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 casgstatus(userG, _Grunning, _Gwaiting)
222 userG.waitreason = 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).
392 // The special itself is a root.
393 scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
395 unlock(&s.speciallock)
400 // gcAssistAlloc performs GC work to make gp's assist debt positive.
401 // gp must be the calling user goroutine.
403 // This must be called with preemption enabled.
404 func gcAssistAlloc(gp *g) {
405 // Don't assist in non-preemptible contexts. These are
406 // generally fragile and won't allow the assist to block.
407 if getg() == gp.m.g0 {
410 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
416 // Compute the amount of scan work we need to do to make the
417 // balance positive. When the required amount of work is low,
418 // we over-assist to build up credit for future allocations
419 // and amortize the cost of assisting.
420 assistWorkPerByte := gcController.assistWorkPerByte.Load()
421 assistBytesPerWork := gcController.assistBytesPerWork.Load()
422 debtBytes := -gp.gcAssistBytes
423 scanWork := int64(assistWorkPerByte * float64(debtBytes))
424 if scanWork < gcOverAssistWork {
425 scanWork = gcOverAssistWork
426 debtBytes = int64(assistBytesPerWork * float64(scanWork))
429 // Steal as much credit as we can from the background GC's
430 // scan credit. This is racy and may drop the background
431 // credit below 0 if two mutators steal at the same time. This
432 // will just cause steals to fail until credit is accumulated
433 // again, so in the long run it doesn't really matter, but we
434 // do have to handle the negative credit case.
435 bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
437 if bgScanCredit > 0 {
438 if bgScanCredit < scanWork {
439 stolen = bgScanCredit
440 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
443 gp.gcAssistBytes += debtBytes
445 atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
450 // We were able to steal all of the credit we
453 traceGCMarkAssistDone()
459 if trace.enabled && !traced {
461 traceGCMarkAssistStart()
464 // Perform assist work
466 gcAssistAlloc1(gp, scanWork)
467 // The user stack may have moved, so this can't touch
468 // anything on it until it returns from systemstack.
471 completed := gp.param != nil
477 if gp.gcAssistBytes < 0 {
478 // We were unable steal enough credit or perform
479 // enough work to pay off the assist debt. We need to
480 // do one of these before letting the mutator allocate
481 // more to prevent over-allocation.
483 // If this is because we were preempted, reschedule
484 // and try some more.
490 // Add this G to an assist queue and park. When the GC
491 // has more background credit, it will satisfy queued
492 // assists before flushing to the global credit pool.
494 // Note that this does *not* get woken up when more
495 // work is added to the work list. The theory is that
496 // there wasn't enough work to do anyway, so we might
497 // as well let background marking take care of the
498 // work that is available.
503 // At this point either background GC has satisfied
504 // this G's assist debt, or the GC cycle is over.
507 traceGCMarkAssistDone()
511 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
512 // stack. This is a separate function to make it easier to see that
513 // we're not capturing anything from the user stack, since the user
514 // stack may move while we're in this function.
516 // gcAssistAlloc1 indicates whether this assist completed the mark
517 // phase by setting gp.param to non-nil. This can't be communicated on
518 // the stack since it may move.
521 func gcAssistAlloc1(gp *g, scanWork int64) {
522 // Clear the flag indicating that this assist completed the
526 if atomic.Load(&gcBlackenEnabled) == 0 {
527 // The gcBlackenEnabled check in malloc races with the
528 // store that clears it but an atomic check in every malloc
529 // would be a performance hit.
530 // Instead we recheck it here on the non-preemptable system
531 // stack to determine if we should perform an assist.
533 // GC is done, so ignore any remaining debt.
537 // Track time spent in this assist. Since we're on the
538 // system stack, this is non-preemptible, so we can
539 // just measure start and end time.
540 startTime := nanotime()
542 decnwait := atomic.Xadd(&work.nwait, -1)
543 if decnwait == work.nproc {
544 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
545 throw("nwait > work.nprocs")
548 // gcDrainN requires the caller to be preemptible.
549 casgstatus(gp, _Grunning, _Gwaiting)
550 gp.waitreason = waitReasonGCAssistMarking
552 // drain own cached work first in the hopes that it
553 // will be more cache friendly.
554 gcw := &getg().m.p.ptr().gcw
555 workDone := gcDrainN(gcw, scanWork)
557 casgstatus(gp, _Gwaiting, _Grunning)
559 // Record that we did this much scan work.
561 // Back out the number of bytes of assist credit that
562 // this scan work counts for. The "1+" is a poor man's
563 // round-up, to ensure this adds credit even if
564 // assistBytesPerWork is very low.
565 assistBytesPerWork := gcController.assistBytesPerWork.Load()
566 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
568 // If this is the last worker and we ran out of work,
569 // signal a completion point.
570 incnwait := atomic.Xadd(&work.nwait, +1)
571 if incnwait > work.nproc {
572 println("runtime: work.nwait=", incnwait,
573 "work.nproc=", work.nproc)
574 throw("work.nwait > work.nproc")
577 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
578 // This has reached a background completion point. Set
579 // gp.param to a non-nil value to indicate this. It
580 // doesn't matter what we set it to (it just has to be
582 gp.param = unsafe.Pointer(gp)
584 duration := nanotime() - startTime
586 _p_.gcAssistTime += duration
587 if _p_.gcAssistTime > gcAssistTimeSlack {
588 atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
593 // gcWakeAllAssists wakes all currently blocked assists. This is used
594 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
595 // new assists from going to sleep after this point.
596 func gcWakeAllAssists() {
597 lock(&work.assistQueue.lock)
598 list := work.assistQueue.q.popList()
600 unlock(&work.assistQueue.lock)
603 // gcParkAssist puts the current goroutine on the assist queue and parks.
605 // gcParkAssist reports whether the assist is now satisfied. If it
606 // returns false, the caller must retry the assist.
607 func gcParkAssist() bool {
608 lock(&work.assistQueue.lock)
609 // If the GC cycle finished while we were getting the lock,
610 // exit the assist. The cycle can't finish while we hold the
612 if atomic.Load(&gcBlackenEnabled) == 0 {
613 unlock(&work.assistQueue.lock)
618 oldList := work.assistQueue.q
619 work.assistQueue.q.pushBack(gp)
621 // Recheck for background credit now that this G is in
622 // the queue, but can still back out. This avoids a
623 // race in case background marking has flushed more
624 // credit since we checked above.
625 if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
626 work.assistQueue.q = oldList
627 if oldList.tail != 0 {
628 oldList.tail.ptr().schedlink.set(nil)
630 unlock(&work.assistQueue.lock)
634 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
638 // gcFlushBgCredit flushes scanWork units of background scan work
639 // credit. This first satisfies blocked assists on the
640 // work.assistQueue and then flushes any remaining credit to
641 // gcController.bgScanCredit.
643 // Write barriers are disallowed because this is used by gcDrain after
644 // it has ensured that all work is drained and this must preserve that
647 //go:nowritebarrierrec
648 func gcFlushBgCredit(scanWork int64) {
649 if work.assistQueue.q.empty() {
650 // Fast path; there are no blocked assists. There's a
651 // small window here where an assist may add itself to
652 // the blocked queue and park. If that happens, we'll
653 // just get it on the next flush.
654 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
658 assistBytesPerWork := gcController.assistBytesPerWork.Load()
659 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
661 lock(&work.assistQueue.lock)
662 for !work.assistQueue.q.empty() && scanBytes > 0 {
663 gp := work.assistQueue.q.pop()
664 // Note that gp.gcAssistBytes is negative because gp
665 // is in debt. Think carefully about the signs below.
666 if scanBytes+gp.gcAssistBytes >= 0 {
667 // Satisfy this entire assist debt.
668 scanBytes += gp.gcAssistBytes
670 // It's important that we *not* put gp in
671 // runnext. Otherwise, it's possible for user
672 // code to exploit the GC worker's high
673 // scheduler priority to get itself always run
674 // before other goroutines and always in the
675 // fresh quantum started by GC.
678 // Partially satisfy this assist.
679 gp.gcAssistBytes += scanBytes
681 // As a heuristic, we move this assist to the
682 // back of the queue so that large assists
683 // can't clog up the assist queue and
684 // substantially delay small assists.
685 work.assistQueue.q.pushBack(gp)
691 // Convert from scan bytes back to work.
692 assistWorkPerByte := gcController.assistWorkPerByte.Load()
693 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
694 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
696 unlock(&work.assistQueue.lock)
699 // scanstack scans gp's stack, greying all pointers found on the stack.
701 // Returns the amount of scan work performed, but doesn't update
702 // gcController.stackScanWork or flush any credit. Any background credit produced
703 // by this function should be flushed by its caller. scanstack itself can't
704 // safely flush because it may result in trying to wake up a goroutine that
705 // was just scanned, resulting in a self-deadlock.
707 // scanstack will also shrink the stack if it is safe to do so. If it
708 // is not, it schedules a stack shrink for the next synchronous safe
711 // scanstack is marked go:systemstack because it must not be preempted
712 // while using a workbuf.
716 func scanstack(gp *g, gcw *gcWork) int64 {
717 if readgstatus(gp)&_Gscan == 0 {
718 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
719 throw("scanstack - bad status")
722 switch readgstatus(gp) &^ _Gscan {
724 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
725 throw("mark - bad status")
729 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
730 throw("scanstack: goroutine not stopped")
731 case _Grunnable, _Gsyscall, _Gwaiting:
736 throw("can't scan our own stack")
739 // stackSize is the amount of work we'll be reporting.
741 // We report the total stack size, more than we scan,
742 // because this number needs to line up with gcControllerState's
743 // stackScan and scannableStackSize fields.
745 // See the documentation on those fields for more information.
746 stackSize := gp.stack.hi - gp.stack.lo
748 if isShrinkStackSafe(gp) {
749 // Shrink the stack if not much of it is being used.
752 // Otherwise, shrink the stack at the next sync safe point.
753 gp.preemptShrink = true
756 var state stackScanState
757 state.stack = gp.stack
760 println("stack trace goroutine", gp.goid)
763 if debugScanConservative && gp.asyncSafePoint {
764 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
767 // Scan the saved context register. This is effectively a live
768 // register that gets moved back and forth between the
769 // register and sched.ctxt without a write barrier.
770 if gp.sched.ctxt != nil {
771 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
774 // Scan the stack. Accumulate a list of stack objects.
775 scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
776 scanframeworker(frame, &state, gcw)
779 gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
781 // Find additional pointers that point into the stack from the heap.
782 // Currently this includes defers and panics. See also function copystack.
784 // Find and trace other pointers in defer records.
785 for d := gp._defer; d != nil; d = d.link {
787 // Scan the func value, which could be a stack allocated closure.
789 scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
792 // The link field of a stack-allocated defer record might point
793 // to a heap-allocated defer record. Keep that heap record live.
794 scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
796 // Retain defers records themselves.
797 // Defer records might not be reachable from the G through regular heap
798 // tracing because the defer linked list might weave between the stack and the heap.
800 scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
803 if gp._panic != nil {
804 // Panics are always stack allocated.
805 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
808 // Find and scan all reachable stack objects.
810 // The state's pointer queue prioritizes precise pointers over
811 // conservative pointers so that we'll prefer scanning stack
812 // objects precisely.
815 p, conservative := state.getPtr()
819 obj := state.findObject(p)
825 // We've already scanned this object.
828 obj.setRecord(nil) // Don't scan it again.
831 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
833 print(" (conservative)")
841 // This path is pretty unlikely, an object large enough
842 // to have a GC program allocated on the stack.
843 // We need some space to unpack the program into a straight
844 // bitmask, which we allocate/free here.
845 // TODO: it would be nice if there were a way to run a GC
846 // program without having to store all its bits. We'd have
847 // to change from a Lempel-Ziv style program to something else.
848 // Or we can forbid putting objects on stacks if they require
849 // a gc program (see issue 27447).
850 s = materializeGCProg(r.ptrdata(), gcdata)
851 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
854 b := state.stack.lo + uintptr(obj.off)
856 scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
858 scanblock(b, r.ptrdata(), gcdata, gcw, &state)
862 dematerializeGCProg(s)
866 // Deallocate object buffers.
867 // (Pointer buffers were all deallocated in the loop above.)
868 for state.head != nil {
872 for i := 0; i < x.nobj; i++ {
874 if obj.r == nil { // reachable
877 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
878 // Note: not necessarily really dead - only reachable-from-ptr dead.
882 putempty((*workbuf)(unsafe.Pointer(x)))
884 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
885 throw("remaining pointer buffers")
887 return int64(stackSize)
890 // Scan a stack frame: local variables and function arguments/results.
893 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
894 if _DebugGC > 1 && frame.continpc != 0 {
895 print("scanframe ", funcname(frame.fn), "\n")
898 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
899 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
900 if state.conservative || isAsyncPreempt || isDebugCall {
901 if debugScanConservative {
902 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
905 // Conservatively scan the frame. Unlike the precise
906 // case, this includes the outgoing argument space
907 // since we may have stopped while this function was
908 // setting up a call.
910 // TODO: We could narrow this down if the compiler
911 // produced a single map per function of stack slots
912 // and registers that ever contain a pointer.
914 size := frame.varp - frame.sp
916 scanConservative(frame.sp, size, nil, gcw, state)
920 // Scan arguments to this frame.
921 if frame.arglen != 0 {
922 // TODO: We could pass the entry argument map
923 // to narrow this down further.
924 scanConservative(frame.argp, frame.arglen, nil, gcw, state)
927 if isAsyncPreempt || isDebugCall {
928 // This function's frame contained the
929 // registers for the asynchronously stopped
930 // parent frame. Scan the parent
932 state.conservative = true
934 // We only wanted to scan those two frames
935 // conservatively. Clear the flag for future
937 state.conservative = false
942 locals, args, objs := getStackMap(frame, &state.cache, false)
944 // Scan local variables if stack frame has been allocated.
946 size := uintptr(locals.n) * goarch.PtrSize
947 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
952 scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
955 // Add all stack objects to the stack object list.
957 // varp is 0 for defers, where there are no locals.
958 // In that case, there can't be a pointer to its args, either.
959 // (And all args would be scanned above anyway.)
960 for i := range objs {
963 base := frame.varp // locals base pointer
965 base = frame.argp // arguments and return values base pointer
967 ptr := base + uintptr(off)
969 // object hasn't been allocated in the frame yet.
973 println("stkobj at", hex(ptr), "of size", obj.size)
975 state.addObject(ptr, obj)
980 type gcDrainFlags int
983 gcDrainUntilPreempt gcDrainFlags = 1 << iota
989 // gcDrain scans roots and objects in work buffers, blackening grey
990 // objects until it is unable to get more work. It may return before
991 // GC is done; it's the caller's responsibility to balance work from
994 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
997 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
1000 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
1001 // pollFractionalWorkerExit() returns true. This implies
1004 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
1005 // credit to gcController.bgScanCredit every gcCreditSlack units of
1008 // gcDrain will always return if there is a pending STW.
1011 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
1012 if !writeBarrier.needed {
1013 throw("gcDrain phase incorrect")
1017 preemptible := flags&gcDrainUntilPreempt != 0
1018 flushBgCredit := flags&gcDrainFlushBgCredit != 0
1019 idle := flags&gcDrainIdle != 0
1021 initScanWork := gcw.heapScanWork
1023 // checkWork is the scan work before performing the next
1024 // self-preempt check.
1025 checkWork := int64(1<<63 - 1)
1026 var check func() bool
1027 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
1028 checkWork = initScanWork + drainCheckThreshold
1031 } else if flags&gcDrainFractional != 0 {
1032 check = pollFractionalWorkerExit
1036 // Drain root marking jobs.
1037 if work.markrootNext < work.markrootJobs {
1038 // Stop if we're preemptible or if someone wants to STW.
1039 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1040 job := atomic.Xadd(&work.markrootNext, +1) - 1
1041 if job >= work.markrootJobs {
1044 markroot(gcw, job, flushBgCredit)
1045 if check != nil && check() {
1051 // Drain heap marking jobs.
1052 // Stop if we're preemptible or if someone wants to STW.
1053 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1054 // Try to keep work available on the global queue. We used to
1055 // check if there were waiting workers, but it's better to
1056 // just keep work available than to make workers wait. In the
1057 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1063 b := gcw.tryGetFast()
1067 // Flush the write barrier
1068 // buffer; this may create
1075 // Unable to get work.
1080 // Flush background scan work credit to the global
1081 // account if we've accumulated enough locally so
1082 // mutator assists can draw on it.
1083 if gcw.heapScanWork >= gcCreditSlack {
1084 gcController.heapScanWork.Add(gcw.heapScanWork)
1086 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1089 checkWork -= gcw.heapScanWork
1090 gcw.heapScanWork = 0
1093 checkWork += drainCheckThreshold
1094 if check != nil && check() {
1102 // Flush remaining scan work credit.
1103 if gcw.heapScanWork > 0 {
1104 gcController.heapScanWork.Add(gcw.heapScanWork)
1106 gcFlushBgCredit(gcw.heapScanWork - initScanWork)
1108 gcw.heapScanWork = 0
1112 // gcDrainN blackens grey objects until it has performed roughly
1113 // scanWork units of scan work or the G is preempted. This is
1114 // best-effort, so it may perform less work if it fails to get a work
1115 // buffer. Otherwise, it will perform at least n units of work, but
1116 // may perform more because scanning is always done in whole object
1117 // increments. It returns the amount of scan work performed.
1119 // The caller goroutine must be in a preemptible state (e.g.,
1120 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1121 // consequence, this must be called on the system stack.
1125 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1126 if !writeBarrier.needed {
1127 throw("gcDrainN phase incorrect")
1130 // There may already be scan work on the gcw, which we don't
1131 // want to claim was done by this call.
1132 workFlushed := -gcw.heapScanWork
1135 for !gp.preempt && workFlushed+gcw.heapScanWork < scanWork {
1136 // See gcDrain comment.
1141 b := gcw.tryGetFast()
1145 // Flush the write barrier buffer;
1146 // this may create more work.
1153 // Try to do a root job.
1154 if work.markrootNext < work.markrootJobs {
1155 job := atomic.Xadd(&work.markrootNext, +1) - 1
1156 if job < work.markrootJobs {
1157 workFlushed += markroot(gcw, job, false)
1161 // No heap or root jobs.
1167 // Flush background scan work credit.
1168 if gcw.heapScanWork >= gcCreditSlack {
1169 gcController.heapScanWork.Add(gcw.heapScanWork)
1170 workFlushed += gcw.heapScanWork
1171 gcw.heapScanWork = 0
1175 // Unlike gcDrain, there's no need to flush remaining work
1176 // here because this never flushes to bgScanCredit and
1177 // gcw.dispose will flush any remaining work to scanWork.
1179 return workFlushed + gcw.heapScanWork
1182 // scanblock scans b as scanobject would, but using an explicit
1183 // pointer bitmap instead of the heap bitmap.
1185 // This is used to scan non-heap roots, so it does not update
1186 // gcw.bytesMarked or gcw.heapScanWork.
1188 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1191 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1192 // Use local copies of original parameters, so that a stack trace
1193 // due to one of the throws below shows the original block
1198 for i := uintptr(0); i < n; {
1199 // Find bits for the next word.
1200 bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
1202 i += goarch.PtrSize * 8
1205 for j := 0; j < 8 && i < n; j++ {
1207 // Same work as in scanobject; see comments there.
1208 p := *(*uintptr)(unsafe.Pointer(b + i))
1210 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1211 greyobject(obj, b, i, span, gcw, objIndex)
1212 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1213 stk.putPtr(p, false)
1223 // scanobject scans the object starting at b, adding pointers to gcw.
1224 // b must point to the beginning of a heap object or an oblet.
1225 // scanobject consults the GC bitmap for the pointer mask and the
1226 // spans for the size of the object.
1229 func scanobject(b uintptr, gcw *gcWork) {
1230 // Prefetch object before we scan it.
1232 // This will overlap fetching the beginning of the object with initial
1233 // setup before we start scanning the object.
1236 // Find the bits for b and the size of the object at b.
1238 // b is either the beginning of an object, in which case this
1239 // is the size of the object to scan, or it points to an
1240 // oblet, in which case we compute the size to scan below.
1241 hbits := heapBitsForAddr(b)
1242 s := spanOfUnchecked(b)
1245 throw("scanobject n == 0")
1248 if n > maxObletBytes {
1249 // Large object. Break into oblets for better
1250 // parallelism and lower latency.
1252 // It's possible this is a noscan object (not
1253 // from greyobject, but from other code
1254 // paths), in which case we must *not* enqueue
1255 // oblets since their bitmaps will be
1257 if s.spanclass.noscan() {
1258 // Bypass the whole scan.
1259 gcw.bytesMarked += uint64(n)
1263 // Enqueue the other oblets to scan later.
1264 // Some oblets may be in b's scalar tail, but
1265 // these will be marked as "no more pointers",
1266 // so we'll drop out immediately when we go to
1268 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1269 if !gcw.putFast(oblet) {
1275 // Compute the size of the oblet. Since this object
1276 // must be a large object, s.base() is the beginning
1278 n = s.base() + s.elemsize - b
1279 if n > maxObletBytes {
1285 for i = 0; i < n; i, hbits = i+goarch.PtrSize, hbits.next() {
1286 // Load bits once. See CL 22712 and issue 16973 for discussion.
1287 bits := hbits.bits()
1288 if bits&bitScan == 0 {
1289 break // no more pointers in this object
1291 if bits&bitPointer == 0 {
1292 continue // not a pointer
1295 // Work here is duplicated in scanblock and above.
1296 // If you make changes here, make changes there too.
1297 obj := *(*uintptr)(unsafe.Pointer(b + i))
1299 // At this point we have extracted the next potential pointer.
1300 // Quickly filter out nil and pointers back to the current object.
1301 if obj != 0 && obj-b >= n {
1302 // Test if obj points into the Go heap and, if so,
1305 // Note that it's possible for findObject to
1306 // fail if obj points to a just-allocated heap
1307 // object because of a race with growing the
1308 // heap. In this case, we know the object was
1309 // just allocated and hence will be marked by
1310 // allocation itself.
1311 if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
1312 greyobject(obj, b, i, span, gcw, objIndex)
1316 gcw.bytesMarked += uint64(n)
1317 gcw.heapScanWork += int64(i)
1320 // scanConservative scans block [b, b+n) conservatively, treating any
1321 // pointer-like value in the block as a pointer.
1323 // If ptrmask != nil, only words that are marked in ptrmask are
1324 // considered as potential pointers.
1326 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1327 // and may contain pointers to stack objects.
1328 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1329 if debugScanConservative {
1331 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1332 hexdumpWords(b, b+n, func(p uintptr) byte {
1334 word := (p - b) / goarch.PtrSize
1335 bits := *addb(ptrmask, word/8)
1336 if (bits>>(word%8))&1 == 0 {
1341 val := *(*uintptr)(unsafe.Pointer(p))
1342 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1346 span := spanOfHeap(val)
1350 idx := span.objIndex(val)
1351 if span.isFree(idx) {
1359 for i := uintptr(0); i < n; i += goarch.PtrSize {
1361 word := i / goarch.PtrSize
1362 bits := *addb(ptrmask, word/8)
1364 // Skip 8 words (the loop increment will do the 8th)
1366 // This must be the first time we've
1367 // seen this word of ptrmask, so i
1368 // must be 8-word-aligned, but check
1369 // our reasoning just in case.
1370 if i%(goarch.PtrSize*8) != 0 {
1371 throw("misaligned mask")
1373 i += goarch.PtrSize*8 - goarch.PtrSize
1376 if (bits>>(word%8))&1 == 0 {
1381 val := *(*uintptr)(unsafe.Pointer(b + i))
1383 // Check if val points into the stack.
1384 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1385 // val may point to a stack object. This
1386 // object may be dead from last cycle and
1387 // hence may contain pointers to unallocated
1388 // objects, but unlike heap objects we can't
1389 // tell if it's already dead. Hence, if all
1390 // pointers to this object are from
1391 // conservative scanning, we have to scan it
1392 // defensively, too.
1393 state.putPtr(val, true)
1397 // Check if val points to a heap span.
1398 span := spanOfHeap(val)
1403 // Check if val points to an allocated object.
1404 idx := span.objIndex(val)
1405 if span.isFree(idx) {
1409 // val points to an allocated object. Mark it.
1410 obj := span.base() + idx*span.elemsize
1411 greyobject(obj, b, i, span, gcw, idx)
1415 // Shade the object if it isn't already.
1416 // The object is not nil and known to be in the heap.
1417 // Preemption must be disabled.
1420 func shade(b uintptr) {
1421 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1422 gcw := &getg().m.p.ptr().gcw
1423 greyobject(obj, 0, 0, span, gcw, objIndex)
1427 // obj is the start of an object with mark mbits.
1428 // If it isn't already marked, mark it and enqueue into gcw.
1429 // base and off are for debugging only and could be removed.
1431 // See also wbBufFlush1, which partially duplicates this logic.
1433 //go:nowritebarrierrec
1434 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1435 // obj should be start of allocation, and so must be at least pointer-aligned.
1436 if obj&(goarch.PtrSize-1) != 0 {
1437 throw("greyobject: obj not pointer-aligned")
1439 mbits := span.markBitsForIndex(objIndex)
1442 if setCheckmark(obj, base, off, mbits) {
1447 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1448 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1449 gcDumpObject("base", base, off)
1450 gcDumpObject("obj", obj, ^uintptr(0))
1451 getg().m.traceback = 2
1452 throw("marking free object")
1455 // If marked we have nothing to do.
1456 if mbits.isMarked() {
1462 arena, pageIdx, pageMask := pageIndexOf(span.base())
1463 if arena.pageMarks[pageIdx]&pageMask == 0 {
1464 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1467 // If this is a noscan object, fast-track it to black
1468 // instead of greying it.
1469 if span.spanclass.noscan() {
1470 gcw.bytesMarked += uint64(span.elemsize)
1475 // We're adding obj to P's local workbuf, so it's likely
1476 // this object will be processed soon by the same P.
1477 // Even if the workbuf gets flushed, there will likely still be
1478 // some benefit on platforms with inclusive shared caches.
1480 // Queue the obj for scanning.
1481 if !gcw.putFast(obj) {
1486 // gcDumpObject dumps the contents of obj for debugging and marks the
1487 // field at byte offset off in obj.
1488 func gcDumpObject(label string, obj, off uintptr) {
1490 print(label, "=", hex(obj))
1495 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1496 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1497 print(mSpanStateNames[state], "\n")
1499 print("unknown(", state, ")\n")
1504 if s.state.get() == mSpanManual && size == 0 {
1505 // We're printing something from a stack frame. We
1506 // don't know how big it is, so just show up to an
1508 size = off + goarch.PtrSize
1510 for i := uintptr(0); i < size; i += goarch.PtrSize {
1511 // For big objects, just print the beginning (because
1512 // that usually hints at the object's type) and the
1513 // fields around off.
1514 if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
1522 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1533 // gcmarknewobject marks a newly allocated object black. obj must
1534 // not contain any non-nil pointers.
1536 // This is nosplit so it can manipulate a gcWork without preemption.
1540 func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
1541 if useCheckmark { // The world should be stopped so this should not happen.
1542 throw("gcmarknewobject called while doing checkmark")
1546 objIndex := span.objIndex(obj)
1547 span.markBitsForIndex(objIndex).setMarked()
1550 arena, pageIdx, pageMask := pageIndexOf(span.base())
1551 if arena.pageMarks[pageIdx]&pageMask == 0 {
1552 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1555 gcw := &getg().m.p.ptr().gcw
1556 gcw.bytesMarked += uint64(size)
1559 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1561 // The world must be stopped.
1562 func gcMarkTinyAllocs() {
1563 assertWorldStopped()
1565 for _, p := range allp {
1567 if c == nil || c.tiny == 0 {
1570 _, span, objIndex := findObject(c.tiny, 0, 0)
1572 greyobject(c.tiny, 0, 0, span, gcw, objIndex)