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
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 casGToWaiting(userG, _Grunning, waitReasonGarbageCollectionScan)
225 // TODO: suspendG blocks (and spins) until gp
226 // stops, which may take a while for
227 // running goroutines. Consider doing this in
228 // two phases where the first is non-blocking:
229 // we scan the stacks we can and ask running
230 // goroutines to scan themselves; and the
232 stopped := suspendG(gp)
238 throw("g already scanned")
240 workDone += scanstack(gp, gcw)
245 casgstatus(userG, _Gwaiting, _Grunning)
249 if workCounter != nil && workDone != 0 {
250 workCounter.Add(workDone)
252 gcFlushBgCredit(workDone)
258 // markrootBlock scans the shard'th shard of the block of memory [b0,
259 // b0+n0), with the given pointer mask.
261 // Returns the amount of work done.
264 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) int64 {
265 if rootBlockBytes%(8*goarch.PtrSize) != 0 {
266 // This is necessary to pick byte offsets in ptrmask0.
267 throw("rootBlockBytes must be a multiple of 8*ptrSize")
270 // Note that if b0 is toward the end of the address space,
271 // then b0 + rootBlockBytes might wrap around.
272 // These tests are written to avoid any possible overflow.
273 off := uintptr(shard) * rootBlockBytes
278 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
279 n := uintptr(rootBlockBytes)
285 scanblock(b, n, ptrmask, gcw, nil)
289 // markrootFreeGStacks frees stacks of dead Gs.
291 // This does not free stacks of dead Gs cached on Ps, but having a few
292 // cached stacks around isn't a problem.
293 func markrootFreeGStacks() {
294 // Take list of dead Gs with stacks.
295 lock(&sched.gFree.lock)
296 list := sched.gFree.stack
297 sched.gFree.stack = gList{}
298 unlock(&sched.gFree.lock)
304 q := gQueue{list.head, list.head}
305 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
309 // Manipulate the queue directly since the Gs are
310 // already all linked the right way.
314 // Put Gs back on the free list.
315 lock(&sched.gFree.lock)
316 sched.gFree.noStack.pushAll(q)
317 unlock(&sched.gFree.lock)
320 // markrootSpans marks roots for one shard of markArenas.
323 func markrootSpans(gcw *gcWork, shard int) {
324 // Objects with finalizers have two GC-related invariants:
326 // 1) Everything reachable from the object must be marked.
327 // This ensures that when we pass the object to its finalizer,
328 // everything the finalizer can reach will be retained.
330 // 2) Finalizer specials (which are not in the garbage
331 // collected heap) are roots. In practice, this means the fn
332 // field must be scanned.
333 sg := mheap_.sweepgen
335 // Find the arena and page index into that arena for this shard.
336 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
337 ha := mheap_.arenas[ai.l1()][ai.l2()]
338 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
340 // Construct slice of bitmap which we'll iterate over.
341 specialsbits := ha.pageSpecials[arenaPage/8:]
342 specialsbits = specialsbits[:pagesPerSpanRoot/8]
343 for i := range specialsbits {
344 // Find set bits, which correspond to spans with specials.
345 specials := atomic.Load8(&specialsbits[i])
349 for j := uint(0); j < 8; j++ {
350 if specials&(1<<j) == 0 {
353 // Find the span for this bit.
355 // This value is guaranteed to be non-nil because having
356 // specials implies that the span is in-use, and since we're
357 // currently marking we can be sure that we don't have to worry
358 // about the span being freed and re-used.
359 s := ha.spans[arenaPage+uint(i)*8+j]
361 // The state must be mSpanInUse if the specials bit is set, so
362 // sanity check that.
363 if state := s.state.get(); state != mSpanInUse {
364 print("s.state = ", state, "\n")
365 throw("non in-use span found with specials bit set")
367 // Check that this span was swept (it may be cached or uncached).
368 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
369 // sweepgen was updated (+2) during non-checkmark GC pass
370 print("sweep ", s.sweepgen, " ", sg, "\n")
371 throw("gc: unswept span")
374 // Lock the specials to prevent a special from being
375 // removed from the list while we're traversing it.
377 for sp := s.specials; sp != nil; sp = sp.next {
378 if sp.kind != _KindSpecialFinalizer {
381 // don't mark finalized object, but scan it so we
382 // retain everything it points to.
383 spf := (*specialfinalizer)(unsafe.Pointer(sp))
384 // A finalizer can be set for an inner byte of an object, find object beginning.
385 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
387 // Mark everything that can be reached from
388 // the object (but *not* the object itself or
389 // we'll never collect it).
390 if !s.spanclass.noscan() {
394 // The special itself is a root.
395 scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
397 unlock(&s.speciallock)
402 // gcAssistAlloc performs GC work to make gp's assist debt positive.
403 // gp must be the calling user goroutine.
405 // This must be called with preemption enabled.
406 func gcAssistAlloc(gp *g) {
407 // Don't assist in non-preemptible contexts. These are
408 // generally fragile and won't allow the assist to block.
409 if getg() == gp.m.g0 {
412 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
418 if gcCPULimiter.limiting() {
419 // If the CPU limiter is enabled, intentionally don't
420 // assist to reduce the amount of CPU time spent in the GC.
422 traceGCMarkAssistDone()
426 // Compute the amount of scan work we need to do to make the
427 // balance positive. When the required amount of work is low,
428 // we over-assist to build up credit for future allocations
429 // and amortize the cost of assisting.
430 assistWorkPerByte := gcController.assistWorkPerByte.Load()
431 assistBytesPerWork := gcController.assistBytesPerWork.Load()
432 debtBytes := -gp.gcAssistBytes
433 scanWork := int64(assistWorkPerByte * float64(debtBytes))
434 if scanWork < gcOverAssistWork {
435 scanWork = gcOverAssistWork
436 debtBytes = int64(assistBytesPerWork * float64(scanWork))
439 // Steal as much credit as we can from the background GC's
440 // scan credit. This is racy and may drop the background
441 // credit below 0 if two mutators steal at the same time. This
442 // will just cause steals to fail until credit is accumulated
443 // again, so in the long run it doesn't really matter, but we
444 // do have to handle the negative credit case.
445 bgScanCredit := gcController.bgScanCredit.Load()
447 if bgScanCredit > 0 {
448 if bgScanCredit < scanWork {
449 stolen = bgScanCredit
450 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
453 gp.gcAssistBytes += debtBytes
455 gcController.bgScanCredit.Add(-stolen)
460 // We were able to steal all of the credit we
463 traceGCMarkAssistDone()
469 if traceEnabled() && !traced {
471 traceGCMarkAssistStart()
474 // Perform assist work
476 gcAssistAlloc1(gp, scanWork)
477 // The user stack may have moved, so this can't touch
478 // anything on it until it returns from systemstack.
481 completed := gp.param != nil
487 if gp.gcAssistBytes < 0 {
488 // We were unable steal enough credit or perform
489 // enough work to pay off the assist debt. We need to
490 // do one of these before letting the mutator allocate
491 // more to prevent over-allocation.
493 // If this is because we were preempted, reschedule
494 // and try some more.
500 // Add this G to an assist queue and park. When the GC
501 // has more background credit, it will satisfy queued
502 // assists before flushing to the global credit pool.
504 // Note that this does *not* get woken up when more
505 // work is added to the work list. The theory is that
506 // there wasn't enough work to do anyway, so we might
507 // as well let background marking take care of the
508 // work that is available.
513 // At this point either background GC has satisfied
514 // this G's assist debt, or the GC cycle is over.
517 traceGCMarkAssistDone()
521 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
522 // stack. This is a separate function to make it easier to see that
523 // we're not capturing anything from the user stack, since the user
524 // stack may move while we're in this function.
526 // gcAssistAlloc1 indicates whether this assist completed the mark
527 // phase by setting gp.param to non-nil. This can't be communicated on
528 // the stack since it may move.
531 func gcAssistAlloc1(gp *g, scanWork int64) {
532 // Clear the flag indicating that this assist completed the
536 if atomic.Load(&gcBlackenEnabled) == 0 {
537 // The gcBlackenEnabled check in malloc races with the
538 // store that clears it but an atomic check in every malloc
539 // would be a performance hit.
540 // Instead we recheck it here on the non-preemptible system
541 // stack to determine if we should perform an assist.
543 // GC is done, so ignore any remaining debt.
547 // Track time spent in this assist. Since we're on the
548 // system stack, this is non-preemptible, so we can
549 // just measure start and end time.
551 // Limiter event tracking might be disabled if we end up here
552 // while on a mark worker.
553 startTime := nanotime()
554 trackLimiterEvent := gp.m.p.ptr().limiterEvent.start(limiterEventMarkAssist, startTime)
556 decnwait := atomic.Xadd(&work.nwait, -1)
557 if decnwait == work.nproc {
558 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
559 throw("nwait > work.nprocs")
562 // gcDrainN requires the caller to be preemptible.
563 casGToWaiting(gp, _Grunning, waitReasonGCAssistMarking)
565 // drain own cached work first in the hopes that it
566 // will be more cache friendly.
567 gcw := &getg().m.p.ptr().gcw
568 workDone := gcDrainN(gcw, scanWork)
570 casgstatus(gp, _Gwaiting, _Grunning)
572 // Record that we did this much scan work.
574 // Back out the number of bytes of assist credit that
575 // this scan work counts for. The "1+" is a poor man's
576 // round-up, to ensure this adds credit even if
577 // assistBytesPerWork is very low.
578 assistBytesPerWork := gcController.assistBytesPerWork.Load()
579 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
581 // If this is the last worker and we ran out of work,
582 // signal a completion point.
583 incnwait := atomic.Xadd(&work.nwait, +1)
584 if incnwait > work.nproc {
585 println("runtime: work.nwait=", incnwait,
586 "work.nproc=", work.nproc)
587 throw("work.nwait > work.nproc")
590 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
591 // This has reached a background completion point. Set
592 // gp.param to a non-nil value to indicate this. It
593 // doesn't matter what we set it to (it just has to be
595 gp.param = unsafe.Pointer(gp)
598 duration := now - startTime
600 pp.gcAssistTime += duration
601 if trackLimiterEvent {
602 pp.limiterEvent.stop(limiterEventMarkAssist, now)
604 if pp.gcAssistTime > gcAssistTimeSlack {
605 gcController.assistTime.Add(pp.gcAssistTime)
606 gcCPULimiter.update(now)
611 // gcWakeAllAssists wakes all currently blocked assists. This is used
612 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
613 // new assists from going to sleep after this point.
614 func gcWakeAllAssists() {
615 lock(&work.assistQueue.lock)
616 list := work.assistQueue.q.popList()
618 unlock(&work.assistQueue.lock)
621 // gcParkAssist puts the current goroutine on the assist queue and parks.
623 // gcParkAssist reports whether the assist is now satisfied. If it
624 // returns false, the caller must retry the assist.
625 func gcParkAssist() bool {
626 lock(&work.assistQueue.lock)
627 // If the GC cycle finished while we were getting the lock,
628 // exit the assist. The cycle can't finish while we hold the
630 if atomic.Load(&gcBlackenEnabled) == 0 {
631 unlock(&work.assistQueue.lock)
636 oldList := work.assistQueue.q
637 work.assistQueue.q.pushBack(gp)
639 // Recheck for background credit now that this G is in
640 // the queue, but can still back out. This avoids a
641 // race in case background marking has flushed more
642 // credit since we checked above.
643 if gcController.bgScanCredit.Load() > 0 {
644 work.assistQueue.q = oldList
645 if oldList.tail != 0 {
646 oldList.tail.ptr().schedlink.set(nil)
648 unlock(&work.assistQueue.lock)
652 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceBlockGCMarkAssist, 2)
656 // gcFlushBgCredit flushes scanWork units of background scan work
657 // credit. This first satisfies blocked assists on the
658 // work.assistQueue and then flushes any remaining credit to
659 // gcController.bgScanCredit.
661 // Write barriers are disallowed because this is used by gcDrain after
662 // it has ensured that all work is drained and this must preserve that
665 //go:nowritebarrierrec
666 func gcFlushBgCredit(scanWork int64) {
667 if work.assistQueue.q.empty() {
668 // Fast path; there are no blocked assists. There's a
669 // small window here where an assist may add itself to
670 // the blocked queue and park. If that happens, we'll
671 // just get it on the next flush.
672 gcController.bgScanCredit.Add(scanWork)
676 assistBytesPerWork := gcController.assistBytesPerWork.Load()
677 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
679 lock(&work.assistQueue.lock)
680 for !work.assistQueue.q.empty() && scanBytes > 0 {
681 gp := work.assistQueue.q.pop()
682 // Note that gp.gcAssistBytes is negative because gp
683 // is in debt. Think carefully about the signs below.
684 if scanBytes+gp.gcAssistBytes >= 0 {
685 // Satisfy this entire assist debt.
686 scanBytes += gp.gcAssistBytes
688 // It's important that we *not* put gp in
689 // runnext. Otherwise, it's possible for user
690 // code to exploit the GC worker's high
691 // scheduler priority to get itself always run
692 // before other goroutines and always in the
693 // fresh quantum started by GC.
696 // Partially satisfy this assist.
697 gp.gcAssistBytes += scanBytes
699 // As a heuristic, we move this assist to the
700 // back of the queue so that large assists
701 // can't clog up the assist queue and
702 // substantially delay small assists.
703 work.assistQueue.q.pushBack(gp)
709 // Convert from scan bytes back to work.
710 assistWorkPerByte := gcController.assistWorkPerByte.Load()
711 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
712 gcController.bgScanCredit.Add(scanWork)
714 unlock(&work.assistQueue.lock)
717 // scanstack scans gp's stack, greying all pointers found on the stack.
719 // Returns the amount of scan work performed, but doesn't update
720 // gcController.stackScanWork or flush any credit. Any background credit produced
721 // by this function should be flushed by its caller. scanstack itself can't
722 // safely flush because it may result in trying to wake up a goroutine that
723 // was just scanned, resulting in a self-deadlock.
725 // scanstack will also shrink the stack if it is safe to do so. If it
726 // is not, it schedules a stack shrink for the next synchronous safe
729 // scanstack is marked go:systemstack because it must not be preempted
730 // while using a workbuf.
734 func scanstack(gp *g, gcw *gcWork) int64 {
735 if readgstatus(gp)&_Gscan == 0 {
736 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
737 throw("scanstack - bad status")
740 switch readgstatus(gp) &^ _Gscan {
742 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
743 throw("mark - bad status")
747 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
748 throw("scanstack: goroutine not stopped")
749 case _Grunnable, _Gsyscall, _Gwaiting:
754 throw("can't scan our own stack")
757 // scannedSize is the amount of work we'll be reporting.
759 // It is less than the allocated size (which is hi-lo).
761 if gp.syscallsp != 0 {
762 sp = gp.syscallsp // If in a system call this is the stack pointer (gp.sched.sp can be 0 in this case on Windows).
766 scannedSize := gp.stack.hi - sp
768 // Keep statistics for initial stack size calculation.
769 // Note that this accumulates the scanned size, not the allocated size.
770 p := getg().m.p.ptr()
771 p.scannedStackSize += uint64(scannedSize)
774 if isShrinkStackSafe(gp) {
775 // Shrink the stack if not much of it is being used.
778 // Otherwise, shrink the stack at the next sync safe point.
779 gp.preemptShrink = true
782 var state stackScanState
783 state.stack = gp.stack
786 println("stack trace goroutine", gp.goid)
789 if debugScanConservative && gp.asyncSafePoint {
790 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
793 // Scan the saved context register. This is effectively a live
794 // register that gets moved back and forth between the
795 // register and sched.ctxt without a write barrier.
796 if gp.sched.ctxt != nil {
797 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
800 // Scan the stack. Accumulate a list of stack objects.
802 for u.init(gp, 0); u.valid(); u.next() {
803 scanframeworker(&u.frame, &state, gcw)
806 // Find additional pointers that point into the stack from the heap.
807 // Currently this includes defers and panics. See also function copystack.
809 // Find and trace other pointers in defer records.
810 for d := gp._defer; d != nil; d = d.link {
812 // Scan the func value, which could be a stack allocated closure.
814 scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
817 // The link field of a stack-allocated defer record might point
818 // to a heap-allocated defer record. Keep that heap record live.
819 scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
821 // Retain defers records themselves.
822 // Defer records might not be reachable from the G through regular heap
823 // tracing because the defer linked list might weave between the stack and the heap.
825 scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
828 if gp._panic != nil {
829 // Panics are always stack allocated.
830 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
833 // Find and scan all reachable stack objects.
835 // The state's pointer queue prioritizes precise pointers over
836 // conservative pointers so that we'll prefer scanning stack
837 // objects precisely.
840 p, conservative := state.getPtr()
844 obj := state.findObject(p)
850 // We've already scanned this object.
853 obj.setRecord(nil) // Don't scan it again.
856 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
858 print(" (conservative)")
866 // This path is pretty unlikely, an object large enough
867 // to have a GC program allocated on the stack.
868 // We need some space to unpack the program into a straight
869 // bitmask, which we allocate/free here.
870 // TODO: it would be nice if there were a way to run a GC
871 // program without having to store all its bits. We'd have
872 // to change from a Lempel-Ziv style program to something else.
873 // Or we can forbid putting objects on stacks if they require
874 // a gc program (see issue 27447).
875 s = materializeGCProg(r.ptrdata(), gcdata)
876 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
879 b := state.stack.lo + uintptr(obj.off)
881 scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
883 scanblock(b, r.ptrdata(), gcdata, gcw, &state)
887 dematerializeGCProg(s)
891 // Deallocate object buffers.
892 // (Pointer buffers were all deallocated in the loop above.)
893 for state.head != nil {
897 for i := 0; i < x.nobj; i++ {
899 if obj.r == nil { // reachable
902 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
903 // Note: not necessarily really dead - only reachable-from-ptr dead.
907 putempty((*workbuf)(unsafe.Pointer(x)))
909 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
910 throw("remaining pointer buffers")
912 return int64(scannedSize)
915 // Scan a stack frame: local variables and function arguments/results.
918 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
919 if _DebugGC > 1 && frame.continpc != 0 {
920 print("scanframe ", funcname(frame.fn), "\n")
923 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == abi.FuncID_asyncPreempt
924 isDebugCall := frame.fn.valid() && frame.fn.funcID == abi.FuncID_debugCallV2
925 if state.conservative || isAsyncPreempt || isDebugCall {
926 if debugScanConservative {
927 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
930 // Conservatively scan the frame. Unlike the precise
931 // case, this includes the outgoing argument space
932 // since we may have stopped while this function was
933 // setting up a call.
935 // TODO: We could narrow this down if the compiler
936 // produced a single map per function of stack slots
937 // and registers that ever contain a pointer.
939 size := frame.varp - frame.sp
941 scanConservative(frame.sp, size, nil, gcw, state)
945 // Scan arguments to this frame.
946 if n := frame.argBytes(); n != 0 {
947 // TODO: We could pass the entry argument map
948 // to narrow this down further.
949 scanConservative(frame.argp, n, nil, gcw, state)
952 if isAsyncPreempt || isDebugCall {
953 // This function's frame contained the
954 // registers for the asynchronously stopped
955 // parent frame. Scan the parent
957 state.conservative = true
959 // We only wanted to scan those two frames
960 // conservatively. Clear the flag for future
962 state.conservative = false
967 locals, args, objs := frame.getStackMap(false)
969 // Scan local variables if stack frame has been allocated.
971 size := uintptr(locals.n) * goarch.PtrSize
972 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
977 scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
980 // Add all stack objects to the stack object list.
982 // varp is 0 for defers, where there are no locals.
983 // In that case, there can't be a pointer to its args, either.
984 // (And all args would be scanned above anyway.)
985 for i := range objs {
988 base := frame.varp // locals base pointer
990 base = frame.argp // arguments and return values base pointer
992 ptr := base + uintptr(off)
994 // object hasn't been allocated in the frame yet.
998 println("stkobj at", hex(ptr), "of size", obj.size)
1000 state.addObject(ptr, obj)
1005 type gcDrainFlags int
1008 gcDrainUntilPreempt gcDrainFlags = 1 << iota
1009 gcDrainFlushBgCredit
1014 // gcDrain scans roots and objects in work buffers, blackening grey
1015 // objects until it is unable to get more work. It may return before
1016 // GC is done; it's the caller's responsibility to balance work from
1019 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
1022 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
1025 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
1026 // pollFractionalWorkerExit() returns true. This implies
1029 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
1030 // credit to gcController.bgScanCredit every gcCreditSlack units of
1033 // gcDrain will always return if there is a pending STW.
1036 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
1037 if !writeBarrier.needed {
1038 throw("gcDrain phase incorrect")
1042 preemptible := flags&gcDrainUntilPreempt != 0
1043 flushBgCredit := flags&gcDrainFlushBgCredit != 0
1044 idle := flags&gcDrainIdle != 0
1046 initScanWork := gcw.heapScanWork
1048 // checkWork is the scan work before performing the next
1049 // self-preempt check.
1050 checkWork := int64(1<<63 - 1)
1051 var check func() bool
1052 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
1053 checkWork = initScanWork + drainCheckThreshold
1056 } else if flags&gcDrainFractional != 0 {
1057 check = pollFractionalWorkerExit
1061 // Drain root marking jobs.
1062 if work.markrootNext < work.markrootJobs {
1063 // Stop if we're preemptible or if someone wants to STW.
1064 for !(gp.preempt && (preemptible || sched.gcwaiting.Load())) {
1065 job := atomic.Xadd(&work.markrootNext, +1) - 1
1066 if job >= work.markrootJobs {
1069 markroot(gcw, job, flushBgCredit)
1070 if check != nil && check() {
1076 // Drain heap marking jobs.
1077 // Stop if we're preemptible or if someone wants to STW.
1078 for !(gp.preempt && (preemptible || sched.gcwaiting.Load())) {
1079 // Try to keep work available on the global queue. We used to
1080 // check if there were waiting workers, but it's better to
1081 // just keep work available than to make workers wait. In the
1082 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1088 b := gcw.tryGetFast()
1092 // Flush the write barrier
1093 // buffer; this may create
1100 // Unable to get work.
1105 // Flush background scan work credit to the global
1106 // account if we've accumulated enough locally so
1107 // mutator assists can draw on it.
1108 if gcw.heapScanWork >= gcCreditSlack {
1109 gcController.heapScanWork.Add(gcw.heapScanWork)
1111 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1114 checkWork -= gcw.heapScanWork
1115 gcw.heapScanWork = 0
1118 checkWork += drainCheckThreshold
1119 if check != nil && check() {
1127 // Flush remaining scan work credit.
1128 if gcw.heapScanWork > 0 {
1129 gcController.heapScanWork.Add(gcw.heapScanWork)
1131 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1133 gcw.heapScanWork = 0
1137 // gcDrainN blackens grey objects until it has performed roughly
1138 // scanWork units of scan work or the G is preempted. This is
1139 // best-effort, so it may perform less work if it fails to get a work
1140 // buffer. Otherwise, it will perform at least n units of work, but
1141 // may perform more because scanning is always done in whole object
1142 // increments. It returns the amount of scan work performed.
1144 // The caller goroutine must be in a preemptible state (e.g.,
1145 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1146 // consequence, this must be called on the system stack.
1150 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1151 if !writeBarrier.needed {
1152 throw("gcDrainN phase incorrect")
1155 // There may already be scan work on the gcw, which we don't
1156 // want to claim was done by this call.
1157 workFlushed := -gcw.heapScanWork
1159 // In addition to backing out because of a preemption, back out
1160 // if the GC CPU limiter is enabled.
1162 for !gp.preempt && !gcCPULimiter.limiting() && workFlushed+gcw.heapScanWork < scanWork {
1163 // See gcDrain comment.
1168 b := gcw.tryGetFast()
1172 // Flush the write barrier buffer;
1173 // this may create more work.
1180 // Try to do a root job.
1181 if work.markrootNext < work.markrootJobs {
1182 job := atomic.Xadd(&work.markrootNext, +1) - 1
1183 if job < work.markrootJobs {
1184 workFlushed += markroot(gcw, job, false)
1188 // No heap or root jobs.
1194 // Flush background scan work credit.
1195 if gcw.heapScanWork >= gcCreditSlack {
1196 gcController.heapScanWork.Add(gcw.heapScanWork)
1197 workFlushed += gcw.heapScanWork
1198 gcw.heapScanWork = 0
1202 // Unlike gcDrain, there's no need to flush remaining work
1203 // here because this never flushes to bgScanCredit and
1204 // gcw.dispose will flush any remaining work to scanWork.
1206 return workFlushed + gcw.heapScanWork
1209 // scanblock scans b as scanobject would, but using an explicit
1210 // pointer bitmap instead of the heap bitmap.
1212 // This is used to scan non-heap roots, so it does not update
1213 // gcw.bytesMarked or gcw.heapScanWork.
1215 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1218 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1219 // Use local copies of original parameters, so that a stack trace
1220 // due to one of the throws below shows the original block
1225 for i := uintptr(0); i < n; {
1226 // Find bits for the next word.
1227 bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
1229 i += goarch.PtrSize * 8
1232 for j := 0; j < 8 && i < n; j++ {
1234 // Same work as in scanobject; see comments there.
1235 p := *(*uintptr)(unsafe.Pointer(b + i))
1237 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1238 greyobject(obj, b, i, span, gcw, objIndex)
1239 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1240 stk.putPtr(p, false)
1250 // scanobject scans the object starting at b, adding pointers to gcw.
1251 // b must point to the beginning of a heap object or an oblet.
1252 // scanobject consults the GC bitmap for the pointer mask and the
1253 // spans for the size of the object.
1256 func scanobject(b uintptr, gcw *gcWork) {
1257 // Prefetch object before we scan it.
1259 // This will overlap fetching the beginning of the object with initial
1260 // setup before we start scanning the object.
1263 // Find the bits for b and the size of the object at b.
1265 // b is either the beginning of an object, in which case this
1266 // is the size of the object to scan, or it points to an
1267 // oblet, in which case we compute the size to scan below.
1268 s := spanOfUnchecked(b)
1271 throw("scanobject n == 0")
1273 if s.spanclass.noscan() {
1274 // Correctness-wise this is ok, but it's inefficient
1275 // if noscan objects reach here.
1276 throw("scanobject of a noscan object")
1279 if n > maxObletBytes {
1280 // Large object. Break into oblets for better
1281 // parallelism and lower latency.
1283 // Enqueue the other oblets to scan later.
1284 // Some oblets may be in b's scalar tail, but
1285 // these will be marked as "no more pointers",
1286 // so we'll drop out immediately when we go to
1288 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1289 if !gcw.putFast(oblet) {
1295 // Compute the size of the oblet. Since this object
1296 // must be a large object, s.base() is the beginning
1298 n = s.base() + s.elemsize - b
1299 if n > maxObletBytes {
1304 hbits := heapBitsForAddr(b, n)
1305 var scanSize uintptr
1308 if hbits, addr = hbits.nextFast(); addr == 0 {
1309 if hbits, addr = hbits.next(); addr == 0 {
1314 // Keep track of farthest pointer we found, so we can
1315 // update heapScanWork. TODO: is there a better metric,
1316 // now that we can skip scalar portions pretty efficiently?
1317 scanSize = addr - b + goarch.PtrSize
1319 // Work here is duplicated in scanblock and above.
1320 // If you make changes here, make changes there too.
1321 obj := *(*uintptr)(unsafe.Pointer(addr))
1323 // At this point we have extracted the next potential pointer.
1324 // Quickly filter out nil and pointers back to the current object.
1325 if obj != 0 && obj-b >= n {
1326 // Test if obj points into the Go heap and, if so,
1329 // Note that it's possible for findObject to
1330 // fail if obj points to a just-allocated heap
1331 // object because of a race with growing the
1332 // heap. In this case, we know the object was
1333 // just allocated and hence will be marked by
1334 // allocation itself.
1335 if obj, span, objIndex := findObject(obj, b, addr-b); obj != 0 {
1336 greyobject(obj, b, addr-b, span, gcw, objIndex)
1340 gcw.bytesMarked += uint64(n)
1341 gcw.heapScanWork += int64(scanSize)
1344 // scanConservative scans block [b, b+n) conservatively, treating any
1345 // pointer-like value in the block as a pointer.
1347 // If ptrmask != nil, only words that are marked in ptrmask are
1348 // considered as potential pointers.
1350 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1351 // and may contain pointers to stack objects.
1352 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1353 if debugScanConservative {
1355 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1356 hexdumpWords(b, b+n, func(p uintptr) byte {
1358 word := (p - b) / goarch.PtrSize
1359 bits := *addb(ptrmask, word/8)
1360 if (bits>>(word%8))&1 == 0 {
1365 val := *(*uintptr)(unsafe.Pointer(p))
1366 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1370 span := spanOfHeap(val)
1374 idx := span.objIndex(val)
1375 if span.isFree(idx) {
1383 for i := uintptr(0); i < n; i += goarch.PtrSize {
1385 word := i / goarch.PtrSize
1386 bits := *addb(ptrmask, word/8)
1388 // Skip 8 words (the loop increment will do the 8th)
1390 // This must be the first time we've
1391 // seen this word of ptrmask, so i
1392 // must be 8-word-aligned, but check
1393 // our reasoning just in case.
1394 if i%(goarch.PtrSize*8) != 0 {
1395 throw("misaligned mask")
1397 i += goarch.PtrSize*8 - goarch.PtrSize
1400 if (bits>>(word%8))&1 == 0 {
1405 val := *(*uintptr)(unsafe.Pointer(b + i))
1407 // Check if val points into the stack.
1408 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1409 // val may point to a stack object. This
1410 // object may be dead from last cycle and
1411 // hence may contain pointers to unallocated
1412 // objects, but unlike heap objects we can't
1413 // tell if it's already dead. Hence, if all
1414 // pointers to this object are from
1415 // conservative scanning, we have to scan it
1416 // defensively, too.
1417 state.putPtr(val, true)
1421 // Check if val points to a heap span.
1422 span := spanOfHeap(val)
1427 // Check if val points to an allocated object.
1428 idx := span.objIndex(val)
1429 if span.isFree(idx) {
1433 // val points to an allocated object. Mark it.
1434 obj := span.base() + idx*span.elemsize
1435 greyobject(obj, b, i, span, gcw, idx)
1439 // Shade the object if it isn't already.
1440 // The object is not nil and known to be in the heap.
1441 // Preemption must be disabled.
1444 func shade(b uintptr) {
1445 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1446 gcw := &getg().m.p.ptr().gcw
1447 greyobject(obj, 0, 0, span, gcw, objIndex)
1451 // obj is the start of an object with mark mbits.
1452 // If it isn't already marked, mark it and enqueue into gcw.
1453 // base and off are for debugging only and could be removed.
1455 // See also wbBufFlush1, which partially duplicates this logic.
1457 //go:nowritebarrierrec
1458 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1459 // obj should be start of allocation, and so must be at least pointer-aligned.
1460 if obj&(goarch.PtrSize-1) != 0 {
1461 throw("greyobject: obj not pointer-aligned")
1463 mbits := span.markBitsForIndex(objIndex)
1466 if setCheckmark(obj, base, off, mbits) {
1471 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1472 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1473 gcDumpObject("base", base, off)
1474 gcDumpObject("obj", obj, ^uintptr(0))
1475 getg().m.traceback = 2
1476 throw("marking free object")
1479 // If marked we have nothing to do.
1480 if mbits.isMarked() {
1486 arena, pageIdx, pageMask := pageIndexOf(span.base())
1487 if arena.pageMarks[pageIdx]&pageMask == 0 {
1488 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1491 // If this is a noscan object, fast-track it to black
1492 // instead of greying it.
1493 if span.spanclass.noscan() {
1494 gcw.bytesMarked += uint64(span.elemsize)
1499 // We're adding obj to P's local workbuf, so it's likely
1500 // this object will be processed soon by the same P.
1501 // Even if the workbuf gets flushed, there will likely still be
1502 // some benefit on platforms with inclusive shared caches.
1504 // Queue the obj for scanning.
1505 if !gcw.putFast(obj) {
1510 // gcDumpObject dumps the contents of obj for debugging and marks the
1511 // field at byte offset off in obj.
1512 func gcDumpObject(label string, obj, off uintptr) {
1514 print(label, "=", hex(obj))
1519 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1520 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1521 print(mSpanStateNames[state], "\n")
1523 print("unknown(", state, ")\n")
1528 if s.state.get() == mSpanManual && size == 0 {
1529 // We're printing something from a stack frame. We
1530 // don't know how big it is, so just show up to an
1532 size = off + goarch.PtrSize
1534 for i := uintptr(0); i < size; i += goarch.PtrSize {
1535 // For big objects, just print the beginning (because
1536 // that usually hints at the object's type) and the
1537 // fields around off.
1538 if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
1546 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1557 // gcmarknewobject marks a newly allocated object black. obj must
1558 // not contain any non-nil pointers.
1560 // This is nosplit so it can manipulate a gcWork without preemption.
1564 func gcmarknewobject(span *mspan, obj, size uintptr) {
1565 if useCheckmark { // The world should be stopped so this should not happen.
1566 throw("gcmarknewobject called while doing checkmark")
1570 objIndex := span.objIndex(obj)
1571 span.markBitsForIndex(objIndex).setMarked()
1574 arena, pageIdx, pageMask := pageIndexOf(span.base())
1575 if arena.pageMarks[pageIdx]&pageMask == 0 {
1576 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1579 gcw := &getg().m.p.ptr().gcw
1580 gcw.bytesMarked += uint64(size)
1583 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1585 // The world must be stopped.
1586 func gcMarkTinyAllocs() {
1587 assertWorldStopped()
1589 for _, p := range allp {
1591 if c == nil || c.tiny == 0 {
1594 _, span, objIndex := findObject(c.tiny, 0, 0)
1596 greyobject(c.tiny, 0, 0, span, gcw, objIndex)