]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2016 John Spray <john.spray@redhat.com> | |
7 | * | |
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. | |
12 | */ | |
13 | ||
224ce89b WB |
14 | #include <boost/tokenizer.hpp> |
15 | ||
7c673cae FG |
16 | #include "messages/MMgrBeacon.h" |
17 | #include "messages/MMgrMap.h" | |
18 | #include "messages/MMgrDigest.h" | |
19 | ||
31f18b77 | 20 | #include "PGStatService.h" |
7c673cae FG |
21 | #include "include/stringify.h" |
22 | #include "mgr/MgrContext.h" | |
c07f9fc5 | 23 | #include "mgr/mgr_commands.h" |
7c673cae FG |
24 | #include "OSDMonitor.h" |
25 | ||
26 | #include "MgrMonitor.h" | |
27 | ||
c07f9fc5 FG |
28 | #define MGR_METADATA_PREFIX "mgr_metadata" |
29 | ||
7c673cae FG |
30 | #define dout_subsys ceph_subsys_mon |
31 | #undef dout_prefix | |
32 | #define dout_prefix _prefix(_dout, mon, map) | |
33 | static ostream& _prefix(std::ostream *_dout, Monitor *mon, | |
34 | const MgrMap& mgrmap) { | |
35 | return *_dout << "mon." << mon->name << "@" << mon->rank | |
36 | << "(" << mon->get_state_name() | |
37 | << ").mgr e" << mgrmap.get_epoch() << " "; | |
38 | } | |
39 | ||
c07f9fc5 FG |
40 | // Prefix for mon store of active mgr's command descriptions |
41 | const static std::string command_descs_prefix = "mgr_command_descs"; | |
42 | ||
31f18b77 | 43 | |
7c673cae FG |
44 | void MgrMonitor::create_initial() |
45 | { | |
224ce89b WB |
46 | boost::tokenizer<> tok(g_conf->mgr_initial_modules); |
47 | for (auto& m : tok) { | |
48 | pending_map.modules.insert(m); | |
49 | } | |
c07f9fc5 FG |
50 | pending_command_descs = mgr_commands; |
51 | dout(10) << __func__ << " initial modules " << pending_map.modules | |
52 | << ", " << pending_command_descs.size() << " commands" | |
53 | << dendl; | |
7c673cae FG |
54 | } |
55 | ||
56 | void MgrMonitor::update_from_paxos(bool *need_bootstrap) | |
57 | { | |
58 | version_t version = get_last_committed(); | |
59 | if (version != map.epoch) { | |
60 | dout(4) << "loading version " << version << dendl; | |
61 | ||
62 | bufferlist bl; | |
63 | int err = get_version(version, bl); | |
64 | assert(err == 0); | |
65 | ||
c07f9fc5 FG |
66 | bool old_available = map.get_available(); |
67 | uint64_t old_gid = map.get_active_gid(); | |
68 | ||
7c673cae FG |
69 | bufferlist::iterator p = bl.begin(); |
70 | map.decode(p); | |
71 | ||
72 | dout(4) << "active server: " << map.active_addr | |
73 | << "(" << map.active_gid << ")" << dendl; | |
74 | ||
224ce89b WB |
75 | ever_had_active_mgr = get_value("ever_had_active_mgr"); |
76 | ||
77 | load_health(); | |
78 | ||
7c673cae FG |
79 | if (map.available) { |
80 | first_seen_inactive = utime_t(); | |
81 | } else { | |
82 | first_seen_inactive = ceph_clock_now(); | |
83 | } | |
84 | ||
85 | check_subs(); | |
c07f9fc5 FG |
86 | |
87 | if (version == 1 | |
88 | || (map.get_available() | |
89 | && (!old_available || old_gid != map.get_active_gid()))) { | |
90 | dout(4) << "mkfs or daemon transitioned to available, loading commands" | |
91 | << dendl; | |
92 | bufferlist loaded_commands; | |
93 | int r = mon->store->get(command_descs_prefix, "", loaded_commands); | |
94 | if (r < 0) { | |
95 | derr << "Failed to load mgr commands: " << cpp_strerror(r) << dendl; | |
96 | } else { | |
97 | auto p = loaded_commands.begin(); | |
98 | ::decode(command_descs, p); | |
99 | } | |
100 | } | |
7c673cae FG |
101 | } |
102 | ||
103 | // feed our pet MgrClient | |
104 | mon->mgr_client.ms_dispatch(new MMgrMap(map)); | |
105 | } | |
106 | ||
107 | void MgrMonitor::create_pending() | |
108 | { | |
109 | pending_map = map; | |
110 | pending_map.epoch++; | |
111 | } | |
112 | ||
224ce89b WB |
113 | health_status_t MgrMonitor::should_warn_about_mgr_down() |
114 | { | |
115 | utime_t now = ceph_clock_now(); | |
116 | // we warn if | |
117 | // - we've ever had an active mgr, or | |
118 | // - we have osds AND we've exceeded the grace period | |
119 | // which means a new mon cluster and be HEALTH_OK indefinitely as long as | |
120 | // no OSDs are ever created. | |
121 | if (ever_had_active_mgr || | |
122 | (mon->osdmon()->osdmap.get_num_osds() > 0 && | |
123 | now > mon->monmap->created + g_conf->mon_mgr_mkfs_grace)) { | |
124 | health_status_t level = HEALTH_WARN; | |
125 | if (first_seen_inactive != utime_t() && | |
126 | now - first_seen_inactive > g_conf->mon_mgr_inactive_grace) { | |
127 | level = HEALTH_ERR; | |
128 | } | |
129 | return level; | |
130 | } | |
131 | return HEALTH_OK; | |
132 | } | |
133 | ||
7c673cae FG |
134 | void MgrMonitor::encode_pending(MonitorDBStore::TransactionRef t) |
135 | { | |
136 | dout(10) << __func__ << " " << pending_map << dendl; | |
137 | bufferlist bl; | |
138 | pending_map.encode(bl, mon->get_quorum_con_features()); | |
139 | put_version(t, pending_map.epoch, bl); | |
140 | put_last_committed(t, pending_map.epoch); | |
224ce89b | 141 | |
c07f9fc5 FG |
142 | for (auto& p : pending_metadata) { |
143 | dout(10) << __func__ << " set metadata for " << p.first << dendl; | |
144 | t->put(MGR_METADATA_PREFIX, p.first, p.second); | |
145 | } | |
146 | for (auto& name : pending_metadata_rm) { | |
147 | dout(10) << __func__ << " rm metadata for " << name << dendl; | |
148 | t->erase(MGR_METADATA_PREFIX, name); | |
149 | } | |
150 | pending_metadata.clear(); | |
151 | pending_metadata_rm.clear(); | |
152 | ||
224ce89b WB |
153 | health_check_map_t next; |
154 | if (pending_map.active_gid == 0) { | |
155 | auto level = should_warn_about_mgr_down(); | |
156 | if (level != HEALTH_OK) { | |
157 | next.add("MGR_DOWN", level, "no active mgr"); | |
158 | } else { | |
159 | dout(10) << __func__ << " no health warning (never active and new cluster)" | |
160 | << dendl; | |
161 | } | |
162 | } else { | |
163 | put_value(t, "ever_had_active_mgr", 1); | |
164 | } | |
165 | encode_health(next, t); | |
c07f9fc5 FG |
166 | |
167 | if (pending_command_descs.size()) { | |
168 | dout(4) << __func__ << " encoding " << pending_command_descs.size() | |
169 | << " command_descs" << dendl; | |
170 | for (auto& p : pending_command_descs) { | |
171 | p.set_flag(MonCommand::FLAG_MGR); | |
172 | } | |
173 | bufferlist bl; | |
174 | ::encode(pending_command_descs, bl); | |
175 | t->put(command_descs_prefix, "", bl); | |
176 | pending_command_descs.clear(); | |
177 | } | |
7c673cae FG |
178 | } |
179 | ||
180 | bool MgrMonitor::check_caps(MonOpRequestRef op, const uuid_d& fsid) | |
181 | { | |
182 | // check permissions | |
183 | MonSession *session = op->get_session(); | |
184 | if (!session) | |
185 | return false; | |
186 | if (!session->is_capable("mgr", MON_CAP_X)) { | |
187 | dout(1) << __func__ << " insufficient caps " << session->caps << dendl; | |
188 | return false; | |
189 | } | |
190 | if (fsid != mon->monmap->fsid) { | |
191 | dout(1) << __func__ << " op fsid " << fsid | |
192 | << " != " << mon->monmap->fsid << dendl; | |
193 | return false; | |
194 | } | |
195 | return true; | |
196 | } | |
197 | ||
198 | bool MgrMonitor::preprocess_query(MonOpRequestRef op) | |
199 | { | |
200 | PaxosServiceMessage *m = static_cast<PaxosServiceMessage*>(op->get_req()); | |
201 | switch (m->get_type()) { | |
202 | case MSG_MGR_BEACON: | |
203 | return preprocess_beacon(op); | |
204 | case MSG_MON_COMMAND: | |
205 | return preprocess_command(op); | |
206 | default: | |
207 | mon->no_reply(op); | |
208 | derr << "Unhandled message type " << m->get_type() << dendl; | |
209 | return true; | |
210 | } | |
211 | } | |
212 | ||
213 | bool MgrMonitor::prepare_update(MonOpRequestRef op) | |
214 | { | |
215 | PaxosServiceMessage *m = static_cast<PaxosServiceMessage*>(op->get_req()); | |
216 | switch (m->get_type()) { | |
217 | case MSG_MGR_BEACON: | |
218 | return prepare_beacon(op); | |
219 | ||
220 | case MSG_MON_COMMAND: | |
221 | return prepare_command(op); | |
222 | ||
223 | default: | |
224 | mon->no_reply(op); | |
225 | derr << "Unhandled message type " << m->get_type() << dendl; | |
226 | return true; | |
227 | } | |
228 | } | |
229 | ||
230 | ||
231 | ||
232 | class C_Updated : public Context { | |
233 | MgrMonitor *mm; | |
234 | MonOpRequestRef op; | |
235 | public: | |
236 | C_Updated(MgrMonitor *a, MonOpRequestRef c) : | |
237 | mm(a), op(c) {} | |
238 | void finish(int r) override { | |
239 | if (r >= 0) { | |
240 | // Success | |
241 | } else if (r == -ECANCELED) { | |
242 | mm->mon->no_reply(op); | |
243 | } else { | |
244 | mm->dispatch(op); // try again | |
245 | } | |
246 | } | |
247 | }; | |
248 | ||
249 | bool MgrMonitor::preprocess_beacon(MonOpRequestRef op) | |
250 | { | |
251 | MMgrBeacon *m = static_cast<MMgrBeacon*>(op->get_req()); | |
252 | dout(4) << "beacon from " << m->get_gid() << dendl; | |
253 | ||
254 | if (!check_caps(op, m->get_fsid())) { | |
255 | // drop it on the floor | |
256 | return true; | |
257 | } | |
258 | ||
259 | // always send this to the leader's prepare_beacon() | |
260 | return false; | |
261 | } | |
262 | ||
263 | bool MgrMonitor::prepare_beacon(MonOpRequestRef op) | |
264 | { | |
265 | MMgrBeacon *m = static_cast<MMgrBeacon*>(op->get_req()); | |
266 | dout(4) << "beacon from " << m->get_gid() << dendl; | |
267 | ||
268 | // See if we are seeing same name, new GID for the active daemon | |
269 | if (m->get_name() == pending_map.active_name | |
270 | && m->get_gid() != pending_map.active_gid) | |
271 | { | |
272 | dout(4) << "Active daemon restart (mgr." << m->get_name() << ")" << dendl; | |
224ce89b WB |
273 | mon->clog->info() << "Active manager daemon " << m->get_name() |
274 | << " restarted"; | |
7c673cae FG |
275 | drop_active(); |
276 | } | |
277 | ||
278 | // See if we are seeing same name, new GID for any standbys | |
279 | for (const auto &i : pending_map.standbys) { | |
280 | const StandbyInfo &s = i.second; | |
281 | if (s.name == m->get_name() && s.gid != m->get_gid()) { | |
282 | dout(4) << "Standby daemon restart (mgr." << m->get_name() << ")" << dendl; | |
224ce89b WB |
283 | mon->clog->debug() << "Standby manager daemon " << m->get_name() |
284 | << " restarted"; | |
7c673cae FG |
285 | drop_standby(i.first); |
286 | break; | |
287 | } | |
288 | } | |
289 | ||
31f18b77 | 290 | last_beacon[m->get_gid()] = ceph::coarse_mono_clock::now(); |
7c673cae FG |
291 | |
292 | // Track whether we modified pending_map | |
293 | bool updated = false; | |
294 | ||
295 | if (pending_map.active_gid == m->get_gid()) { | |
296 | // A beacon from the currently active daemon | |
297 | if (pending_map.active_addr != m->get_server_addr()) { | |
298 | dout(4) << "learned address " << m->get_server_addr() | |
299 | << " (was " << pending_map.active_addr << ")" << dendl; | |
300 | pending_map.active_addr = m->get_server_addr(); | |
301 | updated = true; | |
302 | } | |
303 | ||
304 | if (pending_map.get_available() != m->get_available()) { | |
305 | dout(4) << "available " << m->get_gid() << dendl; | |
224ce89b WB |
306 | mon->clog->info() << "Manager daemon " << pending_map.active_name |
307 | << " is now available"; | |
c07f9fc5 FG |
308 | |
309 | // This beacon should include command descriptions | |
310 | pending_command_descs = m->get_command_descs(); | |
311 | if (pending_command_descs.empty()) { | |
312 | // This should not happen, but it also isn't fatal: we just | |
313 | // won't successfully update our list of commands. | |
314 | dout(4) << "First available beacon from " << pending_map.active_name | |
315 | << "(" << m->get_gid() << ") does not include command descs" | |
316 | << dendl; | |
317 | } else { | |
318 | dout(4) << "First available beacon from " << pending_map.active_name | |
319 | << "(" << m->get_gid() << ") includes " | |
320 | << pending_command_descs.size() << " command descs" << dendl; | |
321 | } | |
322 | ||
7c673cae FG |
323 | pending_map.available = m->get_available(); |
324 | updated = true; | |
325 | } | |
224ce89b WB |
326 | if (pending_map.available_modules != m->get_available_modules()) { |
327 | dout(4) << "available_modules " << m->get_available_modules() | |
328 | << " (was " << pending_map.available_modules << ")" << dendl; | |
329 | pending_map.available_modules = m->get_available_modules(); | |
330 | updated = true; | |
331 | } | |
7c673cae FG |
332 | } else if (pending_map.active_gid == 0) { |
333 | // There is no currently active daemon, select this one. | |
334 | if (pending_map.standbys.count(m->get_gid())) { | |
335 | drop_standby(m->get_gid()); | |
336 | } | |
337 | dout(4) << "selecting new active " << m->get_gid() | |
338 | << " " << m->get_name() | |
339 | << " (was " << pending_map.active_gid << " " | |
340 | << pending_map.active_name << ")" << dendl; | |
341 | pending_map.active_gid = m->get_gid(); | |
342 | pending_map.active_name = m->get_name(); | |
224ce89b | 343 | pending_map.available_modules = m->get_available_modules(); |
c07f9fc5 FG |
344 | ::encode(m->get_metadata(), pending_metadata[m->get_name()]); |
345 | pending_metadata_rm.erase(m->get_name()); | |
224ce89b WB |
346 | |
347 | mon->clog->info() << "Activating manager daemon " | |
348 | << pending_map.active_name; | |
7c673cae FG |
349 | |
350 | updated = true; | |
351 | } else { | |
352 | if (pending_map.standbys.count(m->get_gid()) > 0) { | |
353 | dout(10) << "from existing standby " << m->get_gid() << dendl; | |
224ce89b WB |
354 | if (pending_map.standbys[m->get_gid()].available_modules != |
355 | m->get_available_modules()) { | |
356 | dout(10) << "existing standby " << m->get_gid() << " available_modules " | |
357 | << m->get_available_modules() << " (was " | |
358 | << pending_map.standbys[m->get_gid()].available_modules << ")" | |
359 | << dendl; | |
360 | pending_map.standbys[m->get_gid()].available_modules = | |
361 | m->get_available_modules(); | |
362 | updated = true; | |
363 | } | |
7c673cae FG |
364 | } else { |
365 | dout(10) << "new standby " << m->get_gid() << dendl; | |
224ce89b WB |
366 | mon->clog->debug() << "Standby manager daemon " << m->get_name() |
367 | << " started"; | |
c07f9fc5 FG |
368 | pending_map.standbys[m->get_gid()] = {m->get_gid(), m->get_name(), |
369 | m->get_available_modules()}; | |
370 | ::encode(m->get_metadata(), pending_metadata[m->get_name()]); | |
371 | pending_metadata_rm.erase(m->get_name()); | |
7c673cae FG |
372 | updated = true; |
373 | } | |
374 | } | |
375 | ||
376 | if (updated) { | |
377 | dout(4) << "updating map" << dendl; | |
378 | wait_for_finished_proposal(op, new C_Updated(this, op)); | |
379 | } else { | |
380 | dout(10) << "no change" << dendl; | |
381 | } | |
382 | ||
383 | return updated; | |
384 | } | |
385 | ||
386 | void MgrMonitor::check_subs() | |
387 | { | |
388 | const std::string type = "mgrmap"; | |
389 | if (mon->session_map.subs.count(type) == 0) | |
390 | return; | |
391 | for (auto sub : *(mon->session_map.subs[type])) { | |
392 | check_sub(sub); | |
393 | } | |
394 | } | |
395 | ||
396 | void MgrMonitor::check_sub(Subscription *sub) | |
397 | { | |
398 | if (sub->type == "mgrmap") { | |
399 | if (sub->next <= map.get_epoch()) { | |
224ce89b WB |
400 | dout(20) << "Sending map to subscriber " << sub->session->con |
401 | << " " << sub->session->con->get_peer_addr() << dendl; | |
7c673cae FG |
402 | sub->session->con->send_message(new MMgrMap(map)); |
403 | if (sub->onetime) { | |
404 | mon->session_map.remove_sub(sub); | |
405 | } else { | |
406 | sub->next = map.get_epoch() + 1; | |
407 | } | |
408 | } | |
409 | } else { | |
410 | assert(sub->type == "mgrdigest"); | |
c07f9fc5 FG |
411 | if (sub->next == 0) { |
412 | // new registration; cancel previous timer | |
413 | cancel_timer(); | |
414 | } | |
31f18b77 | 415 | if (digest_event == nullptr) { |
7c673cae FG |
416 | send_digests(); |
417 | } | |
418 | } | |
419 | } | |
420 | ||
421 | /** | |
422 | * Handle digest subscriptions separately (outside of check_sub) because | |
423 | * they are going to be periodic rather than version-driven. | |
424 | */ | |
425 | void MgrMonitor::send_digests() | |
426 | { | |
31f18b77 FG |
427 | cancel_timer(); |
428 | ||
429 | if (!is_active()) { | |
430 | return; | |
431 | } | |
224ce89b | 432 | dout(10) << __func__ << dendl; |
7c673cae FG |
433 | |
434 | const std::string type = "mgrdigest"; | |
435 | if (mon->session_map.subs.count(type) == 0) | |
436 | return; | |
437 | ||
438 | for (auto sub : *(mon->session_map.subs[type])) { | |
224ce89b WB |
439 | dout(10) << __func__ << " sending digest to subscriber " << sub->session->con |
440 | << " " << sub->session->con->get_peer_addr() << dendl; | |
7c673cae FG |
441 | MMgrDigest *mdigest = new MMgrDigest; |
442 | ||
443 | JSONFormatter f; | |
224ce89b | 444 | mon->get_health_status(true, &f, nullptr, nullptr, nullptr); |
7c673cae FG |
445 | f.flush(mdigest->health_json); |
446 | f.reset(); | |
447 | ||
448 | std::ostringstream ss; | |
449 | mon->get_mon_status(&f, ss); | |
450 | f.flush(mdigest->mon_status_json); | |
451 | f.reset(); | |
452 | ||
453 | sub->session->con->send_message(mdigest); | |
454 | } | |
455 | ||
31f18b77 | 456 | digest_event = new C_MonContext(mon, [this](int){ |
7c673cae FG |
457 | send_digests(); |
458 | }); | |
31f18b77 FG |
459 | mon->timer.add_event_after(g_conf->mon_mgr_digest_period, digest_event); |
460 | } | |
461 | ||
462 | void MgrMonitor::cancel_timer() | |
463 | { | |
464 | if (digest_event) { | |
465 | mon->timer.cancel_event(digest_event); | |
466 | digest_event = nullptr; | |
467 | } | |
7c673cae FG |
468 | } |
469 | ||
470 | void MgrMonitor::on_active() | |
471 | { | |
224ce89b WB |
472 | if (mon->is_leader()) { |
473 | mon->clog->debug() << "mgrmap e" << map.epoch << ": " << map; | |
474 | } | |
7c673cae FG |
475 | } |
476 | ||
477 | void MgrMonitor::get_health( | |
478 | list<pair<health_status_t,string> >& summary, | |
479 | list<pair<health_status_t,string> > *detail, | |
480 | CephContext *cct) const | |
481 | { | |
482 | // start mgr warnings as soon as the mons and osds are all upgraded, | |
483 | // but before the require_luminous osdmap flag is set. this way the | |
484 | // user gets some warning before the osd flag is set and mgr is | |
485 | // actually *required*. | |
486 | if (!mon->monmap->get_required_features().contains_all( | |
487 | ceph::features::mon::FEATURE_LUMINOUS) || | |
488 | !HAVE_FEATURE(mon->osdmon()->osdmap.get_up_osd_features(), | |
489 | SERVER_LUMINOUS)) { | |
490 | return; | |
491 | } | |
492 | ||
224ce89b | 493 | if (map.active_gid == 0) { |
7c673cae FG |
494 | auto level = HEALTH_WARN; |
495 | // do not escalate to ERR if they are still upgrading to jewel. | |
31f18b77 | 496 | if (mon->osdmon()->osdmap.require_osd_release >= CEPH_RELEASE_LUMINOUS) { |
7c673cae FG |
497 | utime_t now = ceph_clock_now(); |
498 | if (first_seen_inactive != utime_t() && | |
499 | now - first_seen_inactive > g_conf->mon_mgr_inactive_grace) { | |
500 | level = HEALTH_ERR; | |
501 | } | |
502 | } | |
503 | summary.push_back(make_pair(level, "no active mgr")); | |
504 | } | |
505 | } | |
506 | ||
507 | void MgrMonitor::tick() | |
508 | { | |
509 | if (!is_active() || !mon->is_leader()) | |
510 | return; | |
511 | ||
31f18b77 FG |
512 | const auto now = ceph::coarse_mono_clock::now(); |
513 | const auto cutoff = now - std::chrono::seconds(g_conf->mon_mgr_beacon_grace); | |
7c673cae FG |
514 | |
515 | // Populate any missing beacons (i.e. no beacon since MgrMonitor | |
516 | // instantiation) with the current time, so that they will | |
517 | // eventually look laggy if they fail to give us a beacon. | |
518 | if (pending_map.active_gid != 0 | |
519 | && last_beacon.count(pending_map.active_gid) == 0) { | |
520 | last_beacon[pending_map.active_gid] = now; | |
521 | } | |
522 | for (auto s : pending_map.standbys) { | |
523 | if (last_beacon.count(s.first) == 0) { | |
524 | last_beacon[s.first] = now; | |
525 | } | |
526 | } | |
527 | ||
528 | // Cull standbys first so that any remaining standbys | |
529 | // will be eligible to take over from the active if we cull him. | |
530 | std::list<uint64_t> dead_standbys; | |
531 | for (const auto &i : pending_map.standbys) { | |
532 | auto last_beacon_time = last_beacon.at(i.first); | |
533 | if (last_beacon_time < cutoff) { | |
534 | dead_standbys.push_back(i.first); | |
535 | } | |
536 | } | |
537 | ||
538 | bool propose = false; | |
539 | ||
540 | for (auto i : dead_standbys) { | |
541 | dout(4) << "Dropping laggy standby " << i << dendl; | |
542 | drop_standby(i); | |
543 | propose = true; | |
544 | } | |
545 | ||
546 | if (pending_map.active_gid != 0 | |
547 | && last_beacon.at(pending_map.active_gid) < cutoff) { | |
224ce89b | 548 | const std::string old_active_name = pending_map.active_name; |
7c673cae FG |
549 | drop_active(); |
550 | propose = true; | |
551 | dout(4) << "Dropping active" << pending_map.active_gid << dendl; | |
552 | if (promote_standby()) { | |
553 | dout(4) << "Promoted standby " << pending_map.active_gid << dendl; | |
224ce89b WB |
554 | mon->clog->info() << "Manager daemon " << old_active_name |
555 | << " is unresponsive, replacing it with standby" | |
556 | << " daemon " << pending_map.active_name; | |
7c673cae FG |
557 | } else { |
558 | dout(4) << "Active is laggy but have no standbys to replace it" << dendl; | |
224ce89b WB |
559 | mon->clog->warn() << "Manager daemon " << old_active_name |
560 | << " is unresponsive. No standby daemons available."; | |
7c673cae FG |
561 | } |
562 | } else if (pending_map.active_gid == 0) { | |
563 | if (promote_standby()) { | |
564 | dout(4) << "Promoted standby " << pending_map.active_gid << dendl; | |
224ce89b WB |
565 | mon->clog->info() << "Activating manager daemon " |
566 | << pending_map.active_name; | |
7c673cae FG |
567 | propose = true; |
568 | } | |
569 | } | |
570 | ||
224ce89b | 571 | if (!pending_map.available && |
c07f9fc5 | 572 | !ever_had_active_mgr && |
224ce89b WB |
573 | should_warn_about_mgr_down() != HEALTH_OK) { |
574 | dout(10) << " exceeded mon_mgr_mkfs_grace " << g_conf->mon_mgr_mkfs_grace | |
575 | << " seconds" << dendl; | |
576 | propose = true; | |
577 | } | |
578 | ||
7c673cae FG |
579 | if (propose) { |
580 | propose_pending(); | |
581 | } | |
582 | } | |
583 | ||
224ce89b WB |
584 | void MgrMonitor::on_restart() |
585 | { | |
586 | // Clear out the leader-specific state. | |
587 | last_beacon.clear(); | |
588 | } | |
589 | ||
590 | ||
7c673cae FG |
591 | bool MgrMonitor::promote_standby() |
592 | { | |
593 | assert(pending_map.active_gid == 0); | |
594 | if (pending_map.standbys.size()) { | |
595 | // Promote a replacement (arbitrary choice of standby) | |
596 | auto replacement_gid = pending_map.standbys.begin()->first; | |
597 | pending_map.active_gid = replacement_gid; | |
598 | pending_map.active_name = pending_map.standbys.at(replacement_gid).name; | |
599 | pending_map.available = false; | |
600 | pending_map.active_addr = entity_addr_t(); | |
601 | ||
602 | drop_standby(replacement_gid); | |
603 | return true; | |
604 | } else { | |
605 | return false; | |
606 | } | |
607 | } | |
608 | ||
609 | void MgrMonitor::drop_active() | |
610 | { | |
611 | if (last_beacon.count(pending_map.active_gid) > 0) { | |
612 | last_beacon.erase(pending_map.active_gid); | |
613 | } | |
614 | ||
c07f9fc5 FG |
615 | pending_metadata_rm.insert(pending_map.active_name); |
616 | pending_metadata.erase(pending_map.active_name); | |
7c673cae FG |
617 | pending_map.active_name = ""; |
618 | pending_map.active_gid = 0; | |
619 | pending_map.available = false; | |
620 | pending_map.active_addr = entity_addr_t(); | |
224ce89b WB |
621 | |
622 | // So that when new active mgr subscribes to mgrdigest, it will | |
623 | // get an immediate response instead of waiting for next timer | |
624 | cancel_timer(); | |
7c673cae FG |
625 | } |
626 | ||
627 | void MgrMonitor::drop_standby(uint64_t gid) | |
628 | { | |
c07f9fc5 FG |
629 | pending_metadata_rm.insert(pending_map.standbys[gid].name); |
630 | pending_metadata.erase(pending_map.standbys[gid].name); | |
7c673cae FG |
631 | pending_map.standbys.erase(gid); |
632 | if (last_beacon.count(gid) > 0) { | |
633 | last_beacon.erase(gid); | |
634 | } | |
7c673cae FG |
635 | } |
636 | ||
637 | bool MgrMonitor::preprocess_command(MonOpRequestRef op) | |
638 | { | |
31f18b77 FG |
639 | MMonCommand *m = static_cast<MMonCommand*>(op->get_req()); |
640 | std::stringstream ss; | |
641 | bufferlist rdata; | |
642 | ||
643 | std::map<std::string, cmd_vartype> cmdmap; | |
644 | if (!cmdmap_from_json(m->cmd, &cmdmap, ss)) { | |
645 | string rs = ss.str(); | |
646 | mon->reply_command(op, -EINVAL, rs, rdata, get_last_committed()); | |
647 | return true; | |
648 | } | |
7c673cae | 649 | |
31f18b77 FG |
650 | MonSession *session = m->get_session(); |
651 | if (!session) { | |
652 | mon->reply_command(op, -EACCES, "access denied", rdata, | |
653 | get_last_committed()); | |
654 | return true; | |
655 | } | |
656 | ||
657 | string format; | |
658 | cmd_getval(g_ceph_context, cmdmap, "format", format, string("json-pretty")); | |
659 | boost::scoped_ptr<Formatter> f(Formatter::create(format)); | |
660 | ||
661 | string prefix; | |
662 | cmd_getval(g_ceph_context, cmdmap, "prefix", prefix); | |
663 | int r = 0; | |
664 | ||
665 | if (prefix == "mgr dump") { | |
666 | int64_t epoch = 0; | |
667 | cmd_getval(g_ceph_context, cmdmap, "epoch", epoch, (int64_t)map.get_epoch()); | |
668 | if (epoch == (int64_t)map.get_epoch()) { | |
669 | f->dump_object("mgrmap", map); | |
670 | } else { | |
671 | bufferlist bl; | |
672 | int err = get_version(epoch, bl); | |
673 | if (err == -ENOENT) { | |
674 | r = -ENOENT; | |
675 | ss << "there is no map for epoch " << epoch; | |
676 | goto reply; | |
677 | } | |
678 | MgrMap m; | |
679 | auto p = bl.begin(); | |
680 | m.decode(p); | |
681 | f->dump_object("mgrmap", m); | |
682 | } | |
683 | f->flush(rdata); | |
224ce89b WB |
684 | } else if (prefix == "mgr module ls") { |
685 | f->open_array_section("modules"); | |
686 | for (auto& p : map.modules) { | |
687 | f->dump_string("module", p); | |
688 | } | |
689 | f->close_section(); | |
690 | f->flush(rdata); | |
c07f9fc5 FG |
691 | } else if (prefix == "mgr metadata") { |
692 | string name; | |
693 | cmd_getval(g_ceph_context, cmdmap, "id", name); | |
694 | if (name.size() > 0 && !map.have_name(name)) { | |
695 | ss << "mgr." << name << " does not exist"; | |
696 | r = -ENOENT; | |
697 | goto reply; | |
698 | } | |
699 | string format; | |
700 | cmd_getval(g_ceph_context, cmdmap, "format", format); | |
701 | boost::scoped_ptr<Formatter> f(Formatter::create(format, "json-pretty", "json-pretty")); | |
702 | if (name.size()) { | |
703 | f->open_object_section("mgr_metadata"); | |
704 | f->dump_string("id", name); | |
705 | r = dump_metadata(name, f.get(), &ss); | |
706 | if (r < 0) | |
707 | goto reply; | |
708 | f->close_section(); | |
709 | } else { | |
710 | r = 0; | |
711 | f->open_array_section("mgr_metadata"); | |
712 | for (auto& i : map.get_all_names()) { | |
713 | f->open_object_section("mgr"); | |
714 | f->dump_string("id", i); | |
715 | r = dump_metadata(i, f.get(), NULL); | |
716 | if (r == -EINVAL || r == -ENOENT) { | |
717 | // Drop error, continue to get other daemons' metadata | |
718 | dout(4) << "No metadata for mgr." << i << dendl; | |
719 | r = 0; | |
720 | } else if (r < 0) { | |
721 | // Unexpected error | |
722 | goto reply; | |
723 | } | |
724 | f->close_section(); | |
725 | } | |
726 | f->close_section(); | |
727 | } | |
728 | f->flush(rdata); | |
729 | } else if (prefix == "mgr versions") { | |
730 | if (!f) | |
731 | f.reset(Formatter::create("json-pretty")); | |
732 | count_metadata("ceph_version", f.get()); | |
733 | f->flush(rdata); | |
734 | r = 0; | |
735 | } else if (prefix == "mgr count-metadata") { | |
736 | if (!f) | |
737 | f.reset(Formatter::create("json-pretty")); | |
738 | string field; | |
739 | cmd_getval(g_ceph_context, cmdmap, "property", field); | |
740 | count_metadata(field, f.get()); | |
741 | f->flush(rdata); | |
742 | r = 0; | |
31f18b77 FG |
743 | } else { |
744 | return false; | |
745 | } | |
746 | ||
747 | reply: | |
748 | string rs; | |
749 | getline(ss, rs); | |
750 | mon->reply_command(op, r, rs, rdata, get_last_committed()); | |
751 | return true; | |
7c673cae FG |
752 | } |
753 | ||
754 | bool MgrMonitor::prepare_command(MonOpRequestRef op) | |
755 | { | |
756 | MMonCommand *m = static_cast<MMonCommand*>(op->get_req()); | |
757 | ||
758 | std::stringstream ss; | |
759 | bufferlist rdata; | |
760 | ||
761 | std::map<std::string, cmd_vartype> cmdmap; | |
762 | if (!cmdmap_from_json(m->cmd, &cmdmap, ss)) { | |
763 | string rs = ss.str(); | |
764 | mon->reply_command(op, -EINVAL, rs, rdata, get_last_committed()); | |
765 | return true; | |
766 | } | |
767 | ||
768 | MonSession *session = m->get_session(); | |
769 | if (!session) { | |
770 | mon->reply_command(op, -EACCES, "access denied", rdata, get_last_committed()); | |
771 | return true; | |
772 | } | |
773 | ||
224ce89b WB |
774 | string format; |
775 | cmd_getval(g_ceph_context, cmdmap, "format", format, string("plain")); | |
776 | boost::scoped_ptr<Formatter> f(Formatter::create(format)); | |
777 | ||
7c673cae FG |
778 | string prefix; |
779 | cmd_getval(g_ceph_context, cmdmap, "prefix", prefix); | |
780 | ||
781 | int r = 0; | |
782 | ||
783 | if (prefix == "mgr fail") { | |
784 | string who; | |
785 | cmd_getval(g_ceph_context, cmdmap, "who", who); | |
786 | ||
787 | std::string err; | |
788 | uint64_t gid = strict_strtol(who.c_str(), 10, &err); | |
789 | bool changed = false; | |
790 | if (!err.empty()) { | |
791 | // Does not parse as a gid, treat it as a name | |
792 | if (pending_map.active_name == who) { | |
793 | drop_active(); | |
794 | changed = true; | |
795 | } else { | |
796 | gid = 0; | |
797 | for (const auto &i : pending_map.standbys) { | |
798 | if (i.second.name == who) { | |
799 | gid = i.first; | |
800 | break; | |
801 | } | |
802 | } | |
803 | if (gid != 0) { | |
804 | drop_standby(gid); | |
805 | changed = true; | |
806 | } else { | |
807 | ss << "Daemon not found '" << who << "', already failed?"; | |
808 | } | |
809 | } | |
810 | } else { | |
811 | if (pending_map.active_gid == gid) { | |
812 | drop_active(); | |
813 | changed = true; | |
814 | } else if (pending_map.standbys.count(gid) > 0) { | |
815 | drop_standby(gid); | |
816 | changed = true; | |
817 | } else { | |
818 | ss << "Daemon not found '" << gid << "', already failed?"; | |
819 | } | |
820 | } | |
821 | ||
822 | if (changed && pending_map.active_gid == 0) { | |
823 | promote_standby(); | |
824 | } | |
224ce89b WB |
825 | } else if (prefix == "mgr module enable") { |
826 | string module; | |
827 | cmd_getval(g_ceph_context, cmdmap, "module", module); | |
828 | if (module.empty()) { | |
829 | r = -EINVAL; | |
830 | goto out; | |
831 | } | |
832 | string force; | |
833 | cmd_getval(g_ceph_context, cmdmap, "force", force); | |
834 | if (!pending_map.all_support_module(module) && | |
835 | force != "--force") { | |
836 | ss << "all mgr daemons do not support module '" << module << "', pass " | |
837 | << "--force to force enablement"; | |
838 | r = -ENOENT; | |
839 | goto out; | |
840 | } | |
841 | pending_map.modules.insert(module); | |
842 | } else if (prefix == "mgr module disable") { | |
843 | string module; | |
844 | cmd_getval(g_ceph_context, cmdmap, "module", module); | |
845 | if (module.empty()) { | |
846 | r = -EINVAL; | |
847 | goto out; | |
848 | } | |
849 | pending_map.modules.erase(module); | |
7c673cae | 850 | } else { |
224ce89b | 851 | ss << "Command '" << prefix << "' not implemented!"; |
7c673cae FG |
852 | r = -ENOSYS; |
853 | } | |
854 | ||
224ce89b | 855 | out: |
7c673cae FG |
856 | dout(4) << __func__ << " done, r=" << r << dendl; |
857 | /* Compose response */ | |
858 | string rs; | |
859 | getline(ss, rs); | |
860 | ||
861 | if (r >= 0) { | |
862 | // success.. delay reply | |
863 | wait_for_finished_proposal(op, new Monitor::C_Command(mon, op, r, rs, | |
864 | get_last_committed() + 1)); | |
865 | return true; | |
866 | } else { | |
867 | // reply immediately | |
868 | mon->reply_command(op, r, rs, rdata, get_last_committed()); | |
869 | return false; | |
870 | } | |
871 | } | |
872 | ||
873 | void MgrMonitor::init() | |
874 | { | |
31f18b77 | 875 | if (digest_event == nullptr) { |
7c673cae FG |
876 | send_digests(); // To get it to schedule its own event |
877 | } | |
878 | } | |
879 | ||
880 | void MgrMonitor::on_shutdown() | |
881 | { | |
31f18b77 | 882 | cancel_timer(); |
7c673cae FG |
883 | } |
884 | ||
c07f9fc5 FG |
885 | int MgrMonitor::load_metadata(const string& name, std::map<string, string>& m, |
886 | ostream *err) | |
887 | { | |
888 | bufferlist bl; | |
889 | int r = mon->store->get(MGR_METADATA_PREFIX, name, bl); | |
890 | if (r < 0) | |
891 | return r; | |
892 | try { | |
893 | bufferlist::iterator p = bl.begin(); | |
894 | ::decode(m, p); | |
895 | } | |
896 | catch (buffer::error& e) { | |
897 | if (err) | |
898 | *err << "mgr." << name << " metadata is corrupt"; | |
899 | return -EIO; | |
900 | } | |
901 | return 0; | |
902 | } | |
903 | ||
904 | void MgrMonitor::count_metadata(const string& field, std::map<string,int> *out) | |
905 | { | |
906 | std::set<string> ls = map.get_all_names(); | |
907 | for (auto& name : ls) { | |
908 | std::map<string,string> meta; | |
909 | load_metadata(name, meta, nullptr); | |
910 | auto p = meta.find(field); | |
911 | if (p == meta.end()) { | |
912 | (*out)["unknown"]++; | |
913 | } else { | |
914 | (*out)[p->second]++; | |
915 | } | |
916 | } | |
917 | } | |
918 | ||
919 | void MgrMonitor::count_metadata(const string& field, Formatter *f) | |
920 | { | |
921 | std::map<string,int> by_val; | |
922 | count_metadata(field, &by_val); | |
923 | f->open_object_section(field.c_str()); | |
924 | for (auto& p : by_val) { | |
925 | f->dump_int(p.first.c_str(), p.second); | |
926 | } | |
927 | f->close_section(); | |
928 | } | |
929 | ||
930 | int MgrMonitor::dump_metadata(const string& name, Formatter *f, ostream *err) | |
931 | { | |
932 | std::map<string,string> m; | |
933 | if (int r = load_metadata(name, m, err)) | |
934 | return r; | |
935 | for (auto& p : m) { | |
936 | f->dump_string(p.first.c_str(), p.second); | |
937 | } | |
938 | return 0; | |
939 | } | |
31f18b77 | 940 |