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"
23 #define dout_context cct
24 #define dout_subsys ceph_subsys_mds
26 #define dout_prefix _prefix(_dout, rank) << __func__ << ": "
27 static ostream
& _prefix(std::ostream
*_dout
, mds_rank_t rank
) {
28 return *_dout
<< "mds." << rank
<< ".purge_queue ";
31 void PurgeItem::encode(bufferlist
&bl
) const
33 ENCODE_START(1, 1, bl
);
34 ::encode((uint8_t)action
, bl
);
37 ::encode(layout
, bl
, CEPH_FEATURE_FS_FILE_LAYOUT_V2
);
38 ::encode(old_pools
, bl
);
40 ::encode(fragtree
, bl
);
44 void PurgeItem::decode(bufferlist::iterator
&p
)
47 ::decode((uint8_t&)action
, p
);
51 ::decode(old_pools
, p
);
53 ::decode(fragtree
, p
);
57 // TODO: if Objecter has any slow requests, take that as a hint and
58 // slow down our rate of purging (keep accepting pushes though)
59 PurgeQueue::PurgeQueue(
62 const int64_t metadata_pool_
,
69 metadata_pool(metadata_pool_
),
70 finisher(cct
, "PurgeQueue", "PQ_Finisher"),
72 filer(objecter_
, &finisher
),
74 journaler("pq", MDS_INO_PURGE_QUEUE
+ rank
, metadata_pool
,
75 CEPH_FS_ONDISK_MAGIC
, objecter_
, nullptr, 0,
82 delayed_flush(nullptr)
84 assert(cct
!= nullptr);
85 assert(on_error
!= nullptr);
86 assert(objecter
!= nullptr);
87 journaler
.set_write_error_handler(on_error
);
90 PurgeQueue::~PurgeQueue()
93 g_ceph_context
->get_perfcounters_collection()->remove(logger
.get());
97 void PurgeQueue::create_logger()
99 PerfCountersBuilder
pcb(g_ceph_context
,
100 "purge_queue", l_pq_first
, l_pq_last
);
101 pcb
.add_u64(l_pq_executing_ops
, "pq_executing_ops", "Purge queue ops in flight");
102 pcb
.add_u64(l_pq_executing
, "pq_executing", "Purge queue tasks in flight");
103 pcb
.add_u64_counter(l_pq_executed
, "pq_executed", "Purge queue tasks executed", "purg",
104 PerfCountersBuilder::PRIO_INTERESTING
);
106 logger
.reset(pcb
.create_perf_counters());
107 g_ceph_context
->get_perfcounters_collection()->add(logger
.get());
110 void PurgeQueue::init()
112 Mutex::Locker
l(lock
);
114 assert(logger
!= nullptr);
120 void PurgeQueue::activate()
122 Mutex::Locker
l(lock
);
123 if (journaler
.get_read_pos() == journaler
.get_write_pos())
126 if (in_flight
.empty()) {
127 dout(4) << "start work (by drain)" << dendl
;
128 finisher
.queue(new FunctionContext([this](int r
) {
129 Mutex::Locker
l(lock
);
135 void PurgeQueue::shutdown()
137 Mutex::Locker
l(lock
);
139 journaler
.shutdown();
144 void PurgeQueue::open(Context
*completion
)
146 dout(4) << "opening" << dendl
;
148 Mutex::Locker
l(lock
);
150 journaler
.recover(new FunctionContext([this, completion
](int r
){
152 dout(1) << "Purge Queue not found, assuming this is an upgrade and "
153 "creating it." << dendl
;
156 Mutex::Locker
l(lock
);
157 dout(4) << "open complete" << dendl
;
159 // Journaler only guarantees entries before head write_pos have been
160 // fully flushed. Before appending new entries, we need to find and
161 // drop any partial written entry.
162 if (journaler
.last_committed
.write_pos
< journaler
.get_write_pos()) {
163 dout(4) << "recovering write_pos" << dendl
;
164 journaler
.set_read_pos(journaler
.last_committed
.write_pos
);
165 _recover(completion
);
169 journaler
.set_writeable();
170 completion
->complete(0);
172 derr
<< "Error " << r
<< " loading Journaler" << dendl
;
173 on_error
->complete(r
);
179 void PurgeQueue::_recover(Context
*completion
)
181 assert(lock
.is_locked_by_me());
183 // Journaler::is_readable() adjusts write_pos if partial entry is encountered
185 if (!journaler
.is_readable() &&
186 !journaler
.get_error() &&
187 journaler
.get_read_pos() < journaler
.get_write_pos()) {
188 journaler
.wait_for_readable(new FunctionContext([this, completion
](int r
) {
189 Mutex::Locker
l(lock
);
190 _recover(completion
);
195 if (journaler
.get_error()) {
196 int r
= journaler
.get_error();
197 derr
<< "Error " << r
<< " recovering write_pos" << dendl
;
198 on_error
->complete(r
);
202 if (journaler
.get_read_pos() == journaler
.get_write_pos()) {
203 dout(4) << "write_pos recovered" << dendl
;
204 // restore original read_pos
205 journaler
.set_read_pos(journaler
.last_committed
.expire_pos
);
206 journaler
.set_writeable();
207 completion
->complete(0);
212 bool readable
= journaler
.try_read_entry(bl
);
213 assert(readable
); // we checked earlier
217 void PurgeQueue::create(Context
*fin
)
219 dout(4) << "creating" << dendl
;
220 Mutex::Locker
l(lock
);
222 file_layout_t layout
= file_layout_t::get_default();
223 layout
.pool_id
= metadata_pool
;
224 journaler
.set_writeable();
225 journaler
.create(&layout
, JOURNAL_FORMAT_RESILIENT
);
226 journaler
.write_head(fin
);
230 * The `completion` context will always be called back via a Finisher
232 void PurgeQueue::push(const PurgeItem
&pi
, Context
*completion
)
234 dout(4) << "pushing inode 0x" << std::hex
<< pi
.ino
<< std::dec
<< dendl
;
235 Mutex::Locker
l(lock
);
237 // Callers should have waited for open() before using us
238 assert(!journaler
.is_readonly());
243 journaler
.append_entry(bl
);
244 journaler
.wait_for_flush(completion
);
246 // Maybe go ahead and do something with it right away
247 bool could_consume
= _consume();
248 if (!could_consume
) {
249 // Usually, it is not necessary to explicitly flush here, because the reader
250 // will get flushes generated inside Journaler::is_readable. However,
251 // if we remain in a can_consume()==false state for a long period then
252 // we should flush in order to allow MDCache to drop its strays rather
253 // than having them wait for purgequeue to progress.
254 if (!delayed_flush
) {
255 delayed_flush
= new FunctionContext([this](int r
){
256 delayed_flush
= nullptr;
260 timer
.add_event_after(
261 g_conf
->mds_purge_queue_busy_flush_period
,
267 uint32_t PurgeQueue::_calculate_ops(const PurgeItem
&item
) const
269 uint32_t ops_required
= 0;
270 if (item
.action
== PurgeItem::PURGE_DIR
) {
271 // Directory, count dirfrags to be deleted
272 std::list
<frag_t
> ls
;
273 if (!item
.fragtree
.is_leaf(frag_t())) {
274 item
.fragtree
.get_leaves(ls
);
276 // One for the root, plus any leaves
277 ops_required
= 1 + ls
.size();
279 // File, work out concurrent Filer::purge deletes
280 const uint64_t num
= (item
.size
> 0) ?
281 Striper::get_num_objects(item
.layout
, item
.size
) : 1;
283 ops_required
= MIN(num
, g_conf
->filer_max_purge_ops
);
285 // Account for removing (or zeroing) backtrace
288 // Account for deletions for old pools
289 if (item
.action
!= PurgeItem::TRUNCATE_FILE
) {
290 ops_required
+= item
.old_pools
.size();
297 bool PurgeQueue::can_consume()
299 dout(20) << ops_in_flight
<< "/" << max_purge_ops
<< " ops, "
300 << in_flight
.size() << "/" << g_conf
->mds_max_purge_files
301 << " files" << dendl
;
303 if (in_flight
.size() == 0 && cct
->_conf
->mds_max_purge_files
> 0) {
304 // Always permit consumption if nothing is in flight, so that the ops
305 // limit can never be so low as to forbid all progress (unless
306 // administrator has deliberately paused purging by setting max
307 // purge files to zero).
311 if (ops_in_flight
>= max_purge_ops
) {
312 dout(20) << "Throttling on op limit " << ops_in_flight
<< "/"
313 << max_purge_ops
<< dendl
;
317 if (in_flight
.size() >= cct
->_conf
->mds_max_purge_files
) {
318 dout(20) << "Throttling on item limit " << in_flight
.size()
319 << "/" << cct
->_conf
->mds_max_purge_files
<< dendl
;
326 bool PurgeQueue::_consume()
328 assert(lock
.is_locked_by_me());
330 bool could_consume
= false;
331 while(can_consume()) {
332 could_consume
= true;
335 // We are now going to read from the journal, so any proactive
336 // flush is no longer necessary. This is not functionally necessary
337 // but it can avoid generating extra fragmented flush IOs.
338 timer
.cancel_event(delayed_flush
);
339 delayed_flush
= nullptr;
342 if (!journaler
.is_readable()) {
343 dout(10) << " not readable right now" << dendl
;
344 // Because we are the writer and the reader of the journal
345 // via the same Journaler instance, we never need to reread_head
346 if (!journaler
.have_waiter()) {
347 journaler
.wait_for_readable(new FunctionContext([this](int r
) {
348 Mutex::Locker
l(lock
);
355 return could_consume
;
358 // The journaler is readable: consume an entry
360 bool readable
= journaler
.try_read_entry(bl
);
361 assert(readable
); // we checked earlier
363 dout(20) << " decoding entry" << dendl
;
365 bufferlist::iterator q
= bl
.begin();
368 } catch (const buffer::error
&err
) {
369 derr
<< "Decode error at read_pos=0x" << std::hex
370 << journaler
.get_read_pos() << dendl
;
371 on_error
->complete(0);
373 dout(20) << " executing item (0x" << std::hex
<< item
.ino
374 << std::dec
<< ")" << dendl
;
375 _execute_item(item
, journaler
.get_read_pos());
378 dout(10) << " cannot consume right now" << dendl
;
380 return could_consume
;
383 void PurgeQueue::_execute_item(
384 const PurgeItem
&item
,
387 assert(lock
.is_locked_by_me());
389 in_flight
[expire_to
] = item
;
390 logger
->set(l_pq_executing
, in_flight
.size());
391 ops_in_flight
+= _calculate_ops(item
);
392 logger
->set(l_pq_executing_ops
, ops_in_flight
);
394 SnapContext nullsnapc
;
396 C_GatherBuilder
gather(cct
);
397 if (item
.action
== PurgeItem::PURGE_FILE
) {
399 uint64_t num
= Striper::get_num_objects(item
.layout
, item
.size
);
400 dout(10) << " 0~" << item
.size
<< " objects 0~" << num
401 << " snapc " << item
.snapc
<< " on " << item
.ino
<< dendl
;
402 filer
.purge_range(item
.ino
, &item
.layout
, item
.snapc
,
403 0, num
, ceph::real_clock::now(), 0,
407 // remove the backtrace object if it was not purged
408 object_t oid
= CInode::get_object_name(item
.ino
, frag_t(), "");
409 if (!gather
.has_subs() || !item
.layout
.pool_ns
.empty()) {
410 object_locator_t
oloc(item
.layout
.pool_id
);
411 dout(10) << " remove backtrace object " << oid
412 << " pool " << oloc
.pool
<< " snapc " << item
.snapc
<< dendl
;
413 objecter
->remove(oid
, oloc
, item
.snapc
,
414 ceph::real_clock::now(), 0,
418 // remove old backtrace objects
419 for (const auto &p
: item
.old_pools
) {
420 object_locator_t
oloc(p
);
421 dout(10) << " remove backtrace object " << oid
422 << " old pool " << p
<< " snapc " << item
.snapc
<< dendl
;
423 objecter
->remove(oid
, oloc
, item
.snapc
,
424 ceph::real_clock::now(), 0,
427 } else if (item
.action
== PurgeItem::PURGE_DIR
) {
428 object_locator_t
oloc(metadata_pool
);
429 std::list
<frag_t
> frags
;
430 if (!item
.fragtree
.is_leaf(frag_t()))
431 item
.fragtree
.get_leaves(frags
);
432 frags
.push_back(frag_t());
433 for (const auto &frag
: frags
) {
434 object_t oid
= CInode::get_object_name(item
.ino
, frag
, "");
435 dout(10) << " remove dirfrag " << oid
<< dendl
;
436 objecter
->remove(oid
, oloc
, nullsnapc
,
437 ceph::real_clock::now(),
438 0, gather
.new_sub());
440 } else if (item
.action
== PurgeItem::TRUNCATE_FILE
) {
441 const uint64_t num
= Striper::get_num_objects(item
.layout
, item
.size
);
442 dout(10) << " 0~" << item
.size
<< " objects 0~" << num
443 << " snapc " << item
.snapc
<< " on " << item
.ino
<< dendl
;
445 // keep backtrace object
447 filer
.purge_range(item
.ino
, &item
.layout
, item
.snapc
,
448 1, num
- 1, ceph::real_clock::now(),
449 0, gather
.new_sub());
451 filer
.zero(item
.ino
, &item
.layout
, item
.snapc
,
452 0, item
.layout
.object_size
,
453 ceph::real_clock::now(),
454 0, true, gather
.new_sub());
456 derr
<< "Invalid item (action=" << item
.action
<< ") in purge queue, "
457 "dropping it" << dendl
;
458 in_flight
.erase(expire_to
);
459 logger
->set(l_pq_executing
, in_flight
.size());
462 assert(gather
.has_subs());
464 gather
.set_finisher(new C_OnFinisher(
465 new FunctionContext([this, expire_to
](int r
){
466 Mutex::Locker
l(lock
);
467 _execute_item_complete(expire_to
);
471 // Have we gone idle? If so, do an extra write_head now instead of
472 // waiting for next flush after journaler_write_head_interval.
473 // Also do this periodically even if not idle, so that the persisted
474 // expire_pos doesn't fall too far behind our progress when consuming
475 // a very long queue.
476 if (in_flight
.empty() || journaler
.write_head_needed()) {
477 journaler
.write_head(new FunctionContext([this](int r
){
486 void PurgeQueue::_execute_item_complete(
489 assert(lock
.is_locked_by_me());
490 dout(10) << "complete at 0x" << std::hex
<< expire_to
<< std::dec
<< dendl
;
491 assert(in_flight
.count(expire_to
) == 1);
493 auto iter
= in_flight
.find(expire_to
);
494 assert(iter
!= in_flight
.end());
495 if (iter
== in_flight
.begin()) {
496 // This was the lowest journal position in flight, so we can now
497 // safely expire the journal up to here.
498 dout(10) << "expiring to 0x" << std::hex
<< expire_to
<< std::dec
<< dendl
;
499 journaler
.set_expire_pos(expire_to
);
501 // This is completely fine, we're not supposed to purge files in
502 // order when doing them in parallel.
503 dout(10) << "non-sequential completion, not expiring anything" << dendl
;
506 ops_in_flight
-= _calculate_ops(iter
->second
);
507 logger
->set(l_pq_executing_ops
, ops_in_flight
);
509 dout(10) << "completed item for ino 0x" << std::hex
<< iter
->second
.ino
510 << std::dec
<< dendl
;
512 in_flight
.erase(iter
);
513 logger
->set(l_pq_executing
, in_flight
.size());
514 dout(10) << "in_flight.size() now " << in_flight
.size() << dendl
;
516 logger
->inc(l_pq_executed
);
519 void PurgeQueue::update_op_limit(const MDSMap
&mds_map
)
521 Mutex::Locker
l(lock
);
523 uint64_t pg_count
= 0;
524 objecter
->with_osdmap([&](const OSDMap
& o
) {
525 // Number of PGs across all data pools
526 const std::vector
<int64_t> &data_pools
= mds_map
.get_data_pools();
527 for (const auto dp
: data_pools
) {
528 if (o
.get_pg_pool(dp
) == NULL
) {
529 // It is possible that we have an older OSDMap than MDSMap,
530 // because we don't start watching every OSDMap until after
531 // MDSRank is initialized
532 dout(4) << " data pool " << dp
<< " not found in OSDMap" << dendl
;
535 pg_count
+= o
.get_pg_num(dp
);
539 // Work out a limit based on n_pgs / n_mdss, multiplied by the user's
540 // preference for how many ops per PG
541 max_purge_ops
= uint64_t(((double)pg_count
/ (double)mds_map
.get_max_mds()) *
542 cct
->_conf
->mds_max_purge_ops_per_pg
);
544 // User may also specify a hard limit, apply this if so.
545 if (cct
->_conf
->mds_max_purge_ops
) {
546 max_purge_ops
= MIN(max_purge_ops
, cct
->_conf
->mds_max_purge_ops
);
550 void PurgeQueue::handle_conf_change(const struct md_config_t
*conf
,
551 const std::set
<std::string
> &changed
,
552 const MDSMap
&mds_map
)
554 if (changed
.count("mds_max_purge_ops")
555 || changed
.count("mds_max_purge_ops_per_pg")) {
556 update_op_limit(mds_map
);
557 } else if (changed
.count("mds_max_purge_files")) {
558 Mutex::Locker
l(lock
);
560 if (in_flight
.empty()) {
561 // We might have gone from zero to a finite limit, so
562 // might need to kick off consume.
563 dout(4) << "maybe start work again (max_purge_files="
564 << conf
->mds_max_purge_files
<< dendl
;
565 finisher
.queue(new FunctionContext([this](int r
){
566 Mutex::Locker
l(lock
);
573 bool PurgeQueue::drain(
575 uint64_t *progress_total
,
576 size_t *in_flight_count
579 assert(progress
!= nullptr);
580 assert(progress_total
!= nullptr);
581 assert(in_flight_count
!= nullptr);
583 const bool done
= in_flight
.empty() && (
584 journaler
.get_read_pos() == journaler
.get_write_pos());
589 const uint64_t bytes_remaining
= journaler
.get_write_pos()
590 - journaler
.get_read_pos();
593 // Start of draining: remember how much there was outstanding at
594 // this point so that we can give a progress percentage later
597 // Life the op throttle as this daemon now has nothing to do but
598 // drain the purge queue, so do it as fast as we can.
599 max_purge_ops
= 0xffff;
602 drain_initial
= max(bytes_remaining
, drain_initial
);
604 *progress
= drain_initial
- bytes_remaining
;
605 *progress_total
= drain_initial
;
606 *in_flight_count
= in_flight
.size();