1 // Copyright 2009 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
6 // Patterned after tcmalloc's algorithms; shorter code.
11 "runtime/internal/atomic"
15 // NOTE(rsc): Everything here could use cas if contention became an issue.
18 // All memory allocations are local and do not escape outside of the profiler.
19 // The profiler is forbidden from referring to garbage-collected memory.
23 memProfile bucketType = 1 + iota
27 // size of bucket hash table
30 // max depth of stack to record in bucket
36 // A bucket holds per-call-stack profiling information.
37 // The representation is a bit sleazy, inherited from C.
38 // This struct defines the bucket header. It is followed in
39 // memory by the stack words and then the actual record
40 // data, either a memRecord or a blockRecord.
42 // Per-call-stack profiling information.
43 // Lookup by hashing call stack into a linked-list hash table.
51 typ bucketType // memBucket or blockBucket (includes mutexProfile)
57 // A memRecord is the bucket data for a bucket of type memProfile,
58 // part of the memory profile.
59 type memRecord struct {
60 // The following complex 3-stage scheme of stats accumulation
61 // is required to obtain a consistent picture of mallocs and frees
62 // for some point in time.
63 // The problem is that mallocs come in real time, while frees
64 // come only after a GC during concurrent sweeping. So if we would
65 // naively count them, we would get a skew toward mallocs.
67 // Hence, we delay information to get consistent snapshots as
68 // of mark termination. Allocations count toward the next mark
69 // termination's snapshot, while sweep frees count toward the
70 // previous mark termination's snapshot:
74 // .·˙ | .·˙ | .·˙ | .·˙ |
75 // .·˙ | .·˙ | .·˙ | .·˙ |
76 // .·˙ |.·˙ |.·˙ |.·˙ |
86 // Since we can't publish a consistent snapshot until all of
87 // the sweep frees are accounted for, we wait until the next
88 // mark termination ("MT" above) to publish the previous mark
89 // termination's snapshot ("P" above). To do this, allocation
90 // and free events are accounted to *future* heap profile
91 // cycles ("C+n" above) and we only publish a cycle once all
92 // of the events from that cycle must be done. Specifically:
94 // Mallocs are accounted to cycle C+2.
95 // Explicit frees are accounted to cycle C+2.
96 // GC frees (done during sweeping) are accounted to cycle C+1.
98 // After mark termination, we increment the global heap
99 // profile cycle counter and accumulate the stats from cycle C
100 // into the active profile.
102 // active is the currently published profile. A profiling
103 // cycle can be accumulated into active once its complete.
104 active memRecordCycle
106 // future records the profile events we're counting for cycles
107 // that have not yet been published. This is ring buffer
108 // indexed by the global heap profile cycle C and stores
109 // cycles C, C+1, and C+2. Unlike active, these counts are
110 // only for a single cycle; they are not cumulative across
113 // We store cycle C here because there's a window between when
114 // C becomes the active cycle and when we've flushed it to
116 future [3]memRecordCycle
120 type memRecordCycle struct {
121 allocs, frees uintptr
122 alloc_bytes, free_bytes uintptr
125 // add accumulates b into a. It does not zero b.
126 func (a *memRecordCycle) add(b *memRecordCycle) {
129 a.alloc_bytes += b.alloc_bytes
130 a.free_bytes += b.free_bytes
133 // A blockRecord is the bucket data for a bucket of type blockProfile,
134 // which is used in blocking and mutex profiles.
135 type blockRecord struct {
141 mbuckets *bucket // memory profile buckets
142 bbuckets *bucket // blocking profile buckets
143 xbuckets *bucket // mutex profile buckets
144 buckhash *[179999]*bucket
148 // All fields in mProf are protected by proflock.
150 // cycle is the global heap profile cycle. This wraps
151 // at mProfCycleWrap.
153 // flushed indicates that future[cycle] in all buckets
154 // has been flushed to the active profile.
159 const mProfCycleWrap = uint32(len(memRecord{}.future)) * (2 << 24)
161 // newBucket allocates a bucket with the given type and number of stack entries.
162 func newBucket(typ bucketType, nstk int) *bucket {
163 size := unsafe.Sizeof(bucket{}) + uintptr(nstk)*unsafe.Sizeof(uintptr(0))
166 throw("invalid profile bucket type")
168 size += unsafe.Sizeof(memRecord{})
169 case blockProfile, mutexProfile:
170 size += unsafe.Sizeof(blockRecord{})
173 b := (*bucket)(persistentalloc(size, 0, &memstats.buckhash_sys))
176 b.nstk = uintptr(nstk)
180 // stk returns the slice in b holding the stack.
181 func (b *bucket) stk() []uintptr {
182 stk := (*[maxStack]uintptr)(add(unsafe.Pointer(b), unsafe.Sizeof(*b)))
183 return stk[:b.nstk:b.nstk]
186 // mp returns the memRecord associated with the memProfile bucket b.
187 func (b *bucket) mp() *memRecord {
188 if b.typ != memProfile {
189 throw("bad use of bucket.mp")
191 data := add(unsafe.Pointer(b), unsafe.Sizeof(*b)+b.nstk*unsafe.Sizeof(uintptr(0)))
192 return (*memRecord)(data)
195 // bp returns the blockRecord associated with the blockProfile bucket b.
196 func (b *bucket) bp() *blockRecord {
197 if b.typ != blockProfile && b.typ != mutexProfile {
198 throw("bad use of bucket.bp")
200 data := add(unsafe.Pointer(b), unsafe.Sizeof(*b)+b.nstk*unsafe.Sizeof(uintptr(0)))
201 return (*blockRecord)(data)
204 // Return the bucket for stk[0:nstk], allocating new bucket if needed.
205 func stkbucket(typ bucketType, size uintptr, stk []uintptr, alloc bool) *bucket {
207 buckhash = (*[buckHashSize]*bucket)(sysAlloc(unsafe.Sizeof(*buckhash), &memstats.buckhash_sys))
209 throw("runtime: cannot allocate memory")
215 for _, pc := range stk {
228 i := int(h % buckHashSize)
229 for b := buckhash[i]; b != nil; b = b.next {
230 if b.typ == typ && b.hash == h && b.size == size && eqslice(b.stk(), stk) {
239 // Create new bucket.
240 b := newBucket(typ, len(stk))
246 if typ == memProfile {
249 } else if typ == mutexProfile {
259 func eqslice(x, y []uintptr) bool {
260 if len(x) != len(y) {
263 for i, xi := range x {
271 // mProf_NextCycle publishes the next heap profile cycle and creates a
272 // fresh heap profile cycle. This operation is fast and can be done
273 // during STW. The caller must call mProf_Flush before calling
274 // mProf_NextCycle again.
276 // This is called by mark termination during STW so allocations and
277 // frees after the world is started again count towards a new heap
279 func mProf_NextCycle() {
281 // We explicitly wrap mProf.cycle rather than depending on
282 // uint wraparound because the memRecord.future ring does not
283 // itself wrap at a power of two.
284 mProf.cycle = (mProf.cycle + 1) % mProfCycleWrap
285 mProf.flushed = false
289 // mProf_Flush flushes the events from the current heap profiling
290 // cycle into the active profile. After this it is safe to start a new
291 // heap profiling cycle with mProf_NextCycle.
293 // This is called by GC after mark termination starts the world. In
294 // contrast with mProf_NextCycle, this is somewhat expensive, but safe
295 // to do concurrently.
305 func mProf_FlushLocked() {
307 for b := mbuckets; b != nil; b = b.allnext {
310 // Flush cycle C into the published profile and clear
312 mpc := &mp.future[c%uint32(len(mp.future))]
314 *mpc = memRecordCycle{}
318 // mProf_PostSweep records that all sweep frees for this GC cycle have
319 // completed. This has the effect of publishing the heap profile
320 // snapshot as of the last mark termination without advancing the heap
322 func mProf_PostSweep() {
324 // Flush cycle C+1 to the active profile so everything as of
325 // the last mark termination becomes visible. *Don't* advance
326 // the cycle, since we're still accumulating allocs in cycle
327 // C+2, which have to become C+1 in the next mark termination
330 for b := mbuckets; b != nil; b = b.allnext {
332 mpc := &mp.future[(c+1)%uint32(len(mp.future))]
334 *mpc = memRecordCycle{}
339 // Called by malloc to record a profiled block.
340 func mProf_Malloc(p unsafe.Pointer, size uintptr) {
341 var stk [maxStack]uintptr
342 nstk := callers(4, stk[:])
344 b := stkbucket(memProfile, size, stk[:nstk], true)
347 mpc := &mp.future[(c+2)%uint32(len(mp.future))]
349 mpc.alloc_bytes += size
352 // Setprofilebucket locks a bunch of other mutexes, so we call it outside of proflock.
353 // This reduces potential contention and chances of deadlocks.
354 // Since the object must be alive during call to mProf_Malloc,
355 // it's fine to do this non-atomically.
357 setprofilebucket(p, b)
361 // Called when freeing a profiled block.
362 func mProf_Free(b *bucket, size uintptr) {
366 mpc := &mp.future[(c+1)%uint32(len(mp.future))]
368 mpc.free_bytes += size
372 var blockprofilerate uint64 // in CPU ticks
374 // SetBlockProfileRate controls the fraction of goroutine blocking events
375 // that are reported in the blocking profile. The profiler aims to sample
376 // an average of one blocking event per rate nanoseconds spent blocked.
378 // To include every blocking event in the profile, pass rate = 1.
379 // To turn off profiling entirely, pass rate <= 0.
380 func SetBlockProfileRate(rate int) {
383 r = 0 // disable profiling
384 } else if rate == 1 {
385 r = 1 // profile everything
387 // convert ns to cycles, use float64 to prevent overflow during multiplication
388 r = int64(float64(rate) * float64(tickspersecond()) / (1000 * 1000 * 1000))
394 atomic.Store64(&blockprofilerate, uint64(r))
397 func blockevent(cycles int64, skip int) {
401 if blocksampled(cycles) {
402 saveblockevent(cycles, skip+1, blockProfile)
406 func blocksampled(cycles int64) bool {
407 rate := int64(atomic.Load64(&blockprofilerate))
408 if rate <= 0 || (rate > cycles && int64(fastrand())%rate > cycles) {
414 func saveblockevent(cycles int64, skip int, which bucketType) {
417 var stk [maxStack]uintptr
418 if gp.m.curg == nil || gp.m.curg == gp {
419 nstk = callers(skip, stk[:])
421 nstk = gcallers(gp.m.curg, skip, stk[:])
424 b := stkbucket(which, 0, stk[:nstk], true)
426 b.bp().cycles += cycles
430 var mutexprofilerate uint64 // fraction sampled
432 // SetMutexProfileFraction controls the fraction of mutex contention events
433 // that are reported in the mutex profile. On average 1/rate events are
434 // reported. The previous rate is returned.
436 // To turn off profiling entirely, pass rate 0.
437 // To just read the current rate, pass rate < 0.
438 // (For n>1 the details of sampling may change.)
439 func SetMutexProfileFraction(rate int) int {
441 return int(mutexprofilerate)
443 old := mutexprofilerate
444 atomic.Store64(&mutexprofilerate, uint64(rate))
448 //go:linkname mutexevent sync.event
449 func mutexevent(cycles int64, skip int) {
453 rate := int64(atomic.Load64(&mutexprofilerate))
454 // TODO(pjw): measure impact of always calling fastrand vs using something
455 // like malloc.go:nextSample()
456 if rate > 0 && int64(fastrand())%rate == 0 {
457 saveblockevent(cycles, skip+1, mutexProfile)
461 // Go interface to profile data.
463 // A StackRecord describes a single execution stack.
464 type StackRecord struct {
465 Stack0 [32]uintptr // stack trace for this record; ends at first 0 entry
468 // Stack returns the stack trace associated with the record,
469 // a prefix of r.Stack0.
470 func (r *StackRecord) Stack() []uintptr {
471 for i, v := range r.Stack0 {
479 // MemProfileRate controls the fraction of memory allocations
480 // that are recorded and reported in the memory profile.
481 // The profiler aims to sample an average of
482 // one allocation per MemProfileRate bytes allocated.
484 // To include every allocated block in the profile, set MemProfileRate to 1.
485 // To turn off profiling entirely, set MemProfileRate to 0.
487 // The tools that process the memory profiles assume that the
488 // profile rate is constant across the lifetime of the program
489 // and equal to the current value. Programs that change the
490 // memory profiling rate should do so just once, as early as
491 // possible in the execution of the program (for example,
492 // at the beginning of main).
493 var MemProfileRate int = 512 * 1024
495 // A MemProfileRecord describes the live objects allocated
496 // by a particular call sequence (stack trace).
497 type MemProfileRecord struct {
498 AllocBytes, FreeBytes int64 // number of bytes allocated, freed
499 AllocObjects, FreeObjects int64 // number of objects allocated, freed
500 Stack0 [32]uintptr // stack trace for this record; ends at first 0 entry
503 // InUseBytes returns the number of bytes in use (AllocBytes - FreeBytes).
504 func (r *MemProfileRecord) InUseBytes() int64 { return r.AllocBytes - r.FreeBytes }
506 // InUseObjects returns the number of objects in use (AllocObjects - FreeObjects).
507 func (r *MemProfileRecord) InUseObjects() int64 {
508 return r.AllocObjects - r.FreeObjects
511 // Stack returns the stack trace associated with the record,
512 // a prefix of r.Stack0.
513 func (r *MemProfileRecord) Stack() []uintptr {
514 for i, v := range r.Stack0 {
522 // MemProfile returns a profile of memory allocated and freed per allocation
525 // MemProfile returns n, the number of records in the current memory profile.
526 // If len(p) >= n, MemProfile copies the profile into p and returns n, true.
527 // If len(p) < n, MemProfile does not change p and returns n, false.
529 // If inuseZero is true, the profile includes allocation records
530 // where r.AllocBytes > 0 but r.AllocBytes == r.FreeBytes.
531 // These are sites where memory was allocated, but it has all
532 // been released back to the runtime.
534 // The returned profile may be up to two garbage collection cycles old.
535 // This is to avoid skewing the profile toward allocations; because
536 // allocations happen in real time but frees are delayed until the garbage
537 // collector performs sweeping, the profile only accounts for allocations
538 // that have had a chance to be freed by the garbage collector.
540 // Most clients should use the runtime/pprof package or
541 // the testing package's -test.memprofile flag instead
542 // of calling MemProfile directly.
543 func MemProfile(p []MemProfileRecord, inuseZero bool) (n int, ok bool) {
545 // If we're between mProf_NextCycle and mProf_Flush, take care
546 // of flushing to the active profile so we only have to look
547 // at the active profile below.
550 for b := mbuckets; b != nil; b = b.allnext {
552 if inuseZero || mp.active.alloc_bytes != mp.active.free_bytes {
555 if mp.active.allocs != 0 || mp.active.frees != 0 {
560 // Absolutely no data, suggesting that a garbage collection
561 // has not yet happened. In order to allow profiling when
562 // garbage collection is disabled from the beginning of execution,
563 // accumulate all of the cycles, and recount buckets.
565 for b := mbuckets; b != nil; b = b.allnext {
567 for c := range mp.future {
568 mp.active.add(&mp.future[c])
569 mp.future[c] = memRecordCycle{}
571 if inuseZero || mp.active.alloc_bytes != mp.active.free_bytes {
579 for b := mbuckets; b != nil; b = b.allnext {
581 if inuseZero || mp.active.alloc_bytes != mp.active.free_bytes {
591 // Write b's data to r.
592 func record(r *MemProfileRecord, b *bucket) {
594 r.AllocBytes = int64(mp.active.alloc_bytes)
595 r.FreeBytes = int64(mp.active.free_bytes)
596 r.AllocObjects = int64(mp.active.allocs)
597 r.FreeObjects = int64(mp.active.frees)
599 racewriterangepc(unsafe.Pointer(&r.Stack0[0]), unsafe.Sizeof(r.Stack0), getcallerpc(), funcPC(MemProfile))
602 msanwrite(unsafe.Pointer(&r.Stack0[0]), unsafe.Sizeof(r.Stack0))
604 copy(r.Stack0[:], b.stk())
605 for i := int(b.nstk); i < len(r.Stack0); i++ {
610 func iterate_memprof(fn func(*bucket, uintptr, *uintptr, uintptr, uintptr, uintptr)) {
612 for b := mbuckets; b != nil; b = b.allnext {
614 fn(b, b.nstk, &b.stk()[0], b.size, mp.active.allocs, mp.active.frees)
619 // BlockProfileRecord describes blocking events originated
620 // at a particular call sequence (stack trace).
621 type BlockProfileRecord struct {
627 // BlockProfile returns n, the number of records in the current blocking profile.
628 // If len(p) >= n, BlockProfile copies the profile into p and returns n, true.
629 // If len(p) < n, BlockProfile does not change p and returns n, false.
631 // Most clients should use the runtime/pprof package or
632 // the testing package's -test.blockprofile flag instead
633 // of calling BlockProfile directly.
634 func BlockProfile(p []BlockProfileRecord) (n int, ok bool) {
636 for b := bbuckets; b != nil; b = b.allnext {
641 for b := bbuckets; b != nil; b = b.allnext {
647 racewriterangepc(unsafe.Pointer(&r.Stack0[0]), unsafe.Sizeof(r.Stack0), getcallerpc(), funcPC(BlockProfile))
650 msanwrite(unsafe.Pointer(&r.Stack0[0]), unsafe.Sizeof(r.Stack0))
652 i := copy(r.Stack0[:], b.stk())
653 for ; i < len(r.Stack0); i++ {
663 // MutexProfile returns n, the number of records in the current mutex profile.
664 // If len(p) >= n, MutexProfile copies the profile into p and returns n, true.
665 // Otherwise, MutexProfile does not change p, and returns n, false.
667 // Most clients should use the runtime/pprof package
668 // instead of calling MutexProfile directly.
669 func MutexProfile(p []BlockProfileRecord) (n int, ok bool) {
671 for b := xbuckets; b != nil; b = b.allnext {
676 for b := xbuckets; b != nil; b = b.allnext {
679 r.Count = int64(bp.count)
681 i := copy(r.Stack0[:], b.stk())
682 for ; i < len(r.Stack0); i++ {
692 // ThreadCreateProfile returns n, the number of records in the thread creation profile.
693 // If len(p) >= n, ThreadCreateProfile copies the profile into p and returns n, true.
694 // If len(p) < n, ThreadCreateProfile does not change p and returns n, false.
696 // Most clients should use the runtime/pprof package instead
697 // of calling ThreadCreateProfile directly.
698 func ThreadCreateProfile(p []StackRecord) (n int, ok bool) {
699 first := (*m)(atomic.Loadp(unsafe.Pointer(&allm)))
700 for mp := first; mp != nil; mp = mp.alllink {
706 for mp := first; mp != nil; mp = mp.alllink {
707 p[i].Stack0 = mp.createstack
714 //go:linkname runtime_goroutineProfileWithLabels runtime/pprof.runtime_goroutineProfileWithLabels
715 func runtime_goroutineProfileWithLabels(p []StackRecord, labels []unsafe.Pointer) (n int, ok bool) {
716 return goroutineProfileWithLabels(p, labels)
719 // labels may be nil. If labels is non-nil, it must have the same length as p.
720 func goroutineProfileWithLabels(p []StackRecord, labels []unsafe.Pointer) (n int, ok bool) {
721 if labels != nil && len(labels) != len(p) {
726 isOK := func(gp1 *g) bool {
727 // Checking isSystemGoroutine here makes GoroutineProfile
728 // consistent with both NumGoroutine and Stack.
729 return gp1 != gp && readgstatus(gp1) != _Gdead && !isSystemGoroutine(gp1, false)
732 stopTheWorld("profile")
735 for _, gp1 := range allgs {
745 // Save current goroutine.
749 saveg(pc, sp, gp, &r[0])
753 // If we have a place to put our goroutine labelmap, insert it there.
759 // Save other goroutines.
760 for _, gp1 := range allgs {
763 // Should be impossible, but better to return a
764 // truncated profile than to crash the entire process.
767 saveg(^uintptr(0), ^uintptr(0), gp1, &r[0])
781 // GoroutineProfile returns n, the number of records in the active goroutine stack profile.
782 // If len(p) >= n, GoroutineProfile copies the profile into p and returns n, true.
783 // If len(p) < n, GoroutineProfile does not change p and returns n, false.
785 // Most clients should use the runtime/pprof package instead
786 // of calling GoroutineProfile directly.
787 func GoroutineProfile(p []StackRecord) (n int, ok bool) {
789 return goroutineProfileWithLabels(p, nil)
792 func saveg(pc, sp uintptr, gp *g, r *StackRecord) {
793 n := gentraceback(pc, sp, 0, gp, 0, &r.Stack0[0], len(r.Stack0), nil, nil, 0)
794 if n < len(r.Stack0) {
799 // Stack formats a stack trace of the calling goroutine into buf
800 // and returns the number of bytes written to buf.
801 // If all is true, Stack formats stack traces of all other goroutines
802 // into buf after the trace for the current goroutine.
803 func Stack(buf []byte, all bool) int {
805 stopTheWorld("stack trace")
815 // Force traceback=1 to override GOTRACEBACK setting,
816 // so that Stack's results are consistent.
817 // GOTRACEBACK is only about crash dumps.
819 g0.writebuf = buf[0:0:len(buf)]
821 traceback(pc, sp, 0, gp)
837 // Tracing of alloc/free/gc.
841 func tracealloc(p unsafe.Pointer, size uintptr, typ *_type) {
846 print("tracealloc(", p, ", ", hex(size), ")\n")
848 print("tracealloc(", p, ", ", hex(size), ", ", typ.string(), ")\n")
850 if gp.m.curg == nil || gp == gp.m.curg {
855 traceback(pc, sp, 0, gp)
858 goroutineheader(gp.m.curg)
859 traceback(^uintptr(0), ^uintptr(0), 0, gp.m.curg)
866 func tracefree(p unsafe.Pointer, size uintptr) {
870 print("tracefree(", p, ", ", hex(size), ")\n")
875 traceback(pc, sp, 0, gp)
887 // running on m->g0 stack; show all non-g0 goroutines
889 print("end tracegc\n")