Source file src/runtime/mgc.go

     1  // Copyright 2009 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  // Garbage collector (GC).
     6  //
     7  // The GC runs concurrently with mutator threads, is type accurate (aka precise), allows multiple
     8  // GC thread to run in parallel. It is a concurrent mark and sweep that uses a write barrier. It is
     9  // non-generational and non-compacting. Allocation is done using size segregated per P allocation
    10  // areas to minimize fragmentation while eliminating locks in the common case.
    11  //
    12  // The algorithm decomposes into several steps.
    13  // This is a high level description of the algorithm being used. For an overview of GC a good
    14  // place to start is Richard Jones' gchandbook.org.
    15  //
    16  // The algorithm's intellectual heritage includes Dijkstra's on-the-fly algorithm, see
    17  // Edsger W. Dijkstra, Leslie Lamport, A. J. Martin, C. S. Scholten, and E. F. M. Steffens. 1978.
    18  // On-the-fly garbage collection: an exercise in cooperation. Commun. ACM 21, 11 (November 1978),
    19  // 966-975.
    20  // For journal quality proofs that these steps are complete, correct, and terminate see
    21  // Hudson, R., and Moss, J.E.B. Copying Garbage Collection without stopping the world.
    22  // Concurrency and Computation: Practice and Experience 15(3-5), 2003.
    23  //
    24  // 1. GC performs sweep termination.
    25  //
    26  //    a. Stop the world. This causes all Ps to reach a GC safe-point.
    27  //
    28  //    b. Sweep any unswept spans. There will only be unswept spans if
    29  //    this GC cycle was forced before the expected time.
    30  //
    31  // 2. GC performs the mark phase.
    32  //
    33  //    a. Prepare for the mark phase by setting gcphase to _GCmark
    34  //    (from _GCoff), enabling the write barrier, enabling mutator
    35  //    assists, and enqueueing root mark jobs. No objects may be
    36  //    scanned until all Ps have enabled the write barrier, which is
    37  //    accomplished using STW.
    38  //
    39  //    b. Start the world. From this point, GC work is done by mark
    40  //    workers started by the scheduler and by assists performed as
    41  //    part of allocation. The write barrier shades both the
    42  //    overwritten pointer and the new pointer value for any pointer
    43  //    writes (see mbarrier.go for details). Newly allocated objects
    44  //    are immediately marked black.
    45  //
    46  //    c. GC performs root marking jobs. This includes scanning all
    47  //    stacks, shading all globals, and shading any heap pointers in
    48  //    off-heap runtime data structures. Scanning a stack stops a
    49  //    goroutine, shades any pointers found on its stack, and then
    50  //    resumes the goroutine.
    51  //
    52  //    d. GC drains the work queue of grey objects, scanning each grey
    53  //    object to black and shading all pointers found in the object
    54  //    (which in turn may add those pointers to the work queue).
    55  //
    56  //    e. Because GC work is spread across local caches, GC uses a
    57  //    distributed termination algorithm to detect when there are no
    58  //    more root marking jobs or grey objects (see gcMarkDone). At this
    59  //    point, GC transitions to mark termination.
    60  //
    61  // 3. GC performs mark termination.
    62  //
    63  //    a. Stop the world.
    64  //
    65  //    b. Set gcphase to _GCmarktermination, and disable workers and
    66  //    assists.
    67  //
    68  //    c. Perform housekeeping like flushing mcaches.
    69  //
    70  // 4. GC performs the sweep phase.
    71  //
    72  //    a. Prepare for the sweep phase by setting gcphase to _GCoff,
    73  //    setting up sweep state and disabling the write barrier.
    74  //
    75  //    b. Start the world. From this point on, newly allocated objects
    76  //    are white, and allocating sweeps spans before use if necessary.
    77  //
    78  //    c. GC does concurrent sweeping in the background and in response
    79  //    to allocation. See description below.
    80  //
    81  // 5. When sufficient allocation has taken place, replay the sequence
    82  // starting with 1 above. See discussion of GC rate below.
    83  
    84  // Concurrent sweep.
    85  //
    86  // The sweep phase proceeds concurrently with normal program execution.
    87  // The heap is swept span-by-span both lazily (when a goroutine needs another span)
    88  // and concurrently in a background goroutine (this helps programs that are not CPU bound).
    89  // At the end of STW mark termination all spans are marked as "needs sweeping".
    90  //
    91  // The background sweeper goroutine simply sweeps spans one-by-one.
    92  //
    93  // To avoid requesting more OS memory while there are unswept spans, when a
    94  // goroutine needs another span, it first attempts to reclaim that much memory
    95  // by sweeping. When a goroutine needs to allocate a new small-object span, it
    96  // sweeps small-object spans for the same object size until it frees at least
    97  // one object. When a goroutine needs to allocate large-object span from heap,
    98  // it sweeps spans until it frees at least that many pages into heap. There is
    99  // one case where this may not suffice: if a goroutine sweeps and frees two
   100  // nonadjacent one-page spans to the heap, it will allocate a new two-page
   101  // span, but there can still be other one-page unswept spans which could be
   102  // combined into a two-page span.
   103  //
   104  // It's critical to ensure that no operations proceed on unswept spans (that would corrupt
   105  // mark bits in GC bitmap). During GC all mcaches are flushed into the central cache,
   106  // so they are empty. When a goroutine grabs a new span into mcache, it sweeps it.
   107  // When a goroutine explicitly frees an object or sets a finalizer, it ensures that
   108  // the span is swept (either by sweeping it, or by waiting for the concurrent sweep to finish).
   109  // The finalizer goroutine is kicked off only when all spans are swept.
   110  // When the next GC starts, it sweeps all not-yet-swept spans (if any).
   111  
   112  // GC rate.
   113  // Next GC is after we've allocated an extra amount of memory proportional to
   114  // the amount already in use. The proportion is controlled by GOGC environment variable
   115  // (100 by default). If GOGC=100 and we're using 4M, we'll GC again when we get to 8M
   116  // (this mark is computed by the gcController.heapGoal method). This keeps the GC cost in
   117  // linear proportion to the allocation cost. Adjusting GOGC just changes the linear constant
   118  // (and also the amount of extra memory used).
   119  
   120  // Oblets
   121  //
   122  // In order to prevent long pauses while scanning large objects and to
   123  // improve parallelism, the garbage collector breaks up scan jobs for
   124  // objects larger than maxObletBytes into "oblets" of at most
   125  // maxObletBytes. When scanning encounters the beginning of a large
   126  // object, it scans only the first oblet and enqueues the remaining
   127  // oblets as new scan jobs.
   128  
   129  package runtime
   130  
   131  import (
   132  	"internal/cpu"
   133  	"runtime/internal/atomic"
   134  	"unsafe"
   135  )
   136  
   137  const (
   138  	_DebugGC         = 0
   139  	_ConcurrentSweep = true
   140  	_FinBlockSize    = 4 * 1024
   141  
   142  	// debugScanConservative enables debug logging for stack
   143  	// frames that are scanned conservatively.
   144  	debugScanConservative = false
   145  
   146  	// sweepMinHeapDistance is a lower bound on the heap distance
   147  	// (in bytes) reserved for concurrent sweeping between GC
   148  	// cycles.
   149  	sweepMinHeapDistance = 1024 * 1024
   150  )
   151  
   152  func gcinit() {
   153  	if unsafe.Sizeof(workbuf{}) != _WorkbufSize {
   154  		throw("size of Workbuf is suboptimal")
   155  	}
   156  	// No sweep on the first cycle.
   157  	sweep.active.state.Store(sweepDrainedMask)
   158  
   159  	// Initialize GC pacer state.
   160  	// Use the environment variable GOGC for the initial gcPercent value.
   161  	// Use the environment variable GOMEMLIMIT for the initial memoryLimit value.
   162  	gcController.init(readGOGC(), readGOMEMLIMIT())
   163  
   164  	work.startSema = 1
   165  	work.markDoneSema = 1
   166  	lockInit(&work.sweepWaiters.lock, lockRankSweepWaiters)
   167  	lockInit(&work.assistQueue.lock, lockRankAssistQueue)
   168  	lockInit(&work.wbufSpans.lock, lockRankWbufSpans)
   169  }
   170  
   171  // gcenable is called after the bulk of the runtime initialization,
   172  // just before we're about to start letting user code run.
   173  // It kicks off the background sweeper goroutine, the background
   174  // scavenger goroutine, and enables GC.
   175  func gcenable() {
   176  	// Kick off sweeping and scavenging.
   177  	c := make(chan int, 2)
   178  	go bgsweep(c)
   179  	go bgscavenge(c)
   180  	<-c
   181  	<-c
   182  	memstats.enablegc = true // now that runtime is initialized, GC is okay
   183  }
   184  
   185  // Garbage collector phase.
   186  // Indicates to write barrier and synchronization task to perform.
   187  var gcphase uint32
   188  
   189  // The compiler knows about this variable.
   190  // If you change it, you must change builtin/runtime.go, too.
   191  // If you change the first four bytes, you must also change the write
   192  // barrier insertion code.
   193  var writeBarrier struct {
   194  	enabled bool    // compiler emits a check of this before calling write barrier
   195  	pad     [3]byte // compiler uses 32-bit load for "enabled" field
   196  	needed  bool    // whether we need a write barrier for current GC phase
   197  	cgo     bool    // whether we need a write barrier for a cgo check
   198  	alignme uint64  // guarantee alignment so that compiler can use a 32 or 64-bit load
   199  }
   200  
   201  // gcBlackenEnabled is 1 if mutator assists and background mark
   202  // workers are allowed to blacken objects. This must only be set when
   203  // gcphase == _GCmark.
   204  var gcBlackenEnabled uint32
   205  
   206  const (
   207  	_GCoff             = iota // GC not running; sweeping in background, write barrier disabled
   208  	_GCmark                   // GC marking roots and workbufs: allocate black, write barrier ENABLED
   209  	_GCmarktermination        // GC mark termination: allocate black, P's help GC, write barrier ENABLED
   210  )
   211  
   212  //go:nosplit
   213  func setGCPhase(x uint32) {
   214  	atomic.Store(&gcphase, x)
   215  	writeBarrier.needed = gcphase == _GCmark || gcphase == _GCmarktermination
   216  	writeBarrier.enabled = writeBarrier.needed || writeBarrier.cgo
   217  }
   218  
   219  // gcMarkWorkerMode represents the mode that a concurrent mark worker
   220  // should operate in.
   221  //
   222  // Concurrent marking happens through four different mechanisms. One
   223  // is mutator assists, which happen in response to allocations and are
   224  // not scheduled. The other three are variations in the per-P mark
   225  // workers and are distinguished by gcMarkWorkerMode.
   226  type gcMarkWorkerMode int
   227  
   228  const (
   229  	// gcMarkWorkerNotWorker indicates that the next scheduled G is not
   230  	// starting work and the mode should be ignored.
   231  	gcMarkWorkerNotWorker gcMarkWorkerMode = iota
   232  
   233  	// gcMarkWorkerDedicatedMode indicates that the P of a mark
   234  	// worker is dedicated to running that mark worker. The mark
   235  	// worker should run without preemption.
   236  	gcMarkWorkerDedicatedMode
   237  
   238  	// gcMarkWorkerFractionalMode indicates that a P is currently
   239  	// running the "fractional" mark worker. The fractional worker
   240  	// is necessary when GOMAXPROCS*gcBackgroundUtilization is not
   241  	// an integer and using only dedicated workers would result in
   242  	// utilization too far from the target of gcBackgroundUtilization.
   243  	// The fractional worker should run until it is preempted and
   244  	// will be scheduled to pick up the fractional part of
   245  	// GOMAXPROCS*gcBackgroundUtilization.
   246  	gcMarkWorkerFractionalMode
   247  
   248  	// gcMarkWorkerIdleMode indicates that a P is running the mark
   249  	// worker because it has nothing else to do. The idle worker
   250  	// should run until it is preempted and account its time
   251  	// against gcController.idleMarkTime.
   252  	gcMarkWorkerIdleMode
   253  )
   254  
   255  // gcMarkWorkerModeStrings are the strings labels of gcMarkWorkerModes
   256  // to use in execution traces.
   257  var gcMarkWorkerModeStrings = [...]string{
   258  	"Not worker",
   259  	"GC (dedicated)",
   260  	"GC (fractional)",
   261  	"GC (idle)",
   262  }
   263  
   264  // pollFractionalWorkerExit reports whether a fractional mark worker
   265  // should self-preempt. It assumes it is called from the fractional
   266  // worker.
   267  func pollFractionalWorkerExit() bool {
   268  	// This should be kept in sync with the fractional worker
   269  	// scheduler logic in findRunnableGCWorker.
   270  	now := nanotime()
   271  	delta := now - gcController.markStartTime
   272  	if delta <= 0 {
   273  		return true
   274  	}
   275  	p := getg().m.p.ptr()
   276  	selfTime := p.gcFractionalMarkTime + (now - p.gcMarkWorkerStartTime)
   277  	// Add some slack to the utilization goal so that the
   278  	// fractional worker isn't behind again the instant it exits.
   279  	return float64(selfTime)/float64(delta) > 1.2*gcController.fractionalUtilizationGoal
   280  }
   281  
   282  var work workType
   283  
   284  type workType struct {
   285  	full  lfstack          // lock-free list of full blocks workbuf
   286  	empty lfstack          // lock-free list of empty blocks workbuf
   287  	pad0  cpu.CacheLinePad // prevents false-sharing between full/empty and nproc/nwait
   288  
   289  	wbufSpans struct {
   290  		lock mutex
   291  		// free is a list of spans dedicated to workbufs, but
   292  		// that don't currently contain any workbufs.
   293  		free mSpanList
   294  		// busy is a list of all spans containing workbufs on
   295  		// one of the workbuf lists.
   296  		busy mSpanList
   297  	}
   298  
   299  	// Restore 64-bit alignment on 32-bit.
   300  	_ uint32
   301  
   302  	// bytesMarked is the number of bytes marked this cycle. This
   303  	// includes bytes blackened in scanned objects, noscan objects
   304  	// that go straight to black, and permagrey objects scanned by
   305  	// markroot during the concurrent scan phase. This is updated
   306  	// atomically during the cycle. Updates may be batched
   307  	// arbitrarily, since the value is only read at the end of the
   308  	// cycle.
   309  	//
   310  	// Because of benign races during marking, this number may not
   311  	// be the exact number of marked bytes, but it should be very
   312  	// close.
   313  	//
   314  	// Put this field here because it needs 64-bit atomic access
   315  	// (and thus 8-byte alignment even on 32-bit architectures).
   316  	bytesMarked uint64
   317  
   318  	markrootNext uint32 // next markroot job
   319  	markrootJobs uint32 // number of markroot jobs
   320  
   321  	nproc  uint32
   322  	tstart int64
   323  	nwait  uint32
   324  
   325  	// Number of roots of various root types. Set by gcMarkRootPrepare.
   326  	//
   327  	// nStackRoots == len(stackRoots), but we have nStackRoots for
   328  	// consistency.
   329  	nDataRoots, nBSSRoots, nSpanRoots, nStackRoots int
   330  
   331  	// Base indexes of each root type. Set by gcMarkRootPrepare.
   332  	baseData, baseBSS, baseSpans, baseStacks, baseEnd uint32
   333  
   334  	// stackRoots is a snapshot of all of the Gs that existed
   335  	// before the beginning of concurrent marking. The backing
   336  	// store of this must not be modified because it might be
   337  	// shared with allgs.
   338  	stackRoots []*g
   339  
   340  	// Each type of GC state transition is protected by a lock.
   341  	// Since multiple threads can simultaneously detect the state
   342  	// transition condition, any thread that detects a transition
   343  	// condition must acquire the appropriate transition lock,
   344  	// re-check the transition condition and return if it no
   345  	// longer holds or perform the transition if it does.
   346  	// Likewise, any transition must invalidate the transition
   347  	// condition before releasing the lock. This ensures that each
   348  	// transition is performed by exactly one thread and threads
   349  	// that need the transition to happen block until it has
   350  	// happened.
   351  	//
   352  	// startSema protects the transition from "off" to mark or
   353  	// mark termination.
   354  	startSema uint32
   355  	// markDoneSema protects transitions from mark to mark termination.
   356  	markDoneSema uint32
   357  
   358  	bgMarkReady note   // signal background mark worker has started
   359  	bgMarkDone  uint32 // cas to 1 when at a background mark completion point
   360  	// Background mark completion signaling
   361  
   362  	// mode is the concurrency mode of the current GC cycle.
   363  	mode gcMode
   364  
   365  	// userForced indicates the current GC cycle was forced by an
   366  	// explicit user call.
   367  	userForced bool
   368  
   369  	// initialHeapLive is the value of gcController.heapLive at the
   370  	// beginning of this GC cycle.
   371  	initialHeapLive uint64
   372  
   373  	// assistQueue is a queue of assists that are blocked because
   374  	// there was neither enough credit to steal or enough work to
   375  	// do.
   376  	assistQueue struct {
   377  		lock mutex
   378  		q    gQueue
   379  	}
   380  
   381  	// sweepWaiters is a list of blocked goroutines to wake when
   382  	// we transition from mark termination to sweep.
   383  	sweepWaiters struct {
   384  		lock mutex
   385  		list gList
   386  	}
   387  
   388  	// cycles is the number of completed GC cycles, where a GC
   389  	// cycle is sweep termination, mark, mark termination, and
   390  	// sweep. This differs from memstats.numgc, which is
   391  	// incremented at mark termination.
   392  	cycles atomic.Uint32
   393  
   394  	// Timing/utilization stats for this cycle.
   395  	stwprocs, maxprocs                 int32
   396  	tSweepTerm, tMark, tMarkTerm, tEnd int64 // nanotime() of phase start
   397  
   398  	pauseNS    int64 // total STW time this cycle
   399  	pauseStart int64 // nanotime() of last STW
   400  
   401  	// debug.gctrace heap sizes for this cycle.
   402  	heap0, heap1, heap2 uint64
   403  
   404  	// Cumulative estimated CPU usage.
   405  	cpuStats
   406  }
   407  
   408  // GC runs a garbage collection and blocks the caller until the
   409  // garbage collection is complete. It may also block the entire
   410  // program.
   411  func GC() {
   412  	// We consider a cycle to be: sweep termination, mark, mark
   413  	// termination, and sweep. This function shouldn't return
   414  	// until a full cycle has been completed, from beginning to
   415  	// end. Hence, we always want to finish up the current cycle
   416  	// and start a new one. That means:
   417  	//
   418  	// 1. In sweep termination, mark, or mark termination of cycle
   419  	// N, wait until mark termination N completes and transitions
   420  	// to sweep N.
   421  	//
   422  	// 2. In sweep N, help with sweep N.
   423  	//
   424  	// At this point we can begin a full cycle N+1.
   425  	//
   426  	// 3. Trigger cycle N+1 by starting sweep termination N+1.
   427  	//
   428  	// 4. Wait for mark termination N+1 to complete.
   429  	//
   430  	// 5. Help with sweep N+1 until it's done.
   431  	//
   432  	// This all has to be written to deal with the fact that the
   433  	// GC may move ahead on its own. For example, when we block
   434  	// until mark termination N, we may wake up in cycle N+2.
   435  
   436  	// Wait until the current sweep termination, mark, and mark
   437  	// termination complete.
   438  	n := work.cycles.Load()
   439  	gcWaitOnMark(n)
   440  
   441  	// We're now in sweep N or later. Trigger GC cycle N+1, which
   442  	// will first finish sweep N if necessary and then enter sweep
   443  	// termination N+1.
   444  	gcStart(gcTrigger{kind: gcTriggerCycle, n: n + 1})
   445  
   446  	// Wait for mark termination N+1 to complete.
   447  	gcWaitOnMark(n + 1)
   448  
   449  	// Finish sweep N+1 before returning. We do this both to
   450  	// complete the cycle and because runtime.GC() is often used
   451  	// as part of tests and benchmarks to get the system into a
   452  	// relatively stable and isolated state.
   453  	for work.cycles.Load() == n+1 && sweepone() != ^uintptr(0) {
   454  		sweep.nbgsweep++
   455  		Gosched()
   456  	}
   457  
   458  	// Callers may assume that the heap profile reflects the
   459  	// just-completed cycle when this returns (historically this
   460  	// happened because this was a STW GC), but right now the
   461  	// profile still reflects mark termination N, not N+1.
   462  	//
   463  	// As soon as all of the sweep frees from cycle N+1 are done,
   464  	// we can go ahead and publish the heap profile.
   465  	//
   466  	// First, wait for sweeping to finish. (We know there are no
   467  	// more spans on the sweep queue, but we may be concurrently
   468  	// sweeping spans, so we have to wait.)
   469  	for work.cycles.Load() == n+1 && !isSweepDone() {
   470  		Gosched()
   471  	}
   472  
   473  	// Now we're really done with sweeping, so we can publish the
   474  	// stable heap profile. Only do this if we haven't already hit
   475  	// another mark termination.
   476  	mp := acquirem()
   477  	cycle := work.cycles.Load()
   478  	if cycle == n+1 || (gcphase == _GCmark && cycle == n+2) {
   479  		mProf_PostSweep()
   480  	}
   481  	releasem(mp)
   482  }
   483  
   484  // gcWaitOnMark blocks until GC finishes the Nth mark phase. If GC has
   485  // already completed this mark phase, it returns immediately.
   486  func gcWaitOnMark(n uint32) {
   487  	for {
   488  		// Disable phase transitions.
   489  		lock(&work.sweepWaiters.lock)
   490  		nMarks := work.cycles.Load()
   491  		if gcphase != _GCmark {
   492  			// We've already completed this cycle's mark.
   493  			nMarks++
   494  		}
   495  		if nMarks > n {
   496  			// We're done.
   497  			unlock(&work.sweepWaiters.lock)
   498  			return
   499  		}
   500  
   501  		// Wait until sweep termination, mark, and mark
   502  		// termination of cycle N complete.
   503  		work.sweepWaiters.list.push(getg())
   504  		goparkunlock(&work.sweepWaiters.lock, waitReasonWaitForGCCycle, traceEvGoBlock, 1)
   505  	}
   506  }
   507  
   508  // gcMode indicates how concurrent a GC cycle should be.
   509  type gcMode int
   510  
   511  const (
   512  	gcBackgroundMode gcMode = iota // concurrent GC and sweep
   513  	gcForceMode                    // stop-the-world GC now, concurrent sweep
   514  	gcForceBlockMode               // stop-the-world GC now and STW sweep (forced by user)
   515  )
   516  
   517  // A gcTrigger is a predicate for starting a GC cycle. Specifically,
   518  // it is an exit condition for the _GCoff phase.
   519  type gcTrigger struct {
   520  	kind gcTriggerKind
   521  	now  int64  // gcTriggerTime: current time
   522  	n    uint32 // gcTriggerCycle: cycle number to start
   523  }
   524  
   525  type gcTriggerKind int
   526  
   527  const (
   528  	// gcTriggerHeap indicates that a cycle should be started when
   529  	// the heap size reaches the trigger heap size computed by the
   530  	// controller.
   531  	gcTriggerHeap gcTriggerKind = iota
   532  
   533  	// gcTriggerTime indicates that a cycle should be started when
   534  	// it's been more than forcegcperiod nanoseconds since the
   535  	// previous GC cycle.
   536  	gcTriggerTime
   537  
   538  	// gcTriggerCycle indicates that a cycle should be started if
   539  	// we have not yet started cycle number gcTrigger.n (relative
   540  	// to work.cycles).
   541  	gcTriggerCycle
   542  )
   543  
   544  // test reports whether the trigger condition is satisfied, meaning
   545  // that the exit condition for the _GCoff phase has been met. The exit
   546  // condition should be tested when allocating.
   547  func (t gcTrigger) test() bool {
   548  	if !memstats.enablegc || panicking.Load() != 0 || gcphase != _GCoff {
   549  		return false
   550  	}
   551  	switch t.kind {
   552  	case gcTriggerHeap:
   553  		// Non-atomic access to gcController.heapLive for performance. If
   554  		// we are going to trigger on this, this thread just
   555  		// atomically wrote gcController.heapLive anyway and we'll see our
   556  		// own write.
   557  		trigger, _ := gcController.trigger()
   558  		return gcController.heapLive.Load() >= trigger
   559  	case gcTriggerTime:
   560  		if gcController.gcPercent.Load() < 0 {
   561  			return false
   562  		}
   563  		lastgc := int64(atomic.Load64(&memstats.last_gc_nanotime))
   564  		return lastgc != 0 && t.now-lastgc > forcegcperiod
   565  	case gcTriggerCycle:
   566  		// t.n > work.cycles, but accounting for wraparound.
   567  		return int32(t.n-work.cycles.Load()) > 0
   568  	}
   569  	return true
   570  }
   571  
   572  // gcStart starts the GC. It transitions from _GCoff to _GCmark (if
   573  // debug.gcstoptheworld == 0) or performs all of GC (if
   574  // debug.gcstoptheworld != 0).
   575  //
   576  // This may return without performing this transition in some cases,
   577  // such as when called on a system stack or with locks held.
   578  func gcStart(trigger gcTrigger) {
   579  	// Since this is called from malloc and malloc is called in
   580  	// the guts of a number of libraries that might be holding
   581  	// locks, don't attempt to start GC in non-preemptible or
   582  	// potentially unstable situations.
   583  	mp := acquirem()
   584  	if gp := getg(); gp == mp.g0 || mp.locks > 1 || mp.preemptoff != "" {
   585  		releasem(mp)
   586  		return
   587  	}
   588  	releasem(mp)
   589  	mp = nil
   590  
   591  	// Pick up the remaining unswept/not being swept spans concurrently
   592  	//
   593  	// This shouldn't happen if we're being invoked in background
   594  	// mode since proportional sweep should have just finished
   595  	// sweeping everything, but rounding errors, etc, may leave a
   596  	// few spans unswept. In forced mode, this is necessary since
   597  	// GC can be forced at any point in the sweeping cycle.
   598  	//
   599  	// We check the transition condition continuously here in case
   600  	// this G gets delayed in to the next GC cycle.
   601  	for trigger.test() && sweepone() != ^uintptr(0) {
   602  		sweep.nbgsweep++
   603  	}
   604  
   605  	// Perform GC initialization and the sweep termination
   606  	// transition.
   607  	semacquire(&work.startSema)
   608  	// Re-check transition condition under transition lock.
   609  	if !trigger.test() {
   610  		semrelease(&work.startSema)
   611  		return
   612  	}
   613  
   614  	// In gcstoptheworld debug mode, upgrade the mode accordingly.
   615  	// We do this after re-checking the transition condition so
   616  	// that multiple goroutines that detect the heap trigger don't
   617  	// start multiple STW GCs.
   618  	mode := gcBackgroundMode
   619  	if debug.gcstoptheworld == 1 {
   620  		mode = gcForceMode
   621  	} else if debug.gcstoptheworld == 2 {
   622  		mode = gcForceBlockMode
   623  	}
   624  
   625  	// Ok, we're doing it! Stop everybody else
   626  	semacquire(&gcsema)
   627  	semacquire(&worldsema)
   628  
   629  	// For stats, check if this GC was forced by the user.
   630  	// Update it under gcsema to avoid gctrace getting wrong values.
   631  	work.userForced = trigger.kind == gcTriggerCycle
   632  
   633  	if trace.enabled {
   634  		traceGCStart()
   635  	}
   636  
   637  	// Check that all Ps have finished deferred mcache flushes.
   638  	for _, p := range allp {
   639  		if fg := p.mcache.flushGen.Load(); fg != mheap_.sweepgen {
   640  			println("runtime: p", p.id, "flushGen", fg, "!= sweepgen", mheap_.sweepgen)
   641  			throw("p mcache not flushed")
   642  		}
   643  	}
   644  
   645  	gcBgMarkStartWorkers()
   646  
   647  	systemstack(gcResetMarkState)
   648  
   649  	work.stwprocs, work.maxprocs = gomaxprocs, gomaxprocs
   650  	if work.stwprocs > ncpu {
   651  		// This is used to compute CPU time of the STW phases,
   652  		// so it can't be more than ncpu, even if GOMAXPROCS is.
   653  		work.stwprocs = ncpu
   654  	}
   655  	work.heap0 = gcController.heapLive.Load()
   656  	work.pauseNS = 0
   657  	work.mode = mode
   658  
   659  	now := nanotime()
   660  	work.tSweepTerm = now
   661  	work.pauseStart = now
   662  	if trace.enabled {
   663  		traceGCSTWStart(1)
   664  	}
   665  	systemstack(stopTheWorldWithSema)
   666  	// Finish sweep before we start concurrent scan.
   667  	systemstack(func() {
   668  		finishsweep_m()
   669  	})
   670  
   671  	// clearpools before we start the GC. If we wait they memory will not be
   672  	// reclaimed until the next GC cycle.
   673  	clearpools()
   674  
   675  	work.cycles.Add(1)
   676  
   677  	// Assists and workers can start the moment we start
   678  	// the world.
   679  	gcController.startCycle(now, int(gomaxprocs), trigger)
   680  
   681  	// Notify the CPU limiter that assists may begin.
   682  	gcCPULimiter.startGCTransition(true, now)
   683  
   684  	// In STW mode, disable scheduling of user Gs. This may also
   685  	// disable scheduling of this goroutine, so it may block as
   686  	// soon as we start the world again.
   687  	if mode != gcBackgroundMode {
   688  		schedEnableUser(false)
   689  	}
   690  
   691  	// Enter concurrent mark phase and enable
   692  	// write barriers.
   693  	//
   694  	// Because the world is stopped, all Ps will
   695  	// observe that write barriers are enabled by
   696  	// the time we start the world and begin
   697  	// scanning.
   698  	//
   699  	// Write barriers must be enabled before assists are
   700  	// enabled because they must be enabled before
   701  	// any non-leaf heap objects are marked. Since
   702  	// allocations are blocked until assists can
   703  	// happen, we want enable assists as early as
   704  	// possible.
   705  	setGCPhase(_GCmark)
   706  
   707  	gcBgMarkPrepare() // Must happen before assist enable.
   708  	gcMarkRootPrepare()
   709  
   710  	// Mark all active tinyalloc blocks. Since we're
   711  	// allocating from these, they need to be black like
   712  	// other allocations. The alternative is to blacken
   713  	// the tiny block on every allocation from it, which
   714  	// would slow down the tiny allocator.
   715  	gcMarkTinyAllocs()
   716  
   717  	// At this point all Ps have enabled the write
   718  	// barrier, thus maintaining the no white to
   719  	// black invariant. Enable mutator assists to
   720  	// put back-pressure on fast allocating
   721  	// mutators.
   722  	atomic.Store(&gcBlackenEnabled, 1)
   723  
   724  	// In STW mode, we could block the instant systemstack
   725  	// returns, so make sure we're not preemptible.
   726  	mp = acquirem()
   727  
   728  	// Concurrent mark.
   729  	systemstack(func() {
   730  		now = startTheWorldWithSema(trace.enabled)
   731  		work.pauseNS += now - work.pauseStart
   732  		work.tMark = now
   733  		memstats.gcPauseDist.record(now - work.pauseStart)
   734  
   735  		// Release the CPU limiter.
   736  		gcCPULimiter.finishGCTransition(now)
   737  	})
   738  
   739  	// Release the world sema before Gosched() in STW mode
   740  	// because we will need to reacquire it later but before
   741  	// this goroutine becomes runnable again, and we could
   742  	// self-deadlock otherwise.
   743  	semrelease(&worldsema)
   744  	releasem(mp)
   745  
   746  	// Make sure we block instead of returning to user code
   747  	// in STW mode.
   748  	if mode != gcBackgroundMode {
   749  		Gosched()
   750  	}
   751  
   752  	semrelease(&work.startSema)
   753  }
   754  
   755  // gcMarkDoneFlushed counts the number of P's with flushed work.
   756  //
   757  // Ideally this would be a captured local in gcMarkDone, but forEachP
   758  // escapes its callback closure, so it can't capture anything.
   759  //
   760  // This is protected by markDoneSema.
   761  var gcMarkDoneFlushed uint32
   762  
   763  // gcMarkDone transitions the GC from mark to mark termination if all
   764  // reachable objects have been marked (that is, there are no grey
   765  // objects and can be no more in the future). Otherwise, it flushes
   766  // all local work to the global queues where it can be discovered by
   767  // other workers.
   768  //
   769  // This should be called when all local mark work has been drained and
   770  // there are no remaining workers. Specifically, when
   771  //
   772  //	work.nwait == work.nproc && !gcMarkWorkAvailable(p)
   773  //
   774  // The calling context must be preemptible.
   775  //
   776  // Flushing local work is important because idle Ps may have local
   777  // work queued. This is the only way to make that work visible and
   778  // drive GC to completion.
   779  //
   780  // It is explicitly okay to have write barriers in this function. If
   781  // it does transition to mark termination, then all reachable objects
   782  // have been marked, so the write barrier cannot shade any more
   783  // objects.
   784  func gcMarkDone() {
   785  	// Ensure only one thread is running the ragged barrier at a
   786  	// time.
   787  	semacquire(&work.markDoneSema)
   788  
   789  top:
   790  	// Re-check transition condition under transition lock.
   791  	//
   792  	// It's critical that this checks the global work queues are
   793  	// empty before performing the ragged barrier. Otherwise,
   794  	// there could be global work that a P could take after the P
   795  	// has passed the ragged barrier.
   796  	if !(gcphase == _GCmark && work.nwait == work.nproc && !gcMarkWorkAvailable(nil)) {
   797  		semrelease(&work.markDoneSema)
   798  		return
   799  	}
   800  
   801  	// forEachP needs worldsema to execute, and we'll need it to
   802  	// stop the world later, so acquire worldsema now.
   803  	semacquire(&worldsema)
   804  
   805  	// Flush all local buffers and collect flushedWork flags.
   806  	gcMarkDoneFlushed = 0
   807  	systemstack(func() {
   808  		gp := getg().m.curg
   809  		// Mark the user stack as preemptible so that it may be scanned.
   810  		// Otherwise, our attempt to force all P's to a safepoint could
   811  		// result in a deadlock as we attempt to preempt a worker that's
   812  		// trying to preempt us (e.g. for a stack scan).
   813  		casGToWaiting(gp, _Grunning, waitReasonGCMarkTermination)
   814  		forEachP(func(pp *p) {
   815  			// Flush the write barrier buffer, since this may add
   816  			// work to the gcWork.
   817  			wbBufFlush1(pp)
   818  
   819  			// Flush the gcWork, since this may create global work
   820  			// and set the flushedWork flag.
   821  			//
   822  			// TODO(austin): Break up these workbufs to
   823  			// better distribute work.
   824  			pp.gcw.dispose()
   825  			// Collect the flushedWork flag.
   826  			if pp.gcw.flushedWork {
   827  				atomic.Xadd(&gcMarkDoneFlushed, 1)
   828  				pp.gcw.flushedWork = false
   829  			}
   830  		})
   831  		casgstatus(gp, _Gwaiting, _Grunning)
   832  	})
   833  
   834  	if gcMarkDoneFlushed != 0 {
   835  		// More grey objects were discovered since the
   836  		// previous termination check, so there may be more
   837  		// work to do. Keep going. It's possible the
   838  		// transition condition became true again during the
   839  		// ragged barrier, so re-check it.
   840  		semrelease(&worldsema)
   841  		goto top
   842  	}
   843  
   844  	// There was no global work, no local work, and no Ps
   845  	// communicated work since we took markDoneSema. Therefore
   846  	// there are no grey objects and no more objects can be
   847  	// shaded. Transition to mark termination.
   848  	now := nanotime()
   849  	work.tMarkTerm = now
   850  	work.pauseStart = now
   851  	getg().m.preemptoff = "gcing"
   852  	if trace.enabled {
   853  		traceGCSTWStart(0)
   854  	}
   855  	systemstack(stopTheWorldWithSema)
   856  	// The gcphase is _GCmark, it will transition to _GCmarktermination
   857  	// below. The important thing is that the wb remains active until
   858  	// all marking is complete. This includes writes made by the GC.
   859  
   860  	// There is sometimes work left over when we enter mark termination due
   861  	// to write barriers performed after the completion barrier above.
   862  	// Detect this and resume concurrent mark. This is obviously
   863  	// unfortunate.
   864  	//
   865  	// See issue #27993 for details.
   866  	//
   867  	// Switch to the system stack to call wbBufFlush1, though in this case
   868  	// it doesn't matter because we're non-preemptible anyway.
   869  	restart := false
   870  	systemstack(func() {
   871  		for _, p := range allp {
   872  			wbBufFlush1(p)
   873  			if !p.gcw.empty() {
   874  				restart = true
   875  				break
   876  			}
   877  		}
   878  	})
   879  	if restart {
   880  		getg().m.preemptoff = ""
   881  		systemstack(func() {
   882  			now := startTheWorldWithSema(trace.enabled)
   883  			work.pauseNS += now - work.pauseStart
   884  			memstats.gcPauseDist.record(now - work.pauseStart)
   885  		})
   886  		semrelease(&worldsema)
   887  		goto top
   888  	}
   889  
   890  	gcComputeStartingStackSize()
   891  
   892  	// Disable assists and background workers. We must do
   893  	// this before waking blocked assists.
   894  	atomic.Store(&gcBlackenEnabled, 0)
   895  
   896  	// Notify the CPU limiter that GC assists will now cease.
   897  	gcCPULimiter.startGCTransition(false, now)
   898  
   899  	// Wake all blocked assists. These will run when we
   900  	// start the world again.
   901  	gcWakeAllAssists()
   902  
   903  	// Likewise, release the transition lock. Blocked
   904  	// workers and assists will run when we start the
   905  	// world again.
   906  	semrelease(&work.markDoneSema)
   907  
   908  	// In STW mode, re-enable user goroutines. These will be
   909  	// queued to run after we start the world.
   910  	schedEnableUser(true)
   911  
   912  	// endCycle depends on all gcWork cache stats being flushed.
   913  	// The termination algorithm above ensured that up to
   914  	// allocations since the ragged barrier.
   915  	gcController.endCycle(now, int(gomaxprocs), work.userForced)
   916  
   917  	// Perform mark termination. This will restart the world.
   918  	gcMarkTermination()
   919  }
   920  
   921  // World must be stopped and mark assists and background workers must be
   922  // disabled.
   923  func gcMarkTermination() {
   924  	// Start marktermination (write barrier remains enabled for now).
   925  	setGCPhase(_GCmarktermination)
   926  
   927  	work.heap1 = gcController.heapLive.Load()
   928  	startTime := nanotime()
   929  
   930  	mp := acquirem()
   931  	mp.preemptoff = "gcing"
   932  	mp.traceback = 2
   933  	curgp := mp.curg
   934  	casGToWaiting(curgp, _Grunning, waitReasonGarbageCollection)
   935  
   936  	// Run gc on the g0 stack. We do this so that the g stack
   937  	// we're currently running on will no longer change. Cuts
   938  	// the root set down a bit (g0 stacks are not scanned, and
   939  	// we don't need to scan gc's internal state).  We also
   940  	// need to switch to g0 so we can shrink the stack.
   941  	systemstack(func() {
   942  		gcMark(startTime)
   943  		// Must return immediately.
   944  		// The outer function's stack may have moved
   945  		// during gcMark (it shrinks stacks, including the
   946  		// outer function's stack), so we must not refer
   947  		// to any of its variables. Return back to the
   948  		// non-system stack to pick up the new addresses
   949  		// before continuing.
   950  	})
   951  
   952  	systemstack(func() {
   953  		work.heap2 = work.bytesMarked
   954  		if debug.gccheckmark > 0 {
   955  			// Run a full non-parallel, stop-the-world
   956  			// mark using checkmark bits, to check that we
   957  			// didn't forget to mark anything during the
   958  			// concurrent mark process.
   959  			startCheckmarks()
   960  			gcResetMarkState()
   961  			gcw := &getg().m.p.ptr().gcw
   962  			gcDrain(gcw, 0)
   963  			wbBufFlush1(getg().m.p.ptr())
   964  			gcw.dispose()
   965  			endCheckmarks()
   966  		}
   967  
   968  		// marking is complete so we can turn the write barrier off
   969  		setGCPhase(_GCoff)
   970  		gcSweep(work.mode)
   971  	})
   972  
   973  	mp.traceback = 0
   974  	casgstatus(curgp, _Gwaiting, _Grunning)
   975  
   976  	if trace.enabled {
   977  		traceGCDone()
   978  	}
   979  
   980  	// all done
   981  	mp.preemptoff = ""
   982  
   983  	if gcphase != _GCoff {
   984  		throw("gc done but gcphase != _GCoff")
   985  	}
   986  
   987  	// Record heapInUse for scavenger.
   988  	memstats.lastHeapInUse = gcController.heapInUse.load()
   989  
   990  	// Update GC trigger and pacing, as well as downstream consumers
   991  	// of this pacing information, for the next cycle.
   992  	systemstack(gcControllerCommit)
   993  
   994  	// Update timing memstats
   995  	now := nanotime()
   996  	sec, nsec, _ := time_now()
   997  	unixNow := sec*1e9 + int64(nsec)
   998  	work.pauseNS += now - work.pauseStart
   999  	work.tEnd = now
  1000  	memstats.gcPauseDist.record(now - work.pauseStart)
  1001  	atomic.Store64(&memstats.last_gc_unix, uint64(unixNow)) // must be Unix time to make sense to user
  1002  	atomic.Store64(&memstats.last_gc_nanotime, uint64(now)) // monotonic time for us
  1003  	memstats.pause_ns[memstats.numgc%uint32(len(memstats.pause_ns))] = uint64(work.pauseNS)
  1004  	memstats.pause_end[memstats.numgc%uint32(len(memstats.pause_end))] = uint64(unixNow)
  1005  	memstats.pause_total_ns += uint64(work.pauseNS)
  1006  
  1007  	sweepTermCpu := int64(work.stwprocs) * (work.tMark - work.tSweepTerm)
  1008  	// We report idle marking time below, but omit it from the
  1009  	// overall utilization here since it's "free".
  1010  	markAssistCpu := gcController.assistTime.Load()
  1011  	markDedicatedCpu := gcController.dedicatedMarkTime.Load()
  1012  	markFractionalCpu := gcController.fractionalMarkTime.Load()
  1013  	markIdleCpu := gcController.idleMarkTime.Load()
  1014  	markTermCpu := int64(work.stwprocs) * (work.tEnd - work.tMarkTerm)
  1015  	scavAssistCpu := scavenge.assistTime.Load()
  1016  	scavBgCpu := scavenge.backgroundTime.Load()
  1017  
  1018  	// Update cumulative GC CPU stats.
  1019  	work.cpuStats.gcAssistTime += markAssistCpu
  1020  	work.cpuStats.gcDedicatedTime += markDedicatedCpu + markFractionalCpu
  1021  	work.cpuStats.gcIdleTime += markIdleCpu
  1022  	work.cpuStats.gcPauseTime += sweepTermCpu + markTermCpu
  1023  	work.cpuStats.gcTotalTime += sweepTermCpu + markAssistCpu + markDedicatedCpu + markFractionalCpu + markIdleCpu + markTermCpu
  1024  
  1025  	// Update cumulative scavenge CPU stats.
  1026  	work.cpuStats.scavengeAssistTime += scavAssistCpu
  1027  	work.cpuStats.scavengeBgTime += scavBgCpu
  1028  	work.cpuStats.scavengeTotalTime += scavAssistCpu + scavBgCpu
  1029  
  1030  	// Update total CPU.
  1031  	work.cpuStats.totalTime = sched.totaltime + (now-sched.procresizetime)*int64(gomaxprocs)
  1032  	work.cpuStats.idleTime += sched.idleTime.Load()
  1033  
  1034  	// Compute userTime. We compute this indirectly as everything that's not the above.
  1035  	//
  1036  	// Since time spent in _Pgcstop is covered by gcPauseTime, and time spent in _Pidle
  1037  	// is covered by idleTime, what we're left with is time spent in _Prunning and _Psyscall,
  1038  	// the latter of which is fine because the P will either go idle or get used for something
  1039  	// else via sysmon. Meanwhile if we subtract GC time from whatever's left, we get non-GC
  1040  	// _Prunning time. Note that this still leaves time spent in sweeping and in the scheduler,
  1041  	// but that's fine. The overwhelming majority of this time will be actual user time.
  1042  	work.cpuStats.userTime = work.cpuStats.totalTime - (work.cpuStats.gcTotalTime +
  1043  		work.cpuStats.scavengeTotalTime + work.cpuStats.idleTime)
  1044  
  1045  	// Compute overall GC CPU utilization.
  1046  	// Omit idle marking time from the overall utilization here since it's "free".
  1047  	memstats.gc_cpu_fraction = float64(work.cpuStats.gcTotalTime-work.cpuStats.gcIdleTime) / float64(work.cpuStats.totalTime)
  1048  
  1049  	// Reset assist time and background time stats.
  1050  	//
  1051  	// Do this now, instead of at the start of the next GC cycle, because
  1052  	// these two may keep accumulating even if the GC is not active.
  1053  	scavenge.assistTime.Store(0)
  1054  	scavenge.backgroundTime.Store(0)
  1055  
  1056  	// Reset idle time stat.
  1057  	sched.idleTime.Store(0)
  1058  
  1059  	// Reset sweep state.
  1060  	sweep.nbgsweep = 0
  1061  	sweep.npausesweep = 0
  1062  
  1063  	if work.userForced {
  1064  		memstats.numforcedgc++
  1065  	}
  1066  
  1067  	// Bump GC cycle count and wake goroutines waiting on sweep.
  1068  	lock(&work.sweepWaiters.lock)
  1069  	memstats.numgc++
  1070  	injectglist(&work.sweepWaiters.list)
  1071  	unlock(&work.sweepWaiters.lock)
  1072  
  1073  	// Release the CPU limiter.
  1074  	gcCPULimiter.finishGCTransition(now)
  1075  
  1076  	// Finish the current heap profiling cycle and start a new
  1077  	// heap profiling cycle. We do this before starting the world
  1078  	// so events don't leak into the wrong cycle.
  1079  	mProf_NextCycle()
  1080  
  1081  	// There may be stale spans in mcaches that need to be swept.
  1082  	// Those aren't tracked in any sweep lists, so we need to
  1083  	// count them against sweep completion until we ensure all
  1084  	// those spans have been forced out.
  1085  	sl := sweep.active.begin()
  1086  	if !sl.valid {
  1087  		throw("failed to set sweep barrier")
  1088  	}
  1089  
  1090  	systemstack(func() { startTheWorldWithSema(trace.enabled) })
  1091  
  1092  	// Flush the heap profile so we can start a new cycle next GC.
  1093  	// This is relatively expensive, so we don't do it with the
  1094  	// world stopped.
  1095  	mProf_Flush()
  1096  
  1097  	// Prepare workbufs for freeing by the sweeper. We do this
  1098  	// asynchronously because it can take non-trivial time.
  1099  	prepareFreeWorkbufs()
  1100  
  1101  	// Free stack spans. This must be done between GC cycles.
  1102  	systemstack(freeStackSpans)
  1103  
  1104  	// Ensure all mcaches are flushed. Each P will flush its own
  1105  	// mcache before allocating, but idle Ps may not. Since this
  1106  	// is necessary to sweep all spans, we need to ensure all
  1107  	// mcaches are flushed before we start the next GC cycle.
  1108  	systemstack(func() {
  1109  		forEachP(func(pp *p) {
  1110  			pp.mcache.prepareForSweep()
  1111  		})
  1112  	})
  1113  	// Now that we've swept stale spans in mcaches, they don't
  1114  	// count against unswept spans.
  1115  	sweep.active.end(sl)
  1116  
  1117  	// Print gctrace before dropping worldsema. As soon as we drop
  1118  	// worldsema another cycle could start and smash the stats
  1119  	// we're trying to print.
  1120  	if debug.gctrace > 0 {
  1121  		util := int(memstats.gc_cpu_fraction * 100)
  1122  
  1123  		var sbuf [24]byte
  1124  		printlock()
  1125  		print("gc ", memstats.numgc,
  1126  			" @", string(itoaDiv(sbuf[:], uint64(work.tSweepTerm-runtimeInitTime)/1e6, 3)), "s ",
  1127  			util, "%: ")
  1128  		prev := work.tSweepTerm
  1129  		for i, ns := range []int64{work.tMark, work.tMarkTerm, work.tEnd} {
  1130  			if i != 0 {
  1131  				print("+")
  1132  			}
  1133  			print(string(fmtNSAsMS(sbuf[:], uint64(ns-prev))))
  1134  			prev = ns
  1135  		}
  1136  		print(" ms clock, ")
  1137  		for i, ns := range []int64{
  1138  			sweepTermCpu,
  1139  			gcController.assistTime.Load(),
  1140  			gcController.dedicatedMarkTime.Load() + gcController.fractionalMarkTime.Load(),
  1141  			gcController.idleMarkTime.Load(),
  1142  			markTermCpu,
  1143  		} {
  1144  			if i == 2 || i == 3 {
  1145  				// Separate mark time components with /.
  1146  				print("/")
  1147  			} else if i != 0 {
  1148  				print("+")
  1149  			}
  1150  			print(string(fmtNSAsMS(sbuf[:], uint64(ns))))
  1151  		}
  1152  		print(" ms cpu, ",
  1153  			work.heap0>>20, "->", work.heap1>>20, "->", work.heap2>>20, " MB, ",
  1154  			gcController.lastHeapGoal>>20, " MB goal, ",
  1155  			gcController.lastStackScan.Load()>>20, " MB stacks, ",
  1156  			gcController.globalsScan.Load()>>20, " MB globals, ",
  1157  			work.maxprocs, " P")
  1158  		if work.userForced {
  1159  			print(" (forced)")
  1160  		}
  1161  		print("\n")
  1162  		printunlock()
  1163  	}
  1164  
  1165  	// Set any arena chunks that were deferred to fault.
  1166  	lock(&userArenaState.lock)
  1167  	faultList := userArenaState.fault
  1168  	userArenaState.fault = nil
  1169  	unlock(&userArenaState.lock)
  1170  	for _, lc := range faultList {
  1171  		lc.mspan.setUserArenaChunkToFault()
  1172  	}
  1173  
  1174  	semrelease(&worldsema)
  1175  	semrelease(&gcsema)
  1176  	// Careful: another GC cycle may start now.
  1177  
  1178  	releasem(mp)
  1179  	mp = nil
  1180  
  1181  	// now that gc is done, kick off finalizer thread if needed
  1182  	if !concurrentSweep {
  1183  		// give the queued finalizers, if any, a chance to run
  1184  		Gosched()
  1185  	}
  1186  }
  1187  
  1188  // gcBgMarkStartWorkers prepares background mark worker goroutines. These
  1189  // goroutines will not run until the mark phase, but they must be started while
  1190  // the work is not stopped and from a regular G stack. The caller must hold
  1191  // worldsema.
  1192  func gcBgMarkStartWorkers() {
  1193  	// Background marking is performed by per-P G's. Ensure that each P has
  1194  	// a background GC G.
  1195  	//
  1196  	// Worker Gs don't exit if gomaxprocs is reduced. If it is raised
  1197  	// again, we can reuse the old workers; no need to create new workers.
  1198  	for gcBgMarkWorkerCount < gomaxprocs {
  1199  		go gcBgMarkWorker()
  1200  
  1201  		notetsleepg(&work.bgMarkReady, -1)
  1202  		noteclear(&work.bgMarkReady)
  1203  		// The worker is now guaranteed to be added to the pool before
  1204  		// its P's next findRunnableGCWorker.
  1205  
  1206  		gcBgMarkWorkerCount++
  1207  	}
  1208  }
  1209  
  1210  // gcBgMarkPrepare sets up state for background marking.
  1211  // Mutator assists must not yet be enabled.
  1212  func gcBgMarkPrepare() {
  1213  	// Background marking will stop when the work queues are empty
  1214  	// and there are no more workers (note that, since this is
  1215  	// concurrent, this may be a transient state, but mark
  1216  	// termination will clean it up). Between background workers
  1217  	// and assists, we don't really know how many workers there
  1218  	// will be, so we pretend to have an arbitrarily large number
  1219  	// of workers, almost all of which are "waiting". While a
  1220  	// worker is working it decrements nwait. If nproc == nwait,
  1221  	// there are no workers.
  1222  	work.nproc = ^uint32(0)
  1223  	work.nwait = ^uint32(0)
  1224  }
  1225  
  1226  // gcBgMarkWorkerNode is an entry in the gcBgMarkWorkerPool. It points to a single
  1227  // gcBgMarkWorker goroutine.
  1228  type gcBgMarkWorkerNode struct {
  1229  	// Unused workers are managed in a lock-free stack. This field must be first.
  1230  	node lfnode
  1231  
  1232  	// The g of this worker.
  1233  	gp guintptr
  1234  
  1235  	// Release this m on park. This is used to communicate with the unlock
  1236  	// function, which cannot access the G's stack. It is unused outside of
  1237  	// gcBgMarkWorker().
  1238  	m muintptr
  1239  }
  1240  
  1241  func gcBgMarkWorker() {
  1242  	gp := getg()
  1243  
  1244  	// We pass node to a gopark unlock function, so it can't be on
  1245  	// the stack (see gopark). Prevent deadlock from recursively
  1246  	// starting GC by disabling preemption.
  1247  	gp.m.preemptoff = "GC worker init"
  1248  	node := new(gcBgMarkWorkerNode)
  1249  	gp.m.preemptoff = ""
  1250  
  1251  	node.gp.set(gp)
  1252  
  1253  	node.m.set(acquirem())
  1254  	notewakeup(&work.bgMarkReady)
  1255  	// After this point, the background mark worker is generally scheduled
  1256  	// cooperatively by gcController.findRunnableGCWorker. While performing
  1257  	// work on the P, preemption is disabled because we are working on
  1258  	// P-local work buffers. When the preempt flag is set, this puts itself
  1259  	// into _Gwaiting to be woken up by gcController.findRunnableGCWorker
  1260  	// at the appropriate time.
  1261  	//
  1262  	// When preemption is enabled (e.g., while in gcMarkDone), this worker
  1263  	// may be preempted and schedule as a _Grunnable G from a runq. That is
  1264  	// fine; it will eventually gopark again for further scheduling via
  1265  	// findRunnableGCWorker.
  1266  	//
  1267  	// Since we disable preemption before notifying bgMarkReady, we
  1268  	// guarantee that this G will be in the worker pool for the next
  1269  	// findRunnableGCWorker. This isn't strictly necessary, but it reduces
  1270  	// latency between _GCmark starting and the workers starting.
  1271  
  1272  	for {
  1273  		// Go to sleep until woken by
  1274  		// gcController.findRunnableGCWorker.
  1275  		gopark(func(g *g, nodep unsafe.Pointer) bool {
  1276  			node := (*gcBgMarkWorkerNode)(nodep)
  1277  
  1278  			if mp := node.m.ptr(); mp != nil {
  1279  				// The worker G is no longer running; release
  1280  				// the M.
  1281  				//
  1282  				// N.B. it is _safe_ to release the M as soon
  1283  				// as we are no longer performing P-local mark
  1284  				// work.
  1285  				//
  1286  				// However, since we cooperatively stop work
  1287  				// when gp.preempt is set, if we releasem in
  1288  				// the loop then the following call to gopark
  1289  				// would immediately preempt the G. This is
  1290  				// also safe, but inefficient: the G must
  1291  				// schedule again only to enter gopark and park
  1292  				// again. Thus, we defer the release until
  1293  				// after parking the G.
  1294  				releasem(mp)
  1295  			}
  1296  
  1297  			// Release this G to the pool.
  1298  			gcBgMarkWorkerPool.push(&node.node)
  1299  			// Note that at this point, the G may immediately be
  1300  			// rescheduled and may be running.
  1301  			return true
  1302  		}, unsafe.Pointer(node), waitReasonGCWorkerIdle, traceEvGoBlock, 0)
  1303  
  1304  		// Preemption must not occur here, or another G might see
  1305  		// p.gcMarkWorkerMode.
  1306  
  1307  		// Disable preemption so we can use the gcw. If the
  1308  		// scheduler wants to preempt us, we'll stop draining,
  1309  		// dispose the gcw, and then preempt.
  1310  		node.m.set(acquirem())
  1311  		pp := gp.m.p.ptr() // P can't change with preemption disabled.
  1312  
  1313  		if gcBlackenEnabled == 0 {
  1314  			println("worker mode", pp.gcMarkWorkerMode)
  1315  			throw("gcBgMarkWorker: blackening not enabled")
  1316  		}
  1317  
  1318  		if pp.gcMarkWorkerMode == gcMarkWorkerNotWorker {
  1319  			throw("gcBgMarkWorker: mode not set")
  1320  		}
  1321  
  1322  		startTime := nanotime()
  1323  		pp.gcMarkWorkerStartTime = startTime
  1324  		var trackLimiterEvent bool
  1325  		if pp.gcMarkWorkerMode == gcMarkWorkerIdleMode {
  1326  			trackLimiterEvent = pp.limiterEvent.start(limiterEventIdleMarkWork, startTime)
  1327  		}
  1328  
  1329  		decnwait := atomic.Xadd(&work.nwait, -1)
  1330  		if decnwait == work.nproc {
  1331  			println("runtime: work.nwait=", decnwait, "work.nproc=", work.nproc)
  1332  			throw("work.nwait was > work.nproc")
  1333  		}
  1334  
  1335  		systemstack(func() {
  1336  			// Mark our goroutine preemptible so its stack
  1337  			// can be scanned. This lets two mark workers
  1338  			// scan each other (otherwise, they would
  1339  			// deadlock). We must not modify anything on
  1340  			// the G stack. However, stack shrinking is
  1341  			// disabled for mark workers, so it is safe to
  1342  			// read from the G stack.
  1343  			casGToWaiting(gp, _Grunning, waitReasonGCWorkerActive)
  1344  			switch pp.gcMarkWorkerMode {
  1345  			default:
  1346  				throw("gcBgMarkWorker: unexpected gcMarkWorkerMode")
  1347  			case gcMarkWorkerDedicatedMode:
  1348  				gcDrain(&pp.gcw, gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1349  				if gp.preempt {
  1350  					// We were preempted. This is
  1351  					// a useful signal to kick
  1352  					// everything out of the run
  1353  					// queue so it can run
  1354  					// somewhere else.
  1355  					if drainQ, n := runqdrain(pp); n > 0 {
  1356  						lock(&sched.lock)
  1357  						globrunqputbatch(&drainQ, int32(n))
  1358  						unlock(&sched.lock)
  1359  					}
  1360  				}
  1361  				// Go back to draining, this time
  1362  				// without preemption.
  1363  				gcDrain(&pp.gcw, gcDrainFlushBgCredit)
  1364  			case gcMarkWorkerFractionalMode:
  1365  				gcDrain(&pp.gcw, gcDrainFractional|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1366  			case gcMarkWorkerIdleMode:
  1367  				gcDrain(&pp.gcw, gcDrainIdle|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1368  			}
  1369  			casgstatus(gp, _Gwaiting, _Grunning)
  1370  		})
  1371  
  1372  		// Account for time and mark us as stopped.
  1373  		now := nanotime()
  1374  		duration := now - startTime
  1375  		gcController.markWorkerStop(pp.gcMarkWorkerMode, duration)
  1376  		if trackLimiterEvent {
  1377  			pp.limiterEvent.stop(limiterEventIdleMarkWork, now)
  1378  		}
  1379  		if pp.gcMarkWorkerMode == gcMarkWorkerFractionalMode {
  1380  			atomic.Xaddint64(&pp.gcFractionalMarkTime, duration)
  1381  		}
  1382  
  1383  		// Was this the last worker and did we run out
  1384  		// of work?
  1385  		incnwait := atomic.Xadd(&work.nwait, +1)
  1386  		if incnwait > work.nproc {
  1387  			println("runtime: p.gcMarkWorkerMode=", pp.gcMarkWorkerMode,
  1388  				"work.nwait=", incnwait, "work.nproc=", work.nproc)
  1389  			throw("work.nwait > work.nproc")
  1390  		}
  1391  
  1392  		// We'll releasem after this point and thus this P may run
  1393  		// something else. We must clear the worker mode to avoid
  1394  		// attributing the mode to a different (non-worker) G in
  1395  		// traceGoStart.
  1396  		pp.gcMarkWorkerMode = gcMarkWorkerNotWorker
  1397  
  1398  		// If this worker reached a background mark completion
  1399  		// point, signal the main GC goroutine.
  1400  		if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
  1401  			// We don't need the P-local buffers here, allow
  1402  			// preemption because we may schedule like a regular
  1403  			// goroutine in gcMarkDone (block on locks, etc).
  1404  			releasem(node.m.ptr())
  1405  			node.m.set(nil)
  1406  
  1407  			gcMarkDone()
  1408  		}
  1409  	}
  1410  }
  1411  
  1412  // gcMarkWorkAvailable reports whether executing a mark worker
  1413  // on p is potentially useful. p may be nil, in which case it only
  1414  // checks the global sources of work.
  1415  func gcMarkWorkAvailable(p *p) bool {
  1416  	if p != nil && !p.gcw.empty() {
  1417  		return true
  1418  	}
  1419  	if !work.full.empty() {
  1420  		return true // global work available
  1421  	}
  1422  	if work.markrootNext < work.markrootJobs {
  1423  		return true // root scan work available
  1424  	}
  1425  	return false
  1426  }
  1427  
  1428  // gcMark runs the mark (or, for concurrent GC, mark termination)
  1429  // All gcWork caches must be empty.
  1430  // STW is in effect at this point.
  1431  func gcMark(startTime int64) {
  1432  	if debug.allocfreetrace > 0 {
  1433  		tracegc()
  1434  	}
  1435  
  1436  	if gcphase != _GCmarktermination {
  1437  		throw("in gcMark expecting to see gcphase as _GCmarktermination")
  1438  	}
  1439  	work.tstart = startTime
  1440  
  1441  	// Check that there's no marking work remaining.
  1442  	if work.full != 0 || work.markrootNext < work.markrootJobs {
  1443  		print("runtime: full=", hex(work.full), " next=", work.markrootNext, " jobs=", work.markrootJobs, " nDataRoots=", work.nDataRoots, " nBSSRoots=", work.nBSSRoots, " nSpanRoots=", work.nSpanRoots, " nStackRoots=", work.nStackRoots, "\n")
  1444  		panic("non-empty mark queue after concurrent mark")
  1445  	}
  1446  
  1447  	if debug.gccheckmark > 0 {
  1448  		// This is expensive when there's a large number of
  1449  		// Gs, so only do it if checkmark is also enabled.
  1450  		gcMarkRootCheck()
  1451  	}
  1452  	if work.full != 0 {
  1453  		throw("work.full != 0")
  1454  	}
  1455  
  1456  	// Drop allg snapshot. allgs may have grown, in which case
  1457  	// this is the only reference to the old backing store and
  1458  	// there's no need to keep it around.
  1459  	work.stackRoots = nil
  1460  
  1461  	// Clear out buffers and double-check that all gcWork caches
  1462  	// are empty. This should be ensured by gcMarkDone before we
  1463  	// enter mark termination.
  1464  	//
  1465  	// TODO: We could clear out buffers just before mark if this
  1466  	// has a non-negligible impact on STW time.
  1467  	for _, p := range allp {
  1468  		// The write barrier may have buffered pointers since
  1469  		// the gcMarkDone barrier. However, since the barrier
  1470  		// ensured all reachable objects were marked, all of
  1471  		// these must be pointers to black objects. Hence we
  1472  		// can just discard the write barrier buffer.
  1473  		if debug.gccheckmark > 0 {
  1474  			// For debugging, flush the buffer and make
  1475  			// sure it really was all marked.
  1476  			wbBufFlush1(p)
  1477  		} else {
  1478  			p.wbBuf.reset()
  1479  		}
  1480  
  1481  		gcw := &p.gcw
  1482  		if !gcw.empty() {
  1483  			printlock()
  1484  			print("runtime: P ", p.id, " flushedWork ", gcw.flushedWork)
  1485  			if gcw.wbuf1 == nil {
  1486  				print(" wbuf1=<nil>")
  1487  			} else {
  1488  				print(" wbuf1.n=", gcw.wbuf1.nobj)
  1489  			}
  1490  			if gcw.wbuf2 == nil {
  1491  				print(" wbuf2=<nil>")
  1492  			} else {
  1493  				print(" wbuf2.n=", gcw.wbuf2.nobj)
  1494  			}
  1495  			print("\n")
  1496  			throw("P has cached GC work at end of mark termination")
  1497  		}
  1498  		// There may still be cached empty buffers, which we
  1499  		// need to flush since we're going to free them. Also,
  1500  		// there may be non-zero stats because we allocated
  1501  		// black after the gcMarkDone barrier.
  1502  		gcw.dispose()
  1503  	}
  1504  
  1505  	// Flush scanAlloc from each mcache since we're about to modify
  1506  	// heapScan directly. If we were to flush this later, then scanAlloc
  1507  	// might have incorrect information.
  1508  	//
  1509  	// Note that it's not important to retain this information; we know
  1510  	// exactly what heapScan is at this point via scanWork.
  1511  	for _, p := range allp {
  1512  		c := p.mcache
  1513  		if c == nil {
  1514  			continue
  1515  		}
  1516  		c.scanAlloc = 0
  1517  	}
  1518  
  1519  	// Reset controller state.
  1520  	gcController.resetLive(work.bytesMarked)
  1521  }
  1522  
  1523  // gcSweep must be called on the system stack because it acquires the heap
  1524  // lock. See mheap for details.
  1525  //
  1526  // The world must be stopped.
  1527  //
  1528  //go:systemstack
  1529  func gcSweep(mode gcMode) {
  1530  	assertWorldStopped()
  1531  
  1532  	if gcphase != _GCoff {
  1533  		throw("gcSweep being done but phase is not GCoff")
  1534  	}
  1535  
  1536  	lock(&mheap_.lock)
  1537  	mheap_.sweepgen += 2
  1538  	sweep.active.reset()
  1539  	mheap_.pagesSwept.Store(0)
  1540  	mheap_.sweepArenas = mheap_.allArenas
  1541  	mheap_.reclaimIndex.Store(0)
  1542  	mheap_.reclaimCredit.Store(0)
  1543  	unlock(&mheap_.lock)
  1544  
  1545  	sweep.centralIndex.clear()
  1546  
  1547  	if !_ConcurrentSweep || mode == gcForceBlockMode {
  1548  		// Special case synchronous sweep.
  1549  		// Record that no proportional sweeping has to happen.
  1550  		lock(&mheap_.lock)
  1551  		mheap_.sweepPagesPerByte = 0
  1552  		unlock(&mheap_.lock)
  1553  		// Sweep all spans eagerly.
  1554  		for sweepone() != ^uintptr(0) {
  1555  			sweep.npausesweep++
  1556  		}
  1557  		// Free workbufs eagerly.
  1558  		prepareFreeWorkbufs()
  1559  		for freeSomeWbufs(false) {
  1560  		}
  1561  		// All "free" events for this mark/sweep cycle have
  1562  		// now happened, so we can make this profile cycle
  1563  		// available immediately.
  1564  		mProf_NextCycle()
  1565  		mProf_Flush()
  1566  		return
  1567  	}
  1568  
  1569  	// Background sweep.
  1570  	lock(&sweep.lock)
  1571  	if sweep.parked {
  1572  		sweep.parked = false
  1573  		ready(sweep.g, 0, true)
  1574  	}
  1575  	unlock(&sweep.lock)
  1576  }
  1577  
  1578  // gcResetMarkState resets global state prior to marking (concurrent
  1579  // or STW) and resets the stack scan state of all Gs.
  1580  //
  1581  // This is safe to do without the world stopped because any Gs created
  1582  // during or after this will start out in the reset state.
  1583  //
  1584  // gcResetMarkState must be called on the system stack because it acquires
  1585  // the heap lock. See mheap for details.
  1586  //
  1587  //go:systemstack
  1588  func gcResetMarkState() {
  1589  	// This may be called during a concurrent phase, so lock to make sure
  1590  	// allgs doesn't change.
  1591  	forEachG(func(gp *g) {
  1592  		gp.gcscandone = false // set to true in gcphasework
  1593  		gp.gcAssistBytes = 0
  1594  	})
  1595  
  1596  	// Clear page marks. This is just 1MB per 64GB of heap, so the
  1597  	// time here is pretty trivial.
  1598  	lock(&mheap_.lock)
  1599  	arenas := mheap_.allArenas
  1600  	unlock(&mheap_.lock)
  1601  	for _, ai := range arenas {
  1602  		ha := mheap_.arenas[ai.l1()][ai.l2()]
  1603  		for i := range ha.pageMarks {
  1604  			ha.pageMarks[i] = 0
  1605  		}
  1606  	}
  1607  
  1608  	work.bytesMarked = 0
  1609  	work.initialHeapLive = gcController.heapLive.Load()
  1610  }
  1611  
  1612  // Hooks for other packages
  1613  
  1614  var poolcleanup func()
  1615  var boringCaches []unsafe.Pointer // for crypto/internal/boring
  1616  
  1617  //go:linkname sync_runtime_registerPoolCleanup sync.runtime_registerPoolCleanup
  1618  func sync_runtime_registerPoolCleanup(f func()) {
  1619  	poolcleanup = f
  1620  }
  1621  
  1622  //go:linkname boring_registerCache crypto/internal/boring/bcache.registerCache
  1623  func boring_registerCache(p unsafe.Pointer) {
  1624  	boringCaches = append(boringCaches, p)
  1625  }
  1626  
  1627  func clearpools() {
  1628  	// clear sync.Pools
  1629  	if poolcleanup != nil {
  1630  		poolcleanup()
  1631  	}
  1632  
  1633  	// clear boringcrypto caches
  1634  	for _, p := range boringCaches {
  1635  		atomicstorep(p, nil)
  1636  	}
  1637  
  1638  	// Clear central sudog cache.
  1639  	// Leave per-P caches alone, they have strictly bounded size.
  1640  	// Disconnect cached list before dropping it on the floor,
  1641  	// so that a dangling ref to one entry does not pin all of them.
  1642  	lock(&sched.sudoglock)
  1643  	var sg, sgnext *sudog
  1644  	for sg = sched.sudogcache; sg != nil; sg = sgnext {
  1645  		sgnext = sg.next
  1646  		sg.next = nil
  1647  	}
  1648  	sched.sudogcache = nil
  1649  	unlock(&sched.sudoglock)
  1650  
  1651  	// Clear central defer pool.
  1652  	// Leave per-P pools alone, they have strictly bounded size.
  1653  	lock(&sched.deferlock)
  1654  	// disconnect cached list before dropping it on the floor,
  1655  	// so that a dangling ref to one entry does not pin all of them.
  1656  	var d, dlink *_defer
  1657  	for d = sched.deferpool; d != nil; d = dlink {
  1658  		dlink = d.link
  1659  		d.link = nil
  1660  	}
  1661  	sched.deferpool = nil
  1662  	unlock(&sched.deferlock)
  1663  }
  1664  
  1665  // Timing
  1666  
  1667  // itoaDiv formats val/(10**dec) into buf.
  1668  func itoaDiv(buf []byte, val uint64, dec int) []byte {
  1669  	i := len(buf) - 1
  1670  	idec := i - dec
  1671  	for val >= 10 || i >= idec {
  1672  		buf[i] = byte(val%10 + '0')
  1673  		i--
  1674  		if i == idec {
  1675  			buf[i] = '.'
  1676  			i--
  1677  		}
  1678  		val /= 10
  1679  	}
  1680  	buf[i] = byte(val + '0')
  1681  	return buf[i:]
  1682  }
  1683  
  1684  // fmtNSAsMS nicely formats ns nanoseconds as milliseconds.
  1685  func fmtNSAsMS(buf []byte, ns uint64) []byte {
  1686  	if ns >= 10e6 {
  1687  		// Format as whole milliseconds.
  1688  		return itoaDiv(buf, ns/1e6, 0)
  1689  	}
  1690  	// Format two digits of precision, with at most three decimal places.
  1691  	x := ns / 1e3
  1692  	if x == 0 {
  1693  		buf[0] = '0'
  1694  		return buf[:1]
  1695  	}
  1696  	dec := 3
  1697  	for x >= 100 {
  1698  		x /= 10
  1699  		dec--
  1700  	}
  1701  	return itoaDiv(buf, x, dec)
  1702  }
  1703  
  1704  // Helpers for testing GC.
  1705  
  1706  // gcTestMoveStackOnNextCall causes the stack to be moved on a call
  1707  // immediately following the call to this. It may not work correctly
  1708  // if any other work appears after this call (such as returning).
  1709  // Typically the following call should be marked go:noinline so it
  1710  // performs a stack check.
  1711  //
  1712  // In rare cases this may not cause the stack to move, specifically if
  1713  // there's a preemption between this call and the next.
  1714  func gcTestMoveStackOnNextCall() {
  1715  	gp := getg()
  1716  	gp.stackguard0 = stackForceMove
  1717  }
  1718  
  1719  // gcTestIsReachable performs a GC and returns a bit set where bit i
  1720  // is set if ptrs[i] is reachable.
  1721  func gcTestIsReachable(ptrs ...unsafe.Pointer) (mask uint64) {
  1722  	// This takes the pointers as unsafe.Pointers in order to keep
  1723  	// them live long enough for us to attach specials. After
  1724  	// that, we drop our references to them.
  1725  
  1726  	if len(ptrs) > 64 {
  1727  		panic("too many pointers for uint64 mask")
  1728  	}
  1729  
  1730  	// Block GC while we attach specials and drop our references
  1731  	// to ptrs. Otherwise, if a GC is in progress, it could mark
  1732  	// them reachable via this function before we have a chance to
  1733  	// drop them.
  1734  	semacquire(&gcsema)
  1735  
  1736  	// Create reachability specials for ptrs.
  1737  	specials := make([]*specialReachable, len(ptrs))
  1738  	for i, p := range ptrs {
  1739  		lock(&mheap_.speciallock)
  1740  		s := (*specialReachable)(mheap_.specialReachableAlloc.alloc())
  1741  		unlock(&mheap_.speciallock)
  1742  		s.special.kind = _KindSpecialReachable
  1743  		if !addspecial(p, &s.special) {
  1744  			throw("already have a reachable special (duplicate pointer?)")
  1745  		}
  1746  		specials[i] = s
  1747  		// Make sure we don't retain ptrs.
  1748  		ptrs[i] = nil
  1749  	}
  1750  
  1751  	semrelease(&gcsema)
  1752  
  1753  	// Force a full GC and sweep.
  1754  	GC()
  1755  
  1756  	// Process specials.
  1757  	for i, s := range specials {
  1758  		if !s.done {
  1759  			printlock()
  1760  			println("runtime: object", i, "was not swept")
  1761  			throw("IsReachable failed")
  1762  		}
  1763  		if s.reachable {
  1764  			mask |= 1 << i
  1765  		}
  1766  		lock(&mheap_.speciallock)
  1767  		mheap_.specialReachableAlloc.free(unsafe.Pointer(s))
  1768  		unlock(&mheap_.speciallock)
  1769  	}
  1770  
  1771  	return mask
  1772  }
  1773  
  1774  // gcTestPointerClass returns the category of what p points to, one of:
  1775  // "heap", "stack", "data", "bss", "other". This is useful for checking
  1776  // that a test is doing what it's intended to do.
  1777  //
  1778  // This is nosplit simply to avoid extra pointer shuffling that may
  1779  // complicate a test.
  1780  //
  1781  //go:nosplit
  1782  func gcTestPointerClass(p unsafe.Pointer) string {
  1783  	p2 := uintptr(noescape(p))
  1784  	gp := getg()
  1785  	if gp.stack.lo <= p2 && p2 < gp.stack.hi {
  1786  		return "stack"
  1787  	}
  1788  	if base, _, _ := findObject(p2, 0, 0); base != 0 {
  1789  		return "heap"
  1790  	}
  1791  	for _, datap := range activeModules() {
  1792  		if datap.data <= p2 && p2 < datap.edata || datap.noptrdata <= p2 && p2 < datap.enoptrdata {
  1793  			return "data"
  1794  		}
  1795  		if datap.bss <= p2 && p2 < datap.ebss || datap.noptrbss <= p2 && p2 <= datap.enoptrbss {
  1796  			return "bss"
  1797  		}
  1798  	}
  1799  	KeepAlive(p)
  1800  	return "other"
  1801  }
  1802  

View as plain text