]>
Commit | Line | Data |
---|---|---|
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 <string_view> | |
16 | ||
17 | #include "common/debug.h" | |
18 | #include "common/errno.h" | |
19 | ||
20 | #include "messages/MClientRequestForward.h" | |
21 | #include "messages/MMDSLoadTargets.h" | |
22 | #include "messages/MMDSTableRequest.h" | |
23 | ||
24 | #include "MDSDaemon.h" | |
25 | #include "MDSMap.h" | |
26 | #include "SnapClient.h" | |
27 | #include "SnapServer.h" | |
28 | #include "MDBalancer.h" | |
29 | #include "Migrator.h" | |
30 | #include "Locker.h" | |
31 | #include "InoTable.h" | |
32 | #include "mon/MonClient.h" | |
33 | #include "common/HeartbeatMap.h" | |
34 | #include "ScrubStack.h" | |
35 | ||
36 | ||
37 | #include "MDSRank.h" | |
38 | ||
39 | #define dout_context g_ceph_context | |
40 | #define dout_subsys ceph_subsys_mds | |
41 | #undef dout_prefix | |
42 | #define dout_prefix *_dout << "mds." << whoami << '.' << incarnation << ' ' | |
43 | ||
44 | class C_Flush_Journal : public MDSInternalContext { | |
45 | public: | |
46 | C_Flush_Journal(MDCache *mdcache, MDLog *mdlog, MDSRank *mds, | |
47 | std::ostream *ss, Context *on_finish) | |
48 | : MDSInternalContext(mds), | |
49 | mdcache(mdcache), mdlog(mdlog), ss(ss), on_finish(on_finish), | |
50 | whoami(mds->whoami), incarnation(mds->incarnation) { | |
51 | } | |
52 | ||
53 | void send() { | |
54 | assert(mds->mds_lock.is_locked()); | |
55 | ||
56 | dout(20) << __func__ << dendl; | |
57 | ||
58 | if (mdcache->is_readonly()) { | |
59 | dout(5) << __func__ << ": read-only FS" << dendl; | |
60 | complete(-EROFS); | |
61 | return; | |
62 | } | |
63 | ||
64 | if (!mds->is_active()) { | |
65 | dout(5) << __func__ << ": MDS not active, no-op" << dendl; | |
66 | complete(0); | |
67 | return; | |
68 | } | |
69 | ||
70 | flush_mdlog(); | |
71 | } | |
72 | ||
73 | private: | |
74 | ||
75 | void flush_mdlog() { | |
76 | dout(20) << __func__ << dendl; | |
77 | ||
78 | // I need to seal off the current segment, and then mark all | |
79 | // previous segments for expiry | |
80 | mdlog->start_new_segment(); | |
81 | ||
82 | Context *ctx = new FunctionContext([this](int r) { | |
83 | handle_flush_mdlog(r); | |
84 | }); | |
85 | ||
86 | // Flush initially so that all the segments older than our new one | |
87 | // will be elegible for expiry | |
88 | mdlog->flush(); | |
89 | mdlog->wait_for_safe(new MDSInternalContextWrapper(mds, ctx)); | |
90 | } | |
91 | ||
92 | void handle_flush_mdlog(int r) { | |
93 | dout(20) << __func__ << ": r=" << r << dendl; | |
94 | ||
95 | if (r != 0) { | |
96 | *ss << "Error " << r << " (" << cpp_strerror(r) << ") while flushing journal"; | |
97 | complete(r); | |
98 | return; | |
99 | } | |
100 | ||
101 | clear_mdlog(); | |
102 | } | |
103 | ||
104 | void clear_mdlog() { | |
105 | dout(20) << __func__ << dendl; | |
106 | ||
107 | Context *ctx = new FunctionContext([this](int r) { | |
108 | handle_clear_mdlog(r); | |
109 | }); | |
110 | ||
111 | // Because we may not be the last wait_for_safe context on MDLog, | |
112 | // and subsequent contexts might wake up in the middle of our | |
113 | // later trim_all and interfere with expiry (by e.g. marking | |
114 | // dirs/dentries dirty on previous log segments), we run a second | |
115 | // wait_for_safe here. See #10368 | |
116 | mdlog->wait_for_safe(new MDSInternalContextWrapper(mds, ctx)); | |
117 | } | |
118 | ||
119 | void handle_clear_mdlog(int r) { | |
120 | dout(20) << __func__ << ": r=" << r << dendl; | |
121 | ||
122 | if (r != 0) { | |
123 | *ss << "Error " << r << " (" << cpp_strerror(r) << ") while flushing journal"; | |
124 | complete(r); | |
125 | return; | |
126 | } | |
127 | ||
128 | trim_mdlog(); | |
129 | } | |
130 | ||
131 | void trim_mdlog() { | |
132 | // Put all the old log segments into expiring or expired state | |
133 | dout(5) << __func__ << ": beginning segment expiry" << dendl; | |
134 | ||
135 | int ret = mdlog->trim_all(); | |
136 | if (ret != 0) { | |
137 | *ss << "Error " << ret << " (" << cpp_strerror(ret) << ") while trimming log"; | |
138 | complete(ret); | |
139 | return; | |
140 | } | |
141 | ||
142 | expire_segments(); | |
143 | } | |
144 | ||
145 | void expire_segments() { | |
146 | dout(20) << __func__ << dendl; | |
147 | ||
148 | // Attach contexts to wait for all expiring segments to expire | |
149 | MDSGatherBuilder *expiry_gather = new MDSGatherBuilder(g_ceph_context); | |
150 | ||
151 | const auto &expiring_segments = mdlog->get_expiring_segments(); | |
152 | for (auto p : expiring_segments) { | |
153 | p->wait_for_expiry(expiry_gather->new_sub()); | |
154 | } | |
155 | dout(5) << __func__ << ": waiting for " << expiry_gather->num_subs_created() | |
156 | << " segments to expire" << dendl; | |
157 | ||
158 | if (!expiry_gather->has_subs()) { | |
159 | trim_segments(); | |
160 | delete expiry_gather; | |
161 | return; | |
162 | } | |
163 | ||
164 | Context *ctx = new FunctionContext([this](int r) { | |
165 | handle_expire_segments(r); | |
166 | }); | |
167 | expiry_gather->set_finisher(new MDSInternalContextWrapper(mds, ctx)); | |
168 | expiry_gather->activate(); | |
169 | } | |
170 | ||
171 | void handle_expire_segments(int r) { | |
172 | dout(20) << __func__ << ": r=" << r << dendl; | |
173 | ||
174 | ceph_assert(r == 0); // MDLog is not allowed to raise errors via | |
175 | // wait_for_expiry | |
176 | trim_segments(); | |
177 | } | |
178 | ||
179 | void trim_segments() { | |
180 | dout(20) << __func__ << dendl; | |
181 | ||
182 | Context *ctx = new C_OnFinisher(new FunctionContext([this](int _) { | |
183 | std::lock_guard locker(mds->mds_lock); | |
184 | trim_expired_segments(); | |
185 | }), mds->finisher); | |
186 | ctx->complete(0); | |
187 | } | |
188 | ||
189 | void trim_expired_segments() { | |
190 | dout(5) << __func__ << ": expiry complete, expire_pos/trim_pos is now " | |
191 | << std::hex << mdlog->get_journaler()->get_expire_pos() << "/" | |
192 | << mdlog->get_journaler()->get_trimmed_pos() << dendl; | |
193 | ||
194 | // Now everyone I'm interested in is expired | |
195 | mdlog->trim_expired_segments(); | |
196 | ||
197 | dout(5) << __func__ << ": trim complete, expire_pos/trim_pos is now " | |
198 | << std::hex << mdlog->get_journaler()->get_expire_pos() << "/" | |
199 | << mdlog->get_journaler()->get_trimmed_pos() << dendl; | |
200 | ||
201 | write_journal_head(); | |
202 | } | |
203 | ||
204 | void write_journal_head() { | |
205 | dout(20) << __func__ << dendl; | |
206 | ||
207 | Context *ctx = new FunctionContext([this](int r) { | |
208 | std::lock_guard locker(mds->mds_lock); | |
209 | handle_write_head(r); | |
210 | }); | |
211 | // Flush the journal header so that readers will start from after | |
212 | // the flushed region | |
213 | mdlog->get_journaler()->write_head(ctx); | |
214 | } | |
215 | ||
216 | void handle_write_head(int r) { | |
217 | if (r != 0) { | |
218 | *ss << "Error " << r << " (" << cpp_strerror(r) << ") while writing header"; | |
219 | } else { | |
220 | dout(5) << __func__ << ": write_head complete, all done!" << dendl; | |
221 | } | |
222 | ||
223 | complete(r); | |
224 | } | |
225 | ||
226 | void finish(int r) override { | |
227 | dout(20) << __func__ << ": r=" << r << dendl; | |
228 | on_finish->complete(r); | |
229 | } | |
230 | ||
231 | MDCache *mdcache; | |
232 | MDLog *mdlog; | |
233 | std::ostream *ss; | |
234 | Context *on_finish; | |
235 | ||
236 | // so as to use dout | |
237 | mds_rank_t whoami; | |
238 | int incarnation; | |
239 | }; | |
240 | ||
241 | class C_Drop_Cache : public MDSInternalContext { | |
242 | public: | |
243 | C_Drop_Cache(Server *server, MDCache *mdcache, MDLog *mdlog, | |
244 | MDSRank *mds, uint64_t recall_timeout, | |
245 | Formatter *f, Context *on_finish) | |
246 | : MDSInternalContext(mds), | |
247 | server(server), mdcache(mdcache), mdlog(mdlog), | |
248 | recall_timeout(recall_timeout), recall_start(mono_clock::now()), | |
249 | f(f), on_finish(on_finish), | |
250 | whoami(mds->whoami), incarnation(mds->incarnation) { | |
251 | } | |
252 | ||
253 | void send() { | |
254 | // not really a hard requirement here, but lets ensure this in | |
255 | // case we change the logic here. | |
256 | assert(mds->mds_lock.is_locked()); | |
257 | ||
258 | dout(20) << __func__ << dendl; | |
259 | f->open_object_section("result"); | |
260 | recall_client_state(); | |
261 | } | |
262 | ||
263 | private: | |
264 | // context which completes itself (with -ETIMEDOUT) after a specified | |
265 | // timeout or when explicitly completed, whichever comes first. Note | |
266 | // that the context does not detroy itself after completion -- it | |
267 | // needs to be explicitly freed. | |
268 | class C_ContextTimeout : public MDSInternalContext { | |
269 | public: | |
270 | C_ContextTimeout(MDSRank *mds, uint64_t timeout, Context *on_finish) | |
271 | : MDSInternalContext(mds), | |
272 | timeout(timeout), | |
273 | lock("mds::context::timeout", false, true), | |
274 | on_finish(on_finish) { | |
275 | } | |
276 | ~C_ContextTimeout() { | |
277 | ceph_assert(timer_task == nullptr); | |
278 | } | |
279 | ||
280 | void start_timer() { | |
281 | if (!timeout) { | |
282 | return; | |
283 | } | |
284 | ||
285 | timer_task = new FunctionContext([this](int _) { | |
286 | timer_task = nullptr; | |
287 | complete(-ETIMEDOUT); | |
288 | }); | |
289 | mds->timer.add_event_after(timeout, timer_task); | |
290 | } | |
291 | ||
292 | void finish(int r) override { | |
293 | Context *ctx = nullptr; | |
294 | { | |
295 | std::lock_guard locker(lock); | |
296 | std::swap(on_finish, ctx); | |
297 | } | |
298 | if (ctx != nullptr) { | |
299 | ctx->complete(r); | |
300 | } | |
301 | } | |
302 | void complete(int r) override { | |
303 | if (timer_task != nullptr) { | |
304 | mds->timer.cancel_event(timer_task); | |
305 | } | |
306 | ||
307 | finish(r); | |
308 | } | |
309 | ||
310 | uint64_t timeout; | |
311 | Mutex lock; | |
312 | Context *on_finish = nullptr; | |
313 | Context *timer_task = nullptr; | |
314 | }; | |
315 | ||
316 | auto do_trim() { | |
317 | auto [throttled, count] = mdcache->trim(UINT64_MAX); | |
318 | dout(10) << __func__ | |
319 | << (throttled ? " (throttled)" : "") | |
320 | << " trimmed " << count << " caps" << dendl; | |
321 | dentries_trimmed += count; | |
322 | return std::make_pair(throttled, count); | |
323 | } | |
324 | ||
325 | void recall_client_state() { | |
326 | dout(20) << __func__ << dendl; | |
327 | auto now = mono_clock::now(); | |
328 | auto duration = std::chrono::duration<double>(now-recall_start).count(); | |
329 | ||
330 | MDSGatherBuilder *gather = new MDSGatherBuilder(g_ceph_context); | |
331 | auto flags = Server::RecallFlags::STEADY|Server::RecallFlags::TRIM; | |
332 | auto [throttled, count] = server->recall_client_state(gather, flags); | |
333 | dout(10) << __func__ | |
334 | << (throttled ? " (throttled)" : "") | |
335 | << " recalled " << count << " caps" << dendl; | |
336 | ||
337 | caps_recalled += count; | |
338 | if ((throttled || count > 0) && (recall_timeout == 0 || duration < recall_timeout)) { | |
339 | C_ContextTimeout *ctx = new C_ContextTimeout( | |
340 | mds, 1, new FunctionContext([this](int r) { | |
341 | recall_client_state(); | |
342 | })); | |
343 | ctx->start_timer(); | |
344 | gather->set_finisher(new MDSInternalContextWrapper(mds, ctx)); | |
345 | gather->activate(); | |
346 | mdlog->flush(); /* use down-time to incrementally flush log */ | |
347 | do_trim(); /* use down-time to incrementally trim cache */ | |
348 | } else { | |
349 | if (!gather->has_subs()) { | |
350 | delete gather; | |
351 | return handle_recall_client_state(0); | |
352 | } else if (recall_timeout > 0 && duration > recall_timeout) { | |
353 | gather->set_finisher(new C_MDSInternalNoop); | |
354 | gather->activate(); | |
355 | return handle_recall_client_state(-ETIMEDOUT); | |
356 | } else { | |
357 | uint64_t remaining = (recall_timeout == 0 ? 0 : recall_timeout-duration); | |
358 | C_ContextTimeout *ctx = new C_ContextTimeout( | |
359 | mds, remaining, new FunctionContext([this](int r) { | |
360 | handle_recall_client_state(r); | |
361 | })); | |
362 | ||
363 | ctx->start_timer(); | |
364 | gather->set_finisher(new MDSInternalContextWrapper(mds, ctx)); | |
365 | gather->activate(); | |
366 | } | |
367 | } | |
368 | } | |
369 | ||
370 | void handle_recall_client_state(int r) { | |
371 | dout(20) << __func__ << ": r=" << r << dendl; | |
372 | ||
373 | // client recall section | |
374 | f->open_object_section("client_recall"); | |
375 | f->dump_int("return_code", r); | |
376 | f->dump_string("message", cpp_strerror(r)); | |
377 | f->dump_int("recalled", caps_recalled); | |
378 | f->close_section(); | |
379 | ||
380 | // we can still continue after recall timeout | |
381 | flush_journal(); | |
382 | } | |
383 | ||
384 | void flush_journal() { | |
385 | dout(20) << __func__ << dendl; | |
386 | ||
387 | Context *ctx = new FunctionContext([this](int r) { | |
388 | handle_flush_journal(r); | |
389 | }); | |
390 | ||
391 | C_Flush_Journal *flush_journal = new C_Flush_Journal(mdcache, mdlog, mds, &ss, ctx); | |
392 | flush_journal->send(); | |
393 | } | |
394 | ||
395 | void handle_flush_journal(int r) { | |
396 | dout(20) << __func__ << ": r=" << r << dendl; | |
397 | ||
398 | if (r != 0) { | |
399 | cmd_err(f, ss.str()); | |
400 | complete(r); | |
401 | return; | |
402 | } | |
403 | ||
404 | // journal flush section | |
405 | f->open_object_section("flush_journal"); | |
406 | f->dump_int("return_code", r); | |
407 | f->dump_string("message", ss.str()); | |
408 | f->close_section(); | |
409 | ||
410 | trim_cache(); | |
411 | } | |
412 | ||
413 | void trim_cache() { | |
414 | dout(20) << __func__ << dendl; | |
415 | ||
416 | auto [throttled, count] = do_trim(); | |
417 | if (throttled && count > 0) { | |
418 | auto timer = new FunctionContext([this](int _) { | |
419 | trim_cache(); | |
420 | }); | |
421 | mds->timer.add_event_after(1.0, timer); | |
422 | } else { | |
423 | cache_status(); | |
424 | } | |
425 | } | |
426 | ||
427 | void cache_status() { | |
428 | dout(20) << __func__ << dendl; | |
429 | ||
430 | f->open_object_section("trim_cache"); | |
431 | f->dump_int("trimmed", dentries_trimmed); | |
432 | f->close_section(); | |
433 | ||
434 | // cache status section | |
435 | mdcache->cache_status(f); | |
436 | ||
437 | complete(0); | |
438 | } | |
439 | ||
440 | void finish(int r) override { | |
441 | dout(20) << __func__ << ": r=" << r << dendl; | |
442 | ||
443 | auto d = std::chrono::duration<double>(mono_clock::now()-recall_start); | |
444 | f->dump_float("duration", d.count()); | |
445 | ||
446 | f->close_section(); | |
447 | on_finish->complete(r); | |
448 | } | |
449 | ||
450 | Server *server; | |
451 | MDCache *mdcache; | |
452 | MDLog *mdlog; | |
453 | uint64_t recall_timeout; | |
454 | mono_time recall_start; | |
455 | Formatter *f; | |
456 | Context *on_finish; | |
457 | ||
458 | int retval = 0; | |
459 | std::stringstream ss; | |
460 | uint64_t caps_recalled = 0; | |
461 | uint64_t dentries_trimmed = 0; | |
462 | ||
463 | // so as to use dout | |
464 | mds_rank_t whoami; | |
465 | int incarnation; | |
466 | ||
467 | void cmd_err(Formatter *f, std::string_view err) { | |
468 | f->reset(); | |
469 | f->open_object_section("result"); | |
470 | f->dump_string("error", err); | |
471 | f->close_section(); | |
472 | } | |
473 | }; | |
474 | ||
475 | MDSRank::MDSRank( | |
476 | mds_rank_t whoami_, | |
477 | Mutex &mds_lock_, | |
478 | LogChannelRef &clog_, | |
479 | SafeTimer &timer_, | |
480 | Beacon &beacon_, | |
481 | std::unique_ptr<MDSMap>& mdsmap_, | |
482 | Messenger *msgr, | |
483 | MonClient *monc_, | |
484 | Context *respawn_hook_, | |
485 | Context *suicide_hook_) | |
486 | : | |
487 | whoami(whoami_), incarnation(0), | |
488 | mds_lock(mds_lock_), cct(msgr->cct), clog(clog_), timer(timer_), | |
489 | mdsmap(mdsmap_), | |
490 | objecter(new Objecter(g_ceph_context, msgr, monc_, nullptr, 0, 0)), | |
491 | server(NULL), mdcache(NULL), locker(NULL), mdlog(NULL), | |
492 | balancer(NULL), scrubstack(NULL), | |
493 | damage_table(whoami_), | |
494 | inotable(NULL), snapserver(NULL), snapclient(NULL), | |
495 | sessionmap(this), logger(NULL), mlogger(NULL), | |
496 | op_tracker(g_ceph_context, g_conf()->mds_enable_op_tracker, | |
497 | g_conf()->osd_num_op_tracker_shard), | |
498 | last_state(MDSMap::STATE_BOOT), | |
499 | state(MDSMap::STATE_BOOT), | |
500 | cluster_degraded(false), stopping(false), | |
501 | purge_queue(g_ceph_context, whoami_, | |
502 | mdsmap_->get_metadata_pool(), objecter, | |
503 | new FunctionContext( | |
504 | [this](int r){ | |
505 | // Purge Queue operates inside mds_lock when we're calling into | |
506 | // it, and outside when in background, so must handle both cases. | |
507 | if (mds_lock.is_locked_by_me()) { | |
508 | handle_write_error(r); | |
509 | } else { | |
510 | std::lock_guard l(mds_lock); | |
511 | handle_write_error(r); | |
512 | } | |
513 | } | |
514 | ) | |
515 | ), | |
516 | progress_thread(this), dispatch_depth(0), | |
517 | hb(NULL), last_tid(0), osd_epoch_barrier(0), beacon(beacon_), | |
518 | mds_slow_req_count(0), | |
519 | last_client_mdsmap_bcast(0), | |
520 | messenger(msgr), monc(monc_), | |
521 | respawn_hook(respawn_hook_), | |
522 | suicide_hook(suicide_hook_), | |
523 | standby_replaying(false), | |
524 | starttime(mono_clock::now()) | |
525 | { | |
526 | hb = g_ceph_context->get_heartbeat_map()->add_worker("MDSRank", pthread_self()); | |
527 | ||
528 | purge_queue.update_op_limit(*mdsmap); | |
529 | ||
530 | objecter->unset_honor_osdmap_full(); | |
531 | ||
532 | finisher = new Finisher(cct); | |
533 | ||
534 | mdcache = new MDCache(this, purge_queue); | |
535 | mdlog = new MDLog(this); | |
536 | balancer = new MDBalancer(this, messenger, monc); | |
537 | ||
538 | scrubstack = new ScrubStack(mdcache, finisher); | |
539 | ||
540 | inotable = new InoTable(this); | |
541 | snapserver = new SnapServer(this, monc); | |
542 | snapclient = new SnapClient(this); | |
543 | ||
544 | server = new Server(this); | |
545 | locker = new Locker(this, mdcache); | |
546 | ||
547 | op_tracker.set_complaint_and_threshold(cct->_conf->mds_op_complaint_time, | |
548 | cct->_conf->mds_op_log_threshold); | |
549 | op_tracker.set_history_size_and_duration(cct->_conf->mds_op_history_size, | |
550 | cct->_conf->mds_op_history_duration); | |
551 | } | |
552 | ||
553 | MDSRank::~MDSRank() | |
554 | { | |
555 | if (hb) { | |
556 | g_ceph_context->get_heartbeat_map()->remove_worker(hb); | |
557 | } | |
558 | ||
559 | if (scrubstack) { delete scrubstack; scrubstack = NULL; } | |
560 | if (mdcache) { delete mdcache; mdcache = NULL; } | |
561 | if (mdlog) { delete mdlog; mdlog = NULL; } | |
562 | if (balancer) { delete balancer; balancer = NULL; } | |
563 | if (inotable) { delete inotable; inotable = NULL; } | |
564 | if (snapserver) { delete snapserver; snapserver = NULL; } | |
565 | if (snapclient) { delete snapclient; snapclient = NULL; } | |
566 | ||
567 | if (server) { delete server; server = 0; } | |
568 | if (locker) { delete locker; locker = 0; } | |
569 | ||
570 | if (logger) { | |
571 | g_ceph_context->get_perfcounters_collection()->remove(logger); | |
572 | delete logger; | |
573 | logger = 0; | |
574 | } | |
575 | if (mlogger) { | |
576 | g_ceph_context->get_perfcounters_collection()->remove(mlogger); | |
577 | delete mlogger; | |
578 | mlogger = 0; | |
579 | } | |
580 | ||
581 | delete finisher; | |
582 | finisher = NULL; | |
583 | ||
584 | delete suicide_hook; | |
585 | suicide_hook = NULL; | |
586 | ||
587 | delete respawn_hook; | |
588 | respawn_hook = NULL; | |
589 | ||
590 | delete objecter; | |
591 | objecter = nullptr; | |
592 | } | |
593 | ||
594 | void MDSRankDispatcher::init() | |
595 | { | |
596 | objecter->init(); | |
597 | messenger->add_dispatcher_head(objecter); | |
598 | ||
599 | objecter->start(); | |
600 | ||
601 | update_log_config(); | |
602 | create_logger(); | |
603 | ||
604 | // Expose the OSDMap (already populated during MDS::init) to anyone | |
605 | // who is interested in it. | |
606 | handle_osd_map(); | |
607 | ||
608 | progress_thread.create("mds_rank_progr"); | |
609 | ||
610 | purge_queue.init(); | |
611 | ||
612 | finisher->start(); | |
613 | } | |
614 | ||
615 | void MDSRank::update_targets() | |
616 | { | |
617 | // get MonMap's idea of my export_targets | |
618 | const set<mds_rank_t>& map_targets = mdsmap->get_mds_info(get_nodeid()).export_targets; | |
619 | ||
620 | dout(20) << "updating export targets, currently " << map_targets.size() << " ranks are targets" << dendl; | |
621 | ||
622 | bool send = false; | |
623 | set<mds_rank_t> new_map_targets; | |
624 | ||
625 | auto it = export_targets.begin(); | |
626 | while (it != export_targets.end()) { | |
627 | mds_rank_t rank = it->first; | |
628 | auto &counter = it->second; | |
629 | dout(20) << "export target mds." << rank << " is " << counter << dendl; | |
630 | ||
631 | double val = counter.get(); | |
632 | if (val <= 0.01) { | |
633 | dout(15) << "export target mds." << rank << " is no longer an export target" << dendl; | |
634 | export_targets.erase(it++); | |
635 | send = true; | |
636 | continue; | |
637 | } | |
638 | if (!map_targets.count(rank)) { | |
639 | dout(15) << "export target mds." << rank << " not in map's export_targets" << dendl; | |
640 | send = true; | |
641 | } | |
642 | new_map_targets.insert(rank); | |
643 | it++; | |
644 | } | |
645 | if (new_map_targets.size() < map_targets.size()) { | |
646 | dout(15) << "export target map holds stale targets, sending update" << dendl; | |
647 | send = true; | |
648 | } | |
649 | ||
650 | if (send) { | |
651 | dout(15) << "updating export_targets, now " << new_map_targets.size() << " ranks are targets" << dendl; | |
652 | auto m = MMDSLoadTargets::create(mds_gid_t(monc->get_global_id()), new_map_targets); | |
653 | monc->send_mon_message(m.detach()); | |
654 | } | |
655 | } | |
656 | ||
657 | void MDSRank::hit_export_target(mds_rank_t rank, double amount) | |
658 | { | |
659 | double rate = g_conf()->mds_bal_target_decay; | |
660 | if (amount < 0.0) { | |
661 | amount = 100.0/g_conf()->mds_bal_target_decay; /* a good default for "i am trying to keep this export_target active" */ | |
662 | } | |
663 | auto em = export_targets.emplace(std::piecewise_construct, std::forward_as_tuple(rank), std::forward_as_tuple(DecayRate(rate))); | |
664 | auto &counter = em.first->second; | |
665 | counter.hit(amount); | |
666 | if (em.second) { | |
667 | dout(15) << "hit export target (new) is " << counter << dendl; | |
668 | } else { | |
669 | dout(15) << "hit export target is " << counter << dendl; | |
670 | } | |
671 | } | |
672 | ||
673 | class C_MDS_MonCommand : public MDSInternalContext { | |
674 | std::string cmd; | |
675 | public: | |
676 | std::string outs; | |
677 | C_MDS_MonCommand(MDSRank *m, std::string_view c) | |
678 | : MDSInternalContext(m), cmd(c) {} | |
679 | void finish(int r) override { | |
680 | mds->_mon_command_finish(r, cmd, outs); | |
681 | } | |
682 | }; | |
683 | ||
684 | void MDSRank::_mon_command_finish(int r, std::string_view cmd, std::string_view outs) | |
685 | { | |
686 | if (r < 0) { | |
687 | dout(0) << __func__ << ": mon command " << cmd << " failed with errno " << r | |
688 | << " (" << outs << ")" << dendl; | |
689 | } else { | |
690 | dout(1) << __func__ << ": mon command " << cmd << " succeed" << dendl; | |
691 | } | |
692 | } | |
693 | ||
694 | void MDSRank::set_mdsmap_multimds_snaps_allowed() | |
695 | { | |
696 | static bool already_sent = false; | |
697 | if (already_sent) | |
698 | return; | |
699 | ||
700 | stringstream ss; | |
701 | ss << "{\"prefix\":\"fs set\", \"fs_name\":\"" << mdsmap->get_fs_name() << "\", "; | |
702 | ss << "\"var\":\"allow_multimds_snaps\", \"val\":\"true\", "; | |
703 | ss << "\"confirm\":\"--yes-i-am-really-a-mds\"}"; | |
704 | std::vector<std::string> cmd = {ss.str()}; | |
705 | ||
706 | dout(0) << __func__ << ": sending mon command: " << cmd[0] << dendl; | |
707 | ||
708 | C_MDS_MonCommand *fin = new C_MDS_MonCommand(this, cmd[0]); | |
709 | monc->start_mon_command(cmd, {}, nullptr, &fin->outs, new C_IO_Wrapper(this, fin)); | |
710 | ||
711 | already_sent = true; | |
712 | } | |
713 | ||
714 | void MDSRank::mark_base_recursively_scrubbed(inodeno_t ino) | |
715 | { | |
716 | if (mdsmap->get_tableserver() == whoami) | |
717 | snapserver->mark_base_recursively_scrubbed(ino); | |
718 | } | |
719 | ||
720 | void MDSRankDispatcher::tick() | |
721 | { | |
722 | heartbeat_reset(); | |
723 | ||
724 | if (beacon.is_laggy()) { | |
725 | dout(1) << "skipping upkeep work because connection to Monitors appears laggy" << dendl; | |
726 | return; | |
727 | } | |
728 | ||
729 | check_ops_in_flight(); | |
730 | ||
731 | // Wake up thread in case we use to be laggy and have waiting_for_nolaggy | |
732 | // messages to progress. | |
733 | progress_thread.signal(); | |
734 | ||
735 | // make sure mds log flushes, trims periodically | |
736 | mdlog->flush(); | |
737 | ||
738 | // update average session uptime | |
739 | sessionmap.update_average_session_age(); | |
740 | ||
741 | if (is_active() || is_stopping()) { | |
742 | mdlog->trim(); // NOT during recovery! | |
743 | } | |
744 | ||
745 | // ... | |
746 | if (is_cache_trimmable()) { | |
747 | server->find_idle_sessions(); | |
748 | server->evict_cap_revoke_non_responders(); | |
749 | locker->tick(); | |
750 | } | |
751 | ||
752 | // log | |
753 | if (logger) { | |
754 | logger->set(l_mds_subtrees, mdcache->num_subtrees()); | |
755 | mdcache->log_stat(); | |
756 | } | |
757 | ||
758 | if (is_reconnect()) | |
759 | server->reconnect_tick(); | |
760 | ||
761 | if (is_active()) { | |
762 | balancer->tick(); | |
763 | mdcache->find_stale_fragment_freeze(); | |
764 | mdcache->migrator->find_stale_export_freeze(); | |
765 | ||
766 | if (mdsmap->get_tableserver() == whoami) { | |
767 | snapserver->check_osd_map(false); | |
768 | // Filesystem was created by pre-mimic mds. Allow multi-active mds after | |
769 | // all old snapshots are deleted. | |
770 | if (!mdsmap->allows_multimds_snaps() && | |
771 | snapserver->can_allow_multimds_snaps()) { | |
772 | set_mdsmap_multimds_snaps_allowed(); | |
773 | } | |
774 | } | |
775 | } | |
776 | ||
777 | if (is_active() || is_stopping()) { | |
778 | update_targets(); | |
779 | } | |
780 | ||
781 | // shut down? | |
782 | if (is_stopping()) { | |
783 | mdlog->trim(); | |
784 | if (mdcache->shutdown_pass()) { | |
785 | uint64_t pq_progress = 0 ; | |
786 | uint64_t pq_total = 0; | |
787 | size_t pq_in_flight = 0; | |
788 | if (!purge_queue.drain(&pq_progress, &pq_total, &pq_in_flight)) { | |
789 | dout(7) << "shutdown_pass=true, but still waiting for purge queue" | |
790 | << dendl; | |
791 | // This takes unbounded time, so we must indicate progress | |
792 | // to the administrator: we do it in a slightly imperfect way | |
793 | // by sending periodic (tick frequency) clog messages while | |
794 | // in this state. | |
795 | clog->info() << "MDS rank " << whoami << " waiting for purge queue (" | |
796 | << std::dec << pq_progress << "/" << pq_total << " " << pq_in_flight | |
797 | << " files purging" << ")"; | |
798 | } else { | |
799 | dout(7) << "shutdown_pass=true, finished w/ shutdown, moving to " | |
800 | "down:stopped" << dendl; | |
801 | stopping_done(); | |
802 | } | |
803 | } | |
804 | else { | |
805 | dout(7) << "shutdown_pass=false" << dendl; | |
806 | } | |
807 | } | |
808 | ||
809 | // Expose ourselves to Beacon to update health indicators | |
810 | beacon.notify_health(this); | |
811 | } | |
812 | ||
813 | void MDSRankDispatcher::shutdown() | |
814 | { | |
815 | // It should never be possible for shutdown to get called twice, because | |
816 | // anyone picking up mds_lock checks if stopping is true and drops | |
817 | // out if it is. | |
818 | ceph_assert(stopping == false); | |
819 | stopping = true; | |
820 | ||
821 | dout(1) << __func__ << ": shutting down rank " << whoami << dendl; | |
822 | ||
823 | g_conf().remove_observer(this); | |
824 | ||
825 | timer.shutdown(); | |
826 | ||
827 | // MDLog has to shut down before the finisher, because some of its | |
828 | // threads block on IOs that require finisher to complete. | |
829 | mdlog->shutdown(); | |
830 | ||
831 | // shut down cache | |
832 | mdcache->shutdown(); | |
833 | ||
834 | purge_queue.shutdown(); | |
835 | ||
836 | mds_lock.Unlock(); | |
837 | finisher->stop(); // no flushing | |
838 | mds_lock.Lock(); | |
839 | ||
840 | if (objecter->initialized) | |
841 | objecter->shutdown(); | |
842 | ||
843 | monc->shutdown(); | |
844 | ||
845 | op_tracker.on_shutdown(); | |
846 | ||
847 | progress_thread.shutdown(); | |
848 | ||
849 | // release mds_lock for finisher/messenger threads (e.g. | |
850 | // MDSDaemon::ms_handle_reset called from Messenger). | |
851 | mds_lock.Unlock(); | |
852 | ||
853 | // shut down messenger | |
854 | messenger->shutdown(); | |
855 | ||
856 | mds_lock.Lock(); | |
857 | ||
858 | // Workaround unclean shutdown: HeartbeatMap will assert if | |
859 | // worker is not removed (as we do in ~MDS), but ~MDS is not | |
860 | // always called after suicide. | |
861 | if (hb) { | |
862 | g_ceph_context->get_heartbeat_map()->remove_worker(hb); | |
863 | hb = NULL; | |
864 | } | |
865 | } | |
866 | ||
867 | /** | |
868 | * Helper for simple callbacks that call a void fn with no args. | |
869 | */ | |
870 | class C_MDS_VoidFn : public MDSInternalContext | |
871 | { | |
872 | typedef void (MDSRank::*fn_ptr)(); | |
873 | protected: | |
874 | fn_ptr fn; | |
875 | public: | |
876 | C_MDS_VoidFn(MDSRank *mds_, fn_ptr fn_) | |
877 | : MDSInternalContext(mds_), fn(fn_) | |
878 | { | |
879 | ceph_assert(mds_); | |
880 | ceph_assert(fn_); | |
881 | } | |
882 | ||
883 | void finish(int r) override | |
884 | { | |
885 | (mds->*fn)(); | |
886 | } | |
887 | }; | |
888 | ||
889 | int64_t MDSRank::get_metadata_pool() | |
890 | { | |
891 | return mdsmap->get_metadata_pool(); | |
892 | } | |
893 | ||
894 | MDSTableClient *MDSRank::get_table_client(int t) | |
895 | { | |
896 | switch (t) { | |
897 | case TABLE_ANCHOR: return NULL; | |
898 | case TABLE_SNAP: return snapclient; | |
899 | default: ceph_abort(); | |
900 | } | |
901 | } | |
902 | ||
903 | MDSTableServer *MDSRank::get_table_server(int t) | |
904 | { | |
905 | switch (t) { | |
906 | case TABLE_ANCHOR: return NULL; | |
907 | case TABLE_SNAP: return snapserver; | |
908 | default: ceph_abort(); | |
909 | } | |
910 | } | |
911 | ||
912 | void MDSRank::suicide() | |
913 | { | |
914 | if (suicide_hook) { | |
915 | suicide_hook->complete(0); | |
916 | suicide_hook = NULL; | |
917 | } | |
918 | } | |
919 | ||
920 | void MDSRank::respawn() | |
921 | { | |
922 | if (respawn_hook) { | |
923 | respawn_hook->complete(0); | |
924 | respawn_hook = NULL; | |
925 | } | |
926 | } | |
927 | ||
928 | void MDSRank::damaged() | |
929 | { | |
930 | ceph_assert(whoami != MDS_RANK_NONE); | |
931 | ceph_assert(mds_lock.is_locked_by_me()); | |
932 | ||
933 | beacon.set_want_state(*mdsmap, MDSMap::STATE_DAMAGED); | |
934 | monc->flush_log(); // Flush any clog error from before we were called | |
935 | beacon.notify_health(this); // Include latest status in our swan song | |
936 | beacon.send_and_wait(g_conf()->mds_mon_shutdown_timeout); | |
937 | ||
938 | // It's okay if we timed out and the mon didn't get our beacon, because | |
939 | // another daemon (or ourselves after respawn) will eventually take the | |
940 | // rank and report DAMAGED again when it hits same problem we did. | |
941 | ||
942 | respawn(); // Respawn into standby in case mon has other work for us | |
943 | } | |
944 | ||
945 | void MDSRank::damaged_unlocked() | |
946 | { | |
947 | std::lock_guard l(mds_lock); | |
948 | damaged(); | |
949 | } | |
950 | ||
951 | void MDSRank::handle_write_error(int err) | |
952 | { | |
953 | if (err == -EBLACKLISTED) { | |
954 | derr << "we have been blacklisted (fenced), respawning..." << dendl; | |
955 | respawn(); | |
956 | return; | |
957 | } | |
958 | ||
959 | if (g_conf()->mds_action_on_write_error >= 2) { | |
960 | derr << "unhandled write error " << cpp_strerror(err) << ", suicide..." << dendl; | |
961 | respawn(); | |
962 | } else if (g_conf()->mds_action_on_write_error == 1) { | |
963 | derr << "unhandled write error " << cpp_strerror(err) << ", force readonly..." << dendl; | |
964 | mdcache->force_readonly(); | |
965 | } else { | |
966 | // ignore; | |
967 | derr << "unhandled write error " << cpp_strerror(err) << ", ignore..." << dendl; | |
968 | } | |
969 | } | |
970 | ||
971 | void *MDSRank::ProgressThread::entry() | |
972 | { | |
973 | std::lock_guard l(mds->mds_lock); | |
974 | while (true) { | |
975 | while (!mds->stopping && | |
976 | mds->finished_queue.empty() && | |
977 | (mds->waiting_for_nolaggy.empty() || mds->beacon.is_laggy())) { | |
978 | cond.Wait(mds->mds_lock); | |
979 | } | |
980 | ||
981 | if (mds->stopping) { | |
982 | break; | |
983 | } | |
984 | ||
985 | mds->_advance_queues(); | |
986 | } | |
987 | ||
988 | return NULL; | |
989 | } | |
990 | ||
991 | ||
992 | void MDSRank::ProgressThread::shutdown() | |
993 | { | |
994 | ceph_assert(mds->mds_lock.is_locked_by_me()); | |
995 | ceph_assert(mds->stopping); | |
996 | ||
997 | if (am_self()) { | |
998 | // Stopping is set, we will fall out of our main loop naturally | |
999 | } else { | |
1000 | // Kick the thread to notice mds->stopping, and join it | |
1001 | cond.Signal(); | |
1002 | mds->mds_lock.Unlock(); | |
1003 | if (is_started()) | |
1004 | join(); | |
1005 | mds->mds_lock.Lock(); | |
1006 | } | |
1007 | } | |
1008 | ||
1009 | bool MDSRankDispatcher::ms_dispatch(const Message::const_ref &m) | |
1010 | { | |
1011 | if (m->get_source().is_client()) { | |
1012 | Session *session = static_cast<Session*>(m->get_connection()->get_priv().get()); | |
1013 | if (session) | |
1014 | session->last_seen = Session::clock::now(); | |
1015 | } | |
1016 | ||
1017 | inc_dispatch_depth(); | |
1018 | bool ret = _dispatch(m, true); | |
1019 | dec_dispatch_depth(); | |
1020 | return ret; | |
1021 | } | |
1022 | ||
1023 | bool MDSRank::_dispatch(const Message::const_ref &m, bool new_msg) | |
1024 | { | |
1025 | if (is_stale_message(m)) { | |
1026 | return true; | |
1027 | } | |
1028 | ||
1029 | if (beacon.is_laggy()) { | |
1030 | dout(5) << " laggy, deferring " << *m << dendl; | |
1031 | waiting_for_nolaggy.push_back(m); | |
1032 | } else if (new_msg && !waiting_for_nolaggy.empty()) { | |
1033 | dout(5) << " there are deferred messages, deferring " << *m << dendl; | |
1034 | waiting_for_nolaggy.push_back(m); | |
1035 | } else { | |
1036 | if (!handle_deferrable_message(m)) { | |
1037 | return false; | |
1038 | } | |
1039 | ||
1040 | heartbeat_reset(); | |
1041 | } | |
1042 | ||
1043 | if (dispatch_depth > 1) | |
1044 | return true; | |
1045 | ||
1046 | // finish any triggered contexts | |
1047 | _advance_queues(); | |
1048 | ||
1049 | if (beacon.is_laggy()) { | |
1050 | // We've gone laggy during dispatch, don't do any | |
1051 | // more housekeeping | |
1052 | return true; | |
1053 | } | |
1054 | ||
1055 | // hack: thrash exports | |
1056 | static utime_t start; | |
1057 | utime_t now = ceph_clock_now(); | |
1058 | if (start == utime_t()) | |
1059 | start = now; | |
1060 | /*double el = now - start; | |
1061 | if (el > 30.0 && | |
1062 | el < 60.0)*/ | |
1063 | for (int i=0; i<g_conf()->mds_thrash_exports; i++) { | |
1064 | set<mds_rank_t> s; | |
1065 | if (!is_active()) break; | |
1066 | mdsmap->get_mds_set(s, MDSMap::STATE_ACTIVE); | |
1067 | if (s.size() < 2 || CInode::count() < 10) | |
1068 | break; // need peers for this to work. | |
1069 | if (mdcache->migrator->get_num_exporting() > g_conf()->mds_thrash_exports * 5 || | |
1070 | mdcache->migrator->get_export_queue_size() > g_conf()->mds_thrash_exports * 10) | |
1071 | break; | |
1072 | ||
1073 | dout(7) << "mds thrashing exports pass " << (i+1) << "/" << g_conf()->mds_thrash_exports << dendl; | |
1074 | ||
1075 | // pick a random dir inode | |
1076 | CInode *in = mdcache->hack_pick_random_inode(); | |
1077 | ||
1078 | list<CDir*> ls; | |
1079 | in->get_dirfrags(ls); | |
1080 | if (!ls.empty()) { // must be an open dir. | |
1081 | list<CDir*>::iterator p = ls.begin(); | |
1082 | int n = rand() % ls.size(); | |
1083 | while (n--) | |
1084 | ++p; | |
1085 | CDir *dir = *p; | |
1086 | if (!dir->get_parent_dir()) continue; // must be linked. | |
1087 | if (!dir->is_auth()) continue; // must be auth. | |
1088 | ||
1089 | mds_rank_t dest; | |
1090 | do { | |
1091 | int k = rand() % s.size(); | |
1092 | set<mds_rank_t>::iterator p = s.begin(); | |
1093 | while (k--) ++p; | |
1094 | dest = *p; | |
1095 | } while (dest == whoami); | |
1096 | mdcache->migrator->export_dir_nicely(dir,dest); | |
1097 | } | |
1098 | } | |
1099 | // hack: thrash fragments | |
1100 | for (int i=0; i<g_conf()->mds_thrash_fragments; i++) { | |
1101 | if (!is_active()) break; | |
1102 | if (mdcache->get_num_fragmenting_dirs() > 5 * g_conf()->mds_thrash_fragments) break; | |
1103 | dout(7) << "mds thrashing fragments pass " << (i+1) << "/" << g_conf()->mds_thrash_fragments << dendl; | |
1104 | ||
1105 | // pick a random dir inode | |
1106 | CInode *in = mdcache->hack_pick_random_inode(); | |
1107 | ||
1108 | list<CDir*> ls; | |
1109 | in->get_dirfrags(ls); | |
1110 | if (ls.empty()) continue; // must be an open dir. | |
1111 | CDir *dir = ls.front(); | |
1112 | if (!dir->get_parent_dir()) continue; // must be linked. | |
1113 | if (!dir->is_auth()) continue; // must be auth. | |
1114 | frag_t fg = dir->get_frag(); | |
1115 | if ((fg == frag_t() || (rand() % (1 << fg.bits()) == 0))) { | |
1116 | mdcache->split_dir(dir, 1); | |
1117 | } else { | |
1118 | balancer->queue_merge(dir); | |
1119 | } | |
1120 | } | |
1121 | ||
1122 | // hack: force hash root? | |
1123 | /* | |
1124 | if (false && | |
1125 | mdcache->get_root() && | |
1126 | mdcache->get_root()->dir && | |
1127 | !(mdcache->get_root()->dir->is_hashed() || | |
1128 | mdcache->get_root()->dir->is_hashing())) { | |
1129 | dout(0) << "hashing root" << dendl; | |
1130 | mdcache->migrator->hash_dir(mdcache->get_root()->dir); | |
1131 | } | |
1132 | */ | |
1133 | ||
1134 | update_mlogger(); | |
1135 | return true; | |
1136 | } | |
1137 | ||
1138 | void MDSRank::update_mlogger() | |
1139 | { | |
1140 | if (mlogger) { | |
1141 | mlogger->set(l_mdm_ino, CInode::count()); | |
1142 | mlogger->set(l_mdm_dir, CDir::count()); | |
1143 | mlogger->set(l_mdm_dn, CDentry::count()); | |
1144 | mlogger->set(l_mdm_cap, Capability::count()); | |
1145 | mlogger->set(l_mdm_inoa, CInode::increments()); | |
1146 | mlogger->set(l_mdm_inos, CInode::decrements()); | |
1147 | mlogger->set(l_mdm_dira, CDir::increments()); | |
1148 | mlogger->set(l_mdm_dirs, CDir::decrements()); | |
1149 | mlogger->set(l_mdm_dna, CDentry::increments()); | |
1150 | mlogger->set(l_mdm_dns, CDentry::decrements()); | |
1151 | mlogger->set(l_mdm_capa, Capability::increments()); | |
1152 | mlogger->set(l_mdm_caps, Capability::decrements()); | |
1153 | } | |
1154 | } | |
1155 | ||
1156 | /* | |
1157 | * lower priority messages we defer if we seem laggy | |
1158 | */ | |
1159 | bool MDSRank::handle_deferrable_message(const Message::const_ref &m) | |
1160 | { | |
1161 | int port = m->get_type() & 0xff00; | |
1162 | ||
1163 | switch (port) { | |
1164 | case MDS_PORT_CACHE: | |
1165 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
1166 | mdcache->dispatch(m); | |
1167 | break; | |
1168 | ||
1169 | case MDS_PORT_MIGRATOR: | |
1170 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
1171 | mdcache->migrator->dispatch(m); | |
1172 | break; | |
1173 | ||
1174 | default: | |
1175 | switch (m->get_type()) { | |
1176 | // SERVER | |
1177 | case CEPH_MSG_CLIENT_SESSION: | |
1178 | case CEPH_MSG_CLIENT_RECONNECT: | |
1179 | case CEPH_MSG_CLIENT_RECLAIM: | |
1180 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_CLIENT); | |
1181 | // fall-thru | |
1182 | case CEPH_MSG_CLIENT_REQUEST: | |
1183 | server->dispatch(m); | |
1184 | break; | |
1185 | case MSG_MDS_SLAVE_REQUEST: | |
1186 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
1187 | server->dispatch(m); | |
1188 | break; | |
1189 | ||
1190 | case MSG_MDS_HEARTBEAT: | |
1191 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
1192 | balancer->proc_message(m); | |
1193 | break; | |
1194 | ||
1195 | case MSG_MDS_TABLE_REQUEST: | |
1196 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
1197 | { | |
1198 | const MMDSTableRequest::const_ref &req = MMDSTableRequest::msgref_cast(m); | |
1199 | if (req->op < 0) { | |
1200 | MDSTableClient *client = get_table_client(req->table); | |
1201 | client->handle_request(req); | |
1202 | } else { | |
1203 | MDSTableServer *server = get_table_server(req->table); | |
1204 | server->handle_request(req); | |
1205 | } | |
1206 | } | |
1207 | break; | |
1208 | ||
1209 | case MSG_MDS_LOCK: | |
1210 | case MSG_MDS_INODEFILECAPS: | |
1211 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_MDS); | |
1212 | locker->dispatch(m); | |
1213 | break; | |
1214 | ||
1215 | case CEPH_MSG_CLIENT_CAPS: | |
1216 | case CEPH_MSG_CLIENT_CAPRELEASE: | |
1217 | case CEPH_MSG_CLIENT_LEASE: | |
1218 | ALLOW_MESSAGES_FROM(CEPH_ENTITY_TYPE_CLIENT); | |
1219 | locker->dispatch(m); | |
1220 | break; | |
1221 | ||
1222 | default: | |
1223 | return false; | |
1224 | } | |
1225 | } | |
1226 | ||
1227 | return true; | |
1228 | } | |
1229 | ||
1230 | /** | |
1231 | * Advance finished_queue and waiting_for_nolaggy. | |
1232 | * | |
1233 | * Usually drain both queues, but may not drain waiting_for_nolaggy | |
1234 | * if beacon is currently laggy. | |
1235 | */ | |
1236 | void MDSRank::_advance_queues() | |
1237 | { | |
1238 | ceph_assert(mds_lock.is_locked_by_me()); | |
1239 | ||
1240 | if (!finished_queue.empty()) { | |
1241 | dout(7) << "mds has " << finished_queue.size() << " queued contexts" << dendl; | |
1242 | while (!finished_queue.empty()) { | |
1243 | auto fin = finished_queue.front(); | |
1244 | finished_queue.pop_front(); | |
1245 | ||
1246 | dout(10) << " finish " << fin << dendl; | |
1247 | fin->complete(0); | |
1248 | ||
1249 | heartbeat_reset(); | |
1250 | } | |
1251 | } | |
1252 | ||
1253 | while (!waiting_for_nolaggy.empty()) { | |
1254 | // stop if we're laggy now! | |
1255 | if (beacon.is_laggy()) | |
1256 | break; | |
1257 | ||
1258 | Message::const_ref old = waiting_for_nolaggy.front(); | |
1259 | waiting_for_nolaggy.pop_front(); | |
1260 | ||
1261 | if (!is_stale_message(old)) { | |
1262 | dout(7) << " processing laggy deferred " << *old << dendl; | |
1263 | if (!handle_deferrable_message(old)) { | |
1264 | dout(0) << "unrecognized message " << *old << dendl; | |
1265 | } | |
1266 | } | |
1267 | ||
1268 | heartbeat_reset(); | |
1269 | } | |
1270 | } | |
1271 | ||
1272 | /** | |
1273 | * Call this when you take mds_lock, or periodically if you're going to | |
1274 | * hold the lock for a long time (e.g. iterating over clients/inodes) | |
1275 | */ | |
1276 | void MDSRank::heartbeat_reset() | |
1277 | { | |
1278 | // Any thread might jump into mds_lock and call us immediately | |
1279 | // after a call to suicide() completes, in which case MDSRank::hb | |
1280 | // has been freed and we are a no-op. | |
1281 | if (!hb) { | |
1282 | ceph_assert(stopping); | |
1283 | return; | |
1284 | } | |
1285 | ||
1286 | // NB not enabling suicide grace, because the mon takes care of killing us | |
1287 | // (by blacklisting us) when we fail to send beacons, and it's simpler to | |
1288 | // only have one way of dying. | |
1289 | auto grace = g_conf().get_val<double>("mds_heartbeat_grace"); | |
1290 | g_ceph_context->get_heartbeat_map()->reset_timeout(hb, grace, 0); | |
1291 | } | |
1292 | ||
1293 | bool MDSRank::is_stale_message(const Message::const_ref &m) const | |
1294 | { | |
1295 | // from bad mds? | |
1296 | if (m->get_source().is_mds()) { | |
1297 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
1298 | bool bad = false; | |
1299 | if (mdsmap->is_down(from)) { | |
1300 | bad = true; | |
1301 | } else { | |
1302 | // FIXME: this is a convoluted check. we should be maintaining a nice | |
1303 | // clean map of current ConnectionRefs for current mdses!!! | |
1304 | auto c = messenger->connect_to(CEPH_ENTITY_TYPE_MDS, | |
1305 | mdsmap->get_addrs(from)); | |
1306 | if (c != m->get_connection()) { | |
1307 | bad = true; | |
1308 | dout(5) << " mds." << from << " should be " << c << " " | |
1309 | << c->get_peer_addrs() << " but this message is " | |
1310 | << m->get_connection() << " " << m->get_source_addrs() | |
1311 | << dendl; | |
1312 | } | |
1313 | } | |
1314 | if (bad) { | |
1315 | // bogus mds? | |
1316 | if (m->get_type() == CEPH_MSG_MDS_MAP) { | |
1317 | dout(5) << "got " << *m << " from old/bad/imposter mds " << m->get_source() | |
1318 | << ", but it's an mdsmap, looking at it" << dendl; | |
1319 | } else if (m->get_type() == MSG_MDS_CACHEEXPIRE && | |
1320 | mdsmap->get_addrs(from) == m->get_source_addrs()) { | |
1321 | dout(5) << "got " << *m << " from down mds " << m->get_source() | |
1322 | << ", but it's a cache_expire, looking at it" << dendl; | |
1323 | } else { | |
1324 | dout(5) << "got " << *m << " from down/old/bad/imposter mds " << m->get_source() | |
1325 | << ", dropping" << dendl; | |
1326 | return true; | |
1327 | } | |
1328 | } | |
1329 | } | |
1330 | return false; | |
1331 | } | |
1332 | ||
1333 | Session *MDSRank::get_session(const Message::const_ref &m) | |
1334 | { | |
1335 | // do not carry ref | |
1336 | auto session = static_cast<Session *>(m->get_connection()->get_priv().get()); | |
1337 | if (session) { | |
1338 | dout(20) << "get_session have " << session << " " << session->info.inst | |
1339 | << " state " << session->get_state_name() << dendl; | |
1340 | // Check if we've imported an open session since (new sessions start closed) | |
1341 | if (session->is_closed()) { | |
1342 | Session *imported_session = sessionmap.get_session(session->info.inst.name); | |
1343 | if (imported_session && imported_session != session) { | |
1344 | dout(10) << __func__ << " replacing connection bootstrap session " | |
1345 | << session << " with imported session " << imported_session | |
1346 | << dendl; | |
1347 | imported_session->info.auth_name = session->info.auth_name; | |
1348 | //assert(session->info.auth_name == imported_session->info.auth_name); | |
1349 | ceph_assert(session->info.inst == imported_session->info.inst); | |
1350 | imported_session->set_connection(session->get_connection().get()); | |
1351 | // send out any queued messages | |
1352 | while (!session->preopen_out_queue.empty()) { | |
1353 | imported_session->get_connection()->send_message2(std::move(session->preopen_out_queue.front())); | |
1354 | session->preopen_out_queue.pop_front(); | |
1355 | } | |
1356 | imported_session->auth_caps = session->auth_caps; | |
1357 | imported_session->last_seen = session->last_seen; | |
1358 | ceph_assert(session->get_nref() == 1); | |
1359 | imported_session->get_connection()->set_priv(imported_session->get()); | |
1360 | session = imported_session; | |
1361 | } | |
1362 | } | |
1363 | } else { | |
1364 | dout(20) << "get_session dne for " << m->get_source_inst() << dendl; | |
1365 | } | |
1366 | return session; | |
1367 | } | |
1368 | ||
1369 | void MDSRank::send_message(const Message::ref& m, const ConnectionRef& c) | |
1370 | { | |
1371 | ceph_assert(c); | |
1372 | c->send_message2(m); | |
1373 | } | |
1374 | ||
1375 | ||
1376 | void MDSRank::send_message_mds(const Message::ref& m, mds_rank_t mds) | |
1377 | { | |
1378 | if (!mdsmap->is_up(mds)) { | |
1379 | dout(10) << "send_message_mds mds." << mds << " not up, dropping " << *m << dendl; | |
1380 | return; | |
1381 | } | |
1382 | ||
1383 | // send mdsmap first? | |
1384 | if (mds != whoami && peer_mdsmap_epoch[mds] < mdsmap->get_epoch()) { | |
1385 | auto _m = MMDSMap::create(monc->get_fsid(), *mdsmap); | |
1386 | messenger->send_to_mds(_m.detach(), mdsmap->get_addrs(mds)); | |
1387 | peer_mdsmap_epoch[mds] = mdsmap->get_epoch(); | |
1388 | } | |
1389 | ||
1390 | // send message | |
1391 | messenger->send_to_mds(Message::ref(m).detach(), mdsmap->get_addrs(mds)); | |
1392 | } | |
1393 | ||
1394 | void MDSRank::forward_message_mds(const MClientRequest::const_ref& m, mds_rank_t mds) | |
1395 | { | |
1396 | ceph_assert(mds != whoami); | |
1397 | ||
1398 | /* | |
1399 | * don't actually forward if non-idempotent! | |
1400 | * client has to do it. although the MDS will ignore duplicate requests, | |
1401 | * the affected metadata may migrate, in which case the new authority | |
1402 | * won't have the metareq_id in the completed request map. | |
1403 | */ | |
1404 | // NEW: always make the client resend! | |
1405 | bool client_must_resend = true; //!creq->can_forward(); | |
1406 | ||
1407 | // tell the client where it should go | |
1408 | auto session = get_session(m); | |
1409 | auto f = MClientRequestForward::create(m->get_tid(), mds, m->get_num_fwd()+1, client_must_resend); | |
1410 | send_message_client(f, session); | |
1411 | } | |
1412 | ||
1413 | void MDSRank::send_message_client_counted(const Message::ref& m, client_t client) | |
1414 | { | |
1415 | Session *session = sessionmap.get_session(entity_name_t::CLIENT(client.v)); | |
1416 | if (session) { | |
1417 | send_message_client_counted(m, session); | |
1418 | } else { | |
1419 | dout(10) << "send_message_client_counted no session for client." << client << " " << *m << dendl; | |
1420 | } | |
1421 | } | |
1422 | ||
1423 | void MDSRank::send_message_client_counted(const Message::ref& m, const ConnectionRef& connection) | |
1424 | { | |
1425 | // do not carry ref | |
1426 | auto session = static_cast<Session *>(connection->get_priv().get()); | |
1427 | if (session) { | |
1428 | send_message_client_counted(m, session); | |
1429 | } else { | |
1430 | dout(10) << "send_message_client_counted has no session for " << m->get_source_inst() << dendl; | |
1431 | // another Connection took over the Session | |
1432 | } | |
1433 | } | |
1434 | ||
1435 | void MDSRank::send_message_client_counted(const Message::ref& m, Session* session) | |
1436 | { | |
1437 | version_t seq = session->inc_push_seq(); | |
1438 | dout(10) << "send_message_client_counted " << session->info.inst.name << " seq " | |
1439 | << seq << " " << *m << dendl; | |
1440 | if (session->get_connection()) { | |
1441 | session->get_connection()->send_message2(m); | |
1442 | } else { | |
1443 | session->preopen_out_queue.push_back(m); | |
1444 | } | |
1445 | } | |
1446 | ||
1447 | void MDSRank::send_message_client(const Message::ref& m, Session* session) | |
1448 | { | |
1449 | dout(10) << "send_message_client " << session->info.inst << " " << *m << dendl; | |
1450 | if (session->get_connection()) { | |
1451 | session->get_connection()->send_message2(m); | |
1452 | } else { | |
1453 | session->preopen_out_queue.push_back(m); | |
1454 | } | |
1455 | } | |
1456 | ||
1457 | /** | |
1458 | * This is used whenever a RADOS operation has been cancelled | |
1459 | * or a RADOS client has been blacklisted, to cause the MDS and | |
1460 | * any clients to wait for this OSD epoch before using any new caps. | |
1461 | * | |
1462 | * See doc/cephfs/eviction | |
1463 | */ | |
1464 | void MDSRank::set_osd_epoch_barrier(epoch_t e) | |
1465 | { | |
1466 | dout(4) << __func__ << ": epoch=" << e << dendl; | |
1467 | osd_epoch_barrier = e; | |
1468 | } | |
1469 | ||
1470 | void MDSRank::retry_dispatch(const Message::const_ref &m) | |
1471 | { | |
1472 | inc_dispatch_depth(); | |
1473 | _dispatch(m, false); | |
1474 | dec_dispatch_depth(); | |
1475 | } | |
1476 | ||
1477 | double MDSRank::get_dispatch_queue_max_age(utime_t now) const | |
1478 | { | |
1479 | return messenger->get_dispatch_queue_max_age(now); | |
1480 | } | |
1481 | ||
1482 | bool MDSRank::is_daemon_stopping() const | |
1483 | { | |
1484 | return stopping; | |
1485 | } | |
1486 | ||
1487 | void MDSRank::request_state(MDSMap::DaemonState s) | |
1488 | { | |
1489 | dout(3) << "request_state " << ceph_mds_state_name(s) << dendl; | |
1490 | beacon.set_want_state(*mdsmap, s); | |
1491 | beacon.send(); | |
1492 | } | |
1493 | ||
1494 | ||
1495 | class C_MDS_BootStart : public MDSInternalContext { | |
1496 | MDSRank::BootStep nextstep; | |
1497 | public: | |
1498 | C_MDS_BootStart(MDSRank *m, MDSRank::BootStep n) | |
1499 | : MDSInternalContext(m), nextstep(n) {} | |
1500 | void finish(int r) override { | |
1501 | mds->boot_start(nextstep, r); | |
1502 | } | |
1503 | }; | |
1504 | ||
1505 | ||
1506 | void MDSRank::boot_start(BootStep step, int r) | |
1507 | { | |
1508 | // Handle errors from previous step | |
1509 | if (r < 0) { | |
1510 | if (is_standby_replay() && (r == -EAGAIN)) { | |
1511 | dout(0) << "boot_start encountered an error EAGAIN" | |
1512 | << ", respawning since we fell behind journal" << dendl; | |
1513 | respawn(); | |
1514 | } else if (r == -EINVAL || r == -ENOENT) { | |
1515 | // Invalid or absent data, indicates damaged on-disk structures | |
1516 | clog->error() << "Error loading MDS rank " << whoami << ": " | |
1517 | << cpp_strerror(r); | |
1518 | damaged(); | |
1519 | ceph_assert(r == 0); // Unreachable, damaged() calls respawn() | |
1520 | } else if (r == -EROFS) { | |
1521 | dout(0) << "boot error forcing transition to read-only; MDS will try to continue" << dendl; | |
1522 | } else { | |
1523 | // Completely unexpected error, give up and die | |
1524 | dout(0) << "boot_start encountered an error, failing" << dendl; | |
1525 | suicide(); | |
1526 | return; | |
1527 | } | |
1528 | } | |
1529 | ||
1530 | ceph_assert(is_starting() || is_any_replay()); | |
1531 | ||
1532 | switch(step) { | |
1533 | case MDS_BOOT_INITIAL: | |
1534 | { | |
1535 | mdcache->init_layouts(); | |
1536 | ||
1537 | MDSGatherBuilder gather(g_ceph_context, | |
1538 | new C_MDS_BootStart(this, MDS_BOOT_OPEN_ROOT)); | |
1539 | dout(2) << "Booting: " << step << ": opening inotable" << dendl; | |
1540 | inotable->set_rank(whoami); | |
1541 | inotable->load(gather.new_sub()); | |
1542 | ||
1543 | dout(2) << "Booting: " << step << ": opening sessionmap" << dendl; | |
1544 | sessionmap.set_rank(whoami); | |
1545 | sessionmap.load(gather.new_sub()); | |
1546 | ||
1547 | dout(2) << "Booting: " << step << ": opening mds log" << dendl; | |
1548 | mdlog->open(gather.new_sub()); | |
1549 | ||
1550 | if (is_starting()) { | |
1551 | dout(2) << "Booting: " << step << ": opening purge queue" << dendl; | |
1552 | purge_queue.open(new C_IO_Wrapper(this, gather.new_sub())); | |
1553 | } else if (!standby_replaying) { | |
1554 | dout(2) << "Booting: " << step << ": opening purge queue (async)" << dendl; | |
1555 | purge_queue.open(NULL); | |
1556 | dout(2) << "Booting: " << step << ": loading open file table (async)" << dendl; | |
1557 | mdcache->open_file_table.load(nullptr); | |
1558 | } | |
1559 | ||
1560 | if (mdsmap->get_tableserver() == whoami) { | |
1561 | dout(2) << "Booting: " << step << ": opening snap table" << dendl; | |
1562 | snapserver->set_rank(whoami); | |
1563 | snapserver->load(gather.new_sub()); | |
1564 | } | |
1565 | ||
1566 | gather.activate(); | |
1567 | } | |
1568 | break; | |
1569 | case MDS_BOOT_OPEN_ROOT: | |
1570 | { | |
1571 | dout(2) << "Booting: " << step << ": loading/discovering base inodes" << dendl; | |
1572 | ||
1573 | MDSGatherBuilder gather(g_ceph_context, | |
1574 | new C_MDS_BootStart(this, MDS_BOOT_PREPARE_LOG)); | |
1575 | ||
1576 | if (is_starting()) { | |
1577 | // load mydir frag for the first log segment (creating subtree map) | |
1578 | mdcache->open_mydir_frag(gather.new_sub()); | |
1579 | } else { | |
1580 | mdcache->open_mydir_inode(gather.new_sub()); | |
1581 | } | |
1582 | ||
1583 | mdcache->create_global_snaprealm(); | |
1584 | ||
1585 | if (whoami == mdsmap->get_root()) { // load root inode off disk if we are auth | |
1586 | mdcache->open_root_inode(gather.new_sub()); | |
1587 | } else if (is_any_replay()) { | |
1588 | // replay. make up fake root inode to start with | |
1589 | mdcache->create_root_inode(); | |
1590 | } | |
1591 | gather.activate(); | |
1592 | } | |
1593 | break; | |
1594 | case MDS_BOOT_PREPARE_LOG: | |
1595 | if (is_any_replay()) { | |
1596 | dout(2) << "Booting: " << step << ": replaying mds log" << dendl; | |
1597 | MDSGatherBuilder gather(g_ceph_context, | |
1598 | new C_MDS_BootStart(this, MDS_BOOT_REPLAY_DONE)); | |
1599 | ||
1600 | if (!standby_replaying) { | |
1601 | dout(2) << "Booting: " << step << ": waiting for purge queue recovered" << dendl; | |
1602 | purge_queue.wait_for_recovery(new C_IO_Wrapper(this, gather.new_sub())); | |
1603 | } | |
1604 | ||
1605 | mdlog->replay(gather.new_sub()); | |
1606 | gather.activate(); | |
1607 | } else { | |
1608 | dout(2) << "Booting: " << step << ": positioning at end of old mds log" << dendl; | |
1609 | mdlog->append(); | |
1610 | starting_done(); | |
1611 | } | |
1612 | break; | |
1613 | case MDS_BOOT_REPLAY_DONE: | |
1614 | ceph_assert(is_any_replay()); | |
1615 | ||
1616 | // Sessiontable and inotable should be in sync after replay, validate | |
1617 | // that they are consistent. | |
1618 | validate_sessions(); | |
1619 | ||
1620 | replay_done(); | |
1621 | break; | |
1622 | } | |
1623 | } | |
1624 | ||
1625 | void MDSRank::validate_sessions() | |
1626 | { | |
1627 | ceph_assert(mds_lock.is_locked_by_me()); | |
1628 | bool valid = true; | |
1629 | ||
1630 | // Identify any sessions which have state inconsistent with other, | |
1631 | // after they have been loaded from rados during startup. | |
1632 | // Mitigate bugs like: http://tracker.ceph.com/issues/16842 | |
1633 | for (const auto &i : sessionmap.get_sessions()) { | |
1634 | Session *session = i.second; | |
1635 | interval_set<inodeno_t> badones; | |
1636 | if (inotable->intersects_free(session->info.prealloc_inos, &badones)) { | |
1637 | clog->error() << "client " << *session | |
1638 | << "loaded with preallocated inodes that are inconsistent with inotable"; | |
1639 | valid = false; | |
1640 | } | |
1641 | } | |
1642 | ||
1643 | if (!valid) { | |
1644 | damaged(); | |
1645 | ceph_assert(valid); | |
1646 | } | |
1647 | } | |
1648 | ||
1649 | void MDSRank::starting_done() | |
1650 | { | |
1651 | dout(3) << "starting_done" << dendl; | |
1652 | ceph_assert(is_starting()); | |
1653 | request_state(MDSMap::STATE_ACTIVE); | |
1654 | ||
1655 | mdlog->start_new_segment(); | |
1656 | ||
1657 | // sync snaptable cache | |
1658 | snapclient->sync(new C_MDSInternalNoop); | |
1659 | } | |
1660 | ||
1661 | ||
1662 | void MDSRank::calc_recovery_set() | |
1663 | { | |
1664 | // initialize gather sets | |
1665 | set<mds_rank_t> rs; | |
1666 | mdsmap->get_recovery_mds_set(rs); | |
1667 | rs.erase(whoami); | |
1668 | mdcache->set_recovery_set(rs); | |
1669 | ||
1670 | dout(1) << " recovery set is " << rs << dendl; | |
1671 | } | |
1672 | ||
1673 | ||
1674 | void MDSRank::replay_start() | |
1675 | { | |
1676 | dout(1) << "replay_start" << dendl; | |
1677 | ||
1678 | if (is_standby_replay()) | |
1679 | standby_replaying = true; | |
1680 | ||
1681 | calc_recovery_set(); | |
1682 | ||
1683 | // Check if we need to wait for a newer OSD map before starting | |
1684 | Context *fin = new C_IO_Wrapper(this, new C_MDS_BootStart(this, MDS_BOOT_INITIAL)); | |
1685 | bool const ready = objecter->wait_for_map( | |
1686 | mdsmap->get_last_failure_osd_epoch(), | |
1687 | fin); | |
1688 | ||
1689 | if (ready) { | |
1690 | delete fin; | |
1691 | boot_start(); | |
1692 | } else { | |
1693 | dout(1) << " waiting for osdmap " << mdsmap->get_last_failure_osd_epoch() | |
1694 | << " (which blacklists prior instance)" << dendl; | |
1695 | } | |
1696 | } | |
1697 | ||
1698 | ||
1699 | class MDSRank::C_MDS_StandbyReplayRestartFinish : public MDSIOContext { | |
1700 | uint64_t old_read_pos; | |
1701 | public: | |
1702 | C_MDS_StandbyReplayRestartFinish(MDSRank *mds_, uint64_t old_read_pos_) : | |
1703 | MDSIOContext(mds_), old_read_pos(old_read_pos_) {} | |
1704 | void finish(int r) override { | |
1705 | mds->_standby_replay_restart_finish(r, old_read_pos); | |
1706 | } | |
1707 | void print(ostream& out) const override { | |
1708 | out << "standby_replay_restart"; | |
1709 | } | |
1710 | }; | |
1711 | ||
1712 | void MDSRank::_standby_replay_restart_finish(int r, uint64_t old_read_pos) | |
1713 | { | |
1714 | if (old_read_pos < mdlog->get_journaler()->get_trimmed_pos()) { | |
1715 | dout(0) << "standby MDS fell behind active MDS journal's expire_pos, restarting" << dendl; | |
1716 | respawn(); /* we're too far back, and this is easier than | |
1717 | trying to reset everything in the cache, etc */ | |
1718 | } else { | |
1719 | mdlog->standby_trim_segments(); | |
1720 | boot_start(MDS_BOOT_PREPARE_LOG, r); | |
1721 | } | |
1722 | } | |
1723 | ||
1724 | class MDSRank::C_MDS_StandbyReplayRestart : public MDSInternalContext { | |
1725 | public: | |
1726 | explicit C_MDS_StandbyReplayRestart(MDSRank *m) : MDSInternalContext(m) {} | |
1727 | void finish(int r) override { | |
1728 | ceph_assert(!r); | |
1729 | mds->standby_replay_restart(); | |
1730 | } | |
1731 | }; | |
1732 | ||
1733 | void MDSRank::standby_replay_restart() | |
1734 | { | |
1735 | if (standby_replaying) { | |
1736 | /* Go around for another pass of replaying in standby */ | |
1737 | dout(5) << "Restarting replay as standby-replay" << dendl; | |
1738 | mdlog->get_journaler()->reread_head_and_probe( | |
1739 | new C_MDS_StandbyReplayRestartFinish( | |
1740 | this, | |
1741 | mdlog->get_journaler()->get_read_pos())); | |
1742 | } else { | |
1743 | /* We are transitioning out of standby: wait for OSD map update | |
1744 | before making final pass */ | |
1745 | dout(1) << "standby_replay_restart (final takeover pass)" << dendl; | |
1746 | Context *fin = new C_IO_Wrapper(this, new C_MDS_StandbyReplayRestart(this)); | |
1747 | bool ready = objecter->wait_for_map(mdsmap->get_last_failure_osd_epoch(), fin); | |
1748 | if (ready) { | |
1749 | delete fin; | |
1750 | mdlog->get_journaler()->reread_head_and_probe( | |
1751 | new C_MDS_StandbyReplayRestartFinish( | |
1752 | this, | |
1753 | mdlog->get_journaler()->get_read_pos())); | |
1754 | ||
1755 | dout(1) << " opening purge_queue (async)" << dendl; | |
1756 | purge_queue.open(NULL); | |
1757 | dout(1) << " opening open_file_table (async)" << dendl; | |
1758 | mdcache->open_file_table.load(nullptr); | |
1759 | } else { | |
1760 | dout(1) << " waiting for osdmap " << mdsmap->get_last_failure_osd_epoch() | |
1761 | << " (which blacklists prior instance)" << dendl; | |
1762 | } | |
1763 | } | |
1764 | } | |
1765 | ||
1766 | void MDSRank::replay_done() | |
1767 | { | |
1768 | if (!standby_replaying) { | |
1769 | dout(1) << "Finished replaying journal" << dendl; | |
1770 | } else { | |
1771 | dout(5) << "Finished replaying journal as standby-replay" << dendl; | |
1772 | } | |
1773 | ||
1774 | if (is_standby_replay()) { | |
1775 | // The replay was done in standby state, and we are still in that state | |
1776 | ceph_assert(standby_replaying); | |
1777 | dout(10) << "setting replay timer" << dendl; | |
1778 | timer.add_event_after(g_conf()->mds_replay_interval, | |
1779 | new C_MDS_StandbyReplayRestart(this)); | |
1780 | return; | |
1781 | } else if (standby_replaying) { | |
1782 | // The replay was done in standby state, we have now _left_ that state | |
1783 | dout(10) << " last replay pass was as a standby; making final pass" << dendl; | |
1784 | standby_replaying = false; | |
1785 | standby_replay_restart(); | |
1786 | return; | |
1787 | } else { | |
1788 | // Replay is complete, journal read should be up to date | |
1789 | ceph_assert(mdlog->get_journaler()->get_read_pos() == mdlog->get_journaler()->get_write_pos()); | |
1790 | ceph_assert(!is_standby_replay()); | |
1791 | ||
1792 | // Reformat and come back here | |
1793 | if (mdlog->get_journaler()->get_stream_format() < g_conf()->mds_journal_format) { | |
1794 | dout(4) << "reformatting journal on standby-replay->replay transition" << dendl; | |
1795 | mdlog->reopen(new C_MDS_BootStart(this, MDS_BOOT_REPLAY_DONE)); | |
1796 | return; | |
1797 | } | |
1798 | } | |
1799 | ||
1800 | dout(1) << "making mds journal writeable" << dendl; | |
1801 | mdlog->get_journaler()->set_writeable(); | |
1802 | mdlog->get_journaler()->trim_tail(); | |
1803 | ||
1804 | if (mdsmap->get_tableserver() == whoami && | |
1805 | snapserver->upgrade_format()) { | |
1806 | dout(1) << "upgrading snaptable format" << dendl; | |
1807 | snapserver->save(new C_MDSInternalNoop); | |
1808 | } | |
1809 | ||
1810 | if (g_conf()->mds_wipe_sessions) { | |
1811 | dout(1) << "wiping out client sessions" << dendl; | |
1812 | sessionmap.wipe(); | |
1813 | sessionmap.save(new C_MDSInternalNoop); | |
1814 | } | |
1815 | if (g_conf()->mds_wipe_ino_prealloc) { | |
1816 | dout(1) << "wiping out ino prealloc from sessions" << dendl; | |
1817 | sessionmap.wipe_ino_prealloc(); | |
1818 | sessionmap.save(new C_MDSInternalNoop); | |
1819 | } | |
1820 | if (g_conf()->mds_skip_ino) { | |
1821 | inodeno_t i = g_conf()->mds_skip_ino; | |
1822 | dout(1) << "skipping " << i << " inodes" << dendl; | |
1823 | inotable->skip_inos(i); | |
1824 | inotable->save(new C_MDSInternalNoop); | |
1825 | } | |
1826 | ||
1827 | if (mdsmap->get_num_in_mds() == 1 && | |
1828 | mdsmap->get_num_failed_mds() == 0) { // just me! | |
1829 | dout(2) << "i am alone, moving to state reconnect" << dendl; | |
1830 | request_state(MDSMap::STATE_RECONNECT); | |
1831 | // sync snaptable cache | |
1832 | snapclient->sync(new C_MDSInternalNoop); | |
1833 | } else { | |
1834 | dout(2) << "i am not alone, moving to state resolve" << dendl; | |
1835 | request_state(MDSMap::STATE_RESOLVE); | |
1836 | } | |
1837 | } | |
1838 | ||
1839 | void MDSRank::reopen_log() | |
1840 | { | |
1841 | dout(1) << "reopen_log" << dendl; | |
1842 | mdcache->rollback_uncommitted_fragments(); | |
1843 | } | |
1844 | ||
1845 | void MDSRank::resolve_start() | |
1846 | { | |
1847 | dout(1) << "resolve_start" << dendl; | |
1848 | ||
1849 | reopen_log(); | |
1850 | ||
1851 | mdcache->resolve_start(new C_MDS_VoidFn(this, &MDSRank::resolve_done)); | |
1852 | finish_contexts(g_ceph_context, waiting_for_resolve); | |
1853 | } | |
1854 | ||
1855 | void MDSRank::resolve_done() | |
1856 | { | |
1857 | dout(1) << "resolve_done" << dendl; | |
1858 | request_state(MDSMap::STATE_RECONNECT); | |
1859 | // sync snaptable cache | |
1860 | snapclient->sync(new C_MDSInternalNoop); | |
1861 | } | |
1862 | ||
1863 | void MDSRank::reconnect_start() | |
1864 | { | |
1865 | dout(1) << "reconnect_start" << dendl; | |
1866 | ||
1867 | if (last_state == MDSMap::STATE_REPLAY) { | |
1868 | reopen_log(); | |
1869 | } | |
1870 | ||
1871 | // Drop any blacklisted clients from the SessionMap before going | |
1872 | // into reconnect, so that we don't wait for them. | |
1873 | objecter->enable_blacklist_events(); | |
1874 | std::set<entity_addr_t> blacklist; | |
1875 | epoch_t epoch = 0; | |
1876 | objecter->with_osdmap([&blacklist, &epoch](const OSDMap& o) { | |
1877 | o.get_blacklist(&blacklist); | |
1878 | epoch = o.get_epoch(); | |
1879 | }); | |
1880 | auto killed = server->apply_blacklist(blacklist); | |
1881 | dout(4) << "reconnect_start: killed " << killed << " blacklisted sessions (" | |
1882 | << blacklist.size() << " blacklist entries, " | |
1883 | << sessionmap.get_sessions().size() << ")" << dendl; | |
1884 | if (killed) { | |
1885 | set_osd_epoch_barrier(epoch); | |
1886 | } | |
1887 | ||
1888 | server->reconnect_clients(new C_MDS_VoidFn(this, &MDSRank::reconnect_done)); | |
1889 | finish_contexts(g_ceph_context, waiting_for_reconnect); | |
1890 | } | |
1891 | void MDSRank::reconnect_done() | |
1892 | { | |
1893 | dout(1) << "reconnect_done" << dendl; | |
1894 | request_state(MDSMap::STATE_REJOIN); // move to rejoin state | |
1895 | } | |
1896 | ||
1897 | void MDSRank::rejoin_joint_start() | |
1898 | { | |
1899 | dout(1) << "rejoin_joint_start" << dendl; | |
1900 | mdcache->rejoin_send_rejoins(); | |
1901 | } | |
1902 | void MDSRank::rejoin_start() | |
1903 | { | |
1904 | dout(1) << "rejoin_start" << dendl; | |
1905 | mdcache->rejoin_start(new C_MDS_VoidFn(this, &MDSRank::rejoin_done)); | |
1906 | finish_contexts(g_ceph_context, waiting_for_rejoin); | |
1907 | } | |
1908 | void MDSRank::rejoin_done() | |
1909 | { | |
1910 | dout(1) << "rejoin_done" << dendl; | |
1911 | mdcache->show_subtrees(); | |
1912 | mdcache->show_cache(); | |
1913 | ||
1914 | // funny case: is our cache empty? no subtrees? | |
1915 | if (!mdcache->is_subtrees()) { | |
1916 | if (whoami == 0) { | |
1917 | // The root should always have a subtree! | |
1918 | clog->error() << "No subtrees found for root MDS rank!"; | |
1919 | damaged(); | |
1920 | ceph_assert(mdcache->is_subtrees()); | |
1921 | } else { | |
1922 | dout(1) << " empty cache, no subtrees, leaving cluster" << dendl; | |
1923 | request_state(MDSMap::STATE_STOPPED); | |
1924 | } | |
1925 | return; | |
1926 | } | |
1927 | ||
1928 | if (replay_queue.empty() && !server->get_num_pending_reclaim()) { | |
1929 | request_state(MDSMap::STATE_ACTIVE); | |
1930 | } else { | |
1931 | replaying_requests_done = replay_queue.empty(); | |
1932 | request_state(MDSMap::STATE_CLIENTREPLAY); | |
1933 | } | |
1934 | } | |
1935 | ||
1936 | void MDSRank::clientreplay_start() | |
1937 | { | |
1938 | dout(1) << "clientreplay_start" << dendl; | |
1939 | finish_contexts(g_ceph_context, waiting_for_replay); // kick waiters | |
1940 | mdcache->start_files_to_recover(); | |
1941 | queue_one_replay(); | |
1942 | } | |
1943 | ||
1944 | bool MDSRank::queue_one_replay() | |
1945 | { | |
1946 | if (!replay_queue.empty()) { | |
1947 | queue_waiter(replay_queue.front()); | |
1948 | replay_queue.pop_front(); | |
1949 | return true; | |
1950 | } | |
1951 | if (!replaying_requests_done) { | |
1952 | replaying_requests_done = true; | |
1953 | mdlog->flush(); | |
1954 | } | |
1955 | maybe_clientreplay_done(); | |
1956 | return false; | |
1957 | } | |
1958 | ||
1959 | void MDSRank::maybe_clientreplay_done() | |
1960 | { | |
1961 | if (is_clientreplay() && get_want_state() == MDSMap::STATE_CLIENTREPLAY) { | |
1962 | ||
1963 | // don't go to active if there are session waiting for being reclaimed | |
1964 | if (replaying_requests_done && !server->get_num_pending_reclaim()) { | |
1965 | mdlog->wait_for_safe(new C_MDS_VoidFn(this, &MDSRank::clientreplay_done)); | |
1966 | return; | |
1967 | } | |
1968 | ||
1969 | dout(1) << " still have " << replay_queue.size() + (int)!replaying_requests_done | |
1970 | << " requests need to be replayed, " << server->get_num_pending_reclaim() | |
1971 | << " sessions need to be reclaimed" << dendl; | |
1972 | } | |
1973 | } | |
1974 | ||
1975 | void MDSRank::clientreplay_done() | |
1976 | { | |
1977 | dout(1) << "clientreplay_done" << dendl; | |
1978 | request_state(MDSMap::STATE_ACTIVE); | |
1979 | } | |
1980 | ||
1981 | void MDSRank::active_start() | |
1982 | { | |
1983 | dout(1) << "active_start" << dendl; | |
1984 | ||
1985 | if (last_state == MDSMap::STATE_CREATING || | |
1986 | last_state == MDSMap::STATE_STARTING) { | |
1987 | mdcache->open_root(); | |
1988 | } | |
1989 | ||
1990 | mdcache->clean_open_file_lists(); | |
1991 | mdcache->export_remaining_imported_caps(); | |
1992 | finish_contexts(g_ceph_context, waiting_for_replay); // kick waiters | |
1993 | mdcache->start_files_to_recover(); | |
1994 | ||
1995 | mdcache->reissue_all_caps(); | |
1996 | ||
1997 | finish_contexts(g_ceph_context, waiting_for_active); // kick waiters | |
1998 | } | |
1999 | ||
2000 | void MDSRank::recovery_done(int oldstate) | |
2001 | { | |
2002 | dout(1) << "recovery_done -- successful recovery!" << dendl; | |
2003 | ceph_assert(is_clientreplay() || is_active()); | |
2004 | ||
2005 | if (oldstate == MDSMap::STATE_CREATING) | |
2006 | return; | |
2007 | ||
2008 | mdcache->start_recovered_truncates(); | |
2009 | mdcache->do_file_recover(); | |
2010 | ||
2011 | // tell connected clients | |
2012 | //bcast_mds_map(); // not anymore, they get this from the monitor | |
2013 | ||
2014 | mdcache->populate_mydir(); | |
2015 | } | |
2016 | ||
2017 | void MDSRank::creating_done() | |
2018 | { | |
2019 | dout(1)<< "creating_done" << dendl; | |
2020 | request_state(MDSMap::STATE_ACTIVE); | |
2021 | // sync snaptable cache | |
2022 | snapclient->sync(new C_MDSInternalNoop); | |
2023 | } | |
2024 | ||
2025 | void MDSRank::boot_create() | |
2026 | { | |
2027 | dout(3) << "boot_create" << dendl; | |
2028 | ||
2029 | MDSGatherBuilder fin(g_ceph_context, new C_MDS_VoidFn(this, &MDSRank::creating_done)); | |
2030 | ||
2031 | mdcache->init_layouts(); | |
2032 | ||
2033 | inotable->set_rank(whoami); | |
2034 | sessionmap.set_rank(whoami); | |
2035 | ||
2036 | // start with a fresh journal | |
2037 | dout(10) << "boot_create creating fresh journal" << dendl; | |
2038 | mdlog->create(fin.new_sub()); | |
2039 | ||
2040 | // open new journal segment, but do not journal subtree map (yet) | |
2041 | mdlog->prepare_new_segment(); | |
2042 | ||
2043 | if (whoami == mdsmap->get_root()) { | |
2044 | dout(3) << "boot_create creating fresh hierarchy" << dendl; | |
2045 | mdcache->create_empty_hierarchy(fin.get()); | |
2046 | } | |
2047 | ||
2048 | dout(3) << "boot_create creating mydir hierarchy" << dendl; | |
2049 | mdcache->create_mydir_hierarchy(fin.get()); | |
2050 | ||
2051 | dout(3) << "boot_create creating global snaprealm" << dendl; | |
2052 | mdcache->create_global_snaprealm(); | |
2053 | ||
2054 | // fixme: fake out inotable (reset, pretend loaded) | |
2055 | dout(10) << "boot_create creating fresh inotable table" << dendl; | |
2056 | inotable->reset(); | |
2057 | inotable->save(fin.new_sub()); | |
2058 | ||
2059 | // write empty sessionmap | |
2060 | sessionmap.save(fin.new_sub()); | |
2061 | ||
2062 | // Create empty purge queue | |
2063 | purge_queue.create(new C_IO_Wrapper(this, fin.new_sub())); | |
2064 | ||
2065 | // initialize tables | |
2066 | if (mdsmap->get_tableserver() == whoami) { | |
2067 | dout(10) << "boot_create creating fresh snaptable" << dendl; | |
2068 | snapserver->set_rank(whoami); | |
2069 | snapserver->reset(); | |
2070 | snapserver->save(fin.new_sub()); | |
2071 | } | |
2072 | ||
2073 | ceph_assert(g_conf()->mds_kill_create_at != 1); | |
2074 | ||
2075 | // ok now journal it | |
2076 | mdlog->journal_segment_subtree_map(fin.new_sub()); | |
2077 | mdlog->flush(); | |
2078 | ||
2079 | // Usually we do this during reconnect, but creation skips that. | |
2080 | objecter->enable_blacklist_events(); | |
2081 | ||
2082 | fin.activate(); | |
2083 | } | |
2084 | ||
2085 | void MDSRank::stopping_start() | |
2086 | { | |
2087 | dout(2) << "Stopping..." << dendl; | |
2088 | ||
2089 | if (mdsmap->get_num_in_mds() == 1 && !sessionmap.empty()) { | |
2090 | std::vector<Session*> victims; | |
2091 | const auto& sessions = sessionmap.get_sessions(); | |
2092 | for (const auto& p : sessions) { | |
2093 | if (!p.first.is_client()) { | |
2094 | continue; | |
2095 | } | |
2096 | ||
2097 | Session *s = p.second; | |
2098 | victims.push_back(s); | |
2099 | } | |
2100 | ||
2101 | dout(20) << __func__ << " matched " << victims.size() << " sessions" << dendl; | |
2102 | ceph_assert(!victims.empty()); | |
2103 | ||
2104 | C_GatherBuilder gather(g_ceph_context, new C_MDSInternalNoop); | |
2105 | for (const auto &s : victims) { | |
2106 | std::stringstream ss; | |
2107 | evict_client(s->get_client().v, false, | |
2108 | g_conf()->mds_session_blacklist_on_evict, ss, gather.new_sub()); | |
2109 | } | |
2110 | gather.activate(); | |
2111 | } | |
2112 | ||
2113 | mdcache->shutdown_start(); | |
2114 | } | |
2115 | ||
2116 | void MDSRank::stopping_done() | |
2117 | { | |
2118 | dout(2) << "Finished stopping..." << dendl; | |
2119 | ||
2120 | // tell monitor we shut down cleanly. | |
2121 | request_state(MDSMap::STATE_STOPPED); | |
2122 | } | |
2123 | ||
2124 | void MDSRankDispatcher::handle_mds_map( | |
2125 | const MMDSMap::const_ref &m, | |
2126 | const MDSMap &oldmap) | |
2127 | { | |
2128 | // I am only to be passed MDSMaps in which I hold a rank | |
2129 | ceph_assert(whoami != MDS_RANK_NONE); | |
2130 | ||
2131 | MDSMap::DaemonState oldstate = state; | |
2132 | mds_gid_t mds_gid = mds_gid_t(monc->get_global_id()); | |
2133 | state = mdsmap->get_state_gid(mds_gid); | |
2134 | if (state != oldstate) { | |
2135 | last_state = oldstate; | |
2136 | incarnation = mdsmap->get_inc_gid(mds_gid); | |
2137 | } | |
2138 | ||
2139 | version_t epoch = m->get_epoch(); | |
2140 | ||
2141 | // note source's map version | |
2142 | if (m->get_source().is_mds() && | |
2143 | peer_mdsmap_epoch[mds_rank_t(m->get_source().num())] < epoch) { | |
2144 | dout(15) << " peer " << m->get_source() | |
2145 | << " has mdsmap epoch >= " << epoch | |
2146 | << dendl; | |
2147 | peer_mdsmap_epoch[mds_rank_t(m->get_source().num())] = epoch; | |
2148 | } | |
2149 | ||
2150 | // Validate state transitions while I hold a rank | |
2151 | if (!MDSMap::state_transition_valid(oldstate, state)) { | |
2152 | derr << "Invalid state transition " << ceph_mds_state_name(oldstate) | |
2153 | << "->" << ceph_mds_state_name(state) << dendl; | |
2154 | respawn(); | |
2155 | } | |
2156 | ||
2157 | if (oldstate != state) { | |
2158 | // update messenger. | |
2159 | if (state == MDSMap::STATE_STANDBY_REPLAY) { | |
2160 | dout(1) << "handle_mds_map i am now mds." << mds_gid << "." << incarnation | |
2161 | << " replaying mds." << whoami << "." << incarnation << dendl; | |
2162 | messenger->set_myname(entity_name_t::MDS(mds_gid)); | |
2163 | } else { | |
2164 | dout(1) << "handle_mds_map i am now mds." << whoami << "." << incarnation << dendl; | |
2165 | messenger->set_myname(entity_name_t::MDS(whoami)); | |
2166 | } | |
2167 | } | |
2168 | ||
2169 | // tell objecter my incarnation | |
2170 | if (objecter->get_client_incarnation() != incarnation) | |
2171 | objecter->set_client_incarnation(incarnation); | |
2172 | ||
2173 | if (oldmap.get_min_compat_client() != mdsmap->get_min_compat_client()) | |
2174 | server->update_required_client_features(); | |
2175 | ||
2176 | // for debug | |
2177 | if (g_conf()->mds_dump_cache_on_map) | |
2178 | mdcache->dump_cache(); | |
2179 | ||
2180 | cluster_degraded = mdsmap->is_degraded(); | |
2181 | ||
2182 | // mdsmap and oldmap can be discontinuous. failover might happen in the missing mdsmap. | |
2183 | // the 'restart' set tracks ranks that have restarted since the old mdsmap | |
2184 | set<mds_rank_t> restart; | |
2185 | // replaying mds does not communicate with other ranks | |
2186 | if (state >= MDSMap::STATE_RESOLVE) { | |
2187 | // did someone fail? | |
2188 | // new down? | |
2189 | set<mds_rank_t> olddown, down; | |
2190 | oldmap.get_down_mds_set(&olddown); | |
2191 | mdsmap->get_down_mds_set(&down); | |
2192 | for (const auto& r : down) { | |
2193 | if (oldmap.have_inst(r) && olddown.count(r) == 0) { | |
2194 | messenger->mark_down_addrs(oldmap.get_addrs(r)); | |
2195 | handle_mds_failure(r); | |
2196 | } | |
2197 | } | |
2198 | ||
2199 | // did someone fail? | |
2200 | // did their addr/inst change? | |
2201 | set<mds_rank_t> up; | |
2202 | mdsmap->get_up_mds_set(up); | |
2203 | for (const auto& r : up) { | |
2204 | auto& info = mdsmap->get_info(r); | |
2205 | if (oldmap.have_inst(r)) { | |
2206 | auto& oldinfo = oldmap.get_info(r); | |
2207 | if (info.inc != oldinfo.inc) { | |
2208 | messenger->mark_down_addrs(oldinfo.get_addrs()); | |
2209 | if (info.state == MDSMap::STATE_REPLAY || | |
2210 | info.state == MDSMap::STATE_RESOLVE) { | |
2211 | restart.insert(r); | |
2212 | handle_mds_failure(r); | |
2213 | } else { | |
2214 | ceph_assert(info.state == MDSMap::STATE_STARTING || | |
2215 | info.state == MDSMap::STATE_ACTIVE); | |
2216 | // -> stopped (missing) -> starting -> active | |
2217 | restart.insert(r); | |
2218 | mdcache->migrator->handle_mds_failure_or_stop(r); | |
2219 | if (mdsmap->get_tableserver() == whoami) | |
2220 | snapserver->handle_mds_failure_or_stop(r); | |
2221 | } | |
2222 | } | |
2223 | } else { | |
2224 | if (info.state == MDSMap::STATE_REPLAY || | |
2225 | info.state == MDSMap::STATE_RESOLVE) { | |
2226 | // -> starting/creating (missing) -> active (missing) -> replay -> resolve | |
2227 | restart.insert(r); | |
2228 | handle_mds_failure(r); | |
2229 | } else { | |
2230 | ceph_assert(info.state == MDSMap::STATE_CREATING || | |
2231 | info.state == MDSMap::STATE_STARTING || | |
2232 | info.state == MDSMap::STATE_ACTIVE); | |
2233 | } | |
2234 | } | |
2235 | } | |
2236 | } | |
2237 | ||
2238 | // did it change? | |
2239 | if (oldstate != state) { | |
2240 | dout(1) << "handle_mds_map state change " | |
2241 | << ceph_mds_state_name(oldstate) << " --> " | |
2242 | << ceph_mds_state_name(state) << dendl; | |
2243 | beacon.set_want_state(*mdsmap, state); | |
2244 | ||
2245 | if (oldstate == MDSMap::STATE_STANDBY_REPLAY) { | |
2246 | dout(10) << "Monitor activated us! Deactivating replay loop" << dendl; | |
2247 | assert (state == MDSMap::STATE_REPLAY); | |
2248 | } else { | |
2249 | // did i just recover? | |
2250 | if ((is_active() || is_clientreplay()) && | |
2251 | (oldstate == MDSMap::STATE_CREATING || | |
2252 | oldstate == MDSMap::STATE_REJOIN || | |
2253 | oldstate == MDSMap::STATE_RECONNECT)) | |
2254 | recovery_done(oldstate); | |
2255 | ||
2256 | if (is_active()) { | |
2257 | active_start(); | |
2258 | } else if (is_any_replay()) { | |
2259 | replay_start(); | |
2260 | } else if (is_resolve()) { | |
2261 | resolve_start(); | |
2262 | } else if (is_reconnect()) { | |
2263 | reconnect_start(); | |
2264 | } else if (is_rejoin()) { | |
2265 | rejoin_start(); | |
2266 | } else if (is_clientreplay()) { | |
2267 | clientreplay_start(); | |
2268 | } else if (is_creating()) { | |
2269 | boot_create(); | |
2270 | } else if (is_starting()) { | |
2271 | boot_start(); | |
2272 | } else if (is_stopping()) { | |
2273 | ceph_assert(oldstate == MDSMap::STATE_ACTIVE); | |
2274 | stopping_start(); | |
2275 | } | |
2276 | } | |
2277 | } | |
2278 | ||
2279 | // RESOLVE | |
2280 | // is someone else newly resolving? | |
2281 | if (state >= MDSMap::STATE_RESOLVE) { | |
2282 | // recover snaptable | |
2283 | if (mdsmap->get_tableserver() == whoami) { | |
2284 | if (oldstate < MDSMap::STATE_RESOLVE) { | |
2285 | set<mds_rank_t> s; | |
2286 | mdsmap->get_mds_set_lower_bound(s, MDSMap::STATE_RESOLVE); | |
2287 | snapserver->finish_recovery(s); | |
2288 | } else { | |
2289 | set<mds_rank_t> old_set, new_set; | |
2290 | oldmap.get_mds_set_lower_bound(old_set, MDSMap::STATE_RESOLVE); | |
2291 | mdsmap->get_mds_set_lower_bound(new_set, MDSMap::STATE_RESOLVE); | |
2292 | for (const auto& r : new_set) { | |
2293 | if (r == whoami) | |
2294 | continue; // not me | |
2295 | if (!old_set.count(r) || restart.count(r)) { // newly so? | |
2296 | snapserver->handle_mds_recovery(r); | |
2297 | } | |
2298 | } | |
2299 | } | |
2300 | } | |
2301 | ||
2302 | if ((!oldmap.is_resolving() || !restart.empty()) && mdsmap->is_resolving()) { | |
2303 | set<mds_rank_t> resolve; | |
2304 | mdsmap->get_mds_set(resolve, MDSMap::STATE_RESOLVE); | |
2305 | dout(10) << " resolve set is " << resolve << dendl; | |
2306 | calc_recovery_set(); | |
2307 | mdcache->send_resolves(); | |
2308 | } | |
2309 | } | |
2310 | ||
2311 | // REJOIN | |
2312 | // is everybody finally rejoining? | |
2313 | if (state >= MDSMap::STATE_REJOIN) { | |
2314 | // did we start? | |
2315 | if (!oldmap.is_rejoining() && mdsmap->is_rejoining()) | |
2316 | rejoin_joint_start(); | |
2317 | ||
2318 | // did we finish? | |
2319 | if (g_conf()->mds_dump_cache_after_rejoin && | |
2320 | oldmap.is_rejoining() && !mdsmap->is_rejoining()) | |
2321 | mdcache->dump_cache(); // for DEBUG only | |
2322 | ||
2323 | if (oldstate >= MDSMap::STATE_REJOIN || | |
2324 | oldstate == MDSMap::STATE_STARTING) { | |
2325 | // ACTIVE|CLIENTREPLAY|REJOIN => we can discover from them. | |
2326 | set<mds_rank_t> olddis, dis; | |
2327 | oldmap.get_mds_set_lower_bound(olddis, MDSMap::STATE_REJOIN); | |
2328 | mdsmap->get_mds_set_lower_bound(dis, MDSMap::STATE_REJOIN); | |
2329 | for (const auto& r : dis) { | |
2330 | if (r == whoami) | |
2331 | continue; // not me | |
2332 | if (!olddis.count(r) || restart.count(r)) { // newly so? | |
2333 | mdcache->kick_discovers(r); | |
2334 | mdcache->kick_open_ino_peers(r); | |
2335 | } | |
2336 | } | |
2337 | } | |
2338 | } | |
2339 | ||
2340 | if (oldmap.is_degraded() && !cluster_degraded && state >= MDSMap::STATE_ACTIVE) { | |
2341 | dout(1) << "cluster recovered." << dendl; | |
2342 | auto it = waiting_for_active_peer.find(MDS_RANK_NONE); | |
2343 | if (it != waiting_for_active_peer.end()) { | |
2344 | queue_waiters(it->second); | |
2345 | waiting_for_active_peer.erase(it); | |
2346 | } | |
2347 | } | |
2348 | ||
2349 | // did someone go active? | |
2350 | if (state >= MDSMap::STATE_CLIENTREPLAY && | |
2351 | oldstate >= MDSMap::STATE_CLIENTREPLAY) { | |
2352 | set<mds_rank_t> oldactive, active; | |
2353 | oldmap.get_mds_set_lower_bound(oldactive, MDSMap::STATE_CLIENTREPLAY); | |
2354 | mdsmap->get_mds_set_lower_bound(active, MDSMap::STATE_CLIENTREPLAY); | |
2355 | for (const auto& r : active) { | |
2356 | if (r == whoami) | |
2357 | continue; // not me | |
2358 | if (!oldactive.count(r) || restart.count(r)) // newly so? | |
2359 | handle_mds_recovery(r); | |
2360 | } | |
2361 | } | |
2362 | ||
2363 | if (is_clientreplay() || is_active() || is_stopping()) { | |
2364 | // did anyone stop? | |
2365 | set<mds_rank_t> oldstopped, stopped; | |
2366 | oldmap.get_stopped_mds_set(oldstopped); | |
2367 | mdsmap->get_stopped_mds_set(stopped); | |
2368 | for (const auto& r : stopped) | |
2369 | if (oldstopped.count(r) == 0) { // newly so? | |
2370 | mdcache->migrator->handle_mds_failure_or_stop(r); | |
2371 | if (mdsmap->get_tableserver() == whoami) | |
2372 | snapserver->handle_mds_failure_or_stop(r); | |
2373 | } | |
2374 | } | |
2375 | ||
2376 | { | |
2377 | map<epoch_t,MDSContext::vec >::iterator p = waiting_for_mdsmap.begin(); | |
2378 | while (p != waiting_for_mdsmap.end() && p->first <= mdsmap->get_epoch()) { | |
2379 | MDSContext::vec ls; | |
2380 | ls.swap(p->second); | |
2381 | waiting_for_mdsmap.erase(p++); | |
2382 | queue_waiters(ls); | |
2383 | } | |
2384 | } | |
2385 | ||
2386 | if (is_active()) { | |
2387 | // Before going active, set OSD epoch barrier to latest (so that | |
2388 | // we don't risk handing out caps to clients with old OSD maps that | |
2389 | // might not include barriers from the previous incarnation of this MDS) | |
2390 | set_osd_epoch_barrier(objecter->with_osdmap( | |
2391 | std::mem_fn(&OSDMap::get_epoch))); | |
2392 | ||
2393 | /* Now check if we should hint to the OSD that a read may follow */ | |
2394 | if (mdsmap->has_standby_replay(whoami)) | |
2395 | mdlog->set_write_iohint(0); | |
2396 | else | |
2397 | mdlog->set_write_iohint(CEPH_OSD_OP_FLAG_FADVISE_DONTNEED); | |
2398 | } | |
2399 | ||
2400 | if (oldmap.get_max_mds() != mdsmap->get_max_mds()) { | |
2401 | purge_queue.update_op_limit(*mdsmap); | |
2402 | } | |
2403 | ||
2404 | if (scrubstack->is_scrubbing()) { | |
2405 | if (mdsmap->get_max_mds() > 1) { | |
2406 | auto c = new C_MDSInternalNoop; | |
2407 | scrubstack->scrub_abort(c); | |
2408 | } | |
2409 | } | |
2410 | mdcache->handle_mdsmap(*mdsmap); | |
2411 | } | |
2412 | ||
2413 | void MDSRank::handle_mds_recovery(mds_rank_t who) | |
2414 | { | |
2415 | dout(5) << "handle_mds_recovery mds." << who << dendl; | |
2416 | ||
2417 | mdcache->handle_mds_recovery(who); | |
2418 | ||
2419 | queue_waiters(waiting_for_active_peer[who]); | |
2420 | waiting_for_active_peer.erase(who); | |
2421 | } | |
2422 | ||
2423 | void MDSRank::handle_mds_failure(mds_rank_t who) | |
2424 | { | |
2425 | if (who == whoami) { | |
2426 | dout(5) << "handle_mds_failure for myself; not doing anything" << dendl; | |
2427 | return; | |
2428 | } | |
2429 | dout(5) << "handle_mds_failure mds." << who << dendl; | |
2430 | ||
2431 | mdcache->handle_mds_failure(who); | |
2432 | ||
2433 | if (mdsmap->get_tableserver() == whoami) | |
2434 | snapserver->handle_mds_failure_or_stop(who); | |
2435 | ||
2436 | snapclient->handle_mds_failure(who); | |
2437 | } | |
2438 | ||
2439 | bool MDSRankDispatcher::handle_asok_command(std::string_view command, | |
2440 | const cmdmap_t& cmdmap, | |
2441 | Formatter *f, | |
2442 | std::ostream& ss) | |
2443 | { | |
2444 | if (command == "dump_ops_in_flight" || | |
2445 | command == "ops") { | |
2446 | if (!op_tracker.dump_ops_in_flight(f)) { | |
2447 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
2448 | please enable \"mds_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
2449 | } | |
2450 | } else if (command == "dump_blocked_ops") { | |
2451 | if (!op_tracker.dump_ops_in_flight(f, true)) { | |
2452 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
2453 | Please enable \"mds_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
2454 | } | |
2455 | } else if (command == "dump_historic_ops") { | |
2456 | if (!op_tracker.dump_historic_ops(f)) { | |
2457 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
2458 | please enable \"mds_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
2459 | } | |
2460 | } else if (command == "dump_historic_ops_by_duration") { | |
2461 | if (!op_tracker.dump_historic_ops(f, true)) { | |
2462 | ss << "op_tracker tracking is not enabled now, so no ops are tracked currently, even those get stuck. \ | |
2463 | please enable \"mds_enable_op_tracker\", and the tracker will start to track new ops received afterwards."; | |
2464 | } | |
2465 | } else if (command == "osdmap barrier") { | |
2466 | int64_t target_epoch = 0; | |
2467 | bool got_val = cmd_getval(g_ceph_context, cmdmap, "target_epoch", target_epoch); | |
2468 | ||
2469 | if (!got_val) { | |
2470 | ss << "no target epoch given"; | |
2471 | return true; | |
2472 | } | |
2473 | ||
2474 | mds_lock.Lock(); | |
2475 | set_osd_epoch_barrier(target_epoch); | |
2476 | mds_lock.Unlock(); | |
2477 | ||
2478 | C_SaferCond cond; | |
2479 | bool already_got = objecter->wait_for_map(target_epoch, &cond); | |
2480 | if (!already_got) { | |
2481 | dout(4) << __func__ << ": waiting for OSD epoch " << target_epoch << dendl; | |
2482 | cond.wait(); | |
2483 | } | |
2484 | } else if (command == "session ls") { | |
2485 | std::lock_guard l(mds_lock); | |
2486 | ||
2487 | heartbeat_reset(); | |
2488 | ||
2489 | dump_sessions(SessionFilter(), f); | |
2490 | } else if (command == "session evict") { | |
2491 | std::string client_id; | |
2492 | const bool got_arg = cmd_getval(g_ceph_context, cmdmap, "client_id", client_id); | |
2493 | if(!got_arg) { | |
2494 | ss << "Invalid client_id specified"; | |
2495 | return true; | |
2496 | } | |
2497 | ||
2498 | mds_lock.Lock(); | |
2499 | std::stringstream dss; | |
2500 | bool evicted = evict_client(strtol(client_id.c_str(), 0, 10), true, | |
2501 | g_conf()->mds_session_blacklist_on_evict, dss); | |
2502 | if (!evicted) { | |
2503 | dout(15) << dss.str() << dendl; | |
2504 | ss << dss.str(); | |
2505 | } | |
2506 | mds_lock.Unlock(); | |
2507 | } else if (command == "session config") { | |
2508 | int64_t client_id; | |
2509 | std::string option; | |
2510 | std::string value; | |
2511 | ||
2512 | cmd_getval(g_ceph_context, cmdmap, "client_id", client_id); | |
2513 | cmd_getval(g_ceph_context, cmdmap, "option", option); | |
2514 | bool got_value = cmd_getval(g_ceph_context, cmdmap, "value", value); | |
2515 | ||
2516 | mds_lock.Lock(); | |
2517 | config_client(client_id, !got_value, option, value, ss); | |
2518 | mds_lock.Unlock(); | |
2519 | } else if (command == "scrub_path") { | |
2520 | string path; | |
2521 | vector<string> scrubop_vec; | |
2522 | cmd_getval(g_ceph_context, cmdmap, "scrubops", scrubop_vec); | |
2523 | cmd_getval(g_ceph_context, cmdmap, "path", path); | |
2524 | ||
2525 | /* Multiple MDS scrub is not currently supported. See also: https://tracker.ceph.com/issues/12274 */ | |
2526 | if (mdsmap->get_max_mds() > 1) { | |
2527 | ss << "Scrub is not currently supported for multiple active MDS. Please reduce max_mds to 1 and then scrub."; | |
2528 | return true; | |
2529 | } | |
2530 | ||
2531 | C_SaferCond cond; | |
2532 | command_scrub_start(f, path, "", scrubop_vec, &cond); | |
2533 | cond.wait(); | |
2534 | } else if (command == "tag path") { | |
2535 | string path; | |
2536 | cmd_getval(g_ceph_context, cmdmap, "path", path); | |
2537 | string tag; | |
2538 | cmd_getval(g_ceph_context, cmdmap, "tag", tag); | |
2539 | command_tag_path(f, path, tag); | |
2540 | } else if (command == "flush_path") { | |
2541 | string path; | |
2542 | cmd_getval(g_ceph_context, cmdmap, "path", path); | |
2543 | command_flush_path(f, path); | |
2544 | } else if (command == "flush journal") { | |
2545 | command_flush_journal(f); | |
2546 | } else if (command == "get subtrees") { | |
2547 | command_get_subtrees(f); | |
2548 | } else if (command == "export dir") { | |
2549 | string path; | |
2550 | if(!cmd_getval(g_ceph_context, cmdmap, "path", path)) { | |
2551 | ss << "malformed path"; | |
2552 | return true; | |
2553 | } | |
2554 | int64_t rank; | |
2555 | if(!cmd_getval(g_ceph_context, cmdmap, "rank", rank)) { | |
2556 | ss << "malformed rank"; | |
2557 | return true; | |
2558 | } | |
2559 | command_export_dir(f, path, (mds_rank_t)rank); | |
2560 | } else if (command == "dump cache") { | |
2561 | std::lock_guard l(mds_lock); | |
2562 | string path; | |
2563 | int r; | |
2564 | if(!cmd_getval(g_ceph_context, cmdmap, "path", path)) { | |
2565 | r = mdcache->dump_cache(f); | |
2566 | } else { | |
2567 | r = mdcache->dump_cache(path); | |
2568 | } | |
2569 | ||
2570 | if (r != 0) { | |
2571 | ss << "Failed to dump cache: " << cpp_strerror(r); | |
2572 | f->reset(); | |
2573 | } | |
2574 | } else if (command == "cache status") { | |
2575 | std::lock_guard l(mds_lock); | |
2576 | mdcache->cache_status(f); | |
2577 | } else if (command == "dump tree") { | |
2578 | command_dump_tree(cmdmap, ss, f); | |
2579 | } else if (command == "dump loads") { | |
2580 | std::lock_guard l(mds_lock); | |
2581 | int r = balancer->dump_loads(f); | |
2582 | if (r != 0) { | |
2583 | ss << "Failed to dump loads: " << cpp_strerror(r); | |
2584 | f->reset(); | |
2585 | } | |
2586 | } else if (command == "dump snaps") { | |
2587 | std::lock_guard l(mds_lock); | |
2588 | string server; | |
2589 | cmd_getval(g_ceph_context, cmdmap, "server", server); | |
2590 | if (server == "--server") { | |
2591 | if (mdsmap->get_tableserver() == whoami) { | |
2592 | snapserver->dump(f); | |
2593 | } else { | |
2594 | ss << "Not snapserver"; | |
2595 | } | |
2596 | } else { | |
2597 | int r = snapclient->dump_cache(f); | |
2598 | if (r != 0) { | |
2599 | ss << "Failed to dump snapclient: " << cpp_strerror(r); | |
2600 | f->reset(); | |
2601 | } | |
2602 | } | |
2603 | } else if (command == "force_readonly") { | |
2604 | std::lock_guard l(mds_lock); | |
2605 | mdcache->force_readonly(); | |
2606 | } else if (command == "dirfrag split") { | |
2607 | command_dirfrag_split(cmdmap, ss); | |
2608 | } else if (command == "dirfrag merge") { | |
2609 | command_dirfrag_merge(cmdmap, ss); | |
2610 | } else if (command == "dirfrag ls") { | |
2611 | command_dirfrag_ls(cmdmap, ss, f); | |
2612 | } else if (command == "openfiles ls") { | |
2613 | command_openfiles_ls(f); | |
2614 | } else if (command == "dump inode") { | |
2615 | command_dump_inode(f, cmdmap, ss); | |
2616 | } else { | |
2617 | return false; | |
2618 | } | |
2619 | ||
2620 | return true; | |
2621 | } | |
2622 | ||
2623 | class C_MDS_Send_Command_Reply : public MDSInternalContext { | |
2624 | protected: | |
2625 | MCommand::const_ref m; | |
2626 | public: | |
2627 | C_MDS_Send_Command_Reply(MDSRank *_mds, const MCommand::const_ref &_m) : | |
2628 | MDSInternalContext(_mds), m(_m) {} | |
2629 | ||
2630 | void send(int r, std::string_view ss) { | |
2631 | std::stringstream ds; | |
2632 | send(r, ss, ds); | |
2633 | } | |
2634 | ||
2635 | void send(int r, std::string_view ss, std::stringstream &ds) { | |
2636 | bufferlist bl; | |
2637 | bl.append(ds); | |
2638 | MDSDaemon::send_command_reply(m, mds, r, bl, ss); | |
2639 | } | |
2640 | ||
2641 | void finish(int r) override { | |
2642 | send(r, ""); | |
2643 | } | |
2644 | }; | |
2645 | ||
2646 | class C_ExecAndReply : public C_MDS_Send_Command_Reply { | |
2647 | public: | |
2648 | C_ExecAndReply(MDSRank *mds, const MCommand::const_ref &m) | |
2649 | : C_MDS_Send_Command_Reply(mds, m), f(true) { | |
2650 | } | |
2651 | ||
2652 | void finish(int r) override { | |
2653 | std::stringstream ds; | |
2654 | std::stringstream ss; | |
2655 | if (r != 0) { | |
2656 | f.flush(ss); | |
2657 | } else { | |
2658 | f.flush(ds); | |
2659 | } | |
2660 | ||
2661 | send(r, ss.str(), ds); | |
2662 | } | |
2663 | ||
2664 | virtual void exec() = 0; | |
2665 | ||
2666 | protected: | |
2667 | JSONFormatter f; | |
2668 | }; | |
2669 | ||
2670 | class C_CacheDropExecAndReply : public C_ExecAndReply { | |
2671 | public: | |
2672 | C_CacheDropExecAndReply(MDSRank *mds, const MCommand::const_ref &m, | |
2673 | uint64_t timeout) | |
2674 | : C_ExecAndReply(mds, m), timeout(timeout) { | |
2675 | } | |
2676 | ||
2677 | void exec() override { | |
2678 | mds->command_cache_drop(timeout, &f, this); | |
2679 | } | |
2680 | ||
2681 | private: | |
2682 | uint64_t timeout; | |
2683 | }; | |
2684 | ||
2685 | class C_ScrubExecAndReply : public C_ExecAndReply { | |
2686 | public: | |
2687 | C_ScrubExecAndReply(MDSRank *mds, const MCommand::const_ref &m, | |
2688 | const std::string &path, const std::string &tag, | |
2689 | const std::vector<std::string> &scrubop) | |
2690 | : C_ExecAndReply(mds, m), path(path), tag(tag), scrubop(scrubop) { | |
2691 | } | |
2692 | ||
2693 | void exec() override { | |
2694 | mds->command_scrub_start(&f, path, tag, scrubop, this); | |
2695 | } | |
2696 | ||
2697 | private: | |
2698 | std::string path; | |
2699 | std::string tag; | |
2700 | std::vector<std::string> scrubop; | |
2701 | }; | |
2702 | ||
2703 | class C_ScrubControlExecAndReply : public C_ExecAndReply { | |
2704 | public: | |
2705 | C_ScrubControlExecAndReply(MDSRank *mds, const MCommand::const_ref &m, | |
2706 | const std::string &command) | |
2707 | : C_ExecAndReply(mds, m), command(command) { | |
2708 | } | |
2709 | ||
2710 | void exec() override { | |
2711 | if (command == "abort") { | |
2712 | mds->command_scrub_abort(&f, this); | |
2713 | } else if (command == "pause") { | |
2714 | mds->command_scrub_pause(&f, this); | |
2715 | } else { | |
2716 | ceph_abort(); | |
2717 | } | |
2718 | } | |
2719 | ||
2720 | void finish(int r) override { | |
2721 | f.open_object_section("result"); | |
2722 | f.dump_int("return_code", r); | |
2723 | f.close_section(); | |
2724 | C_ExecAndReply::finish(r); | |
2725 | } | |
2726 | ||
2727 | private: | |
2728 | std::string command; | |
2729 | }; | |
2730 | ||
2731 | /** | |
2732 | * This function drops the mds_lock, so don't do anything with | |
2733 | * MDSRank after calling it (we could have gone into shutdown): just | |
2734 | * send your result back to the calling client and finish. | |
2735 | */ | |
2736 | void MDSRankDispatcher::evict_clients(const SessionFilter &filter, const MCommand::const_ref &m) | |
2737 | { | |
2738 | C_MDS_Send_Command_Reply *reply = new C_MDS_Send_Command_Reply(this, m); | |
2739 | ||
2740 | if (is_any_replay()) { | |
2741 | reply->send(-EAGAIN, "MDS is replaying log"); | |
2742 | delete reply; | |
2743 | return; | |
2744 | } | |
2745 | ||
2746 | std::vector<Session*> victims; | |
2747 | const auto& sessions = sessionmap.get_sessions(); | |
2748 | for (const auto& p : sessions) { | |
2749 | if (!p.first.is_client()) { | |
2750 | continue; | |
2751 | } | |
2752 | ||
2753 | Session *s = p.second; | |
2754 | ||
2755 | if (filter.match(*s, std::bind(&Server::waiting_for_reconnect, server, std::placeholders::_1))) { | |
2756 | victims.push_back(s); | |
2757 | } | |
2758 | } | |
2759 | ||
2760 | dout(20) << __func__ << " matched " << victims.size() << " sessions" << dendl; | |
2761 | ||
2762 | if (victims.empty()) { | |
2763 | reply->send(0, ""); | |
2764 | delete reply; | |
2765 | return; | |
2766 | } | |
2767 | ||
2768 | C_GatherBuilder gather(g_ceph_context, reply); | |
2769 | for (const auto s : victims) { | |
2770 | std::stringstream ss; | |
2771 | evict_client(s->get_client().v, false, | |
2772 | g_conf()->mds_session_blacklist_on_evict, ss, gather.new_sub()); | |
2773 | } | |
2774 | gather.activate(); | |
2775 | } | |
2776 | ||
2777 | void MDSRankDispatcher::dump_sessions(const SessionFilter &filter, Formatter *f) const | |
2778 | { | |
2779 | // Dump sessions, decorated with recovery/replay status | |
2780 | f->open_array_section("sessions"); | |
2781 | for (auto& [name, s] : sessionmap.get_sessions()) { | |
2782 | if (!name.is_client()) { | |
2783 | continue; | |
2784 | } | |
2785 | ||
2786 | if (!filter.match(*s, std::bind(&Server::waiting_for_reconnect, server, std::placeholders::_1))) { | |
2787 | continue; | |
2788 | } | |
2789 | ||
2790 | f->dump_object("session", *s); | |
2791 | } | |
2792 | f->close_section(); // sessions | |
2793 | } | |
2794 | ||
2795 | void MDSRank::command_scrub_start(Formatter *f, | |
2796 | std::string_view path, std::string_view tag, | |
2797 | const vector<string>& scrubop_vec, Context *on_finish) | |
2798 | { | |
2799 | bool force = false; | |
2800 | bool recursive = false; | |
2801 | bool repair = false; | |
2802 | for (auto &op : scrubop_vec) { | |
2803 | if (op == "force") | |
2804 | force = true; | |
2805 | else if (op == "recursive") | |
2806 | recursive = true; | |
2807 | else if (op == "repair") | |
2808 | repair = true; | |
2809 | } | |
2810 | ||
2811 | std::lock_guard l(mds_lock); | |
2812 | mdcache->enqueue_scrub(path, tag, force, recursive, repair, f, on_finish); | |
2813 | // scrub_dentry() finishers will dump the data for us; we're done! | |
2814 | } | |
2815 | ||
2816 | void MDSRank::command_tag_path(Formatter *f, | |
2817 | std::string_view path, std::string_view tag) | |
2818 | { | |
2819 | C_SaferCond scond; | |
2820 | { | |
2821 | std::lock_guard l(mds_lock); | |
2822 | mdcache->enqueue_scrub(path, tag, true, true, false, f, &scond); | |
2823 | } | |
2824 | scond.wait(); | |
2825 | } | |
2826 | ||
2827 | void MDSRank::command_scrub_abort(Formatter *f, Context *on_finish) { | |
2828 | std::lock_guard l(mds_lock); | |
2829 | scrubstack->scrub_abort(on_finish); | |
2830 | } | |
2831 | ||
2832 | void MDSRank::command_scrub_pause(Formatter *f, Context *on_finish) { | |
2833 | std::lock_guard l(mds_lock); | |
2834 | scrubstack->scrub_pause(on_finish); | |
2835 | } | |
2836 | ||
2837 | void MDSRank::command_scrub_resume(Formatter *f) { | |
2838 | int r = scrubstack->scrub_resume(); | |
2839 | ||
2840 | f->open_object_section("result"); | |
2841 | f->dump_int("return_code", r); | |
2842 | f->close_section(); | |
2843 | } | |
2844 | ||
2845 | void MDSRank::command_scrub_status(Formatter *f) { | |
2846 | scrubstack->scrub_status(f); | |
2847 | } | |
2848 | ||
2849 | void MDSRank::command_flush_path(Formatter *f, std::string_view path) | |
2850 | { | |
2851 | C_SaferCond scond; | |
2852 | { | |
2853 | std::lock_guard l(mds_lock); | |
2854 | mdcache->flush_dentry(path, &scond); | |
2855 | } | |
2856 | int r = scond.wait(); | |
2857 | f->open_object_section("results"); | |
2858 | f->dump_int("return_code", r); | |
2859 | f->close_section(); // results | |
2860 | } | |
2861 | ||
2862 | // synchronous wrapper around "journal flush" asynchronous context | |
2863 | // execution. | |
2864 | void MDSRank::command_flush_journal(Formatter *f) { | |
2865 | ceph_assert(f != NULL); | |
2866 | ||
2867 | C_SaferCond cond; | |
2868 | std::stringstream ss; | |
2869 | { | |
2870 | std::lock_guard locker(mds_lock); | |
2871 | C_Flush_Journal *flush_journal = new C_Flush_Journal(mdcache, mdlog, this, &ss, &cond); | |
2872 | flush_journal->send(); | |
2873 | } | |
2874 | int r = cond.wait(); | |
2875 | ||
2876 | f->open_object_section("result"); | |
2877 | f->dump_string("message", ss.str()); | |
2878 | f->dump_int("return_code", r); | |
2879 | f->close_section(); | |
2880 | } | |
2881 | ||
2882 | void MDSRank::command_get_subtrees(Formatter *f) | |
2883 | { | |
2884 | ceph_assert(f != NULL); | |
2885 | std::lock_guard l(mds_lock); | |
2886 | ||
2887 | std::vector<CDir*> subtrees; | |
2888 | mdcache->get_subtrees(subtrees); | |
2889 | ||
2890 | f->open_array_section("subtrees"); | |
2891 | for (const auto& dir : subtrees) { | |
2892 | f->open_object_section("subtree"); | |
2893 | { | |
2894 | f->dump_bool("is_auth", dir->is_auth()); | |
2895 | f->dump_int("auth_first", dir->get_dir_auth().first); | |
2896 | f->dump_int("auth_second", dir->get_dir_auth().second); | |
2897 | f->dump_int("export_pin", dir->inode->get_export_pin()); | |
2898 | f->open_object_section("dir"); | |
2899 | dir->dump(f); | |
2900 | f->close_section(); | |
2901 | } | |
2902 | f->close_section(); | |
2903 | } | |
2904 | f->close_section(); | |
2905 | } | |
2906 | ||
2907 | ||
2908 | void MDSRank::command_export_dir(Formatter *f, | |
2909 | std::string_view path, | |
2910 | mds_rank_t target) | |
2911 | { | |
2912 | int r = _command_export_dir(path, target); | |
2913 | f->open_object_section("results"); | |
2914 | f->dump_int("return_code", r); | |
2915 | f->close_section(); // results | |
2916 | } | |
2917 | ||
2918 | int MDSRank::_command_export_dir( | |
2919 | std::string_view path, | |
2920 | mds_rank_t target) | |
2921 | { | |
2922 | std::lock_guard l(mds_lock); | |
2923 | filepath fp(path); | |
2924 | ||
2925 | if (target == whoami || !mdsmap->is_up(target) || !mdsmap->is_in(target)) { | |
2926 | derr << "bad MDS target " << target << dendl; | |
2927 | return -ENOENT; | |
2928 | } | |
2929 | ||
2930 | CInode *in = mdcache->cache_traverse(fp); | |
2931 | if (!in) { | |
2932 | derr << "Bath path '" << path << "'" << dendl; | |
2933 | return -ENOENT; | |
2934 | } | |
2935 | CDir *dir = in->get_dirfrag(frag_t()); | |
2936 | if (!dir || !(dir->is_auth())) { | |
2937 | derr << "bad export_dir path dirfrag frag_t() or dir not auth" << dendl; | |
2938 | return -EINVAL; | |
2939 | } | |
2940 | ||
2941 | mdcache->migrator->export_dir(dir, target); | |
2942 | return 0; | |
2943 | } | |
2944 | ||
2945 | void MDSRank::command_dump_tree(const cmdmap_t &cmdmap, std::ostream &ss, Formatter *f) | |
2946 | { | |
2947 | std::string root; | |
2948 | int64_t depth; | |
2949 | cmd_getval(g_ceph_context, cmdmap, "root", root); | |
2950 | if (!cmd_getval(g_ceph_context, cmdmap, "depth", depth)) | |
2951 | depth = -1; | |
2952 | std::lock_guard l(mds_lock); | |
2953 | CInode *in = mdcache->cache_traverse(filepath(root.c_str())); | |
2954 | if (!in) { | |
2955 | ss << "root inode is not in cache"; | |
2956 | return; | |
2957 | } | |
2958 | f->open_array_section("inodes"); | |
2959 | mdcache->dump_tree(in, 0, depth, f); | |
2960 | f->close_section(); | |
2961 | } | |
2962 | ||
2963 | CDir *MDSRank::_command_dirfrag_get( | |
2964 | const cmdmap_t &cmdmap, | |
2965 | std::ostream &ss) | |
2966 | { | |
2967 | std::string path; | |
2968 | bool got = cmd_getval(g_ceph_context, cmdmap, "path", path); | |
2969 | if (!got) { | |
2970 | ss << "missing path argument"; | |
2971 | return NULL; | |
2972 | } | |
2973 | ||
2974 | std::string frag_str; | |
2975 | if (!cmd_getval(g_ceph_context, cmdmap, "frag", frag_str)) { | |
2976 | ss << "missing frag argument"; | |
2977 | return NULL; | |
2978 | } | |
2979 | ||
2980 | CInode *in = mdcache->cache_traverse(filepath(path.c_str())); | |
2981 | if (!in) { | |
2982 | // TODO really we should load something in if it's not in cache, | |
2983 | // but the infrastructure is harder, and we might still be unable | |
2984 | // to act on it if someone else is auth. | |
2985 | ss << "directory '" << path << "' inode not in cache"; | |
2986 | return NULL; | |
2987 | } | |
2988 | ||
2989 | frag_t fg; | |
2990 | ||
2991 | if (!fg.parse(frag_str.c_str())) { | |
2992 | ss << "frag " << frag_str << " failed to parse"; | |
2993 | return NULL; | |
2994 | } | |
2995 | ||
2996 | CDir *dir = in->get_dirfrag(fg); | |
2997 | if (!dir) { | |
2998 | ss << "frag " << in->ino() << "/" << fg << " not in cache (" | |
2999 | "use `dirfrag ls` to see if it should exist)"; | |
3000 | return NULL; | |
3001 | } | |
3002 | ||
3003 | if (!dir->is_auth()) { | |
3004 | ss << "frag " << dir->dirfrag() << " not auth (auth = " | |
3005 | << dir->authority() << ")"; | |
3006 | return NULL; | |
3007 | } | |
3008 | ||
3009 | return dir; | |
3010 | } | |
3011 | ||
3012 | bool MDSRank::command_dirfrag_split( | |
3013 | cmdmap_t cmdmap, | |
3014 | std::ostream &ss) | |
3015 | { | |
3016 | std::lock_guard l(mds_lock); | |
3017 | int64_t by = 0; | |
3018 | if (!cmd_getval(g_ceph_context, cmdmap, "bits", by)) { | |
3019 | ss << "missing bits argument"; | |
3020 | return false; | |
3021 | } | |
3022 | ||
3023 | if (by <= 0) { | |
3024 | ss << "must split by >0 bits"; | |
3025 | return false; | |
3026 | } | |
3027 | ||
3028 | CDir *dir = _command_dirfrag_get(cmdmap, ss); | |
3029 | if (!dir) { | |
3030 | return false; | |
3031 | } | |
3032 | ||
3033 | mdcache->split_dir(dir, by); | |
3034 | ||
3035 | return true; | |
3036 | } | |
3037 | ||
3038 | bool MDSRank::command_dirfrag_merge( | |
3039 | cmdmap_t cmdmap, | |
3040 | std::ostream &ss) | |
3041 | { | |
3042 | std::lock_guard l(mds_lock); | |
3043 | std::string path; | |
3044 | bool got = cmd_getval(g_ceph_context, cmdmap, "path", path); | |
3045 | if (!got) { | |
3046 | ss << "missing path argument"; | |
3047 | return false; | |
3048 | } | |
3049 | ||
3050 | std::string frag_str; | |
3051 | if (!cmd_getval(g_ceph_context, cmdmap, "frag", frag_str)) { | |
3052 | ss << "missing frag argument"; | |
3053 | return false; | |
3054 | } | |
3055 | ||
3056 | CInode *in = mdcache->cache_traverse(filepath(path.c_str())); | |
3057 | if (!in) { | |
3058 | ss << "directory '" << path << "' inode not in cache"; | |
3059 | return false; | |
3060 | } | |
3061 | ||
3062 | frag_t fg; | |
3063 | if (!fg.parse(frag_str.c_str())) { | |
3064 | ss << "frag " << frag_str << " failed to parse"; | |
3065 | return false; | |
3066 | } | |
3067 | ||
3068 | mdcache->merge_dir(in, fg); | |
3069 | ||
3070 | return true; | |
3071 | } | |
3072 | ||
3073 | bool MDSRank::command_dirfrag_ls( | |
3074 | cmdmap_t cmdmap, | |
3075 | std::ostream &ss, | |
3076 | Formatter *f) | |
3077 | { | |
3078 | std::lock_guard l(mds_lock); | |
3079 | std::string path; | |
3080 | bool got = cmd_getval(g_ceph_context, cmdmap, "path", path); | |
3081 | if (!got) { | |
3082 | ss << "missing path argument"; | |
3083 | return false; | |
3084 | } | |
3085 | ||
3086 | CInode *in = mdcache->cache_traverse(filepath(path.c_str())); | |
3087 | if (!in) { | |
3088 | ss << "directory inode not in cache"; | |
3089 | return false; | |
3090 | } | |
3091 | ||
3092 | f->open_array_section("frags"); | |
3093 | frag_vec_t leaves; | |
3094 | // NB using get_leaves_under instead of get_dirfrags to give | |
3095 | // you the list of what dirfrags may exist, not which are in cache | |
3096 | in->dirfragtree.get_leaves_under(frag_t(), leaves); | |
3097 | for (const auto& leaf : leaves) { | |
3098 | f->open_object_section("frag"); | |
3099 | f->dump_int("value", leaf.value()); | |
3100 | f->dump_int("bits", leaf.bits()); | |
3101 | CachedStackStringStream css; | |
3102 | *css << std::hex << leaf.value() << "/" << std::dec << leaf.bits(); | |
3103 | f->dump_string("str", css->strv()); | |
3104 | f->close_section(); | |
3105 | } | |
3106 | f->close_section(); | |
3107 | ||
3108 | return true; | |
3109 | } | |
3110 | ||
3111 | void MDSRank::command_openfiles_ls(Formatter *f) | |
3112 | { | |
3113 | std::lock_guard l(mds_lock); | |
3114 | mdcache->dump_openfiles(f); | |
3115 | } | |
3116 | ||
3117 | void MDSRank::command_dump_inode(Formatter *f, const cmdmap_t &cmdmap, std::ostream &ss) | |
3118 | { | |
3119 | std::lock_guard l(mds_lock); | |
3120 | int64_t number; | |
3121 | bool got = cmd_getval(g_ceph_context, cmdmap, "number", number); | |
3122 | if (!got) { | |
3123 | ss << "missing inode number"; | |
3124 | return; | |
3125 | } | |
3126 | ||
3127 | bool success = mdcache->dump_inode(f, number); | |
3128 | if (!success) { | |
3129 | ss << "dump inode failed, wrong inode number or the inode is not cached"; | |
3130 | } | |
3131 | } | |
3132 | ||
3133 | void MDSRank::dump_status(Formatter *f) const | |
3134 | { | |
3135 | if (state == MDSMap::STATE_REPLAY || | |
3136 | state == MDSMap::STATE_STANDBY_REPLAY) { | |
3137 | mdlog->dump_replay_status(f); | |
3138 | } else if (state == MDSMap::STATE_RESOLVE) { | |
3139 | mdcache->dump_resolve_status(f); | |
3140 | } else if (state == MDSMap::STATE_RECONNECT) { | |
3141 | server->dump_reconnect_status(f); | |
3142 | } else if (state == MDSMap::STATE_REJOIN) { | |
3143 | mdcache->dump_rejoin_status(f); | |
3144 | } else if (state == MDSMap::STATE_CLIENTREPLAY) { | |
3145 | dump_clientreplay_status(f); | |
3146 | } | |
3147 | f->dump_float("rank_uptime", get_uptime().count()); | |
3148 | } | |
3149 | ||
3150 | void MDSRank::dump_clientreplay_status(Formatter *f) const | |
3151 | { | |
3152 | f->open_object_section("clientreplay_status"); | |
3153 | f->dump_unsigned("clientreplay_queue", replay_queue.size()); | |
3154 | f->dump_unsigned("active_replay", mdcache->get_num_client_requests()); | |
3155 | f->close_section(); | |
3156 | } | |
3157 | ||
3158 | void MDSRankDispatcher::update_log_config() | |
3159 | { | |
3160 | map<string,string> log_to_monitors; | |
3161 | map<string,string> log_to_syslog; | |
3162 | map<string,string> log_channel; | |
3163 | map<string,string> log_prio; | |
3164 | map<string,string> log_to_graylog; | |
3165 | map<string,string> log_to_graylog_host; | |
3166 | map<string,string> log_to_graylog_port; | |
3167 | uuid_d fsid; | |
3168 | string host; | |
3169 | ||
3170 | if (parse_log_client_options(g_ceph_context, log_to_monitors, log_to_syslog, | |
3171 | log_channel, log_prio, log_to_graylog, | |
3172 | log_to_graylog_host, log_to_graylog_port, | |
3173 | fsid, host) == 0) | |
3174 | clog->update_config(log_to_monitors, log_to_syslog, | |
3175 | log_channel, log_prio, log_to_graylog, | |
3176 | log_to_graylog_host, log_to_graylog_port, | |
3177 | fsid, host); | |
3178 | dout(10) << __func__ << " log_to_monitors " << log_to_monitors << dendl; | |
3179 | } | |
3180 | ||
3181 | void MDSRank::create_logger() | |
3182 | { | |
3183 | dout(10) << "create_logger" << dendl; | |
3184 | { | |
3185 | PerfCountersBuilder mds_plb(g_ceph_context, "mds", l_mds_first, l_mds_last); | |
3186 | ||
3187 | // super useful (high prio) perf stats | |
3188 | mds_plb.add_u64_counter(l_mds_request, "request", "Requests", "req", | |
3189 | PerfCountersBuilder::PRIO_CRITICAL); | |
3190 | mds_plb.add_time_avg(l_mds_reply_latency, "reply_latency", "Reply latency", "rlat", | |
3191 | PerfCountersBuilder::PRIO_CRITICAL); | |
3192 | mds_plb.add_u64(l_mds_inodes, "inodes", "Inodes", "inos", | |
3193 | PerfCountersBuilder::PRIO_CRITICAL); | |
3194 | mds_plb.add_u64_counter(l_mds_forward, "forward", "Forwarding request", "fwd", | |
3195 | PerfCountersBuilder::PRIO_INTERESTING); | |
3196 | mds_plb.add_u64(l_mds_caps, "caps", "Capabilities", "caps", | |
3197 | PerfCountersBuilder::PRIO_INTERESTING); | |
3198 | mds_plb.add_u64_counter(l_mds_exported_inodes, "exported_inodes", "Exported inodes", | |
3199 | "exi", PerfCountersBuilder::PRIO_INTERESTING); | |
3200 | mds_plb.add_u64_counter(l_mds_imported_inodes, "imported_inodes", "Imported inodes", | |
3201 | "imi", PerfCountersBuilder::PRIO_INTERESTING); | |
3202 | ||
3203 | // useful dir/inode/subtree stats | |
3204 | mds_plb.set_prio_default(PerfCountersBuilder::PRIO_USEFUL); | |
3205 | mds_plb.add_u64(l_mds_root_rfiles, "root_rfiles", "root inode rfiles"); | |
3206 | mds_plb.add_u64(l_mds_root_rbytes, "root_rbytes", "root inode rbytes"); | |
3207 | mds_plb.add_u64(l_mds_root_rsnaps, "root_rsnaps", "root inode rsnaps"); | |
3208 | mds_plb.add_u64_counter(l_mds_dir_fetch, "dir_fetch", "Directory fetch"); | |
3209 | mds_plb.add_u64_counter(l_mds_dir_commit, "dir_commit", "Directory commit"); | |
3210 | mds_plb.add_u64_counter(l_mds_dir_split, "dir_split", "Directory split"); | |
3211 | mds_plb.add_u64_counter(l_mds_dir_merge, "dir_merge", "Directory merge"); | |
3212 | mds_plb.add_u64(l_mds_inode_max, "inode_max", "Max inodes, cache size"); | |
3213 | mds_plb.add_u64(l_mds_inodes_pinned, "inodes_pinned", "Inodes pinned"); | |
3214 | mds_plb.add_u64(l_mds_inodes_expired, "inodes_expired", "Inodes expired"); | |
3215 | mds_plb.add_u64(l_mds_inodes_with_caps, "inodes_with_caps", | |
3216 | "Inodes with capabilities"); | |
3217 | mds_plb.add_u64(l_mds_subtrees, "subtrees", "Subtrees"); | |
3218 | mds_plb.add_u64(l_mds_load_cent, "load_cent", "Load per cent"); | |
3219 | mds_plb.add_u64_counter(l_mds_openino_dir_fetch, "openino_dir_fetch", | |
3220 | "OpenIno incomplete directory fetchings"); | |
3221 | ||
3222 | // low prio stats | |
3223 | mds_plb.set_prio_default(PerfCountersBuilder::PRIO_DEBUGONLY); | |
3224 | mds_plb.add_u64_counter(l_mds_reply, "reply", "Replies"); | |
3225 | mds_plb.add_u64(l_mds_inodes_top, "inodes_top", "Inodes on top"); | |
3226 | mds_plb.add_u64(l_mds_inodes_bottom, "inodes_bottom", "Inodes on bottom"); | |
3227 | mds_plb.add_u64( | |
3228 | l_mds_inodes_pin_tail, "inodes_pin_tail", "Inodes on pin tail"); | |
3229 | mds_plb.add_u64_counter(l_mds_traverse, "traverse", "Traverses"); | |
3230 | mds_plb.add_u64_counter(l_mds_traverse_hit, "traverse_hit", "Traverse hits"); | |
3231 | mds_plb.add_u64_counter(l_mds_traverse_forward, "traverse_forward", | |
3232 | "Traverse forwards"); | |
3233 | mds_plb.add_u64_counter(l_mds_traverse_discover, "traverse_discover", | |
3234 | "Traverse directory discovers"); | |
3235 | mds_plb.add_u64_counter(l_mds_traverse_dir_fetch, "traverse_dir_fetch", | |
3236 | "Traverse incomplete directory content fetchings"); | |
3237 | mds_plb.add_u64_counter(l_mds_traverse_remote_ino, "traverse_remote_ino", | |
3238 | "Traverse remote dentries"); | |
3239 | mds_plb.add_u64_counter(l_mds_traverse_lock, "traverse_lock", | |
3240 | "Traverse locks"); | |
3241 | mds_plb.add_u64(l_mds_dispatch_queue_len, "q", "Dispatch queue length"); | |
3242 | mds_plb.add_u64_counter(l_mds_exported, "exported", "Exports"); | |
3243 | mds_plb.add_u64_counter(l_mds_imported, "imported", "Imports"); | |
3244 | mds_plb.add_u64_counter(l_mds_openino_backtrace_fetch, "openino_backtrace_fetch", | |
3245 | "OpenIno backtrace fetchings"); | |
3246 | mds_plb.add_u64_counter(l_mds_openino_peer_discover, "openino_peer_discover", | |
3247 | "OpenIno peer inode discovers"); | |
3248 | ||
3249 | logger = mds_plb.create_perf_counters(); | |
3250 | g_ceph_context->get_perfcounters_collection()->add(logger); | |
3251 | } | |
3252 | ||
3253 | { | |
3254 | PerfCountersBuilder mdm_plb(g_ceph_context, "mds_mem", l_mdm_first, l_mdm_last); | |
3255 | mdm_plb.add_u64(l_mdm_ino, "ino", "Inodes", "ino", | |
3256 | PerfCountersBuilder::PRIO_INTERESTING); | |
3257 | mdm_plb.add_u64(l_mdm_dn, "dn", "Dentries", "dn", | |
3258 | PerfCountersBuilder::PRIO_INTERESTING); | |
3259 | ||
3260 | mdm_plb.set_prio_default(PerfCountersBuilder::PRIO_USEFUL); | |
3261 | mdm_plb.add_u64_counter(l_mdm_inoa, "ino+", "Inodes opened"); | |
3262 | mdm_plb.add_u64_counter(l_mdm_inos, "ino-", "Inodes closed"); | |
3263 | mdm_plb.add_u64(l_mdm_dir, "dir", "Directories"); | |
3264 | mdm_plb.add_u64_counter(l_mdm_dira, "dir+", "Directories opened"); | |
3265 | mdm_plb.add_u64_counter(l_mdm_dirs, "dir-", "Directories closed"); | |
3266 | mdm_plb.add_u64_counter(l_mdm_dna, "dn+", "Dentries opened"); | |
3267 | mdm_plb.add_u64_counter(l_mdm_dns, "dn-", "Dentries closed"); | |
3268 | mdm_plb.add_u64(l_mdm_cap, "cap", "Capabilities"); | |
3269 | mdm_plb.add_u64_counter(l_mdm_capa, "cap+", "Capabilities added"); | |
3270 | mdm_plb.add_u64_counter(l_mdm_caps, "cap-", "Capabilities removed"); | |
3271 | mdm_plb.add_u64(l_mdm_heap, "heap", "Heap size"); | |
3272 | ||
3273 | mdm_plb.set_prio_default(PerfCountersBuilder::PRIO_DEBUGONLY); | |
3274 | mdm_plb.add_u64(l_mdm_rss, "rss", "RSS"); | |
3275 | ||
3276 | mlogger = mdm_plb.create_perf_counters(); | |
3277 | g_ceph_context->get_perfcounters_collection()->add(mlogger); | |
3278 | } | |
3279 | ||
3280 | mdlog->create_logger(); | |
3281 | server->create_logger(); | |
3282 | purge_queue.create_logger(); | |
3283 | sessionmap.register_perfcounters(); | |
3284 | mdcache->register_perfcounters(); | |
3285 | } | |
3286 | ||
3287 | void MDSRank::check_ops_in_flight() | |
3288 | { | |
3289 | string summary; | |
3290 | vector<string> warnings; | |
3291 | int slow = 0; | |
3292 | if (op_tracker.check_ops_in_flight(&summary, warnings, &slow)) { | |
3293 | clog->warn() << summary; | |
3294 | for (const auto& warning : warnings) { | |
3295 | clog->warn() << warning; | |
3296 | } | |
3297 | } | |
3298 | ||
3299 | // set mds slow request count | |
3300 | mds_slow_req_count = slow; | |
3301 | return; | |
3302 | } | |
3303 | ||
3304 | void MDSRankDispatcher::handle_osd_map() | |
3305 | { | |
3306 | if (is_active() && | |
3307 | mdsmap->get_tableserver() == whoami) { | |
3308 | snapserver->check_osd_map(true); | |
3309 | } | |
3310 | ||
3311 | server->handle_osd_map(); | |
3312 | ||
3313 | purge_queue.update_op_limit(*mdsmap); | |
3314 | ||
3315 | std::set<entity_addr_t> newly_blacklisted; | |
3316 | objecter->consume_blacklist_events(&newly_blacklisted); | |
3317 | auto epoch = objecter->with_osdmap([](const OSDMap &o){return o.get_epoch();}); | |
3318 | dout(4) << "handle_osd_map epoch " << epoch << ", " | |
3319 | << newly_blacklisted.size() << " new blacklist entries" << dendl; | |
3320 | auto victims = server->apply_blacklist(newly_blacklisted); | |
3321 | if (victims) { | |
3322 | set_osd_epoch_barrier(epoch); | |
3323 | } | |
3324 | ||
3325 | ||
3326 | // By default the objecter only requests OSDMap updates on use, | |
3327 | // we would like to always receive the latest maps in order to | |
3328 | // apply policy based on the FULL flag. | |
3329 | objecter->maybe_request_map(); | |
3330 | } | |
3331 | ||
3332 | int MDSRank::config_client(int64_t session_id, bool remove, | |
3333 | const std::string& option, const std::string& value, | |
3334 | std::ostream& ss) | |
3335 | { | |
3336 | Session *session = sessionmap.get_session(entity_name_t(CEPH_ENTITY_TYPE_CLIENT, session_id)); | |
3337 | if (!session) { | |
3338 | ss << "session " << session_id << " not in sessionmap!"; | |
3339 | return -ENOENT; | |
3340 | } | |
3341 | ||
3342 | if (option == "timeout") { | |
3343 | if (remove) { | |
3344 | auto it = session->info.client_metadata.find("timeout"); | |
3345 | if (it == session->info.client_metadata.end()) { | |
3346 | ss << "Nonexistent config: " << option; | |
3347 | return -ENODATA; | |
3348 | } | |
3349 | session->info.client_metadata.erase(it); | |
3350 | } else { | |
3351 | char *end; | |
3352 | strtoul(value.c_str(), &end, 0); | |
3353 | if (*end) { | |
3354 | ss << "Invalid config for timeout: " << value; | |
3355 | return -EINVAL; | |
3356 | } | |
3357 | session->info.client_metadata[option] = value; | |
3358 | } | |
3359 | //sessionmap._mark_dirty(session, true); | |
3360 | } else { | |
3361 | ss << "Invalid config option: " << option; | |
3362 | return -EINVAL; | |
3363 | } | |
3364 | ||
3365 | return 0; | |
3366 | } | |
3367 | ||
3368 | bool MDSRank::evict_client(int64_t session_id, | |
3369 | bool wait, bool blacklist, std::ostream& err_ss, | |
3370 | Context *on_killed) | |
3371 | { | |
3372 | ceph_assert(mds_lock.is_locked_by_me()); | |
3373 | ||
3374 | // Mutually exclusive args | |
3375 | ceph_assert(!(wait && on_killed != nullptr)); | |
3376 | ||
3377 | if (is_any_replay()) { | |
3378 | err_ss << "MDS is replaying log"; | |
3379 | return false; | |
3380 | } | |
3381 | ||
3382 | Session *session = sessionmap.get_session( | |
3383 | entity_name_t(CEPH_ENTITY_TYPE_CLIENT, session_id)); | |
3384 | if (!session) { | |
3385 | err_ss << "session " << session_id << " not in sessionmap!"; | |
3386 | return false; | |
3387 | } | |
3388 | ||
3389 | auto& addr = session->info.inst.addr; | |
3390 | { | |
3391 | CachedStackStringStream css; | |
3392 | *css << "Evicting " << (blacklist ? "(and blacklisting) " : "") | |
3393 | << "client session " << session_id << " (" << addr << ")"; | |
3394 | dout(1) << css->strv() << dendl; | |
3395 | clog->info() << css->strv(); | |
3396 | } | |
3397 | ||
3398 | dout(4) << "Preparing blacklist command... (wait=" << wait << ")" << dendl; | |
3399 | stringstream ss; | |
3400 | ss << "{\"prefix\":\"osd blacklist\", \"blacklistop\":\"add\","; | |
3401 | ss << "\"addr\":\""; | |
3402 | ss << addr; | |
3403 | ss << "\"}"; | |
3404 | std::string tmp = ss.str(); | |
3405 | std::vector<std::string> cmd = {tmp}; | |
3406 | ||
3407 | auto kill_client_session = [this, session_id, wait, on_killed](){ | |
3408 | ceph_assert(mds_lock.is_locked_by_me()); | |
3409 | Session *session = sessionmap.get_session( | |
3410 | entity_name_t(CEPH_ENTITY_TYPE_CLIENT, session_id)); | |
3411 | if (session) { | |
3412 | if (on_killed || !wait) { | |
3413 | server->kill_session(session, on_killed); | |
3414 | } else { | |
3415 | C_SaferCond on_safe; | |
3416 | server->kill_session(session, &on_safe); | |
3417 | ||
3418 | mds_lock.Unlock(); | |
3419 | on_safe.wait(); | |
3420 | mds_lock.Lock(); | |
3421 | } | |
3422 | } else { | |
3423 | dout(1) << "session " << session_id << " was removed while we waited " | |
3424 | "for blacklist" << dendl; | |
3425 | ||
3426 | // Even though it wasn't us that removed it, kick our completion | |
3427 | // as the session has been removed. | |
3428 | if (on_killed) { | |
3429 | on_killed->complete(0); | |
3430 | } | |
3431 | } | |
3432 | }; | |
3433 | ||
3434 | auto apply_blacklist = [this, cmd](std::function<void ()> fn){ | |
3435 | ceph_assert(mds_lock.is_locked_by_me()); | |
3436 | ||
3437 | Context *on_blacklist_done = new FunctionContext([this, fn](int r) { | |
3438 | objecter->wait_for_latest_osdmap( | |
3439 | new C_OnFinisher( | |
3440 | new FunctionContext([this, fn](int r) { | |
3441 | std::lock_guard l(mds_lock); | |
3442 | auto epoch = objecter->with_osdmap([](const OSDMap &o){ | |
3443 | return o.get_epoch(); | |
3444 | }); | |
3445 | ||
3446 | set_osd_epoch_barrier(epoch); | |
3447 | ||
3448 | fn(); | |
3449 | }), finisher) | |
3450 | ); | |
3451 | }); | |
3452 | ||
3453 | dout(4) << "Sending mon blacklist command: " << cmd[0] << dendl; | |
3454 | monc->start_mon_command(cmd, {}, nullptr, nullptr, on_blacklist_done); | |
3455 | }; | |
3456 | ||
3457 | if (wait) { | |
3458 | if (blacklist) { | |
3459 | C_SaferCond inline_ctx; | |
3460 | apply_blacklist([&inline_ctx](){inline_ctx.complete(0);}); | |
3461 | mds_lock.Unlock(); | |
3462 | inline_ctx.wait(); | |
3463 | mds_lock.Lock(); | |
3464 | } | |
3465 | ||
3466 | // We dropped mds_lock, so check that session still exists | |
3467 | session = sessionmap.get_session(entity_name_t(CEPH_ENTITY_TYPE_CLIENT, | |
3468 | session_id)); | |
3469 | if (!session) { | |
3470 | dout(1) << "session " << session_id << " was removed while we waited " | |
3471 | "for blacklist" << dendl; | |
3472 | return true; | |
3473 | } | |
3474 | kill_client_session(); | |
3475 | } else { | |
3476 | if (blacklist) { | |
3477 | apply_blacklist(kill_client_session); | |
3478 | } else { | |
3479 | kill_client_session(); | |
3480 | } | |
3481 | } | |
3482 | ||
3483 | return true; | |
3484 | } | |
3485 | ||
3486 | void MDSRank::bcast_mds_map() | |
3487 | { | |
3488 | dout(7) << "bcast_mds_map " << mdsmap->get_epoch() << dendl; | |
3489 | ||
3490 | // share the map with mounted clients | |
3491 | set<Session*> clients; | |
3492 | sessionmap.get_client_session_set(clients); | |
3493 | for (const auto &session : clients) { | |
3494 | auto m = MMDSMap::create(monc->get_fsid(), *mdsmap); | |
3495 | session->get_connection()->send_message2(std::move(m)); | |
3496 | } | |
3497 | last_client_mdsmap_bcast = mdsmap->get_epoch(); | |
3498 | } | |
3499 | ||
3500 | Context *MDSRank::create_async_exec_context(C_ExecAndReply *ctx) { | |
3501 | return new C_OnFinisher(new FunctionContext([ctx](int _) { | |
3502 | ctx->exec(); | |
3503 | }), finisher); | |
3504 | } | |
3505 | ||
3506 | MDSRankDispatcher::MDSRankDispatcher( | |
3507 | mds_rank_t whoami_, | |
3508 | Mutex &mds_lock_, | |
3509 | LogChannelRef &clog_, | |
3510 | SafeTimer &timer_, | |
3511 | Beacon &beacon_, | |
3512 | std::unique_ptr<MDSMap> &mdsmap_, | |
3513 | Messenger *msgr, | |
3514 | MonClient *monc_, | |
3515 | Context *respawn_hook_, | |
3516 | Context *suicide_hook_) | |
3517 | : MDSRank(whoami_, mds_lock_, clog_, timer_, beacon_, mdsmap_, | |
3518 | msgr, monc_, respawn_hook_, suicide_hook_) | |
3519 | { | |
3520 | g_conf().add_observer(this); | |
3521 | } | |
3522 | ||
3523 | bool MDSRankDispatcher::handle_command( | |
3524 | const cmdmap_t &cmdmap, | |
3525 | const MCommand::const_ref &m, | |
3526 | int *r, | |
3527 | std::stringstream *ds, | |
3528 | std::stringstream *ss, | |
3529 | Context **run_later, | |
3530 | bool *need_reply) | |
3531 | { | |
3532 | ceph_assert(r != nullptr); | |
3533 | ceph_assert(ds != nullptr); | |
3534 | ceph_assert(ss != nullptr); | |
3535 | ||
3536 | *need_reply = true; | |
3537 | ||
3538 | std::string prefix; | |
3539 | cmd_getval(g_ceph_context, cmdmap, "prefix", prefix); | |
3540 | ||
3541 | if (prefix == "session ls" || prefix == "client ls") { | |
3542 | std::vector<std::string> filter_args; | |
3543 | cmd_getval(g_ceph_context, cmdmap, "filters", filter_args); | |
3544 | ||
3545 | SessionFilter filter; | |
3546 | *r = filter.parse(filter_args, ss); | |
3547 | if (*r != 0) { | |
3548 | return true; | |
3549 | } | |
3550 | ||
3551 | JSONFormatter f(true); | |
3552 | dump_sessions(filter, &f); | |
3553 | f.flush(*ds); | |
3554 | return true; | |
3555 | } else if (prefix == "session evict" || prefix == "client evict") { | |
3556 | std::vector<std::string> filter_args; | |
3557 | cmd_getval(g_ceph_context, cmdmap, "filters", filter_args); | |
3558 | ||
3559 | SessionFilter filter; | |
3560 | *r = filter.parse(filter_args, ss); | |
3561 | if (*r != 0) { | |
3562 | return true; | |
3563 | } | |
3564 | ||
3565 | evict_clients(filter, m); | |
3566 | ||
3567 | *need_reply = false; | |
3568 | return true; | |
3569 | } else if (prefix == "session config" || prefix == "client config") { | |
3570 | int64_t client_id; | |
3571 | std::string option; | |
3572 | std::string value; | |
3573 | ||
3574 | cmd_getval(g_ceph_context, cmdmap, "client_id", client_id); | |
3575 | cmd_getval(g_ceph_context, cmdmap, "option", option); | |
3576 | bool got_value = cmd_getval(g_ceph_context, cmdmap, "value", value); | |
3577 | ||
3578 | *r = config_client(client_id, !got_value, option, value, *ss); | |
3579 | return true; | |
3580 | } else if (prefix == "damage ls") { | |
3581 | JSONFormatter f(true); | |
3582 | damage_table.dump(&f); | |
3583 | f.flush(*ds); | |
3584 | return true; | |
3585 | } else if (prefix == "damage rm") { | |
3586 | damage_entry_id_t id = 0; | |
3587 | bool got = cmd_getval(g_ceph_context, cmdmap, "damage_id", (int64_t&)id); | |
3588 | if (!got) { | |
3589 | *r = -EINVAL; | |
3590 | return true; | |
3591 | } | |
3592 | ||
3593 | damage_table.erase(id); | |
3594 | return true; | |
3595 | } else if (prefix == "cache drop") { | |
3596 | int64_t timeout; | |
3597 | if (!cmd_getval(g_ceph_context, cmdmap, "timeout", timeout)) { | |
3598 | timeout = 0; | |
3599 | } | |
3600 | ||
3601 | *need_reply = false; | |
3602 | *run_later = create_async_exec_context(new C_CacheDropExecAndReply | |
3603 | (this, m, (uint64_t)timeout)); | |
3604 | return true; | |
3605 | } else if (prefix == "scrub start") { | |
3606 | string path; | |
3607 | string tag; | |
3608 | vector<string> scrubop_vec; | |
3609 | cmd_getval(g_ceph_context, cmdmap, "scrubops", scrubop_vec); | |
3610 | cmd_getval(g_ceph_context, cmdmap, "path", path); | |
3611 | cmd_getval(g_ceph_context, cmdmap, "tag", tag); | |
3612 | ||
3613 | /* Multiple MDS scrub is not currently supported. See also: https://tracker.ceph.com/issues/12274 */ | |
3614 | if (mdsmap->get_max_mds() > 1) { | |
3615 | *ss << "Scrub is not currently supported for multiple active MDS. Please reduce max_mds to 1 and then scrub."; | |
3616 | *r = ENOTSUP; | |
3617 | return true; | |
3618 | } | |
3619 | ||
3620 | *need_reply = false; | |
3621 | *run_later = create_async_exec_context(new C_ScrubExecAndReply | |
3622 | (this, m, path, tag, scrubop_vec)); | |
3623 | return true; | |
3624 | } else if (prefix == "scrub abort") { | |
3625 | *need_reply = false; | |
3626 | *run_later = create_async_exec_context(new C_ScrubControlExecAndReply | |
3627 | (this, m, "abort")); | |
3628 | return true; | |
3629 | } else if (prefix == "scrub pause") { | |
3630 | *need_reply = false; | |
3631 | *run_later = create_async_exec_context(new C_ScrubControlExecAndReply | |
3632 | (this, m, "pause")); | |
3633 | return true; | |
3634 | } else if (prefix == "scrub resume") { | |
3635 | JSONFormatter f(true); | |
3636 | command_scrub_resume(&f); | |
3637 | f.flush(*ds); | |
3638 | return true; | |
3639 | } else if (prefix == "scrub status") { | |
3640 | JSONFormatter f(true); | |
3641 | command_scrub_status(&f); | |
3642 | f.flush(*ds); | |
3643 | return true; | |
3644 | } else { | |
3645 | return false; | |
3646 | } | |
3647 | } | |
3648 | ||
3649 | void MDSRank::command_cache_drop(uint64_t timeout, Formatter *f, Context *on_finish) { | |
3650 | dout(20) << __func__ << dendl; | |
3651 | ||
3652 | std::lock_guard locker(mds_lock); | |
3653 | C_Drop_Cache *request = new C_Drop_Cache(server, mdcache, mdlog, this, | |
3654 | timeout, f, on_finish); | |
3655 | request->send(); | |
3656 | } | |
3657 | ||
3658 | epoch_t MDSRank::get_osd_epoch() const | |
3659 | { | |
3660 | return objecter->with_osdmap(std::mem_fn(&OSDMap::get_epoch)); | |
3661 | } | |
3662 | ||
3663 | const char** MDSRankDispatcher::get_tracked_conf_keys() const | |
3664 | { | |
3665 | static const char* KEYS[] = { | |
3666 | "clog_to_graylog", | |
3667 | "clog_to_graylog_host", | |
3668 | "clog_to_graylog_port", | |
3669 | "clog_to_monitors", | |
3670 | "clog_to_syslog", | |
3671 | "clog_to_syslog_facility", | |
3672 | "clog_to_syslog_level", | |
3673 | "fsid", | |
3674 | "host", | |
3675 | "mds_bal_fragment_dirs", | |
3676 | "mds_bal_fragment_interval", | |
3677 | "mds_cache_memory_limit", | |
3678 | "mds_cache_mid", | |
3679 | "mds_cache_reservation", | |
3680 | "mds_cache_size", | |
3681 | "mds_cache_trim_decay_rate", | |
3682 | "mds_cap_revoke_eviction_timeout", | |
3683 | "mds_dump_cache_threshold_file", | |
3684 | "mds_dump_cache_threshold_formatter", | |
3685 | "mds_enable_op_tracker", | |
3686 | "mds_health_cache_threshold", | |
3687 | "mds_inject_migrator_session_race", | |
3688 | "mds_log_pause", | |
3689 | "mds_max_export_size", | |
3690 | "mds_max_purge_files", | |
3691 | "mds_max_purge_ops", | |
3692 | "mds_max_purge_ops_per_pg", | |
3693 | "mds_op_complaint_time", | |
3694 | "mds_op_history_duration", | |
3695 | "mds_op_history_size", | |
3696 | "mds_op_log_threshold", | |
3697 | "mds_recall_max_decay_rate", | |
3698 | "mds_recall_warning_decay_rate", | |
3699 | "mds_request_load_average_decay_rate", | |
3700 | "mds_session_cache_liveness_decay_rate", | |
3701 | "mds_replay_unsafe_with_closed_session", | |
3702 | NULL | |
3703 | }; | |
3704 | return KEYS; | |
3705 | } | |
3706 | ||
3707 | void MDSRankDispatcher::handle_conf_change(const ConfigProxy& conf, const std::set<std::string>& changed) | |
3708 | { | |
3709 | // XXX with or without mds_lock! | |
3710 | ||
3711 | if (changed.count("mds_op_complaint_time") || changed.count("mds_op_log_threshold")) { | |
3712 | op_tracker.set_complaint_and_threshold(conf->mds_op_complaint_time, conf->mds_op_log_threshold); | |
3713 | } | |
3714 | if (changed.count("mds_op_history_size") || changed.count("mds_op_history_duration")) { | |
3715 | op_tracker.set_history_size_and_duration(conf->mds_op_history_size, conf->mds_op_history_duration); | |
3716 | } | |
3717 | if (changed.count("mds_enable_op_tracker")) { | |
3718 | op_tracker.set_tracking(conf->mds_enable_op_tracker); | |
3719 | } | |
3720 | if (changed.count("clog_to_monitors") || | |
3721 | changed.count("clog_to_syslog") || | |
3722 | changed.count("clog_to_syslog_level") || | |
3723 | changed.count("clog_to_syslog_facility") || | |
3724 | changed.count("clog_to_graylog") || | |
3725 | changed.count("clog_to_graylog_host") || | |
3726 | changed.count("clog_to_graylog_port") || | |
3727 | changed.count("host") || | |
3728 | changed.count("fsid")) { | |
3729 | update_log_config(); | |
3730 | } | |
3731 | ||
3732 | finisher->queue(new FunctionContext([this, changed](int r) { | |
3733 | std::scoped_lock lock(mds_lock); | |
3734 | ||
3735 | if (changed.count("mds_log_pause") && !g_conf()->mds_log_pause) { | |
3736 | mdlog->kick_submitter(); | |
3737 | } | |
3738 | sessionmap.handle_conf_change(changed); | |
3739 | server->handle_conf_change(changed); | |
3740 | mdcache->handle_conf_change(changed, *mdsmap); | |
3741 | purge_queue.handle_conf_change(changed, *mdsmap); | |
3742 | })); | |
3743 | } |