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
10 "runtime/internal/atomic"
11 "runtime/internal/sys"
16 fixedRootFinalizers = iota
20 // rootBlockBytes is the number of bytes to scan per data or
22 rootBlockBytes = 256 << 10
24 // maxObletBytes is the maximum bytes of an object to scan at
25 // once. Larger objects will be split up into "oblets" of at
26 // most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
27 // scan preemption at ~100 µs.
29 // This must be > _MaxSmallSize so that the object base is the
31 maxObletBytes = 128 << 10
33 // drainCheckThreshold specifies how many units of work to do
34 // between self-preemption checks in gcDrain. Assuming a scan
35 // rate of 1 MB/ms, this is ~100 µs. Lower values have higher
36 // overhead in the scan loop (the scheduler check may perform
37 // a syscall, so its overhead is nontrivial). Higher values
38 // make the system less responsive to incoming work.
39 drainCheckThreshold = 100000
41 // pagesPerSpanRoot indicates how many pages to scan from a span root
42 // at a time. Used by special root marking.
44 // Higher values improve throughput by increasing locality, but
45 // increase the minimum latency of a marking operation.
47 // Must be a multiple of the pageInUse bitmap element size and
48 // must also evenly divide pagesPerArena.
49 pagesPerSpanRoot = 512
52 // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
53 // some miscellany) and initializes scanning-related state.
55 // The world must be stopped.
56 func gcMarkRootPrepare() {
59 work.nFlushCacheRoots = 0
61 // Compute how many data and BSS root blocks there are.
62 nBlocks := func(bytes uintptr) int {
63 return int(divRoundUp(bytes, rootBlockBytes))
70 for _, datap := range activeModules() {
71 nDataRoots := nBlocks(datap.edata - datap.data)
72 if nDataRoots > work.nDataRoots {
73 work.nDataRoots = nDataRoots
77 for _, datap := range activeModules() {
78 nBSSRoots := nBlocks(datap.ebss - datap.bss)
79 if nBSSRoots > work.nBSSRoots {
80 work.nBSSRoots = nBSSRoots
84 // Scan span roots for finalizer specials.
86 // We depend on addfinalizer to mark objects that get
87 // finalizers after root marking.
89 // We're going to scan the whole heap (that was available at the time the
90 // mark phase started, i.e. markArenas) for in-use spans which have specials.
92 // Break up the work into arenas, and further into chunks.
94 // Snapshot allArenas as markArenas. This snapshot is safe because allArenas
96 mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
97 work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
101 // Gs may be created after this point, but it's okay that we
102 // ignore them because they begin life without any roots, so
103 // there's nothing to scan, and any roots they create during
104 // the concurrent phase will be caught by the write barrier.
105 work.nStackRoots = int(atomic.Loaduintptr(&allglen))
107 work.markrootNext = 0
108 work.markrootJobs = uint32(fixedRootCount + work.nFlushCacheRoots + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
110 // Calculate base indexes of each root type
111 work.baseFlushCache = uint32(fixedRootCount)
112 work.baseData = work.baseFlushCache + uint32(work.nFlushCacheRoots)
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 // nowritebarrier is only advisory here.
159 func markroot(gcw *gcWork, i uint32) {
160 // Note: if you add a case here, please also update heapdump.go:dumproots.
162 case work.baseFlushCache <= i && i < work.baseData:
163 flushmcache(int(i - work.baseFlushCache))
165 case work.baseData <= i && i < work.baseBSS:
166 for _, datap := range activeModules() {
167 markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
170 case work.baseBSS <= i && i < work.baseSpans:
171 for _, datap := range activeModules() {
172 markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
175 case i == fixedRootFinalizers:
176 for fb := allfin; fb != nil; fb = fb.alllink {
177 cnt := uintptr(atomic.Load(&fb.cnt))
178 scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
181 case i == fixedRootFreeGStacks:
182 // Switch to the system stack so we can call
184 systemstack(markrootFreeGStacks)
186 case work.baseSpans <= i && i < work.baseStacks:
187 // mark mspan.specials
188 markrootSpans(gcw, int(i-work.baseSpans))
191 // the rest is scanning goroutine stacks
193 if work.baseStacks <= i && i < work.baseEnd {
194 // N.B. Atomic read of allglen in gcMarkRootPrepare
195 // acts as a barrier to ensure that allgs must be large
196 // enough to contain all relevant Gs.
197 gp = allgs[i-work.baseStacks]
199 throw("markroot: bad index")
202 // remember when we've first observed the G blocked
203 // needed only to output in traceback
204 status := readgstatus(gp) // We are not in a scan state
205 if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
206 gp.waitsince = work.tstart
209 // scanstack must be done on the system stack in case
210 // we're trying to scan our own stack.
212 // If this is a self-scan, put the user G in
213 // _Gwaiting to prevent self-deadlock. It may
214 // already be in _Gwaiting if this is a mark
215 // worker or we're in mark termination.
216 userG := getg().m.curg
217 selfScan := gp == userG && readgstatus(userG) == _Grunning
219 casgstatus(userG, _Grunning, _Gwaiting)
220 userG.waitreason = waitReasonGarbageCollectionScan
223 // TODO: suspendG blocks (and spins) until gp
224 // stops, which may take a while for
225 // running goroutines. Consider doing this in
226 // two phases where the first is non-blocking:
227 // we scan the stacks we can and ask running
228 // goroutines to scan themselves; and the
230 stopped := suspendG(gp)
236 throw("g already scanned")
243 casgstatus(userG, _Gwaiting, _Grunning)
249 // markrootBlock scans the shard'th shard of the block of memory [b0,
250 // b0+n0), with the given pointer mask.
253 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) {
254 if rootBlockBytes%(8*sys.PtrSize) != 0 {
255 // This is necessary to pick byte offsets in ptrmask0.
256 throw("rootBlockBytes must be a multiple of 8*ptrSize")
259 // Note that if b0 is toward the end of the address space,
260 // then b0 + rootBlockBytes might wrap around.
261 // These tests are written to avoid any possible overflow.
262 off := uintptr(shard) * rootBlockBytes
267 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*sys.PtrSize))))
268 n := uintptr(rootBlockBytes)
274 scanblock(b, n, ptrmask, gcw, nil)
277 // markrootFreeGStacks frees stacks of dead Gs.
279 // This does not free stacks of dead Gs cached on Ps, but having a few
280 // cached stacks around isn't a problem.
281 func markrootFreeGStacks() {
282 // Take list of dead Gs with stacks.
283 lock(&sched.gFree.lock)
284 list := sched.gFree.stack
285 sched.gFree.stack = gList{}
286 unlock(&sched.gFree.lock)
292 q := gQueue{list.head, list.head}
293 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
297 // Manipulate the queue directly since the Gs are
298 // already all linked the right way.
302 // Put Gs back on the free list.
303 lock(&sched.gFree.lock)
304 sched.gFree.noStack.pushAll(q)
305 unlock(&sched.gFree.lock)
308 // markrootSpans marks roots for one shard of markArenas.
311 func markrootSpans(gcw *gcWork, shard int) {
312 // Objects with finalizers have two GC-related invariants:
314 // 1) Everything reachable from the object must be marked.
315 // This ensures that when we pass the object to its finalizer,
316 // everything the finalizer can reach will be retained.
318 // 2) Finalizer specials (which are not in the garbage
319 // collected heap) are roots. In practice, this means the fn
320 // field must be scanned.
321 sg := mheap_.sweepgen
323 // Find the arena and page index into that arena for this shard.
324 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
325 ha := mheap_.arenas[ai.l1()][ai.l2()]
326 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
328 // Construct slice of bitmap which we'll iterate over.
329 specialsbits := ha.pageSpecials[arenaPage/8:]
330 specialsbits = specialsbits[:pagesPerSpanRoot/8]
331 for i := range specialsbits {
332 // Find set bits, which correspond to spans with specials.
333 specials := atomic.Load8(&specialsbits[i])
337 for j := uint(0); j < 8; j++ {
338 if specials&(1<<j) == 0 {
341 // Find the span for this bit.
343 // This value is guaranteed to be non-nil because having
344 // specials implies that the span is in-use, and since we're
345 // currently marking we can be sure that we don't have to worry
346 // about the span being freed and re-used.
347 s := ha.spans[arenaPage+uint(i)*8+j]
349 // The state must be mSpanInUse if the specials bit is set, so
350 // sanity check that.
351 if state := s.state.get(); state != mSpanInUse {
352 print("s.state = ", state, "\n")
353 throw("non in-use span found with specials bit set")
355 // Check that this span was swept (it may be cached or uncached).
356 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
357 // sweepgen was updated (+2) during non-checkmark GC pass
358 print("sweep ", s.sweepgen, " ", sg, "\n")
359 throw("gc: unswept span")
362 // Lock the specials to prevent a special from being
363 // removed from the list while we're traversing it.
365 for sp := s.specials; sp != nil; sp = sp.next {
366 if sp.kind != _KindSpecialFinalizer {
369 // don't mark finalized object, but scan it so we
370 // retain everything it points to.
371 spf := (*specialfinalizer)(unsafe.Pointer(sp))
372 // A finalizer can be set for an inner byte of an object, find object beginning.
373 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
375 // Mark everything that can be reached from
376 // the object (but *not* the object itself or
377 // we'll never collect it).
380 // The special itself is a root.
381 scanblock(uintptr(unsafe.Pointer(&spf.fn)), sys.PtrSize, &oneptrmask[0], gcw, nil)
383 unlock(&s.speciallock)
388 // gcAssistAlloc performs GC work to make gp's assist debt positive.
389 // gp must be the calling user gorountine.
391 // This must be called with preemption enabled.
392 func gcAssistAlloc(gp *g) {
393 // Don't assist in non-preemptible contexts. These are
394 // generally fragile and won't allow the assist to block.
395 if getg() == gp.m.g0 {
398 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
404 // Compute the amount of scan work we need to do to make the
405 // balance positive. When the required amount of work is low,
406 // we over-assist to build up credit for future allocations
407 // and amortize the cost of assisting.
408 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
409 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
410 debtBytes := -gp.gcAssistBytes
411 scanWork := int64(assistWorkPerByte * float64(debtBytes))
412 if scanWork < gcOverAssistWork {
413 scanWork = gcOverAssistWork
414 debtBytes = int64(assistBytesPerWork * float64(scanWork))
417 // Steal as much credit as we can from the background GC's
418 // scan credit. This is racy and may drop the background
419 // credit below 0 if two mutators steal at the same time. This
420 // will just cause steals to fail until credit is accumulated
421 // again, so in the long run it doesn't really matter, but we
422 // do have to handle the negative credit case.
423 bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
425 if bgScanCredit > 0 {
426 if bgScanCredit < scanWork {
427 stolen = bgScanCredit
428 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
431 gp.gcAssistBytes += debtBytes
433 atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
438 // We were able to steal all of the credit we
441 traceGCMarkAssistDone()
447 if trace.enabled && !traced {
449 traceGCMarkAssistStart()
452 // Perform assist work
454 gcAssistAlloc1(gp, scanWork)
455 // The user stack may have moved, so this can't touch
456 // anything on it until it returns from systemstack.
459 completed := gp.param != nil
465 if gp.gcAssistBytes < 0 {
466 // We were unable steal enough credit or perform
467 // enough work to pay off the assist debt. We need to
468 // do one of these before letting the mutator allocate
469 // more to prevent over-allocation.
471 // If this is because we were preempted, reschedule
472 // and try some more.
478 // Add this G to an assist queue and park. When the GC
479 // has more background credit, it will satisfy queued
480 // assists before flushing to the global credit pool.
482 // Note that this does *not* get woken up when more
483 // work is added to the work list. The theory is that
484 // there wasn't enough work to do anyway, so we might
485 // as well let background marking take care of the
486 // work that is available.
491 // At this point either background GC has satisfied
492 // this G's assist debt, or the GC cycle is over.
495 traceGCMarkAssistDone()
499 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
500 // stack. This is a separate function to make it easier to see that
501 // we're not capturing anything from the user stack, since the user
502 // stack may move while we're in this function.
504 // gcAssistAlloc1 indicates whether this assist completed the mark
505 // phase by setting gp.param to non-nil. This can't be communicated on
506 // the stack since it may move.
509 func gcAssistAlloc1(gp *g, scanWork int64) {
510 // Clear the flag indicating that this assist completed the
514 if atomic.Load(&gcBlackenEnabled) == 0 {
515 // The gcBlackenEnabled check in malloc races with the
516 // store that clears it but an atomic check in every malloc
517 // would be a performance hit.
518 // Instead we recheck it here on the non-preemptable system
519 // stack to determine if we should perform an assist.
521 // GC is done, so ignore any remaining debt.
525 // Track time spent in this assist. Since we're on the
526 // system stack, this is non-preemptible, so we can
527 // just measure start and end time.
528 startTime := nanotime()
530 decnwait := atomic.Xadd(&work.nwait, -1)
531 if decnwait == work.nproc {
532 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
533 throw("nwait > work.nprocs")
536 // gcDrainN requires the caller to be preemptible.
537 casgstatus(gp, _Grunning, _Gwaiting)
538 gp.waitreason = waitReasonGCAssistMarking
540 // drain own cached work first in the hopes that it
541 // will be more cache friendly.
542 gcw := &getg().m.p.ptr().gcw
543 workDone := gcDrainN(gcw, scanWork)
545 casgstatus(gp, _Gwaiting, _Grunning)
547 // Record that we did this much scan work.
549 // Back out the number of bytes of assist credit that
550 // this scan work counts for. The "1+" is a poor man's
551 // round-up, to ensure this adds credit even if
552 // assistBytesPerWork is very low.
553 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
554 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
556 // If this is the last worker and we ran out of work,
557 // signal a completion point.
558 incnwait := atomic.Xadd(&work.nwait, +1)
559 if incnwait > work.nproc {
560 println("runtime: work.nwait=", incnwait,
561 "work.nproc=", work.nproc)
562 throw("work.nwait > work.nproc")
565 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
566 // This has reached a background completion point. Set
567 // gp.param to a non-nil value to indicate this. It
568 // doesn't matter what we set it to (it just has to be
570 gp.param = unsafe.Pointer(gp)
572 duration := nanotime() - startTime
574 _p_.gcAssistTime += duration
575 if _p_.gcAssistTime > gcAssistTimeSlack {
576 atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
581 // gcWakeAllAssists wakes all currently blocked assists. This is used
582 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
583 // new assists from going to sleep after this point.
584 func gcWakeAllAssists() {
585 lock(&work.assistQueue.lock)
586 list := work.assistQueue.q.popList()
588 unlock(&work.assistQueue.lock)
591 // gcParkAssist puts the current goroutine on the assist queue and parks.
593 // gcParkAssist reports whether the assist is now satisfied. If it
594 // returns false, the caller must retry the assist.
597 func gcParkAssist() bool {
598 lock(&work.assistQueue.lock)
599 // If the GC cycle finished while we were getting the lock,
600 // exit the assist. The cycle can't finish while we hold the
602 if atomic.Load(&gcBlackenEnabled) == 0 {
603 unlock(&work.assistQueue.lock)
608 oldList := work.assistQueue.q
609 work.assistQueue.q.pushBack(gp)
611 // Recheck for background credit now that this G is in
612 // the queue, but can still back out. This avoids a
613 // race in case background marking has flushed more
614 // credit since we checked above.
615 if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
616 work.assistQueue.q = oldList
617 if oldList.tail != 0 {
618 oldList.tail.ptr().schedlink.set(nil)
620 unlock(&work.assistQueue.lock)
624 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
628 // gcFlushBgCredit flushes scanWork units of background scan work
629 // credit. This first satisfies blocked assists on the
630 // work.assistQueue and then flushes any remaining credit to
631 // gcController.bgScanCredit.
633 // Write barriers are disallowed because this is used by gcDrain after
634 // it has ensured that all work is drained and this must preserve that
637 //go:nowritebarrierrec
638 func gcFlushBgCredit(scanWork int64) {
639 if work.assistQueue.q.empty() {
640 // Fast path; there are no blocked assists. There's a
641 // small window here where an assist may add itself to
642 // the blocked queue and park. If that happens, we'll
643 // just get it on the next flush.
644 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
648 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
649 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
651 lock(&work.assistQueue.lock)
652 for !work.assistQueue.q.empty() && scanBytes > 0 {
653 gp := work.assistQueue.q.pop()
654 // Note that gp.gcAssistBytes is negative because gp
655 // is in debt. Think carefully about the signs below.
656 if scanBytes+gp.gcAssistBytes >= 0 {
657 // Satisfy this entire assist debt.
658 scanBytes += gp.gcAssistBytes
660 // It's important that we *not* put gp in
661 // runnext. Otherwise, it's possible for user
662 // code to exploit the GC worker's high
663 // scheduler priority to get itself always run
664 // before other goroutines and always in the
665 // fresh quantum started by GC.
668 // Partially satisfy this assist.
669 gp.gcAssistBytes += scanBytes
671 // As a heuristic, we move this assist to the
672 // back of the queue so that large assists
673 // can't clog up the assist queue and
674 // substantially delay small assists.
675 work.assistQueue.q.pushBack(gp)
681 // Convert from scan bytes back to work.
682 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
683 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
684 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
686 unlock(&work.assistQueue.lock)
689 // scanstack scans gp's stack, greying all pointers found on the stack.
691 // scanstack will also shrink the stack if it is safe to do so. If it
692 // is not, it schedules a stack shrink for the next synchronous safe
695 // scanstack is marked go:systemstack because it must not be preempted
696 // while using a workbuf.
700 func scanstack(gp *g, gcw *gcWork) {
701 if readgstatus(gp)&_Gscan == 0 {
702 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
703 throw("scanstack - bad status")
706 switch readgstatus(gp) &^ _Gscan {
708 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
709 throw("mark - bad status")
713 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
714 throw("scanstack: goroutine not stopped")
715 case _Grunnable, _Gsyscall, _Gwaiting:
720 throw("can't scan our own stack")
723 if isShrinkStackSafe(gp) {
724 // Shrink the stack if not much of it is being used.
727 // Otherwise, shrink the stack at the next sync safe point.
728 gp.preemptShrink = true
731 var state stackScanState
732 state.stack = gp.stack
735 println("stack trace goroutine", gp.goid)
738 if debugScanConservative && gp.asyncSafePoint {
739 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
742 // Scan the saved context register. This is effectively a live
743 // register that gets moved back and forth between the
744 // register and sched.ctxt without a write barrier.
745 if gp.sched.ctxt != nil {
746 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), sys.PtrSize, &oneptrmask[0], gcw, &state)
749 // Scan the stack. Accumulate a list of stack objects.
750 scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
751 scanframeworker(frame, &state, gcw)
754 gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
756 // Find additional pointers that point into the stack from the heap.
757 // Currently this includes defers and panics. See also function copystack.
759 // Find and trace all defer arguments.
760 tracebackdefers(gp, scanframe, nil)
762 // Find and trace other pointers in defer records.
763 for d := gp._defer; d != nil; d = d.link {
765 // tracebackdefers above does not scan the func value, which could
766 // be a stack allocated closure. See issue 30453.
767 scanblock(uintptr(unsafe.Pointer(&d.fn)), sys.PtrSize, &oneptrmask[0], gcw, &state)
770 // The link field of a stack-allocated defer record might point
771 // to a heap-allocated defer record. Keep that heap record live.
772 scanblock(uintptr(unsafe.Pointer(&d.link)), sys.PtrSize, &oneptrmask[0], gcw, &state)
774 // Retain defers records themselves.
775 // Defer records might not be reachable from the G through regular heap
776 // tracing because the defer linked list might weave between the stack and the heap.
778 scanblock(uintptr(unsafe.Pointer(&d)), sys.PtrSize, &oneptrmask[0], gcw, &state)
781 if gp._panic != nil {
782 // Panics are always stack allocated.
783 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
786 // Find and scan all reachable stack objects.
788 // The state's pointer queue prioritizes precise pointers over
789 // conservative pointers so that we'll prefer scanning stack
790 // objects precisely.
793 p, conservative := state.getPtr()
797 obj := state.findObject(p)
803 // We've already scanned this object.
806 obj.setType(nil) // Don't scan it again.
809 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of type", t.string())
811 print(" (conservative)")
818 if t.kind&kindGCProg != 0 {
819 // This path is pretty unlikely, an object large enough
820 // to have a GC program allocated on the stack.
821 // We need some space to unpack the program into a straight
822 // bitmask, which we allocate/free here.
823 // TODO: it would be nice if there were a way to run a GC
824 // program without having to store all its bits. We'd have
825 // to change from a Lempel-Ziv style program to something else.
826 // Or we can forbid putting objects on stacks if they require
827 // a gc program (see issue 27447).
828 s = materializeGCProg(t.ptrdata, gcdata)
829 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
832 b := state.stack.lo + uintptr(obj.off)
834 scanConservative(b, t.ptrdata, gcdata, gcw, &state)
836 scanblock(b, t.ptrdata, gcdata, gcw, &state)
840 dematerializeGCProg(s)
844 // Deallocate object buffers.
845 // (Pointer buffers were all deallocated in the loop above.)
846 for state.head != nil {
850 for i := 0; i < x.nobj; i++ {
852 if obj.typ == nil { // reachable
855 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of type", obj.typ.string())
856 // Note: not necessarily really dead - only reachable-from-ptr dead.
860 putempty((*workbuf)(unsafe.Pointer(x)))
862 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
863 throw("remaining pointer buffers")
867 // Scan a stack frame: local variables and function arguments/results.
869 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
870 if _DebugGC > 1 && frame.continpc != 0 {
871 print("scanframe ", funcname(frame.fn), "\n")
874 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
875 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV1
876 if state.conservative || isAsyncPreempt || isDebugCall {
877 if debugScanConservative {
878 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
881 // Conservatively scan the frame. Unlike the precise
882 // case, this includes the outgoing argument space
883 // since we may have stopped while this function was
884 // setting up a call.
886 // TODO: We could narrow this down if the compiler
887 // produced a single map per function of stack slots
888 // and registers that ever contain a pointer.
890 size := frame.varp - frame.sp
892 scanConservative(frame.sp, size, nil, gcw, state)
896 // Scan arguments to this frame.
897 if frame.arglen != 0 {
898 // TODO: We could pass the entry argument map
899 // to narrow this down further.
900 scanConservative(frame.argp, frame.arglen, nil, gcw, state)
903 if isAsyncPreempt || isDebugCall {
904 // This function's frame contained the
905 // registers for the asynchronously stopped
906 // parent frame. Scan the parent
908 state.conservative = true
910 // We only wanted to scan those two frames
911 // conservatively. Clear the flag for future
913 state.conservative = false
918 locals, args, objs := getStackMap(frame, &state.cache, false)
920 // Scan local variables if stack frame has been allocated.
922 size := uintptr(locals.n) * sys.PtrSize
923 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
928 scanblock(frame.argp, uintptr(args.n)*sys.PtrSize, args.bytedata, gcw, state)
931 // Add all stack objects to the stack object list.
933 // varp is 0 for defers, where there are no locals.
934 // In that case, there can't be a pointer to its args, either.
935 // (And all args would be scanned above anyway.)
936 for _, obj := range objs {
938 base := frame.varp // locals base pointer
940 base = frame.argp // arguments and return values base pointer
942 ptr := base + uintptr(off)
944 // object hasn't been allocated in the frame yet.
948 println("stkobj at", hex(ptr), "of type", obj.typ.string())
950 state.addObject(ptr, obj.typ)
955 type gcDrainFlags int
958 gcDrainUntilPreempt gcDrainFlags = 1 << iota
964 // gcDrain scans roots and objects in work buffers, blackening grey
965 // objects until it is unable to get more work. It may return before
966 // GC is done; it's the caller's responsibility to balance work from
969 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
972 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
975 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
976 // pollFractionalWorkerExit() returns true. This implies
979 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
980 // credit to gcController.bgScanCredit every gcCreditSlack units of
983 // gcDrain will always return if there is a pending STW.
986 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
987 if !writeBarrier.needed {
988 throw("gcDrain phase incorrect")
992 preemptible := flags&gcDrainUntilPreempt != 0
993 flushBgCredit := flags&gcDrainFlushBgCredit != 0
994 idle := flags&gcDrainIdle != 0
996 initScanWork := gcw.scanWork
998 // checkWork is the scan work before performing the next
999 // self-preempt check.
1000 checkWork := int64(1<<63 - 1)
1001 var check func() bool
1002 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
1003 checkWork = initScanWork + drainCheckThreshold
1006 } else if flags&gcDrainFractional != 0 {
1007 check = pollFractionalWorkerExit
1011 // Drain root marking jobs.
1012 if work.markrootNext < work.markrootJobs {
1013 // Stop if we're preemptible or if someone wants to STW.
1014 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1015 job := atomic.Xadd(&work.markrootNext, +1) - 1
1016 if job >= work.markrootJobs {
1020 if check != nil && check() {
1026 // Drain heap marking jobs.
1027 // Stop if we're preemptible or if someone wants to STW.
1028 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1029 // Try to keep work available on the global queue. We used to
1030 // check if there were waiting workers, but it's better to
1031 // just keep work available than to make workers wait. In the
1032 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1038 b := gcw.tryGetFast()
1042 // Flush the write barrier
1043 // buffer; this may create
1050 // Unable to get work.
1055 // Flush background scan work credit to the global
1056 // account if we've accumulated enough locally so
1057 // mutator assists can draw on it.
1058 if gcw.scanWork >= gcCreditSlack {
1059 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1061 gcFlushBgCredit(gcw.scanWork - initScanWork)
1064 checkWork -= gcw.scanWork
1068 checkWork += drainCheckThreshold
1069 if check != nil && check() {
1077 // Flush remaining scan work credit.
1078 if gcw.scanWork > 0 {
1079 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1081 gcFlushBgCredit(gcw.scanWork - initScanWork)
1087 // gcDrainN blackens grey objects until it has performed roughly
1088 // scanWork units of scan work or the G is preempted. This is
1089 // best-effort, so it may perform less work if it fails to get a work
1090 // buffer. Otherwise, it will perform at least n units of work, but
1091 // may perform more because scanning is always done in whole object
1092 // increments. It returns the amount of scan work performed.
1094 // The caller goroutine must be in a preemptible state (e.g.,
1095 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1096 // consequence, this must be called on the system stack.
1100 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1101 if !writeBarrier.needed {
1102 throw("gcDrainN phase incorrect")
1105 // There may already be scan work on the gcw, which we don't
1106 // want to claim was done by this call.
1107 workFlushed := -gcw.scanWork
1110 for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
1111 // See gcDrain comment.
1116 // This might be a good place to add prefetch code...
1117 // if(wbuf.nobj > 4) {
1118 // PREFETCH(wbuf->obj[wbuf.nobj - 3];
1121 b := gcw.tryGetFast()
1125 // Flush the write barrier buffer;
1126 // this may create more work.
1133 // Try to do a root job.
1135 // TODO: Assists should get credit for this
1137 if work.markrootNext < work.markrootJobs {
1138 job := atomic.Xadd(&work.markrootNext, +1) - 1
1139 if job < work.markrootJobs {
1144 // No heap or root jobs.
1149 // Flush background scan work credit.
1150 if gcw.scanWork >= gcCreditSlack {
1151 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1152 workFlushed += gcw.scanWork
1157 // Unlike gcDrain, there's no need to flush remaining work
1158 // here because this never flushes to bgScanCredit and
1159 // gcw.dispose will flush any remaining work to scanWork.
1161 return workFlushed + gcw.scanWork
1164 // scanblock scans b as scanobject would, but using an explicit
1165 // pointer bitmap instead of the heap bitmap.
1167 // This is used to scan non-heap roots, so it does not update
1168 // gcw.bytesMarked or gcw.scanWork.
1170 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1172 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1173 // Use local copies of original parameters, so that a stack trace
1174 // due to one of the throws below shows the original block
1179 for i := uintptr(0); i < n; {
1180 // Find bits for the next word.
1181 bits := uint32(*addb(ptrmask, i/(sys.PtrSize*8)))
1183 i += sys.PtrSize * 8
1186 for j := 0; j < 8 && i < n; j++ {
1188 // Same work as in scanobject; see comments there.
1189 p := *(*uintptr)(unsafe.Pointer(b + i))
1191 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1192 greyobject(obj, b, i, span, gcw, objIndex)
1193 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1194 stk.putPtr(p, false)
1204 // scanobject scans the object starting at b, adding pointers to gcw.
1205 // b must point to the beginning of a heap object or an oblet.
1206 // scanobject consults the GC bitmap for the pointer mask and the
1207 // spans for the size of the object.
1210 func scanobject(b uintptr, gcw *gcWork) {
1211 // Find the bits for b and the size of the object at b.
1213 // b is either the beginning of an object, in which case this
1214 // is the size of the object to scan, or it points to an
1215 // oblet, in which case we compute the size to scan below.
1216 hbits := heapBitsForAddr(b)
1217 s := spanOfUnchecked(b)
1220 throw("scanobject n == 0")
1223 if n > maxObletBytes {
1224 // Large object. Break into oblets for better
1225 // parallelism and lower latency.
1227 // It's possible this is a noscan object (not
1228 // from greyobject, but from other code
1229 // paths), in which case we must *not* enqueue
1230 // oblets since their bitmaps will be
1232 if s.spanclass.noscan() {
1233 // Bypass the whole scan.
1234 gcw.bytesMarked += uint64(n)
1238 // Enqueue the other oblets to scan later.
1239 // Some oblets may be in b's scalar tail, but
1240 // these will be marked as "no more pointers",
1241 // so we'll drop out immediately when we go to
1243 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1244 if !gcw.putFast(oblet) {
1250 // Compute the size of the oblet. Since this object
1251 // must be a large object, s.base() is the beginning
1253 n = s.base() + s.elemsize - b
1254 if n > maxObletBytes {
1260 for i = 0; i < n; i += sys.PtrSize {
1261 // Find bits for this word.
1263 // Avoid needless hbits.next() on last iteration.
1264 hbits = hbits.next()
1266 // Load bits once. See CL 22712 and issue 16973 for discussion.
1267 bits := hbits.bits()
1268 if bits&bitScan == 0 {
1269 break // no more pointers in this object
1271 if bits&bitPointer == 0 {
1272 continue // not a pointer
1275 // Work here is duplicated in scanblock and above.
1276 // If you make changes here, make changes there too.
1277 obj := *(*uintptr)(unsafe.Pointer(b + i))
1279 // At this point we have extracted the next potential pointer.
1280 // Quickly filter out nil and pointers back to the current object.
1281 if obj != 0 && obj-b >= n {
1282 // Test if obj points into the Go heap and, if so,
1285 // Note that it's possible for findObject to
1286 // fail if obj points to a just-allocated heap
1287 // object because of a race with growing the
1288 // heap. In this case, we know the object was
1289 // just allocated and hence will be marked by
1290 // allocation itself.
1291 if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
1292 greyobject(obj, b, i, span, gcw, objIndex)
1296 gcw.bytesMarked += uint64(n)
1297 gcw.scanWork += int64(i)
1300 // scanConservative scans block [b, b+n) conservatively, treating any
1301 // pointer-like value in the block as a pointer.
1303 // If ptrmask != nil, only words that are marked in ptrmask are
1304 // considered as potential pointers.
1306 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1307 // and may contain pointers to stack objects.
1308 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1309 if debugScanConservative {
1311 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1312 hexdumpWords(b, b+n, func(p uintptr) byte {
1314 word := (p - b) / sys.PtrSize
1315 bits := *addb(ptrmask, word/8)
1316 if (bits>>(word%8))&1 == 0 {
1321 val := *(*uintptr)(unsafe.Pointer(p))
1322 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1326 span := spanOfHeap(val)
1330 idx := span.objIndex(val)
1331 if span.isFree(idx) {
1339 for i := uintptr(0); i < n; i += sys.PtrSize {
1341 word := i / sys.PtrSize
1342 bits := *addb(ptrmask, word/8)
1344 // Skip 8 words (the loop increment will do the 8th)
1346 // This must be the first time we've
1347 // seen this word of ptrmask, so i
1348 // must be 8-word-aligned, but check
1349 // our reasoning just in case.
1350 if i%(sys.PtrSize*8) != 0 {
1351 throw("misaligned mask")
1353 i += sys.PtrSize*8 - sys.PtrSize
1356 if (bits>>(word%8))&1 == 0 {
1361 val := *(*uintptr)(unsafe.Pointer(b + i))
1363 // Check if val points into the stack.
1364 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1365 // val may point to a stack object. This
1366 // object may be dead from last cycle and
1367 // hence may contain pointers to unallocated
1368 // objects, but unlike heap objects we can't
1369 // tell if it's already dead. Hence, if all
1370 // pointers to this object are from
1371 // conservative scanning, we have to scan it
1372 // defensively, too.
1373 state.putPtr(val, true)
1377 // Check if val points to a heap span.
1378 span := spanOfHeap(val)
1383 // Check if val points to an allocated object.
1384 idx := span.objIndex(val)
1385 if span.isFree(idx) {
1389 // val points to an allocated object. Mark it.
1390 obj := span.base() + idx*span.elemsize
1391 greyobject(obj, b, i, span, gcw, idx)
1395 // Shade the object if it isn't already.
1396 // The object is not nil and known to be in the heap.
1397 // Preemption must be disabled.
1399 func shade(b uintptr) {
1400 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1401 gcw := &getg().m.p.ptr().gcw
1402 greyobject(obj, 0, 0, span, gcw, objIndex)
1406 // obj is the start of an object with mark mbits.
1407 // If it isn't already marked, mark it and enqueue into gcw.
1408 // base and off are for debugging only and could be removed.
1410 // See also wbBufFlush1, which partially duplicates this logic.
1412 //go:nowritebarrierrec
1413 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1414 // obj should be start of allocation, and so must be at least pointer-aligned.
1415 if obj&(sys.PtrSize-1) != 0 {
1416 throw("greyobject: obj not pointer-aligned")
1418 mbits := span.markBitsForIndex(objIndex)
1421 if setCheckmark(obj, base, off, mbits) {
1426 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1427 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1428 gcDumpObject("base", base, off)
1429 gcDumpObject("obj", obj, ^uintptr(0))
1430 getg().m.traceback = 2
1431 throw("marking free object")
1434 // If marked we have nothing to do.
1435 if mbits.isMarked() {
1441 arena, pageIdx, pageMask := pageIndexOf(span.base())
1442 if arena.pageMarks[pageIdx]&pageMask == 0 {
1443 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1446 // If this is a noscan object, fast-track it to black
1447 // instead of greying it.
1448 if span.spanclass.noscan() {
1449 gcw.bytesMarked += uint64(span.elemsize)
1454 // Queue the obj for scanning. The PREFETCH(obj) logic has been removed but
1455 // seems like a nice optimization that can be added back in.
1456 // There needs to be time between the PREFETCH and the use.
1457 // Previously we put the obj in an 8 element buffer that is drained at a rate
1458 // to give the PREFETCH time to do its work.
1459 // Use of PREFETCHNTA might be more appropriate than PREFETCH
1460 if !gcw.putFast(obj) {
1465 // gcDumpObject dumps the contents of obj for debugging and marks the
1466 // field at byte offset off in obj.
1467 func gcDumpObject(label string, obj, off uintptr) {
1469 print(label, "=", hex(obj))
1474 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1475 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1476 print(mSpanStateNames[state], "\n")
1478 print("unknown(", state, ")\n")
1483 if s.state.get() == mSpanManual && size == 0 {
1484 // We're printing something from a stack frame. We
1485 // don't know how big it is, so just show up to an
1487 size = off + sys.PtrSize
1489 for i := uintptr(0); i < size; i += sys.PtrSize {
1490 // For big objects, just print the beginning (because
1491 // that usually hints at the object's type) and the
1492 // fields around off.
1493 if !(i < 128*sys.PtrSize || off-16*sys.PtrSize < i && i < off+16*sys.PtrSize) {
1501 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1512 // gcmarknewobject marks a newly allocated object black. obj must
1513 // not contain any non-nil pointers.
1515 // This is nosplit so it can manipulate a gcWork without preemption.
1519 func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
1520 if useCheckmark { // The world should be stopped so this should not happen.
1521 throw("gcmarknewobject called while doing checkmark")
1525 objIndex := span.objIndex(obj)
1526 span.markBitsForIndex(objIndex).setMarked()
1529 arena, pageIdx, pageMask := pageIndexOf(span.base())
1530 if arena.pageMarks[pageIdx]&pageMask == 0 {
1531 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1534 gcw := &getg().m.p.ptr().gcw
1535 gcw.bytesMarked += uint64(size)
1536 gcw.scanWork += int64(scanSize)
1539 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1541 // The world must be stopped.
1542 func gcMarkTinyAllocs() {
1543 assertWorldStopped()
1545 for _, p := range allp {
1547 if c == nil || c.tiny == 0 {
1550 _, span, objIndex := findObject(c.tiny, 0, 0)
1552 greyobject(c.tiny, 0, 0, span, gcw, objIndex)