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) 2015 Red Hat
8 * This is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License version 2.1, as published by the Free Software
11 * Foundation. See file COPYING.
15 #include "common/debug.h"
16 #include "common/errno.h"
18 #include "messages/MClientRequestForward.h"
19 #include "messages/MMDSLoadTargets.h"
20 #include "messages/MMDSMap.h"
21 #include "messages/MMDSTableRequest.h"
22 #include "messages/MCommand.h"
23 #include "messages/MCommandReply.h"
25 #include "MDSDaemon.h"
27 #include "SnapClient.h"
28 #include "SnapServer.h"
29 #include "MDBalancer.h"
33 #include "mon/MonClient.h"
34 #include "common/HeartbeatMap.h"
35 #include "ScrubStack.h"
40 #define dout_context g_ceph_context
41 #define dout_subsys ceph_subsys_mds
43 #define dout_prefix *_dout << "mds." << whoami << '.' << incarnation << ' '
54 Context
*respawn_hook_
,
55 Context
*suicide_hook_
)
57 whoami(whoami_
), incarnation(0),
58 mds_lock(mds_lock_
), cct(msgr
->cct
), clog(clog_
), timer(timer_
),
60 objecter(new Objecter(g_ceph_context
, msgr
, monc_
, nullptr, 0, 0)),
61 server(NULL
), mdcache(NULL
), locker(NULL
), mdlog(NULL
),
62 balancer(NULL
), scrubstack(NULL
),
63 damage_table(whoami_
),
64 inotable(NULL
), snapserver(NULL
), snapclient(NULL
),
65 sessionmap(this), logger(NULL
), mlogger(NULL
),
66 op_tracker(g_ceph_context
, g_conf
->mds_enable_op_tracker
,
67 g_conf
->osd_num_op_tracker_shard
),
68 last_state(MDSMap::STATE_BOOT
),
69 state(MDSMap::STATE_BOOT
),
70 cluster_degraded(false), stopping(false),
71 purge_queue(g_ceph_context
, whoami_
,
72 mdsmap_
->get_metadata_pool(), objecter
,
75 // Purge Queue operates inside mds_lock when we're calling into
76 // it, and outside when in background, so must handle both cases.
77 if (mds_lock
.is_locked_by_me()) {
85 progress_thread(this), dispatch_depth(0),
86 hb(NULL
), last_tid(0), osd_epoch_barrier(0), beacon(beacon_
),
87 mds_slow_req_count(0),
88 last_client_mdsmap_bcast(0),
89 messenger(msgr
), monc(monc_
),
90 respawn_hook(respawn_hook_
),
91 suicide_hook(suicide_hook_
),
92 standby_replaying(false)
94 hb
= g_ceph_context
->get_heartbeat_map()->add_worker("MDSRank", pthread_self());
96 purge_queue
.update_op_limit(*mdsmap
);
98 objecter
->unset_honor_osdmap_full();
100 finisher
= new Finisher(cct
);
102 mdcache
= new MDCache(this, purge_queue
);
103 mdlog
= new MDLog(this);
104 balancer
= new MDBalancer(this, messenger
, monc
);
106 scrubstack
= new ScrubStack(mdcache
, finisher
);
108 inotable
= new InoTable(this);
109 snapserver
= new SnapServer(this, monc
);
110 snapclient
= new SnapClient(this);
112 server
= new Server(this);
113 locker
= new Locker(this, mdcache
);
115 op_tracker
.set_complaint_and_threshold(cct
->_conf
->mds_op_complaint_time
,
116 cct
->_conf
->mds_op_log_threshold
);
117 op_tracker
.set_history_size_and_duration(cct
->_conf
->mds_op_history_size
,
118 cct
->_conf
->mds_op_history_duration
);
124 g_ceph_context
->get_heartbeat_map()->remove_worker(hb
);
127 if (scrubstack
) { delete scrubstack
; scrubstack
= NULL
; }
128 if (mdcache
) { delete mdcache
; mdcache
= NULL
; }
129 if (mdlog
) { delete mdlog
; mdlog
= NULL
; }
130 if (balancer
) { delete balancer
; balancer
= NULL
; }
131 if (inotable
) { delete inotable
; inotable
= NULL
; }
132 if (snapserver
) { delete snapserver
; snapserver
= NULL
; }
133 if (snapclient
) { delete snapclient
; snapclient
= NULL
; }
134 if (mdsmap
) { delete mdsmap
; mdsmap
= 0; }
136 if (server
) { delete server
; server
= 0; }
137 if (locker
) { delete locker
; locker
= 0; }
140 g_ceph_context
->get_perfcounters_collection()->remove(logger
);
145 g_ceph_context
->get_perfcounters_collection()->remove(mlogger
);
163 void MDSRankDispatcher::init()
166 messenger
->add_dispatcher_head(objecter
);
173 // Expose the OSDMap (already populated during MDS::init) to anyone
174 // who is interested in it.
177 progress_thread
.create("mds_rank_progr");
184 void MDSRank::update_targets(utime_t now
)
186 // get MonMap's idea of my export_targets
187 const set
<mds_rank_t
>& map_targets
= mdsmap
->get_mds_info(get_nodeid()).export_targets
;
189 dout(20) << "updating export targets, currently " << map_targets
.size() << " ranks are targets" << dendl
;
192 set
<mds_rank_t
> new_map_targets
;
194 auto it
= export_targets
.begin();
195 while (it
!= export_targets
.end()) {
196 mds_rank_t rank
= it
->first
;
197 double val
= it
->second
.get(now
);
198 dout(20) << "export target mds." << rank
<< " value is " << val
<< " @ " << now
<< dendl
;
201 dout(15) << "export target mds." << rank
<< " is no longer an export target" << dendl
;
202 export_targets
.erase(it
++);
206 if (!map_targets
.count(rank
)) {
207 dout(15) << "export target mds." << rank
<< " not in map's export_targets" << dendl
;
210 new_map_targets
.insert(rank
);
213 if (new_map_targets
.size() < map_targets
.size()) {
214 dout(15) << "export target map holds stale targets, sending update" << dendl
;
219 dout(15) << "updating export_targets, now " << new_map_targets
.size() << " ranks are targets" << dendl
;
220 MMDSLoadTargets
* m
= new MMDSLoadTargets(mds_gid_t(monc
->get_global_id()), new_map_targets
);
221 monc
->send_mon_message(m
);
225 void MDSRank::hit_export_target(utime_t now
, mds_rank_t rank
, double amount
)
227 double rate
= g_conf
->mds_bal_target_decay
;
229 amount
= 100.0/g_conf
->mds_bal_target_decay
; /* a good default for "i am trying to keep this export_target active" */
231 auto em
= export_targets
.emplace(std::piecewise_construct
, std::forward_as_tuple(rank
), std::forward_as_tuple(now
, DecayRate(rate
)));
233 dout(15) << "hit export target (new) " << amount
<< " @ " << now
<< dendl
;
235 dout(15) << "hit export target " << amount
<< " @ " << now
<< dendl
;
237 em
.first
->second
.hit(now
, amount
);
240 void MDSRankDispatcher::tick()
244 if (beacon
.is_laggy()) {
245 dout(5) << "tick bailing out since we seem laggy" << dendl
;
249 check_ops_in_flight();
251 // Wake up thread in case we use to be laggy and have waiting_for_nolaggy
252 // messages to progress.
253 progress_thread
.signal();
255 // make sure mds log flushes, trims periodically
258 if (is_active() || is_stopping()) {
260 mdcache
->trim_client_leases();
261 mdcache
->check_memory_usage();
262 mdlog
->trim(); // NOT during recovery!
266 mds_load_t load
= balancer
->get_load(ceph_clock_now());
269 logger
->set(l_mds_load_cent
, 100 * load
.mds_load());
270 logger
->set(l_mds_dispatch_queue_len
, messenger
->get_dispatch_queue_len());
271 logger
->set(l_mds_subtrees
, mdcache
->num_subtrees());
277 if (is_clientreplay() || is_active() || is_stopping()) {
278 server
->find_idle_sessions();
283 server
->reconnect_tick();
287 mdcache
->find_stale_fragment_freeze();
288 mdcache
->migrator
->find_stale_export_freeze();
290 snapserver
->check_osd_map(false);
293 if (is_active() || is_stopping()) {
294 update_targets(ceph_clock_now());
300 if (mdcache
->shutdown_pass()) {
301 uint64_t pq_progress
= 0 ;
302 uint64_t pq_total
= 0;
303 size_t pq_in_flight
= 0;
304 if (!purge_queue
.drain(&pq_progress
, &pq_total
, &pq_in_flight
)) {
305 dout(7) << "shutdown_pass=true, but still waiting for purge queue"
307 // This takes unbounded time, so we must indicate progress
308 // to the administrator: we do it in a slightly imperfect way
309 // by sending periodic (tick frequency) clog messages while
311 clog
->info() << "MDS rank " << whoami
<< " waiting for purge queue ("
312 << std::dec
<< pq_progress
<< "/" << pq_total
<< " " << pq_in_flight
313 << " files purging" << ")";
315 dout(7) << "shutdown_pass=true, finished w/ shutdown, moving to "
316 "down:stopped" << dendl
;
321 dout(7) << "shutdown_pass=false" << dendl
;
325 // Expose ourselves to Beacon to update health indicators
326 beacon
.notify_health(this);
329 void MDSRankDispatcher::shutdown()
331 // It should never be possible for shutdown to get called twice, because
332 // anyone picking up mds_lock checks if stopping is true and drops
334 assert(stopping
== false);
337 dout(1) << __func__
<< ": shutting down rank " << whoami
<< dendl
;
341 // MDLog has to shut down before the finisher, because some of its
342 // threads block on IOs that require finisher to complete.
348 purge_queue
.shutdown();
351 finisher
->stop(); // no flushing
354 if (objecter
->initialized
)
355 objecter
->shutdown();
359 op_tracker
.on_shutdown();
361 progress_thread
.shutdown();
363 // release mds_lock for finisher/messenger threads (e.g.
364 // MDSDaemon::ms_handle_reset called from Messenger).
367 // shut down messenger
368 messenger
->shutdown();
372 // Workaround unclean shutdown: HeartbeatMap will assert if
373 // worker is not removed (as we do in ~MDS), but ~MDS is not
374 // always called after suicide.
376 g_ceph_context
->get_heartbeat_map()->remove_worker(hb
);
382 * Helper for simple callbacks that call a void fn with no args.
384 class C_MDS_VoidFn
: public MDSInternalContext
386 typedef void (MDSRank::*fn_ptr
)();
390 C_MDS_VoidFn(MDSRank
*mds_
, fn_ptr fn_
)
391 : MDSInternalContext(mds_
), fn(fn_
)
397 void finish(int r
) override
403 int64_t MDSRank::get_metadata_pool()
405 return mdsmap
->get_metadata_pool();
408 MDSTableClient
*MDSRank::get_table_client(int t
)
411 case TABLE_ANCHOR
: return NULL
;
412 case TABLE_SNAP
: return snapclient
;
413 default: ceph_abort();
417 MDSTableServer
*MDSRank::get_table_server(int t
)
420 case TABLE_ANCHOR
: return NULL
;
421 case TABLE_SNAP
: return snapserver
;
422 default: ceph_abort();
426 void MDSRank::suicide()
429 suicide_hook
->complete(0);
434 void MDSRank::respawn()
437 respawn_hook
->complete(0);
442 void MDSRank::damaged()
444 assert(whoami
!= MDS_RANK_NONE
);
445 assert(mds_lock
.is_locked_by_me());
447 beacon
.set_want_state(mdsmap
, MDSMap::STATE_DAMAGED
);
448 monc
->flush_log(); // Flush any clog error from before we were called
449 beacon
.notify_health(this); // Include latest status in our swan song
450 beacon
.send_and_wait(g_conf
->mds_mon_shutdown_timeout
);
452 // It's okay if we timed out and the mon didn't get our beacon, because
453 // another daemon (or ourselves after respawn) will eventually take the
454 // rank and report DAMAGED again when it hits same problem we did.
456 respawn(); // Respawn into standby in case mon has other work for us
459 void MDSRank::damaged_unlocked()
461 Mutex::Locker
l(mds_lock
);
465 void MDSRank::handle_write_error(int err
)
467 if (err
== -EBLACKLISTED
) {
468 derr
<< "we have been blacklisted (fenced), respawning..." << dendl
;
473 if (g_conf
->mds_action_on_write_error
>= 2) {
474 derr
<< "unhandled write error " << cpp_strerror(err
) << ", suicide..." << dendl
;
476 } else if (g_conf
->mds_action_on_write_error
== 1) {
477 derr
<< "unhandled write error " << cpp_strerror(err
) << ", force readonly..." << dendl
;
478 mdcache
->force_readonly();
481 derr
<< "unhandled write error " << cpp_strerror(err
) << ", ignore..." << dendl
;
485 void *MDSRank::ProgressThread::entry()
487 Mutex::Locker
l(mds
->mds_lock
);
489 while (!mds
->stopping
&&
490 mds
->finished_queue
.empty() &&
491 (mds
->waiting_for_nolaggy
.empty() || mds
->beacon
.is_laggy())) {
492 cond
.Wait(mds
->mds_lock
);
499 mds
->_advance_queues();
506 void MDSRank::ProgressThread::shutdown()
508 assert(mds
->mds_lock
.is_locked_by_me());
509 assert(mds
->stopping
);
512 // Stopping is set, we will fall out of our main loop naturally
514 // Kick the thread to notice mds->stopping, and join it
516 mds
->mds_lock
.Unlock();
519 mds
->mds_lock
.Lock();
523 bool MDSRankDispatcher::ms_dispatch(Message
*m
)
526 inc_dispatch_depth();
527 ret
= _dispatch(m
, true);
528 dec_dispatch_depth();
532 /* If this function returns true, it recognizes the message and has taken the
533 * reference. If it returns false, it has done neither. */
534 bool MDSRank::_dispatch(Message
*m
, bool new_msg
)
536 if (is_stale_message(m
)) {
541 if (beacon
.is_laggy()) {
542 dout(10) << " laggy, deferring " << *m
<< dendl
;
543 waiting_for_nolaggy
.push_back(m
);
544 } else if (new_msg
&& !waiting_for_nolaggy
.empty()) {
545 dout(10) << " there are deferred messages, deferring " << *m
<< dendl
;
546 waiting_for_nolaggy
.push_back(m
);
548 if (!handle_deferrable_message(m
)) {
549 dout(0) << "unrecognized message " << *m
<< dendl
;
556 if (dispatch_depth
> 1)
559 // finish any triggered contexts
562 if (beacon
.is_laggy()) {
563 // We've gone laggy during dispatch, don't do any
568 // done with all client replayed requests?
569 if (is_clientreplay() &&
570 mdcache
->is_open() &&
571 replay_queue
.empty() &&
572 beacon
.get_want_state() == MDSMap::STATE_CLIENTREPLAY
) {
573 int num_requests
= mdcache
->get_num_client_requests();
574 dout(10) << " still have " << num_requests
<< " active replay requests" << dendl
;
575 if (num_requests
== 0)
579 // hack: thrash exports
580 static utime_t start
;
581 utime_t now
= ceph_clock_now();
582 if (start
== utime_t())
584 /*double el = now - start;
587 for (int i
=0; i
<g_conf
->mds_thrash_exports
; i
++) {
589 if (!is_active()) break;
590 mdsmap
->get_mds_set(s
, MDSMap::STATE_ACTIVE
);
591 if (s
.size() < 2 || CInode::count() < 10)
592 break; // need peers for this to work.
593 if (mdcache
->migrator
->get_num_exporting() > g_conf
->mds_thrash_exports
* 5 ||
594 mdcache
->migrator
->get_export_queue_size() > g_conf
->mds_thrash_exports
* 10)
597 dout(7) << "mds thrashing exports pass " << (i
+1) << "/" << g_conf
->mds_thrash_exports
<< dendl
;
599 // pick a random dir inode
600 CInode
*in
= mdcache
->hack_pick_random_inode();
603 in
->get_dirfrags(ls
);
604 if (!ls
.empty()) { // must be an open dir.
605 list
<CDir
*>::iterator p
= ls
.begin();
606 int n
= rand() % ls
.size();
610 if (!dir
->get_parent_dir()) continue; // must be linked.
611 if (!dir
->is_auth()) continue; // must be auth.
615 int k
= rand() % s
.size();
616 set
<mds_rank_t
>::iterator p
= s
.begin();
619 } while (dest
== whoami
);
620 mdcache
->migrator
->export_dir_nicely(dir
,dest
);
623 // hack: thrash fragments
624 for (int i
=0; i
<g_conf
->mds_thrash_fragments
; i
++) {
625 if (!is_active()) break;
626 if (mdcache
->get_num_fragmenting_dirs() > 5 * g_conf
->mds_thrash_fragments
) break;
627 dout(7) << "mds thrashing fragments pass " << (i
+1) << "/" << g_conf
->mds_thrash_fragments
<< dendl
;
629 // pick a random dir inode
630 CInode
*in
= mdcache
->hack_pick_random_inode();
633 in
->get_dirfrags(ls
);
634 if (ls
.empty()) continue; // must be an open dir.
635 CDir
*dir
= ls
.front();
636 if (!dir
->get_parent_dir()) continue; // must be linked.
637 if (!dir
->is_auth()) continue; // must be auth.
638 frag_t fg
= dir
->get_frag();
639 if (mdsmap
->allows_dirfrags()) {
640 if ((fg
== frag_t() || (rand() % (1 << fg
.bits()) == 0))) {
641 mdcache
->split_dir(dir
, 1);
643 balancer
->queue_merge(dir
);
648 // hack: force hash root?
651 mdcache->get_root() &&
652 mdcache->get_root()->dir &&
653 !(mdcache->get_root()->dir->is_hashed() ||
654 mdcache->get_root()->dir->is_hashing())) {
655 dout(0) << "hashing root" << dendl;
656 mdcache->migrator->hash_dir(mdcache->get_root()->dir);
664 void MDSRank::update_mlogger()
667 mlogger
->set(l_mdm_ino
, CInode::count());
668 mlogger
->set(l_mdm_dir
, CDir::count());
669 mlogger
->set(l_mdm_dn
, CDentry::count());
670 mlogger
->set(l_mdm_cap
, Capability::count());
671 mlogger
->set(l_mdm_inoa
, CInode::increments());
672 mlogger
->set(l_mdm_inos
, CInode::decrements());
673 mlogger
->set(l_mdm_dira
, CDir::increments());
674 mlogger
->set(l_mdm_dirs
, CDir::decrements());
675 mlogger
->set(l_mdm_dna
, CDentry::increments());
676 mlogger
->set(l_mdm_dns
, CDentry::decrements());
677 mlogger
->set(l_mdm_capa
, Capability::increments());
678 mlogger
->set(l_mdm_caps
, Capability::decrements());
679 mlogger
->set(l_mdm_buf
, buffer::get_total_alloc());
684 * lower priority messages we defer if we seem laggy
686 bool MDSRank::handle_deferrable_message(Message
*m
)
688 int port
= m
->get_type() & 0xff00;
692 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS
);
693 mdcache
->dispatch(m
);
696 case MDS_PORT_MIGRATOR
:
697 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS
);
698 mdcache
->migrator
->dispatch(m
);
702 switch (m
->get_type()) {
704 case CEPH_MSG_CLIENT_SESSION
:
705 case CEPH_MSG_CLIENT_RECONNECT
:
706 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_CLIENT
);
708 case CEPH_MSG_CLIENT_REQUEST
:
711 case MSG_MDS_SLAVE_REQUEST
:
712 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS
);
716 case MSG_MDS_HEARTBEAT
:
717 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS
);
718 balancer
->proc_message(m
);
721 case MSG_MDS_TABLE_REQUEST
:
722 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS
);
724 MMDSTableRequest
*req
= static_cast<MMDSTableRequest
*>(m
);
726 MDSTableClient
*client
= get_table_client(req
->table
);
727 client
->handle_request(req
);
729 MDSTableServer
*server
= get_table_server(req
->table
);
730 server
->handle_request(req
);
736 case MSG_MDS_INODEFILECAPS
:
737 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS
);
741 case CEPH_MSG_CLIENT_CAPS
:
742 case CEPH_MSG_CLIENT_CAPRELEASE
:
743 case CEPH_MSG_CLIENT_LEASE
:
744 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_CLIENT
);
757 * Advance finished_queue and waiting_for_nolaggy.
759 * Usually drain both queues, but may not drain waiting_for_nolaggy
760 * if beacon is currently laggy.
762 void MDSRank::_advance_queues()
764 assert(mds_lock
.is_locked_by_me());
766 while (!finished_queue
.empty()) {
767 dout(7) << "mds has " << finished_queue
.size() << " queued contexts" << dendl
;
768 dout(10) << finished_queue
<< dendl
;
769 list
<MDSInternalContextBase
*> ls
;
770 ls
.swap(finished_queue
);
771 while (!ls
.empty()) {
772 dout(10) << " finish " << ls
.front() << dendl
;
773 ls
.front()->complete(0);
780 while (!waiting_for_nolaggy
.empty()) {
781 // stop if we're laggy now!
782 if (beacon
.is_laggy())
785 Message
*old
= waiting_for_nolaggy
.front();
786 waiting_for_nolaggy
.pop_front();
788 if (is_stale_message(old
)) {
791 dout(7) << " processing laggy deferred " << *old
<< dendl
;
792 if (!handle_deferrable_message(old
)) {
793 dout(0) << "unrecognized message " << *old
<< dendl
;
803 * Call this when you take mds_lock, or periodically if you're going to
804 * hold the lock for a long time (e.g. iterating over clients/inodes)
806 void MDSRank::heartbeat_reset()
808 // Any thread might jump into mds_lock and call us immediately
809 // after a call to suicide() completes, in which case MDSRank::hb
810 // has been freed and we are a no-op.
816 // NB not enabling suicide grace, because the mon takes care of killing us
817 // (by blacklisting us) when we fail to send beacons, and it's simpler to
818 // only have one way of dying.
819 g_ceph_context
->get_heartbeat_map()->reset_timeout(hb
, g_conf
->mds_beacon_grace
, 0);
822 bool MDSRank::is_stale_message(Message
*m
) const
825 if (m
->get_source().is_mds()) {
826 mds_rank_t from
= mds_rank_t(m
->get_source().num());
827 if (!mdsmap
->have_inst(from
) ||
828 mdsmap
->get_inst(from
) != m
->get_source_inst() ||
829 mdsmap
->is_down(from
)) {
831 if (m
->get_type() == CEPH_MSG_MDS_MAP
) {
832 dout(5) << "got " << *m
<< " from old/bad/imposter mds " << m
->get_source()
833 << ", but it's an mdsmap, looking at it" << dendl
;
834 } else if (m
->get_type() == MSG_MDS_CACHEEXPIRE
&&
835 mdsmap
->get_inst(from
) == m
->get_source_inst()) {
836 dout(5) << "got " << *m
<< " from down mds " << m
->get_source()
837 << ", but it's a cache_expire, looking at it" << dendl
;
839 dout(5) << "got " << *m
<< " from down/old/bad/imposter mds " << m
->get_source()
840 << ", dropping" << dendl
;
849 void MDSRank::send_message(Message
*m
, Connection
*c
)
856 void MDSRank::send_message_mds(Message
*m
, mds_rank_t mds
)
858 if (!mdsmap
->is_up(mds
)) {
859 dout(10) << "send_message_mds mds." << mds
<< " not up, dropping " << *m
<< dendl
;
864 // send mdsmap first?
865 if (mds
!= whoami
&& peer_mdsmap_epoch
[mds
] < mdsmap
->get_epoch()) {
866 messenger
->send_message(new MMDSMap(monc
->get_fsid(), mdsmap
),
867 mdsmap
->get_inst(mds
));
868 peer_mdsmap_epoch
[mds
] = mdsmap
->get_epoch();
872 messenger
->send_message(m
, mdsmap
->get_inst(mds
));
875 void MDSRank::forward_message_mds(Message
*m
, mds_rank_t mds
)
877 assert(mds
!= whoami
);
880 if (m
->get_type() == CEPH_MSG_CLIENT_REQUEST
&&
881 (static_cast<MClientRequest
*>(m
))->get_source().is_client()) {
882 MClientRequest
*creq
= static_cast<MClientRequest
*>(m
);
883 creq
->inc_num_fwd(); // inc forward counter
886 * don't actually forward if non-idempotent!
887 * client has to do it. although the MDS will ignore duplicate requests,
888 * the affected metadata may migrate, in which case the new authority
889 * won't have the metareq_id in the completed request map.
891 // NEW: always make the client resend!
892 bool client_must_resend
= true; //!creq->can_forward();
894 // tell the client where it should go
895 messenger
->send_message(new MClientRequestForward(creq
->get_tid(), mds
, creq
->get_num_fwd(),
897 creq
->get_source_inst());
899 if (client_must_resend
) {
905 // these are the only types of messages we should be 'forwarding'; they
906 // explicitly encode their source mds, which gets clobbered when we resend
908 assert(m
->get_type() == MSG_MDS_DIRUPDATE
||
909 m
->get_type() == MSG_MDS_EXPORTDIRDISCOVER
);
911 // send mdsmap first?
912 if (peer_mdsmap_epoch
[mds
] < mdsmap
->get_epoch()) {
913 messenger
->send_message(new MMDSMap(monc
->get_fsid(), mdsmap
),
914 mdsmap
->get_inst(mds
));
915 peer_mdsmap_epoch
[mds
] = mdsmap
->get_epoch();
918 messenger
->send_message(m
, mdsmap
->get_inst(mds
));
923 void MDSRank::send_message_client_counted(Message
*m
, client_t client
)
925 Session
*session
= sessionmap
.get_session(entity_name_t::CLIENT(client
.v
));
927 send_message_client_counted(m
, session
);
929 dout(10) << "send_message_client_counted no session for client." << client
<< " " << *m
<< dendl
;
933 void MDSRank::send_message_client_counted(Message
*m
, Connection
*connection
)
935 Session
*session
= static_cast<Session
*>(connection
->get_priv());
937 session
->put(); // do not carry ref
938 send_message_client_counted(m
, session
);
940 dout(10) << "send_message_client_counted has no session for " << m
->get_source_inst() << dendl
;
941 // another Connection took over the Session
945 void MDSRank::send_message_client_counted(Message
*m
, Session
*session
)
947 version_t seq
= session
->inc_push_seq();
948 dout(10) << "send_message_client_counted " << session
->info
.inst
.name
<< " seq "
949 << seq
<< " " << *m
<< dendl
;
950 if (session
->connection
) {
951 session
->connection
->send_message(m
);
953 session
->preopen_out_queue
.push_back(m
);
957 void MDSRank::send_message_client(Message
*m
, Session
*session
)
959 dout(10) << "send_message_client " << session
->info
.inst
<< " " << *m
<< dendl
;
960 if (session
->connection
) {
961 session
->connection
->send_message(m
);
963 session
->preopen_out_queue
.push_back(m
);
968 * This is used whenever a RADOS operation has been cancelled
969 * or a RADOS client has been blacklisted, to cause the MDS and
970 * any clients to wait for this OSD epoch before using any new caps.
972 * See doc/cephfs/eviction
974 void MDSRank::set_osd_epoch_barrier(epoch_t e
)
976 dout(4) << __func__
<< ": epoch=" << e
<< dendl
;
977 osd_epoch_barrier
= e
;
980 void MDSRank::retry_dispatch(Message
*m
)
982 inc_dispatch_depth();
984 dec_dispatch_depth();
987 utime_t
MDSRank::get_laggy_until() const
989 return beacon
.get_laggy_until();
992 bool MDSRank::is_daemon_stopping() const
997 void MDSRank::request_state(MDSMap::DaemonState s
)
999 dout(3) << "request_state " << ceph_mds_state_name(s
) << dendl
;
1000 beacon
.set_want_state(mdsmap
, s
);
1005 class C_MDS_BootStart
: public MDSInternalContext
{
1006 MDSRank::BootStep nextstep
;
1008 C_MDS_BootStart(MDSRank
*m
, MDSRank::BootStep n
)
1009 : MDSInternalContext(m
), nextstep(n
) {}
1010 void finish(int r
) override
{
1011 mds
->boot_start(nextstep
, r
);
1016 void MDSRank::boot_start(BootStep step
, int r
)
1018 // Handle errors from previous step
1020 if (is_standby_replay() && (r
== -EAGAIN
)) {
1021 dout(0) << "boot_start encountered an error EAGAIN"
1022 << ", respawning since we fell behind journal" << dendl
;
1024 } else if (r
== -EINVAL
|| r
== -ENOENT
) {
1025 // Invalid or absent data, indicates damaged on-disk structures
1026 clog
->error() << "Error loading MDS rank " << whoami
<< ": "
1029 assert(r
== 0); // Unreachable, damaged() calls respawn()
1031 // Completely unexpected error, give up and die
1032 dout(0) << "boot_start encountered an error, failing" << dendl
;
1038 assert(is_starting() || is_any_replay());
1041 case MDS_BOOT_INITIAL
:
1043 mdcache
->init_layouts();
1045 MDSGatherBuilder
gather(g_ceph_context
,
1046 new C_MDS_BootStart(this, MDS_BOOT_OPEN_ROOT
));
1047 dout(2) << "boot_start " << step
<< ": opening inotable" << dendl
;
1048 inotable
->set_rank(whoami
);
1049 inotable
->load(gather
.new_sub());
1051 dout(2) << "boot_start " << step
<< ": opening sessionmap" << dendl
;
1052 sessionmap
.set_rank(whoami
);
1053 sessionmap
.load(gather
.new_sub());
1055 dout(2) << "boot_start " << step
<< ": opening mds log" << dendl
;
1056 mdlog
->open(gather
.new_sub());
1058 if (is_starting()) {
1059 dout(2) << "boot_start " << step
<< ": opening purge queue" << dendl
;
1060 purge_queue
.open(new C_IO_Wrapper(this, gather
.new_sub()));
1061 } else if (!standby_replaying
) {
1062 dout(2) << "boot_start " << step
<< ": opening purge queue (async)" << dendl
;
1063 purge_queue
.open(NULL
);
1066 if (mdsmap
->get_tableserver() == whoami
) {
1067 dout(2) << "boot_start " << step
<< ": opening snap table" << dendl
;
1068 snapserver
->set_rank(whoami
);
1069 snapserver
->load(gather
.new_sub());
1075 case MDS_BOOT_OPEN_ROOT
:
1077 dout(2) << "boot_start " << step
<< ": loading/discovering base inodes" << dendl
;
1079 MDSGatherBuilder
gather(g_ceph_context
,
1080 new C_MDS_BootStart(this, MDS_BOOT_PREPARE_LOG
));
1082 mdcache
->open_mydir_inode(gather
.new_sub());
1084 if (is_starting() ||
1085 whoami
== mdsmap
->get_root()) { // load root inode off disk if we are auth
1086 mdcache
->open_root_inode(gather
.new_sub());
1088 // replay. make up fake root inode to start with
1089 (void)mdcache
->create_root_inode();
1094 case MDS_BOOT_PREPARE_LOG
:
1095 if (is_any_replay()) {
1096 dout(2) << "boot_start " << step
<< ": replaying mds log" << dendl
;
1097 MDSGatherBuilder
gather(g_ceph_context
,
1098 new C_MDS_BootStart(this, MDS_BOOT_REPLAY_DONE
));
1100 if (!standby_replaying
) {
1101 dout(2) << "boot_start " << step
<< ": waiting for purge queue recovered" << dendl
;
1102 purge_queue
.wait_for_recovery(new C_IO_Wrapper(this, gather
.new_sub()));
1105 mdlog
->replay(gather
.new_sub());
1108 dout(2) << "boot_start " << step
<< ": positioning at end of old mds log" << dendl
;
1113 case MDS_BOOT_REPLAY_DONE
:
1114 assert(is_any_replay());
1116 // Sessiontable and inotable should be in sync after replay, validate
1117 // that they are consistent.
1118 validate_sessions();
1125 void MDSRank::validate_sessions()
1127 assert(mds_lock
.is_locked_by_me());
1128 std::vector
<Session
*> victims
;
1130 // Identify any sessions which have state inconsistent with other,
1131 // after they have been loaded from rados during startup.
1132 // Mitigate bugs like: http://tracker.ceph.com/issues/16842
1133 const auto &sessions
= sessionmap
.get_sessions();
1134 for (const auto &i
: sessions
) {
1135 Session
*session
= i
.second
;
1136 interval_set
<inodeno_t
> badones
;
1137 if (inotable
->intersects_free(session
->info
.prealloc_inos
, &badones
)) {
1138 clog
->error() << "Client session loaded with invalid preallocated "
1139 "inodes, evicting session " << *session
;
1141 // Make the session consistent with inotable so that it can
1142 // be cleanly torn down
1143 session
->info
.prealloc_inos
.subtract(badones
);
1145 victims
.push_back(session
);
1149 for (const auto &session
: victims
) {
1150 server
->kill_session(session
, nullptr);
1154 void MDSRank::starting_done()
1156 dout(3) << "starting_done" << dendl
;
1157 assert(is_starting());
1158 request_state(MDSMap::STATE_ACTIVE
);
1160 mdcache
->open_root();
1162 if (mdcache
->is_open()) {
1163 mdlog
->start_new_segment();
1165 mdcache
->wait_for_open(new MDSInternalContextWrapper(this,
1166 new FunctionContext([this] (int r
) {
1167 mdlog
->start_new_segment();
1173 void MDSRank::calc_recovery_set()
1175 // initialize gather sets
1177 mdsmap
->get_recovery_mds_set(rs
);
1179 mdcache
->set_recovery_set(rs
);
1181 dout(1) << " recovery set is " << rs
<< dendl
;
1185 void MDSRank::replay_start()
1187 dout(1) << "replay_start" << dendl
;
1189 if (is_standby_replay())
1190 standby_replaying
= true;
1192 calc_recovery_set();
1194 // Check if we need to wait for a newer OSD map before starting
1195 Context
*fin
= new C_IO_Wrapper(this, new C_MDS_BootStart(this, MDS_BOOT_INITIAL
));
1196 bool const ready
= objecter
->wait_for_map(
1197 mdsmap
->get_last_failure_osd_epoch(),
1204 dout(1) << " waiting for osdmap " << mdsmap
->get_last_failure_osd_epoch()
1205 << " (which blacklists prior instance)" << dendl
;
1210 class MDSRank::C_MDS_StandbyReplayRestartFinish
: public MDSIOContext
{
1211 uint64_t old_read_pos
;
1213 C_MDS_StandbyReplayRestartFinish(MDSRank
*mds_
, uint64_t old_read_pos_
) :
1214 MDSIOContext(mds_
), old_read_pos(old_read_pos_
) {}
1215 void finish(int r
) override
{
1216 mds
->_standby_replay_restart_finish(r
, old_read_pos
);
1220 void MDSRank::_standby_replay_restart_finish(int r
, uint64_t old_read_pos
)
1222 if (old_read_pos
< mdlog
->get_journaler()->get_trimmed_pos()) {
1223 dout(0) << "standby MDS fell behind active MDS journal's expire_pos, restarting" << dendl
;
1224 respawn(); /* we're too far back, and this is easier than
1225 trying to reset everything in the cache, etc */
1227 mdlog
->standby_trim_segments();
1228 boot_start(MDS_BOOT_PREPARE_LOG
, r
);
1232 class MDSRank::C_MDS_StandbyReplayRestart
: public MDSInternalContext
{
1234 explicit C_MDS_StandbyReplayRestart(MDSRank
*m
) : MDSInternalContext(m
) {}
1235 void finish(int r
) override
{
1237 mds
->standby_replay_restart();
1241 void MDSRank::standby_replay_restart()
1243 if (standby_replaying
) {
1244 /* Go around for another pass of replaying in standby */
1245 dout(4) << "standby_replay_restart (as standby)" << dendl
;
1246 mdlog
->get_journaler()->reread_head_and_probe(
1247 new C_MDS_StandbyReplayRestartFinish(
1249 mdlog
->get_journaler()->get_read_pos()));
1251 /* We are transitioning out of standby: wait for OSD map update
1252 before making final pass */
1253 dout(1) << "standby_replay_restart (final takeover pass)" << dendl
;
1254 Context
*fin
= new C_IO_Wrapper(this, new C_MDS_StandbyReplayRestart(this));
1255 bool ready
= objecter
->wait_for_map(mdsmap
->get_last_failure_osd_epoch(), fin
);
1258 mdlog
->get_journaler()->reread_head_and_probe(
1259 new C_MDS_StandbyReplayRestartFinish(
1261 mdlog
->get_journaler()->get_read_pos()));
1263 dout(1) << " opening purge queue (async)" << dendl
;
1264 purge_queue
.open(NULL
);
1266 dout(1) << " waiting for osdmap " << mdsmap
->get_last_failure_osd_epoch()
1267 << " (which blacklists prior instance)" << dendl
;
1272 void MDSRank::replay_done()
1274 dout(1) << "replay_done" << (standby_replaying
? " (as standby)" : "") << dendl
;
1276 if (is_standby_replay()) {
1277 // The replay was done in standby state, and we are still in that state
1278 assert(standby_replaying
);
1279 dout(10) << "setting replay timer" << dendl
;
1280 timer
.add_event_after(g_conf
->mds_replay_interval
,
1281 new C_MDS_StandbyReplayRestart(this));
1283 } else if (standby_replaying
) {
1284 // The replay was done in standby state, we have now _left_ that state
1285 dout(10) << " last replay pass was as a standby; making final pass" << dendl
;
1286 standby_replaying
= false;
1287 standby_replay_restart();
1290 // Replay is complete, journal read should be up to date
1291 assert(mdlog
->get_journaler()->get_read_pos() == mdlog
->get_journaler()->get_write_pos());
1292 assert(!is_standby_replay());
1294 // Reformat and come back here
1295 if (mdlog
->get_journaler()->get_stream_format() < g_conf
->mds_journal_format
) {
1296 dout(4) << "reformatting journal on standbyreplay->replay transition" << dendl
;
1297 mdlog
->reopen(new C_MDS_BootStart(this, MDS_BOOT_REPLAY_DONE
));
1302 dout(1) << "making mds journal writeable" << dendl
;
1303 mdlog
->get_journaler()->set_writeable();
1304 mdlog
->get_journaler()->trim_tail();
1306 if (g_conf
->mds_wipe_sessions
) {
1307 dout(1) << "wiping out client sessions" << dendl
;
1309 sessionmap
.save(new C_MDSInternalNoop
);
1311 if (g_conf
->mds_wipe_ino_prealloc
) {
1312 dout(1) << "wiping out ino prealloc from sessions" << dendl
;
1313 sessionmap
.wipe_ino_prealloc();
1314 sessionmap
.save(new C_MDSInternalNoop
);
1316 if (g_conf
->mds_skip_ino
) {
1317 inodeno_t i
= g_conf
->mds_skip_ino
;
1318 dout(1) << "skipping " << i
<< " inodes" << dendl
;
1319 inotable
->skip_inos(i
);
1320 inotable
->save(new C_MDSInternalNoop
);
1323 if (mdsmap
->get_num_in_mds() == 1 &&
1324 mdsmap
->get_num_failed_mds() == 0) { // just me!
1325 dout(2) << "i am alone, moving to state reconnect" << dendl
;
1326 request_state(MDSMap::STATE_RECONNECT
);
1328 dout(2) << "i am not alone, moving to state resolve" << dendl
;
1329 request_state(MDSMap::STATE_RESOLVE
);
1333 void MDSRank::reopen_log()
1335 dout(1) << "reopen_log" << dendl
;
1336 mdcache
->rollback_uncommitted_fragments();
1340 void MDSRank::resolve_start()
1342 dout(1) << "resolve_start" << dendl
;
1346 mdcache
->resolve_start(new C_MDS_VoidFn(this, &MDSRank::resolve_done
));
1347 finish_contexts(g_ceph_context
, waiting_for_resolve
);
1349 void MDSRank::resolve_done()
1351 dout(1) << "resolve_done" << dendl
;
1352 request_state(MDSMap::STATE_RECONNECT
);
1355 void MDSRank::reconnect_start()
1357 dout(1) << "reconnect_start" << dendl
;
1359 if (last_state
== MDSMap::STATE_REPLAY
) {
1363 // Drop any blacklisted clients from the SessionMap before going
1364 // into reconnect, so that we don't wait for them.
1365 objecter
->enable_blacklist_events();
1366 std::set
<entity_addr_t
> blacklist
;
1368 objecter
->with_osdmap([this, &blacklist
, &epoch
](const OSDMap
& o
) {
1369 o
.get_blacklist(&blacklist
);
1370 epoch
= o
.get_epoch();
1372 auto killed
= server
->apply_blacklist(blacklist
);
1373 dout(4) << "reconnect_start: killed " << killed
<< " blacklisted sessions ("
1374 << blacklist
.size() << " blacklist entries, "
1375 << sessionmap
.get_sessions().size() << ")" << dendl
;
1377 set_osd_epoch_barrier(epoch
);
1380 server
->reconnect_clients(new C_MDS_VoidFn(this, &MDSRank::reconnect_done
));
1381 finish_contexts(g_ceph_context
, waiting_for_reconnect
);
1383 void MDSRank::reconnect_done()
1385 dout(1) << "reconnect_done" << dendl
;
1386 request_state(MDSMap::STATE_REJOIN
); // move to rejoin state
1389 void MDSRank::rejoin_joint_start()
1391 dout(1) << "rejoin_joint_start" << dendl
;
1392 mdcache
->rejoin_send_rejoins();
1394 void MDSRank::rejoin_start()
1396 dout(1) << "rejoin_start" << dendl
;
1397 mdcache
->rejoin_start(new C_MDS_VoidFn(this, &MDSRank::rejoin_done
));
1399 void MDSRank::rejoin_done()
1401 dout(1) << "rejoin_done" << dendl
;
1402 mdcache
->show_subtrees();
1403 mdcache
->show_cache();
1405 // funny case: is our cache empty? no subtrees?
1406 if (!mdcache
->is_subtrees()) {
1408 // The root should always have a subtree!
1409 clog
->error() << "No subtrees found for root MDS rank!";
1411 assert(mdcache
->is_subtrees());
1413 dout(1) << " empty cache, no subtrees, leaving cluster" << dendl
;
1414 request_state(MDSMap::STATE_STOPPED
);
1419 if (replay_queue
.empty())
1420 request_state(MDSMap::STATE_ACTIVE
);
1422 request_state(MDSMap::STATE_CLIENTREPLAY
);
1425 void MDSRank::clientreplay_start()
1427 dout(1) << "clientreplay_start" << dendl
;
1428 finish_contexts(g_ceph_context
, waiting_for_replay
); // kick waiters
1429 mdcache
->start_files_to_recover();
1433 bool MDSRank::queue_one_replay()
1435 if (replay_queue
.empty()) {
1436 mdlog
->wait_for_safe(new C_MDS_VoidFn(this, &MDSRank::clientreplay_done
));
1439 queue_waiter(replay_queue
.front());
1440 replay_queue
.pop_front();
1444 void MDSRank::clientreplay_done()
1446 dout(1) << "clientreplay_done" << dendl
;
1447 request_state(MDSMap::STATE_ACTIVE
);
1450 void MDSRank::active_start()
1452 dout(1) << "active_start" << dendl
;
1454 if (last_state
== MDSMap::STATE_CREATING
) {
1455 mdcache
->open_root();
1458 mdcache
->clean_open_file_lists();
1459 mdcache
->export_remaining_imported_caps();
1460 finish_contexts(g_ceph_context
, waiting_for_replay
); // kick waiters
1461 mdcache
->start_files_to_recover();
1463 mdcache
->reissue_all_caps();
1464 mdcache
->activate_stray_manager();
1466 finish_contexts(g_ceph_context
, waiting_for_active
); // kick waiters
1469 void MDSRank::recovery_done(int oldstate
)
1471 dout(1) << "recovery_done -- successful recovery!" << dendl
;
1472 assert(is_clientreplay() || is_active());
1474 // kick snaptable (resent AGREEs)
1475 if (mdsmap
->get_tableserver() == whoami
) {
1476 set
<mds_rank_t
> active
;
1477 mdsmap
->get_clientreplay_or_active_or_stopping_mds_set(active
);
1478 snapserver
->finish_recovery(active
);
1481 if (oldstate
== MDSMap::STATE_CREATING
)
1484 mdcache
->start_recovered_truncates();
1485 mdcache
->do_file_recover();
1487 // tell connected clients
1488 //bcast_mds_map(); // not anymore, they get this from the monitor
1490 mdcache
->populate_mydir();
1493 void MDSRank::creating_done()
1495 dout(1)<< "creating_done" << dendl
;
1496 request_state(MDSMap::STATE_ACTIVE
);
1499 void MDSRank::boot_create()
1501 dout(3) << "boot_create" << dendl
;
1503 MDSGatherBuilder
fin(g_ceph_context
, new C_MDS_VoidFn(this, &MDSRank::creating_done
));
1505 mdcache
->init_layouts();
1507 snapserver
->set_rank(whoami
);
1508 inotable
->set_rank(whoami
);
1509 sessionmap
.set_rank(whoami
);
1511 // start with a fresh journal
1512 dout(10) << "boot_create creating fresh journal" << dendl
;
1513 mdlog
->create(fin
.new_sub());
1515 // open new journal segment, but do not journal subtree map (yet)
1516 mdlog
->prepare_new_segment();
1518 if (whoami
== mdsmap
->get_root()) {
1519 dout(3) << "boot_create creating fresh hierarchy" << dendl
;
1520 mdcache
->create_empty_hierarchy(fin
.get());
1523 dout(3) << "boot_create creating mydir hierarchy" << dendl
;
1524 mdcache
->create_mydir_hierarchy(fin
.get());
1526 // fixme: fake out inotable (reset, pretend loaded)
1527 dout(10) << "boot_create creating fresh inotable table" << dendl
;
1529 inotable
->save(fin
.new_sub());
1531 // write empty sessionmap
1532 sessionmap
.save(fin
.new_sub());
1534 // Create empty purge queue
1535 purge_queue
.create(new C_IO_Wrapper(this, fin
.new_sub()));
1537 // initialize tables
1538 if (mdsmap
->get_tableserver() == whoami
) {
1539 dout(10) << "boot_create creating fresh snaptable" << dendl
;
1540 snapserver
->reset();
1541 snapserver
->save(fin
.new_sub());
1544 assert(g_conf
->mds_kill_create_at
!= 1);
1546 // ok now journal it
1547 mdlog
->journal_segment_subtree_map(fin
.new_sub());
1550 // Usually we do this during reconnect, but creation skips that.
1551 objecter
->enable_blacklist_events();
1556 void MDSRank::stopping_start()
1558 dout(2) << "stopping_start" << dendl
;
1560 if (mdsmap
->get_num_in_mds() == 1 && !sessionmap
.empty()) {
1561 // we're the only mds up!
1562 dout(0) << "we are the last MDS, and have mounted clients: we cannot flush our journal. suicide!" << dendl
;
1566 mdcache
->shutdown_start();
1569 void MDSRank::stopping_done()
1571 dout(2) << "stopping_done" << dendl
;
1573 // tell monitor we shut down cleanly.
1574 request_state(MDSMap::STATE_STOPPED
);
1577 void MDSRankDispatcher::handle_mds_map(
1581 // I am only to be passed MDSMaps in which I hold a rank
1582 assert(whoami
!= MDS_RANK_NONE
);
1584 MDSMap::DaemonState oldstate
= state
;
1585 mds_gid_t mds_gid
= mds_gid_t(monc
->get_global_id());
1586 state
= mdsmap
->get_state_gid(mds_gid
);
1587 if (state
!= oldstate
) {
1588 last_state
= oldstate
;
1589 incarnation
= mdsmap
->get_inc_gid(mds_gid
);
1592 version_t epoch
= m
->get_epoch();
1594 // note source's map version
1595 if (m
->get_source().is_mds() &&
1596 peer_mdsmap_epoch
[mds_rank_t(m
->get_source().num())] < epoch
) {
1597 dout(15) << " peer " << m
->get_source()
1598 << " has mdsmap epoch >= " << epoch
1600 peer_mdsmap_epoch
[mds_rank_t(m
->get_source().num())] = epoch
;
1603 // Validate state transitions while I hold a rank
1604 if (!MDSMap::state_transition_valid(oldstate
, state
)) {
1605 derr
<< "Invalid state transition " << ceph_mds_state_name(oldstate
)
1606 << "->" << ceph_mds_state_name(state
) << dendl
;
1610 if (oldstate
!= state
) {
1611 // update messenger.
1612 if (state
== MDSMap::STATE_STANDBY_REPLAY
) {
1613 dout(1) << "handle_mds_map i am now mds." << mds_gid
<< "." << incarnation
1614 << " replaying mds." << whoami
<< "." << incarnation
<< dendl
;
1615 messenger
->set_myname(entity_name_t::MDS(mds_gid
));
1617 dout(1) << "handle_mds_map i am now mds." << whoami
<< "." << incarnation
<< dendl
;
1618 messenger
->set_myname(entity_name_t::MDS(whoami
));
1622 // tell objecter my incarnation
1623 if (objecter
->get_client_incarnation() != incarnation
)
1624 objecter
->set_client_incarnation(incarnation
);
1627 if (g_conf
->mds_dump_cache_on_map
)
1628 mdcache
->dump_cache();
1631 if (oldstate
!= state
) {
1632 dout(1) << "handle_mds_map state change "
1633 << ceph_mds_state_name(oldstate
) << " --> "
1634 << ceph_mds_state_name(state
) << dendl
;
1635 beacon
.set_want_state(mdsmap
, state
);
1637 if (oldstate
== MDSMap::STATE_STANDBY_REPLAY
) {
1638 dout(10) << "Monitor activated us! Deactivating replay loop" << dendl
;
1639 assert (state
== MDSMap::STATE_REPLAY
);
1641 // did i just recover?
1642 if ((is_active() || is_clientreplay()) &&
1643 (oldstate
== MDSMap::STATE_CREATING
||
1644 oldstate
== MDSMap::STATE_REJOIN
||
1645 oldstate
== MDSMap::STATE_RECONNECT
))
1646 recovery_done(oldstate
);
1650 } else if (is_any_replay()) {
1652 } else if (is_resolve()) {
1654 } else if (is_reconnect()) {
1656 } else if (is_rejoin()) {
1658 } else if (is_clientreplay()) {
1659 clientreplay_start();
1660 } else if (is_creating()) {
1662 } else if (is_starting()) {
1664 } else if (is_stopping()) {
1665 assert(oldstate
== MDSMap::STATE_ACTIVE
);
1672 // is someone else newly resolving?
1673 if (is_resolve() || is_reconnect() || is_rejoin() ||
1674 is_clientreplay() || is_active() || is_stopping()) {
1675 if (!oldmap
->is_resolving() && mdsmap
->is_resolving()) {
1676 set
<mds_rank_t
> resolve
;
1677 mdsmap
->get_mds_set(resolve
, MDSMap::STATE_RESOLVE
);
1678 dout(10) << " resolve set is " << resolve
<< dendl
;
1679 calc_recovery_set();
1680 mdcache
->send_resolves();
1685 // is everybody finally rejoining?
1686 if (is_starting() || is_rejoin() || is_clientreplay() || is_active() || is_stopping()) {
1688 if (!oldmap
->is_rejoining() && mdsmap
->is_rejoining())
1689 rejoin_joint_start();
1692 if (g_conf
->mds_dump_cache_after_rejoin
&&
1693 oldmap
->is_rejoining() && !mdsmap
->is_rejoining())
1694 mdcache
->dump_cache(); // for DEBUG only
1696 if (oldstate
>= MDSMap::STATE_REJOIN
||
1697 oldstate
== MDSMap::STATE_STARTING
) {
1698 // ACTIVE|CLIENTREPLAY|REJOIN => we can discover from them.
1699 set
<mds_rank_t
> olddis
, dis
;
1700 oldmap
->get_mds_set(olddis
, MDSMap::STATE_ACTIVE
);
1701 oldmap
->get_mds_set(olddis
, MDSMap::STATE_CLIENTREPLAY
);
1702 oldmap
->get_mds_set(olddis
, MDSMap::STATE_REJOIN
);
1703 mdsmap
->get_mds_set(dis
, MDSMap::STATE_ACTIVE
);
1704 mdsmap
->get_mds_set(dis
, MDSMap::STATE_CLIENTREPLAY
);
1705 mdsmap
->get_mds_set(dis
, MDSMap::STATE_REJOIN
);
1706 for (set
<mds_rank_t
>::iterator p
= dis
.begin(); p
!= dis
.end(); ++p
)
1707 if (*p
!= whoami
&& // not me
1708 olddis
.count(*p
) == 0) { // newly so?
1709 mdcache
->kick_discovers(*p
);
1710 mdcache
->kick_open_ino_peers(*p
);
1715 cluster_degraded
= mdsmap
->is_degraded();
1716 if (oldmap
->is_degraded() && !cluster_degraded
&& state
>= MDSMap::STATE_ACTIVE
) {
1717 dout(1) << "cluster recovered." << dendl
;
1718 auto it
= waiting_for_active_peer
.find(MDS_RANK_NONE
);
1719 if (it
!= waiting_for_active_peer
.end()) {
1720 queue_waiters(it
->second
);
1721 waiting_for_active_peer
.erase(it
);
1725 // did someone go active?
1726 if (oldstate
>= MDSMap::STATE_CLIENTREPLAY
&&
1727 (is_clientreplay() || is_active() || is_stopping())) {
1728 set
<mds_rank_t
> oldactive
, active
;
1729 oldmap
->get_mds_set(oldactive
, MDSMap::STATE_ACTIVE
);
1730 oldmap
->get_mds_set(oldactive
, MDSMap::STATE_CLIENTREPLAY
);
1731 mdsmap
->get_mds_set(active
, MDSMap::STATE_ACTIVE
);
1732 mdsmap
->get_mds_set(active
, MDSMap::STATE_CLIENTREPLAY
);
1733 for (set
<mds_rank_t
>::iterator p
= active
.begin(); p
!= active
.end(); ++p
)
1734 if (*p
!= whoami
&& // not me
1735 oldactive
.count(*p
) == 0) // newly so?
1736 handle_mds_recovery(*p
);
1739 // did someone fail?
1742 set
<mds_rank_t
> olddown
, down
;
1743 oldmap
->get_down_mds_set(&olddown
);
1744 mdsmap
->get_down_mds_set(&down
);
1745 for (set
<mds_rank_t
>::iterator p
= down
.begin(); p
!= down
.end(); ++p
) {
1746 if (oldmap
->have_inst(*p
) && olddown
.count(*p
) == 0) {
1747 messenger
->mark_down(oldmap
->get_inst(*p
).addr
);
1748 handle_mds_failure(*p
);
1753 // did someone fail?
1754 // did their addr/inst change?
1757 mdsmap
->get_up_mds_set(up
);
1758 for (set
<mds_rank_t
>::iterator p
= up
.begin(); p
!= up
.end(); ++p
) {
1759 if (oldmap
->have_inst(*p
) &&
1760 oldmap
->get_inst(*p
) != mdsmap
->get_inst(*p
)) {
1761 messenger
->mark_down(oldmap
->get_inst(*p
).addr
);
1762 handle_mds_failure(*p
);
1767 if (is_clientreplay() || is_active() || is_stopping()) {
1769 set
<mds_rank_t
> oldstopped
, stopped
;
1770 oldmap
->get_stopped_mds_set(oldstopped
);
1771 mdsmap
->get_stopped_mds_set(stopped
);
1772 for (set
<mds_rank_t
>::iterator p
= stopped
.begin(); p
!= stopped
.end(); ++p
)
1773 if (oldstopped
.count(*p
) == 0) // newly so?
1774 mdcache
->migrator
->handle_mds_failure_or_stop(*p
);
1778 map
<epoch_t
,list
<MDSInternalContextBase
*> >::iterator p
= waiting_for_mdsmap
.begin();
1779 while (p
!= waiting_for_mdsmap
.end() && p
->first
<= mdsmap
->get_epoch()) {
1780 list
<MDSInternalContextBase
*> ls
;
1782 waiting_for_mdsmap
.erase(p
++);
1783 finish_contexts(g_ceph_context
, ls
);
1788 // Before going active, set OSD epoch barrier to latest (so that
1789 // we don't risk handing out caps to clients with old OSD maps that
1790 // might not include barriers from the previous incarnation of this MDS)
1791 set_osd_epoch_barrier(objecter
->with_osdmap(
1792 std::mem_fn(&OSDMap::get_epoch
)));
1797 MDSMap::mds_info_t info
= mdsmap
->get_info(whoami
);
1799 for (map
<mds_gid_t
,MDSMap::mds_info_t
>::const_iterator p
= mdsmap
->get_mds_info().begin();
1800 p
!= mdsmap
->get_mds_info().end();
1802 if (p
->second
.state
== MDSMap::STATE_STANDBY_REPLAY
&&
1803 (p
->second
.standby_for_rank
== whoami
||(info
.name
.length() && p
->second
.standby_for_name
== info
.name
))) {
1808 mdlog
->set_write_iohint(0);
1810 mdlog
->set_write_iohint(CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
);
1814 if (oldmap
->get_max_mds() != mdsmap
->get_max_mds()) {
1815 purge_queue
.update_op_limit(*mdsmap
);
1819 void MDSRank::handle_mds_recovery(mds_rank_t who
)
1821 dout(5) << "handle_mds_recovery mds." << who
<< dendl
;
1823 mdcache
->handle_mds_recovery(who
);
1825 if (mdsmap
->get_tableserver() == whoami
) {
1826 snapserver
->handle_mds_recovery(who
);
1829 queue_waiters(waiting_for_active_peer
[who
]);
1830 waiting_for_active_peer
.erase(who
);
1833 void MDSRank::handle_mds_failure(mds_rank_t who
)
1835 if (who
== whoami
) {
1836 dout(5) << "handle_mds_failure for myself; not doing anything" << dendl
;
1839 dout(5) << "handle_mds_failure mds." << who
<< dendl
;
1841 mdcache
->handle_mds_failure(who
);
1843 snapclient
->handle_mds_failure(who
);
1846 bool MDSRankDispatcher::handle_asok_command(
1847 std::string command
, cmdmap_t
& cmdmap
, Formatter
*f
,
1850 if (command
== "dump_ops_in_flight" ||
1852 if (!op_tracker
.dump_ops_in_flight(f
)) {
1853 ss
<< "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \
1854 please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards.";
1856 } else if (command
== "dump_blocked_ops") {
1857 if (!op_tracker
.dump_ops_in_flight(f
, true)) {
1858 ss
<< "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \
1859 Please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards.";
1861 } else if (command
== "dump_historic_ops") {
1862 if (!op_tracker
.dump_historic_ops(f
)) {
1863 ss
<< "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \
1864 please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards.";
1866 } else if (command
== "dump_historic_ops_by_duration") {
1867 if (!op_tracker
.dump_historic_ops(f
, true)) {
1868 ss
<< "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \
1869 please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards.";
1871 } else if (command
== "osdmap barrier") {
1872 int64_t target_epoch
= 0;
1873 bool got_val
= cmd_getval(g_ceph_context
, cmdmap
, "target_epoch", target_epoch
);
1876 ss
<< "no target epoch given";
1881 set_osd_epoch_barrier(target_epoch
);
1885 bool already_got
= objecter
->wait_for_map(target_epoch
, &cond
);
1887 dout(4) << __func__
<< ": waiting for OSD epoch " << target_epoch
<< dendl
;
1890 } else if (command
== "session ls") {
1891 Mutex::Locker
l(mds_lock
);
1895 dump_sessions(SessionFilter(), f
);
1896 } else if (command
== "session evict") {
1897 std::string client_id
;
1898 const bool got_arg
= cmd_getval(g_ceph_context
, cmdmap
, "client_id", client_id
);
1900 ss
<< "Invalid client_id specified";
1905 std::stringstream dss
;
1906 bool evicted
= evict_client(strtol(client_id
.c_str(), 0, 10), true,
1907 g_conf
->mds_session_blacklist_on_evict
, dss
);
1909 dout(15) << dss
.str() << dendl
;
1913 } else if (command
== "scrub_path") {
1915 vector
<string
> scrubop_vec
;
1916 cmd_getval(g_ceph_context
, cmdmap
, "scrubops", scrubop_vec
);
1917 cmd_getval(g_ceph_context
, cmdmap
, "path", path
);
1918 command_scrub_path(f
, path
, scrubop_vec
);
1919 } else if (command
== "tag path") {
1921 cmd_getval(g_ceph_context
, cmdmap
, "path", path
);
1923 cmd_getval(g_ceph_context
, cmdmap
, "tag", tag
);
1924 command_tag_path(f
, path
, tag
);
1925 } else if (command
== "flush_path") {
1927 cmd_getval(g_ceph_context
, cmdmap
, "path", path
);
1928 command_flush_path(f
, path
);
1929 } else if (command
== "flush journal") {
1930 command_flush_journal(f
);
1931 } else if (command
== "get subtrees") {
1932 command_get_subtrees(f
);
1933 } else if (command
== "export dir") {
1935 if(!cmd_getval(g_ceph_context
, cmdmap
, "path", path
)) {
1936 ss
<< "malformed path";
1940 if(!cmd_getval(g_ceph_context
, cmdmap
, "rank", rank
)) {
1941 ss
<< "malformed rank";
1944 command_export_dir(f
, path
, (mds_rank_t
)rank
);
1945 } else if (command
== "dump cache") {
1946 Mutex::Locker
l(mds_lock
);
1949 if(!cmd_getval(g_ceph_context
, cmdmap
, "path", path
)) {
1950 r
= mdcache
->dump_cache(f
);
1952 r
= mdcache
->dump_cache(path
);
1956 ss
<< "Failed to dump cache: " << cpp_strerror(r
);
1959 } else if (command
== "cache status") {
1960 Mutex::Locker
l(mds_lock
);
1961 int r
= mdcache
->cache_status(f
);
1963 ss
<< "Failed to get cache status: " << cpp_strerror(r
);
1965 } else if (command
== "dump tree") {
1968 cmd_getval(g_ceph_context
, cmdmap
, "root", root
);
1969 if (!cmd_getval(g_ceph_context
, cmdmap
, "depth", depth
))
1972 Mutex::Locker
l(mds_lock
);
1973 int r
= mdcache
->dump_cache(root
, depth
, f
);
1975 ss
<< "Failed to dump tree: " << cpp_strerror(r
);
1979 } else if (command
== "force_readonly") {
1980 Mutex::Locker
l(mds_lock
);
1981 mdcache
->force_readonly();
1982 } else if (command
== "dirfrag split") {
1983 command_dirfrag_split(cmdmap
, ss
);
1984 } else if (command
== "dirfrag merge") {
1985 command_dirfrag_merge(cmdmap
, ss
);
1986 } else if (command
== "dirfrag ls") {
1987 command_dirfrag_ls(cmdmap
, ss
, f
);
1995 class C_MDS_Send_Command_Reply
: public MDSInternalContext
2000 C_MDS_Send_Command_Reply(MDSRank
*_mds
, MCommand
*_m
) :
2001 MDSInternalContext(_mds
), m(_m
) { m
->get(); }
2002 void send (int r
, const std::string
& out_str
) {
2004 MDSDaemon::send_command_reply(m
, mds
, r
, bl
, out_str
);
2007 void finish (int r
) override
{
2013 * This function drops the mds_lock, so don't do anything with
2014 * MDSRank after calling it (we could have gone into shutdown): just
2015 * send your result back to the calling client and finish.
2017 void MDSRankDispatcher::evict_clients(const SessionFilter
&filter
, MCommand
*m
)
2019 C_MDS_Send_Command_Reply
*reply
= new C_MDS_Send_Command_Reply(this, m
);
2021 if (is_any_replay()) {
2022 reply
->send(-EAGAIN
, "MDS is replaying log");
2027 std::list
<Session
*> victims
;
2028 const auto sessions
= sessionmap
.get_sessions();
2029 for (const auto p
: sessions
) {
2030 if (!p
.first
.is_client()) {
2034 Session
*s
= p
.second
;
2036 if (filter
.match(*s
, std::bind(&Server::waiting_for_reconnect
, server
, std::placeholders::_1
))) {
2037 victims
.push_back(s
);
2041 dout(20) << __func__
<< " matched " << victims
.size() << " sessions" << dendl
;
2043 if (victims
.empty()) {
2049 C_GatherBuilder
gather(g_ceph_context
, reply
);
2050 for (const auto s
: victims
) {
2051 std::stringstream ss
;
2052 evict_client(s
->info
.inst
.name
.num(), false,
2053 g_conf
->mds_session_blacklist_on_evict
, ss
, gather
.new_sub());
2058 void MDSRankDispatcher::dump_sessions(const SessionFilter
&filter
, Formatter
*f
) const
2060 // Dump sessions, decorated with recovery/replay status
2061 f
->open_array_section("sessions");
2062 const ceph::unordered_map
<entity_name_t
, Session
*> session_map
= sessionmap
.get_sessions();
2063 for (ceph::unordered_map
<entity_name_t
,Session
*>::const_iterator p
= session_map
.begin();
2064 p
!= session_map
.end();
2066 if (!p
->first
.is_client()) {
2070 Session
*s
= p
->second
;
2072 if (!filter
.match(*s
, std::bind(&Server::waiting_for_reconnect
, server
, std::placeholders::_1
))) {
2076 f
->open_object_section("session");
2077 f
->dump_int("id", p
->first
.num());
2079 f
->dump_int("num_leases", s
->leases
.size());
2080 f
->dump_int("num_caps", s
->caps
.size());
2082 f
->dump_string("state", s
->get_state_name());
2083 f
->dump_int("replay_requests", is_clientreplay() ? s
->get_request_count() : 0);
2084 f
->dump_unsigned("completed_requests", s
->get_num_completed_requests());
2085 f
->dump_bool("reconnecting", server
->waiting_for_reconnect(p
->first
.num()));
2086 f
->dump_stream("inst") << s
->info
.inst
;
2087 f
->open_object_section("client_metadata");
2088 for (map
<string
, string
>::const_iterator i
= s
->info
.client_metadata
.begin();
2089 i
!= s
->info
.client_metadata
.end(); ++i
) {
2090 f
->dump_string(i
->first
.c_str(), i
->second
);
2092 f
->close_section(); // client_metadata
2093 f
->close_section(); //session
2095 f
->close_section(); //sessions
2098 void MDSRank::command_scrub_path(Formatter
*f
, const string
& path
, vector
<string
>& scrubop_vec
)
2101 bool recursive
= false;
2102 bool repair
= false;
2103 for (vector
<string
>::iterator i
= scrubop_vec
.begin() ; i
!= scrubop_vec
.end(); ++i
) {
2106 else if (*i
== "recursive")
2108 else if (*i
== "repair")
2113 Mutex::Locker
l(mds_lock
);
2114 mdcache
->enqueue_scrub(path
, "", force
, recursive
, repair
, f
, &scond
);
2117 // scrub_dentry() finishers will dump the data for us; we're done!
2120 void MDSRank::command_tag_path(Formatter
*f
,
2121 const string
& path
, const std::string
&tag
)
2125 Mutex::Locker
l(mds_lock
);
2126 mdcache
->enqueue_scrub(path
, tag
, true, true, false, f
, &scond
);
2131 void MDSRank::command_flush_path(Formatter
*f
, const string
& path
)
2135 Mutex::Locker
l(mds_lock
);
2136 mdcache
->flush_dentry(path
, &scond
);
2138 int r
= scond
.wait();
2139 f
->open_object_section("results");
2140 f
->dump_int("return_code", r
);
2141 f
->close_section(); // results
2145 * Wrapper around _command_flush_journal that
2146 * handles serialization of result
2148 void MDSRank::command_flush_journal(Formatter
*f
)
2152 std::stringstream ss
;
2153 const int r
= _command_flush_journal(&ss
);
2154 f
->open_object_section("result");
2155 f
->dump_string("message", ss
.str());
2156 f
->dump_int("return_code", r
);
2161 * Implementation of "flush journal" asok command.
2164 * Optionally populate with a human readable string describing the
2165 * reason for any unexpected return status.
2167 int MDSRank::_command_flush_journal(std::stringstream
*ss
)
2171 Mutex::Locker
l(mds_lock
);
2173 if (mdcache
->is_readonly()) {
2174 dout(5) << __func__
<< ": read-only FS" << dendl
;
2179 dout(5) << __func__
<< ": MDS not active, no-op" << dendl
;
2183 // I need to seal off the current segment, and then mark all previous segments
2185 mdlog
->start_new_segment();
2188 // Flush initially so that all the segments older than our new one
2189 // will be elegible for expiry
2191 C_SaferCond mdlog_flushed
;
2193 mdlog
->wait_for_safe(new MDSInternalContextWrapper(this, &mdlog_flushed
));
2195 r
= mdlog_flushed
.wait();
2198 *ss
<< "Error " << r
<< " (" << cpp_strerror(r
) << ") while flushing journal";
2203 // Because we may not be the last wait_for_safe context on MDLog, and
2204 // subsequent contexts might wake up in the middle of our later trim_all
2205 // and interfere with expiry (by e.g. marking dirs/dentries dirty
2206 // on previous log segments), we run a second wait_for_safe here.
2209 C_SaferCond mdlog_cleared
;
2210 mdlog
->wait_for_safe(new MDSInternalContextWrapper(this, &mdlog_cleared
));
2212 r
= mdlog_cleared
.wait();
2215 *ss
<< "Error " << r
<< " (" << cpp_strerror(r
) << ") while flushing journal";
2220 // Put all the old log segments into expiring or expired state
2221 dout(5) << __func__
<< ": beginning segment expiry" << dendl
;
2222 r
= mdlog
->trim_all();
2224 *ss
<< "Error " << r
<< " (" << cpp_strerror(r
) << ") while trimming log";
2228 // Attach contexts to wait for all expiring segments to expire
2229 MDSGatherBuilder
expiry_gather(g_ceph_context
);
2231 const std::set
<LogSegment
*> &expiring_segments
= mdlog
->get_expiring_segments();
2232 for (std::set
<LogSegment
*>::const_iterator i
= expiring_segments
.begin();
2233 i
!= expiring_segments
.end(); ++i
) {
2234 (*i
)->wait_for_expiry(expiry_gather
.new_sub());
2236 dout(5) << __func__
<< ": waiting for " << expiry_gather
.num_subs_created()
2237 << " segments to expire" << dendl
;
2239 if (expiry_gather
.has_subs()) {
2241 expiry_gather
.set_finisher(new MDSInternalContextWrapper(this, &cond
));
2242 expiry_gather
.activate();
2244 // Drop mds_lock to allow progress until expiry is complete
2246 int r
= cond
.wait();
2249 assert(r
== 0); // MDLog is not allowed to raise errors via wait_for_expiry
2252 dout(5) << __func__
<< ": expiry complete, expire_pos/trim_pos is now " << std::hex
<<
2253 mdlog
->get_journaler()->get_expire_pos() << "/" <<
2254 mdlog
->get_journaler()->get_trimmed_pos() << dendl
;
2256 // Now everyone I'm interested in is expired
2257 mdlog
->trim_expired_segments();
2259 dout(5) << __func__
<< ": trim complete, expire_pos/trim_pos is now " << std::hex
<<
2260 mdlog
->get_journaler()->get_expire_pos() << "/" <<
2261 mdlog
->get_journaler()->get_trimmed_pos() << dendl
;
2263 // Flush the journal header so that readers will start from after the flushed region
2264 C_SaferCond wrote_head
;
2265 mdlog
->get_journaler()->write_head(&wrote_head
);
2266 mds_lock
.Unlock(); // Drop lock to allow messenger dispatch progress
2267 r
= wrote_head
.wait();
2270 *ss
<< "Error " << r
<< " (" << cpp_strerror(r
) << ") while writing header";
2274 dout(5) << __func__
<< ": write_head complete, all done!" << dendl
;
2280 void MDSRank::command_get_subtrees(Formatter
*f
)
2283 Mutex::Locker
l(mds_lock
);
2285 std::list
<CDir
*> subtrees
;
2286 mdcache
->list_subtrees(subtrees
);
2288 f
->open_array_section("subtrees");
2289 for (std::list
<CDir
*>::iterator i
= subtrees
.begin(); i
!= subtrees
.end(); ++i
) {
2290 const CDir
*dir
= *i
;
2292 f
->open_object_section("subtree");
2294 f
->dump_bool("is_auth", dir
->is_auth());
2295 f
->dump_int("auth_first", dir
->get_dir_auth().first
);
2296 f
->dump_int("auth_second", dir
->get_dir_auth().second
);
2297 f
->dump_int("export_pin", dir
->inode
->get_export_pin());
2298 f
->open_object_section("dir");
2308 void MDSRank::command_export_dir(Formatter
*f
,
2309 const std::string
&path
,
2312 int r
= _command_export_dir(path
, target
);
2313 f
->open_object_section("results");
2314 f
->dump_int("return_code", r
);
2315 f
->close_section(); // results
2318 int MDSRank::_command_export_dir(
2319 const std::string
&path
,
2322 Mutex::Locker
l(mds_lock
);
2323 filepath
fp(path
.c_str());
2325 if (target
== whoami
|| !mdsmap
->is_up(target
) || !mdsmap
->is_in(target
)) {
2326 derr
<< "bad MDS target " << target
<< dendl
;
2330 CInode
*in
= mdcache
->cache_traverse(fp
);
2332 derr
<< "Bath path '" << path
<< "'" << dendl
;
2335 CDir
*dir
= in
->get_dirfrag(frag_t());
2336 if (!dir
|| !(dir
->is_auth())) {
2337 derr
<< "bad export_dir path dirfrag frag_t() or dir not auth" << dendl
;
2341 mdcache
->migrator
->export_dir(dir
, target
);
2345 CDir
*MDSRank::_command_dirfrag_get(
2346 const cmdmap_t
&cmdmap
,
2350 bool got
= cmd_getval(g_ceph_context
, cmdmap
, "path", path
);
2352 ss
<< "missing path argument";
2356 std::string frag_str
;
2357 if (!cmd_getval(g_ceph_context
, cmdmap
, "frag", frag_str
)) {
2358 ss
<< "missing frag argument";
2362 CInode
*in
= mdcache
->cache_traverse(filepath(path
.c_str()));
2364 // TODO really we should load something in if it's not in cache,
2365 // but the infrastructure is harder, and we might still be unable
2366 // to act on it if someone else is auth.
2367 ss
<< "directory '" << path
<< "' inode not in cache";
2373 if (!fg
.parse(frag_str
.c_str())) {
2374 ss
<< "frag " << frag_str
<< " failed to parse";
2378 CDir
*dir
= in
->get_dirfrag(fg
);
2380 ss
<< "frag 0x" << std::hex
<< in
->ino() << "/" << fg
<< " not in cache ("
2381 "use `dirfrag ls` to see if it should exist)";
2385 if (!dir
->is_auth()) {
2386 ss
<< "frag " << dir
->dirfrag() << " not auth (auth = "
2387 << dir
->authority() << ")";
2394 bool MDSRank::command_dirfrag_split(
2398 Mutex::Locker
l(mds_lock
);
2399 if (!mdsmap
->allows_dirfrags()) {
2400 ss
<< "dirfrags are disallowed by the mds map!";
2405 if (!cmd_getval(g_ceph_context
, cmdmap
, "bits", by
)) {
2406 ss
<< "missing bits argument";
2411 ss
<< "must split by >0 bits";
2415 CDir
*dir
= _command_dirfrag_get(cmdmap
, ss
);
2420 mdcache
->split_dir(dir
, by
);
2425 bool MDSRank::command_dirfrag_merge(
2429 Mutex::Locker
l(mds_lock
);
2431 bool got
= cmd_getval(g_ceph_context
, cmdmap
, "path", path
);
2433 ss
<< "missing path argument";
2437 std::string frag_str
;
2438 if (!cmd_getval(g_ceph_context
, cmdmap
, "frag", frag_str
)) {
2439 ss
<< "missing frag argument";
2443 CInode
*in
= mdcache
->cache_traverse(filepath(path
.c_str()));
2445 ss
<< "directory '" << path
<< "' inode not in cache";
2450 if (!fg
.parse(frag_str
.c_str())) {
2451 ss
<< "frag " << frag_str
<< " failed to parse";
2455 mdcache
->merge_dir(in
, fg
);
2460 bool MDSRank::command_dirfrag_ls(
2465 Mutex::Locker
l(mds_lock
);
2467 bool got
= cmd_getval(g_ceph_context
, cmdmap
, "path", path
);
2469 ss
<< "missing path argument";
2473 CInode
*in
= mdcache
->cache_traverse(filepath(path
.c_str()));
2475 ss
<< "directory inode not in cache";
2479 f
->open_array_section("frags");
2480 std::list
<frag_t
> frags
;
2481 // NB using get_leaves_under instead of get_dirfrags to give
2482 // you the list of what dirfrags may exist, not which are in cache
2483 in
->dirfragtree
.get_leaves_under(frag_t(), frags
);
2484 for (std::list
<frag_t
>::iterator i
= frags
.begin();
2485 i
!= frags
.end(); ++i
) {
2486 f
->open_object_section("frag");
2487 f
->dump_int("value", i
->value());
2488 f
->dump_int("bits", i
->bits());
2489 std::ostringstream frag_str
;
2490 frag_str
<< std::hex
<< i
->value() << "/" << std::dec
<< i
->bits();
2491 f
->dump_string("str", frag_str
.str());
2499 void MDSRank::dump_status(Formatter
*f
) const
2501 if (state
== MDSMap::STATE_REPLAY
||
2502 state
== MDSMap::STATE_STANDBY_REPLAY
) {
2503 mdlog
->dump_replay_status(f
);
2504 } else if (state
== MDSMap::STATE_RESOLVE
) {
2505 mdcache
->dump_resolve_status(f
);
2506 } else if (state
== MDSMap::STATE_RECONNECT
) {
2507 server
->dump_reconnect_status(f
);
2508 } else if (state
== MDSMap::STATE_REJOIN
) {
2509 mdcache
->dump_rejoin_status(f
);
2510 } else if (state
== MDSMap::STATE_CLIENTREPLAY
) {
2511 dump_clientreplay_status(f
);
2515 void MDSRank::dump_clientreplay_status(Formatter
*f
) const
2517 f
->open_object_section("clientreplay_status");
2518 f
->dump_unsigned("clientreplay_queue", replay_queue
.size());
2519 f
->dump_unsigned("active_replay", mdcache
->get_num_client_requests());
2523 void MDSRankDispatcher::update_log_config()
2525 map
<string
,string
> log_to_monitors
;
2526 map
<string
,string
> log_to_syslog
;
2527 map
<string
,string
> log_channel
;
2528 map
<string
,string
> log_prio
;
2529 map
<string
,string
> log_to_graylog
;
2530 map
<string
,string
> log_to_graylog_host
;
2531 map
<string
,string
> log_to_graylog_port
;
2535 if (parse_log_client_options(g_ceph_context
, log_to_monitors
, log_to_syslog
,
2536 log_channel
, log_prio
, log_to_graylog
,
2537 log_to_graylog_host
, log_to_graylog_port
,
2539 clog
->update_config(log_to_monitors
, log_to_syslog
,
2540 log_channel
, log_prio
, log_to_graylog
,
2541 log_to_graylog_host
, log_to_graylog_port
,
2543 dout(10) << __func__
<< " log_to_monitors " << log_to_monitors
<< dendl
;
2546 void MDSRank::create_logger()
2548 dout(10) << "create_logger" << dendl
;
2550 PerfCountersBuilder
mds_plb(g_ceph_context
, "mds", l_mds_first
, l_mds_last
);
2552 mds_plb
.add_u64_counter(
2553 l_mds_request
, "request", "Requests", "req",
2554 PerfCountersBuilder::PRIO_CRITICAL
);
2555 mds_plb
.add_u64_counter(l_mds_reply
, "reply", "Replies");
2556 mds_plb
.add_time_avg(
2557 l_mds_reply_latency
, "reply_latency", "Reply latency", "rlat",
2558 PerfCountersBuilder::PRIO_CRITICAL
);
2559 mds_plb
.add_u64_counter(
2560 l_mds_forward
, "forward", "Forwarding request", "fwd",
2561 PerfCountersBuilder::PRIO_INTERESTING
);
2562 mds_plb
.add_u64_counter(l_mds_dir_fetch
, "dir_fetch", "Directory fetch");
2563 mds_plb
.add_u64_counter(l_mds_dir_commit
, "dir_commit", "Directory commit");
2564 mds_plb
.add_u64_counter(l_mds_dir_split
, "dir_split", "Directory split");
2565 mds_plb
.add_u64_counter(l_mds_dir_merge
, "dir_merge", "Directory merge");
2567 mds_plb
.add_u64(l_mds_inode_max
, "inode_max", "Max inodes, cache size");
2568 mds_plb
.add_u64(l_mds_inodes
, "inodes", "Inodes", "inos",
2569 PerfCountersBuilder::PRIO_CRITICAL
);
2570 mds_plb
.add_u64(l_mds_inodes_top
, "inodes_top", "Inodes on top");
2571 mds_plb
.add_u64(l_mds_inodes_bottom
, "inodes_bottom", "Inodes on bottom");
2573 l_mds_inodes_pin_tail
, "inodes_pin_tail", "Inodes on pin tail");
2574 mds_plb
.add_u64(l_mds_inodes_pinned
, "inodes_pinned", "Inodes pinned");
2575 mds_plb
.add_u64(l_mds_inodes_expired
, "inodes_expired", "Inodes expired");
2577 l_mds_inodes_with_caps
, "inodes_with_caps", "Inodes with capabilities");
2578 mds_plb
.add_u64(l_mds_caps
, "caps", "Capabilities", "caps",
2579 PerfCountersBuilder::PRIO_INTERESTING
);
2580 mds_plb
.add_u64(l_mds_subtrees
, "subtrees", "Subtrees");
2582 mds_plb
.add_u64_counter(l_mds_traverse
, "traverse", "Traverses");
2583 mds_plb
.add_u64_counter(l_mds_traverse_hit
, "traverse_hit", "Traverse hits");
2584 mds_plb
.add_u64_counter(l_mds_traverse_forward
, "traverse_forward",
2585 "Traverse forwards");
2586 mds_plb
.add_u64_counter(l_mds_traverse_discover
, "traverse_discover",
2587 "Traverse directory discovers");
2588 mds_plb
.add_u64_counter(l_mds_traverse_dir_fetch
, "traverse_dir_fetch",
2589 "Traverse incomplete directory content fetchings");
2590 mds_plb
.add_u64_counter(l_mds_traverse_remote_ino
, "traverse_remote_ino",
2591 "Traverse remote dentries");
2592 mds_plb
.add_u64_counter(l_mds_traverse_lock
, "traverse_lock",
2595 mds_plb
.add_u64(l_mds_load_cent
, "load_cent", "Load per cent");
2596 mds_plb
.add_u64(l_mds_dispatch_queue_len
, "q", "Dispatch queue length");
2598 mds_plb
.add_u64_counter(l_mds_exported
, "exported", "Exports");
2599 mds_plb
.add_u64_counter(
2600 l_mds_exported_inodes
, "exported_inodes", "Exported inodes", "exi",
2601 PerfCountersBuilder::PRIO_INTERESTING
);
2602 mds_plb
.add_u64_counter(l_mds_imported
, "imported", "Imports");
2603 mds_plb
.add_u64_counter(
2604 l_mds_imported_inodes
, "imported_inodes", "Imported inodes", "imi",
2605 PerfCountersBuilder::PRIO_INTERESTING
);
2606 logger
= mds_plb
.create_perf_counters();
2607 g_ceph_context
->get_perfcounters_collection()->add(logger
);
2611 PerfCountersBuilder
mdm_plb(g_ceph_context
, "mds_mem", l_mdm_first
, l_mdm_last
);
2612 mdm_plb
.add_u64(l_mdm_ino
, "ino", "Inodes", "ino",
2613 PerfCountersBuilder::PRIO_INTERESTING
);
2614 mdm_plb
.add_u64_counter(l_mdm_inoa
, "ino+", "Inodes opened");
2615 mdm_plb
.add_u64_counter(l_mdm_inos
, "ino-", "Inodes closed");
2616 mdm_plb
.add_u64(l_mdm_dir
, "dir", "Directories");
2617 mdm_plb
.add_u64_counter(l_mdm_dira
, "dir+", "Directories opened");
2618 mdm_plb
.add_u64_counter(l_mdm_dirs
, "dir-", "Directories closed");
2619 mdm_plb
.add_u64(l_mdm_dn
, "dn", "Dentries", "dn",
2620 PerfCountersBuilder::PRIO_INTERESTING
);
2621 mdm_plb
.add_u64_counter(l_mdm_dna
, "dn+", "Dentries opened");
2622 mdm_plb
.add_u64_counter(l_mdm_dns
, "dn-", "Dentries closed");
2623 mdm_plb
.add_u64(l_mdm_cap
, "cap", "Capabilities");
2624 mdm_plb
.add_u64_counter(l_mdm_capa
, "cap+", "Capabilities added");
2625 mdm_plb
.add_u64_counter(l_mdm_caps
, "cap-", "Capabilities removed");
2626 mdm_plb
.add_u64(l_mdm_rss
, "rss", "RSS");
2627 mdm_plb
.add_u64(l_mdm_heap
, "heap", "Heap size");
2628 mdm_plb
.add_u64(l_mdm_buf
, "buf", "Buffer size");
2629 mlogger
= mdm_plb
.create_perf_counters();
2630 g_ceph_context
->get_perfcounters_collection()->add(mlogger
);
2633 mdlog
->create_logger();
2634 server
->create_logger();
2635 purge_queue
.create_logger();
2636 sessionmap
.register_perfcounters();
2637 mdcache
->register_perfcounters();
2640 void MDSRank::check_ops_in_flight()
2642 vector
<string
> warnings
;
2644 if (op_tracker
.check_ops_in_flight(warnings
, &slow
)) {
2645 for (vector
<string
>::iterator i
= warnings
.begin();
2646 i
!= warnings
.end();
2652 // set mds slow request count
2653 mds_slow_req_count
= slow
;
2657 void MDSRankDispatcher::handle_osd_map()
2659 if (is_active() && snapserver
) {
2660 snapserver
->check_osd_map(true);
2663 server
->handle_osd_map();
2665 purge_queue
.update_op_limit(*mdsmap
);
2667 std::set
<entity_addr_t
> newly_blacklisted
;
2668 objecter
->consume_blacklist_events(&newly_blacklisted
);
2669 auto epoch
= objecter
->with_osdmap([](const OSDMap
&o
){return o
.get_epoch();});
2670 dout(4) << "handle_osd_map epoch " << epoch
<< ", "
2671 << newly_blacklisted
.size() << " new blacklist entries" << dendl
;
2672 auto victims
= server
->apply_blacklist(newly_blacklisted
);
2674 set_osd_epoch_barrier(epoch
);
2678 // By default the objecter only requests OSDMap updates on use,
2679 // we would like to always receive the latest maps in order to
2680 // apply policy based on the FULL flag.
2681 objecter
->maybe_request_map();
2684 bool MDSRank::evict_client(int64_t session_id
,
2685 bool wait
, bool blacklist
, std::stringstream
& err_ss
,
2688 assert(mds_lock
.is_locked_by_me());
2690 // Mutually exclusive args
2691 assert(!(wait
&& on_killed
!= nullptr));
2693 if (is_any_replay()) {
2694 err_ss
<< "MDS is replaying log";
2698 Session
*session
= sessionmap
.get_session(
2699 entity_name_t(CEPH_ENTITY_TYPE_CLIENT
, session_id
));
2701 err_ss
<< "session " << session_id
<< " not in sessionmap!";
2705 dout(4) << "Preparing blacklist command... (wait=" << wait
<< ")" << dendl
;
2707 ss
<< "{\"prefix\":\"osd blacklist\", \"blacklistop\":\"add\",";
2708 ss
<< "\"addr\":\"";
2709 ss
<< session
->info
.inst
.addr
;
2711 std::string tmp
= ss
.str();
2712 std::vector
<std::string
> cmd
= {tmp
};
2714 auto kill_mds_session
= [this, session_id
, on_killed
](){
2715 assert(mds_lock
.is_locked_by_me());
2716 Session
*session
= sessionmap
.get_session(
2717 entity_name_t(CEPH_ENTITY_TYPE_CLIENT
, session_id
));
2720 server
->kill_session(session
, on_killed
);
2722 C_SaferCond on_safe
;
2723 server
->kill_session(session
, &on_safe
);
2730 dout(1) << "session " << session_id
<< " was removed while we waited "
2731 "for blacklist" << dendl
;
2733 // Even though it wasn't us that removed it, kick our completion
2734 // as the session has been removed.
2736 on_killed
->complete(0);
2741 auto background_blacklist
= [this, session_id
, cmd
](std::function
<void ()> fn
){
2742 assert(mds_lock
.is_locked_by_me());
2744 Context
*on_blacklist_done
= new FunctionContext([this, session_id
, fn
](int r
) {
2745 objecter
->wait_for_latest_osdmap(
2747 new FunctionContext([this, session_id
, fn
](int r
) {
2748 Mutex::Locker
l(mds_lock
);
2749 auto epoch
= objecter
->with_osdmap([](const OSDMap
&o
){
2750 return o
.get_epoch();
2753 set_osd_epoch_barrier(epoch
);
2760 dout(4) << "Sending mon blacklist command: " << cmd
[0] << dendl
;
2761 monc
->start_mon_command(cmd
, {}, nullptr, nullptr, on_blacklist_done
);
2764 auto blocking_blacklist
= [this, cmd
, &err_ss
, background_blacklist
](){
2765 C_SaferCond inline_ctx
;
2766 background_blacklist([&inline_ctx
](){inline_ctx
.complete(0);});
2774 blocking_blacklist();
2777 // We dropped mds_lock, so check that session still exists
2778 session
= sessionmap
.get_session(entity_name_t(CEPH_ENTITY_TYPE_CLIENT
,
2781 dout(1) << "session " << session_id
<< " was removed while we waited "
2782 "for blacklist" << dendl
;
2788 background_blacklist(kill_mds_session
);
2797 void MDSRank::bcast_mds_map()
2799 dout(7) << "bcast_mds_map " << mdsmap
->get_epoch() << dendl
;
2801 // share the map with mounted clients
2802 set
<Session
*> clients
;
2803 sessionmap
.get_client_session_set(clients
);
2804 for (set
<Session
*>::const_iterator p
= clients
.begin();
2807 (*p
)->connection
->send_message(new MMDSMap(monc
->get_fsid(), mdsmap
));
2808 last_client_mdsmap_bcast
= mdsmap
->get_epoch();
2811 MDSRankDispatcher::MDSRankDispatcher(
2814 LogChannelRef
&clog_
,
2820 Context
*respawn_hook_
,
2821 Context
*suicide_hook_
)
2822 : MDSRank(whoami_
, mds_lock_
, clog_
, timer_
, beacon_
, mdsmap_
,
2823 msgr
, monc_
, respawn_hook_
, suicide_hook_
)
2826 bool MDSRankDispatcher::handle_command(
2827 const cmdmap_t
&cmdmap
,
2830 std::stringstream
*ds
,
2831 std::stringstream
*ss
,
2834 assert(r
!= nullptr);
2835 assert(ds
!= nullptr);
2836 assert(ss
!= nullptr);
2841 cmd_getval(g_ceph_context
, cmdmap
, "prefix", prefix
);
2843 if (prefix
== "session ls" || prefix
== "client ls") {
2844 std::vector
<std::string
> filter_args
;
2845 cmd_getval(g_ceph_context
, cmdmap
, "filters", filter_args
);
2847 SessionFilter filter
;
2848 *r
= filter
.parse(filter_args
, ss
);
2853 Formatter
*f
= new JSONFormatter(true);
2854 dump_sessions(filter
, f
);
2858 } else if (prefix
== "session evict" || prefix
== "client evict") {
2859 std::vector
<std::string
> filter_args
;
2860 cmd_getval(g_ceph_context
, cmdmap
, "filters", filter_args
);
2862 SessionFilter filter
;
2863 *r
= filter
.parse(filter_args
, ss
);
2868 evict_clients(filter
, m
);
2870 *need_reply
= false;
2872 } else if (prefix
== "damage ls") {
2873 Formatter
*f
= new JSONFormatter(true);
2874 damage_table
.dump(f
);
2878 } else if (prefix
== "damage rm") {
2879 damage_entry_id_t id
= 0;
2880 bool got
= cmd_getval(g_ceph_context
, cmdmap
, "damage_id", (int64_t&)id
);
2886 damage_table
.erase(id
);
2893 epoch_t
MDSRank::get_osd_epoch() const
2895 return objecter
->with_osdmap(std::mem_fn(&OSDMap::get_epoch
));