Source file src/runtime/trace.go

     1  // Copyright 2014 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  //go:build !goexperiment.exectracer2
     6  
     7  // Go execution tracer.
     8  // The tracer captures a wide range of execution events like goroutine
     9  // creation/blocking/unblocking, syscall enter/exit/block, GC-related events,
    10  // changes of heap size, processor start/stop, etc and writes them to a buffer
    11  // in a compact form. A precise nanosecond-precision timestamp and a stack
    12  // trace is captured for most events.
    13  // See https://golang.ir/s/go15trace for more info.
    14  
    15  package runtime
    16  
    17  import (
    18  	"internal/abi"
    19  	"internal/goarch"
    20  	"internal/goos"
    21  	"runtime/internal/atomic"
    22  	"runtime/internal/sys"
    23  	"unsafe"
    24  )
    25  
    26  // Event types in the trace, args are given in square brackets.
    27  const (
    28  	traceEvNone              = 0  // unused
    29  	traceEvBatch             = 1  // start of per-P batch of events [pid, timestamp]
    30  	traceEvFrequency         = 2  // contains tracer timer frequency [frequency (ticks per second)]
    31  	traceEvStack             = 3  // stack [stack id, number of PCs, array of {PC, func string ID, file string ID, line}]
    32  	traceEvGomaxprocs        = 4  // current value of GOMAXPROCS [timestamp, GOMAXPROCS, stack id]
    33  	traceEvProcStart         = 5  // start of P [timestamp, thread id]
    34  	traceEvProcStop          = 6  // stop of P [timestamp]
    35  	traceEvGCStart           = 7  // GC start [timestamp, seq, stack id]
    36  	traceEvGCDone            = 8  // GC done [timestamp]
    37  	traceEvSTWStart          = 9  // STW start [timestamp, kind]
    38  	traceEvSTWDone           = 10 // STW done [timestamp]
    39  	traceEvGCSweepStart      = 11 // GC sweep start [timestamp, stack id]
    40  	traceEvGCSweepDone       = 12 // GC sweep done [timestamp, swept, reclaimed]
    41  	traceEvGoCreate          = 13 // goroutine creation [timestamp, new goroutine id, new stack id, stack id]
    42  	traceEvGoStart           = 14 // goroutine starts running [timestamp, goroutine id, seq]
    43  	traceEvGoEnd             = 15 // goroutine ends [timestamp]
    44  	traceEvGoStop            = 16 // goroutine stops (like in select{}) [timestamp, stack]
    45  	traceEvGoSched           = 17 // goroutine calls Gosched [timestamp, stack]
    46  	traceEvGoPreempt         = 18 // goroutine is preempted [timestamp, stack]
    47  	traceEvGoSleep           = 19 // goroutine calls Sleep [timestamp, stack]
    48  	traceEvGoBlock           = 20 // goroutine blocks [timestamp, stack]
    49  	traceEvGoUnblock         = 21 // goroutine is unblocked [timestamp, goroutine id, seq, stack]
    50  	traceEvGoBlockSend       = 22 // goroutine blocks on chan send [timestamp, stack]
    51  	traceEvGoBlockRecv       = 23 // goroutine blocks on chan recv [timestamp, stack]
    52  	traceEvGoBlockSelect     = 24 // goroutine blocks on select [timestamp, stack]
    53  	traceEvGoBlockSync       = 25 // goroutine blocks on Mutex/RWMutex [timestamp, stack]
    54  	traceEvGoBlockCond       = 26 // goroutine blocks on Cond [timestamp, stack]
    55  	traceEvGoBlockNet        = 27 // goroutine blocks on network [timestamp, stack]
    56  	traceEvGoSysCall         = 28 // syscall enter [timestamp, stack]
    57  	traceEvGoSysExit         = 29 // syscall exit [timestamp, goroutine id, seq, real timestamp]
    58  	traceEvGoSysBlock        = 30 // syscall blocks [timestamp]
    59  	traceEvGoWaiting         = 31 // denotes that goroutine is blocked when tracing starts [timestamp, goroutine id]
    60  	traceEvGoInSyscall       = 32 // denotes that goroutine is in syscall when tracing starts [timestamp, goroutine id]
    61  	traceEvHeapAlloc         = 33 // gcController.heapLive change [timestamp, heap_alloc]
    62  	traceEvHeapGoal          = 34 // gcController.heapGoal() (formerly next_gc) change [timestamp, heap goal in bytes]
    63  	traceEvTimerGoroutine    = 35 // not currently used; previously denoted timer goroutine [timer goroutine id]
    64  	traceEvFutileWakeup      = 36 // not currently used; denotes that the previous wakeup of this goroutine was futile [timestamp]
    65  	traceEvString            = 37 // string dictionary entry [ID, length, string]
    66  	traceEvGoStartLocal      = 38 // goroutine starts running on the same P as the last event [timestamp, goroutine id]
    67  	traceEvGoUnblockLocal    = 39 // goroutine is unblocked on the same P as the last event [timestamp, goroutine id, stack]
    68  	traceEvGoSysExitLocal    = 40 // syscall exit on the same P as the last event [timestamp, goroutine id, real timestamp]
    69  	traceEvGoStartLabel      = 41 // goroutine starts running with label [timestamp, goroutine id, seq, label string id]
    70  	traceEvGoBlockGC         = 42 // goroutine blocks on GC assist [timestamp, stack]
    71  	traceEvGCMarkAssistStart = 43 // GC mark assist start [timestamp, stack]
    72  	traceEvGCMarkAssistDone  = 44 // GC mark assist done [timestamp]
    73  	traceEvUserTaskCreate    = 45 // trace.NewTask [timestamp, internal task id, internal parent task id, name string, stack]
    74  	traceEvUserTaskEnd       = 46 // end of a task [timestamp, internal task id, stack]
    75  	traceEvUserRegion        = 47 // trace.WithRegion [timestamp, internal task id, mode(0:start, 1:end), name string, stack]
    76  	traceEvUserLog           = 48 // trace.Log [timestamp, internal task id, key string id, stack, value string]
    77  	traceEvCPUSample         = 49 // CPU profiling sample [timestamp, real timestamp, real P id (-1 when absent), goroutine id, stack]
    78  	traceEvCount             = 50
    79  	// Byte is used but only 6 bits are available for event type.
    80  	// The remaining 2 bits are used to specify the number of arguments.
    81  	// That means, the max event type value is 63.
    82  )
    83  
    84  // traceBlockReason is an enumeration of reasons a goroutine might block.
    85  // This is the interface the rest of the runtime uses to tell the
    86  // tracer why a goroutine blocked. The tracer then propagates this information
    87  // into the trace however it sees fit.
    88  //
    89  // Note that traceBlockReasons should not be compared, since reasons that are
    90  // distinct by name may *not* be distinct by value.
    91  type traceBlockReason uint8
    92  
    93  // For maximal efficiency, just map the trace block reason directly to a trace
    94  // event.
    95  const (
    96  	traceBlockGeneric         traceBlockReason = traceEvGoBlock
    97  	traceBlockForever                          = traceEvGoStop
    98  	traceBlockNet                              = traceEvGoBlockNet
    99  	traceBlockSelect                           = traceEvGoBlockSelect
   100  	traceBlockCondWait                         = traceEvGoBlockCond
   101  	traceBlockSync                             = traceEvGoBlockSync
   102  	traceBlockChanSend                         = traceEvGoBlockSend
   103  	traceBlockChanRecv                         = traceEvGoBlockRecv
   104  	traceBlockGCMarkAssist                     = traceEvGoBlockGC
   105  	traceBlockGCSweep                          = traceEvGoBlock
   106  	traceBlockSystemGoroutine                  = traceEvGoBlock
   107  	traceBlockPreempted                        = traceEvGoBlock
   108  	traceBlockDebugCall                        = traceEvGoBlock
   109  	traceBlockUntilGCEnds                      = traceEvGoBlock
   110  	traceBlockSleep                            = traceEvGoSleep
   111  )
   112  
   113  const (
   114  	// Timestamps in trace are cputicks/traceTickDiv.
   115  	// This makes absolute values of timestamp diffs smaller,
   116  	// and so they are encoded in less number of bytes.
   117  	// 64 on x86 is somewhat arbitrary (one tick is ~20ns on a 3GHz machine).
   118  	// The suggested increment frequency for PowerPC's time base register is
   119  	// 512 MHz according to Power ISA v2.07 section 6.2, so we use 16 on ppc64
   120  	// and ppc64le.
   121  	traceTimeDiv = 16 + 48*(goarch.Is386|goarch.IsAmd64)
   122  	// Maximum number of PCs in a single stack trace.
   123  	// Since events contain only stack id rather than whole stack trace,
   124  	// we can allow quite large values here.
   125  	traceStackSize = 128
   126  	// Identifier of a fake P that is used when we trace without a real P.
   127  	traceGlobProc = -1
   128  	// Maximum number of bytes to encode uint64 in base-128.
   129  	traceBytesPerNumber = 10
   130  	// Shift of the number of arguments in the first event byte.
   131  	traceArgCountShift = 6
   132  )
   133  
   134  // trace is global tracing context.
   135  var trace struct {
   136  	// trace.lock must only be acquired on the system stack where
   137  	// stack splits cannot happen while it is held.
   138  	lock          mutex       // protects the following members
   139  	enabled       bool        // when set runtime traces events
   140  	shutdown      bool        // set when we are waiting for trace reader to finish after setting enabled to false
   141  	headerWritten bool        // whether ReadTrace has emitted trace header
   142  	footerWritten bool        // whether ReadTrace has emitted trace footer
   143  	shutdownSema  uint32      // used to wait for ReadTrace completion
   144  	seqStart      uint64      // sequence number when tracing was started
   145  	startTicks    int64       // cputicks when tracing was started
   146  	endTicks      int64       // cputicks when tracing was stopped
   147  	startNanotime int64       // nanotime when tracing was started
   148  	endNanotime   int64       // nanotime when tracing was stopped
   149  	startTime     traceTime   // traceClockNow when tracing started
   150  	endTime       traceTime   // traceClockNow when tracing stopped
   151  	seqGC         uint64      // GC start/done sequencer
   152  	reading       traceBufPtr // buffer currently handed off to user
   153  	empty         traceBufPtr // stack of empty buffers
   154  	fullHead      traceBufPtr // queue of full buffers
   155  	fullTail      traceBufPtr
   156  	stackTab      traceStackTable // maps stack traces to unique ids
   157  	// cpuLogRead accepts CPU profile samples from the signal handler where
   158  	// they're generated. It uses a two-word header to hold the IDs of the P and
   159  	// G (respectively) that were active at the time of the sample. Because
   160  	// profBuf uses a record with all zeros in its header to indicate overflow,
   161  	// we make sure to make the P field always non-zero: The ID of a real P will
   162  	// start at bit 1, and bit 0 will be set. Samples that arrive while no P is
   163  	// running (such as near syscalls) will set the first header field to 0b10.
   164  	// This careful handling of the first header field allows us to store ID of
   165  	// the active G directly in the second field, even though that will be 0
   166  	// when sampling g0.
   167  	cpuLogRead *profBuf
   168  	// cpuLogBuf is a trace buffer to hold events corresponding to CPU profile
   169  	// samples, which arrive out of band and not directly connected to a
   170  	// specific P.
   171  	cpuLogBuf traceBufPtr
   172  
   173  	reader atomic.Pointer[g] // goroutine that called ReadTrace, or nil
   174  
   175  	signalLock  atomic.Uint32 // protects use of the following member, only usable in signal handlers
   176  	cpuLogWrite *profBuf      // copy of cpuLogRead for use in signal handlers, set without signalLock
   177  
   178  	// Dictionary for traceEvString.
   179  	//
   180  	// TODO: central lock to access the map is not ideal.
   181  	//   option: pre-assign ids to all user annotation region names and tags
   182  	//   option: per-P cache
   183  	//   option: sync.Map like data structure
   184  	stringsLock mutex
   185  	strings     map[string]uint64
   186  	stringSeq   uint64
   187  
   188  	// markWorkerLabels maps gcMarkWorkerMode to string ID.
   189  	markWorkerLabels [len(gcMarkWorkerModeStrings)]uint64
   190  
   191  	bufLock mutex       // protects buf
   192  	buf     traceBufPtr // global trace buffer, used when running without a p
   193  }
   194  
   195  // gTraceState is per-G state for the tracer.
   196  type gTraceState struct {
   197  	sysExitTime        traceTime // timestamp when syscall has returned
   198  	tracedSyscallEnter bool      // syscall or cgo was entered while trace was enabled or StartTrace has emitted EvGoInSyscall about this goroutine
   199  	seq                uint64    // trace event sequencer
   200  	lastP              puintptr  // last P emitted an event for this goroutine
   201  }
   202  
   203  // Unused; for compatibility with the new tracer.
   204  func (s *gTraceState) reset() {}
   205  
   206  // mTraceState is per-M state for the tracer.
   207  type mTraceState struct {
   208  	startingTrace  bool // this M is in TraceStart, potentially before traceEnabled is true
   209  	tracedSTWStart bool // this M traced a STW start, so it should trace an end
   210  }
   211  
   212  // pTraceState is per-P state for the tracer.
   213  type pTraceState struct {
   214  	buf traceBufPtr
   215  
   216  	// inSweep indicates the sweep events should be traced.
   217  	// This is used to defer the sweep start event until a span
   218  	// has actually been swept.
   219  	inSweep bool
   220  
   221  	// swept and reclaimed track the number of bytes swept and reclaimed
   222  	// by sweeping in the current sweep loop (while inSweep was true).
   223  	swept, reclaimed uintptr
   224  }
   225  
   226  // traceLockInit initializes global trace locks.
   227  func traceLockInit() {
   228  	lockInit(&trace.bufLock, lockRankTraceBuf)
   229  	lockInit(&trace.stringsLock, lockRankTraceStrings)
   230  	lockInit(&trace.lock, lockRankTrace)
   231  	lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
   232  }
   233  
   234  // traceBufHeader is per-P tracing buffer.
   235  type traceBufHeader struct {
   236  	link     traceBufPtr             // in trace.empty/full
   237  	lastTime traceTime               // when we wrote the last event
   238  	pos      int                     // next write offset in arr
   239  	stk      [traceStackSize]uintptr // scratch buffer for traceback
   240  }
   241  
   242  // traceBuf is per-P tracing buffer.
   243  type traceBuf struct {
   244  	_ sys.NotInHeap
   245  	traceBufHeader
   246  	arr [64<<10 - unsafe.Sizeof(traceBufHeader{})]byte // underlying buffer for traceBufHeader.buf
   247  }
   248  
   249  // traceBufPtr is a *traceBuf that is not traced by the garbage
   250  // collector and doesn't have write barriers. traceBufs are not
   251  // allocated from the GC'd heap, so this is safe, and are often
   252  // manipulated in contexts where write barriers are not allowed, so
   253  // this is necessary.
   254  //
   255  // TODO: Since traceBuf is now embedded runtime/internal/sys.NotInHeap, this isn't necessary.
   256  type traceBufPtr uintptr
   257  
   258  func (tp traceBufPtr) ptr() *traceBuf   { return (*traceBuf)(unsafe.Pointer(tp)) }
   259  func (tp *traceBufPtr) set(b *traceBuf) { *tp = traceBufPtr(unsafe.Pointer(b)) }
   260  func traceBufPtrOf(b *traceBuf) traceBufPtr {
   261  	return traceBufPtr(unsafe.Pointer(b))
   262  }
   263  
   264  // traceEnabled returns true if the trace is currently enabled.
   265  //
   266  // nosplit because it's called on the syscall path when stack movement is forbidden.
   267  //
   268  //go:nosplit
   269  func traceEnabled() bool {
   270  	return trace.enabled
   271  }
   272  
   273  // traceShuttingDown returns true if the trace is currently shutting down.
   274  //
   275  //go:nosplit
   276  func traceShuttingDown() bool {
   277  	return trace.shutdown
   278  }
   279  
   280  // traceLocker represents an M writing trace events. While a traceLocker value
   281  // is valid, the tracer observes all operations on the G/M/P or trace events being
   282  // written as happening atomically.
   283  //
   284  // This doesn't do much for the current tracer, because the current tracer doesn't
   285  // need atomicity around non-trace runtime operations. All the state it needs it
   286  // collects carefully during a STW.
   287  type traceLocker struct {
   288  	enabled bool
   289  }
   290  
   291  // traceAcquire prepares this M for writing one or more trace events.
   292  //
   293  // This exists for compatibility with the upcoming new tracer; it doesn't do much
   294  // in the current tracer.
   295  //
   296  // nosplit because it's called on the syscall path when stack movement is forbidden.
   297  //
   298  //go:nosplit
   299  func traceAcquire() traceLocker {
   300  	if !traceEnabled() {
   301  		return traceLocker{false}
   302  	}
   303  	return traceLocker{true}
   304  }
   305  
   306  // ok returns true if the traceLocker is valid (i.e. tracing is enabled).
   307  //
   308  // nosplit because it's called on the syscall path when stack movement is forbidden.
   309  //
   310  //go:nosplit
   311  func (tl traceLocker) ok() bool {
   312  	return tl.enabled
   313  }
   314  
   315  // traceRelease indicates that this M is done writing trace events.
   316  //
   317  // This exists for compatibility with the upcoming new tracer; it doesn't do anything
   318  // in the current tracer.
   319  //
   320  // nosplit because it's called on the syscall path when stack movement is forbidden.
   321  //
   322  //go:nosplit
   323  func traceRelease(tl traceLocker) {
   324  }
   325  
   326  // StartTrace enables tracing for the current process.
   327  // While tracing, the data will be buffered and available via [ReadTrace].
   328  // StartTrace returns an error if tracing is already enabled.
   329  // Most clients should use the [runtime/trace] package or the [testing] package's
   330  // -test.trace flag instead of calling StartTrace directly.
   331  func StartTrace() error {
   332  	// Stop the world so that we can take a consistent snapshot
   333  	// of all goroutines at the beginning of the trace.
   334  	// Do not stop the world during GC so we ensure we always see
   335  	// a consistent view of GC-related events (e.g. a start is always
   336  	// paired with an end).
   337  	stw := stopTheWorldGC(stwStartTrace)
   338  
   339  	// Prevent sysmon from running any code that could generate events.
   340  	lock(&sched.sysmonlock)
   341  
   342  	// We are in stop-the-world, but syscalls can finish and write to trace concurrently.
   343  	// Exitsyscall could check trace.enabled long before and then suddenly wake up
   344  	// and decide to write to trace at a random point in time.
   345  	// However, such syscall will use the global trace.buf buffer, because we've
   346  	// acquired all p's by doing stop-the-world. So this protects us from such races.
   347  	lock(&trace.bufLock)
   348  
   349  	if trace.enabled || trace.shutdown {
   350  		unlock(&trace.bufLock)
   351  		unlock(&sched.sysmonlock)
   352  		startTheWorldGC(stw)
   353  		return errorString("tracing is already enabled")
   354  	}
   355  
   356  	// Can't set trace.enabled yet. While the world is stopped, exitsyscall could
   357  	// already emit a delayed event (see exitTicks in exitsyscall) if we set trace.enabled here.
   358  	// That would lead to an inconsistent trace:
   359  	// - either GoSysExit appears before EvGoInSyscall,
   360  	// - or GoSysExit appears for a goroutine for which we don't emit EvGoInSyscall below.
   361  	// To instruct traceEvent that it must not ignore events below, we set trace.startingTrace.
   362  	// trace.enabled is set afterwards once we have emitted all preliminary events.
   363  	mp := getg().m
   364  	mp.trace.startingTrace = true
   365  
   366  	// Obtain current stack ID to use in all traceEvGoCreate events below.
   367  	stkBuf := make([]uintptr, traceStackSize)
   368  	stackID := traceStackID(mp, stkBuf, 2)
   369  
   370  	profBuf := newProfBuf(2, profBufWordCount, profBufTagCount) // after the timestamp, header is [pp.id, gp.goid]
   371  	trace.cpuLogRead = profBuf
   372  
   373  	// We must not acquire trace.signalLock outside of a signal handler: a
   374  	// profiling signal may arrive at any time and try to acquire it, leading to
   375  	// deadlock. Because we can't use that lock to protect updates to
   376  	// trace.cpuLogWrite (only use of the structure it references), reads and
   377  	// writes of the pointer must be atomic. (And although this field is never
   378  	// the sole pointer to the profBuf value, it's best to allow a write barrier
   379  	// here.)
   380  	atomicstorep(unsafe.Pointer(&trace.cpuLogWrite), unsafe.Pointer(profBuf))
   381  
   382  	// World is stopped, no need to lock.
   383  	forEachGRace(func(gp *g) {
   384  		status := readgstatus(gp)
   385  		if status != _Gdead {
   386  			gp.trace.seq = 0
   387  			gp.trace.lastP = getg().m.p
   388  			// +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum.
   389  			id := trace.stackTab.put([]uintptr{logicalStackSentinel, startPCforTrace(gp.startpc) + sys.PCQuantum})
   390  			traceEvent(traceEvGoCreate, -1, gp.goid, uint64(id), stackID)
   391  		}
   392  		if status == _Gwaiting {
   393  			// traceEvGoWaiting is implied to have seq=1.
   394  			gp.trace.seq++
   395  			traceEvent(traceEvGoWaiting, -1, gp.goid)
   396  		}
   397  		if status == _Gsyscall {
   398  			gp.trace.seq++
   399  			gp.trace.tracedSyscallEnter = true
   400  			traceEvent(traceEvGoInSyscall, -1, gp.goid)
   401  		} else if status == _Gdead && gp.m != nil && gp.m.isextra {
   402  			// Trigger two trace events for the dead g in the extra m,
   403  			// since the next event of the g will be traceEvGoSysExit in exitsyscall,
   404  			// while calling from C thread to Go.
   405  			gp.trace.seq = 0
   406  			gp.trace.lastP = getg().m.p
   407  			// +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum.
   408  			id := trace.stackTab.put([]uintptr{logicalStackSentinel, startPCforTrace(0) + sys.PCQuantum}) // no start pc
   409  			traceEvent(traceEvGoCreate, -1, gp.goid, uint64(id), stackID)
   410  			gp.trace.seq++
   411  			gp.trace.tracedSyscallEnter = true
   412  			traceEvent(traceEvGoInSyscall, -1, gp.goid)
   413  		} else {
   414  			// We need to explicitly clear the flag. A previous trace might have ended with a goroutine
   415  			// not emitting a GoSysExit and clearing the flag, leaving it in a stale state. Clearing
   416  			// it here makes it unambiguous to any goroutine exiting a syscall racing with us that
   417  			// no EvGoInSyscall event was emitted for it. (It's not racy to set this flag here, because
   418  			// it'll only get checked when the goroutine runs again, which will be after the world starts
   419  			// again.)
   420  			gp.trace.tracedSyscallEnter = false
   421  		}
   422  	})
   423  	// Use a dummy traceLocker. The trace isn't enabled yet, but we can still write events.
   424  	tl := traceLocker{}
   425  	tl.ProcStart()
   426  	tl.GoStart()
   427  	// Note: startTicks needs to be set after we emit traceEvGoInSyscall events.
   428  	// If we do it the other way around, it is possible that exitsyscall will
   429  	// query sysExitTime after startTicks but before traceEvGoInSyscall timestamp.
   430  	// It will lead to a false conclusion that cputicks is broken.
   431  	trace.startTime = traceClockNow()
   432  	trace.startTicks = cputicks()
   433  	trace.startNanotime = nanotime()
   434  	trace.headerWritten = false
   435  	trace.footerWritten = false
   436  
   437  	// string to id mapping
   438  	//  0 : reserved for an empty string
   439  	//  remaining: other strings registered by traceString
   440  	trace.stringSeq = 0
   441  	trace.strings = make(map[string]uint64)
   442  
   443  	trace.seqGC = 0
   444  	mp.trace.startingTrace = false
   445  	trace.enabled = true
   446  
   447  	// Register runtime goroutine labels.
   448  	_, pid, bufp := traceAcquireBuffer()
   449  	for i, label := range gcMarkWorkerModeStrings[:] {
   450  		trace.markWorkerLabels[i], bufp = traceString(bufp, pid, label)
   451  	}
   452  	traceReleaseBuffer(mp, pid)
   453  
   454  	unlock(&trace.bufLock)
   455  
   456  	unlock(&sched.sysmonlock)
   457  
   458  	// Record the current state of HeapGoal to avoid information loss in trace.
   459  	//
   460  	// Use the same dummy trace locker. The trace can't end until after we start
   461  	// the world, and we can safely trace from here.
   462  	tl.HeapGoal()
   463  
   464  	startTheWorldGC(stw)
   465  	return nil
   466  }
   467  
   468  // StopTrace stops tracing, if it was previously enabled.
   469  // StopTrace only returns after all the reads for the trace have completed.
   470  func StopTrace() {
   471  	// Stop the world so that we can collect the trace buffers from all p's below,
   472  	// and also to avoid races with traceEvent.
   473  	stw := stopTheWorldGC(stwStopTrace)
   474  
   475  	// See the comment in StartTrace.
   476  	lock(&sched.sysmonlock)
   477  
   478  	// See the comment in StartTrace.
   479  	lock(&trace.bufLock)
   480  
   481  	if !trace.enabled {
   482  		unlock(&trace.bufLock)
   483  		unlock(&sched.sysmonlock)
   484  		startTheWorldGC(stw)
   485  		return
   486  	}
   487  
   488  	// Trace GoSched for us, and use a dummy locker. The world is stopped
   489  	// and we control whether the trace is enabled, so this is safe.
   490  	tl := traceLocker{}
   491  	tl.GoSched()
   492  
   493  	atomicstorep(unsafe.Pointer(&trace.cpuLogWrite), nil)
   494  	trace.cpuLogRead.close()
   495  	traceReadCPU()
   496  
   497  	// Loop over all allocated Ps because dead Ps may still have
   498  	// trace buffers.
   499  	for _, p := range allp[:cap(allp)] {
   500  		buf := p.trace.buf
   501  		if buf != 0 {
   502  			traceFullQueue(buf)
   503  			p.trace.buf = 0
   504  		}
   505  	}
   506  	if trace.buf != 0 {
   507  		buf := trace.buf
   508  		trace.buf = 0
   509  		if buf.ptr().pos != 0 {
   510  			traceFullQueue(buf)
   511  		}
   512  	}
   513  	if trace.cpuLogBuf != 0 {
   514  		buf := trace.cpuLogBuf
   515  		trace.cpuLogBuf = 0
   516  		if buf.ptr().pos != 0 {
   517  			traceFullQueue(buf)
   518  		}
   519  	}
   520  
   521  	// Wait for startNanotime != endNanotime. On Windows the default interval between
   522  	// system clock ticks is typically between 1 and 15 milliseconds, which may not
   523  	// have passed since the trace started. Without nanotime moving forward, trace
   524  	// tooling has no way of identifying how much real time each cputicks time deltas
   525  	// represent.
   526  	for {
   527  		trace.endTime = traceClockNow()
   528  		trace.endTicks = cputicks()
   529  		trace.endNanotime = nanotime()
   530  
   531  		if trace.endNanotime != trace.startNanotime || faketime != 0 {
   532  			break
   533  		}
   534  		osyield()
   535  	}
   536  
   537  	trace.enabled = false
   538  	trace.shutdown = true
   539  	unlock(&trace.bufLock)
   540  
   541  	unlock(&sched.sysmonlock)
   542  
   543  	startTheWorldGC(stw)
   544  
   545  	// The world is started but we've set trace.shutdown, so new tracing can't start.
   546  	// Wait for the trace reader to flush pending buffers and stop.
   547  	semacquire(&trace.shutdownSema)
   548  	if raceenabled {
   549  		raceacquire(unsafe.Pointer(&trace.shutdownSema))
   550  	}
   551  
   552  	systemstack(func() {
   553  		// The lock protects us from races with StartTrace/StopTrace because they do stop-the-world.
   554  		lock(&trace.lock)
   555  		for _, p := range allp[:cap(allp)] {
   556  			if p.trace.buf != 0 {
   557  				throw("trace: non-empty trace buffer in proc")
   558  			}
   559  		}
   560  		if trace.buf != 0 {
   561  			throw("trace: non-empty global trace buffer")
   562  		}
   563  		if trace.fullHead != 0 || trace.fullTail != 0 {
   564  			throw("trace: non-empty full trace buffer")
   565  		}
   566  		if trace.reading != 0 || trace.reader.Load() != nil {
   567  			throw("trace: reading after shutdown")
   568  		}
   569  		for trace.empty != 0 {
   570  			buf := trace.empty
   571  			trace.empty = buf.ptr().link
   572  			sysFree(unsafe.Pointer(buf), unsafe.Sizeof(*buf.ptr()), &memstats.other_sys)
   573  		}
   574  		trace.strings = nil
   575  		trace.shutdown = false
   576  		trace.cpuLogRead = nil
   577  		unlock(&trace.lock)
   578  	})
   579  }
   580  
   581  // ReadTrace returns the next chunk of binary tracing data, blocking until data
   582  // is available. If tracing is turned off and all the data accumulated while it
   583  // was on has been returned, ReadTrace returns nil. The caller must copy the
   584  // returned data before calling ReadTrace again.
   585  // ReadTrace must be called from one goroutine at a time.
   586  func ReadTrace() []byte {
   587  top:
   588  	var buf []byte
   589  	var park bool
   590  	systemstack(func() {
   591  		buf, park = readTrace0()
   592  	})
   593  	if park {
   594  		gopark(func(gp *g, _ unsafe.Pointer) bool {
   595  			if !trace.reader.CompareAndSwapNoWB(nil, gp) {
   596  				// We're racing with another reader.
   597  				// Wake up and handle this case.
   598  				return false
   599  			}
   600  
   601  			if g2 := traceReader(); gp == g2 {
   602  				// New data arrived between unlocking
   603  				// and the CAS and we won the wake-up
   604  				// race, so wake up directly.
   605  				return false
   606  			} else if g2 != nil {
   607  				printlock()
   608  				println("runtime: got trace reader", g2, g2.goid)
   609  				throw("unexpected trace reader")
   610  			}
   611  
   612  			return true
   613  		}, nil, waitReasonTraceReaderBlocked, traceBlockSystemGoroutine, 2)
   614  		goto top
   615  	}
   616  
   617  	return buf
   618  }
   619  
   620  // readTrace0 is ReadTrace's continuation on g0. This must run on the
   621  // system stack because it acquires trace.lock.
   622  //
   623  //go:systemstack
   624  func readTrace0() (buf []byte, park bool) {
   625  	if raceenabled {
   626  		// g0 doesn't have a race context. Borrow the user G's.
   627  		if getg().racectx != 0 {
   628  			throw("expected racectx == 0")
   629  		}
   630  		getg().racectx = getg().m.curg.racectx
   631  		// (This defer should get open-coded, which is safe on
   632  		// the system stack.)
   633  		defer func() { getg().racectx = 0 }()
   634  	}
   635  
   636  	// Optimistically look for CPU profile samples. This may write new stack
   637  	// records, and may write new tracing buffers. This must be done with the
   638  	// trace lock not held. footerWritten and shutdown are safe to access
   639  	// here. They are only mutated by this goroutine or during a STW.
   640  	if !trace.footerWritten && !trace.shutdown {
   641  		traceReadCPU()
   642  	}
   643  
   644  	// This function must not allocate while holding trace.lock:
   645  	// allocation can call heap allocate, which will try to emit a trace
   646  	// event while holding heap lock.
   647  	lock(&trace.lock)
   648  
   649  	if trace.reader.Load() != nil {
   650  		// More than one goroutine reads trace. This is bad.
   651  		// But we rather do not crash the program because of tracing,
   652  		// because tracing can be enabled at runtime on prod servers.
   653  		unlock(&trace.lock)
   654  		println("runtime: ReadTrace called from multiple goroutines simultaneously")
   655  		return nil, false
   656  	}
   657  	// Recycle the old buffer.
   658  	if buf := trace.reading; buf != 0 {
   659  		buf.ptr().link = trace.empty
   660  		trace.empty = buf
   661  		trace.reading = 0
   662  	}
   663  	// Write trace header.
   664  	if !trace.headerWritten {
   665  		trace.headerWritten = true
   666  		unlock(&trace.lock)
   667  		return []byte("go 1.21 trace\x00\x00\x00"), false
   668  	}
   669  	// Wait for new data.
   670  	if trace.fullHead == 0 && !trace.shutdown {
   671  		// We don't simply use a note because the scheduler
   672  		// executes this goroutine directly when it wakes up
   673  		// (also a note would consume an M).
   674  		unlock(&trace.lock)
   675  		return nil, true
   676  	}
   677  newFull:
   678  	assertLockHeld(&trace.lock)
   679  	// Write a buffer.
   680  	if trace.fullHead != 0 {
   681  		buf := traceFullDequeue()
   682  		trace.reading = buf
   683  		unlock(&trace.lock)
   684  		return buf.ptr().arr[:buf.ptr().pos], false
   685  	}
   686  
   687  	// Write footer with timer frequency.
   688  	if !trace.footerWritten {
   689  		trace.footerWritten = true
   690  		freq := (float64(trace.endTicks-trace.startTicks) / traceTimeDiv) / (float64(trace.endNanotime-trace.startNanotime) / 1e9)
   691  		if freq <= 0 {
   692  			throw("trace: ReadTrace got invalid frequency")
   693  		}
   694  		unlock(&trace.lock)
   695  
   696  		// Write frequency event.
   697  		bufp := traceFlush(0, 0)
   698  		buf := bufp.ptr()
   699  		buf.byte(traceEvFrequency | 0<<traceArgCountShift)
   700  		buf.varint(uint64(freq))
   701  
   702  		// Dump stack table.
   703  		// This will emit a bunch of full buffers, we will pick them up
   704  		// on the next iteration.
   705  		bufp = trace.stackTab.dump(bufp)
   706  
   707  		// Flush final buffer.
   708  		lock(&trace.lock)
   709  		traceFullQueue(bufp)
   710  		goto newFull // trace.lock should be held at newFull
   711  	}
   712  	// Done.
   713  	if trace.shutdown {
   714  		unlock(&trace.lock)
   715  		if raceenabled {
   716  			// Model synchronization on trace.shutdownSema, which race
   717  			// detector does not see. This is required to avoid false
   718  			// race reports on writer passed to trace.Start.
   719  			racerelease(unsafe.Pointer(&trace.shutdownSema))
   720  		}
   721  		// trace.enabled is already reset, so can call traceable functions.
   722  		semrelease(&trace.shutdownSema)
   723  		return nil, false
   724  	}
   725  	// Also bad, but see the comment above.
   726  	unlock(&trace.lock)
   727  	println("runtime: spurious wakeup of trace reader")
   728  	return nil, false
   729  }
   730  
   731  // traceReader returns the trace reader that should be woken up, if any.
   732  // Callers should first check that trace.enabled or trace.shutdown is set.
   733  //
   734  // This must run on the system stack because it acquires trace.lock.
   735  //
   736  //go:systemstack
   737  func traceReader() *g {
   738  	// Optimistic check first
   739  	if traceReaderAvailable() == nil {
   740  		return nil
   741  	}
   742  	lock(&trace.lock)
   743  	gp := traceReaderAvailable()
   744  	if gp == nil || !trace.reader.CompareAndSwapNoWB(gp, nil) {
   745  		unlock(&trace.lock)
   746  		return nil
   747  	}
   748  	unlock(&trace.lock)
   749  	return gp
   750  }
   751  
   752  // traceReaderAvailable returns the trace reader if it is not currently
   753  // scheduled and should be. Callers should first check that trace.enabled
   754  // or trace.shutdown is set.
   755  func traceReaderAvailable() *g {
   756  	if trace.fullHead != 0 || trace.shutdown {
   757  		return trace.reader.Load()
   758  	}
   759  	return nil
   760  }
   761  
   762  // traceProcFree frees trace buffer associated with pp.
   763  //
   764  // This must run on the system stack because it acquires trace.lock.
   765  //
   766  //go:systemstack
   767  func traceProcFree(pp *p) {
   768  	buf := pp.trace.buf
   769  	pp.trace.buf = 0
   770  	if buf == 0 {
   771  		return
   772  	}
   773  	lock(&trace.lock)
   774  	traceFullQueue(buf)
   775  	unlock(&trace.lock)
   776  }
   777  
   778  // ThreadDestroy is a no-op. It exists as a stub to support the new tracer.
   779  //
   780  // This must run on the system stack, just to match the new tracer.
   781  func traceThreadDestroy(_ *m) {
   782  	// No-op in old tracer.
   783  }
   784  
   785  // traceFullQueue queues buf into queue of full buffers.
   786  func traceFullQueue(buf traceBufPtr) {
   787  	buf.ptr().link = 0
   788  	if trace.fullHead == 0 {
   789  		trace.fullHead = buf
   790  	} else {
   791  		trace.fullTail.ptr().link = buf
   792  	}
   793  	trace.fullTail = buf
   794  }
   795  
   796  // traceFullDequeue dequeues from queue of full buffers.
   797  func traceFullDequeue() traceBufPtr {
   798  	buf := trace.fullHead
   799  	if buf == 0 {
   800  		return 0
   801  	}
   802  	trace.fullHead = buf.ptr().link
   803  	if trace.fullHead == 0 {
   804  		trace.fullTail = 0
   805  	}
   806  	buf.ptr().link = 0
   807  	return buf
   808  }
   809  
   810  // traceEvent writes a single event to trace buffer, flushing the buffer if necessary.
   811  // ev is event type.
   812  // If skip > 0, write current stack id as the last argument (skipping skip top frames).
   813  // If skip = 0, this event type should contain a stack, but we don't want
   814  // to collect and remember it for this particular call.
   815  func traceEvent(ev byte, skip int, args ...uint64) {
   816  	mp, pid, bufp := traceAcquireBuffer()
   817  	// Double-check trace.enabled now that we've done m.locks++ and acquired bufLock.
   818  	// This protects from races between traceEvent and StartTrace/StopTrace.
   819  
   820  	// The caller checked that trace.enabled == true, but trace.enabled might have been
   821  	// turned off between the check and now. Check again. traceLockBuffer did mp.locks++,
   822  	// StopTrace does stopTheWorld, and stopTheWorld waits for mp.locks to go back to zero,
   823  	// so if we see trace.enabled == true now, we know it's true for the rest of the function.
   824  	// Exitsyscall can run even during stopTheWorld. The race with StartTrace/StopTrace
   825  	// during tracing in exitsyscall is resolved by locking trace.bufLock in traceLockBuffer.
   826  	//
   827  	// Note trace_userTaskCreate runs the same check.
   828  	if !trace.enabled && !mp.trace.startingTrace {
   829  		traceReleaseBuffer(mp, pid)
   830  		return
   831  	}
   832  
   833  	if skip > 0 {
   834  		if getg() == mp.curg {
   835  			skip++ // +1 because stack is captured in traceEventLocked.
   836  		}
   837  	}
   838  	traceEventLocked(0, mp, pid, bufp, ev, 0, skip, args...)
   839  	traceReleaseBuffer(mp, pid)
   840  }
   841  
   842  // traceEventLocked writes a single event of type ev to the trace buffer bufp,
   843  // flushing the buffer if necessary. pid is the id of the current P, or
   844  // traceGlobProc if we're tracing without a real P.
   845  //
   846  // Preemption is disabled, and if running without a real P the global tracing
   847  // buffer is locked.
   848  //
   849  // Events types that do not include a stack set skip to -1. Event types that
   850  // include a stack may explicitly reference a stackID from the trace.stackTab
   851  // (obtained by an earlier call to traceStackID). Without an explicit stackID,
   852  // this function will automatically capture the stack of the goroutine currently
   853  // running on mp, skipping skip top frames or, if skip is 0, writing out an
   854  // empty stack record.
   855  //
   856  // It records the event's args to the traceBuf, and also makes an effort to
   857  // reserve extraBytes bytes of additional space immediately following the event,
   858  // in the same traceBuf.
   859  func traceEventLocked(extraBytes int, mp *m, pid int32, bufp *traceBufPtr, ev byte, stackID uint32, skip int, args ...uint64) {
   860  	buf := bufp.ptr()
   861  	// TODO: test on non-zero extraBytes param.
   862  	maxSize := 2 + 5*traceBytesPerNumber + extraBytes // event type, length, sequence, timestamp, stack id and two add params
   863  	if buf == nil || len(buf.arr)-buf.pos < maxSize {
   864  		systemstack(func() {
   865  			buf = traceFlush(traceBufPtrOf(buf), pid).ptr()
   866  		})
   867  		bufp.set(buf)
   868  	}
   869  
   870  	ts := traceClockNow()
   871  	if ts <= buf.lastTime {
   872  		ts = buf.lastTime + 1
   873  	}
   874  	tsDiff := uint64(ts - buf.lastTime)
   875  	buf.lastTime = ts
   876  	narg := byte(len(args))
   877  	if stackID != 0 || skip >= 0 {
   878  		narg++
   879  	}
   880  	// We have only 2 bits for number of arguments.
   881  	// If number is >= 3, then the event type is followed by event length in bytes.
   882  	if narg > 3 {
   883  		narg = 3
   884  	}
   885  	startPos := buf.pos
   886  	buf.byte(ev | narg<<traceArgCountShift)
   887  	var lenp *byte
   888  	if narg == 3 {
   889  		// Reserve the byte for length assuming that length < 128.
   890  		buf.varint(0)
   891  		lenp = &buf.arr[buf.pos-1]
   892  	}
   893  	buf.varint(tsDiff)
   894  	for _, a := range args {
   895  		buf.varint(a)
   896  	}
   897  	if stackID != 0 {
   898  		buf.varint(uint64(stackID))
   899  	} else if skip == 0 {
   900  		buf.varint(0)
   901  	} else if skip > 0 {
   902  		buf.varint(traceStackID(mp, buf.stk[:], skip))
   903  	}
   904  	evSize := buf.pos - startPos
   905  	if evSize > maxSize {
   906  		throw("invalid length of trace event")
   907  	}
   908  	if lenp != nil {
   909  		// Fill in actual length.
   910  		*lenp = byte(evSize - 2)
   911  	}
   912  }
   913  
   914  // traceCPUSample writes a CPU profile sample stack to the execution tracer's
   915  // profiling buffer. It is called from a signal handler, so is limited in what
   916  // it can do.
   917  func traceCPUSample(gp *g, _ *m, pp *p, stk []uintptr) {
   918  	if !traceEnabled() {
   919  		// Tracing is usually turned off; don't spend time acquiring the signal
   920  		// lock unless it's active.
   921  		return
   922  	}
   923  
   924  	// Match the clock used in traceEventLocked
   925  	now := traceClockNow()
   926  	// The "header" here is the ID of the P that was running the profiled code,
   927  	// followed by the ID of the goroutine. (For normal CPU profiling, it's
   928  	// usually the number of samples with the given stack.) Near syscalls, pp
   929  	// may be nil. Reporting goid of 0 is fine for either g0 or a nil gp.
   930  	var hdr [2]uint64
   931  	if pp != nil {
   932  		// Overflow records in profBuf have all header values set to zero. Make
   933  		// sure that real headers have at least one bit set.
   934  		hdr[0] = uint64(pp.id)<<1 | 0b1
   935  	} else {
   936  		hdr[0] = 0b10
   937  	}
   938  	if gp != nil {
   939  		hdr[1] = gp.goid
   940  	}
   941  
   942  	// Allow only one writer at a time
   943  	for !trace.signalLock.CompareAndSwap(0, 1) {
   944  		// TODO: Is it safe to osyield here? https://golang.ir/issue/52672
   945  		osyield()
   946  	}
   947  
   948  	if log := (*profBuf)(atomic.Loadp(unsafe.Pointer(&trace.cpuLogWrite))); log != nil {
   949  		// Note: we don't pass a tag pointer here (how should profiling tags
   950  		// interact with the execution tracer?), but if we did we'd need to be
   951  		// careful about write barriers. See the long comment in profBuf.write.
   952  		log.write(nil, int64(now), hdr[:], stk)
   953  	}
   954  
   955  	trace.signalLock.Store(0)
   956  }
   957  
   958  func traceReadCPU() {
   959  	bufp := &trace.cpuLogBuf
   960  
   961  	for {
   962  		data, tags, _ := trace.cpuLogRead.read(profBufNonBlocking)
   963  		if len(data) == 0 {
   964  			break
   965  		}
   966  		for len(data) > 0 {
   967  			if len(data) < 4 || data[0] > uint64(len(data)) {
   968  				break // truncated profile
   969  			}
   970  			if data[0] < 4 || tags != nil && len(tags) < 1 {
   971  				break // malformed profile
   972  			}
   973  			if len(tags) < 1 {
   974  				break // mismatched profile records and tags
   975  			}
   976  			timestamp := data[1]
   977  			ppid := data[2] >> 1
   978  			if hasP := (data[2] & 0b1) != 0; !hasP {
   979  				ppid = ^uint64(0)
   980  			}
   981  			goid := data[3]
   982  			stk := data[4:data[0]]
   983  			empty := len(stk) == 1 && data[2] == 0 && data[3] == 0
   984  			data = data[data[0]:]
   985  			// No support here for reporting goroutine tags at the moment; if
   986  			// that information is to be part of the execution trace, we'd
   987  			// probably want to see when the tags are applied and when they
   988  			// change, instead of only seeing them when we get a CPU sample.
   989  			tags = tags[1:]
   990  
   991  			if empty {
   992  				// Looks like an overflow record from the profBuf. Not much to
   993  				// do here, we only want to report full records.
   994  				//
   995  				// TODO: should we start a goroutine to drain the profBuf,
   996  				// rather than relying on a high-enough volume of tracing events
   997  				// to keep ReadTrace busy? https://golang.ir/issue/52674
   998  				continue
   999  			}
  1000  
  1001  			buf := bufp.ptr()
  1002  			if buf == nil {
  1003  				systemstack(func() {
  1004  					*bufp = traceFlush(*bufp, 0)
  1005  				})
  1006  				buf = bufp.ptr()
  1007  			}
  1008  			nstk := 1
  1009  			buf.stk[0] = logicalStackSentinel
  1010  			for ; nstk < len(buf.stk) && nstk-1 < len(stk); nstk++ {
  1011  				buf.stk[nstk] = uintptr(stk[nstk-1])
  1012  			}
  1013  			stackID := trace.stackTab.put(buf.stk[:nstk])
  1014  
  1015  			traceEventLocked(0, nil, 0, bufp, traceEvCPUSample, stackID, 1, timestamp, ppid, goid)
  1016  		}
  1017  	}
  1018  }
  1019  
  1020  // logicalStackSentinel is a sentinel value at pcBuf[0] signifying that
  1021  // pcBuf[1:] holds a logical stack requiring no further processing. Any other
  1022  // value at pcBuf[0] represents a skip value to apply to the physical stack in
  1023  // pcBuf[1:] after inline expansion.
  1024  const logicalStackSentinel = ^uintptr(0)
  1025  
  1026  // traceStackID captures a stack trace into pcBuf, registers it in the trace
  1027  // stack table, and returns its unique ID. pcBuf should have a length equal to
  1028  // traceStackSize. skip controls the number of leaf frames to omit in order to
  1029  // hide tracer internals from stack traces, see CL 5523.
  1030  func traceStackID(mp *m, pcBuf []uintptr, skip int) uint64 {
  1031  	gp := getg()
  1032  	curgp := mp.curg
  1033  	nstk := 1
  1034  	if tracefpunwindoff() || mp.hasCgoOnStack() {
  1035  		// Slow path: Unwind using default unwinder. Used when frame pointer
  1036  		// unwinding is unavailable or disabled (tracefpunwindoff), or might
  1037  		// produce incomplete results or crashes (hasCgoOnStack). Note that no
  1038  		// cgo callback related crashes have been observed yet. The main
  1039  		// motivation is to take advantage of a potentially registered cgo
  1040  		// symbolizer.
  1041  		pcBuf[0] = logicalStackSentinel
  1042  		if curgp == gp {
  1043  			nstk += callers(skip+1, pcBuf[1:])
  1044  		} else if curgp != nil {
  1045  			nstk += gcallers(curgp, skip, pcBuf[1:])
  1046  		}
  1047  	} else {
  1048  		// Fast path: Unwind using frame pointers.
  1049  		pcBuf[0] = uintptr(skip)
  1050  		if curgp == gp {
  1051  			nstk += fpTracebackPCs(unsafe.Pointer(getfp()), pcBuf[1:])
  1052  		} else if curgp != nil {
  1053  			// We're called on the g0 stack through mcall(fn) or systemstack(fn). To
  1054  			// behave like gcallers above, we start unwinding from sched.bp, which
  1055  			// points to the caller frame of the leaf frame on g's stack. The return
  1056  			// address of the leaf frame is stored in sched.pc, which we manually
  1057  			// capture here.
  1058  			pcBuf[1] = curgp.sched.pc
  1059  			nstk += 1 + fpTracebackPCs(unsafe.Pointer(curgp.sched.bp), pcBuf[2:])
  1060  		}
  1061  	}
  1062  	if nstk > 0 {
  1063  		nstk-- // skip runtime.goexit
  1064  	}
  1065  	if nstk > 0 && curgp.goid == 1 {
  1066  		nstk-- // skip runtime.main
  1067  	}
  1068  	id := trace.stackTab.put(pcBuf[:nstk])
  1069  	return uint64(id)
  1070  }
  1071  
  1072  // tracefpunwindoff returns true if frame pointer unwinding for the tracer is
  1073  // disabled via GODEBUG or not supported by the architecture.
  1074  // TODO(#60254): support frame pointer unwinding on plan9/amd64.
  1075  func tracefpunwindoff() bool {
  1076  	return debug.tracefpunwindoff != 0 || (goarch.ArchFamily != goarch.AMD64 && goarch.ArchFamily != goarch.ARM64) || goos.IsPlan9 == 1
  1077  }
  1078  
  1079  // fpTracebackPCs populates pcBuf with the return addresses for each frame and
  1080  // returns the number of PCs written to pcBuf. The returned PCs correspond to
  1081  // "physical frames" rather than "logical frames"; that is if A is inlined into
  1082  // B, this will return a PC for only B.
  1083  func fpTracebackPCs(fp unsafe.Pointer, pcBuf []uintptr) (i int) {
  1084  	for i = 0; i < len(pcBuf) && fp != nil; i++ {
  1085  		// return addr sits one word above the frame pointer
  1086  		pcBuf[i] = *(*uintptr)(unsafe.Pointer(uintptr(fp) + goarch.PtrSize))
  1087  		// follow the frame pointer to the next one
  1088  		fp = unsafe.Pointer(*(*uintptr)(fp))
  1089  	}
  1090  	return i
  1091  }
  1092  
  1093  // traceAcquireBuffer returns trace buffer to use and, if necessary, locks it.
  1094  func traceAcquireBuffer() (mp *m, pid int32, bufp *traceBufPtr) {
  1095  	// Any time we acquire a buffer, we may end up flushing it,
  1096  	// but flushes are rare. Record the lock edge even if it
  1097  	// doesn't happen this time.
  1098  	lockRankMayTraceFlush()
  1099  
  1100  	mp = acquirem()
  1101  	if p := mp.p.ptr(); p != nil {
  1102  		return mp, p.id, &p.trace.buf
  1103  	}
  1104  	lock(&trace.bufLock)
  1105  	return mp, traceGlobProc, &trace.buf
  1106  }
  1107  
  1108  // traceReleaseBuffer releases a buffer previously acquired with traceAcquireBuffer.
  1109  func traceReleaseBuffer(mp *m, pid int32) {
  1110  	if pid == traceGlobProc {
  1111  		unlock(&trace.bufLock)
  1112  	}
  1113  	releasem(mp)
  1114  }
  1115  
  1116  // lockRankMayTraceFlush records the lock ranking effects of a
  1117  // potential call to traceFlush.
  1118  func lockRankMayTraceFlush() {
  1119  	lockWithRankMayAcquire(&trace.lock, getLockRank(&trace.lock))
  1120  }
  1121  
  1122  // traceFlush puts buf onto stack of full buffers and returns an empty buffer.
  1123  //
  1124  // This must run on the system stack because it acquires trace.lock.
  1125  //
  1126  //go:systemstack
  1127  func traceFlush(buf traceBufPtr, pid int32) traceBufPtr {
  1128  	lock(&trace.lock)
  1129  	if buf != 0 {
  1130  		traceFullQueue(buf)
  1131  	}
  1132  	if trace.empty != 0 {
  1133  		buf = trace.empty
  1134  		trace.empty = buf.ptr().link
  1135  	} else {
  1136  		buf = traceBufPtr(sysAlloc(unsafe.Sizeof(traceBuf{}), &memstats.other_sys))
  1137  		if buf == 0 {
  1138  			throw("trace: out of memory")
  1139  		}
  1140  	}
  1141  	bufp := buf.ptr()
  1142  	bufp.link.set(nil)
  1143  	bufp.pos = 0
  1144  
  1145  	// initialize the buffer for a new batch
  1146  	ts := traceClockNow()
  1147  	if ts <= bufp.lastTime {
  1148  		ts = bufp.lastTime + 1
  1149  	}
  1150  	bufp.lastTime = ts
  1151  	bufp.byte(traceEvBatch | 1<<traceArgCountShift)
  1152  	bufp.varint(uint64(pid))
  1153  	bufp.varint(uint64(ts))
  1154  
  1155  	unlock(&trace.lock)
  1156  	return buf
  1157  }
  1158  
  1159  // traceString adds a string to the trace.strings and returns the id.
  1160  func traceString(bufp *traceBufPtr, pid int32, s string) (uint64, *traceBufPtr) {
  1161  	if s == "" {
  1162  		return 0, bufp
  1163  	}
  1164  
  1165  	lock(&trace.stringsLock)
  1166  	if raceenabled {
  1167  		// raceacquire is necessary because the map access
  1168  		// below is race annotated.
  1169  		raceacquire(unsafe.Pointer(&trace.stringsLock))
  1170  	}
  1171  
  1172  	if id, ok := trace.strings[s]; ok {
  1173  		if raceenabled {
  1174  			racerelease(unsafe.Pointer(&trace.stringsLock))
  1175  		}
  1176  		unlock(&trace.stringsLock)
  1177  
  1178  		return id, bufp
  1179  	}
  1180  
  1181  	trace.stringSeq++
  1182  	id := trace.stringSeq
  1183  	trace.strings[s] = id
  1184  
  1185  	if raceenabled {
  1186  		racerelease(unsafe.Pointer(&trace.stringsLock))
  1187  	}
  1188  	unlock(&trace.stringsLock)
  1189  
  1190  	// memory allocation in above may trigger tracing and
  1191  	// cause *bufp changes. Following code now works with *bufp,
  1192  	// so there must be no memory allocation or any activities
  1193  	// that causes tracing after this point.
  1194  
  1195  	buf := bufp.ptr()
  1196  	size := 1 + 2*traceBytesPerNumber + len(s)
  1197  	if buf == nil || len(buf.arr)-buf.pos < size {
  1198  		systemstack(func() {
  1199  			buf = traceFlush(traceBufPtrOf(buf), pid).ptr()
  1200  			bufp.set(buf)
  1201  		})
  1202  	}
  1203  	buf.byte(traceEvString)
  1204  	buf.varint(id)
  1205  
  1206  	// double-check the string and the length can fit.
  1207  	// Otherwise, truncate the string.
  1208  	slen := len(s)
  1209  	if room := len(buf.arr) - buf.pos; room < slen+traceBytesPerNumber {
  1210  		slen = room
  1211  	}
  1212  
  1213  	buf.varint(uint64(slen))
  1214  	buf.pos += copy(buf.arr[buf.pos:], s[:slen])
  1215  
  1216  	bufp.set(buf)
  1217  	return id, bufp
  1218  }
  1219  
  1220  // varint appends v to buf in little-endian-base-128 encoding.
  1221  func (buf *traceBuf) varint(v uint64) {
  1222  	pos := buf.pos
  1223  	for ; v >= 0x80; v >>= 7 {
  1224  		buf.arr[pos] = 0x80 | byte(v)
  1225  		pos++
  1226  	}
  1227  	buf.arr[pos] = byte(v)
  1228  	pos++
  1229  	buf.pos = pos
  1230  }
  1231  
  1232  // varintAt writes varint v at byte position pos in buf. This always
  1233  // consumes traceBytesPerNumber bytes. This is intended for when the
  1234  // caller needs to reserve space for a varint but can't populate it
  1235  // until later.
  1236  func (buf *traceBuf) varintAt(pos int, v uint64) {
  1237  	for i := 0; i < traceBytesPerNumber; i++ {
  1238  		if i < traceBytesPerNumber-1 {
  1239  			buf.arr[pos] = 0x80 | byte(v)
  1240  		} else {
  1241  			buf.arr[pos] = byte(v)
  1242  		}
  1243  		v >>= 7
  1244  		pos++
  1245  	}
  1246  }
  1247  
  1248  // byte appends v to buf.
  1249  func (buf *traceBuf) byte(v byte) {
  1250  	buf.arr[buf.pos] = v
  1251  	buf.pos++
  1252  }
  1253  
  1254  // traceStackTable maps stack traces (arrays of PC's) to unique uint32 ids.
  1255  // It is lock-free for reading.
  1256  type traceStackTable struct {
  1257  	lock mutex // Must be acquired on the system stack
  1258  	seq  uint32
  1259  	mem  traceAlloc
  1260  	tab  [1 << 13]traceStackPtr
  1261  }
  1262  
  1263  // traceStack is a single stack in traceStackTable.
  1264  type traceStack struct {
  1265  	link traceStackPtr
  1266  	hash uintptr
  1267  	id   uint32
  1268  	n    int
  1269  	stk  [0]uintptr // real type [n]uintptr
  1270  }
  1271  
  1272  type traceStackPtr uintptr
  1273  
  1274  func (tp traceStackPtr) ptr() *traceStack { return (*traceStack)(unsafe.Pointer(tp)) }
  1275  
  1276  // stack returns slice of PCs.
  1277  func (ts *traceStack) stack() []uintptr {
  1278  	return (*[traceStackSize]uintptr)(unsafe.Pointer(&ts.stk))[:ts.n]
  1279  }
  1280  
  1281  // put returns a unique id for the stack trace pcs and caches it in the table,
  1282  // if it sees the trace for the first time.
  1283  func (tab *traceStackTable) put(pcs []uintptr) uint32 {
  1284  	if len(pcs) == 0 {
  1285  		return 0
  1286  	}
  1287  	hash := memhash(unsafe.Pointer(&pcs[0]), 0, uintptr(len(pcs))*unsafe.Sizeof(pcs[0]))
  1288  	// First, search the hashtable w/o the mutex.
  1289  	if id := tab.find(pcs, hash); id != 0 {
  1290  		return id
  1291  	}
  1292  	// Now, double check under the mutex.
  1293  	// Switch to the system stack so we can acquire tab.lock
  1294  	var id uint32
  1295  	systemstack(func() {
  1296  		lock(&tab.lock)
  1297  		if id = tab.find(pcs, hash); id != 0 {
  1298  			unlock(&tab.lock)
  1299  			return
  1300  		}
  1301  		// Create new record.
  1302  		tab.seq++
  1303  		stk := tab.newStack(len(pcs))
  1304  		stk.hash = hash
  1305  		stk.id = tab.seq
  1306  		id = stk.id
  1307  		stk.n = len(pcs)
  1308  		stkpc := stk.stack()
  1309  		copy(stkpc, pcs)
  1310  		part := int(hash % uintptr(len(tab.tab)))
  1311  		stk.link = tab.tab[part]
  1312  		atomicstorep(unsafe.Pointer(&tab.tab[part]), unsafe.Pointer(stk))
  1313  		unlock(&tab.lock)
  1314  	})
  1315  	return id
  1316  }
  1317  
  1318  // find checks if the stack trace pcs is already present in the table.
  1319  func (tab *traceStackTable) find(pcs []uintptr, hash uintptr) uint32 {
  1320  	part := int(hash % uintptr(len(tab.tab)))
  1321  Search:
  1322  	for stk := tab.tab[part].ptr(); stk != nil; stk = stk.link.ptr() {
  1323  		if stk.hash == hash && stk.n == len(pcs) {
  1324  			for i, stkpc := range stk.stack() {
  1325  				if stkpc != pcs[i] {
  1326  					continue Search
  1327  				}
  1328  			}
  1329  			return stk.id
  1330  		}
  1331  	}
  1332  	return 0
  1333  }
  1334  
  1335  // newStack allocates a new stack of size n.
  1336  func (tab *traceStackTable) newStack(n int) *traceStack {
  1337  	return (*traceStack)(tab.mem.alloc(unsafe.Sizeof(traceStack{}) + uintptr(n)*goarch.PtrSize))
  1338  }
  1339  
  1340  // traceFrames returns the frames corresponding to pcs. It may
  1341  // allocate and may emit trace events.
  1342  func traceFrames(bufp traceBufPtr, pcs []uintptr) ([]traceFrame, traceBufPtr) {
  1343  	frames := make([]traceFrame, 0, len(pcs))
  1344  	ci := CallersFrames(pcs)
  1345  	for {
  1346  		var frame traceFrame
  1347  		f, more := ci.Next()
  1348  		frame, bufp = traceFrameForPC(bufp, 0, f)
  1349  		frames = append(frames, frame)
  1350  		if !more {
  1351  			return frames, bufp
  1352  		}
  1353  	}
  1354  }
  1355  
  1356  // dump writes all previously cached stacks to trace buffers,
  1357  // releases all memory and resets state.
  1358  //
  1359  // This must run on the system stack because it calls traceFlush.
  1360  //
  1361  //go:systemstack
  1362  func (tab *traceStackTable) dump(bufp traceBufPtr) traceBufPtr {
  1363  	for i := range tab.tab {
  1364  		stk := tab.tab[i].ptr()
  1365  		for ; stk != nil; stk = stk.link.ptr() {
  1366  			var frames []traceFrame
  1367  			frames, bufp = traceFrames(bufp, fpunwindExpand(stk.stack()))
  1368  
  1369  			// Estimate the size of this record. This
  1370  			// bound is pretty loose, but avoids counting
  1371  			// lots of varint sizes.
  1372  			maxSize := 1 + traceBytesPerNumber + (2+4*len(frames))*traceBytesPerNumber
  1373  			// Make sure we have enough buffer space.
  1374  			if buf := bufp.ptr(); len(buf.arr)-buf.pos < maxSize {
  1375  				bufp = traceFlush(bufp, 0)
  1376  			}
  1377  
  1378  			// Emit header, with space reserved for length.
  1379  			buf := bufp.ptr()
  1380  			buf.byte(traceEvStack | 3<<traceArgCountShift)
  1381  			lenPos := buf.pos
  1382  			buf.pos += traceBytesPerNumber
  1383  
  1384  			// Emit body.
  1385  			recPos := buf.pos
  1386  			buf.varint(uint64(stk.id))
  1387  			buf.varint(uint64(len(frames)))
  1388  			for _, frame := range frames {
  1389  				buf.varint(uint64(frame.PC))
  1390  				buf.varint(frame.funcID)
  1391  				buf.varint(frame.fileID)
  1392  				buf.varint(frame.line)
  1393  			}
  1394  
  1395  			// Fill in size header.
  1396  			buf.varintAt(lenPos, uint64(buf.pos-recPos))
  1397  		}
  1398  	}
  1399  
  1400  	tab.mem.drop()
  1401  	*tab = traceStackTable{}
  1402  	lockInit(&((*tab).lock), lockRankTraceStackTab)
  1403  
  1404  	return bufp
  1405  }
  1406  
  1407  // fpunwindExpand checks if pcBuf contains logical frames (which include inlined
  1408  // frames) or physical frames (produced by frame pointer unwinding) using a
  1409  // sentinel value in pcBuf[0]. Logical frames are simply returned without the
  1410  // sentinel. Physical frames are turned into logical frames via inline unwinding
  1411  // and by applying the skip value that's stored in pcBuf[0].
  1412  func fpunwindExpand(pcBuf []uintptr) []uintptr {
  1413  	if len(pcBuf) > 0 && pcBuf[0] == logicalStackSentinel {
  1414  		// pcBuf contains logical rather than inlined frames, skip has already been
  1415  		// applied, just return it without the sentinel value in pcBuf[0].
  1416  		return pcBuf[1:]
  1417  	}
  1418  
  1419  	var (
  1420  		lastFuncID = abi.FuncIDNormal
  1421  		newPCBuf   = make([]uintptr, 0, traceStackSize)
  1422  		skip       = pcBuf[0]
  1423  		// skipOrAdd skips or appends retPC to newPCBuf and returns true if more
  1424  		// pcs can be added.
  1425  		skipOrAdd = func(retPC uintptr) bool {
  1426  			if skip > 0 {
  1427  				skip--
  1428  			} else {
  1429  				newPCBuf = append(newPCBuf, retPC)
  1430  			}
  1431  			return len(newPCBuf) < cap(newPCBuf)
  1432  		}
  1433  	)
  1434  
  1435  outer:
  1436  	for _, retPC := range pcBuf[1:] {
  1437  		callPC := retPC - 1
  1438  		fi := findfunc(callPC)
  1439  		if !fi.valid() {
  1440  			// There is no funcInfo if callPC belongs to a C function. In this case
  1441  			// we still keep the pc, but don't attempt to expand inlined frames.
  1442  			if more := skipOrAdd(retPC); !more {
  1443  				break outer
  1444  			}
  1445  			continue
  1446  		}
  1447  
  1448  		u, uf := newInlineUnwinder(fi, callPC)
  1449  		for ; uf.valid(); uf = u.next(uf) {
  1450  			sf := u.srcFunc(uf)
  1451  			if sf.funcID == abi.FuncIDWrapper && elideWrapperCalling(lastFuncID) {
  1452  				// ignore wrappers
  1453  			} else if more := skipOrAdd(uf.pc + 1); !more {
  1454  				break outer
  1455  			}
  1456  			lastFuncID = sf.funcID
  1457  		}
  1458  	}
  1459  	return newPCBuf
  1460  }
  1461  
  1462  type traceFrame struct {
  1463  	PC     uintptr
  1464  	funcID uint64
  1465  	fileID uint64
  1466  	line   uint64
  1467  }
  1468  
  1469  // traceFrameForPC records the frame information.
  1470  // It may allocate memory.
  1471  func traceFrameForPC(buf traceBufPtr, pid int32, f Frame) (traceFrame, traceBufPtr) {
  1472  	bufp := &buf
  1473  	var frame traceFrame
  1474  	frame.PC = f.PC
  1475  
  1476  	fn := f.Function
  1477  	const maxLen = 1 << 10
  1478  	if len(fn) > maxLen {
  1479  		fn = fn[len(fn)-maxLen:]
  1480  	}
  1481  	frame.funcID, bufp = traceString(bufp, pid, fn)
  1482  	frame.line = uint64(f.Line)
  1483  	file := f.File
  1484  	if len(file) > maxLen {
  1485  		file = file[len(file)-maxLen:]
  1486  	}
  1487  	frame.fileID, bufp = traceString(bufp, pid, file)
  1488  	return frame, (*bufp)
  1489  }
  1490  
  1491  // traceAlloc is a non-thread-safe region allocator.
  1492  // It holds a linked list of traceAllocBlock.
  1493  type traceAlloc struct {
  1494  	head traceAllocBlockPtr
  1495  	off  uintptr
  1496  }
  1497  
  1498  // traceAllocBlock is a block in traceAlloc.
  1499  //
  1500  // traceAllocBlock is allocated from non-GC'd memory, so it must not
  1501  // contain heap pointers. Writes to pointers to traceAllocBlocks do
  1502  // not need write barriers.
  1503  type traceAllocBlock struct {
  1504  	_    sys.NotInHeap
  1505  	next traceAllocBlockPtr
  1506  	data [64<<10 - goarch.PtrSize]byte
  1507  }
  1508  
  1509  // TODO: Since traceAllocBlock is now embedded runtime/internal/sys.NotInHeap, this isn't necessary.
  1510  type traceAllocBlockPtr uintptr
  1511  
  1512  func (p traceAllocBlockPtr) ptr() *traceAllocBlock   { return (*traceAllocBlock)(unsafe.Pointer(p)) }
  1513  func (p *traceAllocBlockPtr) set(x *traceAllocBlock) { *p = traceAllocBlockPtr(unsafe.Pointer(x)) }
  1514  
  1515  // alloc allocates n-byte block.
  1516  func (a *traceAlloc) alloc(n uintptr) unsafe.Pointer {
  1517  	n = alignUp(n, goarch.PtrSize)
  1518  	if a.head == 0 || a.off+n > uintptr(len(a.head.ptr().data)) {
  1519  		if n > uintptr(len(a.head.ptr().data)) {
  1520  			throw("trace: alloc too large")
  1521  		}
  1522  		block := (*traceAllocBlock)(sysAlloc(unsafe.Sizeof(traceAllocBlock{}), &memstats.other_sys))
  1523  		if block == nil {
  1524  			throw("trace: out of memory")
  1525  		}
  1526  		block.next.set(a.head.ptr())
  1527  		a.head.set(block)
  1528  		a.off = 0
  1529  	}
  1530  	p := &a.head.ptr().data[a.off]
  1531  	a.off += n
  1532  	return unsafe.Pointer(p)
  1533  }
  1534  
  1535  // drop frees all previously allocated memory and resets the allocator.
  1536  func (a *traceAlloc) drop() {
  1537  	for a.head != 0 {
  1538  		block := a.head.ptr()
  1539  		a.head.set(block.next.ptr())
  1540  		sysFree(unsafe.Pointer(block), unsafe.Sizeof(traceAllocBlock{}), &memstats.other_sys)
  1541  	}
  1542  }
  1543  
  1544  // The following functions write specific events to trace.
  1545  
  1546  func (_ traceLocker) Gomaxprocs(procs int32) {
  1547  	traceEvent(traceEvGomaxprocs, 1, uint64(procs))
  1548  }
  1549  
  1550  func (_ traceLocker) ProcStart() {
  1551  	traceEvent(traceEvProcStart, -1, uint64(getg().m.id))
  1552  }
  1553  
  1554  func (_ traceLocker) ProcStop(pp *p) {
  1555  	// Sysmon and stopTheWorld can stop Ps blocked in syscalls,
  1556  	// to handle this we temporary employ the P.
  1557  	mp := acquirem()
  1558  	oldp := mp.p
  1559  	mp.p.set(pp)
  1560  	traceEvent(traceEvProcStop, -1)
  1561  	mp.p = oldp
  1562  	releasem(mp)
  1563  }
  1564  
  1565  func (_ traceLocker) GCStart() {
  1566  	traceEvent(traceEvGCStart, 3, trace.seqGC)
  1567  	trace.seqGC++
  1568  }
  1569  
  1570  func (_ traceLocker) GCDone() {
  1571  	traceEvent(traceEvGCDone, -1)
  1572  }
  1573  
  1574  func (_ traceLocker) STWStart(reason stwReason) {
  1575  	// Don't trace if this STW is for trace start/stop, since traceEnabled
  1576  	// switches during a STW.
  1577  	if reason == stwStartTrace || reason == stwStopTrace {
  1578  		return
  1579  	}
  1580  	getg().m.trace.tracedSTWStart = true
  1581  	traceEvent(traceEvSTWStart, -1, uint64(reason))
  1582  }
  1583  
  1584  func (_ traceLocker) STWDone() {
  1585  	mp := getg().m
  1586  	if !mp.trace.tracedSTWStart {
  1587  		return
  1588  	}
  1589  	mp.trace.tracedSTWStart = false
  1590  	traceEvent(traceEvSTWDone, -1)
  1591  }
  1592  
  1593  // traceGCSweepStart prepares to trace a sweep loop. This does not
  1594  // emit any events until traceGCSweepSpan is called.
  1595  //
  1596  // traceGCSweepStart must be paired with traceGCSweepDone and there
  1597  // must be no preemption points between these two calls.
  1598  func (_ traceLocker) GCSweepStart() {
  1599  	// Delay the actual GCSweepStart event until the first span
  1600  	// sweep. If we don't sweep anything, don't emit any events.
  1601  	pp := getg().m.p.ptr()
  1602  	if pp.trace.inSweep {
  1603  		throw("double traceGCSweepStart")
  1604  	}
  1605  	pp.trace.inSweep, pp.trace.swept, pp.trace.reclaimed = true, 0, 0
  1606  }
  1607  
  1608  // traceGCSweepSpan traces the sweep of a single page.
  1609  //
  1610  // This may be called outside a traceGCSweepStart/traceGCSweepDone
  1611  // pair; however, it will not emit any trace events in this case.
  1612  func (_ traceLocker) GCSweepSpan(bytesSwept uintptr) {
  1613  	pp := getg().m.p.ptr()
  1614  	if pp.trace.inSweep {
  1615  		if pp.trace.swept == 0 {
  1616  			traceEvent(traceEvGCSweepStart, 1)
  1617  		}
  1618  		pp.trace.swept += bytesSwept
  1619  	}
  1620  }
  1621  
  1622  func (_ traceLocker) GCSweepDone() {
  1623  	pp := getg().m.p.ptr()
  1624  	if !pp.trace.inSweep {
  1625  		throw("missing traceGCSweepStart")
  1626  	}
  1627  	if pp.trace.swept != 0 {
  1628  		traceEvent(traceEvGCSweepDone, -1, uint64(pp.trace.swept), uint64(pp.trace.reclaimed))
  1629  	}
  1630  	pp.trace.inSweep = false
  1631  }
  1632  
  1633  func (_ traceLocker) GCMarkAssistStart() {
  1634  	traceEvent(traceEvGCMarkAssistStart, 1)
  1635  }
  1636  
  1637  func (_ traceLocker) GCMarkAssistDone() {
  1638  	traceEvent(traceEvGCMarkAssistDone, -1)
  1639  }
  1640  
  1641  func (_ traceLocker) GoCreate(newg *g, pc uintptr) {
  1642  	newg.trace.seq = 0
  1643  	newg.trace.lastP = getg().m.p
  1644  	// +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum.
  1645  	id := trace.stackTab.put([]uintptr{logicalStackSentinel, startPCforTrace(pc) + sys.PCQuantum})
  1646  	traceEvent(traceEvGoCreate, 2, newg.goid, uint64(id))
  1647  }
  1648  
  1649  func (_ traceLocker) GoStart() {
  1650  	gp := getg().m.curg
  1651  	pp := gp.m.p
  1652  	gp.trace.seq++
  1653  	if pp.ptr().gcMarkWorkerMode != gcMarkWorkerNotWorker {
  1654  		traceEvent(traceEvGoStartLabel, -1, gp.goid, gp.trace.seq, trace.markWorkerLabels[pp.ptr().gcMarkWorkerMode])
  1655  	} else if gp.trace.lastP == pp {
  1656  		traceEvent(traceEvGoStartLocal, -1, gp.goid)
  1657  	} else {
  1658  		gp.trace.lastP = pp
  1659  		traceEvent(traceEvGoStart, -1, gp.goid, gp.trace.seq)
  1660  	}
  1661  }
  1662  
  1663  func (_ traceLocker) GoEnd() {
  1664  	traceEvent(traceEvGoEnd, -1)
  1665  }
  1666  
  1667  func (_ traceLocker) GoSched() {
  1668  	gp := getg()
  1669  	gp.trace.lastP = gp.m.p
  1670  	traceEvent(traceEvGoSched, 1)
  1671  }
  1672  
  1673  func (_ traceLocker) GoPreempt() {
  1674  	gp := getg()
  1675  	gp.trace.lastP = gp.m.p
  1676  	traceEvent(traceEvGoPreempt, 1)
  1677  }
  1678  
  1679  func (_ traceLocker) GoPark(reason traceBlockReason, skip int) {
  1680  	// Convert the block reason directly to a trace event type.
  1681  	// See traceBlockReason for more information.
  1682  	traceEvent(byte(reason), skip)
  1683  }
  1684  
  1685  func (_ traceLocker) GoUnpark(gp *g, skip int) {
  1686  	pp := getg().m.p
  1687  	gp.trace.seq++
  1688  	if gp.trace.lastP == pp {
  1689  		traceEvent(traceEvGoUnblockLocal, skip, gp.goid)
  1690  	} else {
  1691  		gp.trace.lastP = pp
  1692  		traceEvent(traceEvGoUnblock, skip, gp.goid, gp.trace.seq)
  1693  	}
  1694  }
  1695  
  1696  func (_ traceLocker) GoSysCall() {
  1697  	var skip int
  1698  	switch {
  1699  	case tracefpunwindoff():
  1700  		// Unwind by skipping 1 frame relative to gp.syscallsp which is captured 3
  1701  		// frames above this frame. For frame pointer unwinding we produce the same
  1702  		// results by hard coding the number of frames in between our caller and the
  1703  		// actual syscall, see cases below.
  1704  		// TODO(felixge): Implement gp.syscallbp to avoid this workaround?
  1705  		skip = 1
  1706  	case GOOS == "solaris" || GOOS == "illumos":
  1707  		// These platforms don't use a libc_read_trampoline.
  1708  		skip = 3
  1709  	default:
  1710  		// Skip the extra trampoline frame used on most systems.
  1711  		skip = 4
  1712  	}
  1713  	getg().m.curg.trace.tracedSyscallEnter = true
  1714  	traceEvent(traceEvGoSysCall, skip)
  1715  }
  1716  
  1717  func (_ traceLocker) GoSysExit(lostP bool) {
  1718  	if !lostP {
  1719  		throw("lostP must always be true in the old tracer for GoSysExit")
  1720  	}
  1721  	gp := getg().m.curg
  1722  	if !gp.trace.tracedSyscallEnter {
  1723  		// There was no syscall entry traced for us at all, so there's definitely
  1724  		// no EvGoSysBlock or EvGoInSyscall before us, which EvGoSysExit requires.
  1725  		return
  1726  	}
  1727  	gp.trace.tracedSyscallEnter = false
  1728  	ts := gp.trace.sysExitTime
  1729  	if ts != 0 && ts < trace.startTime {
  1730  		// There is a race between the code that initializes sysExitTimes
  1731  		// (in exitsyscall, which runs without a P, and therefore is not
  1732  		// stopped with the rest of the world) and the code that initializes
  1733  		// a new trace. The recorded sysExitTime must therefore be treated
  1734  		// as "best effort". If they are valid for this trace, then great,
  1735  		// use them for greater accuracy. But if they're not valid for this
  1736  		// trace, assume that the trace was started after the actual syscall
  1737  		// exit (but before we actually managed to start the goroutine,
  1738  		// aka right now), and assign a fresh time stamp to keep the log consistent.
  1739  		ts = 0
  1740  	}
  1741  	gp.trace.sysExitTime = 0
  1742  	gp.trace.seq++
  1743  	gp.trace.lastP = gp.m.p
  1744  	traceEvent(traceEvGoSysExit, -1, gp.goid, gp.trace.seq, uint64(ts))
  1745  }
  1746  
  1747  // nosplit because it's called from exitsyscall without a P.
  1748  //
  1749  //go:nosplit
  1750  func (_ traceLocker) RecordSyscallExitedTime(gp *g, oldp *p) {
  1751  	// Wait till traceGoSysBlock event is emitted.
  1752  	// This ensures consistency of the trace (the goroutine is started after it is blocked).
  1753  	for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
  1754  		osyield()
  1755  	}
  1756  	// We can't trace syscall exit right now because we don't have a P.
  1757  	// Tracing code can invoke write barriers that cannot run without a P.
  1758  	// So instead we remember the syscall exit time and emit the event
  1759  	// in execute when we have a P.
  1760  	gp.trace.sysExitTime = traceClockNow()
  1761  }
  1762  
  1763  func (_ traceLocker) GoSysBlock(pp *p) {
  1764  	// Sysmon and stopTheWorld can declare syscalls running on remote Ps as blocked,
  1765  	// to handle this we temporary employ the P.
  1766  	mp := acquirem()
  1767  	oldp := mp.p
  1768  	mp.p.set(pp)
  1769  	traceEvent(traceEvGoSysBlock, -1)
  1770  	mp.p = oldp
  1771  	releasem(mp)
  1772  }
  1773  
  1774  func (t traceLocker) ProcSteal(pp *p, forMe bool) {
  1775  	t.ProcStop(pp)
  1776  }
  1777  
  1778  func (_ traceLocker) HeapAlloc(live uint64) {
  1779  	traceEvent(traceEvHeapAlloc, -1, live)
  1780  }
  1781  
  1782  func (_ traceLocker) HeapGoal() {
  1783  	heapGoal := gcController.heapGoal()
  1784  	if heapGoal == ^uint64(0) {
  1785  		// Heap-based triggering is disabled.
  1786  		traceEvent(traceEvHeapGoal, -1, 0)
  1787  	} else {
  1788  		traceEvent(traceEvHeapGoal, -1, heapGoal)
  1789  	}
  1790  }
  1791  
  1792  // To access runtime functions from runtime/trace.
  1793  // See runtime/trace/annotation.go
  1794  
  1795  //go:linkname trace_userTaskCreate runtime/trace.userTaskCreate
  1796  func trace_userTaskCreate(id, parentID uint64, taskType string) {
  1797  	if !trace.enabled {
  1798  		return
  1799  	}
  1800  
  1801  	// Same as in traceEvent.
  1802  	mp, pid, bufp := traceAcquireBuffer()
  1803  	if !trace.enabled && !mp.trace.startingTrace {
  1804  		traceReleaseBuffer(mp, pid)
  1805  		return
  1806  	}
  1807  
  1808  	typeStringID, bufp := traceString(bufp, pid, taskType)
  1809  	traceEventLocked(0, mp, pid, bufp, traceEvUserTaskCreate, 0, 3, id, parentID, typeStringID)
  1810  	traceReleaseBuffer(mp, pid)
  1811  }
  1812  
  1813  //go:linkname trace_userTaskEnd runtime/trace.userTaskEnd
  1814  func trace_userTaskEnd(id uint64) {
  1815  	traceEvent(traceEvUserTaskEnd, 2, id)
  1816  }
  1817  
  1818  //go:linkname trace_userRegion runtime/trace.userRegion
  1819  func trace_userRegion(id, mode uint64, name string) {
  1820  	if !trace.enabled {
  1821  		return
  1822  	}
  1823  
  1824  	mp, pid, bufp := traceAcquireBuffer()
  1825  	if !trace.enabled && !mp.trace.startingTrace {
  1826  		traceReleaseBuffer(mp, pid)
  1827  		return
  1828  	}
  1829  
  1830  	nameStringID, bufp := traceString(bufp, pid, name)
  1831  	traceEventLocked(0, mp, pid, bufp, traceEvUserRegion, 0, 3, id, mode, nameStringID)
  1832  	traceReleaseBuffer(mp, pid)
  1833  }
  1834  
  1835  //go:linkname trace_userLog runtime/trace.userLog
  1836  func trace_userLog(id uint64, category, message string) {
  1837  	if !trace.enabled {
  1838  		return
  1839  	}
  1840  
  1841  	mp, pid, bufp := traceAcquireBuffer()
  1842  	if !trace.enabled && !mp.trace.startingTrace {
  1843  		traceReleaseBuffer(mp, pid)
  1844  		return
  1845  	}
  1846  
  1847  	categoryID, bufp := traceString(bufp, pid, category)
  1848  
  1849  	// The log message is recorded after all of the normal trace event
  1850  	// arguments, including the task, category, and stack IDs. We must ask
  1851  	// traceEventLocked to reserve extra space for the length of the message
  1852  	// and the message itself.
  1853  	extraSpace := traceBytesPerNumber + len(message)
  1854  	traceEventLocked(extraSpace, mp, pid, bufp, traceEvUserLog, 0, 3, id, categoryID)
  1855  	buf := bufp.ptr()
  1856  
  1857  	// double-check the message and its length can fit.
  1858  	// Otherwise, truncate the message.
  1859  	slen := len(message)
  1860  	if room := len(buf.arr) - buf.pos; room < slen+traceBytesPerNumber {
  1861  		slen = room
  1862  	}
  1863  	buf.varint(uint64(slen))
  1864  	buf.pos += copy(buf.arr[buf.pos:], message[:slen])
  1865  
  1866  	traceReleaseBuffer(mp, pid)
  1867  }
  1868  
  1869  // the start PC of a goroutine for tracing purposes. If pc is a wrapper,
  1870  // it returns the PC of the wrapped function. Otherwise it returns pc.
  1871  func startPCforTrace(pc uintptr) uintptr {
  1872  	f := findfunc(pc)
  1873  	if !f.valid() {
  1874  		return pc // may happen for locked g in extra M since its pc is 0.
  1875  	}
  1876  	w := funcdata(f, abi.FUNCDATA_WrapInfo)
  1877  	if w == nil {
  1878  		return pc // not a wrapper
  1879  	}
  1880  	return f.datap.textAddr(*(*uint32)(w))
  1881  }
  1882  
  1883  // OneNewExtraM registers the fact that a new extra M was created with
  1884  // the tracer. This matters if the M (which has an attached G) is used while
  1885  // the trace is still active because if it is, we need the fact that it exists
  1886  // to show up in the final trace.
  1887  func (tl traceLocker) OneNewExtraM(gp *g) {
  1888  	// Trigger two trace events for the locked g in the extra m,
  1889  	// since the next event of the g will be traceEvGoSysExit in exitsyscall,
  1890  	// while calling from C thread to Go.
  1891  	tl.GoCreate(gp, 0) // no start pc
  1892  	gp.trace.seq++
  1893  	traceEvent(traceEvGoInSyscall, -1, gp.goid)
  1894  }
  1895  
  1896  // Used only in the new tracer.
  1897  func (tl traceLocker) GoCreateSyscall(gp *g) {
  1898  }
  1899  
  1900  // Used only in the new tracer.
  1901  func (tl traceLocker) GoDestroySyscall() {
  1902  }
  1903  
  1904  // traceTime represents a timestamp for the trace.
  1905  type traceTime uint64
  1906  
  1907  // traceClockNow returns a monotonic timestamp. The clock this function gets
  1908  // the timestamp from is specific to tracing, and shouldn't be mixed with other
  1909  // clock sources.
  1910  //
  1911  // nosplit because it's called from exitsyscall, which is nosplit.
  1912  //
  1913  //go:nosplit
  1914  func traceClockNow() traceTime {
  1915  	return traceTime(cputicks() / traceTimeDiv)
  1916  }
  1917  
  1918  func traceExitingSyscall() {
  1919  }
  1920  
  1921  func traceExitedSyscall() {
  1922  }
  1923  
  1924  // Not used in the old tracer. Defined for compatibility.
  1925  const defaultTraceAdvancePeriod = 0
  1926  

View as plain text