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
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
|
/**
* Copyright (C) 2019-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.
*/
#pragma once
#include <boost/optional.hpp>
#include "mongo/bson/oid.h"
#include "mongo/db/operation_context.h"
#include "mongo/platform/mutex.h"
#include "mongo/util/concurrency/thread_pool_interface.h"
#include "mongo/util/functional.h"
#include "mongo/util/future.h"
#include "mongo/util/invalidating_lru_cache.h"
namespace mongo {
/**
* Serves as a container of the non-templatised parts of the ReadThroughCache class below.
*/
class ReadThroughCacheBase {
ReadThroughCacheBase(const ReadThroughCacheBase&) = delete;
ReadThroughCacheBase& operator=(const ReadThroughCacheBase&) = delete;
protected:
ReadThroughCacheBase(Mutex& mutex, ServiceContext* service, ThreadPoolInterface& threadPool);
virtual ~ReadThroughCacheBase();
/**
* This method is an extension of ThreadPoolInterface::schedule, with the following additions:
* - Creates a client and an operation context and executes the specified 'work' under that
* environment
* - Returns a CancelToken, which can be used to attempt to cancel 'work'
*
* If the task manages to get canceled before it is executed (through a call to tryCancel),
* 'work' will be invoked out-of-line with a non-OK status, set to error code
* ReadThroughCacheLookupCanceled.
*/
class CancelToken {
public:
struct TaskInfo;
CancelToken(std::shared_ptr<TaskInfo> info);
CancelToken(CancelToken&&);
~CancelToken();
void tryCancel();
private:
std::shared_ptr<TaskInfo> _info;
};
using WorkWithOpContext = unique_function<void(OperationContext*, const Status&)>;
CancelToken _asyncWork(WorkWithOpContext work) noexcept;
Date_t _now();
// Service context under which this cache has been instantiated (used for access to service-wide
// functionality, such as client/operation context creation)
ServiceContext* const _serviceContext;
// Thread pool to be used for invoking the blocking 'lookup' calls
ThreadPoolInterface& _threadPool;
// Used to protect the shared state in the child ReadThroughCache template below. Has a lock
// level of 3, meaning that while held, it is only allowed to take '_cancelTokenMutex' below and
// the Client lock.
Mutex& _mutex;
// Used to protect calls to 'tryCancel' above. Has a lock level of 2, meaning what while held,
// it is only allowed to take the Client lock.
Mutex _cancelTokenMutex = MONGO_MAKE_LATCH("ReadThroughCacheBase::_cancelTokenMutex");
};
template <typename Result, typename Key, typename Value, typename Time>
struct ReadThroughCacheLookupFnImpl {
using fn = unique_function<Result(
OperationContext*, const Key&, const Value& cachedValue, const Time& timeInStore)>;
};
template <typename Result, typename Key, typename Value>
struct ReadThroughCacheLookupFnImpl<Result, Key, Value, CacheNotCausallyConsistent> {
using fn = unique_function<Result(OperationContext*, const Key&, const Value& cachedValue)>;
};
/**
* Implements an (optionally) causally consistent read-through cache from Key to Value, built on top
* of InvalidatingLRUCache.
*
* Causal consistency is provided by requiring the backing store to asociate every Value it returns
* with a logical timestamp of type Time.
*/
template <typename Key, typename Value, typename Time = CacheNotCausallyConsistent>
class ReadThroughCache : public ReadThroughCacheBase {
/**
* Data structure wrapping and expanding on the values stored in the cache.
*/
struct StoredValue {
Value value;
// Contains the wallclock time of when the value was fetched from the backing storage. This
// value is not precise and should only be used for diagnostics purposes (i.e., it cannot be
// relied on to perform any recency comparisons for example).
Date_t updateWallClockTime;
};
using Cache = InvalidatingLRUCache<Key, StoredValue, Time>;
public:
/**
* Common type for values returned from the cache.
*/
class ValueHandle {
public:
// The two constructors below are present in order to offset the fact that the cache doesn't
// support pinning items. Their only usage must be in the authorization mananager for the
// internal authentication user.
ValueHandle(Value&& value) : _valueHandle({std::move(value), Date_t::min()}) {}
ValueHandle() = default;
operator bool() const {
return bool(_valueHandle);
}
bool isValid() const {
return _valueHandle.isValid();
}
Value* get() {
return &_valueHandle->value;
}
const Value* get() const {
return &_valueHandle->value;
}
Value& operator*() {
return *get();
}
const Value& operator*() const {
return *get();
}
Value* operator->() {
return get();
}
const Value* operator->() const {
return get();
}
/**
* See the comments for `StoredValue::updateWallClockTime` above.
*/
Date_t updateWallClockTime() const {
return _valueHandle->updateWallClockTime;
}
private:
friend class ReadThroughCache;
ValueHandle(typename Cache::ValueHandle&& valueHandle)
: _valueHandle(std::move(valueHandle)) {}
typename Cache::ValueHandle _valueHandle;
};
/**
* Signature for a blocking function to provide the value for a key when there is a cache miss.
*
* The implementation must throw a uassertion to indicate an error while looking up the value,
* return boost::none if the key is not found, or return an actual value.
*
* See the comments on 'advanceTimeInStore' for additional requirements that this function must
* fulfill with respect to causal consistency.
*/
struct LookupResult {
// The 't' parameter is mandatory for causally-consistent caches, but not needed otherwise
// (since the time never changes). Using a default of '= CacheNotCausallyConsistent()'
// allows non-causally-consistent users to not have to pass a second parameter, but would
// fail compilation if causally-consistent users forget to pass it.
explicit LookupResult(boost::optional<Value>&& v, Time t = CacheNotCausallyConsistent())
: v(std::move(v)), t(std::move(t)) {}
LookupResult(LookupResult&&) = default;
LookupResult& operator=(LookupResult&&) = default;
// If boost::none, it means the '_lookupFn' did not find the key in the store
boost::optional<Value> v;
// If value is boost::none, specifies the time which was passed to '_lookupFn', effectively
// meaning, at least as of 'time', there was no entry in the store for the key. Otherwise
// contains the time that the store returned for the 'value'.
Time t;
};
using LookupFn =
typename ReadThroughCacheLookupFnImpl<LookupResult, Key, ValueHandle, Time>::fn;
// Exposed publicly so it can be unit-tested indepedently of the usages in this class. Must not
// be used independently.
class InProgressLookup;
/**
* If 'key' is found in the cache and it fulfills the requested 'causalConsistency', returns a
* set ValueHandle (its operator bool will be true). Otherwise, either causes the blocking
* 'LookupFn' to be asynchronously invoked to fetch 'key' from the backing store or joins an
* already scheduled invocation) and returns a future which will be signaled when the lookup
* completes.
*
* If the lookup is successful and 'key' is found in the store, it will be cached (so subsequent
* lookups won't have to re-fetch it) and the future will be set. If 'key' is not found in the
* backing store, returns a not-set ValueHandle (it's bool operator will be false). If 'lookup'
* fails, the future will be set to the appropriate exception and nothing will be cached,
* meaning that subsequent calls to 'acquireAsync' will kick-off 'lookup' again.
*
* NOTES:
* The returned value may be invalid by the time the caller gets to access it, if 'invalidate'
* is called for 'key'.
*/
SharedSemiFuture<ValueHandle> acquireAsync(
const Key& key,
CacheCausalConsistency causalConsistency = CacheCausalConsistency::kLatestCached) {
// Fast path
if (auto cachedValue = _cache.get(key, causalConsistency))
return {std::move(cachedValue)};
stdx::unique_lock ul(_mutex);
// Re-check the cache under a mutex, before kicking-off the asynchronous lookup
if (auto cachedValue = _cache.get(key, causalConsistency))
return {std::move(cachedValue)};
// Join an in-progress lookup if one has already been scheduled
if (auto it = _inProgressLookups.find(key); it != _inProgressLookups.end())
return it->second->addWaiter(ul);
// Schedule an asynchronous lookup for the key
auto [cachedValue, timeInStore] = _cache.getCachedValueAndTime(key);
auto [it, emplaced] = _inProgressLookups.emplace(
key,
std::make_unique<InProgressLookup>(
*this, key, ValueHandle(std::move(cachedValue)), timeInStore));
invariant(emplaced);
auto& inProgressLookup = *it->second;
auto sharedFutureToReturn = inProgressLookup.addWaiter(ul);
ul.unlock();
_doLookupWhileNotValid(key, Status(ErrorCodes::Error(461540), "")).getAsync([](auto) {});
return sharedFutureToReturn;
}
/**
* A blocking variant of 'acquireAsync' above - refer to it for more details.
*
* NOTES:
* This is a potentially blocking method.
*/
ValueHandle acquire(
OperationContext* opCtx,
const Key& key,
CacheCausalConsistency causalConsistency = CacheCausalConsistency::kLatestCached) {
return acquireAsync(key, causalConsistency).get(opCtx);
}
/**
* Invalidates the given 'key' and immediately replaces it with a new value.
*/
ValueHandle insertOrAssignAndGet(const Key& key, Value&& newValue, Date_t updateWallClockTime) {
stdx::lock_guard lg(_mutex);
if (auto it = _inProgressLookups.find(key); it != _inProgressLookups.end())
it->second->invalidateAndCancelCurrentLookupRound(lg);
return _cache.insertOrAssignAndGet(key, {std::move(newValue), updateWallClockTime});
}
/**
* Indicates to the cache that the backing store has a newer version of 'key', corresponding to
* 'newTime'. Subsequent calls to 'acquireAsync' with a causal consistency set to 'LatestKnown'
* will block and perform refresh until the cached value reaches 'newTime'.
*
* With respect to causal consistency, the 'LookupFn' used for this cache must provide the
* guarantee that if 'advanceTimeInStore' is called with a 'newTime', a subsequent call to
* 'LookupFn' for 'key' must return at least 'newTime' or later.
*/
void advanceTimeInStore(const Key& key, const Time& newTime) {
stdx::lock_guard lg(_mutex);
if (auto it = _inProgressLookups.find(key); it != _inProgressLookups.end())
it->second->advanceTimeInStore(lg, newTime);
_cache.advanceTimeInStore(key, newTime);
}
/**
* The invalidate methods below guarantee the following:
* - All affected keys already in the cache (or returned to callers) will be invalidated and
* removed from the cache
* - All affected keys, which are in the process of being loaded (i.e., acquireAsync has not
* yet completed) will be internally interrupted and rescheduled again, as if 'acquireAsync'
* was called *after* the call to invalidate
*
* In essence, the invalidate calls serve as a "barrier" for the affected keys.
*/
void invalidate(const Key& key) {
stdx::lock_guard lg(_mutex);
if (auto it = _inProgressLookups.find(key); it != _inProgressLookups.end())
it->second->invalidateAndCancelCurrentLookupRound(lg);
_cache.invalidate(key);
}
template <typename Pred>
void invalidateIf(const Pred& predicate) {
stdx::lock_guard lg(_mutex);
for (auto& entry : _inProgressLookups) {
if (predicate(entry.first))
entry.second->invalidateAndCancelCurrentLookupRound(lg);
}
_cache.invalidateIf([&](const Key& key, const StoredValue*) { return predicate(key); });
}
void invalidateAll() {
invalidateIf([](const Key&) { return true; });
}
/**
* Returns statistics information about the cache for reporting purposes.
*/
std::vector<typename Cache::CachedItemInfo> getCacheInfo() const {
return _cache.getCacheInfo();
}
protected:
/**
* ReadThroughCache constructor, to be called by sub-classes, which implement 'lookup'.
*
* The 'mutex' is for the exclusive usage of the ReadThroughCache and must not be used in any
* way by the implementing class. Having the mutex stored by the sub-class allows latch
* diagnostics to be correctly associated with the sub-class (not the generic ReadThroughCache
* class).
*
* The 'threadPool' can be used for other purposes, but it is mandatory that by the time this
* object is destructed that it is shut down and joined so that there are no more asynchronous
* loading activities going on.
*
* The 'cacheSize' parameter specifies the maximum size of the cache before the least recently
* used entries start getting evicted. It is allowed to be zero, in which case no entries will
* actually be cached, but it doesn't guarantee that every `acquire` call will result in an
* invocation of `lookup`. Specifically, several concurrent invocations of `acquire` for the
* same key may group together for a single `lookup`.
*/
ReadThroughCache(Mutex& mutex,
ServiceContext* service,
ThreadPoolInterface& threadPool,
LookupFn lookupFn,
int cacheSize)
: ReadThroughCacheBase(mutex, service, threadPool),
_lookupFn(std::move(lookupFn)),
_cache(cacheSize) {}
~ReadThroughCache() {
invariant(_inProgressLookups.empty());
}
private:
using InProgressLookupsMap = stdx::unordered_map<Key, std::unique_ptr<InProgressLookup>>;
/**
* This method implements an asynchronous "while (!valid)" loop over 'key', which must be on the
* in-progress map.
*/
Future<LookupResult> _doLookupWhileNotValid(Key key, StatusWith<LookupResult> sw) {
stdx::unique_lock ul(_mutex);
auto it = _inProgressLookups.find(key);
invariant(it != _inProgressLookups.end());
auto& inProgressLookup = *it->second;
auto [promisesToSet, result, mustDoAnotherLoop] = [&] {
// The thread pool is shutting down, so terminate the loop
if (ErrorCodes::isCancelationError(sw.getStatus()))
return std::make_tuple(inProgressLookup.getAllPromisesOnError(ul),
StatusWith<ValueHandle>(sw.getStatus()),
false);
// There was a concurrent call to 'invalidate', so start all over
if (!inProgressLookup.valid(ul))
return std::make_tuple(
std::vector<std::unique_ptr<SharedPromise<ValueHandle>>>{},
StatusWith<ValueHandle>(Status(ErrorCodes::Error(461541), "")),
true);
// Lookup resulted in an error, which is not cancelation
if (!sw.isOK())
return std::make_tuple(inProgressLookup.getAllPromisesOnError(ul),
StatusWith<ValueHandle>(sw.getStatus()),
false);
// Value (or boost::none) was returned by lookup and there was no concurrent call to
// 'invalidate'. Place the value on the cache and return the necessary promises to
// signal (those which are waiting for time < time at the store).
auto& result = sw.getValue();
auto promisesToSet = inProgressLookup.getPromisesLessThanTime(ul, result.t);
auto valueHandleToSet = [&] {
if (result.v) {
ValueHandle valueHandle(
_cache.insertOrAssignAndGet(key, {std::move(*result.v), _now()}, result.t));
// In the case that 'key' was not present in the store up to this lookup's
// completion, it is possible that concurrent callers advanced the time in store
// further than what was returned by the lookup. Because of this, the time in
// the cache must be synchronised with that of the InProgressLookup.
_cache.advanceTimeInStore(key, inProgressLookup.minTimeInStore(ul));
return valueHandle;
}
_cache.invalidate(key);
return ValueHandle();
}();
return std::make_tuple(std::move(promisesToSet),
StatusWith<ValueHandle>(std::move(valueHandleToSet)),
!inProgressLookup.empty(ul));
}();
if (!mustDoAnotherLoop)
_inProgressLookups.erase(it);
ul.unlock();
// The only reason this loop pops the values as it goes and std::moves into the last value
// is to support the CacheSizeZero unit-test, which requires that once the future it waits
// on is set, it contains the last reference on the returned ValueHandle
while (!promisesToSet.empty()) {
auto p(std::move(promisesToSet.back()));
promisesToSet.pop_back();
if (promisesToSet.empty())
p->setFromStatusWith(std::move(result));
else
p->setFromStatusWith(result);
}
return mustDoAnotherLoop
? inProgressLookup.asyncLookupRound().onCompletion(
[this, key](auto sw) { return _doLookupWhileNotValid(key, std::move(sw)); })
: Future<LookupResult>::makeReady(Status(ErrorCodes::Error(461542), ""));
}
// Blocking function which will be invoked to retrieve entries from the backing store
const LookupFn _lookupFn;
// Contains all the currently cached keys. This structure is self-synchronising and doesn't
// require a mutex. However, on cache miss it is accessed under '_mutex', which is safe, because
// _cache's mutex itself is at level 0.
//
// NOTE: From destruction order point of view, because keys first "start" in
// '_inProgressLookups' and then move on to '_cache' the order of these two fields is important.
Cache _cache;
// Keeps track of all the keys, which were attempted to be 'acquireAsync'-ed, weren't found in
// the cache and are currently in the process of being looked up from the backing store. A
// single key may be missing from '_cache', or contain an old 'kLatestCached' and have an active
// lookup on this map for 'kLatestKnown'.
//
// This map is protected by '_mutex'.
InProgressLookupsMap _inProgressLookups;
};
/**
* This class represents an in-progress lookup for a specific key and implements the guarantees of
* the invalidation logic as described in the comments of 'ReadThroughCache::invalidate'.
*
* It is intended to be used in conjunction with the 'ReadThroughCache', which operates on it under
* its '_mutex' and ensures there is always at most a single active instance at a time active for
* each 'key'.
*
* The methods of this class are not thread-safe, unless indicated in the comments.
*
* Its lifecycle is intended to be like this:
*
* inProgressLookups.emplace(inProgress);
* while (true) {
* result = inProgress.asyncLookupRound();
* if (!inProgress.valid()) {
* continue;
* }
*
* inProgressLookups.remove(inProgress)
* cachedValues.insert(result);
* inProgress.signalWaiters(result);
* }
*/
template <typename Key, typename Value, typename Time>
class ReadThroughCache<Key, Value, Time>::InProgressLookup {
public:
InProgressLookup(ReadThroughCache& cache, Key key, ValueHandle cachedValue, Time minTimeInStore)
: _cache(cache),
_key(std::move(key)),
_cachedValue(std::move(cachedValue)),
_minTimeInStore(std::move(minTimeInStore)) {}
Future<LookupResult> asyncLookupRound() {
auto [promise, future] = makePromiseFuture<LookupResult>();
stdx::lock_guard lg(_cache._mutex);
_valid = true;
_cancelToken.emplace(_cache._asyncWork([ this, promise = std::move(promise) ](
OperationContext * opCtx, const Status& status) mutable noexcept {
promise.setWith([&] {
uassertStatusOK(status);
if constexpr (std::is_same_v<Time, CacheNotCausallyConsistent>) {
return _cache._lookupFn(opCtx, _key, _cachedValue);
} else {
return _cache._lookupFn(opCtx, _key, _cachedValue, _minTimeInStore);
}
});
}));
return std::move(future);
}
SharedSemiFuture<ValueHandle> addWaiter(WithLock) {
auto [it, unusedEmplaced] = _outstanding.try_emplace(
_minTimeInStore, std::make_unique<SharedPromise<ValueHandle>>());
return it->second->getFuture();
}
Time minTimeInStore(WithLock) const {
return _minTimeInStore;
}
bool valid(WithLock) const {
return _valid;
}
std::vector<std::unique_ptr<SharedPromise<ValueHandle>>> getAllPromisesOnError(WithLock) {
invariant(_valid);
std::vector<std::unique_ptr<SharedPromise<ValueHandle>>> ret;
for (auto it = _outstanding.begin(); it != _outstanding.end();) {
ret.emplace_back(std::move(it->second));
it = _outstanding.erase(it);
}
return ret;
}
std::vector<std::unique_ptr<SharedPromise<ValueHandle>>> getPromisesLessThanTime(WithLock,
Time time) {
invariant(_valid);
std::vector<std::unique_ptr<SharedPromise<ValueHandle>>> ret;
for (auto it = _outstanding.begin(); it != _outstanding.end();) {
if (it->first > time)
break;
ret.emplace_back(std::move(it->second));
it = _outstanding.erase(it);
}
return ret;
}
bool empty(WithLock) const {
invariant(_valid);
return _outstanding.empty();
}
void advanceTimeInStore(WithLock, const Time& newTime) {
if (newTime > _minTimeInStore)
_minTimeInStore = newTime;
}
void invalidateAndCancelCurrentLookupRound(WithLock) {
_valid = false;
if (_cancelToken)
_cancelToken->tryCancel();
}
private:
// The owning cache, from which mutex, lookupFn, async task scheduling, etc. will be used. It is
// the responsibility of the owning cache to join all outstanding lookups at destruction time.
ReadThroughCache& _cache;
const Key _key;
bool _valid{false};
boost::optional<CancelToken> _cancelToken;
ValueHandle _cachedValue;
Time _minTimeInStore;
using TimeAndPromiseMap = std::map<Time, std::unique_ptr<SharedPromise<ValueHandle>>>;
TimeAndPromiseMap _outstanding;
};
} // namespace mongo
|