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"
11 _ "unsafe" // for go:linkname
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 // maxStackScanSlack is the bytes of stack space allocated or freed
67 // that can accumulate on a P before updating gcController.stackSize.
68 maxStackScanSlack = 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
77 // gcController implements the GC pacing controller that determines
78 // when to trigger concurrent garbage collection and how much marking
79 // work to do in mutator assists and background marking.
81 // It calculates the ratio between the allocation rate (in terms of CPU
82 // time) and the GC scan throughput to determine the heap size at which to
83 // trigger a GC cycle such that no GC assists are required to finish on time.
84 // This algorithm thus optimizes GC CPU utilization to the dedicated background
85 // mark utilization of 25% of GOMAXPROCS by minimizing GC assists.
86 // GOMAXPROCS. The high-level design of this algorithm is documented
87 // at https://github.com/golang/proposal/blob/master/design/44167-gc-pacer-redesign.md.
88 // See https://golang.org/s/go15gcpacing for additional historical context.
89 var gcController gcControllerState
91 type gcControllerState struct {
92 // Initialized from GOGC. GOGC=off means no GC.
93 gcPercent atomic.Int32
95 // memoryLimit is the soft memory limit in bytes.
97 // Initialized from GOMEMLIMIT. GOMEMLIMIT=off is equivalent to MaxInt64
98 // which means no soft memory limit in practice.
100 // This is an int64 instead of a uint64 to more easily maintain parity with
101 // the SetMemoryLimit API, which sets a maximum at MaxInt64. This value
102 // should never be negative.
103 memoryLimit atomic.Int64
105 // heapMinimum is the minimum heap size at which to trigger GC.
106 // For small heaps, this overrides the usual GOGC*live set rule.
108 // When there is a very small live set but a lot of allocation, simply
109 // collecting when the heap reaches GOGC*live results in many GC
110 // cycles and high total per-GC overhead. This minimum amortizes this
111 // per-GC overhead while keeping the heap reasonably small.
113 // During initialization this is set to 4MB*GOGC/100. In the case of
114 // GOGC==0, this will set heapMinimum to 0, resulting in constant
115 // collection even when the heap size is small, which is useful for
119 // runway is the amount of runway in heap bytes allocated by the
120 // application that we want to give the GC once it starts.
122 // This is computed from consMark during mark termination.
125 // consMark is the estimated per-CPU consMark ratio for the application.
127 // It represents the ratio between the application's allocation
128 // rate, as bytes allocated per CPU-time, and the GC's scan rate,
129 // as bytes scanned per CPU-time.
130 // The units of this ratio are (B / cpu-ns) / (B / cpu-ns).
132 // At a high level, this value is computed as the bytes of memory
133 // allocated (cons) per unit of scan work completed (mark) in a GC
134 // cycle, divided by the CPU time spent on each activity.
136 // Updated at the end of each GC cycle, in endCycle.
139 // consMarkController holds the state for the mark-cons ratio
140 // estimation over time.
142 // Its purpose is to smooth out noisiness in the computation of
143 // consMark; see consMark for details.
144 consMarkController piController
146 // gcPercentHeapGoal is the goal heapLive for when next GC ends derived
149 // Set to ^uint64(0) if gcPercent is disabled.
150 gcPercentHeapGoal atomic.Uint64
152 // sweepDistMinTrigger is the minimum trigger to ensure a minimum
155 // This bound is also special because it applies to both the trigger
156 // *and* the goal (all other trigger bounds must be based *on* the goal).
158 // It is computed ahead of time, at commit time. The theory is that,
159 // absent a sudden change to a parameter like gcPercent, the trigger
160 // will be chosen to always give the sweeper enough headroom. However,
161 // such a change might dramatically and suddenly move up the trigger,
162 // in which case we need to ensure the sweeper still has enough headroom.
163 sweepDistMinTrigger atomic.Uint64
165 // triggered is the point at which the current GC cycle actually triggered.
166 // Only valid during the mark phase of a GC cycle, otherwise set to ^uint64(0).
168 // Updated while the world is stopped.
171 // lastHeapGoal is the value of heapGoal at the moment the last GC
172 // ended. Note that this is distinct from the last value heapGoal had,
173 // because it could change if e.g. gcPercent changes.
175 // Read and written with the world stopped or with mheap_.lock held.
178 // heapLive is the number of bytes considered live by the GC.
179 // That is: retained by the most recent GC plus allocated
180 // since then. heapLive ≤ memstats.totalAlloc-memstats.totalFree, since
181 // heapAlloc includes unmarked objects that have not yet been swept (and
182 // hence goes up as we allocate and down as we sweep) while heapLive
183 // excludes these objects (and hence only goes up between GCs).
185 // To reduce contention, this is updated only when obtaining a span
186 // from an mcentral and at this point it counts all of the unallocated
187 // slots in that span (which will be allocated before that mcache
188 // obtains another span from that mcentral). Hence, it slightly
189 // overestimates the "true" live heap size. It's better to overestimate
190 // than to underestimate because 1) this triggers the GC earlier than
191 // necessary rather than potentially too late and 2) this leads to a
192 // conservative GC rate rather than a GC rate that is potentially too
195 // Whenever this is updated, call traceHeapAlloc() and
196 // this gcControllerState's revise() method.
197 heapLive atomic.Uint64
199 // heapScan is the number of bytes of "scannable" heap. This is the
200 // live heap (as counted by heapLive), but omitting no-scan objects and
201 // no-scan tails of objects.
203 // This value is fixed at the start of a GC cycle. It represents the
204 // maximum scannable heap.
205 heapScan atomic.Uint64
207 // lastHeapScan is the number of bytes of heap that were scanned
208 // last GC cycle. It is the same as heapMarked, but only
209 // includes the "scannable" parts of objects.
211 // Updated when the world is stopped.
214 // lastStackScan is the number of bytes of stack that were scanned
216 lastStackScan atomic.Uint64
218 // maxStackScan is the amount of allocated goroutine stack space in
219 // use by goroutines.
221 // This number tracks allocated goroutine stack space rather than used
222 // goroutine stack space (i.e. what is actually scanned) because used
223 // goroutine stack space is much harder to measure cheaply. By using
224 // allocated space, we make an overestimate; this is OK, it's better
225 // to conservatively overcount than undercount.
226 maxStackScan atomic.Uint64
228 // globalsScan is the total amount of global variable space
229 // that is scannable.
230 globalsScan atomic.Uint64
232 // heapMarked is the number of bytes marked by the previous
233 // GC. After mark termination, heapLive == heapMarked, but
234 // unlike heapLive, heapMarked does not change until the
235 // next mark termination.
238 // heapScanWork is the total heap scan work performed this cycle.
239 // stackScanWork is the total stack scan work performed this cycle.
240 // globalsScanWork is the total globals scan work performed this cycle.
242 // These are updated atomically during the cycle. Updates occur in
243 // bounded batches, since they are both written and read
244 // throughout the cycle. At the end of the cycle, heapScanWork is how
245 // much of the retained heap is scannable.
247 // Currently these are measured in bytes. For most uses, this is an
248 // opaque unit of work, but for estimation the definition is important.
250 // Note that stackScanWork includes only stack space scanned, not all
251 // of the allocated stack.
252 heapScanWork atomic.Int64
253 stackScanWork atomic.Int64
254 globalsScanWork atomic.Int64
256 // bgScanCredit is the scan work credit accumulated by the concurrent
257 // background scan. This credit is accumulated by the background scan
258 // and stolen by mutator assists. Updates occur in bounded batches,
259 // since it is both written and read throughout the cycle.
260 bgScanCredit atomic.Int64
262 // assistTime is the nanoseconds spent in mutator assists
263 // during this cycle. This is updated atomically, and must also
264 // be updated atomically even during a STW, because it is read
265 // by sysmon. Updates occur in bounded batches, since it is both
266 // written and read throughout the cycle.
267 assistTime atomic.Int64
269 // dedicatedMarkTime is the nanoseconds spent in dedicated mark workers
270 // during this cycle. This is updated at the end of the concurrent mark
272 dedicatedMarkTime atomic.Int64
274 // fractionalMarkTime is the nanoseconds spent in the fractional mark
275 // worker during this cycle. This is updated throughout the cycle and
276 // will be up-to-date if the fractional mark worker is not currently
278 fractionalMarkTime atomic.Int64
280 // idleMarkTime is the nanoseconds spent in idle marking during this
281 // cycle. This is updated throughout the cycle.
282 idleMarkTime atomic.Int64
284 // markStartTime is the absolute start time in nanoseconds
285 // that assists and background mark workers started.
288 // dedicatedMarkWorkersNeeded is the number of dedicated mark workers
289 // that need to be started. This is computed at the beginning of each
290 // cycle and decremented as dedicated mark workers get started.
291 dedicatedMarkWorkersNeeded atomic.Int64
293 // idleMarkWorkers is two packed int32 values in a single uint64.
294 // These two values are always updated simultaneously.
296 // The bottom int32 is the current number of idle mark workers executing.
298 // The top int32 is the maximum number of idle mark workers allowed to
299 // execute concurrently. Normally, this number is just gomaxprocs. However,
300 // during periodic GC cycles it is set to 0 because the system is idle
301 // anyway; there's no need to go full blast on all of GOMAXPROCS.
303 // The maximum number of idle mark workers is used to prevent new workers
304 // from starting, but it is not a hard maximum. It is possible (but
305 // exceedingly rare) for the current number of idle mark workers to
306 // transiently exceed the maximum. This could happen if the maximum changes
307 // just after a GC ends, and an M with no P.
309 // Note that if we have no dedicated mark workers, we set this value to
310 // 1 in this case we only have fractional GC workers which aren't scheduled
311 // strictly enough to ensure GC progress. As a result, idle-priority mark
312 // workers are vital to GC progress in these situations.
314 // For example, consider a situation in which goroutines block on the GC
315 // (such as via runtime.GOMAXPROCS) and only fractional mark workers are
316 // scheduled (e.g. GOMAXPROCS=1). Without idle-priority mark workers, the
317 // last running M might skip scheduling a fractional mark worker if its
318 // utilization goal is met, such that once it goes to sleep (because there's
319 // nothing to do), there will be nothing else to spin up a new M for the
320 // fractional worker in the future, stalling GC progress and causing a
321 // deadlock. However, idle-priority workers will *always* run when there is
322 // nothing left to do, ensuring the GC makes progress.
324 // See github.com/golang/go/issues/44163 for more details.
325 idleMarkWorkers atomic.Uint64
327 // assistWorkPerByte is the ratio of scan work to allocated
328 // bytes that should be performed by mutator assists. This is
329 // computed at the beginning of each cycle and updated every
330 // time heapScan is updated.
331 assistWorkPerByte atomic.Float64
333 // assistBytesPerWork is 1/assistWorkPerByte.
335 // Note that because this is read and written independently
336 // from assistWorkPerByte users may notice a skew between
337 // the two values, and such a state should be safe.
338 assistBytesPerWork atomic.Float64
340 // fractionalUtilizationGoal is the fraction of wall clock
341 // time that should be spent in the fractional mark worker on
342 // each P that isn't running a dedicated worker.
344 // For example, if the utilization goal is 25% and there are
345 // no dedicated workers, this will be 0.25. If the goal is
346 // 25%, there is one dedicated worker, and GOMAXPROCS is 5,
347 // this will be 0.05 to make up the missing 5%.
349 // If this is zero, no fractional workers are needed.
350 fractionalUtilizationGoal float64
352 // These memory stats are effectively duplicates of fields from
353 // memstats.heapStats but are updated atomically or with the world
354 // stopped and don't provide the same consistency guarantees.
356 // Because the runtime is responsible for managing a memory limit, it's
357 // useful to couple these stats more tightly to the gcController, which
358 // is intimately connected to how that memory limit is maintained.
359 heapInUse sysMemStat // bytes in mSpanInUse spans
360 heapReleased sysMemStat // bytes released to the OS
361 heapFree sysMemStat // bytes not in any span, but not released to the OS
362 totalAlloc atomic.Uint64 // total bytes allocated
363 totalFree atomic.Uint64 // total bytes freed
364 mappedReady atomic.Uint64 // total virtual memory in the Ready state (see mem.go).
366 // test indicates that this is a test-only copy of gcControllerState.
372 func (c *gcControllerState) init(gcPercent int32, memoryLimit int64) {
373 c.heapMinimum = defaultHeapMinimum
374 c.triggered = ^uint64(0)
376 c.consMarkController = piController{
377 // Tuned first via the Ziegler-Nichols process in simulation,
378 // then the integral time was manually tuned against real-world
379 // applications to deal with noisiness in the measured cons/mark
384 // Set a high reset time in GC cycles.
385 // This is inversely proportional to the rate at which we
386 // accumulate error from clipping. By making this very high
387 // we make the accumulation slow. In general, clipping is
388 // OK in our situation, hence the choice.
390 // Tune this if we get unintended effects from clipping for
397 c.setGCPercent(gcPercent)
398 c.setMemoryLimit(memoryLimit)
399 c.commit(true) // No sweep phase in the first GC cycle.
400 // N.B. Don't bother calling traceHeapGoal. Tracing is never enabled at
401 // initialization time.
402 // N.B. No need to call revise; there's no GC enabled during
406 // startCycle resets the GC controller's state and computes estimates
407 // for a new GC cycle. The caller must hold worldsema and the world
409 func (c *gcControllerState) startCycle(markStartTime int64, procs int, trigger gcTrigger) {
410 c.heapScanWork.Store(0)
411 c.stackScanWork.Store(0)
412 c.globalsScanWork.Store(0)
413 c.bgScanCredit.Store(0)
414 c.assistTime.Store(0)
415 c.dedicatedMarkTime.Store(0)
416 c.fractionalMarkTime.Store(0)
417 c.idleMarkTime.Store(0)
418 c.markStartTime = markStartTime
420 // TODO(mknyszek): This is supposed to be the actual trigger point for the heap, but
421 // causes regressions in memory use. The cause is that the PI controller used to smooth
422 // the cons/mark ratio measurements tends to flail when using the less accurate precomputed
423 // trigger for the cons/mark calculation, and this results in the controller being more
424 // conservative about steady-states it tries to find in the future.
426 // This conservatism is transient, but these transient states tend to matter for short-lived
427 // programs, especially because the PI controller is overdamped, partially because it is
428 // configured with a relatively large time constant.
430 // Ultimately, I think this is just two mistakes piled on one another: the choice of a swingy
431 // smoothing function that recalls a fairly long history (due to its overdamped time constant)
432 // coupled with an inaccurate cons/mark calculation. It just so happens this works better
433 // today, and it makes it harder to change things in the future.
435 // This is described in #53738. Fix this for #53892 by changing back to the actual trigger
436 // point and simplifying the smoothing function.
437 heapTrigger, heapGoal := c.trigger()
438 c.triggered = heapTrigger
440 // Compute the background mark utilization goal. In general,
441 // this may not come out exactly. We round the number of
442 // dedicated workers so that the utilization is closest to
443 // 25%. For small GOMAXPROCS, this would introduce too much
444 // error, so we add fractional workers in that case.
445 totalUtilizationGoal := float64(procs) * gcBackgroundUtilization
446 dedicatedMarkWorkersNeeded := int64(totalUtilizationGoal + 0.5)
447 utilError := float64(dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
448 const maxUtilError = 0.3
449 if utilError < -maxUtilError || utilError > maxUtilError {
450 // Rounding put us more than 30% off our goal. With
451 // gcBackgroundUtilization of 25%, this happens for
452 // GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
453 // workers to compensate.
454 if float64(dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
455 // Too many dedicated workers.
456 dedicatedMarkWorkersNeeded--
458 c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(dedicatedMarkWorkersNeeded)) / float64(procs)
460 c.fractionalUtilizationGoal = 0
463 // In STW mode, we just want dedicated workers.
464 if debug.gcstoptheworld > 0 {
465 dedicatedMarkWorkersNeeded = int64(procs)
466 c.fractionalUtilizationGoal = 0
470 for _, p := range allp {
472 p.gcFractionalMarkTime = 0
475 if trigger.kind == gcTriggerTime {
476 // During a periodic GC cycle, reduce the number of idle mark workers
477 // required. However, we need at least one dedicated mark worker or
478 // idle GC worker to ensure GC progress in some scenarios (see comment
479 // on maxIdleMarkWorkers).
480 if dedicatedMarkWorkersNeeded > 0 {
481 c.setMaxIdleMarkWorkers(0)
483 // TODO(mknyszek): The fundamental reason why we need this is because
484 // we can't count on the fractional mark worker to get scheduled.
485 // Fix that by ensuring it gets scheduled according to its quota even
486 // if the rest of the application is idle.
487 c.setMaxIdleMarkWorkers(1)
490 // N.B. gomaxprocs and dedicatedMarkWorkersNeeded are guaranteed not to
491 // change during a GC cycle.
492 c.setMaxIdleMarkWorkers(int32(procs) - int32(dedicatedMarkWorkersNeeded))
495 // Compute initial values for controls that are updated
496 // throughout the cycle.
497 c.dedicatedMarkWorkersNeeded.Store(dedicatedMarkWorkersNeeded)
500 if debug.gcpacertrace > 0 {
501 assistRatio := c.assistWorkPerByte.Load()
502 print("pacer: assist ratio=", assistRatio,
503 " (scan ", gcController.heapScan.Load()>>20, " MB in ",
504 work.initialHeapLive>>20, "->",
505 heapGoal>>20, " MB)",
506 " workers=", dedicatedMarkWorkersNeeded,
507 "+", c.fractionalUtilizationGoal, "\n")
511 // revise updates the assist ratio during the GC cycle to account for
512 // improved estimates. This should be called whenever gcController.heapScan,
513 // gcController.heapLive, or if any inputs to gcController.heapGoal are
514 // updated. It is safe to call concurrently, but it may race with other
517 // The result of this race is that the two assist ratio values may not line
518 // up or may be stale. In practice this is OK because the assist ratio
519 // moves slowly throughout a GC cycle, and the assist ratio is a best-effort
520 // heuristic anyway. Furthermore, no part of the heuristic depends on
521 // the two assist ratio values being exact reciprocals of one another, since
522 // the two values are used to convert values from different sources.
524 // The worst case result of this raciness is that we may miss a larger shift
525 // in the ratio (say, if we decide to pace more aggressively against the
526 // hard heap goal) but even this "hard goal" is best-effort (see #40460).
527 // The dedicated GC should ensure we don't exceed the hard goal by too much
528 // in the rare case we do exceed it.
530 // It should only be called when gcBlackenEnabled != 0 (because this
531 // is when assists are enabled and the necessary statistics are
533 func (c *gcControllerState) revise() {
534 gcPercent := c.gcPercent.Load()
536 // If GC is disabled but we're running a forced GC,
537 // act like GOGC is huge for the below calculations.
540 live := c.heapLive.Load()
541 scan := c.heapScan.Load()
542 work := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
544 // Assume we're under the soft goal. Pace GC to complete at
545 // heapGoal assuming the heap is in steady-state.
546 heapGoal := int64(c.heapGoal())
548 // The expected scan work is computed as the amount of bytes scanned last
549 // GC cycle (both heap and stack), plus our estimate of globals work for this cycle.
550 scanWorkExpected := int64(c.lastHeapScan + c.lastStackScan.Load() + c.globalsScan.Load())
552 // maxScanWork is a worst-case estimate of the amount of scan work that
553 // needs to be performed in this GC cycle. Specifically, it represents
554 // the case where *all* scannable memory turns out to be live, and
555 // *all* allocated stack space is scannable.
556 maxStackScan := c.maxStackScan.Load()
557 maxScanWork := int64(scan + maxStackScan + c.globalsScan.Load())
558 if work > scanWorkExpected {
559 // We've already done more scan work than expected. Because our expectation
560 // is based on a steady-state scannable heap size, we assume this means our
561 // heap is growing. Compute a new heap goal that takes our existing runway
562 // computed for scanWorkExpected and extrapolates it to maxScanWork, the worst-case
563 // scan work. This keeps our assist ratio stable if the heap continues to grow.
565 // The effect of this mechanism is that assists stay flat in the face of heap
566 // growths. It's OK to use more memory this cycle to scan all the live heap,
567 // because the next GC cycle is inevitably going to use *at least* that much
569 extHeapGoal := int64(float64(heapGoal-int64(c.triggered))/float64(scanWorkExpected)*float64(maxScanWork)) + int64(c.triggered)
570 scanWorkExpected = maxScanWork
572 // hardGoal is a hard limit on the amount that we're willing to push back the
573 // heap goal, and that's twice the heap goal (i.e. if GOGC=100 and the heap and/or
574 // stacks and/or globals grow to twice their size, this limits the current GC cycle's
575 // growth to 4x the original live heap's size).
577 // This maintains the invariant that we use no more memory than the next GC cycle
579 hardGoal := int64((1.0 + float64(gcPercent)/100.0) * float64(heapGoal))
580 if extHeapGoal > hardGoal {
581 extHeapGoal = hardGoal
583 heapGoal = extHeapGoal
585 if int64(live) > heapGoal {
586 // We're already past our heap goal, even the extrapolated one.
587 // Leave ourselves some extra runway, so in the worst case we
588 // finish by that point.
589 const maxOvershoot = 1.1
590 heapGoal = int64(float64(heapGoal) * maxOvershoot)
592 // Compute the upper bound on the scan work remaining.
593 scanWorkExpected = maxScanWork
596 // Compute the remaining scan work estimate.
598 // Note that we currently count allocations during GC as both
599 // scannable heap (heapScan) and scan work completed
600 // (scanWork), so allocation will change this difference
601 // slowly in the soft regime and not at all in the hard
603 scanWorkRemaining := scanWorkExpected - work
604 if scanWorkRemaining < 1000 {
605 // We set a somewhat arbitrary lower bound on
606 // remaining scan work since if we aim a little high,
607 // we can miss by a little.
609 // We *do* need to enforce that this is at least 1,
610 // since marking is racy and double-scanning objects
611 // may legitimately make the remaining scan work
612 // negative, even in the hard goal regime.
613 scanWorkRemaining = 1000
616 // Compute the heap distance remaining.
617 heapRemaining := heapGoal - int64(live)
618 if heapRemaining <= 0 {
619 // This shouldn't happen, but if it does, avoid
620 // dividing by zero or setting the assist negative.
624 // Compute the mutator assist ratio so by the time the mutator
625 // allocates the remaining heap bytes up to heapGoal, it will
626 // have done (or stolen) the remaining amount of scan work.
627 // Note that the assist ratio values are updated atomically
628 // but not together. This means there may be some degree of
629 // skew between the two values. This is generally OK as the
630 // values shift relatively slowly over the course of a GC
632 assistWorkPerByte := float64(scanWorkRemaining) / float64(heapRemaining)
633 assistBytesPerWork := float64(heapRemaining) / float64(scanWorkRemaining)
634 c.assistWorkPerByte.Store(assistWorkPerByte)
635 c.assistBytesPerWork.Store(assistBytesPerWork)
638 // endCycle computes the consMark estimate for the next cycle.
639 // userForced indicates whether the current GC cycle was forced
640 // by the application.
641 func (c *gcControllerState) endCycle(now int64, procs int, userForced bool) {
642 // Record last heap goal for the scavenger.
643 // We'll be updating the heap goal soon.
644 gcController.lastHeapGoal = c.heapGoal()
646 // Compute the duration of time for which assists were turned on.
647 assistDuration := now - c.markStartTime
649 // Assume background mark hit its utilization goal.
650 utilization := gcBackgroundUtilization
651 // Add assist utilization; avoid divide by zero.
652 if assistDuration > 0 {
653 utilization += float64(c.assistTime.Load()) / float64(assistDuration*int64(procs))
656 if c.heapLive.Load() <= c.triggered {
657 // Shouldn't happen, but let's be very safe about this in case the
658 // GC is somehow extremely short.
660 // In this case though, the only reasonable value for c.heapLive-c.triggered
661 // would be 0, which isn't really all that useful, i.e. the GC was so short
662 // that it didn't matter.
664 // Ignore this case and don't update anything.
667 idleUtilization := 0.0
668 if assistDuration > 0 {
669 idleUtilization = float64(c.idleMarkTime.Load()) / float64(assistDuration*int64(procs))
671 // Determine the cons/mark ratio.
673 // The units we want for the numerator and denominator are both B / cpu-ns.
674 // We get this by taking the bytes allocated or scanned, and divide by the amount of
675 // CPU time it took for those operations. For allocations, that CPU time is
677 // assistDuration * procs * (1 - utilization)
679 // Where utilization includes just background GC workers and assists. It does *not*
680 // include idle GC work time, because in theory the mutator is free to take that at
683 // For scanning, that CPU time is
685 // assistDuration * procs * (utilization + idleUtilization)
687 // In this case, we *include* idle utilization, because that is additional CPU time that the
688 // the GC had available to it.
690 // In effect, idle GC time is sort of double-counted here, but it's very weird compared
691 // to other kinds of GC work, because of how fluid it is. Namely, because the mutator is
692 // *always* free to take it.
694 // So this calculation is really:
695 // (heapLive-trigger) / (assistDuration * procs * (1-utilization)) /
696 // (scanWork) / (assistDuration * procs * (utilization+idleUtilization)
698 // Note that because we only care about the ratio, assistDuration and procs cancel out.
699 scanWork := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
700 currentConsMark := (float64(c.heapLive.Load()-c.triggered) * (utilization + idleUtilization)) /
701 (float64(scanWork) * (1 - utilization))
703 // Update cons/mark controller. The time period for this is 1 GC cycle.
705 // This use of a PI controller might seem strange. So, here's an explanation:
707 // currentConsMark represents the consMark we *should've* had to be perfectly
708 // on-target for this cycle. Given that we assume the next GC will be like this
709 // one in the steady-state, it stands to reason that we should just pick that
710 // as our next consMark. In practice, however, currentConsMark is too noisy:
711 // we're going to be wildly off-target in each GC cycle if we do that.
713 // What we do instead is make a long-term assumption: there is some steady-state
714 // consMark value, but it's obscured by noise. By constantly shooting for this
715 // noisy-but-perfect consMark value, the controller will bounce around a bit,
716 // but its average behavior, in aggregate, should be less noisy and closer to
717 // the true long-term consMark value, provided its tuned to be slightly overdamped.
719 oldConsMark := c.consMark
720 c.consMark, ok = c.consMarkController.next(c.consMark, currentConsMark, 1.0)
722 // The error spiraled out of control. This is incredibly unlikely seeing
723 // as this controller is essentially just a smoothing function, but it might
724 // mean that something went very wrong with how currentConsMark was calculated.
725 // Just reset consMark and keep going.
729 if debug.gcpacertrace > 0 {
731 goal := gcGoalUtilization * 100
732 print("pacer: ", int(utilization*100), "% CPU (", int(goal), " exp.) for ")
733 print(c.heapScanWork.Load(), "+", c.stackScanWork.Load(), "+", c.globalsScanWork.Load(), " B work (", c.lastHeapScan+c.lastStackScan.Load()+c.globalsScan.Load(), " B exp.) ")
734 live := c.heapLive.Load()
735 print("in ", c.triggered, " B -> ", live, " B (∆goal ", int64(live)-int64(c.lastHeapGoal), ", cons/mark ", oldConsMark, ")")
737 print("[controller reset]")
744 // enlistWorker encourages another dedicated mark worker to start on
745 // another P if there are spare worker slots. It is used by putfull
746 // when more work is made available.
749 func (c *gcControllerState) enlistWorker() {
750 // If there are idle Ps, wake one so it will run an idle worker.
751 // NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
753 // if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
758 // There are no idle Ps. If we need more dedicated workers,
759 // try to preempt a running P so it will switch to a worker.
760 if c.dedicatedMarkWorkersNeeded.Load() <= 0 {
763 // Pick a random other P to preempt.
768 if gp == nil || gp.m == nil || gp.m.p == 0 {
771 myID := gp.m.p.ptr().id
772 for tries := 0; tries < 5; tries++ {
773 id := int32(fastrandn(uint32(gomaxprocs - 1)))
778 if p.status != _Prunning {
787 // findRunnableGCWorker returns a background mark worker for pp if it
788 // should be run. This must only be called when gcBlackenEnabled != 0.
789 func (c *gcControllerState) findRunnableGCWorker(pp *p, now int64) (*g, int64) {
790 if gcBlackenEnabled == 0 {
791 throw("gcControllerState.findRunnable: blackening not enabled")
794 // Since we have the current time, check if the GC CPU limiter
795 // hasn't had an update in a while. This check is necessary in
796 // case the limiter is on but hasn't been checked in a while and
797 // so may have left sufficient headroom to turn off again.
801 if gcCPULimiter.needUpdate(now) {
802 gcCPULimiter.update(now)
805 if !gcMarkWorkAvailable(pp) {
806 // No work to be done right now. This can happen at
807 // the end of the mark phase when there are still
808 // assists tapering off. Don't bother running a worker
809 // now because it'll just return immediately.
813 // Grab a worker before we commit to running below.
814 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
816 // There is at least one worker per P, so normally there are
817 // enough workers to run on all Ps, if necessary. However, once
818 // a worker enters gcMarkDone it may park without rejoining the
819 // pool, thus freeing a P with no corresponding worker.
820 // gcMarkDone never depends on another worker doing work, so it
821 // is safe to simply do nothing here.
823 // If gcMarkDone bails out without completing the mark phase,
824 // it will always do so with queued global work. Thus, that P
825 // will be immediately eligible to re-run the worker G it was
826 // just using, ensuring work can complete.
830 decIfPositive := func(val *atomic.Int64) bool {
837 if val.CompareAndSwap(v, v-1) {
843 if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
844 // This P is now dedicated to marking until the end of
845 // the concurrent mark phase.
846 pp.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
847 } else if c.fractionalUtilizationGoal == 0 {
848 // No need for fractional workers.
849 gcBgMarkWorkerPool.push(&node.node)
852 // Is this P behind on the fractional utilization
855 // This should be kept in sync with pollFractionalWorkerExit.
856 delta := now - c.markStartTime
857 if delta > 0 && float64(pp.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
858 // Nope. No need to run a fractional worker.
859 gcBgMarkWorkerPool.push(&node.node)
862 // Run a fractional worker.
863 pp.gcMarkWorkerMode = gcMarkWorkerFractionalMode
866 // Run the background mark worker.
868 casgstatus(gp, _Gwaiting, _Grunnable)
875 // resetLive sets up the controller state for the next mark phase after the end
876 // of the previous one. Must be called after endCycle and before commit, before
877 // the world is started.
879 // The world must be stopped.
880 func (c *gcControllerState) resetLive(bytesMarked uint64) {
881 c.heapMarked = bytesMarked
882 c.heapLive.Store(bytesMarked)
883 c.heapScan.Store(uint64(c.heapScanWork.Load()))
884 c.lastHeapScan = uint64(c.heapScanWork.Load())
885 c.lastStackScan.Store(uint64(c.stackScanWork.Load()))
886 c.triggered = ^uint64(0) // Reset triggered.
888 // heapLive was updated, so emit a trace event.
890 traceHeapAlloc(bytesMarked)
894 // markWorkerStop must be called whenever a mark worker stops executing.
896 // It updates mark work accounting in the controller by a duration of
897 // work in nanoseconds and other bookkeeping.
899 // Safe to execute at any time.
900 func (c *gcControllerState) markWorkerStop(mode gcMarkWorkerMode, duration int64) {
902 case gcMarkWorkerDedicatedMode:
903 c.dedicatedMarkTime.Add(duration)
904 c.dedicatedMarkWorkersNeeded.Add(1)
905 case gcMarkWorkerFractionalMode:
906 c.fractionalMarkTime.Add(duration)
907 case gcMarkWorkerIdleMode:
908 c.idleMarkTime.Add(duration)
909 c.removeIdleMarkWorker()
911 throw("markWorkerStop: unknown mark worker mode")
915 func (c *gcControllerState) update(dHeapLive, dHeapScan int64) {
917 live := gcController.heapLive.Add(dHeapLive)
919 // gcController.heapLive changed.
923 if gcBlackenEnabled == 0 {
924 // Update heapScan when we're not in a current GC. It is fixed
925 // at the beginning of a cycle.
927 gcController.heapScan.Add(dHeapScan)
930 // gcController.heapLive changed.
935 func (c *gcControllerState) addScannableStack(pp *p, amount int64) {
937 c.maxStackScan.Add(amount)
940 pp.maxStackScanDelta += amount
941 if pp.maxStackScanDelta >= maxStackScanSlack || pp.maxStackScanDelta <= -maxStackScanSlack {
942 c.maxStackScan.Add(pp.maxStackScanDelta)
943 pp.maxStackScanDelta = 0
947 func (c *gcControllerState) addGlobals(amount int64) {
948 c.globalsScan.Add(amount)
951 // heapGoal returns the current heap goal.
952 func (c *gcControllerState) heapGoal() uint64 {
953 goal, _ := c.heapGoalInternal()
957 // heapGoalInternal is the implementation of heapGoal which returns additional
958 // information that is necessary for computing the trigger.
960 // The returned minTrigger is always <= goal.
961 func (c *gcControllerState) heapGoalInternal() (goal, minTrigger uint64) {
962 // Start with the goal calculated for gcPercent.
963 goal = c.gcPercentHeapGoal.Load()
965 // Check if the memory-limit-based goal is smaller, and if so, pick that.
966 if newGoal := c.memoryLimitHeapGoal(); go119MemoryLimitSupport && newGoal < goal {
969 // We're not limited by the memory limit goal, so perform a series of
970 // adjustments that might move the goal forward in a variety of circumstances.
972 sweepDistTrigger := c.sweepDistMinTrigger.Load()
973 if sweepDistTrigger > goal {
974 // Set the goal to maintain a minimum sweep distance since
975 // the last call to commit. Note that we never want to do this
976 // if we're in the memory limit regime, because it could push
978 goal = sweepDistTrigger
980 // Since we ignore the sweep distance trigger in the memory
981 // limit regime, we need to ensure we don't propagate it to
982 // the trigger, because it could cause a violation of the
983 // invariant that the trigger < goal.
984 minTrigger = sweepDistTrigger
986 // Ensure that the heap goal is at least a little larger than
987 // the point at which we triggered. This may not be the case if GC
988 // start is delayed or if the allocation that pushed gcController.heapLive
989 // over trigger is large or if the trigger is really close to
990 // GOGC. Assist is proportional to this distance, so enforce a
991 // minimum distance, even if it means going over the GOGC goal
994 // Ignore this if we're in the memory limit regime: we'd prefer to
995 // have the GC respond hard about how close we are to the goal than to
996 // push the goal back in such a manner that it could cause us to exceed
998 const minRunway = 64 << 10
999 if c.triggered != ^uint64(0) && goal < c.triggered+minRunway {
1000 goal = c.triggered + minRunway
1006 // memoryLimitHeapGoal returns a heap goal derived from memoryLimit.
1007 func (c *gcControllerState) memoryLimitHeapGoal() uint64 {
1008 // Start by pulling out some values we'll need. Be careful about overflow.
1009 var heapFree, heapAlloc, mappedReady uint64
1011 heapFree = c.heapFree.load() // Free and unscavenged memory.
1012 heapAlloc = c.totalAlloc.Load() - c.totalFree.Load() // Heap object bytes in use.
1013 mappedReady = c.mappedReady.Load() // Total unreleased mapped memory.
1014 if heapFree+heapAlloc <= mappedReady {
1017 // It is impossible for total unreleased mapped memory to exceed heap memory, but
1018 // because these stats are updated independently, we may observe a partial update
1019 // including only some values. Thus, we appear to break the invariant. However,
1020 // this condition is necessarily transient, so just try again. In the case of a
1021 // persistent accounting error, we'll deadlock here.
1024 // Below we compute a goal from memoryLimit. There are a few things to be aware of.
1025 // Firstly, the memoryLimit does not easily compare to the heap goal: the former
1026 // is total mapped memory by the runtime that hasn't been released, while the latter is
1027 // only heap object memory. Intuitively, the way we convert from one to the other is to
1028 // subtract everything from memoryLimit that both contributes to the memory limit (so,
1029 // ignore scavenged memory) and doesn't contain heap objects. This isn't quite what
1030 // lines up with reality, but it's a good starting point.
1032 // In practice this computation looks like the following:
1034 // memoryLimit - ((mappedReady - heapFree - heapAlloc) + max(mappedReady - memoryLimit, 0)) - memoryLimitHeapGoalHeadroom
1037 // Let's break this down.
1039 // The first term (marker 1) is everything that contributes to the memory limit and isn't
1040 // or couldn't become heap objects. It represents, broadly speaking, non-heap overheads.
1041 // One oddity you may have noticed is that we also subtract out heapFree, i.e. unscavenged
1042 // memory that may contain heap objects in the future.
1044 // Let's take a step back. In an ideal world, this term would look something like just
1045 // the heap goal. That is, we "reserve" enough space for the heap to grow to the heap
1046 // goal, and subtract out everything else. This is of course impossible; the definition
1047 // is circular! However, this impossible definition contains a key insight: the amount
1048 // we're *going* to use matters just as much as whatever we're currently using.
1050 // Consider if the heap shrinks to 1/10th its size, leaving behind lots of free and
1051 // unscavenged memory. mappedReady - heapAlloc will be quite large, because of that free
1052 // and unscavenged memory, pushing the goal down significantly.
1054 // heapFree is also safe to exclude from the memory limit because in the steady-state, it's
1055 // just a pool of memory for future heap allocations, and making new allocations from heapFree
1056 // memory doesn't increase overall memory use. In transient states, the scavenger and the
1057 // allocator actively manage the pool of heapFree memory to maintain the memory limit.
1059 // The second term (marker 2) is the amount of memory we've exceeded the limit by, and is
1060 // intended to help recover from such a situation. By pushing the heap goal down, we also
1061 // push the trigger down, triggering and finishing a GC sooner in order to make room for
1062 // other memory sources. Note that since we're effectively reducing the heap goal by X bytes,
1063 // we're actually giving more than X bytes of headroom back, because the heap goal is in
1064 // terms of heap objects, but it takes more than X bytes (e.g. due to fragmentation) to store
1065 // X bytes worth of objects.
1067 // The third term (marker 3) subtracts an additional memoryLimitHeapGoalHeadroom bytes from the
1068 // heap goal. As the name implies, this is to provide additional headroom in the face of pacing
1069 // inaccuracies. This is a fixed number of bytes because these inaccuracies disproportionately
1070 // affect small heaps: as heaps get smaller, the pacer's inputs get fuzzier. Shorter GC cycles
1071 // and less GC work means noisy external factors like the OS scheduler have a greater impact.
1073 memoryLimit := uint64(c.memoryLimit.Load())
1076 nonHeapMemory := mappedReady - heapFree - heapAlloc
1080 if mappedReady > memoryLimit {
1081 overage = mappedReady - memoryLimit
1084 if nonHeapMemory+overage >= memoryLimit {
1085 // We're at a point where non-heap memory exceeds the memory limit on its own.
1086 // There's honestly not much we can do here but just trigger GCs continuously
1087 // and let the CPU limiter reign that in. Something has to give at this point.
1088 // Set it to heapMarked, the lowest possible goal.
1092 // Compute the goal.
1093 goal := memoryLimit - (nonHeapMemory + overage)
1095 // Apply some headroom to the goal to account for pacing inaccuracies.
1096 // Be careful about small limits.
1097 if goal < memoryLimitHeapGoalHeadroom || goal-memoryLimitHeapGoalHeadroom < memoryLimitHeapGoalHeadroom {
1098 goal = memoryLimitHeapGoalHeadroom
1100 goal = goal - memoryLimitHeapGoalHeadroom
1102 // Don't let us go below the live heap. A heap goal below the live heap doesn't make sense.
1103 if goal < c.heapMarked {
1110 // These constants determine the bounds on the GC trigger as a fraction
1111 // of heap bytes allocated between the start of a GC (heapLive == heapMarked)
1112 // and the end of a GC (heapLive == heapGoal).
1114 // The constants are obscured in this way for efficiency. The denominator
1115 // of the fraction is always a power-of-two for a quick division, so that
1116 // the numerator is a single constant integer multiplication.
1117 triggerRatioDen = 64
1119 // The minimum trigger constant was chosen empirically: given a sufficiently
1120 // fast/scalable allocator with 48 Ps that could drive the trigger ratio
1121 // to <0.05, this constant causes applications to retain the same peak
1122 // RSS compared to not having this allocator.
1123 minTriggerRatioNum = 45 // ~0.7
1125 // The maximum trigger constant is chosen somewhat arbitrarily, but the
1126 // current constant has served us well over the years.
1127 maxTriggerRatioNum = 61 // ~0.95
1130 // trigger returns the current point at which a GC should trigger along with
1133 // The returned value may be compared against heapLive to determine whether
1134 // the GC should trigger. Thus, the GC trigger condition should be (but may
1135 // not be, in the case of small movements for efficiency) checked whenever
1136 // the heap goal may change.
1137 func (c *gcControllerState) trigger() (uint64, uint64) {
1138 goal, minTrigger := c.heapGoalInternal()
1140 // Invariant: the trigger must always be less than the heap goal.
1142 // Note that the memory limit sets a hard maximum on our heap goal,
1143 // but the live heap may grow beyond it.
1145 if c.heapMarked >= goal {
1146 // The goal should never be smaller than heapMarked, but let's be
1147 // defensive about it. The only reasonable trigger here is one that
1148 // causes a continuous GC cycle at heapMarked, but respect the goal
1149 // if it came out as smaller than that.
1153 // Below this point, c.heapMarked < goal.
1155 // heapMarked is our absolute minimum, and it's possible the trigger
1156 // bound we get from heapGoalinternal is less than that.
1157 if minTrigger < c.heapMarked {
1158 minTrigger = c.heapMarked
1161 // If we let the trigger go too low, then if the application
1162 // is allocating very rapidly we might end up in a situation
1163 // where we're allocating black during a nearly always-on GC.
1164 // The result of this is a growing heap and ultimately an
1165 // increase in RSS. By capping us at a point >0, we're essentially
1166 // saying that we're OK using more CPU during the GC to prevent
1167 // this growth in RSS.
1168 triggerLowerBound := uint64(((goal-c.heapMarked)/triggerRatioDen)*minTriggerRatioNum) + c.heapMarked
1169 if minTrigger < triggerLowerBound {
1170 minTrigger = triggerLowerBound
1173 // For small heaps, set the max trigger point at maxTriggerRatio of the way
1174 // from the live heap to the heap goal. This ensures we always have *some*
1175 // headroom when the GC actually starts. For larger heaps, set the max trigger
1176 // point at the goal, minus the minimum heap size.
1178 // This choice follows from the fact that the minimum heap size is chosen
1179 // to reflect the costs of a GC with no work to do. With a large heap but
1180 // very little scan work to perform, this gives us exactly as much runway
1181 // as we would need, in the worst case.
1182 maxTrigger := uint64(((goal-c.heapMarked)/triggerRatioDen)*maxTriggerRatioNum) + c.heapMarked
1183 if goal > defaultHeapMinimum && goal-defaultHeapMinimum > maxTrigger {
1184 maxTrigger = goal - defaultHeapMinimum
1186 if maxTrigger < minTrigger {
1187 maxTrigger = minTrigger
1190 // Compute the trigger from our bounds and the runway stored by commit.
1192 runway := c.runway.Load()
1194 trigger = minTrigger
1196 trigger = goal - runway
1198 if trigger < minTrigger {
1199 trigger = minTrigger
1201 if trigger > maxTrigger {
1202 trigger = maxTrigger
1205 print("trigger=", trigger, " heapGoal=", goal, "\n")
1206 print("minTrigger=", minTrigger, " maxTrigger=", maxTrigger, "\n")
1207 throw("produced a trigger greater than the heap goal")
1209 return trigger, goal
1212 // commit recomputes all pacing parameters needed to derive the
1213 // trigger and the heap goal. Namely, the gcPercent-based heap goal,
1214 // and the amount of runway we want to give the GC this cycle.
1216 // This can be called any time. If GC is the in the middle of a
1217 // concurrent phase, it will adjust the pacing of that phase.
1219 // isSweepDone should be the result of calling isSweepDone(),
1220 // unless we're testing or we know we're executing during a GC cycle.
1222 // This depends on gcPercent, gcController.heapMarked, and
1223 // gcController.heapLive. These must be up to date.
1225 // Callers must call gcControllerState.revise after calling this
1226 // function if the GC is enabled.
1228 // mheap_.lock must be held or the world must be stopped.
1229 func (c *gcControllerState) commit(isSweepDone bool) {
1231 assertWorldStoppedOrLockHeld(&mheap_.lock)
1235 // The sweep is done, so there aren't any restrictions on the trigger
1236 // we need to think about.
1237 c.sweepDistMinTrigger.Store(0)
1239 // Concurrent sweep happens in the heap growth
1240 // from gcController.heapLive to trigger. Make sure we
1241 // give the sweeper some runway if it doesn't have enough.
1242 c.sweepDistMinTrigger.Store(c.heapLive.Load() + sweepMinHeapDistance)
1245 // Compute the next GC goal, which is when the allocated heap
1246 // has grown by GOGC/100 over where it started the last cycle,
1247 // plus additional runway for non-heap sources of GC work.
1248 gcPercentHeapGoal := ^uint64(0)
1249 if gcPercent := c.gcPercent.Load(); gcPercent >= 0 {
1250 gcPercentHeapGoal = c.heapMarked + (c.heapMarked+c.lastStackScan.Load()+c.globalsScan.Load())*uint64(gcPercent)/100
1252 // Apply the minimum heap size here. It's defined in terms of gcPercent
1253 // and is only updated by functions that call commit.
1254 if gcPercentHeapGoal < c.heapMinimum {
1255 gcPercentHeapGoal = c.heapMinimum
1257 c.gcPercentHeapGoal.Store(gcPercentHeapGoal)
1259 // Compute the amount of runway we want the GC to have by using our
1260 // estimate of the cons/mark ratio.
1262 // The idea is to take our expected scan work, and multiply it by
1263 // the cons/mark ratio to determine how long it'll take to complete
1264 // that scan work in terms of bytes allocated. This gives us our GC's
1267 // However, the cons/mark ratio is a ratio of rates per CPU-second, but
1268 // here we care about the relative rates for some division of CPU
1269 // resources among the mutator and the GC.
1271 // To summarize, we have B / cpu-ns, and we want B / ns. We get that
1272 // by multiplying by our desired division of CPU resources. We choose
1273 // to express CPU resources as GOMAPROCS*fraction. Note that because
1274 // we're working with a ratio here, we can omit the number of CPU cores,
1275 // because they'll appear in the numerator and denominator and cancel out.
1276 // As a result, this is basically just "weighing" the cons/mark ratio by
1277 // our desired division of resources.
1279 // Furthermore, by setting the runway so that CPU resources are divided
1280 // this way, assuming that the cons/mark ratio is correct, we make that
1281 // division a reality.
1282 c.runway.Store(uint64((c.consMark * (1 - gcGoalUtilization) / (gcGoalUtilization)) * float64(c.lastHeapScan+c.lastStackScan.Load()+c.globalsScan.Load())))
1285 // setGCPercent updates gcPercent. commit must be called after.
1286 // Returns the old value of gcPercent.
1288 // The world must be stopped, or mheap_.lock must be held.
1289 func (c *gcControllerState) setGCPercent(in int32) int32 {
1291 assertWorldStoppedOrLockHeld(&mheap_.lock)
1294 out := c.gcPercent.Load()
1298 c.heapMinimum = defaultHeapMinimum * uint64(in) / 100
1299 c.gcPercent.Store(in)
1304 //go:linkname setGCPercent runtime/debug.setGCPercent
1305 func setGCPercent(in int32) (out int32) {
1306 // Run on the system stack since we grab the heap lock.
1307 systemstack(func() {
1309 out = gcController.setGCPercent(in)
1310 gcControllerCommit()
1311 unlock(&mheap_.lock)
1314 // If we just disabled GC, wait for any concurrent GC mark to
1315 // finish so we always return with no GC running.
1317 gcWaitOnMark(atomic.Load(&work.cycles))
1323 func readGOGC() int32 {
1324 p := gogetenv("GOGC")
1328 if n, ok := atoi32(p); ok {
1334 // setMemoryLimit updates memoryLimit. commit must be called after
1335 // Returns the old value of memoryLimit.
1337 // The world must be stopped, or mheap_.lock must be held.
1338 func (c *gcControllerState) setMemoryLimit(in int64) int64 {
1340 assertWorldStoppedOrLockHeld(&mheap_.lock)
1343 out := c.memoryLimit.Load()
1345 c.memoryLimit.Store(in)
1351 //go:linkname setMemoryLimit runtime/debug.setMemoryLimit
1352 func setMemoryLimit(in int64) (out int64) {
1353 // Run on the system stack since we grab the heap lock.
1354 systemstack(func() {
1356 out = gcController.setMemoryLimit(in)
1357 if in < 0 || out == in {
1358 // If we're just checking the value or not changing
1359 // it, there's no point in doing the rest.
1360 unlock(&mheap_.lock)
1363 gcControllerCommit()
1364 unlock(&mheap_.lock)
1369 func readGOMEMLIMIT() int64 {
1370 p := gogetenv("GOMEMLIMIT")
1371 if p == "" || p == "off" {
1374 n, ok := parseByteCount(p)
1376 print("GOMEMLIMIT=", p, "\n")
1377 throw("malformed GOMEMLIMIT; see `go doc runtime/debug.SetMemoryLimit`")
1382 type piController struct {
1383 kp float64 // Proportional constant.
1384 ti float64 // Integral time constant.
1385 tt float64 // Reset time.
1387 min, max float64 // Output boundaries.
1389 // PI controller state.
1391 errIntegral float64 // Integral of the error from t=0 to now.
1394 errOverflow bool // Set if errIntegral ever overflowed.
1395 inputOverflow bool // Set if an operation with the input overflowed.
1398 // next provides a new sample to the controller.
1400 // input is the sample, setpoint is the desired point, and period is how much
1401 // time (in whatever unit makes the most sense) has passed since the last sample.
1403 // Returns a new value for the variable it's controlling, and whether the operation
1404 // completed successfully. One reason this might fail is if error has been growing
1405 // in an unbounded manner, to the point of overflow.
1407 // In the specific case of an error overflow occurs, the errOverflow field will be
1408 // set and the rest of the controller's internal state will be fully reset.
1409 func (c *piController) next(input, setpoint, period float64) (float64, bool) {
1410 // Compute the raw output value.
1411 prop := c.kp * (setpoint - input)
1412 rawOutput := prop + c.errIntegral
1414 // Clamp rawOutput into output.
1416 if isInf(output) || isNaN(output) {
1417 // The input had a large enough magnitude that either it was already
1418 // overflowed, or some operation with it overflowed.
1419 // Set a flag and reset. That's the safest thing to do.
1421 c.inputOverflow = true
1426 } else if output > c.max {
1430 // Update the controller's state.
1431 if c.ti != 0 && c.tt != 0 {
1432 c.errIntegral += (c.kp*period/c.ti)*(setpoint-input) + (period/c.tt)*(output-rawOutput)
1433 if isInf(c.errIntegral) || isNaN(c.errIntegral) {
1434 // So much error has accumulated that we managed to overflow.
1435 // The assumptions around the controller have likely broken down.
1436 // Set a flag and reset. That's the safest thing to do.
1438 c.errOverflow = true
1445 // reset resets the controller state, except for controller error flags.
1446 func (c *piController) reset() {
1450 // addIdleMarkWorker attempts to add a new idle mark worker.
1452 // If this returns true, the caller must become an idle mark worker unless
1453 // there's no background mark worker goroutines in the pool. This case is
1454 // harmless because there are already background mark workers running.
1455 // If this returns false, the caller must NOT become an idle mark worker.
1457 // 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.
1488 func (c *gcControllerState) needIdleMarkWorker() bool {
1489 p := c.idleMarkWorkers.Load()
1490 n, max := int32(p&uint64(^uint32(0))), int32(p>>32)
1494 // removeIdleMarkWorker must be called when an new idle mark worker stops executing.
1495 func (c *gcControllerState) removeIdleMarkWorker() {
1497 old := c.idleMarkWorkers.Load()
1498 n, max := int32(old&uint64(^uint32(0))), int32(old>>32)
1500 print("n=", n, " max=", max, "\n")
1501 throw("negative idle mark workers")
1503 new := uint64(uint32(n-1)) | (uint64(max) << 32)
1504 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1510 // setMaxIdleMarkWorkers sets the maximum number of idle mark workers allowed.
1512 // This method is optimistic in that it does not wait for the number of
1513 // idle mark workers to reduce to max before returning; it assumes the workers
1514 // will deschedule themselves.
1515 func (c *gcControllerState) setMaxIdleMarkWorkers(max int32) {
1517 old := c.idleMarkWorkers.Load()
1518 n := int32(old & uint64(^uint32(0)))
1520 print("n=", n, " max=", max, "\n")
1521 throw("negative idle mark workers")
1523 new := uint64(uint32(n)) | (uint64(max) << 32)
1524 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1530 // gcControllerCommit is gcController.commit, but passes arguments from live
1531 // (non-test) data. It also updates any consumers of the GC pacing, such as
1532 // sweep pacing and the background scavenger.
1534 // Calls gcController.commit.
1536 // The heap lock must be held, so this must be executed on the system stack.
1539 func gcControllerCommit() {
1540 assertWorldStoppedOrLockHeld(&mheap_.lock)
1542 gcController.commit(isSweepDone())
1544 // Update mark pacing.
1545 if gcphase != _GCoff {
1546 gcController.revise()
1549 // TODO(mknyszek): This isn't really accurate any longer because the heap
1550 // goal is computed dynamically. Still useful to snapshot, but not as useful.
1555 trigger, heapGoal := gcController.trigger()
1556 gcPaceSweeper(trigger)
1557 gcPaceScavenger(gcController.memoryLimit.Load(), heapGoal, gcController.lastHeapGoal)