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  	"fmt"
    78  	"internal/abi"
    79  	"io"
    80  	"runtime"
    81  	"sort"
    82  	"strings"
    83  	"sync"
    84  	"text/tabwriter"
    85  	"time"
    86  	"unsafe"
    87  )
    88  
    89  // BUG(rsc): Profiles are only as good as the kernel support used to generate them.
    90  // See https://golang.org/issue/13841 for details about known problems.
    91  
    92  // A Profile is a collection of stack traces showing the call sequences
    93  // that led to instances of a particular event, such as allocation.
    94  // Packages can create and maintain their own profiles; the most common
    95  // use is for tracking resources that must be explicitly closed, such as files
    96  // or network connections.
    97  //
    98  // A Profile's methods can be called from multiple goroutines simultaneously.
    99  //
   100  // Each Profile has a unique name. A few profiles are predefined:
   101  //
   102  //	goroutine    - stack traces of all current goroutines
   103  //	heap         - a sampling of memory allocations of live objects
   104  //	allocs       - a sampling of all past memory allocations
   105  //	threadcreate - stack traces that led to the creation of new OS threads
   106  //	block        - stack traces that led to blocking on synchronization primitives
   107  //	mutex        - stack traces of holders of contended mutexes
   108  //
   109  // These predefined profiles maintain themselves and panic on an explicit
   110  // Add or Remove method call.
   111  //
   112  // The heap profile reports statistics as of the most recently completed
   113  // garbage collection; it elides more recent allocation to avoid skewing
   114  // the profile away from live data and toward garbage.
   115  // If there has been no garbage collection at all, the heap profile reports
   116  // all known allocations. This exception helps mainly in programs running
   117  // without garbage collection enabled, usually for debugging purposes.
   118  //
   119  // The heap profile tracks both the allocation sites for all live objects in
   120  // the application memory and for all objects allocated since the program start.
   121  // Pprof's -inuse_space, -inuse_objects, -alloc_space, and -alloc_objects
   122  // flags select which to display, defaulting to -inuse_space (live objects,
   123  // scaled by size).
   124  //
   125  // The allocs profile is the same as the heap profile but changes the default
   126  // pprof display to -alloc_space, the total number of bytes allocated since
   127  // the program began (including garbage-collected bytes).
   128  //
   129  // The CPU profile is not available as a Profile. It has a special API,
   130  // the StartCPUProfile and StopCPUProfile functions, because it streams
   131  // output to a writer during profiling.
   132  type Profile struct {
   133  	name  string
   134  	mu    sync.Mutex
   135  	m     map[any][]uintptr
   136  	count func() int
   137  	write func(io.Writer, int) error
   138  }
   139  
   140  // profiles records all registered profiles.
   141  var profiles struct {
   142  	mu sync.Mutex
   143  	m  map[string]*Profile
   144  }
   145  
   146  var goroutineProfile = &Profile{
   147  	name:  "goroutine",
   148  	count: countGoroutine,
   149  	write: writeGoroutine,
   150  }
   151  
   152  var threadcreateProfile = &Profile{
   153  	name:  "threadcreate",
   154  	count: countThreadCreate,
   155  	write: writeThreadCreate,
   156  }
   157  
   158  var heapProfile = &Profile{
   159  	name:  "heap",
   160  	count: countHeap,
   161  	write: writeHeap,
   162  }
   163  
   164  var allocsProfile = &Profile{
   165  	name:  "allocs",
   166  	count: countHeap, // identical to heap profile
   167  	write: writeAlloc,
   168  }
   169  
   170  var blockProfile = &Profile{
   171  	name:  "block",
   172  	count: countBlock,
   173  	write: writeBlock,
   174  }
   175  
   176  var mutexProfile = &Profile{
   177  	name:  "mutex",
   178  	count: countMutex,
   179  	write: writeMutex,
   180  }
   181  
   182  func lockProfiles() {
   183  	profiles.mu.Lock()
   184  	if profiles.m == nil {
   185  		// Initial built-in profiles.
   186  		profiles.m = map[string]*Profile{
   187  			"goroutine":    goroutineProfile,
   188  			"threadcreate": threadcreateProfile,
   189  			"heap":         heapProfile,
   190  			"allocs":       allocsProfile,
   191  			"block":        blockProfile,
   192  			"mutex":        mutexProfile,
   193  		}
   194  	}
   195  }
   196  
   197  func unlockProfiles() {
   198  	profiles.mu.Unlock()
   199  }
   200  
   201  // NewProfile creates a new profile with the given name.
   202  // If a profile with that name already exists, NewProfile panics.
   203  // The convention is to use a 'import/path.' prefix to create
   204  // separate name spaces for each package.
   205  // For compatibility with various tools that read pprof data,
   206  // profile names should not contain spaces.
   207  func NewProfile(name string) *Profile {
   208  	lockProfiles()
   209  	defer unlockProfiles()
   210  	if name == "" {
   211  		panic("pprof: NewProfile with empty name")
   212  	}
   213  	if profiles.m[name] != nil {
   214  		panic("pprof: NewProfile name already in use: " + name)
   215  	}
   216  	p := &Profile{
   217  		name: name,
   218  		m:    map[any][]uintptr{},
   219  	}
   220  	profiles.m[name] = p
   221  	return p
   222  }
   223  
   224  // Lookup returns the profile with the given name, or nil if no such profile exists.
   225  func Lookup(name string) *Profile {
   226  	lockProfiles()
   227  	defer unlockProfiles()
   228  	return profiles.m[name]
   229  }
   230  
   231  // Profiles returns a slice of all the known profiles, sorted by name.
   232  func Profiles() []*Profile {
   233  	lockProfiles()
   234  	defer unlockProfiles()
   235  
   236  	all := make([]*Profile, 0, len(profiles.m))
   237  	for _, p := range profiles.m {
   238  		all = append(all, p)
   239  	}
   240  
   241  	sort.Slice(all, func(i, j int) bool { return all[i].name < all[j].name })
   242  	return all
   243  }
   244  
   245  // Name returns this profile's name, which can be passed to Lookup to reobtain the profile.
   246  func (p *Profile) Name() string {
   247  	return p.name
   248  }
   249  
   250  // Count returns the number of execution stacks currently in the profile.
   251  func (p *Profile) Count() int {
   252  	p.mu.Lock()
   253  	defer p.mu.Unlock()
   254  	if p.count != nil {
   255  		return p.count()
   256  	}
   257  	return len(p.m)
   258  }
   259  
   260  // Add adds the current execution stack to the profile, associated with value.
   261  // Add stores value in an internal map, so value must be suitable for use as
   262  // a map key and will not be garbage collected until the corresponding
   263  // call to Remove. Add panics if the profile already contains a stack for value.
   264  //
   265  // The skip parameter has the same meaning as runtime.Caller's skip
   266  // and controls where the stack trace begins. Passing skip=0 begins the
   267  // trace in the function calling Add. For example, given this
   268  // execution stack:
   269  //
   270  //	Add
   271  //	called from rpc.NewClient
   272  //	called from mypkg.Run
   273  //	called from main.main
   274  //
   275  // Passing skip=0 begins the stack trace at the call to Add inside rpc.NewClient.
   276  // Passing skip=1 begins the stack trace at the call to NewClient inside mypkg.Run.
   277  func (p *Profile) Add(value any, skip int) {
   278  	if p.name == "" {
   279  		panic("pprof: use of uninitialized Profile")
   280  	}
   281  	if p.write != nil {
   282  		panic("pprof: Add called on built-in Profile " + p.name)
   283  	}
   284  
   285  	stk := make([]uintptr, 32)
   286  	n := runtime.Callers(skip+1, stk[:])
   287  	stk = stk[:n]
   288  	if len(stk) == 0 {
   289  		// The value for skip is too large, and there's no stack trace to record.
   290  		stk = []uintptr{abi.FuncPCABIInternal(lostProfileEvent)}
   291  	}
   292  
   293  	p.mu.Lock()
   294  	defer p.mu.Unlock()
   295  	if p.m[value] != nil {
   296  		panic("pprof: Profile.Add of duplicate value")
   297  	}
   298  	p.m[value] = stk
   299  }
   300  
   301  // Remove removes the execution stack associated with value from the profile.
   302  // It is a no-op if the value is not in the profile.
   303  func (p *Profile) Remove(value any) {
   304  	p.mu.Lock()
   305  	defer p.mu.Unlock()
   306  	delete(p.m, value)
   307  }
   308  
   309  // WriteTo writes a pprof-formatted snapshot of the profile to w.
   310  // If a write to w returns an error, WriteTo returns that error.
   311  // Otherwise, WriteTo returns nil.
   312  //
   313  // The debug parameter enables additional output.
   314  // Passing debug=0 writes the gzip-compressed protocol buffer described
   315  // in https://github.com/google/pprof/tree/master/proto#overview.
   316  // Passing debug=1 writes the legacy text format with comments
   317  // translating addresses to function names and line numbers, so that a
   318  // programmer can read the profile without tools.
   319  //
   320  // The predefined profiles may assign meaning to other debug values;
   321  // for example, when printing the "goroutine" profile, debug=2 means to
   322  // print the goroutine stacks in the same form that a Go program uses
   323  // when dying due to an unrecovered panic.
   324  func (p *Profile) WriteTo(w io.Writer, debug int) error {
   325  	if p.name == "" {
   326  		panic("pprof: use of zero Profile")
   327  	}
   328  	if p.write != nil {
   329  		return p.write(w, debug)
   330  	}
   331  
   332  	// Obtain consistent snapshot under lock; then process without lock.
   333  	p.mu.Lock()
   334  	all := make([][]uintptr, 0, len(p.m))
   335  	for _, stk := range p.m {
   336  		all = append(all, stk)
   337  	}
   338  	p.mu.Unlock()
   339  
   340  	// Map order is non-deterministic; make output deterministic.
   341  	sort.Slice(all, func(i, j int) bool {
   342  		t, u := all[i], all[j]
   343  		for k := 0; k < len(t) && k < len(u); k++ {
   344  			if t[k] != u[k] {
   345  				return t[k] < u[k]
   346  			}
   347  		}
   348  		return len(t) < len(u)
   349  	})
   350  
   351  	return printCountProfile(w, debug, p.name, stackProfile(all))
   352  }
   353  
   354  type stackProfile [][]uintptr
   355  
   356  func (x stackProfile) Len() int              { return len(x) }
   357  func (x stackProfile) Stack(i int) []uintptr { return x[i] }
   358  func (x stackProfile) Label(i int) *labelMap { return nil }
   359  
   360  // A countProfile is a set of stack traces to be printed as counts
   361  // grouped by stack trace. There are multiple implementations:
   362  // all that matters is that we can find out how many traces there are
   363  // and obtain each trace in turn.
   364  type countProfile interface {
   365  	Len() int
   366  	Stack(i int) []uintptr
   367  	Label(i int) *labelMap
   368  }
   369  
   370  // printCountCycleProfile outputs block profile records (for block or mutex profiles)
   371  // as the pprof-proto format output. Translations from cycle count to time duration
   372  // are done because The proto expects count and time (nanoseconds) instead of count
   373  // and the number of cycles for block, contention profiles.
   374  func printCountCycleProfile(w io.Writer, countName, cycleName string, records []runtime.BlockProfileRecord) error {
   375  	// Output profile in protobuf form.
   376  	b := newProfileBuilder(w)
   377  	b.pbValueType(tagProfile_PeriodType, countName, "count")
   378  	b.pb.int64Opt(tagProfile_Period, 1)
   379  	b.pbValueType(tagProfile_SampleType, countName, "count")
   380  	b.pbValueType(tagProfile_SampleType, cycleName, "nanoseconds")
   381  
   382  	cpuGHz := float64(runtime_cyclesPerSecond()) / 1e9
   383  
   384  	values := []int64{0, 0}
   385  	var locs []uint64
   386  	for _, r := range records {
   387  		values[0] = r.Count
   388  		values[1] = int64(float64(r.Cycles) / cpuGHz)
   389  		// For count profiles, all stack addresses are
   390  		// return PCs, which is what appendLocsForStack expects.
   391  		locs = b.appendLocsForStack(locs[:0], r.Stack())
   392  		b.pbSample(values, locs, nil)
   393  	}
   394  	b.build()
   395  	return nil
   396  }
   397  
   398  // printCountProfile prints a countProfile at the specified debug level.
   399  // The profile will be in compressed proto format unless debug is nonzero.
   400  func printCountProfile(w io.Writer, debug int, name string, p countProfile) error {
   401  	// Build count of each stack.
   402  	var buf strings.Builder
   403  	key := func(stk []uintptr, lbls *labelMap) string {
   404  		buf.Reset()
   405  		fmt.Fprintf(&buf, "@")
   406  		for _, pc := range stk {
   407  			fmt.Fprintf(&buf, " %#x", pc)
   408  		}
   409  		if lbls != nil {
   410  			buf.WriteString("\n# labels: ")
   411  			buf.WriteString(lbls.String())
   412  		}
   413  		return buf.String()
   414  	}
   415  	count := map[string]int{}
   416  	index := map[string]int{}
   417  	var keys []string
   418  	n := p.Len()
   419  	for i := 0; i < n; i++ {
   420  		k := key(p.Stack(i), p.Label(i))
   421  		if count[k] == 0 {
   422  			index[k] = i
   423  			keys = append(keys, k)
   424  		}
   425  		count[k]++
   426  	}
   427  
   428  	sort.Sort(&keysByCount{keys, count})
   429  
   430  	if debug > 0 {
   431  		// Print debug profile in legacy format
   432  		tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
   433  		fmt.Fprintf(tw, "%s profile: total %d\n", name, p.Len())
   434  		for _, k := range keys {
   435  			fmt.Fprintf(tw, "%d %s\n", count[k], k)
   436  			printStackRecord(tw, p.Stack(index[k]), false)
   437  		}
   438  		return tw.Flush()
   439  	}
   440  
   441  	// Output profile in protobuf form.
   442  	b := newProfileBuilder(w)
   443  	b.pbValueType(tagProfile_PeriodType, name, "count")
   444  	b.pb.int64Opt(tagProfile_Period, 1)
   445  	b.pbValueType(tagProfile_SampleType, name, "count")
   446  
   447  	values := []int64{0}
   448  	var locs []uint64
   449  	for _, k := range keys {
   450  		values[0] = int64(count[k])
   451  		// For count profiles, all stack addresses are
   452  		// return PCs, which is what appendLocsForStack expects.
   453  		locs = b.appendLocsForStack(locs[:0], p.Stack(index[k]))
   454  		idx := index[k]
   455  		var labels func()
   456  		if p.Label(idx) != nil {
   457  			labels = func() {
   458  				for k, v := range *p.Label(idx) {
   459  					b.pbLabel(tagSample_Label, k, v, 0)
   460  				}
   461  			}
   462  		}
   463  		b.pbSample(values, locs, labels)
   464  	}
   465  	b.build()
   466  	return nil
   467  }
   468  
   469  // keysByCount sorts keys with higher counts first, breaking ties by key string order.
   470  type keysByCount struct {
   471  	keys  []string
   472  	count map[string]int
   473  }
   474  
   475  func (x *keysByCount) Len() int      { return len(x.keys) }
   476  func (x *keysByCount) Swap(i, j int) { x.keys[i], x.keys[j] = x.keys[j], x.keys[i] }
   477  func (x *keysByCount) Less(i, j int) bool {
   478  	ki, kj := x.keys[i], x.keys[j]
   479  	ci, cj := x.count[ki], x.count[kj]
   480  	if ci != cj {
   481  		return ci > cj
   482  	}
   483  	return ki < kj
   484  }
   485  
   486  // printStackRecord prints the function + source line information
   487  // for a single stack trace.
   488  func printStackRecord(w io.Writer, stk []uintptr, allFrames bool) {
   489  	show := allFrames
   490  	frames := runtime.CallersFrames(stk)
   491  	for {
   492  		frame, more := frames.Next()
   493  		name := frame.Function
   494  		if name == "" {
   495  			show = true
   496  			fmt.Fprintf(w, "#\t%#x\n", frame.PC)
   497  		} else if name != "runtime.goexit" && (show || !strings.HasPrefix(name, "runtime.")) {
   498  			// Hide runtime.goexit and any runtime functions at the beginning.
   499  			// This is useful mainly for allocation traces.
   500  			show = true
   501  			fmt.Fprintf(w, "#\t%#x\t%s+%#x\t%s:%d\n", frame.PC, name, frame.PC-frame.Entry, frame.File, frame.Line)
   502  		}
   503  		if !more {
   504  			break
   505  		}
   506  	}
   507  	if !show {
   508  		// We didn't print anything; do it again,
   509  		// and this time include runtime functions.
   510  		printStackRecord(w, stk, true)
   511  		return
   512  	}
   513  	fmt.Fprintf(w, "\n")
   514  }
   515  
   516  // Interface to system profiles.
   517  
   518  // WriteHeapProfile is shorthand for Lookup("heap").WriteTo(w, 0).
   519  // It is preserved for backwards compatibility.
   520  func WriteHeapProfile(w io.Writer) error {
   521  	return writeHeap(w, 0)
   522  }
   523  
   524  // countHeap returns the number of records in the heap profile.
   525  func countHeap() int {
   526  	n, _ := runtime.MemProfile(nil, true)
   527  	return n
   528  }
   529  
   530  // writeHeap writes the current runtime heap profile to w.
   531  func writeHeap(w io.Writer, debug int) error {
   532  	return writeHeapInternal(w, debug, "")
   533  }
   534  
   535  // writeAlloc writes the current runtime heap profile to w
   536  // with the total allocation space as the default sample type.
   537  func writeAlloc(w io.Writer, debug int) error {
   538  	return writeHeapInternal(w, debug, "alloc_space")
   539  }
   540  
   541  func writeHeapInternal(w io.Writer, debug int, defaultSampleType string) error {
   542  	var memStats *runtime.MemStats
   543  	if debug != 0 {
   544  		// Read mem stats first, so that our other allocations
   545  		// do not appear in the statistics.
   546  		memStats = new(runtime.MemStats)
   547  		runtime.ReadMemStats(memStats)
   548  	}
   549  
   550  	// Find out how many records there are (MemProfile(nil, true)),
   551  	// allocate that many records, and get the data.
   552  	// There's a race—more records might be added between
   553  	// the two calls—so allocate a few extra records for safety
   554  	// and also try again if we're very unlucky.
   555  	// The loop should only execute one iteration in the common case.
   556  	var p []runtime.MemProfileRecord
   557  	n, ok := runtime.MemProfile(nil, true)
   558  	for {
   559  		// Allocate room for a slightly bigger profile,
   560  		// in case a few more entries have been added
   561  		// since the call to MemProfile.
   562  		p = make([]runtime.MemProfileRecord, n+50)
   563  		n, ok = runtime.MemProfile(p, true)
   564  		if ok {
   565  			p = p[0:n]
   566  			break
   567  		}
   568  		// Profile grew; try again.
   569  	}
   570  
   571  	if debug == 0 {
   572  		return writeHeapProto(w, p, int64(runtime.MemProfileRate), defaultSampleType)
   573  	}
   574  
   575  	sort.Slice(p, func(i, j int) bool { return p[i].InUseBytes() > p[j].InUseBytes() })
   576  
   577  	b := bufio.NewWriter(w)
   578  	tw := tabwriter.NewWriter(b, 1, 8, 1, '\t', 0)
   579  	w = tw
   580  
   581  	var total runtime.MemProfileRecord
   582  	for i := range p {
   583  		r := &p[i]
   584  		total.AllocBytes += r.AllocBytes
   585  		total.AllocObjects += r.AllocObjects
   586  		total.FreeBytes += r.FreeBytes
   587  		total.FreeObjects += r.FreeObjects
   588  	}
   589  
   590  	// Technically the rate is MemProfileRate not 2*MemProfileRate,
   591  	// but early versions of the C++ heap profiler reported 2*MemProfileRate,
   592  	// so that's what pprof has come to expect.
   593  	rate := 2 * runtime.MemProfileRate
   594  
   595  	// pprof reads a profile with alloc == inuse as being a "2-column" profile
   596  	// (objects and bytes, not distinguishing alloc from inuse),
   597  	// but then such a profile can't be merged using pprof *.prof with
   598  	// other 4-column profiles where alloc != inuse.
   599  	// The easiest way to avoid this bug is to adjust allocBytes so it's never == inuseBytes.
   600  	// pprof doesn't use these header values anymore except for checking equality.
   601  	inUseBytes := total.InUseBytes()
   602  	allocBytes := total.AllocBytes
   603  	if inUseBytes == allocBytes {
   604  		allocBytes++
   605  	}
   606  
   607  	fmt.Fprintf(w, "heap profile: %d: %d [%d: %d] @ heap/%d\n",
   608  		total.InUseObjects(), inUseBytes,
   609  		total.AllocObjects, allocBytes,
   610  		rate)
   611  
   612  	for i := range p {
   613  		r := &p[i]
   614  		fmt.Fprintf(w, "%d: %d [%d: %d] @",
   615  			r.InUseObjects(), r.InUseBytes(),
   616  			r.AllocObjects, r.AllocBytes)
   617  		for _, pc := range r.Stack() {
   618  			fmt.Fprintf(w, " %#x", pc)
   619  		}
   620  		fmt.Fprintf(w, "\n")
   621  		printStackRecord(w, r.Stack(), false)
   622  	}
   623  
   624  	// Print memstats information too.
   625  	// Pprof will ignore, but useful for people
   626  	s := memStats
   627  	fmt.Fprintf(w, "\n# runtime.MemStats\n")
   628  	fmt.Fprintf(w, "# Alloc = %d\n", s.Alloc)
   629  	fmt.Fprintf(w, "# TotalAlloc = %d\n", s.TotalAlloc)
   630  	fmt.Fprintf(w, "# Sys = %d\n", s.Sys)
   631  	fmt.Fprintf(w, "# Lookups = %d\n", s.Lookups)
   632  	fmt.Fprintf(w, "# Mallocs = %d\n", s.Mallocs)
   633  	fmt.Fprintf(w, "# Frees = %d\n", s.Frees)
   634  
   635  	fmt.Fprintf(w, "# HeapAlloc = %d\n", s.HeapAlloc)
   636  	fmt.Fprintf(w, "# HeapSys = %d\n", s.HeapSys)
   637  	fmt.Fprintf(w, "# HeapIdle = %d\n", s.HeapIdle)
   638  	fmt.Fprintf(w, "# HeapInuse = %d\n", s.HeapInuse)
   639  	fmt.Fprintf(w, "# HeapReleased = %d\n", s.HeapReleased)
   640  	fmt.Fprintf(w, "# HeapObjects = %d\n", s.HeapObjects)
   641  
   642  	fmt.Fprintf(w, "# Stack = %d / %d\n", s.StackInuse, s.StackSys)
   643  	fmt.Fprintf(w, "# MSpan = %d / %d\n", s.MSpanInuse, s.MSpanSys)
   644  	fmt.Fprintf(w, "# MCache = %d / %d\n", s.MCacheInuse, s.MCacheSys)
   645  	fmt.Fprintf(w, "# BuckHashSys = %d\n", s.BuckHashSys)
   646  	fmt.Fprintf(w, "# GCSys = %d\n", s.GCSys)
   647  	fmt.Fprintf(w, "# OtherSys = %d\n", s.OtherSys)
   648  
   649  	fmt.Fprintf(w, "# NextGC = %d\n", s.NextGC)
   650  	fmt.Fprintf(w, "# LastGC = %d\n", s.LastGC)
   651  	fmt.Fprintf(w, "# PauseNs = %d\n", s.PauseNs)
   652  	fmt.Fprintf(w, "# PauseEnd = %d\n", s.PauseEnd)
   653  	fmt.Fprintf(w, "# NumGC = %d\n", s.NumGC)
   654  	fmt.Fprintf(w, "# NumForcedGC = %d\n", s.NumForcedGC)
   655  	fmt.Fprintf(w, "# GCCPUFraction = %v\n", s.GCCPUFraction)
   656  	fmt.Fprintf(w, "# DebugGC = %v\n", s.DebugGC)
   657  
   658  	// Also flush out MaxRSS on supported platforms.
   659  	addMaxRSS(w)
   660  
   661  	tw.Flush()
   662  	return b.Flush()
   663  }
   664  
   665  // countThreadCreate returns the size of the current ThreadCreateProfile.
   666  func countThreadCreate() int {
   667  	n, _ := runtime.ThreadCreateProfile(nil)
   668  	return n
   669  }
   670  
   671  // writeThreadCreate writes the current runtime ThreadCreateProfile to w.
   672  func writeThreadCreate(w io.Writer, debug int) error {
   673  	// Until https://golang.org/issues/6104 is addressed, wrap
   674  	// ThreadCreateProfile because there's no point in tracking labels when we
   675  	// don't get any stack-traces.
   676  	return writeRuntimeProfile(w, debug, "threadcreate", func(p []runtime.StackRecord, _ []unsafe.Pointer) (n int, ok bool) {
   677  		return runtime.ThreadCreateProfile(p)
   678  	})
   679  }
   680  
   681  // countGoroutine returns the number of goroutines.
   682  func countGoroutine() int {
   683  	return runtime.NumGoroutine()
   684  }
   685  
   686  // runtime_goroutineProfileWithLabels is defined in runtime/mprof.go
   687  func runtime_goroutineProfileWithLabels(p []runtime.StackRecord, labels []unsafe.Pointer) (n int, ok bool)
   688  
   689  // writeGoroutine writes the current runtime GoroutineProfile to w.
   690  func writeGoroutine(w io.Writer, debug int) error {
   691  	if debug >= 2 {
   692  		return writeGoroutineStacks(w)
   693  	}
   694  	return writeRuntimeProfile(w, debug, "goroutine", runtime_goroutineProfileWithLabels)
   695  }
   696  
   697  func writeGoroutineStacks(w io.Writer) error {
   698  	// We don't know how big the buffer needs to be to collect
   699  	// all the goroutines. Start with 1 MB and try a few times, doubling each time.
   700  	// Give up and use a truncated trace if 64 MB is not enough.
   701  	buf := make([]byte, 1<<20)
   702  	for i := 0; ; i++ {
   703  		n := runtime.Stack(buf, true)
   704  		if n < len(buf) {
   705  			buf = buf[:n]
   706  			break
   707  		}
   708  		if len(buf) >= 64<<20 {
   709  			// Filled 64 MB - stop there.
   710  			break
   711  		}
   712  		buf = make([]byte, 2*len(buf))
   713  	}
   714  	_, err := w.Write(buf)
   715  	return err
   716  }
   717  
   718  func writeRuntimeProfile(w io.Writer, debug int, name string, fetch func([]runtime.StackRecord, []unsafe.Pointer) (int, bool)) error {
   719  	// Find out how many records there are (fetch(nil)),
   720  	// allocate that many records, and get the data.
   721  	// There's a race—more records might be added between
   722  	// the two calls—so allocate a few extra records for safety
   723  	// and also try again if we're very unlucky.
   724  	// The loop should only execute one iteration in the common case.
   725  	var p []runtime.StackRecord
   726  	var labels []unsafe.Pointer
   727  	n, ok := fetch(nil, nil)
   728  	for {
   729  		// Allocate room for a slightly bigger profile,
   730  		// in case a few more entries have been added
   731  		// since the call to ThreadProfile.
   732  		p = make([]runtime.StackRecord, n+10)
   733  		labels = make([]unsafe.Pointer, n+10)
   734  		n, ok = fetch(p, labels)
   735  		if ok {
   736  			p = p[0:n]
   737  			break
   738  		}
   739  		// Profile grew; try again.
   740  	}
   741  
   742  	return printCountProfile(w, debug, name, &runtimeProfile{p, labels})
   743  }
   744  
   745  type runtimeProfile struct {
   746  	stk    []runtime.StackRecord
   747  	labels []unsafe.Pointer
   748  }
   749  
   750  func (p *runtimeProfile) Len() int              { return len(p.stk) }
   751  func (p *runtimeProfile) Stack(i int) []uintptr { return p.stk[i].Stack() }
   752  func (p *runtimeProfile) Label(i int) *labelMap { return (*labelMap)(p.labels[i]) }
   753  
   754  var cpu struct {
   755  	sync.Mutex
   756  	profiling bool
   757  	done      chan bool
   758  }
   759  
   760  // StartCPUProfile enables CPU profiling for the current process.
   761  // While profiling, the profile will be buffered and written to w.
   762  // StartCPUProfile returns an error if profiling is already enabled.
   763  //
   764  // On Unix-like systems, StartCPUProfile does not work by default for
   765  // Go code built with -buildmode=c-archive or -buildmode=c-shared.
   766  // StartCPUProfile relies on the SIGPROF signal, but that signal will
   767  // be delivered to the main program's SIGPROF signal handler (if any)
   768  // not to the one used by Go. To make it work, call os/signal.Notify
   769  // for syscall.SIGPROF, but note that doing so may break any profiling
   770  // being done by the main program.
   771  func StartCPUProfile(w io.Writer) error {
   772  	// The runtime routines allow a variable profiling rate,
   773  	// but in practice operating systems cannot trigger signals
   774  	// at more than about 500 Hz, and our processing of the
   775  	// signal is not cheap (mostly getting the stack trace).
   776  	// 100 Hz is a reasonable choice: it is frequent enough to
   777  	// produce useful data, rare enough not to bog down the
   778  	// system, and a nice round number to make it easy to
   779  	// convert sample counts to seconds. Instead of requiring
   780  	// each client to specify the frequency, we hard code it.
   781  	const hz = 100
   782  
   783  	cpu.Lock()
   784  	defer cpu.Unlock()
   785  	if cpu.done == nil {
   786  		cpu.done = make(chan bool)
   787  	}
   788  	// Double-check.
   789  	if cpu.profiling {
   790  		return fmt.Errorf("cpu profiling already in use")
   791  	}
   792  	cpu.profiling = true
   793  	runtime.SetCPUProfileRate(hz)
   794  	go profileWriter(w)
   795  	return nil
   796  }
   797  
   798  // readProfile, provided by the runtime, returns the next chunk of
   799  // binary CPU profiling stack trace data, blocking until data is available.
   800  // If profiling is turned off and all the profile data accumulated while it was
   801  // on has been returned, readProfile returns eof=true.
   802  // The caller must save the returned data and tags before calling readProfile again.
   803  func readProfile() (data []uint64, tags []unsafe.Pointer, eof bool)
   804  
   805  func profileWriter(w io.Writer) {
   806  	b := newProfileBuilder(w)
   807  	var err error
   808  	for {
   809  		time.Sleep(100 * time.Millisecond)
   810  		data, tags, eof := readProfile()
   811  		if e := b.addCPUData(data, tags); e != nil && err == nil {
   812  			err = e
   813  		}
   814  		if eof {
   815  			break
   816  		}
   817  	}
   818  	if err != nil {
   819  		// The runtime should never produce an invalid or truncated profile.
   820  		// It drops records that can't fit into its log buffers.
   821  		panic("runtime/pprof: converting profile: " + err.Error())
   822  	}
   823  	b.build()
   824  	cpu.done <- true
   825  }
   826  
   827  // StopCPUProfile stops the current CPU profile, if any.
   828  // StopCPUProfile only returns after all the writes for the
   829  // profile have completed.
   830  func StopCPUProfile() {
   831  	cpu.Lock()
   832  	defer cpu.Unlock()
   833  
   834  	if !cpu.profiling {
   835  		return
   836  	}
   837  	cpu.profiling = false
   838  	runtime.SetCPUProfileRate(0)
   839  	<-cpu.done
   840  }
   841  
   842  // countBlock returns the number of records in the blocking profile.
   843  func countBlock() int {
   844  	n, _ := runtime.BlockProfile(nil)
   845  	return n
   846  }
   847  
   848  // countMutex returns the number of records in the mutex profile.
   849  func countMutex() int {
   850  	n, _ := runtime.MutexProfile(nil)
   851  	return n
   852  }
   853  
   854  // writeBlock writes the current blocking profile to w.
   855  func writeBlock(w io.Writer, debug int) error {
   856  	return writeProfileInternal(w, debug, "contention", runtime.BlockProfile)
   857  }
   858  
   859  // writeMutex writes the current mutex profile to w.
   860  func writeMutex(w io.Writer, debug int) error {
   861  	return writeProfileInternal(w, debug, "mutex", runtime.MutexProfile)
   862  }
   863  
   864  // writeProfileInternal writes the current blocking or mutex profile depending on the passed parameters.
   865  func writeProfileInternal(w io.Writer, debug int, name string, runtimeProfile func([]runtime.BlockProfileRecord) (int, bool)) error {
   866  	var p []runtime.BlockProfileRecord
   867  	n, ok := runtimeProfile(nil)
   868  	for {
   869  		p = make([]runtime.BlockProfileRecord, n+50)
   870  		n, ok = runtimeProfile(p)
   871  		if ok {
   872  			p = p[:n]
   873  			break
   874  		}
   875  	}
   876  
   877  	sort.Slice(p, func(i, j int) bool { return p[i].Cycles > p[j].Cycles })
   878  
   879  	if debug <= 0 {
   880  		return printCountCycleProfile(w, "contentions", "delay", p)
   881  	}
   882  
   883  	b := bufio.NewWriter(w)
   884  	tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
   885  	w = tw
   886  
   887  	fmt.Fprintf(w, "--- %v:\n", name)
   888  	fmt.Fprintf(w, "cycles/second=%v\n", runtime_cyclesPerSecond())
   889  	if name == "mutex" {
   890  		fmt.Fprintf(w, "sampling period=%d\n", runtime.SetMutexProfileFraction(-1))
   891  	}
   892  	for i := range p {
   893  		r := &p[i]
   894  		fmt.Fprintf(w, "%v %v @", r.Cycles, r.Count)
   895  		for _, pc := range r.Stack() {
   896  			fmt.Fprintf(w, " %#x", pc)
   897  		}
   898  		fmt.Fprint(w, "\n")
   899  		if debug > 0 {
   900  			printStackRecord(w, r.Stack(), true)
   901  		}
   902  	}
   903  
   904  	if tw != nil {
   905  		tw.Flush()
   906  	}
   907  	return b.Flush()
   908  }
   909  
   910  func runtime_cyclesPerSecond() int64
   911  

View as plain text