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 _ uint32 // padding so following 64-bit values are 8-byte aligned
97 // memoryLimit is the soft memory limit in bytes.
99 // Initialized from GOMEMLIMIT. GOMEMLIMIT=off is equivalent to MaxInt64
100 // which means no soft memory limit in practice.
102 // This is an int64 instead of a uint64 to more easily maintain parity with
103 // the SetMemoryLimit API, which sets a maximum at MaxInt64. This value
104 // should never be negative.
105 memoryLimit atomic.Int64
107 // heapMinimum is the minimum heap size at which to trigger GC.
108 // For small heaps, this overrides the usual GOGC*live set rule.
110 // When there is a very small live set but a lot of allocation, simply
111 // collecting when the heap reaches GOGC*live results in many GC
112 // cycles and high total per-GC overhead. This minimum amortizes this
113 // per-GC overhead while keeping the heap reasonably small.
115 // During initialization this is set to 4MB*GOGC/100. In the case of
116 // GOGC==0, this will set heapMinimum to 0, resulting in constant
117 // collection even when the heap size is small, which is useful for
121 // runway is the amount of runway in heap bytes allocated by the
122 // application that we want to give the GC once it starts.
124 // This is computed from consMark during mark termination.
127 // consMark is the estimated per-CPU consMark ratio for the application.
129 // It represents the ratio between the application's allocation
130 // rate, as bytes allocated per CPU-time, and the GC's scan rate,
131 // as bytes scanned per CPU-time.
132 // The units of this ratio are (B / cpu-ns) / (B / cpu-ns).
134 // At a high level, this value is computed as the bytes of memory
135 // allocated (cons) per unit of scan work completed (mark) in a GC
136 // cycle, divided by the CPU time spent on each activity.
138 // Updated at the end of each GC cycle, in endCycle.
141 // consMarkController holds the state for the mark-cons ratio
142 // estimation over time.
144 // Its purpose is to smooth out noisiness in the computation of
145 // consMark; see consMark for details.
146 consMarkController piController
148 _ uint32 // Padding for atomics on 32-bit platforms.
150 // gcPercentHeapGoal is the goal heapLive for when next GC ends derived
153 // Set to ^uint64(0) if gcPercent is disabled.
154 gcPercentHeapGoal atomic.Uint64
156 // sweepDistMinTrigger is the minimum trigger to ensure a minimum
159 // This bound is also special because it applies to both the trigger
160 // *and* the goal (all other trigger bounds must be based *on* the goal).
162 // It is computed ahead of time, at commit time. The theory is that,
163 // absent a sudden change to a parameter like gcPercent, the trigger
164 // will be chosen to always give the sweeper enough headroom. However,
165 // such a change might dramatically and suddenly move up the trigger,
166 // in which case we need to ensure the sweeper still has enough headroom.
167 sweepDistMinTrigger atomic.Uint64
169 // triggered is the point at which the current GC cycle actually triggered.
170 // Only valid during the mark phase of a GC cycle, otherwise set to ^uint64(0).
172 // Updated while the world is stopped.
175 // lastHeapGoal is the value of heapGoal at the moment the last GC
176 // ended. Note that this is distinct from the last value heapGoal had,
177 // because it could change if e.g. gcPercent changes.
179 // Read and written with the world stopped or with mheap_.lock held.
182 // heapLive is the number of bytes considered live by the GC.
183 // That is: retained by the most recent GC plus allocated
184 // since then. heapLive ≤ memstats.totalAlloc-memstats.totalFree, since
185 // heapAlloc includes unmarked objects that have not yet been swept (and
186 // hence goes up as we allocate and down as we sweep) while heapLive
187 // excludes these objects (and hence only goes up between GCs).
189 // To reduce contention, this is updated only when obtaining a span
190 // from an mcentral and at this point it counts all of the unallocated
191 // slots in that span (which will be allocated before that mcache
192 // obtains another span from that mcentral). Hence, it slightly
193 // overestimates the "true" live heap size. It's better to overestimate
194 // than to underestimate because 1) this triggers the GC earlier than
195 // necessary rather than potentially too late and 2) this leads to a
196 // conservative GC rate rather than a GC rate that is potentially too
199 // Whenever this is updated, call traceHeapAlloc() and
200 // this gcControllerState's revise() method.
201 heapLive atomic.Uint64
203 // heapScan is the number of bytes of "scannable" heap. This is the
204 // live heap (as counted by heapLive), but omitting no-scan objects and
205 // no-scan tails of objects.
207 // This value is fixed at the start of a GC cycle. It represents the
208 // maximum scannable heap.
209 heapScan atomic.Uint64
211 // lastHeapScan is the number of bytes of heap that were scanned
212 // last GC cycle. It is the same as heapMarked, but only
213 // includes the "scannable" parts of objects.
215 // Updated when the world is stopped.
218 // lastStackScan is the number of bytes of stack that were scanned
220 lastStackScan atomic.Uint64
222 // maxStackScan is the amount of allocated goroutine stack space in
223 // use by goroutines.
225 // This number tracks allocated goroutine stack space rather than used
226 // goroutine stack space (i.e. what is actually scanned) because used
227 // goroutine stack space is much harder to measure cheaply. By using
228 // allocated space, we make an overestimate; this is OK, it's better
229 // to conservatively overcount than undercount.
230 maxStackScan atomic.Uint64
232 // globalsScan is the total amount of global variable space
233 // that is scannable.
234 globalsScan atomic.Uint64
236 // heapMarked is the number of bytes marked by the previous
237 // GC. After mark termination, heapLive == heapMarked, but
238 // unlike heapLive, heapMarked does not change until the
239 // next mark termination.
242 // heapScanWork is the total heap scan work performed this cycle.
243 // stackScanWork is the total stack scan work performed this cycle.
244 // globalsScanWork is the total globals scan work performed this cycle.
246 // These are updated atomically during the cycle. Updates occur in
247 // bounded batches, since they are both written and read
248 // throughout the cycle. At the end of the cycle, heapScanWork is how
249 // much of the retained heap is scannable.
251 // Currently these are measured in bytes. For most uses, this is an
252 // opaque unit of work, but for estimation the definition is important.
254 // Note that stackScanWork includes only stack space scanned, not all
255 // of the allocated stack.
256 heapScanWork atomic.Int64
257 stackScanWork atomic.Int64
258 globalsScanWork atomic.Int64
260 // bgScanCredit is the scan work credit accumulated by the concurrent
261 // background scan. This credit is accumulated by the background scan
262 // and stolen by mutator assists. Updates occur in bounded batches,
263 // since it is both written and read throughout the cycle.
264 bgScanCredit atomic.Int64
266 // assistTime is the nanoseconds spent in mutator assists
267 // during this cycle. This is updated atomically, and must also
268 // be updated atomically even during a STW, because it is read
269 // by sysmon. Updates occur in bounded batches, since it is both
270 // written and read throughout the cycle.
271 assistTime atomic.Int64
273 // dedicatedMarkTime is the nanoseconds spent in dedicated mark workers
274 // during this cycle. This is updated at the end of the concurrent mark
276 dedicatedMarkTime atomic.Int64
278 // fractionalMarkTime is the nanoseconds spent in the
279 // fractional mark worker during this cycle. This is updated
280 // atomically throughout the cycle and will be up-to-date if
281 // the fractional mark worker is not currently running.
282 fractionalMarkTime int64
284 // idleMarkTime is the nanoseconds spent in idle marking
285 // during this cycle. This is updated atomically throughout
289 // markStartTime is the absolute start time in nanoseconds
290 // that assists and background mark workers started.
293 // dedicatedMarkWorkersNeeded is the number of dedicated mark
294 // workers that need to be started. This is computed at the
295 // beginning of each cycle and decremented atomically as
296 // dedicated mark workers get started.
297 dedicatedMarkWorkersNeeded int64
299 // idleMarkWorkers is two packed int32 values in a single uint64.
300 // These two values are always updated simultaneously.
302 // The bottom int32 is the current number of idle mark workers executing.
304 // The top int32 is the maximum number of idle mark workers allowed to
305 // execute concurrently. Normally, this number is just gomaxprocs. However,
306 // during periodic GC cycles it is set to 0 because the system is idle
307 // anyway; there's no need to go full blast on all of GOMAXPROCS.
309 // The maximum number of idle mark workers is used to prevent new workers
310 // from starting, but it is not a hard maximum. It is possible (but
311 // exceedingly rare) for the current number of idle mark workers to
312 // transiently exceed the maximum. This could happen if the maximum changes
313 // just after a GC ends, and an M with no P.
315 // Note that if we have no dedicated mark workers, we set this value to
316 // 1 in this case we only have fractional GC workers which aren't scheduled
317 // strictly enough to ensure GC progress. As a result, idle-priority mark
318 // workers are vital to GC progress in these situations.
320 // For example, consider a situation in which goroutines block on the GC
321 // (such as via runtime.GOMAXPROCS) and only fractional mark workers are
322 // scheduled (e.g. GOMAXPROCS=1). Without idle-priority mark workers, the
323 // last running M might skip scheduling a fractional mark worker if its
324 // utilization goal is met, such that once it goes to sleep (because there's
325 // nothing to do), there will be nothing else to spin up a new M for the
326 // fractional worker in the future, stalling GC progress and causing a
327 // deadlock. However, idle-priority workers will *always* run when there is
328 // nothing left to do, ensuring the GC makes progress.
330 // See github.com/golang/go/issues/44163 for more details.
331 idleMarkWorkers atomic.Uint64
333 // assistWorkPerByte is the ratio of scan work to allocated
334 // bytes that should be performed by mutator assists. This is
335 // computed at the beginning of each cycle and updated every
336 // time heapScan is updated.
337 assistWorkPerByte atomic.Float64
339 // assistBytesPerWork is 1/assistWorkPerByte.
341 // Note that because this is read and written independently
342 // from assistWorkPerByte users may notice a skew between
343 // the two values, and such a state should be safe.
344 assistBytesPerWork atomic.Float64
346 // fractionalUtilizationGoal is the fraction of wall clock
347 // time that should be spent in the fractional mark worker on
348 // each P that isn't running a dedicated worker.
350 // For example, if the utilization goal is 25% and there are
351 // no dedicated workers, this will be 0.25. If the goal is
352 // 25%, there is one dedicated worker, and GOMAXPROCS is 5,
353 // this will be 0.05 to make up the missing 5%.
355 // If this is zero, no fractional workers are needed.
356 fractionalUtilizationGoal float64
358 // These memory stats are effectively duplicates of fields from
359 // memstats.heapStats but are updated atomically or with the world
360 // stopped and don't provide the same consistency guarantees.
362 // Because the runtime is responsible for managing a memory limit, it's
363 // useful to couple these stats more tightly to the gcController, which
364 // is intimately connected to how that memory limit is maintained.
365 heapInUse sysMemStat // bytes in mSpanInUse spans
366 heapReleased sysMemStat // bytes released to the OS
367 heapFree sysMemStat // bytes not in any span, but not released to the OS
368 totalAlloc atomic.Uint64 // total bytes allocated
369 totalFree atomic.Uint64 // total bytes freed
370 mappedReady atomic.Uint64 // total virtual memory in the Ready state (see mem.go).
372 // test indicates that this is a test-only copy of gcControllerState.
378 func (c *gcControllerState) init(gcPercent int32, memoryLimit int64) {
379 c.heapMinimum = defaultHeapMinimum
380 c.triggered = ^uint64(0)
382 c.consMarkController = piController{
383 // Tuned first via the Ziegler-Nichols process in simulation,
384 // then the integral time was manually tuned against real-world
385 // applications to deal with noisiness in the measured cons/mark
390 // Set a high reset time in GC cycles.
391 // This is inversely proportional to the rate at which we
392 // accumulate error from clipping. By making this very high
393 // we make the accumulation slow. In general, clipping is
394 // OK in our situation, hence the choice.
396 // Tune this if we get unintended effects from clipping for
403 c.setGCPercent(gcPercent)
404 c.setMemoryLimit(memoryLimit)
405 c.commit(true) // No sweep phase in the first GC cycle.
406 // N.B. Don't bother calling traceHeapGoal. Tracing is never enabled at
407 // initialization time.
408 // N.B. No need to call revise; there's no GC enabled during
412 // startCycle resets the GC controller's state and computes estimates
413 // for a new GC cycle. The caller must hold worldsema and the world
415 func (c *gcControllerState) startCycle(markStartTime int64, procs int, trigger gcTrigger) {
416 c.heapScanWork.Store(0)
417 c.stackScanWork.Store(0)
418 c.globalsScanWork.Store(0)
419 c.bgScanCredit.Store(0)
420 c.assistTime.Store(0)
421 c.dedicatedMarkTime.Store(0)
422 c.fractionalMarkTime = 0
424 c.markStartTime = markStartTime
426 // TODO(mknyszek): This is supposed to be the actual trigger point for the heap, but
427 // causes regressions in memory use. The cause is that the PI controller used to smooth
428 // the cons/mark ratio measurements tends to flail when using the less accurate precomputed
429 // trigger for the cons/mark calculation, and this results in the controller being more
430 // conservative about steady-states it tries to find in the future.
432 // This conservatism is transient, but these transient states tend to matter for short-lived
433 // programs, especially because the PI controller is overdamped, partially because it is
434 // configured with a relatively large time constant.
436 // Ultimately, I think this is just two mistakes piled on one another: the choice of a swingy
437 // smoothing function that recalls a fairly long history (due to its overdamped time constant)
438 // coupled with an inaccurate cons/mark calculation. It just so happens this works better
439 // today, and it makes it harder to change things in the future.
441 // This is described in #53738. Fix this for #53892 by changing back to the actual trigger
442 // point and simplifying the smoothing function.
443 heapTrigger, heapGoal := c.trigger()
444 c.triggered = heapTrigger
446 // Compute the background mark utilization goal. In general,
447 // this may not come out exactly. We round the number of
448 // dedicated workers so that the utilization is closest to
449 // 25%. For small GOMAXPROCS, this would introduce too much
450 // error, so we add fractional workers in that case.
451 totalUtilizationGoal := float64(procs) * gcBackgroundUtilization
452 c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5)
453 utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
454 const maxUtilError = 0.3
455 if utilError < -maxUtilError || utilError > maxUtilError {
456 // Rounding put us more than 30% off our goal. With
457 // gcBackgroundUtilization of 25%, this happens for
458 // GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
459 // workers to compensate.
460 if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
461 // Too many dedicated workers.
462 c.dedicatedMarkWorkersNeeded--
464 c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(procs)
466 c.fractionalUtilizationGoal = 0
469 // In STW mode, we just want dedicated workers.
470 if debug.gcstoptheworld > 0 {
471 c.dedicatedMarkWorkersNeeded = int64(procs)
472 c.fractionalUtilizationGoal = 0
476 for _, p := range allp {
478 p.gcFractionalMarkTime = 0
481 if trigger.kind == gcTriggerTime {
482 // During a periodic GC cycle, reduce the number of idle mark workers
483 // required. However, we need at least one dedicated mark worker or
484 // idle GC worker to ensure GC progress in some scenarios (see comment
485 // on maxIdleMarkWorkers).
486 if c.dedicatedMarkWorkersNeeded > 0 {
487 c.setMaxIdleMarkWorkers(0)
489 // TODO(mknyszek): The fundamental reason why we need this is because
490 // we can't count on the fractional mark worker to get scheduled.
491 // Fix that by ensuring it gets scheduled according to its quota even
492 // if the rest of the application is idle.
493 c.setMaxIdleMarkWorkers(1)
496 // N.B. gomaxprocs and dedicatedMarkWorkersNeeded is guaranteed not to
497 // change during a GC cycle.
498 c.setMaxIdleMarkWorkers(int32(procs) - int32(c.dedicatedMarkWorkersNeeded))
501 // Compute initial values for controls that are updated
502 // throughout the cycle.
505 if debug.gcpacertrace > 0 {
506 assistRatio := c.assistWorkPerByte.Load()
507 print("pacer: assist ratio=", assistRatio,
508 " (scan ", gcController.heapScan.Load()>>20, " MB in ",
509 work.initialHeapLive>>20, "->",
510 heapGoal>>20, " MB)",
511 " workers=", c.dedicatedMarkWorkersNeeded,
512 "+", c.fractionalUtilizationGoal, "\n")
516 // revise updates the assist ratio during the GC cycle to account for
517 // improved estimates. This should be called whenever gcController.heapScan,
518 // gcController.heapLive, or if any inputs to gcController.heapGoal are
519 // updated. It is safe to call concurrently, but it may race with other
522 // The result of this race is that the two assist ratio values may not line
523 // up or may be stale. In practice this is OK because the assist ratio
524 // moves slowly throughout a GC cycle, and the assist ratio is a best-effort
525 // heuristic anyway. Furthermore, no part of the heuristic depends on
526 // the two assist ratio values being exact reciprocals of one another, since
527 // the two values are used to convert values from different sources.
529 // The worst case result of this raciness is that we may miss a larger shift
530 // in the ratio (say, if we decide to pace more aggressively against the
531 // hard heap goal) but even this "hard goal" is best-effort (see #40460).
532 // The dedicated GC should ensure we don't exceed the hard goal by too much
533 // in the rare case we do exceed it.
535 // It should only be called when gcBlackenEnabled != 0 (because this
536 // is when assists are enabled and the necessary statistics are
538 func (c *gcControllerState) revise() {
539 gcPercent := c.gcPercent.Load()
541 // If GC is disabled but we're running a forced GC,
542 // act like GOGC is huge for the below calculations.
545 live := c.heapLive.Load()
546 scan := c.heapScan.Load()
547 work := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load()
549 // Assume we're under the soft goal. Pace GC to complete at
550 // heapGoal assuming the heap is in steady-state.
551 heapGoal := int64(c.heapGoal())
553 // The expected scan work is computed as the amount of bytes scanned last
554 // GC cycle (both heap and stack), plus our estimate of globals work for this cycle.
555 scanWorkExpected := int64(c.lastHeapScan + c.lastStackScan.Load() + c.globalsScan.Load())
557 // maxScanWork is a worst-case estimate of the amount of scan work that
558 // needs to be performed in this GC cycle. Specifically, it represents
559 // the case where *all* scannable memory turns out to be live, and
560 // *all* allocated stack space is scannable.
561 maxStackScan := c.maxStackScan.Load()
562 maxScanWork := int64(scan + maxStackScan + c.globalsScan.Load())
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.Load() <= 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.Load()-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 {
736 goal := gcGoalUtilization * 100
737 print("pacer: ", int(utilization*100), "% CPU (", int(goal), " exp.) for ")
738 print(c.heapScanWork.Load(), "+", c.stackScanWork.Load(), "+", c.globalsScanWork.Load(), " B work (", c.lastHeapScan+c.lastStackScan.Load()+c.globalsScan.Load(), " B exp.) ")
739 live := c.heapLive.Load()
740 print("in ", c.triggered, " B -> ", live, " B (∆goal ", int64(live)-int64(c.lastHeapGoal), ", 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 pp if it
793 // should be run. This must only be called when gcBlackenEnabled != 0.
794 func (c *gcControllerState) findRunnableGCWorker(pp *p, now int64) (*g, int64) {
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.
806 if gcCPULimiter.needUpdate(now) {
807 gcCPULimiter.update(now)
810 if !gcMarkWorkAvailable(pp) {
811 // No work to be done right now. This can happen at
812 // the end of the mark phase when there are still
813 // assists tapering off. Don't bother running a worker
814 // now because it'll just return immediately.
818 // Grab a worker before we commit to running below.
819 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
821 // There is at least one worker per P, so normally there are
822 // enough workers to run on all Ps, if necessary. However, once
823 // a worker enters gcMarkDone it may park without rejoining the
824 // pool, thus freeing a P with no corresponding worker.
825 // gcMarkDone never depends on another worker doing work, so it
826 // is safe to simply do nothing here.
828 // If gcMarkDone bails out without completing the mark phase,
829 // it will always do so with queued global work. Thus, that P
830 // will be immediately eligible to re-run the worker G it was
831 // just using, ensuring work can complete.
835 decIfPositive := func(ptr *int64) bool {
837 v := atomic.Loadint64(ptr)
842 if atomic.Casint64(ptr, v, v-1) {
848 if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
849 // This P is now dedicated to marking until the end of
850 // the concurrent mark phase.
851 pp.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
852 } else if c.fractionalUtilizationGoal == 0 {
853 // No need for fractional workers.
854 gcBgMarkWorkerPool.push(&node.node)
857 // Is this P behind on the fractional utilization
860 // This should be kept in sync with pollFractionalWorkerExit.
861 delta := now - c.markStartTime
862 if delta > 0 && float64(pp.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
863 // Nope. No need to run a fractional worker.
864 gcBgMarkWorkerPool.push(&node.node)
867 // Run a fractional worker.
868 pp.gcMarkWorkerMode = gcMarkWorkerFractionalMode
871 // Run the background mark worker.
873 casgstatus(gp, _Gwaiting, _Grunnable)
880 // resetLive sets up the controller state for the next mark phase after the end
881 // of the previous one. Must be called after endCycle and before commit, before
882 // the world is started.
884 // The world must be stopped.
885 func (c *gcControllerState) resetLive(bytesMarked uint64) {
886 c.heapMarked = bytesMarked
887 c.heapLive.Store(bytesMarked)
888 c.heapScan.Store(uint64(c.heapScanWork.Load()))
889 c.lastHeapScan = uint64(c.heapScanWork.Load())
890 c.lastStackScan.Store(uint64(c.stackScanWork.Load()))
891 c.triggered = ^uint64(0) // Reset triggered.
893 // heapLive was updated, so emit a trace event.
895 traceHeapAlloc(bytesMarked)
899 // markWorkerStop must be called whenever a mark worker stops executing.
901 // It updates mark work accounting in the controller by a duration of
902 // work in nanoseconds and other bookkeeping.
904 // Safe to execute at any time.
905 func (c *gcControllerState) markWorkerStop(mode gcMarkWorkerMode, duration int64) {
907 case gcMarkWorkerDedicatedMode:
908 c.dedicatedMarkTime.Add(duration)
909 atomic.Xaddint64(&c.dedicatedMarkWorkersNeeded, 1)
910 case gcMarkWorkerFractionalMode:
911 atomic.Xaddint64(&c.fractionalMarkTime, duration)
912 case gcMarkWorkerIdleMode:
913 atomic.Xaddint64(&c.idleMarkTime, duration)
914 c.removeIdleMarkWorker()
916 throw("markWorkerStop: unknown mark worker mode")
920 func (c *gcControllerState) update(dHeapLive, dHeapScan int64) {
922 live := gcController.heapLive.Add(dHeapLive)
924 // gcController.heapLive changed.
928 if gcBlackenEnabled == 0 {
929 // Update heapScan when we're not in a current GC. It is fixed
930 // at the beginning of a cycle.
932 gcController.heapScan.Add(dHeapScan)
935 // gcController.heapLive changed.
940 func (c *gcControllerState) addScannableStack(pp *p, amount int64) {
942 c.maxStackScan.Add(amount)
945 pp.maxStackScanDelta += amount
946 if pp.maxStackScanDelta >= maxStackScanSlack || pp.maxStackScanDelta <= -maxStackScanSlack {
947 c.maxStackScan.Add(pp.maxStackScanDelta)
948 pp.maxStackScanDelta = 0
952 func (c *gcControllerState) addGlobals(amount int64) {
953 c.globalsScan.Add(amount)
956 // heapGoal returns the current heap goal.
957 func (c *gcControllerState) heapGoal() uint64 {
958 goal, _ := c.heapGoalInternal()
962 // heapGoalInternal is the implementation of heapGoal which returns additional
963 // information that is necessary for computing the trigger.
965 // The returned minTrigger is always <= goal.
966 func (c *gcControllerState) heapGoalInternal() (goal, minTrigger uint64) {
967 // Start with the goal calculated for gcPercent.
968 goal = c.gcPercentHeapGoal.Load()
970 // Check if the memory-limit-based goal is smaller, and if so, pick that.
971 if newGoal := c.memoryLimitHeapGoal(); go119MemoryLimitSupport && newGoal < goal {
974 // We're not limited by the memory limit goal, so perform a series of
975 // adjustments that might move the goal forward in a variety of circumstances.
977 sweepDistTrigger := c.sweepDistMinTrigger.Load()
978 if sweepDistTrigger > goal {
979 // Set the goal to maintain a minimum sweep distance since
980 // the last call to commit. Note that we never want to do this
981 // if we're in the memory limit regime, because it could push
983 goal = sweepDistTrigger
985 // Since we ignore the sweep distance trigger in the memory
986 // limit regime, we need to ensure we don't propagate it to
987 // the trigger, because it could cause a violation of the
988 // invariant that the trigger < goal.
989 minTrigger = sweepDistTrigger
991 // Ensure that the heap goal is at least a little larger than
992 // the point at which we triggered. This may not be the case if GC
993 // start is delayed or if the allocation that pushed gcController.heapLive
994 // over trigger is large or if the trigger is really close to
995 // GOGC. Assist is proportional to this distance, so enforce a
996 // minimum distance, even if it means going over the GOGC goal
999 // Ignore this if we're in the memory limit regime: we'd prefer to
1000 // have the GC respond hard about how close we are to the goal than to
1001 // push the goal back in such a manner that it could cause us to exceed
1002 // the memory limit.
1003 const minRunway = 64 << 10
1004 if c.triggered != ^uint64(0) && goal < c.triggered+minRunway {
1005 goal = c.triggered + minRunway
1011 // memoryLimitHeapGoal returns a heap goal derived from memoryLimit.
1012 func (c *gcControllerState) memoryLimitHeapGoal() uint64 {
1013 // Start by pulling out some values we'll need. Be careful about overflow.
1014 var heapFree, heapAlloc, mappedReady uint64
1016 heapFree = c.heapFree.load() // Free and unscavenged memory.
1017 heapAlloc = c.totalAlloc.Load() - c.totalFree.Load() // Heap object bytes in use.
1018 mappedReady = c.mappedReady.Load() // Total unreleased mapped memory.
1019 if heapFree+heapAlloc <= mappedReady {
1022 // It is impossible for total unreleased mapped memory to exceed heap memory, but
1023 // because these stats are updated independently, we may observe a partial update
1024 // including only some values. Thus, we appear to break the invariant. However,
1025 // this condition is necessarily transient, so just try again. In the case of a
1026 // persistent accounting error, we'll deadlock here.
1029 // Below we compute a goal from memoryLimit. There are a few things to be aware of.
1030 // Firstly, the memoryLimit does not easily compare to the heap goal: the former
1031 // is total mapped memory by the runtime that hasn't been released, while the latter is
1032 // only heap object memory. Intuitively, the way we convert from one to the other is to
1033 // subtract everything from memoryLimit that both contributes to the memory limit (so,
1034 // ignore scavenged memory) and doesn't contain heap objects. This isn't quite what
1035 // lines up with reality, but it's a good starting point.
1037 // In practice this computation looks like the following:
1039 // memoryLimit - ((mappedReady - heapFree - heapAlloc) + max(mappedReady - memoryLimit, 0)) - memoryLimitHeapGoalHeadroom
1042 // Let's break this down.
1044 // The first term (marker 1) is everything that contributes to the memory limit and isn't
1045 // or couldn't become heap objects. It represents, broadly speaking, non-heap overheads.
1046 // One oddity you may have noticed is that we also subtract out heapFree, i.e. unscavenged
1047 // memory that may contain heap objects in the future.
1049 // Let's take a step back. In an ideal world, this term would look something like just
1050 // the heap goal. That is, we "reserve" enough space for the heap to grow to the heap
1051 // goal, and subtract out everything else. This is of course impossible; the definition
1052 // is circular! However, this impossible definition contains a key insight: the amount
1053 // we're *going* to use matters just as much as whatever we're currently using.
1055 // Consider if the heap shrinks to 1/10th its size, leaving behind lots of free and
1056 // unscavenged memory. mappedReady - heapAlloc will be quite large, because of that free
1057 // and unscavenged memory, pushing the goal down significantly.
1059 // heapFree is also safe to exclude from the memory limit because in the steady-state, it's
1060 // just a pool of memory for future heap allocations, and making new allocations from heapFree
1061 // memory doesn't increase overall memory use. In transient states, the scavenger and the
1062 // allocator actively manage the pool of heapFree memory to maintain the memory limit.
1064 // The second term (marker 2) is the amount of memory we've exceeded the limit by, and is
1065 // intended to help recover from such a situation. By pushing the heap goal down, we also
1066 // push the trigger down, triggering and finishing a GC sooner in order to make room for
1067 // other memory sources. Note that since we're effectively reducing the heap goal by X bytes,
1068 // we're actually giving more than X bytes of headroom back, because the heap goal is in
1069 // terms of heap objects, but it takes more than X bytes (e.g. due to fragmentation) to store
1070 // X bytes worth of objects.
1072 // The third term (marker 3) subtracts an additional memoryLimitHeapGoalHeadroom bytes from the
1073 // heap goal. As the name implies, this is to provide additional headroom in the face of pacing
1074 // inaccuracies. This is a fixed number of bytes because these inaccuracies disproportionately
1075 // affect small heaps: as heaps get smaller, the pacer's inputs get fuzzier. Shorter GC cycles
1076 // and less GC work means noisy external factors like the OS scheduler have a greater impact.
1078 memoryLimit := uint64(c.memoryLimit.Load())
1081 nonHeapMemory := mappedReady - heapFree - heapAlloc
1085 if mappedReady > memoryLimit {
1086 overage = mappedReady - memoryLimit
1089 if nonHeapMemory+overage >= memoryLimit {
1090 // We're at a point where non-heap memory exceeds the memory limit on its own.
1091 // There's honestly not much we can do here but just trigger GCs continuously
1092 // and let the CPU limiter reign that in. Something has to give at this point.
1093 // Set it to heapMarked, the lowest possible goal.
1097 // Compute the goal.
1098 goal := memoryLimit - (nonHeapMemory + overage)
1100 // Apply some headroom to the goal to account for pacing inaccuracies.
1101 // Be careful about small limits.
1102 if goal < memoryLimitHeapGoalHeadroom || goal-memoryLimitHeapGoalHeadroom < memoryLimitHeapGoalHeadroom {
1103 goal = memoryLimitHeapGoalHeadroom
1105 goal = goal - memoryLimitHeapGoalHeadroom
1107 // Don't let us go below the live heap. A heap goal below the live heap doesn't make sense.
1108 if goal < c.heapMarked {
1115 // These constants determine the bounds on the GC trigger as a fraction
1116 // of heap bytes allocated between the start of a GC (heapLive == heapMarked)
1117 // and the end of a GC (heapLive == heapGoal).
1119 // The constants are obscured in this way for efficiency. The denominator
1120 // of the fraction is always a power-of-two for a quick division, so that
1121 // the numerator is a single constant integer multiplication.
1122 triggerRatioDen = 64
1124 // The minimum trigger constant was chosen empirically: given a sufficiently
1125 // fast/scalable allocator with 48 Ps that could drive the trigger ratio
1126 // to <0.05, this constant causes applications to retain the same peak
1127 // RSS compared to not having this allocator.
1128 minTriggerRatioNum = 45 // ~0.7
1130 // The maximum trigger constant is chosen somewhat arbitrarily, but the
1131 // current constant has served us well over the years.
1132 maxTriggerRatioNum = 61 // ~0.95
1135 // trigger returns the current point at which a GC should trigger along with
1138 // The returned value may be compared against heapLive to determine whether
1139 // the GC should trigger. Thus, the GC trigger condition should be (but may
1140 // not be, in the case of small movements for efficiency) checked whenever
1141 // the heap goal may change.
1142 func (c *gcControllerState) trigger() (uint64, uint64) {
1143 goal, minTrigger := c.heapGoalInternal()
1145 // Invariant: the trigger must always be less than the heap goal.
1147 // Note that the memory limit sets a hard maximum on our heap goal,
1148 // but the live heap may grow beyond it.
1150 if c.heapMarked >= goal {
1151 // The goal should never be smaller than heapMarked, but let's be
1152 // defensive about it. The only reasonable trigger here is one that
1153 // causes a continuous GC cycle at heapMarked, but respect the goal
1154 // if it came out as smaller than that.
1158 // Below this point, c.heapMarked < goal.
1160 // heapMarked is our absolute minimum, and it's possible the trigger
1161 // bound we get from heapGoalinternal is less than that.
1162 if minTrigger < c.heapMarked {
1163 minTrigger = c.heapMarked
1166 // If we let the trigger go too low, then if the application
1167 // is allocating very rapidly we might end up in a situation
1168 // where we're allocating black during a nearly always-on GC.
1169 // The result of this is a growing heap and ultimately an
1170 // increase in RSS. By capping us at a point >0, we're essentially
1171 // saying that we're OK using more CPU during the GC to prevent
1172 // this growth in RSS.
1173 triggerLowerBound := uint64(((goal-c.heapMarked)/triggerRatioDen)*minTriggerRatioNum) + c.heapMarked
1174 if minTrigger < triggerLowerBound {
1175 minTrigger = triggerLowerBound
1178 // For small heaps, set the max trigger point at maxTriggerRatio of the way
1179 // from the live heap to the heap goal. This ensures we always have *some*
1180 // headroom when the GC actually starts. For larger heaps, set the max trigger
1181 // point at the goal, minus the minimum heap size.
1183 // This choice follows from the fact that the minimum heap size is chosen
1184 // to reflect the costs of a GC with no work to do. With a large heap but
1185 // very little scan work to perform, this gives us exactly as much runway
1186 // as we would need, in the worst case.
1187 maxTrigger := uint64(((goal-c.heapMarked)/triggerRatioDen)*maxTriggerRatioNum) + c.heapMarked
1188 if goal > defaultHeapMinimum && goal-defaultHeapMinimum > maxTrigger {
1189 maxTrigger = goal - defaultHeapMinimum
1191 if maxTrigger < minTrigger {
1192 maxTrigger = minTrigger
1195 // Compute the trigger from our bounds and the runway stored by commit.
1197 runway := c.runway.Load()
1199 trigger = minTrigger
1201 trigger = goal - runway
1203 if trigger < minTrigger {
1204 trigger = minTrigger
1206 if trigger > maxTrigger {
1207 trigger = maxTrigger
1210 print("trigger=", trigger, " heapGoal=", goal, "\n")
1211 print("minTrigger=", minTrigger, " maxTrigger=", maxTrigger, "\n")
1212 throw("produced a trigger greater than the heap goal")
1214 return trigger, goal
1217 // commit recomputes all pacing parameters needed to derive the
1218 // trigger and the heap goal. Namely, the gcPercent-based heap goal,
1219 // and the amount of runway we want to give the GC this cycle.
1221 // This can be called any time. If GC is the in the middle of a
1222 // concurrent phase, it will adjust the pacing of that phase.
1224 // isSweepDone should be the result of calling isSweepDone(),
1225 // unless we're testing or we know we're executing during a GC cycle.
1227 // This depends on gcPercent, gcController.heapMarked, and
1228 // gcController.heapLive. These must be up to date.
1230 // Callers must call gcControllerState.revise after calling this
1231 // function if the GC is enabled.
1233 // mheap_.lock must be held or the world must be stopped.
1234 func (c *gcControllerState) commit(isSweepDone bool) {
1236 assertWorldStoppedOrLockHeld(&mheap_.lock)
1240 // The sweep is done, so there aren't any restrictions on the trigger
1241 // we need to think about.
1242 c.sweepDistMinTrigger.Store(0)
1244 // Concurrent sweep happens in the heap growth
1245 // from gcController.heapLive to trigger. Make sure we
1246 // give the sweeper some runway if it doesn't have enough.
1247 c.sweepDistMinTrigger.Store(c.heapLive.Load() + sweepMinHeapDistance)
1250 // Compute the next GC goal, which is when the allocated heap
1251 // has grown by GOGC/100 over where it started the last cycle,
1252 // plus additional runway for non-heap sources of GC work.
1253 gcPercentHeapGoal := ^uint64(0)
1254 if gcPercent := c.gcPercent.Load(); gcPercent >= 0 {
1255 gcPercentHeapGoal = c.heapMarked + (c.heapMarked+c.lastStackScan.Load()+c.globalsScan.Load())*uint64(gcPercent)/100
1257 // Apply the minimum heap size here. It's defined in terms of gcPercent
1258 // and is only updated by functions that call commit.
1259 if gcPercentHeapGoal < c.heapMinimum {
1260 gcPercentHeapGoal = c.heapMinimum
1262 c.gcPercentHeapGoal.Store(gcPercentHeapGoal)
1264 // Compute the amount of runway we want the GC to have by using our
1265 // estimate of the cons/mark ratio.
1267 // The idea is to take our expected scan work, and multiply it by
1268 // the cons/mark ratio to determine how long it'll take to complete
1269 // that scan work in terms of bytes allocated. This gives us our GC's
1272 // However, the cons/mark ratio is a ratio of rates per CPU-second, but
1273 // here we care about the relative rates for some division of CPU
1274 // resources among the mutator and the GC.
1276 // To summarize, we have B / cpu-ns, and we want B / ns. We get that
1277 // by multiplying by our desired division of CPU resources. We choose
1278 // to express CPU resources as GOMAPROCS*fraction. Note that because
1279 // we're working with a ratio here, we can omit the number of CPU cores,
1280 // because they'll appear in the numerator and denominator and cancel out.
1281 // As a result, this is basically just "weighing" the cons/mark ratio by
1282 // our desired division of resources.
1284 // Furthermore, by setting the runway so that CPU resources are divided
1285 // this way, assuming that the cons/mark ratio is correct, we make that
1286 // division a reality.
1287 c.runway.Store(uint64((c.consMark * (1 - gcGoalUtilization) / (gcGoalUtilization)) * float64(c.lastHeapScan+c.lastStackScan.Load()+c.globalsScan.Load())))
1290 // setGCPercent updates gcPercent. commit must be called after.
1291 // Returns the old value of gcPercent.
1293 // The world must be stopped, or mheap_.lock must be held.
1294 func (c *gcControllerState) setGCPercent(in int32) int32 {
1296 assertWorldStoppedOrLockHeld(&mheap_.lock)
1299 out := c.gcPercent.Load()
1303 c.heapMinimum = defaultHeapMinimum * uint64(in) / 100
1304 c.gcPercent.Store(in)
1309 //go:linkname setGCPercent runtime/debug.setGCPercent
1310 func setGCPercent(in int32) (out int32) {
1311 // Run on the system stack since we grab the heap lock.
1312 systemstack(func() {
1314 out = gcController.setGCPercent(in)
1315 gcControllerCommit()
1316 unlock(&mheap_.lock)
1319 // If we just disabled GC, wait for any concurrent GC mark to
1320 // finish so we always return with no GC running.
1322 gcWaitOnMark(atomic.Load(&work.cycles))
1328 func readGOGC() int32 {
1329 p := gogetenv("GOGC")
1333 if n, ok := atoi32(p); ok {
1339 // setMemoryLimit updates memoryLimit. commit must be called after
1340 // Returns the old value of memoryLimit.
1342 // The world must be stopped, or mheap_.lock must be held.
1343 func (c *gcControllerState) setMemoryLimit(in int64) int64 {
1345 assertWorldStoppedOrLockHeld(&mheap_.lock)
1348 out := c.memoryLimit.Load()
1350 c.memoryLimit.Store(in)
1356 //go:linkname setMemoryLimit runtime/debug.setMemoryLimit
1357 func setMemoryLimit(in int64) (out int64) {
1358 // Run on the system stack since we grab the heap lock.
1359 systemstack(func() {
1361 out = gcController.setMemoryLimit(in)
1362 if in < 0 || out == in {
1363 // If we're just checking the value or not changing
1364 // it, there's no point in doing the rest.
1365 unlock(&mheap_.lock)
1368 gcControllerCommit()
1369 unlock(&mheap_.lock)
1374 func readGOMEMLIMIT() int64 {
1375 p := gogetenv("GOMEMLIMIT")
1376 if p == "" || p == "off" {
1379 n, ok := parseByteCount(p)
1381 print("GOMEMLIMIT=", p, "\n")
1382 throw("malformed GOMEMLIMIT; see `go doc runtime/debug.SetMemoryLimit`")
1387 type piController struct {
1388 kp float64 // Proportional constant.
1389 ti float64 // Integral time constant.
1390 tt float64 // Reset time.
1392 min, max float64 // Output boundaries.
1394 // PI controller state.
1396 errIntegral float64 // Integral of the error from t=0 to now.
1399 errOverflow bool // Set if errIntegral ever overflowed.
1400 inputOverflow bool // Set if an operation with the input overflowed.
1403 // next provides a new sample to the controller.
1405 // input is the sample, setpoint is the desired point, and period is how much
1406 // time (in whatever unit makes the most sense) has passed since the last sample.
1408 // Returns a new value for the variable it's controlling, and whether the operation
1409 // completed successfully. One reason this might fail is if error has been growing
1410 // in an unbounded manner, to the point of overflow.
1412 // In the specific case of an error overflow occurs, the errOverflow field will be
1413 // set and the rest of the controller's internal state will be fully reset.
1414 func (c *piController) next(input, setpoint, period float64) (float64, bool) {
1415 // Compute the raw output value.
1416 prop := c.kp * (setpoint - input)
1417 rawOutput := prop + c.errIntegral
1419 // Clamp rawOutput into output.
1421 if isInf(output) || isNaN(output) {
1422 // The input had a large enough magnitude that either it was already
1423 // overflowed, or some operation with it overflowed.
1424 // Set a flag and reset. That's the safest thing to do.
1426 c.inputOverflow = true
1431 } else if output > c.max {
1435 // Update the controller's state.
1436 if c.ti != 0 && c.tt != 0 {
1437 c.errIntegral += (c.kp*period/c.ti)*(setpoint-input) + (period/c.tt)*(output-rawOutput)
1438 if isInf(c.errIntegral) || isNaN(c.errIntegral) {
1439 // So much error has accumulated that we managed to overflow.
1440 // The assumptions around the controller have likely broken down.
1441 // Set a flag and reset. That's the safest thing to do.
1443 c.errOverflow = true
1450 // reset resets the controller state, except for controller error flags.
1451 func (c *piController) reset() {
1455 // addIdleMarkWorker attempts to add a new idle mark worker.
1457 // If this returns true, the caller must become an idle mark worker unless
1458 // there's no background mark worker goroutines in the pool. This case is
1459 // harmless because there are already background mark workers running.
1460 // If this returns false, the caller must NOT become an idle mark worker.
1462 // nosplit because it may be called without a P.
1465 func (c *gcControllerState) addIdleMarkWorker() bool {
1467 old := c.idleMarkWorkers.Load()
1468 n, max := int32(old&uint64(^uint32(0))), int32(old>>32)
1470 // See the comment on idleMarkWorkers for why
1471 // n > max is tolerated.
1475 print("n=", n, " max=", max, "\n")
1476 throw("negative idle mark workers")
1478 new := uint64(uint32(n+1)) | (uint64(max) << 32)
1479 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1485 // needIdleMarkWorker is a hint as to whether another idle mark worker is needed.
1487 // The caller must still call addIdleMarkWorker to become one. This is mainly
1488 // useful for a quick check before an expensive operation.
1490 // nosplit because it may be called without a P.
1493 func (c *gcControllerState) needIdleMarkWorker() bool {
1494 p := c.idleMarkWorkers.Load()
1495 n, max := int32(p&uint64(^uint32(0))), int32(p>>32)
1499 // removeIdleMarkWorker must be called when an new idle mark worker stops executing.
1500 func (c *gcControllerState) removeIdleMarkWorker() {
1502 old := c.idleMarkWorkers.Load()
1503 n, max := int32(old&uint64(^uint32(0))), int32(old>>32)
1505 print("n=", n, " max=", max, "\n")
1506 throw("negative idle mark workers")
1508 new := uint64(uint32(n-1)) | (uint64(max) << 32)
1509 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1515 // setMaxIdleMarkWorkers sets the maximum number of idle mark workers allowed.
1517 // This method is optimistic in that it does not wait for the number of
1518 // idle mark workers to reduce to max before returning; it assumes the workers
1519 // will deschedule themselves.
1520 func (c *gcControllerState) setMaxIdleMarkWorkers(max int32) {
1522 old := c.idleMarkWorkers.Load()
1523 n := int32(old & uint64(^uint32(0)))
1525 print("n=", n, " max=", max, "\n")
1526 throw("negative idle mark workers")
1528 new := uint64(uint32(n)) | (uint64(max) << 32)
1529 if c.idleMarkWorkers.CompareAndSwap(old, new) {
1535 // gcControllerCommit is gcController.commit, but passes arguments from live
1536 // (non-test) data. It also updates any consumers of the GC pacing, such as
1537 // sweep pacing and the background scavenger.
1539 // Calls gcController.commit.
1541 // The heap lock must be held, so this must be executed on the system stack.
1544 func gcControllerCommit() {
1545 assertWorldStoppedOrLockHeld(&mheap_.lock)
1547 gcController.commit(isSweepDone())
1549 // Update mark pacing.
1550 if gcphase != _GCoff {
1551 gcController.revise()
1554 // TODO(mknyszek): This isn't really accurate any longer because the heap
1555 // goal is computed dynamically. Still useful to snapshot, but not as useful.
1560 trigger, heapGoal := gcController.trigger()
1561 gcPaceSweeper(trigger)
1562 gcPaceScavenger(gcController.memoryLimit.Load(), heapGoal, gcController.lastHeapGoal)