Source file src/runtime/trace.go

     1  // Copyright 2023 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 execution tracer.
     6  // The tracer captures a wide range of execution events like goroutine
     7  // creation/blocking/unblocking, syscall enter/exit/block, GC-related events,
     8  // changes of heap size, processor start/stop, etc and writes them to a buffer
     9  // in a compact form. A precise nanosecond-precision timestamp and a stack
    10  // trace is captured for most events.
    11  //
    12  // Tracer invariants (to keep the synchronization making sense):
    13  // - An m that has a trace buffer must be on either the allm or sched.freem lists.
    14  // - Any trace buffer mutation must either be happening in traceAdvance or between
    15  //   a traceAcquire and a subsequent traceRelease.
    16  // - traceAdvance cannot return until the previous generation's buffers are all flushed.
    17  //
    18  // See https://go.dev/issue/60773 for a link to the full design.
    19  
    20  package runtime
    21  
    22  import (
    23  	"internal/runtime/atomic"
    24  	"unsafe"
    25  )
    26  
    27  // Trace state.
    28  
    29  // trace is global tracing context.
    30  var trace struct {
    31  	// trace.lock must only be acquired on the system stack where
    32  	// stack splits cannot happen while it is held.
    33  	lock mutex
    34  
    35  	// Trace buffer management.
    36  	//
    37  	// First we check the empty list for any free buffers. If not, buffers
    38  	// are allocated directly from the OS. Once they're filled up and/or
    39  	// flushed, they end up on the full queue for trace.gen%2.
    40  	//
    41  	// The trace reader takes buffers off the full list one-by-one and
    42  	// places them into reading until they're finished being read from.
    43  	// Then they're placed onto the empty list.
    44  	//
    45  	// Protected by trace.lock.
    46  	reading       *traceBuf // buffer currently handed off to user
    47  	empty         *traceBuf // stack of empty buffers
    48  	full          [2]traceBufQueue
    49  	workAvailable atomic.Bool
    50  
    51  	// State for the trace reader goroutine.
    52  	//
    53  	// Protected by trace.lock.
    54  	readerGen     atomic.Uintptr // the generation the reader is currently reading for
    55  	flushedGen    atomic.Uintptr // the last completed generation
    56  	headerWritten bool           // whether ReadTrace has emitted trace header
    57  
    58  	// doneSema is used to synchronize the reader and traceAdvance. Specifically,
    59  	// it notifies traceAdvance that the reader is done with a generation.
    60  	// Both semaphores are 0 by default (so, acquires block). traceAdvance
    61  	// attempts to acquire for gen%2 after flushing the last buffers for gen.
    62  	// Meanwhile the reader releases the sema for gen%2 when it has finished
    63  	// processing gen.
    64  	doneSema [2]uint32
    65  
    66  	// Trace data tables for deduplicating data going into the trace.
    67  	// There are 2 of each: one for gen%2, one for 1-gen%2.
    68  	stackTab  [2]traceStackTable  // maps stack traces to unique ids
    69  	stringTab [2]traceStringTable // maps strings to unique ids
    70  	typeTab   [2]traceTypeTable   // maps type pointers to unique ids
    71  
    72  	// cpuLogRead accepts CPU profile samples from the signal handler where
    73  	// they're generated. There are two profBufs here: one for gen%2, one for
    74  	// 1-gen%2. These profBufs use a three-word header to hold the IDs of the P, G,
    75  	// and M (respectively) that were active at the time of the sample. Because
    76  	// profBuf uses a record with all zeros in its header to indicate overflow,
    77  	// we make sure to make the P field always non-zero: The ID of a real P will
    78  	// start at bit 1, and bit 0 will be set. Samples that arrive while no P is
    79  	// running (such as near syscalls) will set the first header field to 0b10.
    80  	// This careful handling of the first header field allows us to store ID of
    81  	// the active G directly in the second field, even though that will be 0
    82  	// when sampling g0.
    83  	//
    84  	// Initialization and teardown of these fields is protected by traceAdvanceSema.
    85  	cpuLogRead  [2]*profBuf
    86  	signalLock  atomic.Uint32              // protects use of the following member, only usable in signal handlers
    87  	cpuLogWrite [2]atomic.Pointer[profBuf] // copy of cpuLogRead for use in signal handlers, set without signalLock
    88  	cpuSleep    *wakeableSleep
    89  	cpuLogDone  <-chan struct{}
    90  	cpuBuf      [2]*traceBuf
    91  
    92  	reader atomic.Pointer[g] // goroutine that called ReadTrace, or nil
    93  
    94  	// Fast mappings from enumerations to string IDs that are prepopulated
    95  	// in the trace.
    96  	markWorkerLabels [2][len(gcMarkWorkerModeStrings)]traceArg
    97  	goStopReasons    [2][len(traceGoStopReasonStrings)]traceArg
    98  	goBlockReasons   [2][len(traceBlockReasonStrings)]traceArg
    99  
   100  	// enabled indicates whether tracing is enabled, but it is only an optimization,
   101  	// NOT the source of truth on whether tracing is enabled. Tracing is only truly
   102  	// enabled if gen != 0. This is used as an optimistic fast path check.
   103  	//
   104  	// Transitioning this value from true -> false is easy (once gen is 0)
   105  	// because it's OK for enabled to have a stale "true" value. traceAcquire will
   106  	// always double-check gen.
   107  	//
   108  	// Transitioning this value from false -> true is harder. We need to make sure
   109  	// this is observable as true strictly before gen != 0. To maintain this invariant
   110  	// we only make this transition with the world stopped and use the store to gen
   111  	// as a publication barrier.
   112  	enabled bool
   113  
   114  	// enabledWithAllocFree is set if debug.traceallocfree is != 0 when tracing begins.
   115  	// It follows the same synchronization protocol as enabled.
   116  	enabledWithAllocFree bool
   117  
   118  	// Trace generation counter.
   119  	gen            atomic.Uintptr
   120  	lastNonZeroGen uintptr // last non-zero value of gen
   121  
   122  	// shutdown is set when we are waiting for trace reader to finish after setting gen to 0
   123  	//
   124  	// Writes protected by trace.lock.
   125  	shutdown atomic.Bool
   126  
   127  	// Number of goroutines in syscall exiting slow path.
   128  	exitingSyscall atomic.Int32
   129  
   130  	// seqGC is the sequence counter for GC begin/end.
   131  	//
   132  	// Mutated only during stop-the-world.
   133  	seqGC uint64
   134  
   135  	// minPageHeapAddr is the minimum address of the page heap when tracing started.
   136  	minPageHeapAddr uint64
   137  
   138  	// debugMalloc is the value of debug.malloc before tracing began.
   139  	debugMalloc bool
   140  }
   141  
   142  // Trace public API.
   143  
   144  var (
   145  	traceAdvanceSema  uint32 = 1
   146  	traceShutdownSema uint32 = 1
   147  )
   148  
   149  // StartTrace enables tracing for the current process.
   150  // While tracing, the data will be buffered and available via [ReadTrace].
   151  // StartTrace returns an error if tracing is already enabled.
   152  // Most clients should use the [runtime/trace] package or the [testing] package's
   153  // -test.trace flag instead of calling StartTrace directly.
   154  func StartTrace() error {
   155  	if traceEnabled() || traceShuttingDown() {
   156  		return errorString("tracing is already enabled")
   157  	}
   158  	// Block until cleanup of the last trace is done.
   159  	semacquire(&traceShutdownSema)
   160  	semrelease(&traceShutdownSema)
   161  
   162  	// Hold traceAdvanceSema across trace start, since we'll want it on
   163  	// the other side of tracing being enabled globally.
   164  	semacquire(&traceAdvanceSema)
   165  
   166  	// Initialize CPU profile -> trace ingestion.
   167  	traceInitReadCPU()
   168  
   169  	// Compute the first generation for this StartTrace.
   170  	//
   171  	// Note: we start from the last non-zero generation rather than 1 so we
   172  	// can avoid resetting all the arrays indexed by gen%2 or gen%3. There's
   173  	// more than one of each per m, p, and goroutine.
   174  	firstGen := traceNextGen(trace.lastNonZeroGen)
   175  
   176  	// Reset GC sequencer.
   177  	trace.seqGC = 1
   178  
   179  	// Reset trace reader state.
   180  	trace.headerWritten = false
   181  	trace.readerGen.Store(firstGen)
   182  	trace.flushedGen.Store(0)
   183  
   184  	// Register some basic strings in the string tables.
   185  	traceRegisterLabelsAndReasons(firstGen)
   186  
   187  	// Stop the world.
   188  	//
   189  	// The purpose of stopping the world is to make sure that no goroutine is in a
   190  	// context where it could emit an event by bringing all goroutines to a safe point
   191  	// with no opportunity to transition.
   192  	//
   193  	// The exception to this rule are goroutines that are concurrently exiting a syscall.
   194  	// Those will all be forced into the syscalling slow path, and we'll just make sure
   195  	// that we don't observe any goroutines in that critical section before starting
   196  	// the world again.
   197  	//
   198  	// A good follow-up question to this is why stopping the world is necessary at all
   199  	// given that we have traceAcquire and traceRelease. Unfortunately, those only help
   200  	// us when tracing is already active (for performance, so when tracing is off the
   201  	// tracing seqlock is left untouched). The main issue here is subtle: we're going to
   202  	// want to obtain a correct starting status for each goroutine, but there are windows
   203  	// of time in which we could read and emit an incorrect status. Specifically:
   204  	//
   205  	//	trace := traceAcquire()
   206  	//  // <----> problem window
   207  	//	casgstatus(gp, _Gwaiting, _Grunnable)
   208  	//	if trace.ok() {
   209  	//		trace.GoUnpark(gp, 2)
   210  	//		traceRelease(trace)
   211  	//	}
   212  	//
   213  	// More precisely, if we readgstatus for a gp while another goroutine is in the problem
   214  	// window and that goroutine didn't observe that tracing had begun, then we might write
   215  	// a GoStatus(GoWaiting) event for that goroutine, but it won't trace an event marking
   216  	// the transition from GoWaiting to GoRunnable. The trace will then be broken, because
   217  	// future events will be emitted assuming the tracer sees GoRunnable.
   218  	//
   219  	// In short, what we really need here is to make sure that the next time *any goroutine*
   220  	// hits a traceAcquire, it sees that the trace is enabled.
   221  	//
   222  	// Note also that stopping the world is necessary to make sure sweep-related events are
   223  	// coherent. Since the world is stopped and sweeps are non-preemptible, we can never start
   224  	// the world and see an unpaired sweep 'end' event. Other parts of the tracer rely on this.
   225  	stw := stopTheWorld(stwStartTrace)
   226  
   227  	// Prevent sysmon from running any code that could generate events.
   228  	lock(&sched.sysmonlock)
   229  
   230  	// Grab the minimum page heap address. All Ps are stopped, so it's safe to read this since
   231  	// nothing can allocate heap memory.
   232  	trace.minPageHeapAddr = uint64(mheap_.pages.inUse.ranges[0].base.addr())
   233  
   234  	// Reset mSyscallID on all Ps while we have them stationary and the trace is disabled.
   235  	for _, pp := range allp {
   236  		pp.trace.mSyscallID = -1
   237  	}
   238  
   239  	// Start tracing.
   240  	//
   241  	// Set trace.enabled. This is *very* subtle. We need to maintain the invariant that if
   242  	// trace.gen != 0, then trace.enabled is always observed as true. Simultaneously, for
   243  	// performance, we need trace.enabled to be read without any synchronization.
   244  	//
   245  	// We ensure this is safe by stopping the world, which acts a global barrier on almost
   246  	// every M, and explicitly synchronize with any other Ms that could be running concurrently
   247  	// with us. Today, there are only two such cases:
   248  	// - sysmon, which we synchronized with by acquiring sysmonlock.
   249  	// - goroutines exiting syscalls, which we synchronize with via trace.exitingSyscall.
   250  	//
   251  	// After trace.gen is updated, other Ms may start creating trace buffers and emitting
   252  	// data into them.
   253  	trace.enabled = true
   254  	if debug.traceallocfree.Load() != 0 {
   255  		// Enable memory events since the GODEBUG is set.
   256  		trace.debugMalloc = debug.malloc
   257  		trace.enabledWithAllocFree = true
   258  		debug.malloc = true
   259  	}
   260  	trace.gen.Store(firstGen)
   261  
   262  	// Wait for exitingSyscall to drain.
   263  	//
   264  	// It may not monotonically decrease to zero, but in the limit it will always become
   265  	// zero because the world is stopped and there are no available Ps for syscall-exited
   266  	// goroutines to run on.
   267  	//
   268  	// Because we set gen before checking this, and because exitingSyscall is always incremented
   269  	// *before* traceAcquire (which checks gen), we can be certain that when exitingSyscall is zero
   270  	// that any goroutine that goes to exit a syscall from then on *must* observe the new gen as
   271  	// well as trace.enabled being set to true.
   272  	//
   273  	// The critical section on each goroutine here is going to be quite short, so the likelihood
   274  	// that we observe a zero value is high.
   275  	for trace.exitingSyscall.Load() != 0 {
   276  		osyield()
   277  	}
   278  
   279  	// Record some initial pieces of information.
   280  	//
   281  	// N.B. This will also emit a status event for this goroutine.
   282  	tl := traceAcquire()
   283  	tl.Gomaxprocs(gomaxprocs)  // Get this as early in the trace as possible. See comment in traceAdvance.
   284  	tl.STWStart(stwStartTrace) // We didn't trace this above, so trace it now.
   285  
   286  	// Record the fact that a GC is active, if applicable.
   287  	if gcphase == _GCmark || gcphase == _GCmarktermination {
   288  		tl.GCActive()
   289  	}
   290  
   291  	// Dump a snapshot of memory, if enabled.
   292  	if trace.enabledWithAllocFree {
   293  		traceSnapshotMemory(firstGen)
   294  	}
   295  
   296  	// Record the heap goal so we have it at the very beginning of the trace.
   297  	tl.HeapGoal()
   298  
   299  	// Make sure a ProcStatus is emitted for every P, while we're here.
   300  	for _, pp := range allp {
   301  		tl.writer().writeProcStatusForP(pp, pp == tl.mp.p.ptr()).end()
   302  	}
   303  	traceRelease(tl)
   304  
   305  	unlock(&sched.sysmonlock)
   306  	startTheWorld(stw)
   307  
   308  	traceStartReadCPU()
   309  	traceAdvancer.start()
   310  
   311  	semrelease(&traceAdvanceSema)
   312  	return nil
   313  }
   314  
   315  // StopTrace stops tracing, if it was previously enabled.
   316  // StopTrace only returns after all the reads for the trace have completed.
   317  func StopTrace() {
   318  	traceAdvance(true)
   319  }
   320  
   321  // traceAdvance moves tracing to the next generation, and cleans up the current generation,
   322  // ensuring that it's flushed out before returning. If stopTrace is true, it disables tracing
   323  // altogether instead of advancing to the next generation.
   324  //
   325  // traceAdvanceSema must not be held.
   326  //
   327  // traceAdvance is called by golang.org/x/exp/trace using linkname.
   328  //
   329  //go:linkname traceAdvance
   330  func traceAdvance(stopTrace bool) {
   331  	semacquire(&traceAdvanceSema)
   332  
   333  	// Get the gen that we're advancing from. In this function we don't really care much
   334  	// about the generation we're advancing _into_ since we'll do all the cleanup in this
   335  	// generation for the next advancement.
   336  	gen := trace.gen.Load()
   337  	if gen == 0 {
   338  		// We may end up here traceAdvance is called concurrently with StopTrace.
   339  		semrelease(&traceAdvanceSema)
   340  		return
   341  	}
   342  
   343  	// Write an EvFrequency event for this generation.
   344  	//
   345  	// N.B. This may block for quite a while to get a good frequency estimate, so make sure we do
   346  	// this here and not e.g. on the trace reader.
   347  	traceFrequency(gen)
   348  
   349  	// Collect all the untraced Gs.
   350  	type untracedG struct {
   351  		gp           *g
   352  		goid         uint64
   353  		mid          int64
   354  		stackID      uint64
   355  		status       uint32
   356  		waitreason   waitReason
   357  		inMarkAssist bool
   358  	}
   359  	var untracedGs []untracedG
   360  	forEachGRace(func(gp *g) {
   361  		// Make absolutely sure all Gs are ready for the next
   362  		// generation. We need to do this even for dead Gs because
   363  		// they may come alive with a new identity, and its status
   364  		// traced bookkeeping might end up being stale.
   365  		// We may miss totally new goroutines, but they'll always
   366  		// have clean bookkeeping.
   367  		gp.trace.readyNextGen(gen)
   368  		// If the status was traced, nothing else to do.
   369  		if gp.trace.statusWasTraced(gen) {
   370  			return
   371  		}
   372  		// Scribble down information about this goroutine.
   373  		ug := untracedG{gp: gp, mid: -1}
   374  		systemstack(func() {
   375  			me := getg().m.curg
   376  			// We don't have to handle this G status transition because we
   377  			// already eliminated ourselves from consideration above.
   378  			casGToWaitingForGC(me, _Grunning, waitReasonTraceGoroutineStatus)
   379  			// We need to suspend and take ownership of the G to safely read its
   380  			// goid. Note that we can't actually emit the event at this point
   381  			// because we might stop the G in a window where it's unsafe to write
   382  			// events based on the G's status. We need the global trace buffer flush
   383  			// coming up to make sure we're not racing with the G.
   384  			//
   385  			// It should be very unlikely that we try to preempt a running G here.
   386  			// The only situation that we might is that we're racing with a G
   387  			// that's running for the first time in this generation. Therefore,
   388  			// this should be relatively fast.
   389  			s := suspendG(gp)
   390  			if !s.dead {
   391  				ug.goid = s.g.goid
   392  				if s.g.m != nil {
   393  					ug.mid = int64(s.g.m.procid)
   394  				}
   395  				ug.status = readgstatus(s.g) &^ _Gscan
   396  				ug.waitreason = s.g.waitreason
   397  				ug.inMarkAssist = s.g.inMarkAssist
   398  				ug.stackID = traceStack(0, gp, gen)
   399  			}
   400  			resumeG(s)
   401  			casgstatus(me, _Gwaiting, _Grunning)
   402  		})
   403  		if ug.goid != 0 {
   404  			untracedGs = append(untracedGs, ug)
   405  		}
   406  	})
   407  
   408  	if !stopTrace {
   409  		// Re-register runtime goroutine labels and stop/block reasons.
   410  		traceRegisterLabelsAndReasons(traceNextGen(gen))
   411  	}
   412  
   413  	// Now that we've done some of the heavy stuff, prevent the world from stopping.
   414  	// This is necessary to ensure the consistency of the STW events. If we're feeling
   415  	// adventurous we could lift this restriction and add a STWActive event, but the
   416  	// cost of maintaining this consistency is low. We're not going to hold this semaphore
   417  	// for very long and most STW periods are very short.
   418  	// Once we hold worldsema, prevent preemption as well so we're not interrupted partway
   419  	// through this. We want to get this done as soon as possible.
   420  	semacquire(&worldsema)
   421  	mp := acquirem()
   422  
   423  	// Advance the generation or stop the trace.
   424  	trace.lastNonZeroGen = gen
   425  	if stopTrace {
   426  		systemstack(func() {
   427  			// Ordering is important here. Set shutdown first, then disable tracing,
   428  			// so that conditions like (traceEnabled() || traceShuttingDown()) have
   429  			// no opportunity to be false. Hold the trace lock so this update appears
   430  			// atomic to the trace reader.
   431  			lock(&trace.lock)
   432  			trace.shutdown.Store(true)
   433  			trace.gen.Store(0)
   434  			unlock(&trace.lock)
   435  
   436  			// Clear trace.enabled. It is totally OK for this value to be stale,
   437  			// because traceAcquire will always double-check gen.
   438  			trace.enabled = false
   439  		})
   440  	} else {
   441  		trace.gen.Store(traceNextGen(gen))
   442  	}
   443  
   444  	// Emit a ProcsChange event so we have one on record for each generation.
   445  	// Let's emit it as soon as possible so that downstream tools can rely on the value
   446  	// being there fairly soon in a generation.
   447  	//
   448  	// It's important that we do this before allowing stop-the-worlds again,
   449  	// because the procs count could change.
   450  	if !stopTrace {
   451  		tl := traceAcquire()
   452  		tl.Gomaxprocs(gomaxprocs)
   453  		traceRelease(tl)
   454  	}
   455  
   456  	// Emit a GCActive event in the new generation if necessary.
   457  	//
   458  	// It's important that we do this before allowing stop-the-worlds again,
   459  	// because that could emit global GC-related events.
   460  	if !stopTrace && (gcphase == _GCmark || gcphase == _GCmarktermination) {
   461  		tl := traceAcquire()
   462  		tl.GCActive()
   463  		traceRelease(tl)
   464  	}
   465  
   466  	// Preemption is OK again after this. If the world stops or whatever it's fine.
   467  	// We're just cleaning up the last generation after this point.
   468  	//
   469  	// We also don't care if the GC starts again after this for the same reasons.
   470  	releasem(mp)
   471  	semrelease(&worldsema)
   472  
   473  	// Snapshot allm and freem.
   474  	//
   475  	// Snapshotting after the generation counter update is sufficient.
   476  	// Because an m must be on either allm or sched.freem if it has an active trace
   477  	// buffer, new threads added to allm after this point must necessarily observe
   478  	// the new generation number (sched.lock acts as a barrier).
   479  	//
   480  	// Threads that exit before this point and are on neither list explicitly
   481  	// flush their own buffers in traceThreadDestroy.
   482  	//
   483  	// Snapshotting freem is necessary because Ms can continue to emit events
   484  	// while they're still on that list. Removal from sched.freem is serialized with
   485  	// this snapshot, so either we'll capture an m on sched.freem and race with
   486  	// the removal to flush its buffers (resolved by traceThreadDestroy acquiring
   487  	// the thread's seqlock, which one of us must win, so at least its old gen buffer
   488  	// will be flushed in time for the new generation) or it will have flushed its
   489  	// buffers before we snapshotted it to begin with.
   490  	lock(&sched.lock)
   491  	mToFlush := allm
   492  	for mp := mToFlush; mp != nil; mp = mp.alllink {
   493  		mp.trace.link = mp.alllink
   494  	}
   495  	for mp := sched.freem; mp != nil; mp = mp.freelink {
   496  		mp.trace.link = mToFlush
   497  		mToFlush = mp
   498  	}
   499  	unlock(&sched.lock)
   500  
   501  	// Iterate over our snapshot, flushing every buffer until we're done.
   502  	//
   503  	// Because trace writers read the generation while the seqlock is
   504  	// held, we can be certain that when there are no writers there are
   505  	// also no stale generation values left. Therefore, it's safe to flush
   506  	// any buffers that remain in that generation's slot.
   507  	const debugDeadlock = false
   508  	systemstack(func() {
   509  		// Track iterations for some rudimentary deadlock detection.
   510  		i := 0
   511  		detectedDeadlock := false
   512  
   513  		for mToFlush != nil {
   514  			prev := &mToFlush
   515  			for mp := *prev; mp != nil; {
   516  				if mp.trace.seqlock.Load()%2 != 0 {
   517  					// The M is writing. Come back to it later.
   518  					prev = &mp.trace.link
   519  					mp = mp.trace.link
   520  					continue
   521  				}
   522  				// Flush the trace buffer.
   523  				//
   524  				// trace.lock needed for traceBufFlush, but also to synchronize
   525  				// with traceThreadDestroy, which flushes both buffers unconditionally.
   526  				lock(&trace.lock)
   527  				for exp, buf := range mp.trace.buf[gen%2] {
   528  					if buf != nil {
   529  						traceBufFlush(buf, gen)
   530  						mp.trace.buf[gen%2][exp] = nil
   531  					}
   532  				}
   533  				unlock(&trace.lock)
   534  
   535  				// Remove the m from the flush list.
   536  				*prev = mp.trace.link
   537  				mp.trace.link = nil
   538  				mp = *prev
   539  			}
   540  			// Yield only if we're going to be going around the loop again.
   541  			if mToFlush != nil {
   542  				osyield()
   543  			}
   544  
   545  			if debugDeadlock {
   546  				// Try to detect a deadlock. We probably shouldn't loop here
   547  				// this many times.
   548  				if i > 100000 && !detectedDeadlock {
   549  					detectedDeadlock = true
   550  					println("runtime: failing to flush")
   551  					for mp := mToFlush; mp != nil; mp = mp.trace.link {
   552  						print("runtime: m=", mp.id, "\n")
   553  					}
   554  				}
   555  				i++
   556  			}
   557  		}
   558  	})
   559  
   560  	// At this point, the old generation is fully flushed minus stack and string
   561  	// tables, CPU samples, and goroutines that haven't run at all during the last
   562  	// generation.
   563  
   564  	// Check to see if any Gs still haven't had events written out for them.
   565  	statusWriter := unsafeTraceWriter(gen, nil)
   566  	for _, ug := range untracedGs {
   567  		if ug.gp.trace.statusWasTraced(gen) {
   568  			// It was traced, we don't need to do anything.
   569  			continue
   570  		}
   571  		// It still wasn't traced. Because we ensured all Ms stopped writing trace
   572  		// events to the last generation, that must mean the G never had its status
   573  		// traced in gen between when we recorded it and now. If that's true, the goid
   574  		// and status we recorded then is exactly what we want right now.
   575  		status := goStatusToTraceGoStatus(ug.status, ug.waitreason)
   576  		statusWriter = statusWriter.writeGoStatus(ug.goid, ug.mid, status, ug.inMarkAssist, ug.stackID)
   577  	}
   578  	statusWriter.flush().end()
   579  
   580  	// Read everything out of the last gen's CPU profile buffer.
   581  	traceReadCPU(gen)
   582  
   583  	// Flush CPU samples, stacks, and strings for the last generation. This is safe,
   584  	// because we're now certain no M is writing to the last generation.
   585  	//
   586  	// Ordering is important here. traceCPUFlush may generate new stacks and dumping
   587  	// stacks may generate new strings.
   588  	traceCPUFlush(gen)
   589  	trace.stackTab[gen%2].dump(gen)
   590  	trace.typeTab[gen%2].dump(gen)
   591  	trace.stringTab[gen%2].reset(gen)
   592  
   593  	// That's it. This generation is done producing buffers.
   594  	systemstack(func() {
   595  		lock(&trace.lock)
   596  		trace.flushedGen.Store(gen)
   597  		unlock(&trace.lock)
   598  	})
   599  
   600  	// Perform status reset on dead Ps because they just appear as idle.
   601  	//
   602  	// Preventing preemption is sufficient to access allp safely. allp is only
   603  	// mutated by GOMAXPROCS calls, which require a STW.
   604  	//
   605  	// TODO(mknyszek): Consider explicitly emitting ProcCreate and ProcDestroy
   606  	// events to indicate whether a P exists, rather than just making its
   607  	// existence implicit.
   608  	mp = acquirem()
   609  	for _, pp := range allp[len(allp):cap(allp)] {
   610  		pp.trace.readyNextGen(traceNextGen(gen))
   611  	}
   612  	releasem(mp)
   613  
   614  	if stopTrace {
   615  		// Acquire the shutdown sema to begin the shutdown process.
   616  		semacquire(&traceShutdownSema)
   617  
   618  		// Finish off CPU profile reading.
   619  		traceStopReadCPU()
   620  
   621  		// Reset debug.malloc if necessary. Note that this is set in a racy
   622  		// way; that's OK. Some mallocs may still enter into the debug.malloc
   623  		// block, but they won't generate events because tracing is disabled.
   624  		// That is, it's OK if mallocs read a stale debug.malloc or
   625  		// trace.enabledWithAllocFree value.
   626  		if trace.enabledWithAllocFree {
   627  			trace.enabledWithAllocFree = false
   628  			debug.malloc = trace.debugMalloc
   629  		}
   630  	} else {
   631  		// Go over each P and emit a status event for it if necessary.
   632  		//
   633  		// We do this at the beginning of the new generation instead of the
   634  		// end like we do for goroutines because forEachP doesn't give us a
   635  		// hook to skip Ps that have already been traced. Since we have to
   636  		// preempt all Ps anyway, might as well stay consistent with StartTrace
   637  		// which does this during the STW.
   638  		semacquire(&worldsema)
   639  		forEachP(waitReasonTraceProcStatus, func(pp *p) {
   640  			tl := traceAcquire()
   641  			if !pp.trace.statusWasTraced(tl.gen) {
   642  				tl.writer().writeProcStatusForP(pp, false).end()
   643  			}
   644  			traceRelease(tl)
   645  		})
   646  		semrelease(&worldsema)
   647  	}
   648  
   649  	// Block until the trace reader has finished processing the last generation.
   650  	semacquire(&trace.doneSema[gen%2])
   651  	if raceenabled {
   652  		raceacquire(unsafe.Pointer(&trace.doneSema[gen%2]))
   653  	}
   654  
   655  	// Double-check that things look as we expect after advancing and perform some
   656  	// final cleanup if the trace has fully stopped.
   657  	systemstack(func() {
   658  		lock(&trace.lock)
   659  		if !trace.full[gen%2].empty() {
   660  			throw("trace: non-empty full trace buffer for done generation")
   661  		}
   662  		if stopTrace {
   663  			if !trace.full[1-(gen%2)].empty() {
   664  				throw("trace: non-empty full trace buffer for next generation")
   665  			}
   666  			if trace.reading != nil || trace.reader.Load() != nil {
   667  				throw("trace: reading after shutdown")
   668  			}
   669  			// Free all the empty buffers.
   670  			for trace.empty != nil {
   671  				buf := trace.empty
   672  				trace.empty = buf.link
   673  				sysFree(unsafe.Pointer(buf), unsafe.Sizeof(*buf), &memstats.other_sys)
   674  			}
   675  			// Clear trace.shutdown and other flags.
   676  			trace.headerWritten = false
   677  			trace.shutdown.Store(false)
   678  		}
   679  		unlock(&trace.lock)
   680  	})
   681  
   682  	if stopTrace {
   683  		// Clear the sweep state on every P for the next time tracing is enabled.
   684  		//
   685  		// It may be stale in the next trace because we may have ended tracing in
   686  		// the middle of a sweep on a P.
   687  		//
   688  		// It's fine not to call forEachP here because tracing is disabled and we
   689  		// know at this point that nothing is calling into the tracer, but we do
   690  		// need to look at dead Ps too just because GOMAXPROCS could have been called
   691  		// at any point since we stopped tracing, and we have to ensure there's no
   692  		// bad state on dead Ps too. Prevent a STW and a concurrent GOMAXPROCS that
   693  		// might mutate allp by making ourselves briefly non-preemptible.
   694  		mp := acquirem()
   695  		for _, pp := range allp[:cap(allp)] {
   696  			pp.trace.inSweep = false
   697  			pp.trace.maySweep = false
   698  			pp.trace.swept = 0
   699  			pp.trace.reclaimed = 0
   700  		}
   701  		releasem(mp)
   702  	}
   703  
   704  	// Release the advance semaphore. If stopTrace is true we're still holding onto
   705  	// traceShutdownSema.
   706  	//
   707  	// Do a direct handoff. Don't let one caller of traceAdvance starve
   708  	// other calls to traceAdvance.
   709  	semrelease1(&traceAdvanceSema, true, 0)
   710  
   711  	if stopTrace {
   712  		// Stop the traceAdvancer. We can't be holding traceAdvanceSema here because
   713  		// we'll deadlock (we're blocked on the advancer goroutine exiting, but it
   714  		// may be currently trying to acquire traceAdvanceSema).
   715  		traceAdvancer.stop()
   716  		semrelease(&traceShutdownSema)
   717  	}
   718  }
   719  
   720  func traceNextGen(gen uintptr) uintptr {
   721  	if gen == ^uintptr(0) {
   722  		// gen is used both %2 and %3 and we want both patterns to continue when we loop around.
   723  		// ^uint32(0) and ^uint64(0) are both odd and multiples of 3. Therefore the next generation
   724  		// we want is even and one more than a multiple of 3. The smallest such number is 4.
   725  		return 4
   726  	}
   727  	return gen + 1
   728  }
   729  
   730  // traceRegisterLabelsAndReasons re-registers mark worker labels and
   731  // goroutine stop/block reasons in the string table for the provided
   732  // generation. Note: the provided generation must not have started yet.
   733  func traceRegisterLabelsAndReasons(gen uintptr) {
   734  	for i, label := range gcMarkWorkerModeStrings[:] {
   735  		trace.markWorkerLabels[gen%2][i] = traceArg(trace.stringTab[gen%2].put(gen, label))
   736  	}
   737  	for i, str := range traceBlockReasonStrings[:] {
   738  		trace.goBlockReasons[gen%2][i] = traceArg(trace.stringTab[gen%2].put(gen, str))
   739  	}
   740  	for i, str := range traceGoStopReasonStrings[:] {
   741  		trace.goStopReasons[gen%2][i] = traceArg(trace.stringTab[gen%2].put(gen, str))
   742  	}
   743  }
   744  
   745  // ReadTrace returns the next chunk of binary tracing data, blocking until data
   746  // is available. If tracing is turned off and all the data accumulated while it
   747  // was on has been returned, ReadTrace returns nil. The caller must copy the
   748  // returned data before calling ReadTrace again.
   749  // ReadTrace must be called from one goroutine at a time.
   750  func ReadTrace() []byte {
   751  top:
   752  	var buf []byte
   753  	var park bool
   754  	systemstack(func() {
   755  		buf, park = readTrace0()
   756  	})
   757  	if park {
   758  		gopark(func(gp *g, _ unsafe.Pointer) bool {
   759  			if !trace.reader.CompareAndSwapNoWB(nil, gp) {
   760  				// We're racing with another reader.
   761  				// Wake up and handle this case.
   762  				return false
   763  			}
   764  
   765  			if g2 := traceReader(); gp == g2 {
   766  				// New data arrived between unlocking
   767  				// and the CAS and we won the wake-up
   768  				// race, so wake up directly.
   769  				return false
   770  			} else if g2 != nil {
   771  				printlock()
   772  				println("runtime: got trace reader", g2, g2.goid)
   773  				throw("unexpected trace reader")
   774  			}
   775  
   776  			return true
   777  		}, nil, waitReasonTraceReaderBlocked, traceBlockSystemGoroutine, 2)
   778  		goto top
   779  	}
   780  
   781  	return buf
   782  }
   783  
   784  // readTrace0 is ReadTrace's continuation on g0. This must run on the
   785  // system stack because it acquires trace.lock.
   786  //
   787  //go:systemstack
   788  func readTrace0() (buf []byte, park bool) {
   789  	if raceenabled {
   790  		// g0 doesn't have a race context. Borrow the user G's.
   791  		if getg().racectx != 0 {
   792  			throw("expected racectx == 0")
   793  		}
   794  		getg().racectx = getg().m.curg.racectx
   795  		// (This defer should get open-coded, which is safe on
   796  		// the system stack.)
   797  		defer func() { getg().racectx = 0 }()
   798  	}
   799  
   800  	// This function must not allocate while holding trace.lock:
   801  	// allocation can call heap allocate, which will try to emit a trace
   802  	// event while holding heap lock.
   803  	lock(&trace.lock)
   804  
   805  	if trace.reader.Load() != nil {
   806  		// More than one goroutine reads trace. This is bad.
   807  		// But we rather do not crash the program because of tracing,
   808  		// because tracing can be enabled at runtime on prod servers.
   809  		unlock(&trace.lock)
   810  		println("runtime: ReadTrace called from multiple goroutines simultaneously")
   811  		return nil, false
   812  	}
   813  	// Recycle the old buffer.
   814  	if buf := trace.reading; buf != nil {
   815  		buf.link = trace.empty
   816  		trace.empty = buf
   817  		trace.reading = nil
   818  	}
   819  	// Write trace header.
   820  	if !trace.headerWritten {
   821  		trace.headerWritten = true
   822  		unlock(&trace.lock)
   823  		return []byte("go 1.23 trace\x00\x00\x00"), false
   824  	}
   825  
   826  	// Read the next buffer.
   827  
   828  	if trace.readerGen.Load() == 0 {
   829  		trace.readerGen.Store(1)
   830  	}
   831  	var gen uintptr
   832  	for {
   833  		assertLockHeld(&trace.lock)
   834  		gen = trace.readerGen.Load()
   835  
   836  		// Check to see if we need to block for more data in this generation
   837  		// or if we need to move our generation forward.
   838  		if !trace.full[gen%2].empty() {
   839  			break
   840  		}
   841  		// Most of the time readerGen is one generation ahead of flushedGen, as the
   842  		// current generation is being read from. Then, once the last buffer is flushed
   843  		// into readerGen, flushedGen will rise to meet it. At this point, the tracer
   844  		// is waiting on the reader to finish flushing the last generation so that it
   845  		// can continue to advance.
   846  		if trace.flushedGen.Load() == gen {
   847  			if trace.shutdown.Load() {
   848  				unlock(&trace.lock)
   849  
   850  				// Wake up anyone waiting for us to be done with this generation.
   851  				//
   852  				// Do this after reading trace.shutdown, because the thread we're
   853  				// waking up is going to clear trace.shutdown.
   854  				if raceenabled {
   855  					// Model synchronization on trace.doneSema, which te race
   856  					// detector does not see. This is required to avoid false
   857  					// race reports on writer passed to trace.Start.
   858  					racerelease(unsafe.Pointer(&trace.doneSema[gen%2]))
   859  				}
   860  				semrelease(&trace.doneSema[gen%2])
   861  
   862  				// We're shutting down, and the last generation is fully
   863  				// read. We're done.
   864  				return nil, false
   865  			}
   866  			// The previous gen has had all of its buffers flushed, and
   867  			// there's nothing else for us to read. Advance the generation
   868  			// we're reading from and try again.
   869  			trace.readerGen.Store(trace.gen.Load())
   870  			unlock(&trace.lock)
   871  
   872  			// Wake up anyone waiting for us to be done with this generation.
   873  			//
   874  			// Do this after reading gen to make sure we can't have the trace
   875  			// advance until we've read it.
   876  			if raceenabled {
   877  				// See comment above in the shutdown case.
   878  				racerelease(unsafe.Pointer(&trace.doneSema[gen%2]))
   879  			}
   880  			semrelease(&trace.doneSema[gen%2])
   881  
   882  			// Reacquire the lock and go back to the top of the loop.
   883  			lock(&trace.lock)
   884  			continue
   885  		}
   886  		// Wait for new data.
   887  		//
   888  		// We don't simply use a note because the scheduler
   889  		// executes this goroutine directly when it wakes up
   890  		// (also a note would consume an M).
   891  		//
   892  		// Before we drop the lock, clear the workAvailable flag. Work can
   893  		// only be queued with trace.lock held, so this is at least true until
   894  		// we drop the lock.
   895  		trace.workAvailable.Store(false)
   896  		unlock(&trace.lock)
   897  		return nil, true
   898  	}
   899  	// Pull a buffer.
   900  	tbuf := trace.full[gen%2].pop()
   901  	trace.reading = tbuf
   902  	unlock(&trace.lock)
   903  	return tbuf.arr[:tbuf.pos], false
   904  }
   905  
   906  // traceReader returns the trace reader that should be woken up, if any.
   907  // Callers should first check (traceEnabled() || traceShuttingDown()).
   908  //
   909  // This must run on the system stack because it acquires trace.lock.
   910  //
   911  //go:systemstack
   912  func traceReader() *g {
   913  	gp := traceReaderAvailable()
   914  	if gp == nil || !trace.reader.CompareAndSwapNoWB(gp, nil) {
   915  		return nil
   916  	}
   917  	return gp
   918  }
   919  
   920  // traceReaderAvailable returns the trace reader if it is not currently
   921  // scheduled and should be. Callers should first check that
   922  // (traceEnabled() || traceShuttingDown()) is true.
   923  func traceReaderAvailable() *g {
   924  	// There are three conditions under which we definitely want to schedule
   925  	// the reader:
   926  	// - The reader is lagging behind in finishing off the last generation.
   927  	//   In this case, trace buffers could even be empty, but the trace
   928  	//   advancer will be waiting on the reader, so we have to make sure
   929  	//   to schedule the reader ASAP.
   930  	// - The reader has pending work to process for it's reader generation
   931  	//   (assuming readerGen is not lagging behind). Note that we also want
   932  	//   to be careful *not* to schedule the reader if there's no work to do.
   933  	// - The trace is shutting down. The trace stopper blocks on the reader
   934  	//   to finish, much like trace advancement.
   935  	//
   936  	// We also want to be careful not to schedule the reader if there's no
   937  	// reason to.
   938  	if trace.flushedGen.Load() == trace.readerGen.Load() || trace.workAvailable.Load() || trace.shutdown.Load() {
   939  		return trace.reader.Load()
   940  	}
   941  	return nil
   942  }
   943  
   944  // Trace advancer goroutine.
   945  var traceAdvancer traceAdvancerState
   946  
   947  type traceAdvancerState struct {
   948  	timer *wakeableSleep
   949  	done  chan struct{}
   950  }
   951  
   952  // start starts a new traceAdvancer.
   953  func (s *traceAdvancerState) start() {
   954  	// Start a goroutine to periodically advance the trace generation.
   955  	s.done = make(chan struct{})
   956  	s.timer = newWakeableSleep()
   957  	go func() {
   958  		for traceEnabled() {
   959  			// Set a timer to wake us up
   960  			s.timer.sleep(int64(debug.traceadvanceperiod))
   961  
   962  			// Try to advance the trace.
   963  			traceAdvance(false)
   964  		}
   965  		s.done <- struct{}{}
   966  	}()
   967  }
   968  
   969  // stop stops a traceAdvancer and blocks until it exits.
   970  func (s *traceAdvancerState) stop() {
   971  	s.timer.wake()
   972  	<-s.done
   973  	close(s.done)
   974  	s.timer.close()
   975  }
   976  
   977  // traceAdvancePeriod is the approximate period between
   978  // new generations.
   979  const defaultTraceAdvancePeriod = 1e9 // 1 second.
   980  
   981  // wakeableSleep manages a wakeable goroutine sleep.
   982  //
   983  // Users of this type must call init before first use and
   984  // close to free up resources. Once close is called, init
   985  // must be called before another use.
   986  type wakeableSleep struct {
   987  	timer *timer
   988  
   989  	// lock protects access to wakeup, but not send/recv on it.
   990  	lock   mutex
   991  	wakeup chan struct{}
   992  }
   993  
   994  // newWakeableSleep initializes a new wakeableSleep and returns it.
   995  func newWakeableSleep() *wakeableSleep {
   996  	s := new(wakeableSleep)
   997  	lockInit(&s.lock, lockRankWakeableSleep)
   998  	s.wakeup = make(chan struct{}, 1)
   999  	s.timer = new(timer)
  1000  	f := func(s any, _ uintptr, _ int64) {
  1001  		s.(*wakeableSleep).wake()
  1002  	}
  1003  	s.timer.init(f, s)
  1004  	return s
  1005  }
  1006  
  1007  // sleep sleeps for the provided duration in nanoseconds or until
  1008  // another goroutine calls wake.
  1009  //
  1010  // Must not be called by more than one goroutine at a time and
  1011  // must not be called concurrently with close.
  1012  func (s *wakeableSleep) sleep(ns int64) {
  1013  	s.timer.reset(nanotime()+ns, 0)
  1014  	lock(&s.lock)
  1015  	if raceenabled {
  1016  		raceacquire(unsafe.Pointer(&s.lock))
  1017  	}
  1018  	wakeup := s.wakeup
  1019  	if raceenabled {
  1020  		racerelease(unsafe.Pointer(&s.lock))
  1021  	}
  1022  	unlock(&s.lock)
  1023  	<-wakeup
  1024  	s.timer.stop()
  1025  }
  1026  
  1027  // wake awakens any goroutine sleeping on the timer.
  1028  //
  1029  // Safe for concurrent use with all other methods.
  1030  func (s *wakeableSleep) wake() {
  1031  	// Grab the wakeup channel, which may be nil if we're
  1032  	// racing with close.
  1033  	lock(&s.lock)
  1034  	if raceenabled {
  1035  		raceacquire(unsafe.Pointer(&s.lock))
  1036  	}
  1037  	if s.wakeup != nil {
  1038  		// Non-blocking send.
  1039  		//
  1040  		// Others may also write to this channel and we don't
  1041  		// want to block on the receiver waking up. This also
  1042  		// effectively batches together wakeup notifications.
  1043  		select {
  1044  		case s.wakeup <- struct{}{}:
  1045  		default:
  1046  		}
  1047  	}
  1048  	if raceenabled {
  1049  		racerelease(unsafe.Pointer(&s.lock))
  1050  	}
  1051  	unlock(&s.lock)
  1052  }
  1053  
  1054  // close wakes any goroutine sleeping on the timer and prevents
  1055  // further sleeping on it.
  1056  //
  1057  // Once close is called, the wakeableSleep must no longer be used.
  1058  //
  1059  // It must only be called once no goroutine is sleeping on the
  1060  // timer *and* nothing else will call wake concurrently.
  1061  func (s *wakeableSleep) close() {
  1062  	// Set wakeup to nil so that a late timer ends up being a no-op.
  1063  	lock(&s.lock)
  1064  	if raceenabled {
  1065  		raceacquire(unsafe.Pointer(&s.lock))
  1066  	}
  1067  	wakeup := s.wakeup
  1068  	s.wakeup = nil
  1069  
  1070  	// Close the channel.
  1071  	close(wakeup)
  1072  
  1073  	if raceenabled {
  1074  		racerelease(unsafe.Pointer(&s.lock))
  1075  	}
  1076  	unlock(&s.lock)
  1077  	return
  1078  }
  1079  

View as plain text