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(&reserver_finisher
, cct
->_conf
->osd_max_backfills
,
251 cct
->_conf
->osd_min_recovery_priority
),
252 remote_reserver(&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(&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
, vector
<int>& want
)
1039 Mutex::Locker
l(pg_temp_lock
);
1040 map
<pg_t
,vector
<int> >::iterator p
= pg_temp_pending
.find(pgid
);
1041 if (p
== pg_temp_pending
.end() ||
1042 p
->second
!= want
) {
1043 pg_temp_wanted
[pgid
] = want
;
1047 void OSDService::remove_want_pg_temp(pg_t pgid
)
1049 Mutex::Locker
l(pg_temp_lock
);
1050 pg_temp_wanted
.erase(pgid
);
1051 pg_temp_pending
.erase(pgid
);
1054 void OSDService::_sent_pg_temp()
1056 for (map
<pg_t
,vector
<int> >::iterator p
= pg_temp_wanted
.begin();
1057 p
!= pg_temp_wanted
.end();
1059 pg_temp_pending
[p
->first
] = p
->second
;
1060 pg_temp_wanted
.clear();
1063 void OSDService::requeue_pg_temp()
1065 Mutex::Locker
l(pg_temp_lock
);
1066 // wanted overrides pending. note that remove_want_pg_temp
1067 // clears the item out of both.
1068 unsigned old_wanted
= pg_temp_wanted
.size();
1069 unsigned old_pending
= pg_temp_pending
.size();
1071 pg_temp_wanted
.swap(pg_temp_pending
);
1072 dout(10) << __func__
<< " " << old_wanted
<< " + " << old_pending
<< " -> "
1073 << pg_temp_wanted
.size() << dendl
;
1076 void OSDService::send_pg_temp()
1078 Mutex::Locker
l(pg_temp_lock
);
1079 if (pg_temp_wanted
.empty())
1081 dout(10) << "send_pg_temp " << pg_temp_wanted
<< dendl
;
1082 MOSDPGTemp
*m
= new MOSDPGTemp(osdmap
->get_epoch());
1083 m
->pg_temp
= pg_temp_wanted
;
1084 monc
->send_mon_message(m
);
1088 void OSDService::send_pg_created(pg_t pgid
)
1090 dout(20) << __func__
<< dendl
;
1091 if (osdmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
1092 monc
->send_mon_message(new MOSDPGCreated(pgid
));
1096 // --------------------------------------
1099 epoch_t
OSDService::get_peer_epoch(int peer
)
1101 Mutex::Locker
l(peer_map_epoch_lock
);
1102 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1103 if (p
== peer_map_epoch
.end())
1108 epoch_t
OSDService::note_peer_epoch(int peer
, epoch_t e
)
1110 Mutex::Locker
l(peer_map_epoch_lock
);
1111 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1112 if (p
!= peer_map_epoch
.end()) {
1113 if (p
->second
< e
) {
1114 dout(10) << "note_peer_epoch osd." << peer
<< " has " << e
<< dendl
;
1117 dout(30) << "note_peer_epoch osd." << peer
<< " has " << p
->second
<< " >= " << e
<< dendl
;
1121 dout(10) << "note_peer_epoch osd." << peer
<< " now has " << e
<< dendl
;
1122 peer_map_epoch
[peer
] = e
;
1127 void OSDService::forget_peer_epoch(int peer
, epoch_t as_of
)
1129 Mutex::Locker
l(peer_map_epoch_lock
);
1130 map
<int,epoch_t
>::iterator p
= peer_map_epoch
.find(peer
);
1131 if (p
!= peer_map_epoch
.end()) {
1132 if (p
->second
<= as_of
) {
1133 dout(10) << "forget_peer_epoch osd." << peer
<< " as_of " << as_of
1134 << " had " << p
->second
<< dendl
;
1135 peer_map_epoch
.erase(p
);
1137 dout(10) << "forget_peer_epoch osd." << peer
<< " as_of " << as_of
1138 << " has " << p
->second
<< " - not forgetting" << dendl
;
1143 bool OSDService::should_share_map(entity_name_t name
, Connection
*con
,
1144 epoch_t epoch
, const OSDMapRef
& osdmap
,
1145 const epoch_t
*sent_epoch_p
)
1147 dout(20) << "should_share_map "
1148 << name
<< " " << con
->get_peer_addr()
1149 << " " << epoch
<< dendl
;
1151 // does client have old map?
1152 if (name
.is_client()) {
1153 bool message_sendmap
= epoch
< osdmap
->get_epoch();
1154 if (message_sendmap
&& sent_epoch_p
) {
1155 dout(20) << "client session last_sent_epoch: "
1157 << " versus osdmap epoch " << osdmap
->get_epoch() << dendl
;
1158 if (*sent_epoch_p
< osdmap
->get_epoch()) {
1160 } // else we don't need to send it out again
1164 if (con
->get_messenger() == osd
->cluster_messenger
&&
1165 con
!= osd
->cluster_messenger
->get_loopback_connection() &&
1166 osdmap
->is_up(name
.num()) &&
1167 (osdmap
->get_cluster_addr(name
.num()) == con
->get_peer_addr() ||
1168 osdmap
->get_hb_back_addr(name
.num()) == con
->get_peer_addr())) {
1170 epoch_t has
= MAX(get_peer_epoch(name
.num()), epoch
);
1173 if (has
< osdmap
->get_epoch()) {
1174 dout(10) << name
<< " " << con
->get_peer_addr()
1175 << " has old map " << epoch
<< " < "
1176 << osdmap
->get_epoch() << dendl
;
1184 void OSDService::share_map(
1189 epoch_t
*sent_epoch_p
)
1191 dout(20) << "share_map "
1192 << name
<< " " << con
->get_peer_addr()
1193 << " " << epoch
<< dendl
;
1195 if (!osd
->is_active()) {
1196 /*It is safe not to proceed as OSD is not in healthy state*/
1200 bool want_shared
= should_share_map(name
, con
, epoch
,
1201 osdmap
, sent_epoch_p
);
1204 if (name
.is_client()) {
1205 dout(10) << name
<< " has old map " << epoch
1206 << " < " << osdmap
->get_epoch() << dendl
;
1207 // we know the Session is valid or we wouldn't be sending
1209 *sent_epoch_p
= osdmap
->get_epoch();
1211 send_incremental_map(epoch
, con
, osdmap
);
1212 } else if (con
->get_messenger() == osd
->cluster_messenger
&&
1213 osdmap
->is_up(name
.num()) &&
1214 (osdmap
->get_cluster_addr(name
.num()) == con
->get_peer_addr() ||
1215 osdmap
->get_hb_back_addr(name
.num()) == con
->get_peer_addr())) {
1216 dout(10) << name
<< " " << con
->get_peer_addr()
1217 << " has old map " << epoch
<< " < "
1218 << osdmap
->get_epoch() << dendl
;
1219 note_peer_epoch(name
.num(), osdmap
->get_epoch());
1220 send_incremental_map(epoch
, con
, osdmap
);
1225 void OSDService::share_map_peer(int peer
, Connection
*con
, OSDMapRef map
)
1231 epoch_t pe
= get_peer_epoch(peer
);
1233 if (pe
< map
->get_epoch()) {
1234 send_incremental_map(pe
, con
, map
);
1235 note_peer_epoch(peer
, map
->get_epoch());
1237 dout(20) << "share_map_peer " << con
<< " already has epoch " << pe
<< dendl
;
1239 dout(20) << "share_map_peer " << con
<< " don't know epoch, doing nothing" << dendl
;
1240 // no idea about peer's epoch.
1241 // ??? send recent ???
1246 bool OSDService::can_inc_scrubs_pending()
1248 bool can_inc
= false;
1249 Mutex::Locker
l(sched_scrub_lock
);
1251 if (scrubs_pending
+ scrubs_active
< cct
->_conf
->osd_max_scrubs
) {
1252 dout(20) << __func__
<< " " << scrubs_pending
<< " -> " << (scrubs_pending
+1)
1253 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1256 dout(20) << __func__
<< scrubs_pending
<< " + " << scrubs_active
<< " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1262 bool OSDService::inc_scrubs_pending()
1264 bool result
= false;
1266 sched_scrub_lock
.Lock();
1267 if (scrubs_pending
+ scrubs_active
< cct
->_conf
->osd_max_scrubs
) {
1268 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
+1)
1269 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1273 dout(20) << "inc_scrubs_pending " << scrubs_pending
<< " + " << scrubs_active
<< " active >= max " << cct
->_conf
->osd_max_scrubs
<< dendl
;
1275 sched_scrub_lock
.Unlock();
1280 void OSDService::dec_scrubs_pending()
1282 sched_scrub_lock
.Lock();
1283 dout(20) << "dec_scrubs_pending " << scrubs_pending
<< " -> " << (scrubs_pending
-1)
1284 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", active " << scrubs_active
<< ")" << dendl
;
1286 assert(scrubs_pending
>= 0);
1287 sched_scrub_lock
.Unlock();
1290 void OSDService::inc_scrubs_active(bool reserved
)
1292 sched_scrub_lock
.Lock();
1296 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1297 << " (max " << cct
->_conf
->osd_max_scrubs
1298 << ", pending " << (scrubs_pending
+1) << " -> " << scrubs_pending
<< ")" << dendl
;
1299 assert(scrubs_pending
>= 0);
1301 dout(20) << "inc_scrubs_active " << (scrubs_active
-1) << " -> " << scrubs_active
1302 << " (max " << cct
->_conf
->osd_max_scrubs
1303 << ", pending " << scrubs_pending
<< ")" << dendl
;
1305 sched_scrub_lock
.Unlock();
1308 void OSDService::dec_scrubs_active()
1310 sched_scrub_lock
.Lock();
1311 dout(20) << "dec_scrubs_active " << scrubs_active
<< " -> " << (scrubs_active
-1)
1312 << " (max " << cct
->_conf
->osd_max_scrubs
<< ", pending " << scrubs_pending
<< ")" << dendl
;
1314 assert(scrubs_active
>= 0);
1315 sched_scrub_lock
.Unlock();
1318 void OSDService::retrieve_epochs(epoch_t
*_boot_epoch
, epoch_t
*_up_epoch
,
1319 epoch_t
*_bind_epoch
) const
1321 Mutex::Locker
l(epoch_lock
);
1323 *_boot_epoch
= boot_epoch
;
1325 *_up_epoch
= up_epoch
;
1327 *_bind_epoch
= bind_epoch
;
1330 void OSDService::set_epochs(const epoch_t
*_boot_epoch
, const epoch_t
*_up_epoch
,
1331 const epoch_t
*_bind_epoch
)
1333 Mutex::Locker
l(epoch_lock
);
1335 assert(*_boot_epoch
== 0 || *_boot_epoch
>= boot_epoch
);
1336 boot_epoch
= *_boot_epoch
;
1339 assert(*_up_epoch
== 0 || *_up_epoch
>= up_epoch
);
1340 up_epoch
= *_up_epoch
;
1343 assert(*_bind_epoch
== 0 || *_bind_epoch
>= bind_epoch
);
1344 bind_epoch
= *_bind_epoch
;
1348 bool OSDService::prepare_to_stop()
1350 Mutex::Locker
l(is_stopping_lock
);
1351 if (get_state() != NOT_STOPPING
)
1354 OSDMapRef osdmap
= get_osdmap();
1355 if (osdmap
&& osdmap
->is_up(whoami
)) {
1356 dout(0) << __func__
<< " telling mon we are shutting down" << dendl
;
1357 set_state(PREPARING_TO_STOP
);
1358 monc
->send_mon_message(new MOSDMarkMeDown(monc
->get_fsid(),
1359 osdmap
->get_inst(whoami
),
1360 osdmap
->get_epoch(),
1363 utime_t now
= ceph_clock_now();
1365 timeout
.set_from_double(now
+ cct
->_conf
->osd_mon_shutdown_timeout
);
1366 while ((ceph_clock_now() < timeout
) &&
1367 (get_state() != STOPPING
)) {
1368 is_stopping_cond
.WaitUntil(is_stopping_lock
, timeout
);
1371 dout(0) << __func__
<< " starting shutdown" << dendl
;
1372 set_state(STOPPING
);
1376 void OSDService::got_stop_ack()
1378 Mutex::Locker
l(is_stopping_lock
);
1379 if (get_state() == PREPARING_TO_STOP
) {
1380 dout(0) << __func__
<< " starting shutdown" << dendl
;
1381 set_state(STOPPING
);
1382 is_stopping_cond
.Signal();
1384 dout(10) << __func__
<< " ignoring msg" << dendl
;
1388 MOSDMap
*OSDService::build_incremental_map_msg(epoch_t since
, epoch_t to
,
1389 OSDSuperblock
& sblock
)
1391 MOSDMap
*m
= new MOSDMap(monc
->get_fsid());
1392 m
->oldest_map
= max_oldest_map
;
1393 m
->newest_map
= sblock
.newest_map
;
1395 for (epoch_t e
= to
; e
> since
; e
--) {
1397 if (e
> m
->oldest_map
&& get_inc_map_bl(e
, bl
)) {
1398 m
->incremental_maps
[e
].claim(bl
);
1399 } else if (get_map_bl(e
, bl
)) {
1400 m
->maps
[e
].claim(bl
);
1403 derr
<< "since " << since
<< " to " << to
1404 << " oldest " << m
->oldest_map
<< " newest " << m
->newest_map
1414 void OSDService::send_map(MOSDMap
*m
, Connection
*con
)
1416 con
->send_message(m
);
1419 void OSDService::send_incremental_map(epoch_t since
, Connection
*con
,
1422 epoch_t to
= osdmap
->get_epoch();
1423 dout(10) << "send_incremental_map " << since
<< " -> " << to
1424 << " to " << con
<< " " << con
->get_peer_addr() << dendl
;
1428 OSDSuperblock
sblock(get_superblock());
1429 if (since
< sblock
.oldest_map
) {
1430 // just send latest full map
1431 MOSDMap
*m
= new MOSDMap(monc
->get_fsid());
1432 m
->oldest_map
= max_oldest_map
;
1433 m
->newest_map
= sblock
.newest_map
;
1434 get_map_bl(to
, m
->maps
[to
]);
1439 if (to
> since
&& (int64_t)(to
- since
) > cct
->_conf
->osd_map_share_max_epochs
) {
1440 dout(10) << " " << (to
- since
) << " > max " << cct
->_conf
->osd_map_share_max_epochs
1441 << ", only sending most recent" << dendl
;
1442 since
= to
- cct
->_conf
->osd_map_share_max_epochs
;
1445 if (to
- since
> (epoch_t
)cct
->_conf
->osd_map_message_max
)
1446 to
= since
+ cct
->_conf
->osd_map_message_max
;
1447 m
= build_incremental_map_msg(since
, to
, sblock
);
1452 bool OSDService::_get_map_bl(epoch_t e
, bufferlist
& bl
)
1454 bool found
= map_bl_cache
.lookup(e
, &bl
);
1457 logger
->inc(l_osd_map_bl_cache_hit
);
1461 logger
->inc(l_osd_map_bl_cache_miss
);
1462 found
= store
->read(coll_t::meta(),
1463 OSD::get_osdmap_pobject_name(e
), 0, 0, bl
,
1464 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1471 bool OSDService::get_inc_map_bl(epoch_t e
, bufferlist
& bl
)
1473 Mutex::Locker
l(map_cache_lock
);
1474 bool found
= map_bl_inc_cache
.lookup(e
, &bl
);
1477 logger
->inc(l_osd_map_bl_cache_hit
);
1481 logger
->inc(l_osd_map_bl_cache_miss
);
1482 found
= store
->read(coll_t::meta(),
1483 OSD::get_inc_osdmap_pobject_name(e
), 0, 0, bl
,
1484 CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) >= 0;
1486 _add_map_inc_bl(e
, bl
);
1491 void OSDService::_add_map_bl(epoch_t e
, bufferlist
& bl
)
1493 dout(10) << "add_map_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1494 // cache a contiguous buffer
1495 if (bl
.get_num_buffers() > 1) {
1498 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1499 map_bl_cache
.add(e
, bl
);
1502 void OSDService::_add_map_inc_bl(epoch_t e
, bufferlist
& bl
)
1504 dout(10) << "add_map_inc_bl " << e
<< " " << bl
.length() << " bytes" << dendl
;
1505 // cache a contiguous buffer
1506 if (bl
.get_num_buffers() > 1) {
1509 bl
.try_assign_to_mempool(mempool::mempool_osd_mapbl
);
1510 map_bl_inc_cache
.add(e
, bl
);
1513 void OSDService::pin_map_inc_bl(epoch_t e
, bufferlist
&bl
)
1515 Mutex::Locker
l(map_cache_lock
);
1516 // cache a contiguous buffer
1517 if (bl
.get_num_buffers() > 1) {
1520 map_bl_inc_cache
.pin(e
, bl
);
1523 void OSDService::pin_map_bl(epoch_t e
, bufferlist
&bl
)
1525 Mutex::Locker
l(map_cache_lock
);
1526 // cache a contiguous buffer
1527 if (bl
.get_num_buffers() > 1) {
1530 map_bl_cache
.pin(e
, bl
);
1533 void OSDService::clear_map_bl_cache_pins(epoch_t e
)
1535 Mutex::Locker
l(map_cache_lock
);
1536 map_bl_inc_cache
.clear_pinned(e
);
1537 map_bl_cache
.clear_pinned(e
);
1540 OSDMapRef
OSDService::_add_map(OSDMap
*o
)
1542 epoch_t e
= o
->get_epoch();
1544 if (cct
->_conf
->osd_map_dedup
) {
1545 // Dedup against an existing map at a nearby epoch
1546 OSDMapRef for_dedup
= map_cache
.lower_bound(e
);
1548 OSDMap::dedup(for_dedup
.get(), o
);
1552 OSDMapRef l
= map_cache
.add(e
, o
, &existed
);
1559 OSDMapRef
OSDService::try_get_map(epoch_t epoch
)
1561 Mutex::Locker
l(map_cache_lock
);
1562 OSDMapRef retval
= map_cache
.lookup(epoch
);
1564 dout(30) << "get_map " << epoch
<< " -cached" << dendl
;
1566 logger
->inc(l_osd_map_cache_hit
);
1571 logger
->inc(l_osd_map_cache_miss
);
1572 epoch_t lb
= map_cache
.cached_key_lower_bound();
1574 dout(30) << "get_map " << epoch
<< " - miss, below lower bound" << dendl
;
1575 logger
->inc(l_osd_map_cache_miss_low
);
1576 logger
->inc(l_osd_map_cache_miss_low_avg
, lb
- epoch
);
1580 OSDMap
*map
= new OSDMap
;
1582 dout(20) << "get_map " << epoch
<< " - loading and decoding " << map
<< dendl
;
1584 if (!_get_map_bl(epoch
, bl
) || bl
.length() == 0) {
1585 derr
<< "failed to load OSD map for epoch " << epoch
<< ", got " << bl
.length() << " bytes" << dendl
;
1591 dout(20) << "get_map " << epoch
<< " - return initial " << map
<< dendl
;
1593 return _add_map(map
);
1599 void OSDService::reply_op_error(OpRequestRef op
, int err
)
1601 reply_op_error(op
, err
, eversion_t(), 0);
1604 void OSDService::reply_op_error(OpRequestRef op
, int err
, eversion_t v
,
1607 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1608 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1610 flags
= m
->get_flags() & (CEPH_OSD_FLAG_ACK
|CEPH_OSD_FLAG_ONDISK
);
1612 MOSDOpReply
*reply
= new MOSDOpReply(m
, err
, osdmap
->get_epoch(), flags
,
1614 reply
->set_reply_versions(v
, uv
);
1615 m
->get_connection()->send_message(reply
);
1618 void OSDService::handle_misdirected_op(PG
*pg
, OpRequestRef op
)
1620 if (!cct
->_conf
->osd_debug_misdirected_ops
) {
1624 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
1625 assert(m
->get_type() == CEPH_MSG_OSD_OP
);
1627 assert(m
->get_map_epoch() >= pg
->info
.history
.same_primary_since
);
1629 if (pg
->is_ec_pg()) {
1631 * OSD recomputes op target based on current OSDMap. With an EC pg, we
1632 * can get this result:
1633 * 1) client at map 512 sends an op to osd 3, pg_t 3.9 based on mapping
1634 * [CRUSH_ITEM_NONE, 2, 3]/3
1635 * 2) OSD 3 at map 513 remaps op to osd 3, spg_t 3.9s0 based on mapping
1637 * 3) PG 3.9s0 dequeues the op at epoch 512 and notices that it isn't primary
1639 * 4) client resends and this time PG 3.9s0 having caught up to 513 gets
1642 * We can't compute the op target based on the sending map epoch due to
1643 * splitting. The simplest thing is to detect such cases here and drop
1644 * them without an error (the client will resend anyway).
1646 assert(m
->get_map_epoch() <= superblock
.newest_map
);
1647 OSDMapRef opmap
= try_get_map(m
->get_map_epoch());
1649 dout(7) << __func__
<< ": " << *pg
<< " no longer have map for "
1650 << m
->get_map_epoch() << ", dropping" << dendl
;
1653 pg_t _pgid
= m
->get_raw_pg();
1655 if ((m
->get_flags() & CEPH_OSD_FLAG_PGOP
) == 0)
1656 _pgid
= opmap
->raw_pg_to_pg(_pgid
);
1657 if (opmap
->get_primary_shard(_pgid
, &pgid
) &&
1658 pgid
.shard
!= pg
->info
.pgid
.shard
) {
1659 dout(7) << __func__
<< ": " << *pg
<< " primary changed since "
1660 << m
->get_map_epoch() << ", dropping" << dendl
;
1665 dout(7) << *pg
<< " misdirected op in " << m
->get_map_epoch() << dendl
;
1666 clog
->warn() << m
->get_source_inst() << " misdirected " << m
->get_reqid()
1667 << " pg " << m
->get_raw_pg()
1668 << " to osd." << whoami
1669 << " not " << pg
->acting
1670 << " in e" << m
->get_map_epoch() << "/" << osdmap
->get_epoch();
1673 void OSDService::enqueue_back(spg_t pgid
, PGQueueable qi
)
1675 osd
->op_shardedwq
.queue(make_pair(pgid
, qi
));
1678 void OSDService::enqueue_front(spg_t pgid
, PGQueueable qi
)
1680 osd
->op_shardedwq
.queue_front(make_pair(pgid
, qi
));
1683 void OSDService::queue_for_peering(PG
*pg
)
1685 peering_wq
.queue(pg
);
1688 void OSDService::queue_for_snap_trim(PG
*pg
)
1690 dout(10) << "queueing " << *pg
<< " for snaptrim" << dendl
;
1691 osd
->op_shardedwq
.queue(
1695 PGSnapTrim(pg
->get_osdmap()->get_epoch()),
1696 cct
->_conf
->osd_snap_trim_cost
,
1697 cct
->_conf
->osd_snap_trim_priority
,
1700 pg
->get_osdmap()->get_epoch())));
1704 // ====================================================================
1708 #define dout_prefix *_dout
1710 // Commands shared between OSD's console and admin console:
1712 namespace osd_cmds
{
1714 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
);
1716 }} // namespace ceph::osd_cmds
1718 int OSD::mkfs(CephContext
*cct
, ObjectStore
*store
, const string
&dev
,
1719 uuid_d fsid
, int whoami
)
1723 ceph::shared_ptr
<ObjectStore::Sequencer
> osr(
1724 new ObjectStore::Sequencer("mkfs"));
1729 // if we are fed a uuid for this osd, use it.
1730 store
->set_fsid(cct
->_conf
->osd_uuid
);
1732 ret
= store
->mkfs();
1734 derr
<< "OSD::mkfs: ObjectStore::mkfs failed with error "
1735 << cpp_strerror(ret
) << dendl
;
1739 store
->set_cache_shards(1); // doesn't matter for mkfs!
1741 ret
= store
->mount();
1743 derr
<< "OSD::mkfs: couldn't mount ObjectStore: error "
1744 << cpp_strerror(ret
) << dendl
;
1748 ret
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, sbbl
);
1750 /* if we already have superblock, check content of superblock */
1751 dout(0) << " have superblock" << dendl
;
1752 bufferlist::iterator p
;
1755 if (whoami
!= sb
.whoami
) {
1756 derr
<< "provided osd id " << whoami
<< " != superblock's " << sb
.whoami
1761 if (fsid
!= sb
.cluster_fsid
) {
1762 derr
<< "provided cluster fsid " << fsid
1763 << " != superblock's " << sb
.cluster_fsid
<< dendl
;
1768 // create superblock
1769 sb
.cluster_fsid
= fsid
;
1770 sb
.osd_fsid
= store
->get_fsid();
1772 sb
.compat_features
= get_osd_initial_compat_set();
1777 ObjectStore::Transaction t
;
1778 t
.create_collection(coll_t::meta(), 0);
1779 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
1780 ret
= store
->apply_transaction(osr
.get(), std::move(t
));
1782 derr
<< "OSD::mkfs: error while writing OSD_SUPERBLOCK_GOBJECT: "
1783 << "apply_transaction returned " << cpp_strerror(ret
) << dendl
;
1788 if (!osr
->flush_commit(&waiter
)) {
1792 ret
= write_meta(store
, sb
.cluster_fsid
, sb
.osd_fsid
, whoami
);
1794 derr
<< "OSD::mkfs: failed to write fsid file: error "
1795 << cpp_strerror(ret
) << dendl
;
1806 int OSD::write_meta(ObjectStore
*store
, uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int whoami
)
1811 snprintf(val
, sizeof(val
), "%s", CEPH_OSD_ONDISK_MAGIC
);
1812 r
= store
->write_meta("magic", val
);
1816 snprintf(val
, sizeof(val
), "%d", whoami
);
1817 r
= store
->write_meta("whoami", val
);
1821 cluster_fsid
.print(val
);
1822 r
= store
->write_meta("ceph_fsid", val
);
1826 r
= store
->write_meta("ready", "ready");
1833 int OSD::peek_meta(ObjectStore
*store
, std::string
& magic
,
1834 uuid_d
& cluster_fsid
, uuid_d
& osd_fsid
, int& whoami
)
1838 int r
= store
->read_meta("magic", &val
);
1843 r
= store
->read_meta("whoami", &val
);
1846 whoami
= atoi(val
.c_str());
1848 r
= store
->read_meta("ceph_fsid", &val
);
1851 r
= cluster_fsid
.parse(val
.c_str());
1855 r
= store
->read_meta("fsid", &val
);
1857 osd_fsid
= uuid_d();
1859 r
= osd_fsid
.parse(val
.c_str());
1869 #define dout_prefix _prefix(_dout, whoami, get_osdmap_epoch())
1873 OSD::OSD(CephContext
*cct_
, ObjectStore
*store_
,
1875 Messenger
*internal_messenger
,
1876 Messenger
*external_messenger
,
1877 Messenger
*hb_client_front
,
1878 Messenger
*hb_client_back
,
1879 Messenger
*hb_front_serverm
,
1880 Messenger
*hb_back_serverm
,
1881 Messenger
*osdc_messenger
,
1883 const std::string
&dev
, const std::string
&jdev
) :
1885 osd_lock("OSD::osd_lock"),
1886 tick_timer(cct
, osd_lock
),
1887 tick_timer_lock("OSD::tick_timer_lock"),
1888 tick_timer_without_osd_lock(cct
, tick_timer_lock
),
1889 authorize_handler_cluster_registry(new AuthAuthorizeHandlerRegistry(cct
,
1890 cct
->_conf
->auth_supported
.empty() ?
1891 cct
->_conf
->auth_cluster_required
:
1892 cct
->_conf
->auth_supported
)),
1893 authorize_handler_service_registry(new AuthAuthorizeHandlerRegistry(cct
,
1894 cct
->_conf
->auth_supported
.empty() ?
1895 cct
->_conf
->auth_service_required
:
1896 cct
->_conf
->auth_supported
)),
1897 cluster_messenger(internal_messenger
),
1898 client_messenger(external_messenger
),
1899 objecter_messenger(osdc_messenger
),
1901 mgrc(cct_
, client_messenger
),
1903 recoverystate_perf(NULL
),
1905 log_client(cct
, client_messenger
, &mc
->monmap
, LogClient::NO_FLAGS
),
1906 clog(log_client
.create_channel()),
1908 dev_path(dev
), journal_path(jdev
),
1909 store_is_rotational(store
->is_rotational()),
1910 trace_endpoint("0.0.0.0", 0, "osd"),
1912 osd_compat(get_osd_compat_set()),
1913 peering_tp(cct
, "OSD::peering_tp", "tp_peering",
1914 cct
->_conf
->osd_peering_wq_threads
,
1915 "osd_peering_tp_threads"),
1916 osd_op_tp(cct
, "OSD::osd_op_tp", "tp_osd_tp",
1917 get_num_op_threads()),
1918 disk_tp(cct
, "OSD::disk_tp", "tp_osd_disk", cct
->_conf
->osd_disk_threads
, "osd_disk_threads"),
1919 command_tp(cct
, "OSD::command_tp", "tp_osd_cmd", 1),
1920 session_waiting_lock("OSD::session_waiting_lock"),
1921 osdmap_subscribe_lock("OSD::osdmap_subscribe_lock"),
1922 heartbeat_lock("OSD::heartbeat_lock"),
1923 heartbeat_stop(false),
1924 heartbeat_need_update(true),
1925 hb_front_client_messenger(hb_client_front
),
1926 hb_back_client_messenger(hb_client_back
),
1927 hb_front_server_messenger(hb_front_serverm
),
1928 hb_back_server_messenger(hb_back_serverm
),
1930 heartbeat_thread(this),
1931 heartbeat_dispatcher(this),
1932 op_tracker(cct
, cct
->_conf
->osd_enable_op_tracker
,
1933 cct
->_conf
->osd_num_op_tracker_shard
),
1934 test_ops_hook(NULL
),
1935 op_queue(get_io_queue()),
1936 op_prio_cutoff(get_io_prio_cut()),
1938 get_num_op_shards(),
1940 cct
->_conf
->osd_op_thread_timeout
,
1941 cct
->_conf
->osd_op_thread_suicide_timeout
,
1945 cct
->_conf
->osd_op_thread_timeout
,
1946 cct
->_conf
->osd_op_thread_suicide_timeout
,
1948 map_lock("OSD::map_lock"),
1949 pg_map_lock("OSD::pg_map_lock"),
1950 last_pg_create_epoch(0),
1951 mon_report_lock("OSD::mon_report_lock"),
1952 stats_ack_timeout(cct
->_conf
->osd_mon_ack_timeout
),
1954 requested_full_first(0),
1955 requested_full_last(0),
1956 pg_stat_queue_lock("OSD::pg_stat_queue_lock"),
1957 osd_stat_updated(false),
1958 pg_stat_tid(0), pg_stat_tid_flushed(0),
1961 cct
->_conf
->osd_command_thread_timeout
,
1962 cct
->_conf
->osd_command_thread_suicide_timeout
,
1967 cct
->_conf
->osd_remove_thread_timeout
,
1968 cct
->_conf
->osd_remove_thread_suicide_timeout
,
1972 monc
->set_messenger(client_messenger
);
1973 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
1974 cct
->_conf
->osd_op_log_threshold
);
1975 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
1976 cct
->_conf
->osd_op_history_duration
);
1977 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
1978 cct
->_conf
->osd_op_history_slow_op_threshold
);
1980 std::stringstream ss
;
1981 ss
<< "osd." << whoami
;
1982 trace_endpoint
.copy_name(ss
.str());
1988 delete authorize_handler_cluster_registry
;
1989 delete authorize_handler_service_registry
;
1990 delete class_handler
;
1991 cct
->get_perfcounters_collection()->remove(recoverystate_perf
);
1992 cct
->get_perfcounters_collection()->remove(logger
);
1993 delete recoverystate_perf
;
1998 void cls_initialize(ClassHandler
*ch
);
2000 void OSD::handle_signal(int signum
)
2002 assert(signum
== SIGINT
|| signum
== SIGTERM
);
2003 derr
<< "*** Got signal " << sig_str(signum
) << " ***" << dendl
;
2009 Mutex::Locker
lock(osd_lock
);
2013 if (store
->test_mount_in_use()) {
2014 derr
<< "OSD::pre_init: object store '" << dev_path
<< "' is "
2015 << "currently in use. (Is ceph-osd already running?)" << dendl
;
2019 cct
->_conf
->add_observer(this);
2025 class OSDSocketHook
: public AdminSocketHook
{
2028 explicit OSDSocketHook(OSD
*o
) : osd(o
) {}
2029 bool call(std::string admin_command
, cmdmap_t
& cmdmap
, std::string format
,
2030 bufferlist
& out
) override
{
2032 bool r
= osd
->asok_command(admin_command
, cmdmap
, format
, ss
);
2038 bool OSD::asok_command(string admin_command
, cmdmap_t
& cmdmap
, string format
,
2041 Formatter
*f
= Formatter::create(format
, "json-pretty", "json-pretty");
2042 if (admin_command
== "status") {
2043 f
->open_object_section("status");
2044 f
->dump_stream("cluster_fsid") << superblock
.cluster_fsid
;
2045 f
->dump_stream("osd_fsid") << superblock
.osd_fsid
;
2046 f
->dump_unsigned("whoami", superblock
.whoami
);
2047 f
->dump_string("state", get_state_name(get_state()));
2048 f
->dump_unsigned("oldest_map", superblock
.oldest_map
);
2049 f
->dump_unsigned("newest_map", superblock
.newest_map
);
2051 RWLock::RLocker
l(pg_map_lock
);
2052 f
->dump_unsigned("num_pgs", pg_map
.size());
2055 } else if (admin_command
== "flush_journal") {
2056 store
->flush_journal();
2057 } else if (admin_command
== "dump_ops_in_flight" ||
2058 admin_command
== "ops" ||
2059 admin_command
== "dump_blocked_ops" ||
2060 admin_command
== "dump_historic_ops" ||
2061 admin_command
== "dump_historic_ops_by_duration" ||
2062 admin_command
== "dump_historic_slow_ops") {
2064 const string error_str
= "op_tracker tracking is not enabled now, so no ops are tracked currently, \
2065 even those get stuck. Please enable \"osd_enable_op_tracker\", and the tracker \
2066 will start to track new ops received afterwards.";
2068 set
<string
> filters
;
2069 vector
<string
> filter_str
;
2070 if (cmd_getval(cct
, cmdmap
, "filterstr", filter_str
)) {
2071 copy(filter_str
.begin(), filter_str
.end(),
2072 inserter(filters
, filters
.end()));
2075 if (admin_command
== "dump_ops_in_flight" ||
2076 admin_command
== "ops") {
2077 if (!op_tracker
.dump_ops_in_flight(f
, false, filters
)) {
2081 if (admin_command
== "dump_blocked_ops") {
2082 if (!op_tracker
.dump_ops_in_flight(f
, true, filters
)) {
2086 if (admin_command
== "dump_historic_ops") {
2087 if (!op_tracker
.dump_historic_ops(f
, false, filters
)) {
2091 if (admin_command
== "dump_historic_ops_by_duration") {
2092 if (!op_tracker
.dump_historic_ops(f
, true, filters
)) {
2096 if (admin_command
== "dump_historic_slow_ops") {
2097 if (!op_tracker
.dump_historic_slow_ops(f
, filters
)) {
2101 } else if (admin_command
== "dump_op_pq_state") {
2102 f
->open_object_section("pq");
2103 op_shardedwq
.dump(f
);
2105 } else if (admin_command
== "dump_blacklist") {
2106 list
<pair
<entity_addr_t
,utime_t
> > bl
;
2107 OSDMapRef curmap
= service
.get_osdmap();
2109 f
->open_array_section("blacklist");
2110 curmap
->get_blacklist(&bl
);
2111 for (list
<pair
<entity_addr_t
,utime_t
> >::iterator it
= bl
.begin();
2112 it
!= bl
.end(); ++it
) {
2113 f
->open_object_section("entry");
2114 f
->open_object_section("entity_addr_t");
2116 f
->close_section(); //entity_addr_t
2117 it
->second
.localtime(f
->dump_stream("expire_time"));
2118 f
->close_section(); //entry
2120 f
->close_section(); //blacklist
2121 } else if (admin_command
== "dump_watchers") {
2122 list
<obj_watch_item_t
> watchers
;
2125 Mutex::Locker
l(osd_lock
);
2126 RWLock::RLocker
l2(pg_map_lock
);
2127 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2131 list
<obj_watch_item_t
> pg_watchers
;
2132 PG
*pg
= it
->second
;
2134 pg
->get_watchers(pg_watchers
);
2136 watchers
.splice(watchers
.end(), pg_watchers
);
2140 f
->open_array_section("watchers");
2141 for (list
<obj_watch_item_t
>::iterator it
= watchers
.begin();
2142 it
!= watchers
.end(); ++it
) {
2144 f
->open_object_section("watch");
2146 f
->dump_string("namespace", it
->obj
.nspace
);
2147 f
->dump_string("object", it
->obj
.oid
.name
);
2149 f
->open_object_section("entity_name");
2150 it
->wi
.name
.dump(f
);
2151 f
->close_section(); //entity_name_t
2153 f
->dump_unsigned("cookie", it
->wi
.cookie
);
2154 f
->dump_unsigned("timeout", it
->wi
.timeout_seconds
);
2156 f
->open_object_section("entity_addr_t");
2157 it
->wi
.addr
.dump(f
);
2158 f
->close_section(); //entity_addr_t
2160 f
->close_section(); //watch
2163 f
->close_section(); //watchers
2164 } else if (admin_command
== "dump_reservations") {
2165 f
->open_object_section("reservations");
2166 f
->open_object_section("local_reservations");
2167 service
.local_reserver
.dump(f
);
2169 f
->open_object_section("remote_reservations");
2170 service
.remote_reserver
.dump(f
);
2173 } else if (admin_command
== "get_latest_osdmap") {
2174 get_latest_osdmap();
2175 } else if (admin_command
== "heap") {
2176 auto result
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ss
);
2178 // Note: Failed heap profile commands won't necessarily trigger an error:
2179 f
->open_object_section("result");
2180 f
->dump_string("error", cpp_strerror(result
));
2181 f
->dump_bool("success", result
>= 0);
2183 } else if (admin_command
== "set_heap_property") {
2187 bool success
= false;
2188 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2189 error
= "unable to get property";
2191 } else if (!cmd_getval(cct
, cmdmap
, "value", value
)) {
2192 error
= "unable to get value";
2194 } else if (value
< 0) {
2195 error
= "negative value not allowed";
2197 } else if (!ceph_heap_set_numeric_property(property
.c_str(), (size_t)value
)) {
2198 error
= "invalid property";
2203 f
->open_object_section("result");
2204 f
->dump_string("error", error
);
2205 f
->dump_bool("success", success
);
2207 } else if (admin_command
== "get_heap_property") {
2211 bool success
= false;
2212 if (!cmd_getval(cct
, cmdmap
, "property", property
)) {
2213 error
= "unable to get property";
2215 } else if (!ceph_heap_get_numeric_property(property
.c_str(), &value
)) {
2216 error
= "invalid property";
2221 f
->open_object_section("result");
2222 f
->dump_string("error", error
);
2223 f
->dump_bool("success", success
);
2224 f
->dump_int("value", value
);
2226 } else if (admin_command
== "dump_objectstore_kv_stats") {
2227 store
->get_db_statistics(f
);
2228 } else if (admin_command
== "dump_scrubs") {
2229 service
.dumps_scrub(f
);
2230 } else if (admin_command
== "calc_objectstore_db_histogram") {
2231 store
->generate_db_histogram(f
);
2232 } else if (admin_command
== "flush_store_cache") {
2233 store
->flush_cache();
2234 } else if (admin_command
== "dump_pgstate_history") {
2235 f
->open_object_section("pgstate_history");
2236 RWLock::RLocker
l2(pg_map_lock
);
2237 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
2241 PG
*pg
= it
->second
;
2242 f
->dump_stream("pg") << pg
->get_pgid();
2244 pg
->pgstate_history
.dump(f
);
2248 } else if (admin_command
== "compact") {
2249 dout(1) << "triggering manual compaction" << dendl
;
2250 auto start
= ceph::coarse_mono_clock::now();
2252 auto end
= ceph::coarse_mono_clock::now();
2253 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
2254 dout(1) << "finished manual compaction in "
2255 << time_span
.count()
2256 << " seconds" << dendl
;
2257 f
->open_object_section("compact_result");
2258 f
->dump_float("elapsed_time", time_span
.count());
2261 assert(0 == "broken asok registration");
2268 class TestOpsSocketHook
: public AdminSocketHook
{
2269 OSDService
*service
;
2272 TestOpsSocketHook(OSDService
*s
, ObjectStore
*st
) : service(s
), store(st
) {}
2273 bool call(std::string command
, cmdmap_t
& cmdmap
, std::string format
,
2274 bufferlist
& out
) override
{
2276 test_ops(service
, store
, command
, cmdmap
, ss
);
2280 void test_ops(OSDService
*service
, ObjectStore
*store
,
2281 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
);
2285 class OSD::C_Tick
: public Context
{
2288 explicit C_Tick(OSD
*o
) : osd(o
) {}
2289 void finish(int r
) override
{
2294 class OSD::C_Tick_WithoutOSDLock
: public Context
{
2297 explicit C_Tick_WithoutOSDLock(OSD
*o
) : osd(o
) {}
2298 void finish(int r
) override
{
2299 osd
->tick_without_osd_lock();
2303 int OSD::enable_disable_fuse(bool stop
)
2307 string mntpath
= cct
->_conf
->osd_data
+ "/fuse";
2308 if (fuse_store
&& (stop
|| !cct
->_conf
->osd_objectstore_fuse
)) {
2309 dout(1) << __func__
<< " disabling" << dendl
;
2313 r
= ::rmdir(mntpath
.c_str());
2316 derr
<< __func__
<< " failed to rmdir " << mntpath
<< ": "
2317 << cpp_strerror(r
) << dendl
;
2322 if (!fuse_store
&& cct
->_conf
->osd_objectstore_fuse
) {
2323 dout(1) << __func__
<< " enabling" << dendl
;
2324 r
= ::mkdir(mntpath
.c_str(), 0700);
2327 if (r
< 0 && r
!= -EEXIST
) {
2328 derr
<< __func__
<< " unable to create " << mntpath
<< ": "
2329 << cpp_strerror(r
) << dendl
;
2332 fuse_store
= new FuseStore(store
, mntpath
);
2333 r
= fuse_store
->start();
2335 derr
<< __func__
<< " unable to start fuse: " << cpp_strerror(r
) << dendl
;
2341 #endif // HAVE_LIBFUSE
2345 int OSD::get_num_op_shards()
2347 if (cct
->_conf
->osd_op_num_shards
)
2348 return cct
->_conf
->osd_op_num_shards
;
2349 if (store_is_rotational
)
2350 return cct
->_conf
->osd_op_num_shards_hdd
;
2352 return cct
->_conf
->osd_op_num_shards_ssd
;
2355 int OSD::get_num_op_threads()
2357 if (cct
->_conf
->osd_op_num_threads_per_shard
)
2358 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard
;
2359 if (store_is_rotational
)
2360 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_hdd
;
2362 return get_num_op_shards() * cct
->_conf
->osd_op_num_threads_per_shard_ssd
;
2365 float OSD::get_osd_recovery_sleep()
2367 if (cct
->_conf
->osd_recovery_sleep
)
2368 return cct
->_conf
->osd_recovery_sleep
;
2369 if (!store_is_rotational
&& !journal_is_rotational
)
2370 return cct
->_conf
->osd_recovery_sleep_ssd
;
2371 else if (store_is_rotational
&& !journal_is_rotational
)
2372 return cct
->_conf
->get_val
<double>("osd_recovery_sleep_hybrid");
2374 return cct
->_conf
->osd_recovery_sleep_hdd
;
2379 CompatSet initial
, diff
;
2380 Mutex::Locker
lock(osd_lock
);
2385 tick_timer_without_osd_lock
.init();
2386 service
.recovery_request_timer
.init();
2387 service
.recovery_sleep_timer
.init();
2390 dout(2) << "init " << dev_path
2391 << " (looks like " << (store_is_rotational
? "hdd" : "ssd") << ")"
2393 dout(2) << "journal " << journal_path
<< dendl
;
2394 assert(store
); // call pre_init() first!
2396 store
->set_cache_shards(get_num_op_shards());
2398 int r
= store
->mount();
2400 derr
<< "OSD:init: unable to mount object store" << dendl
;
2403 journal_is_rotational
= store
->is_journal_rotational();
2404 dout(2) << "journal looks like " << (journal_is_rotational
? "hdd" : "ssd")
2407 enable_disable_fuse(false);
2409 dout(2) << "boot" << dendl
;
2411 // initialize the daily loadavg with current 15min loadavg
2413 if (getloadavg(loadavgs
, 3) == 3) {
2414 daily_loadavg
= loadavgs
[2];
2416 derr
<< "OSD::init() : couldn't read loadavgs\n" << dendl
;
2417 daily_loadavg
= 1.0;
2420 int rotating_auth_attempts
= 0;
2422 // sanity check long object name handling
2425 l
.oid
.name
= string(cct
->_conf
->osd_max_object_name_len
, 'n');
2426 l
.set_key(string(cct
->_conf
->osd_max_object_name_len
, 'k'));
2427 l
.nspace
= string(cct
->_conf
->osd_max_object_namespace_len
, 's');
2428 r
= store
->validate_hobject_key(l
);
2430 derr
<< "backend (" << store
->get_type() << ") is unable to support max "
2431 << "object name[space] len" << dendl
;
2432 derr
<< " osd max object name len = "
2433 << cct
->_conf
->osd_max_object_name_len
<< dendl
;
2434 derr
<< " osd max object namespace len = "
2435 << cct
->_conf
->osd_max_object_namespace_len
<< dendl
;
2436 derr
<< cpp_strerror(r
) << dendl
;
2437 if (cct
->_conf
->osd_check_max_object_name_len_on_startup
) {
2440 derr
<< "osd_check_max_object_name_len_on_startup = false, starting anyway"
2443 dout(20) << "configured osd_max_object_name[space]_len looks ok" << dendl
;
2448 r
= read_superblock();
2450 derr
<< "OSD::init() : unable to read osd superblock" << dendl
;
2455 if (osd_compat
.compare(superblock
.compat_features
) < 0) {
2456 derr
<< "The disk uses features unsupported by the executable." << dendl
;
2457 derr
<< " ondisk features " << superblock
.compat_features
<< dendl
;
2458 derr
<< " daemon features " << osd_compat
<< dendl
;
2460 if (osd_compat
.writeable(superblock
.compat_features
)) {
2461 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2462 derr
<< "it is still writeable, though. Missing features: " << diff
<< dendl
;
2467 CompatSet diff
= osd_compat
.unsupported(superblock
.compat_features
);
2468 derr
<< "Cannot write to disk! Missing features: " << diff
<< dendl
;
2474 assert_warn(whoami
== superblock
.whoami
);
2475 if (whoami
!= superblock
.whoami
) {
2476 derr
<< "OSD::init: superblock says osd"
2477 << superblock
.whoami
<< " but I am osd." << whoami
<< dendl
;
2482 initial
= get_osd_initial_compat_set();
2483 diff
= superblock
.compat_features
.unsupported(initial
);
2484 if (superblock
.compat_features
.merge(initial
)) {
2485 // We need to persist the new compat_set before we
2487 dout(5) << "Upgrading superblock adding: " << diff
<< dendl
;
2488 ObjectStore::Transaction t
;
2489 write_superblock(t
);
2490 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2495 // make sure snap mapper object exists
2496 if (!store
->exists(coll_t::meta(), OSD::make_snapmapper_oid())) {
2497 dout(10) << "init creating/touching snapmapper object" << dendl
;
2498 ObjectStore::Transaction t
;
2499 t
.touch(coll_t::meta(), OSD::make_snapmapper_oid());
2500 r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
2505 class_handler
= new ClassHandler(cct
);
2506 cls_initialize(class_handler
);
2508 if (cct
->_conf
->osd_open_classes_on_start
) {
2509 int r
= class_handler
->open_all_classes();
2511 dout(1) << "warning: got an error loading one or more classes: " << cpp_strerror(r
) << dendl
;
2514 // load up "current" osdmap
2515 assert_warn(!osdmap
);
2517 derr
<< "OSD::init: unable to read current osdmap" << dendl
;
2521 osdmap
= get_map(superblock
.current_epoch
);
2522 check_osdmap_features(store
);
2524 create_recoverystate_perf();
2527 epoch_t bind_epoch
= osdmap
->get_epoch();
2528 service
.set_epochs(NULL
, NULL
, &bind_epoch
);
2531 clear_temp_objects();
2533 // initialize osdmap references in sharded wq
2534 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
2536 // load up pgs (as they previously existed)
2539 dout(2) << "superblock: I am osd." << superblock
.whoami
<< dendl
;
2540 dout(0) << "using " << op_queue
<< " op queue with priority op cut off at " <<
2541 op_prio_cutoff
<< "." << dendl
;
2546 client_messenger
->add_dispatcher_head(this);
2547 cluster_messenger
->add_dispatcher_head(this);
2549 hb_front_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2550 hb_back_client_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2551 hb_front_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2552 hb_back_server_messenger
->add_dispatcher_head(&heartbeat_dispatcher
);
2554 objecter_messenger
->add_dispatcher_head(service
.objecter
);
2556 monc
->set_want_keys(CEPH_ENTITY_TYPE_MON
| CEPH_ENTITY_TYPE_OSD
2557 | CEPH_ENTITY_TYPE_MGR
);
2563 * FIXME: this is a placeholder implementation that unconditionally
2564 * sends every is_primary PG's stats every time we're called, unlike
2565 * the existing mon PGStats mechanism that uses pg_stat_queue and acks.
2566 * This has equivalent cost to the existing worst case where all
2567 * PGs are busy and their stats are always enqueued for sending.
2569 mgrc
.set_pgstats_cb([this](){
2570 RWLock::RLocker
l(map_lock
);
2572 utime_t had_for
= ceph_clock_now() - had_map_since
;
2573 osd_stat_t cur_stat
= service
.get_osd_stat();
2574 cur_stat
.os_perf_stat
= store
->get_cur_stats();
2576 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
2577 m
->osd_stat
= cur_stat
;
2579 Mutex::Locker lec
{min_last_epoch_clean_lock
};
2580 min_last_epoch_clean
= osdmap
->get_epoch();
2581 min_last_epoch_clean_pgs
.clear();
2582 RWLock::RLocker
lpg(pg_map_lock
);
2583 for (const auto &i
: pg_map
) {
2585 if (!pg
->is_primary()) {
2589 pg
->pg_stats_publish_lock
.Lock();
2590 if (pg
->pg_stats_publish_valid
) {
2591 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
2592 const auto lec
= pg
->pg_stats_publish
.get_effective_last_epoch_clean();
2593 min_last_epoch_clean
= min(min_last_epoch_clean
, lec
);
2594 min_last_epoch_clean_pgs
.push_back(pg
->info
.pgid
.pgid
);
2596 pg
->pg_stats_publish_lock
.Unlock();
2603 client_messenger
->add_dispatcher_head(&mgrc
);
2605 // tell monc about log_client so it will know about mon session resets
2606 monc
->set_log_client(&log_client
);
2607 update_log_config();
2614 set_disk_tp_priority();
2616 // start the heartbeat
2617 heartbeat_thread
.create("osd_srv_heartbt");
2620 tick_timer
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick(this));
2622 Mutex::Locker
l(tick_timer_lock
);
2623 tick_timer_without_osd_lock
.add_event_after(cct
->_conf
->osd_heartbeat_interval
, new C_Tick_WithoutOSDLock(this));
2627 service
.publish_map(osdmap
);
2628 service
.publish_superblock(superblock
);
2629 service
.max_oldest_map
= superblock
.oldest_map
;
2633 r
= monc
->authenticate();
2635 derr
<< __func__
<< " authentication failed: " << cpp_strerror(r
)
2637 osd_lock
.Lock(); // locker is going to unlock this on function exit
2643 while (monc
->wait_auth_rotating(30.0) < 0) {
2644 derr
<< "unable to obtain rotating service keys; retrying" << dendl
;
2645 ++rotating_auth_attempts
;
2646 if (rotating_auth_attempts
> g_conf
->max_rotating_auth_attempts
) {
2647 derr
<< __func__
<< " wait_auth_rotating timed out" << dendl
;
2648 osd_lock
.Lock(); // make locker happy
2649 if (!is_stopping()) {
2656 r
= update_crush_device_class();
2658 derr
<< __func__
<< " unable to update_crush_device_class: "
2659 << cpp_strerror(r
) << dendl
;
2664 r
= update_crush_location();
2666 derr
<< __func__
<< " unable to update_crush_location: "
2667 << cpp_strerror(r
) << dendl
;
2676 // start objecter *after* we have authenticated, so that we don't ignore
2677 // the OSDMaps it requests.
2678 service
.final_init();
2682 dout(10) << "ensuring pgs have consumed prior maps" << dendl
;
2686 dout(0) << "done with init, starting boot process" << dendl
;
2688 // subscribe to any pg creations
2689 monc
->sub_want("osd_pg_creates", last_pg_create_epoch
, 0);
2691 // MgrClient needs this (it doesn't have MonClient reference itself)
2692 monc
->sub_want("mgrmap", 0, 0);
2694 // we don't need to ask for an osdmap here; objecter will
2695 //monc->sub_want("osdmap", osdmap->get_epoch(), CEPH_SUBSCRIBE_ONETIME);
2706 enable_disable_fuse(true);
2713 void OSD::final_init()
2715 AdminSocket
*admin_socket
= cct
->get_admin_socket();
2716 asok_hook
= new OSDSocketHook(this);
2717 int r
= admin_socket
->register_command("status", "status", asok_hook
,
2718 "high-level status of OSD");
2720 r
= admin_socket
->register_command("flush_journal", "flush_journal",
2722 "flush the journal to permanent store");
2724 r
= admin_socket
->register_command("dump_ops_in_flight",
2725 "dump_ops_in_flight " \
2726 "name=filterstr,type=CephString,n=N,req=false",
2728 "show the ops currently in flight");
2730 r
= admin_socket
->register_command("ops",
2732 "name=filterstr,type=CephString,n=N,req=false",
2734 "show the ops currently in flight");
2736 r
= admin_socket
->register_command("dump_blocked_ops",
2737 "dump_blocked_ops " \
2738 "name=filterstr,type=CephString,n=N,req=false",
2740 "show the blocked ops currently in flight");
2742 r
= admin_socket
->register_command("dump_historic_ops",
2743 "dump_historic_ops " \
2744 "name=filterstr,type=CephString,n=N,req=false",
2748 r
= admin_socket
->register_command("dump_historic_slow_ops",
2749 "dump_historic_slow_ops " \
2750 "name=filterstr,type=CephString,n=N,req=false",
2752 "show slowest recent ops");
2754 r
= admin_socket
->register_command("dump_historic_ops_by_duration",
2755 "dump_historic_ops_by_duration " \
2756 "name=filterstr,type=CephString,n=N,req=false",
2758 "show slowest recent ops, sorted by duration");
2760 r
= admin_socket
->register_command("dump_op_pq_state", "dump_op_pq_state",
2762 "dump op priority queue state");
2764 r
= admin_socket
->register_command("dump_blacklist", "dump_blacklist",
2766 "dump blacklisted clients and times");
2768 r
= admin_socket
->register_command("dump_watchers", "dump_watchers",
2770 "show clients which have active watches,"
2771 " and on which objects");
2773 r
= admin_socket
->register_command("dump_reservations", "dump_reservations",
2775 "show recovery reservations");
2777 r
= admin_socket
->register_command("get_latest_osdmap", "get_latest_osdmap",
2779 "force osd to update the latest map from "
2783 r
= admin_socket
->register_command( "heap",
2785 "name=heapcmd,type=CephString",
2787 "show heap usage info (available only if "
2788 "compiled with tcmalloc)");
2791 r
= admin_socket
->register_command("set_heap_property",
2792 "set_heap_property " \
2793 "name=property,type=CephString " \
2794 "name=value,type=CephInt",
2796 "update malloc extension heap property");
2799 r
= admin_socket
->register_command("get_heap_property",
2800 "get_heap_property " \
2801 "name=property,type=CephString",
2803 "get malloc extension heap property");
2806 r
= admin_socket
->register_command("dump_objectstore_kv_stats",
2807 "dump_objectstore_kv_stats",
2809 "print statistics of kvdb which used by bluestore");
2812 r
= admin_socket
->register_command("dump_scrubs",
2815 "print scheduled scrubs");
2818 r
= admin_socket
->register_command("calc_objectstore_db_histogram",
2819 "calc_objectstore_db_histogram",
2821 "Generate key value histogram of kvdb(rocksdb) which used by bluestore");
2824 r
= admin_socket
->register_command("flush_store_cache",
2825 "flush_store_cache",
2827 "Flush bluestore internal cache");
2829 r
= admin_socket
->register_command("dump_pgstate_history", "dump_pgstate_history",
2831 "show recent state history");
2834 r
= admin_socket
->register_command("compact", "compact",
2836 "Commpact object store's omap."
2837 " WARNING: Compaction probably slows your requests");
2840 test_ops_hook
= new TestOpsSocketHook(&(this->service
), this->store
);
2841 // Note: pools are CephString instead of CephPoolname because
2842 // these commands traditionally support both pool names and numbers
2843 r
= admin_socket
->register_command(
2846 "name=pool,type=CephString " \
2847 "name=objname,type=CephObjectname " \
2848 "name=key,type=CephString "\
2849 "name=val,type=CephString",
2853 r
= admin_socket
->register_command(
2856 "name=pool,type=CephString " \
2857 "name=objname,type=CephObjectname " \
2858 "name=key,type=CephString",
2862 r
= admin_socket
->register_command(
2865 "name=pool,type=CephString " \
2866 "name=objname,type=CephObjectname " \
2867 "name=header,type=CephString",
2872 r
= admin_socket
->register_command(
2875 "name=pool,type=CephString " \
2876 "name=objname,type=CephObjectname",
2878 "output entire object map");
2881 r
= admin_socket
->register_command(
2884 "name=pool,type=CephString " \
2885 "name=objname,type=CephObjectname " \
2886 "name=len,type=CephInt",
2888 "truncate object to length");
2891 r
= admin_socket
->register_command(
2894 "name=pool,type=CephString " \
2895 "name=objname,type=CephObjectname " \
2896 "name=shardid,type=CephInt,req=false,range=0|255",
2898 "inject data error to an object");
2901 r
= admin_socket
->register_command(
2904 "name=pool,type=CephString " \
2905 "name=objname,type=CephObjectname " \
2906 "name=shardid,type=CephInt,req=false,range=0|255",
2908 "inject metadata error to an object");
2910 r
= admin_socket
->register_command(
2911 "set_recovery_delay",
2912 "set_recovery_delay " \
2913 "name=utime,type=CephInt,req=false",
2915 "Delay osd recovery by specified seconds");
2917 r
= admin_socket
->register_command(
2920 "name=pgid,type=CephString ",
2922 "Trigger a scheduled scrub ");
2924 r
= admin_socket
->register_command(
2927 "name=type,type=CephString,req=false " \
2928 "name=count,type=CephInt,req=false ",
2930 "Inject a full disk (optional count times)");
2934 void OSD::create_logger()
2936 dout(10) << "create_logger" << dendl
;
2938 PerfCountersBuilder
osd_plb(cct
, "osd", l_osd_first
, l_osd_last
);
2940 // Latency axis configuration for op histograms, values are in nanoseconds
2941 PerfHistogramCommon::axis_config_d op_hist_x_axis_config
{
2943 PerfHistogramCommon::SCALE_LOG2
, ///< Latency in logarithmic scale
2945 100000, ///< Quantization unit is 100usec
2946 32, ///< Enough to cover much longer than slow requests
2949 // Op size axis configuration for op histograms, values are in bytes
2950 PerfHistogramCommon::axis_config_d op_hist_y_axis_config
{
2951 "Request size (bytes)",
2952 PerfHistogramCommon::SCALE_LOG2
, ///< Request size in logarithmic scale
2954 512, ///< Quantization unit is 512 bytes
2955 32, ///< Enough to cover requests larger than GB
2960 l_osd_op_wip
, "op_wip",
2961 "Replication operations currently being processed (primary)");
2962 osd_plb
.add_u64_counter(
2964 "Client operations",
2965 "ops", PerfCountersBuilder::PRIO_CRITICAL
);
2966 osd_plb
.add_u64_counter(
2967 l_osd_op_inb
, "op_in_bytes",
2968 "Client operations total write size",
2969 "wr", PerfCountersBuilder::PRIO_INTERESTING
);
2970 osd_plb
.add_u64_counter(
2971 l_osd_op_outb
, "op_out_bytes",
2972 "Client operations total read size",
2973 "rd", PerfCountersBuilder::PRIO_INTERESTING
);
2974 osd_plb
.add_time_avg(
2975 l_osd_op_lat
, "op_latency",
2976 "Latency of client operations (including queue time)",
2978 osd_plb
.add_time_avg(
2979 l_osd_op_process_lat
, "op_process_latency",
2980 "Latency of client operations (excluding queue time)");
2981 osd_plb
.add_time_avg(
2982 l_osd_op_prepare_lat
, "op_prepare_latency",
2983 "Latency of client operations (excluding queue time and wait for finished)");
2985 osd_plb
.add_u64_counter(
2986 l_osd_op_r
, "op_r", "Client read operations");
2987 osd_plb
.add_u64_counter(
2988 l_osd_op_r_outb
, "op_r_out_bytes", "Client data read");
2989 osd_plb
.add_time_avg(
2990 l_osd_op_r_lat
, "op_r_latency",
2991 "Latency of read operation (including queue time)");
2992 osd_plb
.add_u64_counter_histogram(
2993 l_osd_op_r_lat_outb_hist
, "op_r_latency_out_bytes_histogram",
2994 op_hist_x_axis_config
, op_hist_y_axis_config
,
2995 "Histogram of operation latency (including queue time) + data read");
2996 osd_plb
.add_time_avg(
2997 l_osd_op_r_process_lat
, "op_r_process_latency",
2998 "Latency of read operation (excluding queue time)");
2999 osd_plb
.add_time_avg(
3000 l_osd_op_r_prepare_lat
, "op_r_prepare_latency",
3001 "Latency of read operations (excluding queue time and wait for finished)");
3002 osd_plb
.add_u64_counter(
3003 l_osd_op_w
, "op_w", "Client write operations");
3004 osd_plb
.add_u64_counter(
3005 l_osd_op_w_inb
, "op_w_in_bytes", "Client data written");
3006 osd_plb
.add_time_avg(
3007 l_osd_op_w_lat
, "op_w_latency",
3008 "Latency of write operation (including queue time)");
3009 osd_plb
.add_u64_counter_histogram(
3010 l_osd_op_w_lat_inb_hist
, "op_w_latency_in_bytes_histogram",
3011 op_hist_x_axis_config
, op_hist_y_axis_config
,
3012 "Histogram of operation latency (including queue time) + data written");
3013 osd_plb
.add_time_avg(
3014 l_osd_op_w_process_lat
, "op_w_process_latency",
3015 "Latency of write operation (excluding queue time)");
3016 osd_plb
.add_time_avg(
3017 l_osd_op_w_prepare_lat
, "op_w_prepare_latency",
3018 "Latency of write operations (excluding queue time and wait for finished)");
3019 osd_plb
.add_u64_counter(
3020 l_osd_op_rw
, "op_rw",
3021 "Client read-modify-write operations");
3022 osd_plb
.add_u64_counter(
3023 l_osd_op_rw_inb
, "op_rw_in_bytes",
3024 "Client read-modify-write operations write in");
3025 osd_plb
.add_u64_counter(
3026 l_osd_op_rw_outb
,"op_rw_out_bytes",
3027 "Client read-modify-write operations read out ");
3028 osd_plb
.add_time_avg(
3029 l_osd_op_rw_lat
, "op_rw_latency",
3030 "Latency of read-modify-write operation (including queue time)");
3031 osd_plb
.add_u64_counter_histogram(
3032 l_osd_op_rw_lat_inb_hist
, "op_rw_latency_in_bytes_histogram",
3033 op_hist_x_axis_config
, op_hist_y_axis_config
,
3034 "Histogram of rw operation latency (including queue time) + data written");
3035 osd_plb
.add_u64_counter_histogram(
3036 l_osd_op_rw_lat_outb_hist
, "op_rw_latency_out_bytes_histogram",
3037 op_hist_x_axis_config
, op_hist_y_axis_config
,
3038 "Histogram of rw operation latency (including queue time) + data read");
3039 osd_plb
.add_time_avg(
3040 l_osd_op_rw_process_lat
, "op_rw_process_latency",
3041 "Latency of read-modify-write operation (excluding queue time)");
3042 osd_plb
.add_time_avg(
3043 l_osd_op_rw_prepare_lat
, "op_rw_prepare_latency",
3044 "Latency of read-modify-write operations (excluding queue time and wait for finished)");
3046 osd_plb
.add_time_avg(l_osd_op_before_queue_op_lat
, "op_before_queue_op_lat",
3047 "Latency of IO before calling queue(before really queue into ShardedOpWq)"); // client io before queue op_wq latency
3048 osd_plb
.add_time_avg(l_osd_op_before_dequeue_op_lat
, "op_before_dequeue_op_lat",
3049 "Latency of IO before calling dequeue_op(already dequeued and get PG lock)"); // client io before dequeue_op latency
3051 osd_plb
.add_u64_counter(
3052 l_osd_sop
, "subop", "Suboperations");
3053 osd_plb
.add_u64_counter(
3054 l_osd_sop_inb
, "subop_in_bytes", "Suboperations total size");
3055 osd_plb
.add_time_avg(l_osd_sop_lat
, "subop_latency", "Suboperations latency");
3057 osd_plb
.add_u64_counter(l_osd_sop_w
, "subop_w", "Replicated writes");
3058 osd_plb
.add_u64_counter(
3059 l_osd_sop_w_inb
, "subop_w_in_bytes", "Replicated written data size");
3060 osd_plb
.add_time_avg(
3061 l_osd_sop_w_lat
, "subop_w_latency", "Replicated writes latency");
3062 osd_plb
.add_u64_counter(
3063 l_osd_sop_pull
, "subop_pull", "Suboperations pull requests");
3064 osd_plb
.add_time_avg(
3065 l_osd_sop_pull_lat
, "subop_pull_latency", "Suboperations pull latency");
3066 osd_plb
.add_u64_counter(
3067 l_osd_sop_push
, "subop_push", "Suboperations push messages");
3068 osd_plb
.add_u64_counter(
3069 l_osd_sop_push_inb
, "subop_push_in_bytes", "Suboperations pushed size");
3070 osd_plb
.add_time_avg(
3071 l_osd_sop_push_lat
, "subop_push_latency", "Suboperations push latency");
3073 osd_plb
.add_u64_counter(l_osd_pull
, "pull", "Pull requests sent");
3074 osd_plb
.add_u64_counter(l_osd_push
, "push", "Push messages sent");
3075 osd_plb
.add_u64_counter(l_osd_push_outb
, "push_out_bytes", "Pushed size");
3077 osd_plb
.add_u64_counter(
3078 l_osd_rop
, "recovery_ops",
3079 "Started recovery operations",
3080 "rop", PerfCountersBuilder::PRIO_INTERESTING
);
3082 osd_plb
.add_u64(l_osd_loadavg
, "loadavg", "CPU load");
3083 osd_plb
.add_u64(l_osd_buf
, "buffer_bytes", "Total allocated buffer size");
3084 osd_plb
.add_u64(l_osd_history_alloc_bytes
, "history_alloc_Mbytes");
3085 osd_plb
.add_u64(l_osd_history_alloc_num
, "history_alloc_num");
3087 l_osd_cached_crc
, "cached_crc", "Total number getting crc from crc_cache");
3089 l_osd_cached_crc_adjusted
, "cached_crc_adjusted",
3090 "Total number getting crc from crc_cache with adjusting");
3091 osd_plb
.add_u64(l_osd_missed_crc
, "missed_crc",
3092 "Total number of crc cache misses");
3094 osd_plb
.add_u64(l_osd_pg
, "numpg", "Placement groups",
3095 "pgs", PerfCountersBuilder::PRIO_USEFUL
);
3097 l_osd_pg_primary
, "numpg_primary",
3098 "Placement groups for which this osd is primary");
3100 l_osd_pg_replica
, "numpg_replica",
3101 "Placement groups for which this osd is replica");
3103 l_osd_pg_stray
, "numpg_stray",
3104 "Placement groups ready to be deleted from this osd");
3106 l_osd_hb_to
, "heartbeat_to_peers", "Heartbeat (ping) peers we send to");
3107 osd_plb
.add_u64_counter(l_osd_map
, "map_messages", "OSD map messages");
3108 osd_plb
.add_u64_counter(l_osd_mape
, "map_message_epochs", "OSD map epochs");
3109 osd_plb
.add_u64_counter(
3110 l_osd_mape_dup
, "map_message_epoch_dups", "OSD map duplicates");
3111 osd_plb
.add_u64_counter(
3112 l_osd_waiting_for_map
, "messages_delayed_for_map",
3113 "Operations waiting for OSD map");
3115 osd_plb
.add_u64_counter(
3116 l_osd_map_cache_hit
, "osd_map_cache_hit", "osdmap cache hit");
3117 osd_plb
.add_u64_counter(
3118 l_osd_map_cache_miss
, "osd_map_cache_miss", "osdmap cache miss");
3119 osd_plb
.add_u64_counter(
3120 l_osd_map_cache_miss_low
, "osd_map_cache_miss_low",
3121 "osdmap cache miss below cache lower bound");
3122 osd_plb
.add_u64_avg(
3123 l_osd_map_cache_miss_low_avg
, "osd_map_cache_miss_low_avg",
3124 "osdmap cache miss, avg distance below cache lower bound");
3125 osd_plb
.add_u64_counter(
3126 l_osd_map_bl_cache_hit
, "osd_map_bl_cache_hit",
3127 "OSDMap buffer cache hits");
3128 osd_plb
.add_u64_counter(
3129 l_osd_map_bl_cache_miss
, "osd_map_bl_cache_miss",
3130 "OSDMap buffer cache misses");
3132 osd_plb
.add_u64(l_osd_stat_bytes
, "stat_bytes", "OSD size");
3133 osd_plb
.add_u64(l_osd_stat_bytes_used
, "stat_bytes_used", "Used space");
3134 osd_plb
.add_u64(l_osd_stat_bytes_avail
, "stat_bytes_avail", "Available space");
3136 osd_plb
.add_u64_counter(
3137 l_osd_copyfrom
, "copyfrom", "Rados \"copy-from\" operations");
3139 osd_plb
.add_u64_counter(l_osd_tier_promote
, "tier_promote", "Tier promotions");
3140 osd_plb
.add_u64_counter(l_osd_tier_flush
, "tier_flush", "Tier flushes");
3141 osd_plb
.add_u64_counter(
3142 l_osd_tier_flush_fail
, "tier_flush_fail", "Failed tier flushes");
3143 osd_plb
.add_u64_counter(
3144 l_osd_tier_try_flush
, "tier_try_flush", "Tier flush attempts");
3145 osd_plb
.add_u64_counter(
3146 l_osd_tier_try_flush_fail
, "tier_try_flush_fail",
3147 "Failed tier flush attempts");
3148 osd_plb
.add_u64_counter(
3149 l_osd_tier_evict
, "tier_evict", "Tier evictions");
3150 osd_plb
.add_u64_counter(
3151 l_osd_tier_whiteout
, "tier_whiteout", "Tier whiteouts");
3152 osd_plb
.add_u64_counter(
3153 l_osd_tier_dirty
, "tier_dirty", "Dirty tier flag set");
3154 osd_plb
.add_u64_counter(
3155 l_osd_tier_clean
, "tier_clean", "Dirty tier flag cleaned");
3156 osd_plb
.add_u64_counter(
3157 l_osd_tier_delay
, "tier_delay", "Tier delays (agent waiting)");
3158 osd_plb
.add_u64_counter(
3159 l_osd_tier_proxy_read
, "tier_proxy_read", "Tier proxy reads");
3160 osd_plb
.add_u64_counter(
3161 l_osd_tier_proxy_write
, "tier_proxy_write", "Tier proxy writes");
3163 osd_plb
.add_u64_counter(
3164 l_osd_agent_wake
, "agent_wake", "Tiering agent wake up");
3165 osd_plb
.add_u64_counter(
3166 l_osd_agent_skip
, "agent_skip", "Objects skipped by agent");
3167 osd_plb
.add_u64_counter(
3168 l_osd_agent_flush
, "agent_flush", "Tiering agent flushes");
3169 osd_plb
.add_u64_counter(
3170 l_osd_agent_evict
, "agent_evict", "Tiering agent evictions");
3172 osd_plb
.add_u64_counter(
3173 l_osd_object_ctx_cache_hit
, "object_ctx_cache_hit", "Object context cache hits");
3174 osd_plb
.add_u64_counter(
3175 l_osd_object_ctx_cache_total
, "object_ctx_cache_total", "Object context cache lookups");
3177 osd_plb
.add_u64_counter(l_osd_op_cache_hit
, "op_cache_hit");
3178 osd_plb
.add_time_avg(
3179 l_osd_tier_flush_lat
, "osd_tier_flush_lat", "Object flush latency");
3180 osd_plb
.add_time_avg(
3181 l_osd_tier_promote_lat
, "osd_tier_promote_lat", "Object promote latency");
3182 osd_plb
.add_time_avg(
3183 l_osd_tier_r_lat
, "osd_tier_r_lat", "Object proxy read latency");
3185 osd_plb
.add_u64_counter(
3186 l_osd_pg_info
, "osd_pg_info", "PG updated its info (using any method)");
3187 osd_plb
.add_u64_counter(
3188 l_osd_pg_fastinfo
, "osd_pg_fastinfo",
3189 "PG updated its info using fastinfo attr");
3190 osd_plb
.add_u64_counter(
3191 l_osd_pg_biginfo
, "osd_pg_biginfo", "PG updated its biginfo attr");
3193 logger
= osd_plb
.create_perf_counters();
3194 cct
->get_perfcounters_collection()->add(logger
);
3197 void OSD::create_recoverystate_perf()
3199 dout(10) << "create_recoverystate_perf" << dendl
;
3201 PerfCountersBuilder
rs_perf(cct
, "recoverystate_perf", rs_first
, rs_last
);
3203 rs_perf
.add_time_avg(rs_initial_latency
, "initial_latency", "Initial recovery state latency");
3204 rs_perf
.add_time_avg(rs_started_latency
, "started_latency", "Started recovery state latency");
3205 rs_perf
.add_time_avg(rs_reset_latency
, "reset_latency", "Reset recovery state latency");
3206 rs_perf
.add_time_avg(rs_start_latency
, "start_latency", "Start recovery state latency");
3207 rs_perf
.add_time_avg(rs_primary_latency
, "primary_latency", "Primary recovery state latency");
3208 rs_perf
.add_time_avg(rs_peering_latency
, "peering_latency", "Peering recovery state latency");
3209 rs_perf
.add_time_avg(rs_backfilling_latency
, "backfilling_latency", "Backfilling recovery state latency");
3210 rs_perf
.add_time_avg(rs_waitremotebackfillreserved_latency
, "waitremotebackfillreserved_latency", "Wait remote backfill reserved recovery state latency");
3211 rs_perf
.add_time_avg(rs_waitlocalbackfillreserved_latency
, "waitlocalbackfillreserved_latency", "Wait local backfill reserved recovery state latency");
3212 rs_perf
.add_time_avg(rs_notbackfilling_latency
, "notbackfilling_latency", "Notbackfilling recovery state latency");
3213 rs_perf
.add_time_avg(rs_repnotrecovering_latency
, "repnotrecovering_latency", "Repnotrecovering recovery state latency");
3214 rs_perf
.add_time_avg(rs_repwaitrecoveryreserved_latency
, "repwaitrecoveryreserved_latency", "Rep wait recovery reserved recovery state latency");
3215 rs_perf
.add_time_avg(rs_repwaitbackfillreserved_latency
, "repwaitbackfillreserved_latency", "Rep wait backfill reserved recovery state latency");
3216 rs_perf
.add_time_avg(rs_reprecovering_latency
, "reprecovering_latency", "RepRecovering recovery state latency");
3217 rs_perf
.add_time_avg(rs_activating_latency
, "activating_latency", "Activating recovery state latency");
3218 rs_perf
.add_time_avg(rs_waitlocalrecoveryreserved_latency
, "waitlocalrecoveryreserved_latency", "Wait local recovery reserved recovery state latency");
3219 rs_perf
.add_time_avg(rs_waitremoterecoveryreserved_latency
, "waitremoterecoveryreserved_latency", "Wait remote recovery reserved recovery state latency");
3220 rs_perf
.add_time_avg(rs_recovering_latency
, "recovering_latency", "Recovering recovery state latency");
3221 rs_perf
.add_time_avg(rs_recovered_latency
, "recovered_latency", "Recovered recovery state latency");
3222 rs_perf
.add_time_avg(rs_clean_latency
, "clean_latency", "Clean recovery state latency");
3223 rs_perf
.add_time_avg(rs_active_latency
, "active_latency", "Active recovery state latency");
3224 rs_perf
.add_time_avg(rs_replicaactive_latency
, "replicaactive_latency", "Replicaactive recovery state latency");
3225 rs_perf
.add_time_avg(rs_stray_latency
, "stray_latency", "Stray recovery state latency");
3226 rs_perf
.add_time_avg(rs_getinfo_latency
, "getinfo_latency", "Getinfo recovery state latency");
3227 rs_perf
.add_time_avg(rs_getlog_latency
, "getlog_latency", "Getlog recovery state latency");
3228 rs_perf
.add_time_avg(rs_waitactingchange_latency
, "waitactingchange_latency", "Waitactingchange recovery state latency");
3229 rs_perf
.add_time_avg(rs_incomplete_latency
, "incomplete_latency", "Incomplete recovery state latency");
3230 rs_perf
.add_time_avg(rs_down_latency
, "down_latency", "Down recovery state latency");
3231 rs_perf
.add_time_avg(rs_getmissing_latency
, "getmissing_latency", "Getmissing recovery state latency");
3232 rs_perf
.add_time_avg(rs_waitupthru_latency
, "waitupthru_latency", "Waitupthru recovery state latency");
3233 rs_perf
.add_time_avg(rs_notrecovering_latency
, "notrecovering_latency", "Notrecovering recovery state latency");
3235 recoverystate_perf
= rs_perf
.create_perf_counters();
3236 cct
->get_perfcounters_collection()->add(recoverystate_perf
);
3241 if (!service
.prepare_to_stop())
3242 return 0; // already shutting down
3244 if (is_stopping()) {
3248 derr
<< "shutdown" << dendl
;
3250 set_state(STATE_STOPPING
);
3253 cct
->_conf
->set_val("debug_osd", "100");
3254 cct
->_conf
->set_val("debug_journal", "100");
3255 cct
->_conf
->set_val("debug_filestore", "100");
3256 cct
->_conf
->set_val("debug_ms", "100");
3257 cct
->_conf
->apply_changes(NULL
);
3259 // stop MgrClient earlier as it's more like an internal consumer of OSD
3262 service
.start_shutdown();
3264 // stop sending work to pgs. this just prevents any new work in _process
3265 // from racing with on_shutdown and potentially entering the pg after.
3266 op_shardedwq
.drain();
3270 RWLock::RLocker
l(pg_map_lock
);
3271 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3274 dout(20) << " kicking pg " << p
->first
<< dendl
;
3276 p
->second
->on_shutdown();
3277 p
->second
->unlock();
3278 p
->second
->osr
->flush();
3281 clear_pg_stat_queue();
3283 // drain op queue again (in case PGs requeued something)
3284 op_shardedwq
.drain();
3286 finished
.clear(); // zap waiters (bleh, this is messy)
3289 op_shardedwq
.clear_pg_slots();
3291 // unregister commands
3292 cct
->get_admin_socket()->unregister_command("status");
3293 cct
->get_admin_socket()->unregister_command("flush_journal");
3294 cct
->get_admin_socket()->unregister_command("dump_ops_in_flight");
3295 cct
->get_admin_socket()->unregister_command("ops");
3296 cct
->get_admin_socket()->unregister_command("dump_blocked_ops");
3297 cct
->get_admin_socket()->unregister_command("dump_historic_ops");
3298 cct
->get_admin_socket()->unregister_command("dump_historic_ops_by_duration");
3299 cct
->get_admin_socket()->unregister_command("dump_historic_slow_ops");
3300 cct
->get_admin_socket()->unregister_command("dump_op_pq_state");
3301 cct
->get_admin_socket()->unregister_command("dump_blacklist");
3302 cct
->get_admin_socket()->unregister_command("dump_watchers");
3303 cct
->get_admin_socket()->unregister_command("dump_reservations");
3304 cct
->get_admin_socket()->unregister_command("get_latest_osdmap");
3305 cct
->get_admin_socket()->unregister_command("heap");
3306 cct
->get_admin_socket()->unregister_command("set_heap_property");
3307 cct
->get_admin_socket()->unregister_command("get_heap_property");
3308 cct
->get_admin_socket()->unregister_command("dump_objectstore_kv_stats");
3309 cct
->get_admin_socket()->unregister_command("dump_scrubs");
3310 cct
->get_admin_socket()->unregister_command("calc_objectstore_db_histogram");
3311 cct
->get_admin_socket()->unregister_command("flush_store_cache");
3312 cct
->get_admin_socket()->unregister_command("dump_pgstate_history");
3313 cct
->get_admin_socket()->unregister_command("compact");
3317 cct
->get_admin_socket()->unregister_command("setomapval");
3318 cct
->get_admin_socket()->unregister_command("rmomapkey");
3319 cct
->get_admin_socket()->unregister_command("setomapheader");
3320 cct
->get_admin_socket()->unregister_command("getomap");
3321 cct
->get_admin_socket()->unregister_command("truncobj");
3322 cct
->get_admin_socket()->unregister_command("injectdataerr");
3323 cct
->get_admin_socket()->unregister_command("injectmdataerr");
3324 cct
->get_admin_socket()->unregister_command("set_recovery_delay");
3325 cct
->get_admin_socket()->unregister_command("trigger_scrub");
3326 cct
->get_admin_socket()->unregister_command("injectfull");
3327 delete test_ops_hook
;
3328 test_ops_hook
= NULL
;
3332 heartbeat_lock
.Lock();
3333 heartbeat_stop
= true;
3334 heartbeat_cond
.Signal();
3335 heartbeat_lock
.Unlock();
3336 heartbeat_thread
.join();
3341 dout(10) << "osd tp stopped" << dendl
;
3345 dout(10) << "op sharded tp stopped" << dendl
;
3349 dout(10) << "command tp stopped" << dendl
;
3353 dout(10) << "disk tp paused (new)" << dendl
;
3355 dout(10) << "stopping agent" << dendl
;
3356 service
.agent_stop();
3360 reset_heartbeat_peers();
3362 tick_timer
.shutdown();
3365 Mutex::Locker
l(tick_timer_lock
);
3366 tick_timer_without_osd_lock
.shutdown();
3369 // note unmount epoch
3370 dout(10) << "noting clean unmount in epoch " << osdmap
->get_epoch() << dendl
;
3371 superblock
.mounted
= service
.get_boot_epoch();
3372 superblock
.clean_thru
= osdmap
->get_epoch();
3373 ObjectStore::Transaction t
;
3374 write_superblock(t
);
3375 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3377 derr
<< "OSD::shutdown: error writing superblock: "
3378 << cpp_strerror(r
) << dendl
;
3383 Mutex::Locker
l(pg_stat_queue_lock
);
3384 assert(pg_stat_queue
.empty());
3387 service
.shutdown_reserver();
3390 #ifdef PG_DEBUG_REFS
3391 service
.dump_live_pgids();
3394 RWLock::RLocker
l(pg_map_lock
);
3395 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
3398 dout(20) << " kicking pg " << p
->first
<< dendl
;
3400 if (p
->second
->ref
!= 1) {
3401 derr
<< "pgid " << p
->first
<< " has ref count of "
3402 << p
->second
->ref
<< dendl
;
3403 #ifdef PG_DEBUG_REFS
3404 p
->second
->dump_live_ids();
3406 if (cct
->_conf
->osd_shutdown_pgref_assert
) {
3410 p
->second
->unlock();
3411 p
->second
->put("PGMap");
3415 #ifdef PG_DEBUG_REFS
3416 service
.dump_live_pgids();
3418 cct
->_conf
->remove_observer(this);
3420 dout(10) << "syncing store" << dendl
;
3421 enable_disable_fuse(true);
3423 if (cct
->_conf
->osd_journal_flush_on_shutdown
) {
3424 dout(10) << "flushing journal" << dendl
;
3425 store
->flush_journal();
3431 dout(10) << "Store synced" << dendl
;
3436 osdmap
= OSDMapRef();
3438 op_tracker
.on_shutdown();
3440 class_handler
->shutdown();
3441 client_messenger
->shutdown();
3442 cluster_messenger
->shutdown();
3443 hb_front_client_messenger
->shutdown();
3444 hb_back_client_messenger
->shutdown();
3445 objecter_messenger
->shutdown();
3446 hb_front_server_messenger
->shutdown();
3447 hb_back_server_messenger
->shutdown();
3454 int OSD::mon_cmd_maybe_osd_create(string
&cmd
)
3456 bool created
= false;
3458 dout(10) << __func__
<< " cmd: " << cmd
<< dendl
;
3459 vector
<string
> vcmd
{cmd
};
3463 monc
->start_mon_command(vcmd
, inbl
, NULL
, &outs
, &w
);
3466 if (r
== -ENOENT
&& !created
) {
3467 string newcmd
= "{\"prefix\": \"osd create\", \"id\": " + stringify(whoami
)
3468 + ", \"uuid\": \"" + stringify(superblock
.osd_fsid
) + "\"}";
3469 vector
<string
> vnewcmd
{newcmd
};
3473 monc
->start_mon_command(vnewcmd
, inbl
, NULL
, &outs
, &w
);
3476 derr
<< __func__
<< " fail: osd does not exist and created failed: "
3477 << cpp_strerror(r
) << dendl
;
3483 derr
<< __func__
<< " fail: '" << outs
<< "': " << cpp_strerror(r
) << dendl
;
3492 int OSD::update_crush_location()
3494 if (!cct
->_conf
->osd_crush_update_on_start
) {
3495 dout(10) << __func__
<< " osd_crush_update_on_start = false" << dendl
;
3500 if (cct
->_conf
->osd_crush_initial_weight
>= 0) {
3501 snprintf(weight
, sizeof(weight
), "%.4lf", cct
->_conf
->osd_crush_initial_weight
);
3503 struct store_statfs_t st
;
3504 int r
= store
->statfs(&st
);
3506 derr
<< "statfs: " << cpp_strerror(r
) << dendl
;
3509 snprintf(weight
, sizeof(weight
), "%.4lf",
3511 (double)(st
.total
) /
3512 (double)(1ull << 40 /* TB */)));
3515 std::multimap
<string
,string
> loc
= cct
->crush_location
.get_location();
3516 dout(10) << __func__
<< " crush location is " << loc
<< dendl
;
3519 string("{\"prefix\": \"osd crush create-or-move\", ") +
3520 string("\"id\": ") + stringify(whoami
) + string(", ") +
3521 string("\"weight\":") + weight
+ string(", ") +
3522 string("\"args\": [");
3523 for (multimap
<string
,string
>::iterator p
= loc
.begin(); p
!= loc
.end(); ++p
) {
3524 if (p
!= loc
.begin())
3526 cmd
+= "\"" + p
->first
+ "=" + p
->second
+ "\"";
3530 return mon_cmd_maybe_osd_create(cmd
);
3533 int OSD::update_crush_device_class()
3535 if (!cct
->_conf
->osd_class_update_on_start
) {
3536 dout(10) << __func__
<< " osd_class_update_on_start = false" << dendl
;
3540 string device_class
;
3541 int r
= store
->read_meta("crush_device_class", &device_class
);
3542 if (r
< 0 || device_class
.empty()) {
3543 device_class
= store
->get_default_device_class();
3546 if (device_class
.empty()) {
3547 dout(20) << __func__
<< " no device class stored locally" << dendl
;
3552 string("{\"prefix\": \"osd crush set-device-class\", ") +
3553 string("\"class\": \"") + device_class
+ string("\", ") +
3554 string("\"ids\": [\"") + stringify(whoami
) + string("\"]}");
3556 r
= mon_cmd_maybe_osd_create(cmd
);
3557 // the above cmd can fail for various reasons, e.g.:
3558 // (1) we are connecting to a pre-luminous monitor
3559 // (2) user manually specify a class other than
3560 // 'ceph-disk prepare --crush-device-class'
3561 // simply skip result-checking for now
3565 void OSD::write_superblock(ObjectStore::Transaction
& t
)
3567 dout(10) << "write_superblock " << superblock
<< dendl
;
3569 //hack: at minimum it's using the baseline feature set
3570 if (!superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_BASE
))
3571 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE
);
3574 ::encode(superblock
, bl
);
3575 t
.write(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, bl
.length(), bl
);
3578 int OSD::read_superblock()
3581 int r
= store
->read(coll_t::meta(), OSD_SUPERBLOCK_GOBJECT
, 0, 0, bl
);
3585 bufferlist::iterator p
= bl
.begin();
3586 ::decode(superblock
, p
);
3588 dout(10) << "read_superblock " << superblock
<< dendl
;
3593 void OSD::clear_temp_objects()
3595 dout(10) << __func__
<< dendl
;
3597 store
->list_collections(ls
);
3598 for (vector
<coll_t
>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
3600 if (!p
->is_pg(&pgid
))
3603 // list temp objects
3604 dout(20) << " clearing temps in " << *p
<< " pgid " << pgid
<< dendl
;
3606 vector
<ghobject_t
> temps
;
3609 vector
<ghobject_t
> objects
;
3610 store
->collection_list(*p
, next
, ghobject_t::get_max(),
3611 store
->get_ideal_list_max(),
3613 if (objects
.empty())
3615 vector
<ghobject_t
>::iterator q
;
3616 for (q
= objects
.begin(); q
!= objects
.end(); ++q
) {
3617 // Hammer set pool for temps to -1, so check for clean-up
3618 if (q
->hobj
.is_temp() || (q
->hobj
.pool
== -1)) {
3619 temps
.push_back(*q
);
3624 // If we saw a non-temp object and hit the break above we can
3625 // break out of the while loop too.
3626 if (q
!= objects
.end())
3629 if (!temps
.empty()) {
3630 ObjectStore::Transaction t
;
3632 for (vector
<ghobject_t
>::iterator q
= temps
.begin(); q
!= temps
.end(); ++q
) {
3633 dout(20) << " removing " << *p
<< " object " << *q
<< dendl
;
3635 if (++removed
> cct
->_conf
->osd_target_transaction_size
) {
3636 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3637 t
= ObjectStore::Transaction();
3642 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
3648 void OSD::recursive_remove_collection(CephContext
* cct
,
3649 ObjectStore
*store
, spg_t pgid
,
3655 make_snapmapper_oid());
3657 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
3658 ObjectStore::Sequencer
>("rm"));
3659 ObjectStore::Transaction t
;
3660 SnapMapper
mapper(cct
, &driver
, 0, 0, 0, pgid
.shard
);
3662 vector
<ghobject_t
> objects
;
3663 store
->collection_list(tmp
, ghobject_t(), ghobject_t::get_max(),
3664 INT_MAX
, &objects
, 0);
3665 generic_dout(10) << __func__
<< " " << objects
<< dendl
;
3668 for (vector
<ghobject_t
>::iterator p
= objects
.begin();
3671 OSDriver::OSTransaction
_t(driver
.get_transaction(&t
));
3672 int r
= mapper
.remove_oid(p
->hobj
, &_t
);
3673 if (r
!= 0 && r
!= -ENOENT
)
3676 if (removed
> cct
->_conf
->osd_target_transaction_size
) {
3677 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3679 t
= ObjectStore::Transaction();
3683 t
.remove_collection(tmp
);
3684 int r
= store
->apply_transaction(osr
.get(), std::move(t
));
3688 if (!osr
->flush_commit(&waiter
)) {
3694 // ======================================================
3697 PGPool
OSD::_get_pool(int id
, OSDMapRef createmap
)
3699 if (!createmap
->have_pg_pool(id
)) {
3700 dout(5) << __func__
<< ": the OSDmap does not contain a PG pool with id = "
3705 PGPool p
= PGPool(cct
, createmap
, id
);
3707 dout(10) << "_get_pool " << p
.id
<< dendl
;
3711 PG
*OSD::_open_lock_pg(
3712 OSDMapRef createmap
,
3713 spg_t pgid
, bool no_lockdep_check
)
3715 assert(osd_lock
.is_locked());
3717 PG
* pg
= _make_pg(createmap
, pgid
);
3719 RWLock::WLocker
l(pg_map_lock
);
3720 pg
->lock(no_lockdep_check
);
3722 pg
->get("PGMap"); // because it's in pg_map
3723 service
.pg_add_epoch(pg
->info
.pgid
, createmap
->get_epoch());
3729 OSDMapRef createmap
,
3732 dout(10) << "_open_lock_pg " << pgid
<< dendl
;
3733 PGPool pool
= _get_pool(pgid
.pool(), createmap
);
3737 if (createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_REPLICATED
||
3738 createmap
->get_pg_type(pgid
.pgid
) == pg_pool_t::TYPE_ERASURE
)
3739 pg
= new PrimaryLogPG(&service
, createmap
, pool
, pgid
);
3747 void OSD::add_newly_split_pg(PG
*pg
, PG::RecoveryCtx
*rctx
)
3749 epoch_t
e(service
.get_osdmap()->get_epoch());
3750 pg
->get("PGMap"); // For pg_map
3751 pg_map
[pg
->info
.pgid
] = pg
;
3752 service
.pg_add_epoch(pg
->info
.pgid
, pg
->get_osdmap()->get_epoch());
3754 dout(10) << "Adding newly split pg " << *pg
<< dendl
;
3755 pg
->handle_loaded(rctx
);
3756 pg
->write_if_dirty(*(rctx
->transaction
));
3757 pg
->queue_null(e
, e
);
3758 map
<spg_t
, list
<PG::CephPeeringEvtRef
> >::iterator to_wake
=
3759 peering_wait_for_split
.find(pg
->info
.pgid
);
3760 if (to_wake
!= peering_wait_for_split
.end()) {
3761 for (list
<PG::CephPeeringEvtRef
>::iterator i
=
3762 to_wake
->second
.begin();
3763 i
!= to_wake
->second
.end();
3765 pg
->queue_peering_event(*i
);
3767 peering_wait_for_split
.erase(to_wake
);
3769 if (!service
.get_osdmap()->have_pg_pool(pg
->info
.pgid
.pool()))
3773 OSD::res_result
OSD::_try_resurrect_pg(
3774 OSDMapRef curmap
, spg_t pgid
, spg_t
*resurrected
, PGRef
*old_pg_state
)
3776 assert(resurrected
);
3777 assert(old_pg_state
);
3778 // find nearest ancestor
3779 DeletingStateRef df
;
3782 df
= service
.deleting_pgs
.lookup(cur
);
3787 cur
= cur
.get_parent();
3790 return RES_NONE
; // good to go
3792 df
->old_pg_state
->lock();
3793 OSDMapRef create_map
= df
->old_pg_state
->get_osdmap();
3794 df
->old_pg_state
->unlock();
3796 set
<spg_t
> children
;
3798 if (df
->try_stop_deletion()) {
3799 dout(10) << __func__
<< ": halted deletion on pg " << pgid
<< dendl
;
3801 *old_pg_state
= df
->old_pg_state
;
3802 service
.deleting_pgs
.remove(pgid
); // PG is no longer being removed!
3805 // raced, ensure we don't see DeletingStateRef when we try to
3807 service
.deleting_pgs
.remove(pgid
);
3810 } else if (cur
.is_split(create_map
->get_pg_num(cur
.pool()),
3811 curmap
->get_pg_num(cur
.pool()),
3813 children
.count(pgid
)) {
3814 if (df
->try_stop_deletion()) {
3815 dout(10) << __func__
<< ": halted deletion on ancestor pg " << pgid
3818 *old_pg_state
= df
->old_pg_state
;
3819 service
.deleting_pgs
.remove(cur
); // PG is no longer being removed!
3822 /* this is not a problem, failing to cancel proves that all objects
3823 * have been removed, so no hobject_t overlap is possible
3831 PG
*OSD::_create_lock_pg(
3832 OSDMapRef createmap
,
3837 vector
<int>& up
, int up_primary
,
3838 vector
<int>& acting
, int acting_primary
,
3839 pg_history_t history
,
3840 const PastIntervals
& pi
,
3841 ObjectStore::Transaction
& t
)
3843 assert(osd_lock
.is_locked());
3844 dout(20) << "_create_lock_pg pgid " << pgid
<< dendl
;
3846 PG
*pg
= _open_lock_pg(createmap
, pgid
, true);
3848 service
.init_splits_between(pgid
, pg
->get_osdmap(), service
.get_osdmap());
3861 dout(7) << "_create_lock_pg " << *pg
<< dendl
;
3865 PG
*OSD::_lookup_lock_pg(spg_t pgid
)
3867 RWLock::RLocker
l(pg_map_lock
);
3869 auto pg_map_entry
= pg_map
.find(pgid
);
3870 if (pg_map_entry
== pg_map
.end())
3872 PG
*pg
= pg_map_entry
->second
;
3877 PG
*OSD::lookup_lock_pg(spg_t pgid
)
3879 return _lookup_lock_pg(pgid
);
3882 PG
*OSD::_lookup_lock_pg_with_map_lock_held(spg_t pgid
)
3884 assert(pg_map
.count(pgid
));
3885 PG
*pg
= pg_map
[pgid
];
3890 void OSD::load_pgs()
3892 assert(osd_lock
.is_locked());
3893 dout(0) << "load_pgs" << dendl
;
3895 RWLock::RLocker
l(pg_map_lock
);
3896 assert(pg_map
.empty());
3900 int r
= store
->list_collections(ls
);
3902 derr
<< "failed to list pgs: " << cpp_strerror(-r
) << dendl
;
3905 bool has_upgraded
= false;
3907 for (vector
<coll_t
>::iterator it
= ls
.begin();
3911 if (it
->is_temp(&pgid
) ||
3912 (it
->is_pg(&pgid
) && PG::_has_removal_flag(store
, pgid
))) {
3913 dout(10) << "load_pgs " << *it
<< " clearing temp" << dendl
;
3914 recursive_remove_collection(cct
, store
, pgid
, *it
);
3918 if (!it
->is_pg(&pgid
)) {
3919 dout(10) << "load_pgs ignoring unrecognized " << *it
<< dendl
;
3923 if (pgid
.preferred() >= 0) {
3924 dout(10) << __func__
<< ": skipping localized PG " << pgid
<< dendl
;
3925 // FIXME: delete it too, eventually
3929 dout(10) << "pgid " << pgid
<< " coll " << coll_t(pgid
) << dendl
;
3931 epoch_t map_epoch
= 0;
3932 int r
= PG::peek_map_epoch(store
, pgid
, &map_epoch
, &bl
);
3934 derr
<< __func__
<< " unable to peek at " << pgid
<< " metadata, skipping"
3940 if (map_epoch
> 0) {
3941 OSDMapRef pgosdmap
= service
.try_get_map(map_epoch
);
3943 if (!osdmap
->have_pg_pool(pgid
.pool())) {
3944 derr
<< __func__
<< ": could not find map for epoch " << map_epoch
3945 << " on pg " << pgid
<< ", but the pool is not present in the "
3946 << "current map, so this is probably a result of bug 10617. "
3947 << "Skipping the pg for now, you can use ceph-objectstore-tool "
3948 << "to clean it up later." << dendl
;
3951 derr
<< __func__
<< ": have pgid " << pgid
<< " at epoch "
3952 << map_epoch
<< ", but missing map. Crashing."
3954 assert(0 == "Missing map in load_pgs");
3957 pg
= _open_lock_pg(pgosdmap
, pgid
);
3959 pg
= _open_lock_pg(osdmap
, pgid
);
3961 // there can be no waiters here, so we don't call wake_pg_waiters
3963 pg
->ch
= store
->open_collection(pg
->coll
);
3965 // read pg state, log
3966 pg
->read_state(store
, bl
);
3968 if (pg
->must_upgrade()) {
3969 if (!pg
->can_upgrade()) {
3970 derr
<< "PG needs upgrade, but on-disk data is too old; upgrade to"
3971 << " an older version first." << dendl
;
3972 assert(0 == "PG too old to upgrade");
3974 if (!has_upgraded
) {
3975 derr
<< "PGs are upgrading" << dendl
;
3976 has_upgraded
= true;
3978 dout(10) << "PG " << pg
->info
.pgid
3979 << " must upgrade..." << dendl
;
3983 service
.init_splits_between(pg
->info
.pgid
, pg
->get_osdmap(), osdmap
);
3985 // generate state for PG's current mapping
3986 int primary
, up_primary
;
3987 vector
<int> acting
, up
;
3988 pg
->get_osdmap()->pg_to_up_acting_osds(
3989 pgid
.pgid
, &up
, &up_primary
, &acting
, &primary
);
3990 pg
->init_primary_up_acting(
3995 int role
= OSDMap::calc_pg_role(whoami
, pg
->acting
);
3996 if (pg
->pool
.info
.is_replicated() || role
== pg
->pg_whoami
.shard
)
4001 pg
->reg_next_scrub();
4003 PG::RecoveryCtx
rctx(0, 0, 0, 0, 0, 0);
4004 pg
->handle_loaded(&rctx
);
4006 dout(10) << "load_pgs loaded " << *pg
<< " " << pg
->pg_log
.get_log() << dendl
;
4007 if (pg
->pg_log
.is_dirty()) {
4008 ObjectStore::Transaction t
;
4009 pg
->write_if_dirty(t
);
4010 store
->apply_transaction(pg
->osr
.get(), std::move(t
));
4015 RWLock::RLocker
l(pg_map_lock
);
4016 dout(0) << "load_pgs opened " << pg_map
.size() << " pgs" << dendl
;
4019 // clean up old infos object?
4020 if (has_upgraded
&& store
->exists(coll_t::meta(), OSD::make_infos_oid())) {
4021 dout(1) << __func__
<< " removing legacy infos object" << dendl
;
4022 ObjectStore::Transaction t
;
4023 t
.remove(coll_t::meta(), OSD::make_infos_oid());
4024 int r
= store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4026 derr
<< __func__
<< ": apply_transaction returned "
4027 << cpp_strerror(r
) << dendl
;
4032 build_past_intervals_parallel();
4037 * build past_intervals efficiently on old, degraded, and buried
4038 * clusters. this is important for efficiently catching up osds that
4039 * are way behind on maps to the current cluster state.
4041 * this is a parallel version of PG::generate_past_intervals().
4042 * follow the same logic, but do all pgs at the same time so that we
4043 * can make a single pass across the osdmap history.
4045 void OSD::build_past_intervals_parallel()
4049 vector
<int> old_acting
, old_up
;
4050 epoch_t same_interval_since
;
4054 map
<PG
*,pistate
> pis
;
4056 // calculate junction of map range
4057 epoch_t end_epoch
= superblock
.oldest_map
;
4058 epoch_t cur_epoch
= superblock
.newest_map
;
4060 RWLock::RLocker
l(pg_map_lock
);
4061 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4066 auto rpib
= pg
->get_required_past_interval_bounds(
4068 superblock
.oldest_map
);
4069 if (rpib
.first
>= rpib
.second
&& pg
->past_intervals
.empty()) {
4070 if (pg
->info
.history
.same_interval_since
== 0) {
4071 pg
->info
.history
.same_interval_since
= rpib
.second
;
4075 auto apib
= pg
->past_intervals
.get_bounds();
4076 if (apib
.second
>= rpib
.second
&&
4077 apib
.first
<= rpib
.first
) {
4078 if (pg
->info
.history
.same_interval_since
== 0) {
4079 pg
->info
.history
.same_interval_since
= rpib
.second
;
4085 dout(10) << pg
->info
.pgid
<< " needs " << rpib
.first
<< "-"
4086 << rpib
.second
<< dendl
;
4087 pistate
& p
= pis
[pg
];
4088 p
.start
= rpib
.first
;
4089 p
.end
= rpib
.second
;
4090 p
.same_interval_since
= 0;
4092 if (rpib
.first
< cur_epoch
)
4093 cur_epoch
= rpib
.first
;
4094 if (rpib
.second
> end_epoch
)
4095 end_epoch
= rpib
.second
;
4099 dout(10) << __func__
<< " nothing to build" << dendl
;
4103 dout(1) << __func__
<< " over " << cur_epoch
<< "-" << end_epoch
<< dendl
;
4104 assert(cur_epoch
<= end_epoch
);
4106 OSDMapRef cur_map
, last_map
;
4107 for ( ; cur_epoch
<= end_epoch
; cur_epoch
++) {
4108 dout(10) << __func__
<< " epoch " << cur_epoch
<< dendl
;
4110 cur_map
= get_map(cur_epoch
);
4112 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4114 pistate
& p
= i
->second
;
4116 if (cur_epoch
< p
.start
|| cur_epoch
> p
.end
)
4119 vector
<int> acting
, up
;
4122 pg_t pgid
= pg
->info
.pgid
.pgid
;
4123 if (p
.same_interval_since
&& last_map
->get_pools().count(pgid
.pool()))
4124 pgid
= pgid
.get_ancestor(last_map
->get_pg_num(pgid
.pool()));
4125 cur_map
->pg_to_up_acting_osds(
4126 pgid
, &up
, &up_primary
, &acting
, &primary
);
4128 if (p
.same_interval_since
== 0) {
4129 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4130 << " first map, acting " << acting
4131 << " up " << up
<< ", same_interval_since = " << cur_epoch
<< dendl
;
4132 p
.same_interval_since
= cur_epoch
;
4134 p
.old_acting
= acting
;
4135 p
.primary
= primary
;
4136 p
.up_primary
= up_primary
;
4141 boost::scoped_ptr
<IsPGRecoverablePredicate
> recoverable(
4142 pg
->get_is_recoverable_predicate());
4143 std::stringstream debug
;
4144 bool new_interval
= PastIntervals::check_new_interval(
4147 p
.old_acting
, acting
,
4151 p
.same_interval_since
,
4152 pg
->info
.history
.last_epoch_clean
,
4156 &pg
->past_intervals
,
4159 dout(10) << __func__
<< " epoch " << cur_epoch
<< " pg " << pg
->info
.pgid
4160 << " " << debug
.str() << dendl
;
4162 p
.old_acting
= acting
;
4163 p
.primary
= primary
;
4164 p
.up_primary
= up_primary
;
4165 p
.same_interval_since
= cur_epoch
;
4170 // Now that past_intervals have been recomputed let's fix the same_interval_since
4171 // if it was cleared by import.
4172 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4174 pistate
& p
= i
->second
;
4176 if (pg
->info
.history
.same_interval_since
== 0) {
4177 assert(p
.same_interval_since
);
4178 dout(10) << __func__
<< " fix same_interval_since " << p
.same_interval_since
<< " pg " << *pg
<< dendl
;
4179 dout(10) << __func__
<< " past_intervals " << pg
->past_intervals
<< dendl
;
4181 pg
->info
.history
.same_interval_since
= p
.same_interval_since
;
4185 // write info only at the end. this is necessary because we check
4186 // whether the past_intervals go far enough back or forward in time,
4187 // but we don't check for holes. we could avoid it by discarding
4188 // the previous past_intervals and rebuilding from scratch, or we
4189 // can just do this and commit all our work at the end.
4190 ObjectStore::Transaction t
;
4192 for (map
<PG
*,pistate
>::iterator i
= pis
.begin(); i
!= pis
.end(); ++i
) {
4195 pg
->dirty_big_info
= true;
4196 pg
->dirty_info
= true;
4197 pg
->write_if_dirty(t
);
4200 // don't let the transaction get too big
4201 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
4202 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4203 t
= ObjectStore::Transaction();
4208 store
->apply_transaction(service
.meta_osr
.get(), std::move(t
));
4212 * look up a pg. if we have it, great. if not, consider creating it IF the pg mapping
4213 * hasn't changed since the given epoch and we are the primary.
4215 int OSD::handle_pg_peering_evt(
4217 const pg_history_t
& orig_history
,
4218 const PastIntervals
& pi
,
4220 PG::CephPeeringEvtRef evt
)
4222 if (service
.splitting(pgid
)) {
4223 peering_wait_for_split
[pgid
].push_back(evt
);
4227 PG
*pg
= _lookup_lock_pg(pgid
);
4230 if (!osdmap
->have_pg_pool(pgid
.pool()))
4232 int up_primary
, acting_primary
;
4233 vector
<int> up
, acting
;
4234 osdmap
->pg_to_up_acting_osds(
4235 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4237 pg_history_t history
= orig_history
;
4238 bool valid_history
= project_pg_history(
4239 pgid
, history
, epoch
, up
, up_primary
, acting
, acting_primary
);
4241 if (!valid_history
|| epoch
< history
.same_interval_since
) {
4242 dout(10) << __func__
<< pgid
<< " acting changed in "
4243 << history
.same_interval_since
<< " (msg from " << epoch
<< ")"
4248 if (service
.splitting(pgid
)) {
4252 // do we need to resurrect a deleting pg?
4255 res_result result
= _try_resurrect_pg(
4256 service
.get_osdmap(),
4261 PG::RecoveryCtx rctx
= create_context();
4264 const pg_pool_t
* pp
= osdmap
->get_pg_pool(pgid
.pool());
4265 if (pp
->has_flag(pg_pool_t::FLAG_EC_OVERWRITES
) &&
4266 store
->get_type() != "bluestore") {
4267 clog
->warn() << "pg " << pgid
4268 << " is at risk of silent data corruption: "
4269 << "the pool allows ec overwrites but is not stored in "
4270 << "bluestore, so deep scrubbing will not detect bitrot";
4272 PG::_create(*rctx
.transaction
, pgid
, pgid
.get_split_bits(pp
->get_pg_num()));
4273 PG::_init(*rctx
.transaction
, pgid
, pp
);
4275 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
4276 if (!pp
->is_replicated() && role
!= pgid
.shard
)
4279 pg
= _create_lock_pg(
4284 acting
, acting_primary
,
4287 pg
->handle_create(&rctx
);
4288 pg
->write_if_dirty(*rctx
.transaction
);
4289 dispatch_context(rctx
, pg
, osdmap
);
4291 dout(10) << *pg
<< " is new" << dendl
;
4293 pg
->queue_peering_event(evt
);
4294 wake_pg_waiters(pg
);
4299 old_pg_state
->lock();
4300 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4301 int old_role
= old_pg_state
->role
;
4302 vector
<int> old_up
= old_pg_state
->up
;
4303 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4304 vector
<int> old_acting
= old_pg_state
->acting
;
4305 int old_primary
= old_pg_state
->primary
.osd
;
4306 pg_history_t old_history
= old_pg_state
->info
.history
;
4307 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4308 old_pg_state
->unlock();
4309 pg
= _create_lock_pg(
4322 pg
->handle_create(&rctx
);
4323 pg
->write_if_dirty(*rctx
.transaction
);
4324 dispatch_context(rctx
, pg
, osdmap
);
4326 dout(10) << *pg
<< " is new (resurrected)" << dendl
;
4328 pg
->queue_peering_event(evt
);
4329 wake_pg_waiters(pg
);
4334 assert(old_pg_state
);
4335 old_pg_state
->lock();
4336 OSDMapRef old_osd_map
= old_pg_state
->get_osdmap();
4337 int old_role
= old_pg_state
->role
;
4338 vector
<int> old_up
= old_pg_state
->up
;
4339 int old_up_primary
= old_pg_state
->up_primary
.osd
;
4340 vector
<int> old_acting
= old_pg_state
->acting
;
4341 int old_primary
= old_pg_state
->primary
.osd
;
4342 pg_history_t old_history
= old_pg_state
->info
.history
;
4343 PastIntervals old_past_intervals
= old_pg_state
->past_intervals
;
4344 old_pg_state
->unlock();
4345 PG
*parent
= _create_lock_pg(
4359 parent
->handle_create(&rctx
);
4360 parent
->write_if_dirty(*rctx
.transaction
);
4361 dispatch_context(rctx
, parent
, osdmap
);
4363 dout(10) << *parent
<< " is new" << dendl
;
4365 assert(service
.splitting(pgid
));
4366 peering_wait_for_split
[pgid
].push_back(evt
);
4368 //parent->queue_peering_event(evt);
4369 parent
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
4370 wake_pg_waiters(parent
);
4379 // already had it. did the mapping change?
4380 if (epoch
< pg
->info
.history
.same_interval_since
) {
4381 dout(10) << *pg
<< __func__
<< " acting changed in "
4382 << pg
->info
.history
.same_interval_since
4383 << " (msg from " << epoch
<< ")" << dendl
;
4385 pg
->queue_peering_event(evt
);
4393 void OSD::build_initial_pg_history(
4396 utime_t created_stamp
,
4400 dout(10) << __func__
<< " " << pgid
<< " created " << created
<< dendl
;
4401 h
->epoch_created
= created
;
4402 h
->epoch_pool_created
= created
;
4403 h
->same_interval_since
= created
;
4404 h
->same_up_since
= created
;
4405 h
->same_primary_since
= created
;
4406 h
->last_scrub_stamp
= created_stamp
;
4407 h
->last_deep_scrub_stamp
= created_stamp
;
4408 h
->last_clean_scrub_stamp
= created_stamp
;
4410 OSDMapRef lastmap
= service
.get_map(created
);
4411 int up_primary
, acting_primary
;
4412 vector
<int> up
, acting
;
4413 lastmap
->pg_to_up_acting_osds(
4414 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
4416 ostringstream debug
;
4417 for (epoch_t e
= created
+ 1; e
<= osdmap
->get_epoch(); ++e
) {
4418 OSDMapRef osdmap
= service
.get_map(e
);
4419 int new_up_primary
, new_acting_primary
;
4420 vector
<int> new_up
, new_acting
;
4421 osdmap
->pg_to_up_acting_osds(
4422 pgid
.pgid
, &new_up
, &new_up_primary
, &new_acting
, &new_acting_primary
);
4424 // this is a bit imprecise, but sufficient?
4425 struct min_size_predicate_t
: public IsPGRecoverablePredicate
{
4426 const pg_pool_t
*pi
;
4427 bool operator()(const set
<pg_shard_t
> &have
) const {
4428 return have
.size() >= pi
->min_size
;
4430 min_size_predicate_t(const pg_pool_t
*i
) : pi(i
) {}
4431 } min_size_predicate(osdmap
->get_pg_pool(pgid
.pgid
.pool()));
4433 bool new_interval
= PastIntervals::check_new_interval(
4440 h
->same_interval_since
,
4441 h
->last_epoch_clean
,
4445 &min_size_predicate
,
4449 h
->same_interval_since
= e
;
4451 h
->same_up_since
= e
;
4453 if (acting_primary
!= new_acting_primary
) {
4454 h
->same_primary_since
= e
;
4456 if (pgid
.pgid
.is_split(lastmap
->get_pg_num(pgid
.pgid
.pool()),
4457 osdmap
->get_pg_num(pgid
.pgid
.pool()),
4459 h
->last_epoch_split
= e
;
4462 acting
= new_acting
;
4463 up_primary
= new_up_primary
;
4464 acting_primary
= new_acting_primary
;
4468 dout(20) << __func__
<< " " << debug
.str() << dendl
;
4469 dout(10) << __func__
<< " " << *h
<< " " << *pi
4470 << " [" << (pi
->empty() ? pair
<epoch_t
,epoch_t
>(0,0) :
4471 pi
->get_bounds()) << ")"
4476 * Fill in the passed history so you know same_interval_since, same_up_since,
4477 * and same_primary_since.
4479 bool OSD::project_pg_history(spg_t pgid
, pg_history_t
& h
, epoch_t from
,
4480 const vector
<int>& currentup
,
4481 int currentupprimary
,
4482 const vector
<int>& currentacting
,
4483 int currentactingprimary
)
4485 dout(15) << "project_pg_history " << pgid
4486 << " from " << from
<< " to " << osdmap
->get_epoch()
4491 for (e
= osdmap
->get_epoch();
4494 // verify during intermediate epoch (e-1)
4495 OSDMapRef oldmap
= service
.try_get_map(e
-1);
4497 dout(15) << __func__
<< ": found map gap, returning false" << dendl
;
4500 assert(oldmap
->have_pg_pool(pgid
.pool()));
4502 int upprimary
, actingprimary
;
4503 vector
<int> up
, acting
;
4504 oldmap
->pg_to_up_acting_osds(
4511 // acting set change?
4512 if ((actingprimary
!= currentactingprimary
||
4513 upprimary
!= currentupprimary
||
4514 acting
!= currentacting
||
4515 up
!= currentup
) && e
> h
.same_interval_since
) {
4516 dout(15) << "project_pg_history " << pgid
<< " acting|up changed in " << e
4517 << " from " << acting
<< "/" << up
4518 << " " << actingprimary
<< "/" << upprimary
4519 << " -> " << currentacting
<< "/" << currentup
4520 << " " << currentactingprimary
<< "/" << currentupprimary
4522 h
.same_interval_since
= e
;
4525 if (pgid
.is_split(oldmap
->get_pg_num(pgid
.pool()),
4526 osdmap
->get_pg_num(pgid
.pool()),
4527 0) && e
> h
.same_interval_since
) {
4528 h
.same_interval_since
= e
;
4531 if ((up
!= currentup
|| upprimary
!= currentupprimary
)
4532 && e
> h
.same_up_since
) {
4533 dout(15) << "project_pg_history " << pgid
<< " up changed in " << e
4534 << " from " << up
<< " " << upprimary
4535 << " -> " << currentup
<< " " << currentupprimary
<< dendl
;
4536 h
.same_up_since
= e
;
4540 if (OSDMap::primary_changed(
4543 currentactingprimary
,
4545 e
> h
.same_primary_since
) {
4546 dout(15) << "project_pg_history " << pgid
<< " primary changed in " << e
<< dendl
;
4547 h
.same_primary_since
= e
;
4550 if (h
.same_interval_since
>= e
&& h
.same_up_since
>= e
&& h
.same_primary_since
>= e
)
4554 // base case: these floors should be the pg creation epoch if we didn't
4555 // find any changes.
4556 if (e
== h
.epoch_created
) {
4557 if (!h
.same_interval_since
)
4558 h
.same_interval_since
= e
;
4559 if (!h
.same_up_since
)
4560 h
.same_up_since
= e
;
4561 if (!h
.same_primary_since
)
4562 h
.same_primary_since
= e
;
4565 dout(15) << "project_pg_history end " << h
<< dendl
;
4571 void OSD::_add_heartbeat_peer(int p
)
4577 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(p
);
4578 if (i
== heartbeat_peers
.end()) {
4579 pair
<ConnectionRef
,ConnectionRef
> cons
= service
.get_con_osd_hb(p
, osdmap
->get_epoch());
4582 hi
= &heartbeat_peers
[p
];
4584 HeartbeatSession
*s
= new HeartbeatSession(p
);
4585 hi
->con_back
= cons
.first
.get();
4586 hi
->con_back
->set_priv(s
->get());
4588 hi
->con_front
= cons
.second
.get();
4589 hi
->con_front
->set_priv(s
->get());
4590 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4591 << " " << hi
->con_back
->get_peer_addr()
4592 << " " << hi
->con_front
->get_peer_addr()
4595 hi
->con_front
.reset(NULL
);
4596 dout(10) << "_add_heartbeat_peer: new peer osd." << p
4597 << " " << hi
->con_back
->get_peer_addr()
4604 hi
->epoch
= osdmap
->get_epoch();
4607 void OSD::_remove_heartbeat_peer(int n
)
4609 map
<int,HeartbeatInfo
>::iterator q
= heartbeat_peers
.find(n
);
4610 assert(q
!= heartbeat_peers
.end());
4611 dout(20) << " removing heartbeat peer osd." << n
4612 << " " << q
->second
.con_back
->get_peer_addr()
4613 << " " << (q
->second
.con_front
? q
->second
.con_front
->get_peer_addr() : entity_addr_t())
4615 q
->second
.con_back
->mark_down();
4616 if (q
->second
.con_front
) {
4617 q
->second
.con_front
->mark_down();
4619 heartbeat_peers
.erase(q
);
4622 void OSD::need_heartbeat_peer_update()
4626 dout(20) << "need_heartbeat_peer_update" << dendl
;
4627 heartbeat_set_peers_need_update();
4630 void OSD::maybe_update_heartbeat_peers()
4632 assert(osd_lock
.is_locked());
4634 if (is_waiting_for_healthy()) {
4635 utime_t now
= ceph_clock_now();
4636 if (last_heartbeat_resample
== utime_t()) {
4637 last_heartbeat_resample
= now
;
4638 heartbeat_set_peers_need_update();
4639 } else if (!heartbeat_peers_need_update()) {
4640 utime_t dur
= now
- last_heartbeat_resample
;
4641 if (dur
> cct
->_conf
->osd_heartbeat_grace
) {
4642 dout(10) << "maybe_update_heartbeat_peers forcing update after " << dur
<< " seconds" << dendl
;
4643 heartbeat_set_peers_need_update();
4644 last_heartbeat_resample
= now
;
4645 reset_heartbeat_peers(); // we want *new* peers!
4650 if (!heartbeat_peers_need_update())
4652 heartbeat_clear_peers_need_update();
4654 Mutex::Locker
l(heartbeat_lock
);
4656 dout(10) << "maybe_update_heartbeat_peers updating" << dendl
;
4659 // build heartbeat from set
4661 RWLock::RLocker
l(pg_map_lock
);
4662 for (ceph::unordered_map
<spg_t
, PG
*>::iterator i
= pg_map
.begin();
4666 pg
->heartbeat_peer_lock
.Lock();
4667 dout(20) << i
->first
<< " heartbeat_peers " << pg
->heartbeat_peers
<< dendl
;
4668 for (set
<int>::iterator p
= pg
->heartbeat_peers
.begin();
4669 p
!= pg
->heartbeat_peers
.end();
4671 if (osdmap
->is_up(*p
))
4672 _add_heartbeat_peer(*p
);
4673 for (set
<int>::iterator p
= pg
->probe_targets
.begin();
4674 p
!= pg
->probe_targets
.end();
4676 if (osdmap
->is_up(*p
))
4677 _add_heartbeat_peer(*p
);
4678 pg
->heartbeat_peer_lock
.Unlock();
4682 // include next and previous up osds to ensure we have a fully-connected set
4683 set
<int> want
, extras
;
4684 int next
= osdmap
->get_next_up_osd_after(whoami
);
4687 int prev
= osdmap
->get_previous_up_osd_before(whoami
);
4688 if (prev
>= 0 && prev
!= next
)
4691 for (set
<int>::iterator p
= want
.begin(); p
!= want
.end(); ++p
) {
4692 dout(10) << " adding neighbor peer osd." << *p
<< dendl
;
4694 _add_heartbeat_peer(*p
);
4697 // remove down peers; enumerate extras
4698 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4699 while (p
!= heartbeat_peers
.end()) {
4700 if (!osdmap
->is_up(p
->first
)) {
4703 _remove_heartbeat_peer(o
);
4706 if (p
->second
.epoch
< osdmap
->get_epoch()) {
4707 extras
.insert(p
->first
);
4713 int start
= osdmap
->get_next_up_osd_after(whoami
);
4714 for (int n
= start
; n
>= 0; ) {
4715 if ((int)heartbeat_peers
.size() >= cct
->_conf
->osd_heartbeat_min_peers
)
4717 if (!extras
.count(n
) && !want
.count(n
) && n
!= whoami
) {
4718 dout(10) << " adding random peer osd." << n
<< dendl
;
4720 _add_heartbeat_peer(n
);
4722 n
= osdmap
->get_next_up_osd_after(n
);
4724 break; // came full circle; stop
4728 for (set
<int>::iterator p
= extras
.begin();
4729 (int)heartbeat_peers
.size() > cct
->_conf
->osd_heartbeat_min_peers
&& p
!= extras
.end();
4733 _remove_heartbeat_peer(*p
);
4736 dout(10) << "maybe_update_heartbeat_peers " << heartbeat_peers
.size() << " peers, extras " << extras
<< dendl
;
4739 void OSD::reset_heartbeat_peers()
4741 assert(osd_lock
.is_locked());
4742 dout(10) << "reset_heartbeat_peers" << dendl
;
4743 Mutex::Locker
l(heartbeat_lock
);
4744 while (!heartbeat_peers
.empty()) {
4745 HeartbeatInfo
& hi
= heartbeat_peers
.begin()->second
;
4746 hi
.con_back
->mark_down();
4748 hi
.con_front
->mark_down();
4750 heartbeat_peers
.erase(heartbeat_peers
.begin());
4752 failure_queue
.clear();
4755 void OSD::handle_osd_ping(MOSDPing
*m
)
4757 if (superblock
.cluster_fsid
!= m
->fsid
) {
4758 dout(20) << "handle_osd_ping from " << m
->get_source_inst()
4759 << " bad fsid " << m
->fsid
<< " != " << superblock
.cluster_fsid
<< dendl
;
4764 int from
= m
->get_source().num();
4766 heartbeat_lock
.Lock();
4767 if (is_stopping()) {
4768 heartbeat_lock
.Unlock();
4773 OSDMapRef curmap
= service
.get_osdmap();
4775 heartbeat_lock
.Unlock();
4782 case MOSDPing::PING
:
4784 if (cct
->_conf
->osd_debug_drop_ping_probability
> 0) {
4785 auto heartbeat_drop
= debug_heartbeat_drops_remaining
.find(from
);
4786 if (heartbeat_drop
!= debug_heartbeat_drops_remaining
.end()) {
4787 if (heartbeat_drop
->second
== 0) {
4788 debug_heartbeat_drops_remaining
.erase(heartbeat_drop
);
4790 --heartbeat_drop
->second
;
4791 dout(5) << "Dropping heartbeat from " << from
4792 << ", " << heartbeat_drop
->second
4793 << " remaining to drop" << dendl
;
4796 } else if (cct
->_conf
->osd_debug_drop_ping_probability
>
4797 ((((double)(rand()%100))/100.0))) {
4799 debug_heartbeat_drops_remaining
.insert(std::make_pair(from
,
4800 cct
->_conf
->osd_debug_drop_ping_duration
)).first
;
4801 dout(5) << "Dropping heartbeat from " << from
4802 << ", " << heartbeat_drop
->second
4803 << " remaining to drop" << dendl
;
4808 if (!cct
->get_heartbeat_map()->is_healthy()) {
4809 dout(10) << "internal heartbeat not healthy, dropping ping request" << dendl
;
4813 Message
*r
= new MOSDPing(monc
->get_fsid(),
4814 curmap
->get_epoch(),
4815 MOSDPing::PING_REPLY
, m
->stamp
,
4816 cct
->_conf
->osd_heartbeat_min_size
);
4817 m
->get_connection()->send_message(r
);
4819 if (curmap
->is_up(from
)) {
4820 service
.note_peer_epoch(from
, m
->map_epoch
);
4822 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
4824 service
.share_map_peer(from
, con
.get());
4827 } else if (!curmap
->exists(from
) ||
4828 curmap
->get_down_at(from
) > m
->map_epoch
) {
4829 // tell them they have died
4830 Message
*r
= new MOSDPing(monc
->get_fsid(),
4831 curmap
->get_epoch(),
4834 cct
->_conf
->osd_heartbeat_min_size
);
4835 m
->get_connection()->send_message(r
);
4840 case MOSDPing::PING_REPLY
:
4842 map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.find(from
);
4843 if (i
!= heartbeat_peers
.end()) {
4844 if (m
->get_connection() == i
->second
.con_back
) {
4845 dout(25) << "handle_osd_ping got reply from osd." << from
4846 << " first_tx " << i
->second
.first_tx
4847 << " last_tx " << i
->second
.last_tx
4848 << " last_rx_back " << i
->second
.last_rx_back
<< " -> " << m
->stamp
4849 << " last_rx_front " << i
->second
.last_rx_front
4851 i
->second
.last_rx_back
= m
->stamp
;
4852 // if there is no front con, set both stamps.
4853 if (i
->second
.con_front
== NULL
)
4854 i
->second
.last_rx_front
= m
->stamp
;
4855 } else if (m
->get_connection() == i
->second
.con_front
) {
4856 dout(25) << "handle_osd_ping got reply from osd." << from
4857 << " first_tx " << i
->second
.first_tx
4858 << " last_tx " << i
->second
.last_tx
4859 << " last_rx_back " << i
->second
.last_rx_back
4860 << " last_rx_front " << i
->second
.last_rx_front
<< " -> " << m
->stamp
4862 i
->second
.last_rx_front
= m
->stamp
;
4865 utime_t cutoff
= ceph_clock_now();
4866 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
4867 if (i
->second
.is_healthy(cutoff
)) {
4868 // Cancel false reports
4869 auto failure_queue_entry
= failure_queue
.find(from
);
4870 if (failure_queue_entry
!= failure_queue
.end()) {
4871 dout(10) << "handle_osd_ping canceling queued "
4872 << "failure report for osd." << from
<< dendl
;
4873 failure_queue
.erase(failure_queue_entry
);
4876 auto failure_pending_entry
= failure_pending
.find(from
);
4877 if (failure_pending_entry
!= failure_pending
.end()) {
4878 dout(10) << "handle_osd_ping canceling in-flight "
4879 << "failure report for osd." << from
<< dendl
;
4880 send_still_alive(curmap
->get_epoch(),
4881 failure_pending_entry
->second
.second
);
4882 failure_pending
.erase(failure_pending_entry
);
4888 curmap
->is_up(from
)) {
4889 service
.note_peer_epoch(from
, m
->map_epoch
);
4891 ConnectionRef con
= service
.get_con_osd_cluster(from
, curmap
->get_epoch());
4893 service
.share_map_peer(from
, con
.get());
4900 case MOSDPing::YOU_DIED
:
4901 dout(10) << "handle_osd_ping " << m
->get_source_inst()
4902 << " says i am down in " << m
->map_epoch
<< dendl
;
4903 osdmap_subscribe(curmap
->get_epoch()+1, false);
4907 heartbeat_lock
.Unlock();
4911 void OSD::heartbeat_entry()
4913 Mutex::Locker
l(heartbeat_lock
);
4916 while (!heartbeat_stop
) {
4919 double wait
= .5 + ((float)(rand() % 10)/10.0) * (float)cct
->_conf
->osd_heartbeat_interval
;
4921 w
.set_from_double(wait
);
4922 dout(30) << "heartbeat_entry sleeping for " << wait
<< dendl
;
4923 heartbeat_cond
.WaitInterval(heartbeat_lock
, w
);
4926 dout(30) << "heartbeat_entry woke up" << dendl
;
4930 void OSD::heartbeat_check()
4932 assert(heartbeat_lock
.is_locked());
4933 utime_t now
= ceph_clock_now();
4935 // check for heartbeat replies (move me elsewhere?)
4936 utime_t cutoff
= now
;
4937 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
4938 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4939 p
!= heartbeat_peers
.end();
4942 if (p
->second
.first_tx
== utime_t()) {
4943 dout(25) << "heartbeat_check we haven't sent ping to osd." << p
->first
4944 << "yet, skipping" << dendl
;
4948 dout(25) << "heartbeat_check osd." << p
->first
4949 << " first_tx " << p
->second
.first_tx
4950 << " last_tx " << p
->second
.last_tx
4951 << " last_rx_back " << p
->second
.last_rx_back
4952 << " last_rx_front " << p
->second
.last_rx_front
4954 if (p
->second
.is_unhealthy(cutoff
)) {
4955 if (p
->second
.last_rx_back
== utime_t() ||
4956 p
->second
.last_rx_front
== utime_t()) {
4957 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
4958 << " osd." << p
->first
<< " ever on either front or back, first ping sent "
4959 << p
->second
.first_tx
<< " (cutoff " << cutoff
<< ")" << dendl
;
4961 failure_queue
[p
->first
] = p
->second
.last_tx
;
4963 derr
<< "heartbeat_check: no reply from " << p
->second
.con_front
->get_peer_addr().get_sockaddr()
4964 << " osd." << p
->first
<< " since back " << p
->second
.last_rx_back
4965 << " front " << p
->second
.last_rx_front
4966 << " (cutoff " << cutoff
<< ")" << dendl
;
4968 failure_queue
[p
->first
] = MIN(p
->second
.last_rx_back
, p
->second
.last_rx_front
);
4974 void OSD::heartbeat()
4976 dout(30) << "heartbeat" << dendl
;
4980 int n_samples
= 86400 / cct
->_conf
->osd_heartbeat_interval
;
4981 if (getloadavg(loadavgs
, 1) == 1) {
4982 logger
->set(l_osd_loadavg
, 100 * loadavgs
[0]);
4983 daily_loadavg
= (daily_loadavg
* (n_samples
- 1) + loadavgs
[0]) / n_samples
;
4984 dout(30) << "heartbeat: daily_loadavg " << daily_loadavg
<< dendl
;
4987 dout(30) << "heartbeat checking stats" << dendl
;
4990 vector
<int> hb_peers
;
4991 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
4992 p
!= heartbeat_peers
.end();
4994 hb_peers
.push_back(p
->first
);
4995 service
.update_osd_stat(hb_peers
);
4997 dout(5) << "heartbeat: " << service
.get_osd_stat() << dendl
;
4999 utime_t now
= ceph_clock_now();
5002 for (map
<int,HeartbeatInfo
>::iterator i
= heartbeat_peers
.begin();
5003 i
!= heartbeat_peers
.end();
5005 int peer
= i
->first
;
5006 i
->second
.last_tx
= now
;
5007 if (i
->second
.first_tx
== utime_t())
5008 i
->second
.first_tx
= now
;
5009 dout(30) << "heartbeat sending ping to osd." << peer
<< dendl
;
5010 i
->second
.con_back
->send_message(new MOSDPing(monc
->get_fsid(),
5011 service
.get_osdmap()->get_epoch(),
5012 MOSDPing::PING
, now
,
5013 cct
->_conf
->osd_heartbeat_min_size
));
5015 if (i
->second
.con_front
)
5016 i
->second
.con_front
->send_message(new MOSDPing(monc
->get_fsid(),
5017 service
.get_osdmap()->get_epoch(),
5018 MOSDPing::PING
, now
,
5019 cct
->_conf
->osd_heartbeat_min_size
));
5022 logger
->set(l_osd_hb_to
, heartbeat_peers
.size());
5024 // hmm.. am i all alone?
5025 dout(30) << "heartbeat lonely?" << dendl
;
5026 if (heartbeat_peers
.empty()) {
5027 if (now
- last_mon_heartbeat
> cct
->_conf
->osd_mon_heartbeat_interval
&& is_active()) {
5028 last_mon_heartbeat
= now
;
5029 dout(10) << "i have no heartbeat peers; checking mon for new map" << dendl
;
5030 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5034 dout(30) << "heartbeat done" << dendl
;
5037 bool OSD::heartbeat_reset(Connection
*con
)
5039 HeartbeatSession
*s
= static_cast<HeartbeatSession
*>(con
->get_priv());
5041 heartbeat_lock
.Lock();
5042 if (is_stopping()) {
5043 heartbeat_lock
.Unlock();
5047 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(s
->peer
);
5048 if (p
!= heartbeat_peers
.end() &&
5049 (p
->second
.con_back
== con
||
5050 p
->second
.con_front
== con
)) {
5051 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5052 << ", reopening" << dendl
;
5053 if (con
!= p
->second
.con_back
) {
5054 p
->second
.con_back
->mark_down();
5056 p
->second
.con_back
.reset(NULL
);
5057 if (p
->second
.con_front
&& con
!= p
->second
.con_front
) {
5058 p
->second
.con_front
->mark_down();
5060 p
->second
.con_front
.reset(NULL
);
5061 pair
<ConnectionRef
,ConnectionRef
> newcon
= service
.get_con_osd_hb(p
->second
.peer
, p
->second
.epoch
);
5063 p
->second
.con_back
= newcon
.first
.get();
5064 p
->second
.con_back
->set_priv(s
->get());
5065 if (newcon
.second
) {
5066 p
->second
.con_front
= newcon
.second
.get();
5067 p
->second
.con_front
->set_priv(s
->get());
5070 dout(10) << "heartbeat_reset failed hb con " << con
<< " for osd." << p
->second
.peer
5071 << ", raced with osdmap update, closing out peer" << dendl
;
5072 heartbeat_peers
.erase(p
);
5075 dout(10) << "heartbeat_reset closing (old) failed hb con " << con
<< dendl
;
5077 heartbeat_lock
.Unlock();
5085 // =========================================
5089 assert(osd_lock
.is_locked());
5090 dout(10) << "tick" << dendl
;
5092 if (is_active() || is_waiting_for_healthy()) {
5093 maybe_update_heartbeat_peers();
5096 if (is_waiting_for_healthy()) {
5098 } else if (is_preboot() &&
5099 waiting_for_luminous_mons
&&
5100 monc
->monmap
.get_required_features().contains_all(
5101 ceph::features::mon::FEATURE_LUMINOUS
)) {
5102 // mon upgrade finished!
5108 tick_timer
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick(this));
5111 void OSD::tick_without_osd_lock()
5113 assert(tick_timer_lock
.is_locked());
5114 dout(10) << "tick_without_osd_lock" << dendl
;
5116 logger
->set(l_osd_buf
, buffer::get_total_alloc());
5117 logger
->set(l_osd_history_alloc_bytes
, SHIFT_ROUND_UP(buffer::get_history_alloc_bytes(), 20));
5118 logger
->set(l_osd_history_alloc_num
, buffer::get_history_alloc_num());
5119 logger
->set(l_osd_cached_crc
, buffer::get_cached_crc());
5120 logger
->set(l_osd_cached_crc_adjusted
, buffer::get_cached_crc_adjusted());
5121 logger
->set(l_osd_missed_crc
, buffer::get_missed_crc());
5123 // osd_lock is not being held, which means the OSD state
5124 // might change when doing the monitor report
5125 if (is_active() || is_waiting_for_healthy()) {
5126 heartbeat_lock
.Lock();
5128 heartbeat_lock
.Unlock();
5130 map_lock
.get_read();
5131 Mutex::Locker
l(mon_report_lock
);
5135 bool report
= false;
5136 utime_t now
= ceph_clock_now();
5137 pg_stat_queue_lock
.Lock();
5138 double backoff
= stats_ack_timeout
/ cct
->_conf
->osd_mon_ack_timeout
;
5139 double adjusted_min
= cct
->_conf
->osd_mon_report_interval_min
* backoff
;
5140 // note: we shouldn't adjust max because it must remain < the
5141 // mon's mon_osd_report_timeout (which defaults to 1.5x our
5143 double max
= cct
->_conf
->osd_mon_report_interval_max
;
5144 if (!outstanding_pg_stats
.empty() &&
5145 (now
- stats_ack_timeout
) > last_pg_stats_ack
) {
5146 dout(1) << __func__
<< " mon hasn't acked PGStats in "
5147 << now
- last_pg_stats_ack
5148 << " seconds, reconnecting elsewhere" << dendl
;
5150 last_pg_stats_ack
= now
; // reset clock
5151 last_pg_stats_sent
= utime_t();
5153 MAX(cct
->_conf
->osd_mon_ack_timeout
,
5154 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_factor
);
5155 outstanding_pg_stats
.clear();
5157 if (now
- last_pg_stats_sent
> max
) {
5158 osd_stat_updated
= true;
5160 } else if (service
.need_fullness_update()) {
5162 } else if ((int)outstanding_pg_stats
.size() >=
5163 cct
->_conf
->osd_mon_report_max_in_flight
) {
5164 dout(20) << __func__
<< " have max " << outstanding_pg_stats
5165 << " stats updates in flight" << dendl
;
5167 if (now
- last_mon_report
> adjusted_min
) {
5168 dout(20) << __func__
<< " stats backoff " << backoff
5169 << " adjusted_min " << adjusted_min
<< " - sending report"
5171 osd_stat_updated
= true;
5175 pg_stat_queue_lock
.Unlock();
5178 monc
->reopen_session();
5179 } else if (report
) {
5180 last_mon_report
= now
;
5182 // do any pending reports
5185 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5189 map_lock
.put_read();
5193 if (!scrub_random_backoff()) {
5196 service
.promote_throttle_recalibrate();
5197 bool need_send_beacon
= false;
5198 const auto now
= ceph::coarse_mono_clock::now();
5200 // borrow lec lock to pretect last_sent_beacon from changing
5201 Mutex::Locker l
{min_last_epoch_clean_lock
};
5202 const auto elapsed
= now
- last_sent_beacon
;
5203 if (chrono::duration_cast
<chrono::seconds
>(elapsed
).count() >
5204 cct
->_conf
->osd_beacon_report_interval
) {
5205 need_send_beacon
= true;
5208 if (need_send_beacon
) {
5213 check_ops_in_flight();
5214 service
.kick_recovery_queue();
5215 tick_timer_without_osd_lock
.add_event_after(OSD_TICK_INTERVAL
, new C_Tick_WithoutOSDLock(this));
5218 void OSD::check_ops_in_flight()
5220 vector
<string
> warnings
;
5221 if (op_tracker
.check_ops_in_flight(warnings
)) {
5222 for (vector
<string
>::iterator i
= warnings
.begin();
5223 i
!= warnings
.end();
5231 // setomapval <pool-id> [namespace/]<obj-name> <key> <val>
5232 // rmomapkey <pool-id> [namespace/]<obj-name> <key>
5233 // setomapheader <pool-id> [namespace/]<obj-name> <header>
5234 // getomap <pool> [namespace/]<obj-name>
5235 // truncobj <pool-id> [namespace/]<obj-name> <newlen>
5236 // injectmdataerr [namespace/]<obj-name> [shardid]
5237 // injectdataerr [namespace/]<obj-name> [shardid]
5239 // set_recovery_delay [utime]
5240 void TestOpsSocketHook::test_ops(OSDService
*service
, ObjectStore
*store
,
5241 const std::string
&command
, cmdmap_t
& cmdmap
, ostream
&ss
)
5244 //Support changing the omap on a single osd by using the Admin Socket to
5245 //directly request the osd make a change.
5246 if (command
== "setomapval" || command
== "rmomapkey" ||
5247 command
== "setomapheader" || command
== "getomap" ||
5248 command
== "truncobj" || command
== "injectmdataerr" ||
5249 command
== "injectdataerr"
5253 OSDMapRef curmap
= service
->get_osdmap();
5258 cmd_getval(service
->cct
, cmdmap
, "pool", poolstr
);
5259 pool
= curmap
->lookup_pg_pool_name(poolstr
);
5260 //If we can't find it by name then maybe id specified
5261 if (pool
< 0 && isdigit(poolstr
[0]))
5262 pool
= atoll(poolstr
.c_str());
5264 ss
<< "Invalid pool '" << poolstr
<< "''";
5268 string objname
, nspace
;
5269 cmd_getval(service
->cct
, cmdmap
, "objname", objname
);
5270 std::size_t found
= objname
.find_first_of('/');
5271 if (found
!= string::npos
) {
5272 nspace
= objname
.substr(0, found
);
5273 objname
= objname
.substr(found
+1);
5275 object_locator_t
oloc(pool
, nspace
);
5276 r
= curmap
->object_locator_to_pg(object_t(objname
), oloc
, rawpg
);
5279 ss
<< "Invalid namespace/objname";
5284 cmd_getval(service
->cct
, cmdmap
, "shardid", shardid
, int64_t(shard_id_t::NO_SHARD
));
5285 hobject_t
obj(object_t(objname
), string(""), CEPH_NOSNAP
, rawpg
.ps(), pool
, nspace
);
5286 ghobject_t
gobj(obj
, ghobject_t::NO_GEN
, shard_id_t(uint8_t(shardid
)));
5287 spg_t
pgid(curmap
->raw_pg_to_pg(rawpg
), shard_id_t(shardid
));
5288 if (curmap
->pg_is_ec(rawpg
)) {
5289 if ((command
!= "injectdataerr") && (command
!= "injectmdataerr")) {
5290 ss
<< "Must not call on ec pool, except injectdataerr or injectmdataerr";
5295 ObjectStore::Transaction t
;
5297 if (command
== "setomapval") {
5298 map
<string
, bufferlist
> newattrs
;
5301 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5302 cmd_getval(service
->cct
, cmdmap
, "val", valstr
);
5305 newattrs
[key
] = val
;
5306 t
.omap_setkeys(coll_t(pgid
), ghobject_t(obj
), newattrs
);
5307 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5309 ss
<< "error=" << r
;
5312 } else if (command
== "rmomapkey") {
5315 cmd_getval(service
->cct
, cmdmap
, "key", key
);
5318 t
.omap_rmkeys(coll_t(pgid
), ghobject_t(obj
), keys
);
5319 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5321 ss
<< "error=" << r
;
5324 } else if (command
== "setomapheader") {
5325 bufferlist newheader
;
5328 cmd_getval(service
->cct
, cmdmap
, "header", headerstr
);
5329 newheader
.append(headerstr
);
5330 t
.omap_setheader(coll_t(pgid
), ghobject_t(obj
), newheader
);
5331 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5333 ss
<< "error=" << r
;
5336 } else if (command
== "getomap") {
5337 //Debug: Output entire omap
5339 map
<string
, bufferlist
> keyvals
;
5340 r
= store
->omap_get(coll_t(pgid
), ghobject_t(obj
), &hdrbl
, &keyvals
);
5342 ss
<< "header=" << string(hdrbl
.c_str(), hdrbl
.length());
5343 for (map
<string
, bufferlist
>::iterator it
= keyvals
.begin();
5344 it
!= keyvals
.end(); ++it
)
5345 ss
<< " key=" << (*it
).first
<< " val="
5346 << string((*it
).second
.c_str(), (*it
).second
.length());
5348 ss
<< "error=" << r
;
5350 } else if (command
== "truncobj") {
5352 cmd_getval(service
->cct
, cmdmap
, "len", trunclen
);
5353 t
.truncate(coll_t(pgid
), ghobject_t(obj
), trunclen
);
5354 r
= store
->apply_transaction(service
->meta_osr
.get(), std::move(t
));
5356 ss
<< "error=" << r
;
5359 } else if (command
== "injectdataerr") {
5360 store
->inject_data_error(gobj
);
5362 } else if (command
== "injectmdataerr") {
5363 store
->inject_mdata_error(gobj
);
5368 if (command
== "set_recovery_delay") {
5370 cmd_getval(service
->cct
, cmdmap
, "utime", delay
, (int64_t)0);
5373 int r
= service
->cct
->_conf
->set_val("osd_recovery_delay_start",
5376 ss
<< "set_recovery_delay: error setting "
5377 << "osd_recovery_delay_start to '" << delay
<< "': error "
5381 service
->cct
->_conf
->apply_changes(NULL
);
5382 ss
<< "set_recovery_delay: set osd_recovery_delay_start "
5383 << "to " << service
->cct
->_conf
->osd_recovery_delay_start
;
5386 if (command
== "trigger_scrub") {
5388 OSDMapRef curmap
= service
->get_osdmap();
5392 cmd_getval(service
->cct
, cmdmap
, "pgid", pgidstr
);
5393 if (!pgid
.parse(pgidstr
.c_str())) {
5394 ss
<< "Invalid pgid specified";
5398 PG
*pg
= service
->osd
->_lookup_lock_pg(pgid
);
5399 if (pg
== nullptr) {
5400 ss
<< "Can't find pg " << pgid
;
5404 if (pg
->is_primary()) {
5405 pg
->unreg_next_scrub();
5406 const pg_pool_t
*p
= curmap
->get_pg_pool(pgid
.pool());
5407 double pool_scrub_max_interval
= 0;
5408 p
->opts
.get(pool_opts_t::SCRUB_MAX_INTERVAL
, &pool_scrub_max_interval
);
5409 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
5410 pool_scrub_max_interval
: g_conf
->osd_scrub_max_interval
;
5411 // Instead of marking must_scrub force a schedule scrub
5412 utime_t stamp
= ceph_clock_now();
5413 stamp
-= scrub_max_interval
;
5414 stamp
-= 100.0; // push back last scrub more for good measure
5415 pg
->info
.history
.last_scrub_stamp
= stamp
;
5416 pg
->reg_next_scrub();
5419 ss
<< "Not primary";
5424 if (command
== "injectfull") {
5427 OSDService::s_names state
;
5428 cmd_getval(service
->cct
, cmdmap
, "type", type
, string("full"));
5429 cmd_getval(service
->cct
, cmdmap
, "count", count
, (int64_t)-1);
5430 if (type
== "none" || count
== 0) {
5434 state
= service
->get_full_state(type
);
5435 if (state
== OSDService::s_names::INVALID
) {
5436 ss
<< "Invalid type use (none, nearfull, backfillfull, full, failsafe)";
5439 service
->set_injectfull(state
, count
);
5442 ss
<< "Internal error - command=" << command
;
5445 // =========================================
5448 ObjectStore
*store
, SnapMapper
*mapper
,
5450 ObjectStore::Sequencer
*osr
,
5451 coll_t coll
, DeletingStateRef dstate
,
5453 ThreadPool::TPHandle
&handle
)
5455 vector
<ghobject_t
> olist
;
5457 ObjectStore::Transaction t
;
5459 handle
.reset_tp_timeout();
5460 store
->collection_list(
5463 ghobject_t::get_max(),
5464 store
->get_ideal_list_max(),
5467 generic_dout(10) << __func__
<< " " << olist
<< dendl
;
5468 // default cont to true, this is safe because caller(OSD::RemoveWQ::_process())
5469 // will recheck the answer before it really goes on.
5471 for (vector
<ghobject_t
>::iterator i
= olist
.begin();
5476 OSDriver::OSTransaction
_t(osdriver
->get_transaction(&t
));
5477 int r
= mapper
->remove_oid(i
->hobj
, &_t
);
5478 if (r
!= 0 && r
!= -ENOENT
) {
5482 if (++num
>= cct
->_conf
->osd_target_transaction_size
) {
5484 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5485 cont
= dstate
->pause_clearing();
5486 handle
.suspend_tp_timeout();
5488 handle
.reset_tp_timeout();
5490 cont
= dstate
->resume_clearing();
5493 t
= ObjectStore::Transaction();
5499 store
->queue_transaction(osr
, std::move(t
), &waiter
);
5500 cont
= dstate
->pause_clearing();
5501 handle
.suspend_tp_timeout();
5503 handle
.reset_tp_timeout();
5505 cont
= dstate
->resume_clearing();
5507 // whether there are more objects to remove in the collection
5508 *finished
= next
.is_max();
5512 void OSD::RemoveWQ::_process(
5513 pair
<PGRef
, DeletingStateRef
> item
,
5514 ThreadPool::TPHandle
&handle
)
5517 PGRef
pg(item
.first
);
5518 SnapMapper
&mapper
= pg
->snap_mapper
;
5519 OSDriver
&driver
= pg
->osdriver
;
5520 coll_t coll
= coll_t(pg
->info
.pgid
);
5522 bool finished
= false;
5524 if (!item
.second
->start_or_resume_clearing())
5527 bool cont
= remove_dir(
5528 pg
->cct
, store
, &mapper
, &driver
, pg
->osr
.get(), coll
, item
.second
,
5533 if (item
.second
->pause_clearing())
5538 if (!item
.second
->start_deleting())
5541 ObjectStore::Transaction t
;
5542 PGLog::clear_info_log(pg
->info
.pgid
, &t
);
5544 if (cct
->_conf
->osd_inject_failure_on_pg_removal
) {
5545 generic_derr
<< "osd_inject_failure_on_pg_removal" << dendl
;
5548 t
.remove_collection(coll
);
5550 // We need the sequencer to stick around until the op is complete
5551 store
->queue_transaction(
5556 0, // onreadable sync
5557 new ContainerContext
<PGRef
>(pg
),
5560 item
.second
->finish_deleting();
5562 // =========================================
5564 void OSD::ms_handle_connect(Connection
*con
)
5566 dout(10) << __func__
<< " con " << con
<< dendl
;
5567 if (con
->get_peer_type() == CEPH_ENTITY_TYPE_MON
) {
5568 Mutex::Locker
l(osd_lock
);
5571 dout(10) << __func__
<< " on mon" << dendl
;
5575 } else if (is_booting()) {
5576 _send_boot(); // resend boot message
5578 map_lock
.get_read();
5579 Mutex::Locker
l2(mon_report_lock
);
5581 utime_t now
= ceph_clock_now();
5582 last_mon_report
= now
;
5584 // resend everything, it's a new session
5587 service
.requeue_pg_temp();
5588 service
.send_pg_temp();
5591 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
) {
5595 map_lock
.put_read();
5597 send_beacon(ceph::coarse_mono_clock::now());
5601 // full map requests may happen while active or pre-boot
5602 if (requested_full_first
) {
5603 rerequest_full_maps();
5608 void OSD::ms_handle_fast_connect(Connection
*con
)
5610 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5611 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5612 Session
*s
= static_cast<Session
*>(con
->get_priv());
5614 s
= new Session(cct
);
5615 con
->set_priv(s
->get());
5617 dout(10) << " new session (outgoing) " << s
<< " con=" << s
->con
5618 << " addr=" << s
->con
->get_peer_addr() << dendl
;
5619 // we don't connect to clients
5620 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5621 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5627 void OSD::ms_handle_fast_accept(Connection
*con
)
5629 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_MON
&&
5630 con
->get_peer_type() != CEPH_ENTITY_TYPE_MGR
) {
5631 Session
*s
= static_cast<Session
*>(con
->get_priv());
5633 s
= new Session(cct
);
5634 con
->set_priv(s
->get());
5636 dout(10) << "new session (incoming)" << s
<< " con=" << con
5637 << " addr=" << con
->get_peer_addr()
5638 << " must have raced with connect" << dendl
;
5639 assert(con
->get_peer_type() == CEPH_ENTITY_TYPE_OSD
);
5640 s
->entity_name
.set_type(CEPH_ENTITY_TYPE_OSD
);
5646 bool OSD::ms_handle_reset(Connection
*con
)
5648 Session
*session
= static_cast<Session
*>(con
->get_priv());
5649 dout(2) << "ms_handle_reset con " << con
<< " session " << session
<< dendl
;
5652 session
->wstate
.reset(con
);
5653 session
->con
.reset(NULL
); // break con <-> session ref cycle
5654 // note that we break session->con *before* the session_handle_reset
5655 // cleanup below. this avoids a race between us and
5656 // PG::add_backoff, Session::check_backoff, etc.
5657 session_handle_reset(session
);
5662 bool OSD::ms_handle_refused(Connection
*con
)
5664 if (!cct
->_conf
->osd_fast_fail_on_connection_refused
)
5667 Session
*session
= static_cast<Session
*>(con
->get_priv());
5668 dout(2) << "ms_handle_refused con " << con
<< " session " << session
<< dendl
;
5671 int type
= con
->get_peer_type();
5672 // handle only OSD failures here
5673 if (monc
&& (type
== CEPH_ENTITY_TYPE_OSD
)) {
5674 OSDMapRef osdmap
= get_osdmap();
5676 int id
= osdmap
->identify_osd_on_all_channels(con
->get_peer_addr());
5677 if (id
>= 0 && osdmap
->is_up(id
)) {
5678 // I'm cheating mon heartbeat grace logic, because we know it's not going
5679 // to respawn alone. +1 so we won't hit any boundary case.
5680 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(),
5681 osdmap
->get_inst(id
),
5682 cct
->_conf
->osd_heartbeat_grace
+ 1,
5683 osdmap
->get_epoch(),
5684 MOSDFailure::FLAG_IMMEDIATE
| MOSDFailure::FLAG_FAILED
5693 struct C_OSD_GetVersion
: public Context
{
5695 uint64_t oldest
, newest
;
5696 explicit C_OSD_GetVersion(OSD
*o
) : osd(o
), oldest(0), newest(0) {}
5697 void finish(int r
) override
{
5699 osd
->_got_mon_epochs(oldest
, newest
);
5703 void OSD::start_boot()
5705 if (!_is_healthy()) {
5706 // if we are not healthy, do not mark ourselves up (yet)
5707 dout(1) << "not healthy; waiting to boot" << dendl
;
5708 if (!is_waiting_for_healthy())
5709 start_waiting_for_healthy();
5710 // send pings sooner rather than later
5714 dout(1) << __func__
<< dendl
;
5715 set_state(STATE_PREBOOT
);
5716 waiting_for_luminous_mons
= false;
5717 dout(10) << "start_boot - have maps " << superblock
.oldest_map
5718 << ".." << superblock
.newest_map
<< dendl
;
5719 C_OSD_GetVersion
*c
= new C_OSD_GetVersion(this);
5720 monc
->get_version("osdmap", &c
->newest
, &c
->oldest
, c
);
5723 void OSD::_got_mon_epochs(epoch_t oldest
, epoch_t newest
)
5725 Mutex::Locker
l(osd_lock
);
5727 _preboot(oldest
, newest
);
5731 void OSD::_preboot(epoch_t oldest
, epoch_t newest
)
5733 assert(is_preboot());
5734 dout(10) << __func__
<< " _preboot mon has osdmaps "
5735 << oldest
<< ".." << newest
<< dendl
;
5737 // ensure our local fullness awareness is accurate
5740 // if our map within recent history, try to add ourselves to the osdmap.
5741 if (osdmap
->get_epoch() == 0) {
5742 derr
<< "waiting for initial osdmap" << dendl
;
5743 } else if (osdmap
->is_destroyed(whoami
)) {
5744 derr
<< "osdmap says I am destroyed, exiting" << dendl
;
5746 } else if (osdmap
->test_flag(CEPH_OSDMAP_NOUP
) || osdmap
->is_noup(whoami
)) {
5747 derr
<< "osdmap NOUP flag is set, waiting for it to clear" << dendl
;
5748 } else if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
)) {
5749 derr
<< "osdmap SORTBITWISE OSDMap flag is NOT set; please set it"
5751 } else if (osdmap
->require_osd_release
< CEPH_RELEASE_JEWEL
) {
5752 derr
<< "osdmap REQUIRE_JEWEL OSDMap flag is NOT set; please set it"
5754 } else if (!monc
->monmap
.get_required_features().contains_all(
5755 ceph::features::mon::FEATURE_LUMINOUS
)) {
5756 derr
<< "monmap REQUIRE_LUMINOUS is NOT set; must upgrade all monitors to "
5757 << "Luminous or later before Luminous OSDs will boot" << dendl
;
5758 waiting_for_luminous_mons
= true;
5759 } else if (service
.need_fullness_update()) {
5760 derr
<< "osdmap fullness state needs update" << dendl
;
5762 } else if (osdmap
->get_epoch() >= oldest
- 1 &&
5763 osdmap
->get_epoch() + cct
->_conf
->osd_map_message_max
> newest
) {
5768 // get all the latest maps
5769 if (osdmap
->get_epoch() + 1 >= oldest
)
5770 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5772 osdmap_subscribe(oldest
- 1, true);
5775 void OSD::send_full_update()
5777 if (!service
.need_fullness_update())
5780 if (service
.is_full()) {
5781 state
= CEPH_OSD_FULL
;
5782 } else if (service
.is_backfillfull()) {
5783 state
= CEPH_OSD_BACKFILLFULL
;
5784 } else if (service
.is_nearfull()) {
5785 state
= CEPH_OSD_NEARFULL
;
5788 OSDMap::calc_state_set(state
, s
);
5789 dout(10) << __func__
<< " want state " << s
<< dendl
;
5790 monc
->send_mon_message(new MOSDFull(osdmap
->get_epoch(), state
));
5793 void OSD::start_waiting_for_healthy()
5795 dout(1) << "start_waiting_for_healthy" << dendl
;
5796 set_state(STATE_WAITING_FOR_HEALTHY
);
5797 last_heartbeat_resample
= utime_t();
5799 // subscribe to osdmap updates, in case our peers really are known to be dead
5800 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
5803 bool OSD::_is_healthy()
5805 if (!cct
->get_heartbeat_map()->is_healthy()) {
5806 dout(1) << "is_healthy false -- internal heartbeat failed" << dendl
;
5810 if (is_waiting_for_healthy()) {
5811 Mutex::Locker
l(heartbeat_lock
);
5812 utime_t cutoff
= ceph_clock_now();
5813 cutoff
-= cct
->_conf
->osd_heartbeat_grace
;
5814 int num
= 0, up
= 0;
5815 for (map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.begin();
5816 p
!= heartbeat_peers
.end();
5818 if (p
->second
.is_healthy(cutoff
))
5822 if ((float)up
< (float)num
* cct
->_conf
->osd_heartbeat_min_healthy_ratio
) {
5823 dout(1) << "is_healthy false -- only " << up
<< "/" << num
<< " up peers (less than "
5824 << int(cct
->_conf
->osd_heartbeat_min_healthy_ratio
* 100.0) << "%)" << dendl
;
5832 void OSD::_send_boot()
5834 dout(10) << "_send_boot" << dendl
;
5835 entity_addr_t cluster_addr
= cluster_messenger
->get_myaddr();
5836 Connection
*local_connection
= cluster_messenger
->get_loopback_connection().get();
5837 if (cluster_addr
.is_blank_ip()) {
5838 int port
= cluster_addr
.get_port();
5839 cluster_addr
= client_messenger
->get_myaddr();
5840 cluster_addr
.set_port(port
);
5841 cluster_messenger
->set_addr_unknowns(cluster_addr
);
5842 dout(10) << " assuming cluster_addr ip matches client_addr" << dendl
;
5844 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
5848 cluster_messenger
->ms_deliver_handle_fast_connect(local_connection
);
5851 entity_addr_t hb_back_addr
= hb_back_server_messenger
->get_myaddr();
5852 local_connection
= hb_back_server_messenger
->get_loopback_connection().get();
5853 if (hb_back_addr
.is_blank_ip()) {
5854 int port
= hb_back_addr
.get_port();
5855 hb_back_addr
= cluster_addr
;
5856 hb_back_addr
.set_port(port
);
5857 hb_back_server_messenger
->set_addr_unknowns(hb_back_addr
);
5858 dout(10) << " assuming hb_back_addr ip matches cluster_addr" << dendl
;
5860 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
5864 hb_back_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
5867 entity_addr_t hb_front_addr
= hb_front_server_messenger
->get_myaddr();
5868 local_connection
= hb_front_server_messenger
->get_loopback_connection().get();
5869 if (hb_front_addr
.is_blank_ip()) {
5870 int port
= hb_front_addr
.get_port();
5871 hb_front_addr
= client_messenger
->get_myaddr();
5872 hb_front_addr
.set_port(port
);
5873 hb_front_server_messenger
->set_addr_unknowns(hb_front_addr
);
5874 dout(10) << " assuming hb_front_addr ip matches client_addr" << dendl
;
5876 Session
*s
= static_cast<Session
*>(local_connection
->get_priv());
5880 hb_front_server_messenger
->ms_deliver_handle_fast_connect(local_connection
);
5883 MOSDBoot
*mboot
= new MOSDBoot(superblock
, get_osdmap_epoch(), service
.get_boot_epoch(),
5884 hb_back_addr
, hb_front_addr
, cluster_addr
,
5886 dout(10) << " client_addr " << client_messenger
->get_myaddr()
5887 << ", cluster_addr " << cluster_addr
5888 << ", hb_back_addr " << hb_back_addr
5889 << ", hb_front_addr " << hb_front_addr
5891 _collect_metadata(&mboot
->metadata
);
5892 monc
->send_mon_message(mboot
);
5893 set_state(STATE_BOOTING
);
5896 void OSD::_collect_metadata(map
<string
,string
> *pm
)
5899 (*pm
)["osd_data"] = dev_path
;
5900 if (store
->get_type() == "filestore") {
5901 // not applicable for bluestore
5902 (*pm
)["osd_journal"] = journal_path
;
5904 (*pm
)["front_addr"] = stringify(client_messenger
->get_myaddr());
5905 (*pm
)["back_addr"] = stringify(cluster_messenger
->get_myaddr());
5906 (*pm
)["hb_front_addr"] = stringify(hb_front_server_messenger
->get_myaddr());
5907 (*pm
)["hb_back_addr"] = stringify(hb_back_server_messenger
->get_myaddr());
5910 (*pm
)["osd_objectstore"] = store
->get_type();
5911 (*pm
)["rotational"] = store_is_rotational
? "1" : "0";
5912 (*pm
)["journal_rotational"] = journal_is_rotational
? "1" : "0";
5913 (*pm
)["default_device_class"] = store
->get_default_device_class();
5914 store
->collect_metadata(pm
);
5916 collect_sys_info(pm
, cct
);
5918 std::string front_iface
, back_iface
;
5921 CEPH_PICK_ADDRESS_PUBLIC | CEPH_PICK_ADDRESS_CLUSTER,
5922 &front_iface, &back_iface);
5924 (*pm
)["front_iface"] = pick_iface(cct
,
5925 client_messenger
->get_myaddr().get_sockaddr_storage());
5926 (*pm
)["back_iface"] = pick_iface(cct
,
5927 cluster_messenger
->get_myaddr().get_sockaddr_storage());
5929 dout(10) << __func__
<< " " << *pm
<< dendl
;
5932 void OSD::queue_want_up_thru(epoch_t want
)
5934 map_lock
.get_read();
5935 epoch_t cur
= osdmap
->get_up_thru(whoami
);
5936 Mutex::Locker
l(mon_report_lock
);
5937 if (want
> up_thru_wanted
) {
5938 dout(10) << "queue_want_up_thru now " << want
<< " (was " << up_thru_wanted
<< ")"
5939 << ", currently " << cur
5941 up_thru_wanted
= want
;
5944 dout(10) << "queue_want_up_thru want " << want
<< " <= queued " << up_thru_wanted
5945 << ", currently " << cur
5948 map_lock
.put_read();
5951 void OSD::send_alive()
5953 assert(mon_report_lock
.is_locked());
5954 if (!osdmap
->exists(whoami
))
5956 epoch_t up_thru
= osdmap
->get_up_thru(whoami
);
5957 dout(10) << "send_alive up_thru currently " << up_thru
<< " want " << up_thru_wanted
<< dendl
;
5958 if (up_thru_wanted
> up_thru
) {
5959 dout(10) << "send_alive want " << up_thru_wanted
<< dendl
;
5960 monc
->send_mon_message(new MOSDAlive(osdmap
->get_epoch(), up_thru_wanted
));
5964 void OSD::request_full_map(epoch_t first
, epoch_t last
)
5966 dout(10) << __func__
<< " " << first
<< ".." << last
5967 << ", previously requested "
5968 << requested_full_first
<< ".." << requested_full_last
<< dendl
;
5969 assert(osd_lock
.is_locked());
5970 assert(first
> 0 && last
> 0);
5971 assert(first
<= last
);
5972 assert(first
>= requested_full_first
); // we shouldn't ever ask for older maps
5973 if (requested_full_first
== 0) {
5975 requested_full_first
= first
;
5976 requested_full_last
= last
;
5977 } else if (last
<= requested_full_last
) {
5981 // additional request
5982 first
= requested_full_last
+ 1;
5983 requested_full_last
= last
;
5985 MMonGetOSDMap
*req
= new MMonGetOSDMap
;
5986 req
->request_full(first
, last
);
5987 monc
->send_mon_message(req
);
5990 void OSD::got_full_map(epoch_t e
)
5992 assert(requested_full_first
<= requested_full_last
);
5993 assert(osd_lock
.is_locked());
5994 if (requested_full_first
== 0) {
5995 dout(20) << __func__
<< " " << e
<< ", nothing requested" << dendl
;
5998 if (e
< requested_full_first
) {
5999 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6000 << ".." << requested_full_last
6001 << ", ignoring" << dendl
;
6004 if (e
>= requested_full_last
) {
6005 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6006 << ".." << requested_full_last
<< ", resetting" << dendl
;
6007 requested_full_first
= requested_full_last
= 0;
6011 requested_full_first
= e
+ 1;
6013 dout(10) << __func__
<< " " << e
<< ", requested " << requested_full_first
6014 << ".." << requested_full_last
6015 << ", still need more" << dendl
;
6018 void OSD::requeue_failures()
6020 Mutex::Locker
l(heartbeat_lock
);
6021 unsigned old_queue
= failure_queue
.size();
6022 unsigned old_pending
= failure_pending
.size();
6023 for (map
<int,pair
<utime_t
,entity_inst_t
> >::iterator p
=
6024 failure_pending
.begin();
6025 p
!= failure_pending
.end(); ) {
6026 failure_queue
[p
->first
] = p
->second
.first
;
6027 failure_pending
.erase(p
++);
6029 dout(10) << __func__
<< " " << old_queue
<< " + " << old_pending
<< " -> "
6030 << failure_queue
.size() << dendl
;
6033 void OSD::send_failures()
6035 assert(map_lock
.is_locked());
6036 assert(mon_report_lock
.is_locked());
6037 Mutex::Locker
l(heartbeat_lock
);
6038 utime_t now
= ceph_clock_now();
6039 while (!failure_queue
.empty()) {
6040 int osd
= failure_queue
.begin()->first
;
6041 if (!failure_pending
.count(osd
)) {
6042 entity_inst_t i
= osdmap
->get_inst(osd
);
6043 int failed_for
= (int)(double)(now
- failure_queue
.begin()->second
);
6044 monc
->send_mon_message(new MOSDFailure(monc
->get_fsid(), i
, failed_for
,
6045 osdmap
->get_epoch()));
6046 failure_pending
[osd
] = make_pair(failure_queue
.begin()->second
, i
);
6048 failure_queue
.erase(osd
);
6052 void OSD::send_still_alive(epoch_t epoch
, const entity_inst_t
&i
)
6054 MOSDFailure
*m
= new MOSDFailure(monc
->get_fsid(), i
, 0, epoch
, MOSDFailure::FLAG_ALIVE
);
6055 monc
->send_mon_message(m
);
6058 void OSD::send_pg_stats(const utime_t
&now
)
6060 assert(map_lock
.is_locked());
6061 assert(osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
);
6062 dout(20) << "send_pg_stats" << dendl
;
6064 osd_stat_t cur_stat
= service
.get_osd_stat();
6066 cur_stat
.os_perf_stat
= store
->get_cur_stats();
6068 pg_stat_queue_lock
.Lock();
6070 if (osd_stat_updated
|| !pg_stat_queue
.empty()) {
6071 last_pg_stats_sent
= now
;
6072 osd_stat_updated
= false;
6074 dout(10) << "send_pg_stats - " << pg_stat_queue
.size() << " pgs updated" << dendl
;
6076 utime_t
had_for(now
);
6077 had_for
-= had_map_since
;
6079 MPGStats
*m
= new MPGStats(monc
->get_fsid(), osdmap
->get_epoch(), had_for
);
6081 uint64_t tid
= ++pg_stat_tid
;
6083 m
->osd_stat
= cur_stat
;
6085 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6089 if (!pg
->is_primary()) { // we hold map_lock; role is stable.
6090 pg
->stat_queue_item
.remove_myself();
6091 pg
->put("pg_stat_queue");
6094 pg
->pg_stats_publish_lock
.Lock();
6095 if (pg
->pg_stats_publish_valid
) {
6096 m
->pg_stat
[pg
->info
.pgid
.pgid
] = pg
->pg_stats_publish
;
6097 dout(25) << " sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6098 << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6100 dout(25) << " NOT sending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
<< ":"
6101 << pg
->pg_stats_publish
.reported_seq
<< ", not valid" << dendl
;
6103 pg
->pg_stats_publish_lock
.Unlock();
6106 if (last_pg_stats_ack
== utime_t() || !outstanding_pg_stats
.empty()) {
6107 last_pg_stats_ack
= ceph_clock_now();
6109 outstanding_pg_stats
.insert(tid
);
6110 dout(20) << __func__
<< " updates pending: " << outstanding_pg_stats
<< dendl
;
6112 monc
->send_mon_message(m
);
6115 pg_stat_queue_lock
.Unlock();
6118 void OSD::handle_pg_stats_ack(MPGStatsAck
*ack
)
6120 dout(10) << "handle_pg_stats_ack " << dendl
;
6122 if (!require_mon_peer(ack
)) {
6127 // NOTE: we may get replies from a previous mon even while
6128 // outstanding_pg_stats is empty if reconnecting races with replies
6131 pg_stat_queue_lock
.Lock();
6133 last_pg_stats_ack
= ceph_clock_now();
6135 // decay timeout slowly (analogous to TCP)
6137 MAX(cct
->_conf
->osd_mon_ack_timeout
,
6138 stats_ack_timeout
* cct
->_conf
->osd_stats_ack_timeout_decay
);
6139 dout(20) << __func__
<< " timeout now " << stats_ack_timeout
<< dendl
;
6141 if (ack
->get_tid() > pg_stat_tid_flushed
) {
6142 pg_stat_tid_flushed
= ack
->get_tid();
6143 pg_stat_queue_cond
.Signal();
6146 xlist
<PG
*>::iterator p
= pg_stat_queue
.begin();
6152 auto acked
= ack
->pg_stat
.find(pg
->info
.pgid
.pgid
);
6153 if (acked
!= ack
->pg_stat
.end()) {
6154 pg
->pg_stats_publish_lock
.Lock();
6155 if (acked
->second
.first
== pg
->pg_stats_publish
.reported_seq
&&
6156 acked
->second
.second
== pg
->pg_stats_publish
.reported_epoch
) {
6157 dout(25) << " ack on " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6158 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6159 pg
->stat_queue_item
.remove_myself();
6160 pg
->put("pg_stat_queue");
6162 dout(25) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6163 << ":" << pg
->pg_stats_publish
.reported_seq
<< " > acked "
6164 << acked
->second
<< dendl
;
6166 pg
->pg_stats_publish_lock
.Unlock();
6168 dout(30) << " still pending " << pg
->info
.pgid
<< " " << pg
->pg_stats_publish
.reported_epoch
6169 << ":" << pg
->pg_stats_publish
.reported_seq
<< dendl
;
6173 outstanding_pg_stats
.erase(ack
->get_tid());
6174 dout(20) << __func__
<< " still pending: " << outstanding_pg_stats
<< dendl
;
6176 pg_stat_queue_lock
.Unlock();
6181 void OSD::flush_pg_stats()
6183 dout(10) << "flush_pg_stats" << dendl
;
6185 utime_t now
= ceph_clock_now();
6186 map_lock
.get_read();
6187 mon_report_lock
.Lock();
6189 mon_report_lock
.Unlock();
6190 map_lock
.put_read();
6193 pg_stat_queue_lock
.Lock();
6194 uint64_t tid
= pg_stat_tid
;
6195 dout(10) << "flush_pg_stats waiting for stats tid " << tid
<< " to flush" << dendl
;
6196 while (tid
> pg_stat_tid_flushed
)
6197 pg_stat_queue_cond
.Wait(pg_stat_queue_lock
);
6198 dout(10) << "flush_pg_stats finished waiting for stats tid " << tid
<< " to flush" << dendl
;
6199 pg_stat_queue_lock
.Unlock();
6204 void OSD::send_beacon(const ceph::coarse_mono_clock::time_point
& now
)
6206 const auto& monmap
= monc
->monmap
;
6207 // send beacon to mon even if we are just connected, and the monmap is not
6208 // initialized yet by then.
6209 if (monmap
.epoch
> 0 &&
6210 monmap
.get_required_features().contains_all(
6211 ceph::features::mon::FEATURE_LUMINOUS
)) {
6212 dout(20) << __func__
<< " sending" << dendl
;
6213 MOSDBeacon
* beacon
= nullptr;
6215 Mutex::Locker l
{min_last_epoch_clean_lock
};
6216 beacon
= new MOSDBeacon(osdmap
->get_epoch(), min_last_epoch_clean
);
6217 std::swap(beacon
->pgs
, min_last_epoch_clean_pgs
);
6218 last_sent_beacon
= now
;
6220 monc
->send_mon_message(beacon
);
6222 dout(20) << __func__
<< " not sending" << dendl
;
6226 void OSD::handle_command(MMonCommand
*m
)
6228 if (!require_mon_peer(m
)) {
6233 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), NULL
);
6234 command_wq
.queue(c
);
6238 void OSD::handle_command(MCommand
*m
)
6240 ConnectionRef con
= m
->get_connection();
6241 Session
*session
= static_cast<Session
*>(con
->get_priv());
6243 con
->send_message(new MCommandReply(m
, -EPERM
));
6248 OSDCap
& caps
= session
->caps
;
6251 if (!caps
.allow_all() || m
->get_source().is_mon()) {
6252 con
->send_message(new MCommandReply(m
, -EPERM
));
6257 Command
*c
= new Command(m
->cmd
, m
->get_tid(), m
->get_data(), con
.get());
6258 command_wq
.queue(c
);
6268 string availability
;
6269 } osd_commands
[] = {
6271 #define COMMAND(parsesig, helptext, module, perm, availability) \
6272 {parsesig, helptext, module, perm, availability},
6274 // yes, these are really pg commands, but there's a limit to how
6275 // much work it's worth. The OSD returns all of them. Make this
6276 // form (pg <pgid> <cmd>) valid only for the cli.
6277 // Rest uses "tell <pgid> <cmd>"
6280 "name=pgid,type=CephPgid " \
6281 "name=cmd,type=CephChoices,strings=query", \
6282 "show details of a specific pg", "osd", "r", "cli")
6284 "name=pgid,type=CephPgid " \
6285 "name=cmd,type=CephChoices,strings=mark_unfound_lost " \
6286 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6287 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6290 "name=pgid,type=CephPgid " \
6291 "name=cmd,type=CephChoices,strings=list_missing " \
6292 "name=offset,type=CephString,req=false",
6293 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6296 // new form: tell <pgid> <cmd> for both cli and rest
6299 "show details of a specific pg", "osd", "r", "cli,rest")
6300 COMMAND("mark_unfound_lost " \
6301 "name=mulcmd,type=CephChoices,strings=revert|delete", \
6302 "mark all unfound objects in this pg as lost, either removing or reverting to a prior version if one is available",
6303 "osd", "rw", "cli,rest")
6304 COMMAND("list_missing " \
6305 "name=offset,type=CephString,req=false",
6306 "list missing objects on this pg, perhaps starting at an offset given in JSON",
6307 "osd", "r", "cli,rest")
6308 COMMAND("perf histogram dump "
6309 "name=logger,type=CephString,req=false "
6310 "name=counter,type=CephString,req=false",
6311 "Get histogram data",
6312 "osd", "r", "cli,rest")
6314 // tell <osd.n> commands. Validation of osd.n must be special-cased in client
6315 COMMAND("version", "report version of OSD", "osd", "r", "cli,rest")
6316 COMMAND("get_command_descriptions", "list commands descriptions", "osd", "r", "cli,rest")
6317 COMMAND("injectargs " \
6318 "name=injected_args,type=CephString,n=N",
6319 "inject configuration arguments into running OSD",
6320 "osd", "rw", "cli,rest")
6321 COMMAND("config set " \
6322 "name=key,type=CephString name=value,type=CephString",
6323 "Set a configuration option at runtime (not persistent)",
6324 "osd", "rw", "cli,rest")
6325 COMMAND("cluster_log " \
6326 "name=level,type=CephChoices,strings=error,warning,info,debug " \
6327 "name=message,type=CephString,n=N",
6328 "log a message to the cluster log",
6329 "osd", "rw", "cli,rest")
6331 "name=count,type=CephInt,req=false " \
6332 "name=size,type=CephInt,req=false " \
6333 "name=object_size,type=CephInt,req=false " \
6334 "name=object_num,type=CephInt,req=false ", \
6335 "OSD benchmark: write <count> <size>-byte objects, " \
6336 "(default 1G size 4MB). Results in log.",
6337 "osd", "rw", "cli,rest")
6338 COMMAND("flush_pg_stats", "flush pg stats", "osd", "rw", "cli,rest")
6340 "name=heapcmd,type=CephChoices,strings=dump|start_profiler|stop_profiler|release|stats", \
6341 "show heap usage info (available only if compiled with tcmalloc)", \
6342 "osd", "rw", "cli,rest")
6343 COMMAND("debug dump_missing " \
6344 "name=filename,type=CephFilepath",
6345 "dump missing objects to a named file", "osd", "r", "cli,rest")
6346 COMMAND("debug kick_recovery_wq " \
6347 "name=delay,type=CephInt,range=0",
6348 "set osd_recovery_delay_start to <val>", "osd", "rw", "cli,rest")
6349 COMMAND("cpu_profiler " \
6350 "name=arg,type=CephChoices,strings=status|flush",
6351 "run cpu profiling on daemon", "osd", "rw", "cli,rest")
6352 COMMAND("dump_pg_recovery_stats", "dump pg recovery statistics",
6353 "osd", "r", "cli,rest")
6354 COMMAND("reset_pg_recovery_stats", "reset pg recovery statistics",
6355 "osd", "rw", "cli,rest")
6357 "compact object store's omap. "
6358 "WARNING: Compaction probably slows your requests",
6359 "osd", "rw", "cli,rest")
6362 void OSD::do_command(Connection
*con
, ceph_tid_t tid
, vector
<string
>& cmd
, bufferlist
& data
)
6365 stringstream ss
, ds
;
6369 dout(20) << "do_command tid " << tid
<< " " << cmd
<< dendl
;
6371 map
<string
, cmd_vartype
> cmdmap
;
6375 boost::scoped_ptr
<Formatter
> f
;
6378 ss
<< "no command given";
6382 if (!cmdmap_from_json(cmd
, &cmdmap
, ss
)) {
6387 cmd_getval(cct
, cmdmap
, "prefix", prefix
);
6389 if (prefix
== "get_command_descriptions") {
6391 JSONFormatter
*f
= new JSONFormatter();
6392 f
->open_object_section("command_descriptions");
6393 for (OSDCommand
*cp
= osd_commands
;
6394 cp
< &osd_commands
[ARRAY_SIZE(osd_commands
)]; cp
++) {
6396 ostringstream secname
;
6397 secname
<< "cmd" << setfill('0') << std::setw(3) << cmdnum
;
6398 dump_cmddesc_to_json(f
, secname
.str(), cp
->cmdstring
, cp
->helpstring
,
6399 cp
->module
, cp
->perm
, cp
->availability
, 0);
6402 f
->close_section(); // command_descriptions
6409 cmd_getval(cct
, cmdmap
, "format", format
);
6410 f
.reset(Formatter::create(format
));
6412 if (prefix
== "version") {
6414 f
->open_object_section("version");
6415 f
->dump_string("version", pretty_version_to_str());
6419 ds
<< pretty_version_to_str();
6423 else if (prefix
== "injectargs") {
6424 vector
<string
> argsvec
;
6425 cmd_getval(cct
, cmdmap
, "injected_args", argsvec
);
6427 if (argsvec
.empty()) {
6429 ss
<< "ignoring empty injectargs";
6432 string args
= argsvec
.front();
6433 for (vector
<string
>::iterator a
= ++argsvec
.begin(); a
!= argsvec
.end(); ++a
)
6436 r
= cct
->_conf
->injectargs(args
, &ss
);
6439 else if (prefix
== "config set") {
6442 cmd_getval(cct
, cmdmap
, "key", key
);
6443 cmd_getval(cct
, cmdmap
, "value", val
);
6445 r
= cct
->_conf
->set_val(key
, val
, true, &ss
);
6447 cct
->_conf
->apply_changes(nullptr);
6451 else if (prefix
== "cluster_log") {
6453 cmd_getval(cct
, cmdmap
, "message", msg
);
6456 ss
<< "ignoring empty log message";
6459 string message
= msg
.front();
6460 for (vector
<string
>::iterator a
= ++msg
.begin(); a
!= msg
.end(); ++a
)
6461 message
+= " " + *a
;
6463 cmd_getval(cct
, cmdmap
, "level", lvl
);
6464 clog_type level
= string_to_clog_type(lvl
);
6467 ss
<< "unknown level '" << lvl
<< "'";
6470 clog
->do_log(level
, message
);
6473 // either 'pg <pgid> <command>' or
6474 // 'tell <pgid>' (which comes in without any of that prefix)?
6476 else if (prefix
== "pg" ||
6477 prefix
== "query" ||
6478 prefix
== "mark_unfound_lost" ||
6479 prefix
== "list_missing"
6483 if (!cmd_getval(cct
, cmdmap
, "pgid", pgidstr
)) {
6484 ss
<< "no pgid specified";
6486 } else if (!pgid
.parse(pgidstr
.c_str())) {
6487 ss
<< "couldn't parse pgid '" << pgidstr
<< "'";
6492 if (osdmap
->get_primary_shard(pgid
, &pcand
) &&
6493 (pg
= _lookup_lock_pg(pcand
))) {
6494 if (pg
->is_primary()) {
6495 // simulate pg <pgid> cmd= for pg->do-command
6497 cmd_putval(cct
, cmdmap
, "cmd", prefix
);
6498 r
= pg
->do_command(cmdmap
, ss
, data
, odata
, con
, tid
);
6501 // don't reply, pg will do so async
6505 ss
<< "not primary for pgid " << pgid
;
6507 // send them the latest diff to ensure they realize the mapping
6509 service
.send_incremental_map(osdmap
->get_epoch() - 1, con
, osdmap
);
6511 // do not reply; they will get newer maps and realize they
6518 ss
<< "i don't have pgid " << pgid
;
6524 else if (prefix
== "bench") {
6527 int64_t osize
, onum
;
6528 // default count 1G, size 4MB
6529 cmd_getval(cct
, cmdmap
, "count", count
, (int64_t)1 << 30);
6530 cmd_getval(cct
, cmdmap
, "size", bsize
, (int64_t)4 << 20);
6531 cmd_getval(cct
, cmdmap
, "object_size", osize
, (int64_t)0);
6532 cmd_getval(cct
, cmdmap
, "object_num", onum
, (int64_t)0);
6534 ceph::shared_ptr
<ObjectStore::Sequencer
> osr (std::make_shared
<
6535 ObjectStore::Sequencer
>("bench"));
6537 uint32_t duration
= cct
->_conf
->osd_bench_duration
;
6539 if (bsize
> (int64_t) cct
->_conf
->osd_bench_max_block_size
) {
6540 // let us limit the block size because the next checks rely on it
6541 // having a sane value. If we allow any block size to be set things
6542 // can still go sideways.
6543 ss
<< "block 'size' values are capped at "
6544 << prettybyte_t(cct
->_conf
->osd_bench_max_block_size
) << ". If you wish to use"
6545 << " a higher value, please adjust 'osd_bench_max_block_size'";
6548 } else if (bsize
< (int64_t) (1 << 20)) {
6549 // entering the realm of small block sizes.
6550 // limit the count to a sane value, assuming a configurable amount of
6551 // IOPS and duration, so that the OSD doesn't get hung up on this,
6552 // preventing timeouts from going off
6554 bsize
* duration
* cct
->_conf
->osd_bench_small_size_max_iops
;
6555 if (count
> max_count
) {
6556 ss
<< "'count' values greater than " << max_count
6557 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6558 << cct
->_conf
->osd_bench_small_size_max_iops
<< " IOPS,"
6559 << " for " << duration
<< " seconds,"
6560 << " can cause ill effects on osd. "
6561 << " Please adjust 'osd_bench_small_size_max_iops' with a higher"
6562 << " value if you wish to use a higher 'count'.";
6567 // 1MB block sizes are big enough so that we get more stuff done.
6568 // However, to avoid the osd from getting hung on this and having
6569 // timers being triggered, we are going to limit the count assuming
6570 // a configurable throughput and duration.
6571 // NOTE: max_count is the total amount of bytes that we believe we
6572 // will be able to write during 'duration' for the given
6573 // throughput. The block size hardly impacts this unless it's
6574 // way too big. Given we already check how big the block size
6575 // is, it's safe to assume everything will check out.
6577 cct
->_conf
->osd_bench_large_size_max_throughput
* duration
;
6578 if (count
> max_count
) {
6579 ss
<< "'count' values greater than " << max_count
6580 << " for a block size of " << prettybyte_t(bsize
) << ", assuming "
6581 << prettybyte_t(cct
->_conf
->osd_bench_large_size_max_throughput
) << "/s,"
6582 << " for " << duration
<< " seconds,"
6583 << " can cause ill effects on osd. "
6584 << " Please adjust 'osd_bench_large_size_max_throughput'"
6585 << " with a higher value if you wish to use a higher 'count'.";
6591 if (osize
&& bsize
> osize
)
6594 dout(1) << " bench count " << count
6595 << " bsize " << prettybyte_t(bsize
) << dendl
;
6597 ObjectStore::Transaction cleanupt
;
6599 if (osize
&& onum
) {
6601 bufferptr
bp(osize
);
6603 bl
.push_back(std::move(bp
));
6604 bl
.rebuild_page_aligned();
6605 for (int i
=0; i
<onum
; ++i
) {
6607 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", i
);
6609 hobject_t
soid(sobject_t(oid
, 0));
6610 ObjectStore::Transaction t
;
6611 t
.write(coll_t(), ghobject_t(soid
), 0, osize
, bl
);
6612 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6613 cleanupt
.remove(coll_t(), ghobject_t(soid
));
6618 bufferptr
bp(bsize
);
6620 bl
.push_back(std::move(bp
));
6621 bl
.rebuild_page_aligned();
6625 if (!osr
->flush_commit(&waiter
)) {
6630 utime_t start
= ceph_clock_now();
6631 for (int64_t pos
= 0; pos
< count
; pos
+= bsize
) {
6633 unsigned offset
= 0;
6634 if (onum
&& osize
) {
6635 snprintf(nm
, sizeof(nm
), "disk_bw_test_%d", (int)(rand() % onum
));
6636 offset
= rand() % (osize
/ bsize
) * bsize
;
6638 snprintf(nm
, sizeof(nm
), "disk_bw_test_%lld", (long long)pos
);
6641 hobject_t
soid(sobject_t(oid
, 0));
6642 ObjectStore::Transaction t
;
6643 t
.write(coll_t::meta(), ghobject_t(soid
), offset
, bsize
, bl
);
6644 store
->queue_transaction(osr
.get(), std::move(t
), NULL
);
6645 if (!onum
|| !osize
)
6646 cleanupt
.remove(coll_t::meta(), ghobject_t(soid
));
6651 if (!osr
->flush_commit(&waiter
)) {
6655 utime_t end
= ceph_clock_now();
6658 store
->queue_transaction(osr
.get(), std::move(cleanupt
), NULL
);
6661 if (!osr
->flush_commit(&waiter
)) {
6666 uint64_t rate
= (double)count
/ (end
- start
);
6668 f
->open_object_section("osd_bench_results");
6669 f
->dump_int("bytes_written", count
);
6670 f
->dump_int("blocksize", bsize
);
6671 f
->dump_unsigned("bytes_per_sec", rate
);
6675 ss
<< "bench: wrote " << prettybyte_t(count
)
6676 << " in blocks of " << prettybyte_t(bsize
) << " in "
6677 << (end
-start
) << " sec at " << prettybyte_t(rate
) << "/sec";
6681 else if (prefix
== "flush_pg_stats") {
6682 if (osdmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
6683 mgrc
.send_pgstats();
6684 ds
<< service
.get_osd_stat_seq() << "\n";
6690 else if (prefix
== "heap") {
6691 r
= ceph::osd_cmds::heap(*cct
, cmdmap
, *f
, ds
);
6694 else if (prefix
== "debug dump_missing") {
6696 cmd_getval(cct
, cmdmap
, "filename", file_name
);
6697 std::ofstream
fout(file_name
.c_str());
6698 if (!fout
.is_open()) {
6699 ss
<< "failed to open file '" << file_name
<< "'";
6704 fout
<< "*** osd " << whoami
<< ": dump_missing ***" << std::endl
;
6705 RWLock::RLocker
l(pg_map_lock
);
6706 for (ceph::unordered_map
<spg_t
, PG
*>::const_iterator pg_map_e
= pg_map
.begin();
6707 pg_map_e
!= pg_map
.end(); ++pg_map_e
) {
6708 PG
*pg
= pg_map_e
->second
;
6711 fout
<< *pg
<< std::endl
;
6712 std::map
<hobject_t
, pg_missing_item
>::const_iterator mend
=
6713 pg
->pg_log
.get_missing().get_items().end();
6714 std::map
<hobject_t
, pg_missing_item
>::const_iterator mi
=
6715 pg
->pg_log
.get_missing().get_items().begin();
6716 for (; mi
!= mend
; ++mi
) {
6717 fout
<< mi
->first
<< " -> " << mi
->second
<< std::endl
;
6718 if (!pg
->missing_loc
.needs_recovery(mi
->first
))
6720 if (pg
->missing_loc
.is_unfound(mi
->first
))
6721 fout
<< " unfound ";
6722 const set
<pg_shard_t
> &mls(pg
->missing_loc
.get_locations(mi
->first
));
6725 fout
<< "missing_loc: " << mls
<< std::endl
;
6733 else if (prefix
== "debug kick_recovery_wq") {
6735 cmd_getval(cct
, cmdmap
, "delay", delay
);
6738 r
= cct
->_conf
->set_val("osd_recovery_delay_start", oss
.str().c_str());
6740 ss
<< "kick_recovery_wq: error setting "
6741 << "osd_recovery_delay_start to '" << delay
<< "': error "
6745 cct
->_conf
->apply_changes(NULL
);
6746 ss
<< "kicking recovery queue. set osd_recovery_delay_start "
6747 << "to " << cct
->_conf
->osd_recovery_delay_start
;
6750 else if (prefix
== "cpu_profiler") {
6752 cmd_getval(cct
, cmdmap
, "arg", arg
);
6753 vector
<string
> argvec
;
6754 get_str_vec(arg
, argvec
);
6755 cpu_profiler_handle_command(argvec
, ds
);
6758 else if (prefix
== "dump_pg_recovery_stats") {
6761 pg_recovery_stats
.dump_formatted(f
.get());
6764 pg_recovery_stats
.dump(s
);
6765 ds
<< "dump pg recovery stats: " << s
.str();
6769 else if (prefix
== "reset_pg_recovery_stats") {
6770 ss
<< "reset pg recovery stats";
6771 pg_recovery_stats
.reset();
6774 else if (prefix
== "perf histogram dump") {
6776 std::string counter
;
6777 cmd_getval(cct
, cmdmap
, "logger", logger
);
6778 cmd_getval(cct
, cmdmap
, "counter", counter
);
6780 cct
->get_perfcounters_collection()->dump_formatted_histograms(
6781 f
.get(), false, logger
, counter
);
6786 else if (prefix
== "compact") {
6787 dout(1) << "triggering manual compaction" << dendl
;
6788 auto start
= ceph::coarse_mono_clock::now();
6790 auto end
= ceph::coarse_mono_clock::now();
6791 auto time_span
= chrono::duration_cast
<chrono::duration
<double>>(end
- start
);
6792 dout(1) << "finished manual compaction in "
6793 << time_span
.count()
6794 << " seconds" << dendl
;
6795 ss
<< "compacted omap in " << time_span
.count() << " seconds";
6799 ss
<< "unrecognized command! " << cmd
;
6806 dout(0) << "do_command r=" << r
<< " " << rs
<< dendl
;
6809 MCommandReply
*reply
= new MCommandReply(r
, rs
);
6810 reply
->set_tid(tid
);
6811 reply
->set_data(odata
);
6812 con
->send_message(reply
);
6816 bool OSD::heartbeat_dispatch(Message
*m
)
6818 dout(30) << "heartbeat_dispatch " << m
<< dendl
;
6819 switch (m
->get_type()) {
6822 dout(10) << "ping from " << m
->get_source_inst() << dendl
;
6827 handle_osd_ping(static_cast<MOSDPing
*>(m
));
6831 dout(0) << "dropping unexpected message " << *m
<< " from " << m
->get_source_inst() << dendl
;
6838 bool OSD::ms_dispatch(Message
*m
)
6840 dout(20) << "OSD::ms_dispatch: " << *m
<< dendl
;
6841 if (m
->get_type() == MSG_OSD_MARK_ME_DOWN
) {
6842 service
.got_stop_ack();
6850 if (is_stopping()) {
6864 void OSD::maybe_share_map(
6869 if (!op
->check_send_map
) {
6872 epoch_t last_sent_epoch
= 0;
6874 session
->sent_epoch_lock
.lock();
6875 last_sent_epoch
= session
->last_sent_epoch
;
6876 session
->sent_epoch_lock
.unlock();
6878 const Message
*m
= op
->get_req();
6881 m
->get_connection().get(),
6884 session
? &last_sent_epoch
: NULL
);
6886 session
->sent_epoch_lock
.lock();
6887 if (session
->last_sent_epoch
< last_sent_epoch
) {
6888 session
->last_sent_epoch
= last_sent_epoch
;
6890 session
->sent_epoch_lock
.unlock();
6892 op
->check_send_map
= false;
6895 void OSD::dispatch_session_waiting(Session
*session
, OSDMapRef osdmap
)
6897 assert(session
->session_dispatch_lock
.is_locked());
6899 auto i
= session
->waiting_on_map
.begin();
6900 while (i
!= session
->waiting_on_map
.end()) {
6901 OpRequestRef op
= &(*i
);
6902 assert(ms_can_fast_dispatch(op
->get_req()));
6903 const MOSDFastDispatchOp
*m
= static_cast<const MOSDFastDispatchOp
*>(
6905 if (m
->get_min_epoch() > osdmap
->get_epoch()) {
6908 session
->waiting_on_map
.erase(i
++);
6912 if (m
->get_type() == CEPH_MSG_OSD_OP
) {
6913 pg_t actual_pgid
= osdmap
->raw_pg_to_pg(
6914 static_cast<const MOSDOp
*>(m
)->get_pg());
6915 if (!osdmap
->get_primary_shard(actual_pgid
, &pgid
)) {
6919 pgid
= m
->get_spg();
6921 enqueue_op(pgid
, op
, m
->get_map_epoch());
6924 if (session
->waiting_on_map
.empty()) {
6925 clear_session_waiting_on_map(session
);
6927 register_session_waiting_on_map(session
);
6931 void OSD::ms_fast_dispatch(Message
*m
)
6934 if (service
.is_stopping()) {
6938 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
6941 osd_reqid_t reqid
= op
->get_reqid();
6943 tracepoint(osd
, ms_fast_dispatch
, reqid
.name
._type
,
6944 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
6948 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
6950 // note sender epoch, min req'd epoch
6951 op
->sent_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch();
6952 op
->min_epoch
= static_cast<MOSDFastDispatchOp
*>(m
)->get_min_epoch();
6953 assert(op
->min_epoch
<= op
->sent_epoch
); // sanity check!
6955 service
.maybe_inject_dispatch_delay();
6957 if (m
->get_connection()->has_features(CEPH_FEATUREMASK_RESEND_ON_SPLIT
) ||
6958 m
->get_type() != CEPH_MSG_OSD_OP
) {
6959 // queue it directly
6961 static_cast<MOSDFastDispatchOp
*>(m
)->get_spg(),
6963 static_cast<MOSDFastDispatchOp
*>(m
)->get_map_epoch());
6965 // legacy client, and this is an MOSDOp (the *only* fast dispatch
6966 // message that didn't have an explicit spg_t); we need to map
6967 // them to an spg_t while preserving delivery order.
6968 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
6971 Mutex::Locker
l(session
->session_dispatch_lock
);
6973 session
->waiting_on_map
.push_back(*op
);
6974 OSDMapRef nextmap
= service
.get_nextmap_reserved();
6975 dispatch_session_waiting(session
, nextmap
);
6976 service
.release_map(nextmap
);
6981 OID_EVENT_TRACE_WITH_MSG(m
, "MS_FAST_DISPATCH_END", false);
6984 void OSD::ms_fast_preprocess(Message
*m
)
6986 if (m
->get_connection()->get_peer_type() == CEPH_ENTITY_TYPE_OSD
) {
6987 if (m
->get_type() == CEPH_MSG_OSD_MAP
) {
6988 MOSDMap
*mm
= static_cast<MOSDMap
*>(m
);
6989 Session
*s
= static_cast<Session
*>(m
->get_connection()->get_priv());
6991 s
->received_map_lock
.lock();
6992 s
->received_map_epoch
= mm
->get_last();
6993 s
->received_map_lock
.unlock();
7000 bool OSD::ms_get_authorizer(int dest_type
, AuthAuthorizer
**authorizer
, bool force_new
)
7002 dout(10) << "OSD::ms_get_authorizer type=" << ceph_entity_type_name(dest_type
) << dendl
;
7004 if (is_stopping()) {
7005 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7009 if (dest_type
== CEPH_ENTITY_TYPE_MON
)
7013 /* the MonClient checks keys every tick(), so we should just wait for that cycle
7015 if (monc
->wait_auth_rotating(10) < 0) {
7016 derr
<< "OSD::ms_get_authorizer wait_auth_rotating failed" << dendl
;
7021 *authorizer
= monc
->build_authorizer(dest_type
);
7022 return *authorizer
!= NULL
;
7026 bool OSD::ms_verify_authorizer(Connection
*con
, int peer_type
,
7027 int protocol
, bufferlist
& authorizer_data
, bufferlist
& authorizer_reply
,
7028 bool& isvalid
, CryptoKey
& session_key
)
7030 AuthAuthorizeHandler
*authorize_handler
= 0;
7031 switch (peer_type
) {
7032 case CEPH_ENTITY_TYPE_MDS
:
7034 * note: mds is technically a client from our perspective, but
7035 * this makes the 'cluster' consistent w/ monitor's usage.
7037 case CEPH_ENTITY_TYPE_OSD
:
7038 case CEPH_ENTITY_TYPE_MGR
:
7039 authorize_handler
= authorize_handler_cluster_registry
->get_handler(protocol
);
7042 authorize_handler
= authorize_handler_service_registry
->get_handler(protocol
);
7044 if (!authorize_handler
) {
7045 dout(0) << "No AuthAuthorizeHandler found for protocol " << protocol
<< dendl
;
7050 AuthCapsInfo caps_info
;
7053 uint64_t auid
= CEPH_AUTH_UID_DEFAULT
;
7055 RotatingKeyRing
*keys
= monc
->rotating_secrets
.get();
7057 isvalid
= authorize_handler
->verify_authorizer(
7059 authorizer_data
, authorizer_reply
, name
, global_id
, caps_info
, session_key
,
7062 dout(10) << __func__
<< " no rotating_keys (yet), denied" << dendl
;
7067 Session
*s
= static_cast<Session
*>(con
->get_priv());
7069 s
= new Session(cct
);
7070 con
->set_priv(s
->get());
7072 dout(10) << " new session " << s
<< " con=" << s
->con
<< " addr=" << s
->con
->get_peer_addr() << dendl
;
7075 s
->entity_name
= name
;
7076 if (caps_info
.allow_all
)
7077 s
->caps
.set_allow_all();
7080 if (caps_info
.caps
.length() > 0) {
7081 bufferlist::iterator p
= caps_info
.caps
.begin();
7086 catch (buffer::error
& e
) {
7088 bool success
= s
->caps
.parse(str
);
7090 dout(10) << " session " << s
<< " " << s
->entity_name
<< " has caps " << s
->caps
<< " '" << str
<< "'" << dendl
;
7092 dout(10) << " session " << s
<< " " << s
->entity_name
<< " failed to parse caps '" << str
<< "'" << dendl
;
7100 void OSD::do_waiters()
7102 assert(osd_lock
.is_locked());
7104 dout(10) << "do_waiters -- start" << dendl
;
7105 while (!finished
.empty()) {
7106 OpRequestRef next
= finished
.front();
7107 finished
.pop_front();
7110 dout(10) << "do_waiters -- finish" << dendl
;
7113 void OSD::dispatch_op(OpRequestRef op
)
7115 switch (op
->get_req()->get_type()) {
7117 case MSG_OSD_PG_CREATE
:
7118 handle_pg_create(op
);
7120 case MSG_OSD_PG_NOTIFY
:
7121 handle_pg_notify(op
);
7123 case MSG_OSD_PG_QUERY
:
7124 handle_pg_query(op
);
7126 case MSG_OSD_PG_LOG
:
7129 case MSG_OSD_PG_REMOVE
:
7130 handle_pg_remove(op
);
7132 case MSG_OSD_PG_INFO
:
7135 case MSG_OSD_PG_TRIM
:
7138 case MSG_OSD_BACKFILL_RESERVE
:
7139 handle_pg_backfill_reserve(op
);
7141 case MSG_OSD_RECOVERY_RESERVE
:
7142 handle_pg_recovery_reserve(op
);
7147 void OSD::_dispatch(Message
*m
)
7149 assert(osd_lock
.is_locked());
7150 dout(20) << "_dispatch " << m
<< " " << *m
<< dendl
;
7152 switch (m
->get_type()) {
7154 // -- don't need lock --
7156 dout(10) << "ping from " << m
->get_source() << dendl
;
7160 // -- don't need OSDMap --
7162 // map and replication
7163 case CEPH_MSG_OSD_MAP
:
7164 handle_osd_map(static_cast<MOSDMap
*>(m
));
7168 case MSG_PGSTATSACK
:
7169 handle_pg_stats_ack(static_cast<MPGStatsAck
*>(m
));
7172 case MSG_MON_COMMAND
:
7173 handle_command(static_cast<MMonCommand
*>(m
));
7176 handle_command(static_cast<MCommand
*>(m
));
7180 handle_scrub(static_cast<MOSDScrub
*>(m
));
7183 case MSG_OSD_FORCE_RECOVERY
:
7184 handle_force_recovery(m
);
7187 // -- need OSDMap --
7189 case MSG_OSD_PG_CREATE
:
7190 case MSG_OSD_PG_NOTIFY
:
7191 case MSG_OSD_PG_QUERY
:
7192 case MSG_OSD_PG_LOG
:
7193 case MSG_OSD_PG_REMOVE
:
7194 case MSG_OSD_PG_INFO
:
7195 case MSG_OSD_PG_TRIM
:
7196 case MSG_OSD_BACKFILL_RESERVE
:
7197 case MSG_OSD_RECOVERY_RESERVE
:
7199 OpRequestRef op
= op_tracker
.create_request
<OpRequest
, Message
*>(m
);
7201 op
->osd_trace
.init("osd op", &trace_endpoint
, &m
->trace
);
7202 // no map? starting up?
7204 dout(7) << "no OSDMap, not booted" << dendl
;
7205 logger
->inc(l_osd_waiting_for_map
);
7206 waiting_for_osdmap
.push_back(op
);
7207 op
->mark_delayed("no osdmap");
7217 void OSD::handle_pg_scrub(MOSDScrub
*m
, PG
*pg
)
7220 if (pg
->is_primary()) {
7221 pg
->unreg_next_scrub();
7222 pg
->scrubber
.must_scrub
= true;
7223 pg
->scrubber
.must_deep_scrub
= m
->deep
|| m
->repair
;
7224 pg
->scrubber
.must_repair
= m
->repair
;
7225 pg
->reg_next_scrub();
7226 dout(10) << "marking " << *pg
<< " for scrub" << dendl
;
7231 void OSD::handle_scrub(MOSDScrub
*m
)
7233 dout(10) << "handle_scrub " << *m
<< dendl
;
7234 if (!require_mon_or_mgr_peer(m
)) {
7238 if (m
->fsid
!= monc
->get_fsid()) {
7239 dout(0) << "handle_scrub fsid " << m
->fsid
<< " != " << monc
->get_fsid() << dendl
;
7244 RWLock::RLocker
l(pg_map_lock
);
7245 if (m
->scrub_pgs
.empty()) {
7246 for (ceph::unordered_map
<spg_t
, PG
*>::iterator p
= pg_map
.begin();
7249 handle_pg_scrub(m
, p
->second
);
7251 for (vector
<pg_t
>::iterator p
= m
->scrub_pgs
.begin();
7252 p
!= m
->scrub_pgs
.end();
7255 if (osdmap
->get_primary_shard(*p
, &pcand
)) {
7256 auto pg_map_entry
= pg_map
.find(pcand
);
7257 if (pg_map_entry
!= pg_map
.end()) {
7258 handle_pg_scrub(m
, pg_map_entry
->second
);
7267 bool OSD::scrub_random_backoff()
7269 bool coin_flip
= (rand() / (double)RAND_MAX
>=
7270 cct
->_conf
->osd_scrub_backoff_ratio
);
7272 dout(20) << "scrub_random_backoff lost coin flip, randomly backing off" << dendl
;
7278 OSDService::ScrubJob::ScrubJob(CephContext
* cct
,
7279 const spg_t
& pg
, const utime_t
& timestamp
,
7280 double pool_scrub_min_interval
,
7281 double pool_scrub_max_interval
, bool must
)
7284 sched_time(timestamp
),
7287 // if not explicitly requested, postpone the scrub with a random delay
7289 double scrub_min_interval
= pool_scrub_min_interval
> 0 ?
7290 pool_scrub_min_interval
: cct
->_conf
->osd_scrub_min_interval
;
7291 double scrub_max_interval
= pool_scrub_max_interval
> 0 ?
7292 pool_scrub_max_interval
: cct
->_conf
->osd_scrub_max_interval
;
7294 sched_time
+= scrub_min_interval
;
7295 double r
= rand() / (double)RAND_MAX
;
7297 scrub_min_interval
* cct
->_conf
->osd_scrub_interval_randomize_ratio
* r
;
7298 deadline
+= scrub_max_interval
;
7302 bool OSDService::ScrubJob::ScrubJob::operator<(const OSDService::ScrubJob
& rhs
) const {
7303 if (sched_time
< rhs
.sched_time
)
7305 if (sched_time
> rhs
.sched_time
)
7307 return pgid
< rhs
.pgid
;
7310 bool OSD::scrub_time_permit(utime_t now
)
7313 time_t tt
= now
.sec();
7314 localtime_r(&tt
, &bdt
);
7315 bool time_permit
= false;
7316 if (cct
->_conf
->osd_scrub_begin_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7317 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
&& bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7321 if (bdt
.tm_hour
>= cct
->_conf
->osd_scrub_begin_hour
|| bdt
.tm_hour
< cct
->_conf
->osd_scrub_end_hour
) {
7326 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7327 << " - " << cct
->_conf
->osd_scrub_end_hour
7328 << " now " << bdt
.tm_hour
<< " = no" << dendl
;
7330 dout(20) << __func__
<< " should run between " << cct
->_conf
->osd_scrub_begin_hour
7331 << " - " << cct
->_conf
->osd_scrub_end_hour
7332 << " now " << bdt
.tm_hour
<< " = yes" << dendl
;
7337 bool OSD::scrub_load_below_threshold()
7340 if (getloadavg(loadavgs
, 3) != 3) {
7341 dout(10) << __func__
<< " couldn't read loadavgs\n" << dendl
;
7345 // allow scrub if below configured threshold
7346 if (loadavgs
[0] < cct
->_conf
->osd_scrub_load_threshold
) {
7347 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7348 << " < max " << cct
->_conf
->osd_scrub_load_threshold
7349 << " = yes" << dendl
;
7353 // allow scrub if below daily avg and currently decreasing
7354 if (loadavgs
[0] < daily_loadavg
&& loadavgs
[0] < loadavgs
[2]) {
7355 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7356 << " < daily_loadavg " << daily_loadavg
7357 << " and < 15m avg " << loadavgs
[2]
7358 << " = yes" << dendl
;
7362 dout(20) << __func__
<< " loadavg " << loadavgs
[0]
7363 << " >= max " << cct
->_conf
->osd_scrub_load_threshold
7364 << " and ( >= daily_loadavg " << daily_loadavg
7365 << " or >= 15m avg " << loadavgs
[2]
7366 << ") = no" << dendl
;
7370 void OSD::sched_scrub()
7372 // if not permitted, fail fast
7373 if (!service
.can_inc_scrubs_pending()) {
7376 if (!cct
->_conf
->osd_scrub_during_recovery
&& service
.is_recovery_active()) {
7377 dout(20) << __func__
<< " not scheduling scrubs due to active recovery" << dendl
;
7382 utime_t now
= ceph_clock_now();
7383 bool time_permit
= scrub_time_permit(now
);
7384 bool load_is_low
= scrub_load_below_threshold();
7385 dout(20) << "sched_scrub load_is_low=" << (int)load_is_low
<< dendl
;
7387 OSDService::ScrubJob scrub
;
7388 if (service
.first_scrub_stamp(&scrub
)) {
7390 dout(30) << "sched_scrub examine " << scrub
.pgid
<< " at " << scrub
.sched_time
<< dendl
;
7392 if (scrub
.sched_time
> now
) {
7393 // save ourselves some effort
7394 dout(10) << "sched_scrub " << scrub
.pgid
<< " scheduled at " << scrub
.sched_time
7395 << " > " << now
<< dendl
;
7399 if ((scrub
.deadline
>= now
) && !(time_permit
&& load_is_low
)) {
7400 dout(10) << __func__
<< " not scheduling scrub for " << scrub
.pgid
<< " due to "
7401 << (!time_permit
? "time not permit" : "high load") << dendl
;
7405 PG
*pg
= _lookup_lock_pg(scrub
.pgid
);
7408 if (pg
->get_pgbackend()->scrub_supported() && pg
->is_active()) {
7409 dout(10) << "sched_scrub scrubbing " << scrub
.pgid
<< " at " << scrub
.sched_time
7410 << (pg
->scrubber
.must_scrub
? ", explicitly requested" :
7411 (load_is_low
? ", load_is_low" : " deadline < now"))
7413 if (pg
->sched_scrub()) {
7419 } while (service
.next_scrub_stamp(scrub
, &scrub
));
7421 dout(20) << "sched_scrub done" << dendl
;
7426 // =====================================================
7429 void OSD::wait_for_new_map(OpRequestRef op
)
7432 if (waiting_for_osdmap
.empty()) {
7433 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
7436 logger
->inc(l_osd_waiting_for_map
);
7437 waiting_for_osdmap
.push_back(op
);
7438 op
->mark_delayed("wait for new map");
7443 * assimilate new OSDMap(s). scan pgs, etc.
7446 void OSD::note_down_osd(int peer
)
7448 assert(osd_lock
.is_locked());
7449 cluster_messenger
->mark_down(osdmap
->get_cluster_addr(peer
));
7451 heartbeat_lock
.Lock();
7452 failure_queue
.erase(peer
);
7453 failure_pending
.erase(peer
);
7454 map
<int,HeartbeatInfo
>::iterator p
= heartbeat_peers
.find(peer
);
7455 if (p
!= heartbeat_peers
.end()) {
7456 p
->second
.con_back
->mark_down();
7457 if (p
->second
.con_front
) {
7458 p
->second
.con_front
->mark_down();
7460 heartbeat_peers
.erase(p
);
7462 heartbeat_lock
.Unlock();
7465 void OSD::note_up_osd(int peer
)
7467 service
.forget_peer_epoch(peer
, osdmap
->get_epoch() - 1);
7468 heartbeat_set_peers_need_update();
7471 struct C_OnMapCommit
: public Context
{
7473 epoch_t first
, last
;
7475 C_OnMapCommit(OSD
*o
, epoch_t f
, epoch_t l
, MOSDMap
*m
)
7476 : osd(o
), first(f
), last(l
), msg(m
) {}
7477 void finish(int r
) override
{
7478 osd
->_committed_osd_maps(first
, last
, msg
);
7483 struct C_OnMapApply
: public Context
{
7484 OSDService
*service
;
7485 list
<OSDMapRef
> pinned_maps
;
7487 C_OnMapApply(OSDService
*service
,
7488 const list
<OSDMapRef
> &pinned_maps
,
7490 : service(service
), pinned_maps(pinned_maps
), e(e
) {}
7491 void finish(int r
) override
{
7492 service
->clear_map_bl_cache_pins(e
);
7496 void OSD::osdmap_subscribe(version_t epoch
, bool force_request
)
7498 Mutex::Locker
l(osdmap_subscribe_lock
);
7499 if (latest_subscribed_epoch
>= epoch
&& !force_request
)
7502 latest_subscribed_epoch
= MAX(epoch
, latest_subscribed_epoch
);
7504 if (monc
->sub_want_increment("osdmap", epoch
, CEPH_SUBSCRIBE_ONETIME
) ||
7510 void OSD::trim_maps(epoch_t oldest
, int nreceived
, bool skip_maps
)
7512 epoch_t min
= std::min(oldest
, service
.map_cache
.cached_key_lower_bound());
7513 if (min
<= superblock
.oldest_map
)
7517 ObjectStore::Transaction t
;
7518 for (epoch_t e
= superblock
.oldest_map
; e
< min
; ++e
) {
7519 dout(20) << " removing old osdmap epoch " << e
<< dendl
;
7520 t
.remove(coll_t::meta(), get_osdmap_pobject_name(e
));
7521 t
.remove(coll_t::meta(), get_inc_osdmap_pobject_name(e
));
7522 superblock
.oldest_map
= e
+ 1;
7524 if (num
>= cct
->_conf
->osd_target_transaction_size
&& num
>= nreceived
) {
7525 service
.publish_superblock(superblock
);
7526 write_superblock(t
);
7527 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7531 // skip_maps leaves us with a range of old maps if we fail to remove all
7532 // of them before moving superblock.oldest_map forward to the first map
7533 // in the incoming MOSDMap msg. so we should continue removing them in
7534 // this case, even we could do huge series of delete transactions all at
7541 service
.publish_superblock(superblock
);
7542 write_superblock(t
);
7543 int tr
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), nullptr);
7546 // we should not remove the cached maps
7547 assert(min
<= service
.map_cache
.cached_key_lower_bound());
7550 void OSD::handle_osd_map(MOSDMap
*m
)
7552 assert(osd_lock
.is_locked());
7553 // Keep a ref in the list until we get the newly received map written
7554 // onto disk. This is important because as long as the refs are alive,
7555 // the OSDMaps will be pinned in the cache and we won't try to read it
7556 // off of disk. Otherwise these maps will probably not stay in the cache,
7557 // and reading those OSDMaps before they are actually written can result
7559 list
<OSDMapRef
> pinned_maps
;
7560 if (m
->fsid
!= monc
->get_fsid()) {
7561 dout(0) << "handle_osd_map fsid " << m
->fsid
<< " != "
7562 << monc
->get_fsid() << dendl
;
7566 if (is_initializing()) {
7567 dout(0) << "ignoring osdmap until we have initialized" << dendl
;
7572 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
7573 if (session
&& !(session
->entity_name
.is_mon() ||
7574 session
->entity_name
.is_osd())) {
7576 dout(10) << "got osd map from Session " << session
7577 << " which we can't take maps from (not a mon or osd)" << dendl
;
7585 // share with the objecter
7587 service
.objecter
->handle_osd_map(m
);
7589 epoch_t first
= m
->get_first();
7590 epoch_t last
= m
->get_last();
7591 dout(3) << "handle_osd_map epochs [" << first
<< "," << last
<< "], i have "
7592 << superblock
.newest_map
7593 << ", src has [" << m
->oldest_map
<< "," << m
->newest_map
<< "]"
7596 logger
->inc(l_osd_map
);
7597 logger
->inc(l_osd_mape
, last
- first
+ 1);
7598 if (first
<= superblock
.newest_map
)
7599 logger
->inc(l_osd_mape_dup
, superblock
.newest_map
- first
+ 1);
7600 if (service
.max_oldest_map
< m
->oldest_map
) {
7601 service
.max_oldest_map
= m
->oldest_map
;
7602 assert(service
.max_oldest_map
>= superblock
.oldest_map
);
7605 // make sure there is something new, here, before we bother flushing
7606 // the queues and such
7607 if (last
<= superblock
.newest_map
) {
7608 dout(10) << " no new maps here, dropping" << dendl
;
7614 bool skip_maps
= false;
7615 if (first
> superblock
.newest_map
+ 1) {
7616 dout(10) << "handle_osd_map message skips epochs "
7617 << superblock
.newest_map
+ 1 << ".." << (first
-1) << dendl
;
7618 if (m
->oldest_map
<= superblock
.newest_map
+ 1) {
7619 osdmap_subscribe(superblock
.newest_map
+ 1, false);
7623 // always try to get the full range of maps--as many as we can. this
7624 // 1- is good to have
7625 // 2- is at present the only way to ensure that we get a *full* map as
7627 if (m
->oldest_map
< first
) {
7628 osdmap_subscribe(m
->oldest_map
- 1, true);
7635 ObjectStore::Transaction t
;
7636 uint64_t txn_size
= 0;
7638 // store new maps: queue for disk and put in the osdmap cache
7639 epoch_t start
= MAX(superblock
.newest_map
+ 1, first
);
7640 for (epoch_t e
= start
; e
<= last
; e
++) {
7641 if (txn_size
>= t
.get_num_bytes()) {
7642 derr
<< __func__
<< " transaction size overflowed" << dendl
;
7643 assert(txn_size
< t
.get_num_bytes());
7645 txn_size
= t
.get_num_bytes();
7646 map
<epoch_t
,bufferlist
>::iterator p
;
7647 p
= m
->maps
.find(e
);
7648 if (p
!= m
->maps
.end()) {
7649 dout(10) << "handle_osd_map got full map for epoch " << e
<< dendl
;
7650 OSDMap
*o
= new OSDMap
;
7651 bufferlist
& bl
= p
->second
;
7655 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7656 t
.write(coll_t::meta(), fulloid
, 0, bl
.length(), bl
);
7658 pinned_maps
.push_back(add_map(o
));
7664 p
= m
->incremental_maps
.find(e
);
7665 if (p
!= m
->incremental_maps
.end()) {
7666 dout(10) << "handle_osd_map got inc map for epoch " << e
<< dendl
;
7667 bufferlist
& bl
= p
->second
;
7668 ghobject_t oid
= get_inc_osdmap_pobject_name(e
);
7669 t
.write(coll_t::meta(), oid
, 0, bl
.length(), bl
);
7670 pin_map_inc_bl(e
, bl
);
7672 OSDMap
*o
= new OSDMap
;
7675 bool got
= get_map_bl(e
- 1, obl
);
7680 OSDMap::Incremental inc
;
7681 bufferlist::iterator p
= bl
.begin();
7683 if (o
->apply_incremental(inc
) < 0) {
7684 derr
<< "ERROR: bad fsid? i have " << osdmap
->get_fsid() << " and inc has " << inc
.fsid
<< dendl
;
7685 assert(0 == "bad fsid");
7689 o
->encode(fbl
, inc
.encode_features
| CEPH_FEATURE_RESERVED
);
7691 bool injected_failure
= false;
7692 if (cct
->_conf
->osd_inject_bad_map_crc_probability
> 0 &&
7693 (rand() % 10000) < cct
->_conf
->osd_inject_bad_map_crc_probability
*10000.0) {
7694 derr
<< __func__
<< " injecting map crc failure" << dendl
;
7695 injected_failure
= true;
7698 if ((inc
.have_crc
&& o
->get_crc() != inc
.full_crc
) || injected_failure
) {
7699 dout(2) << "got incremental " << e
7700 << " but failed to encode full with correct crc; requesting"
7702 clog
->warn() << "failed to encode map e" << e
<< " with expected crc";
7703 dout(20) << "my encoded map was:\n";
7704 fbl
.hexdump(*_dout
);
7707 request_full_map(e
, last
);
7713 ghobject_t fulloid
= get_osdmap_pobject_name(e
);
7714 t
.write(coll_t::meta(), fulloid
, 0, fbl
.length(), fbl
);
7716 pinned_maps
.push_back(add_map(o
));
7720 assert(0 == "MOSDMap lied about what maps it had?");
7723 // even if this map isn't from a mon, we may have satisfied our subscription
7724 monc
->sub_got("osdmap", last
);
7726 if (!m
->maps
.empty() && requested_full_first
) {
7727 dout(10) << __func__
<< " still missing full maps " << requested_full_first
7728 << ".." << requested_full_last
<< dendl
;
7729 rerequest_full_maps();
7732 if (superblock
.oldest_map
) {
7733 // make sure we at least keep pace with incoming maps
7734 trim_maps(m
->oldest_map
, last
- first
+ 1, skip_maps
);
7737 if (!superblock
.oldest_map
|| skip_maps
)
7738 superblock
.oldest_map
= first
;
7739 superblock
.newest_map
= last
;
7740 superblock
.current_epoch
= last
;
7742 // note in the superblock that we were clean thru the prior epoch
7743 epoch_t boot_epoch
= service
.get_boot_epoch();
7744 if (boot_epoch
&& boot_epoch
>= superblock
.mounted
) {
7745 superblock
.mounted
= boot_epoch
;
7746 superblock
.clean_thru
= last
;
7749 // superblock and commit
7750 write_superblock(t
);
7751 store
->queue_transaction(
7752 service
.meta_osr
.get(),
7754 new C_OnMapApply(&service
, pinned_maps
, last
),
7755 new C_OnMapCommit(this, start
, last
, m
), 0);
7756 service
.publish_superblock(superblock
);
7759 void OSD::_committed_osd_maps(epoch_t first
, epoch_t last
, MOSDMap
*m
)
7761 dout(10) << __func__
<< " " << first
<< ".." << last
<< dendl
;
7762 if (is_stopping()) {
7763 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7766 Mutex::Locker
l(osd_lock
);
7767 if (is_stopping()) {
7768 dout(10) << __func__
<< " bailing, we are shutting down" << dendl
;
7771 map_lock
.get_write();
7773 bool do_shutdown
= false;
7774 bool do_restart
= false;
7775 bool network_error
= false;
7777 // advance through the new maps
7778 for (epoch_t cur
= first
; cur
<= last
; cur
++) {
7779 dout(10) << " advance to epoch " << cur
7780 << " (<= last " << last
7781 << " <= newest_map " << superblock
.newest_map
7784 OSDMapRef newmap
= get_map(cur
);
7785 assert(newmap
); // we just cached it above!
7787 // start blacklisting messages sent to peers that go down.
7788 service
.pre_publish_map(newmap
);
7790 // kill connections to newly down osds
7791 bool waited_for_reservations
= false;
7793 osdmap
->get_all_osds(old
);
7794 for (set
<int>::iterator p
= old
.begin(); p
!= old
.end(); ++p
) {
7796 osdmap
->is_up(*p
) && // in old map
7797 newmap
->is_down(*p
)) { // but not the new one
7798 if (!waited_for_reservations
) {
7799 service
.await_reserved_maps();
7800 waited_for_reservations
= true;
7803 } else if (*p
!= whoami
&&
7804 osdmap
->is_down(*p
) &&
7805 newmap
->is_up(*p
)) {
7810 if ((osdmap
->test_flag(CEPH_OSDMAP_NOUP
) !=
7811 newmap
->test_flag(CEPH_OSDMAP_NOUP
)) ||
7812 (osdmap
->is_noup(whoami
) != newmap
->is_noup(whoami
))) {
7813 dout(10) << __func__
<< " NOUP flag changed in " << newmap
->get_epoch()
7816 // this captures the case where we sent the boot message while
7817 // NOUP was being set on the mon and our boot request was
7818 // dropped, and then later it is cleared. it imperfectly
7819 // handles the case where our original boot message was not
7820 // dropped and we restart even though we might have booted, but
7821 // that is harmless (boot will just take slightly longer).
7825 if (osdmap
->require_osd_release
< CEPH_RELEASE_LUMINOUS
&&
7826 newmap
->require_osd_release
>= CEPH_RELEASE_LUMINOUS
) {
7827 dout(10) << __func__
<< " require_osd_release reached luminous in "
7828 << newmap
->get_epoch() << dendl
;
7829 clear_pg_stat_queue();
7830 clear_outstanding_pg_stats();
7836 service
.retrieve_epochs(&boot_epoch
, &up_epoch
, NULL
);
7838 osdmap
->is_up(whoami
) &&
7839 osdmap
->get_inst(whoami
) == client_messenger
->get_myinst()) {
7840 up_epoch
= osdmap
->get_epoch();
7841 dout(10) << "up_epoch is " << up_epoch
<< dendl
;
7843 boot_epoch
= osdmap
->get_epoch();
7844 dout(10) << "boot_epoch is " << boot_epoch
<< dendl
;
7846 service
.set_epochs(&boot_epoch
, &up_epoch
, NULL
);
7850 had_map_since
= ceph_clock_now();
7852 epoch_t _bind_epoch
= service
.get_bind_epoch();
7853 if (osdmap
->is_up(whoami
) &&
7854 osdmap
->get_addr(whoami
) == client_messenger
->get_myaddr() &&
7855 _bind_epoch
< osdmap
->get_up_from(whoami
)) {
7858 dout(1) << "state: booting -> active" << dendl
;
7859 set_state(STATE_ACTIVE
);
7861 // set incarnation so that osd_reqid_t's we generate for our
7862 // objecter requests are unique across restarts.
7863 service
.objecter
->set_client_incarnation(osdmap
->get_epoch());
7867 if (osdmap
->get_epoch() > 0 &&
7869 if (!osdmap
->exists(whoami
)) {
7870 dout(0) << "map says i do not exist. shutting down." << dendl
;
7871 do_shutdown
= true; // don't call shutdown() while we have
7872 // everything paused
7873 } else if (!osdmap
->is_up(whoami
) ||
7874 !osdmap
->get_addr(whoami
).probably_equals(
7875 client_messenger
->get_myaddr()) ||
7876 !osdmap
->get_cluster_addr(whoami
).probably_equals(
7877 cluster_messenger
->get_myaddr()) ||
7878 !osdmap
->get_hb_back_addr(whoami
).probably_equals(
7879 hb_back_server_messenger
->get_myaddr()) ||
7880 (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
7881 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
7882 hb_front_server_messenger
->get_myaddr()))) {
7883 if (!osdmap
->is_up(whoami
)) {
7884 if (service
.is_preparing_to_stop() || service
.is_stopping()) {
7885 service
.got_stop_ack();
7887 clog
->warn() << "Monitor daemon marked osd." << whoami
<< " down, "
7888 "but it is still running";
7889 clog
->debug() << "map e" << osdmap
->get_epoch()
7890 << " wrongly marked me down at e"
7891 << osdmap
->get_down_at(whoami
);
7893 } else if (!osdmap
->get_addr(whoami
).probably_equals(
7894 client_messenger
->get_myaddr())) {
7895 clog
->error() << "map e" << osdmap
->get_epoch()
7896 << " had wrong client addr (" << osdmap
->get_addr(whoami
)
7897 << " != my " << client_messenger
->get_myaddr() << ")";
7898 } else if (!osdmap
->get_cluster_addr(whoami
).probably_equals(
7899 cluster_messenger
->get_myaddr())) {
7900 clog
->error() << "map e" << osdmap
->get_epoch()
7901 << " had wrong cluster addr ("
7902 << osdmap
->get_cluster_addr(whoami
)
7903 << " != my " << cluster_messenger
->get_myaddr() << ")";
7904 } else if (!osdmap
->get_hb_back_addr(whoami
).probably_equals(
7905 hb_back_server_messenger
->get_myaddr())) {
7906 clog
->error() << "map e" << osdmap
->get_epoch()
7907 << " had wrong heartbeat back addr ("
7908 << osdmap
->get_hb_back_addr(whoami
)
7909 << " != my " << hb_back_server_messenger
->get_myaddr()
7911 } else if (osdmap
->get_hb_front_addr(whoami
) != entity_addr_t() &&
7912 !osdmap
->get_hb_front_addr(whoami
).probably_equals(
7913 hb_front_server_messenger
->get_myaddr())) {
7914 clog
->error() << "map e" << osdmap
->get_epoch()
7915 << " had wrong heartbeat front addr ("
7916 << osdmap
->get_hb_front_addr(whoami
)
7917 << " != my " << hb_front_server_messenger
->get_myaddr()
7921 if (!service
.is_stopping()) {
7922 epoch_t up_epoch
= 0;
7923 epoch_t bind_epoch
= osdmap
->get_epoch();
7924 service
.set_epochs(NULL
,&up_epoch
, &bind_epoch
);
7928 utime_t now
= ceph_clock_now();
7929 utime_t grace
= utime_t(cct
->_conf
->osd_max_markdown_period
, 0);
7930 osd_markdown_log
.push_back(now
);
7931 //clear all out-of-date log
7932 while (!osd_markdown_log
.empty() &&
7933 osd_markdown_log
.front() + grace
< now
)
7934 osd_markdown_log
.pop_front();
7935 if ((int)osd_markdown_log
.size() > cct
->_conf
->osd_max_markdown_count
) {
7936 dout(0) << __func__
<< " marked down "
7937 << osd_markdown_log
.size()
7938 << " > osd_max_markdown_count "
7939 << cct
->_conf
->osd_max_markdown_count
7940 << " in last " << grace
<< " seconds, shutting down"
7946 start_waiting_for_healthy();
7948 set
<int> avoid_ports
;
7949 #if defined(__FreeBSD__)
7950 // prevent FreeBSD from grabbing the client_messenger port during
7951 // rebinding. In which case a cluster_meesneger will connect also
7953 avoid_ports
.insert(client_messenger
->get_myaddr().get_port());
7955 avoid_ports
.insert(cluster_messenger
->get_myaddr().get_port());
7956 avoid_ports
.insert(hb_back_server_messenger
->get_myaddr().get_port());
7957 avoid_ports
.insert(hb_front_server_messenger
->get_myaddr().get_port());
7959 int r
= cluster_messenger
->rebind(avoid_ports
);
7961 do_shutdown
= true; // FIXME: do_restart?
7962 network_error
= true;
7963 dout(0) << __func__
<< " marked down:"
7964 << " rebind cluster_messenger failed" << dendl
;
7967 r
= hb_back_server_messenger
->rebind(avoid_ports
);
7969 do_shutdown
= true; // FIXME: do_restart?
7970 network_error
= true;
7971 dout(0) << __func__
<< " marked down:"
7972 << " rebind hb_back_server_messenger failed" << dendl
;
7975 r
= hb_front_server_messenger
->rebind(avoid_ports
);
7977 do_shutdown
= true; // FIXME: do_restart?
7978 network_error
= true;
7979 dout(0) << __func__
<< " marked down:"
7980 << " rebind hb_front_server_messenger failed" << dendl
;
7983 hb_front_client_messenger
->mark_down_all();
7984 hb_back_client_messenger
->mark_down_all();
7986 reset_heartbeat_peers();
7991 map_lock
.put_write();
7993 check_osdmap_features(store
);
7998 if (is_active() || is_waiting_for_healthy())
7999 maybe_update_heartbeat_peers();
8002 dout(10) << " not yet active; waiting for peering wq to drain" << dendl
;
8009 if (network_error
) {
8010 Mutex::Locker
l(heartbeat_lock
);
8011 map
<int,pair
<utime_t
,entity_inst_t
>>::iterator it
=
8012 failure_pending
.begin();
8013 while (it
!= failure_pending
.end()) {
8014 dout(10) << "handle_osd_ping canceling in-flight failure report for osd."
8015 << it
->first
<< dendl
;
8016 send_still_alive(osdmap
->get_epoch(), it
->second
.second
);
8017 failure_pending
.erase(it
++);
8020 // trigger shutdown in a different thread
8021 dout(0) << __func__
<< " shutdown OSD via async signal" << dendl
;
8022 queue_async_signal(SIGINT
);
8024 else if (m
->newest_map
&& m
->newest_map
> last
) {
8025 dout(10) << " msg say newest map is " << m
->newest_map
8026 << ", requesting more" << dendl
;
8027 osdmap_subscribe(osdmap
->get_epoch()+1, false);
8029 else if (is_preboot()) {
8030 if (m
->get_source().is_mon())
8031 _preboot(m
->oldest_map
, m
->newest_map
);
8035 else if (do_restart
)
8040 void OSD::check_osdmap_features(ObjectStore
*fs
)
8042 // adjust required feature bits?
8044 // we have to be a bit careful here, because we are accessing the
8045 // Policy structures without taking any lock. in particular, only
8046 // modify integer values that can safely be read by a racing CPU.
8047 // since we are only accessing existing Policy structures a their
8048 // current memory location, and setting or clearing bits in integer
8049 // fields, and we are the only writer, this is not a problem.
8052 Messenger::Policy p
= client_messenger
->get_default_policy();
8054 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_CLIENT
, &mask
);
8055 if ((p
.features_required
& mask
) != features
) {
8056 dout(0) << "crush map has features " << features
8057 << ", adjusting msgr requires for clients" << dendl
;
8058 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8059 client_messenger
->set_default_policy(p
);
8063 Messenger::Policy p
= client_messenger
->get_policy(entity_name_t::TYPE_MON
);
8065 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_MON
, &mask
);
8066 if ((p
.features_required
& mask
) != features
) {
8067 dout(0) << "crush map has features " << features
8068 << " was " << p
.features_required
8069 << ", adjusting msgr requires for mons" << dendl
;
8070 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8071 client_messenger
->set_policy(entity_name_t::TYPE_MON
, p
);
8075 Messenger::Policy p
= cluster_messenger
->get_policy(entity_name_t::TYPE_OSD
);
8077 uint64_t features
= osdmap
->get_features(entity_name_t::TYPE_OSD
, &mask
);
8079 if ((p
.features_required
& mask
) != features
) {
8080 dout(0) << "crush map has features " << features
8081 << ", adjusting msgr requires for osds" << dendl
;
8082 p
.features_required
= (p
.features_required
& ~mask
) | features
;
8083 cluster_messenger
->set_policy(entity_name_t::TYPE_OSD
, p
);
8086 if ((features
& CEPH_FEATURE_OSD_ERASURE_CODES
) &&
8087 !superblock
.compat_features
.incompat
.contains(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
)) {
8088 dout(0) << __func__
<< " enabling on-disk ERASURE CODES compat feature" << dendl
;
8089 superblock
.compat_features
.incompat
.insert(CEPH_OSD_FEATURE_INCOMPAT_SHARDS
);
8090 ObjectStore::Transaction t
;
8091 write_superblock(t
);
8092 int err
= store
->queue_transaction(service
.meta_osr
.get(), std::move(t
), NULL
);
8098 bool OSD::advance_pg(
8099 epoch_t osd_epoch
, PG
*pg
,
8100 ThreadPool::TPHandle
&handle
,
8101 PG::RecoveryCtx
*rctx
,
8102 set
<PGRef
> *new_pgs
)
8104 assert(pg
->is_locked());
8105 epoch_t next_epoch
= pg
->get_osdmap()->get_epoch() + 1;
8106 OSDMapRef lastmap
= pg
->get_osdmap();
8108 if (lastmap
->get_epoch() == osd_epoch
)
8110 assert(lastmap
->get_epoch() < osd_epoch
);
8112 epoch_t min_epoch
= service
.get_min_pg_epoch();
8115 max
= min_epoch
+ cct
->_conf
->osd_map_max_advance
;
8117 max
= next_epoch
+ cct
->_conf
->osd_map_max_advance
;
8121 next_epoch
<= osd_epoch
&& next_epoch
<= max
;
8123 OSDMapRef nextmap
= service
.try_get_map(next_epoch
);
8125 dout(20) << __func__
<< " missing map " << next_epoch
<< dendl
;
8126 // make sure max is bumped up so that we can get past any
8128 max
= MAX(max
, next_epoch
+ cct
->_conf
->osd_map_max_advance
);
8132 vector
<int> newup
, newacting
;
8133 int up_primary
, acting_primary
;
8134 nextmap
->pg_to_up_acting_osds(
8136 &newup
, &up_primary
,
8137 &newacting
, &acting_primary
);
8138 pg
->handle_advance_map(
8139 nextmap
, lastmap
, newup
, up_primary
,
8140 newacting
, acting_primary
, rctx
);
8143 set
<spg_t
> children
;
8144 spg_t
parent(pg
->info
.pgid
);
8145 if (parent
.is_split(
8146 lastmap
->get_pg_num(pg
->pool
.id
),
8147 nextmap
->get_pg_num(pg
->pool
.id
),
8149 service
.mark_split_in_progress(pg
->info
.pgid
, children
);
8151 pg
, children
, new_pgs
, lastmap
, nextmap
,
8156 handle
.reset_tp_timeout();
8158 service
.pg_update_epoch(pg
->info
.pgid
, lastmap
->get_epoch());
8159 pg
->handle_activate_map(rctx
);
8160 if (next_epoch
<= osd_epoch
) {
8161 dout(10) << __func__
<< " advanced to max " << max
8162 << " past min epoch " << min_epoch
8163 << " ... will requeue " << *pg
<< dendl
;
8169 void OSD::consume_map()
8171 assert(osd_lock
.is_locked());
8172 dout(7) << "consume_map version " << osdmap
->get_epoch() << dendl
;
8174 int num_pg_primary
= 0, num_pg_replica
= 0, num_pg_stray
= 0;
8175 list
<PGRef
> to_remove
;
8179 RWLock::RLocker
l(pg_map_lock
);
8180 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8183 PG
*pg
= it
->second
;
8185 if (pg
->is_primary())
8187 else if (pg
->is_replica())
8192 if (!osdmap
->have_pg_pool(pg
->info
.pgid
.pool())) {
8194 to_remove
.push_back(PGRef(pg
));
8196 service
.init_splits_between(it
->first
, service
.get_osdmap(), osdmap
);
8203 for (list
<PGRef
>::iterator i
= to_remove
.begin();
8204 i
!= to_remove
.end();
8205 to_remove
.erase(i
++)) {
8206 RWLock::WLocker
locker(pg_map_lock
);
8212 service
.expand_pg_num(service
.get_osdmap(), osdmap
);
8214 service
.pre_publish_map(osdmap
);
8215 service
.await_reserved_maps();
8216 service
.publish_map(osdmap
);
8218 service
.maybe_inject_dispatch_delay();
8220 dispatch_sessions_waiting_on_map();
8222 service
.maybe_inject_dispatch_delay();
8224 // remove any PGs which we no longer host from the session waiting_for_pg lists
8225 dout(20) << __func__
<< " checking waiting_for_pg" << dendl
;
8226 op_shardedwq
.prune_pg_waiters(osdmap
, whoami
);
8228 service
.maybe_inject_dispatch_delay();
8232 RWLock::RLocker
l(pg_map_lock
);
8233 for (ceph::unordered_map
<spg_t
,PG
*>::iterator it
= pg_map
.begin();
8236 PG
*pg
= it
->second
;
8238 pg
->queue_null(osdmap
->get_epoch(), osdmap
->get_epoch());
8242 logger
->set(l_osd_pg
, pg_map
.size());
8244 logger
->set(l_osd_pg_primary
, num_pg_primary
);
8245 logger
->set(l_osd_pg_replica
, num_pg_replica
);
8246 logger
->set(l_osd_pg_stray
, num_pg_stray
);
8249 void OSD::activate_map()
8251 assert(osd_lock
.is_locked());
8253 dout(7) << "activate_map version " << osdmap
->get_epoch() << dendl
;
8255 if (!osdmap
->test_flag(CEPH_OSDMAP_SORTBITWISE
)) {
8256 derr
<< __func__
<< " SORTBITWISE flag is not set" << dendl
;
8260 if (osdmap
->test_flag(CEPH_OSDMAP_FULL
)) {
8261 dout(10) << " osdmap flagged full, doing onetime osdmap subscribe" << dendl
;
8262 osdmap_subscribe(osdmap
->get_epoch() + 1, false);
8266 if (osdmap
->test_flag(CEPH_OSDMAP_NORECOVER
)) {
8267 if (!service
.recovery_is_paused()) {
8268 dout(1) << "pausing recovery (NORECOVER flag set)" << dendl
;
8269 service
.pause_recovery();
8272 if (service
.recovery_is_paused()) {
8273 dout(1) << "unpausing recovery (NORECOVER flag unset)" << dendl
;
8274 service
.unpause_recovery();
8278 service
.activate_map();
8281 take_waiters(waiting_for_osdmap
);
8284 bool OSD::require_mon_peer(const Message
*m
)
8286 if (!m
->get_connection()->peer_is_mon()) {
8287 dout(0) << "require_mon_peer received from non-mon "
8288 << m
->get_connection()->get_peer_addr()
8289 << " " << *m
<< dendl
;
8295 bool OSD::require_mon_or_mgr_peer(const Message
*m
)
8297 if (!m
->get_connection()->peer_is_mon() &&
8298 !m
->get_connection()->peer_is_mgr()) {
8299 dout(0) << "require_mon_or_mgr_peer received from non-mon, non-mgr "
8300 << m
->get_connection()->get_peer_addr()
8301 << " " << *m
<< dendl
;
8307 bool OSD::require_osd_peer(const Message
*m
)
8309 if (!m
->get_connection()->peer_is_osd()) {
8310 dout(0) << "require_osd_peer received from non-osd "
8311 << m
->get_connection()->get_peer_addr()
8312 << " " << *m
<< dendl
;
8318 bool OSD::require_self_aliveness(const Message
*m
, epoch_t epoch
)
8320 epoch_t up_epoch
= service
.get_up_epoch();
8321 if (epoch
< up_epoch
) {
8322 dout(7) << "from pre-up epoch " << epoch
<< " < " << up_epoch
<< dendl
;
8327 dout(7) << "still in boot state, dropping message " << *m
<< dendl
;
8334 bool OSD::require_same_peer_instance(const Message
*m
, OSDMapRef
& map
,
8335 bool is_fast_dispatch
)
8337 int from
= m
->get_source().num();
8339 if (map
->is_down(from
) ||
8340 (map
->get_cluster_addr(from
) != m
->get_source_inst().addr
)) {
8341 dout(5) << "from dead osd." << from
<< ", marking down, "
8342 << " msg was " << m
->get_source_inst().addr
8343 << " expected " << (map
->is_up(from
) ?
8344 map
->get_cluster_addr(from
) : entity_addr_t())
8346 ConnectionRef con
= m
->get_connection();
8348 Session
*s
= static_cast<Session
*>(con
->get_priv());
8350 if (!is_fast_dispatch
)
8351 s
->session_dispatch_lock
.Lock();
8352 clear_session_waiting_on_map(s
);
8353 con
->set_priv(NULL
); // break ref <-> session cycle, if any
8354 if (!is_fast_dispatch
)
8355 s
->session_dispatch_lock
.Unlock();
8365 * require that we have same (or newer) map, and that
8366 * the source is the pg primary.
8368 bool OSD::require_same_or_newer_map(OpRequestRef
& op
, epoch_t epoch
,
8369 bool is_fast_dispatch
)
8371 const Message
*m
= op
->get_req();
8372 dout(15) << "require_same_or_newer_map " << epoch
8373 << " (i am " << osdmap
->get_epoch() << ") " << m
<< dendl
;
8375 assert(osd_lock
.is_locked());
8377 // do they have a newer map?
8378 if (epoch
> osdmap
->get_epoch()) {
8379 dout(7) << "waiting for newer map epoch " << epoch
8380 << " > my " << osdmap
->get_epoch() << " with " << m
<< dendl
;
8381 wait_for_new_map(op
);
8385 if (!require_self_aliveness(op
->get_req(), epoch
)) {
8389 // ok, our map is same or newer.. do they still exist?
8390 if (m
->get_connection()->get_messenger() == cluster_messenger
&&
8391 !require_same_peer_instance(op
->get_req(), osdmap
, is_fast_dispatch
)) {
8402 // ----------------------------------------
8405 void OSD::split_pgs(
8407 const set
<spg_t
> &childpgids
, set
<PGRef
> *out_pgs
,
8410 PG::RecoveryCtx
*rctx
)
8412 unsigned pg_num
= nextmap
->get_pg_num(
8414 parent
->update_snap_mapper_bits(
8415 parent
->info
.pgid
.get_split_bits(pg_num
)
8418 vector
<object_stat_sum_t
> updated_stats(childpgids
.size() + 1);
8419 parent
->info
.stats
.stats
.sum
.split(updated_stats
);
8421 vector
<object_stat_sum_t
>::iterator stat_iter
= updated_stats
.begin();
8422 for (set
<spg_t
>::const_iterator i
= childpgids
.begin();
8423 i
!= childpgids
.end();
8425 assert(stat_iter
!= updated_stats
.end());
8426 dout(10) << "Splitting " << *parent
<< " into " << *i
<< dendl
;
8427 assert(service
.splitting(*i
));
8428 PG
* child
= _make_pg(nextmap
, *i
);
8430 out_pgs
->insert(child
);
8431 rctx
->created_pgs
.insert(child
);
8433 unsigned split_bits
= i
->get_split_bits(pg_num
);
8434 dout(10) << "pg_num is " << pg_num
<< dendl
;
8435 dout(10) << "m_seed " << i
->ps() << dendl
;
8436 dout(10) << "split_bits is " << split_bits
<< dendl
;
8438 parent
->split_colls(
8448 child
->info
.stats
.stats
.sum
= *stat_iter
;
8450 child
->write_if_dirty(*(rctx
->transaction
));
8453 assert(stat_iter
!= updated_stats
.end());
8454 parent
->info
.stats
.stats
.sum
= *stat_iter
;
8455 parent
->write_if_dirty(*(rctx
->transaction
));
8461 void OSD::handle_pg_create(OpRequestRef op
)
8463 const MOSDPGCreate
*m
= static_cast<const MOSDPGCreate
*>(op
->get_req());
8464 assert(m
->get_type() == MSG_OSD_PG_CREATE
);
8466 dout(10) << "handle_pg_create " << *m
<< dendl
;
8468 if (!require_mon_peer(op
->get_req())) {
8472 if (!require_same_or_newer_map(op
, m
->epoch
, false))
8477 map
<pg_t
,utime_t
>::const_iterator ci
= m
->ctimes
.begin();
8478 for (map
<pg_t
,pg_create_t
>::const_iterator p
= m
->mkpg
.begin();
8481 assert(ci
!= m
->ctimes
.end() && ci
->first
== p
->first
);
8482 epoch_t created
= p
->second
.created
;
8483 if (p
->second
.split_bits
) // Skip split pgs
8487 if (on
.preferred() >= 0) {
8488 dout(20) << "ignoring localized pg " << on
<< dendl
;
8492 if (!osdmap
->have_pg_pool(on
.pool())) {
8493 dout(20) << "ignoring pg on deleted pool " << on
<< dendl
;
8497 dout(20) << "mkpg " << on
<< " e" << created
<< "@" << ci
->second
<< dendl
;
8499 // is it still ours?
8500 vector
<int> up
, acting
;
8501 int up_primary
= -1;
8502 int acting_primary
= -1;
8503 osdmap
->pg_to_up_acting_osds(on
, &up
, &up_primary
, &acting
, &acting_primary
);
8504 int role
= osdmap
->calc_pg_role(whoami
, acting
, acting
.size());
8506 if (acting_primary
!= whoami
) {
8507 dout(10) << "mkpg " << on
<< " not acting_primary (" << acting_primary
8508 << "), my role=" << role
<< ", skipping" << dendl
;
8513 bool mapped
= osdmap
->get_primary_shard(on
, &pgid
);
8517 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
8519 pg_history_t history
;
8520 build_initial_pg_history(pgid
, created
, ci
->second
, &history
, &pi
);
8522 // The mon won't resend unless the primary changed, so
8523 // we ignore same_interval_since. We'll pass this history
8524 // to handle_pg_peering_evt with the current epoch as the
8525 // event -- the project_pg_history check in
8526 // handle_pg_peering_evt will be a noop.
8527 if (history
.same_primary_since
> m
->epoch
) {
8528 dout(10) << __func__
<< ": got obsolete pg create on pgid "
8529 << pgid
<< " from epoch " << m
->epoch
8530 << ", primary changed in " << history
.same_primary_since
8535 if (handle_pg_peering_evt(
8539 osdmap
->get_epoch(),
8540 PG::CephPeeringEvtRef(
8541 new PG::CephPeeringEvt(
8542 osdmap
->get_epoch(),
8543 osdmap
->get_epoch(),
8546 service
.send_pg_created(pgid
.pgid
);
8549 last_pg_create_epoch
= m
->epoch
;
8551 maybe_update_heartbeat_peers();
8555 // ----------------------------------------
8556 // peering and recovery
8558 PG::RecoveryCtx
OSD::create_context()
8560 ObjectStore::Transaction
*t
= new ObjectStore::Transaction
;
8561 C_Contexts
*on_applied
= new C_Contexts(cct
);
8562 C_Contexts
*on_safe
= new C_Contexts(cct
);
8563 map
<int, map
<spg_t
,pg_query_t
> > *query_map
=
8564 new map
<int, map
<spg_t
, pg_query_t
> >;
8565 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *notify_list
=
8566 new map
<int, vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8567 map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > > *info_map
=
8568 new map
<int,vector
<pair
<pg_notify_t
, PastIntervals
> > >;
8569 PG::RecoveryCtx
rctx(query_map
, info_map
, notify_list
,
8570 on_applied
, on_safe
, t
);
8574 struct C_OpenPGs
: public Context
{
8578 C_OpenPGs(set
<PGRef
>& p
, ObjectStore
*s
, OSD
* o
) : store(s
), osd(o
) {
8581 void finish(int r
) override
{
8582 RWLock::RLocker
l(osd
->pg_map_lock
);
8583 for (auto p
: pgs
) {
8584 if (osd
->pg_map
.count(p
->info
.pgid
)) {
8585 p
->ch
= store
->open_collection(p
->coll
);
8592 void OSD::dispatch_context_transaction(PG::RecoveryCtx
&ctx
, PG
*pg
,
8593 ThreadPool::TPHandle
*handle
)
8595 if (!ctx
.transaction
->empty()) {
8596 if (!ctx
.created_pgs
.empty()) {
8597 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8599 int tr
= store
->queue_transaction(
8601 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
,
8602 TrackedOpRef(), handle
);
8603 delete (ctx
.transaction
);
8605 ctx
.transaction
= new ObjectStore::Transaction
;
8606 ctx
.on_applied
= new C_Contexts(cct
);
8607 ctx
.on_safe
= new C_Contexts(cct
);
8611 void OSD::dispatch_context(PG::RecoveryCtx
&ctx
, PG
*pg
, OSDMapRef curmap
,
8612 ThreadPool::TPHandle
*handle
)
8614 if (service
.get_osdmap()->is_up(whoami
) &&
8616 do_notifies(*ctx
.notify_list
, curmap
);
8617 do_queries(*ctx
.query_map
, curmap
);
8618 do_infos(*ctx
.info_map
, curmap
);
8620 delete ctx
.notify_list
;
8621 delete ctx
.query_map
;
8622 delete ctx
.info_map
;
8623 if ((ctx
.on_applied
->empty() &&
8624 ctx
.on_safe
->empty() &&
8625 ctx
.transaction
->empty() &&
8626 ctx
.created_pgs
.empty()) || !pg
) {
8627 delete ctx
.transaction
;
8628 delete ctx
.on_applied
;
8630 assert(ctx
.created_pgs
.empty());
8632 if (!ctx
.created_pgs
.empty()) {
8633 ctx
.on_applied
->add(new C_OpenPGs(ctx
.created_pgs
, store
, this));
8635 int tr
= store
->queue_transaction(
8637 std::move(*ctx
.transaction
), ctx
.on_applied
, ctx
.on_safe
, NULL
, TrackedOpRef(),
8639 delete (ctx
.transaction
);
8645 * Send an MOSDPGNotify to a primary, with a list of PGs that I have
8646 * content for, and they are primary for.
8649 void OSD::do_notifies(
8650 map
<int,vector
<pair
<pg_notify_t
,PastIntervals
> > >& notify_list
,
8654 vector
<pair
<pg_notify_t
,PastIntervals
> > >::iterator it
=
8655 notify_list
.begin();
8656 it
!= notify_list
.end();
8658 if (!curmap
->is_up(it
->first
)) {
8659 dout(20) << __func__
<< " skipping down osd." << it
->first
<< dendl
;
8662 ConnectionRef con
= service
.get_con_osd_cluster(
8663 it
->first
, curmap
->get_epoch());
8665 dout(20) << __func__
<< " skipping osd." << it
->first
8666 << " (NULL con)" << dendl
;
8669 service
.share_map_peer(it
->first
, con
.get(), curmap
);
8670 dout(7) << __func__
<< " osd " << it
->first
8671 << " on " << it
->second
.size() << " PGs" << dendl
;
8672 MOSDPGNotify
*m
= new MOSDPGNotify(curmap
->get_epoch(),
8674 con
->send_message(m
);
8680 * send out pending queries for info | summaries
8682 void OSD::do_queries(map
<int, map
<spg_t
,pg_query_t
> >& query_map
,
8685 for (map
<int, map
<spg_t
,pg_query_t
> >::iterator pit
= query_map
.begin();
8686 pit
!= query_map
.end();
8688 if (!curmap
->is_up(pit
->first
)) {
8689 dout(20) << __func__
<< " skipping down osd." << pit
->first
<< dendl
;
8692 int who
= pit
->first
;
8693 ConnectionRef con
= service
.get_con_osd_cluster(who
, curmap
->get_epoch());
8695 dout(20) << __func__
<< " skipping osd." << who
8696 << " (NULL con)" << dendl
;
8699 service
.share_map_peer(who
, con
.get(), curmap
);
8700 dout(7) << __func__
<< " querying osd." << who
8701 << " on " << pit
->second
.size() << " PGs" << dendl
;
8702 MOSDPGQuery
*m
= new MOSDPGQuery(curmap
->get_epoch(), pit
->second
);
8703 con
->send_message(m
);
8708 void OSD::do_infos(map
<int,
8709 vector
<pair
<pg_notify_t
, PastIntervals
> > >& info_map
,
8713 vector
<pair
<pg_notify_t
, PastIntervals
> > >::iterator p
=
8715 p
!= info_map
.end();
8717 if (!curmap
->is_up(p
->first
)) {
8718 dout(20) << __func__
<< " skipping down osd." << p
->first
<< dendl
;
8721 for (vector
<pair
<pg_notify_t
,PastIntervals
> >::iterator i
= p
->second
.begin();
8722 i
!= p
->second
.end();
8724 dout(20) << __func__
<< " sending info " << i
->first
.info
8725 << " to shard " << p
->first
<< dendl
;
8727 ConnectionRef con
= service
.get_con_osd_cluster(
8728 p
->first
, curmap
->get_epoch());
8730 dout(20) << __func__
<< " skipping osd." << p
->first
8731 << " (NULL con)" << dendl
;
8734 service
.share_map_peer(p
->first
, con
.get(), curmap
);
8735 MOSDPGInfo
*m
= new MOSDPGInfo(curmap
->get_epoch());
8736 m
->pg_list
= p
->second
;
8737 con
->send_message(m
);
8744 * from non-primary to primary
8745 * includes pg_info_t.
8746 * NOTE: called with opqueue active.
8748 void OSD::handle_pg_notify(OpRequestRef op
)
8750 const MOSDPGNotify
*m
= static_cast<const MOSDPGNotify
*>(op
->get_req());
8751 assert(m
->get_type() == MSG_OSD_PG_NOTIFY
);
8753 dout(7) << "handle_pg_notify from " << m
->get_source() << dendl
;
8754 int from
= m
->get_source().num();
8756 if (!require_osd_peer(op
->get_req()))
8759 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
8764 for (auto it
= m
->get_pg_list().begin();
8765 it
!= m
->get_pg_list().end();
8767 if (it
->first
.info
.pgid
.preferred() >= 0) {
8768 dout(20) << "ignoring localized pg " << it
->first
.info
.pgid
<< dendl
;
8772 handle_pg_peering_evt(
8773 spg_t(it
->first
.info
.pgid
.pgid
, it
->first
.to
),
8774 it
->first
.info
.history
, it
->second
,
8775 it
->first
.query_epoch
,
8776 PG::CephPeeringEvtRef(
8777 new PG::CephPeeringEvt(
8778 it
->first
.epoch_sent
, it
->first
.query_epoch
,
8779 PG::MNotifyRec(pg_shard_t(from
, it
->first
.from
), it
->first
,
8780 op
->get_req()->get_connection()->get_features())))
8785 void OSD::handle_pg_log(OpRequestRef op
)
8787 MOSDPGLog
*m
= static_cast<MOSDPGLog
*>(op
->get_nonconst_req());
8788 assert(m
->get_type() == MSG_OSD_PG_LOG
);
8789 dout(7) << "handle_pg_log " << *m
<< " from " << m
->get_source() << dendl
;
8791 if (!require_osd_peer(op
->get_req()))
8794 int from
= m
->get_source().num();
8795 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
8798 if (m
->info
.pgid
.preferred() >= 0) {
8799 dout(10) << "ignoring localized pg " << m
->info
.pgid
<< dendl
;
8804 handle_pg_peering_evt(
8805 spg_t(m
->info
.pgid
.pgid
, m
->to
),
8806 m
->info
.history
, m
->past_intervals
, m
->get_epoch(),
8807 PG::CephPeeringEvtRef(
8808 new PG::CephPeeringEvt(
8809 m
->get_epoch(), m
->get_query_epoch(),
8810 PG::MLogRec(pg_shard_t(from
, m
->from
), m
)))
8814 void OSD::handle_pg_info(OpRequestRef op
)
8816 const MOSDPGInfo
*m
= static_cast<const MOSDPGInfo
*>(op
->get_req());
8817 assert(m
->get_type() == MSG_OSD_PG_INFO
);
8818 dout(7) << "handle_pg_info " << *m
<< " from " << m
->get_source() << dendl
;
8820 if (!require_osd_peer(op
->get_req()))
8823 int from
= m
->get_source().num();
8824 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
8829 for (auto p
= m
->pg_list
.begin();
8830 p
!= m
->pg_list
.end();
8832 if (p
->first
.info
.pgid
.preferred() >= 0) {
8833 dout(10) << "ignoring localized pg " << p
->first
.info
.pgid
<< dendl
;
8837 handle_pg_peering_evt(
8838 spg_t(p
->first
.info
.pgid
.pgid
, p
->first
.to
),
8839 p
->first
.info
.history
, p
->second
, p
->first
.epoch_sent
,
8840 PG::CephPeeringEvtRef(
8841 new PG::CephPeeringEvt(
8842 p
->first
.epoch_sent
, p
->first
.query_epoch
,
8845 from
, p
->first
.from
), p
->first
.info
, p
->first
.epoch_sent
)))
8850 void OSD::handle_pg_trim(OpRequestRef op
)
8852 const MOSDPGTrim
*m
= static_cast<const MOSDPGTrim
*>(op
->get_req());
8853 assert(m
->get_type() == MSG_OSD_PG_TRIM
);
8855 dout(7) << "handle_pg_trim " << *m
<< " from " << m
->get_source() << dendl
;
8857 if (!require_osd_peer(op
->get_req()))
8860 int from
= m
->get_source().num();
8861 if (!require_same_or_newer_map(op
, m
->epoch
, false))
8864 if (m
->pgid
.preferred() >= 0) {
8865 dout(10) << "ignoring localized pg " << m
->pgid
<< dendl
;
8871 PG
*pg
= _lookup_lock_pg(m
->pgid
);
8873 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
8877 if (m
->epoch
< pg
->info
.history
.same_interval_since
) {
8878 dout(10) << *pg
<< " got old trim to " << m
->trim_to
<< ", ignoring" << dendl
;
8883 if (pg
->is_primary()) {
8884 // peer is informing us of their last_complete_ondisk
8885 dout(10) << *pg
<< " replica osd." << from
<< " lcod " << m
->trim_to
<< dendl
;
8886 pg
->peer_last_complete_ondisk
[pg_shard_t(from
, m
->pgid
.shard
)] =
8888 // trim log when the pg is recovered
8889 pg
->calc_min_last_complete_ondisk();
8891 // primary is instructing us to trim
8892 ObjectStore::Transaction t
;
8893 pg
->pg_log
.trim(m
->trim_to
, pg
->info
);
8894 pg
->dirty_info
= true;
8895 pg
->write_if_dirty(t
);
8896 int tr
= store
->queue_transaction(pg
->osr
.get(), std::move(t
), NULL
);
8902 void OSD::handle_pg_backfill_reserve(OpRequestRef op
)
8904 const MBackfillReserve
*m
= static_cast<const MBackfillReserve
*>(op
->get_req());
8905 assert(m
->get_type() == MSG_OSD_BACKFILL_RESERVE
);
8907 if (!require_osd_peer(op
->get_req()))
8909 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
8912 PG::CephPeeringEvtRef evt
;
8913 if (m
->type
== MBackfillReserve::REQUEST
) {
8914 evt
= PG::CephPeeringEvtRef(
8915 new PG::CephPeeringEvt(
8918 PG::RequestBackfillPrio(m
->priority
)));
8919 } else if (m
->type
== MBackfillReserve::GRANT
) {
8920 evt
= PG::CephPeeringEvtRef(
8921 new PG::CephPeeringEvt(
8924 PG::RemoteBackfillReserved()));
8925 } else if (m
->type
== MBackfillReserve::REJECT
) {
8926 evt
= PG::CephPeeringEvtRef(
8927 new PG::CephPeeringEvt(
8930 PG::RemoteReservationRejected()));
8935 if (service
.splitting(m
->pgid
)) {
8936 peering_wait_for_split
[m
->pgid
].push_back(evt
);
8940 PG
*pg
= _lookup_lock_pg(m
->pgid
);
8942 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
8946 pg
->queue_peering_event(evt
);
8950 void OSD::handle_pg_recovery_reserve(OpRequestRef op
)
8952 const MRecoveryReserve
*m
= static_cast<const MRecoveryReserve
*>(op
->get_req());
8953 assert(m
->get_type() == MSG_OSD_RECOVERY_RESERVE
);
8955 if (!require_osd_peer(op
->get_req()))
8957 if (!require_same_or_newer_map(op
, m
->query_epoch
, false))
8960 PG::CephPeeringEvtRef evt
;
8961 if (m
->type
== MRecoveryReserve::REQUEST
) {
8962 evt
= PG::CephPeeringEvtRef(
8963 new PG::CephPeeringEvt(
8966 PG::RequestRecovery()));
8967 } else if (m
->type
== MRecoveryReserve::GRANT
) {
8968 evt
= PG::CephPeeringEvtRef(
8969 new PG::CephPeeringEvt(
8972 PG::RemoteRecoveryReserved()));
8973 } else if (m
->type
== MRecoveryReserve::RELEASE
) {
8974 evt
= PG::CephPeeringEvtRef(
8975 new PG::CephPeeringEvt(
8978 PG::RecoveryDone()));
8983 if (service
.splitting(m
->pgid
)) {
8984 peering_wait_for_split
[m
->pgid
].push_back(evt
);
8988 PG
*pg
= _lookup_lock_pg(m
->pgid
);
8990 dout(10) << " don't have pg " << m
->pgid
<< dendl
;
8994 pg
->queue_peering_event(evt
);
8998 void OSD::handle_force_recovery(Message
*m
)
9000 MOSDForceRecovery
*msg
= static_cast<MOSDForceRecovery
*>(m
);
9001 assert(msg
->get_type() == MSG_OSD_FORCE_RECOVERY
);
9003 vector
<PGRef
> local_pgs
;
9004 local_pgs
.reserve(msg
->forced_pgs
.size());
9007 RWLock::RLocker
l(pg_map_lock
);
9008 for (auto& i
: msg
->forced_pgs
) {
9010 if (osdmap
->get_primary_shard(i
, &locpg
)) {
9011 auto pg_map_entry
= pg_map
.find(locpg
);
9012 if (pg_map_entry
!= pg_map
.end()) {
9013 local_pgs
.push_back(pg_map_entry
->second
);
9019 if (local_pgs
.size()) {
9020 service
.adjust_pg_priorities(local_pgs
, msg
->options
);
9027 * from primary to replica | stray
9028 * NOTE: called with opqueue active.
9030 void OSD::handle_pg_query(OpRequestRef op
)
9032 assert(osd_lock
.is_locked());
9034 const MOSDPGQuery
*m
= static_cast<const MOSDPGQuery
*>(op
->get_req());
9035 assert(m
->get_type() == MSG_OSD_PG_QUERY
);
9037 if (!require_osd_peer(op
->get_req()))
9040 dout(7) << "handle_pg_query from " << m
->get_source() << " epoch " << m
->get_epoch() << dendl
;
9041 int from
= m
->get_source().num();
9043 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9048 map
< int, vector
<pair
<pg_notify_t
, PastIntervals
> > > notify_list
;
9050 for (auto it
= m
->pg_list
.begin();
9051 it
!= m
->pg_list
.end();
9053 spg_t pgid
= it
->first
;
9055 if (pgid
.preferred() >= 0) {
9056 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9060 if (service
.splitting(pgid
)) {
9061 peering_wait_for_split
[pgid
].push_back(
9062 PG::CephPeeringEvtRef(
9063 new PG::CephPeeringEvt(
9064 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9065 PG::MQuery(pg_shard_t(from
, it
->second
.from
),
9066 it
->second
, it
->second
.epoch_sent
))));
9071 RWLock::RLocker
l(pg_map_lock
);
9072 if (pg_map
.count(pgid
)) {
9074 pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9076 it
->second
.epoch_sent
, it
->second
.epoch_sent
,
9077 pg_shard_t(from
, it
->second
.from
), it
->second
);
9083 if (!osdmap
->have_pg_pool(pgid
.pool()))
9086 // get active crush mapping
9087 int up_primary
, acting_primary
;
9088 vector
<int> up
, acting
;
9089 osdmap
->pg_to_up_acting_osds(
9090 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9093 pg_history_t history
= it
->second
.history
;
9094 bool valid_history
= project_pg_history(
9095 pgid
, history
, it
->second
.epoch_sent
,
9096 up
, up_primary
, acting
, acting_primary
);
9098 if (!valid_history
||
9099 it
->second
.epoch_sent
< history
.same_interval_since
) {
9100 dout(10) << " pg " << pgid
<< " dne, and pg has changed in "
9101 << history
.same_interval_since
9102 << " (msg from " << it
->second
.epoch_sent
<< ")" << dendl
;
9106 dout(10) << " pg " << pgid
<< " dne" << dendl
;
9107 pg_info_t
empty(spg_t(pgid
.pgid
, it
->second
.to
));
9108 /* This is racy, but that should be ok: if we complete the deletion
9109 * before the pg is recreated, we'll just start it off backfilling
9110 * instead of just empty */
9111 if (service
.deleting_pgs
.lookup(pgid
))
9112 empty
.set_last_backfill(hobject_t());
9113 if (it
->second
.type
== pg_query_t::LOG
||
9114 it
->second
.type
== pg_query_t::FULLLOG
) {
9115 ConnectionRef con
= service
.get_con_osd_cluster(from
, osdmap
->get_epoch());
9117 MOSDPGLog
*mlog
= new MOSDPGLog(
9118 it
->second
.from
, it
->second
.to
,
9119 osdmap
->get_epoch(), empty
,
9120 it
->second
.epoch_sent
);
9121 service
.share_map_peer(from
, con
.get(), osdmap
);
9122 con
->send_message(mlog
);
9125 notify_list
[from
].push_back(
9128 it
->second
.from
, it
->second
.to
,
9129 it
->second
.epoch_sent
,
9130 osdmap
->get_epoch(),
9133 osdmap
->get_pools().at(pgid
.pool()).ec_pool(),
9137 do_notifies(notify_list
, osdmap
);
9141 void OSD::handle_pg_remove(OpRequestRef op
)
9143 const MOSDPGRemove
*m
= static_cast<const MOSDPGRemove
*>(op
->get_req());
9144 assert(m
->get_type() == MSG_OSD_PG_REMOVE
);
9145 assert(osd_lock
.is_locked());
9147 if (!require_osd_peer(op
->get_req()))
9150 dout(7) << "handle_pg_remove from " << m
->get_source() << " on "
9151 << m
->pg_list
.size() << " pgs" << dendl
;
9153 if (!require_same_or_newer_map(op
, m
->get_epoch(), false))
9158 for (auto it
= m
->pg_list
.begin();
9159 it
!= m
->pg_list
.end();
9162 if (pgid
.preferred() >= 0) {
9163 dout(10) << "ignoring localized pg " << pgid
<< dendl
;
9167 RWLock::WLocker
l(pg_map_lock
);
9168 if (pg_map
.count(pgid
) == 0) {
9169 dout(10) << " don't have pg " << pgid
<< dendl
;
9172 dout(5) << "queue_pg_for_deletion: " << pgid
<< dendl
;
9173 PG
*pg
= _lookup_lock_pg_with_map_lock_held(pgid
);
9174 pg_history_t history
= pg
->info
.history
;
9175 int up_primary
, acting_primary
;
9176 vector
<int> up
, acting
;
9177 osdmap
->pg_to_up_acting_osds(
9178 pgid
.pgid
, &up
, &up_primary
, &acting
, &acting_primary
);
9179 bool valid_history
= project_pg_history(
9180 pg
->info
.pgid
, history
, pg
->get_osdmap()->get_epoch(),
9181 up
, up_primary
, acting
, acting_primary
);
9182 if (valid_history
&&
9183 history
.same_interval_since
<= m
->get_epoch()) {
9184 assert(pg
->get_primary().osd
== m
->get_source().num());
9189 dout(10) << *pg
<< " ignoring remove request, pg changed in epoch "
9190 << history
.same_interval_since
9191 << " > " << m
->get_epoch() << dendl
;
9197 void OSD::_remove_pg(PG
*pg
)
9199 ObjectStore::Transaction rmt
;
9201 // on_removal, which calls remove_watchers_and_notifies, and the erasure from
9202 // the pg_map must be done together without unlocking the pg lock,
9203 // to avoid racing with watcher cleanup in ms_handle_reset
9204 // and handle_notify_timeout
9205 pg
->on_removal(&rmt
);
9207 service
.cancel_pending_splits_for_parent(pg
->info
.pgid
);
9208 int tr
= store
->queue_transaction(
9209 pg
->osr
.get(), std::move(rmt
), NULL
,
9210 new ContainerContext
<
9211 SequencerRef
>(pg
->osr
));
9214 DeletingStateRef deleting
= service
.deleting_pgs
.lookup_or_create(
9220 remove_wq
.queue(make_pair(PGRef(pg
), deleting
));
9222 service
.pg_remove_epoch(pg
->info
.pgid
);
9224 // dereference from op_wq
9225 op_shardedwq
.clear_pg_pointer(pg
->info
.pgid
);
9228 pg_map
.erase(pg
->info
.pgid
);
9229 pg
->put("PGMap"); // since we've taken it out of map
9233 // =========================================================
9236 void OSDService::_maybe_queue_recovery() {
9237 assert(recovery_lock
.is_locked_by_me());
9238 uint64_t available_pushes
;
9239 while (!awaiting_throttle
.empty() &&
9240 _recover_now(&available_pushes
)) {
9241 uint64_t to_start
= MIN(
9243 cct
->_conf
->osd_recovery_max_single_start
);
9244 _queue_for_recovery(awaiting_throttle
.front(), to_start
);
9245 awaiting_throttle
.pop_front();
9246 recovery_ops_reserved
+= to_start
;
9250 bool OSDService::_recover_now(uint64_t *available_pushes
)
9252 if (available_pushes
)
9253 *available_pushes
= 0;
9255 if (ceph_clock_now() < defer_recovery_until
) {
9256 dout(15) << __func__
<< " defer until " << defer_recovery_until
<< dendl
;
9260 if (recovery_paused
) {
9261 dout(15) << __func__
<< " paused" << dendl
;
9265 uint64_t max
= cct
->_conf
->osd_recovery_max_active
;
9266 if (max
<= recovery_ops_active
+ recovery_ops_reserved
) {
9267 dout(15) << __func__
<< " active " << recovery_ops_active
9268 << " + reserved " << recovery_ops_reserved
9269 << " >= max " << max
<< dendl
;
9273 if (available_pushes
)
9274 *available_pushes
= max
- recovery_ops_active
- recovery_ops_reserved
;
9280 void OSDService::adjust_pg_priorities(const vector
<PGRef
>& pgs
, int newflags
)
9282 if (!pgs
.size() || !(newflags
& (OFR_BACKFILL
| OFR_RECOVERY
)))
9286 if (newflags
& OFR_BACKFILL
) {
9287 newstate
= PG_STATE_FORCED_BACKFILL
;
9288 } else if (newflags
& OFR_RECOVERY
) {
9289 newstate
= PG_STATE_FORCED_RECOVERY
;
9292 // debug output here may get large, don't generate it if debug level is below
9293 // 10 and use abbreviated pg ids otherwise
9294 if ((cct
)->_conf
->subsys
.should_gather(ceph_subsys_osd
, 10)) {
9297 for (auto& i
: pgs
) {
9298 ss
<< i
->get_pgid() << " ";
9301 dout(10) << __func__
<< " working on " << ss
.str() << dendl
;
9304 if (newflags
& OFR_CANCEL
) {
9305 for (auto& i
: pgs
) {
9307 i
->_change_recovery_force_mode(newstate
, true);
9311 for (auto& i
: pgs
) {
9312 // make sure the PG is in correct state before forcing backfill or recovery, or
9313 // else we'll make PG keeping FORCE_* flag forever, requiring osds restart
9314 // or forcing somehow recovery/backfill.
9316 int pgstate
= i
->get_state();
9317 if ( ((newstate
== PG_STATE_FORCED_RECOVERY
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_RECOVERY_WAIT
| PG_STATE_RECOVERING
))) ||
9318 ((newstate
== PG_STATE_FORCED_BACKFILL
) && (pgstate
& (PG_STATE_DEGRADED
| PG_STATE_BACKFILL_WAIT
| PG_STATE_BACKFILL
))) )
9319 i
->_change_recovery_force_mode(newstate
, false);
9325 void OSD::do_recovery(
9326 PG
*pg
, epoch_t queued
, uint64_t reserved_pushes
,
9327 ThreadPool::TPHandle
&handle
)
9329 uint64_t started
= 0;
9332 * When the value of osd_recovery_sleep is set greater than zero, recovery
9333 * ops are scheduled after osd_recovery_sleep amount of time from the previous
9334 * recovery event's schedule time. This is done by adding a
9335 * recovery_requeue_callback event, which re-queues the recovery op using
9336 * queue_recovery_after_sleep.
9338 float recovery_sleep
= get_osd_recovery_sleep();
9339 if (recovery_sleep
> 0 && service
.recovery_needs_sleep
) {
9341 auto recovery_requeue_callback
= new FunctionContext([this, pgref
, queued
, reserved_pushes
](int r
) {
9342 dout(20) << "do_recovery wake up at "
9344 << ", re-queuing recovery" << dendl
;
9345 service
.recovery_needs_sleep
= false;
9346 service
.queue_recovery_after_sleep(pgref
.get(), queued
, reserved_pushes
);
9348 Mutex::Locker
l(service
.recovery_sleep_lock
);
9350 // This is true for the first recovery op and when the previous recovery op
9351 // has been scheduled in the past. The next recovery op is scheduled after
9352 // completing the sleep from now.
9353 if (service
.recovery_schedule_time
< ceph_clock_now()) {
9354 service
.recovery_schedule_time
= ceph_clock_now();
9356 service
.recovery_schedule_time
+= recovery_sleep
;
9357 service
.recovery_sleep_timer
.add_event_at(service
.recovery_schedule_time
,
9358 recovery_requeue_callback
);
9359 dout(20) << "Recovery event scheduled at "
9360 << service
.recovery_schedule_time
<< dendl
;
9365 service
.recovery_needs_sleep
= true;
9366 if (pg
->pg_has_reset_since(queued
)) {
9370 assert(!pg
->deleting
);
9371 assert(pg
->is_peered() && pg
->is_primary());
9373 assert(pg
->recovery_queued
);
9374 pg
->recovery_queued
= false;
9376 dout(10) << "do_recovery starting " << reserved_pushes
<< " " << *pg
<< dendl
;
9377 #ifdef DEBUG_RECOVERY_OIDS
9378 dout(20) << " active was " << service
.recovery_oids
[pg
->info
.pgid
] << dendl
;
9381 bool more
= pg
->start_recovery_ops(reserved_pushes
, handle
, &started
);
9382 dout(10) << "do_recovery started " << started
<< "/" << reserved_pushes
9383 << " on " << *pg
<< dendl
;
9385 // If no recovery op is started, don't bother to manipulate the RecoveryCtx
9386 if (!started
&& (more
|| !pg
->have_unfound())) {
9390 PG::RecoveryCtx rctx
= create_context();
9391 rctx
.handle
= &handle
;
9394 * if we couldn't start any recovery ops and things are still
9395 * unfound, see if we can discover more missing object locations.
9396 * It may be that our initial locations were bad and we errored
9397 * out while trying to pull.
9399 if (!more
&& pg
->have_unfound()) {
9400 pg
->discover_all_missing(*rctx
.query_map
);
9401 if (rctx
.query_map
->empty()) {
9403 if (pg
->state_test(PG_STATE_BACKFILL
)) {
9404 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9407 PG::CancelBackfill()));
9408 pg
->queue_peering_event(evt
);
9409 action
= "in backfill";
9410 } else if (pg
->state_test(PG_STATE_RECOVERING
)) {
9411 auto evt
= PG::CephPeeringEvtRef(new PG::CephPeeringEvt(
9414 PG::CancelRecovery()));
9415 pg
->queue_peering_event(evt
);
9416 action
= "in recovery";
9418 action
= "already out of recovery/backfill";
9420 dout(10) << __func__
<< ": no luck, giving up on this pg for now (" << action
<< ")" << dendl
;
9422 dout(10) << __func__
<< ": no luck, giving up on this pg for now (queue_recovery)" << dendl
;
9423 pg
->queue_recovery();
9427 pg
->write_if_dirty(*rctx
.transaction
);
9428 OSDMapRef curmap
= pg
->get_osdmap();
9429 dispatch_context(rctx
, pg
, curmap
);
9433 assert(started
<= reserved_pushes
);
9434 service
.release_reserved_pushes(reserved_pushes
);
9437 void OSDService::start_recovery_op(PG
*pg
, const hobject_t
& soid
)
9439 Mutex::Locker
l(recovery_lock
);
9440 dout(10) << "start_recovery_op " << *pg
<< " " << soid
9441 << " (" << recovery_ops_active
<< "/"
9442 << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9444 recovery_ops_active
++;
9446 #ifdef DEBUG_RECOVERY_OIDS
9447 dout(20) << " active was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9448 assert(recovery_oids
[pg
->info
.pgid
].count(soid
) == 0);
9449 recovery_oids
[pg
->info
.pgid
].insert(soid
);
9453 void OSDService::finish_recovery_op(PG
*pg
, const hobject_t
& soid
, bool dequeue
)
9455 Mutex::Locker
l(recovery_lock
);
9456 dout(10) << "finish_recovery_op " << *pg
<< " " << soid
9457 << " dequeue=" << dequeue
9458 << " (" << recovery_ops_active
<< "/" << cct
->_conf
->osd_recovery_max_active
<< " rops)"
9462 assert(recovery_ops_active
> 0);
9463 recovery_ops_active
--;
9465 #ifdef DEBUG_RECOVERY_OIDS
9466 dout(20) << " active oids was " << recovery_oids
[pg
->info
.pgid
] << dendl
;
9467 assert(recovery_oids
[pg
->info
.pgid
].count(soid
));
9468 recovery_oids
[pg
->info
.pgid
].erase(soid
);
9471 _maybe_queue_recovery();
9474 bool OSDService::is_recovery_active()
9476 return local_reserver
.has_reservation() || remote_reserver
.has_reservation();
9479 // =========================================================
9482 bool OSD::op_is_discardable(const MOSDOp
*op
)
9484 // drop client request if they are not connected and can't get the
9486 if (!op
->get_connection()->is_connected()) {
9492 void OSD::enqueue_op(spg_t pg
, OpRequestRef
& op
, epoch_t epoch
)
9494 utime_t latency
= ceph_clock_now() - op
->get_req()->get_recv_stamp();
9495 dout(15) << "enqueue_op " << op
<< " prio " << op
->get_req()->get_priority()
9496 << " cost " << op
->get_req()->get_cost()
9497 << " latency " << latency
9498 << " epoch " << epoch
9499 << " " << *(op
->get_req()) << dendl
;
9500 op
->osd_trace
.event("enqueue op");
9501 op
->osd_trace
.keyval("priority", op
->get_req()->get_priority());
9502 op
->osd_trace
.keyval("cost", op
->get_req()->get_cost());
9503 op
->mark_queued_for_pg();
9504 logger
->tinc(l_osd_op_before_queue_op_lat
, latency
);
9505 op_shardedwq
.queue(make_pair(pg
, PGQueueable(op
, epoch
)));
9511 * NOTE: dequeue called in worker thread, with pg lock
9513 void OSD::dequeue_op(
9514 PGRef pg
, OpRequestRef op
,
9515 ThreadPool::TPHandle
&handle
)
9518 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_BEGIN", false);
9520 utime_t now
= ceph_clock_now();
9521 op
->set_dequeued_time(now
);
9522 utime_t latency
= now
- op
->get_req()->get_recv_stamp();
9523 dout(10) << "dequeue_op " << op
<< " prio " << op
->get_req()->get_priority()
9524 << " cost " << op
->get_req()->get_cost()
9525 << " latency " << latency
9526 << " " << *(op
->get_req())
9527 << " pg " << *pg
<< dendl
;
9529 logger
->tinc(l_osd_op_before_dequeue_op_lat
, latency
);
9531 Session
*session
= static_cast<Session
*>(
9532 op
->get_req()->get_connection()->get_priv());
9534 maybe_share_map(session
, op
, pg
->get_osdmap());
9541 op
->mark_reached_pg();
9542 op
->osd_trace
.event("dequeue_op");
9544 pg
->do_request(op
, handle
);
9547 dout(10) << "dequeue_op " << op
<< " finish" << dendl
;
9548 OID_EVENT_TRACE_WITH_MSG(op
->get_req(), "DEQUEUE_OP_END", false);
9552 struct C_CompleteSplits
: public Context
{
9555 C_CompleteSplits(OSD
*osd
, const set
<PGRef
> &in
)
9556 : osd(osd
), pgs(in
) {}
9557 void finish(int r
) override
{
9558 Mutex::Locker
l(osd
->osd_lock
);
9559 if (osd
->is_stopping())
9561 PG::RecoveryCtx rctx
= osd
->create_context();
9562 for (set
<PGRef
>::iterator i
= pgs
.begin();
9565 osd
->pg_map_lock
.get_write();
9568 osd
->add_newly_split_pg(pg
, &rctx
);
9569 if (!((*i
)->deleting
)) {
9570 set
<spg_t
> to_complete
;
9571 to_complete
.insert((*i
)->info
.pgid
);
9572 osd
->service
.complete_split(to_complete
);
9574 osd
->pg_map_lock
.put_write();
9575 osd
->dispatch_context_transaction(rctx
, pg
);
9576 osd
->wake_pg_waiters(*i
);
9580 osd
->dispatch_context(rctx
, 0, osd
->service
.get_osdmap());
9584 void OSD::process_peering_events(
9585 const list
<PG
*> &pgs
,
9586 ThreadPool::TPHandle
&handle
9589 bool need_up_thru
= false;
9590 epoch_t same_interval_since
= 0;
9592 PG::RecoveryCtx rctx
= create_context();
9593 rctx
.handle
= &handle
;
9594 for (list
<PG
*>::const_iterator i
= pgs
.begin();
9597 set
<PGRef
> split_pgs
;
9599 pg
->lock_suspend_timeout(handle
);
9600 curmap
= service
.get_osdmap();
9605 if (!advance_pg(curmap
->get_epoch(), pg
, handle
, &rctx
, &split_pgs
)) {
9606 // we need to requeue the PG explicitly since we didn't actually
9608 peering_wq
.queue(pg
);
9610 assert(!pg
->peering_queue
.empty());
9611 PG::CephPeeringEvtRef evt
= pg
->peering_queue
.front();
9612 pg
->peering_queue
.pop_front();
9613 pg
->handle_peering_event(evt
, &rctx
);
9615 need_up_thru
= pg
->need_up_thru
|| need_up_thru
;
9616 same_interval_since
= MAX(pg
->info
.history
.same_interval_since
,
9617 same_interval_since
);
9618 pg
->write_if_dirty(*rctx
.transaction
);
9619 if (!split_pgs
.empty()) {
9620 rctx
.on_applied
->add(new C_CompleteSplits(this, split_pgs
));
9623 dispatch_context_transaction(rctx
, pg
, &handle
);
9627 queue_want_up_thru(same_interval_since
);
9628 dispatch_context(rctx
, 0, curmap
, &handle
);
9630 service
.send_pg_temp();
9633 // --------------------------------
9635 const char** OSD::get_tracked_conf_keys() const
9637 static const char* KEYS
[] = {
9638 "osd_max_backfills",
9639 "osd_min_recovery_priority",
9640 "osd_max_trimming_pgs",
9641 "osd_op_complaint_time",
9642 "osd_op_log_threshold",
9643 "osd_op_history_size",
9644 "osd_op_history_duration",
9645 "osd_op_history_slow_op_size",
9646 "osd_op_history_slow_op_threshold",
9647 "osd_enable_op_tracker",
9648 "osd_map_cache_size",
9649 "osd_map_max_advance",
9650 "osd_pg_epoch_persisted_max_stale",
9651 "osd_disk_thread_ioprio_class",
9652 "osd_disk_thread_ioprio_priority",
9653 // clog & admin clog
9656 "clog_to_syslog_facility",
9657 "clog_to_syslog_level",
9658 "osd_objectstore_fuse",
9660 "clog_to_graylog_host",
9661 "clog_to_graylog_port",
9664 "osd_recovery_delay_start",
9665 "osd_client_message_size_cap",
9666 "osd_client_message_cap",
9667 "osd_heartbeat_min_size",
9668 "osd_heartbeat_interval",
9674 void OSD::handle_conf_change(const struct md_config_t
*conf
,
9675 const std::set
<std::string
> &changed
)
9677 if (changed
.count("osd_max_backfills")) {
9678 service
.local_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9679 service
.remote_reserver
.set_max(cct
->_conf
->osd_max_backfills
);
9681 if (changed
.count("osd_min_recovery_priority")) {
9682 service
.local_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9683 service
.remote_reserver
.set_min_priority(cct
->_conf
->osd_min_recovery_priority
);
9685 if (changed
.count("osd_max_trimming_pgs")) {
9686 service
.snap_reserver
.set_max(cct
->_conf
->osd_max_trimming_pgs
);
9688 if (changed
.count("osd_op_complaint_time") ||
9689 changed
.count("osd_op_log_threshold")) {
9690 op_tracker
.set_complaint_and_threshold(cct
->_conf
->osd_op_complaint_time
,
9691 cct
->_conf
->osd_op_log_threshold
);
9693 if (changed
.count("osd_op_history_size") ||
9694 changed
.count("osd_op_history_duration")) {
9695 op_tracker
.set_history_size_and_duration(cct
->_conf
->osd_op_history_size
,
9696 cct
->_conf
->osd_op_history_duration
);
9698 if (changed
.count("osd_op_history_slow_op_size") ||
9699 changed
.count("osd_op_history_slow_op_threshold")) {
9700 op_tracker
.set_history_slow_op_size_and_threshold(cct
->_conf
->osd_op_history_slow_op_size
,
9701 cct
->_conf
->osd_op_history_slow_op_threshold
);
9703 if (changed
.count("osd_enable_op_tracker")) {
9704 op_tracker
.set_tracking(cct
->_conf
->osd_enable_op_tracker
);
9706 if (changed
.count("osd_disk_thread_ioprio_class") ||
9707 changed
.count("osd_disk_thread_ioprio_priority")) {
9708 set_disk_tp_priority();
9710 if (changed
.count("osd_map_cache_size")) {
9711 service
.map_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9712 service
.map_bl_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9713 service
.map_bl_inc_cache
.set_size(cct
->_conf
->osd_map_cache_size
);
9715 if (changed
.count("clog_to_monitors") ||
9716 changed
.count("clog_to_syslog") ||
9717 changed
.count("clog_to_syslog_level") ||
9718 changed
.count("clog_to_syslog_facility") ||
9719 changed
.count("clog_to_graylog") ||
9720 changed
.count("clog_to_graylog_host") ||
9721 changed
.count("clog_to_graylog_port") ||
9722 changed
.count("host") ||
9723 changed
.count("fsid")) {
9724 update_log_config();
9728 if (changed
.count("osd_objectstore_fuse")) {
9730 enable_disable_fuse(false);
9735 if (changed
.count("osd_recovery_delay_start")) {
9736 service
.defer_recovery(cct
->_conf
->osd_recovery_delay_start
);
9737 service
.kick_recovery_queue();
9740 if (changed
.count("osd_client_message_cap")) {
9741 uint64_t newval
= cct
->_conf
->osd_client_message_cap
;
9742 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
9743 if (pol
.throttler_messages
&& newval
> 0) {
9744 pol
.throttler_messages
->reset_max(newval
);
9747 if (changed
.count("osd_client_message_size_cap")) {
9748 uint64_t newval
= cct
->_conf
->osd_client_message_size_cap
;
9749 Messenger::Policy pol
= client_messenger
->get_policy(entity_name_t::TYPE_CLIENT
);
9750 if (pol
.throttler_bytes
&& newval
> 0) {
9751 pol
.throttler_bytes
->reset_max(newval
);
9758 void OSD::update_log_config()
9760 map
<string
,string
> log_to_monitors
;
9761 map
<string
,string
> log_to_syslog
;
9762 map
<string
,string
> log_channel
;
9763 map
<string
,string
> log_prio
;
9764 map
<string
,string
> log_to_graylog
;
9765 map
<string
,string
> log_to_graylog_host
;
9766 map
<string
,string
> log_to_graylog_port
;
9770 if (parse_log_client_options(cct
, log_to_monitors
, log_to_syslog
,
9771 log_channel
, log_prio
, log_to_graylog
,
9772 log_to_graylog_host
, log_to_graylog_port
,
9774 clog
->update_config(log_to_monitors
, log_to_syslog
,
9775 log_channel
, log_prio
, log_to_graylog
,
9776 log_to_graylog_host
, log_to_graylog_port
,
9778 derr
<< "log_to_monitors " << log_to_monitors
<< dendl
;
9781 void OSD::check_config()
9783 // some sanity checks
9784 if (cct
->_conf
->osd_map_cache_size
<= cct
->_conf
->osd_map_max_advance
+ 2) {
9785 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
9786 << " is not > osd_map_max_advance ("
9787 << cct
->_conf
->osd_map_max_advance
<< ")";
9789 if (cct
->_conf
->osd_map_cache_size
<= (int)cct
->_conf
->osd_pg_epoch_persisted_max_stale
+ 2) {
9790 clog
->warn() << "osd_map_cache_size (" << cct
->_conf
->osd_map_cache_size
<< ")"
9791 << " is not > osd_pg_epoch_persisted_max_stale ("
9792 << cct
->_conf
->osd_pg_epoch_persisted_max_stale
<< ")";
9796 void OSD::set_disk_tp_priority()
9798 dout(10) << __func__
9799 << " class " << cct
->_conf
->osd_disk_thread_ioprio_class
9800 << " priority " << cct
->_conf
->osd_disk_thread_ioprio_priority
9802 if (cct
->_conf
->osd_disk_thread_ioprio_class
.empty() ||
9803 cct
->_conf
->osd_disk_thread_ioprio_priority
< 0)
9806 ceph_ioprio_string_to_class(cct
->_conf
->osd_disk_thread_ioprio_class
);
9808 derr
<< __func__
<< cpp_strerror(cls
) << ": "
9809 << "osd_disk_thread_ioprio_class is " << cct
->_conf
->osd_disk_thread_ioprio_class
9810 << " but only the following values are allowed: idle, be or rt" << dendl
;
9812 disk_tp
.set_ioprio(cls
, cct
->_conf
->osd_disk_thread_ioprio_priority
);
9815 // --------------------------------
9817 void OSD::get_latest_osdmap()
9819 dout(10) << __func__
<< " -- start" << dendl
;
9822 service
.objecter
->wait_for_latest_osdmap(&cond
);
9825 dout(10) << __func__
<< " -- finish" << dendl
;
9828 // --------------------------------
9830 int OSD::init_op_flags(OpRequestRef
& op
)
9832 const MOSDOp
*m
= static_cast<const MOSDOp
*>(op
->get_req());
9833 vector
<OSDOp
>::const_iterator iter
;
9835 // client flags have no bearing on whether an op is a read, write, etc.
9838 if (m
->has_flag(CEPH_OSD_FLAG_RWORDERED
)) {
9839 op
->set_force_rwordered();
9842 // set bits based on op codes, called methods.
9843 for (iter
= m
->ops
.begin(); iter
!= m
->ops
.end(); ++iter
) {
9844 if ((iter
->op
.op
== CEPH_OSD_OP_WATCH
&&
9845 iter
->op
.watch
.op
== CEPH_OSD_WATCH_OP_PING
)) {
9846 /* This a bit odd. PING isn't actually a write. It can't
9847 * result in an update to the object_info. PINGs also aren'ty
9848 * resent, so there's no reason to write out a log entry
9850 * However, we pipeline them behind writes, so let's force
9851 * the write_ordered flag.
9853 op
->set_force_rwordered();
9855 if (ceph_osd_op_mode_modify(iter
->op
.op
))
9858 if (ceph_osd_op_mode_read(iter
->op
.op
))
9861 // set READ flag if there are src_oids
9862 if (iter
->soid
.oid
.name
.length())
9865 // set PGOP flag if there are PG ops
9866 if (ceph_osd_op_type_pg(iter
->op
.op
))
9869 if (ceph_osd_op_mode_cache(iter
->op
.op
))
9872 // check for ec base pool
9873 int64_t poolid
= m
->get_pg().pool();
9874 const pg_pool_t
*pool
= osdmap
->get_pg_pool(poolid
);
9875 if (pool
&& pool
->is_tier()) {
9876 const pg_pool_t
*base_pool
= osdmap
->get_pg_pool(pool
->tier_of
);
9877 if (base_pool
&& base_pool
->require_rollback()) {
9878 if ((iter
->op
.op
!= CEPH_OSD_OP_READ
) &&
9879 (iter
->op
.op
!= CEPH_OSD_OP_CHECKSUM
) &&
9880 (iter
->op
.op
!= CEPH_OSD_OP_CMPEXT
) &&
9881 (iter
->op
.op
!= CEPH_OSD_OP_STAT
) &&
9882 (iter
->op
.op
!= CEPH_OSD_OP_ISDIRTY
) &&
9883 (iter
->op
.op
!= CEPH_OSD_OP_UNDIRTY
) &&
9884 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTR
) &&
9885 (iter
->op
.op
!= CEPH_OSD_OP_GETXATTRS
) &&
9886 (iter
->op
.op
!= CEPH_OSD_OP_CMPXATTR
) &&
9887 (iter
->op
.op
!= CEPH_OSD_OP_ASSERT_VER
) &&
9888 (iter
->op
.op
!= CEPH_OSD_OP_LIST_WATCHERS
) &&
9889 (iter
->op
.op
!= CEPH_OSD_OP_LIST_SNAPS
) &&
9890 (iter
->op
.op
!= CEPH_OSD_OP_SETALLOCHINT
) &&
9891 (iter
->op
.op
!= CEPH_OSD_OP_WRITEFULL
) &&
9892 (iter
->op
.op
!= CEPH_OSD_OP_ROLLBACK
) &&
9893 (iter
->op
.op
!= CEPH_OSD_OP_CREATE
) &&
9894 (iter
->op
.op
!= CEPH_OSD_OP_DELETE
) &&
9895 (iter
->op
.op
!= CEPH_OSD_OP_SETXATTR
) &&
9896 (iter
->op
.op
!= CEPH_OSD_OP_RMXATTR
) &&
9897 (iter
->op
.op
!= CEPH_OSD_OP_STARTSYNC
) &&
9898 (iter
->op
.op
!= CEPH_OSD_OP_COPY_GET
) &&
9899 (iter
->op
.op
!= CEPH_OSD_OP_COPY_FROM
)) {
9905 switch (iter
->op
.op
) {
9906 case CEPH_OSD_OP_CALL
:
9908 bufferlist::iterator bp
= const_cast<bufferlist
&>(iter
->indata
).begin();
9909 int is_write
, is_read
;
9910 string cname
, mname
;
9911 bp
.copy(iter
->op
.cls
.class_len
, cname
);
9912 bp
.copy(iter
->op
.cls
.method_len
, mname
);
9914 ClassHandler::ClassData
*cls
;
9915 int r
= class_handler
->open_class(cname
, &cls
);
9917 derr
<< "class " << cname
<< " open got " << cpp_strerror(r
) << dendl
;
9920 else if (r
!= -EPERM
) // propagate permission errors
9924 int flags
= cls
->get_method_flags(mname
.c_str());
9926 if (flags
== -ENOENT
)
9932 is_read
= flags
& CLS_METHOD_RD
;
9933 is_write
= flags
& CLS_METHOD_WR
;
9934 bool is_promote
= flags
& CLS_METHOD_PROMOTE
;
9936 dout(10) << "class " << cname
<< " method " << mname
<< " "
9937 << "flags=" << (is_read
? "r" : "")
9938 << (is_write
? "w" : "")
9939 << (is_promote
? "p" : "")
9942 op
->set_class_read();
9944 op
->set_class_write();
9947 op
->add_class(cname
, is_read
, is_write
, cls
->whitelisted
);
9951 case CEPH_OSD_OP_WATCH
:
9952 // force the read bit for watch since it is depends on previous
9953 // watch state (and may return early if the watch exists) or, in
9954 // the case of ping, is simply a read op.
9957 case CEPH_OSD_OP_NOTIFY
:
9958 case CEPH_OSD_OP_NOTIFY_ACK
:
9964 case CEPH_OSD_OP_DELETE
:
9965 // if we get a delete with FAILOK we can skip handle cache. without
9966 // FAILOK we still need to promote (or do something smarter) to
9967 // determine whether to return ENOENT or 0.
9968 if (iter
== m
->ops
.begin() &&
9969 iter
->op
.flags
== CEPH_OSD_OP_FLAG_FAILOK
) {
9970 op
->set_skip_handle_cache();
9972 // skip promotion when proxying a delete op
9973 if (m
->ops
.size() == 1) {
9974 op
->set_skip_promote();
9978 case CEPH_OSD_OP_CACHE_TRY_FLUSH
:
9979 case CEPH_OSD_OP_CACHE_FLUSH
:
9980 case CEPH_OSD_OP_CACHE_EVICT
:
9981 // If try_flush/flush/evict is the only op, can skip handle cache.
9982 if (m
->ops
.size() == 1) {
9983 op
->set_skip_handle_cache();
9987 case CEPH_OSD_OP_READ
:
9988 case CEPH_OSD_OP_SYNC_READ
:
9989 case CEPH_OSD_OP_SPARSE_READ
:
9990 case CEPH_OSD_OP_CHECKSUM
:
9991 case CEPH_OSD_OP_WRITEFULL
:
9992 if (m
->ops
.size() == 1 &&
9993 (iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
||
9994 iter
->op
.flags
& CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
)) {
9995 op
->set_skip_promote();
9999 // force promotion when pin an object in cache tier
10000 case CEPH_OSD_OP_CACHE_PIN
:
10009 if (op
->rmw_flags
== 0)
10015 void OSD::PeeringWQ::_dequeue(list
<PG
*> *out
) {
10016 for (list
<PG
*>::iterator i
= peering_queue
.begin();
10017 i
!= peering_queue
.end() &&
10018 out
->size() < osd
->cct
->_conf
->osd_peering_wq_batch_size
;
10020 if (in_use
.count(*i
)) {
10023 out
->push_back(*i
);
10024 peering_queue
.erase(i
++);
10027 in_use
.insert(out
->begin(), out
->end());
10031 // =============================================================
10033 #undef dout_context
10034 #define dout_context osd->cct
10036 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq "
10038 void OSD::ShardedOpWQ::wake_pg_waiters(spg_t pgid
)
10040 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10041 auto sdata
= shard_list
[shard_index
];
10042 bool queued
= false;
10043 unsigned pushes_to_free
= 0;
10045 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10046 auto p
= sdata
->pg_slots
.find(pgid
);
10047 if (p
!= sdata
->pg_slots
.end()) {
10048 dout(20) << __func__
<< " " << pgid
10049 << " to_process " << p
->second
.to_process
10050 << " waiting_for_pg=" << (int)p
->second
.waiting_for_pg
<< dendl
;
10051 for (auto i
= p
->second
.to_process
.rbegin();
10052 i
!= p
->second
.to_process
.rend();
10054 sdata
->_enqueue_front(make_pair(pgid
, *i
), osd
->op_prio_cutoff
);
10056 for (auto& q
: p
->second
.to_process
) {
10057 pushes_to_free
+= q
.get_reserved_pushes();
10059 p
->second
.to_process
.clear();
10060 p
->second
.waiting_for_pg
= false;
10061 ++p
->second
.requeue_seq
;
10065 if (pushes_to_free
> 0) {
10066 osd
->service
.release_reserved_pushes(pushes_to_free
);
10069 sdata
->sdata_lock
.Lock();
10070 sdata
->sdata_cond
.SignalOne();
10071 sdata
->sdata_lock
.Unlock();
10075 void OSD::ShardedOpWQ::prune_pg_waiters(OSDMapRef osdmap
, int whoami
)
10077 unsigned pushes_to_free
= 0;
10078 for (auto sdata
: shard_list
) {
10079 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10080 sdata
->waiting_for_pg_osdmap
= osdmap
;
10081 auto p
= sdata
->pg_slots
.begin();
10082 while (p
!= sdata
->pg_slots
.end()) {
10083 ShardData::pg_slot
& slot
= p
->second
;
10084 if (!slot
.to_process
.empty() && slot
.num_running
== 0) {
10085 if (osdmap
->is_up_acting_osd_shard(p
->first
, whoami
)) {
10086 dout(20) << __func__
<< " " << p
->first
<< " maps to us, keeping"
10091 while (!slot
.to_process
.empty() &&
10092 slot
.to_process
.front().get_map_epoch() <= osdmap
->get_epoch()) {
10093 auto& qi
= slot
.to_process
.front();
10094 dout(20) << __func__
<< " " << p
->first
10096 << " epoch " << qi
.get_map_epoch()
10097 << " <= " << osdmap
->get_epoch()
10098 << ", stale, dropping" << dendl
;
10099 pushes_to_free
+= qi
.get_reserved_pushes();
10100 slot
.to_process
.pop_front();
10103 if (slot
.to_process
.empty() &&
10104 slot
.num_running
== 0 &&
10106 dout(20) << __func__
<< " " << p
->first
<< " empty, pruning" << dendl
;
10107 p
= sdata
->pg_slots
.erase(p
);
10113 if (pushes_to_free
> 0) {
10114 osd
->service
.release_reserved_pushes(pushes_to_free
);
10118 void OSD::ShardedOpWQ::clear_pg_pointer(spg_t pgid
)
10120 uint32_t shard_index
= pgid
.hash_to_shard(shard_list
.size());
10121 auto sdata
= shard_list
[shard_index
];
10122 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10123 auto p
= sdata
->pg_slots
.find(pgid
);
10124 if (p
!= sdata
->pg_slots
.end()) {
10125 auto& slot
= p
->second
;
10126 dout(20) << __func__
<< " " << pgid
<< " pg " << slot
.pg
<< dendl
;
10127 assert(!slot
.pg
|| slot
.pg
->deleting
);
10132 void OSD::ShardedOpWQ::clear_pg_slots()
10134 for (auto sdata
: shard_list
) {
10135 Mutex::Locker
l(sdata
->sdata_op_ordering_lock
);
10136 sdata
->pg_slots
.clear();
10137 sdata
->waiting_for_pg_osdmap
.reset();
10138 // don't bother with reserved pushes; we are shutting down
10143 #define dout_prefix *_dout << "osd." << osd->whoami << " op_wq(" << shard_index << ") "
10145 void OSD::ShardedOpWQ::_process(uint32_t thread_index
, heartbeat_handle_d
*hb
)
10147 uint32_t shard_index
= thread_index
% num_shards
;
10148 ShardData
*sdata
= shard_list
[shard_index
];
10149 assert(NULL
!= sdata
);
10152 sdata
->sdata_op_ordering_lock
.Lock();
10153 if (sdata
->pqueue
->empty()) {
10154 dout(20) << __func__
<< " empty q, waiting" << dendl
;
10155 // optimistically sleep a moment; maybe another work item will come along.
10156 osd
->cct
->get_heartbeat_map()->reset_timeout(hb
,
10157 osd
->cct
->_conf
->threadpool_default_timeout
, 0);
10158 sdata
->sdata_lock
.Lock();
10159 sdata
->sdata_op_ordering_lock
.Unlock();
10160 sdata
->sdata_cond
.WaitInterval(sdata
->sdata_lock
,
10161 utime_t(osd
->cct
->_conf
->threadpool_empty_queue_max_wait
, 0));
10162 sdata
->sdata_lock
.Unlock();
10163 sdata
->sdata_op_ordering_lock
.Lock();
10164 if (sdata
->pqueue
->empty()) {
10165 sdata
->sdata_op_ordering_lock
.Unlock();
10169 pair
<spg_t
, PGQueueable
> item
= sdata
->pqueue
->dequeue();
10170 if (osd
->is_stopping()) {
10171 sdata
->sdata_op_ordering_lock
.Unlock();
10172 return; // OSD shutdown, discard.
10175 uint64_t requeue_seq
;
10177 auto& slot
= sdata
->pg_slots
[item
.first
];
10178 dout(30) << __func__
<< " " << item
.first
10179 << " to_process " << slot
.to_process
10180 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10181 slot
.to_process
.push_back(item
.second
);
10182 // note the requeue seq now...
10183 requeue_seq
= slot
.requeue_seq
;
10184 if (slot
.waiting_for_pg
) {
10185 // save ourselves a bit of effort
10186 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10187 << " queued, waiting_for_pg" << dendl
;
10188 sdata
->sdata_op_ordering_lock
.Unlock();
10192 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10193 << " queued" << dendl
;
10194 ++slot
.num_running
;
10196 sdata
->sdata_op_ordering_lock
.Unlock();
10198 osd
->service
.maybe_inject_dispatch_delay();
10200 // [lookup +] lock pg (if we have it)
10202 pg
= osd
->_lookup_lock_pg(item
.first
);
10207 osd
->service
.maybe_inject_dispatch_delay();
10209 boost::optional
<PGQueueable
> qi
;
10211 // we don't use a Mutex::Locker here because of the
10212 // osd->service.release_reserved_pushes() call below
10213 sdata
->sdata_op_ordering_lock
.Lock();
10215 auto q
= sdata
->pg_slots
.find(item
.first
);
10216 assert(q
!= sdata
->pg_slots
.end());
10217 auto& slot
= q
->second
;
10218 --slot
.num_running
;
10220 if (slot
.to_process
.empty()) {
10221 // raced with wake_pg_waiters or prune_pg_waiters
10222 dout(20) << __func__
<< " " << item
.first
<< " nothing queued" << dendl
;
10226 sdata
->sdata_op_ordering_lock
.Unlock();
10229 if (requeue_seq
!= slot
.requeue_seq
) {
10230 dout(20) << __func__
<< " " << item
.first
10231 << " requeue_seq " << slot
.requeue_seq
<< " > our "
10232 << requeue_seq
<< ", we raced with wake_pg_waiters"
10237 sdata
->sdata_op_ordering_lock
.Unlock();
10240 if (pg
&& !slot
.pg
&& !pg
->deleting
) {
10241 dout(20) << __func__
<< " " << item
.first
<< " set pg to " << pg
<< dendl
;
10244 dout(30) << __func__
<< " " << item
.first
<< " to_process " << slot
.to_process
10245 << " waiting_for_pg=" << (int)slot
.waiting_for_pg
<< dendl
;
10247 // make sure we're not already waiting for this pg
10248 if (slot
.waiting_for_pg
) {
10249 dout(20) << __func__
<< " " << item
.first
<< " item " << item
.second
10250 << " slot is waiting_for_pg" << dendl
;
10254 sdata
->sdata_op_ordering_lock
.Unlock();
10259 qi
= slot
.to_process
.front();
10260 slot
.to_process
.pop_front();
10261 dout(20) << __func__
<< " " << item
.first
<< " item " << *qi
10262 << " pg " << pg
<< dendl
;
10265 // should this pg shard exist on this osd in this (or a later) epoch?
10266 OSDMapRef osdmap
= sdata
->waiting_for_pg_osdmap
;
10267 if (osdmap
->is_up_acting_osd_shard(item
.first
, osd
->whoami
)) {
10268 dout(20) << __func__
<< " " << item
.first
10269 << " no pg, should exist, will wait" << " on " << *qi
<< dendl
;
10270 slot
.to_process
.push_front(*qi
);
10271 slot
.waiting_for_pg
= true;
10272 } else if (qi
->get_map_epoch() > osdmap
->get_epoch()) {
10273 dout(20) << __func__
<< " " << item
.first
<< " no pg, item epoch is "
10274 << qi
->get_map_epoch() << " > " << osdmap
->get_epoch()
10275 << ", will wait on " << *qi
<< dendl
;
10276 slot
.to_process
.push_front(*qi
);
10277 slot
.waiting_for_pg
= true;
10279 dout(20) << __func__
<< " " << item
.first
<< " no pg, shouldn't exist,"
10280 << " dropping " << *qi
<< dendl
;
10281 // share map with client?
10282 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10283 Session
*session
= static_cast<Session
*>(
10284 (*_op
)->get_req()->get_connection()->get_priv());
10286 osd
->maybe_share_map(session
, *_op
, sdata
->waiting_for_pg_osdmap
);
10290 unsigned pushes_to_free
= qi
->get_reserved_pushes();
10291 if (pushes_to_free
> 0) {
10292 sdata
->sdata_op_ordering_lock
.Unlock();
10293 osd
->service
.release_reserved_pushes(pushes_to_free
);
10297 sdata
->sdata_op_ordering_lock
.Unlock();
10300 sdata
->sdata_op_ordering_lock
.Unlock();
10303 // osd_opwq_process marks the point at which an operation has been dequeued
10304 // and will begin to be handled by a worker thread.
10308 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10309 reqid
= (*_op
)->get_reqid();
10312 tracepoint(osd
, opwq_process_start
, reqid
.name
._type
,
10313 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10316 lgeneric_subdout(osd
->cct
, osd
, 30) << "dequeue status: ";
10317 Formatter
*f
= Formatter::create("json");
10318 f
->open_object_section("q");
10320 f
->close_section();
10325 ThreadPool::TPHandle
tp_handle(osd
->cct
, hb
, timeout_interval
,
10327 qi
->run(osd
, pg
, tp_handle
);
10332 if (boost::optional
<OpRequestRef
> _op
= qi
->maybe_get_op()) {
10333 reqid
= (*_op
)->get_reqid();
10336 tracepoint(osd
, opwq_process_finish
, reqid
.name
._type
,
10337 reqid
.name
._num
, reqid
.tid
, reqid
.inc
);
10343 void OSD::ShardedOpWQ::_enqueue(pair
<spg_t
, PGQueueable
> item
) {
10344 uint32_t shard_index
=
10345 item
.first
.hash_to_shard(shard_list
.size());
10347 ShardData
* sdata
= shard_list
[shard_index
];
10348 assert (NULL
!= sdata
);
10349 unsigned priority
= item
.second
.get_priority();
10350 unsigned cost
= item
.second
.get_cost();
10351 sdata
->sdata_op_ordering_lock
.Lock();
10353 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10354 if (priority
>= osd
->op_prio_cutoff
)
10355 sdata
->pqueue
->enqueue_strict(
10356 item
.second
.get_owner(), priority
, item
);
10358 sdata
->pqueue
->enqueue(
10359 item
.second
.get_owner(),
10360 priority
, cost
, item
);
10361 sdata
->sdata_op_ordering_lock
.Unlock();
10363 sdata
->sdata_lock
.Lock();
10364 sdata
->sdata_cond
.SignalOne();
10365 sdata
->sdata_lock
.Unlock();
10369 void OSD::ShardedOpWQ::_enqueue_front(pair
<spg_t
, PGQueueable
> item
)
10371 uint32_t shard_index
= item
.first
.hash_to_shard(shard_list
.size());
10372 ShardData
* sdata
= shard_list
[shard_index
];
10373 assert (NULL
!= sdata
);
10374 sdata
->sdata_op_ordering_lock
.Lock();
10375 auto p
= sdata
->pg_slots
.find(item
.first
);
10376 if (p
!= sdata
->pg_slots
.end() && !p
->second
.to_process
.empty()) {
10377 // we may be racing with _process, which has dequeued a new item
10378 // from pqueue, put it on to_process, and is now busy taking the
10379 // pg lock. ensure this old requeued item is ordered before any
10380 // such newer item in to_process.
10381 p
->second
.to_process
.push_front(item
.second
);
10382 item
.second
= p
->second
.to_process
.back();
10383 p
->second
.to_process
.pop_back();
10384 dout(20) << __func__
<< " " << item
.first
10385 << " " << p
->second
.to_process
.front()
10386 << " shuffled w/ " << item
.second
<< dendl
;
10388 dout(20) << __func__
<< " " << item
.first
<< " " << item
.second
<< dendl
;
10390 sdata
->_enqueue_front(item
, osd
->op_prio_cutoff
);
10391 sdata
->sdata_op_ordering_lock
.Unlock();
10392 sdata
->sdata_lock
.Lock();
10393 sdata
->sdata_cond
.SignalOne();
10394 sdata
->sdata_lock
.Unlock();
10398 namespace osd_cmds
{
10400 int heap(CephContext
& cct
, cmdmap_t
& cmdmap
, Formatter
& f
, std::ostream
& os
)
10402 if (!ceph_using_tcmalloc()) {
10403 os
<< "could not issue heap profiler command -- not using tcmalloc!";
10404 return -EOPNOTSUPP
;
10408 if (!cmd_getval(&cct
, cmdmap
, "heapcmd", cmd
)) {
10409 os
<< "unable to get value for command \"" << cmd
<< "\"";
10413 std::vector
<std::string
> cmd_vec
;
10414 get_str_vec(cmd
, cmd_vec
);
10416 ceph_heap_profiler_handle_command(cmd_vec
, os
);
10421 }} // namespace ceph::osd_cmds
10424 std::ostream
& operator<<(std::ostream
& out
, const OSD::io_queue
& q
) {
10426 case OSD::io_queue::prioritized
:
10427 out
<< "prioritized";
10429 case OSD::io_queue::weightedpriority
:
10430 out
<< "weightedpriority";
10432 case OSD::io_queue::mclock_opclass
:
10433 out
<< "mclock_opclass";
10435 case OSD::io_queue::mclock_client
:
10436 out
<< "mclock_client";