]>
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) 2004-2006 Sage Weil <sage@newdream.net> | |
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 | */ | |
14 | ||
15 | #include <sstream> | |
16 | #include <boost/utility.hpp> | |
224ce89b | 17 | #include <boost/regex.hpp> |
7c673cae FG |
18 | |
19 | #include "MDSMonitor.h" | |
20 | #include "FSCommands.h" | |
21 | #include "Monitor.h" | |
22 | #include "MonitorDBStore.h" | |
23 | #include "OSDMonitor.h" | |
24 | #include "PGMonitor.h" | |
25 | ||
26 | #include "common/strtol.h" | |
27 | #include "common/perf_counters.h" | |
28 | #include "common/config.h" | |
29 | #include "common/cmdparse.h" | |
30 | #include "messages/MMDSMap.h" | |
31 | #include "messages/MFSMap.h" | |
32 | #include "messages/MFSMapUser.h" | |
33 | #include "messages/MMDSLoadTargets.h" | |
34 | #include "messages/MMonCommand.h" | |
35 | #include "messages/MGenericMessage.h" | |
36 | ||
37 | #include "include/assert.h" | |
38 | #include "include/str_list.h" | |
39 | #include "include/stringify.h" | |
40 | #include "mds/mdstypes.h" | |
41 | #include "Session.h" | |
42 | ||
43 | #define dout_subsys ceph_subsys_mon | |
44 | #undef dout_prefix | |
28e407b8 AA |
45 | #define dout_prefix _prefix(_dout, mon, get_fsmap()) |
46 | static ostream& _prefix(std::ostream *_dout, Monitor *mon, const FSMap& fsmap) { | |
7c673cae FG |
47 | return *_dout << "mon." << mon->name << "@" << mon->rank |
48 | << "(" << mon->get_state_name() | |
49 | << ").mds e" << fsmap.get_epoch() << " "; | |
50 | } | |
51 | ||
3efd9988 FG |
52 | static const string MDS_METADATA_PREFIX("mds_metadata"); |
53 | static const string MDS_HEALTH_PREFIX("mds_health"); | |
54 | ||
55 | ||
7c673cae FG |
56 | /* |
57 | * Specialized implementation of cmd_getval to allow us to parse | |
58 | * out strongly-typedef'd types | |
59 | */ | |
60 | template<> bool cmd_getval(CephContext *cct, const cmdmap_t& cmdmap, | |
31f18b77 | 61 | const std::string& k, mds_gid_t &val) |
7c673cae FG |
62 | { |
63 | return cmd_getval(cct, cmdmap, k, (int64_t&)val); | |
64 | } | |
65 | ||
66 | template<> bool cmd_getval(CephContext *cct, const cmdmap_t& cmdmap, | |
31f18b77 | 67 | const std::string& k, mds_rank_t &val) |
7c673cae FG |
68 | { |
69 | return cmd_getval(cct, cmdmap, k, (int64_t&)val); | |
70 | } | |
71 | ||
72 | template<> bool cmd_getval(CephContext *cct, const cmdmap_t& cmdmap, | |
31f18b77 | 73 | const std::string& k, MDSMap::DaemonState &val) |
7c673cae FG |
74 | { |
75 | return cmd_getval(cct, cmdmap, k, (int64_t&)val); | |
76 | } | |
77 | ||
7c673cae FG |
78 | // my methods |
79 | ||
28e407b8 | 80 | void MDSMonitor::print_map(const FSMap &m, int dbl) |
7c673cae FG |
81 | { |
82 | dout(dbl) << "print_map\n"; | |
83 | m.print(*_dout); | |
84 | *_dout << dendl; | |
85 | } | |
86 | ||
87 | // service methods | |
88 | void MDSMonitor::create_initial() | |
89 | { | |
90 | dout(10) << "create_initial" << dendl; | |
91 | } | |
92 | ||
3efd9988 FG |
93 | void MDSMonitor::get_store_prefixes(std::set<string>& s) |
94 | { | |
95 | s.insert(service_name); | |
96 | s.insert(MDS_METADATA_PREFIX); | |
97 | s.insert(MDS_HEALTH_PREFIX); | |
98 | } | |
7c673cae FG |
99 | |
100 | void MDSMonitor::update_from_paxos(bool *need_bootstrap) | |
101 | { | |
102 | version_t version = get_last_committed(); | |
28e407b8 | 103 | if (version == get_fsmap().epoch) |
7c673cae FG |
104 | return; |
105 | ||
106 | dout(10) << __func__ << " version " << version | |
28e407b8 AA |
107 | << ", my e " << get_fsmap().epoch << dendl; |
108 | assert(version > get_fsmap().epoch); | |
7c673cae | 109 | |
224ce89b WB |
110 | load_health(); |
111 | ||
7c673cae FG |
112 | // read and decode |
113 | bufferlist fsmap_bl; | |
114 | fsmap_bl.clear(); | |
115 | int err = get_version(version, fsmap_bl); | |
116 | assert(err == 0); | |
117 | ||
118 | assert(fsmap_bl.length() > 0); | |
119 | dout(10) << __func__ << " got " << version << dendl; | |
28e407b8 | 120 | PaxosFSMap::decode(fsmap_bl); |
7c673cae FG |
121 | |
122 | // new map | |
123 | dout(4) << "new map" << dendl; | |
28e407b8 | 124 | print_map(get_fsmap(), 0); |
7c673cae | 125 | if (!g_conf->mon_mds_skip_sanity) { |
28e407b8 | 126 | get_fsmap().sanity(); |
7c673cae FG |
127 | } |
128 | ||
129 | check_subs(); | |
130 | update_logger(); | |
131 | } | |
132 | ||
133 | void MDSMonitor::init() | |
134 | { | |
135 | (void)load_metadata(pending_metadata); | |
136 | } | |
137 | ||
138 | void MDSMonitor::create_pending() | |
139 | { | |
28e407b8 | 140 | auto &fsmap = PaxosFSMap::create_pending(); |
7c673cae | 141 | |
3efd9988 | 142 | if (mon->osdmon()->is_readable()) { |
28e407b8 AA |
143 | const auto &osdmap = mon->osdmon()->osdmap; |
144 | fsmap.sanitize([&osdmap](int64_t pool){return osdmap.have_pg_pool(pool);}); | |
3efd9988 FG |
145 | } |
146 | ||
28e407b8 | 147 | dout(10) << "create_pending e" << fsmap.epoch << dendl; |
7c673cae FG |
148 | } |
149 | ||
150 | void MDSMonitor::encode_pending(MonitorDBStore::TransactionRef t) | |
151 | { | |
28e407b8 AA |
152 | auto &pending = get_pending_fsmap_writeable(); |
153 | auto &epoch = pending.epoch; | |
7c673cae | 154 | |
28e407b8 | 155 | dout(10) << "encode_pending e" << epoch << dendl; |
7c673cae FG |
156 | |
157 | // print map iff 'debug mon = 30' or higher | |
28e407b8 | 158 | print_map(get_pending_fsmap(), 30); |
7c673cae | 159 | if (!g_conf->mon_mds_skip_sanity) { |
28e407b8 | 160 | pending.sanity(); |
7c673cae FG |
161 | } |
162 | ||
163 | // Set 'modified' on maps modified this epoch | |
28e407b8 AA |
164 | for (auto &p : pending.filesystems) { |
165 | if (p.second->mds_map.epoch == epoch) { | |
166 | p.second->mds_map.modified = ceph_clock_now(); | |
7c673cae FG |
167 | } |
168 | } | |
169 | ||
170 | // apply to paxos | |
28e407b8 AA |
171 | assert(get_last_committed() + 1 == pending.epoch); |
172 | bufferlist pending_bl; | |
173 | pending.encode(pending_bl, mon->get_quorum_con_features()); | |
7c673cae FG |
174 | |
175 | /* put everything in the transaction */ | |
28e407b8 AA |
176 | put_version(t, pending.epoch, pending_bl); |
177 | put_last_committed(t, pending.epoch); | |
7c673cae FG |
178 | |
179 | // Encode MDSHealth data | |
180 | for (std::map<uint64_t, MDSHealth>::iterator i = pending_daemon_health.begin(); | |
181 | i != pending_daemon_health.end(); ++i) { | |
182 | bufferlist bl; | |
183 | i->second.encode(bl); | |
184 | t->put(MDS_HEALTH_PREFIX, stringify(i->first), bl); | |
185 | } | |
186 | ||
187 | for (std::set<uint64_t>::iterator i = pending_daemon_health_rm.begin(); | |
188 | i != pending_daemon_health_rm.end(); ++i) { | |
189 | t->erase(MDS_HEALTH_PREFIX, stringify(*i)); | |
190 | } | |
191 | pending_daemon_health_rm.clear(); | |
192 | remove_from_metadata(t); | |
224ce89b WB |
193 | |
194 | // health | |
195 | health_check_map_t new_checks; | |
28e407b8 | 196 | const auto &info_map = pending.get_mds_info(); |
224ce89b WB |
197 | for (const auto &i : info_map) { |
198 | const auto &gid = i.first; | |
199 | const auto &info = i.second; | |
200 | if (pending_daemon_health_rm.count(gid)) { | |
201 | continue; | |
202 | } | |
203 | MDSHealth health; | |
204 | auto p = pending_daemon_health.find(gid); | |
205 | if (p != pending_daemon_health.end()) { | |
206 | health = p->second; | |
207 | } else { | |
208 | bufferlist bl; | |
209 | mon->store->get(MDS_HEALTH_PREFIX, stringify(gid), bl); | |
210 | if (!bl.length()) { | |
211 | derr << "Missing health data for MDS " << gid << dendl; | |
212 | continue; | |
213 | } | |
214 | bufferlist::iterator bl_i = bl.begin(); | |
215 | health.decode(bl_i); | |
216 | } | |
217 | for (const auto &metric : health.metrics) { | |
d2e6a577 | 218 | const int rank = info.rank; |
224ce89b WB |
219 | health_check_t *check = &new_checks.get_or_add( |
220 | mds_metric_name(metric.type), | |
221 | metric.sev, | |
222 | mds_metric_summary(metric.type)); | |
223 | ostringstream ss; | |
224 | ss << "mds" << info.name << "(mds." << rank << "): " << metric.message; | |
28e407b8 AA |
225 | bool first = true; |
226 | for (auto &p : metric.metadata) { | |
227 | if (first) { | |
228 | ss << " "; | |
229 | } else { | |
224ce89b | 230 | ss << ", "; |
28e407b8 AA |
231 | } |
232 | ss << p.first << ": " << p.second; | |
233 | first = false; | |
224ce89b WB |
234 | } |
235 | check->detail.push_back(ss.str()); | |
236 | } | |
237 | } | |
28e407b8 | 238 | pending.get_health_checks(&new_checks); |
224ce89b WB |
239 | for (auto& p : new_checks.checks) { |
240 | p.second.summary = boost::regex_replace( | |
241 | p.second.summary, | |
242 | boost::regex("%num%"), | |
243 | stringify(p.second.detail.size())); | |
244 | p.second.summary = boost::regex_replace( | |
245 | p.second.summary, | |
246 | boost::regex("%plurals%"), | |
247 | p.second.detail.size() > 1 ? "s" : ""); | |
248 | p.second.summary = boost::regex_replace( | |
249 | p.second.summary, | |
250 | boost::regex("%isorare%"), | |
251 | p.second.detail.size() > 1 ? "are" : "is"); | |
181888fb FG |
252 | p.second.summary = boost::regex_replace( |
253 | p.second.summary, | |
254 | boost::regex("%hasorhave%"), | |
255 | p.second.detail.size() > 1 ? "have" : "has"); | |
224ce89b WB |
256 | } |
257 | encode_health(new_checks, t); | |
7c673cae FG |
258 | } |
259 | ||
260 | version_t MDSMonitor::get_trim_to() | |
261 | { | |
262 | version_t floor = 0; | |
263 | if (g_conf->mon_mds_force_trim_to > 0 && | |
264 | g_conf->mon_mds_force_trim_to < (int)get_last_committed()) { | |
265 | floor = g_conf->mon_mds_force_trim_to; | |
266 | dout(10) << __func__ << " explicit mon_mds_force_trim_to = " | |
267 | << floor << dendl; | |
268 | } | |
269 | ||
270 | unsigned max = g_conf->mon_max_mdsmap_epochs; | |
271 | version_t last = get_last_committed(); | |
272 | ||
273 | if (last - get_first_committed() > max && floor < last - max) | |
274 | return last - max; | |
275 | return floor; | |
276 | } | |
277 | ||
278 | void MDSMonitor::update_logger() | |
279 | { | |
280 | dout(10) << "update_logger" << dendl; | |
281 | ||
28e407b8 AA |
282 | const auto &fsmap = get_fsmap(); |
283 | ||
7c673cae FG |
284 | uint64_t up = 0; |
285 | uint64_t in = 0; | |
286 | uint64_t failed = 0; | |
287 | for (const auto &i : fsmap.filesystems) { | |
288 | const MDSMap &mds_map = i.second->mds_map; | |
289 | ||
290 | up += mds_map.get_num_up_mds(); | |
291 | in += mds_map.get_num_in_mds(); | |
292 | failed += mds_map.get_num_failed_mds(); | |
293 | } | |
294 | mon->cluster_logger->set(l_cluster_num_mds_up, up); | |
295 | mon->cluster_logger->set(l_cluster_num_mds_in, in); | |
296 | mon->cluster_logger->set(l_cluster_num_mds_failed, failed); | |
297 | mon->cluster_logger->set(l_cluster_mds_epoch, fsmap.get_epoch()); | |
298 | } | |
299 | ||
300 | bool MDSMonitor::preprocess_query(MonOpRequestRef op) | |
301 | { | |
302 | op->mark_mdsmon_event(__func__); | |
303 | PaxosServiceMessage *m = static_cast<PaxosServiceMessage*>(op->get_req()); | |
304 | dout(10) << "preprocess_query " << *m << " from " << m->get_orig_source_inst() << dendl; | |
305 | ||
306 | switch (m->get_type()) { | |
307 | ||
308 | case MSG_MDS_BEACON: | |
309 | return preprocess_beacon(op); | |
310 | ||
311 | case MSG_MON_COMMAND: | |
312 | return preprocess_command(op); | |
313 | ||
314 | case MSG_MDS_OFFLOAD_TARGETS: | |
315 | return preprocess_offload_targets(op); | |
316 | ||
317 | default: | |
318 | ceph_abort(); | |
319 | return true; | |
320 | } | |
321 | } | |
322 | ||
323 | void MDSMonitor::_note_beacon(MMDSBeacon *m) | |
324 | { | |
325 | mds_gid_t gid = mds_gid_t(m->get_global_id()); | |
326 | version_t seq = m->get_seq(); | |
327 | ||
328 | dout(15) << "_note_beacon " << *m << " noting time" << dendl; | |
329 | last_beacon[gid].stamp = ceph_clock_now(); | |
330 | last_beacon[gid].seq = seq; | |
331 | } | |
332 | ||
333 | bool MDSMonitor::preprocess_beacon(MonOpRequestRef op) | |
334 | { | |
335 | op->mark_mdsmon_event(__func__); | |
336 | MMDSBeacon *m = static_cast<MMDSBeacon*>(op->get_req()); | |
337 | MDSMap::DaemonState state = m->get_state(); | |
338 | mds_gid_t gid = m->get_global_id(); | |
339 | version_t seq = m->get_seq(); | |
340 | MDSMap::mds_info_t info; | |
341 | epoch_t effective_epoch = 0; | |
342 | ||
28e407b8 AA |
343 | const auto &fsmap = get_working_fsmap(); |
344 | ||
7c673cae FG |
345 | // check privileges, ignore if fails |
346 | MonSession *session = m->get_session(); | |
347 | assert(session); | |
348 | if (!session->is_capable("mds", MON_CAP_X)) { | |
349 | dout(0) << "preprocess_beacon got MMDSBeacon from entity with insufficient privileges " | |
350 | << session->caps << dendl; | |
351 | goto ignore; | |
352 | } | |
353 | ||
354 | if (m->get_fsid() != mon->monmap->fsid) { | |
355 | dout(0) << "preprocess_beacon on fsid " << m->get_fsid() << " != " << mon->monmap->fsid << dendl; | |
356 | goto ignore; | |
357 | } | |
358 | ||
359 | dout(12) << "preprocess_beacon " << *m | |
360 | << " from " << m->get_orig_source_inst() | |
361 | << " " << m->get_compat() | |
362 | << dendl; | |
363 | ||
364 | // make sure the address has a port | |
365 | if (m->get_orig_source_addr().get_port() == 0) { | |
366 | dout(1) << " ignoring boot message without a port" << dendl; | |
367 | goto ignore; | |
368 | } | |
369 | ||
370 | // check compat | |
371 | if (!m->get_compat().writeable(fsmap.compat)) { | |
372 | dout(1) << " mds " << m->get_source_inst() << " can't write to fsmap " << fsmap.compat << dendl; | |
373 | goto ignore; | |
374 | } | |
375 | ||
376 | // fw to leader? | |
28e407b8 | 377 | if (!is_leader()) |
7c673cae FG |
378 | return false; |
379 | ||
380 | // booted, but not in map? | |
28e407b8 | 381 | if (!fsmap.gid_exists(gid)) { |
7c673cae FG |
382 | if (state != MDSMap::STATE_BOOT) { |
383 | dout(7) << "mds_beacon " << *m << " is not in fsmap (state " | |
384 | << ceph_mds_state_name(state) << ")" << dendl; | |
385 | ||
386 | MDSMap null_map; | |
387 | null_map.epoch = fsmap.epoch; | |
388 | null_map.compat = fsmap.compat; | |
389 | mon->send_reply(op, new MMDSMap(mon->monmap->fsid, &null_map)); | |
390 | return true; | |
391 | } else { | |
392 | return false; // not booted yet. | |
393 | } | |
394 | } | |
395 | dout(10) << __func__ << ": GID exists in map: " << gid << dendl; | |
28e407b8 | 396 | info = fsmap.get_info_gid(gid); |
7c673cae FG |
397 | |
398 | // old seq? | |
399 | if (info.state_seq > seq) { | |
400 | dout(7) << "mds_beacon " << *m << " has old seq, ignoring" << dendl; | |
401 | goto ignore; | |
402 | } | |
403 | ||
404 | // Work out the latest epoch that this daemon should have seen | |
405 | { | |
28e407b8 | 406 | fs_cluster_id_t fscid = fsmap.mds_roles.at(gid); |
7c673cae | 407 | if (fscid == FS_CLUSTER_ID_NONE) { |
28e407b8 | 408 | effective_epoch = fsmap.standby_epochs.at(gid); |
7c673cae | 409 | } else { |
28e407b8 | 410 | effective_epoch = fsmap.get_filesystem(fscid)->mds_map.epoch; |
7c673cae FG |
411 | } |
412 | if (effective_epoch != m->get_last_epoch_seen()) { | |
413 | dout(10) << "mds_beacon " << *m | |
414 | << " ignoring requested state, because mds hasn't seen latest map" << dendl; | |
415 | goto reply; | |
416 | } | |
417 | } | |
418 | ||
419 | if (info.laggy()) { | |
420 | _note_beacon(m); | |
421 | return false; // no longer laggy, need to update map. | |
422 | } | |
423 | if (state == MDSMap::STATE_BOOT) { | |
424 | // ignore, already booted. | |
425 | goto ignore; | |
426 | } | |
427 | // is there a state change here? | |
428 | if (info.state != state) { | |
429 | // legal state change? | |
430 | if ((info.state == MDSMap::STATE_STANDBY || | |
431 | info.state == MDSMap::STATE_STANDBY_REPLAY) && state > 0) { | |
432 | dout(10) << "mds_beacon mds can't activate itself (" << ceph_mds_state_name(info.state) | |
433 | << " -> " << ceph_mds_state_name(state) << ")" << dendl; | |
434 | goto reply; | |
435 | } | |
436 | ||
437 | if ((state == MDSMap::STATE_STANDBY || state == MDSMap::STATE_STANDBY_REPLAY) | |
438 | && info.rank != MDS_RANK_NONE) | |
439 | { | |
440 | dout(4) << "mds_beacon MDS can't go back into standby after taking rank: " | |
441 | "held rank " << info.rank << " while requesting state " | |
442 | << ceph_mds_state_name(state) << dendl; | |
443 | goto reply; | |
444 | } | |
445 | ||
446 | _note_beacon(m); | |
447 | return false; | |
448 | } | |
449 | ||
450 | // Comparing known daemon health with m->get_health() | |
451 | // and return false (i.e. require proposal) if they | |
452 | // do not match, to update our stored | |
453 | if (!(pending_daemon_health[gid] == m->get_health())) { | |
454 | dout(20) << __func__ << " health metrics for gid " << gid << " were updated" << dendl; | |
455 | _note_beacon(m); | |
456 | return false; | |
457 | } | |
458 | ||
459 | reply: | |
460 | // note time and reply | |
461 | assert(effective_epoch > 0); | |
462 | _note_beacon(m); | |
463 | mon->send_reply(op, | |
464 | new MMDSBeacon(mon->monmap->fsid, m->get_global_id(), m->get_name(), | |
465 | effective_epoch, state, seq, | |
466 | CEPH_FEATURES_SUPPORTED_DEFAULT)); | |
467 | return true; | |
468 | ||
469 | ignore: | |
470 | // I won't reply this beacon, drop it. | |
471 | mon->no_reply(op); | |
472 | return true; | |
473 | } | |
474 | ||
475 | bool MDSMonitor::preprocess_offload_targets(MonOpRequestRef op) | |
476 | { | |
477 | op->mark_mdsmon_event(__func__); | |
478 | MMDSLoadTargets *m = static_cast<MMDSLoadTargets*>(op->get_req()); | |
479 | dout(10) << "preprocess_offload_targets " << *m << " from " << m->get_orig_source() << dendl; | |
28e407b8 AA |
480 | |
481 | auto &fsmap = get_working_fsmap(); | |
7c673cae FG |
482 | |
483 | // check privileges, ignore message if fails | |
484 | MonSession *session = m->get_session(); | |
485 | if (!session) | |
486 | goto done; | |
487 | if (!session->is_capable("mds", MON_CAP_X)) { | |
488 | dout(0) << "preprocess_offload_targets got MMDSLoadTargets from entity with insufficient caps " | |
489 | << session->caps << dendl; | |
490 | goto done; | |
491 | } | |
492 | ||
493 | if (fsmap.gid_exists(m->global_id) && | |
494 | m->targets == fsmap.get_info_gid(m->global_id).export_targets) | |
495 | goto done; | |
496 | ||
497 | return false; | |
498 | ||
499 | done: | |
500 | return true; | |
501 | } | |
502 | ||
503 | ||
504 | bool MDSMonitor::prepare_update(MonOpRequestRef op) | |
505 | { | |
506 | op->mark_mdsmon_event(__func__); | |
507 | PaxosServiceMessage *m = static_cast<PaxosServiceMessage*>(op->get_req()); | |
508 | dout(7) << "prepare_update " << *m << dendl; | |
509 | ||
510 | switch (m->get_type()) { | |
511 | ||
512 | case MSG_MDS_BEACON: | |
513 | return prepare_beacon(op); | |
514 | ||
515 | case MSG_MON_COMMAND: | |
516 | return prepare_command(op); | |
517 | ||
518 | case MSG_MDS_OFFLOAD_TARGETS: | |
519 | return prepare_offload_targets(op); | |
520 | ||
521 | default: | |
522 | ceph_abort(); | |
523 | } | |
524 | ||
525 | return true; | |
526 | } | |
527 | ||
528 | bool MDSMonitor::prepare_beacon(MonOpRequestRef op) | |
529 | { | |
530 | op->mark_mdsmon_event(__func__); | |
531 | MMDSBeacon *m = static_cast<MMDSBeacon*>(op->get_req()); | |
532 | // -- this is an update -- | |
533 | dout(12) << "prepare_beacon " << *m << " from " << m->get_orig_source_inst() << dendl; | |
534 | entity_addr_t addr = m->get_orig_source_inst().addr; | |
535 | mds_gid_t gid = m->get_global_id(); | |
536 | MDSMap::DaemonState state = m->get_state(); | |
537 | version_t seq = m->get_seq(); | |
538 | ||
28e407b8 AA |
539 | auto &pending = get_pending_fsmap_writeable(); |
540 | ||
7c673cae FG |
541 | dout(20) << __func__ << " got health from gid " << gid << " with " << m->get_health().metrics.size() << " metrics." << dendl; |
542 | ||
543 | // Calculate deltas of health metrics created and removed | |
544 | // Do this by type rather than MDSHealthMetric equality, because messages can | |
545 | // change a lot when they include e.g. a number of items. | |
546 | const auto &old_health = pending_daemon_health[gid].metrics; | |
547 | const auto &new_health = m->get_health().metrics; | |
548 | ||
549 | std::set<mds_metric_t> old_types; | |
550 | for (const auto &i : old_health) { | |
551 | old_types.insert(i.type); | |
552 | } | |
553 | ||
554 | std::set<mds_metric_t> new_types; | |
555 | for (const auto &i : new_health) { | |
556 | new_types.insert(i.type); | |
557 | } | |
558 | ||
559 | for (const auto &new_metric: new_health) { | |
560 | if (old_types.count(new_metric.type) == 0) { | |
28e407b8 AA |
561 | dout(10) << "MDS health message (" << m->get_orig_source_inst().name |
562 | << "): " << new_metric.sev << " " << new_metric.message << dendl; | |
7c673cae FG |
563 | } |
564 | } | |
565 | ||
566 | // Log the disappearance of health messages at INFO | |
567 | for (const auto &old_metric : old_health) { | |
568 | if (new_types.count(old_metric.type) == 0) { | |
569 | mon->clog->info() << "MDS health message cleared (" | |
570 | << m->get_orig_source_inst().name << "): " << old_metric.message; | |
571 | } | |
572 | } | |
573 | ||
574 | // Store health | |
575 | pending_daemon_health[gid] = m->get_health(); | |
576 | ||
577 | // boot? | |
578 | if (state == MDSMap::STATE_BOOT) { | |
579 | // zap previous instance of this name? | |
580 | if (g_conf->mds_enforce_unique_name) { | |
581 | bool failed_mds = false; | |
28e407b8 | 582 | while (mds_gid_t existing = pending.find_mds_gid_by_name(m->get_name())) { |
7c673cae FG |
583 | if (!mon->osdmon()->is_writeable()) { |
584 | mon->osdmon()->wait_for_writeable(op, new C_RetryMessage(this, op)); | |
585 | return false; | |
586 | } | |
d2e6a577 | 587 | const MDSMap::mds_info_t &existing_info = |
28e407b8 | 588 | pending.get_info_gid(existing); |
d2e6a577 | 589 | mon->clog->info() << existing_info.human_name() << " restarted"; |
7c673cae FG |
590 | fail_mds_gid(existing); |
591 | failed_mds = true; | |
592 | } | |
593 | if (failed_mds) { | |
594 | assert(mon->osdmon()->is_writeable()); | |
595 | request_proposal(mon->osdmon()); | |
596 | } | |
597 | } | |
598 | ||
599 | // Add this daemon to the map | |
28e407b8 | 600 | if (pending.mds_roles.count(gid) == 0) { |
7c673cae FG |
601 | MDSMap::mds_info_t new_info; |
602 | new_info.global_id = gid; | |
603 | new_info.name = m->get_name(); | |
604 | new_info.addr = addr; | |
605 | new_info.mds_features = m->get_mds_features(); | |
606 | new_info.state = MDSMap::STATE_STANDBY; | |
607 | new_info.state_seq = seq; | |
608 | new_info.standby_for_rank = m->get_standby_for_rank(); | |
609 | new_info.standby_for_name = m->get_standby_for_name(); | |
610 | new_info.standby_for_fscid = m->get_standby_for_fscid(); | |
611 | new_info.standby_replay = m->get_standby_replay(); | |
28e407b8 | 612 | pending.insert(new_info); |
7c673cae FG |
613 | } |
614 | ||
615 | // Resolve standby_for_name to a rank | |
28e407b8 | 616 | const MDSMap::mds_info_t &info = pending.get_info_gid(gid); |
7c673cae | 617 | if (!info.standby_for_name.empty()) { |
28e407b8 | 618 | const MDSMap::mds_info_t *leaderinfo = pending.find_by_name( |
7c673cae FG |
619 | info.standby_for_name); |
620 | if (leaderinfo && (leaderinfo->rank >= 0)) { | |
28e407b8 AA |
621 | const auto &fscid = pending.mds_roles.at(leaderinfo->global_id); |
622 | const auto &fs = pending.get_filesystem(fscid); | |
7c673cae | 623 | |
28e407b8 | 624 | pending.modify_daemon(gid, [fscid, leaderinfo]( |
7c673cae FG |
625 | MDSMap::mds_info_t *info) { |
626 | info->standby_for_rank = leaderinfo->rank; | |
627 | info->standby_for_fscid = fscid; | |
628 | }); | |
629 | } | |
630 | } | |
631 | ||
632 | // initialize the beacon timer | |
633 | last_beacon[gid].stamp = ceph_clock_now(); | |
634 | last_beacon[gid].seq = seq; | |
635 | ||
636 | // new incompat? | |
28e407b8 AA |
637 | if (!pending.compat.writeable(m->get_compat())) { |
638 | dout(10) << " fsmap " << pending.compat | |
7c673cae FG |
639 | << " can't write to new mds' " << m->get_compat() |
640 | << ", updating fsmap and killing old mds's" | |
641 | << dendl; | |
28e407b8 | 642 | pending.update_compat(m->get_compat()); |
7c673cae FG |
643 | } |
644 | ||
645 | update_metadata(m->get_global_id(), m->get_sys_info()); | |
646 | } else { | |
647 | // state update | |
28e407b8 | 648 | const MDSMap::mds_info_t &info = pending.get_info_gid(gid); |
7c673cae FG |
649 | // Old MDS daemons don't mention that they're standby replay until |
650 | // after they've sent their boot beacon, so update this field. | |
651 | if (info.standby_replay != m->get_standby_replay()) { | |
28e407b8 | 652 | pending.modify_daemon(info.global_id, [&m]( |
7c673cae FG |
653 | MDSMap::mds_info_t *i) |
654 | { | |
655 | i->standby_replay = m->get_standby_replay(); | |
656 | }); | |
657 | } | |
658 | ||
659 | if (info.state == MDSMap::STATE_STOPPING && state != MDSMap::STATE_STOPPED ) { | |
660 | // we can't transition to any other states from STOPPING | |
661 | dout(0) << "got beacon for MDS in STATE_STOPPING, ignoring requested state change" | |
662 | << dendl; | |
663 | _note_beacon(m); | |
664 | return true; | |
665 | } | |
666 | ||
667 | if (info.laggy()) { | |
668 | dout(10) << "prepare_beacon clearing laggy flag on " << addr << dendl; | |
28e407b8 | 669 | pending.modify_daemon(info.global_id, [](MDSMap::mds_info_t *info) |
7c673cae FG |
670 | { |
671 | info->clear_laggy(); | |
672 | } | |
673 | ); | |
674 | } | |
675 | ||
676 | dout(10) << "prepare_beacon mds." << info.rank | |
677 | << " " << ceph_mds_state_name(info.state) | |
678 | << " -> " << ceph_mds_state_name(state) | |
679 | << " standby_for_rank=" << m->get_standby_for_rank() | |
680 | << dendl; | |
681 | if (state == MDSMap::STATE_STOPPED) { | |
28e407b8 AA |
682 | const auto fscid = pending.mds_roles.at(gid); |
683 | const auto &fs = pending.get_filesystem(fscid); | |
181888fb | 684 | |
d2e6a577 FG |
685 | mon->clog->info() << info.human_name() << " finished " |
686 | << "deactivating rank " << info.rank << " in filesystem " | |
687 | << fs->mds_map.fs_name << " (now has " | |
181888fb | 688 | << fs->mds_map.get_num_in_mds() - 1 << " ranks)"; |
d2e6a577 | 689 | |
28e407b8 | 690 | auto erased = pending.stop(gid); |
7c673cae FG |
691 | erased.push_back(gid); |
692 | ||
693 | for (const auto &erased_gid : erased) { | |
694 | last_beacon.erase(erased_gid); | |
695 | if (pending_daemon_health.count(erased_gid)) { | |
696 | pending_daemon_health.erase(erased_gid); | |
697 | pending_daemon_health_rm.insert(erased_gid); | |
698 | } | |
699 | } | |
d2e6a577 FG |
700 | |
701 | ||
7c673cae FG |
702 | } else if (state == MDSMap::STATE_DAMAGED) { |
703 | if (!mon->osdmon()->is_writeable()) { | |
704 | dout(4) << __func__ << ": DAMAGED from rank " << info.rank | |
705 | << " waiting for osdmon writeable to blacklist it" << dendl; | |
706 | mon->osdmon()->wait_for_writeable(op, new C_RetryMessage(this, op)); | |
707 | return false; | |
708 | } | |
709 | ||
710 | // Record this MDS rank as damaged, so that other daemons | |
711 | // won't try to run it. | |
712 | dout(4) << __func__ << ": marking rank " | |
713 | << info.rank << " damaged" << dendl; | |
714 | ||
715 | utime_t until = ceph_clock_now(); | |
b32b8144 | 716 | until += g_conf->get_val<double>("mon_mds_blacklist_interval"); |
7c673cae FG |
717 | const auto blacklist_epoch = mon->osdmon()->blacklist(info.addr, until); |
718 | request_proposal(mon->osdmon()); | |
28e407b8 | 719 | pending.damaged(gid, blacklist_epoch); |
7c673cae FG |
720 | last_beacon.erase(gid); |
721 | ||
722 | // Respond to MDS, so that it knows it can continue to shut down | |
723 | mon->send_reply(op, | |
724 | new MMDSBeacon( | |
725 | mon->monmap->fsid, m->get_global_id(), | |
28e407b8 | 726 | m->get_name(), pending.get_epoch(), state, seq, |
7c673cae FG |
727 | CEPH_FEATURES_SUPPORTED_DEFAULT)); |
728 | } else if (state == MDSMap::STATE_DNE) { | |
729 | if (!mon->osdmon()->is_writeable()) { | |
730 | dout(4) << __func__ << ": DNE from rank " << info.rank | |
731 | << " waiting for osdmon writeable to blacklist it" << dendl; | |
732 | mon->osdmon()->wait_for_writeable(op, new C_RetryMessage(this, op)); | |
733 | return false; | |
734 | } | |
735 | ||
736 | fail_mds_gid(gid); | |
737 | assert(mon->osdmon()->is_writeable()); | |
738 | request_proposal(mon->osdmon()); | |
739 | ||
740 | // Respond to MDS, so that it knows it can continue to shut down | |
741 | mon->send_reply(op, | |
742 | new MMDSBeacon( | |
743 | mon->monmap->fsid, m->get_global_id(), | |
28e407b8 | 744 | m->get_name(), pending.get_epoch(), state, seq, |
7c673cae FG |
745 | CEPH_FEATURES_SUPPORTED_DEFAULT)); |
746 | } else if (info.state == MDSMap::STATE_STANDBY && state != info.state) { | |
747 | // Standby daemons should never modify their own | |
748 | // state. Reject any attempts to do so. | |
749 | derr << "standby " << gid << " attempted to change state to " | |
750 | << ceph_mds_state_name(state) << ", rejecting" << dendl; | |
751 | return true; | |
752 | } else if (info.state != MDSMap::STATE_STANDBY && state != info.state && | |
753 | !MDSMap::state_transition_valid(info.state, state)) { | |
754 | // Validate state transitions for daemons that hold a rank | |
755 | derr << "daemon " << gid << " (rank " << info.rank << ") " | |
756 | << "reported invalid state transition " | |
757 | << ceph_mds_state_name(info.state) << " -> " | |
758 | << ceph_mds_state_name(state) << dendl; | |
759 | return true; | |
760 | } else { | |
b32b8144 | 761 | if (info.state != MDSMap::STATE_ACTIVE && state == MDSMap::STATE_ACTIVE) { |
28e407b8 AA |
762 | const auto &fscid = pending.mds_roles.at(gid); |
763 | const auto &fs = pending.get_filesystem(fscid); | |
d2e6a577 FG |
764 | mon->clog->info() << info.human_name() << " is now active in " |
765 | << "filesystem " << fs->mds_map.fs_name << " as rank " | |
766 | << info.rank; | |
767 | } | |
b32b8144 FG |
768 | |
769 | // Made it through special cases and validations, record the | |
770 | // daemon's reported state to the FSMap. | |
28e407b8 | 771 | pending.modify_daemon(gid, [state, seq](MDSMap::mds_info_t *info) { |
b32b8144 FG |
772 | info->state = state; |
773 | info->state_seq = seq; | |
774 | }); | |
7c673cae FG |
775 | } |
776 | } | |
777 | ||
778 | dout(7) << "prepare_beacon pending map now:" << dendl; | |
28e407b8 | 779 | print_map(pending); |
7c673cae FG |
780 | |
781 | wait_for_finished_proposal(op, new FunctionContext([op, this](int r){ | |
782 | if (r >= 0) | |
783 | _updated(op); // success | |
784 | else if (r == -ECANCELED) { | |
785 | mon->no_reply(op); | |
786 | } else { | |
787 | dispatch(op); // try again | |
788 | } | |
789 | })); | |
790 | ||
791 | return true; | |
792 | } | |
793 | ||
794 | bool MDSMonitor::prepare_offload_targets(MonOpRequestRef op) | |
795 | { | |
28e407b8 AA |
796 | auto &pending = get_pending_fsmap_writeable(); |
797 | ||
7c673cae FG |
798 | op->mark_mdsmon_event(__func__); |
799 | MMDSLoadTargets *m = static_cast<MMDSLoadTargets*>(op->get_req()); | |
800 | mds_gid_t gid = m->global_id; | |
28e407b8 | 801 | if (pending.gid_has_rank(gid)) { |
7c673cae | 802 | dout(10) << "prepare_offload_targets " << gid << " " << m->targets << dendl; |
28e407b8 | 803 | pending.update_export_targets(gid, m->targets); |
7c673cae FG |
804 | } else { |
805 | dout(10) << "prepare_offload_targets " << gid << " not in map" << dendl; | |
806 | } | |
807 | return true; | |
808 | } | |
809 | ||
810 | bool MDSMonitor::should_propose(double& delay) | |
811 | { | |
812 | // delegate to PaxosService to assess whether we should propose | |
813 | return PaxosService::should_propose(delay); | |
814 | } | |
815 | ||
816 | void MDSMonitor::_updated(MonOpRequestRef op) | |
817 | { | |
28e407b8 | 818 | const auto &fsmap = get_fsmap(); |
7c673cae FG |
819 | op->mark_mdsmon_event(__func__); |
820 | MMDSBeacon *m = static_cast<MMDSBeacon*>(op->get_req()); | |
821 | dout(10) << "_updated " << m->get_orig_source() << " " << *m << dendl; | |
d2e6a577 | 822 | mon->clog->debug() << m->get_orig_source_inst() << " " |
7c673cae FG |
823 | << ceph_mds_state_name(m->get_state()); |
824 | ||
825 | if (m->get_state() == MDSMap::STATE_STOPPED) { | |
826 | // send the map manually (they're out of the map, so they won't get it automatic) | |
827 | MDSMap null_map; | |
828 | null_map.epoch = fsmap.epoch; | |
829 | null_map.compat = fsmap.compat; | |
830 | mon->send_reply(op, new MMDSMap(mon->monmap->fsid, &null_map)); | |
831 | } else { | |
832 | mon->send_reply(op, new MMDSBeacon(mon->monmap->fsid, | |
833 | m->get_global_id(), | |
834 | m->get_name(), | |
835 | fsmap.get_epoch(), | |
836 | m->get_state(), | |
837 | m->get_seq(), | |
838 | CEPH_FEATURES_SUPPORTED_DEFAULT)); | |
839 | } | |
840 | } | |
841 | ||
842 | void MDSMonitor::on_active() | |
843 | { | |
844 | tick(); | |
845 | update_logger(); | |
846 | ||
28e407b8 AA |
847 | if (is_leader()) { |
848 | mon->clog->debug() << "fsmap " << get_fsmap(); | |
224ce89b | 849 | } |
7c673cae FG |
850 | } |
851 | ||
852 | void MDSMonitor::get_health(list<pair<health_status_t, string> >& summary, | |
853 | list<pair<health_status_t, string> > *detail, | |
854 | CephContext* cct) const | |
855 | { | |
28e407b8 AA |
856 | const auto &fsmap = get_fsmap(); |
857 | ||
7c673cae FG |
858 | fsmap.get_health(summary, detail); |
859 | ||
860 | // For each MDS GID... | |
28e407b8 | 861 | const auto &info_map = fsmap.get_mds_info(); |
7c673cae FG |
862 | for (const auto &i : info_map) { |
863 | const auto &gid = i.first; | |
864 | const auto &info = i.second; | |
865 | ||
866 | // Decode MDSHealth | |
867 | bufferlist bl; | |
868 | mon->store->get(MDS_HEALTH_PREFIX, stringify(gid), bl); | |
869 | if (!bl.length()) { | |
870 | derr << "Missing health data for MDS " << gid << dendl; | |
871 | continue; | |
872 | } | |
873 | MDSHealth health; | |
874 | bufferlist::iterator bl_i = bl.begin(); | |
875 | health.decode(bl_i); | |
876 | ||
877 | for (const auto &metric : health.metrics) { | |
d2e6a577 | 878 | const int rank = info.rank; |
7c673cae FG |
879 | std::ostringstream message; |
880 | message << "mds" << rank << ": " << metric.message; | |
881 | summary.push_back(std::make_pair(metric.sev, message.str())); | |
882 | ||
883 | if (detail) { | |
884 | // There is no way for us to clealy associate detail entries with summary entries (#7192), so | |
885 | // we duplicate the summary message in the detail string and tag the metadata on. | |
886 | std::ostringstream detail_message; | |
887 | detail_message << message.str(); | |
888 | if (metric.metadata.size()) { | |
889 | detail_message << "("; | |
890 | auto k = metric.metadata.begin(); | |
891 | while (k != metric.metadata.end()) { | |
892 | detail_message << k->first << ": " << k->second; | |
893 | if (boost::next(k) != metric.metadata.end()) { | |
894 | detail_message << ", "; | |
895 | } | |
896 | ++k; | |
897 | } | |
898 | detail_message << ")"; | |
899 | } | |
900 | detail->push_back(std::make_pair(metric.sev, detail_message.str())); | |
901 | } | |
902 | } | |
903 | } | |
904 | } | |
905 | ||
906 | void MDSMonitor::dump_info(Formatter *f) | |
907 | { | |
908 | f->open_object_section("fsmap"); | |
28e407b8 | 909 | get_fsmap().dump(f); |
7c673cae FG |
910 | f->close_section(); |
911 | ||
912 | f->dump_unsigned("mdsmap_first_committed", get_first_committed()); | |
913 | f->dump_unsigned("mdsmap_last_committed", get_last_committed()); | |
914 | } | |
915 | ||
916 | bool MDSMonitor::preprocess_command(MonOpRequestRef op) | |
917 | { | |
918 | op->mark_mdsmon_event(__func__); | |
919 | MMonCommand *m = static_cast<MMonCommand*>(op->get_req()); | |
920 | int r = -1; | |
921 | bufferlist rdata; | |
922 | stringstream ss, ds; | |
923 | ||
924 | map<string, cmd_vartype> cmdmap; | |
28e407b8 AA |
925 | const auto &fsmap = get_working_fsmap(); |
926 | ||
7c673cae FG |
927 | if (!cmdmap_from_json(m->cmd, &cmdmap, ss)) { |
928 | // ss has reason for failure | |
929 | string rs = ss.str(); | |
930 | mon->reply_command(op, -EINVAL, rs, rdata, get_last_committed()); | |
931 | return true; | |
932 | } | |
933 | ||
934 | string prefix; | |
935 | cmd_getval(g_ceph_context, cmdmap, "prefix", prefix); | |
936 | string format; | |
937 | cmd_getval(g_ceph_context, cmdmap, "format", format, string("plain")); | |
938 | boost::scoped_ptr<Formatter> f(Formatter::create(format)); | |
939 | ||
940 | MonSession *session = m->get_session(); | |
941 | if (!session) { | |
942 | mon->reply_command(op, -EACCES, "access denied", rdata, get_last_committed()); | |
943 | return true; | |
944 | } | |
945 | ||
946 | if (prefix == "mds stat") { | |
947 | if (f) { | |
948 | f->open_object_section("mds_stat"); | |
949 | dump_info(f.get()); | |
950 | f->close_section(); | |
951 | f->flush(ds); | |
952 | } else { | |
953 | ds << fsmap; | |
954 | } | |
955 | r = 0; | |
956 | } else if (prefix == "mds dump") { | |
957 | int64_t epocharg; | |
958 | epoch_t epoch; | |
959 | ||
28e407b8 AA |
960 | const FSMap *fsmapp = &get_fsmap(); |
961 | FSMap dummy; | |
7c673cae FG |
962 | if (cmd_getval(g_ceph_context, cmdmap, "epoch", epocharg)) { |
963 | epoch = epocharg; | |
964 | bufferlist b; | |
965 | int err = get_version(epoch, b); | |
966 | if (err == -ENOENT) { | |
7c673cae | 967 | r = -ENOENT; |
28e407b8 | 968 | goto out; |
7c673cae FG |
969 | } else { |
970 | assert(err == 0); | |
971 | assert(b.length()); | |
28e407b8 AA |
972 | dummy.decode(b); |
973 | fsmapp = &dummy; | |
7c673cae FG |
974 | } |
975 | } | |
c07f9fc5 | 976 | |
28e407b8 AA |
977 | stringstream ds; |
978 | const MDSMap *mdsmapp = nullptr; | |
979 | MDSMap blank; | |
980 | blank.epoch = fsmapp->epoch; | |
981 | if (fsmapp->legacy_client_fscid != FS_CLUSTER_ID_NONE) { | |
982 | mdsmapp = &fsmapp->filesystems.at(fsmapp->legacy_client_fscid)->mds_map; | |
983 | } else { | |
984 | mdsmapp = ␣ | |
7c673cae | 985 | } |
28e407b8 AA |
986 | if (f != NULL) { |
987 | f->open_object_section("mdsmap"); | |
988 | mdsmapp->dump(f.get()); | |
989 | f->close_section(); | |
990 | f->flush(ds); | |
991 | r = 0; | |
992 | } else { | |
993 | mdsmapp->print(ds); | |
994 | r = 0; | |
995 | } | |
996 | ||
997 | rdata.append(ds); | |
998 | ss << "dumped fsmap epoch " << fsmapp->get_epoch(); | |
7c673cae FG |
999 | } else if (prefix == "fs dump") { |
1000 | int64_t epocharg; | |
1001 | epoch_t epoch; | |
1002 | ||
28e407b8 AA |
1003 | const FSMap *fsmapp = &get_fsmap(); |
1004 | FSMap dummy; | |
7c673cae FG |
1005 | if (cmd_getval(g_ceph_context, cmdmap, "epoch", epocharg)) { |
1006 | epoch = epocharg; | |
1007 | bufferlist b; | |
1008 | int err = get_version(epoch, b); | |
1009 | if (err == -ENOENT) { | |
7c673cae | 1010 | r = -ENOENT; |
28e407b8 | 1011 | goto out; |
7c673cae FG |
1012 | } else { |
1013 | assert(err == 0); | |
1014 | assert(b.length()); | |
28e407b8 AA |
1015 | dummy.decode(b); |
1016 | fsmapp = &dummy; | |
7c673cae FG |
1017 | } |
1018 | } | |
c07f9fc5 | 1019 | |
28e407b8 AA |
1020 | stringstream ds; |
1021 | if (f != NULL) { | |
1022 | f->open_object_section("fsmap"); | |
1023 | fsmapp->dump(f.get()); | |
1024 | f->close_section(); | |
1025 | f->flush(ds); | |
1026 | r = 0; | |
1027 | } else { | |
1028 | fsmapp->print(ds); | |
1029 | r = 0; | |
7c673cae | 1030 | } |
28e407b8 AA |
1031 | |
1032 | rdata.append(ds); | |
1033 | ss << "dumped fsmap epoch " << fsmapp->get_epoch(); | |
7c673cae FG |
1034 | } else if (prefix == "mds metadata") { |
1035 | if (!f) | |
1036 | f.reset(Formatter::create("json-pretty")); | |
1037 | ||
1038 | string who; | |
1039 | bool all = !cmd_getval(g_ceph_context, cmdmap, "who", who); | |
1040 | dout(1) << "all = " << all << dendl; | |
1041 | if (all) { | |
1042 | r = 0; | |
1043 | // Dump all MDSs' metadata | |
1044 | const auto all_info = fsmap.get_mds_info(); | |
1045 | ||
1046 | f->open_array_section("mds_metadata"); | |
1047 | for(const auto &i : all_info) { | |
1048 | const auto &info = i.second; | |
1049 | ||
1050 | f->open_object_section("mds"); | |
1051 | f->dump_string("name", info.name); | |
1052 | std::ostringstream get_err; | |
1053 | r = dump_metadata(info.name, f.get(), get_err); | |
1054 | if (r == -EINVAL || r == -ENOENT) { | |
1055 | // Drop error, list what metadata we do have | |
1056 | dout(1) << get_err.str() << dendl; | |
1057 | r = 0; | |
1058 | } else if (r != 0) { | |
1059 | derr << "Unexpected error reading metadata: " << cpp_strerror(r) | |
1060 | << dendl; | |
1061 | ss << get_err.str(); | |
c07f9fc5 | 1062 | f->close_section(); |
7c673cae FG |
1063 | break; |
1064 | } | |
1065 | f->close_section(); | |
1066 | } | |
1067 | f->close_section(); | |
1068 | } else { | |
1069 | // Dump a single daemon's metadata | |
1070 | f->open_object_section("mds_metadata"); | |
1071 | r = dump_metadata(who, f.get(), ss); | |
1072 | f->close_section(); | |
1073 | } | |
1074 | f->flush(ds); | |
31f18b77 FG |
1075 | } else if (prefix == "mds versions") { |
1076 | if (!f) | |
1077 | f.reset(Formatter::create("json-pretty")); | |
1078 | count_metadata("ceph_version", f.get()); | |
1079 | f->flush(ds); | |
1080 | r = 0; | |
1081 | } else if (prefix == "mds count-metadata") { | |
1082 | if (!f) | |
1083 | f.reset(Formatter::create("json-pretty")); | |
1084 | string field; | |
1085 | cmd_getval(g_ceph_context, cmdmap, "property", field); | |
1086 | count_metadata(field, f.get()); | |
1087 | f->flush(ds); | |
1088 | r = 0; | |
7c673cae FG |
1089 | } else if (prefix == "mds getmap") { |
1090 | epoch_t e; | |
1091 | int64_t epocharg; | |
1092 | bufferlist b; | |
1093 | if (cmd_getval(g_ceph_context, cmdmap, "epoch", epocharg)) { | |
1094 | e = epocharg; | |
1095 | int err = get_version(e, b); | |
1096 | if (err == -ENOENT) { | |
1097 | r = -ENOENT; | |
1098 | } else { | |
1099 | assert(err == 0); | |
1100 | assert(b.length()); | |
1101 | FSMap mm; | |
1102 | mm.decode(b); | |
1103 | mm.encode(rdata, m->get_connection()->get_features()); | |
1104 | ss << "got fsmap epoch " << mm.get_epoch(); | |
1105 | r = 0; | |
1106 | } | |
1107 | } else { | |
1108 | fsmap.encode(rdata, m->get_connection()->get_features()); | |
1109 | ss << "got fsmap epoch " << fsmap.get_epoch(); | |
1110 | r = 0; | |
1111 | } | |
1112 | } else if (prefix == "mds compat show") { | |
1113 | if (f) { | |
1114 | f->open_object_section("mds_compat"); | |
1115 | fsmap.compat.dump(f.get()); | |
1116 | f->close_section(); | |
1117 | f->flush(ds); | |
1118 | } else { | |
1119 | ds << fsmap.compat; | |
1120 | } | |
1121 | r = 0; | |
1122 | } else if (prefix == "fs get") { | |
1123 | string fs_name; | |
1124 | cmd_getval(g_ceph_context, cmdmap, "fs_name", fs_name); | |
28e407b8 | 1125 | const auto &fs = fsmap.get_filesystem(fs_name); |
7c673cae FG |
1126 | if (fs == nullptr) { |
1127 | ss << "filesystem '" << fs_name << "' not found"; | |
1128 | r = -ENOENT; | |
1129 | } else { | |
1130 | if (f != nullptr) { | |
1131 | f->open_object_section("filesystem"); | |
1132 | fs->dump(f.get()); | |
1133 | f->close_section(); | |
1134 | f->flush(ds); | |
1135 | r = 0; | |
1136 | } else { | |
1137 | fs->print(ds); | |
1138 | r = 0; | |
1139 | } | |
1140 | } | |
1141 | } else if (prefix == "fs ls") { | |
1142 | if (f) { | |
1143 | f->open_array_section("filesystems"); | |
1144 | { | |
28e407b8 AA |
1145 | for (const auto &p : fsmap.filesystems) { |
1146 | const auto &fs = p.second; | |
7c673cae FG |
1147 | f->open_object_section("filesystem"); |
1148 | { | |
1149 | const MDSMap &mds_map = fs->mds_map; | |
1150 | f->dump_string("name", mds_map.fs_name); | |
1151 | /* Output both the names and IDs of pools, for use by | |
1152 | * humans and machines respectively */ | |
1153 | f->dump_string("metadata_pool", mon->osdmon()->osdmap.get_pool_name( | |
1154 | mds_map.metadata_pool)); | |
1155 | f->dump_int("metadata_pool_id", mds_map.metadata_pool); | |
1156 | f->open_array_section("data_pool_ids"); | |
1157 | { | |
1158 | for (auto dpi = mds_map.data_pools.begin(); | |
1159 | dpi != mds_map.data_pools.end(); ++dpi) { | |
1160 | f->dump_int("data_pool_id", *dpi); | |
1161 | } | |
1162 | } | |
1163 | f->close_section(); | |
1164 | ||
1165 | f->open_array_section("data_pools"); | |
1166 | { | |
1167 | for (auto dpi = mds_map.data_pools.begin(); | |
1168 | dpi != mds_map.data_pools.end(); ++dpi) { | |
1169 | const auto &name = mon->osdmon()->osdmap.get_pool_name( | |
1170 | *dpi); | |
1171 | f->dump_string("data_pool", name); | |
1172 | } | |
1173 | } | |
1174 | ||
1175 | f->close_section(); | |
1176 | } | |
1177 | f->close_section(); | |
1178 | } | |
1179 | } | |
1180 | f->close_section(); | |
1181 | f->flush(ds); | |
1182 | } else { | |
28e407b8 AA |
1183 | for (const auto &p : fsmap.filesystems) { |
1184 | const auto &fs = p.second; | |
7c673cae FG |
1185 | const MDSMap &mds_map = fs->mds_map; |
1186 | const string &md_pool_name = mon->osdmon()->osdmap.get_pool_name( | |
1187 | mds_map.metadata_pool); | |
1188 | ||
1189 | ds << "name: " << mds_map.fs_name << ", metadata pool: " | |
1190 | << md_pool_name << ", data pools: ["; | |
31f18b77 FG |
1191 | for (auto dpi : mds_map.data_pools) { |
1192 | const string &pool_name = mon->osdmon()->osdmap.get_pool_name(dpi); | |
7c673cae FG |
1193 | ds << pool_name << " "; |
1194 | } | |
1195 | ds << "]" << std::endl; | |
1196 | } | |
1197 | ||
1198 | if (fsmap.filesystems.empty()) { | |
1199 | ds << "No filesystems enabled" << std::endl; | |
1200 | } | |
1201 | } | |
1202 | r = 0; | |
1203 | } | |
1204 | ||
28e407b8 | 1205 | out: |
7c673cae FG |
1206 | if (r != -1) { |
1207 | rdata.append(ds); | |
1208 | string rs; | |
1209 | getline(ss, rs); | |
1210 | mon->reply_command(op, r, rs, rdata, get_last_committed()); | |
1211 | return true; | |
1212 | } else | |
1213 | return false; | |
1214 | } | |
1215 | ||
1216 | bool MDSMonitor::fail_mds_gid(mds_gid_t gid) | |
1217 | { | |
28e407b8 AA |
1218 | auto &pending = get_pending_fsmap_writeable(); |
1219 | ||
1220 | const MDSMap::mds_info_t &info = pending.get_info_gid(gid); | |
7c673cae FG |
1221 | dout(10) << "fail_mds_gid " << gid << " mds." << info.name << " role " << info.rank << dendl; |
1222 | ||
1223 | epoch_t blacklist_epoch = 0; | |
1224 | if (info.rank >= 0 && info.state != MDSMap::STATE_STANDBY_REPLAY) { | |
1225 | utime_t until = ceph_clock_now(); | |
b32b8144 | 1226 | until += g_conf->get_val<double>("mon_mds_blacklist_interval"); |
7c673cae FG |
1227 | blacklist_epoch = mon->osdmon()->blacklist(info.addr, until); |
1228 | } | |
1229 | ||
28e407b8 | 1230 | pending.erase(gid, blacklist_epoch); |
7c673cae FG |
1231 | last_beacon.erase(gid); |
1232 | if (pending_daemon_health.count(gid)) { | |
1233 | pending_daemon_health.erase(gid); | |
1234 | pending_daemon_health_rm.insert(gid); | |
1235 | } | |
1236 | ||
1237 | return blacklist_epoch != 0; | |
1238 | } | |
1239 | ||
1240 | mds_gid_t MDSMonitor::gid_from_arg(const std::string& arg, std::ostream &ss) | |
1241 | { | |
28e407b8 | 1242 | const auto &fsmap = get_working_fsmap(); |
31f18b77 | 1243 | |
7c673cae FG |
1244 | // Try parsing as a role |
1245 | mds_role_t role; | |
1246 | std::ostringstream ignore_err; // Don't spam 'ss' with parse_role errors | |
1247 | int r = parse_role(arg, &role, ignore_err); | |
1248 | if (r == 0) { | |
1249 | // See if a GID is assigned to this role | |
28e407b8 | 1250 | const auto &fs = fsmap.get_filesystem(role.fscid); |
7c673cae FG |
1251 | assert(fs != nullptr); // parse_role ensures it exists |
1252 | if (fs->mds_map.is_up(role.rank)) { | |
1253 | dout(10) << __func__ << ": validated rank/GID " << role | |
1254 | << " as a rank" << dendl; | |
1255 | return fs->mds_map.get_mds_info(role.rank).global_id; | |
1256 | } | |
1257 | } | |
1258 | ||
1259 | // Try parsing as a gid | |
1260 | std::string err; | |
1261 | unsigned long long maybe_gid = strict_strtoll(arg.c_str(), 10, &err); | |
1262 | if (!err.empty()) { | |
1263 | // Not a role or a GID, try as a daemon name | |
28e407b8 | 1264 | const MDSMap::mds_info_t *mds_info = fsmap.find_by_name(arg); |
7c673cae FG |
1265 | if (!mds_info) { |
1266 | ss << "MDS named '" << arg | |
1267 | << "' does not exist, or is not up"; | |
1268 | return MDS_GID_NONE; | |
1269 | } | |
1270 | dout(10) << __func__ << ": resolved MDS name '" << arg | |
1271 | << "' to GID " << mds_info->global_id << dendl; | |
1272 | return mds_info->global_id; | |
1273 | } else { | |
1274 | // Not a role, but parses as a an integer, might be a GID | |
1275 | dout(10) << __func__ << ": treating MDS reference '" << arg | |
1276 | << "' as an integer " << maybe_gid << dendl; | |
31f18b77 | 1277 | |
28e407b8 | 1278 | if (fsmap.gid_exists(mds_gid_t(maybe_gid))) { |
31f18b77 | 1279 | return mds_gid_t(maybe_gid); |
7c673cae FG |
1280 | } |
1281 | } | |
1282 | ||
1283 | dout(1) << __func__ << ": rank/GID " << arg | |
1284 | << " not a existent rank or GID" << dendl; | |
1285 | return MDS_GID_NONE; | |
1286 | } | |
1287 | ||
d2e6a577 FG |
1288 | int MDSMonitor::fail_mds(std::ostream &ss, const std::string &arg, |
1289 | MDSMap::mds_info_t *failed_info) | |
7c673cae | 1290 | { |
d2e6a577 FG |
1291 | assert(failed_info != nullptr); |
1292 | ||
7c673cae FG |
1293 | mds_gid_t gid = gid_from_arg(arg, ss); |
1294 | if (gid == MDS_GID_NONE) { | |
1295 | return 0; | |
1296 | } | |
1297 | if (!mon->osdmon()->is_writeable()) { | |
1298 | return -EAGAIN; | |
1299 | } | |
d2e6a577 FG |
1300 | |
1301 | // Take a copy of the info before removing the MDS from the map, | |
1302 | // so that the caller knows which mds (if any) they ended up removing. | |
28e407b8 | 1303 | *failed_info = get_pending_fsmap().get_info_gid(gid); |
d2e6a577 | 1304 | |
7c673cae FG |
1305 | fail_mds_gid(gid); |
1306 | ss << "failed mds gid " << gid; | |
1307 | assert(mon->osdmon()->is_writeable()); | |
1308 | request_proposal(mon->osdmon()); | |
1309 | return 0; | |
1310 | } | |
1311 | ||
1312 | bool MDSMonitor::prepare_command(MonOpRequestRef op) | |
1313 | { | |
1314 | op->mark_mdsmon_event(__func__); | |
1315 | MMonCommand *m = static_cast<MMonCommand*>(op->get_req()); | |
1316 | int r = -EINVAL; | |
1317 | stringstream ss; | |
1318 | bufferlist rdata; | |
1319 | ||
1320 | map<string, cmd_vartype> cmdmap; | |
1321 | if (!cmdmap_from_json(m->cmd, &cmdmap, ss)) { | |
1322 | string rs = ss.str(); | |
1323 | mon->reply_command(op, -EINVAL, rs, rdata, get_last_committed()); | |
1324 | return true; | |
1325 | } | |
1326 | ||
1327 | string prefix; | |
1328 | cmd_getval(g_ceph_context, cmdmap, "prefix", prefix); | |
1329 | ||
1330 | /* Refuse access if message not associated with a valid session */ | |
1331 | MonSession *session = m->get_session(); | |
1332 | if (!session) { | |
1333 | mon->reply_command(op, -EACCES, "access denied", rdata, get_last_committed()); | |
1334 | return true; | |
1335 | } | |
1336 | ||
28e407b8 AA |
1337 | auto &pending = get_pending_fsmap_writeable(); |
1338 | ||
c07f9fc5 | 1339 | bool batched_propose = false; |
28e407b8 | 1340 | for (const auto &h : handlers) { |
7c673cae | 1341 | if (h->can_handle(prefix)) { |
c07f9fc5 FG |
1342 | batched_propose = h->batched_propose(); |
1343 | if (batched_propose) { | |
1344 | paxos->plug(); | |
1345 | } | |
28e407b8 | 1346 | r = h->handle(mon, pending, op, cmdmap, ss); |
c07f9fc5 FG |
1347 | if (batched_propose) { |
1348 | paxos->unplug(); | |
1349 | } | |
1350 | ||
7c673cae FG |
1351 | if (r == -EAGAIN) { |
1352 | // message has been enqueued for retry; return. | |
1353 | dout(4) << __func__ << " enqueue for retry by prepare_command" << dendl; | |
1354 | return false; | |
1355 | } else { | |
1356 | if (r == 0) { | |
1357 | // On successful updates, print the updated map | |
28e407b8 | 1358 | print_map(pending); |
7c673cae FG |
1359 | } |
1360 | // Successful or not, we're done: respond. | |
1361 | goto out; | |
1362 | } | |
1363 | } | |
1364 | } | |
1365 | ||
1366 | r = filesystem_command(op, prefix, cmdmap, ss); | |
1367 | if (r >= 0) { | |
1368 | goto out; | |
1369 | } else if (r == -EAGAIN) { | |
1370 | // Do not reply, the message has been enqueued for retry | |
1371 | dout(4) << __func__ << " enqueue for retry by filesystem_command" << dendl; | |
1372 | return false; | |
1373 | } else if (r != -ENOSYS) { | |
1374 | goto out; | |
1375 | } | |
1376 | ||
1377 | // Only handle legacy commands if there is a filesystem configured | |
28e407b8 AA |
1378 | if (pending.legacy_client_fscid == FS_CLUSTER_ID_NONE) { |
1379 | if (pending.filesystems.size() == 0) { | |
7c673cae FG |
1380 | ss << "No filesystem configured: use `ceph fs new` to create a filesystem"; |
1381 | } else { | |
1382 | ss << "No filesystem set for use with legacy commands"; | |
1383 | } | |
1384 | r = -EINVAL; | |
1385 | goto out; | |
1386 | } | |
1387 | ||
1388 | r = legacy_filesystem_command(op, prefix, cmdmap, ss); | |
1389 | ||
1390 | if (r == -ENOSYS && ss.str().empty()) { | |
1391 | ss << "unrecognized command"; | |
1392 | } | |
1393 | ||
1394 | out: | |
1395 | dout(4) << __func__ << " done, r=" << r << dendl; | |
1396 | /* Compose response */ | |
1397 | string rs; | |
1398 | getline(ss, rs); | |
1399 | ||
1400 | if (r >= 0) { | |
1401 | // success.. delay reply | |
1402 | wait_for_finished_proposal(op, new Monitor::C_Command(mon, op, r, rs, | |
1403 | get_last_committed() + 1)); | |
c07f9fc5 FG |
1404 | if (batched_propose) { |
1405 | force_immediate_propose(); | |
1406 | } | |
7c673cae FG |
1407 | return true; |
1408 | } else { | |
1409 | // reply immediately | |
1410 | mon->reply_command(op, r, rs, rdata, get_last_committed()); | |
1411 | return false; | |
1412 | } | |
1413 | } | |
1414 | ||
1415 | ||
1416 | /** | |
1417 | * Given one of the following forms: | |
1418 | * <fs name>:<rank> | |
1419 | * <fs id>:<rank> | |
1420 | * <rank> | |
1421 | * | |
1422 | * Parse into a mds_role_t. The rank-only form is only valid | |
1423 | * if legacy_client_ns is set. | |
1424 | */ | |
1425 | int MDSMonitor::parse_role( | |
1426 | const std::string &role_str, | |
1427 | mds_role_t *role, | |
1428 | std::ostream &ss) | |
1429 | { | |
28e407b8 | 1430 | return get_working_fsmap().parse_role(role_str, role, ss); |
7c673cae FG |
1431 | } |
1432 | ||
1433 | int MDSMonitor::filesystem_command( | |
1434 | MonOpRequestRef op, | |
1435 | std::string const &prefix, | |
1436 | map<string, cmd_vartype> &cmdmap, | |
1437 | std::stringstream &ss) | |
1438 | { | |
1439 | dout(4) << __func__ << " prefix='" << prefix << "'" << dendl; | |
1440 | op->mark_mdsmon_event(__func__); | |
1441 | int r = 0; | |
1442 | string whostr; | |
1443 | cmd_getval(g_ceph_context, cmdmap, "who", whostr); | |
1444 | ||
28e407b8 | 1445 | auto &pending = get_pending_fsmap_writeable(); |
7c673cae FG |
1446 | if (prefix == "mds stop" || |
1447 | prefix == "mds deactivate") { | |
7c673cae FG |
1448 | mds_role_t role; |
1449 | r = parse_role(whostr, &role, ss); | |
1450 | if (r < 0 ) { | |
1451 | return r; | |
1452 | } | |
28e407b8 | 1453 | const auto &fs = pending.get_filesystem(role.fscid); |
7c673cae FG |
1454 | |
1455 | if (!fs->mds_map.is_active(role.rank)) { | |
1456 | r = -EEXIST; | |
1457 | ss << "mds." << role << " not active (" | |
1458 | << ceph_mds_state_name(fs->mds_map.get_state(role.rank)) << ")"; | |
1459 | } else if (fs->mds_map.get_root() == role.rank || | |
1460 | fs->mds_map.get_tableserver() == role.rank) { | |
1461 | r = -EINVAL; | |
1462 | ss << "can't tell the root (" << fs->mds_map.get_root() | |
1463 | << ") or tableserver (" << fs->mds_map.get_tableserver() | |
1464 | << ") to deactivate"; | |
31f18b77 FG |
1465 | } else if (role.rank != fs->mds_map.get_last_in_mds()) { |
1466 | r = -EINVAL; | |
1467 | ss << "mds." << role << " doesn't have the max rank (" | |
1468 | << fs->mds_map.get_last_in_mds() << ")"; | |
7c673cae FG |
1469 | } else if (fs->mds_map.get_num_in_mds() <= size_t(fs->mds_map.get_max_mds())) { |
1470 | r = -EBUSY; | |
1471 | ss << "must decrease max_mds or else MDS will immediately reactivate"; | |
1472 | } else { | |
1473 | r = 0; | |
1474 | mds_gid_t gid = fs->mds_map.up.at(role.rank); | |
1475 | ss << "telling mds." << role << " " | |
28e407b8 | 1476 | << pending.get_info_gid(gid).addr << " to deactivate"; |
7c673cae | 1477 | |
28e407b8 | 1478 | pending.modify_daemon(gid, [](MDSMap::mds_info_t *info) { |
7c673cae FG |
1479 | info->state = MDSMap::STATE_STOPPING; |
1480 | }); | |
1481 | } | |
1482 | } else if (prefix == "mds set_state") { | |
1483 | mds_gid_t gid; | |
1484 | if (!cmd_getval(g_ceph_context, cmdmap, "gid", gid)) { | |
1485 | ss << "error parsing 'gid' value '" | |
1486 | << cmd_vartype_stringify(cmdmap["gid"]) << "'"; | |
1487 | return -EINVAL; | |
1488 | } | |
1489 | MDSMap::DaemonState state; | |
1490 | if (!cmd_getval(g_ceph_context, cmdmap, "state", state)) { | |
1491 | ss << "error parsing 'state' string value '" | |
1492 | << cmd_vartype_stringify(cmdmap["state"]) << "'"; | |
1493 | return -EINVAL; | |
1494 | } | |
28e407b8 AA |
1495 | if (pending.gid_exists(gid)) { |
1496 | pending.modify_daemon(gid, [state](MDSMap::mds_info_t *info) { | |
7c673cae FG |
1497 | info->state = state; |
1498 | }); | |
1499 | ss << "set mds gid " << gid << " to state " << state << " " | |
1500 | << ceph_mds_state_name(state); | |
1501 | return 0; | |
1502 | } | |
1503 | } else if (prefix == "mds fail") { | |
1504 | string who; | |
1505 | cmd_getval(g_ceph_context, cmdmap, "who", who); | |
d2e6a577 FG |
1506 | |
1507 | MDSMap::mds_info_t failed_info; | |
1508 | r = fail_mds(ss, who, &failed_info); | |
7c673cae FG |
1509 | if (r < 0 && r == -EAGAIN) { |
1510 | mon->osdmon()->wait_for_writeable(op, new C_RetryMessage(this, op)); | |
1511 | return -EAGAIN; // don't propose yet; wait for message to be retried | |
d2e6a577 FG |
1512 | } else if (r == 0) { |
1513 | // Only log if we really did something (not when was already gone) | |
1514 | if (failed_info.global_id != MDS_GID_NONE) { | |
1515 | mon->clog->info() << failed_info.human_name() << " marked failed by " | |
1516 | << op->get_session()->entity_name; | |
1517 | } | |
7c673cae FG |
1518 | } |
1519 | } else if (prefix == "mds rm") { | |
1520 | mds_gid_t gid; | |
1521 | if (!cmd_getval(g_ceph_context, cmdmap, "gid", gid)) { | |
1522 | ss << "error parsing 'gid' value '" | |
1523 | << cmd_vartype_stringify(cmdmap["gid"]) << "'"; | |
1524 | return -EINVAL; | |
1525 | } | |
28e407b8 | 1526 | if (!pending.gid_exists(gid)) { |
7c673cae FG |
1527 | ss << "mds gid " << gid << " dne"; |
1528 | r = 0; | |
1529 | } else { | |
28e407b8 AA |
1530 | const auto &info = pending.get_info_gid(gid); |
1531 | MDSMap::DaemonState state = info.state; | |
7c673cae | 1532 | if (state > 0) { |
28e407b8 AA |
1533 | ss << "cannot remove active mds." << info.name |
1534 | << " rank " << info.rank; | |
7c673cae FG |
1535 | return -EBUSY; |
1536 | } else { | |
28e407b8 | 1537 | pending.erase(gid, {}); |
7c673cae FG |
1538 | ss << "removed mds gid " << gid; |
1539 | return 0; | |
1540 | } | |
1541 | } | |
1542 | } else if (prefix == "mds rmfailed") { | |
1543 | string confirm; | |
1544 | if (!cmd_getval(g_ceph_context, cmdmap, "confirm", confirm) || | |
1545 | confirm != "--yes-i-really-mean-it") { | |
1546 | ss << "WARNING: this can make your filesystem inaccessible! " | |
1547 | "Add --yes-i-really-mean-it if you are sure you wish to continue."; | |
1548 | return -EPERM; | |
1549 | } | |
1550 | ||
1551 | std::string role_str; | |
1552 | cmd_getval(g_ceph_context, cmdmap, "who", role_str); | |
1553 | mds_role_t role; | |
1554 | int r = parse_role(role_str, &role, ss); | |
1555 | if (r < 0) { | |
1556 | ss << "invalid role '" << role_str << "'"; | |
1557 | return -EINVAL; | |
1558 | } | |
1559 | ||
28e407b8 | 1560 | pending.modify_filesystem( |
7c673cae FG |
1561 | role.fscid, |
1562 | [role](std::shared_ptr<Filesystem> fs) | |
1563 | { | |
1564 | fs->mds_map.failed.erase(role.rank); | |
1565 | }); | |
1566 | ||
1567 | ss << "removed failed mds." << role; | |
1568 | return 0; | |
1569 | } else if (prefix == "mds compat rm_compat") { | |
1570 | int64_t f; | |
1571 | if (!cmd_getval(g_ceph_context, cmdmap, "feature", f)) { | |
1572 | ss << "error parsing feature value '" | |
1573 | << cmd_vartype_stringify(cmdmap["feature"]) << "'"; | |
1574 | return -EINVAL; | |
1575 | } | |
28e407b8 | 1576 | if (pending.compat.compat.contains(f)) { |
7c673cae | 1577 | ss << "removing compat feature " << f; |
28e407b8 | 1578 | CompatSet modified = pending.compat; |
7c673cae | 1579 | modified.compat.remove(f); |
28e407b8 | 1580 | pending.update_compat(modified); |
7c673cae | 1581 | } else { |
28e407b8 | 1582 | ss << "compat feature " << f << " not present in " << pending.compat; |
7c673cae FG |
1583 | } |
1584 | r = 0; | |
1585 | } else if (prefix == "mds compat rm_incompat") { | |
1586 | int64_t f; | |
1587 | if (!cmd_getval(g_ceph_context, cmdmap, "feature", f)) { | |
1588 | ss << "error parsing feature value '" | |
1589 | << cmd_vartype_stringify(cmdmap["feature"]) << "'"; | |
1590 | return -EINVAL; | |
1591 | } | |
28e407b8 | 1592 | if (pending.compat.incompat.contains(f)) { |
7c673cae | 1593 | ss << "removing incompat feature " << f; |
28e407b8 | 1594 | CompatSet modified = pending.compat; |
7c673cae | 1595 | modified.incompat.remove(f); |
28e407b8 | 1596 | pending.update_compat(modified); |
7c673cae | 1597 | } else { |
28e407b8 | 1598 | ss << "incompat feature " << f << " not present in " << pending.compat; |
7c673cae FG |
1599 | } |
1600 | r = 0; | |
1601 | } else if (prefix == "mds repaired") { | |
1602 | std::string role_str; | |
1603 | cmd_getval(g_ceph_context, cmdmap, "rank", role_str); | |
1604 | mds_role_t role; | |
1605 | r = parse_role(role_str, &role, ss); | |
1606 | if (r < 0) { | |
1607 | return r; | |
1608 | } | |
1609 | ||
28e407b8 | 1610 | bool modified = pending.undamaged(role.fscid, role.rank); |
7c673cae FG |
1611 | if (modified) { |
1612 | dout(4) << "repaired: restoring rank " << role << dendl; | |
1613 | } else { | |
1614 | dout(4) << "repaired: no-op on rank " << role << dendl; | |
1615 | } | |
1616 | ||
1617 | r = 0; | |
1618 | } else { | |
1619 | return -ENOSYS; | |
1620 | } | |
1621 | ||
1622 | return r; | |
1623 | } | |
1624 | ||
1625 | /** | |
1626 | * Helper to legacy_filesystem_command | |
1627 | */ | |
1628 | void MDSMonitor::modify_legacy_filesystem( | |
1629 | std::function<void(std::shared_ptr<Filesystem> )> fn) | |
1630 | { | |
28e407b8 | 1631 | auto &pending_fsmap = get_pending_fsmap_writeable(); |
7c673cae FG |
1632 | pending_fsmap.modify_filesystem( |
1633 | pending_fsmap.legacy_client_fscid, | |
1634 | fn | |
1635 | ); | |
1636 | } | |
1637 | ||
1638 | ||
1639 | ||
1640 | /** | |
1641 | * Handle a command that affects the filesystem (i.e. a filesystem | |
1642 | * must exist for the command to act upon). | |
1643 | * | |
1644 | * @retval 0 Command was successfully handled and has side effects | |
1645 | * @retval -EAGAIN Messages has been requeued for retry | |
1646 | * @retval -ENOSYS Unknown command | |
1647 | * @retval < 0 An error has occurred; **ss** may have been set. | |
1648 | */ | |
1649 | int MDSMonitor::legacy_filesystem_command( | |
1650 | MonOpRequestRef op, | |
1651 | std::string const &prefix, | |
1652 | map<string, cmd_vartype> &cmdmap, | |
1653 | std::stringstream &ss) | |
1654 | { | |
1655 | dout(4) << __func__ << " prefix='" << prefix << "'" << dendl; | |
1656 | op->mark_mdsmon_event(__func__); | |
1657 | int r = 0; | |
1658 | string whostr; | |
1659 | cmd_getval(g_ceph_context, cmdmap, "who", whostr); | |
1660 | ||
28e407b8 AA |
1661 | auto &pending_fsmap = get_pending_fsmap_writeable(); |
1662 | ||
7c673cae FG |
1663 | assert (pending_fsmap.legacy_client_fscid != FS_CLUSTER_ID_NONE); |
1664 | ||
1665 | if (prefix == "mds set_max_mds") { | |
1666 | // NOTE: deprecated by "fs set max_mds" | |
1667 | int64_t maxmds; | |
1668 | if (!cmd_getval(g_ceph_context, cmdmap, "maxmds", maxmds) || maxmds <= 0) { | |
1669 | return -EINVAL; | |
1670 | } | |
1671 | ||
1672 | const MDSMap& mdsmap = | |
1673 | pending_fsmap.filesystems.at(pending_fsmap.legacy_client_fscid)->mds_map; | |
1674 | ||
1675 | if (!mdsmap.allows_multimds() && | |
1676 | maxmds > mdsmap.get_max_mds() && | |
1677 | maxmds > 1) { | |
1678 | ss << "multi-MDS clusters are not enabled; set 'allow_multimds' to enable"; | |
1679 | return -EINVAL; | |
1680 | } | |
1681 | ||
1682 | if (maxmds > MAX_MDS) { | |
1683 | ss << "may not have more than " << MAX_MDS << " MDS ranks"; | |
1684 | return -EINVAL; | |
1685 | } | |
1686 | ||
1687 | modify_legacy_filesystem( | |
1688 | [maxmds](std::shared_ptr<Filesystem> fs) | |
1689 | { | |
1690 | fs->mds_map.set_max_mds(maxmds); | |
1691 | }); | |
1692 | ||
1693 | r = 0; | |
1694 | ss << "max_mds = " << maxmds; | |
1695 | } else if (prefix == "mds cluster_down") { | |
1696 | // NOTE: deprecated by "fs set cluster_down" | |
1697 | modify_legacy_filesystem( | |
1698 | [](std::shared_ptr<Filesystem> fs) | |
1699 | { | |
1700 | fs->mds_map.set_flag(CEPH_MDSMAP_DOWN); | |
1701 | }); | |
1702 | ss << "marked fsmap DOWN"; | |
1703 | r = 0; | |
1704 | } else if (prefix == "mds cluster_up") { | |
1705 | // NOTE: deprecated by "fs set cluster_up" | |
1706 | modify_legacy_filesystem( | |
1707 | [](std::shared_ptr<Filesystem> fs) | |
1708 | { | |
1709 | fs->mds_map.clear_flag(CEPH_MDSMAP_DOWN); | |
1710 | }); | |
1711 | ss << "unmarked fsmap DOWN"; | |
1712 | r = 0; | |
1713 | } else { | |
1714 | return -ENOSYS; | |
1715 | } | |
1716 | ||
1717 | return r; | |
1718 | } | |
1719 | ||
1720 | ||
1721 | void MDSMonitor::check_subs() | |
1722 | { | |
1723 | std::list<std::string> types; | |
1724 | ||
1725 | // Subscriptions may be to "mdsmap" (MDS and legacy clients), | |
1726 | // "mdsmap.<namespace>", or to "fsmap" for the full state of all | |
1727 | // filesystems. Build a list of all the types we service | |
1728 | // subscriptions for. | |
1729 | types.push_back("fsmap"); | |
1730 | types.push_back("fsmap.user"); | |
1731 | types.push_back("mdsmap"); | |
28e407b8 AA |
1732 | for (const auto &p : get_fsmap().filesystems) { |
1733 | const auto &fscid = p.first; | |
7c673cae FG |
1734 | std::ostringstream oss; |
1735 | oss << "mdsmap." << fscid; | |
1736 | types.push_back(oss.str()); | |
1737 | } | |
1738 | ||
1739 | for (const auto &type : types) { | |
1740 | if (mon->session_map.subs.count(type) == 0) | |
1741 | continue; | |
1742 | xlist<Subscription*>::iterator p = mon->session_map.subs[type]->begin(); | |
1743 | while (!p.end()) { | |
1744 | Subscription *sub = *p; | |
1745 | ++p; | |
1746 | check_sub(sub); | |
1747 | } | |
1748 | } | |
1749 | } | |
1750 | ||
1751 | ||
1752 | void MDSMonitor::check_sub(Subscription *sub) | |
1753 | { | |
1754 | dout(20) << __func__ << ": " << sub->type << dendl; | |
1755 | ||
28e407b8 AA |
1756 | const auto &fsmap = get_fsmap(); |
1757 | ||
7c673cae FG |
1758 | if (sub->type == "fsmap") { |
1759 | if (sub->next <= fsmap.get_epoch()) { | |
1760 | sub->session->con->send_message(new MFSMap(mon->monmap->fsid, fsmap)); | |
1761 | if (sub->onetime) { | |
1762 | mon->session_map.remove_sub(sub); | |
1763 | } else { | |
1764 | sub->next = fsmap.get_epoch() + 1; | |
1765 | } | |
1766 | } | |
1767 | } else if (sub->type == "fsmap.user") { | |
1768 | if (sub->next <= fsmap.get_epoch()) { | |
1769 | FSMapUser fsmap_u; | |
1770 | fsmap_u.epoch = fsmap.get_epoch(); | |
1771 | fsmap_u.legacy_client_fscid = fsmap.legacy_client_fscid; | |
28e407b8 AA |
1772 | for (const auto &p : fsmap.filesystems) { |
1773 | FSMapUser::fs_info_t& fs_info = fsmap_u.filesystems[p.second->fscid]; | |
1774 | fs_info.cid = p.second->fscid; | |
1775 | fs_info.name = p.second->mds_map.fs_name; | |
7c673cae FG |
1776 | } |
1777 | sub->session->con->send_message(new MFSMapUser(mon->monmap->fsid, fsmap_u)); | |
1778 | if (sub->onetime) { | |
1779 | mon->session_map.remove_sub(sub); | |
1780 | } else { | |
1781 | sub->next = fsmap.get_epoch() + 1; | |
1782 | } | |
1783 | } | |
1784 | } else if (sub->type.compare(0, 6, "mdsmap") == 0) { | |
1785 | if (sub->next > fsmap.get_epoch()) { | |
1786 | return; | |
1787 | } | |
1788 | ||
1789 | const bool is_mds = sub->session->inst.name.is_mds(); | |
1790 | mds_gid_t mds_gid = MDS_GID_NONE; | |
1791 | fs_cluster_id_t fscid = FS_CLUSTER_ID_NONE; | |
1792 | if (is_mds) { | |
1793 | // What (if any) namespace are you assigned to? | |
1794 | auto mds_info = fsmap.get_mds_info(); | |
1795 | for (const auto &i : mds_info) { | |
1796 | if (i.second.addr == sub->session->inst.addr) { | |
1797 | mds_gid = i.first; | |
1798 | fscid = fsmap.mds_roles.at(mds_gid); | |
1799 | } | |
1800 | } | |
1801 | } else { | |
1802 | // You're a client. Did you request a particular | |
1803 | // namespace? | |
1804 | if (sub->type.find("mdsmap.") == 0) { | |
1805 | auto namespace_id_str = sub->type.substr(std::string("mdsmap.").size()); | |
1806 | dout(10) << __func__ << ": namespace_id " << namespace_id_str << dendl; | |
1807 | std::string err; | |
1808 | fscid = strict_strtoll(namespace_id_str.c_str(), 10, &err); | |
1809 | if (!err.empty()) { | |
1810 | // Client asked for a non-existent namespace, send them nothing | |
1811 | dout(1) << "Invalid client subscription '" << sub->type | |
1812 | << "'" << dendl; | |
1813 | return; | |
1814 | } | |
1815 | if (fsmap.filesystems.count(fscid) == 0) { | |
1816 | // Client asked for a non-existent namespace, send them nothing | |
1817 | // TODO: something more graceful for when a client has a filesystem | |
1818 | // mounted, and the fileysstem is deleted. Add a "shut down you fool" | |
1819 | // flag to MMDSMap? | |
1820 | dout(1) << "Client subscribed to non-existent namespace '" << | |
1821 | fscid << "'" << dendl; | |
1822 | return; | |
1823 | } | |
1824 | } else { | |
1825 | // Unqualified request for "mdsmap": give it the one marked | |
1826 | // for use by legacy clients. | |
1827 | if (fsmap.legacy_client_fscid != FS_CLUSTER_ID_NONE) { | |
1828 | fscid = fsmap.legacy_client_fscid; | |
1829 | } else { | |
1830 | dout(1) << "Client subscribed for legacy filesystem but " | |
1831 | "none is configured" << dendl; | |
1832 | return; | |
1833 | } | |
1834 | } | |
1835 | } | |
1836 | dout(10) << __func__ << ": is_mds=" << is_mds << ", fscid= " << fscid << dendl; | |
1837 | ||
1838 | // Work out the effective latest epoch | |
28e407b8 | 1839 | const MDSMap *mds_map = nullptr; |
7c673cae FG |
1840 | MDSMap null_map; |
1841 | null_map.compat = fsmap.compat; | |
1842 | if (fscid == FS_CLUSTER_ID_NONE) { | |
1843 | // For a client, we should have already dropped out | |
1844 | assert(is_mds); | |
1845 | ||
28e407b8 AA |
1846 | auto it = fsmap.standby_daemons.find(mds_gid); |
1847 | if (it != fsmap.standby_daemons.end()) { | |
7c673cae | 1848 | // For an MDS, we need to feed it an MDSMap with its own state in |
28e407b8 AA |
1849 | null_map.mds_info[mds_gid] = it->second; |
1850 | null_map.epoch = fsmap.standby_epochs.at(mds_gid); | |
7c673cae FG |
1851 | } else { |
1852 | null_map.epoch = fsmap.epoch; | |
1853 | } | |
1854 | mds_map = &null_map; | |
1855 | } else { | |
1856 | // Check the effective epoch | |
28e407b8 | 1857 | mds_map = &fsmap.get_filesystem(fscid)->mds_map; |
7c673cae FG |
1858 | } |
1859 | ||
1860 | assert(mds_map != nullptr); | |
1861 | dout(10) << __func__ << " selected MDS map epoch " << | |
1862 | mds_map->epoch << " for namespace " << fscid << " for subscriber " | |
1863 | << sub->session->inst.name << " who wants epoch " << sub->next << dendl; | |
1864 | ||
1865 | if (sub->next > mds_map->epoch) { | |
1866 | return; | |
1867 | } | |
1868 | auto msg = new MMDSMap(mon->monmap->fsid, mds_map); | |
1869 | ||
1870 | sub->session->con->send_message(msg); | |
1871 | if (sub->onetime) { | |
1872 | mon->session_map.remove_sub(sub); | |
1873 | } else { | |
1874 | sub->next = mds_map->get_epoch() + 1; | |
1875 | } | |
1876 | } | |
1877 | } | |
1878 | ||
1879 | ||
1880 | void MDSMonitor::update_metadata(mds_gid_t gid, | |
1881 | const map<string, string>& metadata) | |
1882 | { | |
1883 | if (metadata.empty()) { | |
1884 | return; | |
1885 | } | |
1886 | pending_metadata[gid] = metadata; | |
1887 | ||
1888 | MonitorDBStore::TransactionRef t = paxos->get_pending_transaction(); | |
1889 | bufferlist bl; | |
1890 | ::encode(pending_metadata, bl); | |
1891 | t->put(MDS_METADATA_PREFIX, "last_metadata", bl); | |
1892 | paxos->trigger_propose(); | |
1893 | } | |
1894 | ||
1895 | void MDSMonitor::remove_from_metadata(MonitorDBStore::TransactionRef t) | |
1896 | { | |
1897 | bool update = false; | |
1898 | for (map<mds_gid_t, Metadata>::iterator i = pending_metadata.begin(); | |
1899 | i != pending_metadata.end(); ) { | |
28e407b8 | 1900 | if (!get_pending_fsmap().gid_exists(i->first)) { |
7c673cae FG |
1901 | pending_metadata.erase(i++); |
1902 | update = true; | |
1903 | } else { | |
1904 | ++i; | |
1905 | } | |
1906 | } | |
1907 | if (!update) | |
1908 | return; | |
1909 | bufferlist bl; | |
1910 | ::encode(pending_metadata, bl); | |
1911 | t->put(MDS_METADATA_PREFIX, "last_metadata", bl); | |
1912 | } | |
1913 | ||
1914 | int MDSMonitor::load_metadata(map<mds_gid_t, Metadata>& m) | |
1915 | { | |
1916 | bufferlist bl; | |
1917 | int r = mon->store->get(MDS_METADATA_PREFIX, "last_metadata", bl); | |
1918 | if (r) { | |
1919 | dout(1) << "Unable to load 'last_metadata'" << dendl; | |
1920 | return r; | |
1921 | } | |
1922 | ||
1923 | bufferlist::iterator it = bl.begin(); | |
1924 | ::decode(m, it); | |
1925 | return 0; | |
1926 | } | |
1927 | ||
c07f9fc5 | 1928 | void MDSMonitor::count_metadata(const string& field, map<string,int> *out) |
31f18b77 | 1929 | { |
31f18b77 FG |
1930 | map<mds_gid_t,Metadata> meta; |
1931 | load_metadata(meta); | |
1932 | for (auto& p : meta) { | |
1933 | auto q = p.second.find(field); | |
1934 | if (q == p.second.end()) { | |
c07f9fc5 | 1935 | (*out)["unknown"]++; |
31f18b77 | 1936 | } else { |
c07f9fc5 | 1937 | (*out)[q->second]++; |
31f18b77 FG |
1938 | } |
1939 | } | |
c07f9fc5 FG |
1940 | } |
1941 | ||
1942 | void MDSMonitor::count_metadata(const string& field, Formatter *f) | |
1943 | { | |
1944 | map<string,int> by_val; | |
1945 | count_metadata(field, &by_val); | |
31f18b77 FG |
1946 | f->open_object_section(field.c_str()); |
1947 | for (auto& p : by_val) { | |
1948 | f->dump_int(p.first.c_str(), p.second); | |
1949 | } | |
1950 | f->close_section(); | |
1951 | } | |
1952 | ||
7c673cae FG |
1953 | int MDSMonitor::dump_metadata(const std::string &who, Formatter *f, ostream& err) |
1954 | { | |
1955 | assert(f); | |
1956 | ||
1957 | mds_gid_t gid = gid_from_arg(who, err); | |
1958 | if (gid == MDS_GID_NONE) { | |
1959 | return -EINVAL; | |
1960 | } | |
1961 | ||
1962 | map<mds_gid_t, Metadata> metadata; | |
1963 | if (int r = load_metadata(metadata)) { | |
1964 | err << "Unable to load 'last_metadata'"; | |
1965 | return r; | |
1966 | } | |
1967 | ||
1968 | if (!metadata.count(gid)) { | |
1969 | return -ENOENT; | |
1970 | } | |
1971 | const Metadata& m = metadata[gid]; | |
1972 | for (Metadata::const_iterator p = m.begin(); p != m.end(); ++p) { | |
1973 | f->dump_string(p->first.c_str(), p->second); | |
1974 | } | |
1975 | return 0; | |
1976 | } | |
1977 | ||
1978 | int MDSMonitor::print_nodes(Formatter *f) | |
1979 | { | |
1980 | assert(f); | |
1981 | ||
1982 | map<mds_gid_t, Metadata> metadata; | |
1983 | if (int r = load_metadata(metadata)) { | |
1984 | return r; | |
1985 | } | |
1986 | ||
1987 | map<string, list<int> > mdses; // hostname => rank | |
1988 | for (map<mds_gid_t, Metadata>::iterator it = metadata.begin(); | |
1989 | it != metadata.end(); ++it) { | |
1990 | const Metadata& m = it->second; | |
1991 | Metadata::const_iterator hostname = m.find("hostname"); | |
1992 | if (hostname == m.end()) { | |
1993 | // not likely though | |
1994 | continue; | |
1995 | } | |
1996 | const mds_gid_t gid = it->first; | |
28e407b8 | 1997 | if (!get_fsmap().gid_exists(gid)) { |
7c673cae FG |
1998 | dout(5) << __func__ << ": GID " << gid << " not existent" << dendl; |
1999 | continue; | |
2000 | } | |
28e407b8 | 2001 | const MDSMap::mds_info_t& mds_info = get_fsmap().get_info_gid(gid); |
7c673cae FG |
2002 | // FIXME: include filesystem name with rank here |
2003 | mdses[hostname->second].push_back(mds_info.rank); | |
2004 | } | |
2005 | ||
2006 | dump_services(f, mdses, "mds"); | |
2007 | return 0; | |
2008 | } | |
2009 | ||
2010 | /** | |
2011 | * If a cluster is undersized (with respect to max_mds), then | |
2012 | * attempt to find daemons to grow it. | |
2013 | */ | |
28e407b8 | 2014 | bool MDSMonitor::maybe_expand_cluster(std::shared_ptr<Filesystem> &fs) |
7c673cae FG |
2015 | { |
2016 | bool do_propose = false; | |
28e407b8 | 2017 | auto &pending = get_pending_fsmap_writeable(); |
7c673cae FG |
2018 | |
2019 | if (fs->mds_map.test_flag(CEPH_MDSMAP_DOWN)) { | |
2020 | return do_propose; | |
2021 | } | |
2022 | ||
2023 | while (fs->mds_map.get_num_in_mds() < size_t(fs->mds_map.get_max_mds()) && | |
2024 | !fs->mds_map.is_degraded()) { | |
2025 | mds_rank_t mds = mds_rank_t(0); | |
2026 | string name; | |
2027 | while (fs->mds_map.is_in(mds)) { | |
2028 | mds++; | |
2029 | } | |
28e407b8 | 2030 | mds_gid_t newgid = pending.find_replacement_for({fs->fscid, mds}, |
7c673cae FG |
2031 | name, g_conf->mon_force_standby_active); |
2032 | if (newgid == MDS_GID_NONE) { | |
2033 | break; | |
2034 | } | |
2035 | ||
28e407b8 | 2036 | const auto &new_info = pending.get_info_gid(newgid); |
d2e6a577 | 2037 | dout(1) << "assigned standby " << new_info.addr |
7c673cae | 2038 | << " as mds." << mds << dendl; |
d2e6a577 FG |
2039 | |
2040 | mon->clog->info() << new_info.human_name() << " assigned to " | |
2041 | "filesystem " << fs->mds_map.fs_name << " as rank " | |
181888fb | 2042 | << mds << " (now has " << fs->mds_map.get_num_in_mds() + 1 |
d2e6a577 | 2043 | << " ranks)"; |
28e407b8 | 2044 | pending.promote(newgid, fs, mds); |
7c673cae FG |
2045 | do_propose = true; |
2046 | } | |
2047 | ||
2048 | return do_propose; | |
2049 | } | |
2050 | ||
2051 | ||
2052 | /** | |
2053 | * If a daemon is laggy, and a suitable replacement | |
2054 | * is available, fail this daemon (remove from map) and pass its | |
2055 | * role to another daemon. | |
2056 | */ | |
c07f9fc5 | 2057 | void MDSMonitor::maybe_replace_gid(mds_gid_t gid, const MDSMap::mds_info_t& info, |
7c673cae FG |
2058 | bool *mds_propose, bool *osd_propose) |
2059 | { | |
2060 | assert(mds_propose != nullptr); | |
2061 | assert(osd_propose != nullptr); | |
2062 | ||
28e407b8 AA |
2063 | auto &pending = get_pending_fsmap_writeable(); |
2064 | const auto fscid = pending.mds_roles.at(gid); | |
7c673cae | 2065 | |
31f18b77 FG |
2066 | // We will only take decisive action (replacing/removing a daemon) |
2067 | // if we have some indicating that some other daemon(s) are successfully | |
2068 | // getting beacons through recently. | |
2069 | utime_t latest_beacon; | |
2070 | for (const auto & i : last_beacon) { | |
2071 | latest_beacon = MAX(i.second.stamp, latest_beacon); | |
2072 | } | |
2073 | const bool may_replace = latest_beacon > | |
2074 | (ceph_clock_now() - | |
2075 | MAX(g_conf->mds_beacon_interval, g_conf->mds_beacon_grace * 0.5)); | |
2076 | ||
7c673cae FG |
2077 | // are we in? |
2078 | // and is there a non-laggy standby that can take over for us? | |
2079 | mds_gid_t sgid; | |
2080 | if (info.rank >= 0 && | |
2081 | info.state != MDSMap::STATE_STANDBY && | |
2082 | info.state != MDSMap::STATE_STANDBY_REPLAY && | |
31f18b77 | 2083 | may_replace && |
28e407b8 AA |
2084 | !pending.get_filesystem(fscid)->mds_map.test_flag(CEPH_MDSMAP_DOWN) && |
2085 | (sgid = pending.find_replacement_for({fscid, info.rank}, info.name, | |
7c673cae FG |
2086 | g_conf->mon_force_standby_active)) != MDS_GID_NONE) |
2087 | { | |
2088 | ||
28e407b8 | 2089 | MDSMap::mds_info_t si = pending.get_info_gid(sgid); |
7c673cae FG |
2090 | dout(10) << " replacing " << gid << " " << info.addr << " mds." |
2091 | << info.rank << "." << info.inc | |
2092 | << " " << ceph_mds_state_name(info.state) | |
2093 | << " with " << sgid << "/" << si.name << " " << si.addr << dendl; | |
2094 | ||
d2e6a577 | 2095 | mon->clog->warn() << info.human_name() |
31f18b77 FG |
2096 | << " is not responding, replacing it " |
2097 | << "as rank " << info.rank | |
d2e6a577 | 2098 | << " with standby " << si.human_name(); |
31f18b77 | 2099 | |
7c673cae | 2100 | // Remember what NS the old one was in |
28e407b8 | 2101 | const fs_cluster_id_t fscid = pending.mds_roles.at(gid); |
7c673cae FG |
2102 | |
2103 | // Remove the old one | |
2104 | *osd_propose |= fail_mds_gid(gid); | |
2105 | ||
2106 | // Promote the replacement | |
28e407b8 AA |
2107 | auto fs = pending.filesystems.at(fscid); |
2108 | pending.promote(sgid, fs, info.rank); | |
7c673cae FG |
2109 | |
2110 | *mds_propose = true; | |
31f18b77 FG |
2111 | } else if ((info.state == MDSMap::STATE_STANDBY_REPLAY || |
2112 | info.state == MDSMap::STATE_STANDBY) && may_replace) { | |
7c673cae FG |
2113 | dout(10) << " failing and removing " << gid << " " << info.addr << " mds." << info.rank |
2114 | << "." << info.inc << " " << ceph_mds_state_name(info.state) | |
2115 | << dendl; | |
d2e6a577 FG |
2116 | mon->clog->info() << "Standby " << info.human_name() << " is not " |
2117 | "responding, dropping it"; | |
7c673cae FG |
2118 | fail_mds_gid(gid); |
2119 | *mds_propose = true; | |
2120 | } else if (!info.laggy()) { | |
2121 | dout(10) << " marking " << gid << " " << info.addr << " mds." << info.rank << "." << info.inc | |
2122 | << " " << ceph_mds_state_name(info.state) | |
2123 | << " laggy" << dendl; | |
28e407b8 | 2124 | pending.modify_daemon(info.global_id, [](MDSMap::mds_info_t *info) { |
7c673cae FG |
2125 | info->laggy_since = ceph_clock_now(); |
2126 | }); | |
2127 | *mds_propose = true; | |
2128 | } | |
2129 | } | |
2130 | ||
28e407b8 | 2131 | bool MDSMonitor::maybe_promote_standby(std::shared_ptr<Filesystem> &fs) |
7c673cae FG |
2132 | { |
2133 | assert(!fs->mds_map.test_flag(CEPH_MDSMAP_DOWN)); | |
2134 | ||
28e407b8 AA |
2135 | auto &pending = get_pending_fsmap_writeable(); |
2136 | ||
7c673cae FG |
2137 | bool do_propose = false; |
2138 | ||
2139 | // have a standby take over? | |
2140 | set<mds_rank_t> failed; | |
2141 | fs->mds_map.get_failed_mds_set(failed); | |
2142 | if (!failed.empty()) { | |
2143 | set<mds_rank_t>::iterator p = failed.begin(); | |
2144 | while (p != failed.end()) { | |
2145 | mds_rank_t f = *p++; | |
28e407b8 | 2146 | mds_gid_t sgid = pending.find_replacement_for({fs->fscid, f}, {}, |
7c673cae FG |
2147 | g_conf->mon_force_standby_active); |
2148 | if (sgid) { | |
28e407b8 | 2149 | const MDSMap::mds_info_t si = pending.get_info_gid(sgid); |
7c673cae FG |
2150 | dout(0) << " taking over failed mds." << f << " with " << sgid |
2151 | << "/" << si.name << " " << si.addr << dendl; | |
d2e6a577 FG |
2152 | mon->clog->info() << "Standby " << si.human_name() |
2153 | << " assigned to filesystem " << fs->mds_map.fs_name | |
2154 | << " as rank " << f; | |
2155 | ||
28e407b8 | 2156 | pending.promote(sgid, fs, f); |
7c673cae FG |
2157 | do_propose = true; |
2158 | } | |
2159 | } | |
2160 | } else { | |
2161 | // There were no failures to replace, so try using any available standbys | |
2162 | // as standby-replay daemons. | |
2163 | ||
2164 | // Take a copy of the standby GIDs so that we can iterate over | |
2165 | // them while perhaps-modifying standby_daemons during the loop | |
2166 | // (if we promote anyone they are removed from standby_daemons) | |
2167 | std::vector<mds_gid_t> standby_gids; | |
28e407b8 | 2168 | for (const auto &j : pending.standby_daemons) { |
7c673cae FG |
2169 | standby_gids.push_back(j.first); |
2170 | } | |
2171 | ||
2172 | for (const auto &gid : standby_gids) { | |
28e407b8 | 2173 | const auto &info = pending.standby_daemons.at(gid); |
7c673cae FG |
2174 | assert(info.state == MDSMap::STATE_STANDBY); |
2175 | ||
2176 | if (!info.standby_replay) { | |
2177 | continue; | |
2178 | } | |
2179 | ||
2180 | /* | |
2181 | * This mds is standby but has no rank assigned. | |
2182 | * See if we can find it somebody to shadow | |
2183 | */ | |
2184 | dout(20) << "gid " << gid << " is standby and following nobody" << dendl; | |
2185 | ||
2186 | // standby for someone specific? | |
2187 | if (info.standby_for_rank >= 0) { | |
2188 | // The mds_info_t may or may not tell us exactly which filesystem | |
2189 | // the standby_for_rank refers to: lookup via legacy_client_fscid | |
2190 | mds_role_t target_role = { | |
2191 | info.standby_for_fscid == FS_CLUSTER_ID_NONE ? | |
28e407b8 | 2192 | pending.legacy_client_fscid : info.standby_for_fscid, |
7c673cae FG |
2193 | info.standby_for_rank}; |
2194 | ||
2195 | // It is possible that the map contains a standby_for_fscid | |
2196 | // that doesn't correspond to an existing filesystem, especially | |
2197 | // if we loaded from a version with a bug (#17466) | |
2198 | if (info.standby_for_fscid != FS_CLUSTER_ID_NONE | |
28e407b8 | 2199 | && !pending.filesystem_exists(info.standby_for_fscid)) { |
7c673cae FG |
2200 | derr << "gid " << gid << " has invalid standby_for_fscid " |
2201 | << info.standby_for_fscid << dendl; | |
2202 | continue; | |
2203 | } | |
2204 | ||
2205 | // If we managed to resolve a full target role | |
2206 | if (target_role.fscid != FS_CLUSTER_ID_NONE) { | |
28e407b8 | 2207 | const auto &fs = pending.get_filesystem(target_role.fscid); |
7c673cae FG |
2208 | if (fs->mds_map.is_followable(target_role.rank)) { |
2209 | do_propose |= try_standby_replay( | |
2210 | info, | |
2211 | *fs, | |
2212 | fs->mds_map.get_info(target_role.rank)); | |
2213 | } | |
2214 | } | |
2215 | ||
2216 | continue; | |
2217 | } | |
2218 | ||
2219 | // check everyone | |
28e407b8 AA |
2220 | for (const auto &p : pending.filesystems) { |
2221 | if (info.standby_for_fscid != FS_CLUSTER_ID_NONE && | |
2222 | info.standby_for_fscid != p.first) | |
2223 | continue; | |
2224 | ||
2225 | bool assigned = false; | |
2226 | const auto &fs = p.second; | |
2227 | const MDSMap &mds_map = fs->mds_map; | |
2228 | for (const auto &mds_i : mds_map.mds_info) { | |
2229 | const MDSMap::mds_info_t &cand_info = mds_i.second; | |
7c673cae FG |
2230 | if (cand_info.rank >= 0 && mds_map.is_followable(cand_info.rank)) { |
2231 | if ((info.standby_for_name.length() && info.standby_for_name != cand_info.name) || | |
2232 | info.standby_for_rank != MDS_RANK_NONE) { | |
2233 | continue; // we're supposed to follow someone else | |
2234 | } | |
2235 | ||
28e407b8 AA |
2236 | if (try_standby_replay(info, *fs, cand_info)) { |
2237 | assigned = true; | |
7c673cae FG |
2238 | break; |
2239 | } | |
7c673cae FG |
2240 | } |
2241 | } | |
28e407b8 AA |
2242 | if (assigned) { |
2243 | do_propose = true; | |
2244 | break; | |
2245 | } | |
7c673cae FG |
2246 | } |
2247 | } | |
2248 | } | |
2249 | ||
2250 | return do_propose; | |
2251 | } | |
2252 | ||
2253 | void MDSMonitor::tick() | |
2254 | { | |
2255 | // make sure mds's are still alive | |
2256 | // ...if i am an active leader | |
28e407b8 | 2257 | |
7c673cae FG |
2258 | if (!is_active()) return; |
2259 | ||
28e407b8 | 2260 | dout(10) << get_working_fsmap() << dendl; |
7c673cae | 2261 | |
28e407b8 AA |
2262 | if (!is_leader()) return; |
2263 | ||
2264 | auto &pending = get_pending_fsmap_writeable(); | |
7c673cae | 2265 | |
28e407b8 | 2266 | bool do_propose = false; |
7c673cae | 2267 | |
28e407b8 | 2268 | do_propose |= pending.check_health(); |
7c673cae FG |
2269 | |
2270 | // expand mds cluster (add new nodes to @in)? | |
28e407b8 AA |
2271 | for (auto &p : pending.filesystems) { |
2272 | do_propose |= maybe_expand_cluster(p.second); | |
7c673cae FG |
2273 | } |
2274 | ||
2275 | const auto now = ceph_clock_now(); | |
2276 | if (last_tick.is_zero()) { | |
2277 | last_tick = now; | |
2278 | } | |
2279 | ||
2280 | if (now - last_tick > (g_conf->mds_beacon_grace - g_conf->mds_beacon_interval)) { | |
2281 | // This case handles either local slowness (calls being delayed | |
2282 | // for whatever reason) or cluster election slowness (a long gap | |
2283 | // between calls while an election happened) | |
2284 | dout(4) << __func__ << ": resetting beacon timeouts due to mon delay " | |
2285 | "(slow election?) of " << now - last_tick << " seconds" << dendl; | |
2286 | for (auto &i : last_beacon) { | |
2287 | i.second.stamp = now; | |
2288 | } | |
2289 | } | |
2290 | ||
2291 | last_tick = now; | |
2292 | ||
2293 | // check beacon timestamps | |
2294 | utime_t cutoff = now; | |
2295 | cutoff -= g_conf->mds_beacon_grace; | |
2296 | ||
2297 | // make sure last_beacon is fully populated | |
28e407b8 | 2298 | for (auto &p : pending.mds_roles) { |
7c673cae FG |
2299 | auto &gid = p.first; |
2300 | if (last_beacon.count(gid) == 0) { | |
2301 | last_beacon[gid].stamp = now; | |
2302 | last_beacon[gid].seq = 0; | |
2303 | } | |
2304 | } | |
2305 | ||
c07f9fc5 FG |
2306 | bool propose_osdmap = false; |
2307 | bool osdmap_writeable = mon->osdmon()->is_writeable(); | |
2308 | auto p = last_beacon.begin(); | |
2309 | while (p != last_beacon.end()) { | |
2310 | mds_gid_t gid = p->first; | |
2311 | auto beacon_info = p->second; | |
2312 | ++p; | |
7c673cae | 2313 | |
28e407b8 | 2314 | if (!pending.gid_exists(gid)) { |
c07f9fc5 FG |
2315 | // clean it out |
2316 | last_beacon.erase(gid); | |
2317 | continue; | |
7c673cae FG |
2318 | } |
2319 | ||
c07f9fc5 | 2320 | if (beacon_info.stamp < cutoff) { |
28e407b8 | 2321 | auto &info = pending.get_info_gid(gid); |
c07f9fc5 FG |
2322 | dout(1) << "no beacon from mds." << info.rank << "." << info.inc |
2323 | << " (gid: " << gid << " addr: " << info.addr | |
2324 | << " state: " << ceph_mds_state_name(info.state) << ")" | |
2325 | << " since " << beacon_info.stamp << dendl; | |
2326 | // If the OSDMap is writeable, we can blacklist things, so we can | |
2327 | // try failing any laggy MDS daemons. Consider each one for failure. | |
2328 | if (osdmap_writeable) { | |
2329 | maybe_replace_gid(gid, info, &do_propose, &propose_osdmap); | |
2330 | } | |
7c673cae FG |
2331 | } |
2332 | } | |
c07f9fc5 FG |
2333 | if (propose_osdmap) { |
2334 | request_proposal(mon->osdmon()); | |
2335 | } | |
7c673cae | 2336 | |
28e407b8 AA |
2337 | for (auto &p : pending.filesystems) { |
2338 | auto &fs = p.second; | |
7c673cae FG |
2339 | if (!fs->mds_map.test_flag(CEPH_MDSMAP_DOWN)) { |
2340 | do_propose |= maybe_promote_standby(fs); | |
2341 | } | |
2342 | } | |
2343 | ||
2344 | if (do_propose) { | |
2345 | propose_pending(); | |
2346 | } | |
2347 | } | |
2348 | ||
2349 | /** | |
2350 | * finfo: the would-be follower | |
2351 | * leader_fs: the Filesystem containing the would-be leader | |
2352 | * ainfo: the would-be leader | |
2353 | */ | |
2354 | bool MDSMonitor::try_standby_replay( | |
2355 | const MDSMap::mds_info_t& finfo, | |
2356 | const Filesystem &leader_fs, | |
2357 | const MDSMap::mds_info_t& ainfo) | |
2358 | { | |
2359 | // someone else already following? | |
2360 | if (leader_fs.has_standby_replay(ainfo.global_id)) { | |
2361 | dout(20) << " mds." << ainfo.rank << " already has a follower" << dendl; | |
2362 | return false; | |
2363 | } else { | |
2364 | // Assign the new role to the standby | |
2365 | dout(10) << " setting to follow mds rank " << ainfo.rank << dendl; | |
28e407b8 | 2366 | get_pending_fsmap_writeable().assign_standby_replay(finfo.global_id, leader_fs.fscid, ainfo.rank); |
7c673cae FG |
2367 | return true; |
2368 | } | |
2369 | } | |
2370 | ||
2371 | MDSMonitor::MDSMonitor(Monitor *mn, Paxos *p, string service_name) | |
2372 | : PaxosService(mn, p, service_name) | |
2373 | { | |
c07f9fc5 | 2374 | handlers = FileSystemCommandHandler::load(p); |
7c673cae FG |
2375 | } |
2376 | ||
2377 | void MDSMonitor::on_restart() | |
2378 | { | |
2379 | // Clear out the leader-specific state. | |
2380 | last_tick = utime_t(); | |
2381 | last_beacon.clear(); | |
2382 | } | |
2383 |