osds(pg
->first
, nullptr, nullptr, &acting
, nullptr);
4555 service
.queue_want_pg_temp(pg
->first
, twiddle(acting
), true);
4556 pg
= pending_creates_from_osd
.erase(pg
);
4557 do_sub_pg_creates
= true;
4560 have_pending_creates
= (pending_creates_from_mon
> 0 ||
4561 !pending_creates_from_osd
.empty());
4564 bool do_renew_subs
= false;
4565 if (do_sub_pg_creates
) {
4566 if (monc
->sub_want("osd_pg_creates", last_pg_create_epoch
, 0)) {
4567 dout(4) << __func__
<< ": resolicit pg creates from mon since "
4568 << last_pg_create_epoch
<< dendl
;
4569 do_renew_subs
= true;
4572 version_t start
= osdmap
->get_epoch() + 1;
4573 if (have_pending_creates
) {
4574 // don't miss any new osdmap deleting PGs
4575 if (monc
->sub_want("osdmap", start
, 0)) {
4576 dout(4) << __func__
<< ": resolicit osdmap from mon since "
4578 do_renew_subs
= true;
4580 } else if (do_sub_pg_creates
) {
4581 // no need to subscribe the osdmap continuously anymore
4582 // once the pgtemp and/or mon_subscribe(pg_creates) is sent
4583 if (monc
->sub_want_increment("osdmap", start
, CEPH_SUBSCRIBE_ONETIME
)) {
4584 dout(4) << __func__
<< ": re-subscribe osdmap(onetime) since"
4586 do_renew_subs
= true;
4590 if (do_renew_subs
) {
4594 service
.send_pg_temp();
4597 void OSD::build_initial_pg_history(
4600 utime_t created_stamp
,
4604 dout(10) << __func__
<< " " << pgid
<< " created " << created
<< dendl
;
4605 h
->epoch_created
= created
;
4606 h
->epoch_pool_created
= created
;
4607 h
->same_interval_since
= created
;
4608 h
->same_up_since
= created
;
4609 h
->same_primary_since
= created
;
4610 h
->last_scrub_stamp
= created_stamp
;
4611 h
->last_deep_scrub_stamp
= created_stamp
;
4612 h
->last_clean_scrub_stamp
= created_stamp
;
4614 OSDMapRef lastmap
= service
.get_map(created
);
4615 int up_primary
, acting_primary
;
4616 vector
<int> up
, acting
;
4617 lastmap
->pg_to_up_acting_osds(
4618 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4620 ostringstream debug
;
4621 for (epoch_t e
= created
+ 1; e
<= osdmap
->get_epoch(); ++e
) {
4622 OSDMapRef osdmap
= service
.get_map(e
);
4623 int new_up_primary
, new_acting_primary
;
4624 vector
<int> new_up
, new_acting
;
4625 osdmap
->pg_to_up_acting_osds(
4626 pgid
.pgid
, &new_up
, &new_up_primary
, &new_acting
, &new_acting_primary
);
4628 // this is a bit imprecise, but sufficient?
4629 struct min_size_predicate_t
: public IsPGRecoverablePredicate
{
4630 const pg_pool_t
*pi
;
4631 bool operator()(const set
<pg_shard_t
> &have
) const {
4632 return have
.size() >= pi
->min_size
;
4634 min_size_predicate_t(const pg_pool_t
*i
) : pi(i
) {}
4635 } min_size_predicate(osdmap
->get_pg_pool(pgid
.pgid
.pool()));
4637 bool new_interval
= PastIntervals::check_new_interval(
4644 h
->same_interval_since
,
4645 h
->last_epoch_clean
,
4649 &min_size_predicate
,
4653 h
->same_interval_since
= e
;
4655 h
->same_up_since
= e
;
4657 if (acting_primary
!= new_acting_primary
) {
4658 h
->same_primary_since
= e
;
4660 if (pgid
.pgid
.is_split(lastmap
->get_pg_num(pgid
.pgid
.pool()),
4661 osdmap
->get_pg_num(pgid
.pgid
.pool()),
4663 h
->last_epoch_split
= e
;
4666 acting
= new_acting
;
4667 up_primary
= new_up_primary
;
4668 acting_primary
= new_acting_primary
;
4672 dout(20) << __func__
<< " " << debug
.str() << dendl
;
4673 dout(10) << __func__
<< " " << *h
<< " " << *pi
4674 << " [" << (pi
->empty() ? pair
<epoch_t
,epoch_t
>(0,0) :
4675 pi
->get_bounds()) << ")"
4680 * Fill in the passed history so you know same_interval_since, same_up_since,
4681 * and same_primary_since.
4683 bool OSD::project_pg_history(spg_t pgid
, pg_history_t
& h
, epoch_t from
,
4684 const vector
<int>& currentup
,
4685 int currentupprimary
,
4686 const vector
<int>& currentacting
,
4687 int currentactingprimary
)
4689 dout(15) << "project_pg_history " << pgid
4690 << " from " << from
<< " to " << osdmap
->get_epoch()
4695 for (e
= osdmap
->get_epoch();
4698 // verify during intermediate epoch (e-1)
4699 OSDMapRef oldmap
= service
.try_get_map(e
-1);
4701 dout(15) << __func__
<< ": found map gap, returning false" << dendl
;
4704 assert(oldmap
->have_pg_pool(pgid
.pool()));
4706 int upprimary
, actingprimary
;
4707 vector
<int> up
, acting
;
4708 oldmap
->pg_to_up_acting_osds(
4715 // acting set change?
4716 if ((actingprimary
!= currentactingprimary
||
4717 upprimary
!= currentupprimary
||
4718 acting
!= currentacting
||
4719 up
!= currentup
) && e
> h
.same_interval_since
) {
4720 dout(15) << "project_pg_history " << pgid
<< " acting|up changed in " << e
4721 << " from " << acting
<< "/" << up
4722 << " " << actingprimary
<< "/" << upprimary
4723 << " -> " << currentacting
<< "/" << currentup
4724 << " " << currentactingprimary
<< "/" << currentupprimary
4726 h
.same_interval_since
= e
;
4729 if (pgid
.is_split(oldmap
->get_pg_num(pgid
.pool()),
4730 osdmap
->get_pg_num(pgid
.pool()),
4731 0) && e
> h
.same_interval_since
) {
4732 h
.same_interval_since
= e
;
4735 if ((up
!= currentup
|| upprimary
!= currentupprimary
)
4736 && e
> h
.same_up_since
) {
4737 dout(15) << "project_pg_history " << pgid
<< " up changed in " << e
4738 << " from " << up
<< " " << upprimary
4739 << " -> " << currentup
<< " " << currentupprimary
<< dendl
;
4740 h
.same_up_since
= e
;
4744 if (OSDMap::primary_changed(
4747 currentactingprimary
,
4749 e
> h
.same_primary_since
) {
4750 dout(15) << "project_pg_history " << pgid
<< " primary changed in " << e
<< dendl
;
4751 h
.same_primary_since
= e
;
4754 if (h
.same_interval_since
>= e
&& h
.same_up_since
>= e
&& h
.same_primary_since
>= e
)
4758 // base case: these floors should be the pg creation epoch if we didn't
4759 // find any changes.
4760 if (e
== h
.epoch_created
) {
4761 if (!h
.same_interval_since
)
4762 h
.same_interval_since
= e
;
4763 if (!h
.same_up_since
)
4764 h
.same_up_since
= e
;
4765 if (!h
.same_primary_since
)
4766 h
.same_primary_since
= e
;
4769 dout(15) << "project_pg_history end " << h
<< dendl
;
4775 void OSD::_add_heartbeat_peer(int p
)
4781 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(p
);
4782 if (i
== heartbeat_peers
.end()) {
4783 pair
<ConnectionRef
,ConnectionRef
> cons
= service
.get_con_osd_hb(p
, osdmap
->get_epoch());
4786 hi
= &heartbeat_peers
[p
];
4788 HeartbeatSession
*s
= new HeartbeatSession(p
);
4789 hi
->con_back
= cons
.first
.get();
4790 hi
->con_back
->set_priv(s
->get());
4792 hi
->con_front
= cons
.second
.get();
4793 hi
->con_front
->set_priv(s
->get());
4794 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4795 << " " << hi
->con_back
->get_peer_addr()
4796 << " " << hi
->con_front
->get_peer_addr()
4799 hi
->con_front
.reset(NULL
);
4800 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4801 << " " << hi
->con_back
->get_peer_addr()
4808 hi
->epoch
= osdmap
->get_epoch();
4811 void OSD::_remove_heartbeat_peer(int n
)
4813 map
<int,HeartbeatInfo
>::iterator q
= heartbeat_peers
.find(n
);
4814 assert(q
!= heartbeat_peers
.end());
4815 dout(20) << " removing heartbeat peer osd." << n
4816 << " " << q
->second
.con_back
->get_peer_addr()
4817 << " " << (q
->second
.con_front
? q
->second
.con_front
->get_peer_addr() : entity_addr_t())
4819 q
->second
.con_back
->mark_down();
4820 if (q
->second
.con_front
) {
4821 q
->second
.con_front
->mark_down();
4823 heartbeat_peers
.erase(q
);
4826 void OSD::need_heartbeat_peer_update()
4830 dout(20) << "need_heartbeat_peer_update" << dendl
;
4831 heartbeat_set_peers_need_update();
4834 void OSD::maybe_update_heartbeat_peers()
4836 assert(osd_lock
.is_locked());
4838 if (is_waiting_for_healthy()) {
4839 utime_t now
= ceph_clock_now();
4840 if (last_heartbeat_resample
== utime_t()) {
4841 last_heartbeat_resample
= now
;
4842 heartbeat_set_peers_need_update();
4843 } else if (!heartbeat_peers_need_update()) {
4844 utime_t dur
= now
- last_heartbeat_resample
;
4845 if (dur
> cct
->_conf
->osd_heartbeat_grace
) {
4846 dout(10) << "maybe_update_heartbeat_peers forcing update after " << dur
<< " seconds" << dendl
;
4847 heartbeat_set_peers_need_update();
4848 last_heartbeat_resample
= now
;
4849 reset_heartbeat_peers(); // we want *new* peers!
4854 if (!heartbeat_peers_need_update())
4856 heartbeat_clear_peers_need_update();
4858 Mutex::Locker
l(heartbeat_lock
);
4860 dout(10) << "maybe_update_heartbeat_peers updating" << dendl
;
4863 // build heartbeat from set
4865 RWLock::RLocker
l(pg_map_lock
);
4866 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4870 pg
->heartbeat_peer_lock
.Lock();
4871 dout(20) << i
->first
<< " heartbeat_peers " << pg
->heartbeat_peers
<< dendl
;
4872 for (set
<int>::iterator p
= pg
->heartbeat_peers
.begin();
4873 p
!= pg
->heartbeat_peers
.end();
4875 if (osdmap
->is_up(*p
))
4876 _add_heartbeat_peer(*p
);
4877 for (set
<int>::iterator p
= pg
->probe_targets
.begin();
4878 p
!= pg
->probe_targets
.end();
4880 if (osdmap
->is_up(*p
))
4881 _add_heartbeat_peer(*p
);
4882 pg
->heartbeat_peer_lock
.Unlock();
4886 // include next and previous up osds to ensure we have a fully-connected set
4887 set
<int> want
, extras
;
4888 int next
= osdmap
->get_next_up_osd_after(whoami
);
4891 int prev
= osdmap
->get_previous_up_osd_before(whoami
);
4892 if (prev
>= 0 && prev
!= next
)
4895 for (set
<int>::iterator p
= want
.begin(); p
!= want
.end(); ++p
) {
4896 dout(10) << " adding neighbor peer osd." << *p
<< dendl
;
4898 _add_heartbeat_peer(*p
);
4901 // remove down peers; enumerate extras
4902 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4903 while (p
!= heartbeat_peers
.end()) {
4904 if (!osdmap
->is_up(p
->first
)) {
4907 _remove_heartbeat_peer(o
);
4910 if (p
->second
.epoch
< osdmap
->get_epoch()) {
4911 extras
.insert(p
->first
);
4917 int start
= osdmap
->get_next_up_osd_after(whoami
);
4918 for (int n
= start
; n
>= 0; ) {
4919 if ((int)heartbeat_peers
.size() >= cct
->_conf
->osd_heartbeat_min_peers
)
4921 if (!extras
.count(n
) && !want
.count(n
) && n
!= whoami
) {
4922 dout(10) << " adding random peer osd." << n
<< dendl
;
4924 _add_heartbeat_peer(n
);
4926 n
= osdmap
->get_next_up_osd_after(n
);
4928 break; // came full circle; stop
4932 for (set
<int>::iterator p
= extras
.begin();
4933 (int)heartbeat_peers
.size() > cct
->_conf
->osd_heartbeat_min_peers
&& p
!= extras
.end();
4937 _remove_heartbeat_peer(*p
);
4940 dout(10) << "maybe_update_heartbeat_peers " << heartbeat_peers
.size() << " peers, extras " << extras
<< dendl
;
4943 void OSD::reset_heartbeat_peers()
4945 assert(osd_lock
.is_locked());
4946 dout(10) << "reset_heartbeat_peers" << dendl
;
4947 Mutex::Locker
l(heartbeat_lock
);
4948 while (!heartbeat_peers
.empty()) {
4949 HeartbeatInfo
& hi
= heartbeat_peers
.begin()->second
;
4950 hi
.con_back
->mark_down();
4952 hi
.con_front
->mark_down();
4954 heartbeat_peers
.erase(heartbeat_peers
.begin());
4956 failure_queue
.clear();
4959 void OSD::handle_osd_ping(MOSDPing
*m
)
4961 if (superblock
.cluster_fsid
!= m
->fsid
) {
4962 dout(20) << "handle_osd_ping from " << m
->get_source_inst()
4963 << " bad fsid " << m
->fsid
<< " != " << superblock
.cluster_fsid
<< dendl
;
4968 int from
= m
->get_source().num();
4970 heartbeat_lock
.Lock();
4971 if (is_stopping()) {
4972 heartbeat_lock
.Unlock();
4977 OSDMapRef curmap
= service
.get_osdmap();
4979 heartbeat_lock
.Unlock();
4986 case MOSDPing::PING
:
4988 if (cct
->_conf
->osd_debug_drop_ping_probability
> 0) {
4989 auto heartbeat_drop
= debug_heartbeat_drops_remaining
.find(from
);
4990 if (heartbeat_drop
!= debug_heartbeat_drops_remaining
.end()) {
4991 if (heartbeat_drop
->second
== 0) {
4992 debug_heartbeat_drops_remaining
.erase(heartbeat_drop
);
4994 --heartbeat_drop
->second
;
4995 dout(5) << "Dropping heartbeat from " << from
4996 << ", " << heartbeat_drop
->second
4997 << " remaining to drop" << dendl
;
5000 } else if (cct
->_conf
->osd_debug_drop_ping_probability
>
5001 ((((double)(rand()%100))/100.0))) {
5003 debug_heartbeat_drops_remaining
.insert(std::make_pair(from
,
5004 cct
->_conf
->osd_debug_drop_ping_duration
)).first
;
5005 dout(5) << "Dropping heartbeat from " << from
5006 << ", " << heartbeat_drop
->second
5007 << " remaining to drop" << dendl
;
5012 if (!cct
->get_heartbeat_map()->is_healthy()) {
5013 dout(10) << "internal heartbeat not healthy, dropping ping request" << dendl
;
5017 Message
*r
= new MOSDPing(monc
->get_fsid(),
5018 curmap
->get_epoch(),
5019 MOSDPing::PING_REPLY
, m
->stamp
,
5020 cct
->_conf
->osd_heartbeat_min_size
);
5021 m
->get_connection()->send_message(r
);
5023 if (curmap
->is_up(from
)) {
5024 service
.note_peer_epoch(from
, m
->map_epoch
);
5026 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
5028 service
.share_map_peer(from
, con
.get());
5031 } else if (!curmap
->exists(from
) ||
5032 curmap
->get_down_at(from
) > m
->map_epoch
) {
5033 // tell them they have died
5034 Message
*r
= new MOSDPing(monc
->get_fsid(),
5035 curmap
->get_epoch(),
5038 cct
->_conf
->osd_heartbeat_min_size
);
5039 m
->get_connection()->send_message(r
);
5044 case MOSDPing::PING_REPLY
:
5046 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(from
);
5047 if (i
!= heartbeat_peers
.end()) {
5048 if (m
->get_connection() == i
->second
.con_back
) {
5049 dout(25) << "handle_osd_ping got reply from osd." << from
5050 << " first_tx " << i
->second
.first_tx
5051 << " last_tx " << i
->second
.last_tx
5052 << " last_rx_back " << i
->second
.last_rx_back
<< " -> " << m
->stamp
5053 << " last_rx_front " << i
->second
.last_rx_front
5055 i
->second
.last_rx_back
= m
->stamp
;
5056 // if there is no front con, set both stamps.
5057 if (i
->second
.con_front
== NULL
)
5058 i
->second
.last_rx_front
= m
->stamp
;
5059 } else if (m
->get_connection() == i
->second
.con_front
) {
5060 dout(25) << "handle_osd_ping got reply from osd." << from
5061 << " first_tx " << i
->second
.first_tx
5062 << " last_tx " << i
->second
.last_tx
5063 << " last_rx_back " << i
->second
.last_rx_back
5064 << " last_rx_front " << i
->second
.last_rx_front
<< " -> " << m
->stamp
5066 i
->second
.last_rx_front
= m
->stamp
;
5069 utime_t cutoff
= ceph_clock_now();
5070 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
5071 if (i
->second
.is_healthy(cutoff
)) {
5072 // Cancel false reports
5073 auto failure_queue_entry
= failure_queue
.find(from
);
5074 if (failure_queue_entry
!= failure_queue
.end()) {
5075 dout(10) << "handle_osd_ping canceling queued "
5076 << "failure report for osd." << from
<< dendl
;
5077 failure_queue
.erase(failure_queue_entry
);
5080 auto failure_pending_entry
= failure_pending
.find(from
);
5081 if (failure_pending_entry
!= failure_pending
.end()) {
5082 dout(10) << "handle_osd_ping canceling in-flight "
5083 << "failure report for osd." << from
<< dendl
;
5084 send_still_alive(curmap
->get_epoch(),
5085 failure_pending_entry
->second
.second
);
5086 failure_pending
.erase(failure_pending_entry
);
5092 curmap
->is_up(from
)) {
5093 service
.note_peer_epoch(from
, m
->map_epoch
);
5095 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
5097 service
.share_map_peer(from
, con
.get());
5104 case MOSDPing::YOU_DIED
:
5105 dout(10) << "handle_osd_ping " << m
->get_source_inst()
5106 << " says i am down in " << m
->map_epoch
<< dendl
;
5107 osdmap_subscribe(curmap
->get_epoch()+1, false);
5111 heartbeat_lock
.Unlock();
5115 void OSD::heartbeat_entry()
5117 Mutex::Locker
l(heartbeat_lock
);
5120 while (!heartbeat_stop
) {
5123 double wait
= .5 + ((float)(rand() % 10)/10.0) * (float)cct
->_conf
->osd_heartbeat_interval
;
5125 w
.set_from_double(wait
);
5126 dout(30) << "heartbeat_entry sleeping for " << wait
<< dendl
;
5127 heartbeat_cond
.WaitInterval(heartbeat_lock
, w
);
5130 dout(30) << "heartbeat_entry woke up" << dendl
;
5134 void OSD::heartbeat_check()
5136 assert(heartbeat_lock
.is_locked());
5137 utime_t now
= ceph_clock_now();
5139 // check for heartbeat replies (move me elsewhere?)
5140 utime_t cutoff
= now
;
5141 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
5142 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
5143 p
!= heartbeat_peers
.end();
5146 if (p
->second
.first_tx
== utime_t()) {
5147 dout(25) << "heartbeat_check we haven't sent ping to osd." << p
->first
5148 << "yet, skipping" << dendl
;
5152 dout(25) << "heartbeat_check osd." << p
->first
5153 << " first_tx " << p
->second
.first_tx
5154 << " last_tx " << p
->second
.last_tx
5155 << " last_rx_back " << p
->second
.last_rx_back
5156 << " last_rx_front " << p
->second
.last_rx_front
5158 if (p
->second
.is_unhealthy(cutoff
)) {
5159 if (p
->second
.last_rx_back
== utime_t() ||
5160 p
->second
.last_rx_front
== utime_t()) {
5161 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
5162 << " osd." << p
->first
<< " ever on either front or back, first ping sent "
5163 << p
->second
.first_tx
<< " (cutoff " << cutoff
<< ")" << dendl
;
5165 failure_queue
[p
->first
] = p
->second
.last_tx
;
5167 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
5168 << " osd." << p
->first
<< " since back " << p
->second
.last_rx_back
5169 << " front " << p
->second
.last_rx_front
5170 << " (cutoff " << cutoff
<< ")" << dendl
;
5172 failure_queue
[p
->first
] = MIN(p
->second
.last_rx_back
, p
->second
.last_rx_front
);
5178 void OSD::heartbeat()
5180 dout(30) << "heartbeat" << dendl
;
5184 int n_samples
= 86400 / cct
->_conf
->osd_heartbeat_interval
;
5185 if (getloadavg(loadavgs
, 1) == 1) {
5186 logger
->set(l_osd_loadavg
, 100 * loadavgs
[0]);
5187 daily_loadavg
= (daily_loadavg
* (n_samples
- 1) + loadavgs
[0]) / n_samples
;
5188 dout(30) << "heartbeat: daily_loadavg " << daily_loadavg
<< dendl
;
5191 dout(30) << "heartbeat checking stats" << dendl
;
5194 vector
<int> hb_peers
;
5195 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
5196 p
!= heartbeat_peers
.end();
5198 hb_peers
.push_back(p
->first
);
5199 service
.update_osd_stat(hb_peers
);
5201 dout(5) << "heartbeat: " << service
.get_osd_stat() << dendl
;
5203 utime_t now
= ceph_clock_now();
5206 for (map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.begin();
5207 i
!= heartbeat_peers
.end();
5209 int peer
= i
->first
;
5210 i
->second
.last_tx
= now
;
5211 if (i
->second
.first_tx
== utime_t())
5212 i
->second
.first_tx
= now
;
5213 dout(30) << "heartbeat sending ping to osd." << peer
<< dendl
;
5214 i
->second
.con_back
->send_message(new MOSDPing(monc
->get_fsid(),
5215 service
.get_osdmap()->get_epoch(),
5216 MOSDPing::PING
, now
,
5217 cct
->_conf
->osd_heartbeat_min_size
));
5219 if (i
->second
.con_front
)
5220 i
->second
.con_front
->send_message(new MOSDPing(monc
->get_fsid(),
5221 service
.get_osdmap()->get_epoch(),
5222 MOSDPing::PING
, now
,
5223 cct
->_conf
->osd_heartbeat_min_size
));
5226 logger
->set(l_osd_hb_to
, heartbeat_peers
.size());
5228 // hmm.. am i all alone?
5229 dout(30) << "heartbeat lonely?" << dendl
;
5230 if (heartbeat_peers
.empty()) {
5231 if (now
- last_mon_heartbeat
> cct
->_conf
->osd_mon_heartbeat_interval
&& is_active()) {
5232 last_mon_heartbeat
= now
;
5233 dout(10) << "i have no heartbeat peers; checking mon for new map" << dendl
;
5234 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5238 dout(30) << "heartbeat done" << dendl
;
5241 bool OSD::heartbeat_reset(Connection
*con
)
5243 HeartbeatSession
*s
= static_cast<HeartbeatSession
*>(con
->get_priv());
5245 heartbeat_lock
.Lock();
5246 if (is_stopping()) {
5247 heartbeat_lock
.Unlock();
5251 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(s
->peer
);
5252 if (p
!= heartbeat_peers
.end() &&
5253 (p
->second
.con_back
== con
||
5254 p
->second
.con_front
== con
)) {
5255 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5256 << ", reopening" << dendl
;
5257 if (con
!= p
->second
.con_back
) {
5258 p
->second
.con_back
->mark_down();
5260 p
->second
.con_back
.reset(NULL
);
5261 if (p
->second
.con_front
&& con
!= p
->second
.con_front
) {
5262 p
->second
.con_front
->mark_down();
5264 p
->second
.con_front
.reset(NULL
);
5265 pair
<ConnectionRef
,ConnectionRef
> newcon
= service
.get_con_osd_hb(p
->second
.peer
, p
->second
.epoch
);
5267 p
->second
.con_back
= newcon
.first
.get();
5268 p
->second
.con_back
->set_priv(s
->get());
5269 if (newcon
.second
) {
5270 p
->second
.con_front
= newcon
.second
.get();
5271 p
->second
.con_front
->set_priv(s
->get());
5274 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5275 << ", raced with osdmap update, closing out peer" << dendl
;
5276 heartbeat_peers
.erase(p
);
5279 dout(10) << "heartbeat_reset closing (old) failed hb con " << con
<< dendl
;
5281 heartbeat_lock
.Unlock();
5289 // =========================================
5293 assert(osd_lock
.is_locked());
5294 dout(10) << "tick" << dendl
;
5296 if (is_active() || is_waiting_for_healthy()) {
5297 maybe_update_heartbeat_peers();
5300 if (is_waiting_for_healthy()) {
5302 } else if (is_preboot() &&
5303 waiting_for_luminous_mons
&&
5304 monc
->monmap
.get_required_features().contains_all(
5305 ceph::features::mon::FEATURE_LUMINOUS
)) {
5306 // mon upgrade finished!
5312 tick_timer
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick(this));
5315 void OSD::tick_without_osd_lock()
5317 assert(tick_timer_lock
.is_locked());
5318 dout(10) << "tick_without_osd_lock" << dendl
;
5320 logger
->set(l_osd_buf
, buffer::get_total_alloc());
5321 logger
->set(l_osd_history_alloc_bytes
, SHIFT_ROUND_UP(buffer::get_history_alloc_bytes(), 20));
5322 logger
->set(l_osd_history_alloc_num
, buffer::get_history_alloc_num());
5323 logger
->set(l_osd_cached_crc
, buffer::get_cached_crc());
5324 logger
->set(l_osd_cached_crc_adjusted
, buffer::get_cached_crc_adjusted());
5325 logger
->set(l_osd_missed_crc
, buffer::get_missed_crc());
5326 logger
->set(l_osd_pg_removing
, remove_wq
.get_remove_queue_len());
5328 // osd_lock is not being held, which means the OSD state
5329 // might change when doing the monitor report
5330 if (is_active() || is_waiting_for_healthy()) {
5331 heartbeat_lock
.Lock();
5333 heartbeat_lock
.Unlock();
5335 map_lock
.get_read();
5336 Mutex::Locker
l(mon_report_lock
);
5340 bool report
= false;
5341 utime_t now
= ceph_clock_now();
5342 pg_stat_queue_lock
.Lock();
5343 double backoff
= stats_ack_timeout
/ cct
->_conf
->osd_mon_ack_timeout
;
5344 double adjusted_min
= cct
->_conf
->osd_mon_report_interval_min
* backoff
;
5345 // note: we shouldn't adjust max because it must remain < the
5346 // mon's mon_osd_report_timeout (which defaults to 1.5x our
5348 double max
= cct
->_conf
->osd_mon_report_interval_max
;
5349 if (!outstanding_pg_stats
.empty() &&
5350 (now
- stats_ack_timeout
) > last_pg_stats_ack
) {
5351 dout(1) << __func__
<< " mon hasn't acked PGStats in "
5352 << now
- last_pg_stats_ack
5353 << " seconds, reconnecting elsewhere" << dendl
;
5355 last_pg_stats_ack
= now
; // reset clock
5356 last_pg_stats_sent
= utime_t();
5358 MAX(cct
->_conf
->osd_mon_ack_timeout
,
5359 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_factor
);
5360 outstanding_pg_stats
.clear();
5362 if (now
- last_pg_stats_sent
> max
) {
5363 osd_stat_updated
= true;
5365 } else if (service
.need_fullness_update()) {
5367 } else if ((int)outstanding_pg_stats
.size() >=
5368 cct
->_conf
->osd_mon_report_max_in_flight
) {
5369 dout(20) << __func__
<< " have max " << outstanding_pg_stats
5370 << " stats updates in flight" << dendl
;
5372 if (now
- last_mon_report
> adjusted_min
) {
5373 dout(20) << __func__
<< " stats backoff " << backoff
5374 << " adjusted_min " << adjusted_min
<< " - sending report"
5376 osd_stat_updated
= true;
5380 pg_stat_queue_lock
.Unlock();
5383 monc
->reopen_session();
5384 } else if (report
) {
5385 last_mon_report
= now
;
5387 // do any pending reports
5390 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5394 map_lock
.put_read();
5398 if (!scrub_random_backoff()) {
5401 service
.promote_throttle_recalibrate();
5402 resume_creating_pg();
5403 bool need_send_beacon
= false;
5404 const auto now
= ceph::coarse_mono_clock::now();
5406 // borrow lec lock to pretect last_sent_beacon from changing
5407 Mutex::Locker l
{min_last_epoch_clean_lock
};
5408 const auto elapsed
= now
- last_sent_beacon
;
5409 if (chrono::duration_cast
<chrono::seconds
>(elapsed
).count() >
5410 cct
->_conf
->osd_beacon_report_interval
) {
5411 need_send_beacon
= true;
5414 if (need_send_beacon
) {
5419 mgrc
.update_osd_health(get_health_metrics());
5420 service
.kick_recovery_queue();
5421 tick_timer_without_osd_lock
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick_WithoutOSDLock(this));
5424 void OSD::check_ops_in_flight()
5426 vector
<string
> warnings
;
5427 if (op_tracker
.check_ops_in_flight(warnings
)) {
5428 for (vector
<string
>::iterator i
= warnings
.begin();
5429 i
!= warnings
.end();
5437 // setomapval <pool-id> [namespace/]<obj-name> <key> <val>
5438 // rmomapkey <pool-id> [namespace/]<obj-name> <key>
5439 // setomapheader <pool-id> [namespace/]<obj-name> <header>
5440 // getomap <pool> [namespace/]<obj-name>
5441 // truncobj <pool-id> [namespace/]<obj-name> <newlen>
5442 // injectmdataerr [namespace/]<obj-name> [shardid]
5443 // injectdataerr [namespace/]<obj-name> [shardid]
5445 // set_recovery_delay [utime]
5446 void TestOpsSocketHook::test_ops(OSDService
*service
, ObjectStore
*store
,
5447 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
)
5450 //Support changing the omap on a single osd by using the Admin Socket to
5451 //directly request the osd make a change.
5452 if (command
== "setomapval" || command
== "rmomapkey" ||
5453 command
== "setomapheader" || command
== "getomap" ||
5454 command
== "truncobj" || command
== "injectmdataerr" ||
5455 command
== "injectdataerr"
5459 OSDMapRef curmap
= service
->get_osdmap();
5464 cmd_getval(service
->cct
, cmdmap
, "pool", poolstr
);
5465 pool
= curmap
->lookup_pg_pool_name(poolstr
);
5466 //If we can't find it by name then maybe id specified
5467 if (pool
< 0 && isdigit(poolstr
[0]))
5468 pool
= atoll(poolstr
.c_str());
5470 ss
<< "Invalid pool '" << poolstr
<< "''";
5474 string objname
, nspace
;
5475 cmd_getval(service
->cct
, cmdmap
, "objname", objname
);
5476 std::size_t found
= objname
.find_first_of('/');
5477 if (found
!= string::npos
) {
5478 nspace
= objname
.substr(0, found
);
5479 objname
= objname
.substr(found
+1);
5481 object_locator_t
oloc(pool
, nspace
);
5482 r
= curmap
->object_locator_to_pg(object_t(objname
), oloc
, rawpg
);
5485 ss
<< "Invalid namespace/objname";
5490 cmd_getval(service
->cct
, cmdmap
, "shardid", shardid
, int64_t(shard_id_t::NO_SHARD
));
5491 hobject_t
obj(object_t(objname
), string(""), CEPH_NOSNAP
, rawpg
.ps(), pool
, nspace
);
5492 ghobject_t
gobj(obj
, ghobject_t::NO_GEN
, shard_id_t(uint8_t(shardid
)));
5493 spg_t
pgid(curmap
->raw_pg_to_pg(rawpg
), shard_id_t(shardid
));
5494 if (curmap
->pg_is_ec(rawpg
)) {
5495 if ((command
!= "injectdataerr") && (command
!= "injectmdataerr")) {
5496 ss
<< "Must not call on ec pool, except injectdataerr or injectmdataerr";
5501 ObjectStore::Transaction t
;
5503 if (command
== "setomapval") {
5504 map
<string
, bufferlist
> newattrs
;
5507 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5508 cmd_getval(service
->cct
, cmdmap
, "val", valstr
);
5511 newattrs
[key
] = val
;
5512 t
.omap_setkeys(coll_t(pgid
), ghobject_t(obj
), newattrs
);
5513 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5515 ss
<< "error=" << r
;
5518 } else if (command
== "rmomapkey") {
5521 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5524 t
.omap_rmkeys(coll_t(pgid
), ghobject_t(obj
), keys
);
5525 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5527 ss
<< "error=" << r
;
5530 } else if (command
== "setomapheader") {
5531 bufferlist newheader
;
5534 cmd_getval(service
->cct
, cmdmap
, "header", headerstr
);
5535 newheader
.append(headerstr
);
5536 t
.omap_setheader(coll_t(pgid
), ghobject_t(obj
), newheader
);
5537 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5539 ss
<< "error=" << r
;
5542 } else if (command
== "getomap") {
5543 //Debug: Output entire omap
5545 map
<string
, bufferlist
> keyvals
;
5546 r
= store
->omap_get(coll_t(pgid
), ghobject_t(obj
), &hdrbl
, &keyvals
);
5548 ss
<< "header=" << string(hdrbl
.c_str(), hdrbl
.length());
5549 for (map
<string
, bufferlist
>::iterator it
= keyvals
.begin();
5550 it
!= keyvals
.end(); ++it
)
5551 ss
<< " key=" << (*it
).first
<< " val="
5552 << string((*it
).second
.c_str(), (*it
).second
.length());
5554 ss
<< "error=" << r
;
5556 } else if (command
== "truncobj") {
5558 cmd_getval(service
->cct
, cmdmap
, "len", trunclen
);
5559 t
.truncate(coll_t(pgid
), ghobject_t(obj
), trunclen
);
5560 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5562 ss
<< "error=" << r
;
5565 } else if (command
== "injectdataerr") {
5566 store
->inject_data_error(gobj
);
5568 } else if (command
== "injectmdataerr") {
5569 store
->inject_mdata_error(gobj
);
5574 if (command
== "set_recovery_delay") {
5576 cmd_getval(service
->cct
, cmdmap
, "utime", delay
, (int64_t)0);
5579 int r
= service
->cct
->_conf
->set_val("osd_recovery_delay_start",
5582 ss
<< "set_recovery_delay: error setting "
5583 << "osd_recovery_delay_start to '" << delay
<< "': error "
5587 service
->cct
->_conf
->apply_changes(NULL
);
5588 ss
<< "set_recovery_delay: set osd_recovery_delay_start "
5589 << "to " << service
->cct
->_conf
->osd_recovery_delay_start
;
5592 if (command
== "trigger_scrub") {
5594 OSDMapRef curmap
= service
->get_osdmap();
5598 cmd_getval(service
->cct
, cmdmap
, "pgid", pgidstr
);
5599 if (!pgid
.parse(pgidstr
.c_str())) {
5600 ss
<< "Invalid pgid specified";
5604 PG
*pg
= service
->osd
->_lookup_lock_pg(pgid
);
5605 if (pg
== nullptr) {
5606 ss
<< "Can't find pg " << pgid
;
5610 if (pg
->is_primary()) {
5611 pg
->unreg_next_scrub();
5612 const pg_pool_t
*p
= curmap
->get_pg_pool(pgid
.pool());
5613 double pool_scrub_max_interval
= 0;
5614 p
->opts
.get(pool_opts_t::SCRUB_MAX_INTERVAL
, &pool_scrub_max_interval
);
5615 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
5616 pool_scrub_max_interval
: g_conf
->osd_scrub_max_interval
;
5617 // Instead of marking must_scrub force a schedule scrub
5618 utime_t stamp
= ceph_clock_now();
5619 stamp
-= scrub_max_interval
;
5620 stamp
-= 100.0; // push back last scrub more for good measure
5621 pg
->info
.history
.last_scrub_stamp
= stamp
;
5622 pg
->reg_next_scrub();
5625 ss
<< "Not primary";
5630 if (command
== "injectfull") {
5633 OSDService::s_names state
;
5634 cmd_getval(service
->cct
, cmdmap
, "type", type
, string("full"));
5635 cmd_getval(service
->cct
, cmdmap
, "count", count
, (int64_t)-1);
5636 if (type
== "none" || count
== 0) {
5640 state
= service
->get_full_state(type
);
5641 if (state
== OSDService::s_names::INVALID
) {
5642 ss
<< "Invalid type use (none, nearfull, backfillfull, full, failsafe)";
5645 service
->set_injectfull(state
, count
);
5648 ss
<< "Internal error - command=" << command
;
5651 // =========================================
5654 ObjectStore
*store
, SnapMapper
*mapper
,
5656 ObjectStore::Sequencer
*osr
,
5657 coll_t coll
, DeletingStateRef dstate
,
5659 ThreadPool::TPHandle
&handle
)
5661 vector
<ghobject_t
> olist
;
5663 ObjectStore::Transaction t
;
5665 handle
.reset_tp_timeout();
5666 store
->collection_list(
5669 ghobject_t::get_max(),
5670 store
->get_ideal_list_max(),
5673 generic_dout(10) << __func__
<< " " << olist
<< dendl
;
5674 // default cont to true, this is safe because caller(OSD::RemoveWQ::_process())
5675 // will recheck the answer before it really goes on.
5677 for (vector
<ghobject_t
>::iterator i
= olist
.begin();
5682 OSDriver::OSTransaction
_t(osdriver
->get_transaction(&t
));
5683 int r
= mapper
->remove_oid(i
->hobj
, &_t
);
5684 if (r
!= 0 && r
!= -ENOENT
) {
5688 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
5690 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5691 cont
= dstate
->pause_clearing();
5692 handle
.suspend_tp_timeout();
5694 handle
.reset_tp_timeout();
5696 cont
= dstate
->resume_clearing();
5699 t
= ObjectStore::Transaction();
5705 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5706 cont
= dstate
->pause_clearing();
5707 handle
.suspend_tp_timeout();
5709 handle
.reset_tp_timeout();
5711 cont
= dstate
->resume_clearing();
5713 // whether there are more objects to remove in the collection
5714 *finished
= next
.is_max();
5718 void OSD::RemoveWQ::_process(
5719 pair
<PGRef
, DeletingStateRef
> item
,
5720 ThreadPool::TPHandle
&handle
)
5723 PGRef
pg(item
.first
);
5724 SnapMapper
&mapper
= pg
->snap_mapper
;
5725 OSDriver
&driver
= pg
->osdriver
;
5726 coll_t coll
= coll_t(pg
->info
.pgid
);
5728 bool finished
= false;
5730 if (!item
.second
->start_or_resume_clearing())
5733 bool cont
= remove_dir(
5734 pg
->cct
, store
, &mapper
, &driver
, pg
->osr
.get(), coll
, item
.second
,
5739 if (item
.second
->pause_clearing())
5744 if (!item
.second
->start_deleting())
5747 ObjectStore::Transaction t
;
5748 PGLog::clear_info_log(pg
->info
.pgid
, &t
);
5750 if (cct
->_conf
->osd_inject_failure_on_pg_removal
) {
5751 generic_derr
<< "osd_inject_failure_on_pg_removal" << dendl
;
5754 t
.remove_collection(coll
);
5756 // We need the sequencer to stick around until the op is complete
5757 store
->queue_transaction(
5762 0, // onreadable sync
5763 new ContainerContext
<PGRef
>(pg
),
5766 item
.second
->finish_deleting();
5768 // =========================================
5770 void OSD::ms_handle_connect(Connection
*con
)
5772 dout(10) << __func__
<< " con " << con
<< dendl
;
5773 if (con
->get_peer_type() == CEPH_ENTITY_TYPE_MON
) {
5774 Mutex::Locker
l(osd_lock
);
5777 dout(10) << __func__
<< " on mon" << dendl
;
5781 } else if (is_booting()) {
5782 _send_boot(); // resend boot message
5784 map_lock
.get_read();
5785 Mutex::Locker
l2(mon_report_lock
);
5787 utime_t now
= ceph_clock_now();
5788 last_mon_report
= now
;
5790 // resend everything, it's a new session
5793 service
.requeue_pg_temp();
5794 service
.send_pg_temp();
5797 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5801 map_lock
.put_read();
5803 send_beacon(ceph::coarse_mono_clock::now());
5807 // full map requests may happen while active or pre-boot
5808 if (requested_full_first
) {
5809 rerequest_full_maps();
5814 void OSD::ms_handle_fast_connect(Connection
*con
)
5816 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5817 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5818 Session
*s
= static_cast<Session
*>(con
->get_priv());
5820 s
= new Session(cct
);
5821 con
->set_priv(s
->get());
5823 dout(10) << " new session (outgoing) " << s
<< " con=" << s
->con
5824 << " addr=" << s
->con
->get_peer_addr() << dendl
;
5825 // we don't connect to clients
5826 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5827 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5833 void OSD::ms_handle_fast_accept(Connection
*con
)
5835 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5836 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5837 Session
*s
= static_cast<Session
*>(con
->get_priv());
5839 s
= new Session(cct
);
5840 con
->set_priv(s
->get());
5842 dout(10) << "new session (incoming)" << s
<< " con=" << con
5843 << " addr=" << con
->get_peer_addr()
5844 << " must have raced with connect" << dendl
;
5845 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5846 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5852 bool OSD::ms_handle_reset(Connection
*con
)
5854 Session
*session
= static_cast<Session
*>(con
->get_priv());
5855 dout(2) << "ms_handle_reset con " << con
<< " session " << session
<< dendl
;
5858 session
->wstate
.reset(con
);
5859 session
->con
.reset(NULL
); // break con <-> session ref cycle
5860 // note that we break session->con *before* the session_handle_reset
5861 // cleanup below. this avoids a race between us and
5862 // PG::add_backoff, Session::check_backoff, etc.
5863 session_handle_reset(session
);
5868 bool OSD::ms_handle_refused(Connection
*con
)
5870 if (!cct
->_conf
->osd_fast_fail_on_connection_refused
)
5873 Session
*session
= static_cast<Session
*>(con
->get_priv());
5874 dout(2) << "ms_handle_refused con " << con
<< " session " << session
<< dendl
;
5877 int type
= con
->get_peer_type();
5878 // handle only OSD failures here
5879 if (monc
&& (type
== CEPH_ENTITY_TYPE_OSD
)) {
5880 OSDMapRef osdmap
= get_osdmap();
5882 int id
= osdmap
->identify_osd_on_all_channels(con
->get_peer_addr());
5883 if (id
>= 0 && osdmap
->is_up(id
)) {
5884 // I'm cheating mon heartbeat grace logic, because we know it's not going
5885 // to respawn alone. +1 so we won't hit any boundary case.
5886 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(),
5887 osdmap
->get_inst(id
),
5888 cct
->_conf
->osd_heartbeat_grace
+ 1,
5889 osdmap
->get_epoch(),
5890 MOSDFailure::FLAG_IMMEDIATE
| MOSDFailure::FLAG_FAILED
5899 struct C_OSD_GetVersion
: public Context
{
5901 uint64_t oldest
, newest
;
5902 explicit C_OSD_GetVersion(OSD
*o
) : osd(o
), oldest(0), newest(0) {}
5903 void finish(int r
) override
{
5905 osd
->_got_mon_epochs(oldest
, newest
);
5909 void OSD::start_boot()
5911 if (!_is_healthy()) {
5912 // if we are not healthy, do not mark ourselves up (yet)
5913 dout(1) << "not healthy; waiting to boot" << dendl
;
5914 if (!is_waiting_for_healthy())
5915 start_waiting_for_healthy();
5916 // send pings sooner rather than later
5920 dout(1) << __func__
<< dendl
;
5921 set_state(STATE_PREBOOT
);
5922 waiting_for_luminous_mons
= false;
5923 dout(10) << "start_boot - have maps " << superblock
.oldest_map
5924 << ".." << superblock
.newest_map
<< dendl
;
5925 C_OSD_GetVersion
*c
= new C_OSD_GetVersion(this);
5926 monc
->get_version("osdmap", &c
->newest
, &c
->oldest
, c
);
5929 void OSD::_got_mon_epochs(epoch_t oldest
, epoch_t newest
)
5931 Mutex::Locker
l(osd_lock
);
5933 _preboot(oldest
, newest
);
5937 void OSD::_preboot(epoch_t oldest
, epoch_t newest
)
5939 assert(is_preboot());
5940 dout(10) << __func__
<< " _preboot mon has osdmaps "
5941 << oldest
<< ".." << newest
<< dendl
;
5943 // ensure our local fullness awareness is accurate
5946 // if our map within recent history, try to add ourselves to the osdmap.
5947 if (osdmap
->get_epoch() == 0) {
5948 derr
<< "waiting for initial osdmap" << dendl
;
5949 } else if (osdmap
->is_destroyed(whoami
)) {
5950 derr
<< "osdmap says I am destroyed" << dendl
;
5951 // provide a small margin so we don't livelock seeing if we
5952 // un-destroyed ourselves.
5953 if (osdmap
->get_epoch() > newest
- 1) {
5956 } else if (osdmap
->test_flag(CEPH_OSDMAP_NOUP
) || osdmap
->is_noup(whoami
)) {
5957 derr
<< "osdmap NOUP flag is set, waiting for it to clear" << dendl
;
5958 } else if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
)) {
5959 derr
<< "osdmap SORTBITWISE OSDMap flag is NOT set; please set it"
5961 } else if (osdmap
->require_osd_release
< CEPH_RELEASE_JEWEL
) {
5962 derr
<< "osdmap REQUIRE_JEWEL OSDMap flag is NOT set; please set it"
5964 } else if (!monc
->monmap
.get_required_features().contains_all(
5965 ceph::features::mon::FEATURE_LUMINOUS
)) {
5966 derr
<< "monmap REQUIRE_LUMINOUS is NOT set; must upgrade all monitors to "
5967 << "Luminous or later before Luminous OSDs will boot" << dendl
;
5968 waiting_for_luminous_mons
= true;
5969 } else if (service
.need_fullness_update()) {
5970 derr
<< "osdmap fullness state needs update" << dendl
;
5972 } else if (osdmap
->get_epoch() >= oldest
- 1 &&
5973 osdmap
->get_epoch() + cct
->_conf
->osd_map_message_max
> newest
) {
5978 // get all the latest maps
5979 if (osdmap
->get_epoch() + 1 >= oldest
)
5980 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5982 osdmap_subscribe(oldest
- 1, true);
5985 void OSD::send_full_update()
5987 if (!service
.need_fullness_update())
5990 if (service
.is_full()) {
5991 state
= CEPH_OSD_FULL
;
5992 } else if (service
.is_backfillfull()) {
5993 state
= CEPH_OSD_BACKFILLFULL
;
5994 } else if (service
.is_nearfull()) {
5995 state
= CEPH_OSD_NEARFULL
;
5998 OSDMap::calc_state_set(state
, s
);
5999 dout(10) << __func__
<< " want state " << s
<< dendl
;
6000 monc
->send_mon_message(new MOSDFull(osdmap
->get_epoch(), state
));
6003 void OSD::start_waiting_for_healthy()
6005 dout(1) << "start_waiting_for_healthy" << dendl
;
6006 set_state(STATE_WAITING_FOR_HEALTHY
);
6007 last_heartbeat_resample
= utime_t();
6009 // subscribe to osdmap updates, in case our peers really are known to be dead
6010 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
6013 bool OSD::_is_healthy()
6015 if (!cct
->get_heartbeat_map()->is_healthy()) {
6016 dout(1) << "is_healthy false -- internal heartbeat failed" << dendl
;
6020 if (is_waiting_for_healthy()) {
6021 Mutex::Locker
l(heartbeat_lock
);
6022 utime_t cutoff
= ceph_clock_now();
6023 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
6024 int num
= 0, up
= 0;
6025 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
6026 p
!= heartbeat_peers
.end();
6028 if (p
->second
.is_healthy(cutoff
))
6032 if ((float)up
< (float)num
* cct
->_conf
->osd_heartbeat_min_healthy_ratio
) {
6033 dout(1) << "is_healthy false -- only " << up
<< "/" << num
<< " up peers (less than "
6034 << int(cct
->_conf
->osd_heartbeat_min_healthy_ratio
* 100.0) << "%)" << dendl
;
6042 void OSD::_send_boot()
6044 dout(10) << "_send_boot" << dendl
;
6045 entity_addr_t cluster_addr
= cluster_messenger
->get_myaddr();
6046 Connection
*local_connection
= cluster_messenger
->get_loopback_connection().get();
6047 if (cluster_addr
.is_blank_ip()) {
6048 int port
= cluster_addr
.get_port();
6049 cluster_addr
= client_messenger
->get_myaddr();
6050 cluster_addr
.set_port(port
);
6051 cluster_messenger
->set_addr_unknowns(cluster_addr
);
6052 dout(10) << " assuming cluster_addr ip matches client_addr" << dendl
;
6054 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
6058 cluster_messenger
->ms_deliver_handle_fast_connect(local_connection
);
6061 entity_addr_t hb_back_addr
= hb_back_server_messenger
->get_myaddr();
6062 local_connection
= hb_back_server_messenger
->get_loopback_connection().get();
6063 if (hb_back_addr
.is_blank_ip()) {
6064 int port
= hb_back_addr
.get_port();
6065 hb_back_addr
= cluster_addr
;
6066 hb_back_addr
.set_port(port
);
6067 hb_back_server_messenger
->set_addr_unknowns(hb_back_addr
);
6068 dout(10) << " assuming hb_back_addr ip matches cluster_addr" << dendl
;
6070 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
6074 hb_back_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
6077 entity_addr_t hb_front_addr
= hb_front_server_messenger
->get_myaddr();
6078 local_connection
= hb_front_server_messenger
->get_loopback_connection().get();
6079 if (hb_front_addr
.is_blank_ip()) {
6080 int port
= hb_front_addr
.get_port();
6081 hb_front_addr
= client_messenger
->get_myaddr();
6082 hb_front_addr
.set_port(port
);
6083 hb_front_server_messenger
->set_addr_unknowns(hb_front_addr
);
6084 dout(10) << " assuming hb_front_addr ip matches client_addr" << dendl
;
6086 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
6090 hb_front_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
6093 MOSDBoot
*mboot
= new MOSDBoot(superblock
, get_osdmap_epoch(), service
.get_boot_epoch(),
6094 hb_back_addr
, hb_front_addr
, cluster_addr
,
6096 dout(10) << " client_addr " << client_messenger
->get_myaddr()
6097 << ", cluster_addr " << cluster_addr
6098 << ", hb_back_addr " << hb_back_addr
6099 << ", hb_front_addr " << hb_front_addr
6101 _collect_metadata(&mboot
->metadata
);
6102 monc
->send_mon_message(mboot
);
6103 set_state(STATE_BOOTING
);
6106 void OSD::_collect_metadata(map
<string
,string
> *pm
)
6109 (*pm
)["osd_data"] = dev_path
;
6110 if (store
->get_type() == "filestore") {
6111 // not applicable for bluestore
6112 (*pm
)["osd_journal"] = journal_path
;
6114 (*pm
)["front_addr"] = stringify(client_messenger
->get_myaddr());
6115 (*pm
)["back_addr"] = stringify(cluster_messenger
->get_myaddr());
6116 (*pm
)["hb_front_addr"] = stringify(hb_front_server_messenger
->get_myaddr());
6117 (*pm
)["hb_back_addr"] = stringify(hb_back_server_messenger
->get_myaddr());
6120 (*pm
)["osd_objectstore"] = store
->get_type();
6121 (*pm
)["rotational"] = store_is_rotational
? "1" : "0";
6122 (*pm
)["journal_rotational"] = journal_is_rotational
? "1" : "0";
6123 (*pm
)["default_device_class"] = store
->get_default_device_class();
6124 store
->collect_metadata(pm
);
6126 collect_sys_info(pm
, cct
);
6128 std::string front_iface
, back_iface
;
6131 CEPH_PICK_ADDRESS_PUBLIC | CEPH_PICK_ADDRESS_CLUSTER,
6132 &front_iface, &back_iface);
6134 (*pm
)["front_iface"] = pick_iface(cct
,
6135 client_messenger
->get_myaddr().get_sockaddr_storage());
6136 (*pm
)["back_iface"] = pick_iface(cct
,
6137 cluster_messenger
->get_myaddr().get_sockaddr_storage());
6139 dout(10) << __func__
<< " " << *pm
<< dendl
;
6142 void OSD::queue_want_up_thru(epoch_t want
)
6144 map_lock
.get_read();
6145 epoch_t cur
= osdmap
->get_up_thru(whoami
);
6146 Mutex::Locker
l(mon_report_lock
);
6147 if (want
> up_thru_wanted
) {
6148 dout(10) << "queue_want_up_thru now " << want
<< " (was " << up_thru_wanted
<< ")"
6149 << ", currently " << cur
6151 up_thru_wanted
= want
;
6154 dout(10) << "queue_want_up_thru want " << want
<< " <= queued " << up_thru_wanted
6155 << ", currently " << cur
6158 map_lock
.put_read();
6161 void OSD::send_alive()
6163 assert(mon_report_lock
.is_locked());
6164 if (!osdmap
->exists(whoami
))
6166 epoch_t up_thru
= osdmap
->get_up_thru(whoami
);
6167 dout(10) << "send_alive up_thru currently " << up_thru
<< " want " << up_thru_wanted
<< dendl
;
6168 if (up_thru_wanted
> up_thru
) {
6169 dout(10) << "send_alive want " << up_thru_wanted
<< dendl
;
6170 monc
->send_mon_message(new MOSDAlive(osdmap
->get_epoch(), up_thru_wanted
));
6174 void OSD::request_full_map(epoch_t first
, epoch_t last
)
6176 dout(10) << __func__
<< " " << first
<< ".." << last
6177 << ", previously requested "
6178 << requested_full_first
<< ".." << requested_full_last
<< dendl
;
6179 assert(osd_lock
.is_locked());
6180 assert(first
> 0 && last
> 0);
6181 assert(first
<= last
);
6182 assert(first
>= requested_full_first
); // we shouldn't ever ask for older maps
6183 if (requested_full_first
== 0) {
6185 requested_full_first
= first
;
6186 requested_full_last
= last
;
6187 } else if (last
<= requested_full_last
) {
6191 // additional request
6192 first
= requested_full_last
+ 1;
6193 requested_full_last
= last
;
6195 MMonGetOSDMap
*req
= new MMonGetOSDMap
;
6196 req
->request_full(first
, last
);
6197 monc
->send_mon_message(req
);
6200 void OSD::got_full_map(epoch_t e
)
6202 assert(requested_full_first
<= requested_full_last
);
6203 assert(osd_lock
.is_locked());
6204 if (requested_full_first
== 0) {
6205 dout(20) << __func__
<< " " << e
<< ", nothing requested" << dendl
;
6208 if (e
< requested_full_first
) {
6209 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6210 << ".." << requested_full_last
6211 << ", ignoring" << dendl
;
6214 if (e
>= requested_full_last
) {
6215 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6216 << ".." << requested_full_last
<< ", resetting" << dendl
;
6217 requested_full_first
= requested_full_last
= 0;
6221 requested_full_first
= e
+ 1;
6223 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6224 << ".." << requested_full_last
6225 << ", still need more" << dendl
;
6228 void OSD::requeue_failures()
6230 Mutex::Locker
l(heartbeat_lock
);
6231 unsigned old_queue
= failure_queue
.size();
6232 unsigned old_pending
= failure_pending
.size();
6233 for (map
<int,pair
<utime_t
,entity_inst_t
> >::iterator p
=
6234 failure_pending
.begin();
6235 p
!= failure_pending
.end(); ) {
6236 failure_queue
[p
->first
] = p
->second
.first
;
6237 failure_pending
.erase(p
++);
6239 dout(10) << __func__
<< " " << old_queue
<< " + " << old_pending
<< " -> "
6240 << failure_queue
.size() << dendl
;
6243 void OSD::send_failures()
6245 assert(map_lock
.is_locked());
6246 assert(mon_report_lock
.is_locked());
6247 Mutex::Locker
l(heartbeat_lock
);
6248 utime_t now
= ceph_clock_now();
6249 while (!failure_queue
.empty()) {
6250 int osd
= failure_queue
.begin()->first
;
6251 if (!failure_pending
.count(osd
)) {
6252 entity_inst_t i
= osdmap
->get_inst(osd
);
6253 int failed_for
= (int)(double)(now
- failure_queue
.begin()->second
);
6254 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(), i
, failed_for
,
6255 osdmap
->get_epoch()));
6256 failure_pending
[osd
] = make_pair(failure_queue
.begin()->second
, i
);
6258 failure_queue
.erase(osd
);
6262 void OSD::send_still_alive(epoch_t epoch
, const entity_inst_t
&i
)
6264 MOSDFailure
*m
= new MOSDFailure(monc
->get_fsid(), i
, 0, epoch
, MOSDFailure::FLAG_ALIVE
);
6265 monc
->send_mon_message(m
);
6268 void OSD::send_pg_stats(const utime_t
&now
)
6270 assert(map_lock
.is_locked());
6271 assert(osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
);
6272 dout(20) << "send_pg_stats" << dendl
;
6274 osd_stat_t cur_stat
= service
.get_osd_stat();
6276 cur_stat
.os_perf_stat
= store
->get_cur_stats();
6278 pg_stat_queue_lock
.Lock();
6280 if (osd_stat_updated
|| !pg_stat_queue
.empty()) {
6281 last_pg_stats_sent
= now
;
6282 osd_stat_updated
= false;
6284 dout(10) << "send_pg_stats - " << pg_stat_queue
.size() << " pgs updated" << dendl
;
6286 utime_t
had_for(now
);
6287 had_for
-= had_map_since
;
6289 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
6291 uint64_t tid
= ++pg_stat_tid
;
6293 m
->osd_stat
= cur_stat
;
6295 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6299 if (!pg
->is_primary()) { // we hold map_lock; role is stable.
6300 pg
->stat_queue_item
.remove_myself();
6301 pg
->put("pg_stat_queue");
6304 pg
->pg_stats_publish_lock
.Lock();
6305 if (pg
->pg_stats_publish_valid
) {
6306 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
6307 dout(25) << " sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6308 << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6310 dout(25) << " NOT sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6311 << pg
->pg_stats_publish
.reported_seq
<< ", not valid" << dendl
;
6313 pg
->pg_stats_publish_lock
.Unlock();
6316 if (last_pg_stats_ack
== utime_t() || !outstanding_pg_stats
.empty()) {
6317 last_pg_stats_ack
= ceph_clock_now();
6319 outstanding_pg_stats
.insert(tid
);
6320 dout(20) << __func__
<< " updates pending: " << outstanding_pg_stats
<< dendl
;
6322 monc
->send_mon_message(m
);
6325 pg_stat_queue_lock
.Unlock();
6328 void OSD::handle_pg_stats_ack(MPGStatsAck
*ack
)
6330 dout(10) << "handle_pg_stats_ack " << dendl
;
6332 if (!require_mon_peer(ack
)) {
6337 // NOTE: we may get replies from a previous mon even while
6338 // outstanding_pg_stats is empty if reconnecting races with replies
6341 pg_stat_queue_lock
.Lock();
6343 last_pg_stats_ack
= ceph_clock_now();
6345 // decay timeout slowly (analogous to TCP)
6347 MAX(cct
->_conf
->osd_mon_ack_timeout
,
6348 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_decay
);
6349 dout(20) << __func__
<< " timeout now " << stats_ack_timeout
<< dendl
;
6351 if (ack
->get_tid() > pg_stat_tid_flushed
) {
6352 pg_stat_tid_flushed
= ack
->get_tid();
6353 pg_stat_queue_cond
.Signal();
6356 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6362 auto acked
= ack
->pg_stat
.find(pg
->info
.pgid
.pgid
);
6363 if (acked
!= ack
->pg_stat
.end()) {
6364 pg
->pg_stats_publish_lock
.Lock();
6365 if (acked
->second
.first
== pg
->pg_stats_publish
.reported_seq
&&
6366 acked
->second
.second
== pg
->pg_stats_publish
.reported_epoch
) {
6367 dout(25) << " ack on " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6368 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6369 pg
->stat_queue_item
.remove_myself();
6370 pg
->put("pg_stat_queue");
6372 dout(25) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6373 << ":" << pg
->pg_stats_publish
.reported_seq
<< " > acked "
6374 << acked
->second
<< dendl
;
6376 pg
->pg_stats_publish_lock
.Unlock();
6378 dout(30) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6379 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6383 outstanding_pg_stats
.erase(ack
->get_tid());
6384 dout(20) << __func__
<< " still pending: " << outstanding_pg_stats
<< dendl
;
6386 pg_stat_queue_lock
.Unlock();
6391 void OSD::flush_pg_stats()
6393 dout(10) << "flush_pg_stats" << dendl
;
6395 utime_t now
= ceph_clock_now();
6396 map_lock
.get_read();
6397 mon_report_lock
.Lock();
6399 mon_report_lock
.Unlock();
6400 map_lock
.put_read();
6403 pg_stat_queue_lock
.Lock();
6404 uint64_t tid
= pg_stat_tid
;
6405 dout(10) << "flush_pg_stats waiting for stats tid " << tid
<< " to flush" << dendl
;
6406 while (tid
> pg_stat_tid_flushed
)
6407 pg_stat_queue_cond
.Wait(pg_stat_queue_lock
);
6408 dout(10) << "flush_pg_stats finished waiting for stats tid " << tid
<< " to flush" << dendl
;
6409 pg_stat_queue_lock
.Unlock();
6414 void OSD::send_beacon(const ceph::coarse_mono_clock::time_point
& now
)
6416 const auto& monmap
= monc
->monmap
;
6417 // send beacon to mon even if we are just connected, and the monmap is not
6418 // initialized yet by then.
6419 if (monmap
.epoch
> 0 &&
6420 monmap
.get_required_features().contains_all(
6421 ceph::features::mon::FEATURE_LUMINOUS
)) {
6422 dout(20) << __func__
<< " sending" << dendl
;
6423 MOSDBeacon
* beacon
= nullptr;
6425 Mutex::Locker l
{min_last_epoch_clean_lock
};
6426 beacon
= new MOSDBeacon(osdmap
->get_epoch(), min_last_epoch_clean
);
6427 std::swap(beacon
->pgs
, min_last_epoch_clean_pgs
);
6428 last_sent_beacon
= now
;
6430 monc
->send_mon_message(beacon
);
6432 dout(20) << __func__
<< " not sending" << dendl
;
6436 void OSD::handle_command(MMonCommand
*m
)
6438 if (!require_mon_peer(m
)) {
6443 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), NULL
);
6444 command_wq
.queue(c
);
6448 void OSD::handle_command(MCommand
*m
)
6450 ConnectionRef con
= m
->get_connection();
6451 Session
*session
= static_cast<Session
*>(con
->get_priv());
6453 con
->send_message(new MCommandReply(m
, -EPERM
));
6458 OSDCap
& caps
= session
->caps
;
6461 if (!caps
.allow_all() || m
->get_source().is_mon()) {
6462 con
->send_message(new MCommandReply(m
, -EPERM
));
6467 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), con
.get());
6468 command_wq
.queue(c
);
6478 string availability
;
6479 } osd_commands
[] = {
6481 #define COMMAND(parsesig, helptext, module, perm, availability) \
6482 {parsesig, helptext, module, perm, availability},
6484 // yes, these are really pg commands, but there's a limit to how
6485 // much work it's worth. The OSD returns all of them. Make this
6486 // form (pg <pgid> <cmd>) valid only for the cli.
6487 // Rest uses "tell <pgid> <cmd>"
6490 "name=pgid,type=CephPgid " \
6491 "name=cmd,type=CephChoices,strings=query", \
6492 "show details of a specific pg", "osd", "r", "cli")
6494 "name=pgid,type=CephPgid " \
6495 "name=cmd,type=CephChoices,strings=mark_unfound_lost " \
6496 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6497 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6500 "name=pgid,type=CephPgid " \
6501 "name=cmd,type=CephChoices,strings=list_missing " \
6502 "name=offset,type=CephString,req=false",
6503 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6506 // new form: tell <pgid> <cmd> for both cli and rest
6509 "show details of a specific pg", "osd", "r", "cli,rest")
6510 COMMAND("mark_unfound_lost " \
6511 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6512 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6513 "osd", "rw", "cli,rest")
6514 COMMAND("list_missing " \
6515 "name=offset,type=CephString,req=false",
6516 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6517 "osd", "r", "cli,rest")
6518 COMMAND("perf histogram dump "
6519 "name=logger,type=CephString,req=false "
6520 "name=counter,type=CephString,req=false",
6521 "Get histogram data",
6522 "osd", "r", "cli,rest")
6524 // tell <osd.n> commands. Validation of osd.n must be special-cased in client
6525 COMMAND("version", "report version of OSD", "osd", "r", "cli,rest")
6526 COMMAND("get_command_descriptions", "list commands descriptions", "osd", "r", "cli,rest")
6527 COMMAND("injectargs " \
6528 "name=injected_args,type=CephString,n=N",
6529 "inject configuration arguments into running OSD",
6530 "osd", "rw", "cli,rest")
6531 COMMAND("config set " \
6532 "name=key,type=CephString name=value,type=CephString",
6533 "Set a configuration option at runtime (not persistent)",
6534 "osd", "rw", "cli,rest")
6535 COMMAND("cluster_log " \
6536 "name=level,type=CephChoices,strings=error,warning,info,debug " \
6537 "name=message,type=CephString,n=N",
6538 "log a message to the cluster log",
6539 "osd", "rw", "cli,rest")
6541 "name=count,type=CephInt,req=false " \
6542 "name=size,type=CephInt,req=false " \
6543 "name=object_size,type=CephInt,req=false " \
6544 "name=object_num,type=CephInt,req=false ", \
6545 "OSD benchmark: write <count> <size>-byte objects, " \
6546 "(default 1G size 4MB). Results in log.",
6547 "osd", "rw", "cli,rest")
6548 COMMAND("flush_pg_stats", "flush pg stats", "osd", "rw", "cli,rest")
6550 "name=heapcmd,type=CephChoices,strings=dump|start_profiler|stop_profiler|release|stats", \
6551 "show heap usage info (available only if compiled with tcmalloc)", \
6552 "osd", "rw", "cli,rest")
6553 COMMAND("debug dump_missing " \
6554 "name=filename,type=CephFilepath",
6555 "dump missing objects to a named file", "osd", "r", "cli,rest")
6556 COMMAND("debug kick_recovery_wq " \
6557 "name=delay,type=CephInt,range=0",
6558 "set osd_recovery_delay_start to <val>", "osd", "rw", "cli,rest")
6559 COMMAND("cpu_profiler " \
6560 "name=arg,type=CephChoices,strings=status|flush",
6561 "run cpu profiling on daemon", "osd", "rw", "cli,rest")
6562 COMMAND("dump_pg_recovery_stats", "dump pg recovery statistics",
6563 "osd", "r", "cli,rest")
6564 COMMAND("reset_pg_recovery_stats", "reset pg recovery statistics",
6565 "osd", "rw", "cli,rest")
6567 "compact object store's omap. "
6568 "WARNING: Compaction probably slows your requests",
6569 "osd", "rw", "cli,rest")
6572 void OSD::do_command(Connection
*con
, ceph_tid_t tid
, vector
<string
>& cmd
, bufferlist
& data
)
6575 stringstream ss
, ds
;
6579 dout(20) << "do_command tid " << tid
<< " " << cmd
<< dendl
;
6581 map
<string
, cmd_vartype
> cmdmap
;
6585 boost::scoped_ptr
<Formatter
> f
;
6588 ss
<< "no command given";
6592 if (!cmdmap_from_json(cmd
, &cmdmap
, ss
)) {
6597 cmd_getval(cct
, cmdmap
, "prefix", prefix
);
6599 if (prefix
== "get_command_descriptions") {
6601 JSONFormatter
*f
= new JSONFormatter();
6602 f
->open_object_section("command_descriptions");
6603 for (OSDCommand
*cp
= osd_commands
;
6604 cp
< &osd_commands
[ARRAY_SIZE(osd_commands
)]; cp
++) {
6606 ostringstream secname
;
6607 secname
<< "cmd" << setfill('0') << std::setw(3) << cmdnum
;
6608 dump_cmddesc_to_json(f
, secname
.str(), cp
->cmdstring
, cp
->helpstring
,
6609 cp
->module
, cp
->perm
, cp
->availability
, 0);
6612 f
->close_section(); // command_descriptions
6619 cmd_getval(cct
, cmdmap
, "format", format
);
6620 f
.reset(Formatter::create(format
));
6622 if (prefix
== "version") {
6624 f
->open_object_section("version");
6625 f
->dump_string("version", pretty_version_to_str());
6629 ds
<< pretty_version_to_str();
6633 else if (prefix
== "injectargs") {
6634 vector
<string
> argsvec
;
6635 cmd_getval(cct
, cmdmap
, "injected_args", argsvec
);
6637 if (argsvec
.empty()) {
6639 ss
<< "ignoring empty injectargs";
6642 string args
= argsvec
.front();
6643 for (vector
<string
>::iterator a
= ++argsvec
.begin(); a
!= argsvec
.end(); ++a
)
6646 r
= cct
->_conf
->injectargs(args
, &ss
);
6649 else if (prefix
== "config set") {
6652 cmd_getval(cct
, cmdmap
, "key", key
);
6653 cmd_getval(cct
, cmdmap
, "value", val
);
6655 r
= cct
->_conf
->set_val(key
, val
, true, &ss
);
6657 cct
->_conf
->apply_changes(nullptr);
6661 else if (prefix
== "cluster_log") {
6663 cmd_getval(cct
, cmdmap
, "message", msg
);
6666 ss
<< "ignoring empty log message";
6669 string message
= msg
.front();
6670 for (vector
<string
>::iterator a
= ++msg
.begin(); a
!= msg
.end(); ++a
)
6671 message
+= " " + *a
;
6673 cmd_getval(cct
, cmdmap
, "level", lvl
);
6674 clog_type level
= string_to_clog_type(lvl
);
6677 ss
<< "unknown level '" << lvl
<< "'";
6680 clog
->do_log(level
, message
);
6683 // either 'pg <pgid> <command>' or
6684 // 'tell <pgid>' (which comes in without any of that prefix)?
6686 else if (prefix
== "pg" ||
6687 prefix
== "query" ||
6688 prefix
== "mark_unfound_lost" ||
6689 prefix
== "list_missing"
6693 if (!cmd_getval(cct
, cmdmap
, "pgid", pgidstr
)) {
6694 ss
<< "no pgid specified";
6696 } else if (!pgid
.parse(pgidstr
.c_str())) {
6697 ss
<< "couldn't parse pgid '" << pgidstr
<< "'";
6702 if (osdmap
->get_primary_shard(pgid
, &pcand
) &&
6703 (pg
= _lookup_lock_pg(pcand
))) {
6704 if (pg
->is_primary()) {
6705 // simulate pg <pgid> cmd= for pg->do-command
6707 cmd_putval(cct
, cmdmap
, "cmd", prefix
);
6708 r
= pg
->do_command(cmdmap
, ss
, data
, odata
, con
, tid
);
6711 // don't reply, pg will do so async
6715 ss
<< "not primary for pgid " << pgid
;
6717 // send them the latest diff to ensure they realize the mapping
6719 service
.send_incremental_map(osdmap
->get_epoch() - 1, con
, osdmap
);
6721 // do not reply; they will get newer maps and realize they
6728 ss
<< "i don't have pgid " << pgid
;
6734 else if (prefix
== "bench") {
6737 int64_t osize
, onum
;
6738 // default count 1G, size 4MB
6739 cmd_getval(cct
, cmdmap
, "count", count
, (int64_t)1 << 30);
6740 cmd_getval(cct
, cmdmap
, "size", bsize
, (int64_t)4 << 20);
6741 cmd_getval(cct
, cmdmap
, "object_size", osize
, (int64_t)0);
6742 cmd_getval(cct
, cmdmap
, "object_num", onum
, (int64_t)0);
6744 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
6745 ObjectStore::Sequencer
>("bench"));
6747 uint32_t duration
= cct
->_conf
->osd_bench_duration
;
6749 if (bsize
> (int64_t) cct
->_conf
->osd_bench_max_block_size
) {
6750 // let us limit the block size because the next checks rely on it
6751 // having a sane value. If we allow any block size to be set things
6752 // can still go sideways.
6753 ss
<< "block 'size' values are capped at "
6754 << prettybyte_t(cct
->_conf
->osd_bench_max_block_size
) << ". If you wish to use"
6755 << " a higher value, please adjust 'osd_bench_max_block_size'";
6758 } else if (bsize
< (int64_t) (1 << 20)) {
6759 // entering the realm of small block sizes.
6760 // limit the count to a sane value, assuming a configurable amount of
6761 // IOPS and duration, so that the OSD doesn't get hung up on this,
6762 // preventing timeouts from going off
6764 bsize
* duration
* cct
->_conf
->osd_bench_small_size_max_iops
;
6765 if (count
> max_count
) {
6766 ss
<< "'count' values greater than " << max_count
6767 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6768 << cct
->_conf
->osd_bench_small_size_max_iops
<< " IOPS,"
6769 << " for " << duration
<< " seconds,"
6770 << " can cause ill effects on osd. "
6771 << " Please adjust 'osd_bench_small_size_max_iops' with a higher"
6772 << " value if you wish to use a higher 'count'.";
6777 // 1MB block sizes are big enough so that we get more stuff done.
6778 // However, to avoid the osd from getting hung on this and having
6779 // timers being triggered, we are going to limit the count assuming
6780 // a configurable throughput and duration.
6781 // NOTE: max_count is the total amount of bytes that we believe we
6782 // will be able to write during 'duration' for the given
6783 // throughput. The block size hardly impacts this unless it's
6784 // way too big. Given we already check how big the block size
6785 // is, it's safe to assume everything will check out.
6787 cct
->_conf
->osd_bench_large_size_max_throughput
* duration
;
6788 if (count
> max_count
) {
6789 ss
<< "'count' values greater than " << max_count
6790 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6791 << prettybyte_t(cct
->_conf
->osd_bench_large_size_max_throughput
) << "/s,"
6792 << " for " << duration
<< " seconds,"
6793 << " can cause ill effects on osd. "
6794 << " Please adjust 'osd_bench_large_size_max_throughput'"
6795 << " with a higher value if you wish to use a higher 'count'.";
6801 if (osize
&& bsize
> osize
)
6804 dout(1) << " bench count " << count
6805 << " bsize " << prettybyte_t(bsize
) << dendl
;
6807 ObjectStore::Transaction cleanupt
;
6809 if (osize
&& onum
) {
6811 bufferptr
bp(osize
);
6813 bl
.push_back(std::move(bp
));
6814 bl
.rebuild_page_aligned();
6815 for (int i
=0; i
<onum
; ++i
) {
6817 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", i
);
6819 hobject_t
soid(sobject_t(oid
, 0));
6820 ObjectStore::Transaction t
;
6821 t
.write(coll_t(), ghobject_t(soid
), 0, osize
, bl
);
6822 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6823 cleanupt
.remove(coll_t(), ghobject_t(soid
));
6828 bufferptr
bp(bsize
);
6830 bl
.push_back(std::move(bp
));
6831 bl
.rebuild_page_aligned();
6835 if (!osr
->flush_commit(&waiter
)) {
6840 utime_t start
= ceph_clock_now();
6841 for (int64_t pos
= 0; pos
< count
; pos
+= bsize
) {
6843 unsigned offset
= 0;
6844 if (onum
&& osize
) {
6845 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", (int)(rand() % onum
));
6846 offset
= rand() % (osize
/ bsize
) * bsize
;
6848 snprintf(nm
, sizeof(nm
), "disk_bw_test_%lld", (long long)pos
);
6851 hobject_t
soid(sobject_t(oid
, 0));
6852 ObjectStore::Transaction t
;
6853 t
.write(coll_t::meta(), ghobject_t(soid
), offset
, bsize
, bl
);
6854 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6855 if (!onum
|| !osize
)
6856 cleanupt
.remove(coll_t::meta(), ghobject_t(soid
));
6861 if (!osr
->flush_commit(&waiter
)) {
6865 utime_t end
= ceph_clock_now();
6868 store
->queue_transaction(osr
.get(), std::move(cleanupt
), NULL
);
6871 if (!osr
->flush_commit(&waiter
)) {
6876 uint64_t rate
= (double)count
/ (end
- start
);
6878 f
->open_object_section("osd_bench_results");
6879 f
->dump_int("bytes_written", count
);
6880 f
->dump_int("blocksize", bsize
);
6881 f
->dump_unsigned("bytes_per_sec", rate
);
6885 ss
<< "bench: wrote " << prettybyte_t(count
)
6886 << " in blocks of " << prettybyte_t(bsize
) << " in "
6887 << (end
-start
) << " sec at " << prettybyte_t(rate
) << "/sec";
6891 else if (prefix
== "flush_pg_stats") {
6892 if (osdmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
6893 mgrc
.send_pgstats();
6894 ds
<< service
.get_osd_stat_seq() << "\n";
6900 else if (prefix
== "heap") {
6901 r
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ds
);
6904 else if (prefix
== "debug dump_missing") {
6906 cmd_getval(cct
, cmdmap
, "filename", file_name
);
6907 std::ofstream
fout(file_name
.c_str());
6908 if (!fout
.is_open()) {
6909 ss
<< "failed to open file '" << file_name
<< "'";
6914 fout
<< "*** osd " << whoami
<< ": dump_missing ***" << std::endl
;
6915 RWLock::RLocker
l(pg_map_lock
);
6916 for (ceph::unordered_map
<spg_t
, PG
*>::const_iterator pg_map_e
= pg_map
.begin();
6917 pg_map_e
!= pg_map
.end(); ++pg_map_e
) {
6918 PG
*pg
= pg_map_e
->second
;
6921 fout
<< *pg
<< std::endl
;
6922 std::map
<hobject_t
, pg_missing_item
>::const_iterator mend
=
6923 pg
->pg_log
.get_missing().get_items().end();
6924 std::map
<hobject_t
, pg_missing_item
>::const_iterator mi
=
6925 pg
->pg_log
.get_missing().get_items().begin();
6926 for (; mi
!= mend
; ++mi
) {
6927 fout
<< mi
->first
<< " -> " << mi
->second
<< std::endl
;
6928 if (!pg
->missing_loc
.needs_recovery(mi
->first
))
6930 if (pg
->missing_loc
.is_unfound(mi
->first
))
6931 fout
<< " unfound ";
6932 const set
<pg_shard_t
> &mls(pg
->missing_loc
.get_locations(mi
->first
));
6935 fout
<< "missing_loc: " << mls
<< std::endl
;
6943 else if (prefix
== "debug kick_recovery_wq") {
6945 cmd_getval(cct
, cmdmap
, "delay", delay
);
6948 r
= cct
->_conf
->set_val("osd_recovery_delay_start", oss
.str().c_str());
6950 ss
<< "kick_recovery_wq: error setting "
6951 << "osd_recovery_delay_start to '" << delay
<< "': error "
6955 cct
->_conf
->apply_changes(NULL
);
6956 ss
<< "kicking recovery queue. set osd_recovery_delay_start "
6957 << "to " << cct
->_conf
->osd_recovery_delay_start
;
6960 else if (prefix
== "cpu_profiler") {
6962 cmd_getval(cct
, cmdmap
, "arg", arg
);
6963 vector
<string
> argvec
;
6964 get_str_vec(arg
, argvec
);
6965 cpu_profiler_handle_command(argvec
, ds
);
6968 else if (prefix
== "dump_pg_recovery_stats") {
6971 pg_recovery_stats
.dump_formatted(f
.get());
6974 pg_recovery_stats
.dump(s
);
6975 ds
<< "dump pg recovery stats: " << s
.str();
6979 else if (prefix
== "reset_pg_recovery_stats") {
6980 ss
<< "reset pg recovery stats";
6981 pg_recovery_stats
.reset();
6984 else if (prefix
== "perf histogram dump") {
6986 std::string counter
;
6987 cmd_getval(cct
, cmdmap
, "logger", logger
);
6988 cmd_getval(cct
, cmdmap
, "counter", counter
);
6990 cct
->get_perfcounters_collection()->dump_formatted_histograms(
6991 f
.get(), false, logger
, counter
);
6996 else if (prefix
== "compact") {
6997 dout(1) << "triggering manual compaction" << dendl
;
6998 auto start
= ceph::coarse_mono_clock::now();
7000 auto end
= ceph::coarse_mono_clock::now();
7001 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
7002 dout(1) << "finished manual compaction in "
7003 << time_span
.count()
7004 << " seconds" << dendl
;
7005 ss
<< "compacted omap in " << time_span
.count() << " seconds";
7009 ss
<< "unrecognized command! " << cmd
;
7016 dout(0) << "do_command r=" << r
<< " " << rs
<< dendl
;
7019 MCommandReply
*reply
= new MCommandReply(r
, rs
);
7020 reply
->set_tid(tid
);
7021 reply
->set_data(odata
);
7022 con
->send_message(reply
);
7026 bool OSD::heartbeat_dispatch(Message
*m
)
7028 dout(30) << "heartbeat_dispatch " << m
<< dendl
;
7029 switch (m
->get_type()) {
7032 dout(10) << "ping from " << m
->get_source_inst() << dendl
;
7037 handle_osd_ping(static_cast<MOSDPing
*>(m
));
7041 dout(0) << "dropping unexpected message " << *m
<< " from " << m
->get_source_inst() << dendl
;
7048 bool OSD::ms_dispatch(Message
*m
)
7050 dout(20) << "OSD::ms_dispatch: " << *m
<< dendl
;
7051 if (m
->get_type() == MSG_OSD_MARK_ME_DOWN
) {
7052 service
.got_stop_ack();
7060 if (is_stopping()) {
7074 void OSD::maybe_share_map(
7079 if (!op
->check_send_map
) {
7082 epoch_t last_sent_epoch
= 0;
7084 session
->sent_epoch_lock
.lock();
7085 last_sent_epoch
= session
->last_sent_epoch
;
7086 session
->sent_epoch_lock
.unlock();
7088 const Message
*m
= op
->get_req();
7091 m
->get_connection().get(),
7094 session
? &last_sent_epoch
: NULL
);
7096 session
->sent_epoch_lock
.lock();
7097 if (session
->last_sent_epoch
< last_sent_epoch
) {
7098 session
->last_sent_epoch
= last_sent_epoch
;
7100 session
->sent_epoch_lock
.unlock();
7102 op
->check_send_map
= false;
7105 void OSD::dispatch_session_waiting(Session
*session
, OSDMapRef osdmap
)
7107 assert(session
->session_dispatch_lock
.is_locked());
7109 auto i
= session
->waiting_on_map
.begin();
7110 while (i
!= session
->waiting_on_map
.end()) {
7111 OpRequestRef op
= &(*i
);
7112 assert(ms_can_fast_dispatch(op
->get_req()));
7113 const MOSDFastDispatchOp
*m
= static_cast<const MOSDFastDispatchOp
*>(
7115 if (m
->get_min_epoch() > osdmap
->get_epoch()) {
7118 session
->waiting_on_map
.erase(i
++);
7122 if (m
->get_type() == CEPH_MSG_OSD_OP
) {
7123 pg_t actual_pgid
= osdmap
->raw_pg_to_pg(
7124 static_cast<const MOSDOp
*>(m
)->get_pg());
7125 if (!osdmap
->get_primary_shard(actual_pgid
, &pgid
)) {
7129 pgid
= m
->get_spg();
7131 enqueue_op(pgid
, op
, m
->get_map_epoch());
7134 if (session
->waiting_on_map
.empty()) {
7135 clear_session_waiting_on_map(session
);
7137 register_session_waiting_on_map(session
);
7141 void OSD::ms_fast_dispatch(Message
*m
)
7144 if (service
.is_stopping()) {
7148 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
7151 osd_reqid_t reqid
= op
->get_reqid();
7153 tracepoint(osd
, ms_fast_dispatch
, reqid
.name
._type
,
7154 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
7158 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
7160 // note sender epoch, min req'd epoch
7161 op
->sent_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch();
7162 op
->min_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_min_epoch();
7163 assert(op
->min_epoch
<= op
->sent_epoch
); // sanity check!
7165 service
.maybe_inject_dispatch_delay();
7167 if (m
->get_connection()->has_features(CEPH_FEATUREMASK_RESEND_ON_SPLIT
) ||
7168 m
->get_type() != CEPH_MSG_OSD_OP
) {
7169 // queue it directly
7171 static_cast<MOSDFastDispatchOp
*>(m
)->get_spg(),
7173 static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch());
7175 // legacy client, and this is an MOSDOp (the *only* fast dispatch
7176 // message that didn't have an explicit spg_t); we need to map
7177 // them to an spg_t while preserving delivery order.
7178 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
7181 Mutex::Locker
l(session
->session_dispatch_lock
);
7183 session
->waiting_on_map
.push_back(*op
);
7184 OSDMapRef nextmap
= service
.get_nextmap_reserved();
7185 dispatch_session_waiting(session
, nextmap
);
7186 service
.release_map(nextmap
);
7191 OID_EVENT_TRACE_WITH_MSG(m
, "MS_FAST_DISPATCH_END", false);
7194 void OSD::ms_fast_preprocess(Message
*m
)
7196 if (m
->get_connection()->get_peer_type() == CEPH_ENTITY_TYPE_OSD
) {
7197 if (m
->get_type() == CEPH_MSG_OSD_MAP
) {
7198 MOSDMap
*mm
= static_cast<MOSDMap
*>(m
);
7199 Session
*s
= static_cast<Session
*>(m
->get_connection()->get_priv());
7201 s
->received_map_lock
.lock();
7202 s
->received_map_epoch
= mm
->get_last();
7203 s
->received_map_lock
.unlock();
7210 bool OSD::ms_get_authorizer(int dest_type
, AuthAuthorizer
**authorizer
, bool force_new
)
7212 dout(10) << "OSD::ms_get_authorizer type=" << ceph_entity_type_name(dest_type
) << dendl
;
7214 if (is_stopping()) {
7215 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7219 if (dest_type
== CEPH_ENTITY_TYPE_MON
)
7223 /* the MonClient checks keys every tick(), so we should just wait for that cycle
7225 if (monc
->wait_auth_rotating(10) < 0) {
7226 derr
<< "OSD::ms_get_authorizer wait_auth_rotating failed" << dendl
;
7231 *authorizer
= monc
->build_authorizer(dest_type
);
7232 return *authorizer
!= NULL
;
7236 bool OSD::ms_verify_authorizer(Connection
*con
, int peer_type
,
7237 int protocol
, bufferlist
& authorizer_data
, bufferlist
& authorizer_reply
,
7238 bool& isvalid
, CryptoKey
& session_key
)
7240 AuthAuthorizeHandler
*authorize_handler
= 0;
7241 switch (peer_type
) {
7242 case CEPH_ENTITY_TYPE_MDS
:
7244 * note: mds is technically a client from our perspective, but
7245 * this makes the 'cluster' consistent w/ monitor's usage.
7247 case CEPH_ENTITY_TYPE_OSD
:
7248 case CEPH_ENTITY_TYPE_MGR
:
7249 authorize_handler
= authorize_handler_cluster_registry
->get_handler(protocol
);
7252 authorize_handler
= authorize_handler_service_registry
->get_handler(protocol
);
7254 if (!authorize_handler
) {
7255 dout(0) << "No AuthAuthorizeHandler found for protocol " << protocol
<< dendl
;
7260 AuthCapsInfo caps_info
;
7263 uint64_t auid
= CEPH_AUTH_UID_DEFAULT
;
7265 RotatingKeyRing
*keys
= monc
->rotating_secrets
.get();
7267 isvalid
= authorize_handler
->verify_authorizer(
7269 authorizer_data
, authorizer_reply
, name
, global_id
, caps_info
, session_key
,
7272 dout(10) << __func__
<< " no rotating_keys (yet), denied" << dendl
;
7277 Session
*s
= static_cast<Session
*>(con
->get_priv());
7279 s
= new Session(cct
);
7280 con
->set_priv(s
->get());
7282 dout(10) << " new session " << s
<< " con=" << s
->con
<< " addr=" << s
->con
->get_peer_addr() << dendl
;
7285 s
->entity_name
= name
;
7286 if (caps_info
.allow_all
)
7287 s
->caps
.set_allow_all();
7290 if (caps_info
.caps
.length() > 0) {
7291 bufferlist::iterator p
= caps_info
.caps
.begin();
7296 catch (buffer::error
& e
) {
7298 bool success
= s
->caps
.parse(str
);
7300 dout(10) << " session " << s
<< " " << s
->entity_name
<< " has caps " << s
->caps
<< " '" << str
<< "'" << dendl
;
7302 dout(10) << " session " << s
<< " " << s
->entity_name
<< " failed to parse caps '" << str
<< "'" << dendl
;
7310 void OSD::do_waiters()
7312 assert(osd_lock
.is_locked());
7314 dout(10) << "do_waiters -- start" << dendl
;
7315 while (!finished
.empty()) {
7316 OpRequestRef next
= finished
.front();
7317 finished
.pop_front();
7320 dout(10) << "do_waiters -- finish" << dendl
;
7323 void OSD::dispatch_op(OpRequestRef op
)
7325 switch (op
->get_req()->get_type()) {
7327 case MSG_OSD_PG_CREATE
:
7328 handle_pg_create(op
);
7330 case MSG_OSD_PG_NOTIFY
:
7331 handle_pg_notify(op
);
7333 case MSG_OSD_PG_QUERY
:
7334 handle_pg_query(op
);
7336 case MSG_OSD_PG_LOG
:
7339 case MSG_OSD_PG_REMOVE
:
7340 handle_pg_remove(op
);
7342 case MSG_OSD_PG_INFO
:
7345 case MSG_OSD_PG_TRIM
:
7348 case MSG_OSD_BACKFILL_RESERVE
:
7349 handle_pg_backfill_reserve(op
);
7351 case MSG_OSD_RECOVERY_RESERVE
:
7352 handle_pg_recovery_reserve(op
);
7357 void OSD::_dispatch(Message
*m
)
7359 assert(osd_lock
.is_locked());
7360 dout(20) << "_dispatch " << m
<< " " << *m
<< dendl
;
7362 switch (m
->get_type()) {
7364 // -- don't need lock --
7366 dout(10) << "ping from " << m
->get_source() << dendl
;
7370 // -- don't need OSDMap --
7372 // map and replication
7373 case CEPH_MSG_OSD_MAP
:
7374 handle_osd_map(static_cast<MOSDMap
*>(m
));
7378 case MSG_PGSTATSACK
:
7379 handle_pg_stats_ack(static_cast<MPGStatsAck
*>(m
));
7382 case MSG_MON_COMMAND
:
7383 handle_command(static_cast<MMonCommand
*>(m
));
7386 handle_command(static_cast<MCommand
*>(m
));
7390 handle_scrub(static_cast<MOSDScrub
*>(m
));
7393 case MSG_OSD_FORCE_RECOVERY
:
7394 handle_force_recovery(m
);
7397 // -- need OSDMap --
7399 case MSG_OSD_PG_CREATE
:
7400 case MSG_OSD_PG_NOTIFY
:
7401 case MSG_OSD_PG_QUERY
:
7402 case MSG_OSD_PG_LOG
:
7403 case MSG_OSD_PG_REMOVE
:
7404 case MSG_OSD_PG_INFO
:
7405 case MSG_OSD_PG_TRIM
:
7406 case MSG_OSD_BACKFILL_RESERVE
:
7407 case MSG_OSD_RECOVERY_RESERVE
:
7409 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
7411 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
7412 // no map? starting up?
7414 dout(7) << "no OSDMap, not booted" << dendl
;
7415 logger
->inc(l_osd_waiting_for_map
);
7416 waiting_for_osdmap
.push_back(op
);
7417 op
->mark_delayed("no osdmap");
7427 void OSD::handle_pg_scrub(MOSDScrub
*m
, PG
*pg
)
7430 if (pg
->is_primary()) {
7431 pg
->unreg_next_scrub();
7432 pg
->scrubber
.must_scrub
= true;
7433 pg
->scrubber
.must_deep_scrub
= m
->deep
|| m
->repair
;
7434 pg
->scrubber
.must_repair
= m
->repair
;
7435 pg
->reg_next_scrub();
7436 dout(10) << "marking " << *pg
<< " for scrub" << dendl
;
7441 void OSD::handle_scrub(MOSDScrub
*m
)
7443 dout(10) << "handle_scrub " << *m
<< dendl
;
7444 if (!require_mon_or_mgr_peer(m
)) {
7448 if (m
->fsid
!= monc
->get_fsid()) {
7449 dout(0) << "handle_scrub fsid " << m
->fsid
<< " != " << monc
->get_fsid() << dendl
;
7454 RWLock::RLocker
l(pg_map_lock
);
7455 if (m
->scrub_pgs
.empty()) {
7456 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
7459 handle_pg_scrub(m
, p
->second
);
7461 for (vector
<pg_t
>::iterator p
= m
->scrub_pgs
.begin();
7462 p
!= m
->scrub_pgs
.end();
7465 if (osdmap
->get_primary_shard(*p
, &pcand
)) {
7466 auto pg_map_entry
= pg_map
.find(pcand
);
7467 if (pg_map_entry
!= pg_map
.end()) {
7468 handle_pg_scrub(m
, pg_map_entry
->second
);
7477 bool OSD::scrub_random_backoff()
7479 bool coin_flip
= (rand() / (double)RAND_MAX
>=
7480 cct
->_conf
->osd_scrub_backoff_ratio
);
7482 dout(20) << "scrub_random_backoff lost coin flip, randomly backing off" << dendl
;
7488 OSDService::ScrubJob::ScrubJob(CephContext
* cct
,
7489 const spg_t
& pg
, const utime_t
& timestamp
,
7490 double pool_scrub_min_interval
,
7491 double pool_scrub_max_interval
, bool must
)
7494 sched_time(timestamp
),
7497 // if not explicitly requested, postpone the scrub with a random delay
7499 double scrub_min_interval
= pool_scrub_min_interval
> 0 ?
7500 pool_scrub_min_interval
: cct
->_conf
->osd_scrub_min_interval
;
7501 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
7502 pool_scrub_max_interval
: cct
->_conf
->osd_scrub_max_interval
;
7504 sched_time
+= scrub_min_interval
;
7505 double r
= rand() / (double)RAND_MAX
;
7507 scrub_min_interval
* cct
->_conf
->osd_scrub_interval_randomize_ratio
* r
;
7508 deadline
+= scrub_max_interval
;
7512 bool OSDService::ScrubJob::ScrubJob::operator<(const OSDService::ScrubJob
& rhs
) const {
7513 if (sched_time
< rhs
.sched_time
)
7515 if (sched_time
> rhs
.sched_time
)
7517 return pgid
< rhs
.pgid
;
7520 bool OSD::scrub_time_permit(utime_t now
)
7523 time_t tt
= now
.sec();
7524 localtime_r(&tt
, &bdt
);
7525 bool time_permit
= false;
7526 if (cct
->_conf
->osd_scrub_begin_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7527 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
&& bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7531 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
|| bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7536 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7537 << " - " << cct
->_conf
->osd_scrub_end_hour
7538 << " now " << bdt
.tm_hour
<< " = no" << dendl
;
7540 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7541 << " - " << cct
->_conf
->osd_scrub_end_hour
7542 << " now " << bdt
.tm_hour
<< " = yes" << dendl
;
7547 bool OSD::scrub_load_below_threshold()
7550 if (getloadavg(loadavgs
, 3) != 3) {
7551 dout(10) << __func__
<< " couldn't read loadavgs\n" << dendl
;
7555 // allow scrub if below configured threshold
7556 if (loadavgs
[0] < cct
->_conf
->osd_scrub_load_threshold
) {
7557 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7558 << " < max " << cct
->_conf
->osd_scrub_load_threshold
7559 << " = yes" << dendl
;
7563 // allow scrub if below daily avg and currently decreasing
7564 if (loadavgs
[0] < daily_loadavg
&& loadavgs
[0] < loadavgs
[2]) {
7565 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7566 << " < daily_loadavg " << daily_loadavg
7567 << " and < 15m avg " << loadavgs
[2]
7568 << " = yes" << dendl
;
7572 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7573 << " >= max " << cct
->_conf
->osd_scrub_load_threshold
7574 << " and ( >= daily_loadavg " << daily_loadavg
7575 << " or >= 15m avg " << loadavgs
[2]
7576 << ") = no" << dendl
;
7580 void OSD::sched_scrub()
7582 // if not permitted, fail fast
7583 if (!service
.can_inc_scrubs_pending()) {
7586 if (!cct
->_conf
->osd_scrub_during_recovery
&& service
.is_recovery_active()) {
7587 dout(20) << __func__
<< " not scheduling scrubs due to active recovery" << dendl
;
7592 utime_t now
= ceph_clock_now();
7593 bool time_permit
= scrub_time_permit(now
);
7594 bool load_is_low
= scrub_load_below_threshold();
7595 dout(20) << "sched_scrub load_is_low=" << (int)load_is_low
<< dendl
;
7597 OSDService::ScrubJob scrub
;
7598 if (service
.first_scrub_stamp(&scrub
)) {
7600 dout(30) << "sched_scrub examine " << scrub
.pgid
<< " at " << scrub
.sched_time
<< dendl
;
7602 if (scrub
.sched_time
> now
) {
7603 // save ourselves some effort
7604 dout(10) << "sched_scrub " << scrub
.pgid
<< " scheduled at " << scrub
.sched_time
7605 << " > " << now
<< dendl
;
7609 if ((scrub
.deadline
>= now
) && !(time_permit
&& load_is_low
)) {
7610 dout(10) << __func__
<< " not scheduling scrub for " << scrub
.pgid
<< " due to "
7611 << (!time_permit
? "time not permit" : "high load") << dendl
;
7615 PG
*pg
= _lookup_lock_pg(scrub
.pgid
);
7618 if (pg
->get_pgbackend()->scrub_supported() && pg
->is_active()) {
7619 dout(10) << "sched_scrub scrubbing " << scrub
.pgid
<< " at " << scrub
.sched_time
7620 << (pg
->scrubber
.must_scrub
? ", explicitly requested" :
7621 (load_is_low
? ", load_is_low" : " deadline < now"))
7623 if (pg
->sched_scrub()) {
7629 } while (service
.next_scrub_stamp(scrub
, &scrub
));
7631 dout(20) << "sched_scrub done" << dendl
;
7636 vector
<OSDHealthMetric
> OSD::get_health_metrics()
7638 vector
<OSDHealthMetric
> metrics
;
7639 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
7640 auto n_primaries
= pending_creates_from_mon
;
7641 for (const auto& create
: pending_creates_from_osd
) {
7642 if (create
.second
) {
7646 metrics
.emplace_back(osd_metric::PENDING_CREATING_PGS
, n_primaries
);
7650 // =====================================================
7653 void OSD::wait_for_new_map(OpRequestRef op
)
7656 if (waiting_for_osdmap
.empty()) {
7657 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
7660 logger
->inc(l_osd_waiting_for_map
);
7661 waiting_for_osdmap
.push_back(op
);
7662 op
->mark_delayed("wait for new map");
7667 * assimilate new OSDMap(s). scan pgs, etc.
7670 void OSD::note_down_osd(int peer
)
7672 assert(osd_lock
.is_locked());
7673 cluster_messenger
->mark_down(osdmap
->get_cluster_addr(peer
));
7675 heartbeat_lock
.Lock();
7676 failure_queue
.erase(peer
);
7677 failure_pending
.erase(peer
);
7678 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(peer
);
7679 if (p
!= heartbeat_peers
.end()) {
7680 p
->second
.con_back
->mark_down();
7681 if (p
->second
.con_front
) {
7682 p
->second
.con_front
->mark_down();
7684 heartbeat_peers
.erase(p
);
7686 heartbeat_lock
.Unlock();
7689 void OSD::note_up_osd(int peer
)
7691 service
.forget_peer_epoch(peer
, osdmap
->get_epoch() - 1);
7692 heartbeat_set_peers_need_update();
7695 struct C_OnMapCommit
: public Context
{
7697 epoch_t first
, last
;
7699 C_OnMapCommit(OSD
*o
, epoch_t f
, epoch_t l
, MOSDMap
*m
)
7700 : osd(o
), first(f
), last(l
), msg(m
) {}
7701 void finish(int r
) override
{
7702 osd
->_committed_osd_maps(first
, last
, msg
);
7707 struct C_OnMapApply
: public Context
{
7708 OSDService
*service
;
7709 list
<OSDMapRef
> pinned_maps
;
7711 C_OnMapApply(OSDService
*service
,
7712 const list
<OSDMapRef
> &pinned_maps
,
7714 : service(service
), pinned_maps(pinned_maps
), e(e
) {}
7715 void finish(int r
) override
{
7716 service
->clear_map_bl_cache_pins(e
);
7720 void OSD::osdmap_subscribe(version_t epoch
, bool force_request
)
7722 Mutex::Locker
l(osdmap_subscribe_lock
);
7723 if (latest_subscribed_epoch
>= epoch
&& !force_request
)
7726 latest_subscribed_epoch
= MAX(epoch
, latest_subscribed_epoch
);
7728 if (monc
->sub_want_increment("osdmap", epoch
, CEPH_SUBSCRIBE_ONETIME
) ||
7734 void OSD::trim_maps(epoch_t oldest
, int nreceived
, bool skip_maps
)
7736 epoch_t min
= std::min(oldest
, service
.map_cache
.cached_key_lower_bound());
7737 if (min
<= superblock
.oldest_map
)
7741 ObjectStore::Transaction t
;
7742 for (epoch_t e
= superblock
.oldest_map
; e
< min
; ++e
) {
7743 dout(20) << " removing old osdmap epoch " << e
<< dendl
;
7744 t
.remove(coll_t::meta(), get_osdmap_pobject_name(e
));
7745 t
.remove(coll_t::meta(), get_inc_osdmap_pobject_name(e
));
7746 superblock
.oldest_map
= e
+ 1;
7748 if (num
>= cct
->_conf
->osd_target_transaction_size
&& num
>= nreceived
) {
7749 service
.publish_superblock(superblock
);
7750 write_superblock(t
);
7751 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7755 // skip_maps leaves us with a range of old maps if we fail to remove all
7756 // of them before moving superblock.oldest_map forward to the first map
7757 // in the incoming MOSDMap msg. so we should continue removing them in
7758 // this case, even we could do huge series of delete transactions all at
7765 service
.publish_superblock(superblock
);
7766 write_superblock(t
);
7767 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7770 // we should not remove the cached maps
7771 assert(min
<= service
.map_cache
.cached_key_lower_bound());
7774 void OSD::handle_osd_map(MOSDMap
*m
)
7776 assert(osd_lock
.is_locked());
7777 // Keep a ref in the list until we get the newly received map written
7778 // onto disk. This is important because as long as the refs are alive,
7779 // the OSDMaps will be pinned in the cache and we won't try to read it
7780 // off of disk. Otherwise these maps will probably not stay in the cache,
7781 // and reading those OSDMaps before they are actually written can result
7783 list
<OSDMapRef
> pinned_maps
;
7784 if (m
->fsid
!= monc
->get_fsid()) {
7785 dout(0) << "handle_osd_map fsid " << m
->fsid
<< " != "
7786 << monc
->get_fsid() << dendl
;
7790 if (is_initializing()) {
7791 dout(0) << "ignoring osdmap until we have initialized" << dendl
;
7796 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
7797 if (session
&& !(session
->entity_name
.is_mon() ||
7798 session
->entity_name
.is_osd())) {
7800 dout(10) << "got osd map from Session " << session
7801 << " which we can't take maps from (not a mon or osd)" << dendl
;
7809 // share with the objecter
7811 service
.objecter
->handle_osd_map(m
);
7813 epoch_t first
= m
->get_first();
7814 epoch_t last
= m
->get_last();
7815 dout(3) << "handle_osd_map epochs [" << first
<< "," << last
<< "], i have "
7816 << superblock
.newest_map
7817 << ", src has [" << m
->oldest_map
<< "," << m
->newest_map
<< "]"
7820 logger
->inc(l_osd_map
);
7821 logger
->inc(l_osd_mape
, last
- first
+ 1);
7822 if (first
<= superblock
.newest_map
)
7823 logger
->inc(l_osd_mape_dup
, superblock
.newest_map
- first
+ 1);
7824 if (service
.max_oldest_map
< m
->oldest_map
) {
7825 service
.max_oldest_map
= m
->oldest_map
;
7826 assert(service
.max_oldest_map
>= superblock
.oldest_map
);
7829 // make sure there is something new, here, before we bother flushing
7830 // the queues and such
7831 if (last
<= superblock
.newest_map
) {
7832 dout(10) << " no new maps here, dropping" << dendl
;
7838 bool skip_maps
= false;
7839 if (first
> superblock
.newest_map
+ 1) {
7840 dout(10) << "handle_osd_map message skips epochs "
7841 << superblock
.newest_map
+ 1 << ".." << (first
-1) << dendl
;
7842 if (m
->oldest_map
<= superblock
.newest_map
+ 1) {
7843 osdmap_subscribe(superblock
.newest_map
+ 1, false);
7847 // always try to get the full range of maps--as many as we can. this
7848 // 1- is good to have
7849 // 2- is at present the only way to ensure that we get a *full* map as
7851 if (m
->oldest_map
< first
) {
7852 osdmap_subscribe(m
->oldest_map
- 1, true);
7859 ObjectStore::Transaction t
;
7860 uint64_t txn_size
= 0;
7862 // store new maps: queue for disk and put in the osdmap cache
7863 epoch_t start
= MAX(superblock
.newest_map
+ 1, first
);
7864 for (epoch_t e
= start
; e
<= last
; e
++) {
7865 if (txn_size
>= t
.get_num_bytes()) {
7866 derr
<< __func__
<< " transaction size overflowed" << dendl
;
7867 assert(txn_size
< t
.get_num_bytes());
7869 txn_size
= t
.get_num_bytes();
7870 map
<epoch_t
,bufferlist
>::iterator p
;
7871 p
= m
->maps
.find(e
);
7872 if (p
!= m
->maps
.end()) {
7873 dout(10) << "handle_osd_map got full map for epoch " << e
<< dendl
;
7874 OSDMap
*o
= new OSDMap
;
7875 bufferlist
& bl
= p
->second
;
7879 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7880 t
.write(coll_t::meta(), fulloid
, 0, bl
.length(), bl
);
7882 pinned_maps
.push_back(add_map(o
));
7888 p
= m
->incremental_maps
.find(e
);
7889 if (p
!= m
->incremental_maps
.end()) {
7890 dout(10) << "handle_osd_map got inc map for epoch " << e
<< dendl
;
7891 bufferlist
& bl
= p
->second
;
7892 ghobject_t oid
= get_inc_osdmap_pobject_name(e
);
7893 t
.write(coll_t::meta(), oid
, 0, bl
.length(), bl
);
7894 pin_map_inc_bl(e
, bl
);
7896 OSDMap
*o
= new OSDMap
;
7899 bool got
= get_map_bl(e
- 1, obl
);
7904 OSDMap::Incremental inc
;
7905 bufferlist::iterator p
= bl
.begin();
7907 if (o
->apply_incremental(inc
) < 0) {
7908 derr
<< "ERROR: bad fsid? i have " << osdmap
->get_fsid() << " and inc has " << inc
.fsid
<< dendl
;
7909 assert(0 == "bad fsid");
7913 o
->encode(fbl
, inc
.encode_features
| CEPH_FEATURE_RESERVED
);
7915 bool injected_failure
= false;
7916 if (cct
->_conf
->osd_inject_bad_map_crc_probability
> 0 &&
7917 (rand() % 10000) < cct
->_conf
->osd_inject_bad_map_crc_probability
*10000.0) {
7918 derr
<< __func__
<< " injecting map crc failure" << dendl
;
7919 injected_failure
= true;
7922 if ((inc
.have_crc
&& o
->get_crc() != inc
.full_crc
) || injected_failure
) {
7923 dout(2) << "got incremental " << e
7924 << " but failed to encode full with correct crc; requesting"
7926 clog
->warn() << "failed to encode map e" << e
<< " with expected crc";
7927 dout(20) << "my encoded map was:\n";
7928 fbl
.hexdump(*_dout
);
7931 request_full_map(e
, last
);
7937 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7938 t
.write(coll_t::meta(), fulloid
, 0, fbl
.length(), fbl
);
7940 pinned_maps
.push_back(add_map(o
));
7944 assert(0 == "MOSDMap lied about what maps it had?");
7947 // even if this map isn't from a mon, we may have satisfied our subscription
7948 monc
->sub_got("osdmap", last
);
7950 if (!m
->maps
.empty() && requested_full_first
) {
7951 dout(10) << __func__
<< " still missing full maps " << requested_full_first
7952 << ".." << requested_full_last
<< dendl
;
7953 rerequest_full_maps();
7956 if (superblock
.oldest_map
) {
7957 // make sure we at least keep pace with incoming maps
7958 trim_maps(m
->oldest_map
, last
- first
+ 1, skip_maps
);
7961 if (!superblock
.oldest_map
|| skip_maps
)
7962 superblock
.oldest_map
= first
;
7963 superblock
.newest_map
= last
;
7964 superblock
.current_epoch
= last
;
7966 // note in the superblock that we were clean thru the prior epoch
7967 epoch_t boot_epoch
= service
.get_boot_epoch();
7968 if (boot_epoch
&& boot_epoch
>= superblock
.mounted
) {
7969 superblock
.mounted
= boot_epoch
;
7970 superblock
.clean_thru
= last
;
7973 // superblock and commit
7974 write_superblock(t
);
7975 store
->queue_transaction(
7976 service
.meta_osr
.get(),
7978 new C_OnMapApply(&service
, pinned_maps
, last
),
7979 new C_OnMapCommit(this, start
, last
, m
), 0);
7980 service
.publish_superblock(superblock
);
7983 void OSD::_committed_osd_maps(epoch_t first
, epoch_t last
, MOSDMap
*m
)
7985 dout(10) << __func__
<< " " << first
<< ".." << last
<< dendl
;
7986 if (is_stopping()) {
7987 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7990 Mutex::Locker
l(osd_lock
);
7991 if (is_stopping()) {
7992 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7995 map_lock
.get_write();
7997 bool do_shutdown
= false;
7998 bool do_restart
= false;
7999 bool network_error
= false;
8001 // advance through the new maps
8002 for (epoch_t cur
= first
; cur
<= last
; cur
++) {
8003 dout(10) << " advance to epoch " << cur
8004 << " (<= last " << last
8005 << " <= newest_map " << superblock
.newest_map
8008 OSDMapRef newmap
= get_map(cur
);
8009 assert(newmap
); // we just cached it above!
8011 // start blacklisting messages sent to peers that go down.
8012 service
.pre_publish_map(newmap
);
8014 // kill connections to newly down osds
8015 bool waited_for_reservations
= false;
8017 osdmap
->get_all_osds(old
);
8018 for (set
<int>::iterator p
= old
.begin(); p
!= old
.end(); ++p
) {
8020 osdmap
->is_up(*p
) && // in old map
8021 newmap
->is_down(*p
)) { // but not the new one
8022 if (!waited_for_reservations
) {
8023 service
.await_reserved_maps();
8024 waited_for_reservations
= true;
8027 } else if (*p
!= whoami
&&
8028 osdmap
->is_down(*p
) &&
8029 newmap
->is_up(*p
)) {
8034 if ((osdmap
->test_flag(CEPH_OSDMAP_NOUP
) !=
8035 newmap
->test_flag(CEPH_OSDMAP_NOUP
)) ||
8036 (osdmap
->is_noup(whoami
) != newmap
->is_noup(whoami
))) {
8037 dout(10) << __func__
<< " NOUP flag changed in " << newmap
->get_epoch()
8040 // this captures the case where we sent the boot message while
8041 // NOUP was being set on the mon and our boot request was
8042 // dropped, and then later it is cleared. it imperfectly
8043 // handles the case where our original boot message was not
8044 // dropped and we restart even though we might have booted, but
8045 // that is harmless (boot will just take slightly longer).
8049 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
&&
8050 newmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
8051 dout(10) << __func__
<< " require_osd_release reached luminous in "
8052 << newmap
->get_epoch() << dendl
;
8053 clear_pg_stat_queue();
8054 clear_outstanding_pg_stats();
8060 service
.retrieve_epochs(&boot_epoch
, &up_epoch
, NULL
);
8062 osdmap
->is_up(whoami
) &&
8063 osdmap
->get_inst(whoami
) == client_messenger
->get_myinst()) {
8064 up_epoch
= osdmap
->get_epoch();
8065 dout(10) << "up_epoch is " << up_epoch
<< dendl
;
8067 boot_epoch
= osdmap
->get_epoch();
8068 dout(10) << "boot_epoch is " << boot_epoch
<< dendl
;
8070 service
.set_epochs(&boot_epoch
, &up_epoch
, NULL
);
8074 had_map_since
= ceph_clock_now();
8076 epoch_t _bind_epoch
= service
.get_bind_epoch();
8077 if (osdmap
->is_up(whoami
) &&
8078 osdmap
->get_addr(whoami
) == client_messenger
->get_myaddr() &&
8079 _bind_epoch
< osdmap
->get_up_from(whoami
)) {
8082 dout(1) << "state: booting -> active" << dendl
;
8083 set_state(STATE_ACTIVE
);
8085 // set incarnation so that osd_reqid_t's we generate for our
8086 // objecter requests are unique across restarts.
8087 service
.objecter
->set_client_incarnation(osdmap
->get_epoch());
8091 if (osdmap
->get_epoch() > 0 &&
8093 if (!osdmap
->exists(whoami
)) {
8094 dout(0) << "map says i do not exist. shutting down." << dendl
;
8095 do_shutdown
= true; // don't call shutdown() while we have
8096 // everything paused
8097 } else if (!osdmap
->is_up(whoami
) ||
8098 !osdmap
->get_addr(whoami
).probably_equals(
8099 client_messenger
->get_myaddr()) ||
8100 !osdmap
->get_cluster_addr(whoami
).probably_equals(
8101 cluster_messenger
->get_myaddr()) ||
8102 !osdmap
->get_hb_back_addr(whoami
).probably_equals(
8103 hb_back_server_messenger
->get_myaddr()) ||
8104 (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
8105 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
8106 hb_front_server_messenger
->get_myaddr()))) {
8107 if (!osdmap
->is_up(whoami
)) {
8108 if (service
.is_preparing_to_stop() || service
.is_stopping()) {
8109 service
.got_stop_ack();
8111 clog
->warn() << "Monitor daemon marked osd." << whoami
<< " down, "
8112 "but it is still running";
8113 clog
->debug() << "map e" << osdmap
->get_epoch()
8114 << " wrongly marked me down at e"
8115 << osdmap
->get_down_at(whoami
);
8117 } else if (!osdmap
->get_addr(whoami
).probably_equals(
8118 client_messenger
->get_myaddr())) {
8119 clog
->error() << "map e" << osdmap
->get_epoch()
8120 << " had wrong client addr (" << osdmap
->get_addr(whoami
)
8121 << " != my " << client_messenger
->get_myaddr() << ")";
8122 } else if (!osdmap
->get_cluster_addr(whoami
).probably_equals(
8123 cluster_messenger
->get_myaddr())) {
8124 clog
->error() << "map e" << osdmap
->get_epoch()
8125 << " had wrong cluster addr ("
8126 << osdmap
->get_cluster_addr(whoami
)
8127 << " != my " << cluster_messenger
->get_myaddr() << ")";
8128 } else if (!osdmap
->get_hb_back_addr(whoami
).probably_equals(
8129 hb_back_server_messenger
->get_myaddr())) {
8130 clog
->error() << "map e" << osdmap
->get_epoch()
8131 << " had wrong heartbeat back addr ("
8132 << osdmap
->get_hb_back_addr(whoami
)
8133 << " != my " << hb_back_server_messenger
->get_myaddr()
8135 } else if (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
8136 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
8137 hb_front_server_messenger
->get_myaddr())) {
8138 clog
->error() << "map e" << osdmap
->get_epoch()
8139 << " had wrong heartbeat front addr ("
8140 << osdmap
->get_hb_front_addr(whoami
)
8141 << " != my " << hb_front_server_messenger
->get_myaddr()
8145 if (!service
.is_stopping()) {
8146 epoch_t up_epoch
= 0;
8147 epoch_t bind_epoch
= osdmap
->get_epoch();
8148 service
.set_epochs(NULL
,&up_epoch
, &bind_epoch
);
8152 utime_t now
= ceph_clock_now();
8153 utime_t grace
= utime_t(cct
->_conf
->osd_max_markdown_period
, 0);
8154 osd_markdown_log
.push_back(now
);
8155 //clear all out-of-date log
8156 while (!osd_markdown_log
.empty() &&
8157 osd_markdown_log
.front() + grace
< now
)
8158 osd_markdown_log
.pop_front();
8159 if ((int)osd_markdown_log
.size() > cct
->_conf
->osd_max_markdown_count
) {
8160 dout(0) << __func__
<< " marked down "
8161 << osd_markdown_log
.size()
8162 << " > osd_max_markdown_count "
8163 << cct
->_conf
->osd_max_markdown_count
8164 << " in last " << grace
<< " seconds, shutting down"
8170 start_waiting_for_healthy();
8172 set
<int> avoid_ports
;
8173 #if defined(__FreeBSD__)
8174 // prevent FreeBSD from grabbing the client_messenger port during
8175 // rebinding. In which case a cluster_meesneger will connect also
8177 avoid_ports
.insert(client_messenger
->get_myaddr().get_port());
8179 avoid_ports
.insert(cluster_messenger
->get_myaddr().get_port());
8180 avoid_ports
.insert(hb_back_server_messenger
->get_myaddr().get_port());
8181 avoid_ports
.insert(hb_front_server_messenger
->get_myaddr().get_port());
8183 int r
= cluster_messenger
->rebind(avoid_ports
);
8185 do_shutdown
= true; // FIXME: do_restart?
8186 network_error
= true;
8187 dout(0) << __func__
<< " marked down:"
8188 << " rebind cluster_messenger failed" << dendl
;
8191 r
= hb_back_server_messenger
->rebind(avoid_ports
);
8193 do_shutdown
= true; // FIXME: do_restart?
8194 network_error
= true;
8195 dout(0) << __func__
<< " marked down:"
8196 << " rebind hb_back_server_messenger failed" << dendl
;
8199 r
= hb_front_server_messenger
->rebind(avoid_ports
);
8201 do_shutdown
= true; // FIXME: do_restart?
8202 network_error
= true;
8203 dout(0) << __func__
<< " marked down:"
8204 << " rebind hb_front_server_messenger failed" << dendl
;
8207 hb_front_client_messenger
->mark_down_all();
8208 hb_back_client_messenger
->mark_down_all();
8210 reset_heartbeat_peers();
8215 map_lock
.put_write();
8217 check_osdmap_features(store
);
8222 if (is_active() || is_waiting_for_healthy())
8223 maybe_update_heartbeat_peers();
8226 dout(10) << " not yet active; waiting for peering wq to drain" << dendl
;
8233 if (network_error
) {
8234 Mutex::Locker
l(heartbeat_lock
);
8235 map
<int,pair
<utime_t
,entity_inst_t
>>::iterator it
=
8236 failure_pending
.begin();
8237 while (it
!= failure_pending
.end()) {
8238 dout(10) << "handle_osd_ping canceling in-flight failure report for osd."
8239 << it
->first
<< dendl
;
8240 send_still_alive(osdmap
->get_epoch(), it
->second
.second
);
8241 failure_pending
.erase(it
++);
8244 // trigger shutdown in a different thread
8245 dout(0) << __func__
<< " shutdown OSD via async signal" << dendl
;
8246 queue_async_signal(SIGINT
);
8248 else if (m
->newest_map
&& m
->newest_map
> last
) {
8249 dout(10) << " msg say newest map is " << m
->newest_map
8250 << ", requesting more" << dendl
;
8251 osdmap_subscribe(osdmap
->get_epoch()+1, false);
8253 else if (is_preboot()) {
8254 if (m
->get_source().is_mon())
8255 _preboot(m
->oldest_map
, m
->newest_map
);
8259 else if (do_restart
)
8264 void OSD::check_osdmap_features(ObjectStore
*fs
)
8266 // adjust required feature bits?
8268 // we have to be a bit careful here, because we are accessing the
8269 // Policy structures without taking any lock. in particular, only
8270 // modify integer values that can safely be read by a racing CPU.
8271 // since we are only accessing existing Policy structures a their
8272 // current memory location, and setting or clearing bits in integer
8273 // fields, and we are the only writer, this is not a problem.
8276 Messenger::Policy p
= client_messenger
->get_default_policy();
8278 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_CLIENT
, &mask
);
8279 if ((p
.features_required
& mask
) != features
) {
8280 dout(0) << "crush map has features " << features
8281 << ", adjusting msgr requires for clients" << dendl
;
8282 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8283 client_messenger
->set_default_policy(p
);
8287 Messenger::Policy p
= client_messenger
->get_policy(entity_name_t::TYPE_MON
);
8289 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_MON
, &mask
);
8290 if ((p
.features_required
& mask
) != features
) {
8291 dout(0) << "crush map has features " << features
8292 << " was " << p
.features_required
8293 << ", adjusting msgr requires for mons" << dendl
;
8294 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8295 client_messenger
->set_policy(entity_name_t::TYPE_MON
, p
);
8299 Messenger::Policy p
= cluster_messenger
->get_policy(entity_name_t::TYPE_OSD
);
8301 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_OSD
, &mask
);
8303 if ((p
.features_required
& mask
) != features
) {
8304 dout(0) << "crush map has features " << features
8305 << ", adjusting msgr requires for osds" << dendl
;
8306 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8307 cluster_messenger
->set_policy(entity_name_t::TYPE_OSD
, p
);
8310 if ((features
& CEPH_FEATURE_OSD_ERASURE_CODES
) &&
8311 !superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
)) {
8312 dout(0) << __func__
<< " enabling on-disk ERASURE CODES compat feature" << dendl
;
8313 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
);
8314 ObjectStore::Transaction t
;
8315 write_superblock(t
);
8316 int err
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), NULL
);
8322 bool OSD::advance_pg(
8323 epoch_t osd_epoch
, PG
*pg
,
8324 ThreadPool::TPHandle
&handle
,
8325 PG::RecoveryCtx
*rctx
,
8326 set
<PGRef
> *new_pgs
)
8328 assert(pg
->is_locked());
8329 epoch_t next_epoch
= pg
->get_osdmap()->get_epoch() + 1;
8330 OSDMapRef lastmap
= pg
->get_osdmap();
8332 if (lastmap
->get_epoch() == osd_epoch
)
8334 assert(lastmap
->get_epoch() < osd_epoch
);
8336 epoch_t min_epoch
= service
.get_min_pg_epoch();
8339 max
= min_epoch
+ cct
->_conf
->osd_map_max_advance
;
8341 max
= next_epoch
+ cct
->_conf
->osd_map_max_advance
;
8345 next_epoch
<= osd_epoch
&& next_epoch
<= max
;
8347 OSDMapRef nextmap
= service
.try_get_map(next_epoch
);
8349 dout(20) << __func__
<< " missing map " << next_epoch
<< dendl
;
8350 // make sure max is bumped up so that we can get past any
8352 max
= MAX(max
, next_epoch
+ cct
->_conf
->osd_map_max_advance
);
8356 vector
<int> newup
, newacting
;
8357 int up_primary
, acting_primary
;
8358 nextmap
->pg_to_up_acting_osds(
8360 &newup
, &up_primary
,
8361 &newacting
, &acting_primary
);
8362 pg
->handle_advance_map(
8363 nextmap
, lastmap
, newup
, up_primary
,
8364 newacting
, acting_primary
, rctx
);
8367 set
<spg_t
> children
;
8368 spg_t
parent(pg
->info
.pgid
);
8369 if (parent
.is_split(
8370 lastmap
->get_pg_num(pg
->pool
.id
),
8371 nextmap
->get_pg_num(pg
->pool
.id
),
8373 service
.mark_split_in_progress(pg
->info
.pgid
, children
);
8375 pg
, children
, new_pgs
, lastmap
, nextmap
,
8380 handle
.reset_tp_timeout();
8382 service
.pg_update_epoch(pg
->info
.pgid
, lastmap
->get_epoch());
8383 pg
->handle_activate_map(rctx
);
8384 if (next_epoch
<= osd_epoch
) {
8385 dout(10) << __func__
<< " advanced to max " << max
8386 << " past min epoch " << min_epoch
8387 << " ... will requeue " << *pg
<< dendl
;
8393 void OSD::consume_map()
8395 assert(osd_lock
.is_locked());
8396 dout(7) << "consume_map version " << osdmap
->get_epoch() << dendl
;
8398 /** make sure the cluster is speaking in SORTBITWISE, because we don't
8399 * speak the older sorting version any more. Be careful not to force
8400 * a shutdown if we are merely processing old maps, though.
8402 if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
) && is_active()) {
8403 derr
<< __func__
<< " SORTBITWISE flag is not set" << dendl
;
8407 int num_pg_primary
= 0, num_pg_replica
= 0, num_pg_stray
= 0;
8408 list
<PGRef
> to_remove
;
8412 RWLock::RLocker
l(pg_map_lock
);
8413 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8416 PG
*pg
= it
->second
;
8418 if (pg
->is_primary())
8420 else if (pg
->is_replica())
8425 if (!osdmap
->have_pg_pool(pg
->info
.pgid
.pool())) {
8427 to_remove
.push_back(PGRef(pg
));
8429 service
.init_splits_between(it
->first
, service
.get_osdmap(), osdmap
);
8435 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
8436 for (auto pg
= pending_creates_from_osd
.cbegin();
8437 pg
!= pending_creates_from_osd
.cend();) {
8438 if (osdmap
->get_pg_acting_rank(pg
->first
, whoami
) < 0) {
8439 pg
= pending_creates_from_osd
.erase(pg
);
8446 for (list
<PGRef
>::iterator i
= to_remove
.begin();
8447 i
!= to_remove
.end();
8448 to_remove
.erase(i
++)) {
8449 RWLock::WLocker
locker(pg_map_lock
);
8455 service
.expand_pg_num(service
.get_osdmap(), osdmap
);
8457 service
.pre_publish_map(osdmap
);
8458 service
.await_reserved_maps();
8459 service
.publish_map(osdmap
);
8461 service
.maybe_inject_dispatch_delay();
8463 dispatch_sessions_waiting_on_map();
8465 service
.maybe_inject_dispatch_delay();
8467 // remove any PGs which we no longer host from the session waiting_for_pg lists
8468 dout(20) << __func__
<< " checking waiting_for_pg" << dendl
;
8469 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
8471 service
.maybe_inject_dispatch_delay();
8475 RWLock::RLocker
l(pg_map_lock
);
8476 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8479 PG
*pg
= it
->second
;
8481 pg
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
8485 logger
->set(l_osd_pg
, pg_map
.size());
8487 logger
->set(l_osd_pg_primary
, num_pg_primary
);
8488 logger
->set(l_osd_pg_replica
, num_pg_replica
);
8489 logger
->set(l_osd_pg_stray
, num_pg_stray
);
8490 logger
->set(l_osd_pg_removing
, remove_wq
.get_remove_queue_len());
8493 void OSD::activate_map()
8495 assert(osd_lock
.is_locked());
8497 dout(7) << "activate_map version " << osdmap
->get_epoch() << dendl
;
8499 if (osdmap
->test_flag(CEPH_OSDMAP_FULL
)) {
8500 dout(10) << " osdmap flagged full, doing onetime osdmap subscribe" << dendl
;
8501 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
8505 if (osdmap
->test_flag(CEPH_OSDMAP_NORECOVER
)) {
8506 if (!service
.recovery_is_paused()) {
8507 dout(1) << "pausing recovery (NORECOVER flag set)" << dendl
;
8508 service
.pause_recovery();
8511 if (service
.recovery_is_paused()) {
8512 dout(1) << "unpausing recovery (NORECOVER flag unset)" << dendl
;
8513 service
.unpause_recovery();
8517 service
.activate_map();
8520 take_waiters(waiting_for_osdmap
);
8523 bool OSD::require_mon_peer(const Message
*m
)
8525 if (!m
->get_connection()->peer_is_mon()) {
8526 dout(0) << "require_mon_peer received from non-mon "
8527 << m
->get_connection()->get_peer_addr()
8528 << " " << *m
<< dendl
;
8534 bool OSD::require_mon_or_mgr_peer(const Message
*m
)
8536 if (!m
->get_connection()->peer_is_mon() &&
8537 !m
->get_connection()->peer_is_mgr()) {
8538 dout(0) << "require_mon_or_mgr_peer received from non-mon, non-mgr "
8539 << m
->get_connection()->get_peer_addr()
8540 << " " << *m
<< dendl
;
8546 bool OSD::require_osd_peer(const Message
*m
)
8548 if (!m
->get_connection()->peer_is_osd()) {
8549 dout(0) << "require_osd_peer received from non-osd "
8550 << m
->get_connection()->get_peer_addr()
8551 << " " << *m
<< dendl
;
8557 bool OSD::require_self_aliveness(const Message
*m
, epoch_t epoch
)
8559 epoch_t up_epoch
= service
.get_up_epoch();
8560 if (epoch
< up_epoch
) {
8561 dout(7) << "from pre-up epoch " << epoch
<< " < " << up_epoch
<< dendl
;
8566 dout(7) << "still in boot state, dropping message " << *m
<< dendl
;
8573 bool OSD::require_same_peer_instance(const Message
*m
, OSDMapRef
& map
,
8574 bool is_fast_dispatch
)
8576 int from
= m
->get_source().num();
8578 if (map
->is_down(from
) ||
8579 (map
->get_cluster_addr(from
) != m
->get_source_inst().addr
)) {
8580 dout(5) << "from dead osd." << from
<< ", marking down, "
8581 << " msg was " << m
->get_source_inst().addr
8582 << " expected " << (map
->is_up(from
) ?
8583 map
->get_cluster_addr(from
) : entity_addr_t())
8585 ConnectionRef con
= m
->get_connection();
8587 Session
*s
= static_cast<Session
*>(con
->get_priv());
8589 if (!is_fast_dispatch
)
8590 s
->session_dispatch_lock
.Lock();
8591 clear_session_waiting_on_map(s
);
8592 con
->set_priv(NULL
); // break ref <-> session cycle, if any
8593 if (!is_fast_dispatch
)
8594 s
->session_dispatch_lock
.Unlock();
8604 * require that we have same (or newer) map, and that
8605 * the source is the pg primary.
8607 bool OSD::require_same_or_newer_map(OpRequestRef
& op
, epoch_t epoch
,
8608 bool is_fast_dispatch
)
8610 const Message
*m
= op
->get_req();
8611 dout(15) << "require_same_or_newer_map " << epoch
8612 << " (i am " << osdmap
->get_epoch() << ") " << m
<< dendl
;
8614 assert(osd_lock
.is_locked());
8616 // do they have a newer map?
8617 if (epoch
> osdmap
->get_epoch()) {
8618 dout(7) << "waiting for newer map epoch " << epoch
8619 << " > my " << osdmap
->get_epoch() << " with " << m
<< dendl
;
8620 wait_for_new_map(op
);
8624 if (!require_self_aliveness(op
->get_req(), epoch
)) {
8628 // ok, our map is same or newer.. do they still exist?
8629 if (m
->get_connection()->get_messenger() == cluster_messenger
&&
8630 !require_same_peer_instance(op
->get_req(), osdmap
, is_fast_dispatch
)) {
8641 // ----------------------------------------
8644 void OSD::split_pgs(
8646 const set
<spg_t
> &childpgids
, set
<PGRef
> *out_pgs
,
8649 PG::RecoveryCtx
*rctx
)
8651 unsigned pg_num
= nextmap
->get_pg_num(
8653 parent
->update_snap_mapper_bits(
8654 parent
->info
.pgid
.get_split_bits(pg_num
)
8657 vector
<object_stat_sum_t
> updated_stats(childpgids
.size() + 1);
8658 parent
->info
.stats
.stats
.sum
.split(updated_stats
);
8660 vector
<object_stat_sum_t
>::iterator stat_iter
= updated_stats
.begin();
8661 for (set
<spg_t
>::const_iterator i
= childpgids
.begin();
8662 i
!= childpgids
.end();
8664 assert(stat_iter
!= updated_stats
.end());
8665 dout(10) << "Splitting " << *parent
<< " into " << *i
<< dendl
;
8666 assert(service
.splitting(*i
));
8667 PG
* child
= _make_pg(nextmap
, *i
);
8669 out_pgs
->insert(child
);
8670 rctx
->created_pgs
.insert(child
);
8672 unsigned split_bits
= i
->get_split_bits(pg_num
);
8673 dout(10) << "pg_num is " << pg_num
<< dendl
;
8674 dout(10) << "m_seed " << i
->ps() << dendl
;
8675 dout(10) << "split_bits is " << split_bits
<< dendl
;
8677 parent
->split_colls(
8687 child
->info
.stats
.stats
.sum
= *stat_iter
;
8689 child
->write_if_dirty(*(rctx
->transaction
));
8692 assert(stat_iter
!= updated_stats
.end());
8693 parent
->info
.stats
.stats
.sum
= *stat_iter
;
8694 parent
->write_if_dirty(*(rctx
->transaction
));
8700 void OSD::handle_pg_create(OpRequestRef op
)
8702 const MOSDPGCreate
*m
= static_cast<const MOSDPGCreate
*>(op
->get_req());
8703 assert(m
->get_type() == MSG_OSD_PG_CREATE
);
8705 dout(10) << "handle_pg_create " << *m
<< dendl
;
8707 if (!require_mon_peer(op
->get_req())) {
8711 if (!require_same_or_newer_map(op
, m
->epoch
, false))
8716 map
<pg_t
,utime_t
>::const_iterator ci
= m
->ctimes
.begin();
8717 for (map
<pg_t
,pg_create_t
>::const_iterator p
= m
->mkpg
.begin();
8720 assert(ci
!= m
->ctimes
.end() && ci
->first
== p
->first
);
8721 epoch_t created
= p
->second
.created
;
8722 if (p
->second
.split_bits
) // Skip split pgs
8726 if (on
.preferred() >= 0) {
8727 dout(20) << "ignoring localized pg " << on
<< dendl
;
8731 if (!osdmap
->have_pg_pool(on
.pool())) {
8732 dout(20) << "ignoring pg on deleted pool " << on
<< dendl
;
8736 dout(20) << "mkpg " << on
<< " e" << created
<< "@" << ci
->second
<< dendl
;
8738 // is it still ours?
8739 vector
<int> up
, acting
;
8740 int up_primary
= -1;
8741 int acting_primary
= -1;
8742 osdmap
->pg_to_up_acting_osds(on
, &up
, &up_primary
, &acting
, &acting_primary
);
8743 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
8745 if (acting_primary
!= whoami
) {
8746 dout(10) << "mkpg " << on
<< " not acting_primary (" << acting_primary
8747 << "), my role=" << role
<< ", skipping" << dendl
;
8752 bool mapped
= osdmap
->get_primary_shard(on
, &pgid
);
8756 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
8758 pg_history_t history
;
8759 build_initial_pg_history(pgid
, created
, ci
->second
, &history
, &pi
);
8761 // The mon won't resend unless the primary changed, so
8762 // we ignore same_interval_since. We'll pass this history
8763 // to handle_pg_peering_evt with the current epoch as the
8764 // event -- the project_pg_history check in
8765 // handle_pg_peering_evt will be a noop.
8766 if (history
.same_primary_since
> m
->epoch
) {
8767 dout(10) << __func__
<< ": got obsolete pg create on pgid "
8768 << pgid
<< " from epoch " << m
->epoch
8769 << ", primary changed in " << history
.same_primary_since
8773 if (handle_pg_peering_evt(
8777 osdmap
->get_epoch(),
8778 PG::CephPeeringEvtRef(
8779 new PG::CephPeeringEvt(
8780 osdmap
->get_epoch(),
8781 osdmap
->get_epoch(),
8784 service
.send_pg_created(pgid
.pgid
);
8789 lock_guard
<mutex
> pending_creates_locker
{pending_creates_lock
};
8790 if (pending_creates_from_mon
== 0) {
8791 last_pg_create_epoch
= m
->epoch
;
8794 maybe_update_heartbeat_peers();
8798 // ----------------------------------------
8799 // peering and recovery
8801 PG::RecoveryCtx
OSD::create_context()
8803 ObjectStore::Transaction
*t
= new ObjectStore::Transaction
;
8804 C_Contexts
*on_applied
= new C_Contexts(cct
);
8805 C_Contexts
*on_safe
= new C_Contexts(cct
);
8806 map
<int, map
<spg_t
,pg_query_t
> > *query_map
=
8807 new map
<int, map
<spg_t
, pg_query_t
> >;
8808 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *notify_list
=
8809 new map
<int, vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8810 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *info_map
=
8811 new map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8812 PG::RecoveryCtx
rctx(query_map
, info_map
, notify_list
,
8813 on_applied
, on_safe
, t
);
8817 struct C_OpenPGs
: public Context
{
8821 C_OpenPGs(set
<PGRef
>& p
, ObjectStore
*s
, OSD
* o
) : store(s
), osd(o
) {
8824 void finish(int r
) override
{
8825 RWLock::RLocker
l(osd
->pg_map_lock
);
8826 for (auto p
: pgs
) {
8827 if (osd
->pg_map
.count(p
->info
.pgid
)) {
8828 p
->ch
= store
->open_collection(p
->coll
);
8835 void OSD::dispatch_context_transaction(PG::RecoveryCtx
&ctx
, PG
*pg
,
8836 ThreadPool::TPHandle
*handle
)
8838 if (!ctx
.transaction
->empty()) {
8839 if (!ctx
.created_pgs
.empty()) {
8840 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8842 int tr
= store
->queue_transaction(
8844 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
,
8845 TrackedOpRef(), handle
);
8846 delete (ctx
.transaction
);
8848 ctx
.transaction
= new ObjectStore::Transaction
;
8849 ctx
.on_applied
= new C_Contexts(cct
);
8850 ctx
.on_safe
= new C_Contexts(cct
);
8854 void OSD::dispatch_context(PG::RecoveryCtx
&ctx
, PG
*pg
, OSDMapRef curmap
,
8855 ThreadPool::TPHandle
*handle
)
8857 if (service
.get_osdmap()->is_up(whoami
) &&
8859 do_notifies(*ctx
.notify_list
, curmap
);
8860 do_queries(*ctx
.query_map
, curmap
);
8861 do_infos(*ctx
.info_map
, curmap
);
8863 delete ctx
.notify_list
;
8864 delete ctx
.query_map
;
8865 delete ctx
.info_map
;
8866 if ((ctx
.on_applied
->empty() &&
8867 ctx
.on_safe
->empty() &&
8868 ctx
.transaction
->empty() &&
8869 ctx
.created_pgs
.empty()) || !pg
) {
8870 delete ctx
.transaction
;
8871 delete ctx
.on_applied
;
8873 assert(ctx
.created_pgs
.empty());
8875 if (!ctx
.created_pgs
.empty()) {
8876 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8878 int tr
= store
->queue_transaction(
8880 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
, TrackedOpRef(),
8882 delete (ctx
.transaction
);
8888 * Send an MOSDPGNotify to a primary, with a list of PGs that I have
8889 * content for, and they are primary for.
8892 void OSD::do_notifies(
8893 map
<int,vector
<pair
<pg_notify_t
,PastIntervals
> > >& notify_list
,
8897 vector
<pair
<pg_notify_t
,PastIntervals
> > >::iterator it
=
8898 notify_list
.begin();
8899 it
!= notify_list
.end();
8901 if (!curmap
->is_up(it
->first
)) {
8902 dout(20) << __func__
<< " skipping down osd." << it
->first
<< dendl
;
8905 ConnectionRef con
= service
.get_con_osd_cluster(
8906 it
->first
, curmap
->get_epoch());
8908 dout(20) << __func__
<< " skipping osd." << it
->first
8909 << " (NULL con)" << dendl
;
8912 service
.share_map_peer(it
->first
, con
.get(), curmap
);
8913 dout(7) << __func__
<< " osd." << it
->first
8914 << " on " << it
->second
.size() << " PGs" << dendl
;
8915 MOSDPGNotify
*m
= new MOSDPGNotify(curmap
->get_epoch(),
8917 con
->send_message(m
);
8923 * send out pending queries for info | summaries
8925 void OSD::do_queries(map
<int, map
<spg_t
,pg_query_t
> >& query_map
,
8928 for (map
<int, map
<spg_t
,pg_query_t
> >::iterator pit
= query_map
.begin();
8929 pit
!= query_map
.end();
8931 if (!curmap
->is_up(pit
->first
)) {
8932 dout(20) << __func__
<< " skipping down osd." << pit
->first
<< dendl
;
8935 int who
= pit
->first
;
8936 ConnectionRef con
= service
.get_con_osd_cluster(who
, curmap
->get_epoch());
8938 dout(20) << __func__
<< " skipping osd." << who
8939 << " (NULL con)" << dendl
;
8942 service
.share_map_peer(who
, con
.get(), curmap
);
8943 dout(7) << __func__
<< " querying osd." << who
8944 << " on " << pit
->second
.size() << " PGs" << dendl
;
8945 MOSDPGQuery
*m
= new MOSDPGQuery(curmap
->get_epoch(), pit
->second
);
8946 con
->send_message(m
);
8951 void OSD::do_infos(map
<int,
8952 vector
<pair
<pg_notify_t
, PastIntervals
> > >& info_map
,
8956 vector
<pair
<pg_notify_t
, PastIntervals
> > >::iterator p
=
8958 p
!= info_map
.end();
8960 if (!curmap
->is_up(p
->first
)) {
8961 dout(20) << __func__
<< " skipping down osd." << p
->first
<< dendl
;
8964 for (vector
<pair
<pg_notify_t
,PastIntervals
> >::iterator i
= p
->second
.begin();
8965 i
!= p
->second
.end();
8967 dout(20) << __func__
<< " sending info " << i
->first
.info
8968 << " to shard " << p
->first
<< dendl
;
8970 ConnectionRef con
= service
.get_con_osd_cluster(
8971 p
->first
, curmap
->get_epoch());
8973 dout(20) << __func__
<< " skipping osd." << p
->first
8974 << " (NULL con)" << dendl
;
8977 service
.share_map_peer(p
->first
, con
.get(), curmap
);
8978 MOSDPGInfo
*m
= new MOSDPGInfo(curmap
->get_epoch());
8979 m
->pg_list
= p
->second
;
8980 con
->send_message(m
);
8987 * from non-primary to primary
8988 * includes pg_info_t.
8989 * NOTE: called with opqueue active.
8991 void OSD::handle_pg_notify(OpRequestRef op
)
8993 const MOSDPGNotify
*m
= static_cast<const MOSDPGNotify
*>(op
->get_req());
8994 assert(m
->get_type() == MSG_OSD_PG_NOTIFY
);
8996 dout(7) << "handle_pg_notify from " << m
->get_source() << dendl
;
8997 int from
= m
->get_source().num();
8999 if (!require_osd_peer(op
->get_req()))
9002 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9007 for (auto it
= m
->get_pg_list().begin();
9008 it
!= m
->get_pg_list().end();
9010 if (it
->first
.info
.pgid
.preferred() >= 0) {
9011 dout(20) << "ignoring localized pg " << it
->first
.info
.pgid
<< dendl
;
9015 handle_pg_peering_evt(
9016 spg_t(it
->first
.info
.pgid
.pgid
, it
->first
.to
),
9017 it
->first
.info
.history
, it
->second
,
9018 it
->first
.query_epoch
,
9019 PG::CephPeeringEvtRef(
9020 new PG::CephPeeringEvt(
9021 it
->first
.epoch_sent
, it
->first
.query_epoch
,
9022 PG::MNotifyRec(pg_shard_t(from
, it
->first
.from
), it
->first
,
9023 op
->get_req()->get_connection()->get_features())))
9028 void OSD::handle_pg_log(OpRequestRef op
)
9030 MOSDPGLog
*m
= static_cast<MOSDPGLog
*>(op
->get_nonconst_req());
9031 assert(m
->get_type() == MSG_OSD_PG_LOG
);
9032 dout(7) << "handle_pg_log " << *m
<< " from " << m
->get_source() << dendl
;
9034 if (!require_osd_peer(op
->get_req()))
9037 int from
= m
->get_source().num();
9038 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9041 if (m
->info
.pgid
.preferred() >= 0) {
9042 dout(10) << "ignoring localized pg " << m
->info
.pgid
<< dendl
;
9047 handle_pg_peering_evt(
9048 spg_t(m
->info
.pgid
.pgid
, m
->to
),
9049 m
->info
.history
, m
->past_intervals
, m
->get_epoch(),
9050 PG::CephPeeringEvtRef(
9051 new PG::CephPeeringEvt(
9052 m
->get_epoch(), m
->get_query_epoch(),
9053 PG::MLogRec(pg_shard_t(from
, m
->from
), m
)))
9057 void OSD::handle_pg_info(OpRequestRef op
)
9059 const MOSDPGInfo
*m
= static_cast<const MOSDPGInfo
*>(op
->get_req());
9060 assert(m
->get_type() == MSG_OSD_PG_INFO
);
9061 dout(7) << "handle_pg_info " << *m
<< " from " << m
->get_source() << dendl
;
9063 if (!require_osd_peer(op
->get_req()))
9066 int from
= m
->get_source().num();
9067 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9072 for (auto p
= m
->pg_list
.begin();
9073 p
!= m
->pg_list
.end();
9075 if (p
->first
.info
.pgid
.preferred() >= 0) {
9076 dout(10) << "ignoring localized pg " << p
->first
.info
.pgid
<< dendl
;
9080 handle_pg_peering_evt(
9081 spg_t(p
->first
.info
.pgid
.pgid
, p
->first
.to
),
9082 p
->first
.info
.history
, p
->second
, p
->first
.epoch_sent
,
9083 PG::CephPeeringEvtRef(
9084 new PG::CephPeeringEvt(
9085 p
->first
.epoch_sent
, p
->first
.query_epoch
,
9088 from
, p
->first
.from
), p
->first
.info
, p
->first
.epoch_sent
)))
9093 void OSD::handle_pg_trim(OpRequestRef op
)
9095 const MOSDPGTrim
*m
= static_cast<const MOSDPGTrim
*>(op
->get_req());
9096 assert(m
->get_type() == MSG_OSD_PG_TRIM
);
9098 dout(7) << "handle_pg_trim " << *m
<< " from " << m
->get_source() << dendl
;
9100 if (!require_osd_peer(op
->get_req()))
9103 int from
= m
->get_source().num();
9104 if (!require_same_or_newer_map(op
, m
->epoch
, false))
9107 if (m
->pgid
.preferred() >= 0) {
9108 dout(10) << "ignoring localized pg " << m
->pgid
<< dendl
;
9114 PG
*pg
= _lookup_lock_pg(m
->pgid
);
9116 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
9120 if (m
->epoch
< pg
->info
.history
.same_interval_since
) {
9121 dout(10) << *pg
<< " got old trim to " << m
->trim_to
<< ", ignoring" << dendl
;
9126 if (pg
->is_primary()) {
9127 // peer is informing us of their last_complete_ondisk
9128 dout(10) << *pg
<< " replica osd." << from
<< " lcod " << m
->trim_to
<< dendl
;
9129 pg
->peer_last_complete_ondisk
[pg_shard_t(from
, m
->pgid
.shard
)] =
9131 // trim log when the pg is recovered
9132 pg
->calc_min_last_complete_ondisk();
9134 // primary is instructing us to trim
9135 ObjectStore::Transaction t
;
9136 pg
->pg_log
.trim(m
->trim_to
, pg
->info
);
9137 pg
->dirty_info
= true;
9138 pg
->write_if_dirty(t
);
9139 int tr
= store
->queue_transaction(pg
->osr
.get(), std::move(t
), NULL
);
9145 void OSD::handle_pg_backfill_reserve(OpRequestRef op
)
9147 const MBackfillReserve
*m
= static_cast<const MBackfillReserve
*>(op
->get_req());
9148 assert(m
->get_type() == MSG_OSD_BACKFILL_RESERVE
);
9150 if (!require_osd_peer(op
->get_req()))
9152 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
9155 PG::CephPeeringEvtRef evt
;
9156 if (m
->type
== MBackfillReserve::REQUEST
) {
9157 evt
= PG::CephPeeringEvtRef(
9158 new PG::CephPeeringEvt(
9161 PG::RequestBackfillPrio(m
->priority
)));
9162 } else if (m
->type
== MBackfillReserve::GRANT
) {
9163 evt
= PG::CephPeeringEvtRef(
9164 new PG::CephPeeringEvt(
9167 PG::RemoteBackfillReserved()));
9168 } else if (m
->type
== MBackfillReserve::REJECT
) {
9169 // NOTE: this is replica -> primary "i reject your request"
9170 // and also primary -> replica "cancel my previously-granted request"
9171 evt
= PG::CephPeeringEvtRef(
9172 new PG::CephPeeringEvt(
9175 PG::RemoteReservationRejected()));
9180 if (service
.splitting(m
->pgid
)) {
9181 peering_wait_for_split
[m
->pgid
].push_back(evt
);
9185 PG
*pg
= _lookup_lock_pg(m
->pgid
);
9187 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
9191 pg
->queue_peering_event(evt
);
9195 void OSD::handle_pg_recovery_reserve(OpRequestRef op
)
9197 const MRecoveryReserve
*m
= static_cast<const MRecoveryReserve
*>(op
->get_req());
9198 assert(m
->get_type() == MSG_OSD_RECOVERY_RESERVE
);
9200 if (!require_osd_peer(op
->get_req()))
9202 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
9205 PG::CephPeeringEvtRef evt
;
9206 if (m
->type
== MRecoveryReserve::REQUEST
) {
9207 evt
= PG::CephPeeringEvtRef(
9208 new PG::CephPeeringEvt(
9211 PG::RequestRecovery()));
9212 } else if (m
->type
== MRecoveryReserve::GRANT
) {
9213 evt
= PG::CephPeeringEvtRef(
9214 new PG::CephPeeringEvt(
9217 PG::RemoteRecoveryReserved()));
9218 } else if (m
->type
== MRecoveryReserve::RELEASE
) {
9219 evt
= PG::CephPeeringEvtRef(
9220 new PG::CephPeeringEvt(
9223 PG::RecoveryDone()));
9228 if (service
.splitting(m
->pgid
)) {
9229 peering_wait_for_split
[m
->pgid
].push_back(evt
);
9233 PG
*pg
= _lookup_lock_pg(m
->pgid
);
9235 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
9239 pg
->queue_peering_event(evt
);
9243 void OSD::handle_force_recovery(Message
*m
)
9245 MOSDForceRecovery
*msg
= static_cast<MOSDForceRecovery
*>(m
);
9246 assert(msg
->get_type() == MSG_OSD_FORCE_RECOVERY
);
9248 vector
<PGRef
> local_pgs
;
9249 local_pgs
.reserve(msg
->forced_pgs
.size());
9252 RWLock::RLocker
l(pg_map_lock
);
9253 for (auto& i
: msg
->forced_pgs
) {
9255 if (osdmap
->get_primary_shard(i
, &locpg
)) {
9256 auto pg_map_entry
= pg_map
.find(locpg
);
9257 if (pg_map_entry
!= pg_map
.end()) {
9258 local_pgs
.push_back(pg_map_entry
->second
);
9264 if (local_pgs
.size()) {
9265 service
.adjust_pg_priorities(local_pgs
, msg
->options
);
9272 * from primary to replica | stray
9273 * NOTE: called with opqueue active.
9275 void OSD::handle_pg_query(OpRequestRef op
)
9277 assert(osd_lock
.is_locked());
9279 const MOSDPGQuery
*m
= static_cast<const MOSDPGQuery
*>(op
->get_req());
9280 assert(m
->get_type() == MSG_OSD_PG_QUERY
);
9282 if (!require_osd_peer(op
->get_req()))
9285 dout(7) << "handle_pg_query from " << m
->get_source() << " epoch " << m
->get_epoch() << dendl
;
9286 int from
= m
->get_source().num();
9288 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9293 map
< int, vector
<pair
<pg_notify_t
, PastIntervals
> > > notify_list
;
9295 for (auto it
= m
->pg_list
.begin();
9296 it
!= m
->pg_list
.end();
9298 spg_t pgid
= it
->first
;
9300 if (pgid
.preferred() >= 0) {
9301 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9305 if (service
.splitting(pgid
)) {
9306 peering_wait_for_split
[pgid
].push_back(
9307 PG::CephPeeringEvtRef(
9308 new PG::CephPeeringEvt(
9309 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9310 PG::MQuery(pg_shard_t(from
, it
->second
.from
),
9311 it
->second
, it
->second
.epoch_sent
))));
9316 RWLock::RLocker
l(pg_map_lock
);
9317 if (pg_map
.count(pgid
)) {
9319 pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9321 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9322 pg_shard_t(from
, it
->second
.from
), it
->second
);
9328 if (!osdmap
->have_pg_pool(pgid
.pool()))
9331 // get active crush mapping
9332 int up_primary
, acting_primary
;
9333 vector
<int> up
, acting
;
9334 osdmap
->pg_to_up_acting_osds(
9335 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9338 pg_history_t history
= it
->second
.history
;
9339 bool valid_history
= project_pg_history(
9340 pgid
, history
, it
->second
.epoch_sent
,
9341 up
, up_primary
, acting
, acting_primary
);
9343 if (!valid_history
||
9344 it
->second
.epoch_sent
< history
.same_interval_since
) {
9345 dout(10) << " pg " << pgid
<< " dne, and pg has changed in "
9346 << history
.same_interval_since
9347 << " (msg from " << it
->second
.epoch_sent
<< ")" << dendl
;
9351 dout(10) << " pg " << pgid
<< " dne" << dendl
;
9352 pg_info_t
empty(spg_t(pgid
.pgid
, it
->second
.to
));
9353 /* This is racy, but that should be ok: if we complete the deletion
9354 * before the pg is recreated, we'll just start it off backfilling
9355 * instead of just empty */
9356 if (service
.deleting_pgs
.lookup(pgid
))
9357 empty
.set_last_backfill(hobject_t());
9358 if (it
->second
.type
== pg_query_t::LOG
||
9359 it
->second
.type
== pg_query_t::FULLLOG
) {
9360 ConnectionRef con
= service
.get_con_osd_cluster(from
, osdmap
->get_epoch());
9362 MOSDPGLog
*mlog
= new MOSDPGLog(
9363 it
->second
.from
, it
->second
.to
,
9364 osdmap
->get_epoch(), empty
,
9365 it
->second
.epoch_sent
);
9366 service
.share_map_peer(from
, con
.get(), osdmap
);
9367 con
->send_message(mlog
);
9370 notify_list
[from
].push_back(
9373 it
->second
.from
, it
->second
.to
,
9374 it
->second
.epoch_sent
,
9375 osdmap
->get_epoch(),
9378 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
9382 do_notifies(notify_list
, osdmap
);
9386 void OSD::handle_pg_remove(OpRequestRef op
)
9388 const MOSDPGRemove
*m
= static_cast<const MOSDPGRemove
*>(op
->get_req());
9389 assert(m
->get_type() == MSG_OSD_PG_REMOVE
);
9390 assert(osd_lock
.is_locked());
9392 if (!require_osd_peer(op
->get_req()))
9395 dout(7) << "handle_pg_remove from " << m
->get_source() << " on "
9396 << m
->pg_list
.size() << " pgs" << dendl
;
9398 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9403 for (auto it
= m
->pg_list
.begin();
9404 it
!= m
->pg_list
.end();
9407 if (pgid
.preferred() >= 0) {
9408 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9412 RWLock::WLocker
l(pg_map_lock
);
9413 if (pg_map
.count(pgid
) == 0) {
9414 dout(10) << " don't have pg " << pgid
<< dendl
;
9417 dout(5) << "queue_pg_for_deletion: " << pgid
<< dendl
;
9418 PG
*pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9419 pg_history_t history
= pg
->info
.history
;
9420 int up_primary
, acting_primary
;
9421 vector
<int> up
, acting
;
9422 osdmap
->pg_to_up_acting_osds(
9423 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9424 bool valid_history
= project_pg_history(
9425 pg
->info
.pgid
, history
, pg
->get_osdmap()->get_epoch(),
9426 up
, up_primary
, acting
, acting_primary
);
9427 if (valid_history
&&
9428 history
.same_interval_since
<= m
->get_epoch()) {
9429 assert(pg
->get_primary().osd
== m
->get_source().num());
9434 dout(10) << *pg
<< " ignoring remove request, pg changed in epoch "
9435 << history
.same_interval_since
9436 << " > " << m
->get_epoch() << dendl
;
9442 void OSD::_remove_pg(PG
*pg
)
9444 ObjectStore::Transaction rmt
;
9446 // on_removal, which calls remove_watchers_and_notifies, and the erasure from
9447 // the pg_map must be done together without unlocking the pg lock,
9448 // to avoid racing with watcher cleanup in ms_handle_reset
9449 // and handle_notify_timeout
9450 pg
->on_removal(&rmt
);
9452 service
.cancel_pending_splits_for_parent(pg
->info
.pgid
);
9453 int tr
= store
->queue_transaction(
9454 pg
->osr
.get(), std::move(rmt
), NULL
,
9455 new ContainerContext
<
9456 SequencerRef
>(pg
->osr
));
9459 DeletingStateRef deleting
= service
.deleting_pgs
.lookup_or_create(
9465 remove_wq
.queue(make_pair(PGRef(pg
), deleting
));
9467 service
.pg_remove_epoch(pg
->info
.pgid
);
9469 // dereference from op_wq
9470 op_shardedwq
.clear_pg_pointer(pg
->info
.pgid
);
9473 pg_map
.erase(pg
->info
.pgid
);
9474 pg
->put("PGMap"); // since we've taken it out of map
9477 // =========================================================
9480 void OSDService::_maybe_queue_recovery() {
9481 assert(recovery_lock
.is_locked_by_me());
9482 uint64_t available_pushes
;
9483 while (!awaiting_throttle
.empty() &&
9484 _recover_now(&available_pushes
)) {
9485 uint64_t to_start
= MIN(
9487 cct
->_conf
->osd_recovery_max_single_start
);
9488 _queue_for_recovery(awaiting_throttle
.front(), to_start
);
9489 awaiting_throttle
.pop_front();
9490 recovery_ops_reserved
+= to_start
;
9494 bool OSDService::_recover_now(uint64_t *available_pushes
)
9496 if (available_pushes
)
9497 *available_pushes
= 0;
9499 if (ceph_clock_now() < defer_recovery_until
) {
9500 dout(15) << __func__
<< " defer until " << defer_recovery_until
<< dendl
;
9504 if (recovery_paused
) {
9505 dout(15) << __func__
<< " paused" << dendl
;
9509 uint64_t max
= cct
->_conf
->osd_recovery_max_active
;
9510 if (max
<= recovery_ops_active
+ recovery_ops_reserved
) {
9511 dout(15) << __func__
<< " active " << recovery_ops_active
9512 << " + reserved " << recovery_ops_reserved
9513 << " >= max " << max
<< dendl
;
9517 if (available_pushes
)
9518 *available_pushes
= max
- recovery_ops_active
- recovery_ops_reserved
;
9524 void OSDService::adjust_pg_priorities(const vector
<PGRef
>& pgs
, int newflags
)
9526 if (!pgs
.size() || !(newflags
& (OFR_BACKFILL
| OFR_RECOVERY
)))
9530 if (newflags
& OFR_BACKFILL
) {
9531 newstate
= PG_STATE_FORCED_BACKFILL
;
9532 } else if (newflags
& OFR_RECOVERY
) {
9533 newstate
= PG_STATE_FORCED_RECOVERY
;
9536 // debug output here may get large, don't generate it if debug level is below
9537 // 10 and use abbreviated pg ids otherwise
9538 if ((cct
)->_conf
->subsys
.should_gather(ceph_subsys_osd
, 10)) {
9541 for (auto& i
: pgs
) {
9542 ss
<< i
->get_pgid() << " ";
9545 dout(10) << __func__
<< " working on " << ss
.str() << dendl
;
9548 if (newflags
& OFR_CANCEL
) {
9549 for (auto& i
: pgs
) {
9551 i
->_change_recovery_force_mode(newstate
, true);
9555 for (auto& i
: pgs
) {
9556 // make sure the PG is in correct state before forcing backfill or recovery, or
9557 // else we'll make PG keeping FORCE_* flag forever, requiring osds restart
9558 // or forcing somehow recovery/backfill.
9560 int pgstate
= i
->get_state();
9561 if ( ((newstate
== PG_STATE_FORCED_RECOVERY
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_RECOVERY_WAIT
| PG_STATE_RECOVERING
))) ||
9562 ((newstate
== PG_STATE_FORCED_BACKFILL
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_BACKFILL_WAIT
| PG_STATE_BACKFILLING
))) )
9563 i
->_change_recovery_force_mode(newstate
, false);
9569 void OSD::do_recovery(
9570 PG
*pg
, epoch_t queued
, uint64_t reserved_pushes
,
9571 ThreadPool::TPHandle
&handle
)
9573 uint64_t started
= 0;
9576 * When the value of osd_recovery_sleep is set greater than zero, recovery
9577 * ops are scheduled after osd_recovery_sleep amount of time from the previous
9578 * recovery event's schedule time. This is done by adding a
9579 * recovery_requeue_callback event, which re-queues the recovery op using
9580 * queue_recovery_after_sleep.
9582 float recovery_sleep
= get_osd_recovery_sleep();
9584 Mutex::Locker
l(service
.recovery_sleep_lock
);
9585 if (recovery_sleep
> 0 && service
.recovery_needs_sleep
) {
9587 auto recovery_requeue_callback
= new FunctionContext([this, pgref
, queued
, reserved_pushes
](int r
) {
9588 dout(20) << "do_recovery wake up at "
9590 << ", re-queuing recovery" << dendl
;
9591 Mutex::Locker
l(service
.recovery_sleep_lock
);
9592 service
.recovery_needs_sleep
= false;
9593 service
.queue_recovery_after_sleep(pgref
.get(), queued
, reserved_pushes
);
9596 // This is true for the first recovery op and when the previous recovery op
9597 // has been scheduled in the past. The next recovery op is scheduled after
9598 // completing the sleep from now.
9599 if (service
.recovery_schedule_time
< ceph_clock_now()) {
9600 service
.recovery_schedule_time
= ceph_clock_now();
9602 service
.recovery_schedule_time
+= recovery_sleep
;
9603 service
.recovery_sleep_timer
.add_event_at(service
.recovery_schedule_time
,
9604 recovery_requeue_callback
);
9605 dout(20) << "Recovery event scheduled at "
9606 << service
.recovery_schedule_time
<< dendl
;
9613 Mutex::Locker
l(service
.recovery_sleep_lock
);
9614 service
.recovery_needs_sleep
= true;
9617 if (pg
->pg_has_reset_since(queued
)) {
9621 assert(!pg
->deleting
);
9622 assert(pg
->is_peered() && pg
->is_primary());
9624 assert(pg
->recovery_queued
);
9625 pg
->recovery_queued
= false;
9627 dout(10) << "do_recovery starting " << reserved_pushes
<< " " << *pg
<< dendl
;
9628 #ifdef DEBUG_RECOVERY_OIDS
9629 dout(20) << " active was " << service
.recovery_oids
[pg
->info
.pgid
] << dendl
;
9632 bool more
= pg
->start_recovery_ops(reserved_pushes
, handle
, &started
);
9633 dout(10) << "do_recovery started " << started
<< "/" << reserved_pushes
9634 << " on " << *pg
<< dendl
;
9636 // If no recovery op is started, don't bother to manipulate the RecoveryCtx
9637 if (!started
&& (more
|| !pg
->have_unfound())) {
9641 PG::RecoveryCtx rctx
= create_context();
9642 rctx
.handle
= &handle
;
9645 * if we couldn't start any recovery ops and things are still
9646 * unfound, see if we can discover more missing object locations.
9647 * It may be that our initial locations were bad and we errored
9648 * out while trying to pull.
9650 if (!more
&& pg
->have_unfound()) {
9651 pg
->discover_all_missing(*rctx
.query_map
);
9652 if (rctx
.query_map
->empty()) {
9654 if (pg
->state_test(PG_STATE_BACKFILLING
)) {
9655 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9658 PG::DeferBackfill(cct
->_conf
->osd_recovery_retry_interval
)));
9659 pg
->queue_peering_event(evt
);
9660 action
= "in backfill";
9661 } else if (pg
->state_test(PG_STATE_RECOVERING
)) {
9662 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9665 PG::DeferRecovery(cct
->_conf
->osd_recovery_retry_interval
)));
9666 pg
->queue_peering_event(evt
);
9667 action
= "in recovery";
9669 action
= "already out of recovery/backfill";
9671 dout(10) << __func__
<< ": no luck, giving up on this pg for now (" << action
<< ")" << dendl
;
9673 dout(10) << __func__
<< ": no luck, giving up on this pg for now (queue_recovery)" << dendl
;
9674 pg
->queue_recovery();
9678 pg
->write_if_dirty(*rctx
.transaction
);
9679 OSDMapRef curmap
= pg
->get_osdmap();
9680 dispatch_context(rctx
, pg
, curmap
);
9684 assert(started
<= reserved_pushes
);
9685 service
.release_reserved_pushes(reserved_pushes
);
9688 void OSDService::start_recovery_op(PG
*pg
, const hobject_t
& soid
)
9690 Mutex::Locker
l(recovery_lock
);
9691 dout(10) << "start_recovery_op " << *pg
<< " " << soid
9692 << " (" << recovery_ops_active
<< "/"
9693 << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9695 recovery_ops_active
++;
9697 #ifdef DEBUG_RECOVERY_OIDS
9698 dout(20) << " active was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9699 assert(recovery_oids
[pg
->info
.pgid
].count(soid
) == 0);
9700 recovery_oids
[pg
->info
.pgid
].insert(soid
);
9704 void OSDService::finish_recovery_op(PG
*pg
, const hobject_t
& soid
, bool dequeue
)
9706 Mutex::Locker
l(recovery_lock
);
9707 dout(10) << "finish_recovery_op " << *pg
<< " " << soid
9708 << " dequeue=" << dequeue
9709 << " (" << recovery_ops_active
<< "/" << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9713 assert(recovery_ops_active
> 0);
9714 recovery_ops_active
--;
9716 #ifdef DEBUG_RECOVERY_OIDS
9717 dout(20) << " active oids was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9718 assert(recovery_oids
[pg
->info
.pgid
].count(soid
));
9719 recovery_oids
[pg
->info
.pgid
].erase(soid
);
9722 _maybe_queue_recovery();
9725 bool OSDService::is_recovery_active()
9727 return local_reserver
.has_reservation() || remote_reserver
.has_reservation();
9730 // =========================================================
9733 bool OSD::op_is_discardable(const MOSDOp
*op
)
9735 // drop client request if they are not connected and can't get the
9737 if (!op
->get_connection()->is_connected()) {
9743 void OSD::enqueue_op(spg_t pg
, OpRequestRef
& op
, epoch_t epoch
)
9745 utime_t latency
= ceph_clock_now() - op
->get_req()->get_recv_stamp();
9746 dout(15) << "enqueue_op " << op
<< " prio " << op
->get_req()->get_priority()
9747 << " cost " << op
->get_req()->get_cost()
9748 << " latency " << latency
9749 << " epoch " << epoch
9750 << " " << *(op
->get_req()) << dendl
;
9751 op
->osd_trace
.event("enqueue op");
9752 op
->osd_trace
.keyval("priority", op
->get_req()->get_priority());
9753 op
->osd_trace
.keyval("cost", op
->get_req()->get_cost());
9754 op
->mark_queued_for_pg();
9755 logger
->tinc(l_osd_op_before_queue_op_lat
, latency
);
9756 op_shardedwq
.queue(make_pair(pg
, PGQueueable(op
, epoch
)));
9762 * NOTE: dequeue called in worker thread, with pg lock
9764 void OSD::dequeue_op(
9765 PGRef pg
, OpRequestRef op
,
9766 ThreadPool::TPHandle
&handle
)
9769 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_BEGIN", false);
9771 utime_t now
= ceph_clock_now();
9772 op
->set_dequeued_time(now
);
9773 utime_t latency
= now
- op
->get_req()->get_recv_stamp();
9774 dout(10) << "dequeue_op " << op
<< " prio " << op
->get_req()->get_priority()
9775 << " cost " << op
->get_req()->get_cost()
9776 << " latency " << latency
9777 << " " << *(op
->get_req())
9778 << " pg " << *pg
<< dendl
;
9780 logger
->tinc(l_osd_op_before_dequeue_op_lat
, latency
);
9782 Session
*session
= static_cast<Session
*>(
9783 op
->get_req()->get_connection()->get_priv());
9785 maybe_share_map(session
, op
, pg
->get_osdmap());
9792 op
->mark_reached_pg();
9793 op
->osd_trace
.event("dequeue_op");
9795 pg
->do_request(op
, handle
);
9798 dout(10) << "dequeue_op " << op
<< " finish" << dendl
;
9799 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_END", false);
9803 struct C_CompleteSplits
: public Context
{
9806 C_CompleteSplits(OSD
*osd
, const set
<PGRef
> &in
)
9807 : osd(osd
), pgs(in
) {}
9808 void finish(int r
) override
{
9809 Mutex::Locker
l(osd
->osd_lock
);
9810 if (osd
->is_stopping())
9812 PG::RecoveryCtx rctx
= osd
->create_context();
9813 for (set
<PGRef
>::iterator i
= pgs
.begin();
9816 osd
->pg_map_lock
.get_write();
9819 osd
->add_newly_split_pg(pg
, &rctx
);
9820 if (!((*i
)->deleting
)) {
9821 set
<spg_t
> to_complete
;
9822 to_complete
.insert((*i
)->info
.pgid
);
9823 osd
->service
.complete_split(to_complete
);
9825 osd
->pg_map_lock
.put_write();
9826 osd
->dispatch_context_transaction(rctx
, pg
);
9827 osd
->wake_pg_waiters(*i
);
9831 osd
->dispatch_context(rctx
, 0, osd
->service
.get_osdmap());
9835 void OSD::process_peering_events(
9836 const list
<PG
*> &pgs
,
9837 ThreadPool::TPHandle
&handle
9840 bool need_up_thru
= false;
9841 epoch_t same_interval_since
= 0;
9843 PG::RecoveryCtx rctx
= create_context();
9844 rctx
.handle
= &handle
;
9845 for (list
<PG
*>::const_iterator i
= pgs
.begin();
9848 set
<PGRef
> split_pgs
;
9850 pg
->lock_suspend_timeout(handle
);
9851 curmap
= service
.get_osdmap();
9856 if (!advance_pg(curmap
->get_epoch(), pg
, handle
, &rctx
, &split_pgs
)) {
9857 // we need to requeue the PG explicitly since we didn't actually
9859 peering_wq
.queue(pg
);
9861 assert(!pg
->peering_queue
.empty());
9862 PG::CephPeeringEvtRef evt
= pg
->peering_queue
.front();
9863 pg
->peering_queue
.pop_front();
9864 pg
->handle_peering_event(evt
, &rctx
);
9866 need_up_thru
= pg
->need_up_thru
|| need_up_thru
;
9867 same_interval_since
= MAX(pg
->info
.history
.same_interval_since
,
9868 same_interval_since
);
9869 pg
->write_if_dirty(*rctx
.transaction
);
9870 if (!split_pgs
.empty()) {
9871 rctx
.on_applied
->add(new C_CompleteSplits(this, split_pgs
));
9874 dispatch_context_transaction(rctx
, pg
, &handle
);
9878 queue_want_up_thru(same_interval_since
);
9879 dispatch_context(rctx
, 0, curmap
, &handle
);
9881 service
.send_pg_temp();
9884 // --------------------------------
9886 const char** OSD::get_tracked_conf_keys() const
9888 static const char* KEYS
[] = {
9889 "osd_max_backfills",
9890 "osd_min_recovery_priority",
9891 "osd_max_trimming_pgs",
9892 "osd_op_complaint_time",
9893 "osd_op_log_threshold",
9894 "osd_op_history_size",
9895 "osd_op_history_duration",
9896 "osd_op_history_slow_op_size",
9897 "osd_op_history_slow_op_threshold",
9898 "osd_enable_op_tracker",
9899 "osd_map_cache_size",
9900 "osd_map_max_advance",
9901 "osd_pg_epoch_persisted_max_stale",
9902 "osd_disk_thread_ioprio_class",
9903 "osd_disk_thread_ioprio_priority",
9904 // clog & admin clog
9907 "clog_to_syslog_facility",
9908 "clog_to_syslog_level",
9909 "osd_objectstore_fuse",
9911 "clog_to_graylog_host",
9912 "clog_to_graylog_port",
9915 "osd_recovery_delay_start",
9916 "osd_client_message_size_cap",
9917 "osd_client_message_cap",
9918 "osd_heartbeat_min_size",
9919 "osd_heartbeat_interval",
9925 void OSD::handle_conf_change(const struct md_config_t
*conf
,
9926 const std::set
<std::string
> &changed
)
9928 if (changed
.count("osd_max_backfills")) {
9929 service
.local_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9930 service
.remote_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9932 if (changed
.count("osd_min_recovery_priority")) {
9933 service
.local_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9934 service
.remote_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9936 if (changed
.count("osd_max_trimming_pgs")) {
9937 service
.snap_reserver
.set_max(cct
->_conf
->osd_max_trimming_pgs
);
9939 if (changed
.count("osd_op_complaint_time") ||
9940 changed
.count("osd_op_log_threshold")) {
9941 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
9942 cct
->_conf
->osd_op_log_threshold
);
9944 if (changed
.count("osd_op_history_size") ||
9945 changed
.count("osd_op_history_duration")) {
9946 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
9947 cct
->_conf
->osd_op_history_duration
);
9949 if (changed
.count("osd_op_history_slow_op_size") ||
9950 changed
.count("osd_op_history_slow_op_threshold")) {
9951 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
9952 cct
->_conf
->osd_op_history_slow_op_threshold
);
9954 if (changed
.count("osd_enable_op_tracker")) {
9955 op_tracker
.set_tracking(cct
->_conf
->osd_enable_op_tracker
);
9957 if (changed
.count("osd_disk_thread_ioprio_class") ||
9958 changed
.count("osd_disk_thread_ioprio_priority")) {
9959 set_disk_tp_priority();
9961 if (changed
.count("osd_map_cache_size")) {
9962 service
.map_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9963 service
.map_bl_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9964 service
.map_bl_inc_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9966 if (changed
.count("clog_to_monitors") ||
9967 changed
.count("clog_to_syslog") ||
9968 changed
.count("clog_to_syslog_level") ||
9969 changed
.count("clog_to_syslog_facility") ||
9970 changed
.count("clog_to_graylog") ||
9971 changed
.count("clog_to_graylog_host") ||
9972 changed
.count("clog_to_graylog_port") ||
9973 changed
.count("host") ||
9974 changed
.count("fsid")) {
9975 update_log_config();
9979 if (changed
.count("osd_objectstore_fuse")) {
9981 enable_disable_fuse(false);
9986 if (changed
.count("osd_recovery_delay_start")) {
9987 service
.defer_recovery(cct
->_conf
->osd_recovery_delay_start
);
9988 service
.kick_recovery_queue();
9991 if (changed
.count("osd_client_message_cap")) {
9992 uint64_t newval
= cct
->_conf
->osd_client_message_cap
;
9993 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
9994 if (pol
.throttler_messages
&& newval
> 0) {
9995 pol
.throttler_messages
->reset_max(newval
);
9998 if (changed
.count("osd_client_message_size_cap")) {
9999 uint64_t newval
= cct
->_conf
->osd_client_message_size_cap
;
10000 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
10001 if (pol
.throttler_bytes
&& newval
> 0) {
10002 pol
.throttler_bytes
->reset_max(newval
);
10009 void OSD::update_log_config()
10011 map
<string
,string
> log_to_monitors
;
10012 map
<string
,string
> log_to_syslog
;
10013 map
<string
,string
> log_channel
;
10014 map
<string
,string
> log_prio
;
10015 map
<string
,string
> log_to_graylog
;
10016 map
<string
,string
> log_to_graylog_host
;
10017 map
<string
,string
> log_to_graylog_port
;
10021 if (parse_log_client_options(cct
, log_to_monitors
, log_to_syslog
,
10022 log_channel
, log_prio
, log_to_graylog
,
10023 log_to_graylog_host
, log_to_graylog_port
,
10025 clog
->update_config(log_to_monitors
, log_to_syslog
,
10026 log_channel
, log_prio
, log_to_graylog
,
10027 log_to_graylog_host
, log_to_graylog_port
,
10029 derr
<< "log_to_monitors " << log_to_monitors
<< dendl
;
10032 void OSD::check_config()
10034 // some sanity checks
10035 if (cct
->_conf
->osd_map_cache_size
<= cct
->_conf
->osd_map_max_advance
+ 2) {
10036 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
10037 << " is not > osd_map_max_advance ("
10038 << cct
->_conf
->osd_map_max_advance
<< ")";
10040 if (cct
->_conf
->osd_map_cache_size
<= (int)cct
->_conf
->osd_pg_epoch_persisted_max_stale
+ 2) {
10041 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
10042 << " is not > osd_pg_epoch_persisted_max_stale ("
10043 << cct
->_conf
->osd_pg_epoch_persisted_max_stale
<< ")";
10047 void OSD::set_disk_tp_priority()
10049 dout(10) << __func__
10050 << " class " << cct
->_conf
->osd_disk_thread_ioprio_class
10051 << " priority " << cct
->_conf
->osd_disk_thread_ioprio_priority
10053 if (cct
->_conf
->osd_disk_thread_ioprio_class
.empty() ||
10054 cct
->_conf
->osd_disk_thread_ioprio_priority
< 0)
10057 ceph_ioprio_string_to_class(cct
->_conf
->osd_disk_thread_ioprio_class
);
10059 derr
<< __func__
<< cpp_strerror(cls
) << ": "
10060 << "osd_disk_thread_ioprio_class is " << cct
->_conf
->osd_disk_thread_ioprio_class
10061 << " but only the following values are allowed: idle, be or rt" << dendl
;
10063 disk_tp
.set_ioprio(cls
, cct
->_conf
->osd_disk_thread_ioprio_priority
);
10066 // --------------------------------
10068 void OSD::get_latest_osdmap()
10070 dout(10) << __func__
<< " -- start" << dendl
;
10073 service
.objecter
->wait_for_latest_osdmap(&cond
);
10076 dout(10) << __func__
<< " -- finish" << dendl
;
10079 // --------------------------------
10081 int OSD::init_op_flags(OpRequestRef
& op
)
10083 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
10084 vector
<OSDOp
>::const_iterator iter
;
10086 // client flags have no bearing on whether an op is a read, write, etc.
10089 if (m
->has_flag(CEPH_OSD_FLAG_RWORDERED
)) {
10090 op
->set_force_rwordered();
10093 // set bits based on op codes, called methods.
10094 for (iter
= m
->ops
.begin(); iter
!= m
->ops
.end(); ++iter
) {
10095 if ((iter
->op
.op
== CEPH_OSD_OP_WATCH
&&
10096 iter
->op
.watch
.op
== CEPH_OSD_WATCH_OP_PING
)) {
10097 /* This a bit odd. PING isn't actually a write. It can't
10098 * result in an update to the object_info. PINGs also aren'ty
10099 * resent, so there's no reason to write out a log entry
10101 * However, we pipeline them behind writes, so let's force
10102 * the write_ordered flag.
10104 op
->set_force_rwordered();
10106 if (ceph_osd_op_mode_modify(iter
->op
.op
))
10109 if (ceph_osd_op_mode_read(iter
->op
.op
))
10112 // set READ flag if there are src_oids
10113 if (iter
->soid
.oid
.name
.length())
10116 // set PGOP flag if there are PG ops
10117 if (ceph_osd_op_type_pg(iter
->op
.op
))
10120 if (ceph_osd_op_mode_cache(iter
->op
.op
))
10123 // check for ec base pool
10124 int64_t poolid
= m
->get_pg().pool();
10125 const pg_pool_t
*pool
= osdmap
->get_pg_pool(poolid
);
10126 if (pool
&& pool
->is_tier()) {
10127 const pg_pool_t
*base_pool
= osdmap
->get_pg_pool(pool
->tier_of
);
10128 if (base_pool
&& base_pool
->require_rollback()) {
10129 if ((iter
->op
.op
!= CEPH_OSD_OP_READ
) &&
10130 (iter
->op
.op
!= CEPH_OSD_OP_CHECKSUM
) &&
10131 (iter
->op
.op
!= CEPH_OSD_OP_CMPEXT
) &&
10132 (iter
->op
.op
!= CEPH_OSD_OP_STAT
) &&
10133 (iter
->op
.op
!= CEPH_OSD_OP_ISDIRTY
) &&
10134 (iter
->op
.op
!= CEPH_OSD_OP_UNDIRTY
) &&
10135 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTR
) &&
10136 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTRS
) &&
10137 (iter
->op
.op
!= CEPH_OSD_OP_CMPXATTR
) &&
10138 (iter
->op
.op
!= CEPH_OSD_OP_ASSERT_VER
) &&
10139 (iter
->op
.op
!= CEPH_OSD_OP_LIST_WATCHERS
) &&
10140 (iter
->op
.op
!= CEPH_OSD_OP_LIST_SNAPS
) &&
10141 (iter
->op
.op
!= CEPH_OSD_OP_SETALLOCHINT
) &&
10142 (iter
->op
.op
!= CEPH_OSD_OP_WRITEFULL
) &&
10143 (iter
->op
.op
!= CEPH_OSD_OP_ROLLBACK
) &&
10144 (iter
->op
.op
!= CEPH_OSD_OP_CREATE
) &&
10145 (iter
->op
.op
!= CEPH_OSD_OP_DELETE
) &&
10146 (iter
->op
.op
!= CEPH_OSD_OP_SETXATTR
) &&
10147 (iter
->op
.op
!= CEPH_OSD_OP_RMXATTR
) &&
10148 (iter
->op
.op
!= CEPH_OSD_OP_STARTSYNC
) &&
10149 (iter
->op
.op
!= CEPH_OSD_OP_COPY_GET
) &&
10150 (iter
->op
.op
!= CEPH_OSD_OP_COPY_FROM
)) {
10156 switch (iter
->op
.op
) {
10157 case CEPH_OSD_OP_CALL
:
10159 bufferlist::iterator bp
= const_cast<bufferlist
&>(iter
->indata
).begin();
10160 int is_write
, is_read
;
10161 string cname
, mname
;
10162 bp
.copy(iter
->op
.cls
.class_len
, cname
);
10163 bp
.copy(iter
->op
.cls
.method_len
, mname
);
10165 ClassHandler::ClassData
*cls
;
10166 int r
= class_handler
->open_class(cname
, &cls
);
10168 derr
<< "class " << cname
<< " open got " << cpp_strerror(r
) << dendl
;
10171 else if (r
!= -EPERM
) // propagate permission errors
10175 int flags
= cls
->get_method_flags(mname
.c_str());
10177 if (flags
== -ENOENT
)
10183 is_read
= flags
& CLS_METHOD_RD
;
10184 is_write
= flags
& CLS_METHOD_WR
;
10185 bool is_promote
= flags
& CLS_METHOD_PROMOTE
;
10187 dout(10) << "class " << cname
<< " method " << mname
<< " "
10188 << "flags=" << (is_read
? "r" : "")
10189 << (is_write
? "w" : "")
10190 << (is_promote
? "p" : "")
10193 op
->set_class_read();
10195 op
->set_class_write();
10198 op
->add_class(cname
, is_read
, is_write
, cls
->whitelisted
);
10202 case CEPH_OSD_OP_WATCH
:
10203 // force the read bit for watch since it is depends on previous
10204 // watch state (and may return early if the watch exists) or, in
10205 // the case of ping, is simply a read op.
10208 case CEPH_OSD_OP_NOTIFY
:
10209 case CEPH_OSD_OP_NOTIFY_ACK
:
10215 case CEPH_OSD_OP_DELETE
:
10216 // if we get a delete with FAILOK we can skip handle cache. without
10217 // FAILOK we still need to promote (or do something smarter) to
10218 // determine whether to return ENOENT or 0.
10219 if (iter
== m
->ops
.begin() &&
10220 iter
->op
.flags
== CEPH_OSD_OP_FLAG_FAILOK
) {
10221 op
->set_skip_handle_cache();
10223 // skip promotion when proxying a delete op
10224 if (m
->ops
.size() == 1) {
10225 op
->set_skip_promote();
10229 case CEPH_OSD_OP_CACHE_TRY_FLUSH
:
10230 case CEPH_OSD_OP_CACHE_FLUSH
:
10231 case CEPH_OSD_OP_CACHE_EVICT
:
10232 // If try_flush/flush/evict is the only op, can skip handle cache.
10233 if (m
->ops
.size() == 1) {
10234 op
->set_skip_handle_cache();
10238 case CEPH_OSD_OP_READ
:
10239 case CEPH_OSD_OP_SYNC_READ
:
10240 case CEPH_OSD_OP_SPARSE_READ
:
10241 case CEPH_OSD_OP_CHECKSUM
:
10242 case CEPH_OSD_OP_WRITEFULL
:
10243 if (m
->ops
.size() == 1 &&
10244 (iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
||
10245 iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
)) {
10246 op
->set_skip_promote();
10250 // force promotion when pin an object in cache tier
10251 case CEPH_OSD_OP_CACHE_PIN
:
10260 if (op
->rmw_flags
== 0)
10266 void OSD::PeeringWQ::_dequeue(list
<PG
*> *out
) {
10267 for (list
<PG
*>::iterator i
= peering_queue
.begin();
10268 i
!= peering_queue
.end() &&
10269 out
->size() < osd
->cct
->_conf
->osd_peering_wq_batch_size
;
10271 if (in_use
.count(*i
)) {
10274 out
->push_back(*i
);
10275 peering_queue
.erase(i
++);
10278 in_use
.insert(out
->begin(), out
->end());
10282 // =============================================================
10284 #undef dout_context
10285 #define dout_context osd->cct
10287 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq "
10289 void OSD::ShardedOpWQ::wake_pg_waiters(spg_t pgid
)
10291 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10292 auto sdata
= shard_list
[shard_index
];
10293 bool queued
= false;
10295 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10296 auto p
= sdata
->pg_slots
.find(pgid
);
10297 if (p
!= sdata
->pg_slots
.end()) {
10298 dout(20) << __func__
<< " " << pgid
10299 << " to_process " << p
->second
.to_process
10300 << " waiting_for_pg=" << (int)p
->second
.waiting_for_pg
<< dendl
;
10301 for (auto i
= p
->second
.to_process
.rbegin();
10302 i
!= p
->second
.to_process
.rend();
10304 sdata
->_enqueue_front(make_pair(pgid
, *i
), osd
->op_prio_cutoff
);
10306 p
->second
.to_process
.clear();
10307 p
->second
.waiting_for_pg
= false;
10308 ++p
->second
.requeue_seq
;
10313 sdata
->sdata_lock
.Lock();
10314 sdata
->sdata_cond
.SignalOne();
10315 sdata
->sdata_lock
.Unlock();
10319 void OSD::ShardedOpWQ::prune_pg_waiters(OSDMapRef osdmap
, int whoami
)
10321 unsigned pushes_to_free
= 0;
10322 for (auto sdata
: shard_list
) {
10323 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10324 sdata
->waiting_for_pg_osdmap
= osdmap
;
10325 auto p
= sdata
->pg_slots
.begin();
10326 while (p
!= sdata
->pg_slots
.end()) {
10327 ShardData::pg_slot
& slot
= p
->second
;
10328 if (!slot
.to_process
.empty() && slot
.num_running
== 0) {
10329 if (osdmap
->is_up_acting_osd_shard(p
->first
, whoami
)) {
10330 dout(20) << __func__
<< " " << p
->first
<< " maps to us, keeping"
10335 while (!slot
.to_process
.empty() &&
10336 slot
.to_process
.front().get_map_epoch() <= osdmap
->get_epoch()) {
10337 auto& qi
= slot
.to_process
.front();
10338 dout(20) << __func__
<< " " << p
->first
10340 << " epoch " << qi
.get_map_epoch()
10341 << " <= " << osdmap
->get_epoch()
10342 << ", stale, dropping" << dendl
;
10343 pushes_to_free
+= qi
.get_reserved_pushes();
10344 slot
.to_process
.pop_front();
10347 if (slot
.to_process
.empty() &&
10348 slot
.num_running
== 0 &&
10350 dout(20) << __func__
<< " " << p
->first
<< " empty, pruning" << dendl
;
10351 p
= sdata
->pg_slots
.erase(p
);
10357 if (pushes_to_free
> 0) {
10358 osd
->service
.release_reserved_pushes(pushes_to_free
);
10362 void OSD::ShardedOpWQ::clear_pg_pointer(spg_t pgid
)
10364 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10365 auto sdata
= shard_list
[shard_index
];
10366 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10367 auto p
= sdata
->pg_slots
.find(pgid
);
10368 if (p
!= sdata
->pg_slots
.end()) {
10369 auto& slot
= p
->second
;
10370 dout(20) << __func__
<< " " << pgid
<< " pg " << slot
.pg
<< dendl
;
10371 assert(!slot
.pg
|| slot
.pg
->deleting
);
10376 void OSD::ShardedOpWQ::clear_pg_slots()
10378 for (auto sdata
: shard_list
) {
10379 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10380 sdata
->pg_slots
.clear();
10381 sdata
->waiting_for_pg_osdmap
.reset();
10382 // don't bother with reserved pushes; we are shutting down
10387 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq(" << shard_index << ") "
10389 void OSD::ShardedOpWQ::_process(uint32_t thread_index
, heartbeat_handle_d
*hb
)
10391 uint32_t shard_index
= thread_index
% num_shards
;
10392 ShardData
*sdata
= shard_list
[shard_index
];
10393 assert(NULL
!= sdata
);
10396 sdata
->sdata_op_ordering_lock
.Lock();
10397 if (sdata
->pqueue
->empty()) {
10398 dout(20) << __func__
<< " empty q, waiting" << dendl
;
10399 // optimistically sleep a moment; maybe another work item will come along.
10400 osd
->cct
->get_heartbeat_map()->reset_timeout(hb
,
10401 osd
->cct
->_conf
->threadpool_default_timeout
, 0);
10402 sdata
->sdata_lock
.Lock();
10403 sdata
->sdata_op_ordering_lock
.Unlock();
10404 sdata
->sdata_cond
.WaitInterval(sdata
->sdata_lock
,
10405 utime_t(osd
->cct
->_conf
->threadpool_empty_queue_max_wait
, 0));
10406 sdata
->sdata_lock
.Unlock();
10407 sdata
->sdata_op_ordering_lock
.Lock();
10408 if (sdata
->pqueue
->empty()) {
10409 sdata
->sdata_op_ordering_lock
.Unlock();
10413 pair
<spg_t
, PGQueueable
> item
= sdata
->pqueue
->dequeue();
10414 if (osd
->is_stopping()) {
10415 sdata
->sdata_op_ordering_lock
.Unlock();
10416 return; // OSD shutdown, discard.
10419 uint64_t requeue_seq
;
10421 auto& slot
= sdata
->pg_slots
[item
.first
];
10422 dout(30) << __func__
<< " " << item
.first
10423 << " to_process " << slot
.to_process
10424 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10425 slot
.to_process
.push_back(item
.second
);
10426 // note the requeue seq now...
10427 requeue_seq
= slot
.requeue_seq
;
10428 if (slot
.waiting_for_pg
) {
10429 // save ourselves a bit of effort
10430 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10431 << " queued, waiting_for_pg" << dendl
;
10432 sdata
->sdata_op_ordering_lock
.Unlock();
10436 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10437 << " queued" << dendl
;
10438 ++slot
.num_running
;
10440 sdata
->sdata_op_ordering_lock
.Unlock();
10442 osd
->service
.maybe_inject_dispatch_delay();
10444 // [lookup +] lock pg (if we have it)
10446 pg
= osd
->_lookup_lock_pg(item
.first
);
10451 osd
->service
.maybe_inject_dispatch_delay();
10453 boost::optional
<PGQueueable
> qi
;
10455 // we don't use a Mutex::Locker here because of the
10456 // osd->service.release_reserved_pushes() call below
10457 sdata
->sdata_op_ordering_lock
.Lock();
10459 auto q
= sdata
->pg_slots
.find(item
.first
);
10460 assert(q
!= sdata
->pg_slots
.end());
10461 auto& slot
= q
->second
;
10462 --slot
.num_running
;
10464 if (slot
.to_process
.empty()) {
10465 // raced with wake_pg_waiters or prune_pg_waiters
10466 dout(20) << __func__
<< " " << item
.first
<< " nothing queued" << dendl
;
10470 sdata
->sdata_op_ordering_lock
.Unlock();
10473 if (requeue_seq
!= slot
.requeue_seq
) {
10474 dout(20) << __func__
<< " " << item
.first
10475 << " requeue_seq " << slot
.requeue_seq
<< " > our "
10476 << requeue_seq
<< ", we raced with wake_pg_waiters"
10481 sdata
->sdata_op_ordering_lock
.Unlock();
10484 if (pg
&& !slot
.pg
&& !pg
->deleting
) {
10485 dout(20) << __func__
<< " " << item
.first
<< " set pg to " << pg
<< dendl
;
10488 dout(30) << __func__
<< " " << item
.first
<< " to_process " << slot
.to_process
10489 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10491 // make sure we're not already waiting for this pg
10492 if (slot
.waiting_for_pg
) {
10493 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10494 << " slot is waiting_for_pg" << dendl
;
10498 sdata
->sdata_op_ordering_lock
.Unlock();
10503 qi
= slot
.to_process
.front();
10504 slot
.to_process
.pop_front();
10505 dout(20) << __func__
<< " " << item
.first
<< " item " << *qi
10506 << " pg " << pg
<< dendl
;
10509 // should this pg shard exist on this osd in this (or a later) epoch?
10510 OSDMapRef osdmap
= sdata
->waiting_for_pg_osdmap
;
10511 if (osdmap
->is_up_acting_osd_shard(item
.first
, osd
->whoami
)) {
10512 dout(20) << __func__
<< " " << item
.first
10513 << " no pg, should exist, will wait" << " on " << *qi
<< dendl
;
10514 slot
.to_process
.push_front(*qi
);
10515 slot
.waiting_for_pg
= true;
10516 } else if (qi
->get_map_epoch() > osdmap
->get_epoch()) {
10517 dout(20) << __func__
<< " " << item
.first
<< " no pg, item epoch is "
10518 << qi
->get_map_epoch() << " > " << osdmap
->get_epoch()
10519 << ", will wait on " << *qi
<< dendl
;
10520 slot
.to_process
.push_front(*qi
);
10521 slot
.waiting_for_pg
= true;
10523 dout(20) << __func__
<< " " << item
.first
<< " no pg, shouldn't exist,"
10524 << " dropping " << *qi
<< dendl
;
10525 // share map with client?
10526 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10527 Session
*session
= static_cast<Session
*>(
10528 (*_op
)->get_req()->get_connection()->get_priv());
10530 osd
->maybe_share_map(session
, *_op
, sdata
->waiting_for_pg_osdmap
);
10534 unsigned pushes_to_free
= qi
->get_reserved_pushes();
10535 if (pushes_to_free
> 0) {
10536 sdata
->sdata_op_ordering_lock
.Unlock();
10537 osd
->service
.release_reserved_pushes(pushes_to_free
);
10541 sdata
->sdata_op_ordering_lock
.Unlock();
10544 sdata
->sdata_op_ordering_lock
.Unlock();
10547 // osd_opwq_process marks the point at which an operation has been dequeued
10548 // and will begin to be handled by a worker thread.
10552 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10553 reqid
= (*_op
)->get_reqid();
10556 tracepoint(osd
, opwq_process_start
, reqid
.name
._type
,
10557 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10560 lgeneric_subdout(osd
->cct
, osd
, 30) << "dequeue status: ";
10561 Formatter
*f
= Formatter::create("json");
10562 f
->open_object_section("q");
10564 f
->close_section();
10569 ThreadPool::TPHandle
tp_handle(osd
->cct
, hb
, timeout_interval
,
10571 qi
->run(osd
, pg
, tp_handle
);
10576 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10577 reqid
= (*_op
)->get_reqid();
10580 tracepoint(osd
, opwq_process_finish
, reqid
.name
._type
,
10581 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10587 void OSD::ShardedOpWQ::_enqueue(pair
<spg_t
, PGQueueable
> item
) {
10588 uint32_t shard_index
=
10589 item
.first
.hash_to_shard(shard_list
.size());
10591 ShardData
* sdata
= shard_list
[shard_index
];
10592 assert (NULL
!= sdata
);
10593 unsigned priority
= item
.second
.get_priority();
10594 unsigned cost
= item
.second
.get_cost();
10595 sdata
->sdata_op_ordering_lock
.Lock();
10597 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10598 if (priority
>= osd
->op_prio_cutoff
)
10599 sdata
->pqueue
->enqueue_strict(
10600 item
.second
.get_owner(), priority
, item
);
10602 sdata
->pqueue
->enqueue(
10603 item
.second
.get_owner(),
10604 priority
, cost
, item
);
10605 sdata
->sdata_op_ordering_lock
.Unlock();
10607 sdata
->sdata_lock
.Lock();
10608 sdata
->sdata_cond
.SignalOne();
10609 sdata
->sdata_lock
.Unlock();
10613 void OSD::ShardedOpWQ::_enqueue_front(pair
<spg_t
, PGQueueable
> item
)
10615 uint32_t shard_index
= item
.first
.hash_to_shard(shard_list
.size());
10616 ShardData
* sdata
= shard_list
[shard_index
];
10617 assert (NULL
!= sdata
);
10618 sdata
->sdata_op_ordering_lock
.Lock();
10619 auto p
= sdata
->pg_slots
.find(item
.first
);
10620 if (p
!= sdata
->pg_slots
.end() && !p
->second
.to_process
.empty()) {
10621 // we may be racing with _process, which has dequeued a new item
10622 // from pqueue, put it on to_process, and is now busy taking the
10623 // pg lock. ensure this old requeued item is ordered before any
10624 // such newer item in to_process.
10625 p
->second
.to_process
.push_front(item
.second
);
10626 item
.second
= p
->second
.to_process
.back();
10627 p
->second
.to_process
.pop_back();
10628 dout(20) << __func__
<< " " << item
.first
10629 << " " << p
->second
.to_process
.front()
10630 << " shuffled w/ " << item
.second
<< dendl
;
10632 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10634 sdata
->_enqueue_front(item
, osd
->op_prio_cutoff
);
10635 sdata
->sdata_op_ordering_lock
.Unlock();
10636 sdata
->sdata_lock
.Lock();
10637 sdata
->sdata_cond
.SignalOne();
10638 sdata
->sdata_lock
.Unlock();
10642 namespace osd_cmds
{
10644 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
)
10646 if (!ceph_using_tcmalloc()) {
10647 os
<< "could not issue heap profiler command -- not using tcmalloc!";
10648 return -EOPNOTSUPP
;
10652 if (!cmd_getval(&cct
, cmdmap
, "heapcmd", cmd
)) {
10653 os
<< "unable to get value for command \"" << cmd
<< "\"";
10657 std::vector
<std::string
> cmd_vec
;
10658 get_str_vec(cmd
, cmd_vec
);
10660 ceph_heap_profiler_handle_command(cmd_vec
, os
);
10665 }} // namespace ceph::osd_cmds
10668 std::ostream
& operator<<(std::ostream
& out
, const OSD::io_queue
& q
) {
10670 case OSD::io_queue::prioritized
:
10671 out
<< "prioritized";
10673 case OSD::io_queue::weightedpriority
:
10674 out
<< "weightedpriority";
10676 case OSD::io_queue::mclock_opclass
:
10677 out
<< "mclock_opclass";
10679 case OSD::io_queue::mclock_client
:
10680 out
<< "mclock_client";