forked from golang/go
/
trace.go
1925 lines (1727 loc) · 62.8 KB
/
trace.go
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 2014 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
//go:build !goexperiment.exectracer2
// Go execution tracer.
// The tracer captures a wide range of execution events like goroutine
// creation/blocking/unblocking, syscall enter/exit/block, GC-related events,
// changes of heap size, processor start/stop, etc and writes them to a buffer
// in a compact form. A precise nanosecond-precision timestamp and a stack
// trace is captured for most events.
// See https://golang.org/s/go15trace for more info.
package runtime
import (
"internal/abi"
"internal/goarch"
"internal/goos"
"runtime/internal/atomic"
"runtime/internal/sys"
"unsafe"
)
// Event types in the trace, args are given in square brackets.
const (
traceEvNone = 0 // unused
traceEvBatch = 1 // start of per-P batch of events [pid, timestamp]
traceEvFrequency = 2 // contains tracer timer frequency [frequency (ticks per second)]
traceEvStack = 3 // stack [stack id, number of PCs, array of {PC, func string ID, file string ID, line}]
traceEvGomaxprocs = 4 // current value of GOMAXPROCS [timestamp, GOMAXPROCS, stack id]
traceEvProcStart = 5 // start of P [timestamp, thread id]
traceEvProcStop = 6 // stop of P [timestamp]
traceEvGCStart = 7 // GC start [timestamp, seq, stack id]
traceEvGCDone = 8 // GC done [timestamp]
traceEvSTWStart = 9 // STW start [timestamp, kind]
traceEvSTWDone = 10 // STW done [timestamp]
traceEvGCSweepStart = 11 // GC sweep start [timestamp, stack id]
traceEvGCSweepDone = 12 // GC sweep done [timestamp, swept, reclaimed]
traceEvGoCreate = 13 // goroutine creation [timestamp, new goroutine id, new stack id, stack id]
traceEvGoStart = 14 // goroutine starts running [timestamp, goroutine id, seq]
traceEvGoEnd = 15 // goroutine ends [timestamp]
traceEvGoStop = 16 // goroutine stops (like in select{}) [timestamp, stack]
traceEvGoSched = 17 // goroutine calls Gosched [timestamp, stack]
traceEvGoPreempt = 18 // goroutine is preempted [timestamp, stack]
traceEvGoSleep = 19 // goroutine calls Sleep [timestamp, stack]
traceEvGoBlock = 20 // goroutine blocks [timestamp, stack]
traceEvGoUnblock = 21 // goroutine is unblocked [timestamp, goroutine id, seq, stack]
traceEvGoBlockSend = 22 // goroutine blocks on chan send [timestamp, stack]
traceEvGoBlockRecv = 23 // goroutine blocks on chan recv [timestamp, stack]
traceEvGoBlockSelect = 24 // goroutine blocks on select [timestamp, stack]
traceEvGoBlockSync = 25 // goroutine blocks on Mutex/RWMutex [timestamp, stack]
traceEvGoBlockCond = 26 // goroutine blocks on Cond [timestamp, stack]
traceEvGoBlockNet = 27 // goroutine blocks on network [timestamp, stack]
traceEvGoSysCall = 28 // syscall enter [timestamp, stack]
traceEvGoSysExit = 29 // syscall exit [timestamp, goroutine id, seq, real timestamp]
traceEvGoSysBlock = 30 // syscall blocks [timestamp]
traceEvGoWaiting = 31 // denotes that goroutine is blocked when tracing starts [timestamp, goroutine id]
traceEvGoInSyscall = 32 // denotes that goroutine is in syscall when tracing starts [timestamp, goroutine id]
traceEvHeapAlloc = 33 // gcController.heapLive change [timestamp, heap_alloc]
traceEvHeapGoal = 34 // gcController.heapGoal() (formerly next_gc) change [timestamp, heap goal in bytes]
traceEvTimerGoroutine = 35 // not currently used; previously denoted timer goroutine [timer goroutine id]
traceEvFutileWakeup = 36 // not currently used; denotes that the previous wakeup of this goroutine was futile [timestamp]
traceEvString = 37 // string dictionary entry [ID, length, string]
traceEvGoStartLocal = 38 // goroutine starts running on the same P as the last event [timestamp, goroutine id]
traceEvGoUnblockLocal = 39 // goroutine is unblocked on the same P as the last event [timestamp, goroutine id, stack]
traceEvGoSysExitLocal = 40 // syscall exit on the same P as the last event [timestamp, goroutine id, real timestamp]
traceEvGoStartLabel = 41 // goroutine starts running with label [timestamp, goroutine id, seq, label string id]
traceEvGoBlockGC = 42 // goroutine blocks on GC assist [timestamp, stack]
traceEvGCMarkAssistStart = 43 // GC mark assist start [timestamp, stack]
traceEvGCMarkAssistDone = 44 // GC mark assist done [timestamp]
traceEvUserTaskCreate = 45 // trace.NewTask [timestamp, internal task id, internal parent task id, name string, stack]
traceEvUserTaskEnd = 46 // end of a task [timestamp, internal task id, stack]
traceEvUserRegion = 47 // trace.WithRegion [timestamp, internal task id, mode(0:start, 1:end), name string, stack]
traceEvUserLog = 48 // trace.Log [timestamp, internal task id, key string id, stack, value string]
traceEvCPUSample = 49 // CPU profiling sample [timestamp, real timestamp, real P id (-1 when absent), goroutine id, stack]
traceEvCount = 50
// Byte is used but only 6 bits are available for event type.
// The remaining 2 bits are used to specify the number of arguments.
// That means, the max event type value is 63.
)
// traceBlockReason is an enumeration of reasons a goroutine might block.
// This is the interface the rest of the runtime uses to tell the
// tracer why a goroutine blocked. The tracer then propagates this information
// into the trace however it sees fit.
//
// Note that traceBlockReasons should not be compared, since reasons that are
// distinct by name may *not* be distinct by value.
type traceBlockReason uint8
// For maximal efficiency, just map the trace block reason directly to a trace
// event.
const (
traceBlockGeneric traceBlockReason = traceEvGoBlock
traceBlockForever = traceEvGoStop
traceBlockNet = traceEvGoBlockNet
traceBlockSelect = traceEvGoBlockSelect
traceBlockCondWait = traceEvGoBlockCond
traceBlockSync = traceEvGoBlockSync
traceBlockChanSend = traceEvGoBlockSend
traceBlockChanRecv = traceEvGoBlockRecv
traceBlockGCMarkAssist = traceEvGoBlockGC
traceBlockGCSweep = traceEvGoBlock
traceBlockSystemGoroutine = traceEvGoBlock
traceBlockPreempted = traceEvGoBlock
traceBlockDebugCall = traceEvGoBlock
traceBlockUntilGCEnds = traceEvGoBlock
traceBlockSleep = traceEvGoSleep
)
const (
// Timestamps in trace are cputicks/traceTickDiv.
// This makes absolute values of timestamp diffs smaller,
// and so they are encoded in less number of bytes.
// 64 on x86 is somewhat arbitrary (one tick is ~20ns on a 3GHz machine).
// The suggested increment frequency for PowerPC's time base register is
// 512 MHz according to Power ISA v2.07 section 6.2, so we use 16 on ppc64
// and ppc64le.
traceTimeDiv = 16 + 48*(goarch.Is386|goarch.IsAmd64)
// Maximum number of PCs in a single stack trace.
// Since events contain only stack id rather than whole stack trace,
// we can allow quite large values here.
traceStackSize = 128
// Identifier of a fake P that is used when we trace without a real P.
traceGlobProc = -1
// Maximum number of bytes to encode uint64 in base-128.
traceBytesPerNumber = 10
// Shift of the number of arguments in the first event byte.
traceArgCountShift = 6
)
// trace is global tracing context.
var trace struct {
// trace.lock must only be acquired on the system stack where
// stack splits cannot happen while it is held.
lock mutex // protects the following members
enabled bool // when set runtime traces events
shutdown bool // set when we are waiting for trace reader to finish after setting enabled to false
headerWritten bool // whether ReadTrace has emitted trace header
footerWritten bool // whether ReadTrace has emitted trace footer
shutdownSema uint32 // used to wait for ReadTrace completion
seqStart uint64 // sequence number when tracing was started
startTicks int64 // cputicks when tracing was started
endTicks int64 // cputicks when tracing was stopped
startNanotime int64 // nanotime when tracing was started
endNanotime int64 // nanotime when tracing was stopped
startTime traceTime // traceClockNow when tracing started
endTime traceTime // traceClockNow when tracing stopped
seqGC uint64 // GC start/done sequencer
reading traceBufPtr // buffer currently handed off to user
empty traceBufPtr // stack of empty buffers
fullHead traceBufPtr // queue of full buffers
fullTail traceBufPtr
stackTab traceStackTable // maps stack traces to unique ids
// cpuLogRead accepts CPU profile samples from the signal handler where
// they're generated. It uses a two-word header to hold the IDs of the P and
// G (respectively) that were active at the time of the sample. Because
// profBuf uses a record with all zeros in its header to indicate overflow,
// we make sure to make the P field always non-zero: The ID of a real P will
// start at bit 1, and bit 0 will be set. Samples that arrive while no P is
// running (such as near syscalls) will set the first header field to 0b10.
// This careful handling of the first header field allows us to store ID of
// the active G directly in the second field, even though that will be 0
// when sampling g0.
cpuLogRead *profBuf
// cpuLogBuf is a trace buffer to hold events corresponding to CPU profile
// samples, which arrive out of band and not directly connected to a
// specific P.
cpuLogBuf traceBufPtr
reader atomic.Pointer[g] // goroutine that called ReadTrace, or nil
signalLock atomic.Uint32 // protects use of the following member, only usable in signal handlers
cpuLogWrite *profBuf // copy of cpuLogRead for use in signal handlers, set without signalLock
// Dictionary for traceEvString.
//
// TODO: central lock to access the map is not ideal.
// option: pre-assign ids to all user annotation region names and tags
// option: per-P cache
// option: sync.Map like data structure
stringsLock mutex
strings map[string]uint64
stringSeq uint64
// markWorkerLabels maps gcMarkWorkerMode to string ID.
markWorkerLabels [len(gcMarkWorkerModeStrings)]uint64
bufLock mutex // protects buf
buf traceBufPtr // global trace buffer, used when running without a p
}
// gTraceState is per-G state for the tracer.
type gTraceState struct {
sysExitTime traceTime // timestamp when syscall has returned
tracedSyscallEnter bool // syscall or cgo was entered while trace was enabled or StartTrace has emitted EvGoInSyscall about this goroutine
seq uint64 // trace event sequencer
lastP puintptr // last P emitted an event for this goroutine
}
// Unused; for compatibility with the new tracer.
func (s *gTraceState) reset() {}
// mTraceState is per-M state for the tracer.
type mTraceState struct {
startingTrace bool // this M is in TraceStart, potentially before traceEnabled is true
tracedSTWStart bool // this M traced a STW start, so it should trace an end
}
// pTraceState is per-P state for the tracer.
type pTraceState struct {
buf traceBufPtr
// inSweep indicates the sweep events should be traced.
// This is used to defer the sweep start event until a span
// has actually been swept.
inSweep bool
// swept and reclaimed track the number of bytes swept and reclaimed
// by sweeping in the current sweep loop (while inSweep was true).
swept, reclaimed uintptr
}
// traceLockInit initializes global trace locks.
func traceLockInit() {
lockInit(&trace.bufLock, lockRankTraceBuf)
lockInit(&trace.stringsLock, lockRankTraceStrings)
lockInit(&trace.lock, lockRankTrace)
lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
}
// traceBufHeader is per-P tracing buffer.
type traceBufHeader struct {
link traceBufPtr // in trace.empty/full
lastTime traceTime // when we wrote the last event
pos int // next write offset in arr
stk [traceStackSize]uintptr // scratch buffer for traceback
}
// traceBuf is per-P tracing buffer.
type traceBuf struct {
_ sys.NotInHeap
traceBufHeader
arr [64<<10 - unsafe.Sizeof(traceBufHeader{})]byte // underlying buffer for traceBufHeader.buf
}
// traceBufPtr is a *traceBuf that is not traced by the garbage
// collector and doesn't have write barriers. traceBufs are not
// allocated from the GC'd heap, so this is safe, and are often
// manipulated in contexts where write barriers are not allowed, so
// this is necessary.
//
// TODO: Since traceBuf is now embedded runtime/internal/sys.NotInHeap, this isn't necessary.
type traceBufPtr uintptr
func (tp traceBufPtr) ptr() *traceBuf { return (*traceBuf)(unsafe.Pointer(tp)) }
func (tp *traceBufPtr) set(b *traceBuf) { *tp = traceBufPtr(unsafe.Pointer(b)) }
func traceBufPtrOf(b *traceBuf) traceBufPtr {
return traceBufPtr(unsafe.Pointer(b))
}
// traceEnabled returns true if the trace is currently enabled.
//
// nosplit because it's called on the syscall path when stack movement is forbidden.
//
//go:nosplit
func traceEnabled() bool {
return trace.enabled
}
// traceShuttingDown returns true if the trace is currently shutting down.
//
//go:nosplit
func traceShuttingDown() bool {
return trace.shutdown
}
// traceLocker represents an M writing trace events. While a traceLocker value
// is valid, the tracer observes all operations on the G/M/P or trace events being
// written as happening atomically.
//
// This doesn't do much for the current tracer, because the current tracer doesn't
// need atomicity around non-trace runtime operations. All the state it needs it
// collects carefully during a STW.
type traceLocker struct {
enabled bool
}
// traceAcquire prepares this M for writing one or more trace events.
//
// This exists for compatibility with the upcoming new tracer; it doesn't do much
// in the current tracer.
//
// nosplit because it's called on the syscall path when stack movement is forbidden.
//
//go:nosplit
func traceAcquire() traceLocker {
if !traceEnabled() {
return traceLocker{false}
}
return traceLocker{true}
}
// ok returns true if the traceLocker is valid (i.e. tracing is enabled).
//
// nosplit because it's called on the syscall path when stack movement is forbidden.
//
//go:nosplit
func (tl traceLocker) ok() bool {
return tl.enabled
}
// traceRelease indicates that this M is done writing trace events.
//
// This exists for compatibility with the upcoming new tracer; it doesn't do anything
// in the current tracer.
//
// nosplit because it's called on the syscall path when stack movement is forbidden.
//
//go:nosplit
func traceRelease(tl traceLocker) {
}
// StartTrace enables tracing for the current process.
// While tracing, the data will be buffered and available via [ReadTrace].
// StartTrace returns an error if tracing is already enabled.
// Most clients should use the [runtime/trace] package or the [testing] package's
// -test.trace flag instead of calling StartTrace directly.
func StartTrace() error {
// Stop the world so that we can take a consistent snapshot
// of all goroutines at the beginning of the trace.
// Do not stop the world during GC so we ensure we always see
// a consistent view of GC-related events (e.g. a start is always
// paired with an end).
stw := stopTheWorldGC(stwStartTrace)
// Prevent sysmon from running any code that could generate events.
lock(&sched.sysmonlock)
// We are in stop-the-world, but syscalls can finish and write to trace concurrently.
// Exitsyscall could check trace.enabled long before and then suddenly wake up
// and decide to write to trace at a random point in time.
// However, such syscall will use the global trace.buf buffer, because we've
// acquired all p's by doing stop-the-world. So this protects us from such races.
lock(&trace.bufLock)
if trace.enabled || trace.shutdown {
unlock(&trace.bufLock)
unlock(&sched.sysmonlock)
startTheWorldGC(stw)
return errorString("tracing is already enabled")
}
// Can't set trace.enabled yet. While the world is stopped, exitsyscall could
// already emit a delayed event (see exitTicks in exitsyscall) if we set trace.enabled here.
// That would lead to an inconsistent trace:
// - either GoSysExit appears before EvGoInSyscall,
// - or GoSysExit appears for a goroutine for which we don't emit EvGoInSyscall below.
// To instruct traceEvent that it must not ignore events below, we set trace.startingTrace.
// trace.enabled is set afterwards once we have emitted all preliminary events.
mp := getg().m
mp.trace.startingTrace = true
// Obtain current stack ID to use in all traceEvGoCreate events below.
stkBuf := make([]uintptr, traceStackSize)
stackID := traceStackID(mp, stkBuf, 2)
profBuf := newProfBuf(2, profBufWordCount, profBufTagCount) // after the timestamp, header is [pp.id, gp.goid]
trace.cpuLogRead = profBuf
// We must not acquire trace.signalLock outside of a signal handler: a
// profiling signal may arrive at any time and try to acquire it, leading to
// deadlock. Because we can't use that lock to protect updates to
// trace.cpuLogWrite (only use of the structure it references), reads and
// writes of the pointer must be atomic. (And although this field is never
// the sole pointer to the profBuf value, it's best to allow a write barrier
// here.)
atomicstorep(unsafe.Pointer(&trace.cpuLogWrite), unsafe.Pointer(profBuf))
// World is stopped, no need to lock.
forEachGRace(func(gp *g) {
status := readgstatus(gp)
if status != _Gdead {
gp.trace.seq = 0
gp.trace.lastP = getg().m.p
// +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum.
id := trace.stackTab.put([]uintptr{logicalStackSentinel, startPCforTrace(gp.startpc) + sys.PCQuantum})
traceEvent(traceEvGoCreate, -1, gp.goid, uint64(id), stackID)
}
if status == _Gwaiting {
// traceEvGoWaiting is implied to have seq=1.
gp.trace.seq++
traceEvent(traceEvGoWaiting, -1, gp.goid)
}
if status == _Gsyscall {
gp.trace.seq++
gp.trace.tracedSyscallEnter = true
traceEvent(traceEvGoInSyscall, -1, gp.goid)
} else if status == _Gdead && gp.m != nil && gp.m.isextra {
// Trigger two trace events for the dead g in the extra m,
// since the next event of the g will be traceEvGoSysExit in exitsyscall,
// while calling from C thread to Go.
gp.trace.seq = 0
gp.trace.lastP = getg().m.p
// +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum.
id := trace.stackTab.put([]uintptr{logicalStackSentinel, startPCforTrace(0) + sys.PCQuantum}) // no start pc
traceEvent(traceEvGoCreate, -1, gp.goid, uint64(id), stackID)
gp.trace.seq++
gp.trace.tracedSyscallEnter = true
traceEvent(traceEvGoInSyscall, -1, gp.goid)
} else {
// We need to explicitly clear the flag. A previous trace might have ended with a goroutine
// not emitting a GoSysExit and clearing the flag, leaving it in a stale state. Clearing
// it here makes it unambiguous to any goroutine exiting a syscall racing with us that
// no EvGoInSyscall event was emitted for it. (It's not racy to set this flag here, because
// it'll only get checked when the goroutine runs again, which will be after the world starts
// again.)
gp.trace.tracedSyscallEnter = false
}
})
// Use a dummy traceLocker. The trace isn't enabled yet, but we can still write events.
tl := traceLocker{}
tl.ProcStart()
tl.GoStart()
// Note: startTicks needs to be set after we emit traceEvGoInSyscall events.
// If we do it the other way around, it is possible that exitsyscall will
// query sysExitTime after startTicks but before traceEvGoInSyscall timestamp.
// It will lead to a false conclusion that cputicks is broken.
trace.startTime = traceClockNow()
trace.startTicks = cputicks()
trace.startNanotime = nanotime()
trace.headerWritten = false
trace.footerWritten = false
// string to id mapping
// 0 : reserved for an empty string
// remaining: other strings registered by traceString
trace.stringSeq = 0
trace.strings = make(map[string]uint64)
trace.seqGC = 0
mp.trace.startingTrace = false
trace.enabled = true
// Register runtime goroutine labels.
_, pid, bufp := traceAcquireBuffer()
for i, label := range gcMarkWorkerModeStrings[:] {
trace.markWorkerLabels[i], bufp = traceString(bufp, pid, label)
}
traceReleaseBuffer(mp, pid)
unlock(&trace.bufLock)
unlock(&sched.sysmonlock)
// Record the current state of HeapGoal to avoid information loss in trace.
//
// Use the same dummy trace locker. The trace can't end until after we start
// the world, and we can safely trace from here.
tl.HeapGoal()
startTheWorldGC(stw)
return nil
}
// StopTrace stops tracing, if it was previously enabled.
// StopTrace only returns after all the reads for the trace have completed.
func StopTrace() {
// Stop the world so that we can collect the trace buffers from all p's below,
// and also to avoid races with traceEvent.
stw := stopTheWorldGC(stwStopTrace)
// See the comment in StartTrace.
lock(&sched.sysmonlock)
// See the comment in StartTrace.
lock(&trace.bufLock)
if !trace.enabled {
unlock(&trace.bufLock)
unlock(&sched.sysmonlock)
startTheWorldGC(stw)
return
}
// Trace GoSched for us, and use a dummy locker. The world is stopped
// and we control whether the trace is enabled, so this is safe.
tl := traceLocker{}
tl.GoSched()
atomicstorep(unsafe.Pointer(&trace.cpuLogWrite), nil)
trace.cpuLogRead.close()
traceReadCPU()
// Loop over all allocated Ps because dead Ps may still have
// trace buffers.
for _, p := range allp[:cap(allp)] {
buf := p.trace.buf
if buf != 0 {
traceFullQueue(buf)
p.trace.buf = 0
}
}
if trace.buf != 0 {
buf := trace.buf
trace.buf = 0
if buf.ptr().pos != 0 {
traceFullQueue(buf)
}
}
if trace.cpuLogBuf != 0 {
buf := trace.cpuLogBuf
trace.cpuLogBuf = 0
if buf.ptr().pos != 0 {
traceFullQueue(buf)
}
}
// Wait for startNanotime != endNanotime. On Windows the default interval between
// system clock ticks is typically between 1 and 15 milliseconds, which may not
// have passed since the trace started. Without nanotime moving forward, trace
// tooling has no way of identifying how much real time each cputicks time deltas
// represent.
for {
trace.endTime = traceClockNow()
trace.endTicks = cputicks()
trace.endNanotime = nanotime()
if trace.endNanotime != trace.startNanotime || faketime != 0 {
break
}
osyield()
}
trace.enabled = false
trace.shutdown = true
unlock(&trace.bufLock)
unlock(&sched.sysmonlock)
startTheWorldGC(stw)
// The world is started but we've set trace.shutdown, so new tracing can't start.
// Wait for the trace reader to flush pending buffers and stop.
semacquire(&trace.shutdownSema)
if raceenabled {
raceacquire(unsafe.Pointer(&trace.shutdownSema))
}
systemstack(func() {
// The lock protects us from races with StartTrace/StopTrace because they do stop-the-world.
lock(&trace.lock)
for _, p := range allp[:cap(allp)] {
if p.trace.buf != 0 {
throw("trace: non-empty trace buffer in proc")
}
}
if trace.buf != 0 {
throw("trace: non-empty global trace buffer")
}
if trace.fullHead != 0 || trace.fullTail != 0 {
throw("trace: non-empty full trace buffer")
}
if trace.reading != 0 || trace.reader.Load() != nil {
throw("trace: reading after shutdown")
}
for trace.empty != 0 {
buf := trace.empty
trace.empty = buf.ptr().link
sysFree(unsafe.Pointer(buf), unsafe.Sizeof(*buf.ptr()), &memstats.other_sys)
}
trace.strings = nil
trace.shutdown = false
trace.cpuLogRead = nil
unlock(&trace.lock)
})
}
// ReadTrace returns the next chunk of binary tracing data, blocking until data
// is available. If tracing is turned off and all the data accumulated while it
// was on has been returned, ReadTrace returns nil. The caller must copy the
// returned data before calling ReadTrace again.
// ReadTrace must be called from one goroutine at a time.
func ReadTrace() []byte {
top:
var buf []byte
var park bool
systemstack(func() {
buf, park = readTrace0()
})
if park {
gopark(func(gp *g, _ unsafe.Pointer) bool {
if !trace.reader.CompareAndSwapNoWB(nil, gp) {
// We're racing with another reader.
// Wake up and handle this case.
return false
}
if g2 := traceReader(); gp == g2 {
// New data arrived between unlocking
// and the CAS and we won the wake-up
// race, so wake up directly.
return false
} else if g2 != nil {
printlock()
println("runtime: got trace reader", g2, g2.goid)
throw("unexpected trace reader")
}
return true
}, nil, waitReasonTraceReaderBlocked, traceBlockSystemGoroutine, 2)
goto top
}
return buf
}
// readTrace0 is ReadTrace's continuation on g0. This must run on the
// system stack because it acquires trace.lock.
//
//go:systemstack
func readTrace0() (buf []byte, park bool) {
if raceenabled {
// g0 doesn't have a race context. Borrow the user G's.
if getg().racectx != 0 {
throw("expected racectx == 0")
}
getg().racectx = getg().m.curg.racectx
// (This defer should get open-coded, which is safe on
// the system stack.)
defer func() { getg().racectx = 0 }()
}
// Optimistically look for CPU profile samples. This may write new stack
// records, and may write new tracing buffers. This must be done with the
// trace lock not held. footerWritten and shutdown are safe to access
// here. They are only mutated by this goroutine or during a STW.
if !trace.footerWritten && !trace.shutdown {
traceReadCPU()
}
// This function must not allocate while holding trace.lock:
// allocation can call heap allocate, which will try to emit a trace
// event while holding heap lock.
lock(&trace.lock)
if trace.reader.Load() != nil {
// More than one goroutine reads trace. This is bad.
// But we rather do not crash the program because of tracing,
// because tracing can be enabled at runtime on prod servers.
unlock(&trace.lock)
println("runtime: ReadTrace called from multiple goroutines simultaneously")
return nil, false
}
// Recycle the old buffer.
if buf := trace.reading; buf != 0 {
buf.ptr().link = trace.empty
trace.empty = buf
trace.reading = 0
}
// Write trace header.
if !trace.headerWritten {
trace.headerWritten = true
unlock(&trace.lock)
return []byte("go 1.21 trace\x00\x00\x00"), false
}
// Wait for new data.
if trace.fullHead == 0 && !trace.shutdown {
// We don't simply use a note because the scheduler
// executes this goroutine directly when it wakes up
// (also a note would consume an M).
unlock(&trace.lock)
return nil, true
}
newFull:
assertLockHeld(&trace.lock)
// Write a buffer.
if trace.fullHead != 0 {
buf := traceFullDequeue()
trace.reading = buf
unlock(&trace.lock)
return buf.ptr().arr[:buf.ptr().pos], false
}
// Write footer with timer frequency.
if !trace.footerWritten {
trace.footerWritten = true
freq := (float64(trace.endTicks-trace.startTicks) / traceTimeDiv) / (float64(trace.endNanotime-trace.startNanotime) / 1e9)
if freq <= 0 {
throw("trace: ReadTrace got invalid frequency")
}
unlock(&trace.lock)
// Write frequency event.
bufp := traceFlush(0, 0)
buf := bufp.ptr()
buf.byte(traceEvFrequency | 0<<traceArgCountShift)
buf.varint(uint64(freq))
// Dump stack table.
// This will emit a bunch of full buffers, we will pick them up
// on the next iteration.
bufp = trace.stackTab.dump(bufp)
// Flush final buffer.
lock(&trace.lock)
traceFullQueue(bufp)
goto newFull // trace.lock should be held at newFull
}
// Done.
if trace.shutdown {
unlock(&trace.lock)
if raceenabled {
// Model synchronization on trace.shutdownSema, which race
// detector does not see. This is required to avoid false
// race reports on writer passed to trace.Start.
racerelease(unsafe.Pointer(&trace.shutdownSema))
}
// trace.enabled is already reset, so can call traceable functions.
semrelease(&trace.shutdownSema)
return nil, false
}
// Also bad, but see the comment above.
unlock(&trace.lock)
println("runtime: spurious wakeup of trace reader")
return nil, false
}
// traceReader returns the trace reader that should be woken up, if any.
// Callers should first check that trace.enabled or trace.shutdown is set.
//
// This must run on the system stack because it acquires trace.lock.
//
//go:systemstack
func traceReader() *g {
// Optimistic check first
if traceReaderAvailable() == nil {
return nil
}
lock(&trace.lock)
gp := traceReaderAvailable()
if gp == nil || !trace.reader.CompareAndSwapNoWB(gp, nil) {
unlock(&trace.lock)
return nil
}
unlock(&trace.lock)
return gp
}
// traceReaderAvailable returns the trace reader if it is not currently
// scheduled and should be. Callers should first check that trace.enabled
// or trace.shutdown is set.
func traceReaderAvailable() *g {
if trace.fullHead != 0 || trace.shutdown {
return trace.reader.Load()
}
return nil
}
// traceProcFree frees trace buffer associated with pp.
//
// This must run on the system stack because it acquires trace.lock.
//
//go:systemstack
func traceProcFree(pp *p) {
buf := pp.trace.buf
pp.trace.buf = 0
if buf == 0 {
return
}
lock(&trace.lock)
traceFullQueue(buf)
unlock(&trace.lock)
}
// ThreadDestroy is a no-op. It exists as a stub to support the new tracer.
//
// This must run on the system stack, just to match the new tracer.
func traceThreadDestroy(_ *m) {
// No-op in old tracer.
}
// traceFullQueue queues buf into queue of full buffers.
func traceFullQueue(buf traceBufPtr) {
buf.ptr().link = 0
if trace.fullHead == 0 {
trace.fullHead = buf
} else {
trace.fullTail.ptr().link = buf
}
trace.fullTail = buf
}
// traceFullDequeue dequeues from queue of full buffers.
func traceFullDequeue() traceBufPtr {
buf := trace.fullHead
if buf == 0 {
return 0
}
trace.fullHead = buf.ptr().link
if trace.fullHead == 0 {
trace.fullTail = 0
}
buf.ptr().link = 0
return buf
}
// traceEvent writes a single event to trace buffer, flushing the buffer if necessary.
// ev is event type.
// If skip > 0, write current stack id as the last argument (skipping skip top frames).
// If skip = 0, this event type should contain a stack, but we don't want
// to collect and remember it for this particular call.
func traceEvent(ev byte, skip int, args ...uint64) {
mp, pid, bufp := traceAcquireBuffer()
// Double-check trace.enabled now that we've done m.locks++ and acquired bufLock.
// This protects from races between traceEvent and StartTrace/StopTrace.
// The caller checked that trace.enabled == true, but trace.enabled might have been
// turned off between the check and now. Check again. traceLockBuffer did mp.locks++,
// StopTrace does stopTheWorld, and stopTheWorld waits for mp.locks to go back to zero,
// so if we see trace.enabled == true now, we know it's true for the rest of the function.
// Exitsyscall can run even during stopTheWorld. The race with StartTrace/StopTrace
// during tracing in exitsyscall is resolved by locking trace.bufLock in traceLockBuffer.
//
// Note trace_userTaskCreate runs the same check.
if !trace.enabled && !mp.trace.startingTrace {
traceReleaseBuffer(mp, pid)
return
}
if skip > 0 {
if getg() == mp.curg {
skip++ // +1 because stack is captured in traceEventLocked.
}
}
traceEventLocked(0, mp, pid, bufp, ev, 0, skip, args...)
traceReleaseBuffer(mp, pid)
}
// traceEventLocked writes a single event of type ev to the trace buffer bufp,
// flushing the buffer if necessary. pid is the id of the current P, or
// traceGlobProc if we're tracing without a real P.
//
// Preemption is disabled, and if running without a real P the global tracing
// buffer is locked.
//
// Events types that do not include a stack set skip to -1. Event types that
// include a stack may explicitly reference a stackID from the trace.stackTab
// (obtained by an earlier call to traceStackID). Without an explicit stackID,
// this function will automatically capture the stack of the goroutine currently
// running on mp, skipping skip top frames or, if skip is 0, writing out an
// empty stack record.
//
// It records the event's args to the traceBuf, and also makes an effort to
// reserve extraBytes bytes of additional space immediately following the event,
// in the same traceBuf.
func traceEventLocked(extraBytes int, mp *m, pid int32, bufp *traceBufPtr, ev byte, stackID uint32, skip int, args ...uint64) {
buf := bufp.ptr()
// TODO: test on non-zero extraBytes param.
maxSize := 2 + 5*traceBytesPerNumber + extraBytes // event type, length, sequence, timestamp, stack id and two add params
if buf == nil || len(buf.arr)-buf.pos < maxSize {
systemstack(func() {
buf = traceFlush(traceBufPtrOf(buf), pid).ptr()
})
bufp.set(buf)
}
ts := traceClockNow()
if ts <= buf.lastTime {
ts = buf.lastTime + 1
}
tsDiff := uint64(ts - buf.lastTime)
buf.lastTime = ts
narg := byte(len(args))
if stackID != 0 || skip >= 0 {
narg++
}
// We have only 2 bits for number of arguments.
// If number is >= 3, then the event type is followed by event length in bytes.
if narg > 3 {
narg = 3
}
startPos := buf.pos
buf.byte(ev | narg<<traceArgCountShift)
var lenp *byte
if narg == 3 {
// Reserve the byte for length assuming that length < 128.
buf.varint(0)
lenp = &buf.arr[buf.pos-1]
}
buf.varint(tsDiff)
for _, a := range args {
buf.varint(a)
}
if stackID != 0 {
buf.varint(uint64(stackID))
} else if skip == 0 {
buf.varint(0)
} else if skip > 0 {
buf.varint(traceStackID(mp, buf.stk[:], skip))
}
evSize := buf.pos - startPos
if evSize > maxSize {
throw("invalid length of trace event")
}
if lenp != nil {
// Fill in actual length.
*lenp = byte(evSize - 2)
}
}
// traceCPUSample writes a CPU profile sample stack to the execution tracer's
// profiling buffer. It is called from a signal handler, so is limited in what
// it can do.
func traceCPUSample(gp *g, _ *m, pp *p, stk []uintptr) {
if !traceEnabled() {
// Tracing is usually turned off; don't spend time acquiring the signal
// lock unless it's active.
return
}
// Match the clock used in traceEventLocked
now := traceClockNow()
// The "header" here is the ID of the P that was running the profiled code,
// followed by the ID of the goroutine. (For normal CPU profiling, it's
// usually the number of samples with the given stack.) Near syscalls, pp
// may be nil. Reporting goid of 0 is fine for either g0 or a nil gp.
var hdr [2]uint64
if pp != nil {
// Overflow records in profBuf have all header values set to zero. Make
// sure that real headers have at least one bit set.
hdr[0] = uint64(pp.id)<<1 | 0b1
} else {
hdr[0] = 0b10
}
if gp != nil {
hdr[1] = gp.goid
}
// Allow only one writer at a time
for !trace.signalLock.CompareAndSwap(0, 1) {
// TODO: Is it safe to osyield here? https://go.dev/issue/52672
osyield()
}
if log := (*profBuf)(atomic.Loadp(unsafe.Pointer(&trace.cpuLogWrite))); log != nil {
// Note: we don't pass a tag pointer here (how should profiling tags
// interact with the execution tracer?), but if we did we'd need to be
// careful about write barriers. See the long comment in profBuf.write.
log.write(nil, int64(now), hdr[:], stk)
}
trace.signalLock.Store(0)
}
func traceReadCPU() {
bufp := &trace.cpuLogBuf
for {
data, tags, _ := trace.cpuLogRead.read(profBufNonBlocking)
if len(data) == 0 {
break
}
for len(data) > 0 {
if len(data) < 4 || data[0] > uint64(len(data)) {
break // truncated profile
}
if data[0] < 4 || tags != nil && len(tags) < 1 {
break // malformed profile
}
if len(tags) < 1 {
break // mismatched profile records and tags
}
timestamp := data[1]
ppid := data[2] >> 1
if hasP := (data[2] & 0b1) != 0; !hasP {
ppid = ^uint64(0)
}
goid := data[3]
stk := data[4:data[0]]
empty := len(stk) == 1 && data[2] == 0 && data[3] == 0
data = data[data[0]:]
// No support here for reporting goroutine tags at the moment; if
// that information is to be part of the execution trace, we'd
// probably want to see when the tags are applied and when they
// change, instead of only seeing them when we get a CPU sample.
tags = tags[1:]
if empty {
// Looks like an overflow record from the profBuf. Not much to
// do here, we only want to report full records.
//
// TODO: should we start a goroutine to drain the profBuf,
// rather than relying on a high-enough volume of tracing events
// to keep ReadTrace busy? https://go.dev/issue/52674
continue
}