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 throw("scan missed a g")
138 // ptrmask for an allocation containing a single pointer.
139 var oneptrmask = [...]uint8{1}
141 // markroot scans the i'th root.
143 // Preemption must be disabled (because this uses a gcWork).
145 // nowritebarrier is only advisory here.
148 func markroot(gcw *gcWork, i uint32) {
149 // TODO(austin): This is a bit ridiculous. Compute and store
150 // the bases in gcMarkRootPrepare instead of the counts.
151 baseFlushCache := uint32(fixedRootCount)
152 baseData := baseFlushCache + uint32(work.nFlushCacheRoots)
153 baseBSS := baseData + uint32(work.nDataRoots)
154 baseSpans := baseBSS + uint32(work.nBSSRoots)
155 baseStacks := baseSpans + uint32(work.nSpanRoots)
156 end := baseStacks + uint32(work.nStackRoots)
158 // Note: if you add a case here, please also update heapdump.go:dumproots.
160 case baseFlushCache <= i && i < baseData:
161 flushmcache(int(i - baseFlushCache))
163 case baseData <= i && i < baseBSS:
164 for _, datap := range activeModules() {
165 markrootBlock(datap.data, datap.edata-datap.data, datap.gcdatamask.bytedata, gcw, int(i-baseData))
168 case baseBSS <= i && i < baseSpans:
169 for _, datap := range activeModules() {
170 markrootBlock(datap.bss, datap.ebss-datap.bss, datap.gcbssmask.bytedata, gcw, int(i-baseBSS))
173 case i == fixedRootFinalizers:
174 for fb := allfin; fb != nil; fb = fb.alllink {
175 cnt := uintptr(atomic.Load(&fb.cnt))
176 scanblock(uintptr(unsafe.Pointer(&fb.fin[0])), cnt*unsafe.Sizeof(fb.fin[0]), &finptrmask[0], gcw, nil)
179 case i == fixedRootFreeGStacks:
180 // Switch to the system stack so we can call
182 systemstack(markrootFreeGStacks)
184 case baseSpans <= i && i < baseStacks:
185 // mark mspan.specials
186 markrootSpans(gcw, int(i-baseSpans))
189 // the rest is scanning goroutine stacks
191 if baseStacks <= i && i < end {
192 gp = allgs[i-baseStacks]
194 throw("markroot: bad index")
197 // remember when we've first observed the G blocked
198 // needed only to output in traceback
199 status := readgstatus(gp) // We are not in a scan state
200 if (status == _Gwaiting || status == _Gsyscall) && gp.waitsince == 0 {
201 gp.waitsince = work.tstart
204 // scanstack must be done on the system stack in case
205 // we're trying to scan our own stack.
207 // If this is a self-scan, put the user G in
208 // _Gwaiting to prevent self-deadlock. It may
209 // already be in _Gwaiting if this is a mark
210 // worker or we're in mark termination.
211 userG := getg().m.curg
212 selfScan := gp == userG && readgstatus(userG) == _Grunning
214 casgstatus(userG, _Grunning, _Gwaiting)
215 userG.waitreason = waitReasonGarbageCollectionScan
218 // TODO: suspendG blocks (and spins) until gp
219 // stops, which may take a while for
220 // running goroutines. Consider doing this in
221 // two phases where the first is non-blocking:
222 // we scan the stacks we can and ask running
223 // goroutines to scan themselves; and the
225 stopped := suspendG(gp)
231 throw("g already scanned")
238 casgstatus(userG, _Gwaiting, _Grunning)
244 // markrootBlock scans the shard'th shard of the block of memory [b0,
245 // b0+n0), with the given pointer mask.
248 func markrootBlock(b0, n0 uintptr, ptrmask0 *uint8, gcw *gcWork, shard int) {
249 if rootBlockBytes%(8*sys.PtrSize) != 0 {
250 // This is necessary to pick byte offsets in ptrmask0.
251 throw("rootBlockBytes must be a multiple of 8*ptrSize")
254 // Note that if b0 is toward the end of the address space,
255 // then b0 + rootBlockBytes might wrap around.
256 // These tests are written to avoid any possible overflow.
257 off := uintptr(shard) * rootBlockBytes
262 ptrmask := (*uint8)(add(unsafe.Pointer(ptrmask0), uintptr(shard)*(rootBlockBytes/(8*sys.PtrSize))))
263 n := uintptr(rootBlockBytes)
269 scanblock(b, n, ptrmask, gcw, nil)
272 // markrootFreeGStacks frees stacks of dead Gs.
274 // This does not free stacks of dead Gs cached on Ps, but having a few
275 // cached stacks around isn't a problem.
276 func markrootFreeGStacks() {
277 // Take list of dead Gs with stacks.
278 lock(&sched.gFree.lock)
279 list := sched.gFree.stack
280 sched.gFree.stack = gList{}
281 unlock(&sched.gFree.lock)
287 q := gQueue{list.head, list.head}
288 for gp := list.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
292 // Manipulate the queue directly since the Gs are
293 // already all linked the right way.
297 // Put Gs back on the free list.
298 lock(&sched.gFree.lock)
299 sched.gFree.noStack.pushAll(q)
300 unlock(&sched.gFree.lock)
303 // markrootSpans marks roots for one shard of markArenas.
306 func markrootSpans(gcw *gcWork, shard int) {
307 // Objects with finalizers have two GC-related invariants:
309 // 1) Everything reachable from the object must be marked.
310 // This ensures that when we pass the object to its finalizer,
311 // everything the finalizer can reach will be retained.
313 // 2) Finalizer specials (which are not in the garbage
314 // collected heap) are roots. In practice, this means the fn
315 // field must be scanned.
316 sg := mheap_.sweepgen
318 // Find the arena and page index into that arena for this shard.
319 ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
320 ha := mheap_.arenas[ai.l1()][ai.l2()]
321 arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
323 // Construct slice of bitmap which we'll iterate over.
324 specialsbits := ha.pageSpecials[arenaPage/8:]
325 specialsbits = specialsbits[:pagesPerSpanRoot/8]
326 for i := range specialsbits {
327 // Find set bits, which correspond to spans with specials.
328 specials := atomic.Load8(&specialsbits[i])
332 for j := uint(0); j < 8; j++ {
333 if specials&(1<<j) == 0 {
336 // Find the span for this bit.
338 // This value is guaranteed to be non-nil because having
339 // specials implies that the span is in-use, and since we're
340 // currently marking we can be sure that we don't have to worry
341 // about the span being freed and re-used.
342 s := ha.spans[arenaPage+uint(i)*8+j]
344 // The state must be mSpanInUse if the specials bit is set, so
345 // sanity check that.
346 if state := s.state.get(); state != mSpanInUse {
347 print("s.state = ", state, "\n")
348 throw("non in-use span found with specials bit set")
350 // Check that this span was swept (it may be cached or uncached).
351 if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
352 // sweepgen was updated (+2) during non-checkmark GC pass
353 print("sweep ", s.sweepgen, " ", sg, "\n")
354 throw("gc: unswept span")
357 // Lock the specials to prevent a special from being
358 // removed from the list while we're traversing it.
360 for sp := s.specials; sp != nil; sp = sp.next {
361 if sp.kind != _KindSpecialFinalizer {
364 // don't mark finalized object, but scan it so we
365 // retain everything it points to.
366 spf := (*specialfinalizer)(unsafe.Pointer(sp))
367 // A finalizer can be set for an inner byte of an object, find object beginning.
368 p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
370 // Mark everything that can be reached from
371 // the object (but *not* the object itself or
372 // we'll never collect it).
375 // The special itself is a root.
376 scanblock(uintptr(unsafe.Pointer(&spf.fn)), sys.PtrSize, &oneptrmask[0], gcw, nil)
378 unlock(&s.speciallock)
383 // gcAssistAlloc performs GC work to make gp's assist debt positive.
384 // gp must be the calling user gorountine.
386 // This must be called with preemption enabled.
387 func gcAssistAlloc(gp *g) {
388 // Don't assist in non-preemptible contexts. These are
389 // generally fragile and won't allow the assist to block.
390 if getg() == gp.m.g0 {
393 if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
399 // Compute the amount of scan work we need to do to make the
400 // balance positive. When the required amount of work is low,
401 // we over-assist to build up credit for future allocations
402 // and amortize the cost of assisting.
403 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
404 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
405 debtBytes := -gp.gcAssistBytes
406 scanWork := int64(assistWorkPerByte * float64(debtBytes))
407 if scanWork < gcOverAssistWork {
408 scanWork = gcOverAssistWork
409 debtBytes = int64(assistBytesPerWork * float64(scanWork))
412 // Steal as much credit as we can from the background GC's
413 // scan credit. This is racy and may drop the background
414 // credit below 0 if two mutators steal at the same time. This
415 // will just cause steals to fail until credit is accumulated
416 // again, so in the long run it doesn't really matter, but we
417 // do have to handle the negative credit case.
418 bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
420 if bgScanCredit > 0 {
421 if bgScanCredit < scanWork {
422 stolen = bgScanCredit
423 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(stolen))
426 gp.gcAssistBytes += debtBytes
428 atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
433 // We were able to steal all of the credit we
436 traceGCMarkAssistDone()
442 if trace.enabled && !traced {
444 traceGCMarkAssistStart()
447 // Perform assist work
449 gcAssistAlloc1(gp, scanWork)
450 // The user stack may have moved, so this can't touch
451 // anything on it until it returns from systemstack.
454 completed := gp.param != nil
460 if gp.gcAssistBytes < 0 {
461 // We were unable steal enough credit or perform
462 // enough work to pay off the assist debt. We need to
463 // do one of these before letting the mutator allocate
464 // more to prevent over-allocation.
466 // If this is because we were preempted, reschedule
467 // and try some more.
473 // Add this G to an assist queue and park. When the GC
474 // has more background credit, it will satisfy queued
475 // assists before flushing to the global credit pool.
477 // Note that this does *not* get woken up when more
478 // work is added to the work list. The theory is that
479 // there wasn't enough work to do anyway, so we might
480 // as well let background marking take care of the
481 // work that is available.
486 // At this point either background GC has satisfied
487 // this G's assist debt, or the GC cycle is over.
490 traceGCMarkAssistDone()
494 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
495 // stack. This is a separate function to make it easier to see that
496 // we're not capturing anything from the user stack, since the user
497 // stack may move while we're in this function.
499 // gcAssistAlloc1 indicates whether this assist completed the mark
500 // phase by setting gp.param to non-nil. This can't be communicated on
501 // the stack since it may move.
504 func gcAssistAlloc1(gp *g, scanWork int64) {
505 // Clear the flag indicating that this assist completed the
509 if atomic.Load(&gcBlackenEnabled) == 0 {
510 // The gcBlackenEnabled check in malloc races with the
511 // store that clears it but an atomic check in every malloc
512 // would be a performance hit.
513 // Instead we recheck it here on the non-preemptable system
514 // stack to determine if we should perform an assist.
516 // GC is done, so ignore any remaining debt.
520 // Track time spent in this assist. Since we're on the
521 // system stack, this is non-preemptible, so we can
522 // just measure start and end time.
523 startTime := nanotime()
525 decnwait := atomic.Xadd(&work.nwait, -1)
526 if decnwait == work.nproc {
527 println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
528 throw("nwait > work.nprocs")
531 // gcDrainN requires the caller to be preemptible.
532 casgstatus(gp, _Grunning, _Gwaiting)
533 gp.waitreason = waitReasonGCAssistMarking
535 // drain own cached work first in the hopes that it
536 // will be more cache friendly.
537 gcw := &getg().m.p.ptr().gcw
538 workDone := gcDrainN(gcw, scanWork)
540 casgstatus(gp, _Gwaiting, _Grunning)
542 // Record that we did this much scan work.
544 // Back out the number of bytes of assist credit that
545 // this scan work counts for. The "1+" is a poor man's
546 // round-up, to ensure this adds credit even if
547 // assistBytesPerWork is very low.
548 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
549 gp.gcAssistBytes += 1 + int64(assistBytesPerWork*float64(workDone))
551 // If this is the last worker and we ran out of work,
552 // signal a completion point.
553 incnwait := atomic.Xadd(&work.nwait, +1)
554 if incnwait > work.nproc {
555 println("runtime: work.nwait=", incnwait,
556 "work.nproc=", work.nproc)
557 throw("work.nwait > work.nproc")
560 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
561 // This has reached a background completion point. Set
562 // gp.param to a non-nil value to indicate this. It
563 // doesn't matter what we set it to (it just has to be
565 gp.param = unsafe.Pointer(gp)
567 duration := nanotime() - startTime
569 _p_.gcAssistTime += duration
570 if _p_.gcAssistTime > gcAssistTimeSlack {
571 atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
576 // gcWakeAllAssists wakes all currently blocked assists. This is used
577 // at the end of a GC cycle. gcBlackenEnabled must be false to prevent
578 // new assists from going to sleep after this point.
579 func gcWakeAllAssists() {
580 lock(&work.assistQueue.lock)
581 list := work.assistQueue.q.popList()
583 unlock(&work.assistQueue.lock)
586 // gcParkAssist puts the current goroutine on the assist queue and parks.
588 // gcParkAssist reports whether the assist is now satisfied. If it
589 // returns false, the caller must retry the assist.
592 func gcParkAssist() bool {
593 lock(&work.assistQueue.lock)
594 // If the GC cycle finished while we were getting the lock,
595 // exit the assist. The cycle can't finish while we hold the
597 if atomic.Load(&gcBlackenEnabled) == 0 {
598 unlock(&work.assistQueue.lock)
603 oldList := work.assistQueue.q
604 work.assistQueue.q.pushBack(gp)
606 // Recheck for background credit now that this G is in
607 // the queue, but can still back out. This avoids a
608 // race in case background marking has flushed more
609 // credit since we checked above.
610 if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
611 work.assistQueue.q = oldList
612 if oldList.tail != 0 {
613 oldList.tail.ptr().schedlink.set(nil)
615 unlock(&work.assistQueue.lock)
619 goparkunlock(&work.assistQueue.lock, waitReasonGCAssistWait, traceEvGoBlockGC, 2)
623 // gcFlushBgCredit flushes scanWork units of background scan work
624 // credit. This first satisfies blocked assists on the
625 // work.assistQueue and then flushes any remaining credit to
626 // gcController.bgScanCredit.
628 // Write barriers are disallowed because this is used by gcDrain after
629 // it has ensured that all work is drained and this must preserve that
632 //go:nowritebarrierrec
633 func gcFlushBgCredit(scanWork int64) {
634 if work.assistQueue.q.empty() {
635 // Fast path; there are no blocked assists. There's a
636 // small window here where an assist may add itself to
637 // the blocked queue and park. If that happens, we'll
638 // just get it on the next flush.
639 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
643 assistBytesPerWork := float64frombits(atomic.Load64(&gcController.assistBytesPerWork))
644 scanBytes := int64(float64(scanWork) * assistBytesPerWork)
646 lock(&work.assistQueue.lock)
647 for !work.assistQueue.q.empty() && scanBytes > 0 {
648 gp := work.assistQueue.q.pop()
649 // Note that gp.gcAssistBytes is negative because gp
650 // is in debt. Think carefully about the signs below.
651 if scanBytes+gp.gcAssistBytes >= 0 {
652 // Satisfy this entire assist debt.
653 scanBytes += gp.gcAssistBytes
655 // It's important that we *not* put gp in
656 // runnext. Otherwise, it's possible for user
657 // code to exploit the GC worker's high
658 // scheduler priority to get itself always run
659 // before other goroutines and always in the
660 // fresh quantum started by GC.
663 // Partially satisfy this assist.
664 gp.gcAssistBytes += scanBytes
666 // As a heuristic, we move this assist to the
667 // back of the queue so that large assists
668 // can't clog up the assist queue and
669 // substantially delay small assists.
670 work.assistQueue.q.pushBack(gp)
676 // Convert from scan bytes back to work.
677 assistWorkPerByte := float64frombits(atomic.Load64(&gcController.assistWorkPerByte))
678 scanWork = int64(float64(scanBytes) * assistWorkPerByte)
679 atomic.Xaddint64(&gcController.bgScanCredit, scanWork)
681 unlock(&work.assistQueue.lock)
684 // scanstack scans gp's stack, greying all pointers found on the stack.
686 // scanstack will also shrink the stack if it is safe to do so. If it
687 // is not, it schedules a stack shrink for the next synchronous safe
690 // scanstack is marked go:systemstack because it must not be preempted
691 // while using a workbuf.
695 func scanstack(gp *g, gcw *gcWork) {
696 if readgstatus(gp)&_Gscan == 0 {
697 print("runtime:scanstack: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", hex(readgstatus(gp)), "\n")
698 throw("scanstack - bad status")
701 switch readgstatus(gp) &^ _Gscan {
703 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
704 throw("mark - bad status")
708 print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
709 throw("scanstack: goroutine not stopped")
710 case _Grunnable, _Gsyscall, _Gwaiting:
715 throw("can't scan our own stack")
718 if isShrinkStackSafe(gp) {
719 // Shrink the stack if not much of it is being used.
722 // Otherwise, shrink the stack at the next sync safe point.
723 gp.preemptShrink = true
726 var state stackScanState
727 state.stack = gp.stack
730 println("stack trace goroutine", gp.goid)
733 if debugScanConservative && gp.asyncSafePoint {
734 print("scanning async preempted goroutine ", gp.goid, " stack [", hex(gp.stack.lo), ",", hex(gp.stack.hi), ")\n")
737 // Scan the saved context register. This is effectively a live
738 // register that gets moved back and forth between the
739 // register and sched.ctxt without a write barrier.
740 if gp.sched.ctxt != nil {
741 scanblock(uintptr(unsafe.Pointer(&gp.sched.ctxt)), sys.PtrSize, &oneptrmask[0], gcw, &state)
744 // Scan the stack. Accumulate a list of stack objects.
745 scanframe := func(frame *stkframe, unused unsafe.Pointer) bool {
746 scanframeworker(frame, &state, gcw)
749 gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, scanframe, nil, 0)
751 // Find additional pointers that point into the stack from the heap.
752 // Currently this includes defers and panics. See also function copystack.
754 // Find and trace all defer arguments.
755 tracebackdefers(gp, scanframe, nil)
757 // Find and trace other pointers in defer records.
758 for d := gp._defer; d != nil; d = d.link {
760 // tracebackdefers above does not scan the func value, which could
761 // be a stack allocated closure. See issue 30453.
762 scanblock(uintptr(unsafe.Pointer(&d.fn)), sys.PtrSize, &oneptrmask[0], gcw, &state)
765 // The link field of a stack-allocated defer record might point
766 // to a heap-allocated defer record. Keep that heap record live.
767 scanblock(uintptr(unsafe.Pointer(&d.link)), sys.PtrSize, &oneptrmask[0], gcw, &state)
769 // Retain defers records themselves.
770 // Defer records might not be reachable from the G through regular heap
771 // tracing because the defer linked list might weave between the stack and the heap.
773 scanblock(uintptr(unsafe.Pointer(&d)), sys.PtrSize, &oneptrmask[0], gcw, &state)
776 if gp._panic != nil {
777 // Panics are always stack allocated.
778 state.putPtr(uintptr(unsafe.Pointer(gp._panic)), false)
781 // Find and scan all reachable stack objects.
783 // The state's pointer queue prioritizes precise pointers over
784 // conservative pointers so that we'll prefer scanning stack
785 // objects precisely.
788 p, conservative := state.getPtr()
792 obj := state.findObject(p)
798 // We've already scanned this object.
801 obj.setType(nil) // Don't scan it again.
804 print(" live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of type", t.string())
806 print(" (conservative)")
813 if t.kind&kindGCProg != 0 {
814 // This path is pretty unlikely, an object large enough
815 // to have a GC program allocated on the stack.
816 // We need some space to unpack the program into a straight
817 // bitmask, which we allocate/free here.
818 // TODO: it would be nice if there were a way to run a GC
819 // program without having to store all its bits. We'd have
820 // to change from a Lempel-Ziv style program to something else.
821 // Or we can forbid putting objects on stacks if they require
822 // a gc program (see issue 27447).
823 s = materializeGCProg(t.ptrdata, gcdata)
824 gcdata = (*byte)(unsafe.Pointer(s.startAddr))
827 b := state.stack.lo + uintptr(obj.off)
829 scanConservative(b, t.ptrdata, gcdata, gcw, &state)
831 scanblock(b, t.ptrdata, gcdata, gcw, &state)
835 dematerializeGCProg(s)
839 // Deallocate object buffers.
840 // (Pointer buffers were all deallocated in the loop above.)
841 for state.head != nil {
845 for i := 0; i < x.nobj; i++ {
847 if obj.typ == nil { // reachable
850 println(" dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of type", obj.typ.string())
851 // Note: not necessarily really dead - only reachable-from-ptr dead.
855 putempty((*workbuf)(unsafe.Pointer(x)))
857 if state.buf != nil || state.cbuf != nil || state.freeBuf != nil {
858 throw("remaining pointer buffers")
862 // Scan a stack frame: local variables and function arguments/results.
864 func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
865 if _DebugGC > 1 && frame.continpc != 0 {
866 print("scanframe ", funcname(frame.fn), "\n")
869 isAsyncPreempt := frame.fn.valid() && frame.fn.funcID == funcID_asyncPreempt
870 isDebugCall := frame.fn.valid() && frame.fn.funcID == funcID_debugCallV1
871 if state.conservative || isAsyncPreempt || isDebugCall {
872 if debugScanConservative {
873 println("conservatively scanning function", funcname(frame.fn), "at PC", hex(frame.continpc))
876 // Conservatively scan the frame. Unlike the precise
877 // case, this includes the outgoing argument space
878 // since we may have stopped while this function was
879 // setting up a call.
881 // TODO: We could narrow this down if the compiler
882 // produced a single map per function of stack slots
883 // and registers that ever contain a pointer.
885 size := frame.varp - frame.sp
887 scanConservative(frame.sp, size, nil, gcw, state)
891 // Scan arguments to this frame.
892 if frame.arglen != 0 {
893 // TODO: We could pass the entry argument map
894 // to narrow this down further.
895 scanConservative(frame.argp, frame.arglen, nil, gcw, state)
898 if isAsyncPreempt || isDebugCall {
899 // This function's frame contained the
900 // registers for the asynchronously stopped
901 // parent frame. Scan the parent
903 state.conservative = true
905 // We only wanted to scan those two frames
906 // conservatively. Clear the flag for future
908 state.conservative = false
913 locals, args, objs := getStackMap(frame, &state.cache, false)
915 // Scan local variables if stack frame has been allocated.
917 size := uintptr(locals.n) * sys.PtrSize
918 scanblock(frame.varp-size, size, locals.bytedata, gcw, state)
923 scanblock(frame.argp, uintptr(args.n)*sys.PtrSize, args.bytedata, gcw, state)
926 // Add all stack objects to the stack object list.
928 // varp is 0 for defers, where there are no locals.
929 // In that case, there can't be a pointer to its args, either.
930 // (And all args would be scanned above anyway.)
931 for _, obj := range objs {
933 base := frame.varp // locals base pointer
935 base = frame.argp // arguments and return values base pointer
937 ptr := base + uintptr(off)
939 // object hasn't been allocated in the frame yet.
943 println("stkobj at", hex(ptr), "of type", obj.typ.string())
945 state.addObject(ptr, obj.typ)
950 type gcDrainFlags int
953 gcDrainUntilPreempt gcDrainFlags = 1 << iota
959 // gcDrain scans roots and objects in work buffers, blackening grey
960 // objects until it is unable to get more work. It may return before
961 // GC is done; it's the caller's responsibility to balance work from
964 // If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
967 // If flags&gcDrainIdle != 0, gcDrain returns when there is other work
970 // If flags&gcDrainFractional != 0, gcDrain self-preempts when
971 // pollFractionalWorkerExit() returns true. This implies
974 // If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
975 // credit to gcController.bgScanCredit every gcCreditSlack units of
978 // gcDrain will always return if there is a pending STW.
981 func gcDrain(gcw *gcWork, flags gcDrainFlags) {
982 if !writeBarrier.needed {
983 throw("gcDrain phase incorrect")
987 preemptible := flags&gcDrainUntilPreempt != 0
988 flushBgCredit := flags&gcDrainFlushBgCredit != 0
989 idle := flags&gcDrainIdle != 0
991 initScanWork := gcw.scanWork
993 // checkWork is the scan work before performing the next
994 // self-preempt check.
995 checkWork := int64(1<<63 - 1)
996 var check func() bool
997 if flags&(gcDrainIdle|gcDrainFractional) != 0 {
998 checkWork = initScanWork + drainCheckThreshold
1001 } else if flags&gcDrainFractional != 0 {
1002 check = pollFractionalWorkerExit
1006 // Drain root marking jobs.
1007 if work.markrootNext < work.markrootJobs {
1008 // Stop if we're preemptible or if someone wants to STW.
1009 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1010 job := atomic.Xadd(&work.markrootNext, +1) - 1
1011 if job >= work.markrootJobs {
1015 if check != nil && check() {
1021 // Drain heap marking jobs.
1022 // Stop if we're preemptible or if someone wants to STW.
1023 for !(gp.preempt && (preemptible || atomic.Load(&sched.gcwaiting) != 0)) {
1024 // Try to keep work available on the global queue. We used to
1025 // check if there were waiting workers, but it's better to
1026 // just keep work available than to make workers wait. In the
1027 // worst case, we'll do O(log(_WorkbufSize)) unnecessary
1033 b := gcw.tryGetFast()
1037 // Flush the write barrier
1038 // buffer; this may create
1045 // Unable to get work.
1050 // Flush background scan work credit to the global
1051 // account if we've accumulated enough locally so
1052 // mutator assists can draw on it.
1053 if gcw.scanWork >= gcCreditSlack {
1054 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1056 gcFlushBgCredit(gcw.scanWork - initScanWork)
1059 checkWork -= gcw.scanWork
1063 checkWork += drainCheckThreshold
1064 if check != nil && check() {
1072 // Flush remaining scan work credit.
1073 if gcw.scanWork > 0 {
1074 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1076 gcFlushBgCredit(gcw.scanWork - initScanWork)
1082 // gcDrainN blackens grey objects until it has performed roughly
1083 // scanWork units of scan work or the G is preempted. This is
1084 // best-effort, so it may perform less work if it fails to get a work
1085 // buffer. Otherwise, it will perform at least n units of work, but
1086 // may perform more because scanning is always done in whole object
1087 // increments. It returns the amount of scan work performed.
1089 // The caller goroutine must be in a preemptible state (e.g.,
1090 // _Gwaiting) to prevent deadlocks during stack scanning. As a
1091 // consequence, this must be called on the system stack.
1095 func gcDrainN(gcw *gcWork, scanWork int64) int64 {
1096 if !writeBarrier.needed {
1097 throw("gcDrainN phase incorrect")
1100 // There may already be scan work on the gcw, which we don't
1101 // want to claim was done by this call.
1102 workFlushed := -gcw.scanWork
1105 for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
1106 // See gcDrain comment.
1111 // This might be a good place to add prefetch code...
1112 // if(wbuf.nobj > 4) {
1113 // PREFETCH(wbuf->obj[wbuf.nobj - 3];
1116 b := gcw.tryGetFast()
1120 // Flush the write barrier buffer;
1121 // this may create more work.
1128 // Try to do a root job.
1130 // TODO: Assists should get credit for this
1132 if work.markrootNext < work.markrootJobs {
1133 job := atomic.Xadd(&work.markrootNext, +1) - 1
1134 if job < work.markrootJobs {
1139 // No heap or root jobs.
1144 // Flush background scan work credit.
1145 if gcw.scanWork >= gcCreditSlack {
1146 atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
1147 workFlushed += gcw.scanWork
1152 // Unlike gcDrain, there's no need to flush remaining work
1153 // here because this never flushes to bgScanCredit and
1154 // gcw.dispose will flush any remaining work to scanWork.
1156 return workFlushed + gcw.scanWork
1159 // scanblock scans b as scanobject would, but using an explicit
1160 // pointer bitmap instead of the heap bitmap.
1162 // This is used to scan non-heap roots, so it does not update
1163 // gcw.bytesMarked or gcw.scanWork.
1165 // If stk != nil, possible stack pointers are also reported to stk.putPtr.
1167 func scanblock(b0, n0 uintptr, ptrmask *uint8, gcw *gcWork, stk *stackScanState) {
1168 // Use local copies of original parameters, so that a stack trace
1169 // due to one of the throws below shows the original block
1174 for i := uintptr(0); i < n; {
1175 // Find bits for the next word.
1176 bits := uint32(*addb(ptrmask, i/(sys.PtrSize*8)))
1178 i += sys.PtrSize * 8
1181 for j := 0; j < 8 && i < n; j++ {
1183 // Same work as in scanobject; see comments there.
1184 p := *(*uintptr)(unsafe.Pointer(b + i))
1186 if obj, span, objIndex := findObject(p, b, i); obj != 0 {
1187 greyobject(obj, b, i, span, gcw, objIndex)
1188 } else if stk != nil && p >= stk.stack.lo && p < stk.stack.hi {
1189 stk.putPtr(p, false)
1199 // scanobject scans the object starting at b, adding pointers to gcw.
1200 // b must point to the beginning of a heap object or an oblet.
1201 // scanobject consults the GC bitmap for the pointer mask and the
1202 // spans for the size of the object.
1205 func scanobject(b uintptr, gcw *gcWork) {
1206 // Find the bits for b and the size of the object at b.
1208 // b is either the beginning of an object, in which case this
1209 // is the size of the object to scan, or it points to an
1210 // oblet, in which case we compute the size to scan below.
1211 hbits := heapBitsForAddr(b)
1212 s := spanOfUnchecked(b)
1215 throw("scanobject n == 0")
1218 if n > maxObletBytes {
1219 // Large object. Break into oblets for better
1220 // parallelism and lower latency.
1222 // It's possible this is a noscan object (not
1223 // from greyobject, but from other code
1224 // paths), in which case we must *not* enqueue
1225 // oblets since their bitmaps will be
1227 if s.spanclass.noscan() {
1228 // Bypass the whole scan.
1229 gcw.bytesMarked += uint64(n)
1233 // Enqueue the other oblets to scan later.
1234 // Some oblets may be in b's scalar tail, but
1235 // these will be marked as "no more pointers",
1236 // so we'll drop out immediately when we go to
1238 for oblet := b + maxObletBytes; oblet < s.base()+s.elemsize; oblet += maxObletBytes {
1239 if !gcw.putFast(oblet) {
1245 // Compute the size of the oblet. Since this object
1246 // must be a large object, s.base() is the beginning
1248 n = s.base() + s.elemsize - b
1249 if n > maxObletBytes {
1255 for i = 0; i < n; i += sys.PtrSize {
1256 // Find bits for this word.
1258 // Avoid needless hbits.next() on last iteration.
1259 hbits = hbits.next()
1261 // Load bits once. See CL 22712 and issue 16973 for discussion.
1262 bits := hbits.bits()
1263 if bits&bitScan == 0 {
1264 break // no more pointers in this object
1266 if bits&bitPointer == 0 {
1267 continue // not a pointer
1270 // Work here is duplicated in scanblock and above.
1271 // If you make changes here, make changes there too.
1272 obj := *(*uintptr)(unsafe.Pointer(b + i))
1274 // At this point we have extracted the next potential pointer.
1275 // Quickly filter out nil and pointers back to the current object.
1276 if obj != 0 && obj-b >= n {
1277 // Test if obj points into the Go heap and, if so,
1280 // Note that it's possible for findObject to
1281 // fail if obj points to a just-allocated heap
1282 // object because of a race with growing the
1283 // heap. In this case, we know the object was
1284 // just allocated and hence will be marked by
1285 // allocation itself.
1286 if obj, span, objIndex := findObject(obj, b, i); obj != 0 {
1287 greyobject(obj, b, i, span, gcw, objIndex)
1291 gcw.bytesMarked += uint64(n)
1292 gcw.scanWork += int64(i)
1295 // scanConservative scans block [b, b+n) conservatively, treating any
1296 // pointer-like value in the block as a pointer.
1298 // If ptrmask != nil, only words that are marked in ptrmask are
1299 // considered as potential pointers.
1301 // If state != nil, it's assumed that [b, b+n) is a block in the stack
1302 // and may contain pointers to stack objects.
1303 func scanConservative(b, n uintptr, ptrmask *uint8, gcw *gcWork, state *stackScanState) {
1304 if debugScanConservative {
1306 print("conservatively scanning [", hex(b), ",", hex(b+n), ")\n")
1307 hexdumpWords(b, b+n, func(p uintptr) byte {
1309 word := (p - b) / sys.PtrSize
1310 bits := *addb(ptrmask, word/8)
1311 if (bits>>(word%8))&1 == 0 {
1316 val := *(*uintptr)(unsafe.Pointer(p))
1317 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1321 span := spanOfHeap(val)
1325 idx := span.objIndex(val)
1326 if span.isFree(idx) {
1334 for i := uintptr(0); i < n; i += sys.PtrSize {
1336 word := i / sys.PtrSize
1337 bits := *addb(ptrmask, word/8)
1339 // Skip 8 words (the loop increment will do the 8th)
1341 // This must be the first time we've
1342 // seen this word of ptrmask, so i
1343 // must be 8-word-aligned, but check
1344 // our reasoning just in case.
1345 if i%(sys.PtrSize*8) != 0 {
1346 throw("misaligned mask")
1348 i += sys.PtrSize*8 - sys.PtrSize
1351 if (bits>>(word%8))&1 == 0 {
1356 val := *(*uintptr)(unsafe.Pointer(b + i))
1358 // Check if val points into the stack.
1359 if state != nil && state.stack.lo <= val && val < state.stack.hi {
1360 // val may point to a stack object. This
1361 // object may be dead from last cycle and
1362 // hence may contain pointers to unallocated
1363 // objects, but unlike heap objects we can't
1364 // tell if it's already dead. Hence, if all
1365 // pointers to this object are from
1366 // conservative scanning, we have to scan it
1367 // defensively, too.
1368 state.putPtr(val, true)
1372 // Check if val points to a heap span.
1373 span := spanOfHeap(val)
1378 // Check if val points to an allocated object.
1379 idx := span.objIndex(val)
1380 if span.isFree(idx) {
1384 // val points to an allocated object. Mark it.
1385 obj := span.base() + idx*span.elemsize
1386 greyobject(obj, b, i, span, gcw, idx)
1390 // Shade the object if it isn't already.
1391 // The object is not nil and known to be in the heap.
1392 // Preemption must be disabled.
1394 func shade(b uintptr) {
1395 if obj, span, objIndex := findObject(b, 0, 0); obj != 0 {
1396 gcw := &getg().m.p.ptr().gcw
1397 greyobject(obj, 0, 0, span, gcw, objIndex)
1401 // obj is the start of an object with mark mbits.
1402 // If it isn't already marked, mark it and enqueue into gcw.
1403 // base and off are for debugging only and could be removed.
1405 // See also wbBufFlush1, which partially duplicates this logic.
1407 //go:nowritebarrierrec
1408 func greyobject(obj, base, off uintptr, span *mspan, gcw *gcWork, objIndex uintptr) {
1409 // obj should be start of allocation, and so must be at least pointer-aligned.
1410 if obj&(sys.PtrSize-1) != 0 {
1411 throw("greyobject: obj not pointer-aligned")
1413 mbits := span.markBitsForIndex(objIndex)
1416 if setCheckmark(obj, base, off, mbits) {
1421 if debug.gccheckmark > 0 && span.isFree(objIndex) {
1422 print("runtime: marking free object ", hex(obj), " found at *(", hex(base), "+", hex(off), ")\n")
1423 gcDumpObject("base", base, off)
1424 gcDumpObject("obj", obj, ^uintptr(0))
1425 getg().m.traceback = 2
1426 throw("marking free object")
1429 // If marked we have nothing to do.
1430 if mbits.isMarked() {
1436 arena, pageIdx, pageMask := pageIndexOf(span.base())
1437 if arena.pageMarks[pageIdx]&pageMask == 0 {
1438 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1441 // If this is a noscan object, fast-track it to black
1442 // instead of greying it.
1443 if span.spanclass.noscan() {
1444 gcw.bytesMarked += uint64(span.elemsize)
1449 // Queue the obj for scanning. The PREFETCH(obj) logic has been removed but
1450 // seems like a nice optimization that can be added back in.
1451 // There needs to be time between the PREFETCH and the use.
1452 // Previously we put the obj in an 8 element buffer that is drained at a rate
1453 // to give the PREFETCH time to do its work.
1454 // Use of PREFETCHNTA might be more appropriate than PREFETCH
1455 if !gcw.putFast(obj) {
1460 // gcDumpObject dumps the contents of obj for debugging and marks the
1461 // field at byte offset off in obj.
1462 func gcDumpObject(label string, obj, off uintptr) {
1464 print(label, "=", hex(obj))
1469 print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
1470 if state := s.state.get(); 0 <= state && int(state) < len(mSpanStateNames) {
1471 print(mSpanStateNames[state], "\n")
1473 print("unknown(", state, ")\n")
1478 if s.state.get() == mSpanManual && size == 0 {
1479 // We're printing something from a stack frame. We
1480 // don't know how big it is, so just show up to an
1482 size = off + sys.PtrSize
1484 for i := uintptr(0); i < size; i += sys.PtrSize {
1485 // For big objects, just print the beginning (because
1486 // that usually hints at the object's type) and the
1487 // fields around off.
1488 if !(i < 128*sys.PtrSize || off-16*sys.PtrSize < i && i < off+16*sys.PtrSize) {
1496 print(" *(", label, "+", i, ") = ", hex(*(*uintptr)(unsafe.Pointer(obj + i))))
1507 // gcmarknewobject marks a newly allocated object black. obj must
1508 // not contain any non-nil pointers.
1510 // This is nosplit so it can manipulate a gcWork without preemption.
1514 func gcmarknewobject(span *mspan, obj, size, scanSize uintptr) {
1515 if useCheckmark { // The world should be stopped so this should not happen.
1516 throw("gcmarknewobject called while doing checkmark")
1520 objIndex := span.objIndex(obj)
1521 span.markBitsForIndex(objIndex).setMarked()
1524 arena, pageIdx, pageMask := pageIndexOf(span.base())
1525 if arena.pageMarks[pageIdx]&pageMask == 0 {
1526 atomic.Or8(&arena.pageMarks[pageIdx], pageMask)
1529 gcw := &getg().m.p.ptr().gcw
1530 gcw.bytesMarked += uint64(size)
1531 gcw.scanWork += int64(scanSize)
1534 // gcMarkTinyAllocs greys all active tiny alloc blocks.
1536 // The world must be stopped.
1537 func gcMarkTinyAllocs() {
1538 assertWorldStopped()
1540 for _, p := range allp {
1542 if c == nil || c.tiny == 0 {
1545 _, span, objIndex := findObject(c.tiny, 0, 0)
1547 greyobject(c.tiny, 0, 0, span, gcw, objIndex)