1 // Copyright 2021 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.
9 "runtime/internal/atomic"
14 // gcGoalUtilization is the goal CPU utilization for
15 // marking as a fraction of GOMAXPROCS.
16 gcGoalUtilization = 0.30
18 // gcBackgroundUtilization is the fixed CPU utilization for background
19 // marking. It must be <= gcGoalUtilization. The difference between
20 // gcGoalUtilization and gcBackgroundUtilization will be made up by
21 // mark assists. The scheduler will aim to use within 50% of this
24 // Setting this to < gcGoalUtilization avoids saturating the trigger
25 // feedback controller when there are no assists, which allows it to
26 // better control CPU and heap growth. However, the larger the gap,
27 // the more mutator assists are expected to happen, which impact
29 gcBackgroundUtilization = 0.25
31 // gcCreditSlack is the amount of scan work credit that can
32 // accumulate locally before updating gcController.scanWork and,
33 // optionally, gcController.bgScanCredit. Lower values give a more
34 // accurate assist ratio and make it more likely that assists will
35 // successfully steal background credit. Higher values reduce memory
39 // gcAssistTimeSlack is the nanoseconds of mutator assist time that
40 // can accumulate on a P before updating gcController.assistTime.
41 gcAssistTimeSlack = 5000
43 // gcOverAssistWork determines how many extra units of scan work a GC
44 // assist does when an assist happens. This amortizes the cost of an
45 // assist by pre-paying for this many bytes of future allocations.
46 gcOverAssistWork = 64 << 10
48 // defaultHeapMinimum is the value of heapMinimum for GOGC==100.
49 defaultHeapMinimum = 4 << 20
51 // scannableStackSizeSlack is the bytes of stack space allocated or freed
52 // that can accumulate on a P before updating gcController.stackSize.
53 scannableStackSizeSlack = 8 << 10
57 if offset := unsafe.Offsetof(gcController.heapLive); offset%8 != 0 {
59 throw("gcController.heapLive not aligned to 8 bytes")
63 // gcController implements the GC pacing controller that determines
64 // when to trigger concurrent garbage collection and how much marking
65 // work to do in mutator assists and background marking.
67 // It uses a feedback control algorithm to adjust the gcController.trigger
68 // trigger based on the heap growth and GC CPU utilization each cycle.
69 // This algorithm optimizes for heap growth to match GOGC and for CPU
70 // utilization between assist and background marking to be 25% of
71 // GOMAXPROCS. The high-level design of this algorithm is documented
72 // at https://golang.org/s/go15gcpacing.
74 // All fields of gcController are used only during a single mark
76 var gcController gcControllerState
78 type gcControllerState struct {
79 // Initialized from $GOGC. GOGC=off means no GC.
81 // Updated atomically with mheap_.lock held or during a STW.
82 // Safe to read atomically at any time, or non-atomically with
83 // mheap_.lock or STW.
86 _ uint32 // padding so following 64-bit values are 8-byte aligned
88 // heapMinimum is the minimum heap size at which to trigger GC.
89 // For small heaps, this overrides the usual GOGC*live set rule.
91 // When there is a very small live set but a lot of allocation, simply
92 // collecting when the heap reaches GOGC*live results in many GC
93 // cycles and high total per-GC overhead. This minimum amortizes this
94 // per-GC overhead while keeping the heap reasonably small.
96 // During initialization this is set to 4MB*GOGC/100. In the case of
97 // GOGC==0, this will set heapMinimum to 0, resulting in constant
98 // collection even when the heap size is small, which is useful for
102 // triggerRatio is the heap growth ratio that triggers marking.
104 // E.g., if this is 0.6, then GC should start when the live
105 // heap has reached 1.6 times the heap size marked by the
106 // previous cycle. This should be ≤ GOGC/100 so the trigger
107 // heap size is less than the goal heap size. This is set
108 // during mark termination for the next cycle's trigger.
110 // Protected by mheap_.lock or a STW.
113 // trigger is the heap size that triggers marking.
115 // When heapLive ≥ trigger, the mark phase will start.
116 // This is also the heap size by which proportional sweeping
119 // This is computed from triggerRatio during mark termination
120 // for the next cycle's trigger.
122 // Protected by mheap_.lock or a STW.
125 // heapGoal is the goal heapLive for when next GC ends.
126 // Set to ^uint64(0) if disabled.
128 // Read and written atomically, unless the world is stopped.
131 // lastHeapGoal is the value of heapGoal for the previous GC.
132 // Note that this is distinct from the last value heapGoal had,
133 // because it could change if e.g. gcPercent changes.
135 // Read and written with the world stopped or with mheap_.lock held.
138 // heapLive is the number of bytes considered live by the GC.
139 // That is: retained by the most recent GC plus allocated
140 // since then. heapLive ≤ memstats.heapAlloc, since heapAlloc includes
141 // unmarked objects that have not yet been swept (and hence goes up as we
142 // allocate and down as we sweep) while heapLive excludes these
143 // objects (and hence only goes up between GCs).
145 // This is updated atomically without locking. To reduce
146 // contention, this is updated only when obtaining a span from
147 // an mcentral and at this point it counts all of the
148 // unallocated slots in that span (which will be allocated
149 // before that mcache obtains another span from that
150 // mcentral). Hence, it slightly overestimates the "true" live
151 // heap size. It's better to overestimate than to
152 // underestimate because 1) this triggers the GC earlier than
153 // necessary rather than potentially too late and 2) this
154 // leads to a conservative GC rate rather than a GC rate that
155 // is potentially too low.
157 // Reads should likewise be atomic (or during STW).
159 // Whenever this is updated, call traceHeapAlloc() and
160 // this gcControllerState's revise() method.
163 // heapScan is the number of bytes of "scannable" heap. This
164 // is the live heap (as counted by heapLive), but omitting
165 // no-scan objects and no-scan tails of objects.
167 // Whenever this is updated, call this gcControllerState's
170 // Read and written atomically or with the world stopped.
173 // stackScan is a snapshot of scannableStackSize taken at each GC
174 // STW pause and is used in pacing decisions.
176 // Updated only while the world is stopped.
179 // scannableStackSize is the amount of allocated goroutine stack space in
180 // use by goroutines.
182 // Read and updated atomically.
183 scannableStackSize uint64
185 // globalsScan is the total amount of global variable space
186 // that is scannable.
188 // Read and updated atomically.
191 // heapMarked is the number of bytes marked by the previous
192 // GC. After mark termination, heapLive == heapMarked, but
193 // unlike heapLive, heapMarked does not change until the
194 // next mark termination.
197 // scanWork is the total scan work performed this cycle. This
198 // is updated atomically during the cycle. Updates occur in
199 // bounded batches, since it is both written and read
200 // throughout the cycle. At the end of the cycle, this is how
201 // much of the retained heap is scannable.
203 // Currently this is the bytes of heap scanned. For most uses,
204 // this is an opaque unit of work, but for estimation the
205 // definition is important.
208 // bgScanCredit is the scan work credit accumulated by the
209 // concurrent background scan. This credit is accumulated by
210 // the background scan and stolen by mutator assists. This is
211 // updated atomically. Updates occur in bounded batches, since
212 // it is both written and read throughout the cycle.
215 // assistTime is the nanoseconds spent in mutator assists
216 // during this cycle. This is updated atomically. Updates
217 // occur in bounded batches, since it is both written and read
218 // throughout the cycle.
221 // dedicatedMarkTime is the nanoseconds spent in dedicated
222 // mark workers during this cycle. This is updated atomically
223 // at the end of the concurrent mark phase.
224 dedicatedMarkTime int64
226 // fractionalMarkTime is the nanoseconds spent in the
227 // fractional mark worker during this cycle. This is updated
228 // atomically throughout the cycle and will be up-to-date if
229 // the fractional mark worker is not currently running.
230 fractionalMarkTime int64
232 // idleMarkTime is the nanoseconds spent in idle marking
233 // during this cycle. This is updated atomically throughout
237 // markStartTime is the absolute start time in nanoseconds
238 // that assists and background mark workers started.
241 // dedicatedMarkWorkersNeeded is the number of dedicated mark
242 // workers that need to be started. This is computed at the
243 // beginning of each cycle and decremented atomically as
244 // dedicated mark workers get started.
245 dedicatedMarkWorkersNeeded int64
247 // assistWorkPerByte is the ratio of scan work to allocated
248 // bytes that should be performed by mutator assists. This is
249 // computed at the beginning of each cycle and updated every
250 // time heapScan is updated.
251 assistWorkPerByte atomic.Float64
253 // assistBytesPerWork is 1/assistWorkPerByte.
255 // Note that because this is read and written independently
256 // from assistWorkPerByte users may notice a skew between
257 // the two values, and such a state should be safe.
258 assistBytesPerWork atomic.Float64
260 // fractionalUtilizationGoal is the fraction of wall clock
261 // time that should be spent in the fractional mark worker on
262 // each P that isn't running a dedicated worker.
264 // For example, if the utilization goal is 25% and there are
265 // no dedicated workers, this will be 0.25. If the goal is
266 // 25%, there is one dedicated worker, and GOMAXPROCS is 5,
267 // this will be 0.05 to make up the missing 5%.
269 // If this is zero, no fractional workers are needed.
270 fractionalUtilizationGoal float64
275 func (c *gcControllerState) init(gcPercent int32) {
276 c.heapMinimum = defaultHeapMinimum
278 // Set a reasonable initial GC trigger.
279 c.triggerRatio = 7 / 8.0
281 // Fake a heapMarked value so it looks like a trigger at
282 // heapMinimum is the appropriate growth from heapMarked.
283 // This will go into computing the initial GC goal.
284 c.heapMarked = uint64(float64(c.heapMinimum) / (1 + c.triggerRatio))
286 // This will also compute and set the GC trigger and goal.
287 c.setGCPercent(gcPercent)
290 // startCycle resets the GC controller's state and computes estimates
291 // for a new GC cycle. The caller must hold worldsema and the world
293 func (c *gcControllerState) startCycle(markStartTime int64) {
297 c.dedicatedMarkTime = 0
298 c.fractionalMarkTime = 0
300 c.markStartTime = markStartTime
301 c.stackScan = atomic.Load64(&c.scannableStackSize)
303 // Ensure that the heap goal is at least a little larger than
304 // the current live heap size. This may not be the case if GC
305 // start is delayed or if the allocation that pushed gcController.heapLive
306 // over trigger is large or if the trigger is really close to
307 // GOGC. Assist is proportional to this distance, so enforce a
308 // minimum distance, even if it means going over the GOGC goal
310 if c.heapGoal < c.heapLive+1024*1024 {
311 c.heapGoal = c.heapLive + 1024*1024
314 // Compute the background mark utilization goal. In general,
315 // this may not come out exactly. We round the number of
316 // dedicated workers so that the utilization is closest to
317 // 25%. For small GOMAXPROCS, this would introduce too much
318 // error, so we add fractional workers in that case.
319 totalUtilizationGoal := float64(gomaxprocs) * gcBackgroundUtilization
320 c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5)
321 utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
322 const maxUtilError = 0.3
323 if utilError < -maxUtilError || utilError > maxUtilError {
324 // Rounding put us more than 30% off our goal. With
325 // gcBackgroundUtilization of 25%, this happens for
326 // GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
327 // workers to compensate.
328 if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
329 // Too many dedicated workers.
330 c.dedicatedMarkWorkersNeeded--
332 c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(gomaxprocs)
334 c.fractionalUtilizationGoal = 0
337 // In STW mode, we just want dedicated workers.
338 if debug.gcstoptheworld > 0 {
339 c.dedicatedMarkWorkersNeeded = int64(gomaxprocs)
340 c.fractionalUtilizationGoal = 0
344 for _, p := range allp {
346 p.gcFractionalMarkTime = 0
349 // Compute initial values for controls that are updated
350 // throughout the cycle.
353 if debug.gcpacertrace > 0 {
354 assistRatio := c.assistWorkPerByte.Load()
355 print("pacer: assist ratio=", assistRatio,
356 " (scan ", gcController.heapScan>>20, " MB in ",
357 work.initialHeapLive>>20, "->",
358 c.heapGoal>>20, " MB)",
359 " workers=", c.dedicatedMarkWorkersNeeded,
360 "+", c.fractionalUtilizationGoal, "\n")
364 // revise updates the assist ratio during the GC cycle to account for
365 // improved estimates. This should be called whenever gcController.heapScan,
366 // gcController.heapLive, or gcController.heapGoal is updated. It is safe to
367 // call concurrently, but it may race with other calls to revise.
369 // The result of this race is that the two assist ratio values may not line
370 // up or may be stale. In practice this is OK because the assist ratio
371 // moves slowly throughout a GC cycle, and the assist ratio is a best-effort
372 // heuristic anyway. Furthermore, no part of the heuristic depends on
373 // the two assist ratio values being exact reciprocals of one another, since
374 // the two values are used to convert values from different sources.
376 // The worst case result of this raciness is that we may miss a larger shift
377 // in the ratio (say, if we decide to pace more aggressively against the
378 // hard heap goal) but even this "hard goal" is best-effort (see #40460).
379 // The dedicated GC should ensure we don't exceed the hard goal by too much
380 // in the rare case we do exceed it.
382 // It should only be called when gcBlackenEnabled != 0 (because this
383 // is when assists are enabled and the necessary statistics are
385 func (c *gcControllerState) revise() {
386 gcPercent := atomic.Loadint32(&c.gcPercent)
388 // If GC is disabled but we're running a forced GC,
389 // act like GOGC is huge for the below calculations.
392 live := atomic.Load64(&c.heapLive)
393 scan := atomic.Load64(&c.heapScan)
394 work := atomic.Loadint64(&c.scanWork)
396 // Assume we're under the soft goal. Pace GC to complete at
397 // heapGoal assuming the heap is in steady-state.
398 heapGoal := int64(atomic.Load64(&c.heapGoal))
400 // Compute the expected scan work remaining.
402 // This is estimated based on the expected
403 // steady-state scannable heap. For example, with
404 // GOGC=100, only half of the scannable heap is
405 // expected to be live, so that's what we target.
407 // (This is a float calculation to avoid overflowing on
409 scanWorkExpected := int64(float64(scan) * 100 / float64(100+gcPercent))
411 if int64(live) > heapGoal || work > scanWorkExpected {
412 // We're past the soft goal, or we've already done more scan
413 // work than we expected. Pace GC so that in the worst case it
414 // will complete by the hard goal.
415 const maxOvershoot = 1.1
416 heapGoal = int64(float64(heapGoal) * maxOvershoot)
418 // Compute the upper bound on the scan work remaining.
419 scanWorkExpected = int64(scan)
422 // Compute the remaining scan work estimate.
424 // Note that we currently count allocations during GC as both
425 // scannable heap (heapScan) and scan work completed
426 // (scanWork), so allocation will change this difference
427 // slowly in the soft regime and not at all in the hard
429 scanWorkRemaining := scanWorkExpected - work
430 if scanWorkRemaining < 1000 {
431 // We set a somewhat arbitrary lower bound on
432 // remaining scan work since if we aim a little high,
433 // we can miss by a little.
435 // We *do* need to enforce that this is at least 1,
436 // since marking is racy and double-scanning objects
437 // may legitimately make the remaining scan work
438 // negative, even in the hard goal regime.
439 scanWorkRemaining = 1000
442 // Compute the heap distance remaining.
443 heapRemaining := heapGoal - int64(live)
444 if heapRemaining <= 0 {
445 // This shouldn't happen, but if it does, avoid
446 // dividing by zero or setting the assist negative.
450 // Compute the mutator assist ratio so by the time the mutator
451 // allocates the remaining heap bytes up to heapGoal, it will
452 // have done (or stolen) the remaining amount of scan work.
453 // Note that the assist ratio values are updated atomically
454 // but not together. This means there may be some degree of
455 // skew between the two values. This is generally OK as the
456 // values shift relatively slowly over the course of a GC
458 assistWorkPerByte := float64(scanWorkRemaining) / float64(heapRemaining)
459 assistBytesPerWork := float64(heapRemaining) / float64(scanWorkRemaining)
460 c.assistWorkPerByte.Store(assistWorkPerByte)
461 c.assistBytesPerWork.Store(assistBytesPerWork)
464 // endCycle computes the trigger ratio for the next cycle.
465 // userForced indicates whether the current GC cycle was forced
466 // by the application.
467 func (c *gcControllerState) endCycle(userForced bool) float64 {
468 // Record last heap goal for the scavenger.
469 // We'll be updating the heap goal soon.
470 gcController.lastHeapGoal = gcController.heapGoal
473 // Forced GC means this cycle didn't start at the
474 // trigger, so where it finished isn't good
475 // information about how to adjust the trigger.
476 // Just leave it where it is.
477 return c.triggerRatio
480 // Proportional response gain for the trigger controller. Must
481 // be in [0, 1]. Lower values smooth out transient effects but
482 // take longer to respond to phase changes. Higher values
483 // react to phase changes quickly, but are more affected by
484 // transient changes. Values near 1 may be unstable.
485 const triggerGain = 0.5
487 // Compute next cycle trigger ratio. First, this computes the
488 // "error" for this cycle; that is, how far off the trigger
489 // was from what it should have been, accounting for both heap
490 // growth and GC CPU utilization. We compute the actual heap
491 // growth during this cycle and scale that by how far off from
492 // the goal CPU utilization we were (to estimate the heap
493 // growth if we had the desired CPU utilization). The
494 // difference between this estimate and the GOGC-based goal
495 // heap growth is the error.
496 goalGrowthRatio := c.effectiveGrowthRatio()
497 actualGrowthRatio := float64(c.heapLive)/float64(c.heapMarked) - 1
498 assistDuration := nanotime() - c.markStartTime
500 // Assume background mark hit its utilization goal.
501 utilization := gcBackgroundUtilization
502 // Add assist utilization; avoid divide by zero.
503 if assistDuration > 0 {
504 utilization += float64(c.assistTime) / float64(assistDuration*int64(gomaxprocs))
507 triggerError := goalGrowthRatio - c.triggerRatio - utilization/gcGoalUtilization*(actualGrowthRatio-c.triggerRatio)
509 // Finally, we adjust the trigger for next time by this error,
510 // damped by the proportional gain.
511 triggerRatio := c.triggerRatio + triggerGain*triggerError
513 if debug.gcpacertrace > 0 {
514 // Print controller state in terms of the design
516 H_m_prev := c.heapMarked
517 h_t := c.triggerRatio
519 h_a := actualGrowthRatio
521 h_g := goalGrowthRatio
522 H_g := int64(float64(H_m_prev) * (1 + h_g))
524 u_g := gcGoalUtilization
526 print("pacer: H_m_prev=", H_m_prev,
527 " h_t=", h_t, " H_T=", H_T,
528 " h_a=", h_a, " H_a=", H_a,
529 " h_g=", h_g, " H_g=", H_g,
530 " u_a=", u_a, " u_g=", u_g,
532 " goalΔ=", goalGrowthRatio-h_t,
533 " actualΔ=", h_a-h_t,
534 " u_a/u_g=", u_a/u_g,
541 // enlistWorker encourages another dedicated mark worker to start on
542 // another P if there are spare worker slots. It is used by putfull
543 // when more work is made available.
546 func (c *gcControllerState) enlistWorker() {
547 // If there are idle Ps, wake one so it will run an idle worker.
548 // NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
550 // if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
555 // There are no idle Ps. If we need more dedicated workers,
556 // try to preempt a running P so it will switch to a worker.
557 if c.dedicatedMarkWorkersNeeded <= 0 {
560 // Pick a random other P to preempt.
565 if gp == nil || gp.m == nil || gp.m.p == 0 {
568 myID := gp.m.p.ptr().id
569 for tries := 0; tries < 5; tries++ {
570 id := int32(fastrandn(uint32(gomaxprocs - 1)))
575 if p.status != _Prunning {
584 // findRunnableGCWorker returns a background mark worker for _p_ if it
585 // should be run. This must only be called when gcBlackenEnabled != 0.
586 func (c *gcControllerState) findRunnableGCWorker(_p_ *p) *g {
587 if gcBlackenEnabled == 0 {
588 throw("gcControllerState.findRunnable: blackening not enabled")
591 if !gcMarkWorkAvailable(_p_) {
592 // No work to be done right now. This can happen at
593 // the end of the mark phase when there are still
594 // assists tapering off. Don't bother running a worker
595 // now because it'll just return immediately.
599 // Grab a worker before we commit to running below.
600 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
602 // There is at least one worker per P, so normally there are
603 // enough workers to run on all Ps, if necessary. However, once
604 // a worker enters gcMarkDone it may park without rejoining the
605 // pool, thus freeing a P with no corresponding worker.
606 // gcMarkDone never depends on another worker doing work, so it
607 // is safe to simply do nothing here.
609 // If gcMarkDone bails out without completing the mark phase,
610 // it will always do so with queued global work. Thus, that P
611 // will be immediately eligible to re-run the worker G it was
612 // just using, ensuring work can complete.
616 decIfPositive := func(ptr *int64) bool {
618 v := atomic.Loadint64(ptr)
623 if atomic.Casint64(ptr, v, v-1) {
629 if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
630 // This P is now dedicated to marking until the end of
631 // the concurrent mark phase.
632 _p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
633 } else if c.fractionalUtilizationGoal == 0 {
634 // No need for fractional workers.
635 gcBgMarkWorkerPool.push(&node.node)
638 // Is this P behind on the fractional utilization
641 // This should be kept in sync with pollFractionalWorkerExit.
642 delta := nanotime() - c.markStartTime
643 if delta > 0 && float64(_p_.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
644 // Nope. No need to run a fractional worker.
645 gcBgMarkWorkerPool.push(&node.node)
648 // Run a fractional worker.
649 _p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode
652 // Run the background mark worker.
654 casgstatus(gp, _Gwaiting, _Grunnable)
661 // resetLive sets up the controller state for the next mark phase after the end
662 // of the previous one. Must be called after endCycle and before commit, before
663 // the world is started.
665 // The world must be stopped.
666 func (c *gcControllerState) resetLive(bytesMarked uint64) {
667 c.heapMarked = bytesMarked
668 c.heapLive = bytesMarked
669 c.heapScan = uint64(c.scanWork)
671 // heapLive was updated, so emit a trace event.
677 // logWorkTime updates mark work accounting in the controller by a duration of
678 // work in nanoseconds.
680 // Safe to execute at any time.
681 func (c *gcControllerState) logWorkTime(mode gcMarkWorkerMode, duration int64) {
683 case gcMarkWorkerDedicatedMode:
684 atomic.Xaddint64(&c.dedicatedMarkTime, duration)
685 atomic.Xaddint64(&c.dedicatedMarkWorkersNeeded, 1)
686 case gcMarkWorkerFractionalMode:
687 atomic.Xaddint64(&c.fractionalMarkTime, duration)
688 case gcMarkWorkerIdleMode:
689 atomic.Xaddint64(&c.idleMarkTime, duration)
691 throw("logWorkTime: unknown mark worker mode")
695 func (c *gcControllerState) update(dHeapLive, dHeapScan int64) {
697 atomic.Xadd64(&gcController.heapLive, dHeapLive)
699 // gcController.heapLive changed.
704 atomic.Xadd64(&gcController.heapScan, dHeapScan)
706 if gcBlackenEnabled != 0 {
707 // gcController.heapLive and heapScan changed.
712 func (c *gcControllerState) addScannableStack(pp *p, amount int64) {
714 atomic.Xadd64(&c.scannableStackSize, amount)
717 pp.scannableStackSizeDelta += amount
718 if pp.scannableStackSizeDelta >= scannableStackSizeSlack || pp.scannableStackSizeDelta <= -scannableStackSizeSlack {
719 atomic.Xadd64(&c.scannableStackSize, pp.scannableStackSizeDelta)
720 pp.scannableStackSizeDelta = 0
724 func (c *gcControllerState) addGlobals(amount int64) {
725 atomic.Xadd64(&c.globalsScan, amount)
728 // commit sets the trigger ratio and updates everything
729 // derived from it: the absolute trigger, the heap goal, mark pacing,
732 // This can be called any time. If GC is the in the middle of a
733 // concurrent phase, it will adjust the pacing of that phase.
735 // This depends on gcPercent, gcController.heapMarked, and
736 // gcController.heapLive. These must be up to date.
738 // mheap_.lock must be held or the world must be stopped.
739 func (c *gcControllerState) commit(triggerRatio float64) {
740 assertWorldStoppedOrLockHeld(&mheap_.lock)
742 // Compute the next GC goal, which is when the allocated heap
743 // has grown by GOGC/100 over the heap marked by the last
746 if c.gcPercent >= 0 {
747 goal = c.heapMarked + c.heapMarked*uint64(c.gcPercent)/100
750 // Set the trigger ratio, capped to reasonable bounds.
751 if c.gcPercent >= 0 {
752 scalingFactor := float64(c.gcPercent) / 100
753 // Ensure there's always a little margin so that the
754 // mutator assist ratio isn't infinity.
755 maxTriggerRatio := 0.95 * scalingFactor
756 if triggerRatio > maxTriggerRatio {
757 triggerRatio = maxTriggerRatio
760 // If we let triggerRatio go too low, then if the application
761 // is allocating very rapidly we might end up in a situation
762 // where we're allocating black during a nearly always-on GC.
763 // The result of this is a growing heap and ultimately an
764 // increase in RSS. By capping us at a point >0, we're essentially
765 // saying that we're OK using more CPU during the GC to prevent
766 // this growth in RSS.
768 // The current constant was chosen empirically: given a sufficiently
769 // fast/scalable allocator with 48 Ps that could drive the trigger ratio
770 // to <0.05, this constant causes applications to retain the same peak
771 // RSS compared to not having this allocator.
772 minTriggerRatio := 0.6 * scalingFactor
773 if triggerRatio < minTriggerRatio {
774 triggerRatio = minTriggerRatio
776 } else if triggerRatio < 0 {
777 // gcPercent < 0, so just make sure we're not getting a negative
778 // triggerRatio. This case isn't expected to happen in practice,
779 // and doesn't really matter because if gcPercent < 0 then we won't
780 // ever consume triggerRatio further on in this function, but let's
781 // just be defensive here; the triggerRatio being negative is almost
782 // certainly undesirable.
785 c.triggerRatio = triggerRatio
787 // Compute the absolute GC trigger from the trigger ratio.
789 // We trigger the next GC cycle when the allocated heap has
790 // grown by the trigger ratio over the marked heap size.
791 trigger := ^uint64(0)
792 if c.gcPercent >= 0 {
793 trigger = uint64(float64(c.heapMarked) * (1 + triggerRatio))
794 // Don't trigger below the minimum heap size.
795 minTrigger := c.heapMinimum
797 // Concurrent sweep happens in the heap growth
798 // from gcController.heapLive to trigger, so ensure
799 // that concurrent sweep has some heap growth
800 // in which to perform sweeping before we
801 // start the next GC cycle.
802 sweepMin := atomic.Load64(&c.heapLive) + sweepMinHeapDistance
803 if sweepMin > minTrigger {
804 minTrigger = sweepMin
807 if trigger < minTrigger {
810 if int64(trigger) < 0 {
811 print("runtime: heapGoal=", c.heapGoal, " heapMarked=", c.heapMarked, " gcController.heapLive=", c.heapLive, " initialHeapLive=", work.initialHeapLive, "triggerRatio=", triggerRatio, " minTrigger=", minTrigger, "\n")
812 throw("trigger underflow")
815 // The trigger ratio is always less than GOGC/100, but
816 // other bounds on the trigger may have raised it.
817 // Push up the goal, too.
822 // Commit to the trigger and goal.
824 atomic.Store64(&c.heapGoal, goal)
829 // Update mark pacing.
830 if gcphase != _GCoff {
835 // effectiveGrowthRatio returns the current effective heap growth
836 // ratio (GOGC/100) based on heapMarked from the previous GC and
837 // heapGoal for the current GC.
839 // This may differ from gcPercent/100 because of various upper and
840 // lower bounds on gcPercent. For example, if the heap is smaller than
841 // heapMinimum, this can be higher than gcPercent/100.
843 // mheap_.lock must be held or the world must be stopped.
844 func (c *gcControllerState) effectiveGrowthRatio() float64 {
845 assertWorldStoppedOrLockHeld(&mheap_.lock)
847 egogc := float64(atomic.Load64(&c.heapGoal)-c.heapMarked) / float64(c.heapMarked)
849 // Shouldn't happen, but just in case.
855 // setGCPercent updates gcPercent and all related pacer state.
856 // Returns the old value of gcPercent.
858 // Calls gcControllerState.commit.
860 // The world must be stopped, or mheap_.lock must be held.
861 func (c *gcControllerState) setGCPercent(in int32) int32 {
862 assertWorldStoppedOrLockHeld(&mheap_.lock)
868 // Write it atomically so readers like revise() can read it safely.
869 atomic.Storeint32(&c.gcPercent, in)
870 c.heapMinimum = defaultHeapMinimum * uint64(c.gcPercent) / 100
871 // Update pacing in response to gcPercent change.
872 c.commit(c.triggerRatio)
877 //go:linkname setGCPercent runtime/debug.setGCPercent
878 func setGCPercent(in int32) (out int32) {
879 // Run on the system stack since we grab the heap lock.
882 out = gcController.setGCPercent(in)
883 gcPaceSweeper(gcController.trigger)
884 gcPaceScavenger(gcController.heapGoal, gcController.lastHeapGoal)
888 // If we just disabled GC, wait for any concurrent GC mark to
889 // finish so we always return with no GC running.
891 gcWaitOnMark(atomic.Load(&work.cycles))
897 func readGOGC() int32 {
898 p := gogetenv("GOGC")
902 if n, ok := atoi32(p); ok {