Source File
cpuprof.go
Belonging Package
runtime
// Copyright 2011 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.// CPU profiling.//// The signal handler for the profiling clock tick adds a new stack trace// to a log of recent traces. The log is read by a user goroutine that// turns it into formatted profile data. If the reader does not keep up// with the log, those writes will be recorded as a count of lost records.// The actual profile buffer is in profbuf.go.package runtimeimport ()const maxCPUProfStack = 64type cpuProfile struct {lock mutexon bool // profiling is onlog *profBuf // profile events written here// extra holds extra stacks accumulated in addNonGo// corresponding to profiling signals arriving on// non-Go-created threads. Those stacks are written// to log the next time a normal Go thread gets the// signal handler.// Assuming the stacks are 2 words each (we don't get// a full traceback from those threads), plus one word// size for framing, 100 Hz profiling would generate// 300 words per second.// Hopefully a normal Go thread will get the profiling// signal at least once every few seconds.extra [1000]uintptrnumExtra intlostExtra uint64 // count of frames lost because extra is fulllostAtomic uint64 // count of frames lost because of being in atomic64 on mips/arm; updated racily}var cpuprof cpuProfile// SetCPUProfileRate sets the CPU profiling rate to hz samples per second.// If hz <= 0, SetCPUProfileRate turns off profiling.// If the profiler is on, the rate cannot be changed without first turning it off.//// Most clients should use the runtime/pprof package or// the testing package's -test.cpuprofile flag instead of calling// SetCPUProfileRate directly.func ( int) {// Clamp hz to something reasonable.if < 0 {= 0}if > 1000000 {= 1000000}lock(&cpuprof.lock)if > 0 {if cpuprof.on || cpuprof.log != nil {print("runtime: cannot set cpu profile rate until previous profile has finished.\n")unlock(&cpuprof.lock)return}cpuprof.on = truecpuprof.log = newProfBuf(1, 1<<17, 1<<14):= [1]uint64{uint64()}cpuprof.log.write(nil, nanotime(), [:], nil)setcpuprofilerate(int32())} else if cpuprof.on {setcpuprofilerate(0)cpuprof.on = falsecpuprof.addExtra()cpuprof.log.close()}unlock(&cpuprof.lock)}// add adds the stack trace to the profile.// It is called from signal handlers and other limited environments// and cannot allocate memory or acquire locks that might be// held at the time of the signal, nor can it use substantial amounts// of stack.//go:nowritebarrierrecfunc ( *cpuProfile) ( *g, []uintptr) {// Simple cas-lock to coordinate with setcpuprofilerate.for !atomic.Cas(&prof.signalLock, 0, 1) {osyield()}if prof.hz != 0 { // implies cpuprof.log != nilif .numExtra > 0 || .lostExtra > 0 || .lostAtomic > 0 {.addExtra()}:= [1]uint64{1}// Note: write "knows" that the argument is &gp.labels,// because otherwise its write barrier behavior may not// be correct. See the long comment there before// changing the argument here.cpuprof.log.write(&.labels, nanotime(), [:], )}atomic.Store(&prof.signalLock, 0)}// addNonGo adds the non-Go stack trace to the profile.// It is called from a non-Go thread, so we cannot use much stack at all,// nor do anything that needs a g or an m.// In particular, we can't call cpuprof.log.write.// Instead, we copy the stack into cpuprof.extra,// which will be drained the next time a Go thread// gets the signal handling event.//go:nosplit//go:nowritebarrierrecfunc ( *cpuProfile) ( []uintptr) {// Simple cas-lock to coordinate with SetCPUProfileRate.// (Other calls to add or addNonGo should be blocked out// by the fact that only one SIGPROF can be handled by the// process at a time. If not, this lock will serialize those too.)for !atomic.Cas(&prof.signalLock, 0, 1) {osyield()}if cpuprof.numExtra+1+len() < len(cpuprof.extra) {:= cpuprof.numExtracpuprof.extra[] = uintptr(1 + len())copy(cpuprof.extra[+1:], )cpuprof.numExtra += 1 + len()} else {cpuprof.lostExtra++}atomic.Store(&prof.signalLock, 0)}// addExtra adds the "extra" profiling events,// queued by addNonGo, to the profile log.// addExtra is called either from a signal handler on a Go thread// or from an ordinary goroutine; either way it can use stack// and has a g. The world may be stopped, though.func ( *cpuProfile) () {// Copy accumulated non-Go profile events.:= [1]uint64{1}for := 0; < .numExtra; {.log.write(nil, 0, [:], .extra[+1:+int(.extra[])])+= int(.extra[])}.numExtra = 0// Report any lost events.if .lostExtra > 0 {:= [1]uint64{.lostExtra}:= [2]uintptr{funcPC(_LostExternalCode) + sys.PCQuantum,funcPC(_ExternalCode) + sys.PCQuantum,}.log.write(nil, 0, [:], [:]).lostExtra = 0}if .lostAtomic > 0 {:= [1]uint64{.lostAtomic}:= [2]uintptr{funcPC(_LostSIGPROFDuringAtomic64) + sys.PCQuantum,funcPC(_System) + sys.PCQuantum,}.log.write(nil, 0, [:], [:]).lostAtomic = 0}}// CPUProfile panics.// It formerly provided raw access to chunks of// a pprof-format profile generated by the runtime.// The details of generating that format have changed,// so this functionality has been removed.//// Deprecated: Use the runtime/pprof package,// or the handlers in the net/http/pprof package,// or the testing package's -test.cpuprofile flag instead.func () []byte {panic("CPUProfile no longer available")}//go:linkname runtime_pprof_runtime_cyclesPerSecond runtime/pprof.runtime_cyclesPerSecondfunc () int64 {return tickspersecond()}// readProfile, provided to runtime/pprof, returns the next chunk of// binary CPU profiling stack trace data, blocking until data is available.// If profiling is turned off and all the profile data accumulated while it was// on has been returned, readProfile returns eof=true.// The caller must save the returned data and tags before calling readProfile again.////go:linkname runtime_pprof_readProfile runtime/pprof.readProfilefunc () ([]uint64, []unsafe.Pointer, bool) {lock(&cpuprof.lock):= cpuprof.logunlock(&cpuprof.lock), , := .read(profBufBlocking)if len() == 0 && {lock(&cpuprof.lock)cpuprof.log = nilunlock(&cpuprof.lock)}return , ,}