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 "runtime/internal/atomic"
12 "runtime/internal/sys"
17 fixedRootFinalizers = iota
21 // rootBlockBytes is the number of bytes to scan per data or
23 rootBlockBytes = 256 << 10
25 // maxObletBytes is the maximum bytes of an object to scan at
26 // once. Larger objects will be split up into "oblets" of at
27 // most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
28 // scan preemption at ~100 µs.
30 // This must be > _MaxSmallSize so that the object base is the
32 maxObletBytes = 128 << 10
34 // drainCheckThreshold specifies how many units of work to do
35 // between self-preemption checks in gcDrain. Assuming a scan
36 // rate of 1 MB/ms, this is ~100 µs. Lower values have higher
37 // overhead in the scan loop (the scheduler check may perform
38 // a syscall, so its overhead is nontrivial). Higher values
39 // make the system less responsive to incoming work.
40 drainCheckThreshold = 100000
42 // pagesPerSpanRoot indicates how many pages to scan from a span root
43 // at a time. Used by special root marking.
45 // Higher values improve throughput by increasing locality, but
46 // increase the minimum latency of a marking operation.
48 // Must be a multiple of the pageInUse bitmap element size and
49 // must also evenly divide pagesPerArena.
50 pagesPerSpanRoot = 512
53 // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
54 // some miscellany) and initializes scanning-related state.
56 // The world must be stopped.
57 func gcMarkRootPrepare() {
60 // Compute how many data and BSS root blocks there are.
61 nBlocks := func(bytes uintptr) int {
62 return int(divRoundUp(bytes, rootBlockBytes))
69 for _, datap := range activeModules() {
70 nDataRoots := nBlocks(datap.edata - datap.data)
71 if nDataRoots > work.nDataRoots {
72 work.nDataRoots = nDataRoots
76 for _, datap := range activeModules() {
77 nBSSRoots := nBlocks(datap.ebss - datap.bss)
78 if nBSSRoots > work.nBSSRoots {
79 work.nBSSRoots = nBSSRoots
83 // Scan span roots for finalizer specials.
85 // We depend on addfinalizer to mark objects that get
86 // finalizers after root marking.
88 // We're going to scan the whole heap (that was available at the time the
89 // mark phase started, i.e. markArenas) for in-use spans which have specials.
91 // Break up the work into arenas, and further into chunks.
93 // Snapshot allArenas as markArenas. This snapshot is safe because allArenas
95 mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
96 work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
100 // Gs may be created after this point, but it's okay that we
101 // ignore them because they begin life without any roots, so
102 // there's nothing to scan, and any roots they create during
103 // the concurrent phase will be caught by the write barrier.
104 work.stackRoots = allGsSnapshot()
105 work.nStackRoots = len(work.stackRoots)
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
197 if i < work.baseStacks || work.baseEnd <= i {
199 print("runtime: markroot index ", i, " not in stack roots range [", work.baseStacks, ", ", work.baseEnd, ")\n")
200 throw("markroot: bad index")
202 gp := work.stackRoots[i-work.baseStacks]
204 // remember when we've first observed the G blocked
205 // needed only to output in traceback
206 status := readgstatus(gp) // We are not in a scan state
207 if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
208 gp.waitsince = work.tstart
211 // scanstack must be done on the system stack in case
212 // we're trying to scan our own stack.
214 // If this is a self-scan, put the user G in
215 // _Gwaiting to prevent self-deadlock. It may
216 // already be in _Gwaiting if this is a mark
217 // worker or we're in mark termination.
218 userG := getg().m.curg
219 selfScan := gp == userG && readgstatus(userG) == _Grunning
221 casGToWaiting(userG, _Grunning, waitReasonGarbageCollectionScan)
224 // TODO: suspendG blocks (and spins) until gp
225 // stops, which may take a while for
226 // running goroutines. Consider doing this in
227 // two phases where the first is non-blocking:
228 // we scan the stacks we can and ask running
229 // goroutines to scan themselves; and the
231 stopped := suspendG(gp)
237 throw("g already scanned")
239 workDone += scanstack(gp, gcw)
244 casgstatus(userG, _Gwaiting, _Grunning)
248 if workCounter != nil && workDone != 0 {
249 workCounter.Add(workDone)
251 gcFlushBgCredit(workDone)
257 // markrootBlock scans the shard'th shard of the block of memory [b0,
258 // b0+n0), with the given pointer mask.
260 // Returns the amount of work done.
263 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) int64 {
264 if rootBlockBytes%(8*goarch.PtrSize) != 0 {
265 // This is necessary to pick byte offsets in ptrmask0.
266 throw("rootBlockBytes must be a multiple of 8*ptrSize")
269 // Note that if b0 is toward the end of the address space,
270 // then b0 + rootBlockBytes might wrap around.
271 // These tests are written to avoid any possible overflow.
272 off := uintptr(shard) * rootBlockBytes
277 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
278 n := uintptr(rootBlockBytes)
284 scanblock(b, n, ptrmask, gcw, nil)
288 // markrootFreeGStacks frees stacks of dead Gs.
290 // This does not free stacks of dead Gs cached on Ps, but having a few
291 // cached stacks around isn't a problem.
292 func markrootFreeGStacks() {
293 // Take list of dead Gs with stacks.
294 lock(&sched.gFree.lock)
295 list := sched.gFree.stack
296 sched.gFree.stack = gList{}
297 unlock(&sched.gFree.lock)
303 q := gQueue{list.head, list.head}
304 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
308 // Manipulate the queue directly since the Gs are
309 // already all linked the right way.
313 // Put Gs back on the free list.
314 lock(&sched.gFree.lock)
315 sched.gFree.noStack.pushAll(q)
316 unlock(&sched.gFree.lock)
319 // markrootSpans marks roots for one shard of markArenas.
322 func markrootSpans(gcw *gcWork, shard int) {
323 // Objects with finalizers have two GC-related invariants:
325 // 1) Everything reachable from the object must be marked.
326 // This ensures that when we pass the object to its finalizer,
327 // everything the finalizer can reach will be retained.
329 // 2) Finalizer specials (which are not in the garbage
330 // collected heap) are roots. In practice, this means the fn
331 // field must be scanned.
332 sg := mheap_.sweepgen
334 // Find the arena and page index into that arena for this shard.
335 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
336 ha := mheap_.arenas[ai.l1()][ai.l2()]
337 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
339 // Construct slice of bitmap which we'll iterate over.
340 specialsbits := ha.pageSpecials[arenaPage/8:]
341 specialsbits = specialsbits[:pagesPerSpanRoot/8]
342 for i := range specialsbits {
343 // Find set bits, which correspond to spans with specials.
344 specials := atomic.Load8(&specialsbits[i])
348 for j := uint(0); j < 8; j++ {
349 if specials&(1<<j) == 0 {
352 // Find the span for this bit.
354 // This value is guaranteed to be non-nil because having
355 // specials implies that the span is in-use, and since we're
356 // currently marking we can be sure that we don't have to worry
357 // about the span being freed and re-used.
358 s := ha.spans[arenaPage+uint(i)*8+j]
360 // The state must be mSpanInUse if the specials bit is set, so
361 // sanity check that.
362 if state := s.state.get(); state != mSpanInUse {
363 print("s.state = ", state, "\n")
364 throw("non in-use span found with specials bit set")
366 // Check that this span was swept (it may be cached or uncached).
367 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
368 // sweepgen was updated (+2) during non-checkmark GC pass
369 print("sweep ", s.sweepgen, " ", sg, "\n")
370 throw("gc: unswept span")
373 // Lock the specials to prevent a special from being
374 // removed from the list while we're traversing it.
376 for sp := s.specials; sp != nil; sp = sp.next {
377 if sp.kind != _KindSpecialFinalizer {
380 // don't mark finalized object, but scan it so we
381 // retain everything it points to.
382 spf := (*specialfinalizer)(unsafe.Pointer(sp))
383 // A finalizer can be set for an inner byte of an object, find object beginning.
384 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
386 // Mark everything that can be reached from
387 // the object (but *not* the object itself or
388 // we'll never collect it).
389 if !s.spanclass.noscan() {
393 // The special itself is a root.
394 scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
396 unlock(&s.speciallock)
401 // gcAssistAlloc performs GC work to make gp's assist debt positive.
402 // gp must be the calling user goroutine.
404 // This must be called with preemption enabled.
405 func gcAssistAlloc(gp *g) {
406 // Don't assist in non-preemptible contexts. These are
407 // generally fragile and won't allow the assist to block.
408 if getg() == gp.m.g0 {
411 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
417 if gcCPULimiter.limiting() {
418 // If the CPU limiter is enabled, intentionally don't
419 // assist to reduce the amount of CPU time spent in the GC.
421 traceGCMarkAssistDone()
425 // Compute the amount of scan work we need to do to make the
426 // balance positive. When the required amount of work is low,
427 // we over-assist to build up credit for future allocations
428 // and amortize the cost of assisting.
429 assistWorkPerByte := gcController.assistWorkPerByte.Load()
430 assistBytesPerWork := gcController.assistBytesPerWork.Load()
431 debtBytes := -gp.gcAssistBytes
432 scanWork := int64(assistWorkPerByte * float64(debtBytes))
433 if scanWork < gcOverAssistWork {
434 scanWork = gcOverAssistWork
435 debtBytes = int64(assistBytesPerWork * float64(scanWork))
438 // Steal as much credit as we can from the background GC's
439 // scan credit. This is racy and may drop the background
440 // credit below 0 if two mutators steal at the same time. This
441 // will just cause steals to fail until credit is accumulated
442 // again, so in the long run it doesn't really matter, but we
443 // do have to handle the negative credit case.
444 bgScanCredit := gcController.bgScanCredit.Load()
446 if bgScanCredit > 0 {
447 if bgScanCredit < scanWork {
448 stolen = bgScanCredit
449 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
452 gp.gcAssistBytes += debtBytes
454 gcController.bgScanCredit.Add(-stolen)
459 // We were able to steal all of the credit we
462 traceGCMarkAssistDone()
468 if trace.enabled && !traced {
470 traceGCMarkAssistStart()
473 // Perform assist work
475 gcAssistAlloc1(gp, scanWork)
476 // The user stack may have moved, so this can't touch
477 // anything on it until it returns from systemstack.
480 completed := gp.param != nil
486 if gp.gcAssistBytes < 0 {
487 // We were unable steal enough credit or perform
488 // enough work to pay off the assist debt. We need to
489 // do one of these before letting the mutator allocate
490 // more to prevent over-allocation.
492 // If this is because we were preempted, reschedule
493 // and try some more.
499 // Add this G to an assist queue and park. When the GC
500 // has more background credit, it will satisfy queued
501 // assists before flushing to the global credit pool.
503 // Note that this does *not* get woken up when more
504 // work is added to the work list. The theory is that
505 // there wasn't enough work to do anyway, so we might
506 // as well let background marking take care of the
507 // work that is available.
512 // At this point either background GC has satisfied
513 // this G's assist debt, or the GC cycle is over.
516 traceGCMarkAssistDone()
520 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
521 // stack. This is a separate function to make it easier to see that
522 // we're not capturing anything from the user stack, since the user
523 // stack may move while we're in this function.
525 // gcAssistAlloc1 indicates whether this assist completed the mark
526 // phase by setting gp.param to non-nil. This can't be communicated on
527 // the stack since it may move.
530 func gcAssistAlloc1(gp *g, scanWork int64) {
531 // Clear the flag indicating that this assist completed the
535 if atomic.Load(&gcBlackenEnabled) == 0 {
536 // The gcBlackenEnabled check in malloc races with the
537 // store that clears it but an atomic check in every malloc
538 // would be a performance hit.
539 // Instead we recheck it here on the non-preemptable system
540 // stack to determine if we should perform an assist.
542 // GC is done, so ignore any remaining debt.
546 // Track time spent in this assist. Since we're on the
547 // system stack, this is non-preemptible, so we can
548 // just measure start and end time.
550 // Limiter event tracking might be disabled if we end up here
551 // while on a mark worker.
552 startTime := nanotime()
553 trackLimiterEvent := gp.m.p.ptr().limiterEvent.start(limiterEventMarkAssist, startTime)
555 decnwait := atomic.Xadd(&work.nwait, -1)
556 if decnwait == work.nproc {
557 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
558 throw("nwait > work.nprocs")
561 // gcDrainN requires the caller to be preemptible.
562 casGToWaiting(gp, _Grunning, waitReasonGCAssistMarking)
564 // drain own cached work first in the hopes that it
565 // will be more cache friendly.
566 gcw := &getg().m.p.ptr().gcw
567 workDone := gcDrainN(gcw, scanWork)
569 casgstatus(gp, _Gwaiting, _Grunning)
571 // Record that we did this much scan work.
573 // Back out the number of bytes of assist credit that
574 // this scan work counts for. The "1+" is a poor man's
575 // round-up, to ensure this adds credit even if
576 // assistBytesPerWork is very low.
577 assistBytesPerWork := gcController.assistBytesPerWork.Load()
578 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
580 // If this is the last worker and we ran out of work,
581 // signal a completion point.
582 incnwait := atomic.Xadd(&work.nwait, +1)
583 if incnwait > work.nproc {
584 println("runtime: work.nwait=", incnwait,
585 "work.nproc=", work.nproc)
586 throw("work.nwait > work.nproc")
589 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
590 // This has reached a background completion point. Set
591 // gp.param to a non-nil value to indicate this. It
592 // doesn't matter what we set it to (it just has to be
594 gp.param = unsafe.Pointer(gp)
597 duration := now - startTime
599 pp.gcAssistTime += duration
600 if trackLimiterEvent {
601 pp.limiterEvent.stop(limiterEventMarkAssist, now)
603 if pp.gcAssistTime > gcAssistTimeSlack {
604 gcController.assistTime.Add(pp.gcAssistTime)
605 gcCPULimiter.update(now)
610 // gcWakeAllAssists wakes all currently blocked assists. This is used
611 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
612 // new assists from going to sleep after this point.
613 func gcWakeAllAssists() {
614 lock(&work.assistQueue.lock)
615 list := work.assistQueue.q.popList()
617 unlock(&work.assistQueue.lock)
620 // gcParkAssist puts the current goroutine on the assist queue and parks.
622 // gcParkAssist reports whether the assist is now satisfied. If it
623 // returns false, the caller must retry the assist.
624 func gcParkAssist() bool {
625 lock(&work.assistQueue.lock)
626 // If the GC cycle finished while we were getting the lock,
627 // exit the assist. The cycle can't finish while we hold the
629 if atomic.Load(&gcBlackenEnabled) == 0 {
630 unlock(&work.assistQueue.lock)
635 oldList := work.assistQueue.q
636 work.assistQueue.q.pushBack(gp)
638 // Recheck for background credit now that this G is in
639 // the queue, but can still back out. This avoids a
640 // race in case background marking has flushed more
641 // credit since we checked above.
642 if gcController.bgScanCredit.Load() > 0 {
643 work.assistQueue.q = oldList
644 if oldList.tail != 0 {
645 oldList.tail.ptr().schedlink.set(nil)
647 unlock(&work.assistQueue.lock)
651 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
655 // gcFlushBgCredit flushes scanWork units of background scan work
656 // credit. This first satisfies blocked assists on the
657 // work.assistQueue and then flushes any remaining credit to
658 // gcController.bgScanCredit.
660 // Write barriers are disallowed because this is used by gcDrain after
661 // it has ensured that all work is drained and this must preserve that
664 //go:nowritebarrierrec
665 func gcFlushBgCredit(scanWork int64) {
666 if work.assistQueue.q.empty() {
667 // Fast path; there are no blocked assists. There's a
668 // small window here where an assist may add itself to
669 // the blocked queue and park. If that happens, we'll
670 // just get it on the next flush.
671 gcController.bgScanCredit.Add(scanWork)
675 assistBytesPerWork := gcController.assistBytesPerWork.Load()
676 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
678 lock(&work.assistQueue.lock)
679 for !work.assistQueue.q.empty() && scanBytes > 0 {
680 gp := work.assistQueue.q.pop()
681 // Note that gp.gcAssistBytes is negative because gp
682 // is in debt. Think carefully about the signs below.
683 if scanBytes+gp.gcAssistBytes >= 0 {
684 // Satisfy this entire assist debt.
685 scanBytes += gp.gcAssistBytes
687 // It's important that we *not* put gp in
688 // runnext. Otherwise, it's possible for user
689 // code to exploit the GC worker's high
690 // scheduler priority to get itself always run
691 // before other goroutines and always in the
692 // fresh quantum started by GC.
695 // Partially satisfy this assist.
696 gp.gcAssistBytes += scanBytes
698 // As a heuristic, we move this assist to the
699 // back of the queue so that large assists
700 // can't clog up the assist queue and
701 // substantially delay small assists.
702 work.assistQueue.q.pushBack(gp)
708 // Convert from scan bytes back to work.
709 assistWorkPerByte := gcController.assistWorkPerByte.Load()
710 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
711 gcController.bgScanCredit.Add(scanWork)
713 unlock(&work.assistQueue.lock)
716 // scanstack scans gp's stack, greying all pointers found on the stack.
718 // Returns the amount of scan work performed, but doesn't update
719 // gcController.stackScanWork or flush any credit. Any background credit produced
720 // by this function should be flushed by its caller. scanstack itself can't
721 // safely flush because it may result in trying to wake up a goroutine that
722 // was just scanned, resulting in a self-deadlock.
724 // scanstack will also shrink the stack if it is safe to do so. If it
725 // is not, it schedules a stack shrink for the next synchronous safe
728 // scanstack is marked go:systemstack because it must not be preempted
729 // while using a workbuf.
733 func scanstack(gp *g, gcw *gcWork) int64 {
734 if readgstatus(gp)&_Gscan == 0 {
735 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
736 throw("scanstack - bad status")
739 switch readgstatus(gp) &^ _Gscan {
741 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
742 throw("mark - bad status")
746 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
747 throw("scanstack: goroutine not stopped")
748 case _Grunnable, _Gsyscall, _Gwaiting:
753 throw("can't scan our own stack")
756 // scannedSize is the amount of work we'll be reporting.
758 // It is less than the allocated size (which is hi-lo).
760 if gp.syscallsp != 0 {
761 sp = gp.syscallsp // If in a system call this is the stack pointer (gp.sched.sp can be 0 in this case on Windows).
765 scannedSize := gp.stack.hi - sp
767 // Keep statistics for initial stack size calculation.
768 // Note that this accumulates the scanned size, not the allocated size.
769 p := getg().m.p.ptr()
770 p.scannedStackSize += uint64(scannedSize)
773 if isShrinkStackSafe(gp) {
774 // Shrink the stack if not much of it is being used.
777 // Otherwise, shrink the stack at the next sync safe point.
778 gp.preemptShrink = true
781 var state stackScanState
782 state.stack = gp.stack
785 println("stack trace goroutine", gp.goid)
788 if debugScanConservative && gp.asyncSafePoint {
789 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
792 // Scan the saved context register. This is effectively a live
793 // register that gets moved back and forth between the
794 // register and sched.ctxt without a write barrier.
795 if gp.sched.ctxt != nil {
796 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
799 // Scan the stack. Accumulate a list of stack objects.
801 for u.init(gp, 0); u.valid(); u.next() {
802 scanframeworker(&u.frame, &state, gcw)
805 // Find additional pointers that point into the stack from the heap.
806 // Currently this includes defers and panics. See also function copystack.
808 // Find and trace other pointers in defer records.
809 for d := gp._defer; d != nil; d = d.link {
811 // Scan the func value, which could be a stack allocated closure.
813 scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
816 // The link field of a stack-allocated defer record might point
817 // to a heap-allocated defer record. Keep that heap record live.
818 scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
820 // Retain defers records themselves.
821 // Defer records might not be reachable from the G through regular heap
822 // tracing because the defer linked list might weave between the stack and the heap.
824 scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
827 if gp._panic != nil {
828 // Panics are always stack allocated.
829 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
832 // Find and scan all reachable stack objects.
834 // The state's pointer queue prioritizes precise pointers over
835 // conservative pointers so that we'll prefer scanning stack
836 // objects precisely.
839 p, conservative := state.getPtr()
843 obj := state.findObject(p)
849 // We've already scanned this object.
852 obj.setRecord(nil) // Don't scan it again.
855 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
857 print(" (conservative)")
865 // This path is pretty unlikely, an object large enough
866 // to have a GC program allocated on the stack.
867 // We need some space to unpack the program into a straight
868 // bitmask, which we allocate/free here.
869 // TODO: it would be nice if there were a way to run a GC
870 // program without having to store all its bits. We'd have
871 // to change from a Lempel-Ziv style program to something else.
872 // Or we can forbid putting objects on stacks if they require
873 // a gc program (see issue 27447).
874 s = materializeGCProg(r.ptrdata(), gcdata)
875 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
878 b := state.stack.lo + uintptr(obj.off)
880 scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
882 scanblock(b, r.ptrdata(), gcdata, gcw, &state)
886 dematerializeGCProg(s)
890 // Deallocate object buffers.
891 // (Pointer buffers were all deallocated in the loop above.)
892 for state.head != nil {
896 for i := 0; i < x.nobj; i++ {
898 if obj.r == nil { // reachable
901 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
902 // Note: not necessarily really dead - only reachable-from-ptr dead.
906 putempty((*workbuf)(unsafe.Pointer(x)))
908 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
909 throw("remaining pointer buffers")
911 return int64(scannedSize)
914 // Scan a stack frame: local variables and function arguments/results.
917 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
918 if _DebugGC > 1 && frame.continpc != 0 {
919 print("scanframe ", funcname(frame.fn), "\n")
922 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
923 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
924 if state.conservative || isAsyncPreempt || isDebugCall {
925 if debugScanConservative {
926 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
929 // Conservatively scan the frame. Unlike the precise
930 // case, this includes the outgoing argument space
931 // since we may have stopped while this function was
932 // setting up a call.
934 // TODO: We could narrow this down if the compiler
935 // produced a single map per function of stack slots
936 // and registers that ever contain a pointer.
938 size := frame.varp - frame.sp
940 scanConservative(frame.sp, size, nil, gcw, state)
944 // Scan arguments to this frame.
945 if n := frame.argBytes(); n != 0 {
946 // TODO: We could pass the entry argument map
947 // to narrow this down further.
948 scanConservative(frame.argp, n, nil, gcw, state)
951 if isAsyncPreempt || isDebugCall {
952 // This function's frame contained the
953 // registers for the asynchronously stopped
954 // parent frame. Scan the parent
956 state.conservative = true
958 // We only wanted to scan those two frames
959 // conservatively. Clear the flag for future
961 state.conservative = false
966 locals, args, objs := frame.getStackMap(&state.cache, false)
968 // Scan local variables if stack frame has been allocated.
970 size := uintptr(locals.n) * goarch.PtrSize
971 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
976 scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
979 // Add all stack objects to the stack object list.
981 // varp is 0 for defers, where there are no locals.
982 // In that case, there can't be a pointer to its args, either.
983 // (And all args would be scanned above anyway.)
984 for i := range objs {
987 base := frame.varp // locals base pointer
989 base = frame.argp // arguments and return values base pointer
991 ptr := base + uintptr(off)
993 // object hasn't been allocated in the frame yet.
997 println("stkobj at", hex(ptr), "of size", obj.size)
999 state.addObject(ptr, obj)
1004 type gcDrainFlags int
1007 gcDrainUntilPreempt gcDrainFlags = 1 << iota
1008 gcDrainFlushBgCredit
1013 // gcDrain scans roots and objects in work buffers, blackening grey
1014 // objects until it is unable to get more work. It may return before
1015 // GC is done; it's the caller's responsibility to balance work from
1018 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
1021 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
1024 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
1025 // pollFractionalWorkerExit() returns true. This implies
1028 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
1029 // credit to gcController.bgScanCredit every gcCreditSlack units of
1032 // gcDrain will always return if there is a pending STW.
1035 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
1036 if !writeBarrier.needed {
1037 throw("gcDrain phase incorrect")
1041 preemptible := flags&gcDrainUntilPreempt != 0
1042 flushBgCredit := flags&gcDrainFlushBgCredit != 0
1043 idle := flags&gcDrainIdle != 0
1045 initScanWork := gcw.heapScanWork
1047 // checkWork is the scan work before performing the next
1048 // self-preempt check.
1049 checkWork := int64(1<<63 - 1)
1050 var check func() bool
1051 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
1052 checkWork = initScanWork + drainCheckThreshold
1055 } else if flags&gcDrainFractional != 0 {
1056 check = pollFractionalWorkerExit
1060 // Drain root marking jobs.
1061 if work.markrootNext < work.markrootJobs {
1062 // Stop if we're preemptible or if someone wants to STW.
1063 for !(gp.preempt && (preemptible || sched.gcwaiting.Load())) {
1064 job := atomic.Xadd(&work.markrootNext, +1) - 1
1065 if job >= work.markrootJobs {
1068 markroot(gcw, job, flushBgCredit)
1069 if check != nil && check() {
1075 // Drain heap marking jobs.
1076 // Stop if we're preemptible or if someone wants to STW.
1077 for !(gp.preempt && (preemptible || sched.gcwaiting.Load())) {
1078 // Try to keep work available on the global queue. We used to
1079 // check if there were waiting workers, but it's better to
1080 // just keep work available than to make workers wait. In the
1081 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1087 b := gcw.tryGetFast()
1091 // Flush the write barrier
1092 // buffer; this may create
1099 // Unable to get work.
1104 // Flush background scan work credit to the global
1105 // account if we've accumulated enough locally so
1106 // mutator assists can draw on it.
1107 if gcw.heapScanWork >= gcCreditSlack {
1108 gcController.heapScanWork.Add(gcw.heapScanWork)
1110 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1113 checkWork -= gcw.heapScanWork
1114 gcw.heapScanWork = 0
1117 checkWork += drainCheckThreshold
1118 if check != nil && check() {
1126 // Flush remaining scan work credit.
1127 if gcw.heapScanWork > 0 {
1128 gcController.heapScanWork.Add(gcw.heapScanWork)
1130 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1132 gcw.heapScanWork = 0
1136 // gcDrainN blackens grey objects until it has performed roughly
1137 // scanWork units of scan work or the G is preempted. This is
1138 // best-effort, so it may perform less work if it fails to get a work
1139 // buffer. Otherwise, it will perform at least n units of work, but
1140 // may perform more because scanning is always done in whole object
1141 // increments. It returns the amount of scan work performed.
1143 // The caller goroutine must be in a preemptible state (e.g.,
1144 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1145 // consequence, this must be called on the system stack.
1149 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1150 if !writeBarrier.needed {
1151 throw("gcDrainN phase incorrect")
1154 // There may already be scan work on the gcw, which we don't
1155 // want to claim was done by this call.
1156 workFlushed := -gcw.heapScanWork
1158 // In addition to backing out because of a preemption, back out
1159 // if the GC CPU limiter is enabled.
1161 for !gp.preempt && !gcCPULimiter.limiting() && workFlushed+gcw.heapScanWork < scanWork {
1162 // See gcDrain comment.
1167 b := gcw.tryGetFast()
1171 // Flush the write barrier buffer;
1172 // this may create more work.
1179 // Try to do a root job.
1180 if work.markrootNext < work.markrootJobs {
1181 job := atomic.Xadd(&work.markrootNext, +1) - 1
1182 if job < work.markrootJobs {
1183 workFlushed += markroot(gcw, job, false)
1187 // No heap or root jobs.
1193 // Flush background scan work credit.
1194 if gcw.heapScanWork >= gcCreditSlack {
1195 gcController.heapScanWork.Add(gcw.heapScanWork)
1196 workFlushed += gcw.heapScanWork
1197 gcw.heapScanWork = 0
1201 // Unlike gcDrain, there's no need to flush remaining work
1202 // here because this never flushes to bgScanCredit and
1203 // gcw.dispose will flush any remaining work to scanWork.
1205 return workFlushed + gcw.heapScanWork
1208 // scanblock scans b as scanobject would, but using an explicit
1209 // pointer bitmap instead of the heap bitmap.
1211 // This is used to scan non-heap roots, so it does not update
1212 // gcw.bytesMarked or gcw.heapScanWork.
1214 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1217 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1218 // Use local copies of original parameters, so that a stack trace
1219 // due to one of the throws below shows the original block
1224 for i := uintptr(0); i < n; {
1225 // Find bits for the next word.
1226 bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
1228 i += goarch.PtrSize * 8
1231 for j := 0; j < 8 && i < n; j++ {
1233 // Same work as in scanobject; see comments there.
1234 p := *(*uintptr)(unsafe.Pointer(b + i))
1236 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1237 greyobject(obj, b, i, span, gcw, objIndex)
1238 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1239 stk.putPtr(p, false)
1249 // scanobject scans the object starting at b, adding pointers to gcw.
1250 // b must point to the beginning of a heap object or an oblet.
1251 // scanobject consults the GC bitmap for the pointer mask and the
1252 // spans for the size of the object.
1255 func scanobject(b uintptr, gcw *gcWork) {
1256 // Prefetch object before we scan it.
1258 // This will overlap fetching the beginning of the object with initial
1259 // setup before we start scanning the object.
1262 // Find the bits for b and the size of the object at b.
1264 // b is either the beginning of an object, in which case this
1265 // is the size of the object to scan, or it points to an
1266 // oblet, in which case we compute the size to scan below.
1267 s := spanOfUnchecked(b)
1270 throw("scanobject n == 0")
1272 if s.spanclass.noscan() {
1273 // Correctness-wise this is ok, but it's inefficient
1274 // if noscan objects reach here.
1275 throw("scanobject of a noscan object")
1278 if n > maxObletBytes {
1279 // Large object. Break into oblets for better
1280 // parallelism and lower latency.
1282 // Enqueue the other oblets to scan later.
1283 // Some oblets may be in b's scalar tail, but
1284 // these will be marked as "no more pointers",
1285 // so we'll drop out immediately when we go to
1287 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1288 if !gcw.putFast(oblet) {
1294 // Compute the size of the oblet. Since this object
1295 // must be a large object, s.base() is the beginning
1297 n = s.base() + s.elemsize - b
1298 if n > maxObletBytes {
1303 hbits := heapBitsForAddr(b, n)
1304 var scanSize uintptr
1307 if hbits, addr = hbits.nextFast(); addr == 0 {
1308 if hbits, addr = hbits.next(); addr == 0 {
1313 // Keep track of farthest pointer we found, so we can
1314 // update heapScanWork. TODO: is there a better metric,
1315 // now that we can skip scalar portions pretty efficiently?
1316 scanSize = addr - b + goarch.PtrSize
1318 // Work here is duplicated in scanblock and above.
1319 // If you make changes here, make changes there too.
1320 obj := *(*uintptr)(unsafe.Pointer(addr))
1322 // At this point we have extracted the next potential pointer.
1323 // Quickly filter out nil and pointers back to the current object.
1324 if obj != 0 && obj-b >= n {
1325 // Test if obj points into the Go heap and, if so,
1328 // Note that it's possible for findObject to
1329 // fail if obj points to a just-allocated heap
1330 // object because of a race with growing the
1331 // heap. In this case, we know the object was
1332 // just allocated and hence will be marked by
1333 // allocation itself.
1334 if obj, span, objIndex := findObject(obj, b, addr-b); obj != 0 {
1335 greyobject(obj, b, addr-b, span, gcw, objIndex)
1339 gcw.bytesMarked += uint64(n)
1340 gcw.heapScanWork += int64(scanSize)
1343 // scanConservative scans block [b, b+n) conservatively, treating any
1344 // pointer-like value in the block as a pointer.
1346 // If ptrmask != nil, only words that are marked in ptrmask are
1347 // considered as potential pointers.
1349 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1350 // and may contain pointers to stack objects.
1351 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1352 if debugScanConservative {
1354 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1355 hexdumpWords(b, b+n, func(p uintptr) byte {
1357 word := (p - b) / goarch.PtrSize
1358 bits := *addb(ptrmask, word/8)
1359 if (bits>>(word%8))&1 == 0 {
1364 val := *(*uintptr)(unsafe.Pointer(p))
1365 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1369 span := spanOfHeap(val)
1373 idx := span.objIndex(val)
1374 if span.isFree(idx) {
1382 for i := uintptr(0); i < n; i += goarch.PtrSize {
1384 word := i / goarch.PtrSize
1385 bits := *addb(ptrmask, word/8)
1387 // Skip 8 words (the loop increment will do the 8th)
1389 // This must be the first time we've
1390 // seen this word of ptrmask, so i
1391 // must be 8-word-aligned, but check
1392 // our reasoning just in case.
1393 if i%(goarch.PtrSize*8) != 0 {
1394 throw("misaligned mask")
1396 i += goarch.PtrSize*8 - goarch.PtrSize
1399 if (bits>>(word%8))&1 == 0 {
1404 val := *(*uintptr)(unsafe.Pointer(b + i))
1406 // Check if val points into the stack.
1407 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1408 // val may point to a stack object. This
1409 // object may be dead from last cycle and
1410 // hence may contain pointers to unallocated
1411 // objects, but unlike heap objects we can't
1412 // tell if it's already dead. Hence, if all
1413 // pointers to this object are from
1414 // conservative scanning, we have to scan it
1415 // defensively, too.
1416 state.putPtr(val, true)
1420 // Check if val points to a heap span.
1421 span := spanOfHeap(val)
1426 // Check if val points to an allocated object.
1427 idx := span.objIndex(val)
1428 if span.isFree(idx) {
1432 // val points to an allocated object. Mark it.
1433 obj := span.base() + idx*span.elemsize
1434 greyobject(obj, b, i, span, gcw, idx)
1438 // Shade the object if it isn't already.
1439 // The object is not nil and known to be in the heap.
1440 // Preemption must be disabled.
1443 func shade(b uintptr) {
1444 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1445 gcw := &getg().m.p.ptr().gcw
1446 greyobject(obj, 0, 0, span, gcw, objIndex)
1450 // obj is the start of an object with mark mbits.
1451 // If it isn't already marked, mark it and enqueue into gcw.
1452 // base and off are for debugging only and could be removed.
1454 // See also wbBufFlush1, which partially duplicates this logic.
1456 //go:nowritebarrierrec
1457 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1458 // obj should be start of allocation, and so must be at least pointer-aligned.
1459 if obj&(goarch.PtrSize-1) != 0 {
1460 throw("greyobject: obj not pointer-aligned")
1462 mbits := span.markBitsForIndex(objIndex)
1465 if setCheckmark(obj, base, off, mbits) {
1470 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1471 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1472 gcDumpObject("base", base, off)
1473 gcDumpObject("obj", obj, ^uintptr(0))
1474 getg().m.traceback = 2
1475 throw("marking free object")
1478 // If marked we have nothing to do.
1479 if mbits.isMarked() {
1485 arena, pageIdx, pageMask := pageIndexOf(span.base())
1486 if arena.pageMarks[pageIdx]&pageMask == 0 {
1487 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1490 // If this is a noscan object, fast-track it to black
1491 // instead of greying it.
1492 if span.spanclass.noscan() {
1493 gcw.bytesMarked += uint64(span.elemsize)
1498 // We're adding obj to P's local workbuf, so it's likely
1499 // this object will be processed soon by the same P.
1500 // Even if the workbuf gets flushed, there will likely still be
1501 // some benefit on platforms with inclusive shared caches.
1503 // Queue the obj for scanning.
1504 if !gcw.putFast(obj) {
1509 // gcDumpObject dumps the contents of obj for debugging and marks the
1510 // field at byte offset off in obj.
1511 func gcDumpObject(label string, obj, off uintptr) {
1513 print(label, "=", hex(obj))
1518 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1519 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1520 print(mSpanStateNames[state], "\n")
1522 print("unknown(", state, ")\n")
1527 if s.state.get() == mSpanManual && size == 0 {
1528 // We're printing something from a stack frame. We
1529 // don't know how big it is, so just show up to an
1531 size = off + goarch.PtrSize
1533 for i := uintptr(0); i < size; i += goarch.PtrSize {
1534 // For big objects, just print the beginning (because
1535 // that usually hints at the object's type) and the
1536 // fields around off.
1537 if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
1545 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1556 // gcmarknewobject marks a newly allocated object black. obj must
1557 // not contain any non-nil pointers.
1559 // This is nosplit so it can manipulate a gcWork without preemption.
1563 func gcmarknewobject(span *mspan, obj, size uintptr) {
1564 if useCheckmark { // The world should be stopped so this should not happen.
1565 throw("gcmarknewobject called while doing checkmark")
1569 objIndex := span.objIndex(obj)
1570 span.markBitsForIndex(objIndex).setMarked()
1573 arena, pageIdx, pageMask := pageIndexOf(span.base())
1574 if arena.pageMarks[pageIdx]&pageMask == 0 {
1575 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1578 gcw := &getg().m.p.ptr().gcw
1579 gcw.bytesMarked += uint64(size)
1582 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1584 // The world must be stopped.
1585 func gcMarkTinyAllocs() {
1586 assertWorldStopped()
1588 for _, p := range allp {
1590 if c == nil || c.tiny == 0 {
1593 _, span, objIndex := findObject(c.tiny, 0, 0)
1595 greyobject(c.tiny, 0, 0, span, gcw, objIndex)