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  //	        // Lookup("allocs") creates a profile similar to go test -memprofile.
    48  //	        // Alternatively, use Lookup("heap") for a profile
    49  //	        // that has inuse_space as the default index.
    50  //	        if err := pprof.Lookup("allocs").WriteTo(f, 0); err != nil {
    51  //	            log.Fatal("could not write memory profile: ", err)
    52  //	        }
    53  //	    }
    54  //	}
    55  //
    56  // There is also a standard HTTP interface to profiling data. Adding
    57  // the following line will install handlers under the /debug/pprof/
    58  // URL to download live profiles:
    59  //
    60  //	import _ "net/http/pprof"
    61  //
    62  // See the net/http/pprof package for more details.
    63  //
    64  // Profiles can then be visualized with the pprof tool:
    65  //
    66  //	go tool pprof cpu.prof
    67  //
    68  // There are many commands available from the pprof command line.
    69  // Commonly used commands include "top", which prints a summary of the
    70  // top program hot-spots, and "web", which opens an interactive graph
    71  // of hot-spots and their call graphs. Use "help" for information on
    72  // all pprof commands.
    73  //
    74  // For more information about pprof, see
    75  // https://github.com/google/pprof/blob/main/doc/README.md.
    76  package pprof
    77  
    78  import (
    79  	"bufio"
    80  	"cmp"
    81  	"fmt"
    82  	"internal/abi"
    83  	"internal/profilerecord"
    84  	"io"
    85  	"runtime"
    86  	"slices"
    87  	"sort"
    88  	"strings"
    89  	"sync"
    90  	"text/tabwriter"
    91  	"time"
    92  	"unsafe"
    93  )
    94  
    95  // BUG(rsc): Profiles are only as good as the kernel support used to generate them.
    96  // See https://golang.org/issue/13841 for details about known problems.
    97  
    98  // A Profile is a collection of stack traces showing the call sequences
    99  // that led to instances of a particular event, such as allocation.
   100  // Packages can create and maintain their own profiles; the most common
   101  // use is for tracking resources that must be explicitly closed, such as files
   102  // or network connections.
   103  //
   104  // A Profile's methods can be called from multiple goroutines simultaneously.
   105  //
   106  // Each Profile has a unique name. A few profiles are predefined:
   107  //
   108  //	goroutine    - stack traces of all current goroutines
   109  //	heap         - a sampling of memory allocations of live objects
   110  //	allocs       - a sampling of all past memory allocations
   111  //	threadcreate - stack traces that led to the creation of new OS threads
   112  //	block        - stack traces that led to blocking on synchronization primitives
   113  //	mutex        - stack traces of holders of contended mutexes
   114  //
   115  // These predefined profiles maintain themselves and panic on an explicit
   116  // [Profile.Add] or [Profile.Remove] method call.
   117  //
   118  // The CPU profile is not available as a Profile. It has a special API,
   119  // the [StartCPUProfile] and [StopCPUProfile] functions, because it streams
   120  // output to a writer during profiling.
   121  //
   122  // # Heap profile
   123  //
   124  // The heap profile reports statistics as of the most recently completed
   125  // garbage collection; it elides more recent allocation to avoid skewing
   126  // the profile away from live data and toward garbage.
   127  // If there has been no garbage collection at all, the heap profile reports
   128  // all known allocations. This exception helps mainly in programs running
   129  // without garbage collection enabled, usually for debugging purposes.
   130  //
   131  // The heap profile tracks both the allocation sites for all live objects in
   132  // the application memory and for all objects allocated since the program start.
   133  // Pprof's -inuse_space, -inuse_objects, -alloc_space, and -alloc_objects
   134  // flags select which to display, defaulting to -inuse_space (live objects,
   135  // scaled by size).
   136  //
   137  // # Allocs profile
   138  //
   139  // The allocs profile is the same as the heap profile but changes the default
   140  // pprof display to -alloc_space, the total number of bytes allocated since
   141  // the program began (including garbage-collected bytes).
   142  //
   143  // # Block profile
   144  //
   145  // The block profile tracks time spent blocked on synchronization primitives,
   146  // such as [sync.Mutex], [sync.RWMutex], [sync.WaitGroup], [sync.Cond], and
   147  // channel send/receive/select.
   148  //
   149  // Stack traces correspond to the location that blocked (for example,
   150  // [sync.Mutex.Lock]).
   151  //
   152  // Sample values correspond to cumulative time spent blocked at that stack
   153  // trace, subject to time-based sampling specified by
   154  // [runtime.SetBlockProfileRate].
   155  //
   156  // # Mutex profile
   157  //
   158  // The mutex profile tracks contention on mutexes, such as [sync.Mutex],
   159  // [sync.RWMutex], and runtime-internal locks.
   160  //
   161  // Stack traces correspond to the end of the critical section causing
   162  // contention. For example, a lock held for a long time while other goroutines
   163  // are waiting to acquire the lock will report contention when the lock is
   164  // finally unlocked (that is, at [sync.Mutex.Unlock]).
   165  //
   166  // Sample values correspond to the approximate cumulative time other goroutines
   167  // spent blocked waiting for the lock, subject to event-based sampling
   168  // specified by [runtime.SetMutexProfileFraction]. For example, if a caller
   169  // holds a lock for 1s while 5 other goroutines are waiting for the entire
   170  // second to acquire the lock, its unlock call stack will report 5s of
   171  // contention.
   172  type Profile struct {
   173  	name  string
   174  	mu    sync.Mutex
   175  	m     map[any][]uintptr
   176  	count func() int
   177  	write func(io.Writer, int) error
   178  }
   179  
   180  // profiles records all registered profiles.
   181  var profiles struct {
   182  	mu sync.Mutex
   183  	m  map[string]*Profile
   184  }
   185  
   186  var goroutineProfile = &Profile{
   187  	name:  "goroutine",
   188  	count: countGoroutine,
   189  	write: writeGoroutine,
   190  }
   191  
   192  var threadcreateProfile = &Profile{
   193  	name:  "threadcreate",
   194  	count: countThreadCreate,
   195  	write: writeThreadCreate,
   196  }
   197  
   198  var heapProfile = &Profile{
   199  	name:  "heap",
   200  	count: countHeap,
   201  	write: writeHeap,
   202  }
   203  
   204  var allocsProfile = &Profile{
   205  	name:  "allocs",
   206  	count: countHeap, // identical to heap profile
   207  	write: writeAlloc,
   208  }
   209  
   210  var blockProfile = &Profile{
   211  	name:  "block",
   212  	count: countBlock,
   213  	write: writeBlock,
   214  }
   215  
   216  var mutexProfile = &Profile{
   217  	name:  "mutex",
   218  	count: countMutex,
   219  	write: writeMutex,
   220  }
   221  
   222  func lockProfiles() {
   223  	profiles.mu.Lock()
   224  	if profiles.m == nil {
   225  		// Initial built-in profiles.
   226  		profiles.m = map[string]*Profile{
   227  			"goroutine":    goroutineProfile,
   228  			"threadcreate": threadcreateProfile,
   229  			"heap":         heapProfile,
   230  			"allocs":       allocsProfile,
   231  			"block":        blockProfile,
   232  			"mutex":        mutexProfile,
   233  		}
   234  	}
   235  }
   236  
   237  func unlockProfiles() {
   238  	profiles.mu.Unlock()
   239  }
   240  
   241  // NewProfile creates a new profile with the given name.
   242  // If a profile with that name already exists, NewProfile panics.
   243  // The convention is to use a 'import/path.' prefix to create
   244  // separate name spaces for each package.
   245  // For compatibility with various tools that read pprof data,
   246  // profile names should not contain spaces.
   247  func NewProfile(name string) *Profile {
   248  	lockProfiles()
   249  	defer unlockProfiles()
   250  	if name == "" {
   251  		panic("pprof: NewProfile with empty name")
   252  	}
   253  	if profiles.m[name] != nil {
   254  		panic("pprof: NewProfile name already in use: " + name)
   255  	}
   256  	p := &Profile{
   257  		name: name,
   258  		m:    map[any][]uintptr{},
   259  	}
   260  	profiles.m[name] = p
   261  	return p
   262  }
   263  
   264  // Lookup returns the profile with the given name, or nil if no such profile exists.
   265  func Lookup(name string) *Profile {
   266  	lockProfiles()
   267  	defer unlockProfiles()
   268  	return profiles.m[name]
   269  }
   270  
   271  // Profiles returns a slice of all the known profiles, sorted by name.
   272  func Profiles() []*Profile {
   273  	lockProfiles()
   274  	defer unlockProfiles()
   275  
   276  	all := make([]*Profile, 0, len(profiles.m))
   277  	for _, p := range profiles.m {
   278  		all = append(all, p)
   279  	}
   280  
   281  	slices.SortFunc(all, func(a, b *Profile) int {
   282  		return strings.Compare(a.name, b.name)
   283  	})
   284  	return all
   285  }
   286  
   287  // Name returns this profile's name, which can be passed to [Lookup] to reobtain the profile.
   288  func (p *Profile) Name() string {
   289  	return p.name
   290  }
   291  
   292  // Count returns the number of execution stacks currently in the profile.
   293  func (p *Profile) Count() int {
   294  	p.mu.Lock()
   295  	defer p.mu.Unlock()
   296  	if p.count != nil {
   297  		return p.count()
   298  	}
   299  	return len(p.m)
   300  }
   301  
   302  // Add adds the current execution stack to the profile, associated with value.
   303  // Add stores value in an internal map, so value must be suitable for use as
   304  // a map key and will not be garbage collected until the corresponding
   305  // call to [Profile.Remove]. Add panics if the profile already contains a stack for value.
   306  //
   307  // The skip parameter has the same meaning as [runtime.Caller]'s skip
   308  // and controls where the stack trace begins. Passing skip=0 begins the
   309  // trace in the function calling Add. For example, given this
   310  // execution stack:
   311  //
   312  //	Add
   313  //	called from rpc.NewClient
   314  //	called from mypkg.Run
   315  //	called from main.main
   316  //
   317  // Passing skip=0 begins the stack trace at the call to Add inside rpc.NewClient.
   318  // Passing skip=1 begins the stack trace at the call to NewClient inside mypkg.Run.
   319  func (p *Profile) Add(value any, skip int) {
   320  	if p.name == "" {
   321  		panic("pprof: use of uninitialized Profile")
   322  	}
   323  	if p.write != nil {
   324  		panic("pprof: Add called on built-in Profile " + p.name)
   325  	}
   326  
   327  	stk := make([]uintptr, 32)
   328  	n := runtime.Callers(skip+1, stk[:])
   329  	stk = stk[:n]
   330  	if len(stk) == 0 {
   331  		// The value for skip is too large, and there's no stack trace to record.
   332  		stk = []uintptr{abi.FuncPCABIInternal(lostProfileEvent)}
   333  	}
   334  
   335  	p.mu.Lock()
   336  	defer p.mu.Unlock()
   337  	if p.m[value] != nil {
   338  		panic("pprof: Profile.Add of duplicate value")
   339  	}
   340  	p.m[value] = stk
   341  }
   342  
   343  // Remove removes the execution stack associated with value from the profile.
   344  // It is a no-op if the value is not in the profile.
   345  func (p *Profile) Remove(value any) {
   346  	p.mu.Lock()
   347  	defer p.mu.Unlock()
   348  	delete(p.m, value)
   349  }
   350  
   351  // WriteTo writes a pprof-formatted snapshot of the profile to w.
   352  // If a write to w returns an error, WriteTo returns that error.
   353  // Otherwise, WriteTo returns nil.
   354  //
   355  // The debug parameter enables additional output.
   356  // Passing debug=0 writes the gzip-compressed protocol buffer described
   357  // in https://github.com/google/pprof/tree/main/proto#overview.
   358  // Passing debug=1 writes the legacy text format with comments
   359  // translating addresses to function names and line numbers, so that a
   360  // programmer can read the profile without tools.
   361  //
   362  // The predefined profiles may assign meaning to other debug values;
   363  // for example, when printing the "goroutine" profile, debug=2 means to
   364  // print the goroutine stacks in the same form that a Go program uses
   365  // when dying due to an unrecovered panic.
   366  func (p *Profile) WriteTo(w io.Writer, debug int) error {
   367  	if p.name == "" {
   368  		panic("pprof: use of zero Profile")
   369  	}
   370  	if p.write != nil {
   371  		return p.write(w, debug)
   372  	}
   373  
   374  	// Obtain consistent snapshot under lock; then process without lock.
   375  	p.mu.Lock()
   376  	all := make([][]uintptr, 0, len(p.m))
   377  	for _, stk := range p.m {
   378  		all = append(all, stk)
   379  	}
   380  	p.mu.Unlock()
   381  
   382  	// Map order is non-deterministic; make output deterministic.
   383  	slices.SortFunc(all, slices.Compare)
   384  
   385  	return printCountProfile(w, debug, p.name, stackProfile(all))
   386  }
   387  
   388  type stackProfile [][]uintptr
   389  
   390  func (x stackProfile) Len() int              { return len(x) }
   391  func (x stackProfile) Stack(i int) []uintptr { return x[i] }
   392  func (x stackProfile) Label(i int) *labelMap { return nil }
   393  
   394  // A countProfile is a set of stack traces to be printed as counts
   395  // grouped by stack trace. There are multiple implementations:
   396  // all that matters is that we can find out how many traces there are
   397  // and obtain each trace in turn.
   398  type countProfile interface {
   399  	Len() int
   400  	Stack(i int) []uintptr
   401  	Label(i int) *labelMap
   402  }
   403  
   404  // expandInlinedFrames copies the call stack from pcs into dst, expanding any
   405  // PCs corresponding to inlined calls into the corresponding PCs for the inlined
   406  // functions. Returns the number of frames copied to dst.
   407  func expandInlinedFrames(dst, pcs []uintptr) int {
   408  	cf := runtime.CallersFrames(pcs)
   409  	var n int
   410  	for n < len(dst) {
   411  		f, more := cf.Next()
   412  		// f.PC is a "call PC", but later consumers will expect
   413  		// "return PCs"
   414  		dst[n] = f.PC + 1
   415  		n++
   416  		if !more {
   417  			break
   418  		}
   419  	}
   420  	return n
   421  }
   422  
   423  // printCountCycleProfile outputs block profile records (for block or mutex profiles)
   424  // as the pprof-proto format output. Translations from cycle count to time duration
   425  // are done because The proto expects count and time (nanoseconds) instead of count
   426  // and the number of cycles for block, contention profiles.
   427  func printCountCycleProfile(w io.Writer, countName, cycleName string, records []profilerecord.BlockProfileRecord) error {
   428  	// Output profile in protobuf form.
   429  	b := newProfileBuilder(w)
   430  	b.pbValueType(tagProfile_PeriodType, countName, "count")
   431  	b.pb.int64Opt(tagProfile_Period, 1)
   432  	b.pbValueType(tagProfile_SampleType, countName, "count")
   433  	b.pbValueType(tagProfile_SampleType, cycleName, "nanoseconds")
   434  
   435  	cpuGHz := float64(pprof_cyclesPerSecond()) / 1e9
   436  
   437  	values := []int64{0, 0}
   438  	var locs []uint64
   439  	expandedStack := pprof_makeProfStack()
   440  	for _, r := range records {
   441  		values[0] = r.Count
   442  		values[1] = int64(float64(r.Cycles) / cpuGHz)
   443  		// For count profiles, all stack addresses are
   444  		// return PCs, which is what appendLocsForStack expects.
   445  		n := expandInlinedFrames(expandedStack, r.Stack)
   446  		locs = b.appendLocsForStack(locs[:0], expandedStack[:n])
   447  		b.pbSample(values, locs, nil)
   448  	}
   449  	return b.build()
   450  }
   451  
   452  // printCountProfile prints a countProfile at the specified debug level.
   453  // The profile will be in compressed proto format unless debug is nonzero.
   454  func printCountProfile(w io.Writer, debug int, name string, p countProfile) error {
   455  	// Build count of each stack.
   456  	var buf strings.Builder
   457  	key := func(stk []uintptr, lbls *labelMap) string {
   458  		buf.Reset()
   459  		fmt.Fprintf(&buf, "@")
   460  		for _, pc := range stk {
   461  			fmt.Fprintf(&buf, " %#x", pc)
   462  		}
   463  		if lbls != nil {
   464  			buf.WriteString("\n# labels: ")
   465  			buf.WriteString(lbls.String())
   466  		}
   467  		return buf.String()
   468  	}
   469  	count := map[string]int{}
   470  	index := map[string]int{}
   471  	var keys []string
   472  	n := p.Len()
   473  	for i := 0; i < n; i++ {
   474  		k := key(p.Stack(i), p.Label(i))
   475  		if count[k] == 0 {
   476  			index[k] = i
   477  			keys = append(keys, k)
   478  		}
   479  		count[k]++
   480  	}
   481  
   482  	sort.Sort(&keysByCount{keys, count})
   483  
   484  	if debug > 0 {
   485  		// Print debug profile in legacy format
   486  		tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
   487  		fmt.Fprintf(tw, "%s profile: total %d\n", name, p.Len())
   488  		for _, k := range keys {
   489  			fmt.Fprintf(tw, "%d %s\n", count[k], k)
   490  			printStackRecord(tw, p.Stack(index[k]), false)
   491  		}
   492  		return tw.Flush()
   493  	}
   494  
   495  	// Output profile in protobuf form.
   496  	b := newProfileBuilder(w)
   497  	b.pbValueType(tagProfile_PeriodType, name, "count")
   498  	b.pb.int64Opt(tagProfile_Period, 1)
   499  	b.pbValueType(tagProfile_SampleType, name, "count")
   500  
   501  	values := []int64{0}
   502  	var locs []uint64
   503  	for _, k := range keys {
   504  		values[0] = int64(count[k])
   505  		// For count profiles, all stack addresses are
   506  		// return PCs, which is what appendLocsForStack expects.
   507  		locs = b.appendLocsForStack(locs[:0], p.Stack(index[k]))
   508  		idx := index[k]
   509  		var labels func()
   510  		if p.Label(idx) != nil {
   511  			labels = func() {
   512  				for _, lbl := range p.Label(idx).list {
   513  					b.pbLabel(tagSample_Label, lbl.key, lbl.value, 0)
   514  				}
   515  			}
   516  		}
   517  		b.pbSample(values, locs, labels)
   518  	}
   519  	return b.build()
   520  }
   521  
   522  // keysByCount sorts keys with higher counts first, breaking ties by key string order.
   523  type keysByCount struct {
   524  	keys  []string
   525  	count map[string]int
   526  }
   527  
   528  func (x *keysByCount) Len() int      { return len(x.keys) }
   529  func (x *keysByCount) Swap(i, j int) { x.keys[i], x.keys[j] = x.keys[j], x.keys[i] }
   530  func (x *keysByCount) Less(i, j int) bool {
   531  	ki, kj := x.keys[i], x.keys[j]
   532  	ci, cj := x.count[ki], x.count[kj]
   533  	if ci != cj {
   534  		return ci > cj
   535  	}
   536  	return ki < kj
   537  }
   538  
   539  // printStackRecord prints the function + source line information
   540  // for a single stack trace.
   541  func printStackRecord(w io.Writer, stk []uintptr, allFrames bool) {
   542  	show := allFrames
   543  	frames := runtime.CallersFrames(stk)
   544  	for {
   545  		frame, more := frames.Next()
   546  		name := frame.Function
   547  		if name == "" {
   548  			show = true
   549  			fmt.Fprintf(w, "#\t%#x\n", frame.PC)
   550  		} else if name != "runtime.goexit" && (show || !(strings.HasPrefix(name, "runtime.") || strings.HasPrefix(name, "internal/runtime/"))) {
   551  			// Hide runtime.goexit and any runtime functions at the beginning.
   552  			// This is useful mainly for allocation traces.
   553  			show = true
   554  			fmt.Fprintf(w, "#\t%#x\t%s+%#x\t%s:%d\n", frame.PC, name, frame.PC-frame.Entry, frame.File, frame.Line)
   555  		}
   556  		if !more {
   557  			break
   558  		}
   559  	}
   560  	if !show {
   561  		// We didn't print anything; do it again,
   562  		// and this time include runtime functions.
   563  		printStackRecord(w, stk, true)
   564  		return
   565  	}
   566  	fmt.Fprintf(w, "\n")
   567  }
   568  
   569  // Interface to system profiles.
   570  
   571  // WriteHeapProfile is shorthand for [Lookup]("heap").WriteTo(w, 0).
   572  // It is preserved for backwards compatibility.
   573  func WriteHeapProfile(w io.Writer) error {
   574  	return writeHeap(w, 0)
   575  }
   576  
   577  // countHeap returns the number of records in the heap profile.
   578  func countHeap() int {
   579  	n, _ := runtime.MemProfile(nil, true)
   580  	return n
   581  }
   582  
   583  // writeHeap writes the current runtime heap profile to w.
   584  func writeHeap(w io.Writer, debug int) error {
   585  	return writeHeapInternal(w, debug, "")
   586  }
   587  
   588  // writeAlloc writes the current runtime heap profile to w
   589  // with the total allocation space as the default sample type.
   590  func writeAlloc(w io.Writer, debug int) error {
   591  	return writeHeapInternal(w, debug, "alloc_space")
   592  }
   593  
   594  func writeHeapInternal(w io.Writer, debug int, defaultSampleType string) error {
   595  	var memStats *runtime.MemStats
   596  	if debug != 0 {
   597  		// Read mem stats first, so that our other allocations
   598  		// do not appear in the statistics.
   599  		memStats = new(runtime.MemStats)
   600  		runtime.ReadMemStats(memStats)
   601  	}
   602  
   603  	// Find out how many records there are (the call
   604  	// pprof_memProfileInternal(nil, true) below),
   605  	// allocate that many records, and get the data.
   606  	// There's a race—more records might be added between
   607  	// the two calls—so allocate a few extra records for safety
   608  	// and also try again if we're very unlucky.
   609  	// The loop should only execute one iteration in the common case.
   610  	var p []profilerecord.MemProfileRecord
   611  	n, ok := pprof_memProfileInternal(nil, true)
   612  	for {
   613  		// Allocate room for a slightly bigger profile,
   614  		// in case a few more entries have been added
   615  		// since the call to MemProfile.
   616  		p = make([]profilerecord.MemProfileRecord, n+50)
   617  		n, ok = pprof_memProfileInternal(p, true)
   618  		if ok {
   619  			p = p[0:n]
   620  			break
   621  		}
   622  		// Profile grew; try again.
   623  	}
   624  
   625  	if debug == 0 {
   626  		return writeHeapProto(w, p, int64(runtime.MemProfileRate), defaultSampleType)
   627  	}
   628  
   629  	slices.SortFunc(p, func(a, b profilerecord.MemProfileRecord) int {
   630  		return cmp.Compare(a.InUseBytes(), b.InUseBytes())
   631  	})
   632  
   633  	b := bufio.NewWriter(w)
   634  	tw := tabwriter.NewWriter(b, 1, 8, 1, '\t', 0)
   635  	w = tw
   636  
   637  	var total runtime.MemProfileRecord
   638  	for i := range p {
   639  		r := &p[i]
   640  		total.AllocBytes += r.AllocBytes
   641  		total.AllocObjects += r.AllocObjects
   642  		total.FreeBytes += r.FreeBytes
   643  		total.FreeObjects += r.FreeObjects
   644  	}
   645  
   646  	// Technically the rate is MemProfileRate not 2*MemProfileRate,
   647  	// but early versions of the C++ heap profiler reported 2*MemProfileRate,
   648  	// so that's what pprof has come to expect.
   649  	rate := 2 * runtime.MemProfileRate
   650  
   651  	// pprof reads a profile with alloc == inuse as being a "2-column" profile
   652  	// (objects and bytes, not distinguishing alloc from inuse),
   653  	// but then such a profile can't be merged using pprof *.prof with
   654  	// other 4-column profiles where alloc != inuse.
   655  	// The easiest way to avoid this bug is to adjust allocBytes so it's never == inuseBytes.
   656  	// pprof doesn't use these header values anymore except for checking equality.
   657  	inUseBytes := total.InUseBytes()
   658  	allocBytes := total.AllocBytes
   659  	if inUseBytes == allocBytes {
   660  		allocBytes++
   661  	}
   662  
   663  	fmt.Fprintf(w, "heap profile: %d: %d [%d: %d] @ heap/%d\n",
   664  		total.InUseObjects(), inUseBytes,
   665  		total.AllocObjects, allocBytes,
   666  		rate)
   667  
   668  	for i := range p {
   669  		r := &p[i]
   670  		fmt.Fprintf(w, "%d: %d [%d: %d] @",
   671  			r.InUseObjects(), r.InUseBytes(),
   672  			r.AllocObjects, r.AllocBytes)
   673  		for _, pc := range r.Stack {
   674  			fmt.Fprintf(w, " %#x", pc)
   675  		}
   676  		fmt.Fprintf(w, "\n")
   677  		printStackRecord(w, r.Stack, false)
   678  	}
   679  
   680  	// Print memstats information too.
   681  	// Pprof will ignore, but useful for people
   682  	s := memStats
   683  	fmt.Fprintf(w, "\n# runtime.MemStats\n")
   684  	fmt.Fprintf(w, "# Alloc = %d\n", s.Alloc)
   685  	fmt.Fprintf(w, "# TotalAlloc = %d\n", s.TotalAlloc)
   686  	fmt.Fprintf(w, "# Sys = %d\n", s.Sys)
   687  	fmt.Fprintf(w, "# Lookups = %d\n", s.Lookups)
   688  	fmt.Fprintf(w, "# Mallocs = %d\n", s.Mallocs)
   689  	fmt.Fprintf(w, "# Frees = %d\n", s.Frees)
   690  
   691  	fmt.Fprintf(w, "# HeapAlloc = %d\n", s.HeapAlloc)
   692  	fmt.Fprintf(w, "# HeapSys = %d\n", s.HeapSys)
   693  	fmt.Fprintf(w, "# HeapIdle = %d\n", s.HeapIdle)
   694  	fmt.Fprintf(w, "# HeapInuse = %d\n", s.HeapInuse)
   695  	fmt.Fprintf(w, "# HeapReleased = %d\n", s.HeapReleased)
   696  	fmt.Fprintf(w, "# HeapObjects = %d\n", s.HeapObjects)
   697  
   698  	fmt.Fprintf(w, "# Stack = %d / %d\n", s.StackInuse, s.StackSys)
   699  	fmt.Fprintf(w, "# MSpan = %d / %d\n", s.MSpanInuse, s.MSpanSys)
   700  	fmt.Fprintf(w, "# MCache = %d / %d\n", s.MCacheInuse, s.MCacheSys)
   701  	fmt.Fprintf(w, "# BuckHashSys = %d\n", s.BuckHashSys)
   702  	fmt.Fprintf(w, "# GCSys = %d\n", s.GCSys)
   703  	fmt.Fprintf(w, "# OtherSys = %d\n", s.OtherSys)
   704  
   705  	fmt.Fprintf(w, "# NextGC = %d\n", s.NextGC)
   706  	fmt.Fprintf(w, "# LastGC = %d\n", s.LastGC)
   707  	fmt.Fprintf(w, "# PauseNs = %d\n", s.PauseNs)
   708  	fmt.Fprintf(w, "# PauseEnd = %d\n", s.PauseEnd)
   709  	fmt.Fprintf(w, "# NumGC = %d\n", s.NumGC)
   710  	fmt.Fprintf(w, "# NumForcedGC = %d\n", s.NumForcedGC)
   711  	fmt.Fprintf(w, "# GCCPUFraction = %v\n", s.GCCPUFraction)
   712  	fmt.Fprintf(w, "# DebugGC = %v\n", s.DebugGC)
   713  
   714  	// Also flush out MaxRSS on supported platforms.
   715  	addMaxRSS(w)
   716  
   717  	tw.Flush()
   718  	return b.Flush()
   719  }
   720  
   721  // countThreadCreate returns the size of the current ThreadCreateProfile.
   722  func countThreadCreate() int {
   723  	n, _ := runtime.ThreadCreateProfile(nil)
   724  	return n
   725  }
   726  
   727  // writeThreadCreate writes the current runtime ThreadCreateProfile to w.
   728  func writeThreadCreate(w io.Writer, debug int) error {
   729  	// Until https://golang.org/issues/6104 is addressed, wrap
   730  	// ThreadCreateProfile because there's no point in tracking labels when we
   731  	// don't get any stack-traces.
   732  	return writeRuntimeProfile(w, debug, "threadcreate", func(p []profilerecord.StackRecord, _ []unsafe.Pointer) (n int, ok bool) {
   733  		return pprof_threadCreateInternal(p)
   734  	})
   735  }
   736  
   737  // countGoroutine returns the number of goroutines.
   738  func countGoroutine() int {
   739  	return runtime.NumGoroutine()
   740  }
   741  
   742  // writeGoroutine writes the current runtime GoroutineProfile to w.
   743  func writeGoroutine(w io.Writer, debug int) error {
   744  	if debug >= 2 {
   745  		return writeGoroutineStacks(w)
   746  	}
   747  	return writeRuntimeProfile(w, debug, "goroutine", pprof_goroutineProfileWithLabels)
   748  }
   749  
   750  func writeGoroutineStacks(w io.Writer) error {
   751  	// We don't know how big the buffer needs to be to collect
   752  	// all the goroutines. Start with 1 MB and try a few times, doubling each time.
   753  	// Give up and use a truncated trace if 64 MB is not enough.
   754  	buf := make([]byte, 1<<20)
   755  	for i := 0; ; i++ {
   756  		n := runtime.Stack(buf, true)
   757  		if n < len(buf) {
   758  			buf = buf[:n]
   759  			break
   760  		}
   761  		if len(buf) >= 64<<20 {
   762  			// Filled 64 MB - stop there.
   763  			break
   764  		}
   765  		buf = make([]byte, 2*len(buf))
   766  	}
   767  	_, err := w.Write(buf)
   768  	return err
   769  }
   770  
   771  func writeRuntimeProfile(w io.Writer, debug int, name string, fetch func([]profilerecord.StackRecord, []unsafe.Pointer) (int, bool)) error {
   772  	// Find out how many records there are (fetch(nil)),
   773  	// allocate that many records, and get the data.
   774  	// There's a race—more records might be added between
   775  	// the two calls—so allocate a few extra records for safety
   776  	// and also try again if we're very unlucky.
   777  	// The loop should only execute one iteration in the common case.
   778  	var p []profilerecord.StackRecord
   779  	var labels []unsafe.Pointer
   780  	n, ok := fetch(nil, nil)
   781  
   782  	for {
   783  		// Allocate room for a slightly bigger profile,
   784  		// in case a few more entries have been added
   785  		// since the call to ThreadProfile.
   786  		p = make([]profilerecord.StackRecord, n+10)
   787  		labels = make([]unsafe.Pointer, n+10)
   788  		n, ok = fetch(p, labels)
   789  		if ok {
   790  			p = p[0:n]
   791  			break
   792  		}
   793  		// Profile grew; try again.
   794  	}
   795  
   796  	return printCountProfile(w, debug, name, &runtimeProfile{p, labels})
   797  }
   798  
   799  type runtimeProfile struct {
   800  	stk    []profilerecord.StackRecord
   801  	labels []unsafe.Pointer
   802  }
   803  
   804  func (p *runtimeProfile) Len() int              { return len(p.stk) }
   805  func (p *runtimeProfile) Stack(i int) []uintptr { return p.stk[i].Stack }
   806  func (p *runtimeProfile) Label(i int) *labelMap { return (*labelMap)(p.labels[i]) }
   807  
   808  var cpu struct {
   809  	sync.Mutex
   810  	profiling bool
   811  	done      chan bool
   812  }
   813  
   814  // StartCPUProfile enables CPU profiling for the current process.
   815  // While profiling, the profile will be buffered and written to w.
   816  // StartCPUProfile returns an error if profiling is already enabled.
   817  //
   818  // On Unix-like systems, StartCPUProfile does not work by default for
   819  // Go code built with -buildmode=c-archive or -buildmode=c-shared.
   820  // StartCPUProfile relies on the SIGPROF signal, but that signal will
   821  // be delivered to the main program's SIGPROF signal handler (if any)
   822  // not to the one used by Go. To make it work, call [os/signal.Notify]
   823  // for [syscall.SIGPROF], but note that doing so may break any profiling
   824  // being done by the main program.
   825  func StartCPUProfile(w io.Writer) error {
   826  	// The runtime routines allow a variable profiling rate,
   827  	// but in practice operating systems cannot trigger signals
   828  	// at more than about 500 Hz, and our processing of the
   829  	// signal is not cheap (mostly getting the stack trace).
   830  	// 100 Hz is a reasonable choice: it is frequent enough to
   831  	// produce useful data, rare enough not to bog down the
   832  	// system, and a nice round number to make it easy to
   833  	// convert sample counts to seconds. Instead of requiring
   834  	// each client to specify the frequency, we hard code it.
   835  	const hz = 100
   836  
   837  	cpu.Lock()
   838  	defer cpu.Unlock()
   839  	if cpu.done == nil {
   840  		cpu.done = make(chan bool)
   841  	}
   842  	// Double-check.
   843  	if cpu.profiling {
   844  		return fmt.Errorf("cpu profiling already in use")
   845  	}
   846  	cpu.profiling = true
   847  	runtime.SetCPUProfileRate(hz)
   848  	go profileWriter(w)
   849  	return nil
   850  }
   851  
   852  // readProfile, provided by the runtime, returns the next chunk of
   853  // binary CPU profiling stack trace data, blocking until data is available.
   854  // If profiling is turned off and all the profile data accumulated while it was
   855  // on has been returned, readProfile returns eof=true.
   856  // The caller must save the returned data and tags before calling readProfile again.
   857  func readProfile() (data []uint64, tags []unsafe.Pointer, eof bool)
   858  
   859  func profileWriter(w io.Writer) {
   860  	b := newProfileBuilder(w)
   861  	var err error
   862  	for {
   863  		time.Sleep(100 * time.Millisecond)
   864  		data, tags, eof := readProfile()
   865  		if e := b.addCPUData(data, tags); e != nil && err == nil {
   866  			err = e
   867  		}
   868  		if eof {
   869  			break
   870  		}
   871  	}
   872  	if err != nil {
   873  		// The runtime should never produce an invalid or truncated profile.
   874  		// It drops records that can't fit into its log buffers.
   875  		panic("runtime/pprof: converting profile: " + err.Error())
   876  	}
   877  	b.build()
   878  	cpu.done <- true
   879  }
   880  
   881  // StopCPUProfile stops the current CPU profile, if any.
   882  // StopCPUProfile only returns after all the writes for the
   883  // profile have completed.
   884  func StopCPUProfile() {
   885  	cpu.Lock()
   886  	defer cpu.Unlock()
   887  
   888  	if !cpu.profiling {
   889  		return
   890  	}
   891  	cpu.profiling = false
   892  	runtime.SetCPUProfileRate(0)
   893  	<-cpu.done
   894  }
   895  
   896  // countBlock returns the number of records in the blocking profile.
   897  func countBlock() int {
   898  	n, _ := runtime.BlockProfile(nil)
   899  	return n
   900  }
   901  
   902  // countMutex returns the number of records in the mutex profile.
   903  func countMutex() int {
   904  	n, _ := runtime.MutexProfile(nil)
   905  	return n
   906  }
   907  
   908  // writeBlock writes the current blocking profile to w.
   909  func writeBlock(w io.Writer, debug int) error {
   910  	return writeProfileInternal(w, debug, "contention", pprof_blockProfileInternal)
   911  }
   912  
   913  // writeMutex writes the current mutex profile to w.
   914  func writeMutex(w io.Writer, debug int) error {
   915  	return writeProfileInternal(w, debug, "mutex", pprof_mutexProfileInternal)
   916  }
   917  
   918  // writeProfileInternal writes the current blocking or mutex profile depending on the passed parameters.
   919  func writeProfileInternal(w io.Writer, debug int, name string, runtimeProfile func([]profilerecord.BlockProfileRecord) (int, bool)) error {
   920  	var p []profilerecord.BlockProfileRecord
   921  	n, ok := runtimeProfile(nil)
   922  	for {
   923  		p = make([]profilerecord.BlockProfileRecord, n+50)
   924  		n, ok = runtimeProfile(p)
   925  		if ok {
   926  			p = p[:n]
   927  			break
   928  		}
   929  	}
   930  
   931  	slices.SortFunc(p, func(a, b profilerecord.BlockProfileRecord) int {
   932  		return cmp.Compare(b.Cycles, a.Cycles)
   933  	})
   934  
   935  	if debug <= 0 {
   936  		return printCountCycleProfile(w, "contentions", "delay", p)
   937  	}
   938  
   939  	b := bufio.NewWriter(w)
   940  	tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
   941  	w = tw
   942  
   943  	fmt.Fprintf(w, "--- %v:\n", name)
   944  	fmt.Fprintf(w, "cycles/second=%v\n", pprof_cyclesPerSecond())
   945  	if name == "mutex" {
   946  		fmt.Fprintf(w, "sampling period=%d\n", runtime.SetMutexProfileFraction(-1))
   947  	}
   948  	expandedStack := pprof_makeProfStack()
   949  	for i := range p {
   950  		r := &p[i]
   951  		fmt.Fprintf(w, "%v %v @", r.Cycles, r.Count)
   952  		n := expandInlinedFrames(expandedStack, r.Stack)
   953  		stack := expandedStack[:n]
   954  		for _, pc := range stack {
   955  			fmt.Fprintf(w, " %#x", pc)
   956  		}
   957  		fmt.Fprint(w, "\n")
   958  		if debug > 0 {
   959  			printStackRecord(w, stack, true)
   960  		}
   961  	}
   962  
   963  	if tw != nil {
   964  		tw.Flush()
   965  	}
   966  	return b.Flush()
   967  }
   968  
   969  //go:linkname pprof_goroutineProfileWithLabels runtime.pprof_goroutineProfileWithLabels
   970  func pprof_goroutineProfileWithLabels(p []profilerecord.StackRecord, labels []unsafe.Pointer) (n int, ok bool)
   971  
   972  //go:linkname pprof_cyclesPerSecond runtime/pprof.runtime_cyclesPerSecond
   973  func pprof_cyclesPerSecond() int64
   974  
   975  //go:linkname pprof_memProfileInternal runtime.pprof_memProfileInternal
   976  func pprof_memProfileInternal(p []profilerecord.MemProfileRecord, inuseZero bool) (n int, ok bool)
   977  
   978  //go:linkname pprof_blockProfileInternal runtime.pprof_blockProfileInternal
   979  func pprof_blockProfileInternal(p []profilerecord.BlockProfileRecord) (n int, ok bool)
   980  
   981  //go:linkname pprof_mutexProfileInternal runtime.pprof_mutexProfileInternal
   982  func pprof_mutexProfileInternal(p []profilerecord.BlockProfileRecord) (n int, ok bool)
   983  
   984  //go:linkname pprof_threadCreateInternal runtime.pprof_threadCreateInternal
   985  func pprof_threadCreateInternal(p []profilerecord.StackRecord) (n int, ok bool)
   986  
   987  //go:linkname pprof_fpunwindExpand runtime.pprof_fpunwindExpand
   988  func pprof_fpunwindExpand(dst, src []uintptr) int
   989  
   990  //go:linkname pprof_makeProfStack runtime.pprof_makeProfStack
   991  func pprof_makeProfStack() []uintptr
   992  

View as plain text