1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 * Ceph - scalable distributed file system
6 * Copyright (C) 2015 Red Hat
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.
15 #include "common/debug.h"
16 #include "mds/mdstypes.h"
17 #include "mds/CInode.h"
18 #include "mds/MDCache.h"
20 #include "PurgeQueue.h"
24 #define dout_context cct
25 #define dout_subsys ceph_subsys_mds
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 ";
32 void PurgeItem::encode(bufferlist
&bl
) const
34 ENCODE_START(1, 1, bl
);
35 ::encode((uint8_t)action
, bl
);
38 ::encode(layout
, bl
, CEPH_FEATURE_FS_FILE_LAYOUT_V2
);
39 ::encode(old_pools
, bl
);
41 ::encode(fragtree
, bl
);
45 void PurgeItem::decode(bufferlist::iterator
&p
)
48 ::decode((uint8_t&)action
, p
);
52 ::decode(old_pools
, p
);
54 ::decode(fragtree
, p
);
58 // TODO: if Objecter has any slow requests, take that as a hint and
59 // slow down our rate of purging (keep accepting pushes though)
60 PurgeQueue::PurgeQueue(
63 const int64_t metadata_pool_
,
70 metadata_pool(metadata_pool_
),
71 finisher(cct
, "PurgeQueue", "PQ_Finisher"),
73 filer(objecter_
, &finisher
),
75 journaler("pq", MDS_INO_PURGE_QUEUE
+ rank
, metadata_pool
,
76 CEPH_FS_ONDISK_MAGIC
, objecter_
, nullptr, 0,
83 delayed_flush(nullptr),
86 assert(cct
!= nullptr);
87 assert(on_error
!= nullptr);
88 assert(objecter
!= nullptr);
89 journaler
.set_write_error_handler(on_error
);
92 PurgeQueue::~PurgeQueue()
95 g_ceph_context
->get_perfcounters_collection()->remove(logger
.get());
100 void PurgeQueue::create_logger()
102 PerfCountersBuilder
pcb(g_ceph_context
, "purge_queue", l_pq_first
, l_pq_last
);
104 pcb
.add_u64_counter(l_pq_executed
, "pq_executed", "Purge queue tasks executed",
105 "purg", PerfCountersBuilder::PRIO_INTERESTING
);
107 pcb
.set_prio_default(PerfCountersBuilder::PRIO_USEFUL
);
108 pcb
.add_u64(l_pq_executing_ops
, "pq_executing_ops", "Purge queue ops in flight");
109 pcb
.add_u64(l_pq_executing
, "pq_executing", "Purge queue tasks in flight");
111 logger
.reset(pcb
.create_perf_counters());
112 g_ceph_context
->get_perfcounters_collection()->add(logger
.get());
115 void PurgeQueue::init()
117 Mutex::Locker
l(lock
);
119 assert(logger
!= nullptr);
125 void PurgeQueue::activate()
127 Mutex::Locker
l(lock
);
130 dout(10) << "skipping activate: PurgeQueue is readonly" << dendl
;
134 if (journaler
.get_read_pos() == journaler
.get_write_pos())
137 if (in_flight
.empty()) {
138 dout(4) << "start work (by drain)" << dendl
;
139 finisher
.queue(new FunctionContext([this](int r
) {
140 Mutex::Locker
l(lock
);
146 void PurgeQueue::shutdown()
148 Mutex::Locker
l(lock
);
150 journaler
.shutdown();
155 void PurgeQueue::open(Context
*completion
)
157 dout(4) << "opening" << dendl
;
159 Mutex::Locker
l(lock
);
162 waiting_for_recovery
.push_back(completion
);
164 journaler
.recover(new FunctionContext([this](int r
){
166 dout(1) << "Purge Queue not found, assuming this is an upgrade and "
167 "creating it." << dendl
;
170 Mutex::Locker
l(lock
);
171 dout(4) << "open complete" << dendl
;
173 // Journaler only guarantees entries before head write_pos have been
174 // fully flushed. Before appending new entries, we need to find and
175 // drop any partial written entry.
176 if (journaler
.last_committed
.write_pos
< journaler
.get_write_pos()) {
177 dout(4) << "recovering write_pos" << dendl
;
178 journaler
.set_read_pos(journaler
.last_committed
.write_pos
);
183 journaler
.set_writeable();
185 finish_contexts(g_ceph_context
, waiting_for_recovery
);
187 derr
<< "Error " << r
<< " loading Journaler" << dendl
;
193 void PurgeQueue::wait_for_recovery(Context
* c
)
195 Mutex::Locker
l(lock
);
198 } else if (readonly
) {
199 dout(10) << "cannot wait for recovery: PurgeQueue is readonly" << dendl
;
202 waiting_for_recovery
.push_back(c
);
206 void PurgeQueue::_recover()
208 assert(lock
.is_locked_by_me());
210 // Journaler::is_readable() adjusts write_pos if partial entry is encountered
212 if (!journaler
.is_readable() &&
213 !journaler
.get_error() &&
214 journaler
.get_read_pos() < journaler
.get_write_pos()) {
215 journaler
.wait_for_readable(new FunctionContext([this](int r
) {
216 Mutex::Locker
l(lock
);
222 if (journaler
.get_error()) {
223 int r
= journaler
.get_error();
224 derr
<< "Error " << r
<< " recovering write_pos" << dendl
;
229 if (journaler
.get_read_pos() == journaler
.get_write_pos()) {
230 dout(4) << "write_pos recovered" << dendl
;
231 // restore original read_pos
232 journaler
.set_read_pos(journaler
.last_committed
.expire_pos
);
233 journaler
.set_writeable();
235 finish_contexts(g_ceph_context
, waiting_for_recovery
);
240 bool readable
= journaler
.try_read_entry(bl
);
241 assert(readable
); // we checked earlier
245 void PurgeQueue::create(Context
*fin
)
247 dout(4) << "creating" << dendl
;
248 Mutex::Locker
l(lock
);
251 waiting_for_recovery
.push_back(fin
);
253 file_layout_t layout
= file_layout_t::get_default();
254 layout
.pool_id
= metadata_pool
;
255 journaler
.set_writeable();
256 journaler
.create(&layout
, JOURNAL_FORMAT_RESILIENT
);
257 journaler
.write_head(new FunctionContext([this](int r
) {
258 Mutex::Locker
l(lock
);
263 finish_contexts(g_ceph_context
, waiting_for_recovery
);
269 * The `completion` context will always be called back via a Finisher
271 void PurgeQueue::push(const PurgeItem
&pi
, Context
*completion
)
273 dout(4) << "pushing inode 0x" << std::hex
<< pi
.ino
<< std::dec
<< dendl
;
274 Mutex::Locker
l(lock
);
277 dout(10) << "cannot push inode: PurgeQueue is readonly" << dendl
;
278 completion
->complete(-EROFS
);
282 // Callers should have waited for open() before using us
283 assert(!journaler
.is_readonly());
288 journaler
.append_entry(bl
);
289 journaler
.wait_for_flush(completion
);
291 // Maybe go ahead and do something with it right away
292 bool could_consume
= _consume();
293 if (!could_consume
) {
294 // Usually, it is not necessary to explicitly flush here, because the reader
295 // will get flushes generated inside Journaler::is_readable. However,
296 // if we remain in a _can_consume()==false state for a long period then
297 // we should flush in order to allow MDCache to drop its strays rather
298 // than having them wait for purgequeue to progress.
299 if (!delayed_flush
) {
300 delayed_flush
= new FunctionContext([this](int r
){
301 delayed_flush
= nullptr;
305 timer
.add_event_after(
306 g_conf
->mds_purge_queue_busy_flush_period
,
312 uint32_t PurgeQueue::_calculate_ops(const PurgeItem
&item
) const
314 uint32_t ops_required
= 0;
315 if (item
.action
== PurgeItem::PURGE_DIR
) {
316 // Directory, count dirfrags to be deleted
317 std::list
<frag_t
> ls
;
318 if (!item
.fragtree
.is_leaf(frag_t())) {
319 item
.fragtree
.get_leaves(ls
);
321 // One for the root, plus any leaves
322 ops_required
= 1 + ls
.size();
324 // File, work out concurrent Filer::purge deletes
325 const uint64_t num
= (item
.size
> 0) ?
326 Striper::get_num_objects(item
.layout
, item
.size
) : 1;
328 ops_required
= MIN(num
, g_conf
->filer_max_purge_ops
);
330 // Account for removing (or zeroing) backtrace
333 // Account for deletions for old pools
334 if (item
.action
!= PurgeItem::TRUNCATE_FILE
) {
335 ops_required
+= item
.old_pools
.size();
342 bool PurgeQueue::_can_consume()
345 dout(10) << "can't consume: PurgeQueue is readonly" << dendl
;
349 dout(20) << ops_in_flight
<< "/" << max_purge_ops
<< " ops, "
350 << in_flight
.size() << "/" << g_conf
->mds_max_purge_files
351 << " files" << dendl
;
353 if (in_flight
.size() == 0 && cct
->_conf
->mds_max_purge_files
> 0) {
354 // Always permit consumption if nothing is in flight, so that the ops
355 // limit can never be so low as to forbid all progress (unless
356 // administrator has deliberately paused purging by setting max
357 // purge files to zero).
361 if (ops_in_flight
>= max_purge_ops
) {
362 dout(20) << "Throttling on op limit " << ops_in_flight
<< "/"
363 << max_purge_ops
<< dendl
;
367 if (in_flight
.size() >= cct
->_conf
->mds_max_purge_files
) {
368 dout(20) << "Throttling on item limit " << in_flight
.size()
369 << "/" << cct
->_conf
->mds_max_purge_files
<< dendl
;
376 void PurgeQueue::_go_readonly(int r
)
378 if (readonly
) return;
379 dout(1) << "going readonly because internal IO failed: " << strerror(-r
) << dendl
;
381 on_error
->complete(r
);
383 journaler
.set_readonly();
384 finish_contexts(g_ceph_context
, waiting_for_recovery
, r
);
387 bool PurgeQueue::_consume()
389 assert(lock
.is_locked_by_me());
391 bool could_consume
= false;
392 while(_can_consume()) {
395 // We are now going to read from the journal, so any proactive
396 // flush is no longer necessary. This is not functionally necessary
397 // but it can avoid generating extra fragmented flush IOs.
398 timer
.cancel_event(delayed_flush
);
399 delayed_flush
= nullptr;
402 if (int r
= journaler
.get_error()) {
403 derr
<< "Error " << r
<< " recovering write_pos" << dendl
;
405 return could_consume
;
408 if (!journaler
.is_readable()) {
409 dout(10) << " not readable right now" << dendl
;
410 // Because we are the writer and the reader of the journal
411 // via the same Journaler instance, we never need to reread_head
412 if (!journaler
.have_waiter()) {
413 journaler
.wait_for_readable(new FunctionContext([this](int r
) {
414 Mutex::Locker
l(lock
);
417 } else if (r
!= -EAGAIN
) {
423 return could_consume
;
426 could_consume
= true;
427 // The journaler is readable: consume an entry
429 bool readable
= journaler
.try_read_entry(bl
);
430 assert(readable
); // we checked earlier
432 dout(20) << " decoding entry" << dendl
;
434 bufferlist::iterator q
= bl
.begin();
437 } catch (const buffer::error
&err
) {
438 derr
<< "Decode error at read_pos=0x" << std::hex
439 << journaler
.get_read_pos() << dendl
;
442 dout(20) << " executing item (0x" << std::hex
<< item
.ino
443 << std::dec
<< ")" << dendl
;
444 _execute_item(item
, journaler
.get_read_pos());
447 dout(10) << " cannot consume right now" << dendl
;
449 return could_consume
;
452 void PurgeQueue::_execute_item(
453 const PurgeItem
&item
,
456 assert(lock
.is_locked_by_me());
458 in_flight
[expire_to
] = item
;
459 logger
->set(l_pq_executing
, in_flight
.size());
460 auto ops
= _calculate_ops(item
);
461 ops_in_flight
+= ops
;
462 logger
->set(l_pq_executing_ops
, ops_in_flight
);
464 SnapContext nullsnapc
;
466 C_GatherBuilder
gather(cct
);
467 if (item
.action
== PurgeItem::PURGE_FILE
) {
469 uint64_t num
= Striper::get_num_objects(item
.layout
, item
.size
);
470 dout(10) << " 0~" << item
.size
<< " objects 0~" << num
471 << " snapc " << item
.snapc
<< " on " << item
.ino
<< dendl
;
472 filer
.purge_range(item
.ino
, &item
.layout
, item
.snapc
,
473 0, num
, ceph::real_clock::now(), 0,
477 // remove the backtrace object if it was not purged
478 object_t oid
= CInode::get_object_name(item
.ino
, frag_t(), "");
479 if (!gather
.has_subs() || !item
.layout
.pool_ns
.empty()) {
480 object_locator_t
oloc(item
.layout
.pool_id
);
481 dout(10) << " remove backtrace object " << oid
482 << " pool " << oloc
.pool
<< " snapc " << item
.snapc
<< dendl
;
483 objecter
->remove(oid
, oloc
, item
.snapc
,
484 ceph::real_clock::now(), 0,
488 // remove old backtrace objects
489 for (const auto &p
: item
.old_pools
) {
490 object_locator_t
oloc(p
);
491 dout(10) << " remove backtrace object " << oid
492 << " old pool " << p
<< " snapc " << item
.snapc
<< dendl
;
493 objecter
->remove(oid
, oloc
, item
.snapc
,
494 ceph::real_clock::now(), 0,
497 } else if (item
.action
== PurgeItem::PURGE_DIR
) {
498 object_locator_t
oloc(metadata_pool
);
499 std::list
<frag_t
> frags
;
500 if (!item
.fragtree
.is_leaf(frag_t()))
501 item
.fragtree
.get_leaves(frags
);
502 frags
.push_back(frag_t());
503 for (const auto &frag
: frags
) {
504 object_t oid
= CInode::get_object_name(item
.ino
, frag
, "");
505 dout(10) << " remove dirfrag " << oid
<< dendl
;
506 objecter
->remove(oid
, oloc
, nullsnapc
,
507 ceph::real_clock::now(),
508 0, gather
.new_sub());
510 } else if (item
.action
== PurgeItem::TRUNCATE_FILE
) {
511 const uint64_t num
= Striper::get_num_objects(item
.layout
, item
.size
);
512 dout(10) << " 0~" << item
.size
<< " objects 0~" << num
513 << " snapc " << item
.snapc
<< " on " << item
.ino
<< dendl
;
515 // keep backtrace object
517 filer
.purge_range(item
.ino
, &item
.layout
, item
.snapc
,
518 1, num
- 1, ceph::real_clock::now(),
519 0, gather
.new_sub());
521 filer
.zero(item
.ino
, &item
.layout
, item
.snapc
,
522 0, item
.layout
.object_size
,
523 ceph::real_clock::now(),
524 0, true, gather
.new_sub());
526 derr
<< "Invalid item (action=" << item
.action
<< ") in purge queue, "
527 "dropping it" << dendl
;
528 ops_in_flight
-= ops
;
529 logger
->set(l_pq_executing_ops
, ops_in_flight
);
530 in_flight
.erase(expire_to
);
531 logger
->set(l_pq_executing
, in_flight
.size());
534 assert(gather
.has_subs());
536 gather
.set_finisher(new C_OnFinisher(
537 new FunctionContext([this, expire_to
](int r
){
538 Mutex::Locker
l(lock
);
539 _execute_item_complete(expire_to
);
543 // Have we gone idle? If so, do an extra write_head now instead of
544 // waiting for next flush after journaler_write_head_interval.
545 // Also do this periodically even if not idle, so that the persisted
546 // expire_pos doesn't fall too far behind our progress when consuming
547 // a very long queue.
548 if (in_flight
.empty() || journaler
.write_head_needed()) {
549 journaler
.write_head(nullptr);
556 void PurgeQueue::_execute_item_complete(
559 assert(lock
.is_locked_by_me());
560 dout(10) << "complete at 0x" << std::hex
<< expire_to
<< std::dec
<< dendl
;
561 assert(in_flight
.count(expire_to
) == 1);
563 auto iter
= in_flight
.find(expire_to
);
564 assert(iter
!= in_flight
.end());
565 if (iter
== in_flight
.begin()) {
566 // This was the lowest journal position in flight, so we can now
567 // safely expire the journal up to here.
568 dout(10) << "expiring to 0x" << std::hex
<< expire_to
<< std::dec
<< dendl
;
569 journaler
.set_expire_pos(expire_to
);
571 // This is completely fine, we're not supposed to purge files in
572 // order when doing them in parallel.
573 dout(10) << "non-sequential completion, not expiring anything" << dendl
;
576 ops_in_flight
-= _calculate_ops(iter
->second
);
577 logger
->set(l_pq_executing_ops
, ops_in_flight
);
579 dout(10) << "completed item for ino 0x" << std::hex
<< iter
->second
.ino
580 << std::dec
<< dendl
;
582 in_flight
.erase(iter
);
583 logger
->set(l_pq_executing
, in_flight
.size());
584 dout(10) << "in_flight.size() now " << in_flight
.size() << dendl
;
586 logger
->inc(l_pq_executed
);
589 void PurgeQueue::update_op_limit(const MDSMap
&mds_map
)
591 Mutex::Locker
l(lock
);
594 dout(10) << "skipping; PurgeQueue is readonly" << dendl
;
598 uint64_t pg_count
= 0;
599 objecter
->with_osdmap([&](const OSDMap
& o
) {
600 // Number of PGs across all data pools
601 const std::vector
<int64_t> &data_pools
= mds_map
.get_data_pools();
602 for (const auto dp
: data_pools
) {
603 if (o
.get_pg_pool(dp
) == NULL
) {
604 // It is possible that we have an older OSDMap than MDSMap,
605 // because we don't start watching every OSDMap until after
606 // MDSRank is initialized
607 dout(4) << " data pool " << dp
<< " not found in OSDMap" << dendl
;
610 pg_count
+= o
.get_pg_num(dp
);
614 // Work out a limit based on n_pgs / n_mdss, multiplied by the user's
615 // preference for how many ops per PG
616 max_purge_ops
= uint64_t(((double)pg_count
/ (double)mds_map
.get_max_mds()) *
617 cct
->_conf
->mds_max_purge_ops_per_pg
);
619 // User may also specify a hard limit, apply this if so.
620 if (cct
->_conf
->mds_max_purge_ops
) {
621 max_purge_ops
= MIN(max_purge_ops
, cct
->_conf
->mds_max_purge_ops
);
625 void PurgeQueue::handle_conf_change(const struct md_config_t
*conf
,
626 const std::set
<std::string
> &changed
,
627 const MDSMap
&mds_map
)
629 if (changed
.count("mds_max_purge_ops")
630 || changed
.count("mds_max_purge_ops_per_pg")) {
631 update_op_limit(mds_map
);
632 } else if (changed
.count("mds_max_purge_files")) {
633 Mutex::Locker
l(lock
);
635 if (in_flight
.empty()) {
636 // We might have gone from zero to a finite limit, so
637 // might need to kick off consume.
638 dout(4) << "maybe start work again (max_purge_files="
639 << conf
->mds_max_purge_files
<< dendl
;
640 finisher
.queue(new FunctionContext([this](int r
){
641 Mutex::Locker
l(lock
);
648 bool PurgeQueue::drain(
650 uint64_t *progress_total
,
651 size_t *in_flight_count
654 Mutex::Locker
l(lock
);
657 dout(10) << "skipping drain; PurgeQueue is readonly" << dendl
;
661 assert(progress
!= nullptr);
662 assert(progress_total
!= nullptr);
663 assert(in_flight_count
!= nullptr);
665 const bool done
= in_flight
.empty() && (
666 journaler
.get_read_pos() == journaler
.get_write_pos());
671 const uint64_t bytes_remaining
= journaler
.get_write_pos()
672 - journaler
.get_read_pos();
675 // Start of draining: remember how much there was outstanding at
676 // this point so that we can give a progress percentage later
679 // Life the op throttle as this daemon now has nothing to do but
680 // drain the purge queue, so do it as fast as we can.
681 max_purge_ops
= 0xffff;
684 drain_initial
= max(bytes_remaining
, drain_initial
);
686 *progress
= drain_initial
- bytes_remaining
;
687 *progress_total
= drain_initial
;
688 *in_flight_count
= in_flight
.size();