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"
14 // go119MemoryLimitSupport is a feature flag for a number of changes
15 // related to the memory limit feature (#48409). Disabling this flag
16 // disables those features, as well as the memory limit mechanism,
17 // which becomes a no-op.
18 const go119MemoryLimitSupport = true
21 // gcGoalUtilization is the goal CPU utilization for
22 // marking as a fraction of GOMAXPROCS.
24 // Increasing the goal utilization will shorten GC cycles as the GC
25 // has more resources behind it, lessening costs from the write barrier,
26 // but comes at the cost of increasing mutator latency.
27 gcGoalUtilization = gcBackgroundUtilization
29 // gcBackgroundUtilization is the fixed CPU utilization for background
30 // marking. It must be <= gcGoalUtilization. The difference between
31 // gcGoalUtilization and gcBackgroundUtilization will be made up by
32 // mark assists. The scheduler will aim to use within 50% of this
35 // As a general rule, there's little reason to set gcBackgroundUtilization
36 // < gcGoalUtilization. One reason might be in mostly idle applications,
37 // where goroutines are unlikely to assist at all, so the actual
38 // utilization will be lower than the goal. But this is moot point
39 // because the idle mark workers already soak up idle CPU resources.
40 // These two values are still kept separate however because they are
41 // distinct conceptually, and in previous iterations of the pacer the
42 // distinction was more important.
43 gcBackgroundUtilization = 0.25
45 // gcCreditSlack is the amount of scan work credit that can
46 // accumulate locally before updating gcController.heapScanWork and,
47 // optionally, gcController.bgScanCredit. Lower values give a more
48 // accurate assist ratio and make it more likely that assists will
49 // successfully steal background credit. Higher values reduce memory
53 // gcAssistTimeSlack is the nanoseconds of mutator assist time that
54 // can accumulate on a P before updating gcController.assistTime.
55 gcAssistTimeSlack = 5000
57 // gcOverAssistWork determines how many extra units of scan work a GC
58 // assist does when an assist happens. This amortizes the cost of an
59 // assist by pre-paying for this many bytes of future allocations.
60 gcOverAssistWork = 64 << 10
62 // defaultHeapMinimum is the value of heapMinimum for GOGC==100.
63 defaultHeapMinimum = (goexperiment.HeapMinimum512KiBInt)*(512<<10) +
64 (1-goexperiment.HeapMinimum512KiBInt)*(4<<20)
66 // scannableStackSizeSlack is the bytes of stack space allocated or freed
67 // that can accumulate on a P before updating gcController.stackSize.
68 scannableStackSizeSlack = 8 << 10
70 // memoryLimitHeapGoalHeadroom is the amount of headroom the pacer gives to
71 // the heap goal when operating in the memory-limited regime. That is,
72 // it'll reduce the heap goal by this many extra bytes off of the base
74 memoryLimitHeapGoalHeadroom = 1 << 20
78 if offset := unsafe.Offsetof(gcController.heapLive); offset%8 != 0 {
80 throw("gcController.heapLive not aligned to 8 bytes")
84 // gcController implements the GC pacing controller that determines
85 // when to trigger concurrent garbage collection and how much marking
86 // work to do in mutator assists and background marking.
88 // It calculates the ratio between the allocation rate (in terms of CPU
89 // time) and the GC scan throughput to determine the heap size at which to
90 // trigger a GC cycle such that no GC assists are required to finish on time.
91 // This algorithm thus optimizes GC CPU utilization to the dedicated background
92 // mark utilization of 25% of GOMAXPROCS by minimizing GC assists.
93 // GOMAXPROCS. The high-level design of this algorithm is documented
94 // at https://github.com/golang/proposal/blob/master/design/44167-gc-pacer-redesign.md.
95 // See https://golang.org/s/go15gcpacing for additional historical context.
96 var gcController gcControllerState
98 type gcControllerState struct {
99 // Initialized from GOGC. GOGC=off means no GC.
100 gcPercent atomic.Int32
102 _ uint32 // padding so following 64-bit values are 8-byte aligned
104 // memoryLimit is the soft memory limit in bytes.
106 // Initialized from GOMEMLIMIT. GOMEMLIMIT=off is equivalent to MaxInt64
107 // which means no soft memory limit in practice.
109 // This is an int64 instead of a uint64 to more easily maintain parity with
110 // the SetMemoryLimit API, which sets a maximum at MaxInt64. This value
111 // should never be negative.
112 memoryLimit atomic.Int64
114 // heapMinimum is the minimum heap size at which to trigger GC.
115 // For small heaps, this overrides the usual GOGC*live set rule.
117 // When there is a very small live set but a lot of allocation, simply
118 // collecting when the heap reaches GOGC*live results in many GC
119 // cycles and high total per-GC overhead. This minimum amortizes this
120 // per-GC overhead while keeping the heap reasonably small.
122 // During initialization this is set to 4MB*GOGC/100. In the case of
123 // GOGC==0, this will set heapMinimum to 0, resulting in constant
124 // collection even when the heap size is small, which is useful for
128 // runway is the amount of runway in heap bytes allocated by the
129 // application that we want to give the GC once it starts.
131 // This is computed from consMark during mark termination.
134 // consMark is the estimated per-CPU consMark ratio for the application.
136 // It represents the ratio between the application's allocation
137 // rate, as bytes allocated per CPU-time, and the GC's scan rate,
138 // as bytes scanned per CPU-time.
139 // The units of this ratio are (B / cpu-ns) / (B / cpu-ns).
141 // At a high level, this value is computed as the bytes of memory
142 // allocated (cons) per unit of scan work completed (mark) in a GC
143 // cycle, divided by the CPU time spent on each activity.
145 // Updated at the end of each GC cycle, in endCycle.
148 // consMarkController holds the state for the mark-cons ratio
149 // estimation over time.
151 // Its purpose is to smooth out noisiness in the computation of
152 // consMark; see consMark for details.
153 consMarkController piController
155 _ uint32 // Padding for atomics on 32-bit platforms.
157 // gcPercentHeapGoal is the goal heapLive for when next GC ends derived
160 // Set to ^uint64(0) if gcPercent is disabled.
161 gcPercentHeapGoal atomic.Uint64
163 // sweepDistMinTrigger is the minimum trigger to ensure a minimum
166 // This bound is also special because it applies to both the trigger
167 // *and* the goal (all other trigger bounds must be based *on* the goal).
169 // It is computed ahead of time, at commit time. The theory is that,
170 // absent a sudden change to a parameter like gcPercent, the trigger
171 // will be chosen to always give the sweeper enough headroom. However,
172 // such a change might dramatically and suddenly move up the trigger,
173 // in which case we need to ensure the sweeper still has enough headroom.
174 sweepDistMinTrigger atomic.Uint64
176 // triggered is the point at which the current GC cycle actually triggered.
177 // Only valid during the mark phase of a GC cycle, otherwise set to ^uint64(0).
179 // Updated while the world is stopped.
182 // lastHeapGoal is the value of heapGoal at the moment the last GC
183 // ended. Note that this is distinct from the last value heapGoal had,
184 // because it could change if e.g. gcPercent changes.
186 // Read and written with the world stopped or with mheap_.lock held.
189 // heapLive is the number of bytes considered live by the GC.
190 // That is: retained by the most recent GC plus allocated
191 // since then. heapLive ≤ memstats.totalAlloc-memstats.totalFree, since
192 // heapAlloc includes unmarked objects that have not yet been swept (and
193 // hence goes up as we allocate and down as we sweep) while heapLive
194 // excludes these objects (and hence only goes up between GCs).
196 // This is updated atomically without locking. To reduce
197 // contention, this is updated only when obtaining a span from
198 // an mcentral and at this point it counts all of the
199 // unallocated slots in that span (which will be allocated
200 // before that mcache obtains another span from that
201 // mcentral). Hence, it slightly overestimates the "true" live
202 // heap size. It's better to overestimate than to
203 // underestimate because 1) this triggers the GC earlier than
204 // necessary rather than potentially too late and 2) this
205 // leads to a conservative GC rate rather than a GC rate that
206 // is potentially too low.
208 // Reads should likewise be atomic (or during STW).
210 // Whenever this is updated, call traceHeapAlloc() and
211 // this gcControllerState's revise() method.
214 // heapScan is the number of bytes of "scannable" heap. This
215 // is the live heap (as counted by heapLive), but omitting
216 // no-scan objects and no-scan tails of objects.
218 // This value is fixed at the start of a GC cycle, so during a
219 // GC cycle it is safe to read without atomics, and it represents
220 // the maximum scannable heap.
223 // lastHeapScan is the number of bytes of heap that were scanned
224 // last GC cycle. It is the same as heapMarked, but only
225 // includes the "scannable" parts of objects.
227 // Updated when the world is stopped.
230 // stackScan is a snapshot of scannableStackSize taken at each GC
231 // STW pause and is used in pacing decisions.
233 // Updated only while the world is stopped.
236 // scannableStackSize is the amount of allocated goroutine stack space in
237 // use by goroutines.
239 // This number tracks allocated goroutine stack space rather than used
240 // goroutine stack space (i.e. what is actually scanned) because used
241 // goroutine stack space is much harder to measure cheaply. By using
242 // allocated space, we make an overestimate; this is OK, it's better
243 // to conservatively overcount than undercount.
245 // Read and updated atomically.
246 scannableStackSize uint64
248 // globalsScan is the total amount of global variable space
249 // that is scannable.
251 // Read and updated atomically.
254 // heapMarked is the number of bytes marked by the previous
255 // GC. After mark termination, heapLive == heapMarked, but
256 // unlike heapLive, heapMarked does not change until the
257 // next mark termination.
260 // heapScanWork is the total heap scan work performed this cycle.
261 // stackScanWork is the total stack scan work performed this cycle.
262 // globalsScanWork is the total globals scan work performed this cycle.
264 // These are updated atomically during the cycle. Updates occur in
265 // bounded batches, since they are both written and read
266 // throughout the cycle. At the end of the cycle, heapScanWork is how
267 // much of the retained heap is scannable.
269 // Currently these are measured in bytes. For most uses, this is an
270 // opaque unit of work, but for estimation the definition is important.
272 // Note that stackScanWork includes all allocated space, not just the
273 // size of the stack itself, mirroring stackSize.
274 heapScanWork atomic.Int64
275 stackScanWork atomic.Int64
276 globalsScanWork atomic.Int64
278 // bgScanCredit is the scan work credit accumulated by the
279 // concurrent background scan. This credit is accumulated by
280 // the background scan and stolen by mutator assists. This is
281 // updated atomically. Updates occur in bounded batches, since
282 // it is both written and read throughout the cycle.
285 // assistTime is the nanoseconds spent in mutator assists
286 // during this cycle. This is updated atomically, and must also
287 // be updated atomically even during a STW, because it is read
288 // by sysmon. Updates occur in bounded batches, since it is both
289 // written and read throughout the cycle.
290 assistTime atomic.Int64
292 // dedicatedMarkTime is the nanoseconds spent in dedicated
293 // mark workers during this cycle. This is updated atomically
294 // at the end of the concurrent mark phase.
295 dedicatedMarkTime int64
297 // fractionalMarkTime is the nanoseconds spent in the
298 // fractional mark worker during this cycle. This is updated
299 // atomically throughout the cycle and will be up-to-date if
300 // the fractional mark worker is not currently running.
301 fractionalMarkTime int64
303 // idleMarkTime is the nanoseconds spent in idle marking
304 // during this cycle. This is updated atomically throughout
308 // markStartTime is the absolute start time in nanoseconds
309 // that assists and background mark workers started.
312 // dedicatedMarkWorkersNeeded is the number of dedicated mark
313 // workers that need to be started. This is computed at the
314 // beginning of each cycle and decremented atomically as
315 // dedicated mark workers get started.
316 dedicatedMarkWorkersNeeded int64
318 // idleMarkWorkers is two packed int32 values in a single uint64.
319 // These two values are always updated simultaneously.
321 // The bottom int32 is the current number of idle mark workers executing.
323 // The top int32 is the maximum number of idle mark workers allowed to
324 // execute concurrently. Normally, this number is just gomaxprocs. However,
325 // during periodic GC cycles it is set to 0 because the system is idle
326 // anyway; there's no need to go full blast on all of GOMAXPROCS.
328 // The maximum number of idle mark workers is used to prevent new workers
329 // from starting, but it is not a hard maximum. It is possible (but
330 // exceedingly rare) for the current number of idle mark workers to
331 // transiently exceed the maximum. This could happen if the maximum changes
332 // just after a GC ends, and an M with no P.
334 // Note that if we have no dedicated mark workers, we set this value to
335 // 1 in this case we only have fractional GC workers which aren't scheduled
336 // strictly enough to ensure GC progress. As a result, idle-priority mark
337 // workers are vital to GC progress in these situations.
339 // For example, consider a situation in which goroutines block on the GC
340 // (such as via runtime.GOMAXPROCS) and only fractional mark workers are
341 // scheduled (e.g. GOMAXPROCS=1). Without idle-priority mark workers, the
342 // last running M might skip scheduling a fractional mark worker if its
343 // utilization goal is met, such that once it goes to sleep (because there's
344 // nothing to do), there will be nothing else to spin up a new M for the
345 // fractional worker in the future, stalling GC progress and causing a
346 // deadlock. However, idle-priority workers will *always* run when there is
347 // nothing left to do, ensuring the GC makes progress.
349 // See github.com/golang/go/issues/44163 for more details.
350 idleMarkWorkers atomic.Uint64
352 // assistWorkPerByte is the ratio of scan work to allocated
353 // bytes that should be performed by mutator assists. This is
354 // computed at the beginning of each cycle and updated every
355 // time heapScan is updated.
356 assistWorkPerByte atomic.Float64
358 // assistBytesPerWork is 1/assistWorkPerByte.
360 // Note that because this is read and written independently
361 // from assistWorkPerByte users may notice a skew between
362 // the two values, and such a state should be safe.
363 assistBytesPerWork atomic.Float64
365 // fractionalUtilizationGoal is the fraction of wall clock
366 // time that should be spent in the fractional mark worker on
367 // each P that isn't running a dedicated worker.
369 // For example, if the utilization goal is 25% and there are
370 // no dedicated workers, this will be 0.25. If the goal is
371 // 25%, there is one dedicated worker, and GOMAXPROCS is 5,
372 // this will be 0.05 to make up the missing 5%.
374 // If this is zero, no fractional workers are needed.
375 fractionalUtilizationGoal float64
377 // These memory stats are effectively duplicates of fields from
378 // memstats.heapStats but are updated atomically or with the world
379 // stopped and don't provide the same consistency guarantees.
381 // Because the runtime is responsible for managing a memory limit, it's
382 // useful to couple these stats more tightly to the gcController, which
383 // is intimately connected to how that memory limit is maintained.
384 heapInUse sysMemStat // bytes in mSpanInUse spans
385 heapReleased sysMemStat // bytes released to the OS
386 heapFree sysMemStat // bytes not in any span, but not released to the OS
387 totalAlloc atomic.Uint64 // total bytes allocated
388 totalFree atomic.Uint64 // total bytes freed
389 mappedReady atomic.Uint64 // total virtual memory in the Ready state (see mem.go).
391 // test indicates that this is a test-only copy of gcControllerState.
397 func (c *gcControllerState) init(gcPercent int32, memoryLimit int64) {
398 c.heapMinimum = defaultHeapMinimum
399 c.triggered = ^uint64(0)
401 c.consMarkController = piController{
402 // Tuned first via the Ziegler-Nichols process in simulation,
403 // then the integral time was manually tuned against real-world
404 // applications to deal with noisiness in the measured cons/mark
409 // Set a high reset time in GC cycles.
410 // This is inversely proportional to the rate at which we
411 // accumulate error from clipping. By making this very high
412 // we make the accumulation slow. In general, clipping is
413 // OK in our situation, hence the choice.
415 // Tune this if we get unintended effects from clipping for
422 c.setGCPercent(gcPercent)
423 c.setMemoryLimit(memoryLimit)
424 c.commit(true) // No sweep phase in the first GC cycle.
425 // N.B. Don't bother calling traceHeapGoal. Tracing is never enabled at
426 // initialization time.
427 // N.B. No need to call revise; there's no GC enabled during
431 // startCycle resets the GC controller's state and computes estimates
432 // for a new GC cycle. The caller must hold worldsema and the world
434 func (c *gcControllerState) startCycle(markStartTime int64, procs int, trigger gcTrigger) {
435 c.heapScanWork.Store(0)
436 c.stackScanWork.Store(0)
437 c.globalsScanWork.Store(0)
439 c.assistTime.Store(0)
440 c.dedicatedMarkTime = 0
441 c.fractionalMarkTime = 0
443 c.markStartTime = markStartTime
444 c.stackScan = atomic.Load64(&c.scannableStackSize)
445 c.triggered = c.heapLive
447 // Compute the background mark utilization goal. In general,
448 // this may not come out exactly. We round the number of
449 // dedicated workers so that the utilization is closest to
450 // 25%. For small GOMAXPROCS, this would introduce too much
451 // error, so we add fractional workers in that case.
452 totalUtilizationGoal := float64(procs) * gcBackgroundUtilization
453 c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5)
454 utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
455 const maxUtilError = 0.3
456 if utilError < -maxUtilError || utilError > maxUtilError {
457 // Rounding put us more than 30% off our goal. With
458 // gcBackgroundUtilization of 25%, this happens for
459 // GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
460 // workers to compensate.
461 if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
462 // Too many dedicated workers.
463 c.dedicatedMarkWorkersNeeded--
465 c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(procs)
467 c.fractionalUtilizationGoal = 0
470 // In STW mode, we just want dedicated workers.
471 if debug.gcstoptheworld > 0 {
472 c.dedicatedMarkWorkersNeeded = int64(procs)
473 c.fractionalUtilizationGoal = 0
477 for _, p := range allp {
479 p.gcFractionalMarkTime = 0
482 if trigger.kind == gcTriggerTime {
483 // During a periodic GC cycle, reduce the number of idle mark workers
484 // required. However, we need at least one dedicated mark worker or
485 // idle GC worker to ensure GC progress in some scenarios (see comment
486 // on maxIdleMarkWorkers).
487 if c.dedicatedMarkWorkersNeeded > 0 {
488 c.setMaxIdleMarkWorkers(0)
490 // TODO(mknyszek): The fundamental reason why we need this is because
491 // we can't count on the fractional mark worker to get scheduled.
492 // Fix that by ensuring it gets scheduled according to its quota even
493 // if the rest of the application is idle.
494 c.setMaxIdleMarkWorkers(1)
497 // N.B. gomaxprocs and dedicatedMarkWorkersNeeded is guaranteed not to
498 // change during a GC cycle.
499 c.setMaxIdleMarkWorkers(int32(procs) - int32(c.dedicatedMarkWorkersNeeded))
502 // Compute initial values for controls that are updated
503 // throughout the cycle.
506 if debug.gcpacertrace > 0 {
507 heapGoal := c.heapGoal()
508 assistRatio := c.assistWorkPerByte.Load()
509 print("pacer: assist ratio=", assistRatio,
510 " (scan ", gcController.heapScan>>20, " MB in ",
511 work.initialHeapLive>>20, "->",
512 heapGoal>>20, " MB)",
513 " workers=", c.dedicatedMarkWorkersNeeded,
514 "+", c.fractionalUtilizationGoal, "\n")
518 // revise updates the assist ratio during the GC cycle to account for
519 // improved estimates. This should be called whenever gcController.heapScan,
520 // gcController.heapLive, or if any inputs to gcController.heapGoal are
521 // updated. It is safe to call concurrently, but it may race with other
524 // The result of this race is that the two assist ratio values may not line
525 // up or may be stale. In practice this is OK because the assist ratio
526 // moves slowly throughout a GC cycle, and the assist ratio is a best-effort
527 // heuristic anyway. Furthermore, no part of the heuristic depends on
528 // the two assist ratio values being exact reciprocals of one another, since
529 // the two values are used to convert values from different sources.
531 // The worst case result of this raciness is that we may miss a larger shift
532 // in the ratio (say, if we decide to pace more aggressively against the
533 // hard heap goal) but even this "hard goal" is best-effort (see #40460).
534 // The dedicated GC should ensure we don't exceed the hard goal by too much
535 // in the rare case we do exceed it.
537 // It should only be called when gcBlackenEnabled != 0 (because this
538 // is when assists are enabled and the necessary statistics are
540 func (c *gcControllerState) revise() {
541 gcPercent := c.gcPercent.Load()
543 // If GC is disabled but we're running a forced GC,
544 // act like GOGC is huge for the below calculations.
547 live := atomic.Load64(&c.heapLive)
548 scan := atomic.Load64(&c.heapScan)
549 work := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
551 // Assume we're under the soft goal. Pace GC to complete at
552 // heapGoal assuming the heap is in steady-state.
553 heapGoal := int64(c.heapGoal())
555 // The expected scan work is computed as the amount of bytes scanned last
556 // GC cycle, plus our estimate of stacks and globals work for this cycle.
557 scanWorkExpected := int64(c.lastHeapScan + c.stackScan + c.globalsScan)
559 // maxScanWork is a worst-case estimate of the amount of scan work that
560 // needs to be performed in this GC cycle. Specifically, it represents
561 // the case where *all* scannable memory turns out to be live.
562 maxScanWork := int64(scan + c.stackScan + c.globalsScan)
563 if work > scanWorkExpected {
564 // We've already done more scan work than expected. Because our expectation
565 // is based on a steady-state scannable heap size, we assume this means our
566 // heap is growing. Compute a new heap goal that takes our existing runway
567 // computed for scanWorkExpected and extrapolates it to maxScanWork, the worst-case
568 // scan work. This keeps our assist ratio stable if the heap continues to grow.
570 // The effect of this mechanism is that assists stay flat in the face of heap
571 // growths. It's OK to use more memory this cycle to scan all the live heap,
572 // because the next GC cycle is inevitably going to use *at least* that much
574 extHeapGoal := int64(float64(heapGoal-int64(c.triggered))/float64(scanWorkExpected)*float64(maxScanWork)) + int64(c.triggered)
575 scanWorkExpected = maxScanWork
577 // hardGoal is a hard limit on the amount that we're willing to push back the
578 // heap goal, and that's twice the heap goal (i.e. if GOGC=100 and the heap and/or
579 // stacks and/or globals grow to twice their size, this limits the current GC cycle's
580 // growth to 4x the original live heap's size).
582 // This maintains the invariant that we use no more memory than the next GC cycle
584 hardGoal := int64((1.0 + float64(gcPercent)/100.0) * float64(heapGoal))
585 if extHeapGoal > hardGoal {
586 extHeapGoal = hardGoal
588 heapGoal = extHeapGoal
590 if int64(live) > heapGoal {
591 // We're already past our heap goal, even the extrapolated one.
592 // Leave ourselves some extra runway, so in the worst case we
593 // finish by that point.
594 const maxOvershoot = 1.1
595 heapGoal = int64(float64(heapGoal) * maxOvershoot)
597 // Compute the upper bound on the scan work remaining.
598 scanWorkExpected = maxScanWork
601 // Compute the remaining scan work estimate.
603 // Note that we currently count allocations during GC as both
604 // scannable heap (heapScan) and scan work completed
605 // (scanWork), so allocation will change this difference
606 // slowly in the soft regime and not at all in the hard
608 scanWorkRemaining := scanWorkExpected - work
609 if scanWorkRemaining < 1000 {
610 // We set a somewhat arbitrary lower bound on
611 // remaining scan work since if we aim a little high,
612 // we can miss by a little.
614 // We *do* need to enforce that this is at least 1,
615 // since marking is racy and double-scanning objects
616 // may legitimately make the remaining scan work
617 // negative, even in the hard goal regime.
618 scanWorkRemaining = 1000
621 // Compute the heap distance remaining.
622 heapRemaining := heapGoal - int64(live)
623 if heapRemaining <= 0 {
624 // This shouldn't happen, but if it does, avoid
625 // dividing by zero or setting the assist negative.
629 // Compute the mutator assist ratio so by the time the mutator
630 // allocates the remaining heap bytes up to heapGoal, it will
631 // have done (or stolen) the remaining amount of scan work.
632 // Note that the assist ratio values are updated atomically
633 // but not together. This means there may be some degree of
634 // skew between the two values. This is generally OK as the
635 // values shift relatively slowly over the course of a GC
637 assistWorkPerByte := float64(scanWorkRemaining) / float64(heapRemaining)
638 assistBytesPerWork := float64(heapRemaining) / float64(scanWorkRemaining)
639 c.assistWorkPerByte.Store(assistWorkPerByte)
640 c.assistBytesPerWork.Store(assistBytesPerWork)
643 // endCycle computes the consMark estimate for the next cycle.
644 // userForced indicates whether the current GC cycle was forced
645 // by the application.
646 func (c *gcControllerState) endCycle(now int64, procs int, userForced bool) {
647 // Record last heap goal for the scavenger.
648 // We'll be updating the heap goal soon.
649 gcController.lastHeapGoal = c.heapGoal()
651 // Compute the duration of time for which assists were turned on.
652 assistDuration := now - c.markStartTime
654 // Assume background mark hit its utilization goal.
655 utilization := gcBackgroundUtilization
656 // Add assist utilization; avoid divide by zero.
657 if assistDuration > 0 {
658 utilization += float64(c.assistTime.Load()) / float64(assistDuration*int64(procs))
661 if c.heapLive <= c.triggered {
662 // Shouldn't happen, but let's be very safe about this in case the
663 // GC is somehow extremely short.
665 // In this case though, the only reasonable value for c.heapLive-c.triggered
666 // would be 0, which isn't really all that useful, i.e. the GC was so short
667 // that it didn't matter.
669 // Ignore this case and don't update anything.
672 idleUtilization := 0.0
673 if assistDuration > 0 {
674 idleUtilization = float64(c.idleMarkTime) / float64(assistDuration*int64(procs))
676 // Determine the cons/mark ratio.
678 // The units we want for the numerator and denominator are both B / cpu-ns.
679 // We get this by taking the bytes allocated or scanned, and divide by the amount of
680 // CPU time it took for those operations. For allocations, that CPU time is
682 // assistDuration * procs * (1 - utilization)
684 // Where utilization includes just background GC workers and assists. It does *not*
685 // include idle GC work time, because in theory the mutator is free to take that at
688 // For scanning, that CPU time is
690 // assistDuration * procs * (utilization + idleUtilization)
692 // In this case, we *include* idle utilization, because that is additional CPU time that the
693 // the GC had available to it.
695 // In effect, idle GC time is sort of double-counted here, but it's very weird compared
696 // to other kinds of GC work, because of how fluid it is. Namely, because the mutator is
697 // *always* free to take it.
699 // So this calculation is really:
700 // (heapLive-trigger) / (assistDuration * procs * (1-utilization)) /
701 // (scanWork) / (assistDuration * procs * (utilization+idleUtilization)
703 // Note that because we only care about the ratio, assistDuration and procs cancel out.
704 scanWork := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
705 currentConsMark := (float64(c.heapLive-c.triggered) * (utilization + idleUtilization)) /
706 (float64(scanWork) * (1 - utilization))
708 // Update cons/mark controller. The time period for this is 1 GC cycle.
710 // This use of a PI controller might seem strange. So, here's an explanation:
712 // currentConsMark represents the consMark we *should've* had to be perfectly
713 // on-target for this cycle. Given that we assume the next GC will be like this
714 // one in the steady-state, it stands to reason that we should just pick that
715 // as our next consMark. In practice, however, currentConsMark is too noisy:
716 // we're going to be wildly off-target in each GC cycle if we do that.
718 // What we do instead is make a long-term assumption: there is some steady-state
719 // consMark value, but it's obscured by noise. By constantly shooting for this
720 // noisy-but-perfect consMark value, the controller will bounce around a bit,
721 // but its average behavior, in aggregate, should be less noisy and closer to
722 // the true long-term consMark value, provided its tuned to be slightly overdamped.
724 oldConsMark := c.consMark
725 c.consMark, ok = c.consMarkController.next(c.consMark, currentConsMark, 1.0)
727 // The error spiraled out of control. This is incredibly unlikely seeing
728 // as this controller is essentially just a smoothing function, but it might
729 // mean that something went very wrong with how currentConsMark was calculated.
730 // Just reset consMark and keep going.
734 if debug.gcpacertrace > 0 {
735 heapGoal := c.heapGoal()
737 goal := gcGoalUtilization * 100
738 print("pacer: ", int(utilization*100), "% CPU (", int(goal), " exp.) for ")
739 print(c.heapScanWork.Load(), "+", c.stackScanWork.Load(), "+", c.globalsScanWork.Load(), " B work (", c.lastHeapScan+c.stackScan+c.globalsScan, " B exp.) ")
740 print("in ", c.triggered, " B -> ", c.heapLive, " B (∆goal ", int64(c.heapLive)-int64(heapGoal), ", cons/mark ", oldConsMark, ")")
742 print("[controller reset]")
749 // enlistWorker encourages another dedicated mark worker to start on
750 // another P if there are spare worker slots. It is used by putfull
751 // when more work is made available.
754 func (c *gcControllerState) enlistWorker() {
755 // If there are idle Ps, wake one so it will run an idle worker.
756 // NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
758 // if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
763 // There are no idle Ps. If we need more dedicated workers,
764 // try to preempt a running P so it will switch to a worker.
765 if c.dedicatedMarkWorkersNeeded <= 0 {
768 // Pick a random other P to preempt.
773 if gp == nil || gp.m == nil || gp.m.p == 0 {
776 myID := gp.m.p.ptr().id
777 for tries := 0; tries < 5; tries++ {
778 id := int32(fastrandn(uint32(gomaxprocs - 1)))
783 if p.status != _Prunning {
792 // findRunnableGCWorker returns a background mark worker for _p_ if it
793 // should be run. This must only be called when gcBlackenEnabled != 0.
794 func (c *gcControllerState) findRunnableGCWorker(_p_ *p, now int64) *g {
795 if gcBlackenEnabled == 0 {
796 throw("gcControllerState.findRunnable: blackening not enabled")
799 // Since we have the current time, check if the GC CPU limiter
800 // hasn't had an update in a while. This check is necessary in
801 // case the limiter is on but hasn't been checked in a while and
802 // so may have left sufficient headroom to turn off again.
803 if gcCPULimiter.needUpdate(now) {
804 gcCPULimiter.update(gcController.assistTime.Load(), now)
807 if !gcMarkWorkAvailable(_p_) {
808 // No work to be done right now. This can happen at
809 // the end of the mark phase when there are still
810 // assists tapering off. Don't bother running a worker
811 // now because it'll just return immediately.
815 // Grab a worker before we commit to running below.
816 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
818 // There is at least one worker per P, so normally there are
819 // enough workers to run on all Ps, if necessary. However, once
820 // a worker enters gcMarkDone it may park without rejoining the
821 // pool, thus freeing a P with no corresponding worker.
822 // gcMarkDone never depends on another worker doing work, so it
823 // is safe to simply do nothing here.
825 // If gcMarkDone bails out without completing the mark phase,
826 // it will always do so with queued global work. Thus, that P
827 // will be immediately eligible to re-run the worker G it was
828 // just using, ensuring work can complete.
832 decIfPositive := func(ptr *int64) bool {
834 v := atomic.Loadint64(ptr)
839 if atomic.Casint64(ptr, v, v-1) {
845 if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
846 // This P is now dedicated to marking until the end of
847 // the concurrent mark phase.
848 _p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
849 } else if c.fractionalUtilizationGoal == 0 {
850 // No need for fractional workers.
851 gcBgMarkWorkerPool.push(&node.node)
854 // Is this P behind on the fractional utilization
857 // This should be kept in sync with pollFractionalWorkerExit.
858 delta := now - c.markStartTime
859 if delta > 0 && float64(_p_.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
860 // Nope. No need to run a fractional worker.
861 gcBgMarkWorkerPool.push(&node.node)
864 // Run a fractional worker.
865 _p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode
868 // Run the background mark worker.
870 casgstatus(gp, _Gwaiting, _Grunnable)
877 // resetLive sets up the controller state for the next mark phase after the end
878 // of the previous one. Must be called after endCycle and before commit, before
879 // the world is started.
881 // The world must be stopped.
882 func (c *gcControllerState) resetLive(bytesMarked uint64) {
883 c.heapMarked = bytesMarked
884 c.heapLive = bytesMarked
885 c.heapScan = uint64(c.heapScanWork.Load())
886 c.lastHeapScan = uint64(c.heapScanWork.Load())
887 c.triggered = ^uint64(0) // Reset triggered.
889 // heapLive was updated, so emit a trace event.
895 // markWorkerStop must be called whenever a mark worker stops executing.
897 // It updates mark work accounting in the controller by a duration of
898 // work in nanoseconds and other bookkeeping.
900 // Safe to execute at any time.
901 func (c *gcControllerState) markWorkerStop(mode gcMarkWorkerMode, duration int64) {
903 case gcMarkWorkerDedicatedMode:
904 atomic.Xaddint64(&c.dedicatedMarkTime, duration)
905 atomic.Xaddint64(&c.dedicatedMarkWorkersNeeded, 1)
906 case gcMarkWorkerFractionalMode:
907 atomic.Xaddint64(&c.fractionalMarkTime, duration)
908 case gcMarkWorkerIdleMode:
909 atomic.Xaddint64(&c.idleMarkTime, duration)
910 c.removeIdleMarkWorker()
912 throw("markWorkerStop: unknown mark worker mode")
916 func (c *gcControllerState) update(dHeapLive, dHeapScan int64) {
918 atomic.Xadd64(&gcController.heapLive, dHeapLive)
920 // gcController.heapLive changed.
924 if gcBlackenEnabled == 0 {
925 // Update heapScan when we're not in a current GC. It is fixed
926 // at the beginning of a cycle.
928 atomic.Xadd64(&gcController.heapScan, dHeapScan)
931 // gcController.heapLive changed.
936 func (c *gcControllerState) addScannableStack(pp *p, amount int64) {
938 atomic.Xadd64(&c.scannableStackSize, amount)
941 pp.scannableStackSizeDelta += amount
942 if pp.scannableStackSizeDelta >= scannableStackSizeSlack || pp.scannableStackSizeDelta <= -scannableStackSizeSlack {
943 atomic.Xadd64(&c.scannableStackSize, pp.scannableStackSizeDelta)
944 pp.scannableStackSizeDelta = 0
948 func (c *gcControllerState) addGlobals(amount int64) {
949 atomic.Xadd64(&c.globalsScan, amount)
952 // heapGoal returns the current heap goal.
953 func (c *gcControllerState) heapGoal() uint64 {
954 goal, _ := c.heapGoalInternal()
958 // heapGoalInternal is the implementation of heapGoal which returns additional
959 // information that is necessary for computing the trigger.
961 // The returned minTrigger is always <= goal.
962 func (c *gcControllerState) heapGoalInternal() (goal, minTrigger uint64) {
963 // Start with the goal calculated for gcPercent.
964 goal = c.gcPercentHeapGoal.Load()
966 // Check if the memory-limit-based goal is smaller, and if so, pick that.
967 if newGoal := c.memoryLimitHeapGoal(); go119MemoryLimitSupport && newGoal < goal {
970 // We're not limited by the memory limit goal, so perform a series of
971 // adjustments that might move the goal forward in a variety of circumstances.
973 sweepDistTrigger := c.sweepDistMinTrigger.Load()
974 if sweepDistTrigger > goal {
975 // Set the goal to maintain a minimum sweep distance since
976 // the last call to commit. Note that we never want to do this
977 // if we're in the memory limit regime, because it could push
979 goal = sweepDistTrigger
981 // Since we ignore the sweep distance trigger in the memory
982 // limit regime, we need to ensure we don't propagate it to
983 // the trigger, because it could cause a violation of the
984 // invariant that the trigger < goal.
985 minTrigger = sweepDistTrigger
987 // Ensure that the heap goal is at least a little larger than
988 // the point at which we triggered. This may not be the case if GC
989 // start is delayed or if the allocation that pushed gcController.heapLive
990 // over trigger is large or if the trigger is really close to
991 // GOGC. Assist is proportional to this distance, so enforce a
992 // minimum distance, even if it means going over the GOGC goal
995 // Ignore this if we're in the memory limit regime: we'd prefer to
996 // have the GC respond hard about how close we are to the goal than to
997 // push the goal back in such a manner that it could cause us to exceed
999 const minRunway = 64 << 10
1000 if c.triggered != ^uint64(0) && goal < c.triggered+minRunway {
1001 goal = c.triggered + minRunway
1007 // memoryLimitHeapGoal returns a heap goal derived from memoryLimit.
1008 func (c *gcControllerState) memoryLimitHeapGoal() uint64 {
1009 // Start by pulling out some values we'll need. Be careful about overflow.
1010 var heapFree, heapAlloc, mappedReady uint64
1012 heapFree = c.heapFree.load() // Free and unscavenged memory.
1013 heapAlloc = c.totalAlloc.Load() - c.totalFree.Load() // Heap object bytes in use.
1014 mappedReady = c.mappedReady.Load() // Total unreleased mapped memory.
1015 if heapFree+heapAlloc <= mappedReady {
1018 // It is impossible for total unreleased mapped memory to exceed heap memory, but
1019 // because these stats are updated independently, we may observe a partial update
1020 // including only some values. Thus, we appear to break the invariant. However,
1021 // this condition is necessarily transient, so just try again. In the case of a
1022 // persistent accounting error, we'll deadlock here.
1025 // Below we compute a goal from memoryLimit. There are a few things to be aware of.
1026 // Firstly, the memoryLimit does not easily compare to the heap goal: the former
1027 // is total mapped memory by the runtime that hasn't been released, while the latter is
1028 // only heap object memory. Intuitively, the way we convert from one to the other is to
1029 // subtract everything from memoryLimit that both contributes to the memory limit (so,
1030 // ignore scavenged memory) and doesn't contain heap objects. This isn't quite what
1031 // lines up with reality, but it's a good starting point.
1033 // In practice this computation looks like the following:
1035 // memoryLimit - ((mappedReady - heapFree - heapAlloc) + max(mappedReady - memoryLimit, 0)) - memoryLimitHeapGoalHeadroom
1038 // Let's break this down.
1040 // The first term (marker 1) is everything that contributes to the memory limit and isn't
1041 // or couldn't become heap objects. It represents, broadly speaking, non-heap overheads.
1042 // One oddity you may have noticed is that we also subtract out heapFree, i.e. unscavenged
1043 // memory that may contain heap objects in the future.
1045 // Let's take a step back. In an ideal world, this term would look something like just
1046 // the heap goal. That is, we "reserve" enough space for the heap to grow to the heap
1047 // goal, and subtract out everything else. This is of course impossible; the defintion
1048 // is circular! However, this impossible definition contains a key insight: the amount
1049 // we're *going* to use matters just as much as whatever we're currently using.
1051 // Consider if the heap shrinks to 1/10th its size, leaving behind lots of free and
1052 // unscavenged memory. mappedReady - heapAlloc will be quite large, because of that free
1053 // and unscavenged memory, pushing the goal down significantly.
1055 // heapFree is also safe to exclude from the memory limit because in the steady-state, it's
1056 // just a pool of memory for future heap allocations, and making new allocations from heapFree
1057 // memory doesn't increase overall memory use. In transient states, the scavenger and the
1058 // allocator actively manage the pool of heapFree memory to maintain the memory limit.
1060 // The second term (marker 2) is the amount of memory we've exceeded the limit by, and is
1061 // intended to help recover from such a situation. By pushing the heap goal down, we also
1062 // push the trigger down, triggering and finishing a GC sooner in order to make room for
1063 // other memory sources. Note that since we're effectively reducing the heap goal by X bytes,
1064 // we're actually giving more than X bytes of headroom back, because the heap goal is in
1065 // terms of heap objects, but it takes more than X bytes (e.g. due to fragmentation) to store
1066 // X bytes worth of objects.
1068 // The third term (marker 3) subtracts an additional memoryLimitHeapGoalHeadroom bytes from the
1069 // heap goal. As the name implies, this is to provide additional headroom in the face of pacing
1070 // inaccuracies. This is a fixed number of bytes because these inaccuracies disproportionately
1071 // affect small heaps: as heaps get smaller, the pacer's inputs get fuzzier. Shorter GC cycles
1072 // and less GC work means noisy external factors like the OS scheduler have a greater impact.
1074 memoryLimit := uint64(c.memoryLimit.Load())
1077 nonHeapMemory := mappedReady - heapFree - heapAlloc
1081 if mappedReady > memoryLimit {
1082 overage = mappedReady - memoryLimit
1085 if nonHeapMemory+overage >= memoryLimit {
1086 // We're at a point where non-heap memory exceeds the memory limit on its own.
1087 // There's honestly not much we can do here but just trigger GCs continuously
1088 // and let the CPU limiter reign that in. Something has to give at this point.
1089 // Set it to heapMarked, the lowest possible goal.
1093 // Compute the goal.
1094 goal := memoryLimit - (nonHeapMemory + overage)
1096 // Apply some headroom to the goal to account for pacing inaccuracies.
1097 // Be careful about small limits.
1098 if goal < memoryLimitHeapGoalHeadroom || goal-memoryLimitHeapGoalHeadroom < memoryLimitHeapGoalHeadroom {
1099 goal = memoryLimitHeapGoalHeadroom
1101 goal = goal - memoryLimitHeapGoalHeadroom
1103 // Don't let us go below the live heap. A heap goal below the live heap doesn't make sense.
1104 if goal < c.heapMarked {
1111 // These constants determine the bounds on the GC trigger as a fraction
1112 // of heap bytes allocated between the start of a GC (heapLive == heapMarked)
1113 // and the end of a GC (heapLive == heapGoal).
1115 // The constants are obscured in this way for efficiency. The denominator
1116 // of the fraction is always a power-of-two for a quick division, so that
1117 // the numerator is a single constant integer multiplication.
1118 triggerRatioDen = 64
1120 // The minimum trigger constant was chosen empirically: given a sufficiently
1121 // fast/scalable allocator with 48 Ps that could drive the trigger ratio
1122 // to <0.05, this constant causes applications to retain the same peak
1123 // RSS compared to not having this allocator.
1124 minTriggerRatioNum = 45 // ~0.7
1126 // The maximum trigger constant is chosen somewhat arbitrarily, but the
1127 // current constant has served us well over the years.
1128 maxTriggerRatioNum = 61 // ~0.95
1131 // trigger returns the current point at which a GC should trigger along with
1134 // The returned value may be compared against heapLive to determine whether
1135 // the GC should trigger. Thus, the GC trigger condition should be (but may
1136 // not be, in the case of small movements for efficiency) checked whenever
1137 // the heap goal may change.
1138 func (c *gcControllerState) trigger() (uint64, uint64) {
1139 goal, minTrigger := c.heapGoalInternal()
1141 // Invariant: the trigger must always be less than the heap goal.
1143 // Note that the memory limit sets a hard maximum on our heap goal,
1144 // but the live heap may grow beyond it.
1146 if c.heapMarked >= goal {
1147 // The goal should never be smaller than heapMarked, but let's be
1148 // defensive about it. The only reasonable trigger here is one that
1149 // causes a continuous GC cycle at heapMarked, but respect the goal
1150 // if it came out as smaller than that.
1154 // Below this point, c.heapMarked < goal.
1156 // heapMarked is our absolute minumum, and it's possible the trigger
1157 // bound we get from heapGoalinternal is less than that.
1158 if minTrigger < c.heapMarked {
1159 minTrigger = c.heapMarked
1162 // If we let the trigger go too low, then if the application
1163 // is allocating very rapidly we might end up in a situation
1164 // where we're allocating black during a nearly always-on GC.
1165 // The result of this is a growing heap and ultimately an
1166 // increase in RSS. By capping us at a point >0, we're essentially
1167 // saying that we're OK using more CPU during the GC to prevent
1168 // this growth in RSS.
1169 triggerLowerBound := uint64(((goal-c.heapMarked)/triggerRatioDen)*minTriggerRatioNum) + c.heapMarked
1170 if minTrigger < triggerLowerBound {
1171 minTrigger = triggerLowerBound
1174 // For small heaps, set the max trigger point at maxTriggerRatio of the way
1175 // from the live heap to the heap goal. This ensures we always have *some*
1176 // headroom when the GC actually starts. For larger heaps, set the max trigger
1177 // point at the goal, minus the minimum heap size.
1179 // This choice follows from the fact that the minimum heap size is chosen
1180 // to reflect the costs of a GC with no work to do. With a large heap but
1181 // very little scan work to perform, this gives us exactly as much runway
1182 // as we would need, in the worst case.
1183 maxTrigger := uint64(((goal-c.heapMarked)/triggerRatioDen)*maxTriggerRatioNum) + c.heapMarked
1184 if goal > defaultHeapMinimum && goal-defaultHeapMinimum > maxTrigger {
1185 maxTrigger = goal - defaultHeapMinimum
1187 if maxTrigger < minTrigger {
1188 maxTrigger = minTrigger
1191 // Compute the trigger from our bounds and the runway stored by commit.
1193 runway := c.runway.Load()
1195 trigger = minTrigger
1197 trigger = goal - runway
1199 if trigger < minTrigger {
1200 trigger = minTrigger
1202 if trigger > maxTrigger {
1203 trigger = maxTrigger
1206 print("trigger=", trigger, " heapGoal=", goal, "\n")
1207 print("minTrigger=", minTrigger, " maxTrigger=", maxTrigger, "\n")
1208 throw("produced a trigger greater than the heap goal")
1210 return trigger, goal
1213 // commit recomputes all pacing parameters needed to derive the
1214 // trigger and the heap goal. Namely, the gcPercent-based heap goal,
1215 // and the amount of runway we want to give the GC this cycle.
1217 // This can be called any time. If GC is the in the middle of a
1218 // concurrent phase, it will adjust the pacing of that phase.
1220 // isSweepDone should be the result of calling isSweepDone(),
1221 // unless we're testing or we know we're executing during a GC cycle.
1223 // This depends on gcPercent, gcController.heapMarked, and
1224 // gcController.heapLive. These must be up to date.
1226 // Callers must call gcControllerState.revise after calling this
1227 // function if the GC is enabled.
1229 // mheap_.lock must be held or the world must be stopped.
1230 func (c *gcControllerState) commit(isSweepDone bool) {
1232 assertWorldStoppedOrLockHeld(&mheap_.lock)
1236 // The sweep is done, so there aren't any restrictions on the trigger
1237 // we need to think about.
1238 c.sweepDistMinTrigger.Store(0)
1240 // Concurrent sweep happens in the heap growth
1241 // from gcController.heapLive to trigger. Make sure we
1242 // give the sweeper some runway if it doesn't have enough.
1243 c.sweepDistMinTrigger.Store(atomic.Load64(&c.heapLive) + sweepMinHeapDistance)
1246 // Compute the next GC goal, which is when the allocated heap
1247 // has grown by GOGC/100 over where it started the last cycle,
1248 // plus additional runway for non-heap sources of GC work.
1249 gcPercentHeapGoal := ^uint64(0)
1250 if gcPercent := c.gcPercent.Load(); gcPercent >= 0 {
1251 gcPercentHeapGoal = c.heapMarked + (c.heapMarked+atomic.Load64(&c.stackScan)+atomic.Load64(&c.globalsScan))*uint64(gcPercent)/100
1253 // Apply the minimum heap size here. It's defined in terms of gcPercent
1254 // and is only updated by functions that call commit.
1255 if gcPercentHeapGoal < c.heapMinimum {
1256 gcPercentHeapGoal = c.heapMinimum
1258 c.gcPercentHeapGoal.Store(gcPercentHeapGoal)
1260 // Compute the amount of runway we want the GC to have by using our
1261 // estimate of the cons/mark ratio.
1263 // The idea is to take our expected scan work, and multiply it by
1264 // the cons/mark ratio to determine how long it'll take to complete
1265 // that scan work in terms of bytes allocated. This gives us our GC's
1268 // However, the cons/mark ratio is a ratio of rates per CPU-second, but
1269 // here we care about the relative rates for some division of CPU
1270 // resources among the mutator and the GC.
1272 // To summarize, we have B / cpu-ns, and we want B / ns. We get that
1273 // by multiplying by our desired division of CPU resources. We choose
1274 // to express CPU resources as GOMAPROCS*fraction. Note that because
1275 // we're working with a ratio here, we can omit the number of CPU cores,
1276 // because they'll appear in the numerator and denominator and cancel out.
1277 // As a result, this is basically just "weighing" the cons/mark ratio by
1278 // our desired division of resources.
1280 // Furthermore, by setting the runway so that CPU resources are divided
1281 // this way, assuming that the cons/mark ratio is correct, we make that
1282 // division a reality.
1283 c.runway.Store(uint64((c.consMark * (1 - gcGoalUtilization) / (gcGoalUtilization)) * float64(c.lastHeapScan+c.stackScan+c.globalsScan)))
1286 // setGCPercent updates gcPercent. commit must be called after.
1287 // Returns the old value of gcPercent.
1289 // The world must be stopped, or mheap_.lock must be held.
1290 func (c *gcControllerState) setGCPercent(in int32) int32 {
1292 assertWorldStoppedOrLockHeld(&mheap_.lock)
1295 out := c.gcPercent.Load()
1299 c.heapMinimum = defaultHeapMinimum * uint64(in) / 100
1300 c.gcPercent.Store(in)
1305 //go:linkname setGCPercent runtime/debug.setGCPercent
1306 func setGCPercent(in int32) (out int32) {
1307 // Run on the system stack since we grab the heap lock.
1308 systemstack(func() {
1310 out = gcController.setGCPercent(in)
1311 gcControllerCommit()
1312 unlock(&mheap_.lock)
1315 // If we just disabled GC, wait for any concurrent GC mark to
1316 // finish so we always return with no GC running.
1318 gcWaitOnMark(atomic.Load(&work.cycles))
1324 func readGOGC() int32 {
1325 p := gogetenv("GOGC")
1329 if n, ok := atoi32(p); ok {
1335 // setMemoryLimit updates memoryLimit. commit must be called after
1336 // Returns the old value of memoryLimit.
1338 // The world must be stopped, or mheap_.lock must be held.
1339 func (c *gcControllerState) setMemoryLimit(in int64) int64 {
1341 assertWorldStoppedOrLockHeld(&mheap_.lock)
1344 out := c.memoryLimit.Load()
1346 c.memoryLimit.Store(in)
1352 //go:linkname setMemoryLimit runtime/debug.setMemoryLimit
1353 func setMemoryLimit(in int64) (out int64) {
1354 // Run on the system stack since we grab the heap lock.
1355 systemstack(func() {
1357 out = gcController.setMemoryLimit(in)
1358 if in < 0 || out == in {
1359 // If we're just checking the value or not changing
1360 // it, there's no point in doing the rest.
1361 unlock(&mheap_.lock)
1364 gcControllerCommit()
1365 unlock(&mheap_.lock)
1370 func readGOMEMLIMIT() int64 {
1371 p := gogetenv("GOMEMLIMIT")
1372 if p == "" || p == "off" {
1375 n, ok := parseByteCount(p)
1377 print("GOMEMLIMIT=", p, "\n")
1378 throw("malformed GOMEMLIMIT; see `go doc runtime/debug.SetMemoryLimit`")
1383 type piController struct {
1384 kp float64 // Proportional constant.
1385 ti float64 // Integral time constant.
1386 tt float64 // Reset time.
1388 min, max float64 // Output boundaries.
1390 // PI controller state.
1392 errIntegral float64 // Integral of the error from t=0 to now.
1395 errOverflow bool // Set if errIntegral ever overflowed.
1396 inputOverflow bool // Set if an operation with the input overflowed.
1399 // next provides a new sample to the controller.
1401 // input is the sample, setpoint is the desired point, and period is how much
1402 // time (in whatever unit makes the most sense) has passed since the last sample.
1404 // Returns a new value for the variable it's controlling, and whether the operation
1405 // completed successfully. One reason this might fail is if error has been growing
1406 // in an unbounded manner, to the point of overflow.
1408 // In the specific case of an error overflow occurs, the errOverflow field will be
1409 // set and the rest of the controller's internal state will be fully reset.
1410 func (c *piController) next(input, setpoint, period float64) (float64, bool) {
1411 // Compute the raw output value.
1412 prop := c.kp * (setpoint - input)
1413 rawOutput := prop + c.errIntegral
1415 // Clamp rawOutput into output.
1417 if isInf(output) || isNaN(output) {
1418 // The input had a large enough magnitude that either it was already
1419 // overflowed, or some operation with it overflowed.
1420 // Set a flag and reset. That's the safest thing to do.
1422 c.inputOverflow = true
1427 } else if output > c.max {
1431 // Update the controller's state.
1432 if c.ti != 0 && c.tt != 0 {
1433 c.errIntegral += (c.kp*period/c.ti)*(setpoint-input) + (period/c.tt)*(output-rawOutput)
1434 if isInf(c.errIntegral) || isNaN(c.errIntegral) {
1435 // So much error has accumulated that we managed to overflow.
1436 // The assumptions around the controller have likely broken down.
1437 // Set a flag and reset. That's the safest thing to do.
1439 c.errOverflow = true
1446 // reset resets the controller state, except for controller error flags.
1447 func (c *piController) reset() {
1451 // addIdleMarkWorker attempts to add a new idle mark worker.
1453 // If this returns true, the caller must become an idle mark worker unless
1454 // there's no background mark worker goroutines in the pool. This case is
1455 // harmless because there are already background mark workers running.
1456 // If this returns false, the caller must NOT become an idle mark worker.
1458 // nosplit because it may be called without a P.
1460 func (c *gcControllerState) addIdleMarkWorker() bool {
1462 old := c.idleMarkWorkers.Load()
1463 n, max := int32(old&uint64(^uint32(0))), int32(old>>32)
1465 // See the comment on idleMarkWorkers for why
1466 // n > max is tolerated.
1470 print("n=", n, " max=", max, "\n")
1471 throw("negative idle mark workers")
1473 new := uint64(uint32(n+1)) | (uint64(max) << 32)
1474 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1480 // needIdleMarkWorker is a hint as to whether another idle mark worker is needed.
1482 // The caller must still call addIdleMarkWorker to become one. This is mainly
1483 // useful for a quick check before an expensive operation.
1485 // nosplit because it may be called without a P.
1487 func (c *gcControllerState) needIdleMarkWorker() bool {
1488 p := c.idleMarkWorkers.Load()
1489 n, max := int32(p&uint64(^uint32(0))), int32(p>>32)
1493 // removeIdleMarkWorker must be called when an new idle mark worker stops executing.
1494 func (c *gcControllerState) removeIdleMarkWorker() {
1496 old := c.idleMarkWorkers.Load()
1497 n, max := int32(old&uint64(^uint32(0))), int32(old>>32)
1499 print("n=", n, " max=", max, "\n")
1500 throw("negative idle mark workers")
1502 new := uint64(uint32(n-1)) | (uint64(max) << 32)
1503 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1509 // setMaxIdleMarkWorkers sets the maximum number of idle mark workers allowed.
1511 // This method is optimistic in that it does not wait for the number of
1512 // idle mark workers to reduce to max before returning; it assumes the workers
1513 // will deschedule themselves.
1514 func (c *gcControllerState) setMaxIdleMarkWorkers(max int32) {
1516 old := c.idleMarkWorkers.Load()
1517 n := int32(old & uint64(^uint32(0)))
1519 print("n=", n, " max=", max, "\n")
1520 throw("negative idle mark workers")
1522 new := uint64(uint32(n)) | (uint64(max) << 32)
1523 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1529 // gcControllerCommit is gcController.commit, but passes arguments from live
1530 // (non-test) data. It also updates any consumers of the GC pacing, such as
1531 // sweep pacing and the background scavenger.
1533 // Calls gcController.commit.
1535 // The heap lock must be held, so this must be executed on the system stack.
1538 func gcControllerCommit() {
1539 assertWorldStoppedOrLockHeld(&mheap_.lock)
1541 gcController.commit(isSweepDone())
1543 // Update mark pacing.
1544 if gcphase != _GCoff {
1545 gcController.revise()
1548 // TODO(mknyszek): This isn't really accurate any longer because the heap
1549 // goal is computed dynamically. Still useful to snapshot, but not as useful.
1554 trigger, heapGoal := gcController.trigger()
1555 gcPaceSweeper(trigger)
1556 gcPaceScavenger(gcController.memoryLimit.Load(), heapGoal, gcController.lastHeapGoal)