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 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
}
39 void PurgeItem::encode(bufferlist
&bl
) const
41 ENCODE_START(2, 1, bl
);
42 encode((uint8_t)action
, bl
);
45 encode(layout
, bl
, CEPH_FEATURE_FS_FILE_LAYOUT_V2
);
46 encode(old_pools
, bl
);
50 uint8_t static const pad
= 0xff;
51 for (unsigned int i
= 0; i
<pad_size
; i
++) {
57 void PurgeItem::decode(bufferlist::const_iterator
&p
)
64 // bad encoding introduced by v13.2.2
69 decode(raw_action
, p
);
70 action
= (Action
)raw_action
;
77 if (p
.get_off() > struct_end
)
78 throw buffer::end_of_buffer();
80 } catch (const buffer::error
&e
) {
86 decode(raw_action
, p
);
87 action
= (Action
)raw_action
;
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(
106 const int64_t metadata_pool_
,
112 metadata_pool(metadata_pool_
),
113 finisher(cct
, "PurgeQueue", "PQ_Finisher"),
115 filer(objecter_
, &finisher
),
117 journaler("pq", MDS_INO_PURGE_QUEUE
+ rank
, metadata_pool
,
118 CEPH_FS_ONDISK_MAGIC
, objecter_
, nullptr, 0,
122 ceph_assert(cct
!= nullptr);
123 ceph_assert(on_error
!= nullptr);
124 ceph_assert(objecter
!= nullptr);
125 journaler
.set_write_error_handler(on_error
);
128 PurgeQueue::~PurgeQueue()
131 g_ceph_context
->get_perfcounters_collection()->remove(logger
.get());
136 void PurgeQueue::create_logger()
138 PerfCountersBuilder
pcb(g_ceph_context
, "purge_queue", l_pq_first
, l_pq_last
);
140 pcb
.add_u64_counter(l_pq_executed
, "pq_executed", "Purge queue tasks executed",
141 "purg", PerfCountersBuilder::PRIO_INTERESTING
);
143 pcb
.set_prio_default(PerfCountersBuilder::PRIO_USEFUL
);
144 pcb
.add_u64(l_pq_executing_ops
, "pq_executing_ops", "Purge queue ops in flight");
145 pcb
.add_u64(l_pq_executing_ops_high_water
, "pq_executing_ops_high_water", "Maximum number of executing file purge ops");
146 pcb
.add_u64(l_pq_executing
, "pq_executing", "Purge queue tasks in flight");
147 pcb
.add_u64(l_pq_executing_high_water
, "pq_executing_high_water", "Maximum number of executing file purges");
148 pcb
.add_u64(l_pq_item_in_journal
, "pq_item_in_journal", "Purge item left in journal");
150 logger
.reset(pcb
.create_perf_counters());
151 g_ceph_context
->get_perfcounters_collection()->add(logger
.get());
154 void PurgeQueue::init()
156 std::lock_guard
l(lock
);
158 ceph_assert(logger
!= nullptr);
164 void PurgeQueue::activate()
166 std::lock_guard
l(lock
);
172 // calculate purge item serialized size stored in journal
173 // used to count how many items still left in journal later
175 purge_item_journal_size
= bl
.length() + journaler
.get_journal_envelope_size();
179 dout(10) << "skipping activate: PurgeQueue is readonly" << dendl
;
183 if (journaler
.get_read_pos() == journaler
.get_write_pos())
186 if (in_flight
.empty()) {
187 dout(4) << "start work (by drain)" << dendl
;
188 finisher
.queue(new LambdaContext([this](int r
) {
189 std::lock_guard
l(lock
);
195 void PurgeQueue::shutdown()
197 std::lock_guard
l(lock
);
199 journaler
.shutdown();
204 void PurgeQueue::open(Context
*completion
)
206 dout(4) << "opening" << dendl
;
208 std::lock_guard
l(lock
);
211 waiting_for_recovery
.push_back(completion
);
213 journaler
.recover(new LambdaContext([this](int r
){
215 dout(1) << "Purge Queue not found, assuming this is an upgrade and "
216 "creating it." << dendl
;
219 std::lock_guard
l(lock
);
220 dout(4) << "open complete" << dendl
;
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
);
232 journaler
.set_writeable();
234 finish_contexts(g_ceph_context
, waiting_for_recovery
);
236 derr
<< "Error " << r
<< " loading Journaler" << dendl
;
242 void PurgeQueue::wait_for_recovery(Context
* c
)
244 std::lock_guard
l(lock
);
247 } else if (readonly
) {
248 dout(10) << "cannot wait for recovery: PurgeQueue is readonly" << dendl
;
251 waiting_for_recovery
.push_back(c
);
255 void PurgeQueue::_recover()
257 ceph_assert(ceph_mutex_is_locked_by_me(lock
));
259 // Journaler::is_readable() adjusts write_pos if partial entry is encountered
261 if (!journaler
.is_readable() &&
262 !journaler
.get_error() &&
263 journaler
.get_read_pos() < journaler
.get_write_pos()) {
264 journaler
.wait_for_readable(new LambdaContext([this](int r
) {
265 std::lock_guard
l(lock
);
271 if (journaler
.get_error()) {
272 int r
= journaler
.get_error();
273 derr
<< "Error " << r
<< " recovering write_pos" << dendl
;
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();
284 finish_contexts(g_ceph_context
, waiting_for_recovery
);
289 bool readable
= journaler
.try_read_entry(bl
);
290 ceph_assert(readable
); // we checked earlier
294 void PurgeQueue::create(Context
*fin
)
296 dout(4) << "creating" << dendl
;
297 std::lock_guard
l(lock
);
300 waiting_for_recovery
.push_back(fin
);
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
);
306 journaler
.write_head(new LambdaContext([this](int r
) {
307 std::lock_guard
l(lock
);
312 finish_contexts(g_ceph_context
, waiting_for_recovery
);
318 * The `completion` context will always be called back via a Finisher
320 void PurgeQueue::push(const PurgeItem
&pi
, Context
*completion
)
322 dout(4) << "pushing inode " << pi
.ino
<< dendl
;
323 std::lock_guard
l(lock
);
326 dout(10) << "cannot push inode: PurgeQueue is readonly" << dendl
;
327 completion
->complete(-EROFS
);
331 // Callers should have waited for open() before using us
332 ceph_assert(!journaler
.is_readonly());
337 journaler
.append_entry(bl
);
338 journaler
.wait_for_flush(completion
);
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,
345 // if we remain in a _can_consume()==false state for a long period then
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
) {
349 delayed_flush
= new LambdaContext([this](int r
){
350 delayed_flush
= nullptr;
354 timer
.add_event_after(
355 g_conf()->mds_purge_queue_busy_flush_period
,
361 uint32_t PurgeQueue::_calculate_ops(const PurgeItem
&item
) const
363 uint32_t ops_required
= 0;
364 if (item
.action
== PurgeItem::PURGE_DIR
) {
365 // Directory, count dirfrags to be deleted
367 if (!item
.fragtree
.is_leaf(frag_t())) {
368 item
.fragtree
.get_leaves(leaves
);
370 // One for the root, plus any leaves
371 ops_required
= 1 + leaves
.size();
373 // File, work out concurrent Filer::purge deletes
374 // Account for removing (or zeroing) backtrace
375 const uint64_t num
= (item
.size
> 0) ?
376 Striper::get_num_objects(item
.layout
, item
.size
) : 1;
378 ops_required
= std::min(num
, g_conf()->filer_max_purge_ops
);
380 // Account for deletions for old pools
381 if (item
.action
!= PurgeItem::TRUNCATE_FILE
) {
382 ops_required
+= item
.old_pools
.size();
389 bool PurgeQueue::_can_consume()
392 dout(10) << "can't consume: PurgeQueue is readonly" << dendl
;
396 dout(20) << ops_in_flight
<< "/" << max_purge_ops
<< " ops, "
397 << in_flight
.size() << "/" << g_conf()->mds_max_purge_files
398 << " files" << dendl
;
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).
408 if (ops_in_flight
>= max_purge_ops
) {
409 dout(20) << "Throttling on op limit " << ops_in_flight
<< "/"
410 << max_purge_ops
<< dendl
;
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
;
423 void PurgeQueue::_go_readonly(int r
)
425 if (readonly
) return;
426 dout(1) << "going readonly because internal IO failed: " << strerror(-r
) << dendl
;
428 finisher
.queue(on_error
, r
);
430 journaler
.set_readonly();
431 finish_contexts(g_ceph_context
, waiting_for_recovery
, r
);
434 bool PurgeQueue::_consume()
436 ceph_assert(ceph_mutex_is_locked_by_me(lock
));
438 bool could_consume
= false;
439 while(_can_consume()) {
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;
449 if (int r
= journaler
.get_error()) {
450 derr
<< "Error " << r
<< " recovering write_pos" << dendl
;
452 return could_consume
;
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()) {
460 journaler
.wait_for_readable(new LambdaContext([this](int r
) {
461 std::lock_guard
l(lock
);
464 } else if (r
!= -EAGAIN
) {
470 return could_consume
;
473 could_consume
= true;
474 // The journaler is readable: consume an entry
476 bool readable
= journaler
.try_read_entry(bl
);
477 ceph_assert(readable
); // we checked earlier
479 dout(20) << " decoding entry" << dendl
;
481 auto q
= bl
.cbegin();
484 } catch (const buffer::error
&err
) {
485 derr
<< "Decode error at read_pos=0x" << std::hex
486 << journaler
.get_read_pos() << dendl
;
489 dout(20) << " executing item (" << item
.ino
<< ")" << dendl
;
490 _execute_item(item
, journaler
.get_read_pos());
493 dout(10) << " cannot consume right now" << dendl
;
495 return could_consume
;
498 void PurgeQueue::_execute_item(
499 const PurgeItem
&item
,
502 ceph_assert(ceph_mutex_is_locked_by_me(lock
));
504 in_flight
[expire_to
] = item
;
505 logger
->set(l_pq_executing
, in_flight
.size());
506 files_high_water
= std::max(files_high_water
, in_flight
.size());
507 logger
->set(l_pq_executing_high_water
, files_high_water
);
508 auto ops
= _calculate_ops(item
);
509 ops_in_flight
+= ops
;
510 logger
->set(l_pq_executing_ops
, ops_in_flight
);
511 ops_high_water
= std::max(ops_high_water
, ops_in_flight
);
512 logger
->set(l_pq_executing_ops_high_water
, ops_high_water
);
514 SnapContext nullsnapc
;
516 C_GatherBuilder
gather(cct
);
517 if (item
.action
== PurgeItem::PURGE_FILE
) {
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,
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,
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,
547 } else if (item
.action
== PurgeItem::PURGE_DIR
) {
548 object_locator_t
oloc(metadata_pool
);
550 if (!item
.fragtree
.is_leaf(frag_t()))
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
, "");
555 dout(10) << " remove dirfrag " << oid
<< dendl
;
556 objecter
->remove(oid
, oloc
, nullsnapc
,
557 ceph::real_clock::now(),
558 0, gather
.new_sub());
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
;
565 // keep backtrace object
567 filer
.purge_range(item
.ino
, &item
.layout
, item
.snapc
,
568 1, num
- 1, ceph::real_clock::now(),
569 0, gather
.new_sub());
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());
576 derr
<< "Invalid item (action=" << item
.action
<< ") in purge queue, "
577 "dropping it" << dendl
;
578 ops_in_flight
-= ops
;
579 logger
->set(l_pq_executing_ops
, ops_in_flight
);
580 ops_high_water
= std::max(ops_high_water
, ops_in_flight
);
581 logger
->set(l_pq_executing_ops_high_water
, ops_high_water
);
582 in_flight
.erase(expire_to
);
583 logger
->set(l_pq_executing
, in_flight
.size());
584 files_high_water
= std::max(files_high_water
, in_flight
.size());
585 logger
->set(l_pq_executing_high_water
, files_high_water
);
588 ceph_assert(gather
.has_subs());
590 gather
.set_finisher(new C_OnFinisher(
591 new LambdaContext([this, expire_to
](int r
){
592 std::lock_guard
l(lock
);
594 if (r
== -EBLACKLISTED
) {
595 finisher
.queue(on_error
, r
);
600 _execute_item_complete(expire_to
);
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.
609 (in_flight
.empty() || journaler
.write_head_needed())) {
610 journaler
.write_head(nullptr);
617 void PurgeQueue::_execute_item_complete(
620 ceph_assert(ceph_mutex_is_locked_by_me(lock
));
621 dout(10) << "complete at 0x" << std::hex
<< expire_to
<< std::dec
<< dendl
;
622 ceph_assert(in_flight
.count(expire_to
) == 1);
624 auto iter
= in_flight
.find(expire_to
);
625 ceph_assert(iter
!= in_flight
.end());
626 if (iter
== in_flight
.begin()) {
627 uint64_t pos
= expire_to
;
628 if (!pending_expire
.empty()) {
631 if (n
== in_flight
.end()) {
632 pos
= *pending_expire
.rbegin();
633 pending_expire
.clear();
635 auto p
= pending_expire
.begin();
640 pending_expire
.erase(p
++);
641 } while (p
!= pending_expire
.end());
644 dout(10) << "expiring to 0x" << std::hex
<< pos
<< std::dec
<< dendl
;
645 journaler
.set_expire_pos(pos
);
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
;
650 pending_expire
.insert(expire_to
);
653 ops_in_flight
-= _calculate_ops(iter
->second
);
654 logger
->set(l_pq_executing_ops
, ops_in_flight
);
655 ops_high_water
= std::max(ops_high_water
, ops_in_flight
);
656 logger
->set(l_pq_executing_ops_high_water
, ops_high_water
);
658 dout(10) << "completed item for ino " << iter
->second
.ino
<< dendl
;
660 in_flight
.erase(iter
);
661 logger
->set(l_pq_executing
, in_flight
.size());
662 files_high_water
= std::max(files_high_water
, in_flight
.size());
663 logger
->set(l_pq_executing_high_water
, files_high_water
);
664 dout(10) << "in_flight.size() now " << in_flight
.size() << dendl
;
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
;
676 logger
->set(l_pq_item_in_journal
, item_num
);
677 logger
->inc(l_pq_executed
);
680 void PurgeQueue::update_op_limit(const MDSMap
&mds_map
)
682 std::lock_guard
l(lock
);
685 dout(10) << "skipping; PurgeQueue is readonly" << dendl
;
689 uint64_t pg_count
= 0;
690 objecter
->with_osdmap([&](const OSDMap
& o
) {
691 // Number of PGs across all data pools
692 const std::vector
<int64_t> &data_pools
= mds_map
.get_data_pools();
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
;
701 pg_count
+= o
.get_pg_num(dp
);
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
);
710 // User may also specify a hard limit, apply this if so.
711 if (cct
->_conf
->mds_max_purge_ops
) {
712 max_purge_ops
= std::min(max_purge_ops
, cct
->_conf
->mds_max_purge_ops
);
716 void PurgeQueue::handle_conf_change(const std::set
<std::string
>& changed
, const MDSMap
& mds_map
)
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")) {
722 std::lock_guard
l(lock
);
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="
727 << g_conf()->mds_max_purge_files
<< dendl
;
728 finisher
.queue(new LambdaContext([this](int r
){
729 std::lock_guard
l(lock
);
736 bool PurgeQueue::drain(
738 uint64_t *progress_total
,
739 size_t *in_flight_count
742 std::lock_guard
l(lock
);
745 dout(10) << "skipping drain; PurgeQueue is readonly" << dendl
;
749 ceph_assert(progress
!= nullptr);
750 ceph_assert(progress_total
!= nullptr);
751 ceph_assert(in_flight_count
!= nullptr);
753 const bool done
= in_flight
.empty() && (
754 journaler
.get_read_pos() == journaler
.get_write_pos());
759 const uint64_t bytes_remaining
= journaler
.get_write_pos()
760 - journaler
.get_read_pos();
763 // Start of draining: remember how much there was outstanding at
764 // this point so that we can give a progress percentage later
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;
772 drain_initial
= std::max(bytes_remaining
, drain_initial
);
774 *progress
= drain_initial
- bytes_remaining
;
775 *progress_total
= drain_initial
;
776 *in_flight_count
= in_flight
.size();
781 std::string_view
PurgeItem::get_type_str() const
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";