]>
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) 2015 Red Hat | |
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 "common/debug.h" | |
16 | #include "common/errno.h" | |
17 | ||
18 | #include "messages/MClientRequestForward.h" | |
19 | #include "messages/MMDSLoadTargets.h" | |
20 | #include "messages/MMDSMap.h" | |
21 | #include "messages/MMDSTableRequest.h" | |
22 | #include "messages/MCommand.h" | |
23 | #include "messages/MCommandReply.h" | |
24 | ||
25 | #include "MDSDaemon.h" | |
26 | #include "MDSMap.h" | |
27 | #include "SnapClient.h" | |
28 | #include "SnapServer.h" | |
29 | #include "MDBalancer.h" | |
30 | #include "Locker.h" | |
31 | #include "Server.h" | |
32 | #include "InoTable.h" | |
33 | #include "mon/MonClient.h" | |
34 | #include "common/HeartbeatMap.h" | |
35 | #include "ScrubStack.h" | |
36 | ||
37 | ||
38 | #include "MDSRank.h" | |
39 | ||
40 | #define dout_context g_ceph_context | |
41 | #define dout_subsys ceph_subsys_mds | |
42 | #undef dout_prefix | |
43 | #define dout_prefix *_dout << "mds." << whoami << '.' << incarnation << ' ' | |
44 | ||
45 | MDSRank::MDSRank( | |
46 | mds_rank_t whoami_, | |
47 | Mutex &mds_lock_, | |
48 | LogChannelRef &clog_, | |
49 | SafeTimer &timer_, | |
50 | Beacon &beacon_, | |
51 | MDSMap *& mdsmap_, | |
52 | Messenger *msgr, | |
53 | MonClient *monc_, | |
54 | Context *respawn_hook_, | |
55 | Context *suicide_hook_) | |
56 | : | |
57 | whoami(whoami_), incarnation(0), | |
58 | mds_lock(mds_lock_), clog(clog_), timer(timer_), | |
59 | mdsmap(mdsmap_), | |
60 | objecter(new Objecter(g_ceph_context, msgr, monc_, nullptr, 0, 0)), | |
61 | server(NULL), mdcache(NULL), locker(NULL), mdlog(NULL), | |
62 | balancer(NULL), scrubstack(NULL), | |
63 | damage_table(whoami_), | |
64 | inotable(NULL), snapserver(NULL), snapclient(NULL), | |
65 | sessionmap(this), logger(NULL), mlogger(NULL), | |
66 | op_tracker(g_ceph_context, g_conf->mds_enable_op_tracker, | |
67 | g_conf->osd_num_op_tracker_shard), | |
68 | last_state(MDSMap::STATE_BOOT), | |
69 | state(MDSMap::STATE_BOOT), | |
70 | cluster_degraded(false), stopping(false), | |
71 | purge_queue(g_ceph_context, whoami_, | |
72 | mdsmap_->get_metadata_pool(), objecter, | |
73 | new FunctionContext( | |
74 | [this](int r){ | |
75 | // Purge Queue operates inside mds_lock when we're calling into | |
76 | // it, and outside when in background, so must handle both cases. | |
77 | if (mds_lock.is_locked_by_me()) { | |
78 | damaged(); | |
79 | } else { | |
80 | damaged_unlocked(); | |
81 | } | |
82 | } | |
83 | ) | |
84 | ), | |
85 | progress_thread(this), dispatch_depth(0), | |
86 | hb(NULL), last_tid(0), osd_epoch_barrier(0), beacon(beacon_), | |
87 | mds_slow_req_count(0), | |
88 | last_client_mdsmap_bcast(0), | |
89 | messenger(msgr), monc(monc_), | |
90 | respawn_hook(respawn_hook_), | |
91 | suicide_hook(suicide_hook_), | |
92 | standby_replaying(false) | |
93 | { | |
94 | hb = g_ceph_context->get_heartbeat_map()->add_worker("MDSRank", pthread_self()); | |
95 | ||
96 | purge_queue.update_op_limit(*mdsmap); | |
97 | ||
98 | objecter->unset_honor_osdmap_full(); | |
99 | ||
100 | finisher = new Finisher(msgr->cct); | |
101 | ||
102 | mdcache = new MDCache(this, purge_queue); | |
103 | mdlog = new MDLog(this); | |
104 | balancer = new MDBalancer(this, messenger, monc); | |
105 | ||
106 | scrubstack = new ScrubStack(mdcache, finisher); | |
107 | ||
108 | inotable = new InoTable(this); | |
109 | snapserver = new SnapServer(this, monc); | |
110 | snapclient = new SnapClient(this); | |
111 | ||
112 | server = new Server(this); | |
113 | locker = new Locker(this, mdcache); | |
114 | ||
115 | op_tracker.set_complaint_and_threshold(msgr->cct->_conf->mds_op_complaint_time, | |
116 | msgr->cct->_conf->mds_op_log_threshold); | |
117 | op_tracker.set_history_size_and_duration(msgr->cct->_conf->mds_op_history_size, | |
118 | msgr->cct->_conf->mds_op_history_duration); | |
119 | } | |
120 | ||
121 | MDSRank::~MDSRank() | |
122 | { | |
123 | if (hb) { | |
124 | g_ceph_context->get_heartbeat_map()->remove_worker(hb); | |
125 | } | |
126 | ||
127 | if (scrubstack) { delete scrubstack; scrubstack = NULL; } | |
128 | if (mdcache) { delete mdcache; mdcache = NULL; } | |
129 | if (mdlog) { delete mdlog; mdlog = NULL; } | |
130 | if (balancer) { delete balancer; balancer = NULL; } | |
131 | if (inotable) { delete inotable; inotable = NULL; } | |
132 | if (snapserver) { delete snapserver; snapserver = NULL; } | |
133 | if (snapclient) { delete snapclient; snapclient = NULL; } | |
134 | if (mdsmap) { delete mdsmap; mdsmap = 0; } | |
135 | ||
136 | if (server) { delete server; server = 0; } | |
137 | if (locker) { delete locker; locker = 0; } | |
138 | ||
139 | if (logger) { | |
140 | g_ceph_context->get_perfcounters_collection()->remove(logger); | |
141 | delete logger; | |
142 | logger = 0; | |
143 | } | |
144 | if (mlogger) { | |
145 | g_ceph_context->get_perfcounters_collection()->remove(mlogger); | |
146 | delete mlogger; | |
147 | mlogger = 0; | |
148 | } | |
149 | ||
150 | delete finisher; | |
151 | finisher = NULL; | |
152 | ||
153 | delete suicide_hook; | |
154 | suicide_hook = NULL; | |
155 | ||
156 | delete respawn_hook; | |
157 | respawn_hook = NULL; | |
158 | ||
159 | delete objecter; | |
160 | objecter = nullptr; | |
161 | } | |
162 | ||
163 | void MDSRankDispatcher::init() | |
164 | { | |
165 | objecter->init(); | |
166 | messenger->add_dispatcher_head(objecter); | |
167 | ||
168 | objecter->start(); | |
169 | ||
170 | update_log_config(); | |
171 | create_logger(); | |
172 | ||
173 | // Expose the OSDMap (already populated during MDS::init) to anyone | |
174 | // who is interested in it. | |
175 | handle_osd_map(); | |
176 | ||
177 | progress_thread.create("mds_rank_progr"); | |
178 | ||
179 | purge_queue.init(); | |
180 | ||
181 | finisher->start(); | |
182 | } | |
183 | ||
184 | void MDSRank::update_targets(utime_t now) | |
185 | { | |
186 | // get MonMap's idea of my export_targets | |
187 | const set<mds_rank_t>& map_targets = mdsmap->get_mds_info(get_nodeid()).export_targets; | |
188 | ||
189 | dout(20) << "updating export targets, currently " << map_targets.size() << " ranks are targets" << dendl; | |
190 | ||
191 | bool send = false; | |
192 | set<mds_rank_t> new_map_targets; | |
193 | ||
194 | auto it = export_targets.begin(); | |
195 | while (it != export_targets.end()) { | |
196 | mds_rank_t rank = it->first; | |
197 | double val = it->second.get(now); | |
198 | dout(20) << "export target mds." << rank << " value is " << val << " @ " << now << dendl; | |
199 | ||
200 | if (val <= 0.01) { | |
201 | dout(15) << "export target mds." << rank << " is no longer an export target" << dendl; | |
202 | export_targets.erase(it++); | |
203 | send = true; | |
204 | continue; | |
205 | } | |
206 | if (!map_targets.count(rank)) { | |
207 | dout(15) << "export target mds." << rank << " not in map's export_targets" << dendl; | |
208 | send = true; | |
209 | } | |
210 | new_map_targets.insert(rank); | |
211 | it++; | |
212 | } | |
213 | if (new_map_targets.size() < map_targets.size()) { | |
214 | dout(15) << "export target map holds stale targets, sending update" << dendl; | |
215 | send = true; | |
216 | } | |
217 | ||
218 | if (send) { | |
219 | dout(15) << "updating export_targets, now " << new_map_targets.size() << " ranks are targets" << dendl; | |
220 | MMDSLoadTargets* m = new MMDSLoadTargets(mds_gid_t(monc->get_global_id()), new_map_targets); | |
221 | monc->send_mon_message(m); | |
222 | } | |
223 | } | |
224 | ||
225 | void MDSRank::hit_export_target(utime_t now, mds_rank_t rank, double amount) | |
226 | { | |
227 | double rate = g_conf->mds_bal_target_decay; | |
228 | if (amount < 0.0) { | |
229 | amount = 100.0/g_conf->mds_bal_target_decay; /* a good default for "i am trying to keep this export_target active" */ | |
230 | } | |
231 | auto em = export_targets.emplace(std::piecewise_construct, std::forward_as_tuple(rank), std::forward_as_tuple(now, DecayRate(rate))); | |
232 | if (em.second) { | |
233 | dout(15) << "hit export target (new) " << amount << " @ " << now << dendl; | |
234 | } else { | |
235 | dout(15) << "hit export target " << amount << " @ " << now << dendl; | |
236 | } | |
237 | em.first->second.hit(now, amount); | |
238 | } | |
239 | ||
240 | void MDSRankDispatcher::tick() | |
241 | { | |
242 | heartbeat_reset(); | |
243 | ||
244 | if (beacon.is_laggy()) { | |
245 | dout(5) << "tick bailing out since we seem laggy" << dendl; | |
246 | return; | |
247 | } | |
248 | ||
249 | check_ops_in_flight(); | |
250 | ||
251 | // Wake up thread in case we use to be laggy and have waiting_for_nolaggy | |
252 | // messages to progress. | |
253 | progress_thread.signal(); | |
254 | ||
255 | // make sure mds log flushes, trims periodically | |
256 | mdlog->flush(); | |
257 | ||
258 | if (is_active() || is_stopping()) { | |
259 | mdcache->trim(); | |
260 | mdcache->trim_client_leases(); | |
261 | mdcache->check_memory_usage(); | |
262 | mdlog->trim(); // NOT during recovery! | |
263 | } | |
264 | ||
265 | // log | |
266 | mds_load_t load = balancer->get_load(ceph_clock_now()); | |
267 | ||
268 | if (logger) { | |
269 | logger->set(l_mds_load_cent, 100 * load.mds_load()); | |
270 | logger->set(l_mds_dispatch_queue_len, messenger->get_dispatch_queue_len()); | |
271 | logger->set(l_mds_subtrees, mdcache->num_subtrees()); | |
272 | ||
273 | mdcache->log_stat(); | |
274 | } | |
275 | ||
276 | // ... | |
277 | if (is_clientreplay() || is_active() || is_stopping()) { | |
278 | server->find_idle_sessions(); | |
279 | locker->tick(); | |
280 | } | |
281 | ||
282 | if (is_reconnect()) | |
283 | server->reconnect_tick(); | |
284 | ||
285 | if (is_active()) { | |
286 | balancer->tick(); | |
287 | mdcache->find_stale_fragment_freeze(); | |
288 | mdcache->migrator->find_stale_export_freeze(); | |
289 | if (snapserver) | |
290 | snapserver->check_osd_map(false); | |
291 | } | |
292 | ||
293 | if (is_active() || is_stopping()) { | |
294 | update_targets(ceph_clock_now()); | |
295 | } | |
296 | ||
297 | // shut down? | |
298 | if (is_stopping()) { | |
299 | mdlog->trim(); | |
300 | if (mdcache->shutdown_pass()) { | |
301 | uint64_t pq_progress = 0 ; | |
302 | uint64_t pq_total = 0; | |
303 | size_t pq_in_flight = 0; | |
304 | if (!purge_queue.drain(&pq_progress, &pq_total, &pq_in_flight)) { | |
305 | dout(7) << "shutdown_pass=true, but still waiting for purge queue" | |
306 | << dendl; | |
307 | // This takes unbounded time, so we must indicate progress | |
308 | // to the administrator: we do it in a slightly imperfect way | |
309 | // by sending periodic (tick frequency) clog messages while | |
310 | // in this state. | |
311 | clog->info() << "MDS rank " << whoami << " waiting for purge queue (" | |
312 | << std::dec << pq_progress << "/" << pq_total << " " << pq_in_flight | |
313 | << " files purging" << ")"; | |
314 | } else { | |
315 | dout(7) << "shutdown_pass=true, finished w/ shutdown, moving to " | |
316 | "down:stopped" << dendl; | |
317 | stopping_done(); | |
318 | } | |
319 | } | |
320 | else { | |
321 | dout(7) << "shutdown_pass=false" << dendl; | |
322 | } | |
323 | } | |
324 | ||
325 | // Expose ourselves to Beacon to update health indicators | |
326 | beacon.notify_health(this); | |
327 | } | |
328 | ||
329 | void MDSRankDispatcher::shutdown() | |
330 | { | |
331 | // It should never be possible for shutdown to get called twice, because | |
332 | // anyone picking up mds_lock checks if stopping is true and drops | |
333 | // out if it is. | |
334 | assert(stopping == false); | |
335 | stopping = true; | |
336 | ||
337 | dout(1) << __func__ << ": shutting down rank " << whoami << dendl; | |
338 | ||
339 | timer.shutdown(); | |
340 | ||
341 | // MDLog has to shut down before the finisher, because some of its | |
342 | // threads block on IOs that require finisher to complete. | |
343 | mdlog->shutdown(); | |
344 | ||
345 | // shut down cache | |
346 | mdcache->shutdown(); | |
347 | ||
348 | purge_queue.shutdown(); | |
349 | ||
350 | mds_lock.Unlock(); | |
351 | finisher->stop(); // no flushing | |
352 | mds_lock.Lock(); | |
353 | ||
31f18b77 | 354 | if (objecter->initialized) |
7c673cae FG |
355 | objecter->shutdown(); |
356 | ||
357 | monc->shutdown(); | |
358 | ||
359 | op_tracker.on_shutdown(); | |
360 | ||
361 | progress_thread.shutdown(); | |
362 | ||
363 | // release mds_lock for finisher/messenger threads (e.g. | |
364 | // MDSDaemon::ms_handle_reset called from Messenger). | |
365 | mds_lock.Unlock(); | |
366 | ||
367 | // shut down messenger | |
368 | messenger->shutdown(); | |
369 | ||
370 | mds_lock.Lock(); | |
371 | ||
372 | // Workaround unclean shutdown: HeartbeatMap will assert if | |
373 | // worker is not removed (as we do in ~MDS), but ~MDS is not | |
374 | // always called after suicide. | |
375 | if (hb) { | |
376 | g_ceph_context->get_heartbeat_map()->remove_worker(hb); | |
377 | hb = NULL; | |
378 | } | |
379 | } | |
380 | ||
381 | /** | |
382 | * Helper for simple callbacks that call a void fn with no args. | |
383 | */ | |
384 | class C_MDS_VoidFn : public MDSInternalContext | |
385 | { | |
386 | typedef void (MDSRank::*fn_ptr)(); | |
387 | protected: | |
388 | fn_ptr fn; | |
389 | public: | |
390 | C_MDS_VoidFn(MDSRank *mds_, fn_ptr fn_) | |
391 | : MDSInternalContext(mds_), fn(fn_) | |
392 | { | |
393 | assert(mds_); | |
394 | assert(fn_); | |
395 | } | |
396 | ||
397 | void finish(int r) override | |
398 | { | |
399 | (mds->*fn)(); | |
400 | } | |
401 | }; | |
402 | ||
403 | int64_t MDSRank::get_metadata_pool() | |
404 | { | |
405 | return mdsmap->get_metadata_pool(); | |
406 | } | |
407 | ||
408 | MDSTableClient *MDSRank::get_table_client(int t) | |
409 | { | |
410 | switch (t) { | |
411 | case TABLE_ANCHOR: return NULL; | |
412 | case TABLE_SNAP: return snapclient; | |
413 | default: ceph_abort(); | |
414 | } | |
415 | } | |
416 | ||
417 | MDSTableServer *MDSRank::get_table_server(int t) | |
418 | { | |
419 | switch (t) { | |
420 | case TABLE_ANCHOR: return NULL; | |
421 | case TABLE_SNAP: return snapserver; | |
422 | default: ceph_abort(); | |
423 | } | |
424 | } | |
425 | ||
426 | void MDSRank::suicide() | |
427 | { | |
428 | if (suicide_hook) { | |
429 | suicide_hook->complete(0); | |
430 | suicide_hook = NULL; | |
431 | } | |
432 | } | |
433 | ||
434 | void MDSRank::respawn() | |
435 | { | |
436 | if (respawn_hook) { | |
437 | respawn_hook->complete(0); | |
438 | respawn_hook = NULL; | |
439 | } | |
440 | } | |
441 | ||
442 | void MDSRank::damaged() | |
443 | { | |
444 | assert(whoami != MDS_RANK_NONE); | |
445 | assert(mds_lock.is_locked_by_me()); | |
446 | ||
447 | beacon.set_want_state(mdsmap, MDSMap::STATE_DAMAGED); | |
448 | monc->flush_log(); // Flush any clog error from before we were called | |
449 | beacon.notify_health(this); // Include latest status in our swan song | |
450 | beacon.send_and_wait(g_conf->mds_mon_shutdown_timeout); | |
451 | ||
452 | // It's okay if we timed out and the mon didn't get our beacon, because | |
453 | // another daemon (or ourselves after respawn) will eventually take the | |
454 | // rank and report DAMAGED again when it hits same problem we did. | |
455 | ||
456 | respawn(); // Respawn into standby in case mon has other work for us | |
457 | } | |
458 | ||
459 | void MDSRank::damaged_unlocked() | |
460 | { | |
461 | Mutex::Locker l(mds_lock); | |
462 | damaged(); | |
463 | } | |
464 | ||
465 | void MDSRank::handle_write_error(int err) | |
466 | { | |
467 | if (err == -EBLACKLISTED) { | |
468 | derr << "we have been blacklisted (fenced), respawning..." << dendl; | |
469 | respawn(); | |
470 | return; | |
471 | } | |
472 | ||
473 | if (g_conf->mds_action_on_write_error >= 2) { | |
474 | derr << "unhandled write error " << cpp_strerror(err) << ", suicide..." << dendl; | |
475 | respawn(); | |
476 | } else if (g_conf->mds_action_on_write_error == 1) { | |
477 | derr << "unhandled write error " << cpp_strerror(err) << ", force readonly..." << dendl; | |
478 | mdcache->force_readonly(); | |
479 | } else { | |
480 | // ignore; | |
481 | derr << "unhandled write error " << cpp_strerror(err) << ", ignore..." << dendl; | |
482 | } | |
483 | } | |
484 | ||
485 | void *MDSRank::ProgressThread::entry() | |
486 | { | |
487 | Mutex::Locker l(mds->mds_lock); | |
488 | while (true) { | |
489 | while (!mds->stopping && | |
490 | mds->finished_queue.empty() && | |
491 | (mds->waiting_for_nolaggy.empty() || mds->beacon.is_laggy())) { | |
492 | cond.Wait(mds->mds_lock); | |
493 | } | |
494 | ||
495 | if (mds->stopping) { | |
496 | break; | |
497 | } | |
498 | ||
499 | mds->_advance_queues(); | |
500 | } | |
501 | ||
502 | return NULL; | |
503 | } | |
504 | ||
505 | ||
506 | void MDSRank::ProgressThread::shutdown() | |
507 | { | |
508 | assert(mds->mds_lock.is_locked_by_me()); | |
509 | assert(mds->stopping); | |
510 | ||
511 | if (am_self()) { | |
512 | // Stopping is set, we will fall out of our main loop naturally | |
513 | } else { | |
514 | // Kick the thread to notice mds->stopping, and join it | |
515 | cond.Signal(); | |
516 | mds->mds_lock.Unlock(); | |
517 | if (is_started()) | |
518 | join(); | |
519 | mds->mds_lock.Lock(); | |
520 | } | |
521 | } | |
522 | ||
523 | bool MDSRankDispatcher::ms_dispatch(Message *m) | |
524 | { | |
525 | bool ret; | |
526 | inc_dispatch_depth(); | |
527 | ret = _dispatch(m, true); | |
528 | dec_dispatch_depth(); | |
529 | return ret; | |
530 | } | |
531 | ||
532 | /* If this function returns true, it recognizes the message and has taken the | |
533 | * reference. If it returns false, it has done neither. */ | |
534 | bool MDSRank::_dispatch(Message *m, bool new_msg) | |
535 | { | |
536 | if (is_stale_message(m)) { | |
537 | m->put(); | |
538 | return true; | |
539 | } | |
540 | ||
541 | if (beacon.is_laggy()) { | |
542 | dout(10) << " laggy, deferring " << *m << dendl; | |
543 | waiting_for_nolaggy.push_back(m); | |
544 | } else if (new_msg && !waiting_for_nolaggy.empty()) { | |
545 | dout(10) << " there are deferred messages, deferring " << *m << dendl; | |
546 | waiting_for_nolaggy.push_back(m); | |
547 | } else { | |
548 | if (!handle_deferrable_message(m)) { | |
549 | dout(0) << "unrecognized message " << *m << dendl; | |
550 | return false; | |
551 | } | |
552 | ||
553 | heartbeat_reset(); | |
554 | } | |
555 | ||
556 | if (dispatch_depth > 1) | |
557 | return true; | |
558 | ||
559 | // finish any triggered contexts | |
560 | _advance_queues(); | |
561 | ||
562 | if (beacon.is_laggy()) { | |
563 | // We've gone laggy during dispatch, don't do any | |
564 | // more housekeeping | |
565 | return true; | |
566 | } | |
567 | ||
568 | // done with all client replayed requests? | |
569 | if (is_clientreplay() && | |
570 | mdcache->is_open() && | |
571 | replay_queue.empty() && | |
572 | beacon.get_want_state() == MDSMap::STATE_CLIENTREPLAY) { | |
573 | int num_requests = mdcache->get_num_client_requests(); | |
574 | dout(10) << " still have " << num_requests << " active replay requests" << dendl; | |
575 | if (num_requests == 0) | |
576 | clientreplay_done(); | |
577 | } | |
578 | ||
579 | // hack: thrash exports | |
580 | static utime_t start; | |
581 | utime_t now = ceph_clock_now(); | |
582 | if (start == utime_t()) | |
583 | start = now; | |
584 | /*double el = now - start; | |
585 | if (el > 30.0 && | |
586 | el < 60.0)*/ | |
587 | for (int i=0; i<g_conf->mds_thrash_exports; i++) { | |
588 | set<mds_rank_t> s; | |
589 | if (!is_active()) break; | |
590 | mdsmap->get_mds_set(s, MDSMap::STATE_ACTIVE); | |
591 | if (s.size() < 2 || CInode::count() < 10) | |
592 | break; // need peers for this to work. | |
593 | if (mdcache->migrator->get_num_exporting() > g_conf->mds_thrash_exports * 5 || | |
594 | mdcache->migrator->get_export_queue_size() > g_conf->mds_thrash_exports * 10) | |
595 | break; | |
596 | ||
597 | dout(7) << "mds thrashing exports pass " << (i+1) << "/" << g_conf->mds_thrash_exports << dendl; | |
598 | ||
599 | // pick a random dir inode | |
600 | CInode *in = mdcache->hack_pick_random_inode(); | |
601 | ||
602 | list<CDir*> ls; | |
603 | in->get_dirfrags(ls); | |
604 | if (!ls.empty()) { // must be an open dir. | |
605 | list<CDir*>::iterator p = ls.begin(); | |
606 | int n = rand() % ls.size(); | |
607 | while (n--) | |
608 | ++p; | |
609 | CDir *dir = *p; | |
610 | if (!dir->get_parent_dir()) continue; // must be linked. | |
611 | if (!dir->is_auth()) continue; // must be auth. | |
612 | ||
613 | mds_rank_t dest; | |
614 | do { | |
615 | int k = rand() % s.size(); | |
616 | set<mds_rank_t>::iterator p = s.begin(); | |
617 | while (k--) ++p; | |
618 | dest = *p; | |
619 | } while (dest == whoami); | |
620 | mdcache->migrator->export_dir_nicely(dir,dest); | |
621 | } | |
622 | } | |
623 | // hack: thrash fragments | |
624 | for (int i=0; i<g_conf->mds_thrash_fragments; i++) { | |
625 | if (!is_active()) break; | |
626 | if (mdcache->get_num_fragmenting_dirs() > 5 * g_conf->mds_thrash_fragments) break; | |
627 | dout(7) << "mds thrashing fragments pass " << (i+1) << "/" << g_conf->mds_thrash_fragments << dendl; | |
628 | ||
629 | // pick a random dir inode | |
630 | CInode *in = mdcache->hack_pick_random_inode(); | |
631 | ||
632 | list<CDir*> ls; | |
633 | in->get_dirfrags(ls); | |
634 | if (ls.empty()) continue; // must be an open dir. | |
635 | CDir *dir = ls.front(); | |
636 | if (!dir->get_parent_dir()) continue; // must be linked. | |
637 | if (!dir->is_auth()) continue; // must be auth. | |
638 | frag_t fg = dir->get_frag(); | |
639 | if (mdsmap->allows_dirfrags()) { | |
640 | if ((fg == frag_t() || (rand() % (1 << fg.bits()) == 0))) { | |
641 | mdcache->split_dir(dir, 1); | |
642 | } else { | |
643 | balancer->queue_merge(dir); | |
644 | } | |
645 | } | |
646 | } | |
647 | ||
648 | // hack: force hash root? | |
649 | /* | |
650 | if (false && | |
651 | mdcache->get_root() && | |
652 | mdcache->get_root()->dir && | |
653 | !(mdcache->get_root()->dir->is_hashed() || | |
654 | mdcache->get_root()->dir->is_hashing())) { | |
655 | dout(0) << "hashing root" << dendl; | |
656 | mdcache->migrator->hash_dir(mdcache->get_root()->dir); | |
657 | } | |
658 | */ | |
659 | ||
c07f9fc5 FG |
660 | update_mlogger(); |
661 | return true; | |
662 | } | |
663 | ||
664 | void MDSRank::update_mlogger() | |
665 | { | |
7c673cae FG |
666 | if (mlogger) { |
667 | mlogger->set(l_mdm_ino, CInode::count()); | |
668 | mlogger->set(l_mdm_dir, CDir::count()); | |
669 | mlogger->set(l_mdm_dn, CDentry::count()); | |
670 | mlogger->set(l_mdm_cap, Capability::count()); | |
7c673cae FG |
671 | mlogger->set(l_mdm_inoa, CInode::increments()); |
672 | mlogger->set(l_mdm_inos, CInode::decrements()); | |
673 | mlogger->set(l_mdm_dira, CDir::increments()); | |
674 | mlogger->set(l_mdm_dirs, CDir::decrements()); | |
675 | mlogger->set(l_mdm_dna, CDentry::increments()); | |
676 | mlogger->set(l_mdm_dns, CDentry::decrements()); | |
677 | mlogger->set(l_mdm_capa, Capability::increments()); | |
678 | mlogger->set(l_mdm_caps, Capability::decrements()); | |
7c673cae FG |
679 | mlogger->set(l_mdm_buf, buffer::get_total_alloc()); |
680 | } | |
7c673cae FG |
681 | } |
682 | ||
683 | /* | |
684 | * lower priority messages we defer if we seem laggy | |
685 | */ | |
686 | bool MDSRank::handle_deferrable_message(Message *m) | |
687 | { | |
688 | int port = m->get_type() & 0xff00; | |
689 | ||
690 | switch (port) { | |
691 | case MDS_PORT_CACHE: | |
692 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
693 | mdcache->dispatch(m); | |
694 | break; | |
695 | ||
696 | case MDS_PORT_MIGRATOR: | |
697 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
698 | mdcache->migrator->dispatch(m); | |
699 | break; | |
700 | ||
701 | default: | |
702 | switch (m->get_type()) { | |
703 | // SERVER | |
704 | case CEPH_MSG_CLIENT_SESSION: | |
705 | case CEPH_MSG_CLIENT_RECONNECT: | |
706 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_CLIENT); | |
707 | // fall-thru | |
708 | case CEPH_MSG_CLIENT_REQUEST: | |
709 | server->dispatch(m); | |
710 | break; | |
711 | case MSG_MDS_SLAVE_REQUEST: | |
712 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
713 | server->dispatch(m); | |
714 | break; | |
715 | ||
716 | case MSG_MDS_HEARTBEAT: | |
717 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
718 | balancer->proc_message(m); | |
719 | break; | |
720 | ||
721 | case MSG_MDS_TABLE_REQUEST: | |
722 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
723 | { | |
724 | MMDSTableRequest *req = static_cast<MMDSTableRequest*>(m); | |
725 | if (req->op < 0) { | |
726 | MDSTableClient *client = get_table_client(req->table); | |
727 | client->handle_request(req); | |
728 | } else { | |
729 | MDSTableServer *server = get_table_server(req->table); | |
730 | server->handle_request(req); | |
731 | } | |
732 | } | |
733 | break; | |
734 | ||
735 | case MSG_MDS_LOCK: | |
736 | case MSG_MDS_INODEFILECAPS: | |
737 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
738 | locker->dispatch(m); | |
739 | break; | |
740 | ||
741 | case CEPH_MSG_CLIENT_CAPS: | |
742 | case CEPH_MSG_CLIENT_CAPRELEASE: | |
743 | case CEPH_MSG_CLIENT_LEASE: | |
744 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_CLIENT); | |
745 | locker->dispatch(m); | |
746 | break; | |
747 | ||
748 | default: | |
749 | return false; | |
750 | } | |
751 | } | |
752 | ||
753 | return true; | |
754 | } | |
755 | ||
756 | /** | |
757 | * Advance finished_queue and waiting_for_nolaggy. | |
758 | * | |
759 | * Usually drain both queues, but may not drain waiting_for_nolaggy | |
760 | * if beacon is currently laggy. | |
761 | */ | |
762 | void MDSRank::_advance_queues() | |
763 | { | |
764 | assert(mds_lock.is_locked_by_me()); | |
765 | ||
766 | while (!finished_queue.empty()) { | |
767 | dout(7) << "mds has " << finished_queue.size() << " queued contexts" << dendl; | |
768 | dout(10) << finished_queue << dendl; | |
769 | list<MDSInternalContextBase*> ls; | |
770 | ls.swap(finished_queue); | |
771 | while (!ls.empty()) { | |
772 | dout(10) << " finish " << ls.front() << dendl; | |
773 | ls.front()->complete(0); | |
774 | ls.pop_front(); | |
775 | ||
776 | heartbeat_reset(); | |
777 | } | |
778 | } | |
779 | ||
780 | while (!waiting_for_nolaggy.empty()) { | |
781 | // stop if we're laggy now! | |
782 | if (beacon.is_laggy()) | |
783 | break; | |
784 | ||
785 | Message *old = waiting_for_nolaggy.front(); | |
786 | waiting_for_nolaggy.pop_front(); | |
787 | ||
788 | if (is_stale_message(old)) { | |
789 | old->put(); | |
790 | } else { | |
791 | dout(7) << " processing laggy deferred " << *old << dendl; | |
792 | if (!handle_deferrable_message(old)) { | |
793 | dout(0) << "unrecognized message " << *old << dendl; | |
794 | old->put(); | |
795 | } | |
796 | } | |
797 | ||
798 | heartbeat_reset(); | |
799 | } | |
800 | } | |
801 | ||
802 | /** | |
803 | * Call this when you take mds_lock, or periodically if you're going to | |
804 | * hold the lock for a long time (e.g. iterating over clients/inodes) | |
805 | */ | |
806 | void MDSRank::heartbeat_reset() | |
807 | { | |
808 | // Any thread might jump into mds_lock and call us immediately | |
809 | // after a call to suicide() completes, in which case MDSRank::hb | |
810 | // has been freed and we are a no-op. | |
811 | if (!hb) { | |
812 | assert(stopping); | |
813 | return; | |
814 | } | |
815 | ||
816 | // NB not enabling suicide grace, because the mon takes care of killing us | |
817 | // (by blacklisting us) when we fail to send beacons, and it's simpler to | |
818 | // only have one way of dying. | |
819 | g_ceph_context->get_heartbeat_map()->reset_timeout(hb, g_conf->mds_beacon_grace, 0); | |
820 | } | |
821 | ||
822 | bool MDSRank::is_stale_message(Message *m) const | |
823 | { | |
824 | // from bad mds? | |
825 | if (m->get_source().is_mds()) { | |
826 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
827 | if (!mdsmap->have_inst(from) || | |
828 | mdsmap->get_inst(from) != m->get_source_inst() || | |
829 | mdsmap->is_down(from)) { | |
830 | // bogus mds? | |
831 | if (m->get_type() == CEPH_MSG_MDS_MAP) { | |
832 | dout(5) << "got " << *m << " from old/bad/imposter mds " << m->get_source() | |
833 | << ", but it's an mdsmap, looking at it" << dendl; | |
834 | } else if (m->get_type() == MSG_MDS_CACHEEXPIRE && | |
835 | mdsmap->get_inst(from) == m->get_source_inst()) { | |
836 | dout(5) << "got " << *m << " from down mds " << m->get_source() | |
837 | << ", but it's a cache_expire, looking at it" << dendl; | |
838 | } else { | |
839 | dout(5) << "got " << *m << " from down/old/bad/imposter mds " << m->get_source() | |
840 | << ", dropping" << dendl; | |
841 | return true; | |
842 | } | |
843 | } | |
844 | } | |
845 | return false; | |
846 | } | |
847 | ||
848 | ||
849 | void MDSRank::send_message(Message *m, Connection *c) | |
850 | { | |
851 | assert(c); | |
852 | c->send_message(m); | |
853 | } | |
854 | ||
855 | ||
856 | void MDSRank::send_message_mds(Message *m, mds_rank_t mds) | |
857 | { | |
858 | if (!mdsmap->is_up(mds)) { | |
859 | dout(10) << "send_message_mds mds." << mds << " not up, dropping " << *m << dendl; | |
860 | m->put(); | |
861 | return; | |
862 | } | |
863 | ||
864 | // send mdsmap first? | |
865 | if (mds != whoami && peer_mdsmap_epoch[mds] < mdsmap->get_epoch()) { | |
866 | messenger->send_message(new MMDSMap(monc->get_fsid(), mdsmap), | |
867 | mdsmap->get_inst(mds)); | |
868 | peer_mdsmap_epoch[mds] = mdsmap->get_epoch(); | |
869 | } | |
870 | ||
871 | // send message | |
872 | messenger->send_message(m, mdsmap->get_inst(mds)); | |
873 | } | |
874 | ||
875 | void MDSRank::forward_message_mds(Message *m, mds_rank_t mds) | |
876 | { | |
877 | assert(mds != whoami); | |
878 | ||
879 | // client request? | |
880 | if (m->get_type() == CEPH_MSG_CLIENT_REQUEST && | |
881 | (static_cast<MClientRequest*>(m))->get_source().is_client()) { | |
882 | MClientRequest *creq = static_cast<MClientRequest*>(m); | |
883 | creq->inc_num_fwd(); // inc forward counter | |
884 | ||
885 | /* | |
886 | * don't actually forward if non-idempotent! | |
887 | * client has to do it. although the MDS will ignore duplicate requests, | |
888 | * the affected metadata may migrate, in which case the new authority | |
889 | * won't have the metareq_id in the completed request map. | |
890 | */ | |
891 | // NEW: always make the client resend! | |
892 | bool client_must_resend = true; //!creq->can_forward(); | |
893 | ||
894 | // tell the client where it should go | |
895 | messenger->send_message(new MClientRequestForward(creq->get_tid(), mds, creq->get_num_fwd(), | |
896 | client_must_resend), | |
897 | creq->get_source_inst()); | |
898 | ||
899 | if (client_must_resend) { | |
900 | m->put(); | |
901 | return; | |
902 | } | |
903 | } | |
904 | ||
905 | // these are the only types of messages we should be 'forwarding'; they | |
906 | // explicitly encode their source mds, which gets clobbered when we resend | |
907 | // them here. | |
908 | assert(m->get_type() == MSG_MDS_DIRUPDATE || | |
909 | m->get_type() == MSG_MDS_EXPORTDIRDISCOVER); | |
910 | ||
911 | // send mdsmap first? | |
912 | if (peer_mdsmap_epoch[mds] < mdsmap->get_epoch()) { | |
913 | messenger->send_message(new MMDSMap(monc->get_fsid(), mdsmap), | |
914 | mdsmap->get_inst(mds)); | |
915 | peer_mdsmap_epoch[mds] = mdsmap->get_epoch(); | |
916 | } | |
917 | ||
918 | messenger->send_message(m, mdsmap->get_inst(mds)); | |
919 | } | |
920 | ||
921 | ||
922 | ||
923 | void MDSRank::send_message_client_counted(Message *m, client_t client) | |
924 | { | |
925 | Session *session = sessionmap.get_session(entity_name_t::CLIENT(client.v)); | |
926 | if (session) { | |
927 | send_message_client_counted(m, session); | |
928 | } else { | |
929 | dout(10) << "send_message_client_counted no session for client." << client << " " << *m << dendl; | |
930 | } | |
931 | } | |
932 | ||
933 | void MDSRank::send_message_client_counted(Message *m, Connection *connection) | |
934 | { | |
935 | Session *session = static_cast<Session *>(connection->get_priv()); | |
936 | if (session) { | |
937 | session->put(); // do not carry ref | |
938 | send_message_client_counted(m, session); | |
939 | } else { | |
940 | dout(10) << "send_message_client_counted has no session for " << m->get_source_inst() << dendl; | |
941 | // another Connection took over the Session | |
942 | } | |
943 | } | |
944 | ||
945 | void MDSRank::send_message_client_counted(Message *m, Session *session) | |
946 | { | |
947 | version_t seq = session->inc_push_seq(); | |
948 | dout(10) << "send_message_client_counted " << session->info.inst.name << " seq " | |
949 | << seq << " " << *m << dendl; | |
950 | if (session->connection) { | |
951 | session->connection->send_message(m); | |
952 | } else { | |
953 | session->preopen_out_queue.push_back(m); | |
954 | } | |
955 | } | |
956 | ||
957 | void MDSRank::send_message_client(Message *m, Session *session) | |
958 | { | |
959 | dout(10) << "send_message_client " << session->info.inst << " " << *m << dendl; | |
960 | if (session->connection) { | |
961 | session->connection->send_message(m); | |
962 | } else { | |
963 | session->preopen_out_queue.push_back(m); | |
964 | } | |
965 | } | |
966 | ||
967 | /** | |
968 | * This is used whenever a RADOS operation has been cancelled | |
969 | * or a RADOS client has been blacklisted, to cause the MDS and | |
970 | * any clients to wait for this OSD epoch before using any new caps. | |
971 | * | |
972 | * See doc/cephfs/eviction | |
973 | */ | |
974 | void MDSRank::set_osd_epoch_barrier(epoch_t e) | |
975 | { | |
976 | dout(4) << __func__ << ": epoch=" << e << dendl; | |
977 | osd_epoch_barrier = e; | |
978 | } | |
979 | ||
7c673cae FG |
980 | void MDSRank::retry_dispatch(Message *m) |
981 | { | |
982 | inc_dispatch_depth(); | |
983 | _dispatch(m, false); | |
984 | dec_dispatch_depth(); | |
985 | } | |
986 | ||
987 | utime_t MDSRank::get_laggy_until() const | |
988 | { | |
989 | return beacon.get_laggy_until(); | |
990 | } | |
991 | ||
992 | bool MDSRank::is_daemon_stopping() const | |
993 | { | |
994 | return stopping; | |
995 | } | |
996 | ||
997 | void MDSRank::request_state(MDSMap::DaemonState s) | |
998 | { | |
999 | dout(3) << "request_state " << ceph_mds_state_name(s) << dendl; | |
1000 | beacon.set_want_state(mdsmap, s); | |
1001 | beacon.send(); | |
1002 | } | |
1003 | ||
1004 | ||
1005 | class C_MDS_BootStart : public MDSInternalContext { | |
1006 | MDSRank::BootStep nextstep; | |
1007 | public: | |
1008 | C_MDS_BootStart(MDSRank *m, MDSRank::BootStep n) | |
1009 | : MDSInternalContext(m), nextstep(n) {} | |
1010 | void finish(int r) override { | |
1011 | mds->boot_start(nextstep, r); | |
1012 | } | |
1013 | }; | |
1014 | ||
1015 | ||
1016 | void MDSRank::boot_start(BootStep step, int r) | |
1017 | { | |
1018 | // Handle errors from previous step | |
1019 | if (r < 0) { | |
1020 | if (is_standby_replay() && (r == -EAGAIN)) { | |
1021 | dout(0) << "boot_start encountered an error EAGAIN" | |
1022 | << ", respawning since we fell behind journal" << dendl; | |
1023 | respawn(); | |
1024 | } else if (r == -EINVAL || r == -ENOENT) { | |
1025 | // Invalid or absent data, indicates damaged on-disk structures | |
1026 | clog->error() << "Error loading MDS rank " << whoami << ": " | |
1027 | << cpp_strerror(r); | |
1028 | damaged(); | |
1029 | assert(r == 0); // Unreachable, damaged() calls respawn() | |
1030 | } else { | |
1031 | // Completely unexpected error, give up and die | |
1032 | dout(0) << "boot_start encountered an error, failing" << dendl; | |
1033 | suicide(); | |
1034 | return; | |
1035 | } | |
1036 | } | |
1037 | ||
1038 | assert(is_starting() || is_any_replay()); | |
1039 | ||
1040 | switch(step) { | |
1041 | case MDS_BOOT_INITIAL: | |
1042 | { | |
1043 | mdcache->init_layouts(); | |
1044 | ||
1045 | MDSGatherBuilder gather(g_ceph_context, | |
1046 | new C_MDS_BootStart(this, MDS_BOOT_OPEN_ROOT)); | |
1047 | dout(2) << "boot_start " << step << ": opening inotable" << dendl; | |
1048 | inotable->set_rank(whoami); | |
1049 | inotable->load(gather.new_sub()); | |
1050 | ||
1051 | dout(2) << "boot_start " << step << ": opening sessionmap" << dendl; | |
1052 | sessionmap.set_rank(whoami); | |
1053 | sessionmap.load(gather.new_sub()); | |
1054 | ||
1055 | dout(2) << "boot_start " << step << ": opening mds log" << dendl; | |
1056 | mdlog->open(gather.new_sub()); | |
1057 | ||
1058 | if (mdsmap->get_tableserver() == whoami) { | |
1059 | dout(2) << "boot_start " << step << ": opening snap table" << dendl; | |
1060 | snapserver->set_rank(whoami); | |
1061 | snapserver->load(gather.new_sub()); | |
1062 | } | |
1063 | ||
1064 | gather.activate(); | |
1065 | } | |
1066 | break; | |
1067 | case MDS_BOOT_OPEN_ROOT: | |
1068 | { | |
1069 | dout(2) << "boot_start " << step << ": loading/discovering base inodes" << dendl; | |
1070 | ||
1071 | MDSGatherBuilder gather(g_ceph_context, | |
1072 | new C_MDS_BootStart(this, MDS_BOOT_PREPARE_LOG)); | |
1073 | ||
1074 | mdcache->open_mydir_inode(gather.new_sub()); | |
1075 | ||
1076 | purge_queue.open(new C_IO_Wrapper(this, gather.new_sub())); | |
1077 | ||
1078 | if (is_starting() || | |
1079 | whoami == mdsmap->get_root()) { // load root inode off disk if we are auth | |
1080 | mdcache->open_root_inode(gather.new_sub()); | |
1081 | } else { | |
1082 | // replay. make up fake root inode to start with | |
1083 | (void)mdcache->create_root_inode(); | |
1084 | } | |
1085 | gather.activate(); | |
1086 | } | |
1087 | break; | |
1088 | case MDS_BOOT_PREPARE_LOG: | |
1089 | if (is_any_replay()) { | |
1090 | dout(2) << "boot_start " << step << ": replaying mds log" << dendl; | |
1091 | mdlog->replay(new C_MDS_BootStart(this, MDS_BOOT_REPLAY_DONE)); | |
1092 | } else { | |
1093 | dout(2) << "boot_start " << step << ": positioning at end of old mds log" << dendl; | |
1094 | mdlog->append(); | |
1095 | starting_done(); | |
1096 | } | |
1097 | break; | |
1098 | case MDS_BOOT_REPLAY_DONE: | |
1099 | assert(is_any_replay()); | |
1100 | ||
1101 | // Sessiontable and inotable should be in sync after replay, validate | |
1102 | // that they are consistent. | |
1103 | validate_sessions(); | |
1104 | ||
1105 | replay_done(); | |
1106 | break; | |
1107 | } | |
1108 | } | |
1109 | ||
1110 | void MDSRank::validate_sessions() | |
1111 | { | |
1112 | assert(mds_lock.is_locked_by_me()); | |
1113 | std::vector<Session*> victims; | |
1114 | ||
1115 | // Identify any sessions which have state inconsistent with other, | |
1116 | // after they have been loaded from rados during startup. | |
1117 | // Mitigate bugs like: http://tracker.ceph.com/issues/16842 | |
1118 | const auto &sessions = sessionmap.get_sessions(); | |
1119 | for (const auto &i : sessions) { | |
1120 | Session *session = i.second; | |
1121 | interval_set<inodeno_t> badones; | |
1122 | if (inotable->intersects_free(session->info.prealloc_inos, &badones)) { | |
1123 | clog->error() << "Client session loaded with invalid preallocated " | |
1124 | "inodes, evicting session " << *session; | |
1125 | ||
1126 | // Make the session consistent with inotable so that it can | |
1127 | // be cleanly torn down | |
1128 | session->info.prealloc_inos.subtract(badones); | |
1129 | ||
1130 | victims.push_back(session); | |
1131 | } | |
1132 | } | |
1133 | ||
1134 | for (const auto &session: victims) { | |
1135 | server->kill_session(session, nullptr); | |
1136 | } | |
1137 | } | |
1138 | ||
1139 | void MDSRank::starting_done() | |
1140 | { | |
1141 | dout(3) << "starting_done" << dendl; | |
1142 | assert(is_starting()); | |
1143 | request_state(MDSMap::STATE_ACTIVE); | |
1144 | ||
1145 | mdcache->open_root(); | |
1146 | ||
d2e6a577 FG |
1147 | if (mdcache->is_open()) { |
1148 | mdlog->start_new_segment(); | |
1149 | } else { | |
1150 | mdcache->wait_for_open(new MDSInternalContextWrapper(this, | |
1151 | new FunctionContext([this] (int r) { | |
1152 | mdlog->start_new_segment(); | |
1153 | }))); | |
1154 | } | |
7c673cae FG |
1155 | } |
1156 | ||
1157 | ||
1158 | void MDSRank::calc_recovery_set() | |
1159 | { | |
1160 | // initialize gather sets | |
1161 | set<mds_rank_t> rs; | |
1162 | mdsmap->get_recovery_mds_set(rs); | |
1163 | rs.erase(whoami); | |
1164 | mdcache->set_recovery_set(rs); | |
1165 | ||
1166 | dout(1) << " recovery set is " << rs << dendl; | |
1167 | } | |
1168 | ||
1169 | ||
1170 | void MDSRank::replay_start() | |
1171 | { | |
1172 | dout(1) << "replay_start" << dendl; | |
1173 | ||
1174 | if (is_standby_replay()) | |
1175 | standby_replaying = true; | |
1176 | ||
1177 | calc_recovery_set(); | |
1178 | ||
1179 | // Check if we need to wait for a newer OSD map before starting | |
1180 | Context *fin = new C_IO_Wrapper(this, new C_MDS_BootStart(this, MDS_BOOT_INITIAL)); | |
1181 | bool const ready = objecter->wait_for_map( | |
1182 | mdsmap->get_last_failure_osd_epoch(), | |
1183 | fin); | |
1184 | ||
1185 | if (ready) { | |
1186 | delete fin; | |
1187 | boot_start(); | |
1188 | } else { | |
1189 | dout(1) << " waiting for osdmap " << mdsmap->get_last_failure_osd_epoch() | |
1190 | << " (which blacklists prior instance)" << dendl; | |
1191 | } | |
1192 | } | |
1193 | ||
1194 | ||
1195 | class MDSRank::C_MDS_StandbyReplayRestartFinish : public MDSIOContext { | |
1196 | uint64_t old_read_pos; | |
1197 | public: | |
1198 | C_MDS_StandbyReplayRestartFinish(MDSRank *mds_, uint64_t old_read_pos_) : | |
1199 | MDSIOContext(mds_), old_read_pos(old_read_pos_) {} | |
1200 | void finish(int r) override { | |
1201 | mds->_standby_replay_restart_finish(r, old_read_pos); | |
1202 | } | |
1203 | }; | |
1204 | ||
1205 | void MDSRank::_standby_replay_restart_finish(int r, uint64_t old_read_pos) | |
1206 | { | |
1207 | if (old_read_pos < mdlog->get_journaler()->get_trimmed_pos()) { | |
1208 | dout(0) << "standby MDS fell behind active MDS journal's expire_pos, restarting" << dendl; | |
1209 | respawn(); /* we're too far back, and this is easier than | |
1210 | trying to reset everything in the cache, etc */ | |
1211 | } else { | |
1212 | mdlog->standby_trim_segments(); | |
1213 | boot_start(MDS_BOOT_PREPARE_LOG, r); | |
1214 | } | |
1215 | } | |
1216 | ||
1217 | inline void MDSRank::standby_replay_restart() | |
1218 | { | |
1219 | if (standby_replaying) { | |
1220 | /* Go around for another pass of replaying in standby */ | |
1221 | dout(4) << "standby_replay_restart (as standby)" << dendl; | |
1222 | mdlog->get_journaler()->reread_head_and_probe( | |
1223 | new C_MDS_StandbyReplayRestartFinish( | |
1224 | this, | |
1225 | mdlog->get_journaler()->get_read_pos())); | |
1226 | } else { | |
1227 | /* We are transitioning out of standby: wait for OSD map update | |
1228 | before making final pass */ | |
1229 | dout(1) << "standby_replay_restart (final takeover pass)" << dendl; | |
1230 | Context *fin = new C_IO_Wrapper(this, new C_MDS_BootStart(this, MDS_BOOT_PREPARE_LOG)); | |
1231 | bool const ready = | |
1232 | objecter->wait_for_map(mdsmap->get_last_failure_osd_epoch(), fin); | |
1233 | if (ready) { | |
1234 | delete fin; | |
1235 | mdlog->get_journaler()->reread_head_and_probe( | |
1236 | new C_MDS_StandbyReplayRestartFinish( | |
1237 | this, | |
1238 | mdlog->get_journaler()->get_read_pos())); | |
1239 | } else { | |
1240 | dout(1) << " waiting for osdmap " << mdsmap->get_last_failure_osd_epoch() | |
1241 | << " (which blacklists prior instance)" << dendl; | |
1242 | } | |
1243 | } | |
1244 | } | |
1245 | ||
1246 | class MDSRank::C_MDS_StandbyReplayRestart : public MDSInternalContext { | |
1247 | public: | |
1248 | explicit C_MDS_StandbyReplayRestart(MDSRank *m) : MDSInternalContext(m) {} | |
1249 | void finish(int r) override { | |
1250 | assert(!r); | |
1251 | mds->standby_replay_restart(); | |
1252 | } | |
1253 | }; | |
1254 | ||
1255 | void MDSRank::replay_done() | |
1256 | { | |
1257 | dout(1) << "replay_done" << (standby_replaying ? " (as standby)" : "") << dendl; | |
1258 | ||
1259 | if (is_standby_replay()) { | |
1260 | // The replay was done in standby state, and we are still in that state | |
1261 | assert(standby_replaying); | |
1262 | dout(10) << "setting replay timer" << dendl; | |
1263 | timer.add_event_after(g_conf->mds_replay_interval, | |
1264 | new C_MDS_StandbyReplayRestart(this)); | |
1265 | return; | |
1266 | } else if (standby_replaying) { | |
1267 | // The replay was done in standby state, we have now _left_ that state | |
1268 | dout(10) << " last replay pass was as a standby; making final pass" << dendl; | |
1269 | standby_replaying = false; | |
1270 | standby_replay_restart(); | |
1271 | return; | |
1272 | } else { | |
1273 | // Replay is complete, journal read should be up to date | |
1274 | assert(mdlog->get_journaler()->get_read_pos() == mdlog->get_journaler()->get_write_pos()); | |
1275 | assert(!is_standby_replay()); | |
1276 | ||
1277 | // Reformat and come back here | |
1278 | if (mdlog->get_journaler()->get_stream_format() < g_conf->mds_journal_format) { | |
1279 | dout(4) << "reformatting journal on standbyreplay->replay transition" << dendl; | |
1280 | mdlog->reopen(new C_MDS_BootStart(this, MDS_BOOT_REPLAY_DONE)); | |
1281 | return; | |
1282 | } | |
1283 | } | |
1284 | ||
1285 | dout(1) << "making mds journal writeable" << dendl; | |
1286 | mdlog->get_journaler()->set_writeable(); | |
1287 | mdlog->get_journaler()->trim_tail(); | |
1288 | ||
1289 | if (g_conf->mds_wipe_sessions) { | |
1290 | dout(1) << "wiping out client sessions" << dendl; | |
1291 | sessionmap.wipe(); | |
1292 | sessionmap.save(new C_MDSInternalNoop); | |
1293 | } | |
1294 | if (g_conf->mds_wipe_ino_prealloc) { | |
1295 | dout(1) << "wiping out ino prealloc from sessions" << dendl; | |
1296 | sessionmap.wipe_ino_prealloc(); | |
1297 | sessionmap.save(new C_MDSInternalNoop); | |
1298 | } | |
1299 | if (g_conf->mds_skip_ino) { | |
1300 | inodeno_t i = g_conf->mds_skip_ino; | |
1301 | dout(1) << "skipping " << i << " inodes" << dendl; | |
1302 | inotable->skip_inos(i); | |
1303 | inotable->save(new C_MDSInternalNoop); | |
1304 | } | |
1305 | ||
1306 | if (mdsmap->get_num_in_mds() == 1 && | |
1307 | mdsmap->get_num_failed_mds() == 0) { // just me! | |
1308 | dout(2) << "i am alone, moving to state reconnect" << dendl; | |
1309 | request_state(MDSMap::STATE_RECONNECT); | |
1310 | } else { | |
1311 | dout(2) << "i am not alone, moving to state resolve" << dendl; | |
1312 | request_state(MDSMap::STATE_RESOLVE); | |
1313 | } | |
1314 | } | |
1315 | ||
1316 | void MDSRank::reopen_log() | |
1317 | { | |
1318 | dout(1) << "reopen_log" << dendl; | |
1319 | mdcache->rollback_uncommitted_fragments(); | |
1320 | } | |
1321 | ||
1322 | ||
1323 | void MDSRank::resolve_start() | |
1324 | { | |
1325 | dout(1) << "resolve_start" << dendl; | |
1326 | ||
1327 | reopen_log(); | |
1328 | ||
1329 | mdcache->resolve_start(new C_MDS_VoidFn(this, &MDSRank::resolve_done)); | |
1330 | finish_contexts(g_ceph_context, waiting_for_resolve); | |
1331 | } | |
1332 | void MDSRank::resolve_done() | |
1333 | { | |
1334 | dout(1) << "resolve_done" << dendl; | |
1335 | request_state(MDSMap::STATE_RECONNECT); | |
1336 | } | |
1337 | ||
1338 | void MDSRank::reconnect_start() | |
1339 | { | |
1340 | dout(1) << "reconnect_start" << dendl; | |
1341 | ||
1342 | if (last_state == MDSMap::STATE_REPLAY) { | |
1343 | reopen_log(); | |
1344 | } | |
1345 | ||
31f18b77 FG |
1346 | // Drop any blacklisted clients from the SessionMap before going |
1347 | // into reconnect, so that we don't wait for them. | |
1348 | objecter->enable_blacklist_events(); | |
1349 | std::set<entity_addr_t> blacklist; | |
1350 | epoch_t epoch = 0; | |
1351 | objecter->with_osdmap([this, &blacklist, &epoch](const OSDMap& o) { | |
1352 | o.get_blacklist(&blacklist); | |
1353 | epoch = o.get_epoch(); | |
1354 | }); | |
1355 | auto killed = server->apply_blacklist(blacklist); | |
1356 | dout(4) << "reconnect_start: killed " << killed << " blacklisted sessions (" | |
1357 | << blacklist.size() << " blacklist entries, " | |
1358 | << sessionmap.get_sessions().size() << ")" << dendl; | |
1359 | if (killed) { | |
1360 | set_osd_epoch_barrier(epoch); | |
1361 | } | |
1362 | ||
7c673cae FG |
1363 | server->reconnect_clients(new C_MDS_VoidFn(this, &MDSRank::reconnect_done)); |
1364 | finish_contexts(g_ceph_context, waiting_for_reconnect); | |
1365 | } | |
1366 | void MDSRank::reconnect_done() | |
1367 | { | |
1368 | dout(1) << "reconnect_done" << dendl; | |
1369 | request_state(MDSMap::STATE_REJOIN); // move to rejoin state | |
1370 | } | |
1371 | ||
1372 | void MDSRank::rejoin_joint_start() | |
1373 | { | |
1374 | dout(1) << "rejoin_joint_start" << dendl; | |
1375 | mdcache->rejoin_send_rejoins(); | |
1376 | } | |
1377 | void MDSRank::rejoin_start() | |
1378 | { | |
1379 | dout(1) << "rejoin_start" << dendl; | |
1380 | mdcache->rejoin_start(new C_MDS_VoidFn(this, &MDSRank::rejoin_done)); | |
1381 | } | |
1382 | void MDSRank::rejoin_done() | |
1383 | { | |
1384 | dout(1) << "rejoin_done" << dendl; | |
1385 | mdcache->show_subtrees(); | |
1386 | mdcache->show_cache(); | |
1387 | ||
1388 | // funny case: is our cache empty? no subtrees? | |
1389 | if (!mdcache->is_subtrees()) { | |
1390 | if (whoami == 0) { | |
1391 | // The root should always have a subtree! | |
1392 | clog->error() << "No subtrees found for root MDS rank!"; | |
1393 | damaged(); | |
1394 | assert(mdcache->is_subtrees()); | |
1395 | } else { | |
1396 | dout(1) << " empty cache, no subtrees, leaving cluster" << dendl; | |
1397 | request_state(MDSMap::STATE_STOPPED); | |
1398 | } | |
1399 | return; | |
1400 | } | |
1401 | ||
1402 | if (replay_queue.empty()) | |
1403 | request_state(MDSMap::STATE_ACTIVE); | |
1404 | else | |
1405 | request_state(MDSMap::STATE_CLIENTREPLAY); | |
1406 | } | |
1407 | ||
1408 | void MDSRank::clientreplay_start() | |
1409 | { | |
1410 | dout(1) << "clientreplay_start" << dendl; | |
1411 | finish_contexts(g_ceph_context, waiting_for_replay); // kick waiters | |
1412 | mdcache->start_files_to_recover(); | |
1413 | queue_one_replay(); | |
1414 | } | |
1415 | ||
1416 | bool MDSRank::queue_one_replay() | |
1417 | { | |
1418 | if (replay_queue.empty()) { | |
1419 | mdlog->wait_for_safe(new C_MDS_VoidFn(this, &MDSRank::clientreplay_done)); | |
1420 | return false; | |
1421 | } | |
1422 | queue_waiter(replay_queue.front()); | |
1423 | replay_queue.pop_front(); | |
1424 | return true; | |
1425 | } | |
1426 | ||
1427 | void MDSRank::clientreplay_done() | |
1428 | { | |
1429 | dout(1) << "clientreplay_done" << dendl; | |
1430 | request_state(MDSMap::STATE_ACTIVE); | |
1431 | } | |
1432 | ||
1433 | void MDSRank::active_start() | |
1434 | { | |
1435 | dout(1) << "active_start" << dendl; | |
1436 | ||
1437 | if (last_state == MDSMap::STATE_CREATING) { | |
1438 | mdcache->open_root(); | |
1439 | } | |
1440 | ||
1441 | mdcache->clean_open_file_lists(); | |
1442 | mdcache->export_remaining_imported_caps(); | |
1443 | finish_contexts(g_ceph_context, waiting_for_replay); // kick waiters | |
1444 | mdcache->start_files_to_recover(); | |
1445 | ||
1446 | mdcache->reissue_all_caps(); | |
1447 | mdcache->activate_stray_manager(); | |
1448 | ||
1449 | finish_contexts(g_ceph_context, waiting_for_active); // kick waiters | |
1450 | } | |
1451 | ||
1452 | void MDSRank::recovery_done(int oldstate) | |
1453 | { | |
1454 | dout(1) << "recovery_done -- successful recovery!" << dendl; | |
1455 | assert(is_clientreplay() || is_active()); | |
1456 | ||
1457 | // kick snaptable (resent AGREEs) | |
1458 | if (mdsmap->get_tableserver() == whoami) { | |
1459 | set<mds_rank_t> active; | |
1460 | mdsmap->get_clientreplay_or_active_or_stopping_mds_set(active); | |
1461 | snapserver->finish_recovery(active); | |
1462 | } | |
1463 | ||
1464 | if (oldstate == MDSMap::STATE_CREATING) | |
1465 | return; | |
1466 | ||
1467 | mdcache->start_recovered_truncates(); | |
1468 | mdcache->do_file_recover(); | |
1469 | ||
1470 | // tell connected clients | |
1471 | //bcast_mds_map(); // not anymore, they get this from the monitor | |
1472 | ||
1473 | mdcache->populate_mydir(); | |
1474 | } | |
1475 | ||
1476 | void MDSRank::creating_done() | |
1477 | { | |
1478 | dout(1)<< "creating_done" << dendl; | |
1479 | request_state(MDSMap::STATE_ACTIVE); | |
1480 | } | |
1481 | ||
1482 | void MDSRank::boot_create() | |
1483 | { | |
1484 | dout(3) << "boot_create" << dendl; | |
1485 | ||
1486 | MDSGatherBuilder fin(g_ceph_context, new C_MDS_VoidFn(this, &MDSRank::creating_done)); | |
1487 | ||
1488 | mdcache->init_layouts(); | |
1489 | ||
1490 | snapserver->set_rank(whoami); | |
1491 | inotable->set_rank(whoami); | |
1492 | sessionmap.set_rank(whoami); | |
1493 | ||
1494 | // start with a fresh journal | |
1495 | dout(10) << "boot_create creating fresh journal" << dendl; | |
1496 | mdlog->create(fin.new_sub()); | |
1497 | ||
1498 | // open new journal segment, but do not journal subtree map (yet) | |
1499 | mdlog->prepare_new_segment(); | |
1500 | ||
1501 | if (whoami == mdsmap->get_root()) { | |
1502 | dout(3) << "boot_create creating fresh hierarchy" << dendl; | |
1503 | mdcache->create_empty_hierarchy(fin.get()); | |
1504 | } | |
1505 | ||
1506 | dout(3) << "boot_create creating mydir hierarchy" << dendl; | |
1507 | mdcache->create_mydir_hierarchy(fin.get()); | |
1508 | ||
1509 | // fixme: fake out inotable (reset, pretend loaded) | |
1510 | dout(10) << "boot_create creating fresh inotable table" << dendl; | |
1511 | inotable->reset(); | |
1512 | inotable->save(fin.new_sub()); | |
1513 | ||
1514 | // write empty sessionmap | |
1515 | sessionmap.save(fin.new_sub()); | |
1516 | ||
1517 | // Create empty purge queue | |
1518 | purge_queue.create(new C_IO_Wrapper(this, fin.new_sub())); | |
1519 | ||
1520 | // initialize tables | |
1521 | if (mdsmap->get_tableserver() == whoami) { | |
1522 | dout(10) << "boot_create creating fresh snaptable" << dendl; | |
1523 | snapserver->reset(); | |
1524 | snapserver->save(fin.new_sub()); | |
1525 | } | |
1526 | ||
1527 | assert(g_conf->mds_kill_create_at != 1); | |
1528 | ||
1529 | // ok now journal it | |
1530 | mdlog->journal_segment_subtree_map(fin.new_sub()); | |
1531 | mdlog->flush(); | |
1532 | ||
31f18b77 FG |
1533 | // Usually we do this during reconnect, but creation skips that. |
1534 | objecter->enable_blacklist_events(); | |
1535 | ||
7c673cae FG |
1536 | fin.activate(); |
1537 | } | |
1538 | ||
1539 | void MDSRank::stopping_start() | |
1540 | { | |
1541 | dout(2) << "stopping_start" << dendl; | |
1542 | ||
1543 | if (mdsmap->get_num_in_mds() == 1 && !sessionmap.empty()) { | |
1544 | // we're the only mds up! | |
1545 | dout(0) << "we are the last MDS, and have mounted clients: we cannot flush our journal. suicide!" << dendl; | |
1546 | suicide(); | |
1547 | } | |
1548 | ||
1549 | mdcache->shutdown_start(); | |
1550 | } | |
1551 | ||
1552 | void MDSRank::stopping_done() | |
1553 | { | |
1554 | dout(2) << "stopping_done" << dendl; | |
1555 | ||
1556 | // tell monitor we shut down cleanly. | |
1557 | request_state(MDSMap::STATE_STOPPED); | |
1558 | } | |
1559 | ||
1560 | void MDSRankDispatcher::handle_mds_map( | |
1561 | MMDSMap *m, | |
1562 | MDSMap *oldmap) | |
1563 | { | |
1564 | // I am only to be passed MDSMaps in which I hold a rank | |
1565 | assert(whoami != MDS_RANK_NONE); | |
1566 | ||
1567 | MDSMap::DaemonState oldstate = state; | |
1568 | mds_gid_t mds_gid = mds_gid_t(monc->get_global_id()); | |
1569 | state = mdsmap->get_state_gid(mds_gid); | |
1570 | if (state != oldstate) { | |
1571 | last_state = oldstate; | |
1572 | incarnation = mdsmap->get_inc_gid(mds_gid); | |
1573 | } | |
1574 | ||
1575 | version_t epoch = m->get_epoch(); | |
1576 | ||
1577 | // note source's map version | |
1578 | if (m->get_source().is_mds() && | |
1579 | peer_mdsmap_epoch[mds_rank_t(m->get_source().num())] < epoch) { | |
1580 | dout(15) << " peer " << m->get_source() | |
1581 | << " has mdsmap epoch >= " << epoch | |
1582 | << dendl; | |
1583 | peer_mdsmap_epoch[mds_rank_t(m->get_source().num())] = epoch; | |
1584 | } | |
1585 | ||
1586 | // Validate state transitions while I hold a rank | |
1587 | if (!MDSMap::state_transition_valid(oldstate, state)) { | |
1588 | derr << "Invalid state transition " << ceph_mds_state_name(oldstate) | |
1589 | << "->" << ceph_mds_state_name(state) << dendl; | |
1590 | respawn(); | |
1591 | } | |
1592 | ||
1593 | if (oldstate != state) { | |
1594 | // update messenger. | |
1595 | if (state == MDSMap::STATE_STANDBY_REPLAY) { | |
1596 | dout(1) << "handle_mds_map i am now mds." << mds_gid << "." << incarnation | |
1597 | << " replaying mds." << whoami << "." << incarnation << dendl; | |
1598 | messenger->set_myname(entity_name_t::MDS(mds_gid)); | |
1599 | } else { | |
1600 | dout(1) << "handle_mds_map i am now mds." << whoami << "." << incarnation << dendl; | |
1601 | messenger->set_myname(entity_name_t::MDS(whoami)); | |
1602 | } | |
1603 | } | |
1604 | ||
1605 | // tell objecter my incarnation | |
1606 | if (objecter->get_client_incarnation() != incarnation) | |
1607 | objecter->set_client_incarnation(incarnation); | |
1608 | ||
1609 | // for debug | |
1610 | if (g_conf->mds_dump_cache_on_map) | |
1611 | mdcache->dump_cache(); | |
1612 | ||
1613 | // did it change? | |
1614 | if (oldstate != state) { | |
1615 | dout(1) << "handle_mds_map state change " | |
1616 | << ceph_mds_state_name(oldstate) << " --> " | |
1617 | << ceph_mds_state_name(state) << dendl; | |
1618 | beacon.set_want_state(mdsmap, state); | |
1619 | ||
1620 | if (oldstate == MDSMap::STATE_STANDBY_REPLAY) { | |
1621 | dout(10) << "Monitor activated us! Deactivating replay loop" << dendl; | |
1622 | assert (state == MDSMap::STATE_REPLAY); | |
1623 | } else { | |
1624 | // did i just recover? | |
1625 | if ((is_active() || is_clientreplay()) && | |
1626 | (oldstate == MDSMap::STATE_CREATING || | |
1627 | oldstate == MDSMap::STATE_REJOIN || | |
1628 | oldstate == MDSMap::STATE_RECONNECT)) | |
1629 | recovery_done(oldstate); | |
1630 | ||
1631 | if (is_active()) { | |
1632 | active_start(); | |
1633 | } else if (is_any_replay()) { | |
1634 | replay_start(); | |
1635 | } else if (is_resolve()) { | |
1636 | resolve_start(); | |
1637 | } else if (is_reconnect()) { | |
1638 | reconnect_start(); | |
1639 | } else if (is_rejoin()) { | |
1640 | rejoin_start(); | |
1641 | } else if (is_clientreplay()) { | |
1642 | clientreplay_start(); | |
1643 | } else if (is_creating()) { | |
1644 | boot_create(); | |
1645 | } else if (is_starting()) { | |
1646 | boot_start(); | |
1647 | } else if (is_stopping()) { | |
1648 | assert(oldstate == MDSMap::STATE_ACTIVE); | |
1649 | stopping_start(); | |
1650 | } | |
1651 | } | |
1652 | } | |
1653 | ||
1654 | // RESOLVE | |
1655 | // is someone else newly resolving? | |
1656 | if (is_resolve() || is_reconnect() || is_rejoin() || | |
1657 | is_clientreplay() || is_active() || is_stopping()) { | |
1658 | if (!oldmap->is_resolving() && mdsmap->is_resolving()) { | |
1659 | set<mds_rank_t> resolve; | |
1660 | mdsmap->get_mds_set(resolve, MDSMap::STATE_RESOLVE); | |
1661 | dout(10) << " resolve set is " << resolve << dendl; | |
1662 | calc_recovery_set(); | |
1663 | mdcache->send_resolves(); | |
1664 | } | |
1665 | } | |
1666 | ||
1667 | // REJOIN | |
1668 | // is everybody finally rejoining? | |
d2e6a577 | 1669 | if (is_starting() || is_rejoin() || is_clientreplay() || is_active() || is_stopping()) { |
7c673cae FG |
1670 | // did we start? |
1671 | if (!oldmap->is_rejoining() && mdsmap->is_rejoining()) | |
1672 | rejoin_joint_start(); | |
1673 | ||
1674 | // did we finish? | |
1675 | if (g_conf->mds_dump_cache_after_rejoin && | |
1676 | oldmap->is_rejoining() && !mdsmap->is_rejoining()) | |
1677 | mdcache->dump_cache(); // for DEBUG only | |
1678 | ||
d2e6a577 FG |
1679 | if (oldstate >= MDSMap::STATE_REJOIN || |
1680 | oldstate == MDSMap::STATE_STARTING) { | |
7c673cae FG |
1681 | // ACTIVE|CLIENTREPLAY|REJOIN => we can discover from them. |
1682 | set<mds_rank_t> olddis, dis; | |
1683 | oldmap->get_mds_set(olddis, MDSMap::STATE_ACTIVE); | |
1684 | oldmap->get_mds_set(olddis, MDSMap::STATE_CLIENTREPLAY); | |
1685 | oldmap->get_mds_set(olddis, MDSMap::STATE_REJOIN); | |
1686 | mdsmap->get_mds_set(dis, MDSMap::STATE_ACTIVE); | |
1687 | mdsmap->get_mds_set(dis, MDSMap::STATE_CLIENTREPLAY); | |
1688 | mdsmap->get_mds_set(dis, MDSMap::STATE_REJOIN); | |
1689 | for (set<mds_rank_t>::iterator p = dis.begin(); p != dis.end(); ++p) | |
1690 | if (*p != whoami && // not me | |
1691 | olddis.count(*p) == 0) { // newly so? | |
1692 | mdcache->kick_discovers(*p); | |
1693 | mdcache->kick_open_ino_peers(*p); | |
1694 | } | |
1695 | } | |
1696 | } | |
1697 | ||
1698 | cluster_degraded = mdsmap->is_degraded(); | |
1699 | if (oldmap->is_degraded() && !cluster_degraded && state >= MDSMap::STATE_ACTIVE) { | |
1700 | dout(1) << "cluster recovered." << dendl; | |
1701 | auto it = waiting_for_active_peer.find(MDS_RANK_NONE); | |
1702 | if (it != waiting_for_active_peer.end()) { | |
1703 | queue_waiters(it->second); | |
1704 | waiting_for_active_peer.erase(it); | |
1705 | } | |
1706 | } | |
1707 | ||
1708 | // did someone go active? | |
1709 | if (oldstate >= MDSMap::STATE_CLIENTREPLAY && | |
1710 | (is_clientreplay() || is_active() || is_stopping())) { | |
1711 | set<mds_rank_t> oldactive, active; | |
1712 | oldmap->get_mds_set(oldactive, MDSMap::STATE_ACTIVE); | |
1713 | oldmap->get_mds_set(oldactive, MDSMap::STATE_CLIENTREPLAY); | |
1714 | mdsmap->get_mds_set(active, MDSMap::STATE_ACTIVE); | |
1715 | mdsmap->get_mds_set(active, MDSMap::STATE_CLIENTREPLAY); | |
1716 | for (set<mds_rank_t>::iterator p = active.begin(); p != active.end(); ++p) | |
1717 | if (*p != whoami && // not me | |
1718 | oldactive.count(*p) == 0) // newly so? | |
1719 | handle_mds_recovery(*p); | |
1720 | } | |
1721 | ||
1722 | // did someone fail? | |
1723 | // new down? | |
1724 | { | |
1725 | set<mds_rank_t> olddown, down; | |
1726 | oldmap->get_down_mds_set(&olddown); | |
1727 | mdsmap->get_down_mds_set(&down); | |
1728 | for (set<mds_rank_t>::iterator p = down.begin(); p != down.end(); ++p) { | |
1729 | if (oldmap->have_inst(*p) && olddown.count(*p) == 0) { | |
1730 | messenger->mark_down(oldmap->get_inst(*p).addr); | |
1731 | handle_mds_failure(*p); | |
1732 | } | |
1733 | } | |
1734 | } | |
1735 | ||
1736 | // did someone fail? | |
1737 | // did their addr/inst change? | |
1738 | { | |
1739 | set<mds_rank_t> up; | |
1740 | mdsmap->get_up_mds_set(up); | |
1741 | for (set<mds_rank_t>::iterator p = up.begin(); p != up.end(); ++p) { | |
1742 | if (oldmap->have_inst(*p) && | |
1743 | oldmap->get_inst(*p) != mdsmap->get_inst(*p)) { | |
1744 | messenger->mark_down(oldmap->get_inst(*p).addr); | |
1745 | handle_mds_failure(*p); | |
1746 | } | |
1747 | } | |
1748 | } | |
1749 | ||
1750 | if (is_clientreplay() || is_active() || is_stopping()) { | |
1751 | // did anyone stop? | |
1752 | set<mds_rank_t> oldstopped, stopped; | |
1753 | oldmap->get_stopped_mds_set(oldstopped); | |
1754 | mdsmap->get_stopped_mds_set(stopped); | |
1755 | for (set<mds_rank_t>::iterator p = stopped.begin(); p != stopped.end(); ++p) | |
1756 | if (oldstopped.count(*p) == 0) // newly so? | |
1757 | mdcache->migrator->handle_mds_failure_or_stop(*p); | |
1758 | } | |
1759 | ||
1760 | { | |
1761 | map<epoch_t,list<MDSInternalContextBase*> >::iterator p = waiting_for_mdsmap.begin(); | |
1762 | while (p != waiting_for_mdsmap.end() && p->first <= mdsmap->get_epoch()) { | |
1763 | list<MDSInternalContextBase*> ls; | |
1764 | ls.swap(p->second); | |
1765 | waiting_for_mdsmap.erase(p++); | |
1766 | finish_contexts(g_ceph_context, ls); | |
1767 | } | |
1768 | } | |
1769 | ||
1770 | if (is_active()) { | |
1771 | // Before going active, set OSD epoch barrier to latest (so that | |
1772 | // we don't risk handing out caps to clients with old OSD maps that | |
1773 | // might not include barriers from the previous incarnation of this MDS) | |
1774 | set_osd_epoch_barrier(objecter->with_osdmap( | |
1775 | std::mem_fn(&OSDMap::get_epoch))); | |
1776 | } | |
1777 | ||
1778 | if (is_active()) { | |
1779 | bool found = false; | |
1780 | MDSMap::mds_info_t info = mdsmap->get_info(whoami); | |
1781 | ||
1782 | for (map<mds_gid_t,MDSMap::mds_info_t>::const_iterator p = mdsmap->get_mds_info().begin(); | |
1783 | p != mdsmap->get_mds_info().end(); | |
1784 | ++p) { | |
1785 | if (p->second.state == MDSMap::STATE_STANDBY_REPLAY && | |
1786 | (p->second.standby_for_rank == whoami ||(info.name.length() && p->second.standby_for_name == info.name))) { | |
1787 | found = true; | |
1788 | break; | |
1789 | } | |
1790 | if (found) | |
1791 | mdlog->set_write_iohint(0); | |
1792 | else | |
1793 | mdlog->set_write_iohint(CEPH_OSD_OP_FLAG_FADVISE_DONTNEED); | |
1794 | } | |
1795 | } | |
1796 | ||
1797 | if (oldmap->get_max_mds() != mdsmap->get_max_mds()) { | |
1798 | purge_queue.update_op_limit(*mdsmap); | |
1799 | } | |
1800 | } | |
1801 | ||
1802 | void MDSRank::handle_mds_recovery(mds_rank_t who) | |
1803 | { | |
1804 | dout(5) << "handle_mds_recovery mds." << who << dendl; | |
1805 | ||
1806 | mdcache->handle_mds_recovery(who); | |
1807 | ||
1808 | if (mdsmap->get_tableserver() == whoami) { | |
1809 | snapserver->handle_mds_recovery(who); | |
1810 | } | |
1811 | ||
1812 | queue_waiters(waiting_for_active_peer[who]); | |
1813 | waiting_for_active_peer.erase(who); | |
1814 | } | |
1815 | ||
1816 | void MDSRank::handle_mds_failure(mds_rank_t who) | |
1817 | { | |
1818 | if (who == whoami) { | |
1819 | dout(5) << "handle_mds_failure for myself; not doing anything" << dendl; | |
1820 | return; | |
1821 | } | |
1822 | dout(5) << "handle_mds_failure mds." << who << dendl; | |
1823 | ||
1824 | mdcache->handle_mds_failure(who); | |
1825 | ||
1826 | snapclient->handle_mds_failure(who); | |
1827 | } | |
1828 | ||
1829 | bool MDSRankDispatcher::handle_asok_command( | |
1830 | std::string command, cmdmap_t& cmdmap, Formatter *f, | |
1831 | std::ostream& ss) | |
1832 | { | |
1833 | if (command == "dump_ops_in_flight" || | |
1834 | command == "ops") { | |
1835 | if (!op_tracker.dump_ops_in_flight(f)) { | |
1836 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
1837 | please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
1838 | } | |
1839 | } else if (command == "dump_blocked_ops") { | |
1840 | if (!op_tracker.dump_ops_in_flight(f, true)) { | |
1841 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
1842 | Please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
1843 | } | |
1844 | } else if (command == "dump_historic_ops") { | |
1845 | if (!op_tracker.dump_historic_ops(f)) { | |
1846 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
1847 | please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
1848 | } | |
1849 | } else if (command == "dump_historic_ops_by_duration") { | |
1850 | if (!op_tracker.dump_historic_ops(f, true)) { | |
1851 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
1852 | please enable \"osd_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
1853 | } | |
1854 | } else if (command == "osdmap barrier") { | |
1855 | int64_t target_epoch = 0; | |
1856 | bool got_val = cmd_getval(g_ceph_context, cmdmap, "target_epoch", target_epoch); | |
1857 | ||
1858 | if (!got_val) { | |
1859 | ss << "no target epoch given"; | |
1860 | return true; | |
1861 | } | |
1862 | ||
1863 | mds_lock.Lock(); | |
1864 | set_osd_epoch_barrier(target_epoch); | |
1865 | mds_lock.Unlock(); | |
1866 | ||
1867 | C_SaferCond cond; | |
1868 | bool already_got = objecter->wait_for_map(target_epoch, &cond); | |
1869 | if (!already_got) { | |
1870 | dout(4) << __func__ << ": waiting for OSD epoch " << target_epoch << dendl; | |
1871 | cond.wait(); | |
1872 | } | |
1873 | } else if (command == "session ls") { | |
1874 | Mutex::Locker l(mds_lock); | |
1875 | ||
1876 | heartbeat_reset(); | |
1877 | ||
1878 | dump_sessions(SessionFilter(), f); | |
1879 | } else if (command == "session evict") { | |
1880 | std::string client_id; | |
1881 | const bool got_arg = cmd_getval(g_ceph_context, cmdmap, "client_id", client_id); | |
1882 | if(!got_arg) { | |
1883 | ss << "Invalid client_id specified"; | |
1884 | return true; | |
1885 | } | |
1886 | ||
1887 | mds_lock.Lock(); | |
31f18b77 FG |
1888 | std::stringstream dss; |
1889 | bool evicted = evict_client(strtol(client_id.c_str(), 0, 10), true, | |
1890 | g_conf->mds_session_blacklist_on_evict, dss); | |
1891 | if (!evicted) { | |
7c673cae FG |
1892 | dout(15) << dss.str() << dendl; |
1893 | ss << dss.str(); | |
1894 | } | |
1895 | mds_lock.Unlock(); | |
1896 | } else if (command == "scrub_path") { | |
1897 | string path; | |
1898 | vector<string> scrubop_vec; | |
1899 | cmd_getval(g_ceph_context, cmdmap, "scrubops", scrubop_vec); | |
1900 | cmd_getval(g_ceph_context, cmdmap, "path", path); | |
1901 | command_scrub_path(f, path, scrubop_vec); | |
1902 | } else if (command == "tag path") { | |
1903 | string path; | |
1904 | cmd_getval(g_ceph_context, cmdmap, "path", path); | |
1905 | string tag; | |
1906 | cmd_getval(g_ceph_context, cmdmap, "tag", tag); | |
1907 | command_tag_path(f, path, tag); | |
1908 | } else if (command == "flush_path") { | |
1909 | string path; | |
1910 | cmd_getval(g_ceph_context, cmdmap, "path", path); | |
1911 | command_flush_path(f, path); | |
1912 | } else if (command == "flush journal") { | |
1913 | command_flush_journal(f); | |
1914 | } else if (command == "get subtrees") { | |
1915 | command_get_subtrees(f); | |
1916 | } else if (command == "export dir") { | |
1917 | string path; | |
1918 | if(!cmd_getval(g_ceph_context, cmdmap, "path", path)) { | |
1919 | ss << "malformed path"; | |
1920 | return true; | |
1921 | } | |
1922 | int64_t rank; | |
1923 | if(!cmd_getval(g_ceph_context, cmdmap, "rank", rank)) { | |
1924 | ss << "malformed rank"; | |
1925 | return true; | |
1926 | } | |
1927 | command_export_dir(f, path, (mds_rank_t)rank); | |
1928 | } else if (command == "dump cache") { | |
1929 | Mutex::Locker l(mds_lock); | |
1930 | string path; | |
31f18b77 | 1931 | int r; |
7c673cae | 1932 | if(!cmd_getval(g_ceph_context, cmdmap, "path", path)) { |
31f18b77 | 1933 | r = mdcache->dump_cache(f); |
7c673cae | 1934 | } else { |
31f18b77 FG |
1935 | r = mdcache->dump_cache(path); |
1936 | } | |
1937 | ||
1938 | if (r != 0) { | |
1939 | ss << "Failed to dump cache: " << cpp_strerror(r); | |
7c673cae FG |
1940 | } |
1941 | } else if (command == "dump tree") { | |
1942 | string root; | |
1943 | int64_t depth; | |
1944 | cmd_getval(g_ceph_context, cmdmap, "root", root); | |
1945 | if (!cmd_getval(g_ceph_context, cmdmap, "depth", depth)) | |
1946 | depth = -1; | |
1947 | { | |
1948 | Mutex::Locker l(mds_lock); | |
31f18b77 FG |
1949 | int r = mdcache->dump_cache(root, depth, f); |
1950 | if (r != 0) { | |
1951 | ss << "Failed to dump tree: " << cpp_strerror(r); | |
1952 | } | |
7c673cae FG |
1953 | } |
1954 | } else if (command == "force_readonly") { | |
1955 | Mutex::Locker l(mds_lock); | |
1956 | mdcache->force_readonly(); | |
1957 | } else if (command == "dirfrag split") { | |
1958 | command_dirfrag_split(cmdmap, ss); | |
1959 | } else if (command == "dirfrag merge") { | |
1960 | command_dirfrag_merge(cmdmap, ss); | |
1961 | } else if (command == "dirfrag ls") { | |
1962 | command_dirfrag_ls(cmdmap, ss, f); | |
1963 | } else { | |
1964 | return false; | |
1965 | } | |
1966 | ||
1967 | return true; | |
1968 | } | |
1969 | ||
1970 | class C_MDS_Send_Command_Reply : public MDSInternalContext | |
1971 | { | |
1972 | protected: | |
1973 | MCommand *m; | |
1974 | public: | |
1975 | C_MDS_Send_Command_Reply(MDSRank *_mds, MCommand *_m) : | |
1976 | MDSInternalContext(_mds), m(_m) { m->get(); } | |
1977 | void send (int r, const std::string& out_str) { | |
1978 | bufferlist bl; | |
1979 | MDSDaemon::send_command_reply(m, mds, r, bl, out_str); | |
1980 | m->put(); | |
1981 | } | |
1982 | void finish (int r) override { | |
1983 | send(r, ""); | |
1984 | } | |
1985 | }; | |
1986 | ||
1987 | /** | |
1988 | * This function drops the mds_lock, so don't do anything with | |
1989 | * MDSRank after calling it (we could have gone into shutdown): just | |
1990 | * send your result back to the calling client and finish. | |
1991 | */ | |
31f18b77 | 1992 | void MDSRankDispatcher::evict_clients(const SessionFilter &filter, MCommand *m) |
7c673cae FG |
1993 | { |
1994 | C_MDS_Send_Command_Reply *reply = new C_MDS_Send_Command_Reply(this, m); | |
1995 | ||
1996 | if (is_any_replay()) { | |
1997 | reply->send(-EAGAIN, "MDS is replaying log"); | |
1998 | delete reply; | |
1999 | return; | |
2000 | } | |
2001 | ||
2002 | std::list<Session*> victims; | |
2003 | const auto sessions = sessionmap.get_sessions(); | |
2004 | for (const auto p : sessions) { | |
2005 | if (!p.first.is_client()) { | |
2006 | continue; | |
2007 | } | |
2008 | ||
2009 | Session *s = p.second; | |
2010 | ||
2011 | if (filter.match(*s, std::bind(&Server::waiting_for_reconnect, server, std::placeholders::_1))) { | |
2012 | victims.push_back(s); | |
2013 | } | |
2014 | } | |
2015 | ||
2016 | dout(20) << __func__ << " matched " << victims.size() << " sessions" << dendl; | |
2017 | ||
2018 | if (victims.empty()) { | |
2019 | reply->send(0, ""); | |
2020 | delete reply; | |
2021 | return; | |
2022 | } | |
2023 | ||
2024 | C_GatherBuilder gather(g_ceph_context, reply); | |
2025 | for (const auto s : victims) { | |
31f18b77 FG |
2026 | std::stringstream ss; |
2027 | evict_client(s->info.inst.name.num(), false, | |
2028 | g_conf->mds_session_blacklist_on_evict, ss, gather.new_sub()); | |
7c673cae FG |
2029 | } |
2030 | gather.activate(); | |
2031 | } | |
2032 | ||
2033 | void MDSRankDispatcher::dump_sessions(const SessionFilter &filter, Formatter *f) const | |
2034 | { | |
2035 | // Dump sessions, decorated with recovery/replay status | |
2036 | f->open_array_section("sessions"); | |
2037 | const ceph::unordered_map<entity_name_t, Session*> session_map = sessionmap.get_sessions(); | |
2038 | for (ceph::unordered_map<entity_name_t,Session*>::const_iterator p = session_map.begin(); | |
2039 | p != session_map.end(); | |
2040 | ++p) { | |
2041 | if (!p->first.is_client()) { | |
2042 | continue; | |
2043 | } | |
2044 | ||
2045 | Session *s = p->second; | |
2046 | ||
2047 | if (!filter.match(*s, std::bind(&Server::waiting_for_reconnect, server, std::placeholders::_1))) { | |
2048 | continue; | |
2049 | } | |
2050 | ||
2051 | f->open_object_section("session"); | |
2052 | f->dump_int("id", p->first.num()); | |
2053 | ||
2054 | f->dump_int("num_leases", s->leases.size()); | |
2055 | f->dump_int("num_caps", s->caps.size()); | |
2056 | ||
2057 | f->dump_string("state", s->get_state_name()); | |
2058 | f->dump_int("replay_requests", is_clientreplay() ? s->get_request_count() : 0); | |
2059 | f->dump_unsigned("completed_requests", s->get_num_completed_requests()); | |
2060 | f->dump_bool("reconnecting", server->waiting_for_reconnect(p->first.num())); | |
2061 | f->dump_stream("inst") << s->info.inst; | |
2062 | f->open_object_section("client_metadata"); | |
2063 | for (map<string, string>::const_iterator i = s->info.client_metadata.begin(); | |
2064 | i != s->info.client_metadata.end(); ++i) { | |
2065 | f->dump_string(i->first.c_str(), i->second); | |
2066 | } | |
2067 | f->close_section(); // client_metadata | |
2068 | f->close_section(); //session | |
2069 | } | |
2070 | f->close_section(); //sessions | |
2071 | } | |
2072 | ||
2073 | void MDSRank::command_scrub_path(Formatter *f, const string& path, vector<string>& scrubop_vec) | |
2074 | { | |
2075 | bool force = false; | |
2076 | bool recursive = false; | |
2077 | bool repair = false; | |
2078 | for (vector<string>::iterator i = scrubop_vec.begin() ; i != scrubop_vec.end(); ++i) { | |
2079 | if (*i == "force") | |
2080 | force = true; | |
2081 | else if (*i == "recursive") | |
2082 | recursive = true; | |
2083 | else if (*i == "repair") | |
2084 | repair = true; | |
2085 | } | |
2086 | C_SaferCond scond; | |
2087 | { | |
2088 | Mutex::Locker l(mds_lock); | |
2089 | mdcache->enqueue_scrub(path, "", force, recursive, repair, f, &scond); | |
2090 | } | |
2091 | scond.wait(); | |
2092 | // scrub_dentry() finishers will dump the data for us; we're done! | |
2093 | } | |
2094 | ||
2095 | void MDSRank::command_tag_path(Formatter *f, | |
2096 | const string& path, const std::string &tag) | |
2097 | { | |
2098 | C_SaferCond scond; | |
2099 | { | |
2100 | Mutex::Locker l(mds_lock); | |
2101 | mdcache->enqueue_scrub(path, tag, true, true, false, f, &scond); | |
2102 | } | |
2103 | scond.wait(); | |
2104 | } | |
2105 | ||
2106 | void MDSRank::command_flush_path(Formatter *f, const string& path) | |
2107 | { | |
2108 | C_SaferCond scond; | |
2109 | { | |
2110 | Mutex::Locker l(mds_lock); | |
2111 | mdcache->flush_dentry(path, &scond); | |
2112 | } | |
2113 | int r = scond.wait(); | |
2114 | f->open_object_section("results"); | |
2115 | f->dump_int("return_code", r); | |
2116 | f->close_section(); // results | |
2117 | } | |
2118 | ||
2119 | /** | |
2120 | * Wrapper around _command_flush_journal that | |
2121 | * handles serialization of result | |
2122 | */ | |
2123 | void MDSRank::command_flush_journal(Formatter *f) | |
2124 | { | |
2125 | assert(f != NULL); | |
2126 | ||
2127 | std::stringstream ss; | |
2128 | const int r = _command_flush_journal(&ss); | |
2129 | f->open_object_section("result"); | |
2130 | f->dump_string("message", ss.str()); | |
2131 | f->dump_int("return_code", r); | |
2132 | f->close_section(); | |
2133 | } | |
2134 | ||
2135 | /** | |
2136 | * Implementation of "flush journal" asok command. | |
2137 | * | |
2138 | * @param ss | |
2139 | * Optionally populate with a human readable string describing the | |
2140 | * reason for any unexpected return status. | |
2141 | */ | |
2142 | int MDSRank::_command_flush_journal(std::stringstream *ss) | |
2143 | { | |
2144 | assert(ss != NULL); | |
2145 | ||
2146 | Mutex::Locker l(mds_lock); | |
2147 | ||
2148 | if (mdcache->is_readonly()) { | |
2149 | dout(5) << __func__ << ": read-only FS" << dendl; | |
2150 | return -EROFS; | |
2151 | } | |
2152 | ||
2153 | if (!is_active()) { | |
2154 | dout(5) << __func__ << ": MDS not active, no-op" << dendl; | |
2155 | return 0; | |
2156 | } | |
2157 | ||
2158 | // I need to seal off the current segment, and then mark all previous segments | |
2159 | // for expiry | |
2160 | mdlog->start_new_segment(); | |
2161 | int r = 0; | |
2162 | ||
2163 | // Flush initially so that all the segments older than our new one | |
2164 | // will be elegible for expiry | |
2165 | { | |
2166 | C_SaferCond mdlog_flushed; | |
2167 | mdlog->flush(); | |
2168 | mdlog->wait_for_safe(new MDSInternalContextWrapper(this, &mdlog_flushed)); | |
2169 | mds_lock.Unlock(); | |
2170 | r = mdlog_flushed.wait(); | |
2171 | mds_lock.Lock(); | |
2172 | if (r != 0) { | |
2173 | *ss << "Error " << r << " (" << cpp_strerror(r) << ") while flushing journal"; | |
2174 | return r; | |
2175 | } | |
2176 | } | |
2177 | ||
2178 | // Because we may not be the last wait_for_safe context on MDLog, and | |
2179 | // subsequent contexts might wake up in the middle of our later trim_all | |
2180 | // and interfere with expiry (by e.g. marking dirs/dentries dirty | |
2181 | // on previous log segments), we run a second wait_for_safe here. | |
2182 | // See #10368 | |
2183 | { | |
2184 | C_SaferCond mdlog_cleared; | |
2185 | mdlog->wait_for_safe(new MDSInternalContextWrapper(this, &mdlog_cleared)); | |
2186 | mds_lock.Unlock(); | |
2187 | r = mdlog_cleared.wait(); | |
2188 | mds_lock.Lock(); | |
2189 | if (r != 0) { | |
2190 | *ss << "Error " << r << " (" << cpp_strerror(r) << ") while flushing journal"; | |
2191 | return r; | |
2192 | } | |
2193 | } | |
2194 | ||
2195 | // Put all the old log segments into expiring or expired state | |
2196 | dout(5) << __func__ << ": beginning segment expiry" << dendl; | |
2197 | r = mdlog->trim_all(); | |
2198 | if (r != 0) { | |
2199 | *ss << "Error " << r << " (" << cpp_strerror(r) << ") while trimming log"; | |
2200 | return r; | |
2201 | } | |
2202 | ||
2203 | // Attach contexts to wait for all expiring segments to expire | |
2204 | MDSGatherBuilder expiry_gather(g_ceph_context); | |
2205 | ||
2206 | const std::set<LogSegment*> &expiring_segments = mdlog->get_expiring_segments(); | |
2207 | for (std::set<LogSegment*>::const_iterator i = expiring_segments.begin(); | |
2208 | i != expiring_segments.end(); ++i) { | |
2209 | (*i)->wait_for_expiry(expiry_gather.new_sub()); | |
2210 | } | |
2211 | dout(5) << __func__ << ": waiting for " << expiry_gather.num_subs_created() | |
2212 | << " segments to expire" << dendl; | |
2213 | ||
2214 | if (expiry_gather.has_subs()) { | |
2215 | C_SaferCond cond; | |
2216 | expiry_gather.set_finisher(new MDSInternalContextWrapper(this, &cond)); | |
2217 | expiry_gather.activate(); | |
2218 | ||
2219 | // Drop mds_lock to allow progress until expiry is complete | |
2220 | mds_lock.Unlock(); | |
2221 | int r = cond.wait(); | |
2222 | mds_lock.Lock(); | |
2223 | ||
2224 | assert(r == 0); // MDLog is not allowed to raise errors via wait_for_expiry | |
2225 | } | |
2226 | ||
2227 | dout(5) << __func__ << ": expiry complete, expire_pos/trim_pos is now " << std::hex << | |
2228 | mdlog->get_journaler()->get_expire_pos() << "/" << | |
2229 | mdlog->get_journaler()->get_trimmed_pos() << dendl; | |
2230 | ||
2231 | // Now everyone I'm interested in is expired | |
2232 | mdlog->trim_expired_segments(); | |
2233 | ||
2234 | dout(5) << __func__ << ": trim complete, expire_pos/trim_pos is now " << std::hex << | |
2235 | mdlog->get_journaler()->get_expire_pos() << "/" << | |
2236 | mdlog->get_journaler()->get_trimmed_pos() << dendl; | |
2237 | ||
2238 | // Flush the journal header so that readers will start from after the flushed region | |
2239 | C_SaferCond wrote_head; | |
2240 | mdlog->get_journaler()->write_head(&wrote_head); | |
2241 | mds_lock.Unlock(); // Drop lock to allow messenger dispatch progress | |
2242 | r = wrote_head.wait(); | |
2243 | mds_lock.Lock(); | |
2244 | if (r != 0) { | |
2245 | *ss << "Error " << r << " (" << cpp_strerror(r) << ") while writing header"; | |
2246 | return r; | |
2247 | } | |
2248 | ||
2249 | dout(5) << __func__ << ": write_head complete, all done!" << dendl; | |
2250 | ||
2251 | return 0; | |
2252 | } | |
2253 | ||
2254 | ||
2255 | void MDSRank::command_get_subtrees(Formatter *f) | |
2256 | { | |
2257 | assert(f != NULL); | |
2258 | Mutex::Locker l(mds_lock); | |
2259 | ||
2260 | std::list<CDir*> subtrees; | |
2261 | mdcache->list_subtrees(subtrees); | |
2262 | ||
2263 | f->open_array_section("subtrees"); | |
2264 | for (std::list<CDir*>::iterator i = subtrees.begin(); i != subtrees.end(); ++i) { | |
2265 | const CDir *dir = *i; | |
2266 | ||
2267 | f->open_object_section("subtree"); | |
2268 | { | |
2269 | f->dump_bool("is_auth", dir->is_auth()); | |
2270 | f->dump_int("auth_first", dir->get_dir_auth().first); | |
2271 | f->dump_int("auth_second", dir->get_dir_auth().second); | |
d2e6a577 | 2272 | f->dump_int("export_pin", dir->inode->get_export_pin()); |
7c673cae FG |
2273 | f->open_object_section("dir"); |
2274 | dir->dump(f); | |
2275 | f->close_section(); | |
2276 | } | |
2277 | f->close_section(); | |
2278 | } | |
2279 | f->close_section(); | |
2280 | } | |
2281 | ||
2282 | ||
2283 | void MDSRank::command_export_dir(Formatter *f, | |
2284 | const std::string &path, | |
2285 | mds_rank_t target) | |
2286 | { | |
2287 | int r = _command_export_dir(path, target); | |
2288 | f->open_object_section("results"); | |
2289 | f->dump_int("return_code", r); | |
2290 | f->close_section(); // results | |
2291 | } | |
2292 | ||
2293 | int MDSRank::_command_export_dir( | |
2294 | const std::string &path, | |
2295 | mds_rank_t target) | |
2296 | { | |
2297 | Mutex::Locker l(mds_lock); | |
2298 | filepath fp(path.c_str()); | |
2299 | ||
2300 | if (target == whoami || !mdsmap->is_up(target) || !mdsmap->is_in(target)) { | |
2301 | derr << "bad MDS target " << target << dendl; | |
2302 | return -ENOENT; | |
2303 | } | |
2304 | ||
2305 | CInode *in = mdcache->cache_traverse(fp); | |
2306 | if (!in) { | |
2307 | derr << "Bath path '" << path << "'" << dendl; | |
2308 | return -ENOENT; | |
2309 | } | |
2310 | CDir *dir = in->get_dirfrag(frag_t()); | |
2311 | if (!dir || !(dir->is_auth())) { | |
2312 | derr << "bad export_dir path dirfrag frag_t() or dir not auth" << dendl; | |
2313 | return -EINVAL; | |
2314 | } | |
2315 | ||
2316 | mdcache->migrator->export_dir(dir, target); | |
2317 | return 0; | |
2318 | } | |
2319 | ||
2320 | CDir *MDSRank::_command_dirfrag_get( | |
2321 | const cmdmap_t &cmdmap, | |
2322 | std::ostream &ss) | |
2323 | { | |
2324 | std::string path; | |
2325 | bool got = cmd_getval(g_ceph_context, cmdmap, "path", path); | |
2326 | if (!got) { | |
2327 | ss << "missing path argument"; | |
2328 | return NULL; | |
2329 | } | |
2330 | ||
2331 | std::string frag_str; | |
2332 | if (!cmd_getval(g_ceph_context, cmdmap, "frag", frag_str)) { | |
2333 | ss << "missing frag argument"; | |
2334 | return NULL; | |
2335 | } | |
2336 | ||
2337 | CInode *in = mdcache->cache_traverse(filepath(path.c_str())); | |
2338 | if (!in) { | |
2339 | // TODO really we should load something in if it's not in cache, | |
2340 | // but the infrastructure is harder, and we might still be unable | |
2341 | // to act on it if someone else is auth. | |
2342 | ss << "directory '" << path << "' inode not in cache"; | |
2343 | return NULL; | |
2344 | } | |
2345 | ||
2346 | frag_t fg; | |
2347 | ||
2348 | if (!fg.parse(frag_str.c_str())) { | |
2349 | ss << "frag " << frag_str << " failed to parse"; | |
2350 | return NULL; | |
2351 | } | |
2352 | ||
2353 | CDir *dir = in->get_dirfrag(fg); | |
2354 | if (!dir) { | |
2355 | ss << "frag 0x" << std::hex << in->ino() << "/" << fg << " not in cache (" | |
2356 | "use `dirfrag ls` to see if it should exist)"; | |
2357 | return NULL; | |
2358 | } | |
2359 | ||
2360 | if (!dir->is_auth()) { | |
2361 | ss << "frag " << dir->dirfrag() << " not auth (auth = " | |
2362 | << dir->authority() << ")"; | |
2363 | return NULL; | |
2364 | } | |
2365 | ||
2366 | return dir; | |
2367 | } | |
2368 | ||
2369 | bool MDSRank::command_dirfrag_split( | |
2370 | cmdmap_t cmdmap, | |
2371 | std::ostream &ss) | |
2372 | { | |
2373 | Mutex::Locker l(mds_lock); | |
2374 | if (!mdsmap->allows_dirfrags()) { | |
2375 | ss << "dirfrags are disallowed by the mds map!"; | |
2376 | return false; | |
2377 | } | |
2378 | ||
2379 | int64_t by = 0; | |
2380 | if (!cmd_getval(g_ceph_context, cmdmap, "bits", by)) { | |
2381 | ss << "missing bits argument"; | |
2382 | return false; | |
2383 | } | |
2384 | ||
2385 | if (by <= 0) { | |
2386 | ss << "must split by >0 bits"; | |
2387 | return false; | |
2388 | } | |
2389 | ||
2390 | CDir *dir = _command_dirfrag_get(cmdmap, ss); | |
2391 | if (!dir) { | |
2392 | return false; | |
2393 | } | |
2394 | ||
2395 | mdcache->split_dir(dir, by); | |
2396 | ||
2397 | return true; | |
2398 | } | |
2399 | ||
2400 | bool MDSRank::command_dirfrag_merge( | |
2401 | cmdmap_t cmdmap, | |
2402 | std::ostream &ss) | |
2403 | { | |
2404 | Mutex::Locker l(mds_lock); | |
2405 | std::string path; | |
2406 | bool got = cmd_getval(g_ceph_context, cmdmap, "path", path); | |
2407 | if (!got) { | |
2408 | ss << "missing path argument"; | |
2409 | return false; | |
2410 | } | |
2411 | ||
2412 | std::string frag_str; | |
2413 | if (!cmd_getval(g_ceph_context, cmdmap, "frag", frag_str)) { | |
2414 | ss << "missing frag argument"; | |
2415 | return false; | |
2416 | } | |
2417 | ||
2418 | CInode *in = mdcache->cache_traverse(filepath(path.c_str())); | |
2419 | if (!in) { | |
2420 | ss << "directory '" << path << "' inode not in cache"; | |
2421 | return false; | |
2422 | } | |
2423 | ||
2424 | frag_t fg; | |
2425 | if (!fg.parse(frag_str.c_str())) { | |
2426 | ss << "frag " << frag_str << " failed to parse"; | |
2427 | return false; | |
2428 | } | |
2429 | ||
2430 | mdcache->merge_dir(in, fg); | |
2431 | ||
2432 | return true; | |
2433 | } | |
2434 | ||
2435 | bool MDSRank::command_dirfrag_ls( | |
2436 | cmdmap_t cmdmap, | |
2437 | std::ostream &ss, | |
2438 | Formatter *f) | |
2439 | { | |
2440 | Mutex::Locker l(mds_lock); | |
2441 | std::string path; | |
2442 | bool got = cmd_getval(g_ceph_context, cmdmap, "path", path); | |
2443 | if (!got) { | |
2444 | ss << "missing path argument"; | |
2445 | return false; | |
2446 | } | |
2447 | ||
2448 | CInode *in = mdcache->cache_traverse(filepath(path.c_str())); | |
2449 | if (!in) { | |
2450 | ss << "directory inode not in cache"; | |
2451 | return false; | |
2452 | } | |
2453 | ||
2454 | f->open_array_section("frags"); | |
2455 | std::list<frag_t> frags; | |
2456 | // NB using get_leaves_under instead of get_dirfrags to give | |
2457 | // you the list of what dirfrags may exist, not which are in cache | |
2458 | in->dirfragtree.get_leaves_under(frag_t(), frags); | |
2459 | for (std::list<frag_t>::iterator i = frags.begin(); | |
2460 | i != frags.end(); ++i) { | |
2461 | f->open_object_section("frag"); | |
2462 | f->dump_int("value", i->value()); | |
2463 | f->dump_int("bits", i->bits()); | |
2464 | std::ostringstream frag_str; | |
2465 | frag_str << std::hex << i->value() << "/" << std::dec << i->bits(); | |
2466 | f->dump_string("str", frag_str.str()); | |
2467 | f->close_section(); | |
2468 | } | |
2469 | f->close_section(); | |
2470 | ||
2471 | return true; | |
2472 | } | |
2473 | ||
2474 | void MDSRank::dump_status(Formatter *f) const | |
2475 | { | |
2476 | if (state == MDSMap::STATE_REPLAY || | |
2477 | state == MDSMap::STATE_STANDBY_REPLAY) { | |
2478 | mdlog->dump_replay_status(f); | |
2479 | } else if (state == MDSMap::STATE_RESOLVE) { | |
2480 | mdcache->dump_resolve_status(f); | |
2481 | } else if (state == MDSMap::STATE_RECONNECT) { | |
2482 | server->dump_reconnect_status(f); | |
2483 | } else if (state == MDSMap::STATE_REJOIN) { | |
2484 | mdcache->dump_rejoin_status(f); | |
2485 | } else if (state == MDSMap::STATE_CLIENTREPLAY) { | |
2486 | dump_clientreplay_status(f); | |
2487 | } | |
2488 | } | |
2489 | ||
2490 | void MDSRank::dump_clientreplay_status(Formatter *f) const | |
2491 | { | |
2492 | f->open_object_section("clientreplay_status"); | |
2493 | f->dump_unsigned("clientreplay_queue", replay_queue.size()); | |
2494 | f->dump_unsigned("active_replay", mdcache->get_num_client_requests()); | |
2495 | f->close_section(); | |
2496 | } | |
2497 | ||
2498 | void MDSRankDispatcher::update_log_config() | |
2499 | { | |
2500 | map<string,string> log_to_monitors; | |
2501 | map<string,string> log_to_syslog; | |
2502 | map<string,string> log_channel; | |
2503 | map<string,string> log_prio; | |
2504 | map<string,string> log_to_graylog; | |
2505 | map<string,string> log_to_graylog_host; | |
2506 | map<string,string> log_to_graylog_port; | |
2507 | uuid_d fsid; | |
2508 | string host; | |
2509 | ||
2510 | if (parse_log_client_options(g_ceph_context, log_to_monitors, log_to_syslog, | |
2511 | log_channel, log_prio, log_to_graylog, | |
2512 | log_to_graylog_host, log_to_graylog_port, | |
2513 | fsid, host) == 0) | |
2514 | clog->update_config(log_to_monitors, log_to_syslog, | |
2515 | log_channel, log_prio, log_to_graylog, | |
2516 | log_to_graylog_host, log_to_graylog_port, | |
2517 | fsid, host); | |
2518 | dout(10) << __func__ << " log_to_monitors " << log_to_monitors << dendl; | |
2519 | } | |
2520 | ||
2521 | void MDSRank::create_logger() | |
2522 | { | |
2523 | dout(10) << "create_logger" << dendl; | |
2524 | { | |
2525 | PerfCountersBuilder mds_plb(g_ceph_context, "mds", l_mds_first, l_mds_last); | |
2526 | ||
2527 | mds_plb.add_u64_counter( | |
2528 | l_mds_request, "request", "Requests", "req", | |
2529 | PerfCountersBuilder::PRIO_CRITICAL); | |
2530 | mds_plb.add_u64_counter(l_mds_reply, "reply", "Replies"); | |
2531 | mds_plb.add_time_avg( | |
2532 | l_mds_reply_latency, "reply_latency", "Reply latency", "rlat", | |
2533 | PerfCountersBuilder::PRIO_CRITICAL); | |
2534 | mds_plb.add_u64_counter( | |
2535 | l_mds_forward, "forward", "Forwarding request", "fwd", | |
2536 | PerfCountersBuilder::PRIO_INTERESTING); | |
2537 | mds_plb.add_u64_counter(l_mds_dir_fetch, "dir_fetch", "Directory fetch"); | |
2538 | mds_plb.add_u64_counter(l_mds_dir_commit, "dir_commit", "Directory commit"); | |
2539 | mds_plb.add_u64_counter(l_mds_dir_split, "dir_split", "Directory split"); | |
2540 | mds_plb.add_u64_counter(l_mds_dir_merge, "dir_merge", "Directory merge"); | |
2541 | ||
2542 | mds_plb.add_u64(l_mds_inode_max, "inode_max", "Max inodes, cache size"); | |
2543 | mds_plb.add_u64(l_mds_inodes, "inodes", "Inodes", "inos", | |
2544 | PerfCountersBuilder::PRIO_CRITICAL); | |
2545 | mds_plb.add_u64(l_mds_inodes_top, "inodes_top", "Inodes on top"); | |
2546 | mds_plb.add_u64(l_mds_inodes_bottom, "inodes_bottom", "Inodes on bottom"); | |
2547 | mds_plb.add_u64( | |
2548 | l_mds_inodes_pin_tail, "inodes_pin_tail", "Inodes on pin tail"); | |
2549 | mds_plb.add_u64(l_mds_inodes_pinned, "inodes_pinned", "Inodes pinned"); | |
2550 | mds_plb.add_u64(l_mds_inodes_expired, "inodes_expired", "Inodes expired"); | |
2551 | mds_plb.add_u64( | |
2552 | l_mds_inodes_with_caps, "inodes_with_caps", "Inodes with capabilities"); | |
2553 | mds_plb.add_u64(l_mds_caps, "caps", "Capabilities", "caps", | |
2554 | PerfCountersBuilder::PRIO_INTERESTING); | |
2555 | mds_plb.add_u64(l_mds_subtrees, "subtrees", "Subtrees"); | |
2556 | ||
2557 | mds_plb.add_u64_counter(l_mds_traverse, "traverse", "Traverses"); | |
2558 | mds_plb.add_u64_counter(l_mds_traverse_hit, "traverse_hit", "Traverse hits"); | |
2559 | mds_plb.add_u64_counter(l_mds_traverse_forward, "traverse_forward", | |
2560 | "Traverse forwards"); | |
2561 | mds_plb.add_u64_counter(l_mds_traverse_discover, "traverse_discover", | |
2562 | "Traverse directory discovers"); | |
2563 | mds_plb.add_u64_counter(l_mds_traverse_dir_fetch, "traverse_dir_fetch", | |
2564 | "Traverse incomplete directory content fetchings"); | |
2565 | mds_plb.add_u64_counter(l_mds_traverse_remote_ino, "traverse_remote_ino", | |
2566 | "Traverse remote dentries"); | |
2567 | mds_plb.add_u64_counter(l_mds_traverse_lock, "traverse_lock", | |
2568 | "Traverse locks"); | |
2569 | ||
2570 | mds_plb.add_u64(l_mds_load_cent, "load_cent", "Load per cent"); | |
2571 | mds_plb.add_u64(l_mds_dispatch_queue_len, "q", "Dispatch queue length"); | |
2572 | ||
2573 | mds_plb.add_u64_counter(l_mds_exported, "exported", "Exports"); | |
2574 | mds_plb.add_u64_counter( | |
2575 | l_mds_exported_inodes, "exported_inodes", "Exported inodes", "exi", | |
2576 | PerfCountersBuilder::PRIO_INTERESTING); | |
2577 | mds_plb.add_u64_counter(l_mds_imported, "imported", "Imports"); | |
2578 | mds_plb.add_u64_counter( | |
2579 | l_mds_imported_inodes, "imported_inodes", "Imported inodes", "imi", | |
2580 | PerfCountersBuilder::PRIO_INTERESTING); | |
2581 | logger = mds_plb.create_perf_counters(); | |
2582 | g_ceph_context->get_perfcounters_collection()->add(logger); | |
2583 | } | |
2584 | ||
2585 | { | |
2586 | PerfCountersBuilder mdm_plb(g_ceph_context, "mds_mem", l_mdm_first, l_mdm_last); | |
2587 | mdm_plb.add_u64(l_mdm_ino, "ino", "Inodes"); | |
2588 | mdm_plb.add_u64_counter(l_mdm_inoa, "ino+", "Inodes opened"); | |
2589 | mdm_plb.add_u64_counter(l_mdm_inos, "ino-", "Inodes closed"); | |
2590 | mdm_plb.add_u64(l_mdm_dir, "dir", "Directories"); | |
2591 | mdm_plb.add_u64_counter(l_mdm_dira, "dir+", "Directories opened"); | |
2592 | mdm_plb.add_u64_counter(l_mdm_dirs, "dir-", "Directories closed"); | |
2593 | mdm_plb.add_u64(l_mdm_dn, "dn", "Dentries"); | |
2594 | mdm_plb.add_u64_counter(l_mdm_dna, "dn+", "Dentries opened"); | |
2595 | mdm_plb.add_u64_counter(l_mdm_dns, "dn-", "Dentries closed"); | |
2596 | mdm_plb.add_u64(l_mdm_cap, "cap", "Capabilities"); | |
2597 | mdm_plb.add_u64_counter(l_mdm_capa, "cap+", "Capabilities added"); | |
2598 | mdm_plb.add_u64_counter(l_mdm_caps, "cap-", "Capabilities removed"); | |
2599 | mdm_plb.add_u64(l_mdm_rss, "rss", "RSS"); | |
2600 | mdm_plb.add_u64(l_mdm_heap, "heap", "Heap size"); | |
2601 | mdm_plb.add_u64(l_mdm_buf, "buf", "Buffer size"); | |
2602 | mlogger = mdm_plb.create_perf_counters(); | |
2603 | g_ceph_context->get_perfcounters_collection()->add(mlogger); | |
2604 | } | |
2605 | ||
2606 | mdlog->create_logger(); | |
2607 | server->create_logger(); | |
2608 | purge_queue.create_logger(); | |
2609 | sessionmap.register_perfcounters(); | |
2610 | mdcache->register_perfcounters(); | |
2611 | } | |
2612 | ||
2613 | void MDSRank::check_ops_in_flight() | |
2614 | { | |
2615 | vector<string> warnings; | |
2616 | int slow = 0; | |
2617 | if (op_tracker.check_ops_in_flight(warnings, &slow)) { | |
2618 | for (vector<string>::iterator i = warnings.begin(); | |
2619 | i != warnings.end(); | |
2620 | ++i) { | |
2621 | clog->warn() << *i; | |
2622 | } | |
2623 | } | |
2624 | ||
2625 | // set mds slow request count | |
2626 | mds_slow_req_count = slow; | |
2627 | return; | |
2628 | } | |
2629 | ||
2630 | void MDSRankDispatcher::handle_osd_map() | |
2631 | { | |
2632 | if (is_active() && snapserver) { | |
2633 | snapserver->check_osd_map(true); | |
2634 | } | |
2635 | ||
2636 | server->handle_osd_map(); | |
2637 | ||
2638 | purge_queue.update_op_limit(*mdsmap); | |
2639 | ||
31f18b77 FG |
2640 | std::set<entity_addr_t> newly_blacklisted; |
2641 | objecter->consume_blacklist_events(&newly_blacklisted); | |
2642 | auto epoch = objecter->with_osdmap([](const OSDMap &o){return o.get_epoch();}); | |
2643 | dout(4) << "handle_osd_map epoch " << epoch << ", " | |
2644 | << newly_blacklisted.size() << " new blacklist entries" << dendl; | |
2645 | auto victims = server->apply_blacklist(newly_blacklisted); | |
2646 | if (victims) { | |
2647 | set_osd_epoch_barrier(epoch); | |
2648 | } | |
2649 | ||
2650 | ||
7c673cae FG |
2651 | // By default the objecter only requests OSDMap updates on use, |
2652 | // we would like to always receive the latest maps in order to | |
2653 | // apply policy based on the FULL flag. | |
2654 | objecter->maybe_request_map(); | |
2655 | } | |
2656 | ||
31f18b77 FG |
2657 | bool MDSRank::evict_client(int64_t session_id, |
2658 | bool wait, bool blacklist, std::stringstream& err_ss, | |
2659 | Context *on_killed) | |
7c673cae | 2660 | { |
31f18b77 FG |
2661 | assert(mds_lock.is_locked_by_me()); |
2662 | ||
2663 | // Mutually exclusive args | |
2664 | assert(!(wait && on_killed != nullptr)); | |
2665 | ||
7c673cae FG |
2666 | if (is_any_replay()) { |
2667 | err_ss << "MDS is replaying log"; | |
2668 | return false; | |
2669 | } | |
2670 | ||
31f18b77 FG |
2671 | Session *session = sessionmap.get_session( |
2672 | entity_name_t(CEPH_ENTITY_TYPE_CLIENT, session_id)); | |
7c673cae FG |
2673 | if (!session) { |
2674 | err_ss << "session " << session_id << " not in sessionmap!"; | |
2675 | return false; | |
2676 | } | |
7c673cae | 2677 | |
31f18b77 FG |
2678 | dout(4) << "Preparing blacklist command... (wait=" << wait << ")" << dendl; |
2679 | stringstream ss; | |
2680 | ss << "{\"prefix\":\"osd blacklist\", \"blacklistop\":\"add\","; | |
2681 | ss << "\"addr\":\""; | |
2682 | ss << session->info.inst.addr; | |
2683 | ss << "\"}"; | |
2684 | std::string tmp = ss.str(); | |
2685 | std::vector<std::string> cmd = {tmp}; | |
2686 | ||
2687 | auto kill_mds_session = [this, session_id, on_killed](){ | |
2688 | assert(mds_lock.is_locked_by_me()); | |
2689 | Session *session = sessionmap.get_session( | |
2690 | entity_name_t(CEPH_ENTITY_TYPE_CLIENT, session_id)); | |
2691 | if (session) { | |
2692 | if (on_killed) { | |
2693 | server->kill_session(session, on_killed); | |
2694 | } else { | |
2695 | C_SaferCond on_safe; | |
2696 | server->kill_session(session, &on_safe); | |
2697 | ||
2698 | mds_lock.Unlock(); | |
2699 | on_safe.wait(); | |
2700 | mds_lock.Lock(); | |
2701 | } | |
2702 | } else { | |
2703 | dout(1) << "session " << session_id << " was removed while we waited " | |
2704 | "for blacklist" << dendl; | |
2705 | ||
2706 | // Even though it wasn't us that removed it, kick our completion | |
2707 | // as the session has been removed. | |
2708 | if (on_killed) { | |
2709 | on_killed->complete(0); | |
2710 | } | |
2711 | } | |
2712 | }; | |
2713 | ||
2714 | auto background_blacklist = [this, session_id, cmd](std::function<void ()> fn){ | |
2715 | assert(mds_lock.is_locked_by_me()); | |
2716 | ||
2717 | Context *on_blacklist_done = new FunctionContext([this, session_id, fn](int r) { | |
2718 | objecter->wait_for_latest_osdmap( | |
2719 | new C_OnFinisher( | |
2720 | new FunctionContext([this, session_id, fn](int r) { | |
2721 | Mutex::Locker l(mds_lock); | |
2722 | auto epoch = objecter->with_osdmap([](const OSDMap &o){ | |
2723 | return o.get_epoch(); | |
2724 | }); | |
2725 | ||
2726 | set_osd_epoch_barrier(epoch); | |
2727 | ||
2728 | fn(); | |
2729 | }), finisher) | |
2730 | ); | |
2731 | }); | |
2732 | ||
2733 | dout(4) << "Sending mon blacklist command: " << cmd[0] << dendl; | |
2734 | monc->start_mon_command(cmd, {}, nullptr, nullptr, on_blacklist_done); | |
2735 | }; | |
2736 | ||
2737 | auto blocking_blacklist = [this, cmd, &err_ss, background_blacklist](){ | |
2738 | C_SaferCond inline_ctx; | |
2739 | background_blacklist([&inline_ctx](){inline_ctx.complete(0);}); | |
7c673cae | 2740 | mds_lock.Unlock(); |
31f18b77 | 2741 | inline_ctx.wait(); |
7c673cae | 2742 | mds_lock.Lock(); |
31f18b77 FG |
2743 | }; |
2744 | ||
2745 | if (wait) { | |
2746 | if (blacklist) { | |
2747 | blocking_blacklist(); | |
2748 | } | |
2749 | ||
2750 | // We dropped mds_lock, so check that session still exists | |
2751 | session = sessionmap.get_session(entity_name_t(CEPH_ENTITY_TYPE_CLIENT, | |
2752 | session_id)); | |
2753 | if (!session) { | |
2754 | dout(1) << "session " << session_id << " was removed while we waited " | |
2755 | "for blacklist" << dendl; | |
2756 | return true; | |
2757 | } | |
2758 | kill_mds_session(); | |
7c673cae | 2759 | } else { |
31f18b77 FG |
2760 | if (blacklist) { |
2761 | background_blacklist(kill_mds_session); | |
2762 | } else { | |
2763 | kill_mds_session(); | |
2764 | } | |
7c673cae | 2765 | } |
31f18b77 | 2766 | |
7c673cae FG |
2767 | return true; |
2768 | } | |
2769 | ||
2770 | void MDSRank::bcast_mds_map() | |
2771 | { | |
2772 | dout(7) << "bcast_mds_map " << mdsmap->get_epoch() << dendl; | |
2773 | ||
2774 | // share the map with mounted clients | |
2775 | set<Session*> clients; | |
2776 | sessionmap.get_client_session_set(clients); | |
2777 | for (set<Session*>::const_iterator p = clients.begin(); | |
2778 | p != clients.end(); | |
2779 | ++p) | |
2780 | (*p)->connection->send_message(new MMDSMap(monc->get_fsid(), mdsmap)); | |
2781 | last_client_mdsmap_bcast = mdsmap->get_epoch(); | |
2782 | } | |
2783 | ||
2784 | MDSRankDispatcher::MDSRankDispatcher( | |
2785 | mds_rank_t whoami_, | |
2786 | Mutex &mds_lock_, | |
2787 | LogChannelRef &clog_, | |
2788 | SafeTimer &timer_, | |
2789 | Beacon &beacon_, | |
2790 | MDSMap *& mdsmap_, | |
2791 | Messenger *msgr, | |
2792 | MonClient *monc_, | |
2793 | Context *respawn_hook_, | |
2794 | Context *suicide_hook_) | |
2795 | : MDSRank(whoami_, mds_lock_, clog_, timer_, beacon_, mdsmap_, | |
2796 | msgr, monc_, respawn_hook_, suicide_hook_) | |
2797 | {} | |
2798 | ||
2799 | bool MDSRankDispatcher::handle_command( | |
2800 | const cmdmap_t &cmdmap, | |
2801 | MCommand *m, | |
2802 | int *r, | |
2803 | std::stringstream *ds, | |
2804 | std::stringstream *ss, | |
2805 | bool *need_reply) | |
2806 | { | |
2807 | assert(r != nullptr); | |
2808 | assert(ds != nullptr); | |
2809 | assert(ss != nullptr); | |
2810 | ||
2811 | *need_reply = true; | |
2812 | ||
2813 | std::string prefix; | |
2814 | cmd_getval(g_ceph_context, cmdmap, "prefix", prefix); | |
2815 | ||
31f18b77 | 2816 | if (prefix == "session ls" || prefix == "client ls") { |
7c673cae FG |
2817 | std::vector<std::string> filter_args; |
2818 | cmd_getval(g_ceph_context, cmdmap, "filters", filter_args); | |
2819 | ||
2820 | SessionFilter filter; | |
2821 | *r = filter.parse(filter_args, ss); | |
2822 | if (*r != 0) { | |
2823 | return true; | |
2824 | } | |
2825 | ||
2826 | Formatter *f = new JSONFormatter(true); | |
2827 | dump_sessions(filter, f); | |
2828 | f->flush(*ds); | |
2829 | delete f; | |
2830 | return true; | |
31f18b77 | 2831 | } else if (prefix == "session evict" || prefix == "client evict") { |
7c673cae FG |
2832 | std::vector<std::string> filter_args; |
2833 | cmd_getval(g_ceph_context, cmdmap, "filters", filter_args); | |
2834 | ||
2835 | SessionFilter filter; | |
2836 | *r = filter.parse(filter_args, ss); | |
2837 | if (*r != 0) { | |
2838 | return true; | |
2839 | } | |
2840 | ||
31f18b77 | 2841 | evict_clients(filter, m); |
7c673cae FG |
2842 | |
2843 | *need_reply = false; | |
2844 | return true; | |
2845 | } else if (prefix == "damage ls") { | |
2846 | Formatter *f = new JSONFormatter(true); | |
2847 | damage_table.dump(f); | |
2848 | f->flush(*ds); | |
2849 | delete f; | |
2850 | return true; | |
2851 | } else if (prefix == "damage rm") { | |
2852 | damage_entry_id_t id = 0; | |
2853 | bool got = cmd_getval(g_ceph_context, cmdmap, "damage_id", (int64_t&)id); | |
2854 | if (!got) { | |
2855 | *r = -EINVAL; | |
2856 | return true; | |
2857 | } | |
2858 | ||
2859 | damage_table.erase(id); | |
2860 | return true; | |
2861 | } else { | |
2862 | return false; | |
2863 | } | |
2864 | } | |
2865 | ||
2866 | epoch_t MDSRank::get_osd_epoch() const | |
2867 | { | |
2868 | return objecter->with_osdmap(std::mem_fn(&OSDMap::get_epoch)); | |
2869 | } | |
2870 |