// Copyright 2020 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 // Metrics implementation exported to runtime/metrics. import ( "runtime/internal/atomic" "unsafe" ) var ( // metrics is a map of runtime/metrics keys to // data used by the runtime to sample each metric's // value. metricsSema uint32 = 1 metricsInit bool metrics map[string]metricData sizeClassBuckets []float64 timeHistBuckets []float64 ) type metricData struct { // deps is the set of runtime statistics that this metric // depends on. Before compute is called, the statAggregate // which will be passed must ensure() these dependencies. deps statDepSet // compute is a function that populates a metricValue // given a populated statAggregate structure. compute func(in *statAggregate, out *metricValue) } // initMetrics initializes the metrics map if it hasn't been yet. // // metricsSema must be held. func initMetrics() { if metricsInit { return } sizeClassBuckets = make([]float64, _NumSizeClasses, _NumSizeClasses+1) // Skip size class 0 which is a stand-in for large objects, but large // objects are tracked separately (and they actually get placed in // the last bucket, not the first). sizeClassBuckets[0] = 1 // The smallest allocation is 1 byte in size. for i := 1; i < _NumSizeClasses; i++ { // Size classes have an inclusive upper-bound // and exclusive lower bound (e.g. 48-byte size class is // (32, 48]) whereas we want and inclusive lower-bound // and exclusive upper-bound (e.g. 48-byte size class is // [33, 49). We can achieve this by shifting all bucket // boundaries up by 1. // // Also, a float64 can precisely represent integers with // value up to 2^53 and size classes are relatively small // (nowhere near 2^48 even) so this will give us exact // boundaries. sizeClassBuckets[i] = float64(class_to_size[i] + 1) } sizeClassBuckets = append(sizeClassBuckets, float64Inf()) timeHistBuckets = timeHistogramMetricsBuckets() metrics = map[string]metricData{ "/gc/cycles/automatic:gc-cycles": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.gcCyclesDone - in.sysStats.gcCyclesForced }, }, "/gc/cycles/forced:gc-cycles": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.gcCyclesForced }, }, "/gc/cycles/total:gc-cycles": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.gcCyclesDone }, }, "/gc/heap/allocs-by-size:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { hist := out.float64HistOrInit(sizeClassBuckets) hist.counts[len(hist.counts)-1] = uint64(in.heapStats.largeAllocCount) // Cut off the first index which is ostensibly for size class 0, // but large objects are tracked separately so it's actually unused. for i, count := range in.heapStats.smallAllocCount[1:] { hist.counts[i] = uint64(count) } }, }, "/gc/heap/allocs:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.heapStats.totalAllocated }, }, "/gc/heap/allocs:objects": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.heapStats.totalAllocs }, }, "/gc/heap/frees-by-size:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { hist := out.float64HistOrInit(sizeClassBuckets) hist.counts[len(hist.counts)-1] = uint64(in.heapStats.largeFreeCount) // Cut off the first index which is ostensibly for size class 0, // but large objects are tracked separately so it's actually unused. for i, count := range in.heapStats.smallFreeCount[1:] { hist.counts[i] = uint64(count) } }, }, "/gc/heap/frees:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.heapStats.totalFreed }, }, "/gc/heap/frees:objects": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.heapStats.totalFrees }, }, "/gc/heap/goal:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.heapGoal }, }, "/gc/heap/objects:objects": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.heapStats.numObjects }, }, "/gc/heap/tiny/allocs:objects": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(in.heapStats.tinyAllocCount) }, }, "/gc/pauses:seconds": { compute: func(_ *statAggregate, out *metricValue) { hist := out.float64HistOrInit(timeHistBuckets) // The bottom-most bucket, containing negative values, is tracked // as a separately as underflow, so fill that in manually and then // iterate over the rest. hist.counts[0] = atomic.Load64(&memstats.gcPauseDist.underflow) for i := range memstats.gcPauseDist.counts { hist.counts[i+1] = atomic.Load64(&memstats.gcPauseDist.counts[i]) } }, }, "/memory/classes/heap/free:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(in.heapStats.committed - in.heapStats.inHeap - in.heapStats.inStacks - in.heapStats.inWorkBufs - in.heapStats.inPtrScalarBits) }, }, "/memory/classes/heap/objects:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.heapStats.inObjects }, }, "/memory/classes/heap/released:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(in.heapStats.released) }, }, "/memory/classes/heap/stacks:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(in.heapStats.inStacks) }, }, "/memory/classes/heap/unused:bytes": { deps: makeStatDepSet(heapStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(in.heapStats.inHeap) - in.heapStats.inObjects }, }, "/memory/classes/metadata/mcache/free:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.mCacheSys - in.sysStats.mCacheInUse }, }, "/memory/classes/metadata/mcache/inuse:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.mCacheInUse }, }, "/memory/classes/metadata/mspan/free:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.mSpanSys - in.sysStats.mSpanInUse }, }, "/memory/classes/metadata/mspan/inuse:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.mSpanInUse }, }, "/memory/classes/metadata/other:bytes": { deps: makeStatDepSet(heapStatsDep, sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(in.heapStats.inWorkBufs+in.heapStats.inPtrScalarBits) + in.sysStats.gcMiscSys }, }, "/memory/classes/os-stacks:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.stacksSys }, }, "/memory/classes/other:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.otherSys }, }, "/memory/classes/profiling/buckets:bytes": { deps: makeStatDepSet(sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = in.sysStats.buckHashSys }, }, "/memory/classes/total:bytes": { deps: makeStatDepSet(heapStatsDep, sysStatsDep), compute: func(in *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(in.heapStats.committed+in.heapStats.released) + in.sysStats.stacksSys + in.sysStats.mSpanSys + in.sysStats.mCacheSys + in.sysStats.buckHashSys + in.sysStats.gcMiscSys + in.sysStats.otherSys }, }, "/sched/goroutines:goroutines": { compute: func(_ *statAggregate, out *metricValue) { out.kind = metricKindUint64 out.scalar = uint64(gcount()) }, }, "/sched/latencies:seconds": { compute: func(_ *statAggregate, out *metricValue) { hist := out.float64HistOrInit(timeHistBuckets) hist.counts[0] = atomic.Load64(&sched.timeToRun.underflow) for i := range sched.timeToRun.counts { hist.counts[i+1] = atomic.Load64(&sched.timeToRun.counts[i]) } }, }, } metricsInit = true } // statDep is a dependency on a group of statistics // that a metric might have. type statDep uint const ( heapStatsDep statDep = iota // corresponds to heapStatsAggregate sysStatsDep // corresponds to sysStatsAggregate numStatsDeps ) // statDepSet represents a set of statDeps. // // Under the hood, it's a bitmap. type statDepSet [1]uint64 // makeStatDepSet creates a new statDepSet from a list of statDeps. func makeStatDepSet(deps ...statDep) statDepSet { var s statDepSet for _, d := range deps { s[d/64] |= 1 << (d % 64) } return s } // differennce returns set difference of s from b as a new set. func (s statDepSet) difference(b statDepSet) statDepSet { var c statDepSet for i := range s { c[i] = s[i] &^ b[i] } return c } // union returns the union of the two sets as a new set. func (s statDepSet) union(b statDepSet) statDepSet { var c statDepSet for i := range s { c[i] = s[i] | b[i] } return c } // empty returns true if there are no dependencies in the set. func (s *statDepSet) empty() bool { for _, c := range s { if c != 0 { return false } } return true } // has returns true if the set contains a given statDep. func (s *statDepSet) has(d statDep) bool { return s[d/64]&(1<<(d%64)) != 0 } // heapStatsAggregate represents memory stats obtained from the // runtime. This set of stats is grouped together because they // depend on each other in some way to make sense of the runtime's // current heap memory use. They're also sharded across Ps, so it // makes sense to grab them all at once. type heapStatsAggregate struct { heapStatsDelta // Derived from values in heapStatsDelta. // inObjects is the bytes of memory occupied by objects, inObjects uint64 // numObjects is the number of live objects in the heap. numObjects uint64 // totalAllocated is the total bytes of heap objects allocated // over the lifetime of the program. totalAllocated uint64 // totalFreed is the total bytes of heap objects freed // over the lifetime of the program. totalFreed uint64 // totalAllocs is the number of heap objects allocated over // the lifetime of the program. totalAllocs uint64 // totalFrees is the number of heap objects freed over // the lifetime of the program. totalFrees uint64 } // compute populates the heapStatsAggregate with values from the runtime. func (a *heapStatsAggregate) compute() { memstats.heapStats.read(&a.heapStatsDelta) // Calculate derived stats. a.totalAllocs = uint64(a.largeAllocCount) a.totalFrees = uint64(a.largeFreeCount) a.totalAllocated = uint64(a.largeAlloc) a.totalFreed = uint64(a.largeFree) for i := range a.smallAllocCount { na := uint64(a.smallAllocCount[i]) nf := uint64(a.smallFreeCount[i]) a.totalAllocs += na a.totalFrees += nf a.totalAllocated += na * uint64(class_to_size[i]) a.totalFreed += nf * uint64(class_to_size[i]) } a.inObjects = a.totalAllocated - a.totalFreed a.numObjects = a.totalAllocs - a.totalFrees } // sysStatsAggregate represents system memory stats obtained // from the runtime. This set of stats is grouped together because // they're all relatively cheap to acquire and generally independent // of one another and other runtime memory stats. The fact that they // may be acquired at different times, especially with respect to // heapStatsAggregate, means there could be some skew, but because of // these stats are independent, there's no real consistency issue here. type sysStatsAggregate struct { stacksSys uint64 mSpanSys uint64 mSpanInUse uint64 mCacheSys uint64 mCacheInUse uint64 buckHashSys uint64 gcMiscSys uint64 otherSys uint64 heapGoal uint64 gcCyclesDone uint64 gcCyclesForced uint64 } // compute populates the sysStatsAggregate with values from the runtime. func (a *sysStatsAggregate) compute() { a.stacksSys = memstats.stacks_sys.load() a.buckHashSys = memstats.buckhash_sys.load() a.gcMiscSys = memstats.gcMiscSys.load() a.otherSys = memstats.other_sys.load() a.heapGoal = atomic.Load64(&gcController.heapGoal) a.gcCyclesDone = uint64(memstats.numgc) a.gcCyclesForced = uint64(memstats.numforcedgc) systemstack(func() { lock(&mheap_.lock) a.mSpanSys = memstats.mspan_sys.load() a.mSpanInUse = uint64(mheap_.spanalloc.inuse) a.mCacheSys = memstats.mcache_sys.load() a.mCacheInUse = uint64(mheap_.cachealloc.inuse) unlock(&mheap_.lock) }) } // statAggregate is the main driver of the metrics implementation. // // It contains multiple aggregates of runtime statistics, as well // as a set of these aggregates that it has populated. The aggergates // are populated lazily by its ensure method. type statAggregate struct { ensured statDepSet heapStats heapStatsAggregate sysStats sysStatsAggregate } // ensure populates statistics aggregates determined by deps if they // haven't yet been populated. func (a *statAggregate) ensure(deps *statDepSet) { missing := deps.difference(a.ensured) if missing.empty() { return } for i := statDep(0); i < numStatsDeps; i++ { if !missing.has(i) { continue } switch i { case heapStatsDep: a.heapStats.compute() case sysStatsDep: a.sysStats.compute() } } a.ensured = a.ensured.union(missing) } // metricValidKind is a runtime copy of runtime/metrics.ValueKind and // must be kept structurally identical to that type. type metricKind int const ( // These values must be kept identical to their corresponding Kind* values // in the runtime/metrics package. metricKindBad metricKind = iota metricKindUint64 metricKindFloat64 metricKindFloat64Histogram ) // metricSample is a runtime copy of runtime/metrics.Sample and // must be kept structurally identical to that type. type metricSample struct { name string value metricValue } // metricValue is a runtime copy of runtime/metrics.Sample and // must be kept structurally identical to that type. type metricValue struct { kind metricKind scalar uint64 // contains scalar values for scalar Kinds. pointer unsafe.Pointer // contains non-scalar values. } // float64HistOrInit tries to pull out an existing float64Histogram // from the value, but if none exists, then it allocates one with // the given buckets. func (v *metricValue) float64HistOrInit(buckets []float64) *metricFloat64Histogram { var hist *metricFloat64Histogram if v.kind == metricKindFloat64Histogram && v.pointer != nil { hist = (*metricFloat64Histogram)(v.pointer) } else { v.kind = metricKindFloat64Histogram hist = new(metricFloat64Histogram) v.pointer = unsafe.Pointer(hist) } hist.buckets = buckets if len(hist.counts) != len(hist.buckets)-1 { hist.counts = make([]uint64, len(buckets)-1) } return hist } // metricFloat64Histogram is a runtime copy of runtime/metrics.Float64Histogram // and must be kept structurally identical to that type. type metricFloat64Histogram struct { counts []uint64 buckets []float64 } // agg is used by readMetrics, and is protected by metricsSema. // // Managed as a global variable because its pointer will be // an argument to a dynamically-defined function, and we'd // like to avoid it escaping to the heap. var agg statAggregate // readMetrics is the implementation of runtime/metrics.Read. // //go:linkname readMetrics runtime/metrics.runtime_readMetrics func readMetrics(samplesp unsafe.Pointer, len int, cap int) { // Construct a slice from the args. sl := slice{samplesp, len, cap} samples := *(*[]metricSample)(unsafe.Pointer(&sl)) // Acquire the metricsSema but with handoff. This operation // is expensive enough that queueing up goroutines and handing // off between them will be noticeably better-behaved. semacquire1(&metricsSema, true, 0, 0) // Ensure the map is initialized. initMetrics() // Clear agg defensively. agg = statAggregate{} // Sample. for i := range samples { sample := &samples[i] data, ok := metrics[sample.name] if !ok { sample.value.kind = metricKindBad continue } // Ensure we have all the stats we need. // agg is populated lazily. agg.ensure(&data.deps) // Compute the value based on the stats we have. data.compute(&agg, &sample.value) } semrelease(&metricsSema) }