-
Notifications
You must be signed in to change notification settings - Fork 4.1k
/
AbstractParallelEvaluator.java
1138 lines (1065 loc) · 53 KB
/
AbstractParallelEvaluator.java
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
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
// Copyright 2017 The Bazel Authors. All rights reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package com.google.devtools.build.skyframe;
import static com.google.common.base.Preconditions.checkNotNull;
import static com.google.common.base.Preconditions.checkState;
import static com.google.common.base.Throwables.throwIfInstanceOf;
import static com.google.common.base.Throwables.throwIfUnchecked;
import com.github.benmanes.caffeine.cache.Cache;
import com.github.benmanes.caffeine.cache.Caffeine;
import com.github.benmanes.caffeine.cache.RemovalCause;
import com.google.common.collect.ImmutableCollection;
import com.google.common.collect.ImmutableList;
import com.google.common.collect.ImmutableMap;
import com.google.common.collect.ImmutableSet;
import com.google.common.collect.Iterables;
import com.google.common.collect.Sets;
import com.google.common.flogger.GoogleLogger;
import com.google.common.graph.ImmutableGraph;
import com.google.common.graph.Traverser;
import com.google.common.util.concurrent.ListenableFuture;
import com.google.devtools.build.lib.collect.nestedset.NestedSet;
import com.google.devtools.build.lib.concurrent.QuiescingExecutor;
import com.google.devtools.build.lib.events.Event;
import com.google.devtools.build.lib.events.ExtendedEventHandler;
import com.google.devtools.build.lib.events.Reportable;
import com.google.devtools.build.lib.profiler.Profiler;
import com.google.devtools.build.lib.profiler.ProfilerTask;
import com.google.devtools.build.lib.supplier.InterruptibleSupplier;
import com.google.devtools.build.skyframe.EvaluationProgressReceiver.EvaluationState;
import com.google.devtools.build.skyframe.EvaluationProgressReceiver.NodeState;
import com.google.devtools.build.skyframe.NodeEntry.DependencyState;
import com.google.devtools.build.skyframe.NodeEntry.DirtyType;
import com.google.devtools.build.skyframe.NodeEntry.LifecycleState;
import com.google.devtools.build.skyframe.NodeEntry.NodeValueAndRdepsToSignal;
import com.google.devtools.build.skyframe.QueryableGraph.Reason;
import com.google.devtools.build.skyframe.SkyFunction.Environment.SkyKeyComputeState;
import com.google.devtools.build.skyframe.SkyFunction.Reset;
import com.google.devtools.build.skyframe.SkyFunctionEnvironment.UndonePreviouslyRequestedDeps;
import com.google.devtools.build.skyframe.SkyFunctionException.ReifiedSkyFunctionException;
import com.google.devtools.build.skyframe.proto.GraphInconsistency.Inconsistency;
import com.google.errorprone.annotations.CanIgnoreReturnValue;
import java.io.IOException;
import java.util.Collection;
import java.util.List;
import java.util.Set;
import java.util.function.Predicate;
import javax.annotation.Nullable;
/**
* Defines the evaluation action used in the multi-threaded Skyframe evaluation, and constructs the
* {@link ParallelEvaluatorContext} that the actions rely on.
*
* <p>Evaluates a set of given functions ({@code SkyFunction}s) with arguments ({@code SkyKey}s).
* Cycles are not allowed and are detected during the traversal.
*
* <p>This class implements multi-threaded evaluation. This is a fairly complex process that has
* strong consistency requirements between the {@link ProcessableGraph}, the nodes in the graph of
* type {@link NodeEntry}, the work queue, and the set of in-flight nodes.
*
* <p>The basic invariants are:
*
* <p>A node can be in one of three states: ready, waiting, and done. A node is ready if and only if
* all of its dependencies have been signaled. A node is done if it has a value. It is waiting if
* not all of its dependencies have been signaled.
*
* <p>A node must be in the work queue if and only if it is ready. It is an error for a node to be
* in the work queue twice at the same time.
*
* <p>A node is considered in-flight if it has been created, and is not done yet. In case of an
* interrupt, the work queue is discarded, and the in-flight set is used to remove partially
* computed values.
*
* <p>Each evaluation of the graph takes place at a "version," which is currently given by a
* non-negative {@code long}. The version can also be thought of as an "mtime." Each node in the
* graph has a version, which is the last version at which its value changed. This version data is
* used to avoid unnecessary re-evaluation of values. If a node is re-evaluated and found to have
* the same data as before, its version (mtime) remains the same. If all of a node's children's have
* the same version as before, its re-evaluation can be skipped.
*
* <p>This does not implement other parts of Skyframe evaluation setup and post-processing, such as
* translating a set of requested top-level nodes into actions, or constructing an evaluation
* result. Derived classes should do this.
*/
abstract class AbstractParallelEvaluator {
private static final GoogleLogger logger = GoogleLogger.forEnclosingClass();
final ProcessableGraph graph;
final ParallelEvaluatorContext evaluatorContext;
final CycleDetector cycleDetector;
final Cache<SkyKey, SkyKeyComputeState> stateCache =
Caffeine.newBuilder()
.executor(Runnable::run) // run the removalListener immediately in the same thread
.removalListener((SkyKey k, SkyKeyComputeState v, RemovalCause cause) -> v.close())
.build();
AbstractParallelEvaluator(
ProcessableGraph graph,
Version graphVersion,
Version minimalVersion,
ImmutableMap<SkyFunctionName, SkyFunction> skyFunctions,
ExtendedEventHandler reporter,
EmittedEventState emittedEventState,
EventFilter storedEventFilter,
ErrorInfoManager errorInfoManager,
InflightTrackingProgressReceiver progressReceiver,
GraphInconsistencyReceiver graphInconsistencyReceiver,
QuiescingExecutor executor,
CycleDetector cycleDetector,
Predicate<SkyKey> keepGoing) {
this.graph = graph;
this.cycleDetector = cycleDetector;
this.evaluatorContext =
new ParallelEvaluatorContext(
graph,
graphVersion,
minimalVersion,
skyFunctions,
reporter,
emittedEventState,
progressReceiver,
storedEventFilter,
errorInfoManager,
graphInconsistencyReceiver,
executor,
() -> new NodeEntryVisitor(executor, progressReceiver, Evaluate::new, stateCache),
stateCache,
keepGoing);
}
/**
* If the entry is dirty and not already rebuilding, puts it in a state so that it can rebuild.
*/
static void maybeMarkRebuilding(NodeEntry entry) {
if (entry.isDirty() && entry.getLifecycleState() != LifecycleState.REBUILDING) {
entry.markRebuilding();
}
}
enum DirtyOutcome {
ALREADY_PROCESSED,
NEEDS_EVALUATION
}
/** * An action that evaluates a value. */
private final class Evaluate implements Runnable {
private final SkyKey skyKey;
private Evaluate(SkyKey skyKey) {
this.skyKey = skyKey;
}
/**
* Notes the rdep from the parent to the child, and then does the appropriate thing with the
* child or the parent, returning whether the parent has both been signalled and also is ready
* for evaluation.
*/
@CanIgnoreReturnValue
private boolean enqueueChild(
SkyKey skyKey,
NodeEntry entry,
SkyKey child,
NodeEntry childEntry,
boolean depAlreadyExists,
boolean enqueueParentIfReady,
@Nullable SkyFunctionEnvironment environmentIfEnqueuing)
throws InterruptedException {
checkState(!entry.isDone(), "%s %s", skyKey, entry);
DependencyState dependencyState;
try {
dependencyState =
depAlreadyExists
? childEntry.checkIfDoneForDirtyReverseDep(skyKey)
: childEntry.addReverseDepAndCheckIfDone(skyKey);
} catch (IllegalStateException e) {
// Add some more context regarding crashes.
throw new IllegalStateException("child key: " + child + " error: " + e.getMessage(), e);
}
switch (dependencyState) {
case DONE:
if (entry.signalDep(childEntry.getVersion(), child)) {
if (enqueueParentIfReady) {
evaluatorContext.getVisitor().enqueueEvaluation(skyKey, child);
}
return true;
} else {
if (skyKey.supportsPartialReevaluation()
&& environmentIfEnqueuing != null
&& environmentIfEnqueuing.wasNewlyRequestedDepNullForPartialReevaluation(child)) {
// If a dep was observed not-done by its parent when the parent tried to read its
// value, but that dep is now done, then this is the only chance the parent has to be
// signalled by that dep.
evaluatorContext.getVisitor().enqueueEvaluation(skyKey, child);
}
}
break;
case ALREADY_EVALUATING:
break;
case NEEDS_SCHEDULING:
evaluatorContext.getVisitor().enqueueEvaluation(child, null);
break;
}
return false;
}
/**
* Returns true if this depGroup consists of the error transience value and the error transience
* value is newer than the entry, meaning that the entry must be re-evaluated.
*/
private boolean invalidatedByErrorTransience(Collection<SkyKey> depGroup, NodeEntry entry)
throws InterruptedException {
return depGroup.size() == 1
&& depGroup.contains(ErrorTransienceValue.KEY)
&& !graph
.get(null, Reason.OTHER, ErrorTransienceValue.KEY)
.getVersion()
.atMost(entry.getVersion());
}
private DirtyOutcome maybeHandleDirtyNode(NodeEntry nodeEntry) throws InterruptedException {
while (nodeEntry.getLifecycleState() == LifecycleState.CHECK_DEPENDENCIES) {
// Evaluating a dirty node for the first time, and checking its children to see if any
// of them have changed. Note that there must be dirty children for this to happen.
// Check the children group by group -- we don't want to evaluate a value that is no
// longer needed because an earlier dependency changed. For example, //foo:foo depends
// on target //bar:bar and is built. Then foo/BUILD is modified to remove the dependence
// on bar, and bar/BUILD is deleted. Reloading //bar:bar would incorrectly throw an
// exception. To avoid this, we must reload foo/BUILD first, at which point we will
// discover that it has changed, and re-evaluate target //foo:foo from scratch.
// On the other hand, when an action requests all of its inputs, we can safely check all
// of them in parallel on a subsequent build. So we allow checking an entire group in
// parallel here, if the node builder requested a group last build.
// Note: every dep returned here must either have this node re-registered for it (using
// checkIfDoneForDirtyReverseDep) and be registered as a direct dep of this node, or have
// its reverse dep on this node removed. Failing to do either one of these would result in
// a graph inconsistency, where the child had a reverse dep on this node, but this node
// had no kind of dependency on the child.
List<SkyKey> directDepsToCheck = nodeEntry.getNextDirtyDirectDeps();
if (invalidatedByErrorTransience(directDepsToCheck, nodeEntry)) {
// If this dep is the ErrorTransienceValue and the ErrorTransienceValue has been
// updated then we need to force a rebuild. We would like to just signal the entry as
// usual, but we can't, because then the ErrorTransienceValue would remain as a dep,
// which would be incorrect if, for instance, the value re-evaluated to a non-error.
nodeEntry.forceRebuild();
graph.get(skyKey, Reason.RDEP_REMOVAL, ErrorTransienceValue.KEY).removeReverseDep(skyKey);
return DirtyOutcome.NEEDS_EVALUATION;
}
NodeBatch entriesToCheck = null;
if (!evaluatorContext.keepGoing(skyKey)) {
// This check ensures that we maintain the invariant that if a node with an error is
// reached during a no-keep-going build, none of its currently building parents
// finishes building. If the child isn't done building yet, it will detect on its own
// that it has an error (see the VERIFIED_CLEAN case below). On the other hand, if it
// is done, then it is the parent's responsibility to notice that, which we do here.
// We check the deps for errors so that we don't continue building this node if it has
// a child error.
entriesToCheck = graph.getBatch(skyKey, Reason.OTHER, directDepsToCheck);
for (SkyKey keyToCheck : directDepsToCheck) {
NodeEntry nodeEntryToCheck = entriesToCheck.get(keyToCheck);
SkyValue valueMaybeWithMetadata = nodeEntryToCheck.getValueMaybeWithMetadata();
if (valueMaybeWithMetadata == null) {
continue;
}
ErrorInfo maybeErrorInfo = ValueWithMetadata.getMaybeErrorInfo(valueMaybeWithMetadata);
if (maybeErrorInfo == null) {
continue;
}
// This child has an error. We add a dep from this node to it and throw an exception
// coming from it.
nodeEntry.addSingletonTemporaryDirectDep(keyToCheck);
nodeEntryToCheck.checkIfDoneForDirtyReverseDep(skyKey);
// Perform the necessary bookkeeping for any deps that are not being used.
for (SkyKey depKey : directDepsToCheck) {
if (!depKey.equals(keyToCheck)) {
entriesToCheck.get(depKey).removeReverseDep(skyKey);
}
}
if (!evaluatorContext.getVisitor().preventNewEvaluations()) {
// An error was already thrown in the evaluator. Don't do anything here.
return DirtyOutcome.ALREADY_PROCESSED;
}
throw SchedulerException.ofError(maybeErrorInfo, keyToCheck, ImmutableSet.of(skyKey));
}
}
// It is safe to add these deps back to the node -- even if one of them has changed, the
// contract of pruning is that the node will request these deps again when it rebuilds.
// We must add these deps before enqueuing them, so that the node knows that it depends
// on them. If one of these deps is the error transience node, the check we did above
// in #invalidatedByErrorTransience means that the error transience node is not newer
// than this node, so we are going to mark it clean (since the error transience node is
// always the last dep).
nodeEntry.addTemporaryDirectDepGroup(directDepsToCheck);
DepsReport depsReport = graph.analyzeDepsDoneness(skyKey, directDepsToCheck);
Collection<SkyKey> unknownStatusDeps =
depsReport.hasInformation() ? depsReport : directDepsToCheck;
boolean needsScheduling = false;
for (int i = 0; i < directDepsToCheck.size() - unknownStatusDeps.size(); i++) {
// Since all of these nodes were done at an earlier version than this one, we may safely
// signal with the minimal version, since they cannot trigger a re-evaluation.
needsScheduling = nodeEntry.signalDep(Version.minimal(), /* childForDebugging= */ null);
}
if (needsScheduling) {
checkState(
unknownStatusDeps.isEmpty(),
"Ready without all deps checked? %s %s %s",
skyKey,
nodeEntry,
unknownStatusDeps);
continue;
}
if (entriesToCheck == null || depsReport.hasInformation()) {
entriesToCheck = graph.getBatch(skyKey, Reason.ENQUEUING_CHILD, unknownStatusDeps);
}
boolean parentIsSignalledAndReady =
handleKnownChildrenForDirtyNode(
unknownStatusDeps,
entriesToCheck,
nodeEntry,
/* enqueueParentIfReady= */ false,
/* environmentIfEnqueuing= */ null);
if (!parentIsSignalledAndReady
|| evaluatorContext.getVisitor().shouldPreventNewEvaluations()) {
return DirtyOutcome.ALREADY_PROCESSED;
}
// If we're here, then we may proceed to the rest of the method and continue processing
// the node intra-thread. This is a performance optimization: By not enqueuing the node,
// we avoid contention on the queue data structure (between concurrent threads
// enqueueing and dequeueing), and we also save wall time since the node gets processed
// now rather than at some point in the future.
}
switch (nodeEntry.getLifecycleState()) {
case VERIFIED_CLEAN:
// No child has a changed value. This node can be marked done and its parents signaled
// without any re-evaluation.
NodeValueAndRdepsToSignal nodeValueAndRdeps = nodeEntry.markClean();
Set<SkyKey> rDepsToSignal = nodeValueAndRdeps.getRdepsToSignal();
SkyValue valueMaybeWithMetadata = nodeValueAndRdeps.getValue();
// Replay events once change-pruned.
replay(ValueWithMetadata.getEvents(valueMaybeWithMetadata));
// Tell the receiver that the value was not actually changed this run.
evaluatorContext
.getProgressReceiver()
.evaluated(
skyKey,
EvaluationState.get(valueMaybeWithMetadata, /* changed= */ false),
/* newValue= */ null,
/* newError= */ null,
/* directDeps= */ null);
if (!evaluatorContext.keepGoing(skyKey) && nodeEntry.getErrorInfo() != null) {
if (!evaluatorContext.getVisitor().preventNewEvaluations()) {
return DirtyOutcome.ALREADY_PROCESSED;
}
throw SchedulerException.ofError(nodeEntry.getErrorInfo(), skyKey, rDepsToSignal);
}
evaluatorContext.signalParentsAndEnqueueIfReady(
skyKey, rDepsToSignal, nodeEntry.getVersion());
return DirtyOutcome.ALREADY_PROCESSED;
case NEEDS_REBUILDING:
nodeEntry.markRebuilding();
return DirtyOutcome.NEEDS_EVALUATION;
case REBUILDING:
return DirtyOutcome.NEEDS_EVALUATION;
default:
throw new IllegalStateException("key: " + skyKey + ", entry: " + nodeEntry);
}
}
/** Returns whether the parent has both been signalled and also is ready for evaluation. */
@CanIgnoreReturnValue
private boolean handleKnownChildrenForDirtyNode(
Collection<SkyKey> knownChildren,
NodeBatch oldChildren,
NodeEntry nodeEntry,
boolean enqueueParentIfReady,
@Nullable SkyFunctionEnvironment environmentIfEnqueuing)
throws InterruptedException {
boolean parentIsSignalledAndReady = false;
for (SkyKey directDep : knownChildren) {
NodeEntry directDepEntry =
checkNotNull(
oldChildren.get(directDep),
"Dirty parent had missing child (child=%s, parent=%s %s)",
directDep,
skyKey,
nodeEntry);
parentIsSignalledAndReady |=
enqueueChild(
skyKey,
nodeEntry,
directDep,
directDepEntry,
/* depAlreadyExists= */ true,
enqueueParentIfReady,
environmentIfEnqueuing);
}
return parentIsSignalledAndReady;
}
@Override
public void run() {
SkyFunctionEnvironment env = null;
try {
NodeEntry nodeEntry = graph.get(null, Reason.EVALUATION, skyKey);
if (nodeEntry == null || !nodeEntry.isReadyToEvaluate()) {
checkState(skyKey.supportsPartialReevaluation(), "%s %s", skyKey, nodeEntry);
evaluatorContext.getProgressReceiver().removeFromInflight(skyKey);
return;
}
try {
evaluatorContext.getProgressReceiver().stateStarting(skyKey, NodeState.CHECK_DIRTY);
if (maybeHandleDirtyNode(nodeEntry) == DirtyOutcome.ALREADY_PROCESSED) {
return;
}
} finally {
evaluatorContext.getProgressReceiver().stateEnding(skyKey, NodeState.CHECK_DIRTY);
}
ImmutableSet<SkyKey> oldDeps = nodeEntry.getAllRemainingDirtyDirectDeps();
try {
evaluatorContext
.getProgressReceiver()
.stateStarting(skyKey, NodeState.INITIALIZING_ENVIRONMENT);
env =
SkyFunctionEnvironment.create(
skyKey,
nodeEntry.getTemporaryDirectDeps(),
oldDeps,
nodeEntry.getMaxTransitiveSourceVersion(),
evaluatorContext);
} catch (UndonePreviouslyRequestedDeps undonePreviouslyRequestedDeps) {
handleUndonePreviouslyRequestedDep(nodeEntry);
return;
} finally {
evaluatorContext
.getProgressReceiver()
.stateEnding(skyKey, NodeState.INITIALIZING_ENVIRONMENT);
}
SkyFunctionName functionName = skyKey.functionName();
SkyFunction skyFunction =
checkNotNull(
evaluatorContext.getSkyFunctions().get(functionName),
"Unable to find SkyFunction '%s' for node with key %s, %s",
functionName,
skyKey,
nodeEntry);
SkyValue value = null;
try (var s =
Profiler.instance()
.profile(ProfilerTask.SKYFUNCTION, skyKey.functionName().getName())) {
try {
evaluatorContext.getProgressReceiver().stateStarting(skyKey, NodeState.COMPUTE);
value = skyFunction.compute(skyKey, env);
} finally {
evaluatorContext.getProgressReceiver().stateEnding(skyKey, NodeState.COMPUTE);
}
} catch (SkyFunctionException builderException) {
// TODO(b/261604460): invalidating the state cache here appears to be load-bearing for
// error propagation. It ought to be allowed to invalidate it only after the following
// early return checks pass, but something is misusing the state cache, and moving it
// causes tests to fail.
stateCache.invalidate(skyKey);
// In keep-going mode, we do not let SkyFunctions complete with a thrown error if they
// have missing deps. Instead, we wait until their deps are done and restart the
// SkyFunction, so we can have a definitive error and definitive graph structure, thus
// avoiding non-determinism. It's completely reasonable for SkyFunctions to throw eagerly
// because they do not know if they are in keep-going mode.
if (!evaluatorContext.keepGoing(skyKey) || !env.valuesMissing()) {
if (nodeEntry.hasUnsignaledDeps()) {
// This is a partial reevaluation. It is not safe to set the error because a dep may
// yet signal this node. We return (without preventing new evaluations) so that any
// not-yet-complete deps can complete and signal this node.
return;
}
if (maybeHandleRegisteringNewlyDiscoveredDepsForDoneEntry(
skyKey, nodeEntry, oldDeps, env, evaluatorContext.keepGoing(skyKey))) {
// A newly requested dep transitioned from done to dirty before this node finished.
// It is not safe to set the error because the now-dirty dep has not signaled this
// node. We return (without preventing new evaluations) so that the dep can complete
// and signal this node.
return;
}
try {
env.ensurePreviouslyRequestedDepsFetched();
} catch (UndonePreviouslyRequestedDeps e) {
handleUndonePreviouslyRequestedDep(nodeEntry);
return;
}
boolean shouldFailFast =
!evaluatorContext.keepGoing(skyKey) || builderException.isCatastrophic();
if (shouldFailFast) {
// After we commit this error to the graph but before the doMutatingEvaluation call
// completes with the error there is a race-like opportunity for the error to be used,
// either by an in-flight computation or by a future computation.
if (!evaluatorContext.getVisitor().preventNewEvaluations()) {
// This is not the first error encountered, so we ignore it so that we can terminate
// with the first error.
return;
} else {
logger.atWarning().withCause(builderException).log(
"Aborting evaluation while evaluating %s", skyKey);
}
}
ReifiedSkyFunctionException reifiedBuilderException =
new ReifiedSkyFunctionException(builderException);
boolean isTransitivelyTransient =
reifiedBuilderException.isTransient()
|| env.isAnyDirectDepErrorTransitivelyTransient()
|| env.isAnyNewlyRequestedDepErrorTransitivelyTransient();
ErrorInfo errorInfo =
evaluatorContext
.getErrorInfoManager()
.fromException(skyKey, reifiedBuilderException, isTransitivelyTransient);
env.setError(nodeEntry, errorInfo);
Set<SkyKey> rdepsToBubbleUpTo = env.commitAndGetParents(nodeEntry);
if (shouldFailFast) {
evaluatorContext.signalParentsOnAbort(
skyKey, rdepsToBubbleUpTo, nodeEntry.getVersion());
throw SchedulerException.ofError(errorInfo, skyKey, rdepsToBubbleUpTo);
}
evaluatorContext.signalParentsAndEnqueueIfReady(
skyKey, rdepsToBubbleUpTo, nodeEntry.getVersion());
return;
}
} catch (RuntimeException re) {
// Programmer error (most likely NPE or a failed precondition in a SkyFunction). Output
// some context together with the exception.
String msg = prepareCrashMessage(skyKey, nodeEntry.getInProgressReverseDeps());
RuntimeException ex = new RuntimeException(msg, re);
evaluatorContext.getVisitor().noteCrash(ex);
throw ex;
} finally {
env.doneBuilding();
}
// For any `SkyKey`s, regardless of partially evaluated or not, the node's Max Transitive
// Source Version so far is always tracked at the end of a Skyframe restart.
// This effort makes it meaningless to fetch MTSV of all deps during
// INITIALIZE_ENVIRONMENT's batch prefetch, and resolves a blocker to remove batch prefetch
// from INITIALIZE_ENVIRONMENT. Also, `SkyFunctionEnvironment#PartialEvaluation` subclass
// starts to support `getMaxTransitiveSourceVersionSoFar()` method.
// TODO(b/324948927): This comment should be rephrased when batch prefetch is removed from
// INITIALIZE_ENVIRONMENT PHASE.
nodeEntry.setTemporaryMaxTransitiveSourceVersion(env.getMaxTransitiveSourceVersionSoFar());
if (value instanceof Reset) {
if (nodeEntry.hasUnsignaledDeps()) {
// This is a partial reevaluation. It is not safe to reset the node because a dep may
// be racing to signal it.
return;
}
dirtyRewindGraphAndResetEntry(skyKey, nodeEntry, (Reset) value);
stateCache.invalidate(skyKey);
cancelExternalDeps(env);
evaluatorContext.getVisitor().enqueueEvaluation(skyKey, null);
return;
}
Set<SkyKey> newDeps = env.getNewlyRequestedDeps();
if (value != null) {
if (nodeEntry.hasUnsignaledDeps()) {
// This is a partial reevaluation. It is not safe to set the value because a dep may be
// racing to signal this node.
return;
}
try {
env.ensurePreviouslyRequestedDepsFetched();
} catch (UndonePreviouslyRequestedDeps e) {
handleUndonePreviouslyRequestedDep(nodeEntry);
return;
}
checkState(
!env.valuesMissing(),
"Evaluation of %s returned non-null value but requested dependencies that weren't "
+ "computed yet (one of %s), NodeEntry: %s",
skyKey,
newDeps,
nodeEntry);
stateCache.invalidate(skyKey);
try {
evaluatorContext.getProgressReceiver().stateStarting(skyKey, NodeState.COMMIT);
if (maybeHandleRegisteringNewlyDiscoveredDepsForDoneEntry(
skyKey, nodeEntry, oldDeps, env, evaluatorContext.keepGoing(skyKey))) {
// A newly requested dep transitioned from done to dirty before this node finished.
// This node will be signalled again, and so we should return.
return;
}
env.setValue(value);
Set<SkyKey> reverseDeps = env.commitAndGetParents(nodeEntry);
evaluatorContext.signalParentsAndEnqueueIfReady(
skyKey, reverseDeps, nodeEntry.getVersion());
} finally {
evaluatorContext.getProgressReceiver().stateEnding(skyKey, NodeState.COMMIT);
}
return;
}
SkyKey childErrorKey = env.getDepErrorKey();
if (childErrorKey != null) {
checkState(
!evaluatorContext.keepGoing(skyKey), "%s %s %s", skyKey, nodeEntry, childErrorKey);
// We encountered a child error in noKeepGoing mode, so we want to fail fast. But we first
// need to add the edge between the current node and the child error it requested so that
// error bubbling can occur. Note that this edge will subsequently be removed during graph
// cleaning (since the current node will never be committed to the graph).
NodeEntry childErrorEntry =
checkNotNull(
graph.get(skyKey, Reason.OTHER, childErrorKey),
"skyKey: %s, nodeEntry: %s childErrorKey: %s",
skyKey,
nodeEntry,
childErrorKey);
if (newDeps.contains(childErrorKey)) {
// Add this dep if it was just requested. In certain rare race conditions (see
// MemoizingEvaluatorTest.cachedErrorCausesRestart) this dep may have already been
// requested.
nodeEntry.addSingletonTemporaryDirectDep(childErrorKey);
DependencyState childErrorState;
if (oldDeps.contains(childErrorKey)) {
childErrorState = childErrorEntry.checkIfDoneForDirtyReverseDep(skyKey);
} else {
childErrorState = childErrorEntry.addReverseDepAndCheckIfDone(skyKey);
}
if (childErrorState != DependencyState.DONE) {
// The child in error may have transitioned from done to dirty between when this node
// discovered the error and now. Notify the graph inconsistency receiver so that we
// can crash if that's unexpected.
// We don't enqueue the child, even if it returns NEEDS_SCHEDULING, because we are
// about to shut down evaluation.
evaluatorContext
.getGraphInconsistencyReceiver()
.noteInconsistencyAndMaybeThrow(
skyKey,
ImmutableList.of(childErrorKey),
Inconsistency.BUILDING_PARENT_FOUND_UNDONE_CHILD);
}
}
SkyValue childErrorInfoMaybe =
checkNotNull(
env.maybeGetValueFromErrorOrDeps(childErrorKey),
"dep error found but then lost while building: %s %s",
skyKey,
childErrorKey);
ErrorInfo childErrorInfo =
checkNotNull(
ValueWithMetadata.getMaybeErrorInfo(childErrorInfoMaybe),
"dep error found but then wasn't an error while building: %s %s %s",
skyKey,
childErrorKey,
childErrorInfoMaybe);
evaluatorContext.getVisitor().preventNewEvaluations();
// TODO(b/166268889): Remove when fixed.
if (childErrorInfo.getException() instanceof IOException) {
logger.atInfo().withCause(childErrorInfo.getException()).log(
"Child %s with IOException forced abort of %s", childErrorKey, skyKey);
}
throw SchedulerException.ofError(childErrorInfo, childErrorKey, ImmutableSet.of(skyKey));
}
// TODO(bazel-team): This code is not safe to interrupt, because we would lose the state in
// newDirectDeps.
// TODO(bazel-team): An ill-behaved SkyFunction can throw us into an infinite loop where we
// add more dependencies on every run. [skyframe-core]
// Add all the newly requested dependencies to the temporary direct deps. Note that
// newDirectDeps does not contain any elements in common with the already existing temporary
// direct deps. uniqueNewDeps will be the set of unique keys contained in newDirectDeps.
env.addTemporaryDirectDepsTo(nodeEntry);
List<ListenableFuture<?>> externalDeps = env.externalDeps;
// If the key does not support partial reevaluation and there were no newly requested
// dependencies, then at least one of them was in error or there is a bug in the SkyFunction
// implementation. The environment has collected its errors, so we just order it to be
// built.
if (newDeps.isEmpty() && externalDeps == null && !skyKey.supportsPartialReevaluation()) {
checkState(
!env.getChildErrorInfos().isEmpty(),
"Evaluation of SkyKey failed and no dependencies were requested: %s %s",
skyKey,
nodeEntry);
// If the child error was catastrophic, committing this parent to the graph is not
// necessary, but since we don't do error bubbling in catastrophes, it doesn't violate any
// invariants either.
Set<SkyKey> reverseDeps = env.commitAndGetParents(nodeEntry);
evaluatorContext.signalParentsAndEnqueueIfReady(
skyKey, reverseDeps, nodeEntry.getVersion());
return;
}
// If there are external deps, we register that fact on the NodeEntry before we enqueue
// child nodes in order to prevent the current node from being re-enqueued between here and
// the call to registerExternalDeps below.
if (externalDeps != null) {
nodeEntry.addExternalDep();
}
// We want to split apart the dependencies that existed for this node the last time we did
// an evaluation and those that were introduced in this evaluation. To be clear, the prefix
// "newDeps" refers to newly discovered this time around after a SkyFunction#compute call
// and not to be confused with the oldDeps variable which refers to the last evaluation,
// i.e. a prior call to ParallelEvaluator#eval.
Collection<SkyKey> newDepsThatWerentInTheLastEvaluation;
ImmutableList<SkyKey> newDepsThatWereInTheLastEvaluation;
if (oldDeps.isEmpty()) {
// When there are no old deps (clean evaluations), avoid set views which have O(n) size.
newDepsThatWerentInTheLastEvaluation = newDeps;
newDepsThatWereInTheLastEvaluation = ImmutableList.of();
} else {
newDepsThatWerentInTheLastEvaluation =
ImmutableList.copyOf(Sets.difference(newDeps, oldDeps));
newDepsThatWereInTheLastEvaluation =
ImmutableList.copyOf(Sets.intersection(newDeps, oldDeps));
}
InterruptibleSupplier<NodeBatch> newDepsThatWerentInTheLastEvaluationNodes =
graph.createIfAbsentBatchAsync(
skyKey, Reason.RDEP_ADDITION, newDepsThatWerentInTheLastEvaluation);
ImmutableSet<SkyKey> resetDeps = nodeEntry.getResetDirectDeps();
// Due to multi-threading, either the following call to handleKnownChildrenForDirtyNode or
// the enqueueChild loop may cause the current node to be re-enqueued (and evaluated) if all
// new children of this node are already done. Therefore, the rest of this method cannot
// assume that the node is dirty.
handleKnownChildrenForDirtyNode(
newDepsThatWereInTheLastEvaluation,
graph.getBatch(skyKey, Reason.ENQUEUING_CHILD, newDepsThatWereInTheLastEvaluation),
nodeEntry,
/* enqueueParentIfReady= */ true,
env);
NodeBatch newNodes = newDepsThatWerentInTheLastEvaluationNodes.get();
for (SkyKey newDirectDep : newDepsThatWerentInTheLastEvaluation) {
enqueueChild(
skyKey,
nodeEntry,
newDirectDep,
newNodes.get(newDirectDep),
/* depAlreadyExists= */ resetDeps.contains(newDirectDep),
/* enqueueParentIfReady= */ true,
env);
}
if (externalDeps != null) {
// This can cause the current node to be re-enqueued if all futures are already done.
// This is an exception to the rule above that there must not be code below the for
// loop. It is safe because we call nodeEntry.addExternalDep above, which prevents
// re-enqueueing of the current node in the above loop if externalDeps != null.
evaluatorContext.getVisitor().registerExternalDeps(skyKey, nodeEntry, externalDeps);
}
// Do not put any code here! Any code here can race with a re-evaluation of this same node
// in another thread.
} catch (InterruptedException ie) {
// The current thread can be interrupted at various places during evaluation or while
// committing the result in this method. Since we only register the future(s) with the
// underlying AbstractQueueVisitor in the registerExternalDeps call above, we have to make
// sure that any known futures are correctly canceled if we do not reach that call. Note
// that it is safe to cancel a future multiple times.
cancelExternalDeps(env);
// InterruptedException cannot be thrown by Runnable.run, so we must wrap it.
// Interrupts can be caught by both the Evaluator and the AbstractQueueVisitor.
// The former will unwrap the IE and propagate it as is; the latter will throw a new IE.
throw SchedulerException.ofInterruption(ie, skyKey);
}
}
private void handleUndonePreviouslyRequestedDep(NodeEntry nodeEntry) {
// If a previously requested dep is no longer done, restart this node from scratch.
stateCache.invalidate(skyKey);
resetEntry(skyKey, nodeEntry);
evaluatorContext.getVisitor().enqueueEvaluation(skyKey, null);
}
private void cancelExternalDeps(SkyFunctionEnvironment env) {
if (env != null && env.externalDeps != null) {
for (ListenableFuture<?> future : env.externalDeps) {
future.cancel(/* mayInterruptIfRunning= */ true);
}
}
}
private String prepareCrashMessage(SkyKey skyKey, Iterable<SkyKey> reverseDeps) {
StringBuilder reverseDepDump = new StringBuilder();
for (SkyKey key : reverseDeps) {
if (reverseDepDump.length() > MAX_REVERSEDEP_DUMP_LENGTH) {
reverseDepDump.append(", ...");
break;
}
if (reverseDepDump.length() > 0) {
reverseDepDump.append(", ");
}
reverseDepDump.append("'");
reverseDepDump.append(key);
reverseDepDump.append("'");
}
return String.format(
"Unrecoverable error while evaluating node '%s' (requested by nodes %s)",
skyKey, reverseDepDump);
}
private static final int MAX_REVERSEDEP_DUMP_LENGTH = 1000;
}
protected void replay(NestedSet<Reportable> transitiveEvents) {
// Replaying actions is done on a small number of nodes, but potentially over a large dependency
// graph. Under those conditions, using the regular NestedSet flattening with .toList() is more
// efficient than using NestedSetVisitor's custom traversal logic.
evaluatorContext.getReplayingNestedSetEventVisitor().visit(transitiveEvents.toList());
}
/**
* Resets {@code entry}, and the other nodes specified by {@code restart.rewindGraph()} will be
* marked changed via postorder DFS.
*
* <p>{@code restart.rewindGraph()} must be empty or must contain {@code key}.
*
* <p>TODO(b/123993876): this should verify that edges in rewindGraph correspond to deps in the
* Skyframe graph. Will require a safe way of requesting deps for nodes which may not be done.
*/
// Nodes must be marked changed via postorder DFS. To see why, suppose we have this graph:
//
// FailedNode SomeOtherRdepOfR1
// | /
// | -----
// | /
// R1
// |
// R2
//
// Suppose FailedNode (FN) fails and requires that R1 and R2 must be dirtied and run again.
// Suppose they aren't dirtied via postorder DFS, so R1 is dirtied first.
//
// Then, the evaluation thread working on dirtying these nodes is suspended.
//
// On a separate evaluation thread, SomeOtherRdepOfR1 requests R1. R1 is scheduled for evaluation,
// checks its dep R2, and because R2 is done, R1 completes without scheduling R2 for evaluation.
//
// Then, the evaluation thread working on dirtying these nodes continues its work. It dirties
// R2 and schedules FN for evaluation.
//
// When FN next evaluates, it requests R1, and because R1 is done, R2 is not scheduled for
// evaluation, contrary to FN's expectations.
private void dirtyRewindGraphAndResetEntry(SkyKey key, NodeEntry entry, Reset restart)
throws InterruptedException {
ImmutableGraph<SkyKey> rewindGraph = restart.rewindGraph();
checkState(
rewindGraph.nodes().contains(key),
"Rewind graph missing evaluating key %s: %s",
key,
rewindGraph);
ImmutableList.Builder<SkyKey> builder =
ImmutableList.builderWithExpectedSize(rewindGraph.nodes().size() - 1);
for (SkyKey k : Traverser.forGraph(rewindGraph).depthFirstPostOrder(key)) {
if (!k.equals(key)) {
builder.add(k);
}
}
ImmutableList<SkyKey> childrenToRestart = builder.build();
if (!childrenToRestart.isEmpty()) {
evaluatorContext
.getGraphInconsistencyReceiver()
.noteInconsistencyAndMaybeThrow(
key, childrenToRestart, Inconsistency.PARENT_FORCE_REBUILD_OF_CHILD);
NodeBatch children =
evaluatorContext.getGraph().getBatch(key, Reason.REWINDING, childrenToRestart);
for (SkyKey childToRestart : childrenToRestart) {
NodeEntry childEntry =
checkNotNull(
children.get(childToRestart),
"Missing child for rewinding: %s (parent=%s)",
childToRestart,
key);
if (childEntry.markDirty(DirtyType.REWIND) != null) {
evaluatorContext.getProgressReceiver().dirtied(childToRestart, DirtyType.REWIND);
}
}
}
resetEntry(key, entry);
}
private void resetEntry(SkyKey key, NodeEntry entry) {
evaluatorContext
.getGraphInconsistencyReceiver()
.noteInconsistencyAndMaybeThrow(key, /* otherKeys= */ null, Inconsistency.RESET_REQUESTED);
entry.resetEvaluationFromScratch();
}
void propagateEvaluatorContextCrashIfAny() {
if (!evaluatorContext.getVisitor().getCrashes().isEmpty()) {
evaluatorContext
.getReporter()
.handle(Event.error("Crashes detected: " + evaluatorContext.getVisitor().getCrashes()));
throw checkNotNull(Iterables.getFirst(evaluatorContext.getVisitor().getCrashes(), null));
}
}
static void propagateInterruption(SchedulerException e) throws InterruptedException {
boolean mustThrowInterrupt = Thread.interrupted();
if (e.getCause() != null) {
throwIfInstanceOf(e.getCause(), InterruptedException.class);
throwIfUnchecked(e.getCause());
}
if (mustThrowInterrupt) {
// As per the contract of AbstractQueueVisitor#work, if an unchecked exception is thrown and
// the build is interrupted, the thrown exception is what will be rethrown. Since the user
// presumably wanted to interrupt the build, we ignore the thrown SchedulerException (which
// doesn't indicate a programming bug) and throw an InterruptedException.
throw new InterruptedException();
}
}
/**
* Add any newly discovered deps that were registered during the run of a SkyFunction that
* finished by returning a value or throwing an error. SkyFunctions may throw errors even if all
* their deps were not provided -- we trust that a SkyFunction might know it should throw an error
* even if not all of its requested deps are done. However, that means we're assuming the
* SkyFunction would throw that same error if all of its requested deps were done. Unfortunately,
* there is no way to enforce that condition.
*
* <p>Returns {@code true} if any newly discovered dep is dirty when this node registers itself as
* an rdep and if one of those dirty deps will schedule this node for evaluation.
*
* <p>This can happen if a newly discovered dep transitions from done to dirty between when this
* node's evaluation accessed the dep's value and here. Adding this node as an rdep of that dep
* (or checking that this node is an rdep of that dep) will cause this node to be signalled when
* that dep completes.
*
* <p>If this returns {@code true}, this node should not actually finish, and this evaluation
* attempt should make no changes to the node after this method returns, because a completing dep
* may schedule a new evaluation attempt at any time.
*/
private boolean maybeHandleRegisteringNewlyDiscoveredDepsForDoneEntry(
SkyKey skyKey,
NodeEntry entry,
ImmutableSet<SkyKey> oldDeps,
SkyFunctionEnvironment env,
boolean keepGoing)
throws InterruptedException {
// We don't expect any unfinished deps in a keep-going build.
if (!keepGoing) {
env.removeUndoneNewlyRequestedDeps();
}
Set<SkyKey> newDeps = env.getNewlyRequestedDeps();
if (newDeps.isEmpty()) {
return false;
}
env.addTemporaryDirectDepsTo(entry);
// Reset deps is usually empty. Avoid an unnecessary allocation from Sets.union if possible.
ImmutableSet<SkyKey> resetDeps = entry.getResetDirectDeps();
Set<SkyKey> alreadyRegisteredDeps;
if (resetDeps.isEmpty()) {
alreadyRegisteredDeps = oldDeps;
} else if (oldDeps.isEmpty()) {
alreadyRegisteredDeps = resetDeps;
} else {
alreadyRegisteredDeps = Sets.union(oldDeps, resetDeps);
}
Collection<SkyKey> newlyAddedNewDeps;
ImmutableCollection<SkyKey> previouslyRegisteredNewDeps;
if (alreadyRegisteredDeps.isEmpty()) {
newlyAddedNewDeps = newDeps;
previouslyRegisteredNewDeps = ImmutableSet.of();
} else {
newlyAddedNewDeps = ImmutableList.copyOf(Sets.difference(newDeps, alreadyRegisteredDeps));
previouslyRegisteredNewDeps =
ImmutableList.copyOf(Sets.intersection(newDeps, alreadyRegisteredDeps));
}