Source file src/runtime/pprof/pprof.go

     1  // Copyright 2010 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  // Package pprof writes runtime profiling data in the format expected
     6  // by the pprof visualization tool.
     7  //
     8  // # Profiling a Go program
     9  //
    10  // The first step to profiling a Go program is to enable profiling.
    11  // Support for profiling benchmarks built with the standard testing
    12  // package is built into go test. For example, the following command
    13  // runs benchmarks in the current directory and writes the CPU and
    14  // memory profiles to cpu.prof and mem.prof:
    15  //
    16  //	go test -cpuprofile cpu.prof -memprofile mem.prof -bench .
    17  //
    18  // To add equivalent profiling support to a standalone program, add
    19  // code like the following to your main function:
    20  //
    21  //	var cpuprofile = flag.String("cpuprofile", "", "write cpu profile to `file`")
    22  //	var memprofile = flag.String("memprofile", "", "write memory profile to `file`")
    23  //
    24  //	func main() {
    25  //	    flag.Parse()
    26  //	    if *cpuprofile != "" {
    27  //	        f, err := os.Create(*cpuprofile)
    28  //	        if err != nil {
    29  //	            log.Fatal("could not create CPU profile: ", err)
    30  //	        }
    31  //	        defer f.Close() // error handling omitted for example
    32  //	        if err := pprof.StartCPUProfile(f); err != nil {
    33  //	            log.Fatal("could not start CPU profile: ", err)
    34  //	        }
    35  //	        defer pprof.StopCPUProfile()
    36  //	    }
    37  //
    38  //	    // ... rest of the program ...
    39  //
    40  //	    if *memprofile != "" {
    41  //	        f, err := os.Create(*memprofile)
    42  //	        if err != nil {
    43  //	            log.Fatal("could not create memory profile: ", err)
    44  //	        }
    45  //	        defer f.Close() // error handling omitted for example
    46  //	        runtime.GC() // get up-to-date statistics
    47  //	        if err := pprof.WriteHeapProfile(f); err != nil {
    48  //	            log.Fatal("could not write memory profile: ", err)
    49  //	        }
    50  //	    }
    51  //	}
    52  //
    53  // There is also a standard HTTP interface to profiling data. Adding
    54  // the following line will install handlers under the /debug/pprof/
    55  // URL to download live profiles:
    56  //
    57  //	import _ "net/http/pprof"
    58  //
    59  // See the net/http/pprof package for more details.
    60  //
    61  // Profiles can then be visualized with the pprof tool:
    62  //
    63  //	go tool pprof cpu.prof
    64  //
    65  // There are many commands available from the pprof command line.
    66  // Commonly used commands include "top", which prints a summary of the
    67  // top program hot-spots, and "web", which opens an interactive graph
    68  // of hot-spots and their call graphs. Use "help" for information on
    69  // all pprof commands.
    70  //
    71  // For more information about pprof, see
    72  // https://github.com/google/pprof/blob/master/doc/README.md.
    73  package pprof
    74  
    75  import (
    76  	"bufio"
    77  	"bytes"
    78  	"fmt"
    79  	"internal/abi"
    80  	"io"
    81  	"runtime"
    82  	"sort"
    83  	"strings"
    84  	"sync"
    85  	"text/tabwriter"
    86  	"time"
    87  	"unsafe"
    88  )
    89  
    90  // BUG(rsc): Profiles are only as good as the kernel support used to generate them.
    91  // See https://golang.org/issue/13841 for details about known problems.
    92  
    93  // A Profile is a collection of stack traces showing the call sequences
    94  // that led to instances of a particular event, such as allocation.
    95  // Packages can create and maintain their own profiles; the most common
    96  // use is for tracking resources that must be explicitly closed, such as files
    97  // or network connections.
    98  //
    99  // A Profile's methods can be called from multiple goroutines simultaneously.
   100  //
   101  // Each Profile has a unique name. A few profiles are predefined:
   102  //
   103  //	goroutine    - stack traces of all current goroutines
   104  //	heap         - a sampling of memory allocations of live objects
   105  //	allocs       - a sampling of all past memory allocations
   106  //	threadcreate - stack traces that led to the creation of new OS threads
   107  //	block        - stack traces that led to blocking on synchronization primitives
   108  //	mutex        - stack traces of holders of contended mutexes
   109  //
   110  // These predefined profiles maintain themselves and panic on an explicit
   111  // Add or Remove method call.
   112  //
   113  // The heap profile reports statistics as of the most recently completed
   114  // garbage collection; it elides more recent allocation to avoid skewing
   115  // the profile away from live data and toward garbage.
   116  // If there has been no garbage collection at all, the heap profile reports
   117  // all known allocations. This exception helps mainly in programs running
   118  // without garbage collection enabled, usually for debugging purposes.
   119  //
   120  // The heap profile tracks both the allocation sites for all live objects in
   121  // the application memory and for all objects allocated since the program start.
   122  // Pprof's -inuse_space, -inuse_objects, -alloc_space, and -alloc_objects
   123  // flags select which to display, defaulting to -inuse_space (live objects,
   124  // scaled by size).
   125  //
   126  // The allocs profile is the same as the heap profile but changes the default
   127  // pprof display to -alloc_space, the total number of bytes allocated since
   128  // the program began (including garbage-collected bytes).
   129  //
   130  // The CPU profile is not available as a Profile. It has a special API,
   131  // the StartCPUProfile and StopCPUProfile functions, because it streams
   132  // output to a writer during profiling.
   133  type Profile struct {
   134  	name  string
   135  	mu    sync.Mutex
   136  	m     map[any][]uintptr
   137  	count func() int
   138  	write func(io.Writer, int) error
   139  }
   140  
   141  // profiles records all registered profiles.
   142  var profiles struct {
   143  	mu sync.Mutex
   144  	m  map[string]*Profile
   145  }
   146  
   147  var goroutineProfile = &Profile{
   148  	name:  "goroutine",
   149  	count: countGoroutine,
   150  	write: writeGoroutine,
   151  }
   152  
   153  var threadcreateProfile = &Profile{
   154  	name:  "threadcreate",
   155  	count: countThreadCreate,
   156  	write: writeThreadCreate,
   157  }
   158  
   159  var heapProfile = &Profile{
   160  	name:  "heap",
   161  	count: countHeap,
   162  	write: writeHeap,
   163  }
   164  
   165  var allocsProfile = &Profile{
   166  	name:  "allocs",
   167  	count: countHeap, // identical to heap profile
   168  	write: writeAlloc,
   169  }
   170  
   171  var blockProfile = &Profile{
   172  	name:  "block",
   173  	count: countBlock,
   174  	write: writeBlock,
   175  }
   176  
   177  var mutexProfile = &Profile{
   178  	name:  "mutex",
   179  	count: countMutex,
   180  	write: writeMutex,
   181  }
   182  
   183  func lockProfiles() {
   184  	profiles.mu.Lock()
   185  	if profiles.m == nil {
   186  		// Initial built-in profiles.
   187  		profiles.m = map[string]*Profile{
   188  			"goroutine":    goroutineProfile,
   189  			"threadcreate": threadcreateProfile,
   190  			"heap":         heapProfile,
   191  			"allocs":       allocsProfile,
   192  			"block":        blockProfile,
   193  			"mutex":        mutexProfile,
   194  		}
   195  	}
   196  }
   197  
   198  func unlockProfiles() {
   199  	profiles.mu.Unlock()
   200  }
   201  
   202  // NewProfile creates a new profile with the given name.
   203  // If a profile with that name already exists, NewProfile panics.
   204  // The convention is to use a 'import/path.' prefix to create
   205  // separate name spaces for each package.
   206  // For compatibility with various tools that read pprof data,
   207  // profile names should not contain spaces.
   208  func NewProfile(name string) *Profile {
   209  	lockProfiles()
   210  	defer unlockProfiles()
   211  	if name == "" {
   212  		panic("pprof: NewProfile with empty name")
   213  	}
   214  	if profiles.m[name] != nil {
   215  		panic("pprof: NewProfile name already in use: " + name)
   216  	}
   217  	p := &Profile{
   218  		name: name,
   219  		m:    map[any][]uintptr{},
   220  	}
   221  	profiles.m[name] = p
   222  	return p
   223  }
   224  
   225  // Lookup returns the profile with the given name, or nil if no such profile exists.
   226  func Lookup(name string) *Profile {
   227  	lockProfiles()
   228  	defer unlockProfiles()
   229  	return profiles.m[name]
   230  }
   231  
   232  // Profiles returns a slice of all the known profiles, sorted by name.
   233  func Profiles() []*Profile {
   234  	lockProfiles()
   235  	defer unlockProfiles()
   236  
   237  	all := make([]*Profile, 0, len(profiles.m))
   238  	for _, p := range profiles.m {
   239  		all = append(all, p)
   240  	}
   241  
   242  	sort.Slice(all, func(i, j int) bool { return all[i].name < all[j].name })
   243  	return all
   244  }
   245  
   246  // Name returns this profile's name, which can be passed to Lookup to reobtain the profile.
   247  func (p *Profile) Name() string {
   248  	return p.name
   249  }
   250  
   251  // Count returns the number of execution stacks currently in the profile.
   252  func (p *Profile) Count() int {
   253  	p.mu.Lock()
   254  	defer p.mu.Unlock()
   255  	if p.count != nil {
   256  		return p.count()
   257  	}
   258  	return len(p.m)
   259  }
   260  
   261  // Add adds the current execution stack to the profile, associated with value.
   262  // Add stores value in an internal map, so value must be suitable for use as
   263  // a map key and will not be garbage collected until the corresponding
   264  // call to Remove. Add panics if the profile already contains a stack for value.
   265  //
   266  // The skip parameter has the same meaning as runtime.Caller's skip
   267  // and controls where the stack trace begins. Passing skip=0 begins the
   268  // trace in the function calling Add. For example, given this
   269  // execution stack:
   270  //
   271  //	Add
   272  //	called from rpc.NewClient
   273  //	called from mypkg.Run
   274  //	called from main.main
   275  //
   276  // Passing skip=0 begins the stack trace at the call to Add inside rpc.NewClient.
   277  // Passing skip=1 begins the stack trace at the call to NewClient inside mypkg.Run.
   278  func (p *Profile) Add(value any, skip int) {
   279  	if p.name == "" {
   280  		panic("pprof: use of uninitialized Profile")
   281  	}
   282  	if p.write != nil {
   283  		panic("pprof: Add called on built-in Profile " + p.name)
   284  	}
   285  
   286  	stk := make([]uintptr, 32)
   287  	n := runtime.Callers(skip+1, stk[:])
   288  	stk = stk[:n]
   289  	if len(stk) == 0 {
   290  		// The value for skip is too large, and there's no stack trace to record.
   291  		stk = []uintptr{abi.FuncPCABIInternal(lostProfileEvent)}
   292  	}
   293  
   294  	p.mu.Lock()
   295  	defer p.mu.Unlock()
   296  	if p.m[value] != nil {
   297  		panic("pprof: Profile.Add of duplicate value")
   298  	}
   299  	p.m[value] = stk
   300  }
   301  
   302  // Remove removes the execution stack associated with value from the profile.
   303  // It is a no-op if the value is not in the profile.
   304  func (p *Profile) Remove(value any) {
   305  	p.mu.Lock()
   306  	defer p.mu.Unlock()
   307  	delete(p.m, value)
   308  }
   309  
   310  // WriteTo writes a pprof-formatted snapshot of the profile to w.
   311  // If a write to w returns an error, WriteTo returns that error.
   312  // Otherwise, WriteTo returns nil.
   313  //
   314  // The debug parameter enables additional output.
   315  // Passing debug=0 writes the gzip-compressed protocol buffer described
   316  // in https://github.com/google/pprof/tree/master/proto#overview.
   317  // Passing debug=1 writes the legacy text format with comments
   318  // translating addresses to function names and line numbers, so that a
   319  // programmer can read the profile without tools.
   320  //
   321  // The predefined profiles may assign meaning to other debug values;
   322  // for example, when printing the "goroutine" profile, debug=2 means to
   323  // print the goroutine stacks in the same form that a Go program uses
   324  // when dying due to an unrecovered panic.
   325  func (p *Profile) WriteTo(w io.Writer, debug int) error {
   326  	if p.name == "" {
   327  		panic("pprof: use of zero Profile")
   328  	}
   329  	if p.write != nil {
   330  		return p.write(w, debug)
   331  	}
   332  
   333  	// Obtain consistent snapshot under lock; then process without lock.
   334  	p.mu.Lock()
   335  	all := make([][]uintptr, 0, len(p.m))
   336  	for _, stk := range p.m {
   337  		all = append(all, stk)
   338  	}
   339  	p.mu.Unlock()
   340  
   341  	// Map order is non-deterministic; make output deterministic.
   342  	sort.Slice(all, func(i, j int) bool {
   343  		t, u := all[i], all[j]
   344  		for k := 0; k < len(t) && k < len(u); k++ {
   345  			if t[k] != u[k] {
   346  				return t[k] < u[k]
   347  			}
   348  		}
   349  		return len(t) < len(u)
   350  	})
   351  
   352  	return printCountProfile(w, debug, p.name, stackProfile(all))
   353  }
   354  
   355  type stackProfile [][]uintptr
   356  
   357  func (x stackProfile) Len() int              { return len(x) }
   358  func (x stackProfile) Stack(i int) []uintptr { return x[i] }
   359  func (x stackProfile) Label(i int) *labelMap { return nil }
   360  
   361  // A countProfile is a set of stack traces to be printed as counts
   362  // grouped by stack trace. There are multiple implementations:
   363  // all that matters is that we can find out how many traces there are
   364  // and obtain each trace in turn.
   365  type countProfile interface {
   366  	Len() int
   367  	Stack(i int) []uintptr
   368  	Label(i int) *labelMap
   369  }
   370  
   371  // printCountCycleProfile outputs block profile records (for block or mutex profiles)
   372  // as the pprof-proto format output. Translations from cycle count to time duration
   373  // are done because The proto expects count and time (nanoseconds) instead of count
   374  // and the number of cycles for block, contention profiles.
   375  // Possible 'scaler' functions are scaleBlockProfile and scaleMutexProfile.
   376  func printCountCycleProfile(w io.Writer, countName, cycleName string, scaler func(int64, float64) (int64, float64), records []runtime.BlockProfileRecord) error {
   377  	// Output profile in protobuf form.
   378  	b := newProfileBuilder(w)
   379  	b.pbValueType(tagProfile_PeriodType, countName, "count")
   380  	b.pb.int64Opt(tagProfile_Period, 1)
   381  	b.pbValueType(tagProfile_SampleType, countName, "count")
   382  	b.pbValueType(tagProfile_SampleType, cycleName, "nanoseconds")
   383  
   384  	cpuGHz := float64(runtime_cyclesPerSecond()) / 1e9
   385  
   386  	values := []int64{0, 0}
   387  	var locs []uint64
   388  	for _, r := range records {
   389  		count, nanosec := scaler(r.Count, float64(r.Cycles)/cpuGHz)
   390  		values[0] = count
   391  		values[1] = int64(nanosec)
   392  		// For count profiles, all stack addresses are
   393  		// return PCs, which is what appendLocsForStack expects.
   394  		locs = b.appendLocsForStack(locs[:0], r.Stack())
   395  		b.pbSample(values, locs, nil)
   396  	}
   397  	b.build()
   398  	return nil
   399  }
   400  
   401  // printCountProfile prints a countProfile at the specified debug level.
   402  // The profile will be in compressed proto format unless debug is nonzero.
   403  func printCountProfile(w io.Writer, debug int, name string, p countProfile) error {
   404  	// Build count of each stack.
   405  	var buf bytes.Buffer
   406  	key := func(stk []uintptr, lbls *labelMap) string {
   407  		buf.Reset()
   408  		fmt.Fprintf(&buf, "@")
   409  		for _, pc := range stk {
   410  			fmt.Fprintf(&buf, " %#x", pc)
   411  		}
   412  		if lbls != nil {
   413  			buf.WriteString("\n# labels: ")
   414  			buf.WriteString(lbls.String())
   415  		}
   416  		return buf.String()
   417  	}
   418  	count := map[string]int{}
   419  	index := map[string]int{}
   420  	var keys []string
   421  	n := p.Len()
   422  	for i := 0; i < n; i++ {
   423  		k := key(p.Stack(i), p.Label(i))
   424  		if count[k] == 0 {
   425  			index[k] = i
   426  			keys = append(keys, k)
   427  		}
   428  		count[k]++
   429  	}
   430  
   431  	sort.Sort(&keysByCount{keys, count})
   432  
   433  	if debug > 0 {
   434  		// Print debug profile in legacy format
   435  		tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
   436  		fmt.Fprintf(tw, "%s profile: total %d\n", name, p.Len())
   437  		for _, k := range keys {
   438  			fmt.Fprintf(tw, "%d %s\n", count[k], k)
   439  			printStackRecord(tw, p.Stack(index[k]), false)
   440  		}
   441  		return tw.Flush()
   442  	}
   443  
   444  	// Output profile in protobuf form.
   445  	b := newProfileBuilder(w)
   446  	b.pbValueType(tagProfile_PeriodType, name, "count")
   447  	b.pb.int64Opt(tagProfile_Period, 1)
   448  	b.pbValueType(tagProfile_SampleType, name, "count")
   449  
   450  	values := []int64{0}
   451  	var locs []uint64
   452  	for _, k := range keys {
   453  		values[0] = int64(count[k])
   454  		// For count profiles, all stack addresses are
   455  		// return PCs, which is what appendLocsForStack expects.
   456  		locs = b.appendLocsForStack(locs[:0], p.Stack(index[k]))
   457  		idx := index[k]
   458  		var labels func()
   459  		if p.Label(idx) != nil {
   460  			labels = func() {
   461  				for k, v := range *p.Label(idx) {
   462  					b.pbLabel(tagSample_Label, k, v, 0)
   463  				}
   464  			}
   465  		}
   466  		b.pbSample(values, locs, labels)
   467  	}
   468  	b.build()
   469  	return nil
   470  }
   471  
   472  // keysByCount sorts keys with higher counts first, breaking ties by key string order.
   473  type keysByCount struct {
   474  	keys  []string
   475  	count map[string]int
   476  }
   477  
   478  func (x *keysByCount) Len() int      { return len(x.keys) }
   479  func (x *keysByCount) Swap(i, j int) { x.keys[i], x.keys[j] = x.keys[j], x.keys[i] }
   480  func (x *keysByCount) Less(i, j int) bool {
   481  	ki, kj := x.keys[i], x.keys[j]
   482  	ci, cj := x.count[ki], x.count[kj]
   483  	if ci != cj {
   484  		return ci > cj
   485  	}
   486  	return ki < kj
   487  }
   488  
   489  // printStackRecord prints the function + source line information
   490  // for a single stack trace.
   491  func printStackRecord(w io.Writer, stk []uintptr, allFrames bool) {
   492  	show := allFrames
   493  	frames := runtime.CallersFrames(stk)
   494  	for {
   495  		frame, more := frames.Next()
   496  		name := frame.Function
   497  		if name == "" {
   498  			show = true
   499  			fmt.Fprintf(w, "#\t%#x\n", frame.PC)
   500  		} else if name != "runtime.goexit" && (show || !strings.HasPrefix(name, "runtime.")) {
   501  			// Hide runtime.goexit and any runtime functions at the beginning.
   502  			// This is useful mainly for allocation traces.
   503  			show = true
   504  			fmt.Fprintf(w, "#\t%#x\t%s+%#x\t%s:%d\n", frame.PC, name, frame.PC-frame.Entry, frame.File, frame.Line)
   505  		}
   506  		if !more {
   507  			break
   508  		}
   509  	}
   510  	if !show {
   511  		// We didn't print anything; do it again,
   512  		// and this time include runtime functions.
   513  		printStackRecord(w, stk, true)
   514  		return
   515  	}
   516  	fmt.Fprintf(w, "\n")
   517  }
   518  
   519  // Interface to system profiles.
   520  
   521  // WriteHeapProfile is shorthand for Lookup("heap").WriteTo(w, 0).
   522  // It is preserved for backwards compatibility.
   523  func WriteHeapProfile(w io.Writer) error {
   524  	return writeHeap(w, 0)
   525  }
   526  
   527  // countHeap returns the number of records in the heap profile.
   528  func countHeap() int {
   529  	n, _ := runtime.MemProfile(nil, true)
   530  	return n
   531  }
   532  
   533  // writeHeap writes the current runtime heap profile to w.
   534  func writeHeap(w io.Writer, debug int) error {
   535  	return writeHeapInternal(w, debug, "")
   536  }
   537  
   538  // writeAlloc writes the current runtime heap profile to w
   539  // with the total allocation space as the default sample type.
   540  func writeAlloc(w io.Writer, debug int) error {
   541  	return writeHeapInternal(w, debug, "alloc_space")
   542  }
   543  
   544  func writeHeapInternal(w io.Writer, debug int, defaultSampleType string) error {
   545  	var memStats *runtime.MemStats
   546  	if debug != 0 {
   547  		// Read mem stats first, so that our other allocations
   548  		// do not appear in the statistics.
   549  		memStats = new(runtime.MemStats)
   550  		runtime.ReadMemStats(memStats)
   551  	}
   552  
   553  	// Find out how many records there are (MemProfile(nil, true)),
   554  	// allocate that many records, and get the data.
   555  	// There's a race—more records might be added between
   556  	// the two calls—so allocate a few extra records for safety
   557  	// and also try again if we're very unlucky.
   558  	// The loop should only execute one iteration in the common case.
   559  	var p []runtime.MemProfileRecord
   560  	n, ok := runtime.MemProfile(nil, true)
   561  	for {
   562  		// Allocate room for a slightly bigger profile,
   563  		// in case a few more entries have been added
   564  		// since the call to MemProfile.
   565  		p = make([]runtime.MemProfileRecord, n+50)
   566  		n, ok = runtime.MemProfile(p, true)
   567  		if ok {
   568  			p = p[0:n]
   569  			break
   570  		}
   571  		// Profile grew; try again.
   572  	}
   573  
   574  	if debug == 0 {
   575  		return writeHeapProto(w, p, int64(runtime.MemProfileRate), defaultSampleType)
   576  	}
   577  
   578  	sort.Slice(p, func(i, j int) bool { return p[i].InUseBytes() > p[j].InUseBytes() })
   579  
   580  	b := bufio.NewWriter(w)
   581  	tw := tabwriter.NewWriter(b, 1, 8, 1, '\t', 0)
   582  	w = tw
   583  
   584  	var total runtime.MemProfileRecord
   585  	for i := range p {
   586  		r := &p[i]
   587  		total.AllocBytes += r.AllocBytes
   588  		total.AllocObjects += r.AllocObjects
   589  		total.FreeBytes += r.FreeBytes
   590  		total.FreeObjects += r.FreeObjects
   591  	}
   592  
   593  	// Technically the rate is MemProfileRate not 2*MemProfileRate,
   594  	// but early versions of the C++ heap profiler reported 2*MemProfileRate,
   595  	// so that's what pprof has come to expect.
   596  	fmt.Fprintf(w, "heap profile: %d: %d [%d: %d] @ heap/%d\n",
   597  		total.InUseObjects(), total.InUseBytes(),
   598  		total.AllocObjects, total.AllocBytes,
   599  		2*runtime.MemProfileRate)
   600  
   601  	for i := range p {
   602  		r := &p[i]
   603  		fmt.Fprintf(w, "%d: %d [%d: %d] @",
   604  			r.InUseObjects(), r.InUseBytes(),
   605  			r.AllocObjects, r.AllocBytes)
   606  		for _, pc := range r.Stack() {
   607  			fmt.Fprintf(w, " %#x", pc)
   608  		}
   609  		fmt.Fprintf(w, "\n")
   610  		printStackRecord(w, r.Stack(), false)
   611  	}
   612  
   613  	// Print memstats information too.
   614  	// Pprof will ignore, but useful for people
   615  	s := memStats
   616  	fmt.Fprintf(w, "\n# runtime.MemStats\n")
   617  	fmt.Fprintf(w, "# Alloc = %d\n", s.Alloc)
   618  	fmt.Fprintf(w, "# TotalAlloc = %d\n", s.TotalAlloc)
   619  	fmt.Fprintf(w, "# Sys = %d\n", s.Sys)
   620  	fmt.Fprintf(w, "# Lookups = %d\n", s.Lookups)
   621  	fmt.Fprintf(w, "# Mallocs = %d\n", s.Mallocs)
   622  	fmt.Fprintf(w, "# Frees = %d\n", s.Frees)
   623  
   624  	fmt.Fprintf(w, "# HeapAlloc = %d\n", s.HeapAlloc)
   625  	fmt.Fprintf(w, "# HeapSys = %d\n", s.HeapSys)
   626  	fmt.Fprintf(w, "# HeapIdle = %d\n", s.HeapIdle)
   627  	fmt.Fprintf(w, "# HeapInuse = %d\n", s.HeapInuse)
   628  	fmt.Fprintf(w, "# HeapReleased = %d\n", s.HeapReleased)
   629  	fmt.Fprintf(w, "# HeapObjects = %d\n", s.HeapObjects)
   630  
   631  	fmt.Fprintf(w, "# Stack = %d / %d\n", s.StackInuse, s.StackSys)
   632  	fmt.Fprintf(w, "# MSpan = %d / %d\n", s.MSpanInuse, s.MSpanSys)
   633  	fmt.Fprintf(w, "# MCache = %d / %d\n", s.MCacheInuse, s.MCacheSys)
   634  	fmt.Fprintf(w, "# BuckHashSys = %d\n", s.BuckHashSys)
   635  	fmt.Fprintf(w, "# GCSys = %d\n", s.GCSys)
   636  	fmt.Fprintf(w, "# OtherSys = %d\n", s.OtherSys)
   637  
   638  	fmt.Fprintf(w, "# NextGC = %d\n", s.NextGC)
   639  	fmt.Fprintf(w, "# LastGC = %d\n", s.LastGC)
   640  	fmt.Fprintf(w, "# PauseNs = %d\n", s.PauseNs)
   641  	fmt.Fprintf(w, "# PauseEnd = %d\n", s.PauseEnd)
   642  	fmt.Fprintf(w, "# NumGC = %d\n", s.NumGC)
   643  	fmt.Fprintf(w, "# NumForcedGC = %d\n", s.NumForcedGC)
   644  	fmt.Fprintf(w, "# GCCPUFraction = %v\n", s.GCCPUFraction)
   645  	fmt.Fprintf(w, "# DebugGC = %v\n", s.DebugGC)
   646  
   647  	// Also flush out MaxRSS on supported platforms.
   648  	addMaxRSS(w)
   649  
   650  	tw.Flush()
   651  	return b.Flush()
   652  }
   653  
   654  // countThreadCreate returns the size of the current ThreadCreateProfile.
   655  func countThreadCreate() int {
   656  	n, _ := runtime.ThreadCreateProfile(nil)
   657  	return n
   658  }
   659  
   660  // writeThreadCreate writes the current runtime ThreadCreateProfile to w.
   661  func writeThreadCreate(w io.Writer, debug int) error {
   662  	// Until https://golang.org/issues/6104 is addressed, wrap
   663  	// ThreadCreateProfile because there's no point in tracking labels when we
   664  	// don't get any stack-traces.
   665  	return writeRuntimeProfile(w, debug, "threadcreate", func(p []runtime.StackRecord, _ []unsafe.Pointer) (n int, ok bool) {
   666  		return runtime.ThreadCreateProfile(p)
   667  	})
   668  }
   669  
   670  // countGoroutine returns the number of goroutines.
   671  func countGoroutine() int {
   672  	return runtime.NumGoroutine()
   673  }
   674  
   675  // runtime_goroutineProfileWithLabels is defined in runtime/mprof.go
   676  func runtime_goroutineProfileWithLabels(p []runtime.StackRecord, labels []unsafe.Pointer) (n int, ok bool)
   677  
   678  // writeGoroutine writes the current runtime GoroutineProfile to w.
   679  func writeGoroutine(w io.Writer, debug int) error {
   680  	if debug >= 2 {
   681  		return writeGoroutineStacks(w)
   682  	}
   683  	return writeRuntimeProfile(w, debug, "goroutine", runtime_goroutineProfileWithLabels)
   684  }
   685  
   686  func writeGoroutineStacks(w io.Writer) error {
   687  	// We don't know how big the buffer needs to be to collect
   688  	// all the goroutines. Start with 1 MB and try a few times, doubling each time.
   689  	// Give up and use a truncated trace if 64 MB is not enough.
   690  	buf := make([]byte, 1<<20)
   691  	for i := 0; ; i++ {
   692  		n := runtime.Stack(buf, true)
   693  		if n < len(buf) {
   694  			buf = buf[:n]
   695  			break
   696  		}
   697  		if len(buf) >= 64<<20 {
   698  			// Filled 64 MB - stop there.
   699  			break
   700  		}
   701  		buf = make([]byte, 2*len(buf))
   702  	}
   703  	_, err := w.Write(buf)
   704  	return err
   705  }
   706  
   707  func writeRuntimeProfile(w io.Writer, debug int, name string, fetch func([]runtime.StackRecord, []unsafe.Pointer) (int, bool)) error {
   708  	// Find out how many records there are (fetch(nil)),
   709  	// allocate that many records, and get the data.
   710  	// There's a race—more records might be added between
   711  	// the two calls—so allocate a few extra records for safety
   712  	// and also try again if we're very unlucky.
   713  	// The loop should only execute one iteration in the common case.
   714  	var p []runtime.StackRecord
   715  	var labels []unsafe.Pointer
   716  	n, ok := fetch(nil, nil)
   717  	for {
   718  		// Allocate room for a slightly bigger profile,
   719  		// in case a few more entries have been added
   720  		// since the call to ThreadProfile.
   721  		p = make([]runtime.StackRecord, n+10)
   722  		labels = make([]unsafe.Pointer, n+10)
   723  		n, ok = fetch(p, labels)
   724  		if ok {
   725  			p = p[0:n]
   726  			break
   727  		}
   728  		// Profile grew; try again.
   729  	}
   730  
   731  	return printCountProfile(w, debug, name, &runtimeProfile{p, labels})
   732  }
   733  
   734  type runtimeProfile struct {
   735  	stk    []runtime.StackRecord
   736  	labels []unsafe.Pointer
   737  }
   738  
   739  func (p *runtimeProfile) Len() int              { return len(p.stk) }
   740  func (p *runtimeProfile) Stack(i int) []uintptr { return p.stk[i].Stack() }
   741  func (p *runtimeProfile) Label(i int) *labelMap { return (*labelMap)(p.labels[i]) }
   742  
   743  var cpu struct {
   744  	sync.Mutex
   745  	profiling bool
   746  	done      chan bool
   747  }
   748  
   749  // StartCPUProfile enables CPU profiling for the current process.
   750  // While profiling, the profile will be buffered and written to w.
   751  // StartCPUProfile returns an error if profiling is already enabled.
   752  //
   753  // On Unix-like systems, StartCPUProfile does not work by default for
   754  // Go code built with -buildmode=c-archive or -buildmode=c-shared.
   755  // StartCPUProfile relies on the SIGPROF signal, but that signal will
   756  // be delivered to the main program's SIGPROF signal handler (if any)
   757  // not to the one used by Go. To make it work, call os/signal.Notify
   758  // for syscall.SIGPROF, but note that doing so may break any profiling
   759  // being done by the main program.
   760  func StartCPUProfile(w io.Writer) error {
   761  	// The runtime routines allow a variable profiling rate,
   762  	// but in practice operating systems cannot trigger signals
   763  	// at more than about 500 Hz, and our processing of the
   764  	// signal is not cheap (mostly getting the stack trace).
   765  	// 100 Hz is a reasonable choice: it is frequent enough to
   766  	// produce useful data, rare enough not to bog down the
   767  	// system, and a nice round number to make it easy to
   768  	// convert sample counts to seconds. Instead of requiring
   769  	// each client to specify the frequency, we hard code it.
   770  	const hz = 100
   771  
   772  	cpu.Lock()
   773  	defer cpu.Unlock()
   774  	if cpu.done == nil {
   775  		cpu.done = make(chan bool)
   776  	}
   777  	// Double-check.
   778  	if cpu.profiling {
   779  		return fmt.Errorf("cpu profiling already in use")
   780  	}
   781  	cpu.profiling = true
   782  	runtime.SetCPUProfileRate(hz)
   783  	go profileWriter(w)
   784  	return nil
   785  }
   786  
   787  // readProfile, provided by the runtime, returns the next chunk of
   788  // binary CPU profiling stack trace data, blocking until data is available.
   789  // If profiling is turned off and all the profile data accumulated while it was
   790  // on has been returned, readProfile returns eof=true.
   791  // The caller must save the returned data and tags before calling readProfile again.
   792  func readProfile() (data []uint64, tags []unsafe.Pointer, eof bool)
   793  
   794  func profileWriter(w io.Writer) {
   795  	b := newProfileBuilder(w)
   796  	var err error
   797  	for {
   798  		time.Sleep(100 * time.Millisecond)
   799  		data, tags, eof := readProfile()
   800  		if e := b.addCPUData(data, tags); e != nil && err == nil {
   801  			err = e
   802  		}
   803  		if eof {
   804  			break
   805  		}
   806  	}
   807  	if err != nil {
   808  		// The runtime should never produce an invalid or truncated profile.
   809  		// It drops records that can't fit into its log buffers.
   810  		panic("runtime/pprof: converting profile: " + err.Error())
   811  	}
   812  	b.build()
   813  	cpu.done <- true
   814  }
   815  
   816  // StopCPUProfile stops the current CPU profile, if any.
   817  // StopCPUProfile only returns after all the writes for the
   818  // profile have completed.
   819  func StopCPUProfile() {
   820  	cpu.Lock()
   821  	defer cpu.Unlock()
   822  
   823  	if !cpu.profiling {
   824  		return
   825  	}
   826  	cpu.profiling = false
   827  	runtime.SetCPUProfileRate(0)
   828  	<-cpu.done
   829  }
   830  
   831  // countBlock returns the number of records in the blocking profile.
   832  func countBlock() int {
   833  	n, _ := runtime.BlockProfile(nil)
   834  	return n
   835  }
   836  
   837  // countMutex returns the number of records in the mutex profile.
   838  func countMutex() int {
   839  	n, _ := runtime.MutexProfile(nil)
   840  	return n
   841  }
   842  
   843  // writeBlock writes the current blocking profile to w.
   844  func writeBlock(w io.Writer, debug int) error {
   845  	return writeProfileInternal(w, debug, "contention", runtime.BlockProfile, scaleBlockProfile)
   846  }
   847  
   848  func scaleBlockProfile(cnt int64, ns float64) (int64, float64) {
   849  	// Do nothing.
   850  	// The current way of block profile sampling makes it
   851  	// hard to compute the unsampled number. The legacy block
   852  	// profile parse doesn't attempt to scale or unsample.
   853  	return cnt, ns
   854  }
   855  
   856  // writeMutex writes the current mutex profile to w.
   857  func writeMutex(w io.Writer, debug int) error {
   858  	return writeProfileInternal(w, debug, "mutex", runtime.MutexProfile, scaleMutexProfile)
   859  }
   860  
   861  // writeProfileInternal writes the current blocking or mutex profile depending on the passed parameters
   862  func writeProfileInternal(w io.Writer, debug int, name string, runtimeProfile func([]runtime.BlockProfileRecord) (int, bool), scaleProfile func(int64, float64) (int64, float64)) error {
   863  	var p []runtime.BlockProfileRecord
   864  	n, ok := runtimeProfile(nil)
   865  	for {
   866  		p = make([]runtime.BlockProfileRecord, n+50)
   867  		n, ok = runtimeProfile(p)
   868  		if ok {
   869  			p = p[:n]
   870  			break
   871  		}
   872  	}
   873  
   874  	sort.Slice(p, func(i, j int) bool { return p[i].Cycles > p[j].Cycles })
   875  
   876  	if debug <= 0 {
   877  		return printCountCycleProfile(w, "contentions", "delay", scaleProfile, p)
   878  	}
   879  
   880  	b := bufio.NewWriter(w)
   881  	tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
   882  	w = tw
   883  
   884  	fmt.Fprintf(w, "--- %v:\n", name)
   885  	fmt.Fprintf(w, "cycles/second=%v\n", runtime_cyclesPerSecond())
   886  	if name == "mutex" {
   887  		fmt.Fprintf(w, "sampling period=%d\n", runtime.SetMutexProfileFraction(-1))
   888  	}
   889  	for i := range p {
   890  		r := &p[i]
   891  		fmt.Fprintf(w, "%v %v @", r.Cycles, r.Count)
   892  		for _, pc := range r.Stack() {
   893  			fmt.Fprintf(w, " %#x", pc)
   894  		}
   895  		fmt.Fprint(w, "\n")
   896  		if debug > 0 {
   897  			printStackRecord(w, r.Stack(), true)
   898  		}
   899  	}
   900  
   901  	if tw != nil {
   902  		tw.Flush()
   903  	}
   904  	return b.Flush()
   905  }
   906  
   907  func scaleMutexProfile(cnt int64, ns float64) (int64, float64) {
   908  	period := runtime.SetMutexProfileFraction(-1)
   909  	return cnt * int64(period), ns * float64(period)
   910  }
   911  
   912  func runtime_cyclesPerSecond() int64
   913  

View as plain text