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.nStackRoots = int(atomic.Loaduintptr(&allglen))
107 work.markrootNext = 0
108 work.markrootJobs = uint32(fixedRootCount + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
110 // Calculate base indexes of each root type
111 work.baseData = uint32(fixedRootCount)
112 work.baseBSS = work.baseData + uint32(work.nDataRoots)
113 work.baseSpans = work.baseBSS + uint32(work.nBSSRoots)
114 work.baseStacks = work.baseSpans + uint32(work.nSpanRoots)
115 work.baseEnd = work.baseStacks + uint32(work.nStackRoots)
118 // gcMarkRootCheck checks that all roots have been scanned. It is
119 // purely for debugging.
120 func gcMarkRootCheck() {
121 if work.markrootNext < work.markrootJobs {
122 print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
123 throw("left over markroot jobs")
126 // Check that stacks have been scanned.
128 // We only check the first nStackRoots Gs that we should have scanned.
129 // Since we don't care about newer Gs (see comment in
130 // gcMarkRootPrepare), no locking is required.
132 forEachGRace(func(gp *g) {
133 if i >= work.nStackRoots {
138 println("gp", gp, "goid", gp.goid,
139 "status", readgstatus(gp),
140 "gcscandone", gp.gcscandone)
141 throw("scan missed a g")
148 // ptrmask for an allocation containing a single pointer.
149 var oneptrmask = [...]uint8{1}
151 // markroot scans the i'th root.
153 // Preemption must be disabled (because this uses a gcWork).
155 // Returns the amount of GC work credit produced by the operation.
156 // If flushBgCredit is true, then that credit is also flushed
157 // to the background credit pool.
159 // nowritebarrier is only advisory here.
162 func markroot(gcw *gcWork, i uint32, flushBgCredit bool) int64 {
163 // Note: if you add a case here, please also update heapdump.go:dumproots.
165 var workCounter *atomic.Int64
167 case work.baseData <= i && i < work.baseBSS:
168 workCounter = &gcController.globalsScanWork
169 for _, datap := range activeModules() {
170 workDone += markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
173 case work.baseBSS <= i && i < work.baseSpans:
174 workCounter = &gcController.globalsScanWork
175 for _, datap := range activeModules() {
176 workDone += markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
179 case i == fixedRootFinalizers:
180 for fb := allfin; fb != nil; fb = fb.alllink {
181 cnt := uintptr(atomic.Load(&fb.cnt))
182 scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
185 case i == fixedRootFreeGStacks:
186 // Switch to the system stack so we can call
188 systemstack(markrootFreeGStacks)
190 case work.baseSpans <= i && i < work.baseStacks:
191 // mark mspan.specials
192 markrootSpans(gcw, int(i-work.baseSpans))
195 // the rest is scanning goroutine stacks
196 workCounter = &gcController.stackScanWork
198 if work.baseStacks <= i && i < work.baseEnd {
199 // N.B. Atomic read of allglen in gcMarkRootPrepare
200 // acts as a barrier to ensure that allgs must be large
201 // enough to contain all relevant Gs.
202 gp = allgs[i-work.baseStacks]
204 throw("markroot: bad index")
207 // remember when we've first observed the G blocked
208 // needed only to output in traceback
209 status := readgstatus(gp) // We are not in a scan state
210 if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
211 gp.waitsince = work.tstart
214 // scanstack must be done on the system stack in case
215 // we're trying to scan our own stack.
217 // If this is a self-scan, put the user G in
218 // _Gwaiting to prevent self-deadlock. It may
219 // already be in _Gwaiting if this is a mark
220 // worker or we're in mark termination.
221 userG := getg().m.curg
222 selfScan := gp == userG && readgstatus(userG) == _Grunning
224 casgstatus(userG, _Grunning, _Gwaiting)
225 userG.waitreason = waitReasonGarbageCollectionScan
228 // TODO: suspendG blocks (and spins) until gp
229 // stops, which may take a while for
230 // running goroutines. Consider doing this in
231 // two phases where the first is non-blocking:
232 // we scan the stacks we can and ask running
233 // goroutines to scan themselves; and the
235 stopped := suspendG(gp)
241 throw("g already scanned")
243 workDone += scanstack(gp, gcw)
248 casgstatus(userG, _Gwaiting, _Grunning)
252 if goexperiment.PacerRedesign {
253 if workCounter != nil && workDone != 0 {
254 workCounter.Add(workDone)
256 gcFlushBgCredit(workDone)
263 // markrootBlock scans the shard'th shard of the block of memory [b0,
264 // b0+n0), with the given pointer mask.
266 // Returns the amount of work done.
269 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) int64 {
270 if rootBlockBytes%(8*goarch.PtrSize) != 0 {
271 // This is necessary to pick byte offsets in ptrmask0.
272 throw("rootBlockBytes must be a multiple of 8*ptrSize")
275 // Note that if b0 is toward the end of the address space,
276 // then b0 + rootBlockBytes might wrap around.
277 // These tests are written to avoid any possible overflow.
278 off := uintptr(shard) * rootBlockBytes
283 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
284 n := uintptr(rootBlockBytes)
290 scanblock(b, n, ptrmask, gcw, nil)
294 // markrootFreeGStacks frees stacks of dead Gs.
296 // This does not free stacks of dead Gs cached on Ps, but having a few
297 // cached stacks around isn't a problem.
298 func markrootFreeGStacks() {
299 // Take list of dead Gs with stacks.
300 lock(&sched.gFree.lock)
301 list := sched.gFree.stack
302 sched.gFree.stack = gList{}
303 unlock(&sched.gFree.lock)
309 q := gQueue{list.head, list.head}
310 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
314 // Manipulate the queue directly since the Gs are
315 // already all linked the right way.
319 // Put Gs back on the free list.
320 lock(&sched.gFree.lock)
321 sched.gFree.noStack.pushAll(q)
322 unlock(&sched.gFree.lock)
325 // markrootSpans marks roots for one shard of markArenas.
328 func markrootSpans(gcw *gcWork, shard int) {
329 // Objects with finalizers have two GC-related invariants:
331 // 1) Everything reachable from the object must be marked.
332 // This ensures that when we pass the object to its finalizer,
333 // everything the finalizer can reach will be retained.
335 // 2) Finalizer specials (which are not in the garbage
336 // collected heap) are roots. In practice, this means the fn
337 // field must be scanned.
338 sg := mheap_.sweepgen
340 // Find the arena and page index into that arena for this shard.
341 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
342 ha := mheap_.arenas[ai.l1()][ai.l2()]
343 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
345 // Construct slice of bitmap which we'll iterate over.
346 specialsbits := ha.pageSpecials[arenaPage/8:]
347 specialsbits = specialsbits[:pagesPerSpanRoot/8]
348 for i := range specialsbits {
349 // Find set bits, which correspond to spans with specials.
350 specials := atomic.Load8(&specialsbits[i])
354 for j := uint(0); j < 8; j++ {
355 if specials&(1<<j) == 0 {
358 // Find the span for this bit.
360 // This value is guaranteed to be non-nil because having
361 // specials implies that the span is in-use, and since we're
362 // currently marking we can be sure that we don't have to worry
363 // about the span being freed and re-used.
364 s := ha.spans[arenaPage+uint(i)*8+j]
366 // The state must be mSpanInUse if the specials bit is set, so
367 // sanity check that.
368 if state := s.state.get(); state != mSpanInUse {
369 print("s.state = ", state, "\n")
370 throw("non in-use span found with specials bit set")
372 // Check that this span was swept (it may be cached or uncached).
373 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
374 // sweepgen was updated (+2) during non-checkmark GC pass
375 print("sweep ", s.sweepgen, " ", sg, "\n")
376 throw("gc: unswept span")
379 // Lock the specials to prevent a special from being
380 // removed from the list while we're traversing it.
382 for sp := s.specials; sp != nil; sp = sp.next {
383 if sp.kind != _KindSpecialFinalizer {
386 // don't mark finalized object, but scan it so we
387 // retain everything it points to.
388 spf := (*specialfinalizer)(unsafe.Pointer(sp))
389 // A finalizer can be set for an inner byte of an object, find object beginning.
390 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
392 // Mark everything that can be reached from
393 // the object (but *not* the object itself or
394 // we'll never collect it).
397 // The special itself is a root.
398 scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
400 unlock(&s.speciallock)
405 // gcAssistAlloc performs GC work to make gp's assist debt positive.
406 // gp must be the calling user gorountine.
408 // This must be called with preemption enabled.
409 func gcAssistAlloc(gp *g) {
410 // Don't assist in non-preemptible contexts. These are
411 // generally fragile and won't allow the assist to block.
412 if getg() == gp.m.g0 {
415 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
421 // Compute the amount of scan work we need to do to make the
422 // balance positive. When the required amount of work is low,
423 // we over-assist to build up credit for future allocations
424 // and amortize the cost of assisting.
425 assistWorkPerByte := gcController.assistWorkPerByte.Load()
426 assistBytesPerWork := gcController.assistBytesPerWork.Load()
427 debtBytes := -gp.gcAssistBytes
428 scanWork := int64(assistWorkPerByte * float64(debtBytes))
429 if scanWork < gcOverAssistWork {
430 scanWork = gcOverAssistWork
431 debtBytes = int64(assistBytesPerWork * float64(scanWork))
434 // Steal as much credit as we can from the background GC's
435 // scan credit. This is racy and may drop the background
436 // credit below 0 if two mutators steal at the same time. This
437 // will just cause steals to fail until credit is accumulated
438 // again, so in the long run it doesn't really matter, but we
439 // do have to handle the negative credit case.
440 bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
442 if bgScanCredit > 0 {
443 if bgScanCredit < scanWork {
444 stolen = bgScanCredit
445 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
448 gp.gcAssistBytes += debtBytes
450 atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
455 // We were able to steal all of the credit we
458 traceGCMarkAssistDone()
464 if trace.enabled && !traced {
466 traceGCMarkAssistStart()
469 // Perform assist work
471 gcAssistAlloc1(gp, scanWork)
472 // The user stack may have moved, so this can't touch
473 // anything on it until it returns from systemstack.
476 completed := gp.param != nil
482 if gp.gcAssistBytes < 0 {
483 // We were unable steal enough credit or perform
484 // enough work to pay off the assist debt. We need to
485 // do one of these before letting the mutator allocate
486 // more to prevent over-allocation.
488 // If this is because we were preempted, reschedule
489 // and try some more.
495 // Add this G to an assist queue and park. When the GC
496 // has more background credit, it will satisfy queued
497 // assists before flushing to the global credit pool.
499 // Note that this does *not* get woken up when more
500 // work is added to the work list. The theory is that
501 // there wasn't enough work to do anyway, so we might
502 // as well let background marking take care of the
503 // work that is available.
508 // At this point either background GC has satisfied
509 // this G's assist debt, or the GC cycle is over.
512 traceGCMarkAssistDone()
516 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
517 // stack. This is a separate function to make it easier to see that
518 // we're not capturing anything from the user stack, since the user
519 // stack may move while we're in this function.
521 // gcAssistAlloc1 indicates whether this assist completed the mark
522 // phase by setting gp.param to non-nil. This can't be communicated on
523 // the stack since it may move.
526 func gcAssistAlloc1(gp *g, scanWork int64) {
527 // Clear the flag indicating that this assist completed the
531 if atomic.Load(&gcBlackenEnabled) == 0 {
532 // The gcBlackenEnabled check in malloc races with the
533 // store that clears it but an atomic check in every malloc
534 // would be a performance hit.
535 // Instead we recheck it here on the non-preemptable system
536 // stack to determine if we should perform an assist.
538 // GC is done, so ignore any remaining debt.
542 // Track time spent in this assist. Since we're on the
543 // system stack, this is non-preemptible, so we can
544 // just measure start and end time.
545 startTime := nanotime()
547 decnwait := atomic.Xadd(&work.nwait, -1)
548 if decnwait == work.nproc {
549 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
550 throw("nwait > work.nprocs")
553 // gcDrainN requires the caller to be preemptible.
554 casgstatus(gp, _Grunning, _Gwaiting)
555 gp.waitreason = waitReasonGCAssistMarking
557 // drain own cached work first in the hopes that it
558 // will be more cache friendly.
559 gcw := &getg().m.p.ptr().gcw
560 workDone := gcDrainN(gcw, scanWork)
562 casgstatus(gp, _Gwaiting, _Grunning)
564 // Record that we did this much scan work.
566 // Back out the number of bytes of assist credit that
567 // this scan work counts for. The "1+" is a poor man's
568 // round-up, to ensure this adds credit even if
569 // assistBytesPerWork is very low.
570 assistBytesPerWork := gcController.assistBytesPerWork.Load()
571 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
573 // If this is the last worker and we ran out of work,
574 // signal a completion point.
575 incnwait := atomic.Xadd(&work.nwait, +1)
576 if incnwait > work.nproc {
577 println("runtime: work.nwait=", incnwait,
578 "work.nproc=", work.nproc)
579 throw("work.nwait > work.nproc")
582 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
583 // This has reached a background completion point. Set
584 // gp.param to a non-nil value to indicate this. It
585 // doesn't matter what we set it to (it just has to be
587 gp.param = unsafe.Pointer(gp)
589 duration := nanotime() - startTime
591 _p_.gcAssistTime += duration
592 if _p_.gcAssistTime > gcAssistTimeSlack {
593 atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
598 // gcWakeAllAssists wakes all currently blocked assists. This is used
599 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
600 // new assists from going to sleep after this point.
601 func gcWakeAllAssists() {
602 lock(&work.assistQueue.lock)
603 list := work.assistQueue.q.popList()
605 unlock(&work.assistQueue.lock)
608 // gcParkAssist puts the current goroutine on the assist queue and parks.
610 // gcParkAssist reports whether the assist is now satisfied. If it
611 // returns false, the caller must retry the assist.
612 func gcParkAssist() bool {
613 lock(&work.assistQueue.lock)
614 // If the GC cycle finished while we were getting the lock,
615 // exit the assist. The cycle can't finish while we hold the
617 if atomic.Load(&gcBlackenEnabled) == 0 {
618 unlock(&work.assistQueue.lock)
623 oldList := work.assistQueue.q
624 work.assistQueue.q.pushBack(gp)
626 // Recheck for background credit now that this G is in
627 // the queue, but can still back out. This avoids a
628 // race in case background marking has flushed more
629 // credit since we checked above.
630 if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
631 work.assistQueue.q = oldList
632 if oldList.tail != 0 {
633 oldList.tail.ptr().schedlink.set(nil)
635 unlock(&work.assistQueue.lock)
639 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
643 // gcFlushBgCredit flushes scanWork units of background scan work
644 // credit. This first satisfies blocked assists on the
645 // work.assistQueue and then flushes any remaining credit to
646 // gcController.bgScanCredit.
648 // Write barriers are disallowed because this is used by gcDrain after
649 // it has ensured that all work is drained and this must preserve that
652 //go:nowritebarrierrec
653 func gcFlushBgCredit(scanWork int64) {
654 if work.assistQueue.q.empty() {
655 // Fast path; there are no blocked assists. There's a
656 // small window here where an assist may add itself to
657 // the blocked queue and park. If that happens, we'll
658 // just get it on the next flush.
659 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
663 assistBytesPerWork := gcController.assistBytesPerWork.Load()
664 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
666 lock(&work.assistQueue.lock)
667 for !work.assistQueue.q.empty() && scanBytes > 0 {
668 gp := work.assistQueue.q.pop()
669 // Note that gp.gcAssistBytes is negative because gp
670 // is in debt. Think carefully about the signs below.
671 if scanBytes+gp.gcAssistBytes >= 0 {
672 // Satisfy this entire assist debt.
673 scanBytes += gp.gcAssistBytes
675 // It's important that we *not* put gp in
676 // runnext. Otherwise, it's possible for user
677 // code to exploit the GC worker's high
678 // scheduler priority to get itself always run
679 // before other goroutines and always in the
680 // fresh quantum started by GC.
683 // Partially satisfy this assist.
684 gp.gcAssistBytes += scanBytes
686 // As a heuristic, we move this assist to the
687 // back of the queue so that large assists
688 // can't clog up the assist queue and
689 // substantially delay small assists.
690 work.assistQueue.q.pushBack(gp)
696 // Convert from scan bytes back to work.
697 assistWorkPerByte := gcController.assistWorkPerByte.Load()
698 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
699 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
701 unlock(&work.assistQueue.lock)
704 // scanstack scans gp's stack, greying all pointers found on the stack.
706 // For goexperiment.PacerRedesign:
707 // Returns the amount of scan work performed, but doesn't update
708 // gcController.stackScanWork or flush any credit. Any background credit produced
709 // by this function should be flushed by its caller. scanstack itself can't
710 // safely flush because it may result in trying to wake up a goroutine that
711 // was just scanned, resulting in a self-deadlock.
713 // scanstack will also shrink the stack if it is safe to do so. If it
714 // is not, it schedules a stack shrink for the next synchronous safe
717 // scanstack is marked go:systemstack because it must not be preempted
718 // while using a workbuf.
722 func scanstack(gp *g, gcw *gcWork) int64 {
723 if readgstatus(gp)&_Gscan == 0 {
724 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
725 throw("scanstack - bad status")
728 switch readgstatus(gp) &^ _Gscan {
730 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
731 throw("mark - bad status")
735 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
736 throw("scanstack: goroutine not stopped")
737 case _Grunnable, _Gsyscall, _Gwaiting:
742 throw("can't scan our own stack")
745 // stackSize is the amount of work we'll be reporting.
747 // We report the total stack size, more than we scan,
748 // because this number needs to line up with gcControllerState's
749 // stackScan and scannableStackSize fields.
751 // See the documentation on those fields for more information.
752 stackSize := gp.stack.hi - gp.stack.lo
754 if isShrinkStackSafe(gp) {
755 // Shrink the stack if not much of it is being used.
758 // Otherwise, shrink the stack at the next sync safe point.
759 gp.preemptShrink = true
762 var state stackScanState
763 state.stack = gp.stack
766 println("stack trace goroutine", gp.goid)
769 if debugScanConservative && gp.asyncSafePoint {
770 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
773 // Scan the saved context register. This is effectively a live
774 // register that gets moved back and forth between the
775 // register and sched.ctxt without a write barrier.
776 if gp.sched.ctxt != nil {
777 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
780 // Scan the stack. Accumulate a list of stack objects.
781 scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
782 scanframeworker(frame, &state, gcw)
785 gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
787 // Find additional pointers that point into the stack from the heap.
788 // Currently this includes defers and panics. See also function copystack.
790 // Find and trace other pointers in defer records.
791 for d := gp._defer; d != nil; d = d.link {
793 // Scan the func value, which could be a stack allocated closure.
795 scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
798 // The link field of a stack-allocated defer record might point
799 // to a heap-allocated defer record. Keep that heap record live.
800 scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
802 // Retain defers records themselves.
803 // Defer records might not be reachable from the G through regular heap
804 // tracing because the defer linked list might weave between the stack and the heap.
806 scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
809 if gp._panic != nil {
810 // Panics are always stack allocated.
811 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
814 // Find and scan all reachable stack objects.
816 // The state's pointer queue prioritizes precise pointers over
817 // conservative pointers so that we'll prefer scanning stack
818 // objects precisely.
821 p, conservative := state.getPtr()
825 obj := state.findObject(p)
831 // We've already scanned this object.
834 obj.setRecord(nil) // Don't scan it again.
837 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
839 print(" (conservative)")
847 // This path is pretty unlikely, an object large enough
848 // to have a GC program allocated on the stack.
849 // We need some space to unpack the program into a straight
850 // bitmask, which we allocate/free here.
851 // TODO: it would be nice if there were a way to run a GC
852 // program without having to store all its bits. We'd have
853 // to change from a Lempel-Ziv style program to something else.
854 // Or we can forbid putting objects on stacks if they require
855 // a gc program (see issue 27447).
856 s = materializeGCProg(r.ptrdata(), gcdata)
857 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
860 b := state.stack.lo + uintptr(obj.off)
862 scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
864 scanblock(b, r.ptrdata(), gcdata, gcw, &state)
868 dematerializeGCProg(s)
872 // Deallocate object buffers.
873 // (Pointer buffers were all deallocated in the loop above.)
874 for state.head != nil {
878 for i := 0; i < x.nobj; i++ {
880 if obj.r == nil { // reachable
883 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
884 // Note: not necessarily really dead - only reachable-from-ptr dead.
888 putempty((*workbuf)(unsafe.Pointer(x)))
890 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
891 throw("remaining pointer buffers")
893 return int64(stackSize)
896 // Scan a stack frame: local variables and function arguments/results.
898 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
899 if _DebugGC > 1 && frame.continpc != 0 {
900 print("scanframe ", funcname(frame.fn), "\n")
903 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
904 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
905 if state.conservative || isAsyncPreempt || isDebugCall {
906 if debugScanConservative {
907 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
910 // Conservatively scan the frame. Unlike the precise
911 // case, this includes the outgoing argument space
912 // since we may have stopped while this function was
913 // setting up a call.
915 // TODO: We could narrow this down if the compiler
916 // produced a single map per function of stack slots
917 // and registers that ever contain a pointer.
919 size := frame.varp - frame.sp
921 scanConservative(frame.sp, size, nil, gcw, state)
925 // Scan arguments to this frame.
926 if frame.arglen != 0 {
927 // TODO: We could pass the entry argument map
928 // to narrow this down further.
929 scanConservative(frame.argp, frame.arglen, nil, gcw, state)
932 if isAsyncPreempt || isDebugCall {
933 // This function's frame contained the
934 // registers for the asynchronously stopped
935 // parent frame. Scan the parent
937 state.conservative = true
939 // We only wanted to scan those two frames
940 // conservatively. Clear the flag for future
942 state.conservative = false
947 locals, args, objs := getStackMap(frame, &state.cache, false)
949 // Scan local variables if stack frame has been allocated.
951 size := uintptr(locals.n) * goarch.PtrSize
952 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
957 scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
960 // Add all stack objects to the stack object list.
962 // varp is 0 for defers, where there are no locals.
963 // In that case, there can't be a pointer to its args, either.
964 // (And all args would be scanned above anyway.)
965 for i := range objs {
968 base := frame.varp // locals base pointer
970 base = frame.argp // arguments and return values base pointer
972 ptr := base + uintptr(off)
974 // object hasn't been allocated in the frame yet.
978 println("stkobj at", hex(ptr), "of size", obj.size)
980 state.addObject(ptr, obj)
985 type gcDrainFlags int
988 gcDrainUntilPreempt gcDrainFlags = 1 << iota
994 // gcDrain scans roots and objects in work buffers, blackening grey
995 // objects until it is unable to get more work. It may return before
996 // GC is done; it's the caller's responsibility to balance work from
999 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
1002 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
1005 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
1006 // pollFractionalWorkerExit() returns true. This implies
1009 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
1010 // credit to gcController.bgScanCredit every gcCreditSlack units of
1013 // gcDrain will always return if there is a pending STW.
1016 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
1017 if !writeBarrier.needed {
1018 throw("gcDrain phase incorrect")
1022 preemptible := flags&gcDrainUntilPreempt != 0
1023 flushBgCredit := flags&gcDrainFlushBgCredit != 0
1024 idle := flags&gcDrainIdle != 0
1026 initScanWork := gcw.heapScanWork
1028 // checkWork is the scan work before performing the next
1029 // self-preempt check.
1030 checkWork := int64(1<<63 - 1)
1031 var check func() bool
1032 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
1033 checkWork = initScanWork + drainCheckThreshold
1036 } else if flags&gcDrainFractional != 0 {
1037 check = pollFractionalWorkerExit
1041 // Drain root marking jobs.
1042 if work.markrootNext < work.markrootJobs {
1043 // Stop if we're preemptible or if someone wants to STW.
1044 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1045 job := atomic.Xadd(&work.markrootNext, +1) - 1
1046 if job >= work.markrootJobs {
1049 markroot(gcw, job, flushBgCredit)
1050 if check != nil && check() {
1056 // Drain heap marking jobs.
1057 // Stop if we're preemptible or if someone wants to STW.
1058 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1059 // Try to keep work available on the global queue. We used to
1060 // check if there were waiting workers, but it's better to
1061 // just keep work available than to make workers wait. In the
1062 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1068 b := gcw.tryGetFast()
1072 // Flush the write barrier
1073 // buffer; this may create
1080 // Unable to get work.
1085 // Flush background scan work credit to the global
1086 // account if we've accumulated enough locally so
1087 // mutator assists can draw on it.
1088 if gcw.heapScanWork >= gcCreditSlack {
1089 gcController.heapScanWork.Add(gcw.heapScanWork)
1091 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1094 checkWork -= gcw.heapScanWork
1095 gcw.heapScanWork = 0
1098 checkWork += drainCheckThreshold
1099 if check != nil && check() {
1107 // Flush remaining scan work credit.
1108 if gcw.heapScanWork > 0 {
1109 gcController.heapScanWork.Add(gcw.heapScanWork)
1111 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1113 gcw.heapScanWork = 0
1117 // gcDrainN blackens grey objects until it has performed roughly
1118 // scanWork units of scan work or the G is preempted. This is
1119 // best-effort, so it may perform less work if it fails to get a work
1120 // buffer. Otherwise, it will perform at least n units of work, but
1121 // may perform more because scanning is always done in whole object
1122 // increments. It returns the amount of scan work performed.
1124 // The caller goroutine must be in a preemptible state (e.g.,
1125 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1126 // consequence, this must be called on the system stack.
1130 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1131 if !writeBarrier.needed {
1132 throw("gcDrainN phase incorrect")
1135 // There may already be scan work on the gcw, which we don't
1136 // want to claim was done by this call.
1137 workFlushed := -gcw.heapScanWork
1140 for !gp.preempt && workFlushed+gcw.heapScanWork < scanWork {
1141 // See gcDrain comment.
1146 b := gcw.tryGetFast()
1150 // Flush the write barrier buffer;
1151 // this may create more work.
1158 // Try to do a root job.
1159 if work.markrootNext < work.markrootJobs {
1160 job := atomic.Xadd(&work.markrootNext, +1) - 1
1161 if job < work.markrootJobs {
1162 work := markroot(gcw, job, false)
1163 if goexperiment.PacerRedesign {
1169 // No heap or root jobs.
1175 // Flush background scan work credit.
1176 if gcw.heapScanWork >= gcCreditSlack {
1177 gcController.heapScanWork.Add(gcw.heapScanWork)
1178 workFlushed += gcw.heapScanWork
1179 gcw.heapScanWork = 0
1183 // Unlike gcDrain, there's no need to flush remaining work
1184 // here because this never flushes to bgScanCredit and
1185 // gcw.dispose will flush any remaining work to scanWork.
1187 return workFlushed + gcw.heapScanWork
1190 // scanblock scans b as scanobject would, but using an explicit
1191 // pointer bitmap instead of the heap bitmap.
1193 // This is used to scan non-heap roots, so it does not update
1194 // gcw.bytesMarked or gcw.heapScanWork.
1196 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1198 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1199 // Use local copies of original parameters, so that a stack trace
1200 // due to one of the throws below shows the original block
1205 for i := uintptr(0); i < n; {
1206 // Find bits for the next word.
1207 bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
1209 i += goarch.PtrSize * 8
1212 for j := 0; j < 8 && i < n; j++ {
1214 // Same work as in scanobject; see comments there.
1215 p := *(*uintptr)(unsafe.Pointer(b + i))
1217 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1218 greyobject(obj, b, i, span, gcw, objIndex)
1219 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1220 stk.putPtr(p, false)
1230 // scanobject scans the object starting at b, adding pointers to gcw.
1231 // b must point to the beginning of a heap object or an oblet.
1232 // scanobject consults the GC bitmap for the pointer mask and the
1233 // spans for the size of the object.
1236 func scanobject(b uintptr, gcw *gcWork) {
1237 // Prefetch object before we scan it.
1239 // This will overlap fetching the beginning of the object with initial
1240 // setup before we start scanning the object.
1243 // Find the bits for b and the size of the object at b.
1245 // b is either the beginning of an object, in which case this
1246 // is the size of the object to scan, or it points to an
1247 // oblet, in which case we compute the size to scan below.
1248 hbits := heapBitsForAddr(b)
1249 s := spanOfUnchecked(b)
1252 throw("scanobject n == 0")
1255 if n > maxObletBytes {
1256 // Large object. Break into oblets for better
1257 // parallelism and lower latency.
1259 // It's possible this is a noscan object (not
1260 // from greyobject, but from other code
1261 // paths), in which case we must *not* enqueue
1262 // oblets since their bitmaps will be
1264 if s.spanclass.noscan() {
1265 // Bypass the whole scan.
1266 gcw.bytesMarked += uint64(n)
1270 // Enqueue the other oblets to scan later.
1271 // Some oblets may be in b's scalar tail, but
1272 // these will be marked as "no more pointers",
1273 // so we'll drop out immediately when we go to
1275 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1276 if !gcw.putFast(oblet) {
1282 // Compute the size of the oblet. Since this object
1283 // must be a large object, s.base() is the beginning
1285 n = s.base() + s.elemsize - b
1286 if n > maxObletBytes {
1292 for i = 0; i < n; i, hbits = i+goarch.PtrSize, hbits.next() {
1293 // Load bits once. See CL 22712 and issue 16973 for discussion.
1294 bits := hbits.bits()
1295 if bits&bitScan == 0 {
1296 break // no more pointers in this object
1298 if bits&bitPointer == 0 {
1299 continue // not a pointer
1302 // Work here is duplicated in scanblock and above.
1303 // If you make changes here, make changes there too.
1304 obj := *(*uintptr)(unsafe.Pointer(b + i))
1306 // At this point we have extracted the next potential pointer.
1307 // Quickly filter out nil and pointers back to the current object.
1308 if obj != 0 && obj-b >= n {
1309 // Test if obj points into the Go heap and, if so,
1312 // Note that it's possible for findObject to
1313 // fail if obj points to a just-allocated heap
1314 // object because of a race with growing the
1315 // heap. In this case, we know the object was
1316 // just allocated and hence will be marked by
1317 // allocation itself.
1318 if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
1319 greyobject(obj, b, i, span, gcw, objIndex)
1323 gcw.bytesMarked += uint64(n)
1324 gcw.heapScanWork += int64(i)
1327 // scanConservative scans block [b, b+n) conservatively, treating any
1328 // pointer-like value in the block as a pointer.
1330 // If ptrmask != nil, only words that are marked in ptrmask are
1331 // considered as potential pointers.
1333 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1334 // and may contain pointers to stack objects.
1335 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1336 if debugScanConservative {
1338 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1339 hexdumpWords(b, b+n, func(p uintptr) byte {
1341 word := (p - b) / goarch.PtrSize
1342 bits := *addb(ptrmask, word/8)
1343 if (bits>>(word%8))&1 == 0 {
1348 val := *(*uintptr)(unsafe.Pointer(p))
1349 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1353 span := spanOfHeap(val)
1357 idx := span.objIndex(val)
1358 if span.isFree(idx) {
1366 for i := uintptr(0); i < n; i += goarch.PtrSize {
1368 word := i / goarch.PtrSize
1369 bits := *addb(ptrmask, word/8)
1371 // Skip 8 words (the loop increment will do the 8th)
1373 // This must be the first time we've
1374 // seen this word of ptrmask, so i
1375 // must be 8-word-aligned, but check
1376 // our reasoning just in case.
1377 if i%(goarch.PtrSize*8) != 0 {
1378 throw("misaligned mask")
1380 i += goarch.PtrSize*8 - goarch.PtrSize
1383 if (bits>>(word%8))&1 == 0 {
1388 val := *(*uintptr)(unsafe.Pointer(b + i))
1390 // Check if val points into the stack.
1391 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1392 // val may point to a stack object. This
1393 // object may be dead from last cycle and
1394 // hence may contain pointers to unallocated
1395 // objects, but unlike heap objects we can't
1396 // tell if it's already dead. Hence, if all
1397 // pointers to this object are from
1398 // conservative scanning, we have to scan it
1399 // defensively, too.
1400 state.putPtr(val, true)
1404 // Check if val points to a heap span.
1405 span := spanOfHeap(val)
1410 // Check if val points to an allocated object.
1411 idx := span.objIndex(val)
1412 if span.isFree(idx) {
1416 // val points to an allocated object. Mark it.
1417 obj := span.base() + idx*span.elemsize
1418 greyobject(obj, b, i, span, gcw, idx)
1422 // Shade the object if it isn't already.
1423 // The object is not nil and known to be in the heap.
1424 // Preemption must be disabled.
1426 func shade(b uintptr) {
1427 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1428 gcw := &getg().m.p.ptr().gcw
1429 greyobject(obj, 0, 0, span, gcw, objIndex)
1433 // obj is the start of an object with mark mbits.
1434 // If it isn't already marked, mark it and enqueue into gcw.
1435 // base and off are for debugging only and could be removed.
1437 // See also wbBufFlush1, which partially duplicates this logic.
1439 //go:nowritebarrierrec
1440 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1441 // obj should be start of allocation, and so must be at least pointer-aligned.
1442 if obj&(goarch.PtrSize-1) != 0 {
1443 throw("greyobject: obj not pointer-aligned")
1445 mbits := span.markBitsForIndex(objIndex)
1448 if setCheckmark(obj, base, off, mbits) {
1453 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1454 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1455 gcDumpObject("base", base, off)
1456 gcDumpObject("obj", obj, ^uintptr(0))
1457 getg().m.traceback = 2
1458 throw("marking free object")
1461 // If marked we have nothing to do.
1462 if mbits.isMarked() {
1468 arena, pageIdx, pageMask := pageIndexOf(span.base())
1469 if arena.pageMarks[pageIdx]&pageMask == 0 {
1470 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1473 // If this is a noscan object, fast-track it to black
1474 // instead of greying it.
1475 if span.spanclass.noscan() {
1476 gcw.bytesMarked += uint64(span.elemsize)
1481 // We're adding obj to P's local workbuf, so it's likely
1482 // this object will be processed soon by the same P.
1483 // Even if the workbuf gets flushed, there will likely still be
1484 // some benefit on platforms with inclusive shared caches.
1486 // Queue the obj for scanning.
1487 if !gcw.putFast(obj) {
1492 // gcDumpObject dumps the contents of obj for debugging and marks the
1493 // field at byte offset off in obj.
1494 func gcDumpObject(label string, obj, off uintptr) {
1496 print(label, "=", hex(obj))
1501 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1502 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1503 print(mSpanStateNames[state], "\n")
1505 print("unknown(", state, ")\n")
1510 if s.state.get() == mSpanManual && size == 0 {
1511 // We're printing something from a stack frame. We
1512 // don't know how big it is, so just show up to an
1514 size = off + goarch.PtrSize
1516 for i := uintptr(0); i < size; i += goarch.PtrSize {
1517 // For big objects, just print the beginning (because
1518 // that usually hints at the object's type) and the
1519 // fields around off.
1520 if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
1528 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1539 // gcmarknewobject marks a newly allocated object black. obj must
1540 // not contain any non-nil pointers.
1542 // This is nosplit so it can manipulate a gcWork without preemption.
1546 func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
1547 if useCheckmark { // The world should be stopped so this should not happen.
1548 throw("gcmarknewobject called while doing checkmark")
1552 objIndex := span.objIndex(obj)
1553 span.markBitsForIndex(objIndex).setMarked()
1556 arena, pageIdx, pageMask := pageIndexOf(span.base())
1557 if arena.pageMarks[pageIdx]&pageMask == 0 {
1558 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1561 gcw := &getg().m.p.ptr().gcw
1562 gcw.bytesMarked += uint64(size)
1563 if !goexperiment.PacerRedesign {
1564 // The old pacer counts newly allocated memory toward
1565 // heapScanWork because heapScan is continuously updated
1566 // throughout the GC cyle with newly allocated memory. However,
1567 // these objects are never actually scanned, so we need
1568 // to account for them in heapScanWork here, "faking" their work.
1569 // Otherwise the pacer will think it's always behind, potentially
1570 // by a large margin.
1572 // The new pacer doesn't care about this because it ceases to updated
1573 // heapScan once a GC cycle starts, effectively snapshotting it.
1574 gcw.heapScanWork += int64(scanSize)
1578 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1580 // The world must be stopped.
1581 func gcMarkTinyAllocs() {
1582 assertWorldStopped()
1584 for _, p := range allp {
1586 if c == nil || c.tiny == 0 {
1589 _, span, objIndex := findObject(c.tiny, 0, 0)
1591 greyobject(c.tiny, 0, 0, span, gcw, objIndex)