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

View as plain text