1// Copyright 2011 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. 4 5// CPU profiling. 6// 7// The signal handler for the profiling clock tick adds a new stack trace 8// to a log of recent traces. The log is read by a user goroutine that 9// turns it into formatted profile data. If the reader does not keep up 10// with the log, those writes will be recorded as a count of lost records. 11// The actual profile buffer is in profbuf.go. 12 13package runtime 14 15import ( 16 "runtime/internal/atomic" 17 "runtime/internal/sys" 18 "unsafe" 19) 20 21const maxCPUProfStack = 64 22 23type cpuProfile struct { 24 lock mutex 25 on bool // profiling is on 26 log *profBuf // profile events written here 27 28 // extra holds extra stacks accumulated in addNonGo 29 // corresponding to profiling signals arriving on 30 // non-Go-created threads. Those stacks are written 31 // to log the next time a normal Go thread gets the 32 // signal handler. 33 // Assuming the stacks are 2 words each (we don't get 34 // a full traceback from those threads), plus one word 35 // size for framing, 100 Hz profiling would generate 36 // 300 words per second. 37 // Hopefully a normal Go thread will get the profiling 38 // signal at least once every few seconds. 39 extra [1000]uintptr 40 numExtra int 41 lostExtra uint64 // count of frames lost because extra is full 42 lostAtomic uint64 // count of frames lost because of being in atomic64 on mips/arm; updated racily 43} 44 45var cpuprof cpuProfile 46 47// SetCPUProfileRate sets the CPU profiling rate to hz samples per second. 48// If hz <= 0, SetCPUProfileRate turns off profiling. 49// If the profiler is on, the rate cannot be changed without first turning it off. 50// 51// Most clients should use the runtime/pprof package or 52// the testing package's -test.cpuprofile flag instead of calling 53// SetCPUProfileRate directly. 54func SetCPUProfileRate(hz int) { 55 // Clamp hz to something reasonable. 56 if hz < 0 { 57 hz = 0 58 } 59 if hz > 1000000 { 60 hz = 1000000 61 } 62 63 lock(&cpuprof.lock) 64 if hz > 0 { 65 if cpuprof.on || cpuprof.log != nil { 66 print("runtime: cannot set cpu profile rate until previous profile has finished.\n") 67 unlock(&cpuprof.lock) 68 return 69 } 70 71 cpuprof.on = true 72 cpuprof.log = newProfBuf(1, 1<<17, 1<<14) 73 hdr := [1]uint64{uint64(hz)} 74 cpuprof.log.write(nil, nanotime(), hdr[:], nil) 75 setcpuprofilerate(int32(hz)) 76 } else if cpuprof.on { 77 setcpuprofilerate(0) 78 cpuprof.on = false 79 cpuprof.addExtra() 80 cpuprof.log.close() 81 } 82 unlock(&cpuprof.lock) 83} 84 85// add adds the stack trace to the profile. 86// It is called from signal handlers and other limited environments 87// and cannot allocate memory or acquire locks that might be 88// held at the time of the signal, nor can it use substantial amounts 89// of stack. 90//go:nowritebarrierrec 91func (p *cpuProfile) add(gp *g, stk []uintptr) { 92 // Simple cas-lock to coordinate with setcpuprofilerate. 93 for !atomic.Cas(&prof.signalLock, 0, 1) { 94 osyield() 95 } 96 97 if prof.hz != 0 { // implies cpuprof.log != nil 98 if p.numExtra > 0 || p.lostExtra > 0 || p.lostAtomic > 0 { 99 p.addExtra() 100 } 101 hdr := [1]uint64{1} 102 // Note: write "knows" that the argument is &gp.labels, 103 // because otherwise its write barrier behavior may not 104 // be correct. See the long comment there before 105 // changing the argument here. 106 cpuprof.log.write(&gp.labels, nanotime(), hdr[:], stk) 107 } 108 109 atomic.Store(&prof.signalLock, 0) 110} 111 112// addNonGo adds the non-Go stack trace to the profile. 113// It is called from a non-Go thread, so we cannot use much stack at all, 114// nor do anything that needs a g or an m. 115// In particular, we can't call cpuprof.log.write. 116// Instead, we copy the stack into cpuprof.extra, 117// which will be drained the next time a Go thread 118// gets the signal handling event. 119//go:nosplit 120//go:nowritebarrierrec 121func (p *cpuProfile) addNonGo(stk []uintptr) { 122 // Simple cas-lock to coordinate with SetCPUProfileRate. 123 // (Other calls to add or addNonGo should be blocked out 124 // by the fact that only one SIGPROF can be handled by the 125 // process at a time. If not, this lock will serialize those too.) 126 for !atomic.Cas(&prof.signalLock, 0, 1) { 127 osyield() 128 } 129 130 if cpuprof.numExtra+1+len(stk) < len(cpuprof.extra) { 131 i := cpuprof.numExtra 132 cpuprof.extra[i] = uintptr(1 + len(stk)) 133 copy(cpuprof.extra[i+1:], stk) 134 cpuprof.numExtra += 1 + len(stk) 135 } else { 136 cpuprof.lostExtra++ 137 } 138 139 atomic.Store(&prof.signalLock, 0) 140} 141 142// addExtra adds the "extra" profiling events, 143// queued by addNonGo, to the profile log. 144// addExtra is called either from a signal handler on a Go thread 145// or from an ordinary goroutine; either way it can use stack 146// and has a g. The world may be stopped, though. 147func (p *cpuProfile) addExtra() { 148 // Copy accumulated non-Go profile events. 149 hdr := [1]uint64{1} 150 for i := 0; i < p.numExtra; { 151 p.log.write(nil, 0, hdr[:], p.extra[i+1:i+int(p.extra[i])]) 152 i += int(p.extra[i]) 153 } 154 p.numExtra = 0 155 156 // Report any lost events. 157 if p.lostExtra > 0 { 158 hdr := [1]uint64{p.lostExtra} 159 lostStk := [2]uintptr{ 160 _LostExternalCodePC + sys.PCQuantum, 161 _ExternalCodePC + sys.PCQuantum, 162 } 163 p.log.write(nil, 0, hdr[:], lostStk[:]) 164 p.lostExtra = 0 165 } 166 167 if p.lostAtomic > 0 { 168 hdr := [1]uint64{p.lostAtomic} 169 lostStk := [2]uintptr{ 170 _LostSIGPROFDuringAtomic64PC + sys.PCQuantum, 171 _SystemPC + sys.PCQuantum, 172 } 173 p.log.write(nil, 0, hdr[:], lostStk[:]) 174 p.lostAtomic = 0 175 } 176 177} 178 179// CPUProfile panics. 180// It formerly provided raw access to chunks of 181// a pprof-format profile generated by the runtime. 182// The details of generating that format have changed, 183// so this functionality has been removed. 184// 185// Deprecated: Use the runtime/pprof package, 186// or the handlers in the net/http/pprof package, 187// or the testing package's -test.cpuprofile flag instead. 188func CPUProfile() []byte { 189 panic("CPUProfile no longer available") 190} 191 192//go:linkname runtime_pprof_runtime_cyclesPerSecond runtime..z2fpprof.runtime_cyclesPerSecond 193func runtime_pprof_runtime_cyclesPerSecond() int64 { 194 return tickspersecond() 195} 196 197// readProfile, provided to runtime/pprof, returns the next chunk of 198// binary CPU profiling stack trace data, blocking until data is available. 199// If profiling is turned off and all the profile data accumulated while it was 200// on has been returned, readProfile returns eof=true. 201// The caller must save the returned data and tags before calling readProfile again. 202// 203//go:linkname runtime_pprof_readProfile runtime..z2fpprof.readProfile 204func runtime_pprof_readProfile() ([]uint64, []unsafe.Pointer, bool) { 205 lock(&cpuprof.lock) 206 log := cpuprof.log 207 unlock(&cpuprof.lock) 208 data, tags, eof := log.read(profBufBlocking) 209 if len(data) == 0 && eof { 210 lock(&cpuprof.lock) 211 cpuprof.log = nil 212 unlock(&cpuprof.lock) 213 } 214 return data, tags, eof 215} 216