]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2015 Red Hat | |
7 | * | |
8 | * This is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License version 2.1, as published by the Free Software | |
11 | * Foundation. See file COPYING. | |
12 | * | |
13 | */ | |
14 | ||
15 | #include "common/debug.h" | |
16 | #include "mds/mdstypes.h" | |
17 | #include "mds/CInode.h" | |
18 | #include "mds/MDCache.h" | |
19 | ||
20 | #include "PurgeQueue.h" | |
21 | ||
f64942e4 | 22 | #include <string.h> |
7c673cae FG |
23 | |
24 | #define dout_context cct | |
25 | #define dout_subsys ceph_subsys_mds | |
26 | #undef dout_prefix | |
27 | #define dout_prefix _prefix(_dout, rank) << __func__ << ": " | |
28 | static ostream& _prefix(std::ostream *_dout, mds_rank_t rank) { | |
29 | return *_dout << "mds." << rank << ".purge_queue "; | |
30 | } | |
31 | ||
11fdf7f2 TL |
32 | const std::map<std::string, PurgeItem::Action> PurgeItem::actions = { |
33 | {"NONE", PurgeItem::NONE}, | |
34 | {"PURGE_FILE", PurgeItem::PURGE_FILE}, | |
35 | {"TRUNCATE_FILE", PurgeItem::TRUNCATE_FILE}, | |
36 | {"PURGE_DIR", PurgeItem::PURGE_DIR} | |
37 | }; | |
38 | ||
7c673cae FG |
39 | void PurgeItem::encode(bufferlist &bl) const |
40 | { | |
11fdf7f2 TL |
41 | ENCODE_START(2, 1, bl); |
42 | encode((uint8_t)action, bl); | |
43 | encode(ino, bl); | |
44 | encode(size, bl); | |
45 | encode(layout, bl, CEPH_FEATURE_FS_FILE_LAYOUT_V2); | |
46 | encode(old_pools, bl); | |
47 | encode(snapc, bl); | |
48 | encode(fragtree, bl); | |
49 | encode(stamp, bl); | |
50 | uint8_t static const pad = 0xff; | |
51 | for (unsigned int i = 0; i<pad_size; i++) { | |
52 | encode(pad, bl); | |
53 | } | |
7c673cae FG |
54 | ENCODE_FINISH(bl); |
55 | } | |
56 | ||
11fdf7f2 | 57 | void PurgeItem::decode(bufferlist::const_iterator &p) |
7c673cae | 58 | { |
11fdf7f2 | 59 | DECODE_START(2, p); |
9f95a23c TL |
60 | bool done = false; |
61 | if (struct_v == 1) { | |
62 | auto p_start = p; | |
63 | try { | |
64 | // bad encoding introduced by v13.2.2 | |
65 | decode(stamp, p); | |
66 | decode(pad_size, p); | |
67 | p += pad_size; | |
f91f0fd5 TL |
68 | uint8_t raw_action; |
69 | decode(raw_action, p); | |
70 | action = (Action)raw_action; | |
9f95a23c TL |
71 | decode(ino, p); |
72 | decode(size, p); | |
73 | decode(layout, p); | |
74 | decode(old_pools, p); | |
75 | decode(snapc, p); | |
76 | decode(fragtree, p); | |
77 | if (p.get_off() > struct_end) | |
78 | throw buffer::end_of_buffer(); | |
79 | done = true; | |
80 | } catch (const buffer::error &e) { | |
81 | p = p_start; | |
82 | } | |
83 | } | |
84 | if (!done) { | |
f91f0fd5 TL |
85 | uint8_t raw_action; |
86 | decode(raw_action, p); | |
87 | action = (Action)raw_action; | |
9f95a23c TL |
88 | decode(ino, p); |
89 | decode(size, p); | |
90 | decode(layout, p); | |
91 | decode(old_pools, p); | |
92 | decode(snapc, p); | |
93 | decode(fragtree, p); | |
94 | if (struct_v >= 2) { | |
95 | decode(stamp, p); | |
96 | } | |
11fdf7f2 | 97 | } |
7c673cae FG |
98 | DECODE_FINISH(p); |
99 | } | |
100 | ||
101 | // TODO: if Objecter has any slow requests, take that as a hint and | |
102 | // slow down our rate of purging (keep accepting pushes though) | |
103 | PurgeQueue::PurgeQueue( | |
104 | CephContext *cct_, | |
105 | mds_rank_t rank_, | |
106 | const int64_t metadata_pool_, | |
107 | Objecter *objecter_, | |
108 | Context *on_error_) | |
109 | : | |
110 | cct(cct_), | |
111 | rank(rank_), | |
7c673cae FG |
112 | metadata_pool(metadata_pool_), |
113 | finisher(cct, "PurgeQueue", "PQ_Finisher"), | |
114 | timer(cct, lock), | |
115 | filer(objecter_, &finisher), | |
116 | objecter(objecter_), | |
117 | journaler("pq", MDS_INO_PURGE_QUEUE + rank, metadata_pool, | |
118 | CEPH_FS_ONDISK_MAGIC, objecter_, nullptr, 0, | |
119 | &finisher), | |
9f95a23c | 120 | on_error(on_error_) |
7c673cae | 121 | { |
11fdf7f2 TL |
122 | ceph_assert(cct != nullptr); |
123 | ceph_assert(on_error != nullptr); | |
124 | ceph_assert(objecter != nullptr); | |
7c673cae FG |
125 | journaler.set_write_error_handler(on_error); |
126 | } | |
127 | ||
128 | PurgeQueue::~PurgeQueue() | |
129 | { | |
130 | if (logger) { | |
131 | g_ceph_context->get_perfcounters_collection()->remove(logger.get()); | |
132 | } | |
f64942e4 | 133 | delete on_error; |
7c673cae FG |
134 | } |
135 | ||
136 | void PurgeQueue::create_logger() | |
137 | { | |
91327a77 AA |
138 | PerfCountersBuilder pcb(g_ceph_context, "purge_queue", l_pq_first, l_pq_last); |
139 | ||
140 | pcb.add_u64_counter(l_pq_executed, "pq_executed", "Purge queue tasks executed", | |
141 | "purg", PerfCountersBuilder::PRIO_INTERESTING); | |
142 | ||
143 | pcb.set_prio_default(PerfCountersBuilder::PRIO_USEFUL); | |
7c673cae | 144 | pcb.add_u64(l_pq_executing_ops, "pq_executing_ops", "Purge queue ops in flight"); |
92f5a8d4 | 145 | pcb.add_u64(l_pq_executing_ops_high_water, "pq_executing_ops_high_water", "Maximum number of executing file purge ops"); |
7c673cae | 146 | pcb.add_u64(l_pq_executing, "pq_executing", "Purge queue tasks in flight"); |
92f5a8d4 | 147 | pcb.add_u64(l_pq_executing_high_water, "pq_executing_high_water", "Maximum number of executing file purges"); |
9f95a23c | 148 | pcb.add_u64(l_pq_item_in_journal, "pq_item_in_journal", "Purge item left in journal"); |
7c673cae FG |
149 | |
150 | logger.reset(pcb.create_perf_counters()); | |
151 | g_ceph_context->get_perfcounters_collection()->add(logger.get()); | |
152 | } | |
153 | ||
154 | void PurgeQueue::init() | |
155 | { | |
11fdf7f2 | 156 | std::lock_guard l(lock); |
7c673cae | 157 | |
11fdf7f2 | 158 | ceph_assert(logger != nullptr); |
7c673cae FG |
159 | |
160 | finisher.start(); | |
161 | timer.init(); | |
162 | } | |
163 | ||
c07f9fc5 FG |
164 | void PurgeQueue::activate() |
165 | { | |
11fdf7f2 | 166 | std::lock_guard l(lock); |
f64942e4 | 167 | |
9f95a23c TL |
168 | { |
169 | PurgeItem item; | |
170 | bufferlist bl; | |
171 | ||
172 | // calculate purge item serialized size stored in journal | |
173 | // used to count how many items still left in journal later | |
174 | ::encode(item, bl); | |
175 | purge_item_journal_size = bl.length() + journaler.get_journal_envelope_size(); | |
176 | } | |
177 | ||
f64942e4 AA |
178 | if (readonly) { |
179 | dout(10) << "skipping activate: PurgeQueue is readonly" << dendl; | |
180 | return; | |
181 | } | |
182 | ||
c07f9fc5 FG |
183 | if (journaler.get_read_pos() == journaler.get_write_pos()) |
184 | return; | |
185 | ||
186 | if (in_flight.empty()) { | |
187 | dout(4) << "start work (by drain)" << dendl; | |
9f95a23c | 188 | finisher.queue(new LambdaContext([this](int r) { |
11fdf7f2 | 189 | std::lock_guard l(lock); |
c07f9fc5 FG |
190 | _consume(); |
191 | })); | |
192 | } | |
193 | } | |
194 | ||
7c673cae FG |
195 | void PurgeQueue::shutdown() |
196 | { | |
11fdf7f2 | 197 | std::lock_guard l(lock); |
7c673cae FG |
198 | |
199 | journaler.shutdown(); | |
200 | timer.shutdown(); | |
201 | finisher.stop(); | |
202 | } | |
203 | ||
204 | void PurgeQueue::open(Context *completion) | |
205 | { | |
206 | dout(4) << "opening" << dendl; | |
207 | ||
11fdf7f2 | 208 | std::lock_guard l(lock); |
7c673cae | 209 | |
3efd9988 FG |
210 | if (completion) |
211 | waiting_for_recovery.push_back(completion); | |
212 | ||
9f95a23c | 213 | journaler.recover(new LambdaContext([this](int r){ |
7c673cae FG |
214 | if (r == -ENOENT) { |
215 | dout(1) << "Purge Queue not found, assuming this is an upgrade and " | |
216 | "creating it." << dendl; | |
3efd9988 | 217 | create(NULL); |
7c673cae | 218 | } else if (r == 0) { |
11fdf7f2 | 219 | std::lock_guard l(lock); |
7c673cae FG |
220 | dout(4) << "open complete" << dendl; |
221 | ||
222 | // Journaler only guarantees entries before head write_pos have been | |
223 | // fully flushed. Before appending new entries, we need to find and | |
224 | // drop any partial written entry. | |
225 | if (journaler.last_committed.write_pos < journaler.get_write_pos()) { | |
226 | dout(4) << "recovering write_pos" << dendl; | |
227 | journaler.set_read_pos(journaler.last_committed.write_pos); | |
3efd9988 | 228 | _recover(); |
7c673cae FG |
229 | return; |
230 | } | |
231 | ||
232 | journaler.set_writeable(); | |
3efd9988 FG |
233 | recovered = true; |
234 | finish_contexts(g_ceph_context, waiting_for_recovery); | |
7c673cae FG |
235 | } else { |
236 | derr << "Error " << r << " loading Journaler" << dendl; | |
f64942e4 | 237 | _go_readonly(r); |
7c673cae FG |
238 | } |
239 | })); | |
240 | } | |
241 | ||
3efd9988 FG |
242 | void PurgeQueue::wait_for_recovery(Context* c) |
243 | { | |
11fdf7f2 | 244 | std::lock_guard l(lock); |
f64942e4 | 245 | if (recovered) { |
3efd9988 | 246 | c->complete(0); |
f64942e4 AA |
247 | } else if (readonly) { |
248 | dout(10) << "cannot wait for recovery: PurgeQueue is readonly" << dendl; | |
249 | c->complete(-EROFS); | |
250 | } else { | |
3efd9988 | 251 | waiting_for_recovery.push_back(c); |
f64942e4 | 252 | } |
3efd9988 | 253 | } |
7c673cae | 254 | |
3efd9988 | 255 | void PurgeQueue::_recover() |
7c673cae | 256 | { |
9f95a23c | 257 | ceph_assert(ceph_mutex_is_locked_by_me(lock)); |
7c673cae FG |
258 | |
259 | // Journaler::is_readable() adjusts write_pos if partial entry is encountered | |
260 | while (1) { | |
261 | if (!journaler.is_readable() && | |
262 | !journaler.get_error() && | |
263 | journaler.get_read_pos() < journaler.get_write_pos()) { | |
9f95a23c | 264 | journaler.wait_for_readable(new LambdaContext([this](int r) { |
11fdf7f2 | 265 | std::lock_guard l(lock); |
3efd9988 | 266 | _recover(); |
7c673cae FG |
267 | })); |
268 | return; | |
269 | } | |
270 | ||
271 | if (journaler.get_error()) { | |
272 | int r = journaler.get_error(); | |
273 | derr << "Error " << r << " recovering write_pos" << dendl; | |
f64942e4 | 274 | _go_readonly(r); |
7c673cae FG |
275 | return; |
276 | } | |
277 | ||
278 | if (journaler.get_read_pos() == journaler.get_write_pos()) { | |
279 | dout(4) << "write_pos recovered" << dendl; | |
280 | // restore original read_pos | |
281 | journaler.set_read_pos(journaler.last_committed.expire_pos); | |
282 | journaler.set_writeable(); | |
3efd9988 FG |
283 | recovered = true; |
284 | finish_contexts(g_ceph_context, waiting_for_recovery); | |
7c673cae FG |
285 | return; |
286 | } | |
287 | ||
288 | bufferlist bl; | |
289 | bool readable = journaler.try_read_entry(bl); | |
11fdf7f2 | 290 | ceph_assert(readable); // we checked earlier |
7c673cae FG |
291 | } |
292 | } | |
293 | ||
294 | void PurgeQueue::create(Context *fin) | |
295 | { | |
296 | dout(4) << "creating" << dendl; | |
11fdf7f2 | 297 | std::lock_guard l(lock); |
7c673cae | 298 | |
3efd9988 FG |
299 | if (fin) |
300 | waiting_for_recovery.push_back(fin); | |
301 | ||
7c673cae FG |
302 | file_layout_t layout = file_layout_t::get_default(); |
303 | layout.pool_id = metadata_pool; | |
304 | journaler.set_writeable(); | |
305 | journaler.create(&layout, JOURNAL_FORMAT_RESILIENT); | |
9f95a23c | 306 | journaler.write_head(new LambdaContext([this](int r) { |
11fdf7f2 | 307 | std::lock_guard l(lock); |
f64942e4 AA |
308 | if (r) { |
309 | _go_readonly(r); | |
310 | } else { | |
311 | recovered = true; | |
312 | finish_contexts(g_ceph_context, waiting_for_recovery); | |
313 | } | |
3efd9988 | 314 | })); |
7c673cae FG |
315 | } |
316 | ||
317 | /** | |
318 | * The `completion` context will always be called back via a Finisher | |
319 | */ | |
320 | void PurgeQueue::push(const PurgeItem &pi, Context *completion) | |
321 | { | |
11fdf7f2 TL |
322 | dout(4) << "pushing inode " << pi.ino << dendl; |
323 | std::lock_guard l(lock); | |
7c673cae | 324 | |
f64942e4 AA |
325 | if (readonly) { |
326 | dout(10) << "cannot push inode: PurgeQueue is readonly" << dendl; | |
327 | completion->complete(-EROFS); | |
328 | return; | |
329 | } | |
330 | ||
7c673cae | 331 | // Callers should have waited for open() before using us |
11fdf7f2 | 332 | ceph_assert(!journaler.is_readonly()); |
7c673cae FG |
333 | |
334 | bufferlist bl; | |
335 | ||
11fdf7f2 | 336 | encode(pi, bl); |
7c673cae FG |
337 | journaler.append_entry(bl); |
338 | journaler.wait_for_flush(completion); | |
339 | ||
340 | // Maybe go ahead and do something with it right away | |
341 | bool could_consume = _consume(); | |
342 | if (!could_consume) { | |
343 | // Usually, it is not necessary to explicitly flush here, because the reader | |
344 | // will get flushes generated inside Journaler::is_readable. However, | |
f64942e4 | 345 | // if we remain in a _can_consume()==false state for a long period then |
7c673cae FG |
346 | // we should flush in order to allow MDCache to drop its strays rather |
347 | // than having them wait for purgequeue to progress. | |
348 | if (!delayed_flush) { | |
9f95a23c | 349 | delayed_flush = new LambdaContext([this](int r){ |
7c673cae FG |
350 | delayed_flush = nullptr; |
351 | journaler.flush(); | |
352 | }); | |
353 | ||
354 | timer.add_event_after( | |
11fdf7f2 | 355 | g_conf()->mds_purge_queue_busy_flush_period, |
7c673cae FG |
356 | delayed_flush); |
357 | } | |
358 | } | |
359 | } | |
360 | ||
361 | uint32_t PurgeQueue::_calculate_ops(const PurgeItem &item) const | |
362 | { | |
363 | uint32_t ops_required = 0; | |
364 | if (item.action == PurgeItem::PURGE_DIR) { | |
365 | // Directory, count dirfrags to be deleted | |
11fdf7f2 | 366 | frag_vec_t leaves; |
7c673cae | 367 | if (!item.fragtree.is_leaf(frag_t())) { |
11fdf7f2 | 368 | item.fragtree.get_leaves(leaves); |
7c673cae FG |
369 | } |
370 | // One for the root, plus any leaves | |
11fdf7f2 | 371 | ops_required = 1 + leaves.size(); |
7c673cae FG |
372 | } else { |
373 | // File, work out concurrent Filer::purge deletes | |
f91f0fd5 | 374 | // Account for removing (or zeroing) backtrace |
7c673cae FG |
375 | const uint64_t num = (item.size > 0) ? |
376 | Striper::get_num_objects(item.layout, item.size) : 1; | |
377 | ||
11fdf7f2 | 378 | ops_required = std::min(num, g_conf()->filer_max_purge_ops); |
7c673cae | 379 | |
7c673cae FG |
380 | // Account for deletions for old pools |
381 | if (item.action != PurgeItem::TRUNCATE_FILE) { | |
382 | ops_required += item.old_pools.size(); | |
383 | } | |
384 | } | |
385 | ||
386 | return ops_required; | |
387 | } | |
388 | ||
f64942e4 | 389 | bool PurgeQueue::_can_consume() |
7c673cae | 390 | { |
f64942e4 AA |
391 | if (readonly) { |
392 | dout(10) << "can't consume: PurgeQueue is readonly" << dendl; | |
393 | return false; | |
394 | } | |
395 | ||
7c673cae | 396 | dout(20) << ops_in_flight << "/" << max_purge_ops << " ops, " |
11fdf7f2 | 397 | << in_flight.size() << "/" << g_conf()->mds_max_purge_files |
7c673cae FG |
398 | << " files" << dendl; |
399 | ||
400 | if (in_flight.size() == 0 && cct->_conf->mds_max_purge_files > 0) { | |
401 | // Always permit consumption if nothing is in flight, so that the ops | |
402 | // limit can never be so low as to forbid all progress (unless | |
403 | // administrator has deliberately paused purging by setting max | |
404 | // purge files to zero). | |
405 | return true; | |
406 | } | |
407 | ||
408 | if (ops_in_flight >= max_purge_ops) { | |
409 | dout(20) << "Throttling on op limit " << ops_in_flight << "/" | |
410 | << max_purge_ops << dendl; | |
411 | return false; | |
412 | } | |
413 | ||
414 | if (in_flight.size() >= cct->_conf->mds_max_purge_files) { | |
415 | dout(20) << "Throttling on item limit " << in_flight.size() | |
416 | << "/" << cct->_conf->mds_max_purge_files << dendl; | |
417 | return false; | |
418 | } else { | |
419 | return true; | |
420 | } | |
421 | } | |
422 | ||
f64942e4 AA |
423 | void PurgeQueue::_go_readonly(int r) |
424 | { | |
425 | if (readonly) return; | |
426 | dout(1) << "going readonly because internal IO failed: " << strerror(-r) << dendl; | |
427 | readonly = true; | |
9f95a23c | 428 | finisher.queue(on_error, r); |
f64942e4 AA |
429 | on_error = nullptr; |
430 | journaler.set_readonly(); | |
431 | finish_contexts(g_ceph_context, waiting_for_recovery, r); | |
432 | } | |
433 | ||
7c673cae FG |
434 | bool PurgeQueue::_consume() |
435 | { | |
9f95a23c | 436 | ceph_assert(ceph_mutex_is_locked_by_me(lock)); |
7c673cae FG |
437 | |
438 | bool could_consume = false; | |
f64942e4 | 439 | while(_can_consume()) { |
7c673cae FG |
440 | |
441 | if (delayed_flush) { | |
442 | // We are now going to read from the journal, so any proactive | |
443 | // flush is no longer necessary. This is not functionally necessary | |
444 | // but it can avoid generating extra fragmented flush IOs. | |
445 | timer.cancel_event(delayed_flush); | |
446 | delayed_flush = nullptr; | |
447 | } | |
448 | ||
1adf2230 AA |
449 | if (int r = journaler.get_error()) { |
450 | derr << "Error " << r << " recovering write_pos" << dendl; | |
f64942e4 | 451 | _go_readonly(r); |
1adf2230 AA |
452 | return could_consume; |
453 | } | |
454 | ||
7c673cae FG |
455 | if (!journaler.is_readable()) { |
456 | dout(10) << " not readable right now" << dendl; | |
457 | // Because we are the writer and the reader of the journal | |
458 | // via the same Journaler instance, we never need to reread_head | |
459 | if (!journaler.have_waiter()) { | |
9f95a23c | 460 | journaler.wait_for_readable(new LambdaContext([this](int r) { |
11fdf7f2 | 461 | std::lock_guard l(lock); |
7c673cae FG |
462 | if (r == 0) { |
463 | _consume(); | |
1adf2230 | 464 | } else if (r != -EAGAIN) { |
f64942e4 | 465 | _go_readonly(r); |
7c673cae FG |
466 | } |
467 | })); | |
468 | } | |
469 | ||
470 | return could_consume; | |
471 | } | |
472 | ||
28e407b8 | 473 | could_consume = true; |
7c673cae FG |
474 | // The journaler is readable: consume an entry |
475 | bufferlist bl; | |
476 | bool readable = journaler.try_read_entry(bl); | |
11fdf7f2 | 477 | ceph_assert(readable); // we checked earlier |
7c673cae FG |
478 | |
479 | dout(20) << " decoding entry" << dendl; | |
480 | PurgeItem item; | |
11fdf7f2 | 481 | auto q = bl.cbegin(); |
7c673cae | 482 | try { |
11fdf7f2 | 483 | decode(item, q); |
7c673cae FG |
484 | } catch (const buffer::error &err) { |
485 | derr << "Decode error at read_pos=0x" << std::hex | |
486 | << journaler.get_read_pos() << dendl; | |
f64942e4 | 487 | _go_readonly(EIO); |
7c673cae | 488 | } |
11fdf7f2 | 489 | dout(20) << " executing item (" << item.ino << ")" << dendl; |
7c673cae FG |
490 | _execute_item(item, journaler.get_read_pos()); |
491 | } | |
492 | ||
493 | dout(10) << " cannot consume right now" << dendl; | |
494 | ||
495 | return could_consume; | |
496 | } | |
497 | ||
498 | void PurgeQueue::_execute_item( | |
499 | const PurgeItem &item, | |
500 | uint64_t expire_to) | |
501 | { | |
9f95a23c | 502 | ceph_assert(ceph_mutex_is_locked_by_me(lock)); |
7c673cae FG |
503 | |
504 | in_flight[expire_to] = item; | |
505 | logger->set(l_pq_executing, in_flight.size()); | |
92f5a8d4 TL |
506 | files_high_water = std::max(files_high_water, in_flight.size()); |
507 | logger->set(l_pq_executing_high_water, files_high_water); | |
f64942e4 AA |
508 | auto ops = _calculate_ops(item); |
509 | ops_in_flight += ops; | |
7c673cae | 510 | logger->set(l_pq_executing_ops, ops_in_flight); |
92f5a8d4 TL |
511 | ops_high_water = std::max(ops_high_water, ops_in_flight); |
512 | logger->set(l_pq_executing_ops_high_water, ops_high_water); | |
7c673cae FG |
513 | |
514 | SnapContext nullsnapc; | |
515 | ||
516 | C_GatherBuilder gather(cct); | |
517 | if (item.action == PurgeItem::PURGE_FILE) { | |
518 | if (item.size > 0) { | |
519 | uint64_t num = Striper::get_num_objects(item.layout, item.size); | |
520 | dout(10) << " 0~" << item.size << " objects 0~" << num | |
521 | << " snapc " << item.snapc << " on " << item.ino << dendl; | |
522 | filer.purge_range(item.ino, &item.layout, item.snapc, | |
523 | 0, num, ceph::real_clock::now(), 0, | |
524 | gather.new_sub()); | |
525 | } | |
526 | ||
527 | // remove the backtrace object if it was not purged | |
528 | object_t oid = CInode::get_object_name(item.ino, frag_t(), ""); | |
529 | if (!gather.has_subs() || !item.layout.pool_ns.empty()) { | |
530 | object_locator_t oloc(item.layout.pool_id); | |
531 | dout(10) << " remove backtrace object " << oid | |
532 | << " pool " << oloc.pool << " snapc " << item.snapc << dendl; | |
533 | objecter->remove(oid, oloc, item.snapc, | |
534 | ceph::real_clock::now(), 0, | |
535 | gather.new_sub()); | |
536 | } | |
537 | ||
538 | // remove old backtrace objects | |
539 | for (const auto &p : item.old_pools) { | |
540 | object_locator_t oloc(p); | |
541 | dout(10) << " remove backtrace object " << oid | |
542 | << " old pool " << p << " snapc " << item.snapc << dendl; | |
543 | objecter->remove(oid, oloc, item.snapc, | |
544 | ceph::real_clock::now(), 0, | |
545 | gather.new_sub()); | |
546 | } | |
547 | } else if (item.action == PurgeItem::PURGE_DIR) { | |
548 | object_locator_t oloc(metadata_pool); | |
11fdf7f2 | 549 | frag_vec_t leaves; |
7c673cae | 550 | if (!item.fragtree.is_leaf(frag_t())) |
11fdf7f2 TL |
551 | item.fragtree.get_leaves(leaves); |
552 | leaves.push_back(frag_t()); | |
553 | for (const auto &leaf : leaves) { | |
554 | object_t oid = CInode::get_object_name(item.ino, leaf, ""); | |
7c673cae FG |
555 | dout(10) << " remove dirfrag " << oid << dendl; |
556 | objecter->remove(oid, oloc, nullsnapc, | |
557 | ceph::real_clock::now(), | |
558 | 0, gather.new_sub()); | |
559 | } | |
560 | } else if (item.action == PurgeItem::TRUNCATE_FILE) { | |
561 | const uint64_t num = Striper::get_num_objects(item.layout, item.size); | |
562 | dout(10) << " 0~" << item.size << " objects 0~" << num | |
563 | << " snapc " << item.snapc << " on " << item.ino << dendl; | |
564 | ||
565 | // keep backtrace object | |
566 | if (num > 1) { | |
567 | filer.purge_range(item.ino, &item.layout, item.snapc, | |
568 | 1, num - 1, ceph::real_clock::now(), | |
569 | 0, gather.new_sub()); | |
570 | } | |
571 | filer.zero(item.ino, &item.layout, item.snapc, | |
572 | 0, item.layout.object_size, | |
573 | ceph::real_clock::now(), | |
574 | 0, true, gather.new_sub()); | |
575 | } else { | |
576 | derr << "Invalid item (action=" << item.action << ") in purge queue, " | |
577 | "dropping it" << dendl; | |
f64942e4 AA |
578 | ops_in_flight -= ops; |
579 | logger->set(l_pq_executing_ops, ops_in_flight); | |
92f5a8d4 TL |
580 | ops_high_water = std::max(ops_high_water, ops_in_flight); |
581 | logger->set(l_pq_executing_ops_high_water, ops_high_water); | |
7c673cae FG |
582 | in_flight.erase(expire_to); |
583 | logger->set(l_pq_executing, in_flight.size()); | |
92f5a8d4 TL |
584 | files_high_water = std::max(files_high_water, in_flight.size()); |
585 | logger->set(l_pq_executing_high_water, files_high_water); | |
7c673cae FG |
586 | return; |
587 | } | |
11fdf7f2 | 588 | ceph_assert(gather.has_subs()); |
7c673cae | 589 | |
31f18b77 | 590 | gather.set_finisher(new C_OnFinisher( |
9f95a23c | 591 | new LambdaContext([this, expire_to](int r){ |
11fdf7f2 | 592 | std::lock_guard l(lock); |
7c673cae | 593 | |
e306af50 TL |
594 | if (r == -EBLACKLISTED) { |
595 | finisher.queue(on_error, r); | |
596 | on_error = nullptr; | |
597 | return; | |
598 | } | |
599 | ||
600 | _execute_item_complete(expire_to); | |
31f18b77 | 601 | _consume(); |
7c673cae FG |
602 | |
603 | // Have we gone idle? If so, do an extra write_head now instead of | |
604 | // waiting for next flush after journaler_write_head_interval. | |
605 | // Also do this periodically even if not idle, so that the persisted | |
606 | // expire_pos doesn't fall too far behind our progress when consuming | |
607 | // a very long queue. | |
9f95a23c TL |
608 | if (!readonly && |
609 | (in_flight.empty() || journaler.write_head_needed())) { | |
f64942e4 | 610 | journaler.write_head(nullptr); |
7c673cae | 611 | } |
31f18b77 FG |
612 | }), &finisher)); |
613 | ||
7c673cae FG |
614 | gather.activate(); |
615 | } | |
616 | ||
617 | void PurgeQueue::_execute_item_complete( | |
618 | uint64_t expire_to) | |
619 | { | |
9f95a23c | 620 | ceph_assert(ceph_mutex_is_locked_by_me(lock)); |
7c673cae | 621 | dout(10) << "complete at 0x" << std::hex << expire_to << std::dec << dendl; |
11fdf7f2 | 622 | ceph_assert(in_flight.count(expire_to) == 1); |
7c673cae FG |
623 | |
624 | auto iter = in_flight.find(expire_to); | |
11fdf7f2 | 625 | ceph_assert(iter != in_flight.end()); |
7c673cae | 626 | if (iter == in_flight.begin()) { |
11fdf7f2 TL |
627 | uint64_t pos = expire_to; |
628 | if (!pending_expire.empty()) { | |
629 | auto n = iter; | |
630 | ++n; | |
631 | if (n == in_flight.end()) { | |
632 | pos = *pending_expire.rbegin(); | |
633 | pending_expire.clear(); | |
634 | } else { | |
635 | auto p = pending_expire.begin(); | |
636 | do { | |
637 | if (*p >= n->first) | |
638 | break; | |
639 | pos = *p; | |
640 | pending_expire.erase(p++); | |
641 | } while (p != pending_expire.end()); | |
642 | } | |
643 | } | |
644 | dout(10) << "expiring to 0x" << std::hex << pos << std::dec << dendl; | |
645 | journaler.set_expire_pos(pos); | |
7c673cae FG |
646 | } else { |
647 | // This is completely fine, we're not supposed to purge files in | |
648 | // order when doing them in parallel. | |
649 | dout(10) << "non-sequential completion, not expiring anything" << dendl; | |
11fdf7f2 | 650 | pending_expire.insert(expire_to); |
7c673cae FG |
651 | } |
652 | ||
653 | ops_in_flight -= _calculate_ops(iter->second); | |
654 | logger->set(l_pq_executing_ops, ops_in_flight); | |
92f5a8d4 TL |
655 | ops_high_water = std::max(ops_high_water, ops_in_flight); |
656 | logger->set(l_pq_executing_ops_high_water, ops_high_water); | |
7c673cae | 657 | |
11fdf7f2 | 658 | dout(10) << "completed item for ino " << iter->second.ino << dendl; |
7c673cae FG |
659 | |
660 | in_flight.erase(iter); | |
661 | logger->set(l_pq_executing, in_flight.size()); | |
92f5a8d4 TL |
662 | files_high_water = std::max(files_high_water, in_flight.size()); |
663 | logger->set(l_pq_executing_high_water, files_high_water); | |
7c673cae FG |
664 | dout(10) << "in_flight.size() now " << in_flight.size() << dendl; |
665 | ||
9f95a23c TL |
666 | uint64_t write_pos = journaler.get_write_pos(); |
667 | uint64_t read_pos = journaler.get_read_pos(); | |
668 | uint64_t expire_pos = journaler.get_expire_pos(); | |
669 | uint64_t item_num = (write_pos - (in_flight.size() ? expire_pos : read_pos)) | |
670 | / purge_item_journal_size; | |
671 | dout(10) << "left purge items in journal: " << item_num | |
672 | << " (purge_item_journal_size/write_pos/read_pos/expire_pos) now at " | |
673 | << "(" << purge_item_journal_size << "/" << write_pos << "/" << read_pos | |
674 | << "/" << expire_pos << ")" << dendl; | |
675 | ||
676 | logger->set(l_pq_item_in_journal, item_num); | |
7c673cae FG |
677 | logger->inc(l_pq_executed); |
678 | } | |
679 | ||
680 | void PurgeQueue::update_op_limit(const MDSMap &mds_map) | |
681 | { | |
11fdf7f2 | 682 | std::lock_guard l(lock); |
7c673cae | 683 | |
f64942e4 AA |
684 | if (readonly) { |
685 | dout(10) << "skipping; PurgeQueue is readonly" << dendl; | |
686 | return; | |
687 | } | |
688 | ||
7c673cae FG |
689 | uint64_t pg_count = 0; |
690 | objecter->with_osdmap([&](const OSDMap& o) { | |
691 | // Number of PGs across all data pools | |
31f18b77 | 692 | const std::vector<int64_t> &data_pools = mds_map.get_data_pools(); |
7c673cae FG |
693 | for (const auto dp : data_pools) { |
694 | if (o.get_pg_pool(dp) == NULL) { | |
695 | // It is possible that we have an older OSDMap than MDSMap, | |
696 | // because we don't start watching every OSDMap until after | |
697 | // MDSRank is initialized | |
698 | dout(4) << " data pool " << dp << " not found in OSDMap" << dendl; | |
699 | continue; | |
700 | } | |
701 | pg_count += o.get_pg_num(dp); | |
702 | } | |
703 | }); | |
704 | ||
705 | // Work out a limit based on n_pgs / n_mdss, multiplied by the user's | |
706 | // preference for how many ops per PG | |
707 | max_purge_ops = uint64_t(((double)pg_count / (double)mds_map.get_max_mds()) * | |
708 | cct->_conf->mds_max_purge_ops_per_pg); | |
709 | ||
710 | // User may also specify a hard limit, apply this if so. | |
711 | if (cct->_conf->mds_max_purge_ops) { | |
11fdf7f2 | 712 | max_purge_ops = std::min(max_purge_ops, cct->_conf->mds_max_purge_ops); |
7c673cae FG |
713 | } |
714 | } | |
715 | ||
92f5a8d4 | 716 | void PurgeQueue::handle_conf_change(const std::set<std::string>& changed, const MDSMap& mds_map) |
7c673cae FG |
717 | { |
718 | if (changed.count("mds_max_purge_ops") | |
719 | || changed.count("mds_max_purge_ops_per_pg")) { | |
720 | update_op_limit(mds_map); | |
721 | } else if (changed.count("mds_max_purge_files")) { | |
11fdf7f2 | 722 | std::lock_guard l(lock); |
7c673cae FG |
723 | if (in_flight.empty()) { |
724 | // We might have gone from zero to a finite limit, so | |
725 | // might need to kick off consume. | |
726 | dout(4) << "maybe start work again (max_purge_files=" | |
92f5a8d4 | 727 | << g_conf()->mds_max_purge_files << dendl; |
9f95a23c | 728 | finisher.queue(new LambdaContext([this](int r){ |
11fdf7f2 | 729 | std::lock_guard l(lock); |
7c673cae FG |
730 | _consume(); |
731 | })); | |
732 | } | |
733 | } | |
734 | } | |
735 | ||
736 | bool PurgeQueue::drain( | |
737 | uint64_t *progress, | |
738 | uint64_t *progress_total, | |
739 | size_t *in_flight_count | |
740 | ) | |
741 | { | |
11fdf7f2 | 742 | std::lock_guard l(lock); |
f64942e4 AA |
743 | |
744 | if (readonly) { | |
745 | dout(10) << "skipping drain; PurgeQueue is readonly" << dendl; | |
746 | return true; | |
747 | } | |
748 | ||
11fdf7f2 TL |
749 | ceph_assert(progress != nullptr); |
750 | ceph_assert(progress_total != nullptr); | |
751 | ceph_assert(in_flight_count != nullptr); | |
7c673cae FG |
752 | |
753 | const bool done = in_flight.empty() && ( | |
754 | journaler.get_read_pos() == journaler.get_write_pos()); | |
755 | if (done) { | |
756 | return true; | |
757 | } | |
758 | ||
759 | const uint64_t bytes_remaining = journaler.get_write_pos() | |
760 | - journaler.get_read_pos(); | |
761 | ||
762 | if (!draining) { | |
763 | // Start of draining: remember how much there was outstanding at | |
764 | // this point so that we can give a progress percentage later | |
765 | draining = true; | |
766 | ||
767 | // Life the op throttle as this daemon now has nothing to do but | |
768 | // drain the purge queue, so do it as fast as we can. | |
769 | max_purge_ops = 0xffff; | |
770 | } | |
771 | ||
11fdf7f2 | 772 | drain_initial = std::max(bytes_remaining, drain_initial); |
7c673cae FG |
773 | |
774 | *progress = drain_initial - bytes_remaining; | |
775 | *progress_total = drain_initial; | |
776 | *in_flight_count = in_flight.size(); | |
777 | ||
778 | return false; | |
779 | } | |
780 | ||
11fdf7f2 TL |
781 | std::string_view PurgeItem::get_type_str() const |
782 | { | |
783 | switch(action) { | |
784 | case PurgeItem::NONE: return "NONE"; | |
785 | case PurgeItem::PURGE_FILE: return "PURGE_FILE"; | |
786 | case PurgeItem::PURGE_DIR: return "PURGE_DIR"; | |
787 | case PurgeItem::TRUNCATE_FILE: return "TRUNCATE_FILE"; | |
788 | default: | |
789 | return "UNKNOWN"; | |
790 | } | |
791 | } | |
792 |