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>
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.
17 #include "include/compat.h"
18 #include "include/types.h"
19 #include "include/str_list.h"
21 #include "common/Clock.h"
22 #include "common/HeartbeatMap.h"
23 #include "common/Timer.h"
24 #include "common/backport14.h"
25 #include "common/ceph_argparse.h"
26 #include "common/config.h"
27 #include "common/entity_name.h"
28 #include "common/errno.h"
29 #include "common/perf_counters.h"
30 #include "common/signal.h"
31 #include "common/version.h"
33 #include "global/signal_handler.h"
35 #include "msg/Messenger.h"
36 #include "mon/MonClient.h"
38 #include "osdc/Objecter.h"
42 #include "MDSDaemon.h"
46 #include "SnapServer.h"
47 #include "SnapClient.h"
49 #include "events/ESession.h"
50 #include "events/ESubtreeMap.h"
52 #include "messages/MMDSMap.h"
54 #include "messages/MGenericMessage.h"
56 #include "messages/MMonCommand.h"
57 #include "messages/MCommand.h"
58 #include "messages/MCommandReply.h"
60 #include "auth/AuthAuthorizeHandler.h"
61 #include "auth/RotatingKeyRing.h"
62 #include "auth/KeyRing.h"
64 #include "perfglue/cpu_profiler.h"
65 #include "perfglue/heap_profiler.h"
67 #define dout_context g_ceph_context
68 #define dout_subsys ceph_subsys_mds
70 #define dout_prefix *_dout << "mds." << name << ' '
73 class MDSDaemon::C_MDS_Tick
: public Context
{
75 MDSDaemon
*mds_daemon
;
77 explicit C_MDS_Tick(MDSDaemon
*m
) : mds_daemon(m
) {}
78 void finish(int r
) override
{
79 assert(mds_daemon
->mds_lock
.is_locked_by_me());
85 MDSDaemon::MDSDaemon(const std::string
&n
, Messenger
*m
, MonClient
*mc
) :
87 mds_lock("MDSDaemon::mds_lock"),
89 timer(m
->cct
, mds_lock
),
90 beacon(m
->cct
, mc
, n
),
91 authorize_handler_cluster_registry(new AuthAuthorizeHandlerRegistry(m
->cct
,
92 m
->cct
->_conf
->auth_supported
.empty() ?
93 m
->cct
->_conf
->auth_cluster_required
:
94 m
->cct
->_conf
->auth_supported
)),
95 authorize_handler_service_registry(new AuthAuthorizeHandlerRegistry(m
->cct
,
96 m
->cct
->_conf
->auth_supported
.empty() ?
97 m
->cct
->_conf
->auth_service_required
:
98 m
->cct
->_conf
->auth_supported
)),
103 log_client(m
->cct
, messenger
, &mc
->monmap
, LogClient::NO_FLAGS
),
111 clog
= log_client
.create_channel();
113 monc
->set_messenger(messenger
);
118 MDSDaemon::~MDSDaemon() {
119 Mutex::Locker
lock(mds_lock
);
126 delete authorize_handler_service_registry
;
127 delete authorize_handler_cluster_registry
;
130 class MDSSocketHook
: public AdminSocketHook
{
133 explicit MDSSocketHook(MDSDaemon
*m
) : mds(m
) {}
134 bool call(std::string command
, cmdmap_t
& cmdmap
, std::string format
,
135 bufferlist
& out
) override
{
137 bool r
= mds
->asok_command(command
, cmdmap
, format
, ss
);
143 bool MDSDaemon::asok_command(string command
, cmdmap_t
& cmdmap
, string format
,
146 dout(1) << "asok_command: " << command
<< " (starting...)" << dendl
;
148 Formatter
*f
= Formatter::create(format
, "json-pretty", "json-pretty");
149 bool handled
= false;
150 if (command
== "status") {
154 if (mds_rank
== NULL
) {
155 dout(1) << "Can't run that command on an inactive MDS!" << dendl
;
156 f
->dump_string("error", "mds_not_active");
158 handled
= mds_rank
->handle_asok_command(command
, cmdmap
, f
, ss
);
164 dout(1) << "asok_command: " << command
<< " (complete)" << dendl
;
169 void MDSDaemon::dump_status(Formatter
*f
)
171 f
->open_object_section("status");
172 f
->dump_stream("cluster_fsid") << monc
->get_fsid();
174 f
->dump_int("whoami", mds_rank
->get_nodeid());
176 f
->dump_int("whoami", MDS_RANK_NONE
);
179 f
->dump_int("id", monc
->get_global_id());
180 f
->dump_string("want_state", ceph_mds_state_name(beacon
.get_want_state()));
181 f
->dump_string("state", ceph_mds_state_name(mdsmap
->get_state_gid(mds_gid_t(
182 monc
->get_global_id()))));
184 Mutex::Locker
l(mds_lock
);
185 mds_rank
->dump_status(f
);
188 f
->dump_unsigned("mdsmap_epoch", mdsmap
->get_epoch());
190 f
->dump_unsigned("osdmap_epoch", mds_rank
->get_osd_epoch());
191 f
->dump_unsigned("osdmap_epoch_barrier", mds_rank
->get_osd_epoch_barrier());
193 f
->dump_unsigned("osdmap_epoch", 0);
194 f
->dump_unsigned("osdmap_epoch_barrier", 0);
196 f
->close_section(); // status
199 void MDSDaemon::set_up_admin_socket()
202 AdminSocket
*admin_socket
= g_ceph_context
->get_admin_socket();
203 assert(asok_hook
== nullptr);
204 asok_hook
= new MDSSocketHook(this);
205 r
= admin_socket
->register_command("status", "status", asok_hook
,
206 "high-level status of MDS");
208 r
= admin_socket
->register_command("dump_ops_in_flight",
209 "dump_ops_in_flight", asok_hook
,
210 "show the ops currently in flight");
212 r
= admin_socket
->register_command("ops",
214 "show the ops currently in flight");
216 r
= admin_socket
->register_command("dump_blocked_ops", "dump_blocked_ops",
218 "show the blocked ops currently in flight");
220 r
= admin_socket
->register_command("dump_historic_ops", "dump_historic_ops",
222 "show slowest recent ops");
224 r
= admin_socket
->register_command("dump_historic_ops_by_duration", "dump_historic_ops_by_duration",
226 "show slowest recent ops, sorted by op duration");
228 r
= admin_socket
->register_command("scrub_path",
229 "scrub_path name=path,type=CephString "
230 "name=scrubops,type=CephChoices,"
231 "strings=force|recursive|repair,n=N,req=false",
233 "scrub an inode and output results");
235 r
= admin_socket
->register_command("tag path",
236 "tag path name=path,type=CephString"
237 " name=tag,type=CephString",
239 "Apply scrub tag recursively");
241 r
= admin_socket
->register_command("flush_path",
242 "flush_path name=path,type=CephString",
244 "flush an inode (and its dirfrags)");
246 r
= admin_socket
->register_command("export dir",
248 "name=path,type=CephString "
249 "name=rank,type=CephInt",
251 "migrate a subtree to named MDS");
253 r
= admin_socket
->register_command("dump cache",
254 "dump cache name=path,type=CephString,req=false",
256 "dump metadata cache (optionally to a file)");
258 r
= admin_socket
->register_command("cache status",
261 "show cache status");
263 r
= admin_socket
->register_command("dump tree",
265 "name=root,type=CephString,req=true "
266 "name=depth,type=CephInt,req=false ",
268 "dump metadata cache for subtree");
270 r
= admin_socket
->register_command("session evict",
271 "session evict name=client_id,type=CephString",
273 "Evict a CephFS client");
275 r
= admin_socket
->register_command("osdmap barrier",
276 "osdmap barrier name=target_epoch,type=CephInt",
278 "Wait until the MDS has this OSD map epoch");
280 r
= admin_socket
->register_command("session ls",
283 "Enumerate connected CephFS clients");
285 r
= admin_socket
->register_command("flush journal",
288 "Flush the journal to the backing store");
290 r
= admin_socket
->register_command("force_readonly",
293 "Force MDS to read-only mode");
295 r
= admin_socket
->register_command("get subtrees",
298 "Return the subtree map");
300 r
= admin_socket
->register_command("dirfrag split",
302 "name=path,type=CephString,req=true "
303 "name=frag,type=CephString,req=true "
304 "name=bits,type=CephInt,req=true ",
306 "Fragment directory by path");
308 r
= admin_socket
->register_command("dirfrag merge",
310 "name=path,type=CephString,req=true "
311 "name=frag,type=CephString,req=true",
313 "De-fragment directory by path");
315 r
= admin_socket
->register_command("dirfrag ls",
317 "name=path,type=CephString,req=true",
319 "List fragments in directory");
323 void MDSDaemon::clean_up_admin_socket()
325 AdminSocket
*admin_socket
= g_ceph_context
->get_admin_socket();
326 admin_socket
->unregister_command("status");
327 admin_socket
->unregister_command("dump_ops_in_flight");
328 admin_socket
->unregister_command("ops");
329 admin_socket
->unregister_command("dump_blocked_ops");
330 admin_socket
->unregister_command("dump_historic_ops");
331 admin_socket
->unregister_command("dump_historic_ops_by_duration");
332 admin_socket
->unregister_command("scrub_path");
333 admin_socket
->unregister_command("tag path");
334 admin_socket
->unregister_command("flush_path");
335 admin_socket
->unregister_command("export dir");
336 admin_socket
->unregister_command("dump cache");
337 admin_socket
->unregister_command("cache status");
338 admin_socket
->unregister_command("dump tree");
339 admin_socket
->unregister_command("session evict");
340 admin_socket
->unregister_command("osdmap barrier");
341 admin_socket
->unregister_command("session ls");
342 admin_socket
->unregister_command("flush journal");
343 admin_socket
->unregister_command("force_readonly");
344 admin_socket
->unregister_command("get subtrees");
345 admin_socket
->unregister_command("dirfrag split");
346 admin_socket
->unregister_command("dirfrag merge");
347 admin_socket
->unregister_command("dirfrag ls");
352 const char** MDSDaemon::get_tracked_conf_keys() const
354 static const char* KEYS
[] = {
355 "mds_op_complaint_time", "mds_op_log_threshold",
356 "mds_op_history_size", "mds_op_history_duration",
357 "mds_enable_op_tracker",
362 "clog_to_syslog_facility",
363 "clog_to_syslog_level",
366 "mds_max_purge_ops_per_pg",
367 "mds_max_purge_files",
369 "clog_to_graylog_host",
370 "clog_to_graylog_port",
378 void MDSDaemon::handle_conf_change(const struct md_config_t
*conf
,
379 const std::set
<std::string
> &changed
)
381 // We may be called within mds_lock (via `tell`) or outwith the
382 // lock (via admin socket `config set`), so handle either case.
383 const bool initially_locked
= mds_lock
.is_locked_by_me();
384 if (!initially_locked
) {
388 if (changed
.count("mds_op_complaint_time") ||
389 changed
.count("mds_op_log_threshold")) {
391 mds_rank
->op_tracker
.set_complaint_and_threshold(conf
->mds_op_complaint_time
,
392 conf
->mds_op_log_threshold
);
395 if (changed
.count("mds_op_history_size") ||
396 changed
.count("mds_op_history_duration")) {
398 mds_rank
->op_tracker
.set_history_size_and_duration(conf
->mds_op_history_size
,
399 conf
->mds_op_history_duration
);
402 if (changed
.count("mds_enable_op_tracker")) {
404 mds_rank
->op_tracker
.set_tracking(conf
->mds_enable_op_tracker
);
407 if (changed
.count("clog_to_monitors") ||
408 changed
.count("clog_to_syslog") ||
409 changed
.count("clog_to_syslog_level") ||
410 changed
.count("clog_to_syslog_facility") ||
411 changed
.count("clog_to_graylog") ||
412 changed
.count("clog_to_graylog_host") ||
413 changed
.count("clog_to_graylog_port") ||
414 changed
.count("host") ||
415 changed
.count("fsid")) {
417 mds_rank
->update_log_config();
421 if (!g_conf
->mds_log_pause
&& changed
.count("mds_log_pause")) {
423 mds_rank
->mdlog
->kick_submitter();
428 mds_rank
->handle_conf_change(conf
, changed
);
431 if (!initially_locked
) {
437 int MDSDaemon::init()
439 dout(10) << sizeof(MDSCacheObject
) << "\tMDSCacheObject" << dendl
;
440 dout(10) << sizeof(CInode
) << "\tCInode" << dendl
;
441 dout(10) << sizeof(elist
<void*>::item
) << "\t elist<>::item *7=" << 7*sizeof(elist
<void*>::item
) << dendl
;
442 dout(10) << sizeof(inode_t
) << "\t inode_t " << dendl
;
443 dout(10) << sizeof(nest_info_t
) << "\t nest_info_t " << dendl
;
444 dout(10) << sizeof(frag_info_t
) << "\t frag_info_t " << dendl
;
445 dout(10) << sizeof(SimpleLock
) << "\t SimpleLock *5=" << 5*sizeof(SimpleLock
) << dendl
;
446 dout(10) << sizeof(ScatterLock
) << "\t ScatterLock *3=" << 3*sizeof(ScatterLock
) << dendl
;
447 dout(10) << sizeof(CDentry
) << "\tCDentry" << dendl
;
448 dout(10) << sizeof(elist
<void*>::item
) << "\t elist<>::item" << dendl
;
449 dout(10) << sizeof(SimpleLock
) << "\t SimpleLock" << dendl
;
450 dout(10) << sizeof(CDir
) << "\tCDir " << dendl
;
451 dout(10) << sizeof(elist
<void*>::item
) << "\t elist<>::item *2=" << 2*sizeof(elist
<void*>::item
) << dendl
;
452 dout(10) << sizeof(fnode_t
) << "\t fnode_t " << dendl
;
453 dout(10) << sizeof(nest_info_t
) << "\t nest_info_t *2" << dendl
;
454 dout(10) << sizeof(frag_info_t
) << "\t frag_info_t *2" << dendl
;
455 dout(10) << sizeof(Capability
) << "\tCapability " << dendl
;
456 dout(10) << sizeof(xlist
<void*>::item
) << "\t xlist<>::item *2=" << 2*sizeof(xlist
<void*>::item
) << dendl
;
458 messenger
->add_dispatcher_tail(&beacon
);
459 messenger
->add_dispatcher_tail(this);
462 monc
->set_messenger(messenger
);
464 monc
->set_want_keys(CEPH_ENTITY_TYPE_MON
| CEPH_ENTITY_TYPE_OSD
|
465 CEPH_ENTITY_TYPE_MDS
| CEPH_ENTITY_TYPE_MGR
);
469 derr
<< "ERROR: failed to get monmap: " << cpp_strerror(-r
) << dendl
;
476 // tell monc about log_client so it will know about mon session resets
477 monc
->set_log_client(&log_client
);
479 r
= monc
->authenticate();
481 derr
<< "ERROR: failed to authenticate: " << cpp_strerror(-r
) << dendl
;
488 int rotating_auth_attempts
= 0;
489 while (monc
->wait_auth_rotating(30.0) < 0) {
490 if (++rotating_auth_attempts
<= g_conf
->max_rotating_auth_attempts
) {
491 derr
<< "unable to obtain rotating service keys; retrying" << dendl
;
494 derr
<< "ERROR: failed to refresh rotating keys, "
495 << "maximum retry time reached." << dendl
;
503 messenger
->add_dispatcher_head(&mgrc
);
506 if (beacon
.get_want_state() == CEPH_MDS_STATE_DNE
) {
507 dout(4) << __func__
<< ": terminated already, dropping out" << dendl
;
512 monc
->sub_want("mdsmap", 0, 0);
513 monc
->sub_want("mgrmap", 0, 0);
518 // Set up admin socket before taking mds_lock, so that ordering
519 // is consistent (later we take mds_lock within asok callbacks)
520 set_up_admin_socket();
521 g_conf
->add_observer(this);
523 if (beacon
.get_want_state() == MDSMap::STATE_DNE
) {
524 suicide(); // we could do something more graceful here
525 dout(4) << __func__
<< ": terminated already, dropping out" << dendl
;
533 messenger
->set_myname(entity_name_t::MDS(MDS_RANK_NONE
));
542 void MDSDaemon::reset_tick()
545 if (tick_event
) timer
.cancel_event(tick_event
);
548 tick_event
= new C_MDS_Tick(this);
549 timer
.add_event_after(g_conf
->mds_tick_interval
, tick_event
);
552 void MDSDaemon::tick()
557 // Call through to subsystems' tick functions
563 void MDSDaemon::send_command_reply(MCommand
*m
, MDSRank
*mds_rank
,
564 int r
, bufferlist outbl
,
565 const std::string
& outs
)
567 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
568 assert(session
!= NULL
);
569 // If someone is using a closed session for sending commands (e.g.
570 // the ceph CLI) then we should feel free to clean up this connection
571 // as soon as we've sent them a response.
572 const bool live_session
= mds_rank
&&
573 mds_rank
->sessionmap
.get_session(session
->info
.inst
.name
) != nullptr
574 && session
->get_state_seq() > 0;
577 // This session only existed to issue commands, so terminate it
578 // as soon as we can.
579 assert(session
->is_closed());
580 session
->connection
->mark_disposable();
584 MCommandReply
*reply
= new MCommandReply(r
, outs
);
585 reply
->set_tid(m
->get_tid());
586 reply
->set_data(outbl
);
587 m
->get_connection()->send_message(reply
);
590 /* This function DOES put the passed message before returning*/
591 void MDSDaemon::handle_command(MCommand
*m
)
593 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
594 assert(session
!= NULL
);
598 std::stringstream ss
;
601 Context
*run_after
= NULL
;
602 bool need_reply
= true;
604 if (!session
->auth_caps
.allow_all()) {
606 << ": received command from client without `tell` capability: "
607 << m
->get_connection()->peer_addr
<< dendl
;
609 ss
<< "permission denied";
611 } else if (m
->cmd
.empty()) {
613 ss
<< "no command given";
615 } else if (!cmdmap_from_json(m
->cmd
, &cmdmap
, ss
)) {
619 r
= _handle_command(cmdmap
, m
, &outbl
, &outs
, &run_after
, &need_reply
);
623 send_command_reply(m
, mds_rank
, r
, outbl
, outs
);
627 run_after
->complete(0);
642 #define COMMAND(parsesig, helptext, module, perm, availability) \
643 {parsesig, helptext, module, perm, availability},
645 COMMAND("injectargs " \
646 "name=injected_args,type=CephString,n=N",
647 "inject configuration arguments into running MDS",
648 "mds", "*", "cli,rest")
649 COMMAND("config set " \
650 "name=key,type=CephString name=value,type=CephString",
651 "Set a configuration option at runtime (not persistent)",
652 "mds", "*", "cli,rest")
654 "Terminate this MDS",
655 "mds", "*", "cli,rest")
658 "mds", "*", "cli,rest")
659 COMMAND("session kill " \
660 "name=session_id,type=CephInt",
661 "End a client session",
662 "mds", "*", "cli,rest")
663 COMMAND("cpu_profiler " \
664 "name=arg,type=CephChoices,strings=status|flush",
665 "run cpu profiling on daemon", "mds", "rw", "cli,rest")
666 COMMAND("session ls " \
667 "name=filters,type=CephString,n=N,req=false",
668 "List client sessions", "mds", "r", "cli,rest")
669 COMMAND("client ls " \
670 "name=filters,type=CephString,n=N,req=false",
671 "List client sessions", "mds", "r", "cli,rest")
672 COMMAND("session evict " \
673 "name=filters,type=CephString,n=N,req=false",
674 "Evict client session(s)", "mds", "rw", "cli,rest")
675 COMMAND("client evict " \
676 "name=filters,type=CephString,n=N,req=false",
677 "Evict client session(s)", "mds", "rw", "cli,rest")
679 "List detected metadata damage", "mds", "r", "cli,rest")
680 COMMAND("damage rm name=damage_id,type=CephInt",
681 "Remove a damage table entry", "mds", "rw", "cli,rest")
682 COMMAND("version", "report version of MDS", "mds", "r", "cli,rest")
684 "name=heapcmd,type=CephChoices,strings=dump|start_profiler|stop_profiler|release|stats", \
685 "show heap usage info (available only if compiled with tcmalloc)", \
686 "mds", "*", "cli,rest")
690 int MDSDaemon::_handle_command(
691 const cmdmap_t
&cmdmap
,
698 assert(outbl
!= NULL
);
699 assert(outs
!= NULL
);
701 class SuicideLater
: public Context
706 explicit SuicideLater(MDSDaemon
*mds_
) : mds(mds_
) {}
707 void finish(int r
) override
{
708 // Wait a little to improve chances of caller getting
709 // our response before seeing us disappear from mdsmap
717 class RespawnLater
: public Context
723 explicit RespawnLater(MDSDaemon
*mds_
) : mds(mds_
) {}
724 void finish(int r
) override
{
725 // Wait a little to improve chances of caller getting
726 // our response before seeing us disappear from mdsmap
733 std::stringstream ds
;
734 std::stringstream ss
;
737 std::unique_ptr
<Formatter
> f(Formatter::create(format
));
738 cmd_getval(cct
, cmdmap
, "prefix", prefix
);
742 if (prefix
== "get_command_descriptions") {
744 std::unique_ptr
<JSONFormatter
> f(ceph::make_unique
<JSONFormatter
>());
745 f
->open_object_section("command_descriptions");
746 for (MDSCommand
*cp
= mds_commands
;
747 cp
< &mds_commands
[ARRAY_SIZE(mds_commands
)]; cp
++) {
749 ostringstream secname
;
750 secname
<< "cmd" << setfill('0') << std::setw(3) << cmdnum
;
751 dump_cmddesc_to_json(f
.get(), secname
.str(), cp
->cmdstring
, cp
->helpstring
,
752 cp
->module
, cp
->perm
, cp
->availability
, 0);
755 f
->close_section(); // command_descriptions
761 cmd_getval(cct
, cmdmap
, "format", format
);
762 if (prefix
== "version") {
764 f
->open_object_section("version");
765 f
->dump_string("version", pretty_version_to_str());
769 ds
<< pretty_version_to_str();
771 } else if (prefix
== "injectargs") {
772 vector
<string
> argsvec
;
773 cmd_getval(cct
, cmdmap
, "injected_args", argsvec
);
775 if (argsvec
.empty()) {
777 ss
<< "ignoring empty injectargs";
780 string args
= argsvec
.front();
781 for (vector
<string
>::iterator a
= ++argsvec
.begin(); a
!= argsvec
.end(); ++a
)
783 r
= cct
->_conf
->injectargs(args
, &ss
);
784 } else if (prefix
== "config set") {
786 cmd_getval(cct
, cmdmap
, "key", key
);
788 cmd_getval(cct
, cmdmap
, "value", val
);
789 r
= cct
->_conf
->set_val(key
, val
, true, &ss
);
791 cct
->_conf
->apply_changes(nullptr);
793 } else if (prefix
== "exit") {
794 // We will send response before executing
796 *run_later
= new SuicideLater(this);
797 } else if (prefix
== "respawn") {
798 // We will send response before executing
799 ss
<< "Respawning...";
800 *run_later
= new RespawnLater(this);
801 } else if (prefix
== "session kill") {
802 if (mds_rank
== NULL
) {
804 ss
<< "MDS not active";
807 // FIXME harmonize `session kill` with admin socket session evict
808 int64_t session_id
= 0;
809 bool got
= cmd_getval(cct
, cmdmap
, "session_id", session_id
);
811 bool killed
= mds_rank
->evict_client(session_id
, false,
812 g_conf
->mds_session_blacklist_on_evict
,
816 } else if (prefix
== "heap") {
817 if (!ceph_using_tcmalloc()) {
819 ss
<< "could not issue heap profiler command -- not using tcmalloc!";
822 cmd_getval(cct
, cmdmap
, "heapcmd", heapcmd
);
823 vector
<string
> heapcmd_vec
;
824 get_str_vec(heapcmd
, heapcmd_vec
);
825 ceph_heap_profiler_handle_command(heapcmd_vec
, ds
);
827 } else if (prefix
== "cpu_profiler") {
829 cmd_getval(cct
, cmdmap
, "arg", arg
);
830 vector
<string
> argvec
;
831 get_str_vec(arg
, argvec
);
832 cpu_profiler_handle_command(argvec
, ds
);
834 // Give MDSRank a shot at the command
836 bool handled
= mds_rank
->handle_command(cmdmap
, m
, &r
, &ds
, &ss
,
843 // Neither MDSDaemon nor MDSRank know this command
844 std::ostringstream ss
;
845 ss
<< "unrecognized command! " << prefix
;
855 /* This function deletes the passed message before returning. */
857 void MDSDaemon::handle_mds_map(MMDSMap
*m
)
859 version_t epoch
= m
->get_epoch();
860 dout(5) << "handle_mds_map epoch " << epoch
<< " from " << m
->get_source() << dendl
;
863 if (epoch
<= mdsmap
->get_epoch()) {
864 dout(5) << " old map epoch " << epoch
<< " <= " << mdsmap
->get_epoch()
865 << ", discarding" << dendl
;
872 // keep old map, for a moment
873 MDSMap
*oldmap
= mdsmap
;
875 // decode and process
877 mdsmap
->decode(m
->get_encoded());
878 const MDSMap::DaemonState new_state
= mdsmap
->get_state_gid(mds_gid_t(monc
->get_global_id()));
879 const int incarnation
= mdsmap
->get_inc_gid(mds_gid_t(monc
->get_global_id()));
881 monc
->sub_got("mdsmap", mdsmap
->get_epoch());
883 // Calculate my effective rank (either my owned rank or my
884 // standby_for_rank if in standby replay)
885 mds_rank_t whoami
= mdsmap
->get_rank_gid(mds_gid_t(monc
->get_global_id()));
888 CompatSet
mdsmap_compat(get_mdsmap_compat_set_all());
889 dout(10) << " my compat " << mdsmap_compat
<< dendl
;
890 dout(10) << " mdsmap compat " << mdsmap
->compat
<< dendl
;
891 if (!mdsmap_compat
.writeable(mdsmap
->compat
)) {
892 dout(0) << "handle_mds_map mdsmap compatset " << mdsmap
->compat
893 << " not writeable with daemon features " << mdsmap_compat
894 << ", killing myself" << dendl
;
899 // mark down any failed peers
900 for (map
<mds_gid_t
,MDSMap::mds_info_t
>::const_iterator p
= oldmap
->get_mds_info().begin();
901 p
!= oldmap
->get_mds_info().end();
903 if (mdsmap
->get_mds_info().count(p
->first
) == 0) {
904 dout(10) << " peer mds gid " << p
->first
<< " removed from map" << dendl
;
905 messenger
->mark_down(p
->second
.addr
);
909 if (whoami
== MDS_RANK_NONE
&&
910 new_state
== MDSMap::STATE_STANDBY_REPLAY
) {
911 whoami
= mdsmap
->get_mds_info_gid(mds_gid_t(monc
->get_global_id())).standby_for_rank
;
915 addr
= messenger
->get_myaddr();
916 dout(10) << "map says I am " << addr
<< " mds." << whoami
<< "." << incarnation
917 << " state " << ceph_mds_state_name(new_state
) << dendl
;
919 if (whoami
== MDS_RANK_NONE
) {
920 if (mds_rank
!= NULL
) {
921 const auto myid
= monc
->get_global_id();
922 // We have entered a rank-holding state, we shouldn't be back
924 if (g_conf
->mds_enforce_unique_name
) {
925 if (mds_gid_t existing
= mdsmap
->find_mds_gid_by_name(name
)) {
926 const MDSMap::mds_info_t
& i
= mdsmap
->get_info_gid(existing
);
927 if (i
.global_id
> myid
) {
928 dout(1) << "map replaced me with another mds." << whoami
929 << " with gid (" << i
.global_id
<< ") larger than myself ("
930 << myid
<< "); quitting!" << dendl
;
931 // Call suicide() rather than respawn() because if someone else
932 // has taken our ID, we don't want to keep restarting and
933 // fighting them for the ID.
941 dout(1) << "map removed me (mds." << whoami
<< " gid:"
942 << myid
<< ") from cluster due to lost contact; respawning" << dendl
;
945 // MDSRank not active: process the map here to see if we have
946 // been assigned a rank.
947 dout(10) << __func__
<< ": handling map in rankless mode" << dendl
;
948 _handle_mds_map(oldmap
);
951 // Did we already hold a different rank? MDSMonitor shouldn't try
952 // to change that out from under me!
953 if (mds_rank
&& whoami
!= mds_rank
->get_nodeid()) {
954 derr
<< "Invalid rank transition " << mds_rank
->get_nodeid() << "->"
959 // Did I previously not hold a rank? Initialize!
960 if (mds_rank
== NULL
) {
961 mds_rank
= new MDSRankDispatcher(whoami
, mds_lock
, clog
,
962 timer
, beacon
, mdsmap
, messenger
, monc
,
963 new FunctionContext([this](int r
){respawn();}),
964 new FunctionContext([this](int r
){suicide();}));
965 dout(10) << __func__
<< ": initializing MDS rank "
966 << mds_rank
->get_nodeid() << dendl
;
970 // MDSRank is active: let him process the map, we have no say.
971 dout(10) << __func__
<< ": handling map as rank "
972 << mds_rank
->get_nodeid() << dendl
;
973 mds_rank
->handle_mds_map(m
, oldmap
);
977 beacon
.notify_mdsmap(mdsmap
);
982 void MDSDaemon::_handle_mds_map(MDSMap
*oldmap
)
984 MDSMap::DaemonState new_state
= mdsmap
->get_state_gid(mds_gid_t(monc
->get_global_id()));
986 // Normal rankless case, we're marked as standby
987 if (new_state
== MDSMap::STATE_STANDBY
) {
988 beacon
.set_want_state(mdsmap
, new_state
);
989 dout(1) << "handle_mds_map standby" << dendl
;
994 // Case where we thought we were standby, but MDSMap disagrees
995 if (beacon
.get_want_state() == MDSMap::STATE_STANDBY
) {
996 dout(10) << "dropped out of mdsmap, try to re-add myself" << dendl
;
997 new_state
= MDSMap::STATE_BOOT
;
998 beacon
.set_want_state(mdsmap
, new_state
);
1002 // Case where we have sent a boot beacon that isn't reflected yet
1003 if (beacon
.get_want_state() == MDSMap::STATE_BOOT
) {
1004 dout(10) << "not in map yet" << dendl
;
1008 void MDSDaemon::handle_signal(int signum
)
1010 assert(signum
== SIGINT
|| signum
== SIGTERM
);
1011 derr
<< "*** got signal " << sig_str(signum
) << " ***" << dendl
;
1013 Mutex::Locker
l(mds_lock
);
1021 void MDSDaemon::suicide()
1023 assert(mds_lock
.is_locked());
1025 // make sure we don't suicide twice
1026 assert(stopping
== false);
1029 dout(1) << "suicide. wanted state "
1030 << ceph_mds_state_name(beacon
.get_want_state()) << dendl
;
1033 timer
.cancel_event(tick_event
);
1037 //because add_observer is called after set_up_admin_socket
1038 //so we can use asok_hook to avoid assert in the remove_observer
1039 if (asok_hook
!= NULL
)
1040 g_conf
->remove_observer(this);
1042 clean_up_admin_socket();
1044 // Inform MDS we are going away, then shut down beacon
1045 beacon
.set_want_state(mdsmap
, MDSMap::STATE_DNE
);
1046 if (!mdsmap
->is_dne_gid(mds_gid_t(monc
->get_global_id()))) {
1047 // Notify the MDSMonitor that we're dying, so that it doesn't have to
1048 // wait for us to go laggy. Only do this if we're actually in the
1049 // MDSMap, because otherwise the MDSMonitor will drop our message.
1050 beacon
.send_and_wait(1);
1057 mds_rank
->shutdown();
1062 messenger
->shutdown();
1066 void MDSDaemon::respawn()
1068 dout(1) << "respawn" << dendl
;
1070 char *new_argv
[orig_argc
+1];
1071 dout(1) << " e: '" << orig_argv
[0] << "'" << dendl
;
1072 for (int i
=0; i
<orig_argc
; i
++) {
1073 new_argv
[i
] = (char *)orig_argv
[i
];
1074 dout(1) << " " << i
<< ": '" << orig_argv
[i
] << "'" << dendl
;
1076 new_argv
[orig_argc
] = NULL
;
1078 /* Determine the path to our executable, test if Linux /proc/self/exe exists.
1079 * This allows us to exec the same executable even if it has since been
1082 char exe_path
[PATH_MAX
] = "";
1083 if (readlink(PROCPREFIX
"/proc/self/exe", exe_path
, PATH_MAX
-1) == -1) {
1084 /* Print CWD for the user's interest */
1086 char *cwd
= getcwd(buf
, sizeof(buf
));
1088 dout(1) << " cwd " << cwd
<< dendl
;
1090 /* Fall back to a best-effort: just running in our CWD */
1091 strncpy(exe_path
, orig_argv
[0], PATH_MAX
-1);
1093 dout(1) << "respawning with exe " << exe_path
<< dendl
;
1094 strcpy(exe_path
, PROCPREFIX
"/proc/self/exe");
1097 dout(1) << " exe_path " << exe_path
<< dendl
;
1099 unblock_all_signals(NULL
);
1100 execv(exe_path
, new_argv
);
1102 dout(0) << "respawn execv " << orig_argv
[0]
1103 << " failed with " << cpp_strerror(errno
) << dendl
;
1105 // We have to assert out here, because suicide() returns, and callers
1106 // to respawn expect it never to return.
1112 bool MDSDaemon::ms_dispatch(Message
*m
)
1114 Mutex::Locker
l(mds_lock
);
1119 // Drop out early if shutting down
1120 if (beacon
.get_want_state() == CEPH_MDS_STATE_DNE
) {
1121 dout(10) << " stopping, discarding " << *m
<< dendl
;
1126 // First see if it's a daemon message
1127 const bool handled_core
= handle_core_message(m
);
1132 // Not core, try it as a rank message
1134 return mds_rank
->ms_dispatch(m
);
1140 bool MDSDaemon::ms_get_authorizer(int dest_type
, AuthAuthorizer
**authorizer
, bool force_new
)
1142 dout(10) << "MDSDaemon::ms_get_authorizer type="
1143 << ceph_entity_type_name(dest_type
) << dendl
;
1145 /* monitor authorization is being handled on different layer */
1146 if (dest_type
== CEPH_ENTITY_TYPE_MON
)
1150 if (monc
->wait_auth_rotating(10) < 0)
1154 *authorizer
= monc
->build_authorizer(dest_type
);
1155 return *authorizer
!= NULL
;
1160 * high priority messages we always process
1162 bool MDSDaemon::handle_core_message(Message
*m
)
1164 switch (m
->get_type()) {
1165 case CEPH_MSG_MON_MAP
:
1166 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MON
);
1171 case CEPH_MSG_MDS_MAP
:
1172 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MON
| CEPH_ENTITY_TYPE_MDS
);
1173 handle_mds_map(static_cast<MMDSMap
*>(m
));
1178 handle_command(static_cast<MCommand
*>(m
));
1180 case CEPH_MSG_OSD_MAP
:
1181 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MON
| CEPH_ENTITY_TYPE_OSD
);
1184 mds_rank
->handle_osd_map();
1189 case MSG_MON_COMMAND
:
1190 ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MON
);
1191 clog
->warn() << "dropping `mds tell` command from legacy monitor";
1201 void MDSDaemon::ms_handle_connect(Connection
*con
)
1205 bool MDSDaemon::ms_handle_reset(Connection
*con
)
1207 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_CLIENT
)
1210 Mutex::Locker
l(mds_lock
);
1214 dout(5) << "ms_handle_reset on " << con
->get_peer_addr() << dendl
;
1215 if (beacon
.get_want_state() == CEPH_MDS_STATE_DNE
)
1218 Session
*session
= static_cast<Session
*>(con
->get_priv());
1220 if (session
->is_closed()) {
1221 dout(3) << "ms_handle_reset closing connection for session " << session
->info
.inst
<< dendl
;
1223 con
->set_priv(NULL
);
1233 void MDSDaemon::ms_handle_remote_reset(Connection
*con
)
1235 if (con
->get_peer_type() != CEPH_ENTITY_TYPE_CLIENT
)
1238 Mutex::Locker
l(mds_lock
);
1243 dout(5) << "ms_handle_remote_reset on " << con
->get_peer_addr() << dendl
;
1244 if (beacon
.get_want_state() == CEPH_MDS_STATE_DNE
)
1247 Session
*session
= static_cast<Session
*>(con
->get_priv());
1249 if (session
->is_closed()) {
1250 dout(3) << "ms_handle_remote_reset closing connection for session " << session
->info
.inst
<< dendl
;
1252 con
->set_priv(NULL
);
1258 bool MDSDaemon::ms_handle_refused(Connection
*con
)
1260 // do nothing for now
1264 bool MDSDaemon::ms_verify_authorizer(Connection
*con
, int peer_type
,
1265 int protocol
, bufferlist
& authorizer_data
, bufferlist
& authorizer_reply
,
1266 bool& is_valid
, CryptoKey
& session_key
)
1268 Mutex::Locker
l(mds_lock
);
1272 if (beacon
.get_want_state() == CEPH_MDS_STATE_DNE
)
1275 AuthAuthorizeHandler
*authorize_handler
= 0;
1276 switch (peer_type
) {
1277 case CEPH_ENTITY_TYPE_MDS
:
1278 authorize_handler
= authorize_handler_cluster_registry
->get_handler(protocol
);
1281 authorize_handler
= authorize_handler_service_registry
->get_handler(protocol
);
1283 if (!authorize_handler
) {
1284 dout(0) << "No AuthAuthorizeHandler found for protocol " << protocol
<< dendl
;
1289 AuthCapsInfo caps_info
;
1293 RotatingKeyRing
*keys
= monc
->rotating_secrets
.get();
1295 is_valid
= authorize_handler
->verify_authorizer(
1297 authorizer_data
, authorizer_reply
, name
, global_id
, caps_info
,
1300 dout(10) << __func__
<< " no rotating_keys (yet), denied" << dendl
;
1305 entity_name_t
n(con
->get_peer_type(), global_id
);
1307 // We allow connections and assign Session instances to connections
1308 // even if we have not been assigned a rank, because clients with
1309 // "allow *" are allowed to connect and do 'tell' operations before
1313 // If we do hold a rank, see if this is an existing client establishing
1314 // a new connection, rather than a new client
1315 s
= mds_rank
->sessionmap
.get_session(n
);
1318 // Wire up a Session* to this connection
1319 // It doesn't go into a SessionMap instance until it sends an explicit
1320 // request to open a session (initial state of Session is `closed`)
1323 s
->info
.auth_name
= name
;
1324 s
->info
.inst
.addr
= con
->get_peer_addr();
1325 s
->info
.inst
.name
= n
;
1326 dout(10) << " new session " << s
<< " for " << s
->info
.inst
<< " con " << con
<< dendl
;
1328 s
->connection
= con
;
1330 dout(10) << " existing session " << s
<< " for " << s
->info
.inst
<< " existing con " << s
->connection
1331 << ", new/authorizing con " << con
<< dendl
;
1332 con
->set_priv(s
->get());
1336 // Wait until we fully accept the connection before setting
1337 // s->connection. In particular, if there are multiple incoming
1338 // connection attempts, they will all get their authorizer
1339 // validated, but some of them may "lose the race" and get
1340 // dropped. We only want to consider the winner(s). See
1341 // ms_handle_accept(). This is important for Sessions we replay
1342 // from the journal on recovery that don't have established
1343 // messenger state; we want the con from only the winning
1344 // connect attempt(s). (Normal reconnects that don't follow MDS
1345 // recovery are reconnected to the existing con by the
1349 if (caps_info
.allow_all
) {
1350 // Flag for auth providers that don't provide cap strings
1351 s
->auth_caps
.set_allow_all();
1353 bufferlist::iterator p
= caps_info
.caps
.begin();
1354 string auth_cap_str
;
1356 ::decode(auth_cap_str
, p
);
1358 dout(10) << __func__
<< ": parsing auth_cap_str='" << auth_cap_str
<< "'" << dendl
;
1359 std::ostringstream errstr
;
1360 if (!s
->auth_caps
.parse(g_ceph_context
, auth_cap_str
, &errstr
)) {
1361 dout(1) << __func__
<< ": auth cap parse error: " << errstr
.str()
1362 << " parsing '" << auth_cap_str
<< "'" << dendl
;
1363 clog
->warn() << name
<< " mds cap '" << auth_cap_str
1364 << "' does not parse: " << errstr
.str();
1367 } catch (buffer::error
& e
) {
1368 // Assume legacy auth, defaults to:
1369 // * permit all filesystem ops
1370 // * permit no `tell` ops
1371 dout(1) << __func__
<< ": cannot decode auth caps bl of length " << caps_info
.caps
.length() << dendl
;
1377 return true; // we made a decision (see is_valid)
1381 void MDSDaemon::ms_handle_accept(Connection
*con
)
1383 Mutex::Locker
l(mds_lock
);
1388 Session
*s
= static_cast<Session
*>(con
->get_priv());
1389 dout(10) << "ms_handle_accept " << con
->get_peer_addr() << " con " << con
<< " session " << s
<< dendl
;
1391 if (s
->connection
!= con
) {
1392 dout(10) << " session connection " << s
->connection
<< " -> " << con
<< dendl
;
1393 s
->connection
= con
;
1395 // send out any queued messages
1396 while (!s
->preopen_out_queue
.empty()) {
1397 con
->send_message(s
->preopen_out_queue
.front());
1398 s
->preopen_out_queue
.pop_front();
1405 bool MDSDaemon::is_clean_shutdown()
1408 return mds_rank
->is_stopped();