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  	"internal/runtime/atomic"
   134  	"unsafe"
   135  )
   136  
   137  const (
   138  	_DebugGC      = 0
   139  	_FinBlockSize = 4 * 1024
   140  
   141  	// concurrentSweep is a debug flag. Disabling this flag
   142  	// ensures all spans are swept while the world is stopped.
   143  	concurrentSweep = true
   144  
   145  	// debugScanConservative enables debug logging for stack
   146  	// frames that are scanned conservatively.
   147  	debugScanConservative = false
   148  
   149  	// sweepMinHeapDistance is a lower bound on the heap distance
   150  	// (in bytes) reserved for concurrent sweeping between GC
   151  	// cycles.
   152  	sweepMinHeapDistance = 1024 * 1024
   153  )
   154  
   155  // heapObjectsCanMove always returns false in the current garbage collector.
   156  // It exists for go4.org/unsafe/assume-no-moving-gc, which is an
   157  // unfortunate idea that had an even more unfortunate implementation.
   158  // Every time a new Go release happened, the package stopped building,
   159  // and the authors had to add a new file with a new //go:build line, and
   160  // then the entire ecosystem of packages with that as a dependency had to
   161  // explicitly update to the new version. Many packages depend on
   162  // assume-no-moving-gc transitively, through paths like
   163  // inet.af/netaddr -> go4.org/intern -> assume-no-moving-gc.
   164  // This was causing a significant amount of friction around each new
   165  // release, so we added this bool for the package to //go:linkname
   166  // instead. The bool is still unfortunate, but it's not as bad as
   167  // breaking the ecosystem on every new release.
   168  //
   169  // If the Go garbage collector ever does move heap objects, we can set
   170  // this to true to break all the programs using assume-no-moving-gc.
   171  //
   172  //go:linkname heapObjectsCanMove
   173  func heapObjectsCanMove() bool {
   174  	return false
   175  }
   176  
   177  func gcinit() {
   178  	if unsafe.Sizeof(workbuf{}) != _WorkbufSize {
   179  		throw("size of Workbuf is suboptimal")
   180  	}
   181  	// No sweep on the first cycle.
   182  	sweep.active.state.Store(sweepDrainedMask)
   183  
   184  	// Initialize GC pacer state.
   185  	// Use the environment variable GOGC for the initial gcPercent value.
   186  	// Use the environment variable GOMEMLIMIT for the initial memoryLimit value.
   187  	gcController.init(readGOGC(), readGOMEMLIMIT())
   188  
   189  	work.startSema = 1
   190  	work.markDoneSema = 1
   191  	lockInit(&work.sweepWaiters.lock, lockRankSweepWaiters)
   192  	lockInit(&work.assistQueue.lock, lockRankAssistQueue)
   193  	lockInit(&work.strongFromWeak.lock, lockRankStrongFromWeakQueue)
   194  	lockInit(&work.wbufSpans.lock, lockRankWbufSpans)
   195  }
   196  
   197  // gcenable is called after the bulk of the runtime initialization,
   198  // just before we're about to start letting user code run.
   199  // It kicks off the background sweeper goroutine, the background
   200  // scavenger goroutine, and enables GC.
   201  func gcenable() {
   202  	// Kick off sweeping and scavenging.
   203  	c := make(chan int, 2)
   204  	go bgsweep(c)
   205  	go bgscavenge(c)
   206  	<-c
   207  	<-c
   208  	memstats.enablegc = true // now that runtime is initialized, GC is okay
   209  }
   210  
   211  // Garbage collector phase.
   212  // Indicates to write barrier and synchronization task to perform.
   213  var gcphase uint32
   214  
   215  // The compiler knows about this variable.
   216  // If you change it, you must change builtin/runtime.go, too.
   217  // If you change the first four bytes, you must also change the write
   218  // barrier insertion code.
   219  //
   220  // writeBarrier should be an internal detail,
   221  // but widely used packages access it using linkname.
   222  // Notable members of the hall of shame include:
   223  //   - github.com/bytedance/sonic
   224  //
   225  // Do not remove or change the type signature.
   226  // See go.dev/issue/67401.
   227  //
   228  //go:linkname writeBarrier
   229  var writeBarrier struct {
   230  	enabled bool    // compiler emits a check of this before calling write barrier
   231  	pad     [3]byte // compiler uses 32-bit load for "enabled" field
   232  	alignme uint64  // guarantee alignment so that compiler can use a 32 or 64-bit load
   233  }
   234  
   235  // gcBlackenEnabled is 1 if mutator assists and background mark
   236  // workers are allowed to blacken objects. This must only be set when
   237  // gcphase == _GCmark.
   238  var gcBlackenEnabled uint32
   239  
   240  const (
   241  	_GCoff             = iota // GC not running; sweeping in background, write barrier disabled
   242  	_GCmark                   // GC marking roots and workbufs: allocate black, write barrier ENABLED
   243  	_GCmarktermination        // GC mark termination: allocate black, P's help GC, write barrier ENABLED
   244  )
   245  
   246  //go:nosplit
   247  func setGCPhase(x uint32) {
   248  	atomic.Store(&gcphase, x)
   249  	writeBarrier.enabled = gcphase == _GCmark || gcphase == _GCmarktermination
   250  }
   251  
   252  // gcMarkWorkerMode represents the mode that a concurrent mark worker
   253  // should operate in.
   254  //
   255  // Concurrent marking happens through four different mechanisms. One
   256  // is mutator assists, which happen in response to allocations and are
   257  // not scheduled. The other three are variations in the per-P mark
   258  // workers and are distinguished by gcMarkWorkerMode.
   259  type gcMarkWorkerMode int
   260  
   261  const (
   262  	// gcMarkWorkerNotWorker indicates that the next scheduled G is not
   263  	// starting work and the mode should be ignored.
   264  	gcMarkWorkerNotWorker gcMarkWorkerMode = iota
   265  
   266  	// gcMarkWorkerDedicatedMode indicates that the P of a mark
   267  	// worker is dedicated to running that mark worker. The mark
   268  	// worker should run without preemption.
   269  	gcMarkWorkerDedicatedMode
   270  
   271  	// gcMarkWorkerFractionalMode indicates that a P is currently
   272  	// running the "fractional" mark worker. The fractional worker
   273  	// is necessary when GOMAXPROCS*gcBackgroundUtilization is not
   274  	// an integer and using only dedicated workers would result in
   275  	// utilization too far from the target of gcBackgroundUtilization.
   276  	// The fractional worker should run until it is preempted and
   277  	// will be scheduled to pick up the fractional part of
   278  	// GOMAXPROCS*gcBackgroundUtilization.
   279  	gcMarkWorkerFractionalMode
   280  
   281  	// gcMarkWorkerIdleMode indicates that a P is running the mark
   282  	// worker because it has nothing else to do. The idle worker
   283  	// should run until it is preempted and account its time
   284  	// against gcController.idleMarkTime.
   285  	gcMarkWorkerIdleMode
   286  )
   287  
   288  // gcMarkWorkerModeStrings are the strings labels of gcMarkWorkerModes
   289  // to use in execution traces.
   290  var gcMarkWorkerModeStrings = [...]string{
   291  	"Not worker",
   292  	"GC (dedicated)",
   293  	"GC (fractional)",
   294  	"GC (idle)",
   295  }
   296  
   297  // pollFractionalWorkerExit reports whether a fractional mark worker
   298  // should self-preempt. It assumes it is called from the fractional
   299  // worker.
   300  func pollFractionalWorkerExit() bool {
   301  	// This should be kept in sync with the fractional worker
   302  	// scheduler logic in findRunnableGCWorker.
   303  	now := nanotime()
   304  	delta := now - gcController.markStartTime
   305  	if delta <= 0 {
   306  		return true
   307  	}
   308  	p := getg().m.p.ptr()
   309  	selfTime := p.gcFractionalMarkTime + (now - p.gcMarkWorkerStartTime)
   310  	// Add some slack to the utilization goal so that the
   311  	// fractional worker isn't behind again the instant it exits.
   312  	return float64(selfTime)/float64(delta) > 1.2*gcController.fractionalUtilizationGoal
   313  }
   314  
   315  var work workType
   316  
   317  type workType struct {
   318  	full  lfstack          // lock-free list of full blocks workbuf
   319  	_     cpu.CacheLinePad // prevents false-sharing between full and empty
   320  	empty lfstack          // lock-free list of empty blocks workbuf
   321  	_     cpu.CacheLinePad // prevents false-sharing between empty and nproc/nwait
   322  
   323  	wbufSpans struct {
   324  		lock mutex
   325  		// free is a list of spans dedicated to workbufs, but
   326  		// that don't currently contain any workbufs.
   327  		free mSpanList
   328  		// busy is a list of all spans containing workbufs on
   329  		// one of the workbuf lists.
   330  		busy mSpanList
   331  	}
   332  
   333  	// Restore 64-bit alignment on 32-bit.
   334  	_ uint32
   335  
   336  	// bytesMarked is the number of bytes marked this cycle. This
   337  	// includes bytes blackened in scanned objects, noscan objects
   338  	// that go straight to black, objects allocated as black during
   339  	// the cycle, and permagrey objects scanned by markroot during
   340  	// the concurrent scan phase.
   341  	//
   342  	// This is updated atomically during the cycle. Updates may be batched
   343  	// arbitrarily, since the value is only read at the end of the cycle.
   344  	//
   345  	// Because of benign races during marking, this number may not
   346  	// be the exact number of marked bytes, but it should be very
   347  	// close.
   348  	//
   349  	// Put this field here because it needs 64-bit atomic access
   350  	// (and thus 8-byte alignment even on 32-bit architectures).
   351  	bytesMarked uint64
   352  
   353  	markrootNext uint32 // next markroot job
   354  	markrootJobs uint32 // number of markroot jobs
   355  
   356  	nproc  uint32
   357  	tstart int64
   358  	nwait  uint32
   359  
   360  	// Number of roots of various root types. Set by gcMarkRootPrepare.
   361  	//
   362  	// nStackRoots == len(stackRoots), but we have nStackRoots for
   363  	// consistency.
   364  	nDataRoots, nBSSRoots, nSpanRoots, nStackRoots int
   365  
   366  	// Base indexes of each root type. Set by gcMarkRootPrepare.
   367  	baseData, baseBSS, baseSpans, baseStacks, baseEnd uint32
   368  
   369  	// stackRoots is a snapshot of all of the Gs that existed
   370  	// before the beginning of concurrent marking. The backing
   371  	// store of this must not be modified because it might be
   372  	// shared with allgs.
   373  	stackRoots []*g
   374  
   375  	// Each type of GC state transition is protected by a lock.
   376  	// Since multiple threads can simultaneously detect the state
   377  	// transition condition, any thread that detects a transition
   378  	// condition must acquire the appropriate transition lock,
   379  	// re-check the transition condition and return if it no
   380  	// longer holds or perform the transition if it does.
   381  	// Likewise, any transition must invalidate the transition
   382  	// condition before releasing the lock. This ensures that each
   383  	// transition is performed by exactly one thread and threads
   384  	// that need the transition to happen block until it has
   385  	// happened.
   386  	//
   387  	// startSema protects the transition from "off" to mark or
   388  	// mark termination.
   389  	startSema uint32
   390  	// markDoneSema protects transitions from mark to mark termination.
   391  	markDoneSema uint32
   392  
   393  	bgMarkDone uint32 // cas to 1 when at a background mark completion point
   394  	// Background mark completion signaling
   395  
   396  	// mode is the concurrency mode of the current GC cycle.
   397  	mode gcMode
   398  
   399  	// userForced indicates the current GC cycle was forced by an
   400  	// explicit user call.
   401  	userForced bool
   402  
   403  	// initialHeapLive is the value of gcController.heapLive at the
   404  	// beginning of this GC cycle.
   405  	initialHeapLive uint64
   406  
   407  	// assistQueue is a queue of assists that are blocked because
   408  	// there was neither enough credit to steal or enough work to
   409  	// do.
   410  	assistQueue struct {
   411  		lock mutex
   412  		q    gQueue
   413  	}
   414  
   415  	// sweepWaiters is a list of blocked goroutines to wake when
   416  	// we transition from mark termination to sweep.
   417  	sweepWaiters struct {
   418  		lock mutex
   419  		list gList
   420  	}
   421  
   422  	// strongFromWeak controls how the GC interacts with weak->strong
   423  	// pointer conversions.
   424  	strongFromWeak struct {
   425  		// block is a flag set during mark termination that prevents
   426  		// new weak->strong conversions from executing by blocking the
   427  		// goroutine and enqueuing it onto q.
   428  		//
   429  		// Mutated only by one goroutine at a time in gcMarkDone,
   430  		// with globally-synchronizing events like forEachP and
   431  		// stopTheWorld.
   432  		block bool
   433  
   434  		// q is a queue of goroutines that attempted to perform a
   435  		// weak->strong conversion during mark termination.
   436  		//
   437  		// Protected by lock.
   438  		lock mutex
   439  		q    gQueue
   440  	}
   441  
   442  	// cycles is the number of completed GC cycles, where a GC
   443  	// cycle is sweep termination, mark, mark termination, and
   444  	// sweep. This differs from memstats.numgc, which is
   445  	// incremented at mark termination.
   446  	cycles atomic.Uint32
   447  
   448  	// Timing/utilization stats for this cycle.
   449  	stwprocs, maxprocs                 int32
   450  	tSweepTerm, tMark, tMarkTerm, tEnd int64 // nanotime() of phase start
   451  
   452  	// pauseNS is the total STW time this cycle, measured as the time between
   453  	// when stopping began (just before trying to stop Ps) and just after the
   454  	// world started again.
   455  	pauseNS int64
   456  
   457  	// debug.gctrace heap sizes for this cycle.
   458  	heap0, heap1, heap2 uint64
   459  
   460  	// Cumulative estimated CPU usage.
   461  	cpuStats
   462  }
   463  
   464  // GC runs a garbage collection and blocks the caller until the
   465  // garbage collection is complete. It may also block the entire
   466  // program.
   467  func GC() {
   468  	// We consider a cycle to be: sweep termination, mark, mark
   469  	// termination, and sweep. This function shouldn't return
   470  	// until a full cycle has been completed, from beginning to
   471  	// end. Hence, we always want to finish up the current cycle
   472  	// and start a new one. That means:
   473  	//
   474  	// 1. In sweep termination, mark, or mark termination of cycle
   475  	// N, wait until mark termination N completes and transitions
   476  	// to sweep N.
   477  	//
   478  	// 2. In sweep N, help with sweep N.
   479  	//
   480  	// At this point we can begin a full cycle N+1.
   481  	//
   482  	// 3. Trigger cycle N+1 by starting sweep termination N+1.
   483  	//
   484  	// 4. Wait for mark termination N+1 to complete.
   485  	//
   486  	// 5. Help with sweep N+1 until it's done.
   487  	//
   488  	// This all has to be written to deal with the fact that the
   489  	// GC may move ahead on its own. For example, when we block
   490  	// until mark termination N, we may wake up in cycle N+2.
   491  
   492  	// Wait until the current sweep termination, mark, and mark
   493  	// termination complete.
   494  	n := work.cycles.Load()
   495  	gcWaitOnMark(n)
   496  
   497  	// We're now in sweep N or later. Trigger GC cycle N+1, which
   498  	// will first finish sweep N if necessary and then enter sweep
   499  	// termination N+1.
   500  	gcStart(gcTrigger{kind: gcTriggerCycle, n: n + 1})
   501  
   502  	// Wait for mark termination N+1 to complete.
   503  	gcWaitOnMark(n + 1)
   504  
   505  	// Finish sweep N+1 before returning. We do this both to
   506  	// complete the cycle and because runtime.GC() is often used
   507  	// as part of tests and benchmarks to get the system into a
   508  	// relatively stable and isolated state.
   509  	for work.cycles.Load() == n+1 && sweepone() != ^uintptr(0) {
   510  		Gosched()
   511  	}
   512  
   513  	// Callers may assume that the heap profile reflects the
   514  	// just-completed cycle when this returns (historically this
   515  	// happened because this was a STW GC), but right now the
   516  	// profile still reflects mark termination N, not N+1.
   517  	//
   518  	// As soon as all of the sweep frees from cycle N+1 are done,
   519  	// we can go ahead and publish the heap profile.
   520  	//
   521  	// First, wait for sweeping to finish. (We know there are no
   522  	// more spans on the sweep queue, but we may be concurrently
   523  	// sweeping spans, so we have to wait.)
   524  	for work.cycles.Load() == n+1 && !isSweepDone() {
   525  		Gosched()
   526  	}
   527  
   528  	// Now we're really done with sweeping, so we can publish the
   529  	// stable heap profile. Only do this if we haven't already hit
   530  	// another mark termination.
   531  	mp := acquirem()
   532  	cycle := work.cycles.Load()
   533  	if cycle == n+1 || (gcphase == _GCmark && cycle == n+2) {
   534  		mProf_PostSweep()
   535  	}
   536  	releasem(mp)
   537  }
   538  
   539  // gcWaitOnMark blocks until GC finishes the Nth mark phase. If GC has
   540  // already completed this mark phase, it returns immediately.
   541  func gcWaitOnMark(n uint32) {
   542  	for {
   543  		// Disable phase transitions.
   544  		lock(&work.sweepWaiters.lock)
   545  		nMarks := work.cycles.Load()
   546  		if gcphase != _GCmark {
   547  			// We've already completed this cycle's mark.
   548  			nMarks++
   549  		}
   550  		if nMarks > n {
   551  			// We're done.
   552  			unlock(&work.sweepWaiters.lock)
   553  			return
   554  		}
   555  
   556  		// Wait until sweep termination, mark, and mark
   557  		// termination of cycle N complete.
   558  		work.sweepWaiters.list.push(getg())
   559  		goparkunlock(&work.sweepWaiters.lock, waitReasonWaitForGCCycle, traceBlockUntilGCEnds, 1)
   560  	}
   561  }
   562  
   563  // gcMode indicates how concurrent a GC cycle should be.
   564  type gcMode int
   565  
   566  const (
   567  	gcBackgroundMode gcMode = iota // concurrent GC and sweep
   568  	gcForceMode                    // stop-the-world GC now, concurrent sweep
   569  	gcForceBlockMode               // stop-the-world GC now and STW sweep (forced by user)
   570  )
   571  
   572  // A gcTrigger is a predicate for starting a GC cycle. Specifically,
   573  // it is an exit condition for the _GCoff phase.
   574  type gcTrigger struct {
   575  	kind gcTriggerKind
   576  	now  int64  // gcTriggerTime: current time
   577  	n    uint32 // gcTriggerCycle: cycle number to start
   578  }
   579  
   580  type gcTriggerKind int
   581  
   582  const (
   583  	// gcTriggerHeap indicates that a cycle should be started when
   584  	// the heap size reaches the trigger heap size computed by the
   585  	// controller.
   586  	gcTriggerHeap gcTriggerKind = iota
   587  
   588  	// gcTriggerTime indicates that a cycle should be started when
   589  	// it's been more than forcegcperiod nanoseconds since the
   590  	// previous GC cycle.
   591  	gcTriggerTime
   592  
   593  	// gcTriggerCycle indicates that a cycle should be started if
   594  	// we have not yet started cycle number gcTrigger.n (relative
   595  	// to work.cycles).
   596  	gcTriggerCycle
   597  )
   598  
   599  // test reports whether the trigger condition is satisfied, meaning
   600  // that the exit condition for the _GCoff phase has been met. The exit
   601  // condition should be tested when allocating.
   602  func (t gcTrigger) test() bool {
   603  	if !memstats.enablegc || panicking.Load() != 0 || gcphase != _GCoff {
   604  		return false
   605  	}
   606  	switch t.kind {
   607  	case gcTriggerHeap:
   608  		trigger, _ := gcController.trigger()
   609  		return gcController.heapLive.Load() >= trigger
   610  	case gcTriggerTime:
   611  		if gcController.gcPercent.Load() < 0 {
   612  			return false
   613  		}
   614  		lastgc := int64(atomic.Load64(&memstats.last_gc_nanotime))
   615  		return lastgc != 0 && t.now-lastgc > forcegcperiod
   616  	case gcTriggerCycle:
   617  		// t.n > work.cycles, but accounting for wraparound.
   618  		return int32(t.n-work.cycles.Load()) > 0
   619  	}
   620  	return true
   621  }
   622  
   623  // gcStart starts the GC. It transitions from _GCoff to _GCmark (if
   624  // debug.gcstoptheworld == 0) or performs all of GC (if
   625  // debug.gcstoptheworld != 0).
   626  //
   627  // This may return without performing this transition in some cases,
   628  // such as when called on a system stack or with locks held.
   629  func gcStart(trigger gcTrigger) {
   630  	// Since this is called from malloc and malloc is called in
   631  	// the guts of a number of libraries that might be holding
   632  	// locks, don't attempt to start GC in non-preemptible or
   633  	// potentially unstable situations.
   634  	mp := acquirem()
   635  	if gp := getg(); gp == mp.g0 || mp.locks > 1 || mp.preemptoff != "" {
   636  		releasem(mp)
   637  		return
   638  	}
   639  	releasem(mp)
   640  	mp = nil
   641  
   642  	if gp := getg(); gp.syncGroup != nil {
   643  		// Disassociate the G from its synctest bubble while allocating.
   644  		// This is less elegant than incrementing the group's active count,
   645  		// but avoids any contamination between GC and synctest.
   646  		sg := gp.syncGroup
   647  		gp.syncGroup = nil
   648  		defer func() {
   649  			gp.syncGroup = sg
   650  		}()
   651  	}
   652  
   653  	// Pick up the remaining unswept/not being swept spans concurrently
   654  	//
   655  	// This shouldn't happen if we're being invoked in background
   656  	// mode since proportional sweep should have just finished
   657  	// sweeping everything, but rounding errors, etc, may leave a
   658  	// few spans unswept. In forced mode, this is necessary since
   659  	// GC can be forced at any point in the sweeping cycle.
   660  	//
   661  	// We check the transition condition continuously here in case
   662  	// this G gets delayed in to the next GC cycle.
   663  	for trigger.test() && sweepone() != ^uintptr(0) {
   664  	}
   665  
   666  	// Perform GC initialization and the sweep termination
   667  	// transition.
   668  	semacquire(&work.startSema)
   669  	// Re-check transition condition under transition lock.
   670  	if !trigger.test() {
   671  		semrelease(&work.startSema)
   672  		return
   673  	}
   674  
   675  	// In gcstoptheworld debug mode, upgrade the mode accordingly.
   676  	// We do this after re-checking the transition condition so
   677  	// that multiple goroutines that detect the heap trigger don't
   678  	// start multiple STW GCs.
   679  	mode := gcBackgroundMode
   680  	if debug.gcstoptheworld == 1 {
   681  		mode = gcForceMode
   682  	} else if debug.gcstoptheworld == 2 {
   683  		mode = gcForceBlockMode
   684  	}
   685  
   686  	// Ok, we're doing it! Stop everybody else
   687  	semacquire(&gcsema)
   688  	semacquire(&worldsema)
   689  
   690  	// For stats, check if this GC was forced by the user.
   691  	// Update it under gcsema to avoid gctrace getting wrong values.
   692  	work.userForced = trigger.kind == gcTriggerCycle
   693  
   694  	trace := traceAcquire()
   695  	if trace.ok() {
   696  		trace.GCStart()
   697  		traceRelease(trace)
   698  	}
   699  
   700  	// Check that all Ps have finished deferred mcache flushes.
   701  	for _, p := range allp {
   702  		if fg := p.mcache.flushGen.Load(); fg != mheap_.sweepgen {
   703  			println("runtime: p", p.id, "flushGen", fg, "!= sweepgen", mheap_.sweepgen)
   704  			throw("p mcache not flushed")
   705  		}
   706  	}
   707  
   708  	gcBgMarkStartWorkers()
   709  
   710  	systemstack(gcResetMarkState)
   711  
   712  	work.stwprocs, work.maxprocs = gomaxprocs, gomaxprocs
   713  	if work.stwprocs > ncpu {
   714  		// This is used to compute CPU time of the STW phases,
   715  		// so it can't be more than ncpu, even if GOMAXPROCS is.
   716  		work.stwprocs = ncpu
   717  	}
   718  	work.heap0 = gcController.heapLive.Load()
   719  	work.pauseNS = 0
   720  	work.mode = mode
   721  
   722  	now := nanotime()
   723  	work.tSweepTerm = now
   724  	var stw worldStop
   725  	systemstack(func() {
   726  		stw = stopTheWorldWithSema(stwGCSweepTerm)
   727  	})
   728  
   729  	// Accumulate fine-grained stopping time.
   730  	work.cpuStats.accumulateGCPauseTime(stw.stoppingCPUTime, 1)
   731  
   732  	// Finish sweep before we start concurrent scan.
   733  	systemstack(func() {
   734  		finishsweep_m()
   735  	})
   736  
   737  	// clearpools before we start the GC. If we wait the memory will not be
   738  	// reclaimed until the next GC cycle.
   739  	clearpools()
   740  
   741  	work.cycles.Add(1)
   742  
   743  	// Assists and workers can start the moment we start
   744  	// the world.
   745  	gcController.startCycle(now, int(gomaxprocs), trigger)
   746  
   747  	// Notify the CPU limiter that assists may begin.
   748  	gcCPULimiter.startGCTransition(true, now)
   749  
   750  	// In STW mode, disable scheduling of user Gs. This may also
   751  	// disable scheduling of this goroutine, so it may block as
   752  	// soon as we start the world again.
   753  	if mode != gcBackgroundMode {
   754  		schedEnableUser(false)
   755  	}
   756  
   757  	// Enter concurrent mark phase and enable
   758  	// write barriers.
   759  	//
   760  	// Because the world is stopped, all Ps will
   761  	// observe that write barriers are enabled by
   762  	// the time we start the world and begin
   763  	// scanning.
   764  	//
   765  	// Write barriers must be enabled before assists are
   766  	// enabled because they must be enabled before
   767  	// any non-leaf heap objects are marked. Since
   768  	// allocations are blocked until assists can
   769  	// happen, we want to enable assists as early as
   770  	// possible.
   771  	setGCPhase(_GCmark)
   772  
   773  	gcBgMarkPrepare() // Must happen before assists are enabled.
   774  	gcMarkRootPrepare()
   775  
   776  	// Mark all active tinyalloc blocks. Since we're
   777  	// allocating from these, they need to be black like
   778  	// other allocations. The alternative is to blacken
   779  	// the tiny block on every allocation from it, which
   780  	// would slow down the tiny allocator.
   781  	gcMarkTinyAllocs()
   782  
   783  	// At this point all Ps have enabled the write
   784  	// barrier, thus maintaining the no white to
   785  	// black invariant. Enable mutator assists to
   786  	// put back-pressure on fast allocating
   787  	// mutators.
   788  	atomic.Store(&gcBlackenEnabled, 1)
   789  
   790  	// In STW mode, we could block the instant systemstack
   791  	// returns, so make sure we're not preemptible.
   792  	mp = acquirem()
   793  
   794  	// Update the CPU stats pause time.
   795  	//
   796  	// Use maxprocs instead of stwprocs here because the total time
   797  	// computed in the CPU stats is based on maxprocs, and we want them
   798  	// to be comparable.
   799  	work.cpuStats.accumulateGCPauseTime(nanotime()-stw.finishedStopping, work.maxprocs)
   800  
   801  	// Concurrent mark.
   802  	systemstack(func() {
   803  		now = startTheWorldWithSema(0, stw)
   804  		work.pauseNS += now - stw.startedStopping
   805  		work.tMark = now
   806  
   807  		// Release the CPU limiter.
   808  		gcCPULimiter.finishGCTransition(now)
   809  	})
   810  
   811  	// Release the world sema before Gosched() in STW mode
   812  	// because we will need to reacquire it later but before
   813  	// this goroutine becomes runnable again, and we could
   814  	// self-deadlock otherwise.
   815  	semrelease(&worldsema)
   816  	releasem(mp)
   817  
   818  	// Make sure we block instead of returning to user code
   819  	// in STW mode.
   820  	if mode != gcBackgroundMode {
   821  		Gosched()
   822  	}
   823  
   824  	semrelease(&work.startSema)
   825  }
   826  
   827  // gcMarkDoneFlushed counts the number of P's with flushed work.
   828  //
   829  // Ideally this would be a captured local in gcMarkDone, but forEachP
   830  // escapes its callback closure, so it can't capture anything.
   831  //
   832  // This is protected by markDoneSema.
   833  var gcMarkDoneFlushed uint32
   834  
   835  // gcDebugMarkDone contains fields used to debug/test mark termination.
   836  var gcDebugMarkDone struct {
   837  	// spinAfterRaggedBarrier forces gcMarkDone to spin after it executes
   838  	// the ragged barrier.
   839  	spinAfterRaggedBarrier atomic.Bool
   840  
   841  	// restartedDueTo27993 indicates that we restarted mark termination
   842  	// due to the bug described in issue #27993.
   843  	//
   844  	// Protected by worldsema.
   845  	restartedDueTo27993 bool
   846  }
   847  
   848  // gcMarkDone transitions the GC from mark to mark termination if all
   849  // reachable objects have been marked (that is, there are no grey
   850  // objects and can be no more in the future). Otherwise, it flushes
   851  // all local work to the global queues where it can be discovered by
   852  // other workers.
   853  //
   854  // This should be called when all local mark work has been drained and
   855  // there are no remaining workers. Specifically, when
   856  //
   857  //	work.nwait == work.nproc && !gcMarkWorkAvailable(p)
   858  //
   859  // The calling context must be preemptible.
   860  //
   861  // Flushing local work is important because idle Ps may have local
   862  // work queued. This is the only way to make that work visible and
   863  // drive GC to completion.
   864  //
   865  // It is explicitly okay to have write barriers in this function. If
   866  // it does transition to mark termination, then all reachable objects
   867  // have been marked, so the write barrier cannot shade any more
   868  // objects.
   869  func gcMarkDone() {
   870  	// Ensure only one thread is running the ragged barrier at a
   871  	// time.
   872  	semacquire(&work.markDoneSema)
   873  
   874  top:
   875  	// Re-check transition condition under transition lock.
   876  	//
   877  	// It's critical that this checks the global work queues are
   878  	// empty before performing the ragged barrier. Otherwise,
   879  	// there could be global work that a P could take after the P
   880  	// has passed the ragged barrier.
   881  	if !(gcphase == _GCmark && work.nwait == work.nproc && !gcMarkWorkAvailable(nil)) {
   882  		semrelease(&work.markDoneSema)
   883  		return
   884  	}
   885  
   886  	// forEachP needs worldsema to execute, and we'll need it to
   887  	// stop the world later, so acquire worldsema now.
   888  	semacquire(&worldsema)
   889  
   890  	// Prevent weak->strong conversions from generating additional
   891  	// GC work. forEachP will guarantee that it is observed globally.
   892  	work.strongFromWeak.block = true
   893  
   894  	// Flush all local buffers and collect flushedWork flags.
   895  	gcMarkDoneFlushed = 0
   896  	forEachP(waitReasonGCMarkTermination, func(pp *p) {
   897  		// Flush the write barrier buffer, since this may add
   898  		// work to the gcWork.
   899  		wbBufFlush1(pp)
   900  
   901  		// Flush the gcWork, since this may create global work
   902  		// and set the flushedWork flag.
   903  		//
   904  		// TODO(austin): Break up these workbufs to
   905  		// better distribute work.
   906  		pp.gcw.dispose()
   907  		// Collect the flushedWork flag.
   908  		if pp.gcw.flushedWork {
   909  			atomic.Xadd(&gcMarkDoneFlushed, 1)
   910  			pp.gcw.flushedWork = false
   911  		}
   912  	})
   913  
   914  	if gcMarkDoneFlushed != 0 {
   915  		// More grey objects were discovered since the
   916  		// previous termination check, so there may be more
   917  		// work to do. Keep going. It's possible the
   918  		// transition condition became true again during the
   919  		// ragged barrier, so re-check it.
   920  		semrelease(&worldsema)
   921  		goto top
   922  	}
   923  
   924  	// For debugging/testing.
   925  	for gcDebugMarkDone.spinAfterRaggedBarrier.Load() {
   926  	}
   927  
   928  	// There was no global work, no local work, and no Ps
   929  	// communicated work since we took markDoneSema. Therefore
   930  	// there are no grey objects and no more objects can be
   931  	// shaded. Transition to mark termination.
   932  	now := nanotime()
   933  	work.tMarkTerm = now
   934  	getg().m.preemptoff = "gcing"
   935  	var stw worldStop
   936  	systemstack(func() {
   937  		stw = stopTheWorldWithSema(stwGCMarkTerm)
   938  	})
   939  	// The gcphase is _GCmark, it will transition to _GCmarktermination
   940  	// below. The important thing is that the wb remains active until
   941  	// all marking is complete. This includes writes made by the GC.
   942  
   943  	// Accumulate fine-grained stopping time.
   944  	work.cpuStats.accumulateGCPauseTime(stw.stoppingCPUTime, 1)
   945  
   946  	// There is sometimes work left over when we enter mark termination due
   947  	// to write barriers performed after the completion barrier above.
   948  	// Detect this and resume concurrent mark. This is obviously
   949  	// unfortunate.
   950  	//
   951  	// See issue #27993 for details.
   952  	//
   953  	// Switch to the system stack to call wbBufFlush1, though in this case
   954  	// it doesn't matter because we're non-preemptible anyway.
   955  	restart := false
   956  	systemstack(func() {
   957  		for _, p := range allp {
   958  			wbBufFlush1(p)
   959  			if !p.gcw.empty() {
   960  				restart = true
   961  				break
   962  			}
   963  		}
   964  	})
   965  	if restart {
   966  		gcDebugMarkDone.restartedDueTo27993 = true
   967  
   968  		getg().m.preemptoff = ""
   969  		systemstack(func() {
   970  			// Accumulate the time we were stopped before we had to start again.
   971  			work.cpuStats.accumulateGCPauseTime(nanotime()-stw.finishedStopping, work.maxprocs)
   972  
   973  			// Start the world again.
   974  			now := startTheWorldWithSema(0, stw)
   975  			work.pauseNS += now - stw.startedStopping
   976  		})
   977  		semrelease(&worldsema)
   978  		goto top
   979  	}
   980  
   981  	gcComputeStartingStackSize()
   982  
   983  	// Disable assists and background workers. We must do
   984  	// this before waking blocked assists.
   985  	atomic.Store(&gcBlackenEnabled, 0)
   986  
   987  	// Notify the CPU limiter that GC assists will now cease.
   988  	gcCPULimiter.startGCTransition(false, now)
   989  
   990  	// Wake all blocked assists. These will run when we
   991  	// start the world again.
   992  	gcWakeAllAssists()
   993  
   994  	// Wake all blocked weak->strong conversions. These will run
   995  	// when we start the world again.
   996  	work.strongFromWeak.block = false
   997  	gcWakeAllStrongFromWeak()
   998  
   999  	// Likewise, release the transition lock. Blocked
  1000  	// workers and assists will run when we start the
  1001  	// world again.
  1002  	semrelease(&work.markDoneSema)
  1003  
  1004  	// In STW mode, re-enable user goroutines. These will be
  1005  	// queued to run after we start the world.
  1006  	schedEnableUser(true)
  1007  
  1008  	// endCycle depends on all gcWork cache stats being flushed.
  1009  	// The termination algorithm above ensured that up to
  1010  	// allocations since the ragged barrier.
  1011  	gcController.endCycle(now, int(gomaxprocs), work.userForced)
  1012  
  1013  	// Perform mark termination. This will restart the world.
  1014  	gcMarkTermination(stw)
  1015  }
  1016  
  1017  // World must be stopped and mark assists and background workers must be
  1018  // disabled.
  1019  func gcMarkTermination(stw worldStop) {
  1020  	// Start marktermination (write barrier remains enabled for now).
  1021  	setGCPhase(_GCmarktermination)
  1022  
  1023  	work.heap1 = gcController.heapLive.Load()
  1024  	startTime := nanotime()
  1025  
  1026  	mp := acquirem()
  1027  	mp.preemptoff = "gcing"
  1028  	mp.traceback = 2
  1029  	curgp := mp.curg
  1030  	// N.B. The execution tracer is not aware of this status
  1031  	// transition and handles it specially based on the
  1032  	// wait reason.
  1033  	casGToWaitingForGC(curgp, _Grunning, waitReasonGarbageCollection)
  1034  
  1035  	// Run gc on the g0 stack. We do this so that the g stack
  1036  	// we're currently running on will no longer change. Cuts
  1037  	// the root set down a bit (g0 stacks are not scanned, and
  1038  	// we don't need to scan gc's internal state).  We also
  1039  	// need to switch to g0 so we can shrink the stack.
  1040  	systemstack(func() {
  1041  		gcMark(startTime)
  1042  		// Must return immediately.
  1043  		// The outer function's stack may have moved
  1044  		// during gcMark (it shrinks stacks, including the
  1045  		// outer function's stack), so we must not refer
  1046  		// to any of its variables. Return back to the
  1047  		// non-system stack to pick up the new addresses
  1048  		// before continuing.
  1049  	})
  1050  
  1051  	var stwSwept bool
  1052  	systemstack(func() {
  1053  		work.heap2 = work.bytesMarked
  1054  		if debug.gccheckmark > 0 {
  1055  			// Run a full non-parallel, stop-the-world
  1056  			// mark using checkmark bits, to check that we
  1057  			// didn't forget to mark anything during the
  1058  			// concurrent mark process.
  1059  			startCheckmarks()
  1060  			gcResetMarkState()
  1061  			gcw := &getg().m.p.ptr().gcw
  1062  			gcDrain(gcw, 0)
  1063  			wbBufFlush1(getg().m.p.ptr())
  1064  			gcw.dispose()
  1065  			endCheckmarks()
  1066  		}
  1067  
  1068  		// marking is complete so we can turn the write barrier off
  1069  		setGCPhase(_GCoff)
  1070  		stwSwept = gcSweep(work.mode)
  1071  	})
  1072  
  1073  	mp.traceback = 0
  1074  	casgstatus(curgp, _Gwaiting, _Grunning)
  1075  
  1076  	trace := traceAcquire()
  1077  	if trace.ok() {
  1078  		trace.GCDone()
  1079  		traceRelease(trace)
  1080  	}
  1081  
  1082  	// all done
  1083  	mp.preemptoff = ""
  1084  
  1085  	if gcphase != _GCoff {
  1086  		throw("gc done but gcphase != _GCoff")
  1087  	}
  1088  
  1089  	// Record heapInUse for scavenger.
  1090  	memstats.lastHeapInUse = gcController.heapInUse.load()
  1091  
  1092  	// Update GC trigger and pacing, as well as downstream consumers
  1093  	// of this pacing information, for the next cycle.
  1094  	systemstack(gcControllerCommit)
  1095  
  1096  	// Update timing memstats
  1097  	now := nanotime()
  1098  	sec, nsec, _ := time_now()
  1099  	unixNow := sec*1e9 + int64(nsec)
  1100  	work.pauseNS += now - stw.startedStopping
  1101  	work.tEnd = now
  1102  	atomic.Store64(&memstats.last_gc_unix, uint64(unixNow)) // must be Unix time to make sense to user
  1103  	atomic.Store64(&memstats.last_gc_nanotime, uint64(now)) // monotonic time for us
  1104  	memstats.pause_ns[memstats.numgc%uint32(len(memstats.pause_ns))] = uint64(work.pauseNS)
  1105  	memstats.pause_end[memstats.numgc%uint32(len(memstats.pause_end))] = uint64(unixNow)
  1106  	memstats.pause_total_ns += uint64(work.pauseNS)
  1107  
  1108  	// Accumulate CPU stats.
  1109  	//
  1110  	// Use maxprocs instead of stwprocs for GC pause time because the total time
  1111  	// computed in the CPU stats is based on maxprocs, and we want them to be
  1112  	// comparable.
  1113  	//
  1114  	// Pass gcMarkPhase=true to accumulate so we can get all the latest GC CPU stats
  1115  	// in there too.
  1116  	work.cpuStats.accumulateGCPauseTime(now-stw.finishedStopping, work.maxprocs)
  1117  	work.cpuStats.accumulate(now, true)
  1118  
  1119  	// Compute overall GC CPU utilization.
  1120  	// Omit idle marking time from the overall utilization here since it's "free".
  1121  	memstats.gc_cpu_fraction = float64(work.cpuStats.GCTotalTime-work.cpuStats.GCIdleTime) / float64(work.cpuStats.TotalTime)
  1122  
  1123  	// Reset assist time and background time stats.
  1124  	//
  1125  	// Do this now, instead of at the start of the next GC cycle, because
  1126  	// these two may keep accumulating even if the GC is not active.
  1127  	scavenge.assistTime.Store(0)
  1128  	scavenge.backgroundTime.Store(0)
  1129  
  1130  	// Reset idle time stat.
  1131  	sched.idleTime.Store(0)
  1132  
  1133  	if work.userForced {
  1134  		memstats.numforcedgc++
  1135  	}
  1136  
  1137  	// Bump GC cycle count and wake goroutines waiting on sweep.
  1138  	lock(&work.sweepWaiters.lock)
  1139  	memstats.numgc++
  1140  	injectglist(&work.sweepWaiters.list)
  1141  	unlock(&work.sweepWaiters.lock)
  1142  
  1143  	// Increment the scavenge generation now.
  1144  	//
  1145  	// This moment represents peak heap in use because we're
  1146  	// about to start sweeping.
  1147  	mheap_.pages.scav.index.nextGen()
  1148  
  1149  	// Release the CPU limiter.
  1150  	gcCPULimiter.finishGCTransition(now)
  1151  
  1152  	// Finish the current heap profiling cycle and start a new
  1153  	// heap profiling cycle. We do this before starting the world
  1154  	// so events don't leak into the wrong cycle.
  1155  	mProf_NextCycle()
  1156  
  1157  	// There may be stale spans in mcaches that need to be swept.
  1158  	// Those aren't tracked in any sweep lists, so we need to
  1159  	// count them against sweep completion until we ensure all
  1160  	// those spans have been forced out.
  1161  	//
  1162  	// If gcSweep fully swept the heap (for example if the sweep
  1163  	// is not concurrent due to a GODEBUG setting), then we expect
  1164  	// the sweepLocker to be invalid, since sweeping is done.
  1165  	//
  1166  	// N.B. Below we might duplicate some work from gcSweep; this is
  1167  	// fine as all that work is idempotent within a GC cycle, and
  1168  	// we're still holding worldsema so a new cycle can't start.
  1169  	sl := sweep.active.begin()
  1170  	if !stwSwept && !sl.valid {
  1171  		throw("failed to set sweep barrier")
  1172  	} else if stwSwept && sl.valid {
  1173  		throw("non-concurrent sweep failed to drain all sweep queues")
  1174  	}
  1175  
  1176  	systemstack(func() {
  1177  		// The memstats updated above must be updated with the world
  1178  		// stopped to ensure consistency of some values, such as
  1179  		// sched.idleTime and sched.totaltime. memstats also include
  1180  		// the pause time (work,pauseNS), forcing computation of the
  1181  		// total pause time before the pause actually ends.
  1182  		//
  1183  		// Here we reuse the same now for start the world so that the
  1184  		// time added to /sched/pauses/total/gc:seconds will be
  1185  		// consistent with the value in memstats.
  1186  		startTheWorldWithSema(now, stw)
  1187  	})
  1188  
  1189  	// Flush the heap profile so we can start a new cycle next GC.
  1190  	// This is relatively expensive, so we don't do it with the
  1191  	// world stopped.
  1192  	mProf_Flush()
  1193  
  1194  	// Prepare workbufs for freeing by the sweeper. We do this
  1195  	// asynchronously because it can take non-trivial time.
  1196  	prepareFreeWorkbufs()
  1197  
  1198  	// Free stack spans. This must be done between GC cycles.
  1199  	systemstack(freeStackSpans)
  1200  
  1201  	// Ensure all mcaches are flushed. Each P will flush its own
  1202  	// mcache before allocating, but idle Ps may not. Since this
  1203  	// is necessary to sweep all spans, we need to ensure all
  1204  	// mcaches are flushed before we start the next GC cycle.
  1205  	//
  1206  	// While we're here, flush the page cache for idle Ps to avoid
  1207  	// having pages get stuck on them. These pages are hidden from
  1208  	// the scavenger, so in small idle heaps a significant amount
  1209  	// of additional memory might be held onto.
  1210  	//
  1211  	// Also, flush the pinner cache, to avoid leaking that memory
  1212  	// indefinitely.
  1213  	forEachP(waitReasonFlushProcCaches, func(pp *p) {
  1214  		pp.mcache.prepareForSweep()
  1215  		if pp.status == _Pidle {
  1216  			systemstack(func() {
  1217  				lock(&mheap_.lock)
  1218  				pp.pcache.flush(&mheap_.pages)
  1219  				unlock(&mheap_.lock)
  1220  			})
  1221  		}
  1222  		pp.pinnerCache = nil
  1223  	})
  1224  	if sl.valid {
  1225  		// Now that we've swept stale spans in mcaches, they don't
  1226  		// count against unswept spans.
  1227  		//
  1228  		// Note: this sweepLocker may not be valid if sweeping had
  1229  		// already completed during the STW. See the corresponding
  1230  		// begin() call that produced sl.
  1231  		sweep.active.end(sl)
  1232  	}
  1233  
  1234  	// Print gctrace before dropping worldsema. As soon as we drop
  1235  	// worldsema another cycle could start and smash the stats
  1236  	// we're trying to print.
  1237  	if debug.gctrace > 0 {
  1238  		util := int(memstats.gc_cpu_fraction * 100)
  1239  
  1240  		var sbuf [24]byte
  1241  		printlock()
  1242  		print("gc ", memstats.numgc,
  1243  			" @", string(itoaDiv(sbuf[:], uint64(work.tSweepTerm-runtimeInitTime)/1e6, 3)), "s ",
  1244  			util, "%: ")
  1245  		prev := work.tSweepTerm
  1246  		for i, ns := range []int64{work.tMark, work.tMarkTerm, work.tEnd} {
  1247  			if i != 0 {
  1248  				print("+")
  1249  			}
  1250  			print(string(fmtNSAsMS(sbuf[:], uint64(ns-prev))))
  1251  			prev = ns
  1252  		}
  1253  		print(" ms clock, ")
  1254  		for i, ns := range []int64{
  1255  			int64(work.stwprocs) * (work.tMark - work.tSweepTerm),
  1256  			gcController.assistTime.Load(),
  1257  			gcController.dedicatedMarkTime.Load() + gcController.fractionalMarkTime.Load(),
  1258  			gcController.idleMarkTime.Load(),
  1259  			int64(work.stwprocs) * (work.tEnd - work.tMarkTerm),
  1260  		} {
  1261  			if i == 2 || i == 3 {
  1262  				// Separate mark time components with /.
  1263  				print("/")
  1264  			} else if i != 0 {
  1265  				print("+")
  1266  			}
  1267  			print(string(fmtNSAsMS(sbuf[:], uint64(ns))))
  1268  		}
  1269  		print(" ms cpu, ",
  1270  			work.heap0>>20, "->", work.heap1>>20, "->", work.heap2>>20, " MB, ",
  1271  			gcController.lastHeapGoal>>20, " MB goal, ",
  1272  			gcController.lastStackScan.Load()>>20, " MB stacks, ",
  1273  			gcController.globalsScan.Load()>>20, " MB globals, ",
  1274  			work.maxprocs, " P")
  1275  		if work.userForced {
  1276  			print(" (forced)")
  1277  		}
  1278  		print("\n")
  1279  		printunlock()
  1280  	}
  1281  
  1282  	// Set any arena chunks that were deferred to fault.
  1283  	lock(&userArenaState.lock)
  1284  	faultList := userArenaState.fault
  1285  	userArenaState.fault = nil
  1286  	unlock(&userArenaState.lock)
  1287  	for _, lc := range faultList {
  1288  		lc.mspan.setUserArenaChunkToFault()
  1289  	}
  1290  
  1291  	// Enable huge pages on some metadata if we cross a heap threshold.
  1292  	if gcController.heapGoal() > minHeapForMetadataHugePages {
  1293  		systemstack(func() {
  1294  			mheap_.enableMetadataHugePages()
  1295  		})
  1296  	}
  1297  
  1298  	semrelease(&worldsema)
  1299  	semrelease(&gcsema)
  1300  	// Careful: another GC cycle may start now.
  1301  
  1302  	releasem(mp)
  1303  	mp = nil
  1304  
  1305  	// now that gc is done, kick off finalizer thread if needed
  1306  	if !concurrentSweep {
  1307  		// give the queued finalizers, if any, a chance to run
  1308  		Gosched()
  1309  	}
  1310  }
  1311  
  1312  // gcBgMarkStartWorkers prepares background mark worker goroutines. These
  1313  // goroutines will not run until the mark phase, but they must be started while
  1314  // the work is not stopped and from a regular G stack. The caller must hold
  1315  // worldsema.
  1316  func gcBgMarkStartWorkers() {
  1317  	// Background marking is performed by per-P G's. Ensure that each P has
  1318  	// a background GC G.
  1319  	//
  1320  	// Worker Gs don't exit if gomaxprocs is reduced. If it is raised
  1321  	// again, we can reuse the old workers; no need to create new workers.
  1322  	if gcBgMarkWorkerCount >= gomaxprocs {
  1323  		return
  1324  	}
  1325  
  1326  	// Increment mp.locks when allocating. We are called within gcStart,
  1327  	// and thus must not trigger another gcStart via an allocation. gcStart
  1328  	// bails when allocating with locks held, so simulate that for these
  1329  	// allocations.
  1330  	//
  1331  	// TODO(prattmic): cleanup gcStart to use a more explicit "in gcStart"
  1332  	// check for bailing.
  1333  	mp := acquirem()
  1334  	ready := make(chan struct{}, 1)
  1335  	releasem(mp)
  1336  
  1337  	for gcBgMarkWorkerCount < gomaxprocs {
  1338  		mp := acquirem() // See above, we allocate a closure here.
  1339  		go gcBgMarkWorker(ready)
  1340  		releasem(mp)
  1341  
  1342  		// N.B. we intentionally wait on each goroutine individually
  1343  		// rather than starting all in a batch and then waiting once
  1344  		// afterwards. By running one goroutine at a time, we can take
  1345  		// advantage of runnext to bounce back and forth between
  1346  		// workers and this goroutine. In an overloaded application,
  1347  		// this can reduce GC start latency by prioritizing these
  1348  		// goroutines rather than waiting on the end of the run queue.
  1349  		<-ready
  1350  		// The worker is now guaranteed to be added to the pool before
  1351  		// its P's next findRunnableGCWorker.
  1352  
  1353  		gcBgMarkWorkerCount++
  1354  	}
  1355  }
  1356  
  1357  // gcBgMarkPrepare sets up state for background marking.
  1358  // Mutator assists must not yet be enabled.
  1359  func gcBgMarkPrepare() {
  1360  	// Background marking will stop when the work queues are empty
  1361  	// and there are no more workers (note that, since this is
  1362  	// concurrent, this may be a transient state, but mark
  1363  	// termination will clean it up). Between background workers
  1364  	// and assists, we don't really know how many workers there
  1365  	// will be, so we pretend to have an arbitrarily large number
  1366  	// of workers, almost all of which are "waiting". While a
  1367  	// worker is working it decrements nwait. If nproc == nwait,
  1368  	// there are no workers.
  1369  	work.nproc = ^uint32(0)
  1370  	work.nwait = ^uint32(0)
  1371  }
  1372  
  1373  // gcBgMarkWorkerNode is an entry in the gcBgMarkWorkerPool. It points to a single
  1374  // gcBgMarkWorker goroutine.
  1375  type gcBgMarkWorkerNode struct {
  1376  	// Unused workers are managed in a lock-free stack. This field must be first.
  1377  	node lfnode
  1378  
  1379  	// The g of this worker.
  1380  	gp guintptr
  1381  
  1382  	// Release this m on park. This is used to communicate with the unlock
  1383  	// function, which cannot access the G's stack. It is unused outside of
  1384  	// gcBgMarkWorker().
  1385  	m muintptr
  1386  }
  1387  
  1388  func gcBgMarkWorker(ready chan struct{}) {
  1389  	gp := getg()
  1390  
  1391  	// We pass node to a gopark unlock function, so it can't be on
  1392  	// the stack (see gopark). Prevent deadlock from recursively
  1393  	// starting GC by disabling preemption.
  1394  	gp.m.preemptoff = "GC worker init"
  1395  	node := new(gcBgMarkWorkerNode)
  1396  	gp.m.preemptoff = ""
  1397  
  1398  	node.gp.set(gp)
  1399  
  1400  	node.m.set(acquirem())
  1401  
  1402  	ready <- struct{}{}
  1403  	// After this point, the background mark worker is generally scheduled
  1404  	// cooperatively by gcController.findRunnableGCWorker. While performing
  1405  	// work on the P, preemption is disabled because we are working on
  1406  	// P-local work buffers. When the preempt flag is set, this puts itself
  1407  	// into _Gwaiting to be woken up by gcController.findRunnableGCWorker
  1408  	// at the appropriate time.
  1409  	//
  1410  	// When preemption is enabled (e.g., while in gcMarkDone), this worker
  1411  	// may be preempted and schedule as a _Grunnable G from a runq. That is
  1412  	// fine; it will eventually gopark again for further scheduling via
  1413  	// findRunnableGCWorker.
  1414  	//
  1415  	// Since we disable preemption before notifying ready, we guarantee that
  1416  	// this G will be in the worker pool for the next findRunnableGCWorker.
  1417  	// This isn't strictly necessary, but it reduces latency between
  1418  	// _GCmark starting and the workers starting.
  1419  
  1420  	for {
  1421  		// Go to sleep until woken by
  1422  		// gcController.findRunnableGCWorker.
  1423  		gopark(func(g *g, nodep unsafe.Pointer) bool {
  1424  			node := (*gcBgMarkWorkerNode)(nodep)
  1425  
  1426  			if mp := node.m.ptr(); mp != nil {
  1427  				// The worker G is no longer running; release
  1428  				// the M.
  1429  				//
  1430  				// N.B. it is _safe_ to release the M as soon
  1431  				// as we are no longer performing P-local mark
  1432  				// work.
  1433  				//
  1434  				// However, since we cooperatively stop work
  1435  				// when gp.preempt is set, if we releasem in
  1436  				// the loop then the following call to gopark
  1437  				// would immediately preempt the G. This is
  1438  				// also safe, but inefficient: the G must
  1439  				// schedule again only to enter gopark and park
  1440  				// again. Thus, we defer the release until
  1441  				// after parking the G.
  1442  				releasem(mp)
  1443  			}
  1444  
  1445  			// Release this G to the pool.
  1446  			gcBgMarkWorkerPool.push(&node.node)
  1447  			// Note that at this point, the G may immediately be
  1448  			// rescheduled and may be running.
  1449  			return true
  1450  		}, unsafe.Pointer(node), waitReasonGCWorkerIdle, traceBlockSystemGoroutine, 0)
  1451  
  1452  		// Preemption must not occur here, or another G might see
  1453  		// p.gcMarkWorkerMode.
  1454  
  1455  		// Disable preemption so we can use the gcw. If the
  1456  		// scheduler wants to preempt us, we'll stop draining,
  1457  		// dispose the gcw, and then preempt.
  1458  		node.m.set(acquirem())
  1459  		pp := gp.m.p.ptr() // P can't change with preemption disabled.
  1460  
  1461  		if gcBlackenEnabled == 0 {
  1462  			println("worker mode", pp.gcMarkWorkerMode)
  1463  			throw("gcBgMarkWorker: blackening not enabled")
  1464  		}
  1465  
  1466  		if pp.gcMarkWorkerMode == gcMarkWorkerNotWorker {
  1467  			throw("gcBgMarkWorker: mode not set")
  1468  		}
  1469  
  1470  		startTime := nanotime()
  1471  		pp.gcMarkWorkerStartTime = startTime
  1472  		var trackLimiterEvent bool
  1473  		if pp.gcMarkWorkerMode == gcMarkWorkerIdleMode {
  1474  			trackLimiterEvent = pp.limiterEvent.start(limiterEventIdleMarkWork, startTime)
  1475  		}
  1476  
  1477  		decnwait := atomic.Xadd(&work.nwait, -1)
  1478  		if decnwait == work.nproc {
  1479  			println("runtime: work.nwait=", decnwait, "work.nproc=", work.nproc)
  1480  			throw("work.nwait was > work.nproc")
  1481  		}
  1482  
  1483  		systemstack(func() {
  1484  			// Mark our goroutine preemptible so its stack
  1485  			// can be scanned. This lets two mark workers
  1486  			// scan each other (otherwise, they would
  1487  			// deadlock). We must not modify anything on
  1488  			// the G stack. However, stack shrinking is
  1489  			// disabled for mark workers, so it is safe to
  1490  			// read from the G stack.
  1491  			//
  1492  			// N.B. The execution tracer is not aware of this status
  1493  			// transition and handles it specially based on the
  1494  			// wait reason.
  1495  			casGToWaitingForGC(gp, _Grunning, waitReasonGCWorkerActive)
  1496  			switch pp.gcMarkWorkerMode {
  1497  			default:
  1498  				throw("gcBgMarkWorker: unexpected gcMarkWorkerMode")
  1499  			case gcMarkWorkerDedicatedMode:
  1500  				gcDrainMarkWorkerDedicated(&pp.gcw, true)
  1501  				if gp.preempt {
  1502  					// We were preempted. This is
  1503  					// a useful signal to kick
  1504  					// everything out of the run
  1505  					// queue so it can run
  1506  					// somewhere else.
  1507  					if drainQ, n := runqdrain(pp); n > 0 {
  1508  						lock(&sched.lock)
  1509  						globrunqputbatch(&drainQ, int32(n))
  1510  						unlock(&sched.lock)
  1511  					}
  1512  				}
  1513  				// Go back to draining, this time
  1514  				// without preemption.
  1515  				gcDrainMarkWorkerDedicated(&pp.gcw, false)
  1516  			case gcMarkWorkerFractionalMode:
  1517  				gcDrainMarkWorkerFractional(&pp.gcw)
  1518  			case gcMarkWorkerIdleMode:
  1519  				gcDrainMarkWorkerIdle(&pp.gcw)
  1520  			}
  1521  			casgstatus(gp, _Gwaiting, _Grunning)
  1522  		})
  1523  
  1524  		// Account for time and mark us as stopped.
  1525  		now := nanotime()
  1526  		duration := now - startTime
  1527  		gcController.markWorkerStop(pp.gcMarkWorkerMode, duration)
  1528  		if trackLimiterEvent {
  1529  			pp.limiterEvent.stop(limiterEventIdleMarkWork, now)
  1530  		}
  1531  		if pp.gcMarkWorkerMode == gcMarkWorkerFractionalMode {
  1532  			atomic.Xaddint64(&pp.gcFractionalMarkTime, duration)
  1533  		}
  1534  
  1535  		// Was this the last worker and did we run out
  1536  		// of work?
  1537  		incnwait := atomic.Xadd(&work.nwait, +1)
  1538  		if incnwait > work.nproc {
  1539  			println("runtime: p.gcMarkWorkerMode=", pp.gcMarkWorkerMode,
  1540  				"work.nwait=", incnwait, "work.nproc=", work.nproc)
  1541  			throw("work.nwait > work.nproc")
  1542  		}
  1543  
  1544  		// We'll releasem after this point and thus this P may run
  1545  		// something else. We must clear the worker mode to avoid
  1546  		// attributing the mode to a different (non-worker) G in
  1547  		// traceGoStart.
  1548  		pp.gcMarkWorkerMode = gcMarkWorkerNotWorker
  1549  
  1550  		// If this worker reached a background mark completion
  1551  		// point, signal the main GC goroutine.
  1552  		if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
  1553  			// We don't need the P-local buffers here, allow
  1554  			// preemption because we may schedule like a regular
  1555  			// goroutine in gcMarkDone (block on locks, etc).
  1556  			releasem(node.m.ptr())
  1557  			node.m.set(nil)
  1558  
  1559  			gcMarkDone()
  1560  		}
  1561  	}
  1562  }
  1563  
  1564  // gcMarkWorkAvailable reports whether executing a mark worker
  1565  // on p is potentially useful. p may be nil, in which case it only
  1566  // checks the global sources of work.
  1567  func gcMarkWorkAvailable(p *p) bool {
  1568  	if p != nil && !p.gcw.empty() {
  1569  		return true
  1570  	}
  1571  	if !work.full.empty() {
  1572  		return true // global work available
  1573  	}
  1574  	if work.markrootNext < work.markrootJobs {
  1575  		return true // root scan work available
  1576  	}
  1577  	return false
  1578  }
  1579  
  1580  // gcMark runs the mark (or, for concurrent GC, mark termination)
  1581  // All gcWork caches must be empty.
  1582  // STW is in effect at this point.
  1583  func gcMark(startTime int64) {
  1584  	if gcphase != _GCmarktermination {
  1585  		throw("in gcMark expecting to see gcphase as _GCmarktermination")
  1586  	}
  1587  	work.tstart = startTime
  1588  
  1589  	// Check that there's no marking work remaining.
  1590  	if work.full != 0 || work.markrootNext < work.markrootJobs {
  1591  		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")
  1592  		panic("non-empty mark queue after concurrent mark")
  1593  	}
  1594  
  1595  	if debug.gccheckmark > 0 {
  1596  		// This is expensive when there's a large number of
  1597  		// Gs, so only do it if checkmark is also enabled.
  1598  		gcMarkRootCheck()
  1599  	}
  1600  
  1601  	// Drop allg snapshot. allgs may have grown, in which case
  1602  	// this is the only reference to the old backing store and
  1603  	// there's no need to keep it around.
  1604  	work.stackRoots = nil
  1605  
  1606  	// Clear out buffers and double-check that all gcWork caches
  1607  	// are empty. This should be ensured by gcMarkDone before we
  1608  	// enter mark termination.
  1609  	//
  1610  	// TODO: We could clear out buffers just before mark if this
  1611  	// has a non-negligible impact on STW time.
  1612  	for _, p := range allp {
  1613  		// The write barrier may have buffered pointers since
  1614  		// the gcMarkDone barrier. However, since the barrier
  1615  		// ensured all reachable objects were marked, all of
  1616  		// these must be pointers to black objects. Hence we
  1617  		// can just discard the write barrier buffer.
  1618  		if debug.gccheckmark > 0 {
  1619  			// For debugging, flush the buffer and make
  1620  			// sure it really was all marked.
  1621  			wbBufFlush1(p)
  1622  		} else {
  1623  			p.wbBuf.reset()
  1624  		}
  1625  
  1626  		gcw := &p.gcw
  1627  		if !gcw.empty() {
  1628  			printlock()
  1629  			print("runtime: P ", p.id, " flushedWork ", gcw.flushedWork)
  1630  			if gcw.wbuf1 == nil {
  1631  				print(" wbuf1=<nil>")
  1632  			} else {
  1633  				print(" wbuf1.n=", gcw.wbuf1.nobj)
  1634  			}
  1635  			if gcw.wbuf2 == nil {
  1636  				print(" wbuf2=<nil>")
  1637  			} else {
  1638  				print(" wbuf2.n=", gcw.wbuf2.nobj)
  1639  			}
  1640  			print("\n")
  1641  			throw("P has cached GC work at end of mark termination")
  1642  		}
  1643  		// There may still be cached empty buffers, which we
  1644  		// need to flush since we're going to free them. Also,
  1645  		// there may be non-zero stats because we allocated
  1646  		// black after the gcMarkDone barrier.
  1647  		gcw.dispose()
  1648  	}
  1649  
  1650  	// Flush scanAlloc from each mcache since we're about to modify
  1651  	// heapScan directly. If we were to flush this later, then scanAlloc
  1652  	// might have incorrect information.
  1653  	//
  1654  	// Note that it's not important to retain this information; we know
  1655  	// exactly what heapScan is at this point via scanWork.
  1656  	for _, p := range allp {
  1657  		c := p.mcache
  1658  		if c == nil {
  1659  			continue
  1660  		}
  1661  		c.scanAlloc = 0
  1662  	}
  1663  
  1664  	// Reset controller state.
  1665  	gcController.resetLive(work.bytesMarked)
  1666  }
  1667  
  1668  // gcSweep must be called on the system stack because it acquires the heap
  1669  // lock. See mheap for details.
  1670  //
  1671  // Returns true if the heap was fully swept by this function.
  1672  //
  1673  // The world must be stopped.
  1674  //
  1675  //go:systemstack
  1676  func gcSweep(mode gcMode) bool {
  1677  	assertWorldStopped()
  1678  
  1679  	if gcphase != _GCoff {
  1680  		throw("gcSweep being done but phase is not GCoff")
  1681  	}
  1682  
  1683  	lock(&mheap_.lock)
  1684  	mheap_.sweepgen += 2
  1685  	sweep.active.reset()
  1686  	mheap_.pagesSwept.Store(0)
  1687  	mheap_.sweepArenas = mheap_.allArenas
  1688  	mheap_.reclaimIndex.Store(0)
  1689  	mheap_.reclaimCredit.Store(0)
  1690  	unlock(&mheap_.lock)
  1691  
  1692  	sweep.centralIndex.clear()
  1693  
  1694  	if !concurrentSweep || mode == gcForceBlockMode {
  1695  		// Special case synchronous sweep.
  1696  		// Record that no proportional sweeping has to happen.
  1697  		lock(&mheap_.lock)
  1698  		mheap_.sweepPagesPerByte = 0
  1699  		unlock(&mheap_.lock)
  1700  		// Flush all mcaches.
  1701  		for _, pp := range allp {
  1702  			pp.mcache.prepareForSweep()
  1703  		}
  1704  		// Sweep all spans eagerly.
  1705  		for sweepone() != ^uintptr(0) {
  1706  		}
  1707  		// Free workbufs eagerly.
  1708  		prepareFreeWorkbufs()
  1709  		for freeSomeWbufs(false) {
  1710  		}
  1711  		// All "free" events for this mark/sweep cycle have
  1712  		// now happened, so we can make this profile cycle
  1713  		// available immediately.
  1714  		mProf_NextCycle()
  1715  		mProf_Flush()
  1716  		return true
  1717  	}
  1718  
  1719  	// Background sweep.
  1720  	lock(&sweep.lock)
  1721  	if sweep.parked {
  1722  		sweep.parked = false
  1723  		ready(sweep.g, 0, true)
  1724  	}
  1725  	unlock(&sweep.lock)
  1726  	return false
  1727  }
  1728  
  1729  // gcResetMarkState resets global state prior to marking (concurrent
  1730  // or STW) and resets the stack scan state of all Gs.
  1731  //
  1732  // This is safe to do without the world stopped because any Gs created
  1733  // during or after this will start out in the reset state.
  1734  //
  1735  // gcResetMarkState must be called on the system stack because it acquires
  1736  // the heap lock. See mheap for details.
  1737  //
  1738  //go:systemstack
  1739  func gcResetMarkState() {
  1740  	// This may be called during a concurrent phase, so lock to make sure
  1741  	// allgs doesn't change.
  1742  	forEachG(func(gp *g) {
  1743  		gp.gcscandone = false // set to true in gcphasework
  1744  		gp.gcAssistBytes = 0
  1745  	})
  1746  
  1747  	// Clear page marks. This is just 1MB per 64GB of heap, so the
  1748  	// time here is pretty trivial.
  1749  	lock(&mheap_.lock)
  1750  	arenas := mheap_.allArenas
  1751  	unlock(&mheap_.lock)
  1752  	for _, ai := range arenas {
  1753  		ha := mheap_.arenas[ai.l1()][ai.l2()]
  1754  		clear(ha.pageMarks[:])
  1755  	}
  1756  
  1757  	work.bytesMarked = 0
  1758  	work.initialHeapLive = gcController.heapLive.Load()
  1759  }
  1760  
  1761  // Hooks for other packages
  1762  
  1763  var poolcleanup func()
  1764  var boringCaches []unsafe.Pointer  // for crypto/internal/boring
  1765  var uniqueMapCleanup chan struct{} // for unique
  1766  
  1767  // sync_runtime_registerPoolCleanup should be an internal detail,
  1768  // but widely used packages access it using linkname.
  1769  // Notable members of the hall of shame include:
  1770  //   - github.com/bytedance/gopkg
  1771  //   - github.com/songzhibin97/gkit
  1772  //
  1773  // Do not remove or change the type signature.
  1774  // See go.dev/issue/67401.
  1775  //
  1776  //go:linkname sync_runtime_registerPoolCleanup sync.runtime_registerPoolCleanup
  1777  func sync_runtime_registerPoolCleanup(f func()) {
  1778  	poolcleanup = f
  1779  }
  1780  
  1781  //go:linkname boring_registerCache crypto/internal/boring/bcache.registerCache
  1782  func boring_registerCache(p unsafe.Pointer) {
  1783  	boringCaches = append(boringCaches, p)
  1784  }
  1785  
  1786  //go:linkname unique_runtime_registerUniqueMapCleanup unique.runtime_registerUniqueMapCleanup
  1787  func unique_runtime_registerUniqueMapCleanup(f func()) {
  1788  	// Create the channel on the system stack so it doesn't inherit the current G's
  1789  	// synctest bubble (if any).
  1790  	systemstack(func() {
  1791  		uniqueMapCleanup = make(chan struct{}, 1)
  1792  	})
  1793  	// Start the goroutine in the runtime so it's counted as a system goroutine.
  1794  	go func(cleanup func()) {
  1795  		for {
  1796  			<-uniqueMapCleanup
  1797  			cleanup()
  1798  		}
  1799  	}(f)
  1800  }
  1801  
  1802  func clearpools() {
  1803  	// clear sync.Pools
  1804  	if poolcleanup != nil {
  1805  		poolcleanup()
  1806  	}
  1807  
  1808  	// clear boringcrypto caches
  1809  	for _, p := range boringCaches {
  1810  		atomicstorep(p, nil)
  1811  	}
  1812  
  1813  	// clear unique maps
  1814  	if uniqueMapCleanup != nil {
  1815  		select {
  1816  		case uniqueMapCleanup <- struct{}{}:
  1817  		default:
  1818  		}
  1819  	}
  1820  
  1821  	// Clear central sudog cache.
  1822  	// Leave per-P caches alone, they have strictly bounded size.
  1823  	// Disconnect cached list before dropping it on the floor,
  1824  	// so that a dangling ref to one entry does not pin all of them.
  1825  	lock(&sched.sudoglock)
  1826  	var sg, sgnext *sudog
  1827  	for sg = sched.sudogcache; sg != nil; sg = sgnext {
  1828  		sgnext = sg.next
  1829  		sg.next = nil
  1830  	}
  1831  	sched.sudogcache = nil
  1832  	unlock(&sched.sudoglock)
  1833  
  1834  	// Clear central defer pool.
  1835  	// Leave per-P pools alone, they have strictly bounded size.
  1836  	lock(&sched.deferlock)
  1837  	// disconnect cached list before dropping it on the floor,
  1838  	// so that a dangling ref to one entry does not pin all of them.
  1839  	var d, dlink *_defer
  1840  	for d = sched.deferpool; d != nil; d = dlink {
  1841  		dlink = d.link
  1842  		d.link = nil
  1843  	}
  1844  	sched.deferpool = nil
  1845  	unlock(&sched.deferlock)
  1846  }
  1847  
  1848  // Timing
  1849  
  1850  // itoaDiv formats val/(10**dec) into buf.
  1851  func itoaDiv(buf []byte, val uint64, dec int) []byte {
  1852  	i := len(buf) - 1
  1853  	idec := i - dec
  1854  	for val >= 10 || i >= idec {
  1855  		buf[i] = byte(val%10 + '0')
  1856  		i--
  1857  		if i == idec {
  1858  			buf[i] = '.'
  1859  			i--
  1860  		}
  1861  		val /= 10
  1862  	}
  1863  	buf[i] = byte(val + '0')
  1864  	return buf[i:]
  1865  }
  1866  
  1867  // fmtNSAsMS nicely formats ns nanoseconds as milliseconds.
  1868  func fmtNSAsMS(buf []byte, ns uint64) []byte {
  1869  	if ns >= 10e6 {
  1870  		// Format as whole milliseconds.
  1871  		return itoaDiv(buf, ns/1e6, 0)
  1872  	}
  1873  	// Format two digits of precision, with at most three decimal places.
  1874  	x := ns / 1e3
  1875  	if x == 0 {
  1876  		buf[0] = '0'
  1877  		return buf[:1]
  1878  	}
  1879  	dec := 3
  1880  	for x >= 100 {
  1881  		x /= 10
  1882  		dec--
  1883  	}
  1884  	return itoaDiv(buf, x, dec)
  1885  }
  1886  
  1887  // Helpers for testing GC.
  1888  
  1889  // gcTestMoveStackOnNextCall causes the stack to be moved on a call
  1890  // immediately following the call to this. It may not work correctly
  1891  // if any other work appears after this call (such as returning).
  1892  // Typically the following call should be marked go:noinline so it
  1893  // performs a stack check.
  1894  //
  1895  // In rare cases this may not cause the stack to move, specifically if
  1896  // there's a preemption between this call and the next.
  1897  func gcTestMoveStackOnNextCall() {
  1898  	gp := getg()
  1899  	gp.stackguard0 = stackForceMove
  1900  }
  1901  
  1902  // gcTestIsReachable performs a GC and returns a bit set where bit i
  1903  // is set if ptrs[i] is reachable.
  1904  func gcTestIsReachable(ptrs ...unsafe.Pointer) (mask uint64) {
  1905  	// This takes the pointers as unsafe.Pointers in order to keep
  1906  	// them live long enough for us to attach specials. After
  1907  	// that, we drop our references to them.
  1908  
  1909  	if len(ptrs) > 64 {
  1910  		panic("too many pointers for uint64 mask")
  1911  	}
  1912  
  1913  	// Block GC while we attach specials and drop our references
  1914  	// to ptrs. Otherwise, if a GC is in progress, it could mark
  1915  	// them reachable via this function before we have a chance to
  1916  	// drop them.
  1917  	semacquire(&gcsema)
  1918  
  1919  	// Create reachability specials for ptrs.
  1920  	specials := make([]*specialReachable, len(ptrs))
  1921  	for i, p := range ptrs {
  1922  		lock(&mheap_.speciallock)
  1923  		s := (*specialReachable)(mheap_.specialReachableAlloc.alloc())
  1924  		unlock(&mheap_.speciallock)
  1925  		s.special.kind = _KindSpecialReachable
  1926  		if !addspecial(p, &s.special, false) {
  1927  			throw("already have a reachable special (duplicate pointer?)")
  1928  		}
  1929  		specials[i] = s
  1930  		// Make sure we don't retain ptrs.
  1931  		ptrs[i] = nil
  1932  	}
  1933  
  1934  	semrelease(&gcsema)
  1935  
  1936  	// Force a full GC and sweep.
  1937  	GC()
  1938  
  1939  	// Process specials.
  1940  	for i, s := range specials {
  1941  		if !s.done {
  1942  			printlock()
  1943  			println("runtime: object", i, "was not swept")
  1944  			throw("IsReachable failed")
  1945  		}
  1946  		if s.reachable {
  1947  			mask |= 1 << i
  1948  		}
  1949  		lock(&mheap_.speciallock)
  1950  		mheap_.specialReachableAlloc.free(unsafe.Pointer(s))
  1951  		unlock(&mheap_.speciallock)
  1952  	}
  1953  
  1954  	return mask
  1955  }
  1956  
  1957  // gcTestPointerClass returns the category of what p points to, one of:
  1958  // "heap", "stack", "data", "bss", "other". This is useful for checking
  1959  // that a test is doing what it's intended to do.
  1960  //
  1961  // This is nosplit simply to avoid extra pointer shuffling that may
  1962  // complicate a test.
  1963  //
  1964  //go:nosplit
  1965  func gcTestPointerClass(p unsafe.Pointer) string {
  1966  	p2 := uintptr(noescape(p))
  1967  	gp := getg()
  1968  	if gp.stack.lo <= p2 && p2 < gp.stack.hi {
  1969  		return "stack"
  1970  	}
  1971  	if base, _, _ := findObject(p2, 0, 0); base != 0 {
  1972  		return "heap"
  1973  	}
  1974  	for _, datap := range activeModules() {
  1975  		if datap.data <= p2 && p2 < datap.edata || datap.noptrdata <= p2 && p2 < datap.enoptrdata {
  1976  			return "data"
  1977  		}
  1978  		if datap.bss <= p2 && p2 < datap.ebss || datap.noptrbss <= p2 && p2 <= datap.enoptrbss {
  1979  			return "bss"
  1980  		}
  1981  	}
  1982  	KeepAlive(p)
  1983  	return "other"
  1984  }
  1985  

View as plain text