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 "internal/goexperiment"
10 "runtime/internal/atomic"
15 // gcGoalUtilization is the goal CPU utilization for
16 // marking as a fraction of GOMAXPROCS.
18 // Increasing the goal utilization will shorten GC cycles as the GC
19 // has more resources behind it, lessening costs from the write barrier,
20 // but comes at the cost of increasing mutator latency.
21 gcGoalUtilization = gcBackgroundUtilization
23 // gcBackgroundUtilization is the fixed CPU utilization for background
24 // marking. It must be <= gcGoalUtilization. The difference between
25 // gcGoalUtilization and gcBackgroundUtilization will be made up by
26 // mark assists. The scheduler will aim to use within 50% of this
29 // As a general rule, there's little reason to set gcBackgroundUtilization
30 // < gcGoalUtilization. One reason might be in mostly idle applications,
31 // where goroutines are unlikely to assist at all, so the actual
32 // utilization will be lower than the goal. But this is moot point
33 // because the idle mark workers already soak up idle CPU resources.
34 // These two values are still kept separate however because they are
35 // distinct conceptually, and in previous iterations of the pacer the
36 // distinction was more important.
37 gcBackgroundUtilization = 0.25
39 // gcCreditSlack is the amount of scan work credit that can
40 // accumulate locally before updating gcController.heapScanWork and,
41 // optionally, gcController.bgScanCredit. Lower values give a more
42 // accurate assist ratio and make it more likely that assists will
43 // successfully steal background credit. Higher values reduce memory
47 // gcAssistTimeSlack is the nanoseconds of mutator assist time that
48 // can accumulate on a P before updating gcController.assistTime.
49 gcAssistTimeSlack = 5000
51 // gcOverAssistWork determines how many extra units of scan work a GC
52 // assist does when an assist happens. This amortizes the cost of an
53 // assist by pre-paying for this many bytes of future allocations.
54 gcOverAssistWork = 64 << 10
56 // defaultHeapMinimum is the value of heapMinimum for GOGC==100.
57 defaultHeapMinimum = (goexperiment.HeapMinimum512KiBInt)*(512<<10) +
58 (1-goexperiment.HeapMinimum512KiBInt)*(4<<20)
60 // scannableStackSizeSlack is the bytes of stack space allocated or freed
61 // that can accumulate on a P before updating gcController.stackSize.
62 scannableStackSizeSlack = 8 << 10
66 if offset := unsafe.Offsetof(gcController.heapLive); offset%8 != 0 {
68 throw("gcController.heapLive not aligned to 8 bytes")
72 // gcController implements the GC pacing controller that determines
73 // when to trigger concurrent garbage collection and how much marking
74 // work to do in mutator assists and background marking.
76 // It uses a feedback control algorithm to adjust the gcController.trigger
77 // trigger based on the heap growth and GC CPU utilization each cycle.
78 // This algorithm optimizes for heap growth to match GOGC and for CPU
79 // utilization between assist and background marking to be 25% of
80 // GOMAXPROCS. The high-level design of this algorithm is documented
81 // at https://golang.org/s/go15gcpacing.
83 // All fields of gcController are used only during a single mark
85 var gcController gcControllerState
87 type gcControllerState struct {
89 // Initialized from GOGC. GOGC=off means no GC.
90 gcPercent atomic.Int32
92 _ uint32 // padding so following 64-bit values are 8-byte aligned
94 // heapMinimum is the minimum heap size at which to trigger GC.
95 // For small heaps, this overrides the usual GOGC*live set rule.
97 // When there is a very small live set but a lot of allocation, simply
98 // collecting when the heap reaches GOGC*live results in many GC
99 // cycles and high total per-GC overhead. This minimum amortizes this
100 // per-GC overhead while keeping the heap reasonably small.
102 // During initialization this is set to 4MB*GOGC/100. In the case of
103 // GOGC==0, this will set heapMinimum to 0, resulting in constant
104 // collection even when the heap size is small, which is useful for
108 // trigger is the heap size that triggers marking.
110 // When heapLive ≥ trigger, the mark phase will start.
111 // This is also the heap size by which proportional sweeping
114 // This is computed from consMark during mark termination for
115 // the next cycle's trigger.
117 // Protected by mheap_.lock or a STW.
120 // consMark is the estimated per-CPU consMark ratio for the application.
122 // It represents the ratio between the application's allocation
123 // rate, as bytes allocated per CPU-time, and the GC's scan rate,
124 // as bytes scanned per CPU-time.
125 // The units of this ratio are (B / cpu-ns) / (B / cpu-ns).
127 // At a high level, this value is computed as the bytes of memory
128 // allocated (cons) per unit of scan work completed (mark) in a GC
129 // cycle, divided by the CPU time spent on each activity.
131 // Updated at the end of each GC cycle, in endCycle.
134 // consMarkController holds the state for the mark-cons ratio
135 // estimation over time.
137 // Its purpose is to smooth out noisiness in the computation of
138 // consMark; see consMark for details.
139 consMarkController piController
141 _ uint32 // Padding for atomics on 32-bit platforms.
143 // heapGoal is the goal heapLive for when next GC ends.
144 // Set to ^uint64(0) if disabled.
146 // Read and written atomically, unless the world is stopped.
149 // lastHeapGoal is the value of heapGoal for the previous GC.
150 // Note that this is distinct from the last value heapGoal had,
151 // because it could change if e.g. gcPercent changes.
153 // Read and written with the world stopped or with mheap_.lock held.
156 // heapLive is the number of bytes considered live by the GC.
157 // That is: retained by the most recent GC plus allocated
158 // since then. heapLive ≤ memstats.heapAlloc, since heapAlloc includes
159 // unmarked objects that have not yet been swept (and hence goes up as we
160 // allocate and down as we sweep) while heapLive excludes these
161 // objects (and hence only goes up between GCs).
163 // This is updated atomically without locking. To reduce
164 // contention, this is updated only when obtaining a span from
165 // an mcentral and at this point it counts all of the
166 // unallocated slots in that span (which will be allocated
167 // before that mcache obtains another span from that
168 // mcentral). Hence, it slightly overestimates the "true" live
169 // heap size. It's better to overestimate than to
170 // underestimate because 1) this triggers the GC earlier than
171 // necessary rather than potentially too late and 2) this
172 // leads to a conservative GC rate rather than a GC rate that
173 // is potentially too low.
175 // Reads should likewise be atomic (or during STW).
177 // Whenever this is updated, call traceHeapAlloc() and
178 // this gcControllerState's revise() method.
181 // heapScan is the number of bytes of "scannable" heap. This
182 // is the live heap (as counted by heapLive), but omitting
183 // no-scan objects and no-scan tails of objects.
185 // This value is fixed at the start of a GC cycle, so during a
186 // GC cycle it is safe to read without atomics, and it represents
187 // the maximum scannable heap.
190 // lastHeapScan is the number of bytes of heap that were scanned
191 // last GC cycle. It is the same as heapMarked, but only
192 // includes the "scannable" parts of objects.
194 // Updated when the world is stopped.
197 // stackScan is a snapshot of scannableStackSize taken at each GC
198 // STW pause and is used in pacing decisions.
200 // Updated only while the world is stopped.
203 // scannableStackSize is the amount of allocated goroutine stack space in
204 // use by goroutines.
206 // This number tracks allocated goroutine stack space rather than used
207 // goroutine stack space (i.e. what is actually scanned) because used
208 // goroutine stack space is much harder to measure cheaply. By using
209 // allocated space, we make an overestimate; this is OK, it's better
210 // to conservatively overcount than undercount.
212 // Read and updated atomically.
213 scannableStackSize uint64
215 // globalsScan is the total amount of global variable space
216 // that is scannable.
218 // Read and updated atomically.
221 // heapMarked is the number of bytes marked by the previous
222 // GC. After mark termination, heapLive == heapMarked, but
223 // unlike heapLive, heapMarked does not change until the
224 // next mark termination.
227 // heapScanWork is the total heap scan work performed this cycle.
228 // stackScanWork is the total stack scan work performed this cycle.
229 // globalsScanWork is the total globals scan work performed this cycle.
231 // These are updated atomically during the cycle. Updates occur in
232 // bounded batches, since they are both written and read
233 // throughout the cycle. At the end of the cycle, heapScanWork is how
234 // much of the retained heap is scannable.
236 // Currently these are measured in bytes. For most uses, this is an
237 // opaque unit of work, but for estimation the definition is important.
239 // Note that stackScanWork includes all allocated space, not just the
240 // size of the stack itself, mirroring stackSize.
241 heapScanWork atomic.Int64
242 stackScanWork atomic.Int64
243 globalsScanWork atomic.Int64
245 // bgScanCredit is the scan work credit accumulated by the
246 // concurrent background scan. This credit is accumulated by
247 // the background scan and stolen by mutator assists. This is
248 // updated atomically. Updates occur in bounded batches, since
249 // it is both written and read throughout the cycle.
252 // assistTime is the nanoseconds spent in mutator assists
253 // during this cycle. This is updated atomically. Updates
254 // occur in bounded batches, since it is both written and read
255 // throughout the cycle.
258 // dedicatedMarkTime is the nanoseconds spent in dedicated
259 // mark workers during this cycle. This is updated atomically
260 // at the end of the concurrent mark phase.
261 dedicatedMarkTime int64
263 // fractionalMarkTime is the nanoseconds spent in the
264 // fractional mark worker during this cycle. This is updated
265 // atomically throughout the cycle and will be up-to-date if
266 // the fractional mark worker is not currently running.
267 fractionalMarkTime int64
269 // idleMarkTime is the nanoseconds spent in idle marking
270 // during this cycle. This is updated atomically throughout
274 // markStartTime is the absolute start time in nanoseconds
275 // that assists and background mark workers started.
278 // dedicatedMarkWorkersNeeded is the number of dedicated mark
279 // workers that need to be started. This is computed at the
280 // beginning of each cycle and decremented atomically as
281 // dedicated mark workers get started.
282 dedicatedMarkWorkersNeeded int64
284 // assistWorkPerByte is the ratio of scan work to allocated
285 // bytes that should be performed by mutator assists. This is
286 // computed at the beginning of each cycle and updated every
287 // time heapScan is updated.
288 assistWorkPerByte atomic.Float64
290 // assistBytesPerWork is 1/assistWorkPerByte.
292 // Note that because this is read and written independently
293 // from assistWorkPerByte users may notice a skew between
294 // the two values, and such a state should be safe.
295 assistBytesPerWork atomic.Float64
297 // fractionalUtilizationGoal is the fraction of wall clock
298 // time that should be spent in the fractional mark worker on
299 // each P that isn't running a dedicated worker.
301 // For example, if the utilization goal is 25% and there are
302 // no dedicated workers, this will be 0.25. If the goal is
303 // 25%, there is one dedicated worker, and GOMAXPROCS is 5,
304 // this will be 0.05 to make up the missing 5%.
306 // If this is zero, no fractional workers are needed.
307 fractionalUtilizationGoal float64
309 // test indicates that this is a test-only copy of gcControllerState.
315 func (c *gcControllerState) init(gcPercent int32) {
316 c.heapMinimum = defaultHeapMinimum
318 c.consMarkController = piController{
319 // Tuned first via the Ziegler-Nichols process in simulation,
320 // then the integral time was manually tuned against real-world
321 // applications to deal with noisiness in the measured cons/mark
326 // Set a high reset time in GC cycles.
327 // This is inversely proportional to the rate at which we
328 // accumulate error from clipping. By making this very high
329 // we make the accumulation slow. In general, clipping is
330 // OK in our situation, hence the choice.
332 // Tune this if we get unintended effects from clipping for
339 // This will also compute and set the GC trigger and goal.
340 c.setGCPercent(gcPercent)
343 // startCycle resets the GC controller's state and computes estimates
344 // for a new GC cycle. The caller must hold worldsema and the world
346 func (c *gcControllerState) startCycle(markStartTime int64, procs int) {
347 c.heapScanWork.Store(0)
348 c.stackScanWork.Store(0)
349 c.globalsScanWork.Store(0)
352 c.dedicatedMarkTime = 0
353 c.fractionalMarkTime = 0
355 c.markStartTime = markStartTime
356 c.stackScan = atomic.Load64(&c.scannableStackSize)
358 // Ensure that the heap goal is at least a little larger than
359 // the current live heap size. This may not be the case if GC
360 // start is delayed or if the allocation that pushed gcController.heapLive
361 // over trigger is large or if the trigger is really close to
362 // GOGC. Assist is proportional to this distance, so enforce a
363 // minimum distance, even if it means going over the GOGC goal
365 if c.heapGoal < c.heapLive+64<<10 {
366 c.heapGoal = c.heapLive + 64<<10
369 // Compute the background mark utilization goal. In general,
370 // this may not come out exactly. We round the number of
371 // dedicated workers so that the utilization is closest to
372 // 25%. For small GOMAXPROCS, this would introduce too much
373 // error, so we add fractional workers in that case.
374 totalUtilizationGoal := float64(procs) * gcBackgroundUtilization
375 c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5)
376 utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
377 const maxUtilError = 0.3
378 if utilError < -maxUtilError || utilError > maxUtilError {
379 // Rounding put us more than 30% off our goal. With
380 // gcBackgroundUtilization of 25%, this happens for
381 // GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
382 // workers to compensate.
383 if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
384 // Too many dedicated workers.
385 c.dedicatedMarkWorkersNeeded--
387 c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(procs)
389 c.fractionalUtilizationGoal = 0
392 // In STW mode, we just want dedicated workers.
393 if debug.gcstoptheworld > 0 {
394 c.dedicatedMarkWorkersNeeded = int64(procs)
395 c.fractionalUtilizationGoal = 0
399 for _, p := range allp {
401 p.gcFractionalMarkTime = 0
404 // Compute initial values for controls that are updated
405 // throughout the cycle.
408 if debug.gcpacertrace > 0 {
409 assistRatio := c.assistWorkPerByte.Load()
410 print("pacer: assist ratio=", assistRatio,
411 " (scan ", gcController.heapScan>>20, " MB in ",
412 work.initialHeapLive>>20, "->",
413 c.heapGoal>>20, " MB)",
414 " workers=", c.dedicatedMarkWorkersNeeded,
415 "+", c.fractionalUtilizationGoal, "\n")
419 // revise updates the assist ratio during the GC cycle to account for
420 // improved estimates. This should be called whenever gcController.heapScan,
421 // gcController.heapLive, or gcController.heapGoal is updated. It is safe to
422 // call concurrently, but it may race with other calls to revise.
424 // The result of this race is that the two assist ratio values may not line
425 // up or may be stale. In practice this is OK because the assist ratio
426 // moves slowly throughout a GC cycle, and the assist ratio is a best-effort
427 // heuristic anyway. Furthermore, no part of the heuristic depends on
428 // the two assist ratio values being exact reciprocals of one another, since
429 // the two values are used to convert values from different sources.
431 // The worst case result of this raciness is that we may miss a larger shift
432 // in the ratio (say, if we decide to pace more aggressively against the
433 // hard heap goal) but even this "hard goal" is best-effort (see #40460).
434 // The dedicated GC should ensure we don't exceed the hard goal by too much
435 // in the rare case we do exceed it.
437 // It should only be called when gcBlackenEnabled != 0 (because this
438 // is when assists are enabled and the necessary statistics are
440 func (c *gcControllerState) revise() {
441 gcPercent := c.gcPercent.Load()
443 // If GC is disabled but we're running a forced GC,
444 // act like GOGC is huge for the below calculations.
447 live := atomic.Load64(&c.heapLive)
448 scan := atomic.Load64(&c.heapScan)
449 work := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
451 // Assume we're under the soft goal. Pace GC to complete at
452 // heapGoal assuming the heap is in steady-state.
453 heapGoal := int64(atomic.Load64(&c.heapGoal))
455 // The expected scan work is computed as the amount of bytes scanned last
456 // GC cycle, plus our estimate of stacks and globals work for this cycle.
457 scanWorkExpected := int64(c.lastHeapScan + c.stackScan + c.globalsScan)
459 // maxScanWork is a worst-case estimate of the amount of scan work that
460 // needs to be performed in this GC cycle. Specifically, it represents
461 // the case where *all* scannable memory turns out to be live.
462 maxScanWork := int64(scan + c.stackScan + c.globalsScan)
463 if work > scanWorkExpected {
464 // We've already done more scan work than expected. Because our expectation
465 // is based on a steady-state scannable heap size, we assume this means our
466 // heap is growing. Compute a new heap goal that takes our existing runway
467 // computed for scanWorkExpected and extrapolates it to maxScanWork, the worst-case
468 // scan work. This keeps our assist ratio stable if the heap continues to grow.
470 // The effect of this mechanism is that assists stay flat in the face of heap
471 // growths. It's OK to use more memory this cycle to scan all the live heap,
472 // because the next GC cycle is inevitably going to use *at least* that much
474 extHeapGoal := int64(float64(heapGoal-int64(c.trigger))/float64(scanWorkExpected)*float64(maxScanWork)) + int64(c.trigger)
475 scanWorkExpected = maxScanWork
477 // hardGoal is a hard limit on the amount that we're willing to push back the
478 // heap goal, and that's twice the heap goal (i.e. if GOGC=100 and the heap and/or
479 // stacks and/or globals grow to twice their size, this limits the current GC cycle's
480 // growth to 4x the original live heap's size).
482 // This maintains the invariant that we use no more memory than the next GC cycle
484 hardGoal := int64((1.0 + float64(gcPercent)/100.0) * float64(heapGoal))
485 if extHeapGoal > hardGoal {
486 extHeapGoal = hardGoal
488 heapGoal = extHeapGoal
490 if int64(live) > heapGoal {
491 // We're already past our heap goal, even the extrapolated one.
492 // Leave ourselves some extra runway, so in the worst case we
493 // finish by that point.
494 const maxOvershoot = 1.1
495 heapGoal = int64(float64(heapGoal) * maxOvershoot)
497 // Compute the upper bound on the scan work remaining.
498 scanWorkExpected = maxScanWork
501 // Compute the remaining scan work estimate.
503 // Note that we currently count allocations during GC as both
504 // scannable heap (heapScan) and scan work completed
505 // (scanWork), so allocation will change this difference
506 // slowly in the soft regime and not at all in the hard
508 scanWorkRemaining := scanWorkExpected - work
509 if scanWorkRemaining < 1000 {
510 // We set a somewhat arbitrary lower bound on
511 // remaining scan work since if we aim a little high,
512 // we can miss by a little.
514 // We *do* need to enforce that this is at least 1,
515 // since marking is racy and double-scanning objects
516 // may legitimately make the remaining scan work
517 // negative, even in the hard goal regime.
518 scanWorkRemaining = 1000
521 // Compute the heap distance remaining.
522 heapRemaining := heapGoal - int64(live)
523 if heapRemaining <= 0 {
524 // This shouldn't happen, but if it does, avoid
525 // dividing by zero or setting the assist negative.
529 // Compute the mutator assist ratio so by the time the mutator
530 // allocates the remaining heap bytes up to heapGoal, it will
531 // have done (or stolen) the remaining amount of scan work.
532 // Note that the assist ratio values are updated atomically
533 // but not together. This means there may be some degree of
534 // skew between the two values. This is generally OK as the
535 // values shift relatively slowly over the course of a GC
537 assistWorkPerByte := float64(scanWorkRemaining) / float64(heapRemaining)
538 assistBytesPerWork := float64(heapRemaining) / float64(scanWorkRemaining)
539 c.assistWorkPerByte.Store(assistWorkPerByte)
540 c.assistBytesPerWork.Store(assistBytesPerWork)
543 // endCycle computes the consMark estimate for the next cycle.
544 // userForced indicates whether the current GC cycle was forced
545 // by the application.
546 func (c *gcControllerState) endCycle(now int64, procs int, userForced bool) {
547 // Record last heap goal for the scavenger.
548 // We'll be updating the heap goal soon.
549 gcController.lastHeapGoal = gcController.heapGoal
551 // Compute the duration of time for which assists were turned on.
552 assistDuration := now - c.markStartTime
554 // Assume background mark hit its utilization goal.
555 utilization := gcBackgroundUtilization
556 // Add assist utilization; avoid divide by zero.
557 if assistDuration > 0 {
558 utilization += float64(c.assistTime) / float64(assistDuration*int64(procs))
561 if c.heapLive <= c.trigger {
562 // Shouldn't happen, but let's be very safe about this in case the
563 // GC is somehow extremely short.
565 // In this case though, the only reasonable value for c.heapLive-c.trigger
566 // would be 0, which isn't really all that useful, i.e. the GC was so short
567 // that it didn't matter.
569 // Ignore this case and don't update anything.
572 idleUtilization := 0.0
573 if assistDuration > 0 {
574 idleUtilization = float64(c.idleMarkTime) / float64(assistDuration*int64(procs))
576 // Determine the cons/mark ratio.
578 // The units we want for the numerator and denominator are both B / cpu-ns.
579 // We get this by taking the bytes allocated or scanned, and divide by the amount of
580 // CPU time it took for those operations. For allocations, that CPU time is
582 // assistDuration * procs * (1 - utilization)
584 // Where utilization includes just background GC workers and assists. It does *not*
585 // include idle GC work time, because in theory the mutator is free to take that at
588 // For scanning, that CPU time is
590 // assistDuration * procs * (utilization + idleUtilization)
592 // In this case, we *include* idle utilization, because that is additional CPU time that the
593 // the GC had available to it.
595 // In effect, idle GC time is sort of double-counted here, but it's very weird compared
596 // to other kinds of GC work, because of how fluid it is. Namely, because the mutator is
597 // *always* free to take it.
599 // So this calculation is really:
600 // (heapLive-trigger) / (assistDuration * procs * (1-utilization)) /
601 // (scanWork) / (assistDuration * procs * (utilization+idleUtilization)
603 // Note that because we only care about the ratio, assistDuration and procs cancel out.
604 scanWork := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
605 currentConsMark := (float64(c.heapLive-c.trigger) * (utilization + idleUtilization)) /
606 (float64(scanWork) * (1 - utilization))
608 // Update cons/mark controller. The time period for this is 1 GC cycle.
610 // This use of a PI controller might seem strange. So, here's an explanation:
612 // currentConsMark represents the consMark we *should've* had to be perfectly
613 // on-target for this cycle. Given that we assume the next GC will be like this
614 // one in the steady-state, it stands to reason that we should just pick that
615 // as our next consMark. In practice, however, currentConsMark is too noisy:
616 // we're going to be wildly off-target in each GC cycle if we do that.
618 // What we do instead is make a long-term assumption: there is some steady-state
619 // consMark value, but it's obscured by noise. By constantly shooting for this
620 // noisy-but-perfect consMark value, the controller will bounce around a bit,
621 // but its average behavior, in aggregate, should be less noisy and closer to
622 // the true long-term consMark value, provided its tuned to be slightly overdamped.
624 oldConsMark := c.consMark
625 c.consMark, ok = c.consMarkController.next(c.consMark, currentConsMark, 1.0)
627 // The error spiraled out of control. This is incredibly unlikely seeing
628 // as this controller is essentially just a smoothing function, but it might
629 // mean that something went very wrong with how currentConsMark was calculated.
630 // Just reset consMark and keep going.
634 if debug.gcpacertrace > 0 {
636 goal := gcGoalUtilization * 100
637 print("pacer: ", int(utilization*100), "% CPU (", int(goal), " exp.) for ")
638 print(c.heapScanWork.Load(), "+", c.stackScanWork.Load(), "+", c.globalsScanWork.Load(), " B work (", c.lastHeapScan+c.stackScan+c.globalsScan, " B exp.) ")
639 print("in ", c.trigger, " B -> ", c.heapLive, " B (∆goal ", int64(c.heapLive)-int64(c.heapGoal), ", cons/mark ", oldConsMark, ")")
641 print("[controller reset]")
648 // enlistWorker encourages another dedicated mark worker to start on
649 // another P if there are spare worker slots. It is used by putfull
650 // when more work is made available.
653 func (c *gcControllerState) enlistWorker() {
654 // If there are idle Ps, wake one so it will run an idle worker.
655 // NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
657 // if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
662 // There are no idle Ps. If we need more dedicated workers,
663 // try to preempt a running P so it will switch to a worker.
664 if c.dedicatedMarkWorkersNeeded <= 0 {
667 // Pick a random other P to preempt.
672 if gp == nil || gp.m == nil || gp.m.p == 0 {
675 myID := gp.m.p.ptr().id
676 for tries := 0; tries < 5; tries++ {
677 id := int32(fastrandn(uint32(gomaxprocs - 1)))
682 if p.status != _Prunning {
691 // findRunnableGCWorker returns a background mark worker for _p_ if it
692 // should be run. This must only be called when gcBlackenEnabled != 0.
693 func (c *gcControllerState) findRunnableGCWorker(_p_ *p) *g {
694 if gcBlackenEnabled == 0 {
695 throw("gcControllerState.findRunnable: blackening not enabled")
698 if !gcMarkWorkAvailable(_p_) {
699 // No work to be done right now. This can happen at
700 // the end of the mark phase when there are still
701 // assists tapering off. Don't bother running a worker
702 // now because it'll just return immediately.
706 // Grab a worker before we commit to running below.
707 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
709 // There is at least one worker per P, so normally there are
710 // enough workers to run on all Ps, if necessary. However, once
711 // a worker enters gcMarkDone it may park without rejoining the
712 // pool, thus freeing a P with no corresponding worker.
713 // gcMarkDone never depends on another worker doing work, so it
714 // is safe to simply do nothing here.
716 // If gcMarkDone bails out without completing the mark phase,
717 // it will always do so with queued global work. Thus, that P
718 // will be immediately eligible to re-run the worker G it was
719 // just using, ensuring work can complete.
723 decIfPositive := func(ptr *int64) bool {
725 v := atomic.Loadint64(ptr)
730 if atomic.Casint64(ptr, v, v-1) {
736 if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
737 // This P is now dedicated to marking until the end of
738 // the concurrent mark phase.
739 _p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
740 } else if c.fractionalUtilizationGoal == 0 {
741 // No need for fractional workers.
742 gcBgMarkWorkerPool.push(&node.node)
745 // Is this P behind on the fractional utilization
748 // This should be kept in sync with pollFractionalWorkerExit.
749 delta := nanotime() - c.markStartTime
750 if delta > 0 && float64(_p_.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
751 // Nope. No need to run a fractional worker.
752 gcBgMarkWorkerPool.push(&node.node)
755 // Run a fractional worker.
756 _p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode
759 // Run the background mark worker.
761 casgstatus(gp, _Gwaiting, _Grunnable)
768 // resetLive sets up the controller state for the next mark phase after the end
769 // of the previous one. Must be called after endCycle and before commit, before
770 // the world is started.
772 // The world must be stopped.
773 func (c *gcControllerState) resetLive(bytesMarked uint64) {
774 c.heapMarked = bytesMarked
775 c.heapLive = bytesMarked
776 c.heapScan = uint64(c.heapScanWork.Load())
777 c.lastHeapScan = uint64(c.heapScanWork.Load())
779 // heapLive was updated, so emit a trace event.
785 // logWorkTime updates mark work accounting in the controller by a duration of
786 // work in nanoseconds.
788 // Safe to execute at any time.
789 func (c *gcControllerState) logWorkTime(mode gcMarkWorkerMode, duration int64) {
791 case gcMarkWorkerDedicatedMode:
792 atomic.Xaddint64(&c.dedicatedMarkTime, duration)
793 atomic.Xaddint64(&c.dedicatedMarkWorkersNeeded, 1)
794 case gcMarkWorkerFractionalMode:
795 atomic.Xaddint64(&c.fractionalMarkTime, duration)
796 case gcMarkWorkerIdleMode:
797 atomic.Xaddint64(&c.idleMarkTime, duration)
799 throw("logWorkTime: unknown mark worker mode")
803 func (c *gcControllerState) update(dHeapLive, dHeapScan int64) {
805 atomic.Xadd64(&gcController.heapLive, dHeapLive)
807 // gcController.heapLive changed.
811 if gcBlackenEnabled == 0 {
812 // Update heapScan when we're not in a current GC. It is fixed
813 // at the beginning of a cycle.
815 atomic.Xadd64(&gcController.heapScan, dHeapScan)
818 // gcController.heapLive changed.
823 func (c *gcControllerState) addScannableStack(pp *p, amount int64) {
825 atomic.Xadd64(&c.scannableStackSize, amount)
828 pp.scannableStackSizeDelta += amount
829 if pp.scannableStackSizeDelta >= scannableStackSizeSlack || pp.scannableStackSizeDelta <= -scannableStackSizeSlack {
830 atomic.Xadd64(&c.scannableStackSize, pp.scannableStackSizeDelta)
831 pp.scannableStackSizeDelta = 0
835 func (c *gcControllerState) addGlobals(amount int64) {
836 atomic.Xadd64(&c.globalsScan, amount)
839 // commit recomputes all pacing parameters from scratch, namely
840 // absolute trigger, the heap goal, mark pacing, and sweep pacing.
842 // This can be called any time. If GC is the in the middle of a
843 // concurrent phase, it will adjust the pacing of that phase.
845 // This depends on gcPercent, gcController.heapMarked, and
846 // gcController.heapLive. These must be up to date.
848 // mheap_.lock must be held or the world must be stopped.
849 func (c *gcControllerState) commit() {
851 assertWorldStoppedOrLockHeld(&mheap_.lock)
854 // Compute the next GC goal, which is when the allocated heap
855 // has grown by GOGC/100 over where it started the last cycle,
856 // plus additional runway for non-heap sources of GC work.
858 if gcPercent := c.gcPercent.Load(); gcPercent >= 0 {
859 goal = c.heapMarked + (c.heapMarked+atomic.Load64(&c.stackScan)+atomic.Load64(&c.globalsScan))*uint64(gcPercent)/100
862 // Don't trigger below the minimum heap size.
863 minTrigger := c.heapMinimum
865 // Concurrent sweep happens in the heap growth
866 // from gcController.heapLive to trigger, so ensure
867 // that concurrent sweep has some heap growth
868 // in which to perform sweeping before we
869 // start the next GC cycle.
870 sweepMin := atomic.Load64(&c.heapLive) + sweepMinHeapDistance
871 if sweepMin > minTrigger {
872 minTrigger = sweepMin
876 // If we let the trigger go too low, then if the application
877 // is allocating very rapidly we might end up in a situation
878 // where we're allocating black during a nearly always-on GC.
879 // The result of this is a growing heap and ultimately an
880 // increase in RSS. By capping us at a point >0, we're essentially
881 // saying that we're OK using more CPU during the GC to prevent
882 // this growth in RSS.
884 // The current constant was chosen empirically: given a sufficiently
885 // fast/scalable allocator with 48 Ps that could drive the trigger ratio
886 // to <0.05, this constant causes applications to retain the same peak
887 // RSS compared to not having this allocator.
888 if triggerBound := uint64(0.7*float64(goal-c.heapMarked)) + c.heapMarked; minTrigger < triggerBound {
889 minTrigger = triggerBound
892 // For small heaps, set the max trigger point at 95% of the heap goal.
893 // This ensures we always have *some* headroom when the GC actually starts.
894 // For larger heaps, set the max trigger point at the goal, minus the
895 // minimum heap size.
896 // This choice follows from the fact that the minimum heap size is chosen
897 // to reflect the costs of a GC with no work to do. With a large heap but
898 // very little scan work to perform, this gives us exactly as much runway
899 // as we would need, in the worst case.
900 maxRunway := uint64(0.95 * float64(goal-c.heapMarked))
901 if largeHeapMaxRunway := goal - c.heapMinimum; goal > c.heapMinimum && maxRunway < largeHeapMaxRunway {
902 maxRunway = largeHeapMaxRunway
904 maxTrigger := maxRunway + c.heapMarked
905 if maxTrigger < minTrigger {
906 maxTrigger = minTrigger
909 // Compute the trigger by using our estimate of the cons/mark ratio.
911 // The idea is to take our expected scan work, and multiply it by
912 // the cons/mark ratio to determine how long it'll take to complete
913 // that scan work in terms of bytes allocated. This gives us our GC's
916 // However, the cons/mark ratio is a ratio of rates per CPU-second, but
917 // here we care about the relative rates for some division of CPU
918 // resources among the mutator and the GC.
920 // To summarize, we have B / cpu-ns, and we want B / ns. We get that
921 // by multiplying by our desired division of CPU resources. We choose
922 // to express CPU resources as GOMAPROCS*fraction. Note that because
923 // we're working with a ratio here, we can omit the number of CPU cores,
924 // because they'll appear in the numerator and denominator and cancel out.
925 // As a result, this is basically just "weighing" the cons/mark ratio by
926 // our desired division of resources.
928 // Furthermore, by setting the trigger so that CPU resources are divided
929 // this way, assuming that the cons/mark ratio is correct, we make that
930 // division a reality.
932 runway := uint64((c.consMark * (1 - gcGoalUtilization) / (gcGoalUtilization)) * float64(c.lastHeapScan+c.stackScan+c.globalsScan))
936 trigger = goal - runway
938 if trigger < minTrigger {
941 if trigger > maxTrigger {
948 // Commit to the trigger and goal.
950 atomic.Store64(&c.heapGoal, goal)
955 // Update mark pacing.
956 if gcphase != _GCoff {
961 // effectiveGrowthRatio returns the current effective heap growth
962 // ratio (GOGC/100) based on heapMarked from the previous GC and
963 // heapGoal for the current GC.
965 // This may differ from gcPercent/100 because of various upper and
966 // lower bounds on gcPercent. For example, if the heap is smaller than
967 // heapMinimum, this can be higher than gcPercent/100.
969 // mheap_.lock must be held or the world must be stopped.
970 func (c *gcControllerState) effectiveGrowthRatio() float64 {
972 assertWorldStoppedOrLockHeld(&mheap_.lock)
975 egogc := float64(atomic.Load64(&c.heapGoal)-c.heapMarked) / float64(c.heapMarked)
977 // Shouldn't happen, but just in case.
983 // setGCPercent updates gcPercent and all related pacer state.
984 // Returns the old value of gcPercent.
986 // Calls gcControllerState.commit.
988 // The world must be stopped, or mheap_.lock must be held.
989 func (c *gcControllerState) setGCPercent(in int32) int32 {
991 assertWorldStoppedOrLockHeld(&mheap_.lock)
994 out := c.gcPercent.Load()
998 c.heapMinimum = defaultHeapMinimum * uint64(in) / 100
999 c.gcPercent.Store(in)
1000 // Update pacing in response to gcPercent change.
1006 //go:linkname setGCPercent runtime/debug.setGCPercent
1007 func setGCPercent(in int32) (out int32) {
1008 // Run on the system stack since we grab the heap lock.
1009 systemstack(func() {
1011 out = gcController.setGCPercent(in)
1012 gcPaceSweeper(gcController.trigger)
1013 gcPaceScavenger(gcController.heapGoal, gcController.lastHeapGoal)
1014 unlock(&mheap_.lock)
1017 // If we just disabled GC, wait for any concurrent GC mark to
1018 // finish so we always return with no GC running.
1020 gcWaitOnMark(atomic.Load(&work.cycles))
1026 func readGOGC() int32 {
1027 p := gogetenv("GOGC")
1031 if n, ok := atoi32(p); ok {
1037 type piController struct {
1038 kp float64 // Proportional constant.
1039 ti float64 // Integral time constant.
1040 tt float64 // Reset time.
1042 min, max float64 // Output boundaries.
1044 // PI controller state.
1046 errIntegral float64 // Integral of the error from t=0 to now.
1049 errOverflow bool // Set if errIntegral ever overflowed.
1050 inputOverflow bool // Set if an operation with the input overflowed.
1053 // next provides a new sample to the controller.
1055 // input is the sample, setpoint is the desired point, and period is how much
1056 // time (in whatever unit makes the most sense) has passed since the last sample.
1058 // Returns a new value for the variable it's controlling, and whether the operation
1059 // completed successfully. One reason this might fail is if error has been growing
1060 // in an unbounded manner, to the point of overflow.
1062 // In the specific case of an error overflow occurs, the errOverflow field will be
1063 // set and the rest of the controller's internal state will be fully reset.
1064 func (c *piController) next(input, setpoint, period float64) (float64, bool) {
1065 // Compute the raw output value.
1066 prop := c.kp * (setpoint - input)
1067 rawOutput := prop + c.errIntegral
1069 // Clamp rawOutput into output.
1071 if isInf(output) || isNaN(output) {
1072 // The input had a large enough magnitude that either it was already
1073 // overflowed, or some operation with it overflowed.
1074 // Set a flag and reset. That's the safest thing to do.
1076 c.inputOverflow = true
1081 } else if output > c.max {
1085 // Update the controller's state.
1086 if c.ti != 0 && c.tt != 0 {
1087 c.errIntegral += (c.kp*period/c.ti)*(setpoint-input) + (period/c.tt)*(output-rawOutput)
1088 if isInf(c.errIntegral) || isNaN(c.errIntegral) {
1089 // So much error has accumulated that we managed to overflow.
1090 // The assumptions around the controller have likely broken down.
1091 // Set a flag and reset. That's the safest thing to do.
1093 c.errOverflow = true
1100 // reset resets the controller state, except for controller error flags.
1101 func (c *piController) reset() {