Source file src/runtime/cpuprof.go

     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  
    13  package runtime
    14  
    15  import (
    16  	"internal/abi"
    17  	"runtime/internal/atomic"
    18  	"runtime/internal/sys"
    19  	"unsafe"
    20  )
    21  
    22  const (
    23  	maxCPUProfStack = 64
    24  
    25  	// profBufWordCount is the size of the CPU profile buffer's storage for the
    26  	// header and stack of each sample, measured in 64-bit words. Every sample
    27  	// has a required header of two words. With a small additional header (a
    28  	// word or two) and stacks at the profiler's maximum length of 64 frames,
    29  	// that capacity can support 1900 samples or 19 thread-seconds at a 100 Hz
    30  	// sample rate, at a cost of 1 MiB.
    31  	profBufWordCount = 1 << 17
    32  	// profBufTagCount is the size of the CPU profile buffer's storage for the
    33  	// goroutine tags associated with each sample. A capacity of 1<<14 means
    34  	// room for 16k samples, or 160 thread-seconds at a 100 Hz sample rate.
    35  	profBufTagCount = 1 << 14
    36  )
    37  
    38  type cpuProfile struct {
    39  	lock mutex
    40  	on   bool     // profiling is on
    41  	log  *profBuf // profile events written here
    42  
    43  	// extra holds extra stacks accumulated in addNonGo
    44  	// corresponding to profiling signals arriving on
    45  	// non-Go-created threads. Those stacks are written
    46  	// to log the next time a normal Go thread gets the
    47  	// signal handler.
    48  	// Assuming the stacks are 2 words each (we don't get
    49  	// a full traceback from those threads), plus one word
    50  	// size for framing, 100 Hz profiling would generate
    51  	// 300 words per second.
    52  	// Hopefully a normal Go thread will get the profiling
    53  	// signal at least once every few seconds.
    54  	extra      [1000]uintptr
    55  	numExtra   int
    56  	lostExtra  uint64 // count of frames lost because extra is full
    57  	lostAtomic uint64 // count of frames lost because of being in atomic64 on mips/arm; updated racily
    58  }
    59  
    60  var cpuprof cpuProfile
    61  
    62  // SetCPUProfileRate sets the CPU profiling rate to hz samples per second.
    63  // If hz <= 0, SetCPUProfileRate turns off profiling.
    64  // If the profiler is on, the rate cannot be changed without first turning it off.
    65  //
    66  // Most clients should use the runtime/pprof package or
    67  // the testing package's -test.cpuprofile flag instead of calling
    68  // SetCPUProfileRate directly.
    69  func SetCPUProfileRate(hz int) {
    70  	// Clamp hz to something reasonable.
    71  	if hz < 0 {
    72  		hz = 0
    73  	}
    74  	if hz > 1000000 {
    75  		hz = 1000000
    76  	}
    77  
    78  	lock(&cpuprof.lock)
    79  	if hz > 0 {
    80  		if cpuprof.on || cpuprof.log != nil {
    81  			print("runtime: cannot set cpu profile rate until previous profile has finished.\n")
    82  			unlock(&cpuprof.lock)
    83  			return
    84  		}
    85  
    86  		cpuprof.on = true
    87  		cpuprof.log = newProfBuf(1, profBufWordCount, profBufTagCount)
    88  		hdr := [1]uint64{uint64(hz)}
    89  		cpuprof.log.write(nil, nanotime(), hdr[:], nil)
    90  		setcpuprofilerate(int32(hz))
    91  	} else if cpuprof.on {
    92  		setcpuprofilerate(0)
    93  		cpuprof.on = false
    94  		cpuprof.addExtra()
    95  		cpuprof.log.close()
    96  	}
    97  	unlock(&cpuprof.lock)
    98  }
    99  
   100  // add adds the stack trace to the profile.
   101  // It is called from signal handlers and other limited environments
   102  // and cannot allocate memory or acquire locks that might be
   103  // held at the time of the signal, nor can it use substantial amounts
   104  // of stack.
   105  //
   106  //go:nowritebarrierrec
   107  func (p *cpuProfile) add(tagPtr *unsafe.Pointer, stk []uintptr) {
   108  	// Simple cas-lock to coordinate with setcpuprofilerate.
   109  	for !atomic.Cas(&prof.signalLock, 0, 1) {
   110  		// TODO: Is it safe to osyield here? https://go.dev/issue/52672
   111  		osyield()
   112  	}
   113  
   114  	if prof.hz != 0 { // implies cpuprof.log != nil
   115  		if p.numExtra > 0 || p.lostExtra > 0 || p.lostAtomic > 0 {
   116  			p.addExtra()
   117  		}
   118  		hdr := [1]uint64{1}
   119  		// Note: write "knows" that the argument is &gp.labels,
   120  		// because otherwise its write barrier behavior may not
   121  		// be correct. See the long comment there before
   122  		// changing the argument here.
   123  		cpuprof.log.write(tagPtr, nanotime(), hdr[:], stk)
   124  	}
   125  
   126  	atomic.Store(&prof.signalLock, 0)
   127  }
   128  
   129  // addNonGo adds the non-Go stack trace to the profile.
   130  // It is called from a non-Go thread, so we cannot use much stack at all,
   131  // nor do anything that needs a g or an m.
   132  // In particular, we can't call cpuprof.log.write.
   133  // Instead, we copy the stack into cpuprof.extra,
   134  // which will be drained the next time a Go thread
   135  // gets the signal handling event.
   136  //
   137  //go:nosplit
   138  //go:nowritebarrierrec
   139  func (p *cpuProfile) addNonGo(stk []uintptr) {
   140  	// Simple cas-lock to coordinate with SetCPUProfileRate.
   141  	// (Other calls to add or addNonGo should be blocked out
   142  	// by the fact that only one SIGPROF can be handled by the
   143  	// process at a time. If not, this lock will serialize those too.
   144  	// The use of timer_create(2) on Linux to request process-targeted
   145  	// signals may have changed this.)
   146  	for !atomic.Cas(&prof.signalLock, 0, 1) {
   147  		// TODO: Is it safe to osyield here? https://go.dev/issue/52672
   148  		osyield()
   149  	}
   150  
   151  	if cpuprof.numExtra+1+len(stk) < len(cpuprof.extra) {
   152  		i := cpuprof.numExtra
   153  		cpuprof.extra[i] = uintptr(1 + len(stk))
   154  		copy(cpuprof.extra[i+1:], stk)
   155  		cpuprof.numExtra += 1 + len(stk)
   156  	} else {
   157  		cpuprof.lostExtra++
   158  	}
   159  
   160  	atomic.Store(&prof.signalLock, 0)
   161  }
   162  
   163  // addExtra adds the "extra" profiling events,
   164  // queued by addNonGo, to the profile log.
   165  // addExtra is called either from a signal handler on a Go thread
   166  // or from an ordinary goroutine; either way it can use stack
   167  // and has a g. The world may be stopped, though.
   168  func (p *cpuProfile) addExtra() {
   169  	// Copy accumulated non-Go profile events.
   170  	hdr := [1]uint64{1}
   171  	for i := 0; i < p.numExtra; {
   172  		p.log.write(nil, 0, hdr[:], p.extra[i+1:i+int(p.extra[i])])
   173  		i += int(p.extra[i])
   174  	}
   175  	p.numExtra = 0
   176  
   177  	// Report any lost events.
   178  	if p.lostExtra > 0 {
   179  		hdr := [1]uint64{p.lostExtra}
   180  		lostStk := [2]uintptr{
   181  			abi.FuncPCABIInternal(_LostExternalCode) + sys.PCQuantum,
   182  			abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum,
   183  		}
   184  		p.log.write(nil, 0, hdr[:], lostStk[:])
   185  		p.lostExtra = 0
   186  	}
   187  
   188  	if p.lostAtomic > 0 {
   189  		hdr := [1]uint64{p.lostAtomic}
   190  		lostStk := [2]uintptr{
   191  			abi.FuncPCABIInternal(_LostSIGPROFDuringAtomic64) + sys.PCQuantum,
   192  			abi.FuncPCABIInternal(_System) + sys.PCQuantum,
   193  		}
   194  		p.log.write(nil, 0, hdr[:], lostStk[:])
   195  		p.lostAtomic = 0
   196  	}
   197  
   198  }
   199  
   200  // CPUProfile panics.
   201  // It formerly provided raw access to chunks of
   202  // a pprof-format profile generated by the runtime.
   203  // The details of generating that format have changed,
   204  // so this functionality has been removed.
   205  //
   206  // Deprecated: Use the runtime/pprof package,
   207  // or the handlers in the net/http/pprof package,
   208  // or the testing package's -test.cpuprofile flag instead.
   209  func CPUProfile() []byte {
   210  	panic("CPUProfile no longer available")
   211  }
   212  
   213  //go:linkname runtime_pprof_runtime_cyclesPerSecond runtime/pprof.runtime_cyclesPerSecond
   214  func runtime_pprof_runtime_cyclesPerSecond() int64 {
   215  	return tickspersecond()
   216  }
   217  
   218  // readProfile, provided to runtime/pprof, returns the next chunk of
   219  // binary CPU profiling stack trace data, blocking until data is available.
   220  // If profiling is turned off and all the profile data accumulated while it was
   221  // on has been returned, readProfile returns eof=true.
   222  // The caller must save the returned data and tags before calling readProfile again.
   223  // The returned data contains a whole number of records, and tags contains
   224  // exactly one entry per record.
   225  //
   226  //go:linkname runtime_pprof_readProfile runtime/pprof.readProfile
   227  func runtime_pprof_readProfile() ([]uint64, []unsafe.Pointer, bool) {
   228  	lock(&cpuprof.lock)
   229  	log := cpuprof.log
   230  	unlock(&cpuprof.lock)
   231  	data, tags, eof := log.read(profBufBlocking)
   232  	if len(data) == 0 && eof {
   233  		lock(&cpuprof.lock)
   234  		cpuprof.log = nil
   235  		unlock(&cpuprof.lock)
   236  	}
   237  	return data, tags, eof
   238  }
   239  

View as plain text