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)
111 // gcMarkRootCheck checks that all roots have been scanned. It is
112 // purely for debugging.
113 func gcMarkRootCheck() {
114 if work.markrootNext < work.markrootJobs {
115 print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
116 throw("left over markroot jobs")
120 // Check that stacks have been scanned.
122 for i := 0; i < work.nStackRoots; i++ {
132 println("gp", gp, "goid", gp.goid,
133 "status", readgstatus(gp),
134 "gcscandone", gp.gcscandone)
135 unlock(&allglock) // Avoid self-deadlock with traceback.
136 throw("scan missed a g")
139 // ptrmask for an allocation containing a single pointer.
140 var oneptrmask = [...]uint8{1}
142 // markroot scans the i'th root.
144 // Preemption must be disabled (because this uses a gcWork).
146 // nowritebarrier is only advisory here.
149 func markroot(gcw *gcWork, i uint32) {
150 // TODO(austin): This is a bit ridiculous. Compute and store
151 // the bases in gcMarkRootPrepare instead of the counts.
152 baseFlushCache := uint32(fixedRootCount)
153 baseData := baseFlushCache + uint32(work.nFlushCacheRoots)
154 baseBSS := baseData + uint32(work.nDataRoots)
155 baseSpans := baseBSS + uint32(work.nBSSRoots)
156 baseStacks := baseSpans + uint32(work.nSpanRoots)
157 end := baseStacks + uint32(work.nStackRoots)
159 // Note: if you add a case here, please also update heapdump.go:dumproots.
161 case baseFlushCache <= i && i < baseData:
162 flushmcache(int(i - baseFlushCache))
164 case baseData <= i && i < baseBSS:
165 for _, datap := range activeModules() {
166 markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-baseData))
169 case baseBSS <= i && i < baseSpans:
170 for _, datap := range activeModules() {
171 markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-baseBSS))
174 case i == fixedRootFinalizers:
175 for fb := allfin; fb != nil; fb = fb.alllink {
176 cnt := uintptr(atomic.Load(&fb.cnt))
177 scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
180 case i == fixedRootFreeGStacks:
181 // Switch to the system stack so we can call
183 systemstack(markrootFreeGStacks)
185 case baseSpans <= i && i < baseStacks:
186 // mark mspan.specials
187 markrootSpans(gcw, int(i-baseSpans))
190 // the rest is scanning goroutine stacks
192 if baseStacks <= i && i < end {
193 gp = allgs[i-baseStacks]
195 throw("markroot: bad index")
198 // remember when we've first observed the G blocked
199 // needed only to output in traceback
200 status := readgstatus(gp) // We are not in a scan state
201 if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
202 gp.waitsince = work.tstart
205 // scanstack must be done on the system stack in case
206 // we're trying to scan our own stack.
208 // If this is a self-scan, put the user G in
209 // _Gwaiting to prevent self-deadlock. It may
210 // already be in _Gwaiting if this is a mark
211 // worker or we're in mark termination.
212 userG := getg().m.curg
213 selfScan := gp == userG && readgstatus(userG) == _Grunning
215 casgstatus(userG, _Grunning, _Gwaiting)
216 userG.waitreason = waitReasonGarbageCollectionScan
219 // TODO: suspendG blocks (and spins) until gp
220 // stops, which may take a while for
221 // running goroutines. Consider doing this in
222 // two phases where the first is non-blocking:
223 // we scan the stacks we can and ask running
224 // goroutines to scan themselves; and the
226 stopped := suspendG(gp)
232 throw("g already scanned")
239 casgstatus(userG, _Gwaiting, _Grunning)
245 // markrootBlock scans the shard'th shard of the block of memory [b0,
246 // b0+n0), with the given pointer mask.
249 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) {
250 if rootBlockBytes%(8*sys.PtrSize) != 0 {
251 // This is necessary to pick byte offsets in ptrmask0.
252 throw("rootBlockBytes must be a multiple of 8*ptrSize")
255 // Note that if b0 is toward the end of the address space,
256 // then b0 + rootBlockBytes might wrap around.
257 // These tests are written to avoid any possible overflow.
258 off := uintptr(shard) * rootBlockBytes
263 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*sys.PtrSize))))
264 n := uintptr(rootBlockBytes)
270 scanblock(b, n, ptrmask, gcw, nil)
273 // markrootFreeGStacks frees stacks of dead Gs.
275 // This does not free stacks of dead Gs cached on Ps, but having a few
276 // cached stacks around isn't a problem.
277 func markrootFreeGStacks() {
278 // Take list of dead Gs with stacks.
279 lock(&sched.gFree.lock)
280 list := sched.gFree.stack
281 sched.gFree.stack = gList{}
282 unlock(&sched.gFree.lock)
288 q := gQueue{list.head, list.head}
289 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
293 // Manipulate the queue directly since the Gs are
294 // already all linked the right way.
298 // Put Gs back on the free list.
299 lock(&sched.gFree.lock)
300 sched.gFree.noStack.pushAll(q)
301 unlock(&sched.gFree.lock)
304 // markrootSpans marks roots for one shard of markArenas.
307 func markrootSpans(gcw *gcWork, shard int) {
308 // Objects with finalizers have two GC-related invariants:
310 // 1) Everything reachable from the object must be marked.
311 // This ensures that when we pass the object to its finalizer,
312 // everything the finalizer can reach will be retained.
314 // 2) Finalizer specials (which are not in the garbage
315 // collected heap) are roots. In practice, this means the fn
316 // field must be scanned.
317 sg := mheap_.sweepgen
319 // Find the arena and page index into that arena for this shard.
320 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
321 ha := mheap_.arenas[ai.l1()][ai.l2()]
322 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
324 // Construct slice of bitmap which we'll iterate over.
325 specialsbits := ha.pageSpecials[arenaPage/8:]
326 specialsbits = specialsbits[:pagesPerSpanRoot/8]
327 for i := range specialsbits {
328 // Find set bits, which correspond to spans with specials.
329 specials := atomic.Load8(&specialsbits[i])
333 for j := uint(0); j < 8; j++ {
334 if specials&(1<<j) == 0 {
337 // Find the span for this bit.
339 // This value is guaranteed to be non-nil because having
340 // specials implies that the span is in-use, and since we're
341 // currently marking we can be sure that we don't have to worry
342 // about the span being freed and re-used.
343 s := ha.spans[arenaPage+uint(i)*8+j]
345 // The state must be mSpanInUse if the specials bit is set, so
346 // sanity check that.
347 if state := s.state.get(); state != mSpanInUse {
348 print("s.state = ", state, "\n")
349 throw("non in-use span found with specials bit set")
351 // Check that this span was swept (it may be cached or uncached).
352 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
353 // sweepgen was updated (+2) during non-checkmark GC pass
354 print("sweep ", s.sweepgen, " ", sg, "\n")
355 throw("gc: unswept span")
358 // Lock the specials to prevent a special from being
359 // removed from the list while we're traversing it.
361 for sp := s.specials; sp != nil; sp = sp.next {
362 if sp.kind != _KindSpecialFinalizer {
365 // don't mark finalized object, but scan it so we
366 // retain everything it points to.
367 spf := (*specialfinalizer)(unsafe.Pointer(sp))
368 // A finalizer can be set for an inner byte of an object, find object beginning.
369 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
371 // Mark everything that can be reached from
372 // the object (but *not* the object itself or
373 // we'll never collect it).
376 // The special itself is a root.
377 scanblock(uintptr(unsafe.Pointer(&spf.fn)), sys.PtrSize, &oneptrmask[0], gcw, nil)
379 unlock(&s.speciallock)
384 // gcAssistAlloc performs GC work to make gp's assist debt positive.
385 // gp must be the calling user gorountine.
387 // This must be called with preemption enabled.
388 func gcAssistAlloc(gp *g) {
389 // Don't assist in non-preemptible contexts. These are
390 // generally fragile and won't allow the assist to block.
391 if getg() == gp.m.g0 {
394 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
400 // Compute the amount of scan work we need to do to make the
401 // balance positive. When the required amount of work is low,
402 // we over-assist to build up credit for future allocations
403 // and amortize the cost of assisting.
404 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
405 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
406 debtBytes := -gp.gcAssistBytes
407 scanWork := int64(assistWorkPerByte * float64(debtBytes))
408 if scanWork < gcOverAssistWork {
409 scanWork = gcOverAssistWork
410 debtBytes = int64(assistBytesPerWork * float64(scanWork))
413 // Steal as much credit as we can from the background GC's
414 // scan credit. This is racy and may drop the background
415 // credit below 0 if two mutators steal at the same time. This
416 // will just cause steals to fail until credit is accumulated
417 // again, so in the long run it doesn't really matter, but we
418 // do have to handle the negative credit case.
419 bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
421 if bgScanCredit > 0 {
422 if bgScanCredit < scanWork {
423 stolen = bgScanCredit
424 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
427 gp.gcAssistBytes += debtBytes
429 atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
434 // We were able to steal all of the credit we
437 traceGCMarkAssistDone()
443 if trace.enabled && !traced {
445 traceGCMarkAssistStart()
448 // Perform assist work
450 gcAssistAlloc1(gp, scanWork)
451 // The user stack may have moved, so this can't touch
452 // anything on it until it returns from systemstack.
455 completed := gp.param != nil
461 if gp.gcAssistBytes < 0 {
462 // We were unable steal enough credit or perform
463 // enough work to pay off the assist debt. We need to
464 // do one of these before letting the mutator allocate
465 // more to prevent over-allocation.
467 // If this is because we were preempted, reschedule
468 // and try some more.
474 // Add this G to an assist queue and park. When the GC
475 // has more background credit, it will satisfy queued
476 // assists before flushing to the global credit pool.
478 // Note that this does *not* get woken up when more
479 // work is added to the work list. The theory is that
480 // there wasn't enough work to do anyway, so we might
481 // as well let background marking take care of the
482 // work that is available.
487 // At this point either background GC has satisfied
488 // this G's assist debt, or the GC cycle is over.
491 traceGCMarkAssistDone()
495 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
496 // stack. This is a separate function to make it easier to see that
497 // we're not capturing anything from the user stack, since the user
498 // stack may move while we're in this function.
500 // gcAssistAlloc1 indicates whether this assist completed the mark
501 // phase by setting gp.param to non-nil. This can't be communicated on
502 // the stack since it may move.
505 func gcAssistAlloc1(gp *g, scanWork int64) {
506 // Clear the flag indicating that this assist completed the
510 if atomic.Load(&gcBlackenEnabled) == 0 {
511 // The gcBlackenEnabled check in malloc races with the
512 // store that clears it but an atomic check in every malloc
513 // would be a performance hit.
514 // Instead we recheck it here on the non-preemptable system
515 // stack to determine if we should perform an assist.
517 // GC is done, so ignore any remaining debt.
521 // Track time spent in this assist. Since we're on the
522 // system stack, this is non-preemptible, so we can
523 // just measure start and end time.
524 startTime := nanotime()
526 decnwait := atomic.Xadd(&work.nwait, -1)
527 if decnwait == work.nproc {
528 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
529 throw("nwait > work.nprocs")
532 // gcDrainN requires the caller to be preemptible.
533 casgstatus(gp, _Grunning, _Gwaiting)
534 gp.waitreason = waitReasonGCAssistMarking
536 // drain own cached work first in the hopes that it
537 // will be more cache friendly.
538 gcw := &getg().m.p.ptr().gcw
539 workDone := gcDrainN(gcw, scanWork)
541 casgstatus(gp, _Gwaiting, _Grunning)
543 // Record that we did this much scan work.
545 // Back out the number of bytes of assist credit that
546 // this scan work counts for. The "1+" is a poor man's
547 // round-up, to ensure this adds credit even if
548 // assistBytesPerWork is very low.
549 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
550 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
552 // If this is the last worker and we ran out of work,
553 // signal a completion point.
554 incnwait := atomic.Xadd(&work.nwait, +1)
555 if incnwait > work.nproc {
556 println("runtime: work.nwait=", incnwait,
557 "work.nproc=", work.nproc)
558 throw("work.nwait > work.nproc")
561 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
562 // This has reached a background completion point. Set
563 // gp.param to a non-nil value to indicate this. It
564 // doesn't matter what we set it to (it just has to be
566 gp.param = unsafe.Pointer(gp)
568 duration := nanotime() - startTime
570 _p_.gcAssistTime += duration
571 if _p_.gcAssistTime > gcAssistTimeSlack {
572 atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
577 // gcWakeAllAssists wakes all currently blocked assists. This is used
578 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
579 // new assists from going to sleep after this point.
580 func gcWakeAllAssists() {
581 lock(&work.assistQueue.lock)
582 list := work.assistQueue.q.popList()
584 unlock(&work.assistQueue.lock)
587 // gcParkAssist puts the current goroutine on the assist queue and parks.
589 // gcParkAssist reports whether the assist is now satisfied. If it
590 // returns false, the caller must retry the assist.
593 func gcParkAssist() bool {
594 lock(&work.assistQueue.lock)
595 // If the GC cycle finished while we were getting the lock,
596 // exit the assist. The cycle can't finish while we hold the
598 if atomic.Load(&gcBlackenEnabled) == 0 {
599 unlock(&work.assistQueue.lock)
604 oldList := work.assistQueue.q
605 work.assistQueue.q.pushBack(gp)
607 // Recheck for background credit now that this G is in
608 // the queue, but can still back out. This avoids a
609 // race in case background marking has flushed more
610 // credit since we checked above.
611 if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
612 work.assistQueue.q = oldList
613 if oldList.tail != 0 {
614 oldList.tail.ptr().schedlink.set(nil)
616 unlock(&work.assistQueue.lock)
620 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
624 // gcFlushBgCredit flushes scanWork units of background scan work
625 // credit. This first satisfies blocked assists on the
626 // work.assistQueue and then flushes any remaining credit to
627 // gcController.bgScanCredit.
629 // Write barriers are disallowed because this is used by gcDrain after
630 // it has ensured that all work is drained and this must preserve that
633 //go:nowritebarrierrec
634 func gcFlushBgCredit(scanWork int64) {
635 if work.assistQueue.q.empty() {
636 // Fast path; there are no blocked assists. There's a
637 // small window here where an assist may add itself to
638 // the blocked queue and park. If that happens, we'll
639 // just get it on the next flush.
640 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
644 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
645 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
647 lock(&work.assistQueue.lock)
648 for !work.assistQueue.q.empty() && scanBytes > 0 {
649 gp := work.assistQueue.q.pop()
650 // Note that gp.gcAssistBytes is negative because gp
651 // is in debt. Think carefully about the signs below.
652 if scanBytes+gp.gcAssistBytes >= 0 {
653 // Satisfy this entire assist debt.
654 scanBytes += gp.gcAssistBytes
656 // It's important that we *not* put gp in
657 // runnext. Otherwise, it's possible for user
658 // code to exploit the GC worker's high
659 // scheduler priority to get itself always run
660 // before other goroutines and always in the
661 // fresh quantum started by GC.
664 // Partially satisfy this assist.
665 gp.gcAssistBytes += scanBytes
667 // As a heuristic, we move this assist to the
668 // back of the queue so that large assists
669 // can't clog up the assist queue and
670 // substantially delay small assists.
671 work.assistQueue.q.pushBack(gp)
677 // Convert from scan bytes back to work.
678 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
679 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
680 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
682 unlock(&work.assistQueue.lock)
685 // scanstack scans gp's stack, greying all pointers found on the stack.
687 // scanstack will also shrink the stack if it is safe to do so. If it
688 // is not, it schedules a stack shrink for the next synchronous safe
691 // scanstack is marked go:systemstack because it must not be preempted
692 // while using a workbuf.
696 func scanstack(gp *g, gcw *gcWork) {
697 if readgstatus(gp)&_Gscan == 0 {
698 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
699 throw("scanstack - bad status")
702 switch readgstatus(gp) &^ _Gscan {
704 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
705 throw("mark - bad status")
709 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
710 throw("scanstack: goroutine not stopped")
711 case _Grunnable, _Gsyscall, _Gwaiting:
716 throw("can't scan our own stack")
719 if isShrinkStackSafe(gp) {
720 // Shrink the stack if not much of it is being used.
723 // Otherwise, shrink the stack at the next sync safe point.
724 gp.preemptShrink = true
727 var state stackScanState
728 state.stack = gp.stack
731 println("stack trace goroutine", gp.goid)
734 if debugScanConservative && gp.asyncSafePoint {
735 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
738 // Scan the saved context register. This is effectively a live
739 // register that gets moved back and forth between the
740 // register and sched.ctxt without a write barrier.
741 if gp.sched.ctxt != nil {
742 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), sys.PtrSize, &oneptrmask[0], gcw, &state)
745 // Scan the stack. Accumulate a list of stack objects.
746 scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
747 scanframeworker(frame, &state, gcw)
750 gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
752 // Find additional pointers that point into the stack from the heap.
753 // Currently this includes defers and panics. See also function copystack.
755 // Find and trace all defer arguments.
756 tracebackdefers(gp, scanframe, nil)
758 // Find and trace other pointers in defer records.
759 for d := gp._defer; d != nil; d = d.link {
761 // tracebackdefers above does not scan the func value, which could
762 // be a stack allocated closure. See issue 30453.
763 scanblock(uintptr(unsafe.Pointer(&d.fn)), sys.PtrSize, &oneptrmask[0], gcw, &state)
766 // The link field of a stack-allocated defer record might point
767 // to a heap-allocated defer record. Keep that heap record live.
768 scanblock(uintptr(unsafe.Pointer(&d.link)), sys.PtrSize, &oneptrmask[0], gcw, &state)
770 // Retain defers records themselves.
771 // Defer records might not be reachable from the G through regular heap
772 // tracing because the defer linked list might weave between the stack and the heap.
774 scanblock(uintptr(unsafe.Pointer(&d)), sys.PtrSize, &oneptrmask[0], gcw, &state)
777 if gp._panic != nil {
778 // Panics are always stack allocated.
779 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
782 // Find and scan all reachable stack objects.
784 // The state's pointer queue prioritizes precise pointers over
785 // conservative pointers so that we'll prefer scanning stack
786 // objects precisely.
789 p, conservative := state.getPtr()
793 obj := state.findObject(p)
799 // We've already scanned this object.
802 obj.setType(nil) // Don't scan it again.
805 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of type", t.string())
807 print(" (conservative)")
814 if t.kind&kindGCProg != 0 {
815 // This path is pretty unlikely, an object large enough
816 // to have a GC program allocated on the stack.
817 // We need some space to unpack the program into a straight
818 // bitmask, which we allocate/free here.
819 // TODO: it would be nice if there were a way to run a GC
820 // program without having to store all its bits. We'd have
821 // to change from a Lempel-Ziv style program to something else.
822 // Or we can forbid putting objects on stacks if they require
823 // a gc program (see issue 27447).
824 s = materializeGCProg(t.ptrdata, gcdata)
825 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
828 b := state.stack.lo + uintptr(obj.off)
830 scanConservative(b, t.ptrdata, gcdata, gcw, &state)
832 scanblock(b, t.ptrdata, gcdata, gcw, &state)
836 dematerializeGCProg(s)
840 // Deallocate object buffers.
841 // (Pointer buffers were all deallocated in the loop above.)
842 for state.head != nil {
846 for i := 0; i < x.nobj; i++ {
848 if obj.typ == nil { // reachable
851 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of type", obj.typ.string())
852 // Note: not necessarily really dead - only reachable-from-ptr dead.
856 putempty((*workbuf)(unsafe.Pointer(x)))
858 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
859 throw("remaining pointer buffers")
863 // Scan a stack frame: local variables and function arguments/results.
865 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
866 if _DebugGC > 1 && frame.continpc != 0 {
867 print("scanframe ", funcname(frame.fn), "\n")
870 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
871 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV1
872 if state.conservative || isAsyncPreempt || isDebugCall {
873 if debugScanConservative {
874 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
877 // Conservatively scan the frame. Unlike the precise
878 // case, this includes the outgoing argument space
879 // since we may have stopped while this function was
880 // setting up a call.
882 // TODO: We could narrow this down if the compiler
883 // produced a single map per function of stack slots
884 // and registers that ever contain a pointer.
886 size := frame.varp - frame.sp
888 scanConservative(frame.sp, size, nil, gcw, state)
892 // Scan arguments to this frame.
893 if frame.arglen != 0 {
894 // TODO: We could pass the entry argument map
895 // to narrow this down further.
896 scanConservative(frame.argp, frame.arglen, nil, gcw, state)
899 if isAsyncPreempt || isDebugCall {
900 // This function's frame contained the
901 // registers for the asynchronously stopped
902 // parent frame. Scan the parent
904 state.conservative = true
906 // We only wanted to scan those two frames
907 // conservatively. Clear the flag for future
909 state.conservative = false
914 locals, args, objs := getStackMap(frame, &state.cache, false)
916 // Scan local variables if stack frame has been allocated.
918 size := uintptr(locals.n) * sys.PtrSize
919 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
924 scanblock(frame.argp, uintptr(args.n)*sys.PtrSize, args.bytedata, gcw, state)
927 // Add all stack objects to the stack object list.
929 // varp is 0 for defers, where there are no locals.
930 // In that case, there can't be a pointer to its args, either.
931 // (And all args would be scanned above anyway.)
932 for _, obj := range objs {
934 base := frame.varp // locals base pointer
936 base = frame.argp // arguments and return values base pointer
938 ptr := base + uintptr(off)
940 // object hasn't been allocated in the frame yet.
944 println("stkobj at", hex(ptr), "of type", obj.typ.string())
946 state.addObject(ptr, obj.typ)
951 type gcDrainFlags int
954 gcDrainUntilPreempt gcDrainFlags = 1 << iota
960 // gcDrain scans roots and objects in work buffers, blackening grey
961 // objects until it is unable to get more work. It may return before
962 // GC is done; it's the caller's responsibility to balance work from
965 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
968 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
971 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
972 // pollFractionalWorkerExit() returns true. This implies
975 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
976 // credit to gcController.bgScanCredit every gcCreditSlack units of
979 // gcDrain will always return if there is a pending STW.
982 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
983 if !writeBarrier.needed {
984 throw("gcDrain phase incorrect")
988 preemptible := flags&gcDrainUntilPreempt != 0
989 flushBgCredit := flags&gcDrainFlushBgCredit != 0
990 idle := flags&gcDrainIdle != 0
992 initScanWork := gcw.scanWork
994 // checkWork is the scan work before performing the next
995 // self-preempt check.
996 checkWork := int64(1<<63 - 1)
997 var check func() bool
998 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
999 checkWork = initScanWork + drainCheckThreshold
1002 } else if flags&gcDrainFractional != 0 {
1003 check = pollFractionalWorkerExit
1007 // Drain root marking jobs.
1008 if work.markrootNext < work.markrootJobs {
1009 // Stop if we're preemptible or if someone wants to STW.
1010 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1011 job := atomic.Xadd(&work.markrootNext, +1) - 1
1012 if job >= work.markrootJobs {
1016 if check != nil && check() {
1022 // Drain heap marking jobs.
1023 // Stop if we're preemptible or if someone wants to STW.
1024 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1025 // Try to keep work available on the global queue. We used to
1026 // check if there were waiting workers, but it's better to
1027 // just keep work available than to make workers wait. In the
1028 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1034 b := gcw.tryGetFast()
1038 // Flush the write barrier
1039 // buffer; this may create
1046 // Unable to get work.
1051 // Flush background scan work credit to the global
1052 // account if we've accumulated enough locally so
1053 // mutator assists can draw on it.
1054 if gcw.scanWork >= gcCreditSlack {
1055 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1057 gcFlushBgCredit(gcw.scanWork - initScanWork)
1060 checkWork -= gcw.scanWork
1064 checkWork += drainCheckThreshold
1065 if check != nil && check() {
1073 // Flush remaining scan work credit.
1074 if gcw.scanWork > 0 {
1075 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1077 gcFlushBgCredit(gcw.scanWork - initScanWork)
1083 // gcDrainN blackens grey objects until it has performed roughly
1084 // scanWork units of scan work or the G is preempted. This is
1085 // best-effort, so it may perform less work if it fails to get a work
1086 // buffer. Otherwise, it will perform at least n units of work, but
1087 // may perform more because scanning is always done in whole object
1088 // increments. It returns the amount of scan work performed.
1090 // The caller goroutine must be in a preemptible state (e.g.,
1091 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1092 // consequence, this must be called on the system stack.
1096 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1097 if !writeBarrier.needed {
1098 throw("gcDrainN phase incorrect")
1101 // There may already be scan work on the gcw, which we don't
1102 // want to claim was done by this call.
1103 workFlushed := -gcw.scanWork
1106 for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
1107 // See gcDrain comment.
1112 // This might be a good place to add prefetch code...
1113 // if(wbuf.nobj > 4) {
1114 // PREFETCH(wbuf->obj[wbuf.nobj - 3];
1117 b := gcw.tryGetFast()
1121 // Flush the write barrier buffer;
1122 // this may create more work.
1129 // Try to do a root job.
1131 // TODO: Assists should get credit for this
1133 if work.markrootNext < work.markrootJobs {
1134 job := atomic.Xadd(&work.markrootNext, +1) - 1
1135 if job < work.markrootJobs {
1140 // No heap or root jobs.
1145 // Flush background scan work credit.
1146 if gcw.scanWork >= gcCreditSlack {
1147 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1148 workFlushed += gcw.scanWork
1153 // Unlike gcDrain, there's no need to flush remaining work
1154 // here because this never flushes to bgScanCredit and
1155 // gcw.dispose will flush any remaining work to scanWork.
1157 return workFlushed + gcw.scanWork
1160 // scanblock scans b as scanobject would, but using an explicit
1161 // pointer bitmap instead of the heap bitmap.
1163 // This is used to scan non-heap roots, so it does not update
1164 // gcw.bytesMarked or gcw.scanWork.
1166 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1168 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1169 // Use local copies of original parameters, so that a stack trace
1170 // due to one of the throws below shows the original block
1175 for i := uintptr(0); i < n; {
1176 // Find bits for the next word.
1177 bits := uint32(*addb(ptrmask, i/(sys.PtrSize*8)))
1179 i += sys.PtrSize * 8
1182 for j := 0; j < 8 && i < n; j++ {
1184 // Same work as in scanobject; see comments there.
1185 p := *(*uintptr)(unsafe.Pointer(b + i))
1187 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1188 greyobject(obj, b, i, span, gcw, objIndex)
1189 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1190 stk.putPtr(p, false)
1200 // scanobject scans the object starting at b, adding pointers to gcw.
1201 // b must point to the beginning of a heap object or an oblet.
1202 // scanobject consults the GC bitmap for the pointer mask and the
1203 // spans for the size of the object.
1206 func scanobject(b uintptr, gcw *gcWork) {
1207 // Find the bits for b and the size of the object at b.
1209 // b is either the beginning of an object, in which case this
1210 // is the size of the object to scan, or it points to an
1211 // oblet, in which case we compute the size to scan below.
1212 hbits := heapBitsForAddr(b)
1213 s := spanOfUnchecked(b)
1216 throw("scanobject n == 0")
1219 if n > maxObletBytes {
1220 // Large object. Break into oblets for better
1221 // parallelism and lower latency.
1223 // It's possible this is a noscan object (not
1224 // from greyobject, but from other code
1225 // paths), in which case we must *not* enqueue
1226 // oblets since their bitmaps will be
1228 if s.spanclass.noscan() {
1229 // Bypass the whole scan.
1230 gcw.bytesMarked += uint64(n)
1234 // Enqueue the other oblets to scan later.
1235 // Some oblets may be in b's scalar tail, but
1236 // these will be marked as "no more pointers",
1237 // so we'll drop out immediately when we go to
1239 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1240 if !gcw.putFast(oblet) {
1246 // Compute the size of the oblet. Since this object
1247 // must be a large object, s.base() is the beginning
1249 n = s.base() + s.elemsize - b
1250 if n > maxObletBytes {
1256 for i = 0; i < n; i += sys.PtrSize {
1257 // Find bits for this word.
1259 // Avoid needless hbits.next() on last iteration.
1260 hbits = hbits.next()
1262 // Load bits once. See CL 22712 and issue 16973 for discussion.
1263 bits := hbits.bits()
1264 if bits&bitScan == 0 {
1265 break // no more pointers in this object
1267 if bits&bitPointer == 0 {
1268 continue // not a pointer
1271 // Work here is duplicated in scanblock and above.
1272 // If you make changes here, make changes there too.
1273 obj := *(*uintptr)(unsafe.Pointer(b + i))
1275 // At this point we have extracted the next potential pointer.
1276 // Quickly filter out nil and pointers back to the current object.
1277 if obj != 0 && obj-b >= n {
1278 // Test if obj points into the Go heap and, if so,
1281 // Note that it's possible for findObject to
1282 // fail if obj points to a just-allocated heap
1283 // object because of a race with growing the
1284 // heap. In this case, we know the object was
1285 // just allocated and hence will be marked by
1286 // allocation itself.
1287 if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
1288 greyobject(obj, b, i, span, gcw, objIndex)
1292 gcw.bytesMarked += uint64(n)
1293 gcw.scanWork += int64(i)
1296 // scanConservative scans block [b, b+n) conservatively, treating any
1297 // pointer-like value in the block as a pointer.
1299 // If ptrmask != nil, only words that are marked in ptrmask are
1300 // considered as potential pointers.
1302 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1303 // and may contain pointers to stack objects.
1304 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1305 if debugScanConservative {
1307 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1308 hexdumpWords(b, b+n, func(p uintptr) byte {
1310 word := (p - b) / sys.PtrSize
1311 bits := *addb(ptrmask, word/8)
1312 if (bits>>(word%8))&1 == 0 {
1317 val := *(*uintptr)(unsafe.Pointer(p))
1318 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1322 span := spanOfHeap(val)
1326 idx := span.objIndex(val)
1327 if span.isFree(idx) {
1335 for i := uintptr(0); i < n; i += sys.PtrSize {
1337 word := i / sys.PtrSize
1338 bits := *addb(ptrmask, word/8)
1340 // Skip 8 words (the loop increment will do the 8th)
1342 // This must be the first time we've
1343 // seen this word of ptrmask, so i
1344 // must be 8-word-aligned, but check
1345 // our reasoning just in case.
1346 if i%(sys.PtrSize*8) != 0 {
1347 throw("misaligned mask")
1349 i += sys.PtrSize*8 - sys.PtrSize
1352 if (bits>>(word%8))&1 == 0 {
1357 val := *(*uintptr)(unsafe.Pointer(b + i))
1359 // Check if val points into the stack.
1360 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1361 // val may point to a stack object. This
1362 // object may be dead from last cycle and
1363 // hence may contain pointers to unallocated
1364 // objects, but unlike heap objects we can't
1365 // tell if it's already dead. Hence, if all
1366 // pointers to this object are from
1367 // conservative scanning, we have to scan it
1368 // defensively, too.
1369 state.putPtr(val, true)
1373 // Check if val points to a heap span.
1374 span := spanOfHeap(val)
1379 // Check if val points to an allocated object.
1380 idx := span.objIndex(val)
1381 if span.isFree(idx) {
1385 // val points to an allocated object. Mark it.
1386 obj := span.base() + idx*span.elemsize
1387 greyobject(obj, b, i, span, gcw, idx)
1391 // Shade the object if it isn't already.
1392 // The object is not nil and known to be in the heap.
1393 // Preemption must be disabled.
1395 func shade(b uintptr) {
1396 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1397 gcw := &getg().m.p.ptr().gcw
1398 greyobject(obj, 0, 0, span, gcw, objIndex)
1402 // obj is the start of an object with mark mbits.
1403 // If it isn't already marked, mark it and enqueue into gcw.
1404 // base and off are for debugging only and could be removed.
1406 // See also wbBufFlush1, which partially duplicates this logic.
1408 //go:nowritebarrierrec
1409 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1410 // obj should be start of allocation, and so must be at least pointer-aligned.
1411 if obj&(sys.PtrSize-1) != 0 {
1412 throw("greyobject: obj not pointer-aligned")
1414 mbits := span.markBitsForIndex(objIndex)
1417 if setCheckmark(obj, base, off, mbits) {
1422 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1423 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1424 gcDumpObject("base", base, off)
1425 gcDumpObject("obj", obj, ^uintptr(0))
1426 getg().m.traceback = 2
1427 throw("marking free object")
1430 // If marked we have nothing to do.
1431 if mbits.isMarked() {
1437 arena, pageIdx, pageMask := pageIndexOf(span.base())
1438 if arena.pageMarks[pageIdx]&pageMask == 0 {
1439 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1442 // If this is a noscan object, fast-track it to black
1443 // instead of greying it.
1444 if span.spanclass.noscan() {
1445 gcw.bytesMarked += uint64(span.elemsize)
1450 // Queue the obj for scanning. The PREFETCH(obj) logic has been removed but
1451 // seems like a nice optimization that can be added back in.
1452 // There needs to be time between the PREFETCH and the use.
1453 // Previously we put the obj in an 8 element buffer that is drained at a rate
1454 // to give the PREFETCH time to do its work.
1455 // Use of PREFETCHNTA might be more appropriate than PREFETCH
1456 if !gcw.putFast(obj) {
1461 // gcDumpObject dumps the contents of obj for debugging and marks the
1462 // field at byte offset off in obj.
1463 func gcDumpObject(label string, obj, off uintptr) {
1465 print(label, "=", hex(obj))
1470 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1471 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1472 print(mSpanStateNames[state], "\n")
1474 print("unknown(", state, ")\n")
1479 if s.state.get() == mSpanManual && size == 0 {
1480 // We're printing something from a stack frame. We
1481 // don't know how big it is, so just show up to an
1483 size = off + sys.PtrSize
1485 for i := uintptr(0); i < size; i += sys.PtrSize {
1486 // For big objects, just print the beginning (because
1487 // that usually hints at the object's type) and the
1488 // fields around off.
1489 if !(i < 128*sys.PtrSize || off-16*sys.PtrSize < i && i < off+16*sys.PtrSize) {
1497 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1508 // gcmarknewobject marks a newly allocated object black. obj must
1509 // not contain any non-nil pointers.
1511 // This is nosplit so it can manipulate a gcWork without preemption.
1515 func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
1516 if useCheckmark { // The world should be stopped so this should not happen.
1517 throw("gcmarknewobject called while doing checkmark")
1521 objIndex := span.objIndex(obj)
1522 span.markBitsForIndex(objIndex).setMarked()
1525 arena, pageIdx, pageMask := pageIndexOf(span.base())
1526 if arena.pageMarks[pageIdx]&pageMask == 0 {
1527 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1530 gcw := &getg().m.p.ptr().gcw
1531 gcw.bytesMarked += uint64(size)
1532 gcw.scanWork += int64(scanSize)
1535 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1537 // The world must be stopped.
1538 func gcMarkTinyAllocs() {
1539 assertWorldStopped()
1541 for _, p := range allp {
1543 if c == nil || c.tiny == 0 {
1546 _, span, objIndex := findObject(c.tiny, 0, 0)
1548 greyobject(c.tiny, 0, 0, span, gcw, objIndex)