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.
5 // Garbage collector: marking and scanning
11 "internal/goexperiment"
12 "runtime/internal/atomic"
13 "runtime/internal/sys"
18 fixedRootFinalizers = iota
22 // rootBlockBytes is the number of bytes to scan per data or
24 rootBlockBytes = 256 << 10
26 // maxObletBytes is the maximum bytes of an object to scan at
27 // once. Larger objects will be split up into "oblets" of at
28 // most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
29 // scan preemption at ~100 µs.
31 // This must be > _MaxSmallSize so that the object base is the
33 maxObletBytes = 128 << 10
35 // drainCheckThreshold specifies how many units of work to do
36 // between self-preemption checks in gcDrain. Assuming a scan
37 // rate of 1 MB/ms, this is ~100 µs. Lower values have higher
38 // overhead in the scan loop (the scheduler check may perform
39 // a syscall, so its overhead is nontrivial). Higher values
40 // make the system less responsive to incoming work.
41 drainCheckThreshold = 100000
43 // pagesPerSpanRoot indicates how many pages to scan from a span root
44 // at a time. Used by special root marking.
46 // Higher values improve throughput by increasing locality, but
47 // increase the minimum latency of a marking operation.
49 // Must be a multiple of the pageInUse bitmap element size and
50 // must also evenly divide pagesPerArena.
51 pagesPerSpanRoot = 512
54 // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
55 // some miscellany) and initializes scanning-related state.
57 // The world must be stopped.
58 func gcMarkRootPrepare() {
61 // Compute how many data and BSS root blocks there are.
62 nBlocks := func(bytes uintptr) int {
63 return int(divRoundUp(bytes, rootBlockBytes))
70 for _, datap := range activeModules() {
71 nDataRoots := nBlocks(datap.edata - datap.data)
72 if nDataRoots > work.nDataRoots {
73 work.nDataRoots = nDataRoots
77 for _, datap := range activeModules() {
78 nBSSRoots := nBlocks(datap.ebss - datap.bss)
79 if nBSSRoots > work.nBSSRoots {
80 work.nBSSRoots = nBSSRoots
84 // Scan span roots for finalizer specials.
86 // We depend on addfinalizer to mark objects that get
87 // finalizers after root marking.
89 // We're going to scan the whole heap (that was available at the time the
90 // mark phase started, i.e. markArenas) for in-use spans which have specials.
92 // Break up the work into arenas, and further into chunks.
94 // Snapshot allArenas as markArenas. This snapshot is safe because allArenas
96 mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
97 work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
101 // Gs may be created after this point, but it's okay that we
102 // ignore them because they begin life without any roots, so
103 // there's nothing to scan, and any roots they create during
104 // the concurrent phase will be caught by the write barrier.
105 work.stackRoots = allGsSnapshot()
106 work.nStackRoots = len(work.stackRoots)
108 work.markrootNext = 0
109 work.markrootJobs = uint32(fixedRootCount + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
111 // Calculate base indexes of each root type
112 work.baseData = uint32(fixedRootCount)
113 work.baseBSS = work.baseData + uint32(work.nDataRoots)
114 work.baseSpans = work.baseBSS + uint32(work.nBSSRoots)
115 work.baseStacks = work.baseSpans + uint32(work.nSpanRoots)
116 work.baseEnd = work.baseStacks + uint32(work.nStackRoots)
119 // gcMarkRootCheck checks that all roots have been scanned. It is
120 // purely for debugging.
121 func gcMarkRootCheck() {
122 if work.markrootNext < work.markrootJobs {
123 print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
124 throw("left over markroot jobs")
127 // Check that stacks have been scanned.
129 // We only check the first nStackRoots Gs that we should have scanned.
130 // Since we don't care about newer Gs (see comment in
131 // gcMarkRootPrepare), no locking is required.
133 forEachGRace(func(gp *g) {
134 if i >= work.nStackRoots {
139 println("gp", gp, "goid", gp.goid,
140 "status", readgstatus(gp),
141 "gcscandone", gp.gcscandone)
142 throw("scan missed a g")
149 // ptrmask for an allocation containing a single pointer.
150 var oneptrmask = [...]uint8{1}
152 // markroot scans the i'th root.
154 // Preemption must be disabled (because this uses a gcWork).
156 // Returns the amount of GC work credit produced by the operation.
157 // If flushBgCredit is true, then that credit is also flushed
158 // to the background credit pool.
160 // nowritebarrier is only advisory here.
163 func markroot(gcw *gcWork, i uint32, flushBgCredit bool) int64 {
164 // Note: if you add a case here, please also update heapdump.go:dumproots.
166 var workCounter *atomic.Int64
168 case work.baseData <= i && i < work.baseBSS:
169 workCounter = &gcController.globalsScanWork
170 for _, datap := range activeModules() {
171 workDone += markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
174 case work.baseBSS <= i && i < work.baseSpans:
175 workCounter = &gcController.globalsScanWork
176 for _, datap := range activeModules() {
177 workDone += markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
180 case i == fixedRootFinalizers:
181 for fb := allfin; fb != nil; fb = fb.alllink {
182 cnt := uintptr(atomic.Load(&fb.cnt))
183 scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
186 case i == fixedRootFreeGStacks:
187 // Switch to the system stack so we can call
189 systemstack(markrootFreeGStacks)
191 case work.baseSpans <= i && i < work.baseStacks:
192 // mark mspan.specials
193 markrootSpans(gcw, int(i-work.baseSpans))
196 // the rest is scanning goroutine stacks
197 workCounter = &gcController.stackScanWork
198 if i < work.baseStacks || work.baseEnd <= i {
200 print("runtime: markroot index ", i, " not in stack roots range [", work.baseStacks, ", ", work.baseEnd, ")\n")
201 throw("markroot: bad index")
203 gp := work.stackRoots[i-work.baseStacks]
205 // remember when we've first observed the G blocked
206 // needed only to output in traceback
207 status := readgstatus(gp) // We are not in a scan state
208 if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
209 gp.waitsince = work.tstart
212 // scanstack must be done on the system stack in case
213 // we're trying to scan our own stack.
215 // If this is a self-scan, put the user G in
216 // _Gwaiting to prevent self-deadlock. It may
217 // already be in _Gwaiting if this is a mark
218 // worker or we're in mark termination.
219 userG := getg().m.curg
220 selfScan := gp == userG && readgstatus(userG) == _Grunning
222 casgstatus(userG, _Grunning, _Gwaiting)
223 userG.waitreason = waitReasonGarbageCollectionScan
226 // TODO: suspendG blocks (and spins) until gp
227 // stops, which may take a while for
228 // running goroutines. Consider doing this in
229 // two phases where the first is non-blocking:
230 // we scan the stacks we can and ask running
231 // goroutines to scan themselves; and the
233 stopped := suspendG(gp)
239 throw("g already scanned")
241 workDone += scanstack(gp, gcw)
246 casgstatus(userG, _Gwaiting, _Grunning)
250 if goexperiment.PacerRedesign {
251 if workCounter != nil && workDone != 0 {
252 workCounter.Add(workDone)
254 gcFlushBgCredit(workDone)
261 // markrootBlock scans the shard'th shard of the block of memory [b0,
262 // b0+n0), with the given pointer mask.
264 // Returns the amount of work done.
267 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) int64 {
268 if rootBlockBytes%(8*goarch.PtrSize) != 0 {
269 // This is necessary to pick byte offsets in ptrmask0.
270 throw("rootBlockBytes must be a multiple of 8*ptrSize")
273 // Note that if b0 is toward the end of the address space,
274 // then b0 + rootBlockBytes might wrap around.
275 // These tests are written to avoid any possible overflow.
276 off := uintptr(shard) * rootBlockBytes
281 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
282 n := uintptr(rootBlockBytes)
288 scanblock(b, n, ptrmask, gcw, nil)
292 // markrootFreeGStacks frees stacks of dead Gs.
294 // This does not free stacks of dead Gs cached on Ps, but having a few
295 // cached stacks around isn't a problem.
296 func markrootFreeGStacks() {
297 // Take list of dead Gs with stacks.
298 lock(&sched.gFree.lock)
299 list := sched.gFree.stack
300 sched.gFree.stack = gList{}
301 unlock(&sched.gFree.lock)
307 q := gQueue{list.head, list.head}
308 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
312 // Manipulate the queue directly since the Gs are
313 // already all linked the right way.
317 // Put Gs back on the free list.
318 lock(&sched.gFree.lock)
319 sched.gFree.noStack.pushAll(q)
320 unlock(&sched.gFree.lock)
323 // markrootSpans marks roots for one shard of markArenas.
326 func markrootSpans(gcw *gcWork, shard int) {
327 // Objects with finalizers have two GC-related invariants:
329 // 1) Everything reachable from the object must be marked.
330 // This ensures that when we pass the object to its finalizer,
331 // everything the finalizer can reach will be retained.
333 // 2) Finalizer specials (which are not in the garbage
334 // collected heap) are roots. In practice, this means the fn
335 // field must be scanned.
336 sg := mheap_.sweepgen
338 // Find the arena and page index into that arena for this shard.
339 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
340 ha := mheap_.arenas[ai.l1()][ai.l2()]
341 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
343 // Construct slice of bitmap which we'll iterate over.
344 specialsbits := ha.pageSpecials[arenaPage/8:]
345 specialsbits = specialsbits[:pagesPerSpanRoot/8]
346 for i := range specialsbits {
347 // Find set bits, which correspond to spans with specials.
348 specials := atomic.Load8(&specialsbits[i])
352 for j := uint(0); j < 8; j++ {
353 if specials&(1<<j) == 0 {
356 // Find the span for this bit.
358 // This value is guaranteed to be non-nil because having
359 // specials implies that the span is in-use, and since we're
360 // currently marking we can be sure that we don't have to worry
361 // about the span being freed and re-used.
362 s := ha.spans[arenaPage+uint(i)*8+j]
364 // The state must be mSpanInUse if the specials bit is set, so
365 // sanity check that.
366 if state := s.state.get(); state != mSpanInUse {
367 print("s.state = ", state, "\n")
368 throw("non in-use span found with specials bit set")
370 // Check that this span was swept (it may be cached or uncached).
371 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
372 // sweepgen was updated (+2) during non-checkmark GC pass
373 print("sweep ", s.sweepgen, " ", sg, "\n")
374 throw("gc: unswept span")
377 // Lock the specials to prevent a special from being
378 // removed from the list while we're traversing it.
380 for sp := s.specials; sp != nil; sp = sp.next {
381 if sp.kind != _KindSpecialFinalizer {
384 // don't mark finalized object, but scan it so we
385 // retain everything it points to.
386 spf := (*specialfinalizer)(unsafe.Pointer(sp))
387 // A finalizer can be set for an inner byte of an object, find object beginning.
388 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
390 // Mark everything that can be reached from
391 // the object (but *not* the object itself or
392 // we'll never collect it).
395 // The special itself is a root.
396 scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
398 unlock(&s.speciallock)
403 // gcAssistAlloc performs GC work to make gp's assist debt positive.
404 // gp must be the calling user gorountine.
406 // This must be called with preemption enabled.
407 func gcAssistAlloc(gp *g) {
408 // Don't assist in non-preemptible contexts. These are
409 // generally fragile and won't allow the assist to block.
410 if getg() == gp.m.g0 {
413 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
419 // Compute the amount of scan work we need to do to make the
420 // balance positive. When the required amount of work is low,
421 // we over-assist to build up credit for future allocations
422 // and amortize the cost of assisting.
423 assistWorkPerByte := gcController.assistWorkPerByte.Load()
424 assistBytesPerWork := gcController.assistBytesPerWork.Load()
425 debtBytes := -gp.gcAssistBytes
426 scanWork := int64(assistWorkPerByte * float64(debtBytes))
427 if scanWork < gcOverAssistWork {
428 scanWork = gcOverAssistWork
429 debtBytes = int64(assistBytesPerWork * float64(scanWork))
432 // Steal as much credit as we can from the background GC's
433 // scan credit. This is racy and may drop the background
434 // credit below 0 if two mutators steal at the same time. This
435 // will just cause steals to fail until credit is accumulated
436 // again, so in the long run it doesn't really matter, but we
437 // do have to handle the negative credit case.
438 bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
440 if bgScanCredit > 0 {
441 if bgScanCredit < scanWork {
442 stolen = bgScanCredit
443 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
446 gp.gcAssistBytes += debtBytes
448 atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
453 // We were able to steal all of the credit we
456 traceGCMarkAssistDone()
462 if trace.enabled && !traced {
464 traceGCMarkAssistStart()
467 // Perform assist work
469 gcAssistAlloc1(gp, scanWork)
470 // The user stack may have moved, so this can't touch
471 // anything on it until it returns from systemstack.
474 completed := gp.param != nil
480 if gp.gcAssistBytes < 0 {
481 // We were unable steal enough credit or perform
482 // enough work to pay off the assist debt. We need to
483 // do one of these before letting the mutator allocate
484 // more to prevent over-allocation.
486 // If this is because we were preempted, reschedule
487 // and try some more.
493 // Add this G to an assist queue and park. When the GC
494 // has more background credit, it will satisfy queued
495 // assists before flushing to the global credit pool.
497 // Note that this does *not* get woken up when more
498 // work is added to the work list. The theory is that
499 // there wasn't enough work to do anyway, so we might
500 // as well let background marking take care of the
501 // work that is available.
506 // At this point either background GC has satisfied
507 // this G's assist debt, or the GC cycle is over.
510 traceGCMarkAssistDone()
514 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
515 // stack. This is a separate function to make it easier to see that
516 // we're not capturing anything from the user stack, since the user
517 // stack may move while we're in this function.
519 // gcAssistAlloc1 indicates whether this assist completed the mark
520 // phase by setting gp.param to non-nil. This can't be communicated on
521 // the stack since it may move.
524 func gcAssistAlloc1(gp *g, scanWork int64) {
525 // Clear the flag indicating that this assist completed the
529 if atomic.Load(&gcBlackenEnabled) == 0 {
530 // The gcBlackenEnabled check in malloc races with the
531 // store that clears it but an atomic check in every malloc
532 // would be a performance hit.
533 // Instead we recheck it here on the non-preemptable system
534 // stack to determine if we should perform an assist.
536 // GC is done, so ignore any remaining debt.
540 // Track time spent in this assist. Since we're on the
541 // system stack, this is non-preemptible, so we can
542 // just measure start and end time.
543 startTime := nanotime()
545 decnwait := atomic.Xadd(&work.nwait, -1)
546 if decnwait == work.nproc {
547 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
548 throw("nwait > work.nprocs")
551 // gcDrainN requires the caller to be preemptible.
552 casgstatus(gp, _Grunning, _Gwaiting)
553 gp.waitreason = waitReasonGCAssistMarking
555 // drain own cached work first in the hopes that it
556 // will be more cache friendly.
557 gcw := &getg().m.p.ptr().gcw
558 workDone := gcDrainN(gcw, scanWork)
560 casgstatus(gp, _Gwaiting, _Grunning)
562 // Record that we did this much scan work.
564 // Back out the number of bytes of assist credit that
565 // this scan work counts for. The "1+" is a poor man's
566 // round-up, to ensure this adds credit even if
567 // assistBytesPerWork is very low.
568 assistBytesPerWork := gcController.assistBytesPerWork.Load()
569 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
571 // If this is the last worker and we ran out of work,
572 // signal a completion point.
573 incnwait := atomic.Xadd(&work.nwait, +1)
574 if incnwait > work.nproc {
575 println("runtime: work.nwait=", incnwait,
576 "work.nproc=", work.nproc)
577 throw("work.nwait > work.nproc")
580 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
581 // This has reached a background completion point. Set
582 // gp.param to a non-nil value to indicate this. It
583 // doesn't matter what we set it to (it just has to be
585 gp.param = unsafe.Pointer(gp)
587 duration := nanotime() - startTime
589 _p_.gcAssistTime += duration
590 if _p_.gcAssistTime > gcAssistTimeSlack {
591 atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
596 // gcWakeAllAssists wakes all currently blocked assists. This is used
597 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
598 // new assists from going to sleep after this point.
599 func gcWakeAllAssists() {
600 lock(&work.assistQueue.lock)
601 list := work.assistQueue.q.popList()
603 unlock(&work.assistQueue.lock)
606 // gcParkAssist puts the current goroutine on the assist queue and parks.
608 // gcParkAssist reports whether the assist is now satisfied. If it
609 // returns false, the caller must retry the assist.
610 func gcParkAssist() bool {
611 lock(&work.assistQueue.lock)
612 // If the GC cycle finished while we were getting the lock,
613 // exit the assist. The cycle can't finish while we hold the
615 if atomic.Load(&gcBlackenEnabled) == 0 {
616 unlock(&work.assistQueue.lock)
621 oldList := work.assistQueue.q
622 work.assistQueue.q.pushBack(gp)
624 // Recheck for background credit now that this G is in
625 // the queue, but can still back out. This avoids a
626 // race in case background marking has flushed more
627 // credit since we checked above.
628 if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
629 work.assistQueue.q = oldList
630 if oldList.tail != 0 {
631 oldList.tail.ptr().schedlink.set(nil)
633 unlock(&work.assistQueue.lock)
637 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
641 // gcFlushBgCredit flushes scanWork units of background scan work
642 // credit. This first satisfies blocked assists on the
643 // work.assistQueue and then flushes any remaining credit to
644 // gcController.bgScanCredit.
646 // Write barriers are disallowed because this is used by gcDrain after
647 // it has ensured that all work is drained and this must preserve that
650 //go:nowritebarrierrec
651 func gcFlushBgCredit(scanWork int64) {
652 if work.assistQueue.q.empty() {
653 // Fast path; there are no blocked assists. There's a
654 // small window here where an assist may add itself to
655 // the blocked queue and park. If that happens, we'll
656 // just get it on the next flush.
657 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
661 assistBytesPerWork := gcController.assistBytesPerWork.Load()
662 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
664 lock(&work.assistQueue.lock)
665 for !work.assistQueue.q.empty() && scanBytes > 0 {
666 gp := work.assistQueue.q.pop()
667 // Note that gp.gcAssistBytes is negative because gp
668 // is in debt. Think carefully about the signs below.
669 if scanBytes+gp.gcAssistBytes >= 0 {
670 // Satisfy this entire assist debt.
671 scanBytes += gp.gcAssistBytes
673 // It's important that we *not* put gp in
674 // runnext. Otherwise, it's possible for user
675 // code to exploit the GC worker's high
676 // scheduler priority to get itself always run
677 // before other goroutines and always in the
678 // fresh quantum started by GC.
681 // Partially satisfy this assist.
682 gp.gcAssistBytes += scanBytes
684 // As a heuristic, we move this assist to the
685 // back of the queue so that large assists
686 // can't clog up the assist queue and
687 // substantially delay small assists.
688 work.assistQueue.q.pushBack(gp)
694 // Convert from scan bytes back to work.
695 assistWorkPerByte := gcController.assistWorkPerByte.Load()
696 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
697 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
699 unlock(&work.assistQueue.lock)
702 // scanstack scans gp's stack, greying all pointers found on the stack.
704 // For goexperiment.PacerRedesign:
705 // Returns the amount of scan work performed, but doesn't update
706 // gcController.stackScanWork or flush any credit. Any background credit produced
707 // by this function should be flushed by its caller. scanstack itself can't
708 // safely flush because it may result in trying to wake up a goroutine that
709 // was just scanned, resulting in a self-deadlock.
711 // scanstack will also shrink the stack if it is safe to do so. If it
712 // is not, it schedules a stack shrink for the next synchronous safe
715 // scanstack is marked go:systemstack because it must not be preempted
716 // while using a workbuf.
720 func scanstack(gp *g, gcw *gcWork) int64 {
721 if readgstatus(gp)&_Gscan == 0 {
722 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
723 throw("scanstack - bad status")
726 switch readgstatus(gp) &^ _Gscan {
728 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
729 throw("mark - bad status")
733 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
734 throw("scanstack: goroutine not stopped")
735 case _Grunnable, _Gsyscall, _Gwaiting:
740 throw("can't scan our own stack")
743 // stackSize is the amount of work we'll be reporting.
745 // We report the total stack size, more than we scan,
746 // because this number needs to line up with gcControllerState's
747 // stackScan and scannableStackSize fields.
749 // See the documentation on those fields for more information.
750 stackSize := gp.stack.hi - gp.stack.lo
752 if isShrinkStackSafe(gp) {
753 // Shrink the stack if not much of it is being used.
756 // Otherwise, shrink the stack at the next sync safe point.
757 gp.preemptShrink = true
760 var state stackScanState
761 state.stack = gp.stack
764 println("stack trace goroutine", gp.goid)
767 if debugScanConservative && gp.asyncSafePoint {
768 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
771 // Scan the saved context register. This is effectively a live
772 // register that gets moved back and forth between the
773 // register and sched.ctxt without a write barrier.
774 if gp.sched.ctxt != nil {
775 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
778 // Scan the stack. Accumulate a list of stack objects.
779 scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
780 scanframeworker(frame, &state, gcw)
783 gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
785 // Find additional pointers that point into the stack from the heap.
786 // Currently this includes defers and panics. See also function copystack.
788 // Find and trace other pointers in defer records.
789 for d := gp._defer; d != nil; d = d.link {
791 // Scan the func value, which could be a stack allocated closure.
793 scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
796 // The link field of a stack-allocated defer record might point
797 // to a heap-allocated defer record. Keep that heap record live.
798 scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
800 // Retain defers records themselves.
801 // Defer records might not be reachable from the G through regular heap
802 // tracing because the defer linked list might weave between the stack and the heap.
804 scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
807 if gp._panic != nil {
808 // Panics are always stack allocated.
809 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
812 // Find and scan all reachable stack objects.
814 // The state's pointer queue prioritizes precise pointers over
815 // conservative pointers so that we'll prefer scanning stack
816 // objects precisely.
819 p, conservative := state.getPtr()
823 obj := state.findObject(p)
829 // We've already scanned this object.
832 obj.setRecord(nil) // Don't scan it again.
835 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
837 print(" (conservative)")
845 // This path is pretty unlikely, an object large enough
846 // to have a GC program allocated on the stack.
847 // We need some space to unpack the program into a straight
848 // bitmask, which we allocate/free here.
849 // TODO: it would be nice if there were a way to run a GC
850 // program without having to store all its bits. We'd have
851 // to change from a Lempel-Ziv style program to something else.
852 // Or we can forbid putting objects on stacks if they require
853 // a gc program (see issue 27447).
854 s = materializeGCProg(r.ptrdata(), gcdata)
855 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
858 b := state.stack.lo + uintptr(obj.off)
860 scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
862 scanblock(b, r.ptrdata(), gcdata, gcw, &state)
866 dematerializeGCProg(s)
870 // Deallocate object buffers.
871 // (Pointer buffers were all deallocated in the loop above.)
872 for state.head != nil {
876 for i := 0; i < x.nobj; i++ {
878 if obj.r == nil { // reachable
881 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
882 // Note: not necessarily really dead - only reachable-from-ptr dead.
886 putempty((*workbuf)(unsafe.Pointer(x)))
888 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
889 throw("remaining pointer buffers")
891 return int64(stackSize)
894 // Scan a stack frame: local variables and function arguments/results.
896 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
897 if _DebugGC > 1 && frame.continpc != 0 {
898 print("scanframe ", funcname(frame.fn), "\n")
901 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
902 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
903 if state.conservative || isAsyncPreempt || isDebugCall {
904 if debugScanConservative {
905 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
908 // Conservatively scan the frame. Unlike the precise
909 // case, this includes the outgoing argument space
910 // since we may have stopped while this function was
911 // setting up a call.
913 // TODO: We could narrow this down if the compiler
914 // produced a single map per function of stack slots
915 // and registers that ever contain a pointer.
917 size := frame.varp - frame.sp
919 scanConservative(frame.sp, size, nil, gcw, state)
923 // Scan arguments to this frame.
924 if frame.arglen != 0 {
925 // TODO: We could pass the entry argument map
926 // to narrow this down further.
927 scanConservative(frame.argp, frame.arglen, nil, gcw, state)
930 if isAsyncPreempt || isDebugCall {
931 // This function's frame contained the
932 // registers for the asynchronously stopped
933 // parent frame. Scan the parent
935 state.conservative = true
937 // We only wanted to scan those two frames
938 // conservatively. Clear the flag for future
940 state.conservative = false
945 locals, args, objs := getStackMap(frame, &state.cache, false)
947 // Scan local variables if stack frame has been allocated.
949 size := uintptr(locals.n) * goarch.PtrSize
950 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
955 scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
958 // Add all stack objects to the stack object list.
960 // varp is 0 for defers, where there are no locals.
961 // In that case, there can't be a pointer to its args, either.
962 // (And all args would be scanned above anyway.)
963 for i := range objs {
966 base := frame.varp // locals base pointer
968 base = frame.argp // arguments and return values base pointer
970 ptr := base + uintptr(off)
972 // object hasn't been allocated in the frame yet.
976 println("stkobj at", hex(ptr), "of size", obj.size)
978 state.addObject(ptr, obj)
983 type gcDrainFlags int
986 gcDrainUntilPreempt gcDrainFlags = 1 << iota
992 // gcDrain scans roots and objects in work buffers, blackening grey
993 // objects until it is unable to get more work. It may return before
994 // GC is done; it's the caller's responsibility to balance work from
997 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
1000 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
1003 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
1004 // pollFractionalWorkerExit() returns true. This implies
1007 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
1008 // credit to gcController.bgScanCredit every gcCreditSlack units of
1011 // gcDrain will always return if there is a pending STW.
1014 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
1015 if !writeBarrier.needed {
1016 throw("gcDrain phase incorrect")
1020 preemptible := flags&gcDrainUntilPreempt != 0
1021 flushBgCredit := flags&gcDrainFlushBgCredit != 0
1022 idle := flags&gcDrainIdle != 0
1024 initScanWork := gcw.heapScanWork
1026 // checkWork is the scan work before performing the next
1027 // self-preempt check.
1028 checkWork := int64(1<<63 - 1)
1029 var check func() bool
1030 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
1031 checkWork = initScanWork + drainCheckThreshold
1034 } else if flags&gcDrainFractional != 0 {
1035 check = pollFractionalWorkerExit
1039 // Drain root marking jobs.
1040 if work.markrootNext < work.markrootJobs {
1041 // Stop if we're preemptible or if someone wants to STW.
1042 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1043 job := atomic.Xadd(&work.markrootNext, +1) - 1
1044 if job >= work.markrootJobs {
1047 markroot(gcw, job, flushBgCredit)
1048 if check != nil && check() {
1054 // Drain heap marking jobs.
1055 // Stop if we're preemptible or if someone wants to STW.
1056 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1057 // Try to keep work available on the global queue. We used to
1058 // check if there were waiting workers, but it's better to
1059 // just keep work available than to make workers wait. In the
1060 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1066 b := gcw.tryGetFast()
1070 // Flush the write barrier
1071 // buffer; this may create
1078 // Unable to get work.
1083 // Flush background scan work credit to the global
1084 // account if we've accumulated enough locally so
1085 // mutator assists can draw on it.
1086 if gcw.heapScanWork >= gcCreditSlack {
1087 gcController.heapScanWork.Add(gcw.heapScanWork)
1089 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1092 checkWork -= gcw.heapScanWork
1093 gcw.heapScanWork = 0
1096 checkWork += drainCheckThreshold
1097 if check != nil && check() {
1105 // Flush remaining scan work credit.
1106 if gcw.heapScanWork > 0 {
1107 gcController.heapScanWork.Add(gcw.heapScanWork)
1109 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1111 gcw.heapScanWork = 0
1115 // gcDrainN blackens grey objects until it has performed roughly
1116 // scanWork units of scan work or the G is preempted. This is
1117 // best-effort, so it may perform less work if it fails to get a work
1118 // buffer. Otherwise, it will perform at least n units of work, but
1119 // may perform more because scanning is always done in whole object
1120 // increments. It returns the amount of scan work performed.
1122 // The caller goroutine must be in a preemptible state (e.g.,
1123 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1124 // consequence, this must be called on the system stack.
1128 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1129 if !writeBarrier.needed {
1130 throw("gcDrainN phase incorrect")
1133 // There may already be scan work on the gcw, which we don't
1134 // want to claim was done by this call.
1135 workFlushed := -gcw.heapScanWork
1138 for !gp.preempt && workFlushed+gcw.heapScanWork < scanWork {
1139 // See gcDrain comment.
1144 b := gcw.tryGetFast()
1148 // Flush the write barrier buffer;
1149 // this may create more work.
1156 // Try to do a root job.
1157 if work.markrootNext < work.markrootJobs {
1158 job := atomic.Xadd(&work.markrootNext, +1) - 1
1159 if job < work.markrootJobs {
1160 work := markroot(gcw, job, false)
1161 if goexperiment.PacerRedesign {
1167 // No heap or root jobs.
1173 // Flush background scan work credit.
1174 if gcw.heapScanWork >= gcCreditSlack {
1175 gcController.heapScanWork.Add(gcw.heapScanWork)
1176 workFlushed += gcw.heapScanWork
1177 gcw.heapScanWork = 0
1181 // Unlike gcDrain, there's no need to flush remaining work
1182 // here because this never flushes to bgScanCredit and
1183 // gcw.dispose will flush any remaining work to scanWork.
1185 return workFlushed + gcw.heapScanWork
1188 // scanblock scans b as scanobject would, but using an explicit
1189 // pointer bitmap instead of the heap bitmap.
1191 // This is used to scan non-heap roots, so it does not update
1192 // gcw.bytesMarked or gcw.heapScanWork.
1194 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1196 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1197 // Use local copies of original parameters, so that a stack trace
1198 // due to one of the throws below shows the original block
1203 for i := uintptr(0); i < n; {
1204 // Find bits for the next word.
1205 bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
1207 i += goarch.PtrSize * 8
1210 for j := 0; j < 8 && i < n; j++ {
1212 // Same work as in scanobject; see comments there.
1213 p := *(*uintptr)(unsafe.Pointer(b + i))
1215 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1216 greyobject(obj, b, i, span, gcw, objIndex)
1217 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1218 stk.putPtr(p, false)
1228 // scanobject scans the object starting at b, adding pointers to gcw.
1229 // b must point to the beginning of a heap object or an oblet.
1230 // scanobject consults the GC bitmap for the pointer mask and the
1231 // spans for the size of the object.
1234 func scanobject(b uintptr, gcw *gcWork) {
1235 // Prefetch object before we scan it.
1237 // This will overlap fetching the beginning of the object with initial
1238 // setup before we start scanning the object.
1241 // Find the bits for b and the size of the object at b.
1243 // b is either the beginning of an object, in which case this
1244 // is the size of the object to scan, or it points to an
1245 // oblet, in which case we compute the size to scan below.
1246 hbits := heapBitsForAddr(b)
1247 s := spanOfUnchecked(b)
1250 throw("scanobject n == 0")
1253 if n > maxObletBytes {
1254 // Large object. Break into oblets for better
1255 // parallelism and lower latency.
1257 // It's possible this is a noscan object (not
1258 // from greyobject, but from other code
1259 // paths), in which case we must *not* enqueue
1260 // oblets since their bitmaps will be
1262 if s.spanclass.noscan() {
1263 // Bypass the whole scan.
1264 gcw.bytesMarked += uint64(n)
1268 // Enqueue the other oblets to scan later.
1269 // Some oblets may be in b's scalar tail, but
1270 // these will be marked as "no more pointers",
1271 // so we'll drop out immediately when we go to
1273 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1274 if !gcw.putFast(oblet) {
1280 // Compute the size of the oblet. Since this object
1281 // must be a large object, s.base() is the beginning
1283 n = s.base() + s.elemsize - b
1284 if n > maxObletBytes {
1290 for i = 0; i < n; i, hbits = i+goarch.PtrSize, hbits.next() {
1291 // Load bits once. See CL 22712 and issue 16973 for discussion.
1292 bits := hbits.bits()
1293 if bits&bitScan == 0 {
1294 break // no more pointers in this object
1296 if bits&bitPointer == 0 {
1297 continue // not a pointer
1300 // Work here is duplicated in scanblock and above.
1301 // If you make changes here, make changes there too.
1302 obj := *(*uintptr)(unsafe.Pointer(b + i))
1304 // At this point we have extracted the next potential pointer.
1305 // Quickly filter out nil and pointers back to the current object.
1306 if obj != 0 && obj-b >= n {
1307 // Test if obj points into the Go heap and, if so,
1310 // Note that it's possible for findObject to
1311 // fail if obj points to a just-allocated heap
1312 // object because of a race with growing the
1313 // heap. In this case, we know the object was
1314 // just allocated and hence will be marked by
1315 // allocation itself.
1316 if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
1317 greyobject(obj, b, i, span, gcw, objIndex)
1321 gcw.bytesMarked += uint64(n)
1322 gcw.heapScanWork += int64(i)
1325 // scanConservative scans block [b, b+n) conservatively, treating any
1326 // pointer-like value in the block as a pointer.
1328 // If ptrmask != nil, only words that are marked in ptrmask are
1329 // considered as potential pointers.
1331 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1332 // and may contain pointers to stack objects.
1333 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1334 if debugScanConservative {
1336 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1337 hexdumpWords(b, b+n, func(p uintptr) byte {
1339 word := (p - b) / goarch.PtrSize
1340 bits := *addb(ptrmask, word/8)
1341 if (bits>>(word%8))&1 == 0 {
1346 val := *(*uintptr)(unsafe.Pointer(p))
1347 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1351 span := spanOfHeap(val)
1355 idx := span.objIndex(val)
1356 if span.isFree(idx) {
1364 for i := uintptr(0); i < n; i += goarch.PtrSize {
1366 word := i / goarch.PtrSize
1367 bits := *addb(ptrmask, word/8)
1369 // Skip 8 words (the loop increment will do the 8th)
1371 // This must be the first time we've
1372 // seen this word of ptrmask, so i
1373 // must be 8-word-aligned, but check
1374 // our reasoning just in case.
1375 if i%(goarch.PtrSize*8) != 0 {
1376 throw("misaligned mask")
1378 i += goarch.PtrSize*8 - goarch.PtrSize
1381 if (bits>>(word%8))&1 == 0 {
1386 val := *(*uintptr)(unsafe.Pointer(b + i))
1388 // Check if val points into the stack.
1389 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1390 // val may point to a stack object. This
1391 // object may be dead from last cycle and
1392 // hence may contain pointers to unallocated
1393 // objects, but unlike heap objects we can't
1394 // tell if it's already dead. Hence, if all
1395 // pointers to this object are from
1396 // conservative scanning, we have to scan it
1397 // defensively, too.
1398 state.putPtr(val, true)
1402 // Check if val points to a heap span.
1403 span := spanOfHeap(val)
1408 // Check if val points to an allocated object.
1409 idx := span.objIndex(val)
1410 if span.isFree(idx) {
1414 // val points to an allocated object. Mark it.
1415 obj := span.base() + idx*span.elemsize
1416 greyobject(obj, b, i, span, gcw, idx)
1420 // Shade the object if it isn't already.
1421 // The object is not nil and known to be in the heap.
1422 // Preemption must be disabled.
1424 func shade(b uintptr) {
1425 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1426 gcw := &getg().m.p.ptr().gcw
1427 greyobject(obj, 0, 0, span, gcw, objIndex)
1431 // obj is the start of an object with mark mbits.
1432 // If it isn't already marked, mark it and enqueue into gcw.
1433 // base and off are for debugging only and could be removed.
1435 // See also wbBufFlush1, which partially duplicates this logic.
1437 //go:nowritebarrierrec
1438 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1439 // obj should be start of allocation, and so must be at least pointer-aligned.
1440 if obj&(goarch.PtrSize-1) != 0 {
1441 throw("greyobject: obj not pointer-aligned")
1443 mbits := span.markBitsForIndex(objIndex)
1446 if setCheckmark(obj, base, off, mbits) {
1451 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1452 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1453 gcDumpObject("base", base, off)
1454 gcDumpObject("obj", obj, ^uintptr(0))
1455 getg().m.traceback = 2
1456 throw("marking free object")
1459 // If marked we have nothing to do.
1460 if mbits.isMarked() {
1466 arena, pageIdx, pageMask := pageIndexOf(span.base())
1467 if arena.pageMarks[pageIdx]&pageMask == 0 {
1468 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1471 // If this is a noscan object, fast-track it to black
1472 // instead of greying it.
1473 if span.spanclass.noscan() {
1474 gcw.bytesMarked += uint64(span.elemsize)
1479 // We're adding obj to P's local workbuf, so it's likely
1480 // this object will be processed soon by the same P.
1481 // Even if the workbuf gets flushed, there will likely still be
1482 // some benefit on platforms with inclusive shared caches.
1484 // Queue the obj for scanning.
1485 if !gcw.putFast(obj) {
1490 // gcDumpObject dumps the contents of obj for debugging and marks the
1491 // field at byte offset off in obj.
1492 func gcDumpObject(label string, obj, off uintptr) {
1494 print(label, "=", hex(obj))
1499 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1500 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1501 print(mSpanStateNames[state], "\n")
1503 print("unknown(", state, ")\n")
1508 if s.state.get() == mSpanManual && size == 0 {
1509 // We're printing something from a stack frame. We
1510 // don't know how big it is, so just show up to an
1512 size = off + goarch.PtrSize
1514 for i := uintptr(0); i < size; i += goarch.PtrSize {
1515 // For big objects, just print the beginning (because
1516 // that usually hints at the object's type) and the
1517 // fields around off.
1518 if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
1526 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1537 // gcmarknewobject marks a newly allocated object black. obj must
1538 // not contain any non-nil pointers.
1540 // This is nosplit so it can manipulate a gcWork without preemption.
1544 func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
1545 if useCheckmark { // The world should be stopped so this should not happen.
1546 throw("gcmarknewobject called while doing checkmark")
1550 objIndex := span.objIndex(obj)
1551 span.markBitsForIndex(objIndex).setMarked()
1554 arena, pageIdx, pageMask := pageIndexOf(span.base())
1555 if arena.pageMarks[pageIdx]&pageMask == 0 {
1556 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1559 gcw := &getg().m.p.ptr().gcw
1560 gcw.bytesMarked += uint64(size)
1561 if !goexperiment.PacerRedesign {
1562 // The old pacer counts newly allocated memory toward
1563 // heapScanWork because heapScan is continuously updated
1564 // throughout the GC cyle with newly allocated memory. However,
1565 // these objects are never actually scanned, so we need
1566 // to account for them in heapScanWork here, "faking" their work.
1567 // Otherwise the pacer will think it's always behind, potentially
1568 // by a large margin.
1570 // The new pacer doesn't care about this because it ceases to updated
1571 // heapScan once a GC cycle starts, effectively snapshotting it.
1572 gcw.heapScanWork += int64(scanSize)
1576 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1578 // The world must be stopped.
1579 func gcMarkTinyAllocs() {
1580 assertWorldStopped()
1582 for _, p := range allp {
1584 if c == nil || c.tiny == 0 {
1587 _, span, objIndex := findObject(c.tiny, 0, 0)
1589 greyobject(c.tiny, 0, 0, span, gcw, objIndex)