diff options
Diffstat (limited to 'src/runtime/mgcpacer.go')
| -rw-r--r-- | src/runtime/mgcpacer.go | 1348 |
1 files changed, 1348 insertions, 0 deletions
diff --git a/src/runtime/mgcpacer.go b/src/runtime/mgcpacer.go new file mode 100644 index 0000000..d54dbc2 --- /dev/null +++ b/src/runtime/mgcpacer.go @@ -0,0 +1,1348 @@ +// Copyright 2021 The Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +package runtime + +import ( + "internal/cpu" + "internal/goexperiment" + "runtime/internal/atomic" + "unsafe" +) + +const ( + // gcGoalUtilization is the goal CPU utilization for + // marking as a fraction of GOMAXPROCS. + gcGoalUtilization = goexperiment.PacerRedesignInt*gcBackgroundUtilization + + (1-goexperiment.PacerRedesignInt)*(gcBackgroundUtilization+0.05) + + // gcBackgroundUtilization is the fixed CPU utilization for background + // marking. It must be <= gcGoalUtilization. The difference between + // gcGoalUtilization and gcBackgroundUtilization will be made up by + // mark assists. The scheduler will aim to use within 50% of this + // goal. + // + // Setting this to < gcGoalUtilization avoids saturating the trigger + // feedback controller when there are no assists, which allows it to + // better control CPU and heap growth. However, the larger the gap, + // the more mutator assists are expected to happen, which impact + // mutator latency. + // + // If goexperiment.PacerRedesign, the trigger feedback controller + // is replaced with an estimate of the mark/cons ratio that doesn't + // have the same saturation issues, so this is set equal to + // gcGoalUtilization. + gcBackgroundUtilization = 0.25 + + // gcCreditSlack is the amount of scan work credit that can + // accumulate locally before updating gcController.heapScanWork and, + // optionally, gcController.bgScanCredit. Lower values give a more + // accurate assist ratio and make it more likely that assists will + // successfully steal background credit. Higher values reduce memory + // contention. + gcCreditSlack = 2000 + + // gcAssistTimeSlack is the nanoseconds of mutator assist time that + // can accumulate on a P before updating gcController.assistTime. + gcAssistTimeSlack = 5000 + + // gcOverAssistWork determines how many extra units of scan work a GC + // assist does when an assist happens. This amortizes the cost of an + // assist by pre-paying for this many bytes of future allocations. + gcOverAssistWork = 64 << 10 + + // defaultHeapMinimum is the value of heapMinimum for GOGC==100. + defaultHeapMinimum = (goexperiment.HeapMinimum512KiBInt)*(512<<10) + + (1-goexperiment.HeapMinimum512KiBInt)*(4<<20) + + // scannableStackSizeSlack is the bytes of stack space allocated or freed + // that can accumulate on a P before updating gcController.stackSize. + scannableStackSizeSlack = 8 << 10 +) + +func init() { + if offset := unsafe.Offsetof(gcController.heapLive); offset%8 != 0 { + println(offset) + throw("gcController.heapLive not aligned to 8 bytes") + } +} + +// gcController implements the GC pacing controller that determines +// when to trigger concurrent garbage collection and how much marking +// work to do in mutator assists and background marking. +// +// It uses a feedback control algorithm to adjust the gcController.trigger +// trigger based on the heap growth and GC CPU utilization each cycle. +// This algorithm optimizes for heap growth to match GOGC and for CPU +// utilization between assist and background marking to be 25% of +// GOMAXPROCS. The high-level design of this algorithm is documented +// at https://golang.org/s/go15gcpacing. +// +// All fields of gcController are used only during a single mark +// cycle. +var gcController gcControllerState + +type gcControllerState struct { + + // Initialized from GOGC. GOGC=off means no GC. + gcPercent atomic.Int32 + + _ uint32 // padding so following 64-bit values are 8-byte aligned + + // heapMinimum is the minimum heap size at which to trigger GC. + // For small heaps, this overrides the usual GOGC*live set rule. + // + // When there is a very small live set but a lot of allocation, simply + // collecting when the heap reaches GOGC*live results in many GC + // cycles and high total per-GC overhead. This minimum amortizes this + // per-GC overhead while keeping the heap reasonably small. + // + // During initialization this is set to 4MB*GOGC/100. In the case of + // GOGC==0, this will set heapMinimum to 0, resulting in constant + // collection even when the heap size is small, which is useful for + // debugging. + heapMinimum uint64 + + // triggerRatio is the heap growth ratio that triggers marking. + // + // E.g., if this is 0.6, then GC should start when the live + // heap has reached 1.6 times the heap size marked by the + // previous cycle. This should be ≤ GOGC/100 so the trigger + // heap size is less than the goal heap size. This is set + // during mark termination for the next cycle's trigger. + // + // Protected by mheap_.lock or a STW. + // + // Used if !goexperiment.PacerRedesign. + triggerRatio float64 + + // trigger is the heap size that triggers marking. + // + // When heapLive ≥ trigger, the mark phase will start. + // This is also the heap size by which proportional sweeping + // must be complete. + // + // This is computed from triggerRatio during mark termination + // for the next cycle's trigger. + // + // Protected by mheap_.lock or a STW. + trigger uint64 + + // consMark is the estimated per-CPU consMark ratio for the application. + // + // It represents the ratio between the application's allocation + // rate, as bytes allocated per CPU-time, and the GC's scan rate, + // as bytes scanned per CPU-time. + // The units of this ratio are (B / cpu-ns) / (B / cpu-ns). + // + // At a high level, this value is computed as the bytes of memory + // allocated (cons) per unit of scan work completed (mark) in a GC + // cycle, divided by the CPU time spent on each activity. + // + // Updated at the end of each GC cycle, in endCycle. + // + // For goexperiment.PacerRedesign. + consMark float64 + + // consMarkController holds the state for the mark-cons ratio + // estimation over time. + // + // Its purpose is to smooth out noisiness in the computation of + // consMark; see consMark for details. + // + // For goexperiment.PacerRedesign. + consMarkController piController + + _ uint32 // Padding for atomics on 32-bit platforms. + + // heapGoal is the goal heapLive for when next GC ends. + // Set to ^uint64(0) if disabled. + // + // Read and written atomically, unless the world is stopped. + heapGoal uint64 + + // lastHeapGoal is the value of heapGoal for the previous GC. + // Note that this is distinct from the last value heapGoal had, + // because it could change if e.g. gcPercent changes. + // + // Read and written with the world stopped or with mheap_.lock held. + lastHeapGoal uint64 + + // heapLive is the number of bytes considered live by the GC. + // That is: retained by the most recent GC plus allocated + // since then. heapLive ≤ memstats.heapAlloc, since heapAlloc includes + // unmarked objects that have not yet been swept (and hence goes up as we + // allocate and down as we sweep) while heapLive excludes these + // objects (and hence only goes up between GCs). + // + // This is updated atomically without locking. To reduce + // contention, this is updated only when obtaining a span from + // an mcentral and at this point it counts all of the + // unallocated slots in that span (which will be allocated + // before that mcache obtains another span from that + // mcentral). Hence, it slightly overestimates the "true" live + // heap size. It's better to overestimate than to + // underestimate because 1) this triggers the GC earlier than + // necessary rather than potentially too late and 2) this + // leads to a conservative GC rate rather than a GC rate that + // is potentially too low. + // + // Reads should likewise be atomic (or during STW). + // + // Whenever this is updated, call traceHeapAlloc() and + // this gcControllerState's revise() method. + heapLive uint64 + + // heapScan is the number of bytes of "scannable" heap. This + // is the live heap (as counted by heapLive), but omitting + // no-scan objects and no-scan tails of objects. + // + // For !goexperiment.PacerRedesign: Whenever this is updated, + // call this gcControllerState's revise() method. It is read + // and written atomically or with the world stopped. + // + // For goexperiment.PacerRedesign: This value is fixed at the + // start of a GC cycle, so during a GC cycle it is safe to + // read without atomics, and it represents the maximum scannable + // heap. + heapScan uint64 + + // lastHeapScan is the number of bytes of heap that were scanned + // last GC cycle. It is the same as heapMarked, but only + // includes the "scannable" parts of objects. + // + // Updated when the world is stopped. + lastHeapScan uint64 + + // stackScan is a snapshot of scannableStackSize taken at each GC + // STW pause and is used in pacing decisions. + // + // Updated only while the world is stopped. + stackScan uint64 + + // scannableStackSize is the amount of allocated goroutine stack space in + // use by goroutines. + // + // This number tracks allocated goroutine stack space rather than used + // goroutine stack space (i.e. what is actually scanned) because used + // goroutine stack space is much harder to measure cheaply. By using + // allocated space, we make an overestimate; this is OK, it's better + // to conservatively overcount than undercount. + // + // Read and updated atomically. + scannableStackSize uint64 + + // globalsScan is the total amount of global variable space + // that is scannable. + // + // Read and updated atomically. + globalsScan uint64 + + // heapMarked is the number of bytes marked by the previous + // GC. After mark termination, heapLive == heapMarked, but + // unlike heapLive, heapMarked does not change until the + // next mark termination. + heapMarked uint64 + + // heapScanWork is the total heap scan work performed this cycle. + // stackScanWork is the total stack scan work performed this cycle. + // globalsScanWork is the total globals scan work performed this cycle. + // + // These are updated atomically during the cycle. Updates occur in + // bounded batches, since they are both written and read + // throughout the cycle. At the end of the cycle, heapScanWork is how + // much of the retained heap is scannable. + // + // Currently these are measured in bytes. For most uses, this is an + // opaque unit of work, but for estimation the definition is important. + // + // Note that stackScanWork includes all allocated space, not just the + // size of the stack itself, mirroring stackSize. + // + // For !goexperiment.PacerRedesign, stackScanWork and globalsScanWork + // are always zero. + heapScanWork atomic.Int64 + stackScanWork atomic.Int64 + globalsScanWork atomic.Int64 + + // bgScanCredit is the scan work credit accumulated by the + // concurrent background scan. This credit is accumulated by + // the background scan and stolen by mutator assists. This is + // updated atomically. Updates occur in bounded batches, since + // it is both written and read throughout the cycle. + bgScanCredit int64 + + // assistTime is the nanoseconds spent in mutator assists + // during this cycle. This is updated atomically. Updates + // occur in bounded batches, since it is both written and read + // throughout the cycle. + assistTime int64 + + // dedicatedMarkTime is the nanoseconds spent in dedicated + // mark workers during this cycle. This is updated atomically + // at the end of the concurrent mark phase. + dedicatedMarkTime int64 + + // fractionalMarkTime is the nanoseconds spent in the + // fractional mark worker during this cycle. This is updated + // atomically throughout the cycle and will be up-to-date if + // the fractional mark worker is not currently running. + fractionalMarkTime int64 + + // idleMarkTime is the nanoseconds spent in idle marking + // during this cycle. This is updated atomically throughout + // the cycle. + idleMarkTime int64 + + // markStartTime is the absolute start time in nanoseconds + // that assists and background mark workers started. + markStartTime int64 + + // dedicatedMarkWorkersNeeded is the number of dedicated mark + // workers that need to be started. This is computed at the + // beginning of each cycle and decremented atomically as + // dedicated mark workers get started. + dedicatedMarkWorkersNeeded int64 + + // assistWorkPerByte is the ratio of scan work to allocated + // bytes that should be performed by mutator assists. This is + // computed at the beginning of each cycle and updated every + // time heapScan is updated. + assistWorkPerByte atomic.Float64 + + // assistBytesPerWork is 1/assistWorkPerByte. + // + // Note that because this is read and written independently + // from assistWorkPerByte users may notice a skew between + // the two values, and such a state should be safe. + assistBytesPerWork atomic.Float64 + + // fractionalUtilizationGoal is the fraction of wall clock + // time that should be spent in the fractional mark worker on + // each P that isn't running a dedicated worker. + // + // For example, if the utilization goal is 25% and there are + // no dedicated workers, this will be 0.25. If the goal is + // 25%, there is one dedicated worker, and GOMAXPROCS is 5, + // this will be 0.05 to make up the missing 5%. + // + // If this is zero, no fractional workers are needed. + fractionalUtilizationGoal float64 + + // test indicates that this is a test-only copy of gcControllerState. + test bool + + _ cpu.CacheLinePad +} + +func (c *gcControllerState) init(gcPercent int32) { + c.heapMinimum = defaultHeapMinimum + + if goexperiment.PacerRedesign { + c.consMarkController = piController{ + // Tuned first via the Ziegler-Nichols process in simulation, + // then the integral time was manually tuned against real-world + // applications to deal with noisiness in the measured cons/mark + // ratio. + kp: 0.9, + ti: 4.0, + + // Set a high reset time in GC cycles. + // This is inversely proportional to the rate at which we + // accumulate error from clipping. By making this very high + // we make the accumulation slow. In general, clipping is + // OK in our situation, hence the choice. + // + // Tune this if we get unintended effects from clipping for + // a long time. + tt: 1000, + min: -1000, + max: 1000, + } + } else { + // Set a reasonable initial GC trigger. + c.triggerRatio = 7 / 8.0 + + // Fake a heapMarked value so it looks like a trigger at + // heapMinimum is the appropriate growth from heapMarked. + // This will go into computing the initial GC goal. + c.heapMarked = uint64(float64(c.heapMinimum) / (1 + c.triggerRatio)) + } + + // This will also compute and set the GC trigger and goal. + c.setGCPercent(gcPercent) +} + +// startCycle resets the GC controller's state and computes estimates +// for a new GC cycle. The caller must hold worldsema and the world +// must be stopped. +func (c *gcControllerState) startCycle(markStartTime int64, procs int) { + c.heapScanWork.Store(0) + c.stackScanWork.Store(0) + c.globalsScanWork.Store(0) + c.bgScanCredit = 0 + c.assistTime = 0 + c.dedicatedMarkTime = 0 + c.fractionalMarkTime = 0 + c.idleMarkTime = 0 + c.markStartTime = markStartTime + c.stackScan = atomic.Load64(&c.scannableStackSize) + + // Ensure that the heap goal is at least a little larger than + // the current live heap size. This may not be the case if GC + // start is delayed or if the allocation that pushed gcController.heapLive + // over trigger is large or if the trigger is really close to + // GOGC. Assist is proportional to this distance, so enforce a + // minimum distance, even if it means going over the GOGC goal + // by a tiny bit. + if goexperiment.PacerRedesign { + if c.heapGoal < c.heapLive+64<<10 { + c.heapGoal = c.heapLive + 64<<10 + } + } else { + if c.heapGoal < c.heapLive+1<<20 { + c.heapGoal = c.heapLive + 1<<20 + } + } + + // Compute the background mark utilization goal. In general, + // this may not come out exactly. We round the number of + // dedicated workers so that the utilization is closest to + // 25%. For small GOMAXPROCS, this would introduce too much + // error, so we add fractional workers in that case. + totalUtilizationGoal := float64(procs) * gcBackgroundUtilization + c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5) + utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1 + const maxUtilError = 0.3 + if utilError < -maxUtilError || utilError > maxUtilError { + // Rounding put us more than 30% off our goal. With + // gcBackgroundUtilization of 25%, this happens for + // GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional + // workers to compensate. + if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal { + // Too many dedicated workers. + c.dedicatedMarkWorkersNeeded-- + } + c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(procs) + } else { + c.fractionalUtilizationGoal = 0 + } + + // In STW mode, we just want dedicated workers. + if debug.gcstoptheworld > 0 { + c.dedicatedMarkWorkersNeeded = int64(procs) + c.fractionalUtilizationGoal = 0 + } + + // Clear per-P state + for _, p := range allp { + p.gcAssistTime = 0 + p.gcFractionalMarkTime = 0 + } + + // Compute initial values for controls that are updated + // throughout the cycle. + c.revise() + + if debug.gcpacertrace > 0 { + assistRatio := c.assistWorkPerByte.Load() + print("pacer: assist ratio=", assistRatio, + " (scan ", gcController.heapScan>>20, " MB in ", + work.initialHeapLive>>20, "->", + c.heapGoal>>20, " MB)", + " workers=", c.dedicatedMarkWorkersNeeded, + "+", c.fractionalUtilizationGoal, "\n") + } +} + +// revise updates the assist ratio during the GC cycle to account for +// improved estimates. This should be called whenever gcController.heapScan, +// gcController.heapLive, or gcController.heapGoal is updated. It is safe to +// call concurrently, but it may race with other calls to revise. +// +// The result of this race is that the two assist ratio values may not line +// up or may be stale. In practice this is OK because the assist ratio +// moves slowly throughout a GC cycle, and the assist ratio is a best-effort +// heuristic anyway. Furthermore, no part of the heuristic depends on +// the two assist ratio values being exact reciprocals of one another, since +// the two values are used to convert values from different sources. +// +// The worst case result of this raciness is that we may miss a larger shift +// in the ratio (say, if we decide to pace more aggressively against the +// hard heap goal) but even this "hard goal" is best-effort (see #40460). +// The dedicated GC should ensure we don't exceed the hard goal by too much +// in the rare case we do exceed it. +// +// It should only be called when gcBlackenEnabled != 0 (because this +// is when assists are enabled and the necessary statistics are +// available). +func (c *gcControllerState) revise() { + gcPercent := c.gcPercent.Load() + if gcPercent < 0 { + // If GC is disabled but we're running a forced GC, + // act like GOGC is huge for the below calculations. + gcPercent = 100000 + } + live := atomic.Load64(&c.heapLive) + scan := atomic.Load64(&c.heapScan) + work := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load() + + // Assume we're under the soft goal. Pace GC to complete at + // heapGoal assuming the heap is in steady-state. + heapGoal := int64(atomic.Load64(&c.heapGoal)) + + var scanWorkExpected int64 + if goexperiment.PacerRedesign { + // The expected scan work is computed as the amount of bytes scanned last + // GC cycle, plus our estimate of stacks and globals work for this cycle. + scanWorkExpected = int64(c.lastHeapScan + c.stackScan + c.globalsScan) + + // maxScanWork is a worst-case estimate of the amount of scan work that + // needs to be performed in this GC cycle. Specifically, it represents + // the case where *all* scannable memory turns out to be live. + maxScanWork := int64(scan + c.stackScan + c.globalsScan) + if work > scanWorkExpected { + // We've already done more scan work than expected. Because our expectation + // is based on a steady-state scannable heap size, we assume this means our + // heap is growing. Compute a new heap goal that takes our existing runway + // computed for scanWorkExpected and extrapolates it to maxScanWork, the worst-case + // scan work. This keeps our assist ratio stable if the heap continues to grow. + // + // The effect of this mechanism is that assists stay flat in the face of heap + // growths. It's OK to use more memory this cycle to scan all the live heap, + // because the next GC cycle is inevitably going to use *at least* that much + // memory anyway. + extHeapGoal := int64(float64(heapGoal-int64(c.trigger))/float64(scanWorkExpected)*float64(maxScanWork)) + int64(c.trigger) + scanWorkExpected = maxScanWork + + // hardGoal is a hard limit on the amount that we're willing to push back the + // heap goal, and that's twice the heap goal (i.e. if GOGC=100 and the heap and/or + // stacks and/or globals grow to twice their size, this limits the current GC cycle's + // growth to 4x the original live heap's size). + // + // This maintains the invariant that we use no more memory than the next GC cycle + // will anyway. + hardGoal := int64((1.0 + float64(gcPercent)/100.0) * float64(heapGoal)) + if extHeapGoal > hardGoal { + extHeapGoal = hardGoal + } + heapGoal = extHeapGoal + } + if int64(live) > heapGoal { + // We're already past our heap goal, even the extrapolated one. + // Leave ourselves some extra runway, so in the worst case we + // finish by that point. + const maxOvershoot = 1.1 + heapGoal = int64(float64(heapGoal) * maxOvershoot) + + // Compute the upper bound on the scan work remaining. + scanWorkExpected = maxScanWork + } + } else { + // Compute the expected scan work remaining. + // + // This is estimated based on the expected + // steady-state scannable heap. For example, with + // GOGC=100, only half of the scannable heap is + // expected to be live, so that's what we target. + // + // (This is a float calculation to avoid overflowing on + // 100*heapScan.) + scanWorkExpected = int64(float64(scan) * 100 / float64(100+gcPercent)) + if int64(live) > heapGoal || work > scanWorkExpected { + // We're past the soft goal, or we've already done more scan + // work than we expected. Pace GC so that in the worst case it + // will complete by the hard goal. + const maxOvershoot = 1.1 + heapGoal = int64(float64(heapGoal) * maxOvershoot) + + // Compute the upper bound on the scan work remaining. + scanWorkExpected = int64(scan) + } + } + + // Compute the remaining scan work estimate. + // + // Note that we currently count allocations during GC as both + // scannable heap (heapScan) and scan work completed + // (scanWork), so allocation will change this difference + // slowly in the soft regime and not at all in the hard + // regime. + scanWorkRemaining := scanWorkExpected - work + if scanWorkRemaining < 1000 { + // We set a somewhat arbitrary lower bound on + // remaining scan work since if we aim a little high, + // we can miss by a little. + // + // We *do* need to enforce that this is at least 1, + // since marking is racy and double-scanning objects + // may legitimately make the remaining scan work + // negative, even in the hard goal regime. + scanWorkRemaining = 1000 + } + + // Compute the heap distance remaining. + heapRemaining := heapGoal - int64(live) + if heapRemaining <= 0 { + // This shouldn't happen, but if it does, avoid + // dividing by zero or setting the assist negative. + heapRemaining = 1 + } + + // Compute the mutator assist ratio so by the time the mutator + // allocates the remaining heap bytes up to heapGoal, it will + // have done (or stolen) the remaining amount of scan work. + // Note that the assist ratio values are updated atomically + // but not together. This means there may be some degree of + // skew between the two values. This is generally OK as the + // values shift relatively slowly over the course of a GC + // cycle. + assistWorkPerByte := float64(scanWorkRemaining) / float64(heapRemaining) + assistBytesPerWork := float64(heapRemaining) / float64(scanWorkRemaining) + c.assistWorkPerByte.Store(assistWorkPerByte) + c.assistBytesPerWork.Store(assistBytesPerWork) +} + +// endCycle computes the trigger ratio (!goexperiment.PacerRedesign) +// or the consMark estimate (goexperiment.PacerRedesign) for the next cycle. +// Returns the trigger ratio if application, or 0 (goexperiment.PacerRedesign). +// userForced indicates whether the current GC cycle was forced +// by the application. +func (c *gcControllerState) endCycle(now int64, procs int, userForced bool) float64 { + // Record last heap goal for the scavenger. + // We'll be updating the heap goal soon. + gcController.lastHeapGoal = gcController.heapGoal + + // Compute the duration of time for which assists were turned on. + assistDuration := now - c.markStartTime + + // Assume background mark hit its utilization goal. + utilization := gcBackgroundUtilization + // Add assist utilization; avoid divide by zero. + if assistDuration > 0 { + utilization += float64(c.assistTime) / float64(assistDuration*int64(procs)) + } + + if goexperiment.PacerRedesign { + if c.heapLive <= c.trigger { + // Shouldn't happen, but let's be very safe about this in case the + // GC is somehow extremely short. + // + // In this case though, the only reasonable value for c.heapLive-c.trigger + // would be 0, which isn't really all that useful, i.e. the GC was so short + // that it didn't matter. + // + // Ignore this case and don't update anything. + return 0 + } + idleUtilization := 0.0 + if assistDuration > 0 { + idleUtilization = float64(c.idleMarkTime) / float64(assistDuration*int64(procs)) + } + // Determine the cons/mark ratio. + // + // The units we want for the numerator and denominator are both B / cpu-ns. + // We get this by taking the bytes allocated or scanned, and divide by the amount of + // CPU time it took for those operations. For allocations, that CPU time is + // + // assistDuration * procs * (1 - utilization) + // + // Where utilization includes just background GC workers and assists. It does *not* + // include idle GC work time, because in theory the mutator is free to take that at + // any point. + // + // For scanning, that CPU time is + // + // assistDuration * procs * (utilization + idleUtilization) + // + // In this case, we *include* idle utilization, because that is additional CPU time that the + // the GC had available to it. + // + // In effect, idle GC time is sort of double-counted here, but it's very weird compared + // to other kinds of GC work, because of how fluid it is. Namely, because the mutator is + // *always* free to take it. + // + // So this calculation is really: + // (heapLive-trigger) / (assistDuration * procs * (1-utilization)) / + // (scanWork) / (assistDuration * procs * (utilization+idleUtilization) + // + // Note that because we only care about the ratio, assistDuration and procs cancel out. + scanWork := c.heapScanWork.Load() + c.stackScanWork.Load() + c.globalsScanWork.Load() + currentConsMark := (float64(c.heapLive-c.trigger) * (utilization + idleUtilization)) / + (float64(scanWork) * (1 - utilization)) + + // Update cons/mark controller. The time period for this is 1 GC cycle. + // + // This use of a PI controller might seem strange. So, here's an explanation: + // + // currentConsMark represents the consMark we *should've* had to be perfectly + // on-target for this cycle. Given that we assume the next GC will be like this + // one in the steady-state, it stands to reason that we should just pick that + // as our next consMark. In practice, however, currentConsMark is too noisy: + // we're going to be wildly off-target in each GC cycle if we do that. + // + // What we do instead is make a long-term assumption: there is some steady-state + // consMark value, but it's obscured by noise. By constantly shooting for this + // noisy-but-perfect consMark value, the controller will bounce around a bit, + // but its average behavior, in aggregate, should be less noisy and closer to + // the true long-term consMark value, provided its tuned to be slightly overdamped. + var ok bool + oldConsMark := c.consMark + c.consMark, ok = c.consMarkController.next(c.consMark, currentConsMark, 1.0) + if !ok { + // The error spiraled out of control. This is incredibly unlikely seeing + // as this controller is essentially just a smoothing function, but it might + // mean that something went very wrong with how currentConsMark was calculated. + // Just reset consMark and keep going. + c.consMark = 0 + } + + if debug.gcpacertrace > 0 { + printlock() + goal := gcGoalUtilization * 100 + print("pacer: ", int(utilization*100), "% CPU (", int(goal), " exp.) for ") + print(c.heapScanWork.Load(), "+", c.stackScanWork.Load(), "+", c.globalsScanWork.Load(), " B work (", c.lastHeapScan+c.stackScan+c.globalsScan, " B exp.) ") + print("in ", c.trigger, " B -> ", c.heapLive, " B (∆goal ", int64(c.heapLive)-int64(c.heapGoal), ", cons/mark ", oldConsMark, ")") + if !ok { + print("[controller reset]") + } + println() + printunlock() + } + return 0 + } + + // !goexperiment.PacerRedesign below. + + if userForced { + // Forced GC means this cycle didn't start at the + // trigger, so where it finished isn't good + // information about how to adjust the trigger. + // Just leave it where it is. + return c.triggerRatio + } + + // Proportional response gain for the trigger controller. Must + // be in [0, 1]. Lower values smooth out transient effects but + // take longer to respond to phase changes. Higher values + // react to phase changes quickly, but are more affected by + // transient changes. Values near 1 may be unstable. + const triggerGain = 0.5 + + // Compute next cycle trigger ratio. First, this computes the + // "error" for this cycle; that is, how far off the trigger + // was from what it should have been, accounting for both heap + // growth and GC CPU utilization. We compute the actual heap + // growth during this cycle and scale that by how far off from + // the goal CPU utilization we were (to estimate the heap + // growth if we had the desired CPU utilization). The + // difference between this estimate and the GOGC-based goal + // heap growth is the error. + goalGrowthRatio := c.effectiveGrowthRatio() + actualGrowthRatio := float64(c.heapLive)/float64(c.heapMarked) - 1 + triggerError := goalGrowthRatio - c.triggerRatio - utilization/gcGoalUtilization*(actualGrowthRatio-c.triggerRatio) + + // Finally, we adjust the trigger for next time by this error, + // damped by the proportional gain. + triggerRatio := c.triggerRatio + triggerGain*triggerError + + if debug.gcpacertrace > 0 { + // Print controller state in terms of the design + // document. + H_m_prev := c.heapMarked + h_t := c.triggerRatio + H_T := c.trigger + h_a := actualGrowthRatio + H_a := c.heapLive + h_g := goalGrowthRatio + H_g := int64(float64(H_m_prev) * (1 + h_g)) + u_a := utilization + u_g := gcGoalUtilization + W_a := c.heapScanWork.Load() + print("pacer: H_m_prev=", H_m_prev, + " h_t=", h_t, " H_T=", H_T, + " h_a=", h_a, " H_a=", H_a, + " h_g=", h_g, " H_g=", H_g, + " u_a=", u_a, " u_g=", u_g, + " W_a=", W_a, + " goalΔ=", goalGrowthRatio-h_t, + " actualΔ=", h_a-h_t, + " u_a/u_g=", u_a/u_g, + "\n") + } + + return triggerRatio +} + +// enlistWorker encourages another dedicated mark worker to start on +// another P if there are spare worker slots. It is used by putfull +// when more work is made available. +// +//go:nowritebarrier +func (c *gcControllerState) enlistWorker() { + // If there are idle Ps, wake one so it will run an idle worker. + // NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112. + // + // if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 { + // wakep() + // return + // } + + // There are no idle Ps. If we need more dedicated workers, + // try to preempt a running P so it will switch to a worker. + if c.dedicatedMarkWorkersNeeded <= 0 { + return + } + // Pick a random other P to preempt. + if gomaxprocs <= 1 { + return + } + gp := getg() + if gp == nil || gp.m == nil || gp.m.p == 0 { + return + } + myID := gp.m.p.ptr().id + for tries := 0; tries < 5; tries++ { + id := int32(fastrandn(uint32(gomaxprocs - 1))) + if id >= myID { + id++ + } + p := allp[id] + if p.status != _Prunning { + continue + } + if preemptone(p) { + return + } + } +} + +// findRunnableGCWorker returns a background mark worker for _p_ if it +// should be run. This must only be called when gcBlackenEnabled != 0. +func (c *gcControllerState) findRunnableGCWorker(_p_ *p) *g { + if gcBlackenEnabled == 0 { + throw("gcControllerState.findRunnable: blackening not enabled") + } + + if !gcMarkWorkAvailable(_p_) { + // No work to be done right now. This can happen at + // the end of the mark phase when there are still + // assists tapering off. Don't bother running a worker + // now because it'll just return immediately. + return nil + } + + // Grab a worker before we commit to running below. + node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop()) + if node == nil { + // There is at least one worker per P, so normally there are + // enough workers to run on all Ps, if necessary. However, once + // a worker enters gcMarkDone it may park without rejoining the + // pool, thus freeing a P with no corresponding worker. + // gcMarkDone never depends on another worker doing work, so it + // is safe to simply do nothing here. + // + // If gcMarkDone bails out without completing the mark phase, + // it will always do so with queued global work. Thus, that P + // will be immediately eligible to re-run the worker G it was + // just using, ensuring work can complete. + return nil + } + + decIfPositive := func(ptr *int64) bool { + for { + v := atomic.Loadint64(ptr) + if v <= 0 { + return false + } + + if atomic.Casint64(ptr, v, v-1) { + return true + } + } + } + + if decIfPositive(&c.dedicatedMarkWorkersNeeded) { + // This P is now dedicated to marking until the end of + // the concurrent mark phase. + _p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode + } else if c.fractionalUtilizationGoal == 0 { + // No need for fractional workers. + gcBgMarkWorkerPool.push(&node.node) + return nil + } else { + // Is this P behind on the fractional utilization + // goal? + // + // This should be kept in sync with pollFractionalWorkerExit. + delta := nanotime() - c.markStartTime + if delta > 0 && float64(_p_.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal { + // Nope. No need to run a fractional worker. + gcBgMarkWorkerPool.push(&node.node) + return nil + } + // Run a fractional worker. + _p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode + } + + // Run the background mark worker. + gp := node.gp.ptr() + casgstatus(gp, _Gwaiting, _Grunnable) + if trace.enabled { + traceGoUnpark(gp, 0) + } + return gp +} + +// resetLive sets up the controller state for the next mark phase after the end +// of the previous one. Must be called after endCycle and before commit, before +// the world is started. +// +// The world must be stopped. +func (c *gcControllerState) resetLive(bytesMarked uint64) { + c.heapMarked = bytesMarked + c.heapLive = bytesMarked + c.heapScan = uint64(c.heapScanWork.Load()) + c.lastHeapScan = uint64(c.heapScanWork.Load()) + + // heapLive was updated, so emit a trace event. + if trace.enabled { + traceHeapAlloc() + } +} + +// logWorkTime updates mark work accounting in the controller by a duration of +// work in nanoseconds. +// +// Safe to execute at any time. +func (c *gcControllerState) logWorkTime(mode gcMarkWorkerMode, duration int64) { + switch mode { + case gcMarkWorkerDedicatedMode: + atomic.Xaddint64(&c.dedicatedMarkTime, duration) + atomic.Xaddint64(&c.dedicatedMarkWorkersNeeded, 1) + case gcMarkWorkerFractionalMode: + atomic.Xaddint64(&c.fractionalMarkTime, duration) + case gcMarkWorkerIdleMode: + atomic.Xaddint64(&c.idleMarkTime, duration) + default: + throw("logWorkTime: unknown mark worker mode") + } +} + +func (c *gcControllerState) update(dHeapLive, dHeapScan int64) { + if dHeapLive != 0 { + atomic.Xadd64(&gcController.heapLive, dHeapLive) + if trace.enabled { + // gcController.heapLive changed. + traceHeapAlloc() + } + } + // Only update heapScan in the new pacer redesign if we're not + // currently in a GC. + if !goexperiment.PacerRedesign || gcBlackenEnabled == 0 { + if dHeapScan != 0 { + atomic.Xadd64(&gcController.heapScan, dHeapScan) + } + } + if gcBlackenEnabled != 0 { + // gcController.heapLive and heapScan changed. + c.revise() + } +} + +func (c *gcControllerState) addScannableStack(pp *p, amount int64) { + if pp == nil { + atomic.Xadd64(&c.scannableStackSize, amount) + return + } + pp.scannableStackSizeDelta += amount + if pp.scannableStackSizeDelta >= scannableStackSizeSlack || pp.scannableStackSizeDelta <= -scannableStackSizeSlack { + atomic.Xadd64(&c.scannableStackSize, pp.scannableStackSizeDelta) + pp.scannableStackSizeDelta = 0 + } +} + +func (c *gcControllerState) addGlobals(amount int64) { + atomic.Xadd64(&c.globalsScan, amount) +} + +// commit recomputes all pacing parameters from scratch, namely +// absolute trigger, the heap goal, mark pacing, and sweep pacing. +// +// If goexperiment.PacerRedesign is true, triggerRatio is ignored. +// +// This can be called any time. If GC is the in the middle of a +// concurrent phase, it will adjust the pacing of that phase. +// +// This depends on gcPercent, gcController.heapMarked, and +// gcController.heapLive. These must be up to date. +// +// mheap_.lock must be held or the world must be stopped. +func (c *gcControllerState) commit(triggerRatio float64) { + if !c.test { + assertWorldStoppedOrLockHeld(&mheap_.lock) + } + + if !goexperiment.PacerRedesign { + c.oldCommit(triggerRatio) + return + } + + // Compute the next GC goal, which is when the allocated heap + // has grown by GOGC/100 over where it started the last cycle, + // plus additional runway for non-heap sources of GC work. + goal := ^uint64(0) + if gcPercent := c.gcPercent.Load(); gcPercent >= 0 { + goal = c.heapMarked + (c.heapMarked+atomic.Load64(&c.stackScan)+atomic.Load64(&c.globalsScan))*uint64(gcPercent)/100 + } + + // Don't trigger below the minimum heap size. + minTrigger := c.heapMinimum + if !isSweepDone() { + // Concurrent sweep happens in the heap growth + // from gcController.heapLive to trigger, so ensure + // that concurrent sweep has some heap growth + // in which to perform sweeping before we + // start the next GC cycle. + sweepMin := atomic.Load64(&c.heapLive) + sweepMinHeapDistance + if sweepMin > minTrigger { + minTrigger = sweepMin + } + } + + // If we let the trigger go too low, then if the application + // is allocating very rapidly we might end up in a situation + // where we're allocating black during a nearly always-on GC. + // The result of this is a growing heap and ultimately an + // increase in RSS. By capping us at a point >0, we're essentially + // saying that we're OK using more CPU during the GC to prevent + // this growth in RSS. + // + // The current constant was chosen empirically: given a sufficiently + // fast/scalable allocator with 48 Ps that could drive the trigger ratio + // to <0.05, this constant causes applications to retain the same peak + // RSS compared to not having this allocator. + if triggerBound := uint64(0.7*float64(goal-c.heapMarked)) + c.heapMarked; minTrigger < triggerBound { + minTrigger = triggerBound + } + + // For small heaps, set the max trigger point at 95% of the heap goal. + // This ensures we always have *some* headroom when the GC actually starts. + // For larger heaps, set the max trigger point at the goal, minus the + // minimum heap size. + // This choice follows from the fact that the minimum heap size is chosen + // to reflect the costs of a GC with no work to do. With a large heap but + // very little scan work to perform, this gives us exactly as much runway + // as we would need, in the worst case. + maxRunway := uint64(0.95 * float64(goal-c.heapMarked)) + if largeHeapMaxRunway := goal - c.heapMinimum; goal > c.heapMinimum && maxRunway < largeHeapMaxRunway { + maxRunway = largeHeapMaxRunway + } + maxTrigger := maxRunway + c.heapMarked + if maxTrigger < minTrigger { + maxTrigger = minTrigger + } + + // Compute the trigger by using our estimate of the cons/mark ratio. + // + // The idea is to take our expected scan work, and multiply it by + // the cons/mark ratio to determine how long it'll take to complete + // that scan work in terms of bytes allocated. This gives us our GC's + // runway. + // + // However, the cons/mark ratio is a ratio of rates per CPU-second, but + // here we care about the relative rates for some division of CPU + // resources among the mutator and the GC. + // + // To summarize, we have B / cpu-ns, and we want B / ns. We get that + // by multiplying by our desired division of CPU resources. We choose + // to express CPU resources as GOMAPROCS*fraction. Note that because + // we're working with a ratio here, we can omit the number of CPU cores, + // because they'll appear in the numerator and denominator and cancel out. + // As a result, this is basically just "weighing" the cons/mark ratio by + // our desired division of resources. + // + // Furthermore, by setting the trigger so that CPU resources are divided + // this way, assuming that the cons/mark ratio is correct, we make that + // division a reality. + var trigger uint64 + runway := uint64((c.consMark * (1 - gcGoalUtilization) / (gcGoalUtilization)) * float64(c.lastHeapScan+c.stackScan+c.globalsScan)) + if runway > goal { + trigger = minTrigger + } else { + trigger = goal - runway + } + if trigger < minTrigger { + trigger = minTrigger + } + if trigger > maxTrigger { + trigger = maxTrigger + } + if trigger > goal { + goal = trigger + } + + // Commit to the trigger and goal. + c.trigger = trigger + atomic.Store64(&c.heapGoal, goal) + if trace.enabled { + traceHeapGoal() + } + + // Update mark pacing. + if gcphase != _GCoff { + c.revise() + } +} + +// oldCommit sets the trigger ratio and updates everything +// derived from it: the absolute trigger, the heap goal, mark pacing, +// and sweep pacing. +// +// This can be called any time. If GC is the in the middle of a +// concurrent phase, it will adjust the pacing of that phase. +// +// This depends on gcPercent, gcController.heapMarked, and +// gcController.heapLive. These must be up to date. +// +// For !goexperiment.PacerRedesign. +func (c *gcControllerState) oldCommit(triggerRatio float64) { + gcPercent := c.gcPercent.Load() + + // Compute the next GC goal, which is when the allocated heap + // has grown by GOGC/100 over the heap marked by the last + // cycle. + goal := ^uint64(0) + if gcPercent >= 0 { + goal = c.heapMarked + c.heapMarked*uint64(gcPercent)/100 + } + + // Set the trigger ratio, capped to reasonable bounds. + if gcPercent >= 0 { + scalingFactor := float64(gcPercent) / 100 + // Ensure there's always a little margin so that the + // mutator assist ratio isn't infinity. + maxTriggerRatio := 0.95 * scalingFactor + if triggerRatio > maxTriggerRatio { + triggerRatio = maxTriggerRatio + } + + // If we let triggerRatio go too low, then if the application + // is allocating very rapidly we might end up in a situation + // where we're allocating black during a nearly always-on GC. + // The result of this is a growing heap and ultimately an + // increase in RSS. By capping us at a point >0, we're essentially + // saying that we're OK using more CPU during the GC to prevent + // this growth in RSS. + // + // The current constant was chosen empirically: given a sufficiently + // fast/scalable allocator with 48 Ps that could drive the trigger ratio + // to <0.05, this constant causes applications to retain the same peak + // RSS compared to not having this allocator. + minTriggerRatio := 0.6 * scalingFactor + if triggerRatio < minTriggerRatio { + triggerRatio = minTriggerRatio + } + } else if triggerRatio < 0 { + // gcPercent < 0, so just make sure we're not getting a negative + // triggerRatio. This case isn't expected to happen in practice, + // and doesn't really matter because if gcPercent < 0 then we won't + // ever consume triggerRatio further on in this function, but let's + // just be defensive here; the triggerRatio being negative is almost + // certainly undesirable. + triggerRatio = 0 + } + c.triggerRatio = triggerRatio + + // Compute the absolute GC trigger from the trigger ratio. + // + // We trigger the next GC cycle when the allocated heap has + // grown by the trigger ratio over the marked heap size. + trigger := ^uint64(0) + if gcPercent >= 0 { + trigger = uint64(float64(c.heapMarked) * (1 + triggerRatio)) + // Don't trigger below the minimum heap size. + minTrigger := c.heapMinimum + if !isSweepDone() { + // Concurrent sweep happens in the heap growth + // from gcController.heapLive to trigger, so ensure + // that concurrent sweep has some heap growth + // in which to perform sweeping before we + // start the next GC cycle. + sweepMin := atomic.Load64(&c.heapLive) + sweepMinHeapDistance + if sweepMin > minTrigger { + minTrigger = sweepMin + } + } + if trigger < minTrigger { + trigger = minTrigger + } + if int64(trigger) < 0 { + print("runtime: heapGoal=", c.heapGoal, " heapMarked=", c.heapMarked, " gcController.heapLive=", c.heapLive, " initialHeapLive=", work.initialHeapLive, "triggerRatio=", triggerRatio, " minTrigger=", minTrigger, "\n") + throw("trigger underflow") + } + if trigger > goal { + // The trigger ratio is always less than GOGC/100, but + // other bounds on the trigger may have raised it. + // Push up the goal, too. + goal = trigger + } + } + + // Commit to the trigger and goal. + c.trigger = trigger + atomic.Store64(&c.heapGoal, goal) + if trace.enabled { + traceHeapGoal() + } + + // Update mark pacing. + if gcphase != _GCoff { + c.revise() + } +} + +// effectiveGrowthRatio returns the current effective heap growth +// ratio (GOGC/100) based on heapMarked from the previous GC and +// heapGoal for the current GC. +// +// This may differ from gcPercent/100 because of various upper and +// lower bounds on gcPercent. For example, if the heap is smaller than +// heapMinimum, this can be higher than gcPercent/100. +// +// mheap_.lock must be held or the world must be stopped. +func (c *gcControllerState) effectiveGrowthRatio() float64 { + if !c.test { + assertWorldStoppedOrLockHeld(&mheap_.lock) + } + + egogc := float64(atomic.Load64(&c.heapGoal)-c.heapMarked) / float64(c.heapMarked) + if egogc < 0 { + // Shouldn't happen, but just in case. + egogc = 0 + } + return egogc +} + +// setGCPercent updates gcPercent and all related pacer state. +// Returns the old value of gcPercent. +// +// Calls gcControllerState.commit. +// +// The world must be stopped, or mheap_.lock must be held. +func (c *gcControllerState) setGCPercent(in int32) int32 { + if !c.test { + assertWorldStoppedOrLockHeld(&mheap_.lock) + } + + out := c.gcPercent.Load() + if in < 0 { + in = -1 + } + c.heapMinimum = defaultHeapMinimum * uint64(in) / 100 + c.gcPercent.Store(in) + // Update pacing in response to gcPercent change. + c.commit(c.triggerRatio) + + return out +} + +//go:linkname setGCPercent runtime/debug.setGCPercent +func setGCPercent(in int32) (out int32) { + // Run on the system stack since we grab the heap lock. + systemstack(func() { + lock(&mheap_.lock) + out = gcController.setGCPercent(in) + gcPaceSweeper(gcController.trigger) + gcPaceScavenger(gcController.heapGoal, gcController.lastHeapGoal) + unlock(&mheap_.lock) + }) + + // If we just disabled GC, wait for any concurrent GC mark to + // finish so we always return with no GC running. + if in < 0 { + gcWaitOnMark(atomic.Load(&work.cycles)) + } + + return out +} + +func readGOGC() int32 { + p := gogetenv("GOGC") + if p == "off" { + return -1 + } + if n, ok := atoi32(p); ok { + return n + } + return 100 +} + +type piController struct { + kp float64 // Proportional constant. + ti float64 // Integral time constant. + tt float64 // Reset time. + + min, max float64 // Output boundaries. + + // PI controller state. + + errIntegral float64 // Integral of the error from t=0 to now. + + // Error flags. + errOverflow bool // Set if errIntegral ever overflowed. + inputOverflow bool // Set if an operation with the input overflowed. +} + +// next provides a new sample to the controller. +// +// input is the sample, setpoint is the desired point, and period is how much +// time (in whatever unit makes the most sense) has passed since the last sample. +// +// Returns a new value for the variable it's controlling, and whether the operation +// completed successfully. One reason this might fail is if error has been growing +// in an unbounded manner, to the point of overflow. +// +// In the specific case of an error overflow occurs, the errOverflow field will be +// set and the rest of the controller's internal state will be fully reset. +func (c *piController) next(input, setpoint, period float64) (float64, bool) { + // Compute the raw output value. + prop := c.kp * (setpoint - input) + rawOutput := prop + c.errIntegral + + // Clamp rawOutput into output. + output := rawOutput + if isInf(output) || isNaN(output) { + // The input had a large enough magnitude that either it was already + // overflowed, or some operation with it overflowed. + // Set a flag and reset. That's the safest thing to do. + c.reset() + c.inputOverflow = true + return c.min, false + } + if output < c.min { + output = c.min + } else if output > c.max { + output = c.max + } + + // Update the controller's state. + if c.ti != 0 && c.tt != 0 { + c.errIntegral += (c.kp*period/c.ti)*(setpoint-input) + (period/c.tt)*(output-rawOutput) + if isInf(c.errIntegral) || isNaN(c.errIntegral) { + // So much error has accumulated that we managed to overflow. + // The assumptions around the controller have likely broken down. + // Set a flag and reset. That's the safest thing to do. + c.reset() + c.errOverflow = true + return c.min, false + } + } + return output, true +} + +// reset resets the controller state, except for controller error flags. +func (c *piController) reset() { + c.errIntegral = 0 +} |