1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
|
/**
* Copyright (C) 2018-present MongoDB, Inc.
*
* This program is free software: you can redistribute it and/or modify
* it under the terms of the Server Side Public License, version 1,
* as published by MongoDB, Inc.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* Server Side Public License for more details.
*
* You should have received a copy of the Server Side Public License
* along with this program. If not, see
* <http://www.mongodb.com/licensing/server-side-public-license>.
*
* As a special exception, the copyright holders give permission to link the
* code of portions of this program with the OpenSSL library under certain
* conditions as described in each individual source file and distribute
* linked combinations including the program with the OpenSSL library. You
* must comply with the Server Side Public License in all respects for
* all of the code used other than as permitted herein. If you modify file(s)
* with this exception, you may extend this exception to your version of the
* file(s), but you are not obligated to do so. If you do not wish to do so,
* delete this exception statement from your version. If you delete this
* exception statement from all source files in the program, then also delete
* it in the license file.
*/
#define MONGO_LOG_DEFAULT_COMPONENT ::mongo::logger::LogComponent::kStorage
#include "mongo/platform/basic.h"
#include "mongo/db/db_raii.h"
#include "mongo/db/catalog/database_holder.h"
#include "mongo/db/concurrency/locker.h"
#include "mongo/db/curop.h"
#include "mongo/db/db_raii_gen.h"
#include "mongo/db/repl/replication_coordinator.h"
#include "mongo/db/s/collection_sharding_state.h"
#include "mongo/logv2/log.h"
namespace mongo {
namespace {
const boost::optional<int> kDoNotChangeProfilingLevel = boost::none;
// TODO: SERVER-44105 remove
// If set to false, secondary reads should wait behind the PBW lock.
// Does nothing if gAllowSecondaryReadsDuringBatchApplication setting is false.
const auto allowSecondaryReadsDuringBatchApplication_DONT_USE =
OperationContext::declareDecoration<boost::optional<bool>>();
} // namespace
AutoStatsTracker::AutoStatsTracker(OperationContext* opCtx,
const NamespaceString& nss,
Top::LockType lockType,
LogMode logMode,
boost::optional<int> dbProfilingLevel,
Date_t deadline)
: _opCtx(opCtx), _lockType(lockType), _nss(nss), _logMode(logMode) {
if (_logMode == LogMode::kUpdateTop) {
return;
}
if (!dbProfilingLevel) {
// No profiling level was determined, attempt to read the profiling level from the Database
// object. Since we are only reading the in-memory profiling level out of the database
// object (which is configured on a per-node basis and not replicated or persisted), we
// never need to conflict with secondary batch application.
ShouldNotConflictWithSecondaryBatchApplicationBlock noConflict(opCtx->lockState());
AutoGetDb autoDb(_opCtx, _nss.db(), MODE_IS, deadline);
if (autoDb.getDb()) {
dbProfilingLevel = autoDb.getDb()->getProfilingLevel();
}
}
stdx::lock_guard<Client> clientLock(*_opCtx->getClient());
CurOp::get(_opCtx)->enter_inlock(_nss.ns().c_str(), dbProfilingLevel);
}
AutoStatsTracker::~AutoStatsTracker() {
if (_logMode == LogMode::kUpdateCurOp) {
return;
}
auto curOp = CurOp::get(_opCtx);
Top::get(_opCtx->getServiceContext())
.record(_opCtx,
_nss.ns(),
curOp->getLogicalOp(),
_lockType,
durationCount<Microseconds>(curOp->elapsedTimeExcludingPauses()),
curOp->isCommand(),
curOp->getReadWriteType());
}
AutoGetCollectionForRead::AutoGetCollectionForRead(OperationContext* opCtx,
const NamespaceStringOrUUID& nsOrUUID,
AutoGetCollection::ViewMode viewMode,
Date_t deadline) {
// Don't take the ParallelBatchWriterMode lock when the server parameter is set and our
// storage engine supports snapshot reads.
if (gAllowSecondaryReadsDuringBatchApplication.load() &&
allowSecondaryReadsDuringBatchApplication_DONT_USE(opCtx).value_or(true) &&
opCtx->getServiceContext()->getStorageEngine()->supportsReadConcernSnapshot()) {
_shouldNotConflictWithSecondaryBatchApplicationBlock.emplace(opCtx->lockState());
}
const auto collectionLockMode = getLockModeForQuery(opCtx, nsOrUUID.nss());
_autoColl.emplace(opCtx, nsOrUUID, collectionLockMode, viewMode, deadline);
// If the read source is explicitly set to kNoTimestamp, we read the most up to date data and do
// not consider reading at last applied (e.g. FTDC needs that).
if (opCtx->recoveryUnit()->getTimestampReadSource() == RecoveryUnit::ReadSource::kNoTimestamp)
return;
repl::ReplicationCoordinator* const replCoord = repl::ReplicationCoordinator::get(opCtx);
const auto readConcernLevel = repl::ReadConcernArgs::get(opCtx).getLevel();
// If the collection doesn't exist or disappears after releasing locks and waiting, there is no
// need to check for pending catalog changes.
while (auto coll = _autoColl->getCollection()) {
auto readSource = opCtx->recoveryUnit()->getTimestampReadSource();
auto minSnapshot = coll->getMinimumVisibleSnapshot();
auto mySnapshot = opCtx->recoveryUnit()->getPointInTimeReadTimestamp();
// If we are reading at a provided timestamp earlier than the latest catalog changes, then
// we must return an error.
if (readSource == RecoveryUnit::ReadSource::kProvided && minSnapshot &&
(*mySnapshot < *minSnapshot)) {
uasserted(ErrorCodes::SnapshotUnavailable,
str::stream()
<< "Unable to read from a snapshot due to pending collection catalog "
"changes; please retry the operation. Snapshot timestamp is "
<< mySnapshot->toString() << ". Collection minimum is "
<< minSnapshot->toString());
}
// During batch application on secondaries, there is a potential to read inconsistent states
// that would normally be protected by the PBWM lock. In order to serve secondary reads
// during this period, we default to not acquiring the lock (by setting
// _shouldNotConflictWithSecondaryBatchApplicationBlock). On primaries, we always read at a
// consistent time, so not taking the PBWM lock is not a problem. On secondaries, we have to
// guarantee we read at a consistent state, so we must read at the last applied timestamp,
// which is set after each complete batch.
//
// If an attempt to read at the last applied timestamp is unsuccessful because there are
// pending catalog changes that occur after the last applied timestamp, we release our locks
// and try again with the PBWM lock (by unsetting
// _shouldNotConflictWithSecondaryBatchApplicationBlock).
const NamespaceString nss = coll->ns();
bool readAtLastAppliedTimestamp =
_shouldReadAtLastAppliedTimestamp(opCtx, nss, readConcernLevel);
if (readAtLastAppliedTimestamp) {
opCtx->recoveryUnit()->setTimestampReadSource(RecoveryUnit::ReadSource::kLastApplied);
readSource = opCtx->recoveryUnit()->getTimestampReadSource();
}
// This timestamp could be earlier than the timestamp seen when the transaction is opened
// because it is set asynchonously. This is not problematic because holding the collection
// lock guarantees no metadata changes will occur in that time.
auto lastAppliedTimestamp = readAtLastAppliedTimestamp
? boost::optional<Timestamp>(replCoord->getMyLastAppliedOpTime().getTimestamp())
: boost::none;
if (!_conflictingCatalogChanges(opCtx, minSnapshot, lastAppliedTimestamp)) {
return;
}
invariant(lastAppliedTimestamp ||
// The kMajorityCommitted and kNoOverlap read sources already read from timestamps
// that are safe with respect to concurrent secondary batch application.
readSource == RecoveryUnit::ReadSource::kMajorityCommitted ||
readSource == RecoveryUnit::ReadSource::kNoOverlap);
invariant(readConcernLevel != repl::ReadConcernLevel::kSnapshotReadConcern);
// Yield locks in order to do the blocking call below.
_autoColl = boost::none;
// If there are pending catalog changes, we should conflict with any in-progress batches (by
// taking the PBWM lock) and choose not to read from the last applied timestamp by unsetting
// _shouldNotConflictWithSecondaryBatchApplicationBlock. Index builds on secondaries can
// complete at timestamps later than the lastAppliedTimestamp during initial sync. After
// initial sync finishes, if we waited instead of retrying, readers would block indefinitely
// waiting for the lastAppliedTimestamp to move forward. Instead we force the reader take
// the PBWM lock and retry.
if (lastAppliedTimestamp) {
LOGV2(20576,
"tried reading at last-applied time: {lastAppliedTimestamp} on ns: {nss_ns}, but "
"future catalog changes are pending at time {minSnapshot}. Trying again without "
"reading at last-applied time.",
"lastAppliedTimestamp"_attr = *lastAppliedTimestamp,
"nss_ns"_attr = nss.ns(),
"minSnapshot"_attr = *minSnapshot);
// Destructing the block sets _shouldConflictWithSecondaryBatchApplication back to the
// previous value. If the previous value is false (because there is another
// shouldNotConflictWithSecondaryBatchApplicationBlock outside of this function), this
// does not take the PBWM lock.
_shouldNotConflictWithSecondaryBatchApplicationBlock = boost::none;
// As alluded to above, if we are AutoGetting multiple collections, it
// is possible that our "reaquire the PBWM" trick doesn't work, since we've already done
// some reads and locked in our snapshot. At this point, the only way out is to fail
// the operation. The client application will need to retry.
uassert(
ErrorCodes::SnapshotUnavailable,
str::stream() << "Unable to read from a snapshot due to pending collection catalog "
"changes; please retry the operation. Snapshot timestamp is "
<< (mySnapshot ? mySnapshot->toString() : "(none)")
<< ". Collection minimum is " << minSnapshot->toString(),
opCtx->lockState()->shouldConflictWithSecondaryBatchApplication());
// Cannot change ReadSource while a RecoveryUnit is active, which may result from
// calling getPointInTimeReadTimestamp().
opCtx->recoveryUnit()->abandonSnapshot();
opCtx->recoveryUnit()->setTimestampReadSource(RecoveryUnit::ReadSource::kUnset);
}
// If there are pending catalog changes when using a no-overlap read source, we choose to
// take the PBWM lock to conflict with any in-progress batches. This prevents us from idly
// spinning in this loop trying to get a new read timestamp ahead of the minimum visible
// snapshot. This helps us guarantee liveness (i.e. we can eventually get a suitable read
// timestamp) but should not be necessary for correctness.
if (readSource == RecoveryUnit::ReadSource::kNoOverlap) {
invariant(!lastAppliedTimestamp); // no-overlap read source selects its own timestamp.
_shouldNotConflictWithSecondaryBatchApplicationBlock = boost::none;
invariant(opCtx->lockState()->shouldConflictWithSecondaryBatchApplication());
// Abandon our snapshot but don't change our read source, so that we can select a new
// read timestamp on the next loop iteration.
opCtx->recoveryUnit()->abandonSnapshot();
}
if (readSource == RecoveryUnit::ReadSource::kMajorityCommitted) {
replCoord->waitUntilSnapshotCommitted(opCtx, *minSnapshot);
uassertStatusOK(opCtx->recoveryUnit()->obtainMajorityCommittedSnapshot());
}
{
stdx::lock_guard<Client> lk(*opCtx->getClient());
CurOp::get(opCtx)->yielded();
}
_autoColl.emplace(opCtx, nsOrUUID, collectionLockMode, viewMode, deadline);
}
}
bool AutoGetCollectionForRead::_shouldReadAtLastAppliedTimestamp(
OperationContext* opCtx,
const NamespaceString& nss,
repl::ReadConcernLevel readConcernLevel) const {
// If this block is unset, then the operation did not opt-out of the PBWM lock, implying that it
// cannot read at lastApplied. It's important to note that it is possible for this to be set,
// but still be holding the PBWM lock, explained below.
if (!_shouldNotConflictWithSecondaryBatchApplicationBlock) {
return false;
}
// If we are already holding the PBWM lock, do not read at last-applied. This is because once an
// operation reads without a timestamp (effectively seeing all writes), it is no longer safe to
// start reading at a timestamp, as writes or catalog operations may appear to vanish.
// This may occur when multiple collection locks are held concurrently, which is often the case
// when DBDirectClient is used.
if (opCtx->lockState()->isLockHeldForMode(resourceIdParallelBatchWriterMode, MODE_IS)) {
LOGV2_DEBUG(20577, 1, "not reading at last-applied because the PBWM lock is held");
return false;
}
// Majority and snapshot readConcern levels should not read from lastApplied; these read
// concerns already have a designated timestamp to read from.
if (readConcernLevel != repl::ReadConcernLevel::kLocalReadConcern &&
readConcernLevel != repl::ReadConcernLevel::kAvailableReadConcern) {
return false;
}
// If we are in a replication state (like secondary or primary catch-up) where we are not
// accepting writes, we should read at lastApplied. If this node can accept writes, then no
// conflicting replication batches are being applied and we can read from the default snapshot.
if (repl::ReplicationCoordinator::get(opCtx)->canAcceptWritesForDatabase(opCtx, "admin")) {
return false;
}
// Non-replicated collections do not need to read at lastApplied, as those collections are not
// written by the replication system. However, the oplog is special, as it *is* written by the
// replication system.
if (!nss.isReplicated() && !nss.isOplog()) {
return false;
}
return true;
}
bool AutoGetCollectionForRead::_conflictingCatalogChanges(
OperationContext* opCtx,
boost::optional<Timestamp> minSnapshot,
boost::optional<Timestamp> lastAppliedTimestamp) const {
// This is the timestamp of the most recent catalog changes to this collection. If this is
// greater than any point in time read timestamps, we should either wait or return an error.
if (!minSnapshot) {
return false;
}
// If we are reading from the lastAppliedTimestamp and it is up-to-date with any catalog
// changes, we can return.
if (lastAppliedTimestamp &&
(lastAppliedTimestamp->isNull() || *lastAppliedTimestamp >= *minSnapshot)) {
return false;
}
// This can be set when readConcern is "snapshot" or "majority".
auto mySnapshot = opCtx->recoveryUnit()->getPointInTimeReadTimestamp();
// If we do not have a point in time to conflict with minSnapshot, return.
if (!mySnapshot && !lastAppliedTimestamp) {
return false;
}
// Return if there are no conflicting catalog changes with mySnapshot.
if (mySnapshot && *mySnapshot >= *minSnapshot) {
return false;
}
return true;
}
AutoGetCollectionForReadCommand::AutoGetCollectionForReadCommand(
OperationContext* opCtx,
const NamespaceStringOrUUID& nsOrUUID,
AutoGetCollection::ViewMode viewMode,
Date_t deadline,
AutoStatsTracker::LogMode logMode)
: _autoCollForRead(opCtx, nsOrUUID, viewMode, deadline),
_statsTracker(opCtx,
_autoCollForRead.getNss(),
Top::LockType::ReadLocked,
logMode,
_autoCollForRead.getDb() ? _autoCollForRead.getDb()->getProfilingLevel()
: kDoNotChangeProfilingLevel,
deadline) {
// Perform the check early so the query planner would be able to extract the correct
// shard key. Also make sure that version is compatible if query planner decides to
// use an empty plan.
invariant(!_autoCollForRead.getView() || !_autoCollForRead.getCollection());
auto css = CollectionShardingState::get(opCtx, _autoCollForRead.getNss());
css->checkShardVersionOrThrow(opCtx, _autoCollForRead.getCollection());
}
OldClientContext::OldClientContext(OperationContext* opCtx, const std::string& ns, bool doVersion)
: _opCtx(opCtx), _db(DatabaseHolder::get(opCtx)->getDb(opCtx, ns)) {
if (!_db) {
const auto dbName = nsToDatabaseSubstring(ns);
_db = DatabaseHolder::get(opCtx)->openDb(_opCtx, dbName, &_justCreated);
invariant(_db);
}
auto const currentOp = CurOp::get(_opCtx);
if (doVersion) {
switch (currentOp->getNetworkOp()) {
case dbGetMore: // getMore is special and should be handled elsewhere
case dbUpdate: // update & delete check shard version as part of the write executor
case dbDelete: // path, so no need to check them here as well
break;
default:
CollectionShardingState::get(_opCtx, NamespaceString(ns))
->checkShardVersionOrThrow(
_opCtx,
CollectionCatalog::get(opCtx).lookupCollectionByNamespace(
opCtx, NamespaceString(ns)));
break;
}
}
stdx::lock_guard<Client> lk(*_opCtx->getClient());
currentOp->enter_inlock(ns.c_str(), _db->getProfilingLevel());
}
OldClientContext::~OldClientContext() {
// If in an interrupt, don't record any stats.
// It is possible to have no lock after saving the lock state and being interrupted while
// waiting to restore.
if (_opCtx->getKillStatus() != ErrorCodes::OK)
return;
invariant(_opCtx->lockState()->isLocked());
auto currentOp = CurOp::get(_opCtx);
Top::get(_opCtx->getClient()->getServiceContext())
.record(_opCtx,
currentOp->getNS(),
currentOp->getLogicalOp(),
_opCtx->lockState()->isWriteLocked() ? Top::LockType::WriteLocked
: Top::LockType::ReadLocked,
_timer.micros(),
currentOp->isCommand(),
currentOp->getReadWriteType());
}
LockMode getLockModeForQuery(OperationContext* opCtx, const boost::optional<NamespaceString>& nss) {
invariant(opCtx);
// Use IX locks for multi-statement transactions; otherwise, use IS locks.
if (opCtx->inMultiDocumentTransaction()) {
uassert(51071,
"Cannot query system.views within a transaction",
!nss || !nss->isSystemDotViews());
return MODE_IX;
}
return MODE_IS;
}
BlockSecondaryReadsDuringBatchApplication_DONT_USE::
BlockSecondaryReadsDuringBatchApplication_DONT_USE(OperationContext* opCtx)
: _opCtx(opCtx) {
auto allowSecondaryReads = &allowSecondaryReadsDuringBatchApplication_DONT_USE(opCtx);
allowSecondaryReads->swap(_originalSettings);
*allowSecondaryReads = false;
}
BlockSecondaryReadsDuringBatchApplication_DONT_USE::
~BlockSecondaryReadsDuringBatchApplication_DONT_USE() {
auto allowSecondaryReads = &allowSecondaryReadsDuringBatchApplication_DONT_USE(_opCtx);
allowSecondaryReads->swap(_originalSettings);
}
} // namespace mongo
|