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"
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 // Compute how many data and BSS root blocks there are.
60 nBlocks := func(bytes uintptr) int {
61 return int(divRoundUp(bytes, rootBlockBytes))
68 for _, datap := range activeModules() {
69 nDataRoots := nBlocks(datap.edata - datap.data)
70 if nDataRoots > work.nDataRoots {
71 work.nDataRoots = nDataRoots
75 for _, datap := range activeModules() {
76 nBSSRoots := nBlocks(datap.ebss - datap.bss)
77 if nBSSRoots > work.nBSSRoots {
78 work.nBSSRoots = nBSSRoots
82 // Scan span roots for finalizer specials.
84 // We depend on addfinalizer to mark objects that get
85 // finalizers after root marking.
87 // We're going to scan the whole heap (that was available at the time the
88 // mark phase started, i.e. markArenas) for in-use spans which have specials.
90 // Break up the work into arenas, and further into chunks.
92 // Snapshot allArenas as markArenas. This snapshot is safe because allArenas
94 mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
95 work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
99 // Gs may be created after this point, but it's okay that we
100 // ignore them because they begin life without any roots, so
101 // there's nothing to scan, and any roots they create during
102 // the concurrent phase will be caught by the write barrier.
103 work.nStackRoots = int(atomic.Loaduintptr(&allglen))
105 work.markrootNext = 0
106 work.markrootJobs = uint32(fixedRootCount + work.nDataRoots + work.nBSSRoots + work.nSpanRoots + work.nStackRoots)
108 // Calculate base indexes of each root type
109 work.baseData = uint32(fixedRootCount)
110 work.baseBSS = work.baseData + uint32(work.nDataRoots)
111 work.baseSpans = work.baseBSS + uint32(work.nBSSRoots)
112 work.baseStacks = work.baseSpans + uint32(work.nSpanRoots)
113 work.baseEnd = work.baseStacks + uint32(work.nStackRoots)
116 // gcMarkRootCheck checks that all roots have been scanned. It is
117 // purely for debugging.
118 func gcMarkRootCheck() {
119 if work.markrootNext < work.markrootJobs {
120 print(work.markrootNext, " of ", work.markrootJobs, " markroot jobs done\n")
121 throw("left over markroot jobs")
124 // Check that stacks have been scanned.
126 // We only check the first nStackRoots Gs that we should have scanned.
127 // Since we don't care about newer Gs (see comment in
128 // gcMarkRootPrepare), no locking is required.
130 forEachGRace(func(gp *g) {
131 if i >= work.nStackRoots {
136 println("gp", gp, "goid", gp.goid,
137 "status", readgstatus(gp),
138 "gcscandone", gp.gcscandone)
139 throw("scan missed a g")
146 // ptrmask for an allocation containing a single pointer.
147 var oneptrmask = [...]uint8{1}
149 // markroot scans the i'th root.
151 // Preemption must be disabled (because this uses a gcWork).
153 // nowritebarrier is only advisory here.
156 func markroot(gcw *gcWork, i uint32) {
157 // Note: if you add a case here, please also update heapdump.go:dumproots.
159 case work.baseData <= i && i < work.baseBSS:
160 for _, datap := range activeModules() {
161 markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-work.baseData))
164 case work.baseBSS <= i && i < work.baseSpans:
165 for _, datap := range activeModules() {
166 markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-work.baseBSS))
169 case i == fixedRootFinalizers:
170 for fb := allfin; fb != nil; fb = fb.alllink {
171 cnt := uintptr(atomic.Load(&fb.cnt))
172 scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
175 case i == fixedRootFreeGStacks:
176 // Switch to the system stack so we can call
178 systemstack(markrootFreeGStacks)
180 case work.baseSpans <= i && i < work.baseStacks:
181 // mark mspan.specials
182 markrootSpans(gcw, int(i-work.baseSpans))
185 // the rest is scanning goroutine stacks
187 if work.baseStacks <= i && i < work.baseEnd {
188 // N.B. Atomic read of allglen in gcMarkRootPrepare
189 // acts as a barrier to ensure that allgs must be large
190 // enough to contain all relevant Gs.
191 gp = allgs[i-work.baseStacks]
193 throw("markroot: bad index")
196 // remember when we've first observed the G blocked
197 // needed only to output in traceback
198 status := readgstatus(gp) // We are not in a scan state
199 if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
200 gp.waitsince = work.tstart
203 // scanstack must be done on the system stack in case
204 // we're trying to scan our own stack.
206 // If this is a self-scan, put the user G in
207 // _Gwaiting to prevent self-deadlock. It may
208 // already be in _Gwaiting if this is a mark
209 // worker or we're in mark termination.
210 userG := getg().m.curg
211 selfScan := gp == userG && readgstatus(userG) == _Grunning
213 casgstatus(userG, _Grunning, _Gwaiting)
214 userG.waitreason = waitReasonGarbageCollectionScan
217 // TODO: suspendG blocks (and spins) until gp
218 // stops, which may take a while for
219 // running goroutines. Consider doing this in
220 // two phases where the first is non-blocking:
221 // we scan the stacks we can and ask running
222 // goroutines to scan themselves; and the
224 stopped := suspendG(gp)
230 throw("g already scanned")
237 casgstatus(userG, _Gwaiting, _Grunning)
243 // markrootBlock scans the shard'th shard of the block of memory [b0,
244 // b0+n0), with the given pointer mask.
247 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) {
248 if rootBlockBytes%(8*goarch.PtrSize) != 0 {
249 // This is necessary to pick byte offsets in ptrmask0.
250 throw("rootBlockBytes must be a multiple of 8*ptrSize")
253 // Note that if b0 is toward the end of the address space,
254 // then b0 + rootBlockBytes might wrap around.
255 // These tests are written to avoid any possible overflow.
256 off := uintptr(shard) * rootBlockBytes
261 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*goarch.PtrSize))))
262 n := uintptr(rootBlockBytes)
268 scanblock(b, n, ptrmask, gcw, nil)
271 // markrootFreeGStacks frees stacks of dead Gs.
273 // This does not free stacks of dead Gs cached on Ps, but having a few
274 // cached stacks around isn't a problem.
275 func markrootFreeGStacks() {
276 // Take list of dead Gs with stacks.
277 lock(&sched.gFree.lock)
278 list := sched.gFree.stack
279 sched.gFree.stack = gList{}
280 unlock(&sched.gFree.lock)
286 q := gQueue{list.head, list.head}
287 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
291 // Manipulate the queue directly since the Gs are
292 // already all linked the right way.
296 // Put Gs back on the free list.
297 lock(&sched.gFree.lock)
298 sched.gFree.noStack.pushAll(q)
299 unlock(&sched.gFree.lock)
302 // markrootSpans marks roots for one shard of markArenas.
305 func markrootSpans(gcw *gcWork, shard int) {
306 // Objects with finalizers have two GC-related invariants:
308 // 1) Everything reachable from the object must be marked.
309 // This ensures that when we pass the object to its finalizer,
310 // everything the finalizer can reach will be retained.
312 // 2) Finalizer specials (which are not in the garbage
313 // collected heap) are roots. In practice, this means the fn
314 // field must be scanned.
315 sg := mheap_.sweepgen
317 // Find the arena and page index into that arena for this shard.
318 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
319 ha := mheap_.arenas[ai.l1()][ai.l2()]
320 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
322 // Construct slice of bitmap which we'll iterate over.
323 specialsbits := ha.pageSpecials[arenaPage/8:]
324 specialsbits = specialsbits[:pagesPerSpanRoot/8]
325 for i := range specialsbits {
326 // Find set bits, which correspond to spans with specials.
327 specials := atomic.Load8(&specialsbits[i])
331 for j := uint(0); j < 8; j++ {
332 if specials&(1<<j) == 0 {
335 // Find the span for this bit.
337 // This value is guaranteed to be non-nil because having
338 // specials implies that the span is in-use, and since we're
339 // currently marking we can be sure that we don't have to worry
340 // about the span being freed and re-used.
341 s := ha.spans[arenaPage+uint(i)*8+j]
343 // The state must be mSpanInUse if the specials bit is set, so
344 // sanity check that.
345 if state := s.state.get(); state != mSpanInUse {
346 print("s.state = ", state, "\n")
347 throw("non in-use span found with specials bit set")
349 // Check that this span was swept (it may be cached or uncached).
350 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
351 // sweepgen was updated (+2) during non-checkmark GC pass
352 print("sweep ", s.sweepgen, " ", sg, "\n")
353 throw("gc: unswept span")
356 // Lock the specials to prevent a special from being
357 // removed from the list while we're traversing it.
359 for sp := s.specials; sp != nil; sp = sp.next {
360 if sp.kind != _KindSpecialFinalizer {
363 // don't mark finalized object, but scan it so we
364 // retain everything it points to.
365 spf := (*specialfinalizer)(unsafe.Pointer(sp))
366 // A finalizer can be set for an inner byte of an object, find object beginning.
367 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
369 // Mark everything that can be reached from
370 // the object (but *not* the object itself or
371 // we'll never collect it).
374 // The special itself is a root.
375 scanblock(uintptr(unsafe.Pointer(&spf.fn)), goarch.PtrSize, &oneptrmask[0], gcw, nil)
377 unlock(&s.speciallock)
382 // gcAssistAlloc performs GC work to make gp's assist debt positive.
383 // gp must be the calling user gorountine.
385 // This must be called with preemption enabled.
386 func gcAssistAlloc(gp *g) {
387 // Don't assist in non-preemptible contexts. These are
388 // generally fragile and won't allow the assist to block.
389 if getg() == gp.m.g0 {
392 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
398 // Compute the amount of scan work we need to do to make the
399 // balance positive. When the required amount of work is low,
400 // we over-assist to build up credit for future allocations
401 // and amortize the cost of assisting.
402 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
403 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
404 debtBytes := -gp.gcAssistBytes
405 scanWork := int64(assistWorkPerByte * float64(debtBytes))
406 if scanWork < gcOverAssistWork {
407 scanWork = gcOverAssistWork
408 debtBytes = int64(assistBytesPerWork * float64(scanWork))
411 // Steal as much credit as we can from the background GC's
412 // scan credit. This is racy and may drop the background
413 // credit below 0 if two mutators steal at the same time. This
414 // will just cause steals to fail until credit is accumulated
415 // again, so in the long run it doesn't really matter, but we
416 // do have to handle the negative credit case.
417 bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
419 if bgScanCredit > 0 {
420 if bgScanCredit < scanWork {
421 stolen = bgScanCredit
422 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
425 gp.gcAssistBytes += debtBytes
427 atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
432 // We were able to steal all of the credit we
435 traceGCMarkAssistDone()
441 if trace.enabled && !traced {
443 traceGCMarkAssistStart()
446 // Perform assist work
448 gcAssistAlloc1(gp, scanWork)
449 // The user stack may have moved, so this can't touch
450 // anything on it until it returns from systemstack.
453 completed := gp.param != nil
459 if gp.gcAssistBytes < 0 {
460 // We were unable steal enough credit or perform
461 // enough work to pay off the assist debt. We need to
462 // do one of these before letting the mutator allocate
463 // more to prevent over-allocation.
465 // If this is because we were preempted, reschedule
466 // and try some more.
472 // Add this G to an assist queue and park. When the GC
473 // has more background credit, it will satisfy queued
474 // assists before flushing to the global credit pool.
476 // Note that this does *not* get woken up when more
477 // work is added to the work list. The theory is that
478 // there wasn't enough work to do anyway, so we might
479 // as well let background marking take care of the
480 // work that is available.
485 // At this point either background GC has satisfied
486 // this G's assist debt, or the GC cycle is over.
489 traceGCMarkAssistDone()
493 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
494 // stack. This is a separate function to make it easier to see that
495 // we're not capturing anything from the user stack, since the user
496 // stack may move while we're in this function.
498 // gcAssistAlloc1 indicates whether this assist completed the mark
499 // phase by setting gp.param to non-nil. This can't be communicated on
500 // the stack since it may move.
503 func gcAssistAlloc1(gp *g, scanWork int64) {
504 // Clear the flag indicating that this assist completed the
508 if atomic.Load(&gcBlackenEnabled) == 0 {
509 // The gcBlackenEnabled check in malloc races with the
510 // store that clears it but an atomic check in every malloc
511 // would be a performance hit.
512 // Instead we recheck it here on the non-preemptable system
513 // stack to determine if we should perform an assist.
515 // GC is done, so ignore any remaining debt.
519 // Track time spent in this assist. Since we're on the
520 // system stack, this is non-preemptible, so we can
521 // just measure start and end time.
522 startTime := nanotime()
524 decnwait := atomic.Xadd(&work.nwait, -1)
525 if decnwait == work.nproc {
526 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
527 throw("nwait > work.nprocs")
530 // gcDrainN requires the caller to be preemptible.
531 casgstatus(gp, _Grunning, _Gwaiting)
532 gp.waitreason = waitReasonGCAssistMarking
534 // drain own cached work first in the hopes that it
535 // will be more cache friendly.
536 gcw := &getg().m.p.ptr().gcw
537 workDone := gcDrainN(gcw, scanWork)
539 casgstatus(gp, _Gwaiting, _Grunning)
541 // Record that we did this much scan work.
543 // Back out the number of bytes of assist credit that
544 // this scan work counts for. The "1+" is a poor man's
545 // round-up, to ensure this adds credit even if
546 // assistBytesPerWork is very low.
547 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
548 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
550 // If this is the last worker and we ran out of work,
551 // signal a completion point.
552 incnwait := atomic.Xadd(&work.nwait, +1)
553 if incnwait > work.nproc {
554 println("runtime: work.nwait=", incnwait,
555 "work.nproc=", work.nproc)
556 throw("work.nwait > work.nproc")
559 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
560 // This has reached a background completion point. Set
561 // gp.param to a non-nil value to indicate this. It
562 // doesn't matter what we set it to (it just has to be
564 gp.param = unsafe.Pointer(gp)
566 duration := nanotime() - startTime
568 _p_.gcAssistTime += duration
569 if _p_.gcAssistTime > gcAssistTimeSlack {
570 atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
575 // gcWakeAllAssists wakes all currently blocked assists. This is used
576 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
577 // new assists from going to sleep after this point.
578 func gcWakeAllAssists() {
579 lock(&work.assistQueue.lock)
580 list := work.assistQueue.q.popList()
582 unlock(&work.assistQueue.lock)
585 // gcParkAssist puts the current goroutine on the assist queue and parks.
587 // gcParkAssist reports whether the assist is now satisfied. If it
588 // returns false, the caller must retry the assist.
591 func gcParkAssist() bool {
592 lock(&work.assistQueue.lock)
593 // If the GC cycle finished while we were getting the lock,
594 // exit the assist. The cycle can't finish while we hold the
596 if atomic.Load(&gcBlackenEnabled) == 0 {
597 unlock(&work.assistQueue.lock)
602 oldList := work.assistQueue.q
603 work.assistQueue.q.pushBack(gp)
605 // Recheck for background credit now that this G is in
606 // the queue, but can still back out. This avoids a
607 // race in case background marking has flushed more
608 // credit since we checked above.
609 if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
610 work.assistQueue.q = oldList
611 if oldList.tail != 0 {
612 oldList.tail.ptr().schedlink.set(nil)
614 unlock(&work.assistQueue.lock)
618 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
622 // gcFlushBgCredit flushes scanWork units of background scan work
623 // credit. This first satisfies blocked assists on the
624 // work.assistQueue and then flushes any remaining credit to
625 // gcController.bgScanCredit.
627 // Write barriers are disallowed because this is used by gcDrain after
628 // it has ensured that all work is drained and this must preserve that
631 //go:nowritebarrierrec
632 func gcFlushBgCredit(scanWork int64) {
633 if work.assistQueue.q.empty() {
634 // Fast path; there are no blocked assists. There's a
635 // small window here where an assist may add itself to
636 // the blocked queue and park. If that happens, we'll
637 // just get it on the next flush.
638 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
642 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
643 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
645 lock(&work.assistQueue.lock)
646 for !work.assistQueue.q.empty() && scanBytes > 0 {
647 gp := work.assistQueue.q.pop()
648 // Note that gp.gcAssistBytes is negative because gp
649 // is in debt. Think carefully about the signs below.
650 if scanBytes+gp.gcAssistBytes >= 0 {
651 // Satisfy this entire assist debt.
652 scanBytes += gp.gcAssistBytes
654 // It's important that we *not* put gp in
655 // runnext. Otherwise, it's possible for user
656 // code to exploit the GC worker's high
657 // scheduler priority to get itself always run
658 // before other goroutines and always in the
659 // fresh quantum started by GC.
662 // Partially satisfy this assist.
663 gp.gcAssistBytes += scanBytes
665 // As a heuristic, we move this assist to the
666 // back of the queue so that large assists
667 // can't clog up the assist queue and
668 // substantially delay small assists.
669 work.assistQueue.q.pushBack(gp)
675 // Convert from scan bytes back to work.
676 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
677 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
678 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
680 unlock(&work.assistQueue.lock)
683 // scanstack scans gp's stack, greying all pointers found on the stack.
685 // scanstack will also shrink the stack if it is safe to do so. If it
686 // is not, it schedules a stack shrink for the next synchronous safe
689 // scanstack is marked go:systemstack because it must not be preempted
690 // while using a workbuf.
694 func scanstack(gp *g, gcw *gcWork) {
695 if readgstatus(gp)&_Gscan == 0 {
696 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
697 throw("scanstack - bad status")
700 switch readgstatus(gp) &^ _Gscan {
702 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
703 throw("mark - bad status")
707 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
708 throw("scanstack: goroutine not stopped")
709 case _Grunnable, _Gsyscall, _Gwaiting:
714 throw("can't scan our own stack")
717 if isShrinkStackSafe(gp) {
718 // Shrink the stack if not much of it is being used.
721 // Otherwise, shrink the stack at the next sync safe point.
722 gp.preemptShrink = true
725 var state stackScanState
726 state.stack = gp.stack
729 println("stack trace goroutine", gp.goid)
732 if debugScanConservative && gp.asyncSafePoint {
733 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
736 // Scan the saved context register. This is effectively a live
737 // register that gets moved back and forth between the
738 // register and sched.ctxt without a write barrier.
739 if gp.sched.ctxt != nil {
740 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
743 // Scan the stack. Accumulate a list of stack objects.
744 scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
745 scanframeworker(frame, &state, gcw)
748 gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
750 // Find additional pointers that point into the stack from the heap.
751 // Currently this includes defers and panics. See also function copystack.
753 // Find and trace other pointers in defer records.
754 for d := gp._defer; d != nil; d = d.link {
756 // Scan the func value, which could be a stack allocated closure.
758 scanblock(uintptr(unsafe.Pointer(&d.fn)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
761 // The link field of a stack-allocated defer record might point
762 // to a heap-allocated defer record. Keep that heap record live.
763 scanblock(uintptr(unsafe.Pointer(&d.link)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
765 // Retain defers records themselves.
766 // Defer records might not be reachable from the G through regular heap
767 // tracing because the defer linked list might weave between the stack and the heap.
769 scanblock(uintptr(unsafe.Pointer(&d)), goarch.PtrSize, &oneptrmask[0], gcw, &state)
772 if gp._panic != nil {
773 // Panics are always stack allocated.
774 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
777 // Find and scan all reachable stack objects.
779 // The state's pointer queue prioritizes precise pointers over
780 // conservative pointers so that we'll prefer scanning stack
781 // objects precisely.
784 p, conservative := state.getPtr()
788 obj := state.findObject(p)
794 // We've already scanned this object.
797 obj.setRecord(nil) // Don't scan it again.
800 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
802 print(" (conservative)")
810 // This path is pretty unlikely, an object large enough
811 // to have a GC program allocated on the stack.
812 // We need some space to unpack the program into a straight
813 // bitmask, which we allocate/free here.
814 // TODO: it would be nice if there were a way to run a GC
815 // program without having to store all its bits. We'd have
816 // to change from a Lempel-Ziv style program to something else.
817 // Or we can forbid putting objects on stacks if they require
818 // a gc program (see issue 27447).
819 s = materializeGCProg(r.ptrdata(), gcdata)
820 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
823 b := state.stack.lo + uintptr(obj.off)
825 scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
827 scanblock(b, r.ptrdata(), gcdata, gcw, &state)
831 dematerializeGCProg(s)
835 // Deallocate object buffers.
836 // (Pointer buffers were all deallocated in the loop above.)
837 for state.head != nil {
841 for i := 0; i < x.nobj; i++ {
843 if obj.r == nil { // reachable
846 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
847 // Note: not necessarily really dead - only reachable-from-ptr dead.
851 putempty((*workbuf)(unsafe.Pointer(x)))
853 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
854 throw("remaining pointer buffers")
858 // Scan a stack frame: local variables and function arguments/results.
860 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
861 if _DebugGC > 1 && frame.continpc != 0 {
862 print("scanframe ", funcname(frame.fn), "\n")
865 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
866 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV2
867 if state.conservative || isAsyncPreempt || isDebugCall {
868 if debugScanConservative {
869 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
872 // Conservatively scan the frame. Unlike the precise
873 // case, this includes the outgoing argument space
874 // since we may have stopped while this function was
875 // setting up a call.
877 // TODO: We could narrow this down if the compiler
878 // produced a single map per function of stack slots
879 // and registers that ever contain a pointer.
881 size := frame.varp - frame.sp
883 scanConservative(frame.sp, size, nil, gcw, state)
887 // Scan arguments to this frame.
888 if frame.arglen != 0 {
889 // TODO: We could pass the entry argument map
890 // to narrow this down further.
891 scanConservative(frame.argp, frame.arglen, nil, gcw, state)
894 if isAsyncPreempt || isDebugCall {
895 // This function's frame contained the
896 // registers for the asynchronously stopped
897 // parent frame. Scan the parent
899 state.conservative = true
901 // We only wanted to scan those two frames
902 // conservatively. Clear the flag for future
904 state.conservative = false
909 locals, args, objs := getStackMap(frame, &state.cache, false)
911 // Scan local variables if stack frame has been allocated.
913 size := uintptr(locals.n) * goarch.PtrSize
914 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
919 scanblock(frame.argp, uintptr(args.n)*goarch.PtrSize, args.bytedata, gcw, state)
922 // Add all stack objects to the stack object list.
924 // varp is 0 for defers, where there are no locals.
925 // In that case, there can't be a pointer to its args, either.
926 // (And all args would be scanned above anyway.)
927 for i, obj := range objs {
929 base := frame.varp // locals base pointer
931 base = frame.argp // arguments and return values base pointer
933 ptr := base + uintptr(off)
935 // object hasn't been allocated in the frame yet.
939 println("stkobj at", hex(ptr), "of size", obj.size)
941 state.addObject(ptr, &objs[i])
946 type gcDrainFlags int
949 gcDrainUntilPreempt gcDrainFlags = 1 << iota
955 // gcDrain scans roots and objects in work buffers, blackening grey
956 // objects until it is unable to get more work. It may return before
957 // GC is done; it's the caller's responsibility to balance work from
960 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
963 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
966 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
967 // pollFractionalWorkerExit() returns true. This implies
970 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
971 // credit to gcController.bgScanCredit every gcCreditSlack units of
974 // gcDrain will always return if there is a pending STW.
977 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
978 if !writeBarrier.needed {
979 throw("gcDrain phase incorrect")
983 preemptible := flags&gcDrainUntilPreempt != 0
984 flushBgCredit := flags&gcDrainFlushBgCredit != 0
985 idle := flags&gcDrainIdle != 0
987 initScanWork := gcw.scanWork
989 // checkWork is the scan work before performing the next
990 // self-preempt check.
991 checkWork := int64(1<<63 - 1)
992 var check func() bool
993 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
994 checkWork = initScanWork + drainCheckThreshold
997 } else if flags&gcDrainFractional != 0 {
998 check = pollFractionalWorkerExit
1002 // Drain root marking jobs.
1003 if work.markrootNext < work.markrootJobs {
1004 // Stop if we're preemptible or if someone wants to STW.
1005 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1006 job := atomic.Xadd(&work.markrootNext, +1) - 1
1007 if job >= work.markrootJobs {
1011 if check != nil && check() {
1017 // Drain heap marking jobs.
1018 // Stop if we're preemptible or if someone wants to STW.
1019 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1020 // Try to keep work available on the global queue. We used to
1021 // check if there were waiting workers, but it's better to
1022 // just keep work available than to make workers wait. In the
1023 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1029 b := gcw.tryGetFast()
1033 // Flush the write barrier
1034 // buffer; this may create
1041 // Unable to get work.
1046 // Flush background scan work credit to the global
1047 // account if we've accumulated enough locally so
1048 // mutator assists can draw on it.
1049 if gcw.scanWork >= gcCreditSlack {
1050 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1052 gcFlushBgCredit(gcw.scanWork - initScanWork)
1055 checkWork -= gcw.scanWork
1059 checkWork += drainCheckThreshold
1060 if check != nil && check() {
1068 // Flush remaining scan work credit.
1069 if gcw.scanWork > 0 {
1070 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1072 gcFlushBgCredit(gcw.scanWork - initScanWork)
1078 // gcDrainN blackens grey objects until it has performed roughly
1079 // scanWork units of scan work or the G is preempted. This is
1080 // best-effort, so it may perform less work if it fails to get a work
1081 // buffer. Otherwise, it will perform at least n units of work, but
1082 // may perform more because scanning is always done in whole object
1083 // increments. It returns the amount of scan work performed.
1085 // The caller goroutine must be in a preemptible state (e.g.,
1086 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1087 // consequence, this must be called on the system stack.
1091 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1092 if !writeBarrier.needed {
1093 throw("gcDrainN phase incorrect")
1096 // There may already be scan work on the gcw, which we don't
1097 // want to claim was done by this call.
1098 workFlushed := -gcw.scanWork
1101 for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
1102 // See gcDrain comment.
1107 // This might be a good place to add prefetch code...
1108 // if(wbuf.nobj > 4) {
1109 // PREFETCH(wbuf->obj[wbuf.nobj - 3];
1112 b := gcw.tryGetFast()
1116 // Flush the write barrier buffer;
1117 // this may create more work.
1124 // Try to do a root job.
1126 // TODO: Assists should get credit for this
1128 if work.markrootNext < work.markrootJobs {
1129 job := atomic.Xadd(&work.markrootNext, +1) - 1
1130 if job < work.markrootJobs {
1135 // No heap or root jobs.
1140 // Flush background scan work credit.
1141 if gcw.scanWork >= gcCreditSlack {
1142 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1143 workFlushed += gcw.scanWork
1148 // Unlike gcDrain, there's no need to flush remaining work
1149 // here because this never flushes to bgScanCredit and
1150 // gcw.dispose will flush any remaining work to scanWork.
1152 return workFlushed + gcw.scanWork
1155 // scanblock scans b as scanobject would, but using an explicit
1156 // pointer bitmap instead of the heap bitmap.
1158 // This is used to scan non-heap roots, so it does not update
1159 // gcw.bytesMarked or gcw.scanWork.
1161 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1163 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1164 // Use local copies of original parameters, so that a stack trace
1165 // due to one of the throws below shows the original block
1170 for i := uintptr(0); i < n; {
1171 // Find bits for the next word.
1172 bits := uint32(*addb(ptrmask, i/(goarch.PtrSize*8)))
1174 i += goarch.PtrSize * 8
1177 for j := 0; j < 8 && i < n; j++ {
1179 // Same work as in scanobject; see comments there.
1180 p := *(*uintptr)(unsafe.Pointer(b + i))
1182 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1183 greyobject(obj, b, i, span, gcw, objIndex)
1184 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1185 stk.putPtr(p, false)
1195 // scanobject scans the object starting at b, adding pointers to gcw.
1196 // b must point to the beginning of a heap object or an oblet.
1197 // scanobject consults the GC bitmap for the pointer mask and the
1198 // spans for the size of the object.
1201 func scanobject(b uintptr, gcw *gcWork) {
1202 // Find the bits for b and the size of the object at b.
1204 // b is either the beginning of an object, in which case this
1205 // is the size of the object to scan, or it points to an
1206 // oblet, in which case we compute the size to scan below.
1207 hbits := heapBitsForAddr(b)
1208 s := spanOfUnchecked(b)
1211 throw("scanobject n == 0")
1214 if n > maxObletBytes {
1215 // Large object. Break into oblets for better
1216 // parallelism and lower latency.
1218 // It's possible this is a noscan object (not
1219 // from greyobject, but from other code
1220 // paths), in which case we must *not* enqueue
1221 // oblets since their bitmaps will be
1223 if s.spanclass.noscan() {
1224 // Bypass the whole scan.
1225 gcw.bytesMarked += uint64(n)
1229 // Enqueue the other oblets to scan later.
1230 // Some oblets may be in b's scalar tail, but
1231 // these will be marked as "no more pointers",
1232 // so we'll drop out immediately when we go to
1234 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1235 if !gcw.putFast(oblet) {
1241 // Compute the size of the oblet. Since this object
1242 // must be a large object, s.base() is the beginning
1244 n = s.base() + s.elemsize - b
1245 if n > maxObletBytes {
1251 for i = 0; i < n; i, hbits = i+goarch.PtrSize, hbits.next() {
1252 // Load bits once. See CL 22712 and issue 16973 for discussion.
1253 bits := hbits.bits()
1254 if bits&bitScan == 0 {
1255 break // no more pointers in this object
1257 if bits&bitPointer == 0 {
1258 continue // not a pointer
1261 // Work here is duplicated in scanblock and above.
1262 // If you make changes here, make changes there too.
1263 obj := *(*uintptr)(unsafe.Pointer(b + i))
1265 // At this point we have extracted the next potential pointer.
1266 // Quickly filter out nil and pointers back to the current object.
1267 if obj != 0 && obj-b >= n {
1268 // Test if obj points into the Go heap and, if so,
1271 // Note that it's possible for findObject to
1272 // fail if obj points to a just-allocated heap
1273 // object because of a race with growing the
1274 // heap. In this case, we know the object was
1275 // just allocated and hence will be marked by
1276 // allocation itself.
1277 if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
1278 greyobject(obj, b, i, span, gcw, objIndex)
1282 gcw.bytesMarked += uint64(n)
1283 gcw.scanWork += int64(i)
1286 // scanConservative scans block [b, b+n) conservatively, treating any
1287 // pointer-like value in the block as a pointer.
1289 // If ptrmask != nil, only words that are marked in ptrmask are
1290 // considered as potential pointers.
1292 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1293 // and may contain pointers to stack objects.
1294 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1295 if debugScanConservative {
1297 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1298 hexdumpWords(b, b+n, func(p uintptr) byte {
1300 word := (p - b) / goarch.PtrSize
1301 bits := *addb(ptrmask, word/8)
1302 if (bits>>(word%8))&1 == 0 {
1307 val := *(*uintptr)(unsafe.Pointer(p))
1308 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1312 span := spanOfHeap(val)
1316 idx := span.objIndex(val)
1317 if span.isFree(idx) {
1325 for i := uintptr(0); i < n; i += goarch.PtrSize {
1327 word := i / goarch.PtrSize
1328 bits := *addb(ptrmask, word/8)
1330 // Skip 8 words (the loop increment will do the 8th)
1332 // This must be the first time we've
1333 // seen this word of ptrmask, so i
1334 // must be 8-word-aligned, but check
1335 // our reasoning just in case.
1336 if i%(goarch.PtrSize*8) != 0 {
1337 throw("misaligned mask")
1339 i += goarch.PtrSize*8 - goarch.PtrSize
1342 if (bits>>(word%8))&1 == 0 {
1347 val := *(*uintptr)(unsafe.Pointer(b + i))
1349 // Check if val points into the stack.
1350 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1351 // val may point to a stack object. This
1352 // object may be dead from last cycle and
1353 // hence may contain pointers to unallocated
1354 // objects, but unlike heap objects we can't
1355 // tell if it's already dead. Hence, if all
1356 // pointers to this object are from
1357 // conservative scanning, we have to scan it
1358 // defensively, too.
1359 state.putPtr(val, true)
1363 // Check if val points to a heap span.
1364 span := spanOfHeap(val)
1369 // Check if val points to an allocated object.
1370 idx := span.objIndex(val)
1371 if span.isFree(idx) {
1375 // val points to an allocated object. Mark it.
1376 obj := span.base() + idx*span.elemsize
1377 greyobject(obj, b, i, span, gcw, idx)
1381 // Shade the object if it isn't already.
1382 // The object is not nil and known to be in the heap.
1383 // Preemption must be disabled.
1385 func shade(b uintptr) {
1386 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1387 gcw := &getg().m.p.ptr().gcw
1388 greyobject(obj, 0, 0, span, gcw, objIndex)
1392 // obj is the start of an object with mark mbits.
1393 // If it isn't already marked, mark it and enqueue into gcw.
1394 // base and off are for debugging only and could be removed.
1396 // See also wbBufFlush1, which partially duplicates this logic.
1398 //go:nowritebarrierrec
1399 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1400 // obj should be start of allocation, and so must be at least pointer-aligned.
1401 if obj&(goarch.PtrSize-1) != 0 {
1402 throw("greyobject: obj not pointer-aligned")
1404 mbits := span.markBitsForIndex(objIndex)
1407 if setCheckmark(obj, base, off, mbits) {
1412 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1413 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1414 gcDumpObject("base", base, off)
1415 gcDumpObject("obj", obj, ^uintptr(0))
1416 getg().m.traceback = 2
1417 throw("marking free object")
1420 // If marked we have nothing to do.
1421 if mbits.isMarked() {
1427 arena, pageIdx, pageMask := pageIndexOf(span.base())
1428 if arena.pageMarks[pageIdx]&pageMask == 0 {
1429 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1432 // If this is a noscan object, fast-track it to black
1433 // instead of greying it.
1434 if span.spanclass.noscan() {
1435 gcw.bytesMarked += uint64(span.elemsize)
1440 // Queue the obj for scanning. The PREFETCH(obj) logic has been removed but
1441 // seems like a nice optimization that can be added back in.
1442 // There needs to be time between the PREFETCH and the use.
1443 // Previously we put the obj in an 8 element buffer that is drained at a rate
1444 // to give the PREFETCH time to do its work.
1445 // Use of PREFETCHNTA might be more appropriate than PREFETCH
1446 if !gcw.putFast(obj) {
1451 // gcDumpObject dumps the contents of obj for debugging and marks the
1452 // field at byte offset off in obj.
1453 func gcDumpObject(label string, obj, off uintptr) {
1455 print(label, "=", hex(obj))
1460 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1461 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1462 print(mSpanStateNames[state], "\n")
1464 print("unknown(", state, ")\n")
1469 if s.state.get() == mSpanManual && size == 0 {
1470 // We're printing something from a stack frame. We
1471 // don't know how big it is, so just show up to an
1473 size = off + goarch.PtrSize
1475 for i := uintptr(0); i < size; i += goarch.PtrSize {
1476 // For big objects, just print the beginning (because
1477 // that usually hints at the object's type) and the
1478 // fields around off.
1479 if !(i < 128*goarch.PtrSize || off-16*goarch.PtrSize < i && i < off+16*goarch.PtrSize) {
1487 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1498 // gcmarknewobject marks a newly allocated object black. obj must
1499 // not contain any non-nil pointers.
1501 // This is nosplit so it can manipulate a gcWork without preemption.
1505 func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
1506 if useCheckmark { // The world should be stopped so this should not happen.
1507 throw("gcmarknewobject called while doing checkmark")
1511 objIndex := span.objIndex(obj)
1512 span.markBitsForIndex(objIndex).setMarked()
1515 arena, pageIdx, pageMask := pageIndexOf(span.base())
1516 if arena.pageMarks[pageIdx]&pageMask == 0 {
1517 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1520 gcw := &getg().m.p.ptr().gcw
1521 gcw.bytesMarked += uint64(size)
1522 gcw.scanWork += int64(scanSize)
1525 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1527 // The world must be stopped.
1528 func gcMarkTinyAllocs() {
1529 assertWorldStopped()
1531 for _, p := range allp {
1533 if c == nil || c.tiny == 0 {
1536 _, span, objIndex := findObject(c.tiny, 0, 0)
1538 greyobject(c.tiny, 0, 0, span, gcw, objIndex)