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/goexperiment"
    84  	"internal/profilerecord"
    85  	"io"
    86  	"runtime"
    87  	"slices"
    88  	"sort"
    89  	"strings"
    90  	"sync"
    91  	"text/tabwriter"
    92  	"time"
    93  	"unsafe"
    94  )
    95  
    96  // BUG(rsc): Profiles are only as good as the kernel support used to generate them.
    97  // See https://golang.org/issue/13841 for details about known problems.
    98  
    99  // A Profile is a collection of stack traces showing the call sequences
   100  // that led to instances of a particular event, such as allocation.
   101  // Packages can create and maintain their own profiles; the most common
   102  // use is for tracking resources that must be explicitly closed, such as files
   103  // or network connections.
   104  //
   105  // A Profile's methods can be called from multiple goroutines simultaneously.
   106  //
   107  // Each Profile has a unique name. A few profiles are predefined:
   108  //
   109  //	goroutine      - stack traces of all current goroutines
   110  //	goroutineleak  - stack traces of all leaked goroutines
   111  //	allocs         - a sampling of all past memory allocations
   112  //	heap           - a sampling of memory allocations of live objects
   113  //	threadcreate   - stack traces that led to the creation of new OS threads
   114  //	block          - stack traces that led to blocking on synchronization primitives
   115  //	mutex          - stack traces of holders of contended mutexes
   116  //
   117  // These predefined profiles maintain themselves and panic on an explicit
   118  // [Profile.Add] or [Profile.Remove] method call.
   119  //
   120  // The CPU profile is not available as a Profile. It has a special API,
   121  // the [StartCPUProfile] and [StopCPUProfile] functions, because it streams
   122  // output to a writer during profiling.
   123  //
   124  // # Heap profile
   125  //
   126  // The heap profile reports statistics as of the most recently completed
   127  // garbage collection; it elides more recent allocation to avoid skewing
   128  // the profile away from live data and toward garbage.
   129  // If there has been no garbage collection at all, the heap profile reports
   130  // all known allocations. This exception helps mainly in programs running
   131  // without garbage collection enabled, usually for debugging purposes.
   132  //
   133  // The heap profile tracks both the allocation sites for all live objects in
   134  // the application memory and for all objects allocated since the program start.
   135  // Pprof's -inuse_space, -inuse_objects, -alloc_space, and -alloc_objects
   136  // flags select which to display, defaulting to -inuse_space (live objects,
   137  // scaled by size).
   138  //
   139  // # Allocs profile
   140  //
   141  // The allocs profile is the same as the heap profile but changes the default
   142  // pprof display to -alloc_space, the total number of bytes allocated since
   143  // the program began (including garbage-collected bytes).
   144  //
   145  // # Block profile
   146  //
   147  // The block profile tracks time spent blocked on synchronization primitives,
   148  // such as [sync.Mutex], [sync.RWMutex], [sync.WaitGroup], [sync.Cond], and
   149  // channel send/receive/select.
   150  //
   151  // Stack traces correspond to the location that blocked (for example,
   152  // [sync.Mutex.Lock]).
   153  //
   154  // Sample values correspond to cumulative time spent blocked at that stack
   155  // trace, subject to time-based sampling specified by
   156  // [runtime.SetBlockProfileRate].
   157  //
   158  // # Mutex profile
   159  //
   160  // The mutex profile tracks contention on mutexes, such as [sync.Mutex],
   161  // [sync.RWMutex], and runtime-internal locks.
   162  //
   163  // Stack traces correspond to the end of the critical section causing
   164  // contention. For example, a lock held for a long time while other goroutines
   165  // are waiting to acquire the lock will report contention when the lock is
   166  // finally unlocked (that is, at [sync.Mutex.Unlock]).
   167  //
   168  // Sample values correspond to the approximate cumulative time other goroutines
   169  // spent blocked waiting for the lock, subject to event-based sampling
   170  // specified by [runtime.SetMutexProfileFraction]. For example, if a caller
   171  // holds a lock for 1s while 5 other goroutines are waiting for the entire
   172  // second to acquire the lock, its unlock call stack will report 5s of
   173  // contention.
   174  
   175  type Profile struct {
   176  	name  string
   177  	mu    sync.Mutex
   178  	m     map[any][]uintptr
   179  	count func() int
   180  	write func(io.Writer, int) error
   181  }
   182  
   183  // profiles records all registered profiles.
   184  var profiles struct {
   185  	mu sync.Mutex
   186  	m  map[string]*Profile
   187  }
   188  
   189  var goroutineProfile = &Profile{
   190  	name:  "goroutine",
   191  	count: countGoroutine,
   192  	write: writeGoroutine,
   193  }
   194  
   195  var goroutineLeakProfile = &Profile{
   196  	name:  "goroutineleak",
   197  	count: runtime_goroutineleakcount,
   198  	write: writeGoroutineLeak,
   199  }
   200  
   201  var threadcreateProfile = &Profile{
   202  	name:  "threadcreate",
   203  	count: countThreadCreate,
   204  	write: writeThreadCreate,
   205  }
   206  
   207  var heapProfile = &Profile{
   208  	name:  "heap",
   209  	count: countHeap,
   210  	write: writeHeap,
   211  }
   212  
   213  var allocsProfile = &Profile{
   214  	name:  "allocs",
   215  	count: countHeap, // identical to heap profile
   216  	write: writeAlloc,
   217  }
   218  
   219  var blockProfile = &Profile{
   220  	name:  "block",
   221  	count: countBlock,
   222  	write: writeBlock,
   223  }
   224  
   225  var mutexProfile = &Profile{
   226  	name:  "mutex",
   227  	count: countMutex,
   228  	write: writeMutex,
   229  }
   230  
   231  // goroutineLeakProfileLock ensures that the goroutine leak profile writer observes the
   232  // leaked goroutines discovered during the goroutine leak detection GC cycle
   233  // that was triggered by the profile request.
   234  // This prevents a race condition between the garbage collector and the profile writer
   235  // when multiple profile requests are issued concurrently: the status of leaked goroutines
   236  // is reset to _Gwaiting at the beginning of a leak detection cycle, which may lead the
   237  // profile writer of another concurrent request to produce an incomplete profile.
   238  //
   239  // Example trace:
   240  //
   241  //	G1                    | GC                          | G2
   242  //	----------------------+-----------------------------+---------------------
   243  //	Request profile       | .                           | .
   244  //	.                     | .                           | Request profile
   245  //	.                     | [G1] Resets leaked g status | .
   246  //	.                     | [G1] Leaks detected         | .
   247  //	.                     | <New cycle>                 | .
   248  //	.                     | [G2] Resets leaked g status | .
   249  //	Write profile         | .                           | .
   250  //	.                     | [G2] Leaks detected         | .
   251  //	.                     | .                           | Write profile
   252  //	----------------------+-----------------------------+---------------------
   253  //	Incomplete profile    |+++++++++++++++++++++++++++++| Complete profile
   254  var goroutineLeakProfileLock sync.Mutex
   255  
   256  func lockProfiles() {
   257  	profiles.mu.Lock()
   258  	if profiles.m == nil {
   259  		// Initial built-in profiles.
   260  		profiles.m = map[string]*Profile{
   261  			"goroutine":    goroutineProfile,
   262  			"threadcreate": threadcreateProfile,
   263  			"heap":         heapProfile,
   264  			"allocs":       allocsProfile,
   265  			"block":        blockProfile,
   266  			"mutex":        mutexProfile,
   267  		}
   268  		if goexperiment.GoroutineLeakProfile {
   269  			profiles.m["goroutineleak"] = goroutineLeakProfile
   270  		}
   271  	}
   272  }
   273  
   274  func unlockProfiles() {
   275  	profiles.mu.Unlock()
   276  }
   277  
   278  // NewProfile creates a new profile with the given name.
   279  // If a profile with that name already exists, NewProfile panics.
   280  // The convention is to use a 'import/path.' prefix to create
   281  // separate name spaces for each package.
   282  // For compatibility with various tools that read pprof data,
   283  // profile names should not contain spaces.
   284  func NewProfile(name string) *Profile {
   285  	lockProfiles()
   286  	defer unlockProfiles()
   287  	if name == "" {
   288  		panic("pprof: NewProfile with empty name")
   289  	}
   290  	if profiles.m[name] != nil {
   291  		panic("pprof: NewProfile name already in use: " + name)
   292  	}
   293  	p := &Profile{
   294  		name: name,
   295  		m:    map[any][]uintptr{},
   296  	}
   297  	profiles.m[name] = p
   298  	return p
   299  }
   300  
   301  // Lookup returns the profile with the given name, or nil if no such profile exists.
   302  func Lookup(name string) *Profile {
   303  	lockProfiles()
   304  	defer unlockProfiles()
   305  	return profiles.m[name]
   306  }
   307  
   308  // Profiles returns a slice of all the known profiles, sorted by name.
   309  func Profiles() []*Profile {
   310  	lockProfiles()
   311  	defer unlockProfiles()
   312  
   313  	all := make([]*Profile, 0, len(profiles.m))
   314  	for _, p := range profiles.m {
   315  
   316  		all = append(all, p)
   317  	}
   318  
   319  	slices.SortFunc(all, func(a, b *Profile) int {
   320  		return strings.Compare(a.name, b.name)
   321  	})
   322  	return all
   323  }
   324  
   325  // Name returns this profile's name, which can be passed to [Lookup] to reobtain the profile.
   326  func (p *Profile) Name() string {
   327  	return p.name
   328  }
   329  
   330  // Count returns the number of execution stacks currently in the profile.
   331  func (p *Profile) Count() int {
   332  	p.mu.Lock()
   333  	defer p.mu.Unlock()
   334  	if p.count != nil {
   335  		return p.count()
   336  	}
   337  	return len(p.m)
   338  }
   339  
   340  // Add adds the current execution stack to the profile, associated with value.
   341  // Add stores value in an internal map, so value must be suitable for use as
   342  // a map key and will not be garbage collected until the corresponding
   343  // call to [Profile.Remove]. Add panics if the profile already contains a stack for value.
   344  //
   345  // The skip parameter has the same meaning as [runtime.Caller]'s skip
   346  // and controls where the stack trace begins. Passing skip=0 begins the
   347  // trace in the function calling Add. For example, given this
   348  // execution stack:
   349  //
   350  //	Add
   351  //	called from rpc.NewClient
   352  //	called from mypkg.Run
   353  //	called from main.main
   354  //
   355  // Passing skip=0 begins the stack trace at the call to Add inside rpc.NewClient.
   356  // Passing skip=1 begins the stack trace at the call to NewClient inside mypkg.Run.
   357  func (p *Profile) Add(value any, skip int) {
   358  	if p.name == "" {
   359  		panic("pprof: use of uninitialized Profile")
   360  	}
   361  	if p.write != nil {
   362  		panic("pprof: Add called on built-in Profile " + p.name)
   363  	}
   364  
   365  	stk := make([]uintptr, 32)
   366  	n := runtime.Callers(skip+1, stk[:])
   367  	stk = stk[:n]
   368  	if len(stk) == 0 {
   369  		// The value for skip is too large, and there's no stack trace to record.
   370  		stk = []uintptr{abi.FuncPCABIInternal(lostProfileEvent)}
   371  	}
   372  
   373  	p.mu.Lock()
   374  	defer p.mu.Unlock()
   375  	if p.m[value] != nil {
   376  		panic("pprof: Profile.Add of duplicate value")
   377  	}
   378  	p.m[value] = stk
   379  }
   380  
   381  // Remove removes the execution stack associated with value from the profile.
   382  // It is a no-op if the value is not in the profile.
   383  func (p *Profile) Remove(value any) {
   384  	p.mu.Lock()
   385  	defer p.mu.Unlock()
   386  	delete(p.m, value)
   387  }
   388  
   389  // WriteTo writes a pprof-formatted snapshot of the profile to w.
   390  // If a write to w returns an error, WriteTo returns that error.
   391  // Otherwise, WriteTo returns nil.
   392  //
   393  // The debug parameter enables additional output.
   394  // Passing debug=0 writes the gzip-compressed protocol buffer described
   395  // in https://github.com/google/pprof/tree/main/proto#overview.
   396  // Passing debug=1 writes the legacy text format with comments
   397  // translating addresses to function names and line numbers, so that a
   398  // programmer can read the profile without tools.
   399  //
   400  // The predefined profiles may assign meaning to other debug values;
   401  // for example, when printing the "goroutine" profile, debug=2 means to
   402  // print the goroutine stacks in the same form that a Go program uses
   403  // when dying due to an unrecovered panic.
   404  func (p *Profile) WriteTo(w io.Writer, debug int) error {
   405  	if p.name == "" {
   406  		panic("pprof: use of zero Profile")
   407  	}
   408  	if p.write != nil {
   409  		return p.write(w, debug)
   410  	}
   411  
   412  	// Obtain consistent snapshot under lock; then process without lock.
   413  	p.mu.Lock()
   414  	all := make([][]uintptr, 0, len(p.m))
   415  	for _, stk := range p.m {
   416  		all = append(all, stk)
   417  	}
   418  	p.mu.Unlock()
   419  
   420  	// Map order is non-deterministic; make output deterministic.
   421  	slices.SortFunc(all, slices.Compare)
   422  
   423  	return printCountProfile(w, debug, p.name, stackProfile(all))
   424  }
   425  
   426  type stackProfile [][]uintptr
   427  
   428  func (x stackProfile) Len() int              { return len(x) }
   429  func (x stackProfile) Stack(i int) []uintptr { return x[i] }
   430  func (x stackProfile) Label(i int) *labelMap { return nil }
   431  
   432  // A countProfile is a set of stack traces to be printed as counts
   433  // grouped by stack trace. There are multiple implementations:
   434  // all that matters is that we can find out how many traces there are
   435  // and obtain each trace in turn.
   436  type countProfile interface {
   437  	Len() int
   438  	Stack(i int) []uintptr
   439  	Label(i int) *labelMap
   440  }
   441  
   442  // expandInlinedFrames copies the call stack from pcs into dst, expanding any
   443  // PCs corresponding to inlined calls into the corresponding PCs for the inlined
   444  // functions. Returns the number of frames copied to dst.
   445  func expandInlinedFrames(dst, pcs []uintptr) int {
   446  	cf := runtime.CallersFrames(pcs)
   447  	var n int
   448  	for n < len(dst) {
   449  		f, more := cf.Next()
   450  		// f.PC is a "call PC", but later consumers will expect
   451  		// "return PCs"
   452  		dst[n] = f.PC + 1
   453  		n++
   454  		if !more {
   455  			break
   456  		}
   457  	}
   458  	return n
   459  }
   460  
   461  // printCountCycleProfile outputs block profile records (for block or mutex profiles)
   462  // as the pprof-proto format output. Translations from cycle count to time duration
   463  // are done because The proto expects count and time (nanoseconds) instead of count
   464  // and the number of cycles for block, contention profiles.
   465  func printCountCycleProfile(w io.Writer, countName, cycleName string, records []profilerecord.BlockProfileRecord) error {
   466  	// Output profile in protobuf form.
   467  	b := newProfileBuilder(w)
   468  	b.pbValueType(tagProfile_PeriodType, countName, "count")
   469  	b.pb.int64Opt(tagProfile_Period, 1)
   470  	b.pbValueType(tagProfile_SampleType, countName, "count")
   471  	b.pbValueType(tagProfile_SampleType, cycleName, "nanoseconds")
   472  
   473  	cpuGHz := float64(pprof_cyclesPerSecond()) / 1e9
   474  
   475  	values := []int64{0, 0}
   476  	var locs []uint64
   477  	expandedStack := pprof_makeProfStack()
   478  	for _, r := range records {
   479  		values[0] = r.Count
   480  		values[1] = int64(float64(r.Cycles) / cpuGHz)
   481  		// For count profiles, all stack addresses are
   482  		// return PCs, which is what appendLocsForStack expects.
   483  		n := expandInlinedFrames(expandedStack, r.Stack)
   484  		locs = b.appendLocsForStack(locs[:0], expandedStack[:n])
   485  		b.pbSample(values, locs, nil)
   486  	}
   487  	return b.build()
   488  }
   489  
   490  // printCountProfile prints a countProfile at the specified debug level.
   491  // The profile will be in compressed proto format unless debug is nonzero.
   492  func printCountProfile(w io.Writer, debug int, name string, p countProfile) error {
   493  	// Build count of each stack.
   494  	var buf strings.Builder
   495  	key := func(stk []uintptr, lbls *labelMap) string {
   496  		buf.Reset()
   497  		fmt.Fprintf(&buf, "@")
   498  		for _, pc := range stk {
   499  			fmt.Fprintf(&buf, " %#x", pc)
   500  		}
   501  		if lbls != nil {
   502  			buf.WriteString("\n# labels: ")
   503  			buf.WriteString(lbls.String())
   504  		}
   505  		return buf.String()
   506  	}
   507  	count := map[string]int{}
   508  	index := map[string]int{}
   509  	var keys []string
   510  	n := p.Len()
   511  	for i := 0; i < n; i++ {
   512  		k := key(p.Stack(i), p.Label(i))
   513  		if count[k] == 0 {
   514  			index[k] = i
   515  			keys = append(keys, k)
   516  		}
   517  		count[k]++
   518  	}
   519  
   520  	sort.Sort(&keysByCount{keys, count})
   521  
   522  	if debug > 0 {
   523  		// Print debug profile in legacy format
   524  		tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
   525  		fmt.Fprintf(tw, "%s profile: total %d\n", name, p.Len())
   526  		for _, k := range keys {
   527  			fmt.Fprintf(tw, "%d %s\n", count[k], k)
   528  			printStackRecord(tw, p.Stack(index[k]), false)
   529  		}
   530  		return tw.Flush()
   531  	}
   532  
   533  	// Output profile in protobuf form.
   534  	b := newProfileBuilder(w)
   535  	b.pbValueType(tagProfile_PeriodType, name, "count")
   536  	b.pb.int64Opt(tagProfile_Period, 1)
   537  	b.pbValueType(tagProfile_SampleType, name, "count")
   538  
   539  	values := []int64{0}
   540  	var locs []uint64
   541  	for _, k := range keys {
   542  		values[0] = int64(count[k])
   543  		// For count profiles, all stack addresses are
   544  		// return PCs, which is what appendLocsForStack expects.
   545  		locs = b.appendLocsForStack(locs[:0], p.Stack(index[k]))
   546  		idx := index[k]
   547  		var labels func()
   548  		if p.Label(idx) != nil {
   549  			labels = func() {
   550  				for _, lbl := range p.Label(idx).Set.List {
   551  					b.pbLabel(tagSample_Label, lbl.Key, lbl.Value, 0)
   552  				}
   553  			}
   554  		}
   555  		b.pbSample(values, locs, labels)
   556  	}
   557  	return b.build()
   558  }
   559  
   560  // keysByCount sorts keys with higher counts first, breaking ties by key string order.
   561  type keysByCount struct {
   562  	keys  []string
   563  	count map[string]int
   564  }
   565  
   566  func (x *keysByCount) Len() int      { return len(x.keys) }
   567  func (x *keysByCount) Swap(i, j int) { x.keys[i], x.keys[j] = x.keys[j], x.keys[i] }
   568  func (x *keysByCount) Less(i, j int) bool {
   569  	ki, kj := x.keys[i], x.keys[j]
   570  	ci, cj := x.count[ki], x.count[kj]
   571  	if ci != cj {
   572  		return ci > cj
   573  	}
   574  	return ki < kj
   575  }
   576  
   577  // printStackRecord prints the function + source line information
   578  // for a single stack trace.
   579  func printStackRecord(w io.Writer, stk []uintptr, allFrames bool) {
   580  	show := allFrames
   581  	frames := runtime.CallersFrames(stk)
   582  	for {
   583  		frame, more := frames.Next()
   584  		name := frame.Function
   585  		if name == "" {
   586  			show = true
   587  			fmt.Fprintf(w, "#\t%#x\n", frame.PC)
   588  		} else if name != "runtime.goexit" && (show || !(strings.HasPrefix(name, "runtime.") || strings.HasPrefix(name, "internal/runtime/"))) {
   589  			// Hide runtime.goexit and any runtime functions at the beginning.
   590  			// This is useful mainly for allocation traces.
   591  			show = true
   592  			fmt.Fprintf(w, "#\t%#x\t%s+%#x\t%s:%d\n", frame.PC, name, frame.PC-frame.Entry, frame.File, frame.Line)
   593  		}
   594  		if !more {
   595  			break
   596  		}
   597  	}
   598  	if !show {
   599  		// We didn't print anything; do it again,
   600  		// and this time include runtime functions.
   601  		printStackRecord(w, stk, true)
   602  		return
   603  	}
   604  	fmt.Fprintf(w, "\n")
   605  }
   606  
   607  // Interface to system profiles.
   608  
   609  // WriteHeapProfile is shorthand for [Lookup]("heap").WriteTo(w, 0).
   610  // It is preserved for backwards compatibility.
   611  func WriteHeapProfile(w io.Writer) error {
   612  	return writeHeap(w, 0)
   613  }
   614  
   615  // countHeap returns the number of records in the heap profile.
   616  func countHeap() int {
   617  	n, _ := runtime.MemProfile(nil, true)
   618  	return n
   619  }
   620  
   621  // writeHeap writes the current runtime heap profile to w.
   622  func writeHeap(w io.Writer, debug int) error {
   623  	return writeHeapInternal(w, debug, "")
   624  }
   625  
   626  // writeAlloc writes the current runtime heap profile to w
   627  // with the total allocation space as the default sample type.
   628  func writeAlloc(w io.Writer, debug int) error {
   629  	return writeHeapInternal(w, debug, "alloc_space")
   630  }
   631  
   632  func writeHeapInternal(w io.Writer, debug int, defaultSampleType string) error {
   633  	var memStats *runtime.MemStats
   634  	if debug != 0 {
   635  		// Read mem stats first, so that our other allocations
   636  		// do not appear in the statistics.
   637  		memStats = new(runtime.MemStats)
   638  		runtime.ReadMemStats(memStats)
   639  	}
   640  
   641  	// Find out how many records there are (the call
   642  	// pprof_memProfileInternal(nil, true) below),
   643  	// allocate that many records, and get the data.
   644  	// There's a race—more records might be added between
   645  	// the two calls—so allocate a few extra records for safety
   646  	// and also try again if we're very unlucky.
   647  	// The loop should only execute one iteration in the common case.
   648  	var p []profilerecord.MemProfileRecord
   649  	n, ok := pprof_memProfileInternal(nil, true)
   650  	for {
   651  		// Allocate room for a slightly bigger profile,
   652  		// in case a few more entries have been added
   653  		// since the call to MemProfile.
   654  		p = make([]profilerecord.MemProfileRecord, n+50)
   655  		n, ok = pprof_memProfileInternal(p, true)
   656  		if ok {
   657  			p = p[0:n]
   658  			break
   659  		}
   660  		// Profile grew; try again.
   661  	}
   662  
   663  	if debug == 0 {
   664  		return writeHeapProto(w, p, int64(runtime.MemProfileRate), defaultSampleType)
   665  	}
   666  
   667  	slices.SortFunc(p, func(a, b profilerecord.MemProfileRecord) int {
   668  		return cmp.Compare(a.InUseBytes(), b.InUseBytes())
   669  	})
   670  
   671  	b := bufio.NewWriter(w)
   672  	tw := tabwriter.NewWriter(b, 1, 8, 1, '\t', 0)
   673  	w = tw
   674  
   675  	var total runtime.MemProfileRecord
   676  	for i := range p {
   677  		r := &p[i]
   678  		total.AllocBytes += r.AllocBytes
   679  		total.AllocObjects += r.AllocObjects
   680  		total.FreeBytes += r.FreeBytes
   681  		total.FreeObjects += r.FreeObjects
   682  	}
   683  
   684  	// Technically the rate is MemProfileRate not 2*MemProfileRate,
   685  	// but early versions of the C++ heap profiler reported 2*MemProfileRate,
   686  	// so that's what pprof has come to expect.
   687  	rate := 2 * runtime.MemProfileRate
   688  
   689  	// pprof reads a profile with alloc == inuse as being a "2-column" profile
   690  	// (objects and bytes, not distinguishing alloc from inuse),
   691  	// but then such a profile can't be merged using pprof *.prof with
   692  	// other 4-column profiles where alloc != inuse.
   693  	// The easiest way to avoid this bug is to adjust allocBytes so it's never == inuseBytes.
   694  	// pprof doesn't use these header values anymore except for checking equality.
   695  	inUseBytes := total.InUseBytes()
   696  	allocBytes := total.AllocBytes
   697  	if inUseBytes == allocBytes {
   698  		allocBytes++
   699  	}
   700  
   701  	fmt.Fprintf(w, "heap profile: %d: %d [%d: %d] @ heap/%d\n",
   702  		total.InUseObjects(), inUseBytes,
   703  		total.AllocObjects, allocBytes,
   704  		rate)
   705  
   706  	for i := range p {
   707  		r := &p[i]
   708  		fmt.Fprintf(w, "%d: %d [%d: %d] @",
   709  			r.InUseObjects(), r.InUseBytes(),
   710  			r.AllocObjects, r.AllocBytes)
   711  		for _, pc := range r.Stack {
   712  			fmt.Fprintf(w, " %#x", pc)
   713  		}
   714  		fmt.Fprintf(w, "\n")
   715  		printStackRecord(w, r.Stack, false)
   716  	}
   717  
   718  	// Print memstats information too.
   719  	// Pprof will ignore, but useful for people
   720  	s := memStats
   721  	fmt.Fprintf(w, "\n# runtime.MemStats\n")
   722  	fmt.Fprintf(w, "# Alloc = %d\n", s.Alloc)
   723  	fmt.Fprintf(w, "# TotalAlloc = %d\n", s.TotalAlloc)
   724  	fmt.Fprintf(w, "# Sys = %d\n", s.Sys)
   725  	fmt.Fprintf(w, "# Lookups = %d\n", s.Lookups)
   726  	fmt.Fprintf(w, "# Mallocs = %d\n", s.Mallocs)
   727  	fmt.Fprintf(w, "# Frees = %d\n", s.Frees)
   728  
   729  	fmt.Fprintf(w, "# HeapAlloc = %d\n", s.HeapAlloc)
   730  	fmt.Fprintf(w, "# HeapSys = %d\n", s.HeapSys)
   731  	fmt.Fprintf(w, "# HeapIdle = %d\n", s.HeapIdle)
   732  	fmt.Fprintf(w, "# HeapInuse = %d\n", s.HeapInuse)
   733  	fmt.Fprintf(w, "# HeapReleased = %d\n", s.HeapReleased)
   734  	fmt.Fprintf(w, "# HeapObjects = %d\n", s.HeapObjects)
   735  
   736  	fmt.Fprintf(w, "# Stack = %d / %d\n", s.StackInuse, s.StackSys)
   737  	fmt.Fprintf(w, "# MSpan = %d / %d\n", s.MSpanInuse, s.MSpanSys)
   738  	fmt.Fprintf(w, "# MCache = %d / %d\n", s.MCacheInuse, s.MCacheSys)
   739  	fmt.Fprintf(w, "# BuckHashSys = %d\n", s.BuckHashSys)
   740  	fmt.Fprintf(w, "# GCSys = %d\n", s.GCSys)
   741  	fmt.Fprintf(w, "# OtherSys = %d\n", s.OtherSys)
   742  
   743  	fmt.Fprintf(w, "# NextGC = %d\n", s.NextGC)
   744  	fmt.Fprintf(w, "# LastGC = %d\n", s.LastGC)
   745  	fmt.Fprintf(w, "# PauseNs = %d\n", s.PauseNs)
   746  	fmt.Fprintf(w, "# PauseEnd = %d\n", s.PauseEnd)
   747  	fmt.Fprintf(w, "# NumGC = %d\n", s.NumGC)
   748  	fmt.Fprintf(w, "# NumForcedGC = %d\n", s.NumForcedGC)
   749  	fmt.Fprintf(w, "# GCCPUFraction = %v\n", s.GCCPUFraction)
   750  	fmt.Fprintf(w, "# DebugGC = %v\n", s.DebugGC)
   751  
   752  	// Also flush out MaxRSS on supported platforms.
   753  	addMaxRSS(w)
   754  
   755  	tw.Flush()
   756  	return b.Flush()
   757  }
   758  
   759  // countThreadCreate returns the size of the current ThreadCreateProfile.
   760  func countThreadCreate() int {
   761  	n, _ := runtime.ThreadCreateProfile(nil)
   762  	return n
   763  }
   764  
   765  // writeThreadCreate writes the current runtime ThreadCreateProfile to w.
   766  func writeThreadCreate(w io.Writer, debug int) error {
   767  	// Until https://golang.org/issues/6104 is addressed, wrap
   768  	// ThreadCreateProfile because there's no point in tracking labels when we
   769  	// don't get any stack-traces.
   770  	return writeRuntimeProfile(w, debug, "threadcreate", func(p []profilerecord.StackRecord, _ []unsafe.Pointer) (n int, ok bool) {
   771  		return pprof_threadCreateInternal(p)
   772  	})
   773  }
   774  
   775  // countGoroutine returns the number of goroutines.
   776  func countGoroutine() int {
   777  	return runtime.NumGoroutine()
   778  }
   779  
   780  // writeGoroutine writes the current runtime GoroutineProfile to w.
   781  func writeGoroutine(w io.Writer, debug int) error {
   782  	if debug >= 2 {
   783  		return writeGoroutineStacks(w)
   784  	}
   785  	return writeRuntimeProfile(w, debug, "goroutine", pprof_goroutineProfileWithLabels)
   786  }
   787  
   788  // writeGoroutineLeak first invokes a GC cycle that performs goroutine leak detection.
   789  // It then writes the goroutine profile, filtering for leaked goroutines.
   790  func writeGoroutineLeak(w io.Writer, debug int) error {
   791  	// Acquire the goroutine leak detection lock and release
   792  	// it after the goroutine leak profile is written.
   793  	//
   794  	// While the critical section is long, this is needed to prevent
   795  	// a race condition between the garbage collector and the goroutine
   796  	// leak profile writer when multiple profile requests are issued concurrently.
   797  	goroutineLeakProfileLock.Lock()
   798  	defer goroutineLeakProfileLock.Unlock()
   799  
   800  	// Run the GC with leak detection first so that leaked goroutines
   801  	// may transition to the leaked state.
   802  	runtime_goroutineLeakGC()
   803  
   804  	// If the debug flag is set sufficiently high, just defer to writing goroutine stacks
   805  	// like in a regular goroutine profile. Include non-leaked goroutines, too.
   806  	if debug >= 2 {
   807  		return writeGoroutineStacks(w)
   808  	}
   809  
   810  	// Otherwise, write the goroutine leak profile.
   811  	return writeRuntimeProfile(w, debug, "goroutineleak", pprof_goroutineLeakProfileWithLabels)
   812  }
   813  
   814  func writeGoroutineStacks(w io.Writer) error {
   815  	// We don't know how big the buffer needs to be to collect
   816  	// all the goroutines. Start with 1 MB and try a few times, doubling each time.
   817  	// Give up and use a truncated trace if 64 MB is not enough.
   818  	buf := make([]byte, 1<<20)
   819  	for i := 0; ; i++ {
   820  		n := runtime.Stack(buf, true)
   821  		if n < len(buf) {
   822  			buf = buf[:n]
   823  			break
   824  		}
   825  		if len(buf) >= 64<<20 {
   826  			// Filled 64 MB - stop there.
   827  			break
   828  		}
   829  		buf = make([]byte, 2*len(buf))
   830  	}
   831  	_, err := w.Write(buf)
   832  	return err
   833  }
   834  
   835  func writeRuntimeProfile(w io.Writer, debug int, name string, fetch func([]profilerecord.StackRecord, []unsafe.Pointer) (int, bool)) error {
   836  	// Find out how many records there are (fetch(nil)),
   837  	// allocate that many records, and get the data.
   838  	// There's a race—more records might be added between
   839  	// the two calls—so allocate a few extra records for safety
   840  	// and also try again if we're very unlucky.
   841  	// The loop should only execute one iteration in the common case.
   842  	var p []profilerecord.StackRecord
   843  	var labels []unsafe.Pointer
   844  	n, ok := fetch(nil, nil)
   845  
   846  	for {
   847  		// Allocate room for a slightly bigger profile,
   848  		// in case a few more entries have been added
   849  		// since the call to ThreadProfile.
   850  		p = make([]profilerecord.StackRecord, n+10)
   851  		labels = make([]unsafe.Pointer, n+10)
   852  		n, ok = fetch(p, labels)
   853  		if ok {
   854  			p = p[0:n]
   855  			break
   856  		}
   857  		// Profile grew; try again.
   858  	}
   859  
   860  	return printCountProfile(w, debug, name, &runtimeProfile{p, labels})
   861  }
   862  
   863  type runtimeProfile struct {
   864  	stk    []profilerecord.StackRecord
   865  	labels []unsafe.Pointer
   866  }
   867  
   868  func (p *runtimeProfile) Len() int              { return len(p.stk) }
   869  func (p *runtimeProfile) Stack(i int) []uintptr { return p.stk[i].Stack }
   870  func (p *runtimeProfile) Label(i int) *labelMap { return (*labelMap)(p.labels[i]) }
   871  
   872  var cpu struct {
   873  	sync.Mutex
   874  	profiling bool
   875  	done      chan bool
   876  }
   877  
   878  // StartCPUProfile enables CPU profiling for the current process.
   879  // While profiling, the profile will be buffered and written to w.
   880  // StartCPUProfile returns an error if profiling is already enabled.
   881  //
   882  // On Unix-like systems, StartCPUProfile does not work by default for
   883  // Go code built with -buildmode=c-archive or -buildmode=c-shared.
   884  // StartCPUProfile relies on the SIGPROF signal, but that signal will
   885  // be delivered to the main program's SIGPROF signal handler (if any)
   886  // not to the one used by Go. To make it work, call [os/signal.Notify]
   887  // for [syscall.SIGPROF], but note that doing so may break any profiling
   888  // being done by the main program.
   889  func StartCPUProfile(w io.Writer) error {
   890  	// The runtime routines allow a variable profiling rate,
   891  	// but in practice operating systems cannot trigger signals
   892  	// at more than about 500 Hz, and our processing of the
   893  	// signal is not cheap (mostly getting the stack trace).
   894  	// 100 Hz is a reasonable choice: it is frequent enough to
   895  	// produce useful data, rare enough not to bog down the
   896  	// system, and a nice round number to make it easy to
   897  	// convert sample counts to seconds. Instead of requiring
   898  	// each client to specify the frequency, we hard code it.
   899  	const hz = 100
   900  
   901  	cpu.Lock()
   902  	defer cpu.Unlock()
   903  	if cpu.done == nil {
   904  		cpu.done = make(chan bool)
   905  	}
   906  	// Double-check.
   907  	if cpu.profiling {
   908  		return fmt.Errorf("cpu profiling already in use")
   909  	}
   910  	cpu.profiling = true
   911  	runtime.SetCPUProfileRate(hz)
   912  	go profileWriter(w)
   913  	return nil
   914  }
   915  
   916  // readProfile, provided by the runtime, returns the next chunk of
   917  // binary CPU profiling stack trace data, blocking until data is available.
   918  // If profiling is turned off and all the profile data accumulated while it was
   919  // on has been returned, readProfile returns eof=true.
   920  // The caller must save the returned data and tags before calling readProfile again.
   921  func readProfile() (data []uint64, tags []unsafe.Pointer, eof bool)
   922  
   923  func profileWriter(w io.Writer) {
   924  	b := newProfileBuilder(w)
   925  	var err error
   926  	for {
   927  		if runtime.GOOS == "darwin" || runtime.GOOS == "ios" {
   928  			// see runtime_pprof_readProfile
   929  			time.Sleep(100 * time.Millisecond)
   930  		}
   931  		data, tags, eof := readProfile()
   932  		if e := b.addCPUData(data, tags); e != nil && err == nil {
   933  			err = e
   934  		}
   935  		if eof {
   936  			break
   937  		}
   938  	}
   939  	if err != nil {
   940  		// The runtime should never produce an invalid or truncated profile.
   941  		// It drops records that can't fit into its log buffers.
   942  		panic("runtime/pprof: converting profile: " + err.Error())
   943  	}
   944  	b.build()
   945  	cpu.done <- true
   946  }
   947  
   948  // StopCPUProfile stops the current CPU profile, if any.
   949  // StopCPUProfile only returns after all the writes for the
   950  // profile have completed.
   951  func StopCPUProfile() {
   952  	cpu.Lock()
   953  	defer cpu.Unlock()
   954  
   955  	if !cpu.profiling {
   956  		return
   957  	}
   958  	cpu.profiling = false
   959  	runtime.SetCPUProfileRate(0)
   960  	<-cpu.done
   961  }
   962  
   963  // countBlock returns the number of records in the blocking profile.
   964  func countBlock() int {
   965  	n, _ := runtime.BlockProfile(nil)
   966  	return n
   967  }
   968  
   969  // countMutex returns the number of records in the mutex profile.
   970  func countMutex() int {
   971  	n, _ := runtime.MutexProfile(nil)
   972  	return n
   973  }
   974  
   975  // writeBlock writes the current blocking profile to w.
   976  func writeBlock(w io.Writer, debug int) error {
   977  	return writeProfileInternal(w, debug, "contention", pprof_blockProfileInternal)
   978  }
   979  
   980  // writeMutex writes the current mutex profile to w.
   981  func writeMutex(w io.Writer, debug int) error {
   982  	return writeProfileInternal(w, debug, "mutex", pprof_mutexProfileInternal)
   983  }
   984  
   985  // writeProfileInternal writes the current blocking or mutex profile depending on the passed parameters.
   986  func writeProfileInternal(w io.Writer, debug int, name string, runtimeProfile func([]profilerecord.BlockProfileRecord) (int, bool)) error {
   987  	var p []profilerecord.BlockProfileRecord
   988  	n, ok := runtimeProfile(nil)
   989  	for {
   990  		p = make([]profilerecord.BlockProfileRecord, n+50)
   991  		n, ok = runtimeProfile(p)
   992  		if ok {
   993  			p = p[:n]
   994  			break
   995  		}
   996  	}
   997  
   998  	slices.SortFunc(p, func(a, b profilerecord.BlockProfileRecord) int {
   999  		return cmp.Compare(b.Cycles, a.Cycles)
  1000  	})
  1001  
  1002  	if debug <= 0 {
  1003  		return printCountCycleProfile(w, "contentions", "delay", p)
  1004  	}
  1005  
  1006  	b := bufio.NewWriter(w)
  1007  	tw := tabwriter.NewWriter(w, 1, 8, 1, '\t', 0)
  1008  	w = tw
  1009  
  1010  	fmt.Fprintf(w, "--- %v:\n", name)
  1011  	fmt.Fprintf(w, "cycles/second=%v\n", pprof_cyclesPerSecond())
  1012  	if name == "mutex" {
  1013  		fmt.Fprintf(w, "sampling period=%d\n", runtime.SetMutexProfileFraction(-1))
  1014  	}
  1015  	expandedStack := pprof_makeProfStack()
  1016  	for i := range p {
  1017  		r := &p[i]
  1018  		fmt.Fprintf(w, "%v %v @", r.Cycles, r.Count)
  1019  		n := expandInlinedFrames(expandedStack, r.Stack)
  1020  		stack := expandedStack[:n]
  1021  		for _, pc := range stack {
  1022  			fmt.Fprintf(w, " %#x", pc)
  1023  		}
  1024  		fmt.Fprint(w, "\n")
  1025  		if debug > 0 {
  1026  			printStackRecord(w, stack, true)
  1027  		}
  1028  	}
  1029  
  1030  	if tw != nil {
  1031  		tw.Flush()
  1032  	}
  1033  	return b.Flush()
  1034  }
  1035  
  1036  //go:linkname pprof_goroutineProfileWithLabels runtime.pprof_goroutineProfileWithLabels
  1037  func pprof_goroutineProfileWithLabels(p []profilerecord.StackRecord, labels []unsafe.Pointer) (n int, ok bool)
  1038  
  1039  //go:linkname pprof_goroutineLeakProfileWithLabels runtime.pprof_goroutineLeakProfileWithLabels
  1040  func pprof_goroutineLeakProfileWithLabels(p []profilerecord.StackRecord, labels []unsafe.Pointer) (n int, ok bool)
  1041  
  1042  //go:linkname pprof_cyclesPerSecond runtime/pprof.runtime_cyclesPerSecond
  1043  func pprof_cyclesPerSecond() int64
  1044  
  1045  //go:linkname pprof_memProfileInternal runtime.pprof_memProfileInternal
  1046  func pprof_memProfileInternal(p []profilerecord.MemProfileRecord, inuseZero bool) (n int, ok bool)
  1047  
  1048  //go:linkname pprof_blockProfileInternal runtime.pprof_blockProfileInternal
  1049  func pprof_blockProfileInternal(p []profilerecord.BlockProfileRecord) (n int, ok bool)
  1050  
  1051  //go:linkname pprof_mutexProfileInternal runtime.pprof_mutexProfileInternal
  1052  func pprof_mutexProfileInternal(p []profilerecord.BlockProfileRecord) (n int, ok bool)
  1053  
  1054  //go:linkname pprof_threadCreateInternal runtime.pprof_threadCreateInternal
  1055  func pprof_threadCreateInternal(p []profilerecord.StackRecord) (n int, ok bool)
  1056  
  1057  //go:linkname pprof_fpunwindExpand runtime.pprof_fpunwindExpand
  1058  func pprof_fpunwindExpand(dst, src []uintptr) int
  1059  
  1060  //go:linkname pprof_makeProfStack runtime.pprof_makeProfStack
  1061  func pprof_makeProfStack() []uintptr
  1062
View as plain text