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 (GC).
7 // The GC runs concurrently with mutator threads, is type accurate (aka precise), allows multiple
8 // GC thread to run in parallel. It is a concurrent mark and sweep that uses a write barrier. It is
9 // non-generational and non-compacting. Allocation is done using size segregated per P allocation
10 // areas to minimize fragmentation while eliminating locks in the common case.
12 // The algorithm decomposes into several steps.
13 // This is a high level description of the algorithm being used. For an overview of GC a good
14 // place to start is Richard Jones' gchandbook.org.
16 // The algorithm's intellectual heritage includes Dijkstra's on-the-fly algorithm, see
17 // Edsger W. Dijkstra, Leslie Lamport, A. J. Martin, C. S. Scholten, and E. F. M. Steffens. 1978.
18 // On-the-fly garbage collection: an exercise in cooperation. Commun. ACM 21, 11 (November 1978),
20 // For journal quality proofs that these steps are complete, correct, and terminate see
21 // Hudson, R., and Moss, J.E.B. Copying Garbage Collection without stopping the world.
22 // Concurrency and Computation: Practice and Experience 15(3-5), 2003.
24 // 1. GC performs sweep termination.
26 // a. Stop the world. This causes all Ps to reach a GC safe-point.
28 // b. Sweep any unswept spans. There will only be unswept spans if
29 // this GC cycle was forced before the expected time.
31 // 2. GC performs the mark phase.
33 // a. Prepare for the mark phase by setting gcphase to _GCmark
34 // (from _GCoff), enabling the write barrier, enabling mutator
35 // assists, and enqueueing root mark jobs. No objects may be
36 // scanned until all Ps have enabled the write barrier, which is
37 // accomplished using STW.
39 // b. Start the world. From this point, GC work is done by mark
40 // workers started by the scheduler and by assists performed as
41 // part of allocation. The write barrier shades both the
42 // overwritten pointer and the new pointer value for any pointer
43 // writes (see mbarrier.go for details). Newly allocated objects
44 // are immediately marked black.
46 // c. GC performs root marking jobs. This includes scanning all
47 // stacks, shading all globals, and shading any heap pointers in
48 // off-heap runtime data structures. Scanning a stack stops a
49 // goroutine, shades any pointers found on its stack, and then
50 // resumes the goroutine.
52 // d. GC drains the work queue of grey objects, scanning each grey
53 // object to black and shading all pointers found in the object
54 // (which in turn may add those pointers to the work queue).
56 // e. Because GC work is spread across local caches, GC uses a
57 // distributed termination algorithm to detect when there are no
58 // more root marking jobs or grey objects (see gcMarkDone). At this
59 // point, GC transitions to mark termination.
61 // 3. GC performs mark termination.
65 // b. Set gcphase to _GCmarktermination, and disable workers and
68 // c. Perform housekeeping like flushing mcaches.
70 // 4. GC performs the sweep phase.
72 // a. Prepare for the sweep phase by setting gcphase to _GCoff,
73 // setting up sweep state and disabling the write barrier.
75 // b. Start the world. From this point on, newly allocated objects
76 // are white, and allocating sweeps spans before use if necessary.
78 // c. GC does concurrent sweeping in the background and in response
79 // to allocation. See description below.
81 // 5. When sufficient allocation has taken place, replay the sequence
82 // starting with 1 above. See discussion of GC rate below.
86 // The sweep phase proceeds concurrently with normal program execution.
87 // The heap is swept span-by-span both lazily (when a goroutine needs another span)
88 // and concurrently in a background goroutine (this helps programs that are not CPU bound).
89 // At the end of STW mark termination all spans are marked as "needs sweeping".
91 // The background sweeper goroutine simply sweeps spans one-by-one.
93 // To avoid requesting more OS memory while there are unswept spans, when a
94 // goroutine needs another span, it first attempts to reclaim that much memory
95 // by sweeping. When a goroutine needs to allocate a new small-object span, it
96 // sweeps small-object spans for the same object size until it frees at least
97 // one object. When a goroutine needs to allocate large-object span from heap,
98 // it sweeps spans until it frees at least that many pages into heap. There is
99 // one case where this may not suffice: if a goroutine sweeps and frees two
100 // nonadjacent one-page spans to the heap, it will allocate a new two-page
101 // span, but there can still be other one-page unswept spans which could be
102 // combined into a two-page span.
104 // It's critical to ensure that no operations proceed on unswept spans (that would corrupt
105 // mark bits in GC bitmap). During GC all mcaches are flushed into the central cache,
106 // so they are empty. When a goroutine grabs a new span into mcache, it sweeps it.
107 // When a goroutine explicitly frees an object or sets a finalizer, it ensures that
108 // the span is swept (either by sweeping it, or by waiting for the concurrent sweep to finish).
109 // The finalizer goroutine is kicked off only when all spans are swept.
110 // When the next GC starts, it sweeps all not-yet-swept spans (if any).
113 // Next GC is after we've allocated an extra amount of memory proportional to
114 // the amount already in use. The proportion is controlled by GOGC environment variable
115 // (100 by default). If GOGC=100 and we're using 4M, we'll GC again when we get to 8M
116 // (this mark is computed by the gcController.heapGoal method). This keeps the GC cost in
117 // linear proportion to the allocation cost. Adjusting GOGC just changes the linear constant
118 // (and also the amount of extra memory used).
122 // In order to prevent long pauses while scanning large objects and to
123 // improve parallelism, the garbage collector breaks up scan jobs for
124 // objects larger than maxObletBytes into "oblets" of at most
125 // maxObletBytes. When scanning encounters the beginning of a large
126 // object, it scans only the first oblet and enqueues the remaining
127 // oblets as new scan jobs.
133 "runtime/internal/atomic"
139 _ConcurrentSweep = true
140 _FinBlockSize = 4 * 1024
142 // debugScanConservative enables debug logging for stack
143 // frames that are scanned conservatively.
144 debugScanConservative = false
146 // sweepMinHeapDistance is a lower bound on the heap distance
147 // (in bytes) reserved for concurrent sweeping between GC
149 sweepMinHeapDistance = 1024 * 1024
153 if unsafe.Sizeof(workbuf{}) != _WorkbufSize {
154 throw("size of Workbuf is suboptimal")
156 // No sweep on the first cycle.
157 sweep.active.state.Store(sweepDrainedMask)
159 // Initialize GC pacer state.
160 // Use the environment variable GOGC for the initial gcPercent value.
161 // Use the environment variable GOMEMLIMIT for the initial memoryLimit value.
162 gcController.init(readGOGC(), readGOMEMLIMIT())
165 work.markDoneSema = 1
166 lockInit(&work.sweepWaiters.lock, lockRankSweepWaiters)
167 lockInit(&work.assistQueue.lock, lockRankAssistQueue)
168 lockInit(&work.wbufSpans.lock, lockRankWbufSpans)
171 // gcenable is called after the bulk of the runtime initialization,
172 // just before we're about to start letting user code run.
173 // It kicks off the background sweeper goroutine, the background
174 // scavenger goroutine, and enables GC.
176 // Kick off sweeping and scavenging.
177 c := make(chan int, 2)
182 memstats.enablegc = true // now that runtime is initialized, GC is okay
185 // Garbage collector phase.
186 // Indicates to write barrier and synchronization task to perform.
189 // The compiler knows about this variable.
190 // If you change it, you must change builtin/runtime.go, too.
191 // If you change the first four bytes, you must also change the write
192 // barrier insertion code.
193 var writeBarrier struct {
194 enabled bool // compiler emits a check of this before calling write barrier
195 pad [3]byte // compiler uses 32-bit load for "enabled" field
196 needed bool // identical to enabled, for now (TODO: dedup)
197 alignme uint64 // guarantee alignment so that compiler can use a 32 or 64-bit load
200 // gcBlackenEnabled is 1 if mutator assists and background mark
201 // workers are allowed to blacken objects. This must only be set when
202 // gcphase == _GCmark.
203 var gcBlackenEnabled uint32
206 _GCoff = iota // GC not running; sweeping in background, write barrier disabled
207 _GCmark // GC marking roots and workbufs: allocate black, write barrier ENABLED
208 _GCmarktermination // GC mark termination: allocate black, P's help GC, write barrier ENABLED
212 func setGCPhase(x uint32) {
213 atomic.Store(&gcphase, x)
214 writeBarrier.needed = gcphase == _GCmark || gcphase == _GCmarktermination
215 writeBarrier.enabled = writeBarrier.needed
218 // gcMarkWorkerMode represents the mode that a concurrent mark worker
219 // should operate in.
221 // Concurrent marking happens through four different mechanisms. One
222 // is mutator assists, which happen in response to allocations and are
223 // not scheduled. The other three are variations in the per-P mark
224 // workers and are distinguished by gcMarkWorkerMode.
225 type gcMarkWorkerMode int
228 // gcMarkWorkerNotWorker indicates that the next scheduled G is not
229 // starting work and the mode should be ignored.
230 gcMarkWorkerNotWorker gcMarkWorkerMode = iota
232 // gcMarkWorkerDedicatedMode indicates that the P of a mark
233 // worker is dedicated to running that mark worker. The mark
234 // worker should run without preemption.
235 gcMarkWorkerDedicatedMode
237 // gcMarkWorkerFractionalMode indicates that a P is currently
238 // running the "fractional" mark worker. The fractional worker
239 // is necessary when GOMAXPROCS*gcBackgroundUtilization is not
240 // an integer and using only dedicated workers would result in
241 // utilization too far from the target of gcBackgroundUtilization.
242 // The fractional worker should run until it is preempted and
243 // will be scheduled to pick up the fractional part of
244 // GOMAXPROCS*gcBackgroundUtilization.
245 gcMarkWorkerFractionalMode
247 // gcMarkWorkerIdleMode indicates that a P is running the mark
248 // worker because it has nothing else to do. The idle worker
249 // should run until it is preempted and account its time
250 // against gcController.idleMarkTime.
254 // gcMarkWorkerModeStrings are the strings labels of gcMarkWorkerModes
255 // to use in execution traces.
256 var gcMarkWorkerModeStrings = [...]string{
263 // pollFractionalWorkerExit reports whether a fractional mark worker
264 // should self-preempt. It assumes it is called from the fractional
266 func pollFractionalWorkerExit() bool {
267 // This should be kept in sync with the fractional worker
268 // scheduler logic in findRunnableGCWorker.
270 delta := now - gcController.markStartTime
274 p := getg().m.p.ptr()
275 selfTime := p.gcFractionalMarkTime + (now - p.gcMarkWorkerStartTime)
276 // Add some slack to the utilization goal so that the
277 // fractional worker isn't behind again the instant it exits.
278 return float64(selfTime)/float64(delta) > 1.2*gcController.fractionalUtilizationGoal
283 type workType struct {
284 full lfstack // lock-free list of full blocks workbuf
285 empty lfstack // lock-free list of empty blocks workbuf
286 pad0 cpu.CacheLinePad // prevents false-sharing between full/empty and nproc/nwait
290 // free is a list of spans dedicated to workbufs, but
291 // that don't currently contain any workbufs.
293 // busy is a list of all spans containing workbufs on
294 // one of the workbuf lists.
298 // Restore 64-bit alignment on 32-bit.
301 // bytesMarked is the number of bytes marked this cycle. This
302 // includes bytes blackened in scanned objects, noscan objects
303 // that go straight to black, and permagrey objects scanned by
304 // markroot during the concurrent scan phase. This is updated
305 // atomically during the cycle. Updates may be batched
306 // arbitrarily, since the value is only read at the end of the
309 // Because of benign races during marking, this number may not
310 // be the exact number of marked bytes, but it should be very
313 // Put this field here because it needs 64-bit atomic access
314 // (and thus 8-byte alignment even on 32-bit architectures).
317 markrootNext uint32 // next markroot job
318 markrootJobs uint32 // number of markroot jobs
324 // Number of roots of various root types. Set by gcMarkRootPrepare.
326 // nStackRoots == len(stackRoots), but we have nStackRoots for
328 nDataRoots, nBSSRoots, nSpanRoots, nStackRoots int
330 // Base indexes of each root type. Set by gcMarkRootPrepare.
331 baseData, baseBSS, baseSpans, baseStacks, baseEnd uint32
333 // stackRoots is a snapshot of all of the Gs that existed
334 // before the beginning of concurrent marking. The backing
335 // store of this must not be modified because it might be
336 // shared with allgs.
339 // Each type of GC state transition is protected by a lock.
340 // Since multiple threads can simultaneously detect the state
341 // transition condition, any thread that detects a transition
342 // condition must acquire the appropriate transition lock,
343 // re-check the transition condition and return if it no
344 // longer holds or perform the transition if it does.
345 // Likewise, any transition must invalidate the transition
346 // condition before releasing the lock. This ensures that each
347 // transition is performed by exactly one thread and threads
348 // that need the transition to happen block until it has
351 // startSema protects the transition from "off" to mark or
354 // markDoneSema protects transitions from mark to mark termination.
357 bgMarkReady note // signal background mark worker has started
358 bgMarkDone uint32 // cas to 1 when at a background mark completion point
359 // Background mark completion signaling
361 // mode is the concurrency mode of the current GC cycle.
364 // userForced indicates the current GC cycle was forced by an
365 // explicit user call.
368 // initialHeapLive is the value of gcController.heapLive at the
369 // beginning of this GC cycle.
370 initialHeapLive uint64
372 // assistQueue is a queue of assists that are blocked because
373 // there was neither enough credit to steal or enough work to
380 // sweepWaiters is a list of blocked goroutines to wake when
381 // we transition from mark termination to sweep.
382 sweepWaiters struct {
387 // cycles is the number of completed GC cycles, where a GC
388 // cycle is sweep termination, mark, mark termination, and
389 // sweep. This differs from memstats.numgc, which is
390 // incremented at mark termination.
393 // Timing/utilization stats for this cycle.
394 stwprocs, maxprocs int32
395 tSweepTerm, tMark, tMarkTerm, tEnd int64 // nanotime() of phase start
397 pauseNS int64 // total STW time this cycle
398 pauseStart int64 // nanotime() of last STW
400 // debug.gctrace heap sizes for this cycle.
401 heap0, heap1, heap2 uint64
403 // Cumulative estimated CPU usage.
407 // GC runs a garbage collection and blocks the caller until the
408 // garbage collection is complete. It may also block the entire
411 // We consider a cycle to be: sweep termination, mark, mark
412 // termination, and sweep. This function shouldn't return
413 // until a full cycle has been completed, from beginning to
414 // end. Hence, we always want to finish up the current cycle
415 // and start a new one. That means:
417 // 1. In sweep termination, mark, or mark termination of cycle
418 // N, wait until mark termination N completes and transitions
421 // 2. In sweep N, help with sweep N.
423 // At this point we can begin a full cycle N+1.
425 // 3. Trigger cycle N+1 by starting sweep termination N+1.
427 // 4. Wait for mark termination N+1 to complete.
429 // 5. Help with sweep N+1 until it's done.
431 // This all has to be written to deal with the fact that the
432 // GC may move ahead on its own. For example, when we block
433 // until mark termination N, we may wake up in cycle N+2.
435 // Wait until the current sweep termination, mark, and mark
436 // termination complete.
437 n := work.cycles.Load()
440 // We're now in sweep N or later. Trigger GC cycle N+1, which
441 // will first finish sweep N if necessary and then enter sweep
443 gcStart(gcTrigger{kind: gcTriggerCycle, n: n + 1})
445 // Wait for mark termination N+1 to complete.
448 // Finish sweep N+1 before returning. We do this both to
449 // complete the cycle and because runtime.GC() is often used
450 // as part of tests and benchmarks to get the system into a
451 // relatively stable and isolated state.
452 for work.cycles.Load() == n+1 && sweepone() != ^uintptr(0) {
457 // Callers may assume that the heap profile reflects the
458 // just-completed cycle when this returns (historically this
459 // happened because this was a STW GC), but right now the
460 // profile still reflects mark termination N, not N+1.
462 // As soon as all of the sweep frees from cycle N+1 are done,
463 // we can go ahead and publish the heap profile.
465 // First, wait for sweeping to finish. (We know there are no
466 // more spans on the sweep queue, but we may be concurrently
467 // sweeping spans, so we have to wait.)
468 for work.cycles.Load() == n+1 && !isSweepDone() {
472 // Now we're really done with sweeping, so we can publish the
473 // stable heap profile. Only do this if we haven't already hit
474 // another mark termination.
476 cycle := work.cycles.Load()
477 if cycle == n+1 || (gcphase == _GCmark && cycle == n+2) {
483 // gcWaitOnMark blocks until GC finishes the Nth mark phase. If GC has
484 // already completed this mark phase, it returns immediately.
485 func gcWaitOnMark(n uint32) {
487 // Disable phase transitions.
488 lock(&work.sweepWaiters.lock)
489 nMarks := work.cycles.Load()
490 if gcphase != _GCmark {
491 // We've already completed this cycle's mark.
496 unlock(&work.sweepWaiters.lock)
500 // Wait until sweep termination, mark, and mark
501 // termination of cycle N complete.
502 work.sweepWaiters.list.push(getg())
503 goparkunlock(&work.sweepWaiters.lock, waitReasonWaitForGCCycle, traceEvGoBlock, 1)
507 // gcMode indicates how concurrent a GC cycle should be.
511 gcBackgroundMode gcMode = iota // concurrent GC and sweep
512 gcForceMode // stop-the-world GC now, concurrent sweep
513 gcForceBlockMode // stop-the-world GC now and STW sweep (forced by user)
516 // A gcTrigger is a predicate for starting a GC cycle. Specifically,
517 // it is an exit condition for the _GCoff phase.
518 type gcTrigger struct {
520 now int64 // gcTriggerTime: current time
521 n uint32 // gcTriggerCycle: cycle number to start
524 type gcTriggerKind int
527 // gcTriggerHeap indicates that a cycle should be started when
528 // the heap size reaches the trigger heap size computed by the
530 gcTriggerHeap gcTriggerKind = iota
532 // gcTriggerTime indicates that a cycle should be started when
533 // it's been more than forcegcperiod nanoseconds since the
534 // previous GC cycle.
537 // gcTriggerCycle indicates that a cycle should be started if
538 // we have not yet started cycle number gcTrigger.n (relative
543 // test reports whether the trigger condition is satisfied, meaning
544 // that the exit condition for the _GCoff phase has been met. The exit
545 // condition should be tested when allocating.
546 func (t gcTrigger) test() bool {
547 if !memstats.enablegc || panicking.Load() != 0 || gcphase != _GCoff {
552 // Non-atomic access to gcController.heapLive for performance. If
553 // we are going to trigger on this, this thread just
554 // atomically wrote gcController.heapLive anyway and we'll see our
556 trigger, _ := gcController.trigger()
557 return gcController.heapLive.Load() >= trigger
559 if gcController.gcPercent.Load() < 0 {
562 lastgc := int64(atomic.Load64(&memstats.last_gc_nanotime))
563 return lastgc != 0 && t.now-lastgc > forcegcperiod
565 // t.n > work.cycles, but accounting for wraparound.
566 return int32(t.n-work.cycles.Load()) > 0
571 // gcStart starts the GC. It transitions from _GCoff to _GCmark (if
572 // debug.gcstoptheworld == 0) or performs all of GC (if
573 // debug.gcstoptheworld != 0).
575 // This may return without performing this transition in some cases,
576 // such as when called on a system stack or with locks held.
577 func gcStart(trigger gcTrigger) {
578 // Since this is called from malloc and malloc is called in
579 // the guts of a number of libraries that might be holding
580 // locks, don't attempt to start GC in non-preemptible or
581 // potentially unstable situations.
583 if gp := getg(); gp == mp.g0 || mp.locks > 1 || mp.preemptoff != "" {
590 // Pick up the remaining unswept/not being swept spans concurrently
592 // This shouldn't happen if we're being invoked in background
593 // mode since proportional sweep should have just finished
594 // sweeping everything, but rounding errors, etc, may leave a
595 // few spans unswept. In forced mode, this is necessary since
596 // GC can be forced at any point in the sweeping cycle.
598 // We check the transition condition continuously here in case
599 // this G gets delayed in to the next GC cycle.
600 for trigger.test() && sweepone() != ^uintptr(0) {
604 // Perform GC initialization and the sweep termination
606 semacquire(&work.startSema)
607 // Re-check transition condition under transition lock.
609 semrelease(&work.startSema)
613 // In gcstoptheworld debug mode, upgrade the mode accordingly.
614 // We do this after re-checking the transition condition so
615 // that multiple goroutines that detect the heap trigger don't
616 // start multiple STW GCs.
617 mode := gcBackgroundMode
618 if debug.gcstoptheworld == 1 {
620 } else if debug.gcstoptheworld == 2 {
621 mode = gcForceBlockMode
624 // Ok, we're doing it! Stop everybody else
626 semacquire(&worldsema)
628 // For stats, check if this GC was forced by the user.
629 // Update it under gcsema to avoid gctrace getting wrong values.
630 work.userForced = trigger.kind == gcTriggerCycle
636 // Check that all Ps have finished deferred mcache flushes.
637 for _, p := range allp {
638 if fg := p.mcache.flushGen.Load(); fg != mheap_.sweepgen {
639 println("runtime: p", p.id, "flushGen", fg, "!= sweepgen", mheap_.sweepgen)
640 throw("p mcache not flushed")
644 gcBgMarkStartWorkers()
646 systemstack(gcResetMarkState)
648 work.stwprocs, work.maxprocs = gomaxprocs, gomaxprocs
649 if work.stwprocs > ncpu {
650 // This is used to compute CPU time of the STW phases,
651 // so it can't be more than ncpu, even if GOMAXPROCS is.
654 work.heap0 = gcController.heapLive.Load()
659 work.tSweepTerm = now
660 work.pauseStart = now
661 systemstack(func() { stopTheWorldWithSema(stwGCSweepTerm) })
662 // Finish sweep before we start concurrent scan.
667 // clearpools before we start the GC. If we wait they memory will not be
668 // reclaimed until the next GC cycle.
673 // Assists and workers can start the moment we start
675 gcController.startCycle(now, int(gomaxprocs), trigger)
677 // Notify the CPU limiter that assists may begin.
678 gcCPULimiter.startGCTransition(true, now)
680 // In STW mode, disable scheduling of user Gs. This may also
681 // disable scheduling of this goroutine, so it may block as
682 // soon as we start the world again.
683 if mode != gcBackgroundMode {
684 schedEnableUser(false)
687 // Enter concurrent mark phase and enable
690 // Because the world is stopped, all Ps will
691 // observe that write barriers are enabled by
692 // the time we start the world and begin
695 // Write barriers must be enabled before assists are
696 // enabled because they must be enabled before
697 // any non-leaf heap objects are marked. Since
698 // allocations are blocked until assists can
699 // happen, we want enable assists as early as
703 gcBgMarkPrepare() // Must happen before assist enable.
706 // Mark all active tinyalloc blocks. Since we're
707 // allocating from these, they need to be black like
708 // other allocations. The alternative is to blacken
709 // the tiny block on every allocation from it, which
710 // would slow down the tiny allocator.
713 // At this point all Ps have enabled the write
714 // barrier, thus maintaining the no white to
715 // black invariant. Enable mutator assists to
716 // put back-pressure on fast allocating
718 atomic.Store(&gcBlackenEnabled, 1)
720 // In STW mode, we could block the instant systemstack
721 // returns, so make sure we're not preemptible.
726 now = startTheWorldWithSema()
727 work.pauseNS += now - work.pauseStart
729 memstats.gcPauseDist.record(now - work.pauseStart)
731 // Release the CPU limiter.
732 gcCPULimiter.finishGCTransition(now)
735 // Release the world sema before Gosched() in STW mode
736 // because we will need to reacquire it later but before
737 // this goroutine becomes runnable again, and we could
738 // self-deadlock otherwise.
739 semrelease(&worldsema)
742 // Make sure we block instead of returning to user code
744 if mode != gcBackgroundMode {
748 semrelease(&work.startSema)
751 // gcMarkDoneFlushed counts the number of P's with flushed work.
753 // Ideally this would be a captured local in gcMarkDone, but forEachP
754 // escapes its callback closure, so it can't capture anything.
756 // This is protected by markDoneSema.
757 var gcMarkDoneFlushed uint32
759 // gcMarkDone transitions the GC from mark to mark termination if all
760 // reachable objects have been marked (that is, there are no grey
761 // objects and can be no more in the future). Otherwise, it flushes
762 // all local work to the global queues where it can be discovered by
765 // This should be called when all local mark work has been drained and
766 // there are no remaining workers. Specifically, when
768 // work.nwait == work.nproc && !gcMarkWorkAvailable(p)
770 // The calling context must be preemptible.
772 // Flushing local work is important because idle Ps may have local
773 // work queued. This is the only way to make that work visible and
774 // drive GC to completion.
776 // It is explicitly okay to have write barriers in this function. If
777 // it does transition to mark termination, then all reachable objects
778 // have been marked, so the write barrier cannot shade any more
781 // Ensure only one thread is running the ragged barrier at a
783 semacquire(&work.markDoneSema)
786 // Re-check transition condition under transition lock.
788 // It's critical that this checks the global work queues are
789 // empty before performing the ragged barrier. Otherwise,
790 // there could be global work that a P could take after the P
791 // has passed the ragged barrier.
792 if !(gcphase == _GCmark && work.nwait == work.nproc && !gcMarkWorkAvailable(nil)) {
793 semrelease(&work.markDoneSema)
797 // forEachP needs worldsema to execute, and we'll need it to
798 // stop the world later, so acquire worldsema now.
799 semacquire(&worldsema)
801 // Flush all local buffers and collect flushedWork flags.
802 gcMarkDoneFlushed = 0
805 // Mark the user stack as preemptible so that it may be scanned.
806 // Otherwise, our attempt to force all P's to a safepoint could
807 // result in a deadlock as we attempt to preempt a worker that's
808 // trying to preempt us (e.g. for a stack scan).
809 casGToWaiting(gp, _Grunning, waitReasonGCMarkTermination)
810 forEachP(func(pp *p) {
811 // Flush the write barrier buffer, since this may add
812 // work to the gcWork.
815 // Flush the gcWork, since this may create global work
816 // and set the flushedWork flag.
818 // TODO(austin): Break up these workbufs to
819 // better distribute work.
821 // Collect the flushedWork flag.
822 if pp.gcw.flushedWork {
823 atomic.Xadd(&gcMarkDoneFlushed, 1)
824 pp.gcw.flushedWork = false
827 casgstatus(gp, _Gwaiting, _Grunning)
830 if gcMarkDoneFlushed != 0 {
831 // More grey objects were discovered since the
832 // previous termination check, so there may be more
833 // work to do. Keep going. It's possible the
834 // transition condition became true again during the
835 // ragged barrier, so re-check it.
836 semrelease(&worldsema)
840 // There was no global work, no local work, and no Ps
841 // communicated work since we took markDoneSema. Therefore
842 // there are no grey objects and no more objects can be
843 // shaded. Transition to mark termination.
846 work.pauseStart = now
847 getg().m.preemptoff = "gcing"
848 systemstack(func() { stopTheWorldWithSema(stwGCMarkTerm) })
849 // The gcphase is _GCmark, it will transition to _GCmarktermination
850 // below. The important thing is that the wb remains active until
851 // all marking is complete. This includes writes made by the GC.
853 // There is sometimes work left over when we enter mark termination due
854 // to write barriers performed after the completion barrier above.
855 // Detect this and resume concurrent mark. This is obviously
858 // See issue #27993 for details.
860 // Switch to the system stack to call wbBufFlush1, though in this case
861 // it doesn't matter because we're non-preemptible anyway.
864 for _, p := range allp {
873 getg().m.preemptoff = ""
875 now := startTheWorldWithSema()
876 work.pauseNS += now - work.pauseStart
877 memstats.gcPauseDist.record(now - work.pauseStart)
879 semrelease(&worldsema)
883 gcComputeStartingStackSize()
885 // Disable assists and background workers. We must do
886 // this before waking blocked assists.
887 atomic.Store(&gcBlackenEnabled, 0)
889 // Notify the CPU limiter that GC assists will now cease.
890 gcCPULimiter.startGCTransition(false, now)
892 // Wake all blocked assists. These will run when we
893 // start the world again.
896 // Likewise, release the transition lock. Blocked
897 // workers and assists will run when we start the
899 semrelease(&work.markDoneSema)
901 // In STW mode, re-enable user goroutines. These will be
902 // queued to run after we start the world.
903 schedEnableUser(true)
905 // endCycle depends on all gcWork cache stats being flushed.
906 // The termination algorithm above ensured that up to
907 // allocations since the ragged barrier.
908 gcController.endCycle(now, int(gomaxprocs), work.userForced)
910 // Perform mark termination. This will restart the world.
914 // World must be stopped and mark assists and background workers must be
916 func gcMarkTermination() {
917 // Start marktermination (write barrier remains enabled for now).
918 setGCPhase(_GCmarktermination)
920 work.heap1 = gcController.heapLive.Load()
921 startTime := nanotime()
924 mp.preemptoff = "gcing"
927 casGToWaiting(curgp, _Grunning, waitReasonGarbageCollection)
929 // Run gc on the g0 stack. We do this so that the g stack
930 // we're currently running on will no longer change. Cuts
931 // the root set down a bit (g0 stacks are not scanned, and
932 // we don't need to scan gc's internal state). We also
933 // need to switch to g0 so we can shrink the stack.
936 // Must return immediately.
937 // The outer function's stack may have moved
938 // during gcMark (it shrinks stacks, including the
939 // outer function's stack), so we must not refer
940 // to any of its variables. Return back to the
941 // non-system stack to pick up the new addresses
942 // before continuing.
946 work.heap2 = work.bytesMarked
947 if debug.gccheckmark > 0 {
948 // Run a full non-parallel, stop-the-world
949 // mark using checkmark bits, to check that we
950 // didn't forget to mark anything during the
951 // concurrent mark process.
954 gcw := &getg().m.p.ptr().gcw
956 wbBufFlush1(getg().m.p.ptr())
961 // marking is complete so we can turn the write barrier off
967 casgstatus(curgp, _Gwaiting, _Grunning)
976 if gcphase != _GCoff {
977 throw("gc done but gcphase != _GCoff")
980 // Record heapInUse for scavenger.
981 memstats.lastHeapInUse = gcController.heapInUse.load()
983 // Update GC trigger and pacing, as well as downstream consumers
984 // of this pacing information, for the next cycle.
985 systemstack(gcControllerCommit)
987 // Update timing memstats
989 sec, nsec, _ := time_now()
990 unixNow := sec*1e9 + int64(nsec)
991 work.pauseNS += now - work.pauseStart
993 memstats.gcPauseDist.record(now - work.pauseStart)
994 atomic.Store64(&memstats.last_gc_unix, uint64(unixNow)) // must be Unix time to make sense to user
995 atomic.Store64(&memstats.last_gc_nanotime, uint64(now)) // monotonic time for us
996 memstats.pause_ns[memstats.numgc%uint32(len(memstats.pause_ns))] = uint64(work.pauseNS)
997 memstats.pause_end[memstats.numgc%uint32(len(memstats.pause_end))] = uint64(unixNow)
998 memstats.pause_total_ns += uint64(work.pauseNS)
1000 sweepTermCpu := int64(work.stwprocs) * (work.tMark - work.tSweepTerm)
1001 // We report idle marking time below, but omit it from the
1002 // overall utilization here since it's "free".
1003 markAssistCpu := gcController.assistTime.Load()
1004 markDedicatedCpu := gcController.dedicatedMarkTime.Load()
1005 markFractionalCpu := gcController.fractionalMarkTime.Load()
1006 markIdleCpu := gcController.idleMarkTime.Load()
1007 markTermCpu := int64(work.stwprocs) * (work.tEnd - work.tMarkTerm)
1008 scavAssistCpu := scavenge.assistTime.Load()
1009 scavBgCpu := scavenge.backgroundTime.Load()
1011 // Update cumulative GC CPU stats.
1012 work.cpuStats.gcAssistTime += markAssistCpu
1013 work.cpuStats.gcDedicatedTime += markDedicatedCpu + markFractionalCpu
1014 work.cpuStats.gcIdleTime += markIdleCpu
1015 work.cpuStats.gcPauseTime += sweepTermCpu + markTermCpu
1016 work.cpuStats.gcTotalTime += sweepTermCpu + markAssistCpu + markDedicatedCpu + markFractionalCpu + markIdleCpu + markTermCpu
1018 // Update cumulative scavenge CPU stats.
1019 work.cpuStats.scavengeAssistTime += scavAssistCpu
1020 work.cpuStats.scavengeBgTime += scavBgCpu
1021 work.cpuStats.scavengeTotalTime += scavAssistCpu + scavBgCpu
1023 // Update total CPU.
1024 work.cpuStats.totalTime = sched.totaltime + (now-sched.procresizetime)*int64(gomaxprocs)
1025 work.cpuStats.idleTime += sched.idleTime.Load()
1027 // Compute userTime. We compute this indirectly as everything that's not the above.
1029 // Since time spent in _Pgcstop is covered by gcPauseTime, and time spent in _Pidle
1030 // is covered by idleTime, what we're left with is time spent in _Prunning and _Psyscall,
1031 // the latter of which is fine because the P will either go idle or get used for something
1032 // else via sysmon. Meanwhile if we subtract GC time from whatever's left, we get non-GC
1033 // _Prunning time. Note that this still leaves time spent in sweeping and in the scheduler,
1034 // but that's fine. The overwhelming majority of this time will be actual user time.
1035 work.cpuStats.userTime = work.cpuStats.totalTime - (work.cpuStats.gcTotalTime +
1036 work.cpuStats.scavengeTotalTime + work.cpuStats.idleTime)
1038 // Compute overall GC CPU utilization.
1039 // Omit idle marking time from the overall utilization here since it's "free".
1040 memstats.gc_cpu_fraction = float64(work.cpuStats.gcTotalTime-work.cpuStats.gcIdleTime) / float64(work.cpuStats.totalTime)
1042 // Reset assist time and background time stats.
1044 // Do this now, instead of at the start of the next GC cycle, because
1045 // these two may keep accumulating even if the GC is not active.
1046 scavenge.assistTime.Store(0)
1047 scavenge.backgroundTime.Store(0)
1049 // Reset idle time stat.
1050 sched.idleTime.Store(0)
1052 // Reset sweep state.
1054 sweep.npausesweep = 0
1056 if work.userForced {
1057 memstats.numforcedgc++
1060 // Bump GC cycle count and wake goroutines waiting on sweep.
1061 lock(&work.sweepWaiters.lock)
1063 injectglist(&work.sweepWaiters.list)
1064 unlock(&work.sweepWaiters.lock)
1066 // Increment the scavenge generation now.
1068 // This moment represents peak heap in use because we're
1069 // about to start sweeping.
1070 mheap_.pages.scav.index.nextGen()
1072 // Release the CPU limiter.
1073 gcCPULimiter.finishGCTransition(now)
1075 // Finish the current heap profiling cycle and start a new
1076 // heap profiling cycle. We do this before starting the world
1077 // so events don't leak into the wrong cycle.
1080 // There may be stale spans in mcaches that need to be swept.
1081 // Those aren't tracked in any sweep lists, so we need to
1082 // count them against sweep completion until we ensure all
1083 // those spans have been forced out.
1084 sl := sweep.active.begin()
1086 throw("failed to set sweep barrier")
1089 systemstack(func() { startTheWorldWithSema() })
1091 // Flush the heap profile so we can start a new cycle next GC.
1092 // This is relatively expensive, so we don't do it with the
1096 // Prepare workbufs for freeing by the sweeper. We do this
1097 // asynchronously because it can take non-trivial time.
1098 prepareFreeWorkbufs()
1100 // Free stack spans. This must be done between GC cycles.
1101 systemstack(freeStackSpans)
1103 // Ensure all mcaches are flushed. Each P will flush its own
1104 // mcache before allocating, but idle Ps may not. Since this
1105 // is necessary to sweep all spans, we need to ensure all
1106 // mcaches are flushed before we start the next GC cycle.
1108 // While we're here, flush the page cache for idle Ps to avoid
1109 // having pages get stuck on them. These pages are hidden from
1110 // the scavenger, so in small idle heaps a significant amount
1111 // of additional memory might be held onto.
1112 systemstack(func() {
1113 forEachP(func(pp *p) {
1114 pp.mcache.prepareForSweep()
1115 if pp.status == _Pidle {
1116 systemstack(func() {
1118 pp.pcache.flush(&mheap_.pages)
1119 unlock(&mheap_.lock)
1124 // Now that we've swept stale spans in mcaches, they don't
1125 // count against unswept spans.
1126 sweep.active.end(sl)
1128 // Print gctrace before dropping worldsema. As soon as we drop
1129 // worldsema another cycle could start and smash the stats
1130 // we're trying to print.
1131 if debug.gctrace > 0 {
1132 util := int(memstats.gc_cpu_fraction * 100)
1136 print("gc ", memstats.numgc,
1137 " @", string(itoaDiv(sbuf[:], uint64(work.tSweepTerm-runtimeInitTime)/1e6, 3)), "s ",
1139 prev := work.tSweepTerm
1140 for i, ns := range []int64{work.tMark, work.tMarkTerm, work.tEnd} {
1144 print(string(fmtNSAsMS(sbuf[:], uint64(ns-prev))))
1147 print(" ms clock, ")
1148 for i, ns := range []int64{
1150 gcController.assistTime.Load(),
1151 gcController.dedicatedMarkTime.Load() + gcController.fractionalMarkTime.Load(),
1152 gcController.idleMarkTime.Load(),
1155 if i == 2 || i == 3 {
1156 // Separate mark time components with /.
1161 print(string(fmtNSAsMS(sbuf[:], uint64(ns))))
1164 work.heap0>>20, "->", work.heap1>>20, "->", work.heap2>>20, " MB, ",
1165 gcController.lastHeapGoal>>20, " MB goal, ",
1166 gcController.lastStackScan.Load()>>20, " MB stacks, ",
1167 gcController.globalsScan.Load()>>20, " MB globals, ",
1168 work.maxprocs, " P")
1169 if work.userForced {
1176 // Set any arena chunks that were deferred to fault.
1177 lock(&userArenaState.lock)
1178 faultList := userArenaState.fault
1179 userArenaState.fault = nil
1180 unlock(&userArenaState.lock)
1181 for _, lc := range faultList {
1182 lc.mspan.setUserArenaChunkToFault()
1185 // Enable huge pages on some metadata if we cross a heap threshold.
1186 if gcController.heapGoal() > minHeapForMetadataHugePages {
1187 mheap_.enableMetadataHugePages()
1190 semrelease(&worldsema)
1192 // Careful: another GC cycle may start now.
1197 // now that gc is done, kick off finalizer thread if needed
1198 if !concurrentSweep {
1199 // give the queued finalizers, if any, a chance to run
1204 // gcBgMarkStartWorkers prepares background mark worker goroutines. These
1205 // goroutines will not run until the mark phase, but they must be started while
1206 // the work is not stopped and from a regular G stack. The caller must hold
1208 func gcBgMarkStartWorkers() {
1209 // Background marking is performed by per-P G's. Ensure that each P has
1210 // a background GC G.
1212 // Worker Gs don't exit if gomaxprocs is reduced. If it is raised
1213 // again, we can reuse the old workers; no need to create new workers.
1214 for gcBgMarkWorkerCount < gomaxprocs {
1217 notetsleepg(&work.bgMarkReady, -1)
1218 noteclear(&work.bgMarkReady)
1219 // The worker is now guaranteed to be added to the pool before
1220 // its P's next findRunnableGCWorker.
1222 gcBgMarkWorkerCount++
1226 // gcBgMarkPrepare sets up state for background marking.
1227 // Mutator assists must not yet be enabled.
1228 func gcBgMarkPrepare() {
1229 // Background marking will stop when the work queues are empty
1230 // and there are no more workers (note that, since this is
1231 // concurrent, this may be a transient state, but mark
1232 // termination will clean it up). Between background workers
1233 // and assists, we don't really know how many workers there
1234 // will be, so we pretend to have an arbitrarily large number
1235 // of workers, almost all of which are "waiting". While a
1236 // worker is working it decrements nwait. If nproc == nwait,
1237 // there are no workers.
1238 work.nproc = ^uint32(0)
1239 work.nwait = ^uint32(0)
1242 // gcBgMarkWorkerNode is an entry in the gcBgMarkWorkerPool. It points to a single
1243 // gcBgMarkWorker goroutine.
1244 type gcBgMarkWorkerNode struct {
1245 // Unused workers are managed in a lock-free stack. This field must be first.
1248 // The g of this worker.
1251 // Release this m on park. This is used to communicate with the unlock
1252 // function, which cannot access the G's stack. It is unused outside of
1253 // gcBgMarkWorker().
1257 func gcBgMarkWorker() {
1260 // We pass node to a gopark unlock function, so it can't be on
1261 // the stack (see gopark). Prevent deadlock from recursively
1262 // starting GC by disabling preemption.
1263 gp.m.preemptoff = "GC worker init"
1264 node := new(gcBgMarkWorkerNode)
1265 gp.m.preemptoff = ""
1269 node.m.set(acquirem())
1270 notewakeup(&work.bgMarkReady)
1271 // After this point, the background mark worker is generally scheduled
1272 // cooperatively by gcController.findRunnableGCWorker. While performing
1273 // work on the P, preemption is disabled because we are working on
1274 // P-local work buffers. When the preempt flag is set, this puts itself
1275 // into _Gwaiting to be woken up by gcController.findRunnableGCWorker
1276 // at the appropriate time.
1278 // When preemption is enabled (e.g., while in gcMarkDone), this worker
1279 // may be preempted and schedule as a _Grunnable G from a runq. That is
1280 // fine; it will eventually gopark again for further scheduling via
1281 // findRunnableGCWorker.
1283 // Since we disable preemption before notifying bgMarkReady, we
1284 // guarantee that this G will be in the worker pool for the next
1285 // findRunnableGCWorker. This isn't strictly necessary, but it reduces
1286 // latency between _GCmark starting and the workers starting.
1289 // Go to sleep until woken by
1290 // gcController.findRunnableGCWorker.
1291 gopark(func(g *g, nodep unsafe.Pointer) bool {
1292 node := (*gcBgMarkWorkerNode)(nodep)
1294 if mp := node.m.ptr(); mp != nil {
1295 // The worker G is no longer running; release
1298 // N.B. it is _safe_ to release the M as soon
1299 // as we are no longer performing P-local mark
1302 // However, since we cooperatively stop work
1303 // when gp.preempt is set, if we releasem in
1304 // the loop then the following call to gopark
1305 // would immediately preempt the G. This is
1306 // also safe, but inefficient: the G must
1307 // schedule again only to enter gopark and park
1308 // again. Thus, we defer the release until
1309 // after parking the G.
1313 // Release this G to the pool.
1314 gcBgMarkWorkerPool.push(&node.node)
1315 // Note that at this point, the G may immediately be
1316 // rescheduled and may be running.
1318 }, unsafe.Pointer(node), waitReasonGCWorkerIdle, traceEvGoBlock, 0)
1320 // Preemption must not occur here, or another G might see
1321 // p.gcMarkWorkerMode.
1323 // Disable preemption so we can use the gcw. If the
1324 // scheduler wants to preempt us, we'll stop draining,
1325 // dispose the gcw, and then preempt.
1326 node.m.set(acquirem())
1327 pp := gp.m.p.ptr() // P can't change with preemption disabled.
1329 if gcBlackenEnabled == 0 {
1330 println("worker mode", pp.gcMarkWorkerMode)
1331 throw("gcBgMarkWorker: blackening not enabled")
1334 if pp.gcMarkWorkerMode == gcMarkWorkerNotWorker {
1335 throw("gcBgMarkWorker: mode not set")
1338 startTime := nanotime()
1339 pp.gcMarkWorkerStartTime = startTime
1340 var trackLimiterEvent bool
1341 if pp.gcMarkWorkerMode == gcMarkWorkerIdleMode {
1342 trackLimiterEvent = pp.limiterEvent.start(limiterEventIdleMarkWork, startTime)
1345 decnwait := atomic.Xadd(&work.nwait, -1)
1346 if decnwait == work.nproc {
1347 println("runtime: work.nwait=", decnwait, "work.nproc=", work.nproc)
1348 throw("work.nwait was > work.nproc")
1351 systemstack(func() {
1352 // Mark our goroutine preemptible so its stack
1353 // can be scanned. This lets two mark workers
1354 // scan each other (otherwise, they would
1355 // deadlock). We must not modify anything on
1356 // the G stack. However, stack shrinking is
1357 // disabled for mark workers, so it is safe to
1358 // read from the G stack.
1359 casGToWaiting(gp, _Grunning, waitReasonGCWorkerActive)
1360 switch pp.gcMarkWorkerMode {
1362 throw("gcBgMarkWorker: unexpected gcMarkWorkerMode")
1363 case gcMarkWorkerDedicatedMode:
1364 gcDrain(&pp.gcw, gcDrainUntilPreempt|gcDrainFlushBgCredit)
1366 // We were preempted. This is
1367 // a useful signal to kick
1368 // everything out of the run
1369 // queue so it can run
1371 if drainQ, n := runqdrain(pp); n > 0 {
1373 globrunqputbatch(&drainQ, int32(n))
1377 // Go back to draining, this time
1378 // without preemption.
1379 gcDrain(&pp.gcw, gcDrainFlushBgCredit)
1380 case gcMarkWorkerFractionalMode:
1381 gcDrain(&pp.gcw, gcDrainFractional|gcDrainUntilPreempt|gcDrainFlushBgCredit)
1382 case gcMarkWorkerIdleMode:
1383 gcDrain(&pp.gcw, gcDrainIdle|gcDrainUntilPreempt|gcDrainFlushBgCredit)
1385 casgstatus(gp, _Gwaiting, _Grunning)
1388 // Account for time and mark us as stopped.
1390 duration := now - startTime
1391 gcController.markWorkerStop(pp.gcMarkWorkerMode, duration)
1392 if trackLimiterEvent {
1393 pp.limiterEvent.stop(limiterEventIdleMarkWork, now)
1395 if pp.gcMarkWorkerMode == gcMarkWorkerFractionalMode {
1396 atomic.Xaddint64(&pp.gcFractionalMarkTime, duration)
1399 // Was this the last worker and did we run out
1401 incnwait := atomic.Xadd(&work.nwait, +1)
1402 if incnwait > work.nproc {
1403 println("runtime: p.gcMarkWorkerMode=", pp.gcMarkWorkerMode,
1404 "work.nwait=", incnwait, "work.nproc=", work.nproc)
1405 throw("work.nwait > work.nproc")
1408 // We'll releasem after this point and thus this P may run
1409 // something else. We must clear the worker mode to avoid
1410 // attributing the mode to a different (non-worker) G in
1412 pp.gcMarkWorkerMode = gcMarkWorkerNotWorker
1414 // If this worker reached a background mark completion
1415 // point, signal the main GC goroutine.
1416 if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
1417 // We don't need the P-local buffers here, allow
1418 // preemption because we may schedule like a regular
1419 // goroutine in gcMarkDone (block on locks, etc).
1420 releasem(node.m.ptr())
1428 // gcMarkWorkAvailable reports whether executing a mark worker
1429 // on p is potentially useful. p may be nil, in which case it only
1430 // checks the global sources of work.
1431 func gcMarkWorkAvailable(p *p) bool {
1432 if p != nil && !p.gcw.empty() {
1435 if !work.full.empty() {
1436 return true // global work available
1438 if work.markrootNext < work.markrootJobs {
1439 return true // root scan work available
1444 // gcMark runs the mark (or, for concurrent GC, mark termination)
1445 // All gcWork caches must be empty.
1446 // STW is in effect at this point.
1447 func gcMark(startTime int64) {
1448 if debug.allocfreetrace > 0 {
1452 if gcphase != _GCmarktermination {
1453 throw("in gcMark expecting to see gcphase as _GCmarktermination")
1455 work.tstart = startTime
1457 // Check that there's no marking work remaining.
1458 if work.full != 0 || work.markrootNext < work.markrootJobs {
1459 print("runtime: full=", hex(work.full), " next=", work.markrootNext, " jobs=", work.markrootJobs, " nDataRoots=", work.nDataRoots, " nBSSRoots=", work.nBSSRoots, " nSpanRoots=", work.nSpanRoots, " nStackRoots=", work.nStackRoots, "\n")
1460 panic("non-empty mark queue after concurrent mark")
1463 if debug.gccheckmark > 0 {
1464 // This is expensive when there's a large number of
1465 // Gs, so only do it if checkmark is also enabled.
1469 // Drop allg snapshot. allgs may have grown, in which case
1470 // this is the only reference to the old backing store and
1471 // there's no need to keep it around.
1472 work.stackRoots = nil
1474 // Clear out buffers and double-check that all gcWork caches
1475 // are empty. This should be ensured by gcMarkDone before we
1476 // enter mark termination.
1478 // TODO: We could clear out buffers just before mark if this
1479 // has a non-negligible impact on STW time.
1480 for _, p := range allp {
1481 // The write barrier may have buffered pointers since
1482 // the gcMarkDone barrier. However, since the barrier
1483 // ensured all reachable objects were marked, all of
1484 // these must be pointers to black objects. Hence we
1485 // can just discard the write barrier buffer.
1486 if debug.gccheckmark > 0 {
1487 // For debugging, flush the buffer and make
1488 // sure it really was all marked.
1497 print("runtime: P ", p.id, " flushedWork ", gcw.flushedWork)
1498 if gcw.wbuf1 == nil {
1499 print(" wbuf1=<nil>")
1501 print(" wbuf1.n=", gcw.wbuf1.nobj)
1503 if gcw.wbuf2 == nil {
1504 print(" wbuf2=<nil>")
1506 print(" wbuf2.n=", gcw.wbuf2.nobj)
1509 throw("P has cached GC work at end of mark termination")
1511 // There may still be cached empty buffers, which we
1512 // need to flush since we're going to free them. Also,
1513 // there may be non-zero stats because we allocated
1514 // black after the gcMarkDone barrier.
1518 // Flush scanAlloc from each mcache since we're about to modify
1519 // heapScan directly. If we were to flush this later, then scanAlloc
1520 // might have incorrect information.
1522 // Note that it's not important to retain this information; we know
1523 // exactly what heapScan is at this point via scanWork.
1524 for _, p := range allp {
1532 // Reset controller state.
1533 gcController.resetLive(work.bytesMarked)
1536 // gcSweep must be called on the system stack because it acquires the heap
1537 // lock. See mheap for details.
1539 // The world must be stopped.
1542 func gcSweep(mode gcMode) {
1543 assertWorldStopped()
1545 if gcphase != _GCoff {
1546 throw("gcSweep being done but phase is not GCoff")
1550 mheap_.sweepgen += 2
1551 sweep.active.reset()
1552 mheap_.pagesSwept.Store(0)
1553 mheap_.sweepArenas = mheap_.allArenas
1554 mheap_.reclaimIndex.Store(0)
1555 mheap_.reclaimCredit.Store(0)
1556 unlock(&mheap_.lock)
1558 sweep.centralIndex.clear()
1560 if !_ConcurrentSweep || mode == gcForceBlockMode {
1561 // Special case synchronous sweep.
1562 // Record that no proportional sweeping has to happen.
1564 mheap_.sweepPagesPerByte = 0
1565 unlock(&mheap_.lock)
1566 // Sweep all spans eagerly.
1567 for sweepone() != ^uintptr(0) {
1570 // Free workbufs eagerly.
1571 prepareFreeWorkbufs()
1572 for freeSomeWbufs(false) {
1574 // All "free" events for this mark/sweep cycle have
1575 // now happened, so we can make this profile cycle
1576 // available immediately.
1582 // Background sweep.
1585 sweep.parked = false
1586 ready(sweep.g, 0, true)
1591 // gcResetMarkState resets global state prior to marking (concurrent
1592 // or STW) and resets the stack scan state of all Gs.
1594 // This is safe to do without the world stopped because any Gs created
1595 // during or after this will start out in the reset state.
1597 // gcResetMarkState must be called on the system stack because it acquires
1598 // the heap lock. See mheap for details.
1601 func gcResetMarkState() {
1602 // This may be called during a concurrent phase, so lock to make sure
1603 // allgs doesn't change.
1604 forEachG(func(gp *g) {
1605 gp.gcscandone = false // set to true in gcphasework
1606 gp.gcAssistBytes = 0
1609 // Clear page marks. This is just 1MB per 64GB of heap, so the
1610 // time here is pretty trivial.
1612 arenas := mheap_.allArenas
1613 unlock(&mheap_.lock)
1614 for _, ai := range arenas {
1615 ha := mheap_.arenas[ai.l1()][ai.l2()]
1616 for i := range ha.pageMarks {
1621 work.bytesMarked = 0
1622 work.initialHeapLive = gcController.heapLive.Load()
1625 // Hooks for other packages
1627 var poolcleanup func()
1628 var boringCaches []unsafe.Pointer // for crypto/internal/boring
1630 //go:linkname sync_runtime_registerPoolCleanup sync.runtime_registerPoolCleanup
1631 func sync_runtime_registerPoolCleanup(f func()) {
1635 //go:linkname boring_registerCache crypto/internal/boring/bcache.registerCache
1636 func boring_registerCache(p unsafe.Pointer) {
1637 boringCaches = append(boringCaches, p)
1642 if poolcleanup != nil {
1646 // clear boringcrypto caches
1647 for _, p := range boringCaches {
1648 atomicstorep(p, nil)
1651 // Clear central sudog cache.
1652 // Leave per-P caches alone, they have strictly bounded size.
1653 // Disconnect cached list before dropping it on the floor,
1654 // so that a dangling ref to one entry does not pin all of them.
1655 lock(&sched.sudoglock)
1656 var sg, sgnext *sudog
1657 for sg = sched.sudogcache; sg != nil; sg = sgnext {
1661 sched.sudogcache = nil
1662 unlock(&sched.sudoglock)
1664 // Clear central defer pool.
1665 // Leave per-P pools alone, they have strictly bounded size.
1666 lock(&sched.deferlock)
1667 // disconnect cached list before dropping it on the floor,
1668 // so that a dangling ref to one entry does not pin all of them.
1669 var d, dlink *_defer
1670 for d = sched.deferpool; d != nil; d = dlink {
1674 sched.deferpool = nil
1675 unlock(&sched.deferlock)
1680 // itoaDiv formats val/(10**dec) into buf.
1681 func itoaDiv(buf []byte, val uint64, dec int) []byte {
1684 for val >= 10 || i >= idec {
1685 buf[i] = byte(val%10 + '0')
1693 buf[i] = byte(val + '0')
1697 // fmtNSAsMS nicely formats ns nanoseconds as milliseconds.
1698 func fmtNSAsMS(buf []byte, ns uint64) []byte {
1700 // Format as whole milliseconds.
1701 return itoaDiv(buf, ns/1e6, 0)
1703 // Format two digits of precision, with at most three decimal places.
1714 return itoaDiv(buf, x, dec)
1717 // Helpers for testing GC.
1719 // gcTestMoveStackOnNextCall causes the stack to be moved on a call
1720 // immediately following the call to this. It may not work correctly
1721 // if any other work appears after this call (such as returning).
1722 // Typically the following call should be marked go:noinline so it
1723 // performs a stack check.
1725 // In rare cases this may not cause the stack to move, specifically if
1726 // there's a preemption between this call and the next.
1727 func gcTestMoveStackOnNextCall() {
1729 gp.stackguard0 = stackForceMove
1732 // gcTestIsReachable performs a GC and returns a bit set where bit i
1733 // is set if ptrs[i] is reachable.
1734 func gcTestIsReachable(ptrs ...unsafe.Pointer) (mask uint64) {
1735 // This takes the pointers as unsafe.Pointers in order to keep
1736 // them live long enough for us to attach specials. After
1737 // that, we drop our references to them.
1740 panic("too many pointers for uint64 mask")
1743 // Block GC while we attach specials and drop our references
1744 // to ptrs. Otherwise, if a GC is in progress, it could mark
1745 // them reachable via this function before we have a chance to
1749 // Create reachability specials for ptrs.
1750 specials := make([]*specialReachable, len(ptrs))
1751 for i, p := range ptrs {
1752 lock(&mheap_.speciallock)
1753 s := (*specialReachable)(mheap_.specialReachableAlloc.alloc())
1754 unlock(&mheap_.speciallock)
1755 s.special.kind = _KindSpecialReachable
1756 if !addspecial(p, &s.special) {
1757 throw("already have a reachable special (duplicate pointer?)")
1760 // Make sure we don't retain ptrs.
1766 // Force a full GC and sweep.
1769 // Process specials.
1770 for i, s := range specials {
1773 println("runtime: object", i, "was not swept")
1774 throw("IsReachable failed")
1779 lock(&mheap_.speciallock)
1780 mheap_.specialReachableAlloc.free(unsafe.Pointer(s))
1781 unlock(&mheap_.speciallock)
1787 // gcTestPointerClass returns the category of what p points to, one of:
1788 // "heap", "stack", "data", "bss", "other". This is useful for checking
1789 // that a test is doing what it's intended to do.
1791 // This is nosplit simply to avoid extra pointer shuffling that may
1792 // complicate a test.
1795 func gcTestPointerClass(p unsafe.Pointer) string {
1796 p2 := uintptr(noescape(p))
1798 if gp.stack.lo <= p2 && p2 < gp.stack.hi {
1801 if base, _, _ := findObject(p2, 0, 0); base != 0 {
1804 for _, datap := range activeModules() {
1805 if datap.data <= p2 && p2 < datap.edata || datap.noptrdata <= p2 && p2 < datap.enoptrdata {
1808 if datap.bss <= p2 && p2 < datap.ebss || datap.noptrbss <= p2 && p2 <= datap.enoptrbss {