Source file src/runtime/mgc.go

Documentation: runtime

     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 tracked in next_gc variable). This keeps the GC cost in linear
   117  // 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  // heapminimum is the minimum heap size at which to trigger GC.
   153  // For small heaps, this overrides the usual GOGC*live set rule.
   154  //
   155  // When there is a very small live set but a lot of allocation, simply
   156  // collecting when the heap reaches GOGC*live results in many GC
   157  // cycles and high total per-GC overhead. This minimum amortizes this
   158  // per-GC overhead while keeping the heap reasonably small.
   159  //
   160  // During initialization this is set to 4MB*GOGC/100. In the case of
   161  // GOGC==0, this will set heapminimum to 0, resulting in constant
   162  // collection even when the heap size is small, which is useful for
   163  // debugging.
   164  var heapminimum uint64 = defaultHeapMinimum
   165  
   166  // defaultHeapMinimum is the value of heapminimum for GOGC==100.
   167  const defaultHeapMinimum = 4 << 20
   168  
   169  // Initialized from $GOGC.  GOGC=off means no GC.
   170  var gcpercent int32
   171  
   172  func gcinit() {
   173  	if unsafe.Sizeof(workbuf{}) != _WorkbufSize {
   174  		throw("size of Workbuf is suboptimal")
   175  	}
   176  
   177  	// No sweep on the first cycle.
   178  	mheap_.sweepdone = 1
   179  
   180  	// Set a reasonable initial GC trigger.
   181  	memstats.triggerRatio = 7 / 8.0
   182  
   183  	// Fake a heap_marked value so it looks like a trigger at
   184  	// heapminimum is the appropriate growth from heap_marked.
   185  	// This will go into computing the initial GC goal.
   186  	memstats.heap_marked = uint64(float64(heapminimum) / (1 + memstats.triggerRatio))
   187  
   188  	// Set gcpercent from the environment. This will also compute
   189  	// and set the GC trigger and goal.
   190  	_ = setGCPercent(readgogc())
   191  
   192  	work.startSema = 1
   193  	work.markDoneSema = 1
   194  	lockInit(&work.sweepWaiters.lock, lockRankSweepWaiters)
   195  	lockInit(&work.assistQueue.lock, lockRankAssistQueue)
   196  	lockInit(&work.wbufSpans.lock, lockRankWbufSpans)
   197  }
   198  
   199  func readgogc() int32 {
   200  	p := gogetenv("GOGC")
   201  	if p == "off" {
   202  		return -1
   203  	}
   204  	if n, ok := atoi32(p); ok {
   205  		return n
   206  	}
   207  	return 100
   208  }
   209  
   210  // gcenable is called after the bulk of the runtime initialization,
   211  // just before we're about to start letting user code run.
   212  // It kicks off the background sweeper goroutine, the background
   213  // scavenger goroutine, and enables GC.
   214  func gcenable() {
   215  	// Kick off sweeping and scavenging.
   216  	c := make(chan int, 2)
   217  	go bgsweep(c)
   218  	go bgscavenge(c)
   219  	<-c
   220  	<-c
   221  	memstats.enablegc = true // now that runtime is initialized, GC is okay
   222  }
   223  
   224  //go:linkname setGCPercent runtime/debug.setGCPercent
   225  func setGCPercent(in int32) (out int32) {
   226  	// Run on the system stack since we grab the heap lock.
   227  	systemstack(func() {
   228  		lock(&mheap_.lock)
   229  		out = gcpercent
   230  		if in < 0 {
   231  			in = -1
   232  		}
   233  		gcpercent = in
   234  		heapminimum = defaultHeapMinimum * uint64(gcpercent) / 100
   235  		// Update pacing in response to gcpercent change.
   236  		gcSetTriggerRatio(memstats.triggerRatio)
   237  		unlock(&mheap_.lock)
   238  	})
   239  
   240  	// If we just disabled GC, wait for any concurrent GC mark to
   241  	// finish so we always return with no GC running.
   242  	if in < 0 {
   243  		gcWaitOnMark(atomic.Load(&work.cycles))
   244  	}
   245  
   246  	return out
   247  }
   248  
   249  // Garbage collector phase.
   250  // Indicates to write barrier and synchronization task to perform.
   251  var gcphase uint32
   252  
   253  // The compiler knows about this variable.
   254  // If you change it, you must change builtin/runtime.go, too.
   255  // If you change the first four bytes, you must also change the write
   256  // barrier insertion code.
   257  var writeBarrier struct {
   258  	enabled bool    // compiler emits a check of this before calling write barrier
   259  	pad     [3]byte // compiler uses 32-bit load for "enabled" field
   260  	needed  bool    // whether we need a write barrier for current GC phase
   261  	cgo     bool    // whether we need a write barrier for a cgo check
   262  	alignme uint64  // guarantee alignment so that compiler can use a 32 or 64-bit load
   263  }
   264  
   265  // gcBlackenEnabled is 1 if mutator assists and background mark
   266  // workers are allowed to blacken objects. This must only be set when
   267  // gcphase == _GCmark.
   268  var gcBlackenEnabled uint32
   269  
   270  const (
   271  	_GCoff             = iota // GC not running; sweeping in background, write barrier disabled
   272  	_GCmark                   // GC marking roots and workbufs: allocate black, write barrier ENABLED
   273  	_GCmarktermination        // GC mark termination: allocate black, P's help GC, write barrier ENABLED
   274  )
   275  
   276  //go:nosplit
   277  func setGCPhase(x uint32) {
   278  	atomic.Store(&gcphase, x)
   279  	writeBarrier.needed = gcphase == _GCmark || gcphase == _GCmarktermination
   280  	writeBarrier.enabled = writeBarrier.needed || writeBarrier.cgo
   281  }
   282  
   283  // gcMarkWorkerMode represents the mode that a concurrent mark worker
   284  // should operate in.
   285  //
   286  // Concurrent marking happens through four different mechanisms. One
   287  // is mutator assists, which happen in response to allocations and are
   288  // not scheduled. The other three are variations in the per-P mark
   289  // workers and are distinguished by gcMarkWorkerMode.
   290  type gcMarkWorkerMode int
   291  
   292  const (
   293  	// gcMarkWorkerNotWorker indicates that the next scheduled G is not
   294  	// starting work and the mode should be ignored.
   295  	gcMarkWorkerNotWorker gcMarkWorkerMode = iota
   296  
   297  	// gcMarkWorkerDedicatedMode indicates that the P of a mark
   298  	// worker is dedicated to running that mark worker. The mark
   299  	// worker should run without preemption.
   300  	gcMarkWorkerDedicatedMode
   301  
   302  	// gcMarkWorkerFractionalMode indicates that a P is currently
   303  	// running the "fractional" mark worker. The fractional worker
   304  	// is necessary when GOMAXPROCS*gcBackgroundUtilization is not
   305  	// an integer. The fractional worker should run until it is
   306  	// preempted and will be scheduled to pick up the fractional
   307  	// part of GOMAXPROCS*gcBackgroundUtilization.
   308  	gcMarkWorkerFractionalMode
   309  
   310  	// gcMarkWorkerIdleMode indicates that a P is running the mark
   311  	// worker because it has nothing else to do. The idle worker
   312  	// should run until it is preempted and account its time
   313  	// against gcController.idleMarkTime.
   314  	gcMarkWorkerIdleMode
   315  )
   316  
   317  // gcMarkWorkerModeStrings are the strings labels of gcMarkWorkerModes
   318  // to use in execution traces.
   319  var gcMarkWorkerModeStrings = [...]string{
   320  	"Not worker",
   321  	"GC (dedicated)",
   322  	"GC (fractional)",
   323  	"GC (idle)",
   324  }
   325  
   326  // gcController implements the GC pacing controller that determines
   327  // when to trigger concurrent garbage collection and how much marking
   328  // work to do in mutator assists and background marking.
   329  //
   330  // It uses a feedback control algorithm to adjust the memstats.gc_trigger
   331  // trigger based on the heap growth and GC CPU utilization each cycle.
   332  // This algorithm optimizes for heap growth to match GOGC and for CPU
   333  // utilization between assist and background marking to be 25% of
   334  // GOMAXPROCS. The high-level design of this algorithm is documented
   335  // at https://golang.org/s/go15gcpacing.
   336  //
   337  // All fields of gcController are used only during a single mark
   338  // cycle.
   339  var gcController gcControllerState
   340  
   341  type gcControllerState struct {
   342  	// scanWork is the total scan work performed this cycle. This
   343  	// is updated atomically during the cycle. Updates occur in
   344  	// bounded batches, since it is both written and read
   345  	// throughout the cycle. At the end of the cycle, this is how
   346  	// much of the retained heap is scannable.
   347  	//
   348  	// Currently this is the bytes of heap scanned. For most uses,
   349  	// this is an opaque unit of work, but for estimation the
   350  	// definition is important.
   351  	scanWork int64
   352  
   353  	// bgScanCredit is the scan work credit accumulated by the
   354  	// concurrent background scan. This credit is accumulated by
   355  	// the background scan and stolen by mutator assists. This is
   356  	// updated atomically. Updates occur in bounded batches, since
   357  	// it is both written and read throughout the cycle.
   358  	bgScanCredit int64
   359  
   360  	// assistTime is the nanoseconds spent in mutator assists
   361  	// during this cycle. This is updated atomically. Updates
   362  	// occur in bounded batches, since it is both written and read
   363  	// throughout the cycle.
   364  	assistTime int64
   365  
   366  	// dedicatedMarkTime is the nanoseconds spent in dedicated
   367  	// mark workers during this cycle. This is updated atomically
   368  	// at the end of the concurrent mark phase.
   369  	dedicatedMarkTime int64
   370  
   371  	// fractionalMarkTime is the nanoseconds spent in the
   372  	// fractional mark worker during this cycle. This is updated
   373  	// atomically throughout the cycle and will be up-to-date if
   374  	// the fractional mark worker is not currently running.
   375  	fractionalMarkTime int64
   376  
   377  	// idleMarkTime is the nanoseconds spent in idle marking
   378  	// during this cycle. This is updated atomically throughout
   379  	// the cycle.
   380  	idleMarkTime int64
   381  
   382  	// markStartTime is the absolute start time in nanoseconds
   383  	// that assists and background mark workers started.
   384  	markStartTime int64
   385  
   386  	// dedicatedMarkWorkersNeeded is the number of dedicated mark
   387  	// workers that need to be started. This is computed at the
   388  	// beginning of each cycle and decremented atomically as
   389  	// dedicated mark workers get started.
   390  	dedicatedMarkWorkersNeeded int64
   391  
   392  	// assistWorkPerByte is the ratio of scan work to allocated
   393  	// bytes that should be performed by mutator assists. This is
   394  	// computed at the beginning of each cycle and updated every
   395  	// time heap_scan is updated.
   396  	//
   397  	// Stored as a uint64, but it's actually a float64. Use
   398  	// float64frombits to get the value.
   399  	//
   400  	// Read and written atomically.
   401  	assistWorkPerByte uint64
   402  
   403  	// assistBytesPerWork is 1/assistWorkPerByte.
   404  	//
   405  	// Stored as a uint64, but it's actually a float64. Use
   406  	// float64frombits to get the value.
   407  	//
   408  	// Read and written atomically.
   409  	//
   410  	// Note that because this is read and written independently
   411  	// from assistWorkPerByte users may notice a skew between
   412  	// the two values, and such a state should be safe.
   413  	assistBytesPerWork uint64
   414  
   415  	// fractionalUtilizationGoal is the fraction of wall clock
   416  	// time that should be spent in the fractional mark worker on
   417  	// each P that isn't running a dedicated worker.
   418  	//
   419  	// For example, if the utilization goal is 25% and there are
   420  	// no dedicated workers, this will be 0.25. If the goal is
   421  	// 25%, there is one dedicated worker, and GOMAXPROCS is 5,
   422  	// this will be 0.05 to make up the missing 5%.
   423  	//
   424  	// If this is zero, no fractional workers are needed.
   425  	fractionalUtilizationGoal float64
   426  
   427  	_ cpu.CacheLinePad
   428  }
   429  
   430  // startCycle resets the GC controller's state and computes estimates
   431  // for a new GC cycle. The caller must hold worldsema and the world
   432  // must be stopped.
   433  func (c *gcControllerState) startCycle() {
   434  	c.scanWork = 0
   435  	c.bgScanCredit = 0
   436  	c.assistTime = 0
   437  	c.dedicatedMarkTime = 0
   438  	c.fractionalMarkTime = 0
   439  	c.idleMarkTime = 0
   440  
   441  	// Ensure that the heap goal is at least a little larger than
   442  	// the current live heap size. This may not be the case if GC
   443  	// start is delayed or if the allocation that pushed heap_live
   444  	// over gc_trigger is large or if the trigger is really close to
   445  	// GOGC. Assist is proportional to this distance, so enforce a
   446  	// minimum distance, even if it means going over the GOGC goal
   447  	// by a tiny bit.
   448  	if memstats.next_gc < memstats.heap_live+1024*1024 {
   449  		memstats.next_gc = memstats.heap_live + 1024*1024
   450  	}
   451  
   452  	// Compute the background mark utilization goal. In general,
   453  	// this may not come out exactly. We round the number of
   454  	// dedicated workers so that the utilization is closest to
   455  	// 25%. For small GOMAXPROCS, this would introduce too much
   456  	// error, so we add fractional workers in that case.
   457  	totalUtilizationGoal := float64(gomaxprocs) * gcBackgroundUtilization
   458  	c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5)
   459  	utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
   460  	const maxUtilError = 0.3
   461  	if utilError < -maxUtilError || utilError > maxUtilError {
   462  		// Rounding put us more than 30% off our goal. With
   463  		// gcBackgroundUtilization of 25%, this happens for
   464  		// GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
   465  		// workers to compensate.
   466  		if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
   467  			// Too many dedicated workers.
   468  			c.dedicatedMarkWorkersNeeded--
   469  		}
   470  		c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(gomaxprocs)
   471  	} else {
   472  		c.fractionalUtilizationGoal = 0
   473  	}
   474  
   475  	// In STW mode, we just want dedicated workers.
   476  	if debug.gcstoptheworld > 0 {
   477  		c.dedicatedMarkWorkersNeeded = int64(gomaxprocs)
   478  		c.fractionalUtilizationGoal = 0
   479  	}
   480  
   481  	// Clear per-P state
   482  	for _, p := range allp {
   483  		p.gcAssistTime = 0
   484  		p.gcFractionalMarkTime = 0
   485  	}
   486  
   487  	// Compute initial values for controls that are updated
   488  	// throughout the cycle.
   489  	c.revise()
   490  
   491  	if debug.gcpacertrace > 0 {
   492  		assistRatio := float64frombits(atomic.Load64(&c.assistWorkPerByte))
   493  		print("pacer: assist ratio=", assistRatio,
   494  			" (scan ", memstats.heap_scan>>20, " MB in ",
   495  			work.initialHeapLive>>20, "->",
   496  			memstats.next_gc>>20, " MB)",
   497  			" workers=", c.dedicatedMarkWorkersNeeded,
   498  			"+", c.fractionalUtilizationGoal, "\n")
   499  	}
   500  }
   501  
   502  // revise updates the assist ratio during the GC cycle to account for
   503  // improved estimates. This should be called whenever memstats.heap_scan,
   504  // memstats.heap_live, or memstats.next_gc is updated. It is safe to
   505  // call concurrently, but it may race with other calls to revise.
   506  //
   507  // The result of this race is that the two assist ratio values may not line
   508  // up or may be stale. In practice this is OK because the assist ratio
   509  // moves slowly throughout a GC cycle, and the assist ratio is a best-effort
   510  // heuristic anyway. Furthermore, no part of the heuristic depends on
   511  // the two assist ratio values being exact reciprocals of one another, since
   512  // the two values are used to convert values from different sources.
   513  //
   514  // The worst case result of this raciness is that we may miss a larger shift
   515  // in the ratio (say, if we decide to pace more aggressively against the
   516  // hard heap goal) but even this "hard goal" is best-effort (see #40460).
   517  // The dedicated GC should ensure we don't exceed the hard goal by too much
   518  // in the rare case we do exceed it.
   519  //
   520  // It should only be called when gcBlackenEnabled != 0 (because this
   521  // is when assists are enabled and the necessary statistics are
   522  // available).
   523  func (c *gcControllerState) revise() {
   524  	gcpercent := gcpercent
   525  	if gcpercent < 0 {
   526  		// If GC is disabled but we're running a forced GC,
   527  		// act like GOGC is huge for the below calculations.
   528  		gcpercent = 100000
   529  	}
   530  	live := atomic.Load64(&memstats.heap_live)
   531  	scan := atomic.Load64(&memstats.heap_scan)
   532  	work := atomic.Loadint64(&c.scanWork)
   533  
   534  	// Assume we're under the soft goal. Pace GC to complete at
   535  	// next_gc assuming the heap is in steady-state.
   536  	heapGoal := int64(atomic.Load64(&memstats.next_gc))
   537  
   538  	// Compute the expected scan work remaining.
   539  	//
   540  	// This is estimated based on the expected
   541  	// steady-state scannable heap. For example, with
   542  	// GOGC=100, only half of the scannable heap is
   543  	// expected to be live, so that's what we target.
   544  	//
   545  	// (This is a float calculation to avoid overflowing on
   546  	// 100*heap_scan.)
   547  	scanWorkExpected := int64(float64(scan) * 100 / float64(100+gcpercent))
   548  
   549  	if int64(live) > heapGoal || work > scanWorkExpected {
   550  		// We're past the soft goal, or we've already done more scan
   551  		// work than we expected. Pace GC so that in the worst case it
   552  		// will complete by the hard goal.
   553  		const maxOvershoot = 1.1
   554  		heapGoal = int64(float64(heapGoal) * maxOvershoot)
   555  
   556  		// Compute the upper bound on the scan work remaining.
   557  		scanWorkExpected = int64(scan)
   558  	}
   559  
   560  	// Compute the remaining scan work estimate.
   561  	//
   562  	// Note that we currently count allocations during GC as both
   563  	// scannable heap (heap_scan) and scan work completed
   564  	// (scanWork), so allocation will change this difference
   565  	// slowly in the soft regime and not at all in the hard
   566  	// regime.
   567  	scanWorkRemaining := scanWorkExpected - work
   568  	if scanWorkRemaining < 1000 {
   569  		// We set a somewhat arbitrary lower bound on
   570  		// remaining scan work since if we aim a little high,
   571  		// we can miss by a little.
   572  		//
   573  		// We *do* need to enforce that this is at least 1,
   574  		// since marking is racy and double-scanning objects
   575  		// may legitimately make the remaining scan work
   576  		// negative, even in the hard goal regime.
   577  		scanWorkRemaining = 1000
   578  	}
   579  
   580  	// Compute the heap distance remaining.
   581  	heapRemaining := heapGoal - int64(live)
   582  	if heapRemaining <= 0 {
   583  		// This shouldn't happen, but if it does, avoid
   584  		// dividing by zero or setting the assist negative.
   585  		heapRemaining = 1
   586  	}
   587  
   588  	// Compute the mutator assist ratio so by the time the mutator
   589  	// allocates the remaining heap bytes up to next_gc, it will
   590  	// have done (or stolen) the remaining amount of scan work.
   591  	// Note that the assist ratio values are updated atomically
   592  	// but not together. This means there may be some degree of
   593  	// skew between the two values. This is generally OK as the
   594  	// values shift relatively slowly over the course of a GC
   595  	// cycle.
   596  	assistWorkPerByte := float64(scanWorkRemaining) / float64(heapRemaining)
   597  	assistBytesPerWork := float64(heapRemaining) / float64(scanWorkRemaining)
   598  	atomic.Store64(&c.assistWorkPerByte, float64bits(assistWorkPerByte))
   599  	atomic.Store64(&c.assistBytesPerWork, float64bits(assistBytesPerWork))
   600  }
   601  
   602  // endCycle computes the trigger ratio for the next cycle.
   603  func (c *gcControllerState) endCycle() float64 {
   604  	if work.userForced {
   605  		// Forced GC means this cycle didn't start at the
   606  		// trigger, so where it finished isn't good
   607  		// information about how to adjust the trigger.
   608  		// Just leave it where it is.
   609  		return memstats.triggerRatio
   610  	}
   611  
   612  	// Proportional response gain for the trigger controller. Must
   613  	// be in [0, 1]. Lower values smooth out transient effects but
   614  	// take longer to respond to phase changes. Higher values
   615  	// react to phase changes quickly, but are more affected by
   616  	// transient changes. Values near 1 may be unstable.
   617  	const triggerGain = 0.5
   618  
   619  	// Compute next cycle trigger ratio. First, this computes the
   620  	// "error" for this cycle; that is, how far off the trigger
   621  	// was from what it should have been, accounting for both heap
   622  	// growth and GC CPU utilization. We compute the actual heap
   623  	// growth during this cycle and scale that by how far off from
   624  	// the goal CPU utilization we were (to estimate the heap
   625  	// growth if we had the desired CPU utilization). The
   626  	// difference between this estimate and the GOGC-based goal
   627  	// heap growth is the error.
   628  	goalGrowthRatio := gcEffectiveGrowthRatio()
   629  	actualGrowthRatio := float64(memstats.heap_live)/float64(memstats.heap_marked) - 1
   630  	assistDuration := nanotime() - c.markStartTime
   631  
   632  	// Assume background mark hit its utilization goal.
   633  	utilization := gcBackgroundUtilization
   634  	// Add assist utilization; avoid divide by zero.
   635  	if assistDuration > 0 {
   636  		utilization += float64(c.assistTime) / float64(assistDuration*int64(gomaxprocs))
   637  	}
   638  
   639  	triggerError := goalGrowthRatio - memstats.triggerRatio - utilization/gcGoalUtilization*(actualGrowthRatio-memstats.triggerRatio)
   640  
   641  	// Finally, we adjust the trigger for next time by this error,
   642  	// damped by the proportional gain.
   643  	triggerRatio := memstats.triggerRatio + triggerGain*triggerError
   644  
   645  	if debug.gcpacertrace > 0 {
   646  		// Print controller state in terms of the design
   647  		// document.
   648  		H_m_prev := memstats.heap_marked
   649  		h_t := memstats.triggerRatio
   650  		H_T := memstats.gc_trigger
   651  		h_a := actualGrowthRatio
   652  		H_a := memstats.heap_live
   653  		h_g := goalGrowthRatio
   654  		H_g := int64(float64(H_m_prev) * (1 + h_g))
   655  		u_a := utilization
   656  		u_g := gcGoalUtilization
   657  		W_a := c.scanWork
   658  		print("pacer: H_m_prev=", H_m_prev,
   659  			" h_t=", h_t, " H_T=", H_T,
   660  			" h_a=", h_a, " H_a=", H_a,
   661  			" h_g=", h_g, " H_g=", H_g,
   662  			" u_a=", u_a, " u_g=", u_g,
   663  			" W_a=", W_a,
   664  			" goalΔ=", goalGrowthRatio-h_t,
   665  			" actualΔ=", h_a-h_t,
   666  			" u_a/u_g=", u_a/u_g,
   667  			"\n")
   668  	}
   669  
   670  	return triggerRatio
   671  }
   672  
   673  // enlistWorker encourages another dedicated mark worker to start on
   674  // another P if there are spare worker slots. It is used by putfull
   675  // when more work is made available.
   676  //
   677  //go:nowritebarrier
   678  func (c *gcControllerState) enlistWorker() {
   679  	// If there are idle Ps, wake one so it will run an idle worker.
   680  	// NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
   681  	//
   682  	//	if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
   683  	//		wakep()
   684  	//		return
   685  	//	}
   686  
   687  	// There are no idle Ps. If we need more dedicated workers,
   688  	// try to preempt a running P so it will switch to a worker.
   689  	if c.dedicatedMarkWorkersNeeded <= 0 {
   690  		return
   691  	}
   692  	// Pick a random other P to preempt.
   693  	if gomaxprocs <= 1 {
   694  		return
   695  	}
   696  	gp := getg()
   697  	if gp == nil || gp.m == nil || gp.m.p == 0 {
   698  		return
   699  	}
   700  	myID := gp.m.p.ptr().id
   701  	for tries := 0; tries < 5; tries++ {
   702  		id := int32(fastrandn(uint32(gomaxprocs - 1)))
   703  		if id >= myID {
   704  			id++
   705  		}
   706  		p := allp[id]
   707  		if p.status != _Prunning {
   708  			continue
   709  		}
   710  		if preemptone(p) {
   711  			return
   712  		}
   713  	}
   714  }
   715  
   716  // findRunnableGCWorker returns a background mark worker for _p_ if it
   717  // should be run. This must only be called when gcBlackenEnabled != 0.
   718  func (c *gcControllerState) findRunnableGCWorker(_p_ *p) *g {
   719  	if gcBlackenEnabled == 0 {
   720  		throw("gcControllerState.findRunnable: blackening not enabled")
   721  	}
   722  
   723  	if !gcMarkWorkAvailable(_p_) {
   724  		// No work to be done right now. This can happen at
   725  		// the end of the mark phase when there are still
   726  		// assists tapering off. Don't bother running a worker
   727  		// now because it'll just return immediately.
   728  		return nil
   729  	}
   730  
   731  	// Grab a worker before we commit to running below.
   732  	node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
   733  	if node == nil {
   734  		// There is at least one worker per P, so normally there are
   735  		// enough workers to run on all Ps, if necessary. However, once
   736  		// a worker enters gcMarkDone it may park without rejoining the
   737  		// pool, thus freeing a P with no corresponding worker.
   738  		// gcMarkDone never depends on another worker doing work, so it
   739  		// is safe to simply do nothing here.
   740  		//
   741  		// If gcMarkDone bails out without completing the mark phase,
   742  		// it will always do so with queued global work. Thus, that P
   743  		// will be immediately eligible to re-run the worker G it was
   744  		// just using, ensuring work can complete.
   745  		return nil
   746  	}
   747  
   748  	decIfPositive := func(ptr *int64) bool {
   749  		for {
   750  			v := atomic.Loadint64(ptr)
   751  			if v <= 0 {
   752  				return false
   753  			}
   754  
   755  			// TODO: having atomic.Casint64 would be more pleasant.
   756  			if atomic.Cas64((*uint64)(unsafe.Pointer(ptr)), uint64(v), uint64(v-1)) {
   757  				return true
   758  			}
   759  		}
   760  	}
   761  
   762  	if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
   763  		// This P is now dedicated to marking until the end of
   764  		// the concurrent mark phase.
   765  		_p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
   766  	} else if c.fractionalUtilizationGoal == 0 {
   767  		// No need for fractional workers.
   768  		gcBgMarkWorkerPool.push(&node.node)
   769  		return nil
   770  	} else {
   771  		// Is this P behind on the fractional utilization
   772  		// goal?
   773  		//
   774  		// This should be kept in sync with pollFractionalWorkerExit.
   775  		delta := nanotime() - gcController.markStartTime
   776  		if delta > 0 && float64(_p_.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
   777  			// Nope. No need to run a fractional worker.
   778  			gcBgMarkWorkerPool.push(&node.node)
   779  			return nil
   780  		}
   781  		// Run a fractional worker.
   782  		_p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode
   783  	}
   784  
   785  	// Run the background mark worker.
   786  	gp := node.gp.ptr()
   787  	casgstatus(gp, _Gwaiting, _Grunnable)
   788  	if trace.enabled {
   789  		traceGoUnpark(gp, 0)
   790  	}
   791  	return gp
   792  }
   793  
   794  // pollFractionalWorkerExit reports whether a fractional mark worker
   795  // should self-preempt. It assumes it is called from the fractional
   796  // worker.
   797  func pollFractionalWorkerExit() bool {
   798  	// This should be kept in sync with the fractional worker
   799  	// scheduler logic in findRunnableGCWorker.
   800  	now := nanotime()
   801  	delta := now - gcController.markStartTime
   802  	if delta <= 0 {
   803  		return true
   804  	}
   805  	p := getg().m.p.ptr()
   806  	selfTime := p.gcFractionalMarkTime + (now - p.gcMarkWorkerStartTime)
   807  	// Add some slack to the utilization goal so that the
   808  	// fractional worker isn't behind again the instant it exits.
   809  	return float64(selfTime)/float64(delta) > 1.2*gcController.fractionalUtilizationGoal
   810  }
   811  
   812  // gcSetTriggerRatio sets the trigger ratio and updates everything
   813  // derived from it: the absolute trigger, the heap goal, mark pacing,
   814  // and sweep pacing.
   815  //
   816  // This can be called any time. If GC is the in the middle of a
   817  // concurrent phase, it will adjust the pacing of that phase.
   818  //
   819  // This depends on gcpercent, memstats.heap_marked, and
   820  // memstats.heap_live. These must be up to date.
   821  //
   822  // mheap_.lock must be held or the world must be stopped.
   823  func gcSetTriggerRatio(triggerRatio float64) {
   824  	assertWorldStoppedOrLockHeld(&mheap_.lock)
   825  
   826  	// Compute the next GC goal, which is when the allocated heap
   827  	// has grown by GOGC/100 over the heap marked by the last
   828  	// cycle.
   829  	goal := ^uint64(0)
   830  	if gcpercent >= 0 {
   831  		goal = memstats.heap_marked + memstats.heap_marked*uint64(gcpercent)/100
   832  	}
   833  
   834  	// Set the trigger ratio, capped to reasonable bounds.
   835  	if gcpercent >= 0 {
   836  		scalingFactor := float64(gcpercent) / 100
   837  		// Ensure there's always a little margin so that the
   838  		// mutator assist ratio isn't infinity.
   839  		maxTriggerRatio := 0.95 * scalingFactor
   840  		if triggerRatio > maxTriggerRatio {
   841  			triggerRatio = maxTriggerRatio
   842  		}
   843  
   844  		// If we let triggerRatio go too low, then if the application
   845  		// is allocating very rapidly we might end up in a situation
   846  		// where we're allocating black during a nearly always-on GC.
   847  		// The result of this is a growing heap and ultimately an
   848  		// increase in RSS. By capping us at a point >0, we're essentially
   849  		// saying that we're OK using more CPU during the GC to prevent
   850  		// this growth in RSS.
   851  		//
   852  		// The current constant was chosen empirically: given a sufficiently
   853  		// fast/scalable allocator with 48 Ps that could drive the trigger ratio
   854  		// to <0.05, this constant causes applications to retain the same peak
   855  		// RSS compared to not having this allocator.
   856  		minTriggerRatio := 0.6 * scalingFactor
   857  		if triggerRatio < minTriggerRatio {
   858  			triggerRatio = minTriggerRatio
   859  		}
   860  	} else if triggerRatio < 0 {
   861  		// gcpercent < 0, so just make sure we're not getting a negative
   862  		// triggerRatio. This case isn't expected to happen in practice,
   863  		// and doesn't really matter because if gcpercent < 0 then we won't
   864  		// ever consume triggerRatio further on in this function, but let's
   865  		// just be defensive here; the triggerRatio being negative is almost
   866  		// certainly undesirable.
   867  		triggerRatio = 0
   868  	}
   869  	memstats.triggerRatio = triggerRatio
   870  
   871  	// Compute the absolute GC trigger from the trigger ratio.
   872  	//
   873  	// We trigger the next GC cycle when the allocated heap has
   874  	// grown by the trigger ratio over the marked heap size.
   875  	trigger := ^uint64(0)
   876  	if gcpercent >= 0 {
   877  		trigger = uint64(float64(memstats.heap_marked) * (1 + triggerRatio))
   878  		// Don't trigger below the minimum heap size.
   879  		minTrigger := heapminimum
   880  		if !isSweepDone() {
   881  			// Concurrent sweep happens in the heap growth
   882  			// from heap_live to gc_trigger, so ensure
   883  			// that concurrent sweep has some heap growth
   884  			// in which to perform sweeping before we
   885  			// start the next GC cycle.
   886  			sweepMin := atomic.Load64(&memstats.heap_live) + sweepMinHeapDistance
   887  			if sweepMin > minTrigger {
   888  				minTrigger = sweepMin
   889  			}
   890  		}
   891  		if trigger < minTrigger {
   892  			trigger = minTrigger
   893  		}
   894  		if int64(trigger) < 0 {
   895  			print("runtime: next_gc=", memstats.next_gc, " heap_marked=", memstats.heap_marked, " heap_live=", memstats.heap_live, " initialHeapLive=", work.initialHeapLive, "triggerRatio=", triggerRatio, " minTrigger=", minTrigger, "\n")
   896  			throw("gc_trigger underflow")
   897  		}
   898  		if trigger > goal {
   899  			// The trigger ratio is always less than GOGC/100, but
   900  			// other bounds on the trigger may have raised it.
   901  			// Push up the goal, too.
   902  			goal = trigger
   903  		}
   904  	}
   905  
   906  	// Commit to the trigger and goal.
   907  	memstats.gc_trigger = trigger
   908  	atomic.Store64(&memstats.next_gc, goal)
   909  	if trace.enabled {
   910  		traceNextGC()
   911  	}
   912  
   913  	// Update mark pacing.
   914  	if gcphase != _GCoff {
   915  		gcController.revise()
   916  	}
   917  
   918  	// Update sweep pacing.
   919  	if isSweepDone() {
   920  		mheap_.sweepPagesPerByte = 0
   921  	} else {
   922  		// Concurrent sweep needs to sweep all of the in-use
   923  		// pages by the time the allocated heap reaches the GC
   924  		// trigger. Compute the ratio of in-use pages to sweep
   925  		// per byte allocated, accounting for the fact that
   926  		// some might already be swept.
   927  		heapLiveBasis := atomic.Load64(&memstats.heap_live)
   928  		heapDistance := int64(trigger) - int64(heapLiveBasis)
   929  		// Add a little margin so rounding errors and
   930  		// concurrent sweep are less likely to leave pages
   931  		// unswept when GC starts.
   932  		heapDistance -= 1024 * 1024
   933  		if heapDistance < _PageSize {
   934  			// Avoid setting the sweep ratio extremely high
   935  			heapDistance = _PageSize
   936  		}
   937  		pagesSwept := atomic.Load64(&mheap_.pagesSwept)
   938  		pagesInUse := atomic.Load64(&mheap_.pagesInUse)
   939  		sweepDistancePages := int64(pagesInUse) - int64(pagesSwept)
   940  		if sweepDistancePages <= 0 {
   941  			mheap_.sweepPagesPerByte = 0
   942  		} else {
   943  			mheap_.sweepPagesPerByte = float64(sweepDistancePages) / float64(heapDistance)
   944  			mheap_.sweepHeapLiveBasis = heapLiveBasis
   945  			// Write pagesSweptBasis last, since this
   946  			// signals concurrent sweeps to recompute
   947  			// their debt.
   948  			atomic.Store64(&mheap_.pagesSweptBasis, pagesSwept)
   949  		}
   950  	}
   951  
   952  	gcPaceScavenger()
   953  }
   954  
   955  // gcEffectiveGrowthRatio returns the current effective heap growth
   956  // ratio (GOGC/100) based on heap_marked from the previous GC and
   957  // next_gc for the current GC.
   958  //
   959  // This may differ from gcpercent/100 because of various upper and
   960  // lower bounds on gcpercent. For example, if the heap is smaller than
   961  // heapminimum, this can be higher than gcpercent/100.
   962  //
   963  // mheap_.lock must be held or the world must be stopped.
   964  func gcEffectiveGrowthRatio() float64 {
   965  	assertWorldStoppedOrLockHeld(&mheap_.lock)
   966  
   967  	egogc := float64(atomic.Load64(&memstats.next_gc)-memstats.heap_marked) / float64(memstats.heap_marked)
   968  	if egogc < 0 {
   969  		// Shouldn't happen, but just in case.
   970  		egogc = 0
   971  	}
   972  	return egogc
   973  }
   974  
   975  // gcGoalUtilization is the goal CPU utilization for
   976  // marking as a fraction of GOMAXPROCS.
   977  const gcGoalUtilization = 0.30
   978  
   979  // gcBackgroundUtilization is the fixed CPU utilization for background
   980  // marking. It must be <= gcGoalUtilization. The difference between
   981  // gcGoalUtilization and gcBackgroundUtilization will be made up by
   982  // mark assists. The scheduler will aim to use within 50% of this
   983  // goal.
   984  //
   985  // Setting this to < gcGoalUtilization avoids saturating the trigger
   986  // feedback controller when there are no assists, which allows it to
   987  // better control CPU and heap growth. However, the larger the gap,
   988  // the more mutator assists are expected to happen, which impact
   989  // mutator latency.
   990  const gcBackgroundUtilization = 0.25
   991  
   992  // gcCreditSlack is the amount of scan work credit that can
   993  // accumulate locally before updating gcController.scanWork and,
   994  // optionally, gcController.bgScanCredit. Lower values give a more
   995  // accurate assist ratio and make it more likely that assists will
   996  // successfully steal background credit. Higher values reduce memory
   997  // contention.
   998  const gcCreditSlack = 2000
   999  
  1000  // gcAssistTimeSlack is the nanoseconds of mutator assist time that
  1001  // can accumulate on a P before updating gcController.assistTime.
  1002  const gcAssistTimeSlack = 5000
  1003  
  1004  // gcOverAssistWork determines how many extra units of scan work a GC
  1005  // assist does when an assist happens. This amortizes the cost of an
  1006  // assist by pre-paying for this many bytes of future allocations.
  1007  const gcOverAssistWork = 64 << 10
  1008  
  1009  var work struct {
  1010  	full  lfstack          // lock-free list of full blocks workbuf
  1011  	empty lfstack          // lock-free list of empty blocks workbuf
  1012  	pad0  cpu.CacheLinePad // prevents false-sharing between full/empty and nproc/nwait
  1013  
  1014  	wbufSpans struct {
  1015  		lock mutex
  1016  		// free is a list of spans dedicated to workbufs, but
  1017  		// that don't currently contain any workbufs.
  1018  		free mSpanList
  1019  		// busy is a list of all spans containing workbufs on
  1020  		// one of the workbuf lists.
  1021  		busy mSpanList
  1022  	}
  1023  
  1024  	// Restore 64-bit alignment on 32-bit.
  1025  	_ uint32
  1026  
  1027  	// bytesMarked is the number of bytes marked this cycle. This
  1028  	// includes bytes blackened in scanned objects, noscan objects
  1029  	// that go straight to black, and permagrey objects scanned by
  1030  	// markroot during the concurrent scan phase. This is updated
  1031  	// atomically during the cycle. Updates may be batched
  1032  	// arbitrarily, since the value is only read at the end of the
  1033  	// cycle.
  1034  	//
  1035  	// Because of benign races during marking, this number may not
  1036  	// be the exact number of marked bytes, but it should be very
  1037  	// close.
  1038  	//
  1039  	// Put this field here because it needs 64-bit atomic access
  1040  	// (and thus 8-byte alignment even on 32-bit architectures).
  1041  	bytesMarked uint64
  1042  
  1043  	markrootNext uint32 // next markroot job
  1044  	markrootJobs uint32 // number of markroot jobs
  1045  
  1046  	nproc  uint32
  1047  	tstart int64
  1048  	nwait  uint32
  1049  
  1050  	// Number of roots of various root types. Set by gcMarkRootPrepare.
  1051  	nFlushCacheRoots                               int
  1052  	nDataRoots, nBSSRoots, nSpanRoots, nStackRoots int
  1053  
  1054  	// Each type of GC state transition is protected by a lock.
  1055  	// Since multiple threads can simultaneously detect the state
  1056  	// transition condition, any thread that detects a transition
  1057  	// condition must acquire the appropriate transition lock,
  1058  	// re-check the transition condition and return if it no
  1059  	// longer holds or perform the transition if it does.
  1060  	// Likewise, any transition must invalidate the transition
  1061  	// condition before releasing the lock. This ensures that each
  1062  	// transition is performed by exactly one thread and threads
  1063  	// that need the transition to happen block until it has
  1064  	// happened.
  1065  	//
  1066  	// startSema protects the transition from "off" to mark or
  1067  	// mark termination.
  1068  	startSema uint32
  1069  	// markDoneSema protects transitions from mark to mark termination.
  1070  	markDoneSema uint32
  1071  
  1072  	bgMarkReady note   // signal background mark worker has started
  1073  	bgMarkDone  uint32 // cas to 1 when at a background mark completion point
  1074  	// Background mark completion signaling
  1075  
  1076  	// mode is the concurrency mode of the current GC cycle.
  1077  	mode gcMode
  1078  
  1079  	// userForced indicates the current GC cycle was forced by an
  1080  	// explicit user call.
  1081  	userForced bool
  1082  
  1083  	// totaltime is the CPU nanoseconds spent in GC since the
  1084  	// program started if debug.gctrace > 0.
  1085  	totaltime int64
  1086  
  1087  	// initialHeapLive is the value of memstats.heap_live at the
  1088  	// beginning of this GC cycle.
  1089  	initialHeapLive uint64
  1090  
  1091  	// assistQueue is a queue of assists that are blocked because
  1092  	// there was neither enough credit to steal or enough work to
  1093  	// do.
  1094  	assistQueue struct {
  1095  		lock mutex
  1096  		q    gQueue
  1097  	}
  1098  
  1099  	// sweepWaiters is a list of blocked goroutines to wake when
  1100  	// we transition from mark termination to sweep.
  1101  	sweepWaiters struct {
  1102  		lock mutex
  1103  		list gList
  1104  	}
  1105  
  1106  	// cycles is the number of completed GC cycles, where a GC
  1107  	// cycle is sweep termination, mark, mark termination, and
  1108  	// sweep. This differs from memstats.numgc, which is
  1109  	// incremented at mark termination.
  1110  	cycles uint32
  1111  
  1112  	// Timing/utilization stats for this cycle.
  1113  	stwprocs, maxprocs                 int32
  1114  	tSweepTerm, tMark, tMarkTerm, tEnd int64 // nanotime() of phase start
  1115  
  1116  	pauseNS    int64 // total STW time this cycle
  1117  	pauseStart int64 // nanotime() of last STW
  1118  
  1119  	// debug.gctrace heap sizes for this cycle.
  1120  	heap0, heap1, heap2, heapGoal uint64
  1121  }
  1122  
  1123  // GC runs a garbage collection and blocks the caller until the
  1124  // garbage collection is complete. It may also block the entire
  1125  // program.
  1126  func GC() {
  1127  	// We consider a cycle to be: sweep termination, mark, mark
  1128  	// termination, and sweep. This function shouldn't return
  1129  	// until a full cycle has been completed, from beginning to
  1130  	// end. Hence, we always want to finish up the current cycle
  1131  	// and start a new one. That means:
  1132  	//
  1133  	// 1. In sweep termination, mark, or mark termination of cycle
  1134  	// N, wait until mark termination N completes and transitions
  1135  	// to sweep N.
  1136  	//
  1137  	// 2. In sweep N, help with sweep N.
  1138  	//
  1139  	// At this point we can begin a full cycle N+1.
  1140  	//
  1141  	// 3. Trigger cycle N+1 by starting sweep termination N+1.
  1142  	//
  1143  	// 4. Wait for mark termination N+1 to complete.
  1144  	//
  1145  	// 5. Help with sweep N+1 until it's done.
  1146  	//
  1147  	// This all has to be written to deal with the fact that the
  1148  	// GC may move ahead on its own. For example, when we block
  1149  	// until mark termination N, we may wake up in cycle N+2.
  1150  
  1151  	// Wait until the current sweep termination, mark, and mark
  1152  	// termination complete.
  1153  	n := atomic.Load(&work.cycles)
  1154  	gcWaitOnMark(n)
  1155  
  1156  	// We're now in sweep N or later. Trigger GC cycle N+1, which
  1157  	// will first finish sweep N if necessary and then enter sweep
  1158  	// termination N+1.
  1159  	gcStart(gcTrigger{kind: gcTriggerCycle, n: n + 1})
  1160  
  1161  	// Wait for mark termination N+1 to complete.
  1162  	gcWaitOnMark(n + 1)
  1163  
  1164  	// Finish sweep N+1 before returning. We do this both to
  1165  	// complete the cycle and because runtime.GC() is often used
  1166  	// as part of tests and benchmarks to get the system into a
  1167  	// relatively stable and isolated state.
  1168  	for atomic.Load(&work.cycles) == n+1 && sweepone() != ^uintptr(0) {
  1169  		sweep.nbgsweep++
  1170  		Gosched()
  1171  	}
  1172  
  1173  	// Callers may assume that the heap profile reflects the
  1174  	// just-completed cycle when this returns (historically this
  1175  	// happened because this was a STW GC), but right now the
  1176  	// profile still reflects mark termination N, not N+1.
  1177  	//
  1178  	// As soon as all of the sweep frees from cycle N+1 are done,
  1179  	// we can go ahead and publish the heap profile.
  1180  	//
  1181  	// First, wait for sweeping to finish. (We know there are no
  1182  	// more spans on the sweep queue, but we may be concurrently
  1183  	// sweeping spans, so we have to wait.)
  1184  	for atomic.Load(&work.cycles) == n+1 && atomic.Load(&mheap_.sweepers) != 0 {
  1185  		Gosched()
  1186  	}
  1187  
  1188  	// Now we're really done with sweeping, so we can publish the
  1189  	// stable heap profile. Only do this if we haven't already hit
  1190  	// another mark termination.
  1191  	mp := acquirem()
  1192  	cycle := atomic.Load(&work.cycles)
  1193  	if cycle == n+1 || (gcphase == _GCmark && cycle == n+2) {
  1194  		mProf_PostSweep()
  1195  	}
  1196  	releasem(mp)
  1197  }
  1198  
  1199  // gcWaitOnMark blocks until GC finishes the Nth mark phase. If GC has
  1200  // already completed this mark phase, it returns immediately.
  1201  func gcWaitOnMark(n uint32) {
  1202  	for {
  1203  		// Disable phase transitions.
  1204  		lock(&work.sweepWaiters.lock)
  1205  		nMarks := atomic.Load(&work.cycles)
  1206  		if gcphase != _GCmark {
  1207  			// We've already completed this cycle's mark.
  1208  			nMarks++
  1209  		}
  1210  		if nMarks > n {
  1211  			// We're done.
  1212  			unlock(&work.sweepWaiters.lock)
  1213  			return
  1214  		}
  1215  
  1216  		// Wait until sweep termination, mark, and mark
  1217  		// termination of cycle N complete.
  1218  		work.sweepWaiters.list.push(getg())
  1219  		goparkunlock(&work.sweepWaiters.lock, waitReasonWaitForGCCycle, traceEvGoBlock, 1)
  1220  	}
  1221  }
  1222  
  1223  // gcMode indicates how concurrent a GC cycle should be.
  1224  type gcMode int
  1225  
  1226  const (
  1227  	gcBackgroundMode gcMode = iota // concurrent GC and sweep
  1228  	gcForceMode                    // stop-the-world GC now, concurrent sweep
  1229  	gcForceBlockMode               // stop-the-world GC now and STW sweep (forced by user)
  1230  )
  1231  
  1232  // A gcTrigger is a predicate for starting a GC cycle. Specifically,
  1233  // it is an exit condition for the _GCoff phase.
  1234  type gcTrigger struct {
  1235  	kind gcTriggerKind
  1236  	now  int64  // gcTriggerTime: current time
  1237  	n    uint32 // gcTriggerCycle: cycle number to start
  1238  }
  1239  
  1240  type gcTriggerKind int
  1241  
  1242  const (
  1243  	// gcTriggerHeap indicates that a cycle should be started when
  1244  	// the heap size reaches the trigger heap size computed by the
  1245  	// controller.
  1246  	gcTriggerHeap gcTriggerKind = iota
  1247  
  1248  	// gcTriggerTime indicates that a cycle should be started when
  1249  	// it's been more than forcegcperiod nanoseconds since the
  1250  	// previous GC cycle.
  1251  	gcTriggerTime
  1252  
  1253  	// gcTriggerCycle indicates that a cycle should be started if
  1254  	// we have not yet started cycle number gcTrigger.n (relative
  1255  	// to work.cycles).
  1256  	gcTriggerCycle
  1257  )
  1258  
  1259  // test reports whether the trigger condition is satisfied, meaning
  1260  // that the exit condition for the _GCoff phase has been met. The exit
  1261  // condition should be tested when allocating.
  1262  func (t gcTrigger) test() bool {
  1263  	if !memstats.enablegc || panicking != 0 || gcphase != _GCoff {
  1264  		return false
  1265  	}
  1266  	switch t.kind {
  1267  	case gcTriggerHeap:
  1268  		// Non-atomic access to heap_live for performance. If
  1269  		// we are going to trigger on this, this thread just
  1270  		// atomically wrote heap_live anyway and we'll see our
  1271  		// own write.
  1272  		return memstats.heap_live >= memstats.gc_trigger
  1273  	case gcTriggerTime:
  1274  		if gcpercent < 0 {
  1275  			return false
  1276  		}
  1277  		lastgc := int64(atomic.Load64(&memstats.last_gc_nanotime))
  1278  		return lastgc != 0 && t.now-lastgc > forcegcperiod
  1279  	case gcTriggerCycle:
  1280  		// t.n > work.cycles, but accounting for wraparound.
  1281  		return int32(t.n-work.cycles) > 0
  1282  	}
  1283  	return true
  1284  }
  1285  
  1286  // gcStart starts the GC. It transitions from _GCoff to _GCmark (if
  1287  // debug.gcstoptheworld == 0) or performs all of GC (if
  1288  // debug.gcstoptheworld != 0).
  1289  //
  1290  // This may return without performing this transition in some cases,
  1291  // such as when called on a system stack or with locks held.
  1292  func gcStart(trigger gcTrigger) {
  1293  	// Since this is called from malloc and malloc is called in
  1294  	// the guts of a number of libraries that might be holding
  1295  	// locks, don't attempt to start GC in non-preemptible or
  1296  	// potentially unstable situations.
  1297  	mp := acquirem()
  1298  	if gp := getg(); gp == mp.g0 || mp.locks > 1 || mp.preemptoff != "" {
  1299  		releasem(mp)
  1300  		return
  1301  	}
  1302  	releasem(mp)
  1303  	mp = nil
  1304  
  1305  	// Pick up the remaining unswept/not being swept spans concurrently
  1306  	//
  1307  	// This shouldn't happen if we're being invoked in background
  1308  	// mode since proportional sweep should have just finished
  1309  	// sweeping everything, but rounding errors, etc, may leave a
  1310  	// few spans unswept. In forced mode, this is necessary since
  1311  	// GC can be forced at any point in the sweeping cycle.
  1312  	//
  1313  	// We check the transition condition continuously here in case
  1314  	// this G gets delayed in to the next GC cycle.
  1315  	for trigger.test() && sweepone() != ^uintptr(0) {
  1316  		sweep.nbgsweep++
  1317  	}
  1318  
  1319  	// Perform GC initialization and the sweep termination
  1320  	// transition.
  1321  	semacquire(&work.startSema)
  1322  	// Re-check transition condition under transition lock.
  1323  	if !trigger.test() {
  1324  		semrelease(&work.startSema)
  1325  		return
  1326  	}
  1327  
  1328  	// For stats, check if this GC was forced by the user.
  1329  	work.userForced = trigger.kind == gcTriggerCycle
  1330  
  1331  	// In gcstoptheworld debug mode, upgrade the mode accordingly.
  1332  	// We do this after re-checking the transition condition so
  1333  	// that multiple goroutines that detect the heap trigger don't
  1334  	// start multiple STW GCs.
  1335  	mode := gcBackgroundMode
  1336  	if debug.gcstoptheworld == 1 {
  1337  		mode = gcForceMode
  1338  	} else if debug.gcstoptheworld == 2 {
  1339  		mode = gcForceBlockMode
  1340  	}
  1341  
  1342  	// Ok, we're doing it! Stop everybody else
  1343  	semacquire(&gcsema)
  1344  	semacquire(&worldsema)
  1345  
  1346  	if trace.enabled {
  1347  		traceGCStart()
  1348  	}
  1349  
  1350  	// Check that all Ps have finished deferred mcache flushes.
  1351  	for _, p := range allp {
  1352  		if fg := atomic.Load(&p.mcache.flushGen); fg != mheap_.sweepgen {
  1353  			println("runtime: p", p.id, "flushGen", fg, "!= sweepgen", mheap_.sweepgen)
  1354  			throw("p mcache not flushed")
  1355  		}
  1356  	}
  1357  
  1358  	gcBgMarkStartWorkers()
  1359  
  1360  	systemstack(gcResetMarkState)
  1361  
  1362  	work.stwprocs, work.maxprocs = gomaxprocs, gomaxprocs
  1363  	if work.stwprocs > ncpu {
  1364  		// This is used to compute CPU time of the STW phases,
  1365  		// so it can't be more than ncpu, even if GOMAXPROCS is.
  1366  		work.stwprocs = ncpu
  1367  	}
  1368  	work.heap0 = atomic.Load64(&memstats.heap_live)
  1369  	work.pauseNS = 0
  1370  	work.mode = mode
  1371  
  1372  	now := nanotime()
  1373  	work.tSweepTerm = now
  1374  	work.pauseStart = now
  1375  	if trace.enabled {
  1376  		traceGCSTWStart(1)
  1377  	}
  1378  	systemstack(stopTheWorldWithSema)
  1379  	// Finish sweep before we start concurrent scan.
  1380  	systemstack(func() {
  1381  		finishsweep_m()
  1382  	})
  1383  
  1384  	// clearpools before we start the GC. If we wait they memory will not be
  1385  	// reclaimed until the next GC cycle.
  1386  	clearpools()
  1387  
  1388  	work.cycles++
  1389  
  1390  	gcController.startCycle()
  1391  	work.heapGoal = memstats.next_gc
  1392  
  1393  	// In STW mode, disable scheduling of user Gs. This may also
  1394  	// disable scheduling of this goroutine, so it may block as
  1395  	// soon as we start the world again.
  1396  	if mode != gcBackgroundMode {
  1397  		schedEnableUser(false)
  1398  	}
  1399  
  1400  	// Enter concurrent mark phase and enable
  1401  	// write barriers.
  1402  	//
  1403  	// Because the world is stopped, all Ps will
  1404  	// observe that write barriers are enabled by
  1405  	// the time we start the world and begin
  1406  	// scanning.
  1407  	//
  1408  	// Write barriers must be enabled before assists are
  1409  	// enabled because they must be enabled before
  1410  	// any non-leaf heap objects are marked. Since
  1411  	// allocations are blocked until assists can
  1412  	// happen, we want enable assists as early as
  1413  	// possible.
  1414  	setGCPhase(_GCmark)
  1415  
  1416  	gcBgMarkPrepare() // Must happen before assist enable.
  1417  	gcMarkRootPrepare()
  1418  
  1419  	// Mark all active tinyalloc blocks. Since we're
  1420  	// allocating from these, they need to be black like
  1421  	// other allocations. The alternative is to blacken
  1422  	// the tiny block on every allocation from it, which
  1423  	// would slow down the tiny allocator.
  1424  	gcMarkTinyAllocs()
  1425  
  1426  	// At this point all Ps have enabled the write
  1427  	// barrier, thus maintaining the no white to
  1428  	// black invariant. Enable mutator assists to
  1429  	// put back-pressure on fast allocating
  1430  	// mutators.
  1431  	atomic.Store(&gcBlackenEnabled, 1)
  1432  
  1433  	// Assists and workers can start the moment we start
  1434  	// the world.
  1435  	gcController.markStartTime = now
  1436  
  1437  	// In STW mode, we could block the instant systemstack
  1438  	// returns, so make sure we're not preemptible.
  1439  	mp = acquirem()
  1440  
  1441  	// Concurrent mark.
  1442  	systemstack(func() {
  1443  		now = startTheWorldWithSema(trace.enabled)
  1444  		work.pauseNS += now - work.pauseStart
  1445  		work.tMark = now
  1446  		memstats.gcPauseDist.record(now - work.pauseStart)
  1447  	})
  1448  
  1449  	// Release the world sema before Gosched() in STW mode
  1450  	// because we will need to reacquire it later but before
  1451  	// this goroutine becomes runnable again, and we could
  1452  	// self-deadlock otherwise.
  1453  	semrelease(&worldsema)
  1454  	releasem(mp)
  1455  
  1456  	// Make sure we block instead of returning to user code
  1457  	// in STW mode.
  1458  	if mode != gcBackgroundMode {
  1459  		Gosched()
  1460  	}
  1461  
  1462  	semrelease(&work.startSema)
  1463  }
  1464  
  1465  // gcMarkDoneFlushed counts the number of P's with flushed work.
  1466  //
  1467  // Ideally this would be a captured local in gcMarkDone, but forEachP
  1468  // escapes its callback closure, so it can't capture anything.
  1469  //
  1470  // This is protected by markDoneSema.
  1471  var gcMarkDoneFlushed uint32
  1472  
  1473  // gcMarkDone transitions the GC from mark to mark termination if all
  1474  // reachable objects have been marked (that is, there are no grey
  1475  // objects and can be no more in the future). Otherwise, it flushes
  1476  // all local work to the global queues where it can be discovered by
  1477  // other workers.
  1478  //
  1479  // This should be called when all local mark work has been drained and
  1480  // there are no remaining workers. Specifically, when
  1481  //
  1482  //   work.nwait == work.nproc && !gcMarkWorkAvailable(p)
  1483  //
  1484  // The calling context must be preemptible.
  1485  //
  1486  // Flushing local work is important because idle Ps may have local
  1487  // work queued. This is the only way to make that work visible and
  1488  // drive GC to completion.
  1489  //
  1490  // It is explicitly okay to have write barriers in this function. If
  1491  // it does transition to mark termination, then all reachable objects
  1492  // have been marked, so the write barrier cannot shade any more
  1493  // objects.
  1494  func gcMarkDone() {
  1495  	// Ensure only one thread is running the ragged barrier at a
  1496  	// time.
  1497  	semacquire(&work.markDoneSema)
  1498  
  1499  top:
  1500  	// Re-check transition condition under transition lock.
  1501  	//
  1502  	// It's critical that this checks the global work queues are
  1503  	// empty before performing the ragged barrier. Otherwise,
  1504  	// there could be global work that a P could take after the P
  1505  	// has passed the ragged barrier.
  1506  	if !(gcphase == _GCmark && work.nwait == work.nproc && !gcMarkWorkAvailable(nil)) {
  1507  		semrelease(&work.markDoneSema)
  1508  		return
  1509  	}
  1510  
  1511  	// forEachP needs worldsema to execute, and we'll need it to
  1512  	// stop the world later, so acquire worldsema now.
  1513  	semacquire(&worldsema)
  1514  
  1515  	// Flush all local buffers and collect flushedWork flags.
  1516  	gcMarkDoneFlushed = 0
  1517  	systemstack(func() {
  1518  		gp := getg().m.curg
  1519  		// Mark the user stack as preemptible so that it may be scanned.
  1520  		// Otherwise, our attempt to force all P's to a safepoint could
  1521  		// result in a deadlock as we attempt to preempt a worker that's
  1522  		// trying to preempt us (e.g. for a stack scan).
  1523  		casgstatus(gp, _Grunning, _Gwaiting)
  1524  		forEachP(func(_p_ *p) {
  1525  			// Flush the write barrier buffer, since this may add
  1526  			// work to the gcWork.
  1527  			wbBufFlush1(_p_)
  1528  
  1529  			// Flush the gcWork, since this may create global work
  1530  			// and set the flushedWork flag.
  1531  			//
  1532  			// TODO(austin): Break up these workbufs to
  1533  			// better distribute work.
  1534  			_p_.gcw.dispose()
  1535  			// Collect the flushedWork flag.
  1536  			if _p_.gcw.flushedWork {
  1537  				atomic.Xadd(&gcMarkDoneFlushed, 1)
  1538  				_p_.gcw.flushedWork = false
  1539  			}
  1540  		})
  1541  		casgstatus(gp, _Gwaiting, _Grunning)
  1542  	})
  1543  
  1544  	if gcMarkDoneFlushed != 0 {
  1545  		// More grey objects were discovered since the
  1546  		// previous termination check, so there may be more
  1547  		// work to do. Keep going. It's possible the
  1548  		// transition condition became true again during the
  1549  		// ragged barrier, so re-check it.
  1550  		semrelease(&worldsema)
  1551  		goto top
  1552  	}
  1553  
  1554  	// There was no global work, no local work, and no Ps
  1555  	// communicated work since we took markDoneSema. Therefore
  1556  	// there are no grey objects and no more objects can be
  1557  	// shaded. Transition to mark termination.
  1558  	now := nanotime()
  1559  	work.tMarkTerm = now
  1560  	work.pauseStart = now
  1561  	getg().m.preemptoff = "gcing"
  1562  	if trace.enabled {
  1563  		traceGCSTWStart(0)
  1564  	}
  1565  	systemstack(stopTheWorldWithSema)
  1566  	// The gcphase is _GCmark, it will transition to _GCmarktermination
  1567  	// below. The important thing is that the wb remains active until
  1568  	// all marking is complete. This includes writes made by the GC.
  1569  
  1570  	// There is sometimes work left over when we enter mark termination due
  1571  	// to write barriers performed after the completion barrier above.
  1572  	// Detect this and resume concurrent mark. This is obviously
  1573  	// unfortunate.
  1574  	//
  1575  	// See issue #27993 for details.
  1576  	//
  1577  	// Switch to the system stack to call wbBufFlush1, though in this case
  1578  	// it doesn't matter because we're non-preemptible anyway.
  1579  	restart := false
  1580  	systemstack(func() {
  1581  		for _, p := range allp {
  1582  			wbBufFlush1(p)
  1583  			if !p.gcw.empty() {
  1584  				restart = true
  1585  				break
  1586  			}
  1587  		}
  1588  	})
  1589  	if restart {
  1590  		getg().m.preemptoff = ""
  1591  		systemstack(func() {
  1592  			now := startTheWorldWithSema(true)
  1593  			work.pauseNS += now - work.pauseStart
  1594  			memstats.gcPauseDist.record(now - work.pauseStart)
  1595  		})
  1596  		semrelease(&worldsema)
  1597  		goto top
  1598  	}
  1599  
  1600  	// Disable assists and background workers. We must do
  1601  	// this before waking blocked assists.
  1602  	atomic.Store(&gcBlackenEnabled, 0)
  1603  
  1604  	// Wake all blocked assists. These will run when we
  1605  	// start the world again.
  1606  	gcWakeAllAssists()
  1607  
  1608  	// Likewise, release the transition lock. Blocked
  1609  	// workers and assists will run when we start the
  1610  	// world again.
  1611  	semrelease(&work.markDoneSema)
  1612  
  1613  	// In STW mode, re-enable user goroutines. These will be
  1614  	// queued to run after we start the world.
  1615  	schedEnableUser(true)
  1616  
  1617  	// endCycle depends on all gcWork cache stats being flushed.
  1618  	// The termination algorithm above ensured that up to
  1619  	// allocations since the ragged barrier.
  1620  	nextTriggerRatio := gcController.endCycle()
  1621  
  1622  	// Perform mark termination. This will restart the world.
  1623  	gcMarkTermination(nextTriggerRatio)
  1624  }
  1625  
  1626  // World must be stopped and mark assists and background workers must be
  1627  // disabled.
  1628  func gcMarkTermination(nextTriggerRatio float64) {
  1629  	// Start marktermination (write barrier remains enabled for now).
  1630  	setGCPhase(_GCmarktermination)
  1631  
  1632  	work.heap1 = memstats.heap_live
  1633  	startTime := nanotime()
  1634  
  1635  	mp := acquirem()
  1636  	mp.preemptoff = "gcing"
  1637  	_g_ := getg()
  1638  	_g_.m.traceback = 2
  1639  	gp := _g_.m.curg
  1640  	casgstatus(gp, _Grunning, _Gwaiting)
  1641  	gp.waitreason = waitReasonGarbageCollection
  1642  
  1643  	// Run gc on the g0 stack. We do this so that the g stack
  1644  	// we're currently running on will no longer change. Cuts
  1645  	// the root set down a bit (g0 stacks are not scanned, and
  1646  	// we don't need to scan gc's internal state).  We also
  1647  	// need to switch to g0 so we can shrink the stack.
  1648  	systemstack(func() {
  1649  		gcMark(startTime)
  1650  		// Must return immediately.
  1651  		// The outer function's stack may have moved
  1652  		// during gcMark (it shrinks stacks, including the
  1653  		// outer function's stack), so we must not refer
  1654  		// to any of its variables. Return back to the
  1655  		// non-system stack to pick up the new addresses
  1656  		// before continuing.
  1657  	})
  1658  
  1659  	systemstack(func() {
  1660  		work.heap2 = work.bytesMarked
  1661  		if debug.gccheckmark > 0 {
  1662  			// Run a full non-parallel, stop-the-world
  1663  			// mark using checkmark bits, to check that we
  1664  			// didn't forget to mark anything during the
  1665  			// concurrent mark process.
  1666  			startCheckmarks()
  1667  			gcResetMarkState()
  1668  			gcw := &getg().m.p.ptr().gcw
  1669  			gcDrain(gcw, 0)
  1670  			wbBufFlush1(getg().m.p.ptr())
  1671  			gcw.dispose()
  1672  			endCheckmarks()
  1673  		}
  1674  
  1675  		// marking is complete so we can turn the write barrier off
  1676  		setGCPhase(_GCoff)
  1677  		gcSweep(work.mode)
  1678  	})
  1679  
  1680  	_g_.m.traceback = 0
  1681  	casgstatus(gp, _Gwaiting, _Grunning)
  1682  
  1683  	if trace.enabled {
  1684  		traceGCDone()
  1685  	}
  1686  
  1687  	// all done
  1688  	mp.preemptoff = ""
  1689  
  1690  	if gcphase != _GCoff {
  1691  		throw("gc done but gcphase != _GCoff")
  1692  	}
  1693  
  1694  	// Record next_gc and heap_inuse for scavenger.
  1695  	memstats.last_next_gc = memstats.next_gc
  1696  	memstats.last_heap_inuse = memstats.heap_inuse
  1697  
  1698  	// Update GC trigger and pacing for the next cycle.
  1699  	gcSetTriggerRatio(nextTriggerRatio)
  1700  
  1701  	// Update timing memstats
  1702  	now := nanotime()
  1703  	sec, nsec, _ := time_now()
  1704  	unixNow := sec*1e9 + int64(nsec)
  1705  	work.pauseNS += now - work.pauseStart
  1706  	work.tEnd = now
  1707  	memstats.gcPauseDist.record(now - work.pauseStart)
  1708  	atomic.Store64(&memstats.last_gc_unix, uint64(unixNow)) // must be Unix time to make sense to user
  1709  	atomic.Store64(&memstats.last_gc_nanotime, uint64(now)) // monotonic time for us
  1710  	memstats.pause_ns[memstats.numgc%uint32(len(memstats.pause_ns))] = uint64(work.pauseNS)
  1711  	memstats.pause_end[memstats.numgc%uint32(len(memstats.pause_end))] = uint64(unixNow)
  1712  	memstats.pause_total_ns += uint64(work.pauseNS)
  1713  
  1714  	// Update work.totaltime.
  1715  	sweepTermCpu := int64(work.stwprocs) * (work.tMark - work.tSweepTerm)
  1716  	// We report idle marking time below, but omit it from the
  1717  	// overall utilization here since it's "free".
  1718  	markCpu := gcController.assistTime + gcController.dedicatedMarkTime + gcController.fractionalMarkTime
  1719  	markTermCpu := int64(work.stwprocs) * (work.tEnd - work.tMarkTerm)
  1720  	cycleCpu := sweepTermCpu + markCpu + markTermCpu
  1721  	work.totaltime += cycleCpu
  1722  
  1723  	// Compute overall GC CPU utilization.
  1724  	totalCpu := sched.totaltime + (now-sched.procresizetime)*int64(gomaxprocs)
  1725  	memstats.gc_cpu_fraction = float64(work.totaltime) / float64(totalCpu)
  1726  
  1727  	// Reset sweep state.
  1728  	sweep.nbgsweep = 0
  1729  	sweep.npausesweep = 0
  1730  
  1731  	if work.userForced {
  1732  		memstats.numforcedgc++
  1733  	}
  1734  
  1735  	// Bump GC cycle count and wake goroutines waiting on sweep.
  1736  	lock(&work.sweepWaiters.lock)
  1737  	memstats.numgc++
  1738  	injectglist(&work.sweepWaiters.list)
  1739  	unlock(&work.sweepWaiters.lock)
  1740  
  1741  	// Finish the current heap profiling cycle and start a new
  1742  	// heap profiling cycle. We do this before starting the world
  1743  	// so events don't leak into the wrong cycle.
  1744  	mProf_NextCycle()
  1745  
  1746  	systemstack(func() { startTheWorldWithSema(true) })
  1747  
  1748  	// Flush the heap profile so we can start a new cycle next GC.
  1749  	// This is relatively expensive, so we don't do it with the
  1750  	// world stopped.
  1751  	mProf_Flush()
  1752  
  1753  	// Prepare workbufs for freeing by the sweeper. We do this
  1754  	// asynchronously because it can take non-trivial time.
  1755  	prepareFreeWorkbufs()
  1756  
  1757  	// Free stack spans. This must be done between GC cycles.
  1758  	systemstack(freeStackSpans)
  1759  
  1760  	// Ensure all mcaches are flushed. Each P will flush its own
  1761  	// mcache before allocating, but idle Ps may not. Since this
  1762  	// is necessary to sweep all spans, we need to ensure all
  1763  	// mcaches are flushed before we start the next GC cycle.
  1764  	systemstack(func() {
  1765  		forEachP(func(_p_ *p) {
  1766  			_p_.mcache.prepareForSweep()
  1767  		})
  1768  	})
  1769  
  1770  	// Print gctrace before dropping worldsema. As soon as we drop
  1771  	// worldsema another cycle could start and smash the stats
  1772  	// we're trying to print.
  1773  	if debug.gctrace > 0 {
  1774  		util := int(memstats.gc_cpu_fraction * 100)
  1775  
  1776  		var sbuf [24]byte
  1777  		printlock()
  1778  		print("gc ", memstats.numgc,
  1779  			" @", string(itoaDiv(sbuf[:], uint64(work.tSweepTerm-runtimeInitTime)/1e6, 3)), "s ",
  1780  			util, "%: ")
  1781  		prev := work.tSweepTerm
  1782  		for i, ns := range []int64{work.tMark, work.tMarkTerm, work.tEnd} {
  1783  			if i != 0 {
  1784  				print("+")
  1785  			}
  1786  			print(string(fmtNSAsMS(sbuf[:], uint64(ns-prev))))
  1787  			prev = ns
  1788  		}
  1789  		print(" ms clock, ")
  1790  		for i, ns := range []int64{sweepTermCpu, gcController.assistTime, gcController.dedicatedMarkTime + gcController.fractionalMarkTime, gcController.idleMarkTime, markTermCpu} {
  1791  			if i == 2 || i == 3 {
  1792  				// Separate mark time components with /.
  1793  				print("/")
  1794  			} else if i != 0 {
  1795  				print("+")
  1796  			}
  1797  			print(string(fmtNSAsMS(sbuf[:], uint64(ns))))
  1798  		}
  1799  		print(" ms cpu, ",
  1800  			work.heap0>>20, "->", work.heap1>>20, "->", work.heap2>>20, " MB, ",
  1801  			work.heapGoal>>20, " MB goal, ",
  1802  			work.maxprocs, " P")
  1803  		if work.userForced {
  1804  			print(" (forced)")
  1805  		}
  1806  		print("\n")
  1807  		printunlock()
  1808  	}
  1809  
  1810  	semrelease(&worldsema)
  1811  	semrelease(&gcsema)
  1812  	// Careful: another GC cycle may start now.
  1813  
  1814  	releasem(mp)
  1815  	mp = nil
  1816  
  1817  	// now that gc is done, kick off finalizer thread if needed
  1818  	if !concurrentSweep {
  1819  		// give the queued finalizers, if any, a chance to run
  1820  		Gosched()
  1821  	}
  1822  }
  1823  
  1824  // gcBgMarkStartWorkers prepares background mark worker goroutines. These
  1825  // goroutines will not run until the mark phase, but they must be started while
  1826  // the work is not stopped and from a regular G stack. The caller must hold
  1827  // worldsema.
  1828  func gcBgMarkStartWorkers() {
  1829  	// Background marking is performed by per-P G's. Ensure that each P has
  1830  	// a background GC G.
  1831  	//
  1832  	// Worker Gs don't exit if gomaxprocs is reduced. If it is raised
  1833  	// again, we can reuse the old workers; no need to create new workers.
  1834  	for gcBgMarkWorkerCount < gomaxprocs {
  1835  		go gcBgMarkWorker()
  1836  
  1837  		notetsleepg(&work.bgMarkReady, -1)
  1838  		noteclear(&work.bgMarkReady)
  1839  		// The worker is now guaranteed to be added to the pool before
  1840  		// its P's next findRunnableGCWorker.
  1841  
  1842  		gcBgMarkWorkerCount++
  1843  	}
  1844  }
  1845  
  1846  // gcBgMarkPrepare sets up state for background marking.
  1847  // Mutator assists must not yet be enabled.
  1848  func gcBgMarkPrepare() {
  1849  	// Background marking will stop when the work queues are empty
  1850  	// and there are no more workers (note that, since this is
  1851  	// concurrent, this may be a transient state, but mark
  1852  	// termination will clean it up). Between background workers
  1853  	// and assists, we don't really know how many workers there
  1854  	// will be, so we pretend to have an arbitrarily large number
  1855  	// of workers, almost all of which are "waiting". While a
  1856  	// worker is working it decrements nwait. If nproc == nwait,
  1857  	// there are no workers.
  1858  	work.nproc = ^uint32(0)
  1859  	work.nwait = ^uint32(0)
  1860  }
  1861  
  1862  // gcBgMarkWorker is an entry in the gcBgMarkWorkerPool. It points to a single
  1863  // gcBgMarkWorker goroutine.
  1864  type gcBgMarkWorkerNode struct {
  1865  	// Unused workers are managed in a lock-free stack. This field must be first.
  1866  	node lfnode
  1867  
  1868  	// The g of this worker.
  1869  	gp guintptr
  1870  
  1871  	// Release this m on park. This is used to communicate with the unlock
  1872  	// function, which cannot access the G's stack. It is unused outside of
  1873  	// gcBgMarkWorker().
  1874  	m muintptr
  1875  }
  1876  
  1877  func gcBgMarkWorker() {
  1878  	gp := getg()
  1879  
  1880  	// We pass node to a gopark unlock function, so it can't be on
  1881  	// the stack (see gopark). Prevent deadlock from recursively
  1882  	// starting GC by disabling preemption.
  1883  	gp.m.preemptoff = "GC worker init"
  1884  	node := new(gcBgMarkWorkerNode)
  1885  	gp.m.preemptoff = ""
  1886  
  1887  	node.gp.set(gp)
  1888  
  1889  	node.m.set(acquirem())
  1890  	notewakeup(&work.bgMarkReady)
  1891  	// After this point, the background mark worker is generally scheduled
  1892  	// cooperatively by gcController.findRunnableGCWorker. While performing
  1893  	// work on the P, preemption is disabled because we are working on
  1894  	// P-local work buffers. When the preempt flag is set, this puts itself
  1895  	// into _Gwaiting to be woken up by gcController.findRunnableGCWorker
  1896  	// at the appropriate time.
  1897  	//
  1898  	// When preemption is enabled (e.g., while in gcMarkDone), this worker
  1899  	// may be preempted and schedule as a _Grunnable G from a runq. That is
  1900  	// fine; it will eventually gopark again for further scheduling via
  1901  	// findRunnableGCWorker.
  1902  	//
  1903  	// Since we disable preemption before notifying bgMarkReady, we
  1904  	// guarantee that this G will be in the worker pool for the next
  1905  	// findRunnableGCWorker. This isn't strictly necessary, but it reduces
  1906  	// latency between _GCmark starting and the workers starting.
  1907  
  1908  	for {
  1909  		// Go to sleep until woken by
  1910  		// gcController.findRunnableGCWorker.
  1911  		gopark(func(g *g, nodep unsafe.Pointer) bool {
  1912  			node := (*gcBgMarkWorkerNode)(nodep)
  1913  
  1914  			if mp := node.m.ptr(); mp != nil {
  1915  				// The worker G is no longer running; release
  1916  				// the M.
  1917  				//
  1918  				// N.B. it is _safe_ to release the M as soon
  1919  				// as we are no longer performing P-local mark
  1920  				// work.
  1921  				//
  1922  				// However, since we cooperatively stop work
  1923  				// when gp.preempt is set, if we releasem in
  1924  				// the loop then the following call to gopark
  1925  				// would immediately preempt the G. This is
  1926  				// also safe, but inefficient: the G must
  1927  				// schedule again only to enter gopark and park
  1928  				// again. Thus, we defer the release until
  1929  				// after parking the G.
  1930  				releasem(mp)
  1931  			}
  1932  
  1933  			// Release this G to the pool.
  1934  			gcBgMarkWorkerPool.push(&node.node)
  1935  			// Note that at this point, the G may immediately be
  1936  			// rescheduled and may be running.
  1937  			return true
  1938  		}, unsafe.Pointer(node), waitReasonGCWorkerIdle, traceEvGoBlock, 0)
  1939  
  1940  		// Preemption must not occur here, or another G might see
  1941  		// p.gcMarkWorkerMode.
  1942  
  1943  		// Disable preemption so we can use the gcw. If the
  1944  		// scheduler wants to preempt us, we'll stop draining,
  1945  		// dispose the gcw, and then preempt.
  1946  		node.m.set(acquirem())
  1947  		pp := gp.m.p.ptr() // P can't change with preemption disabled.
  1948  
  1949  		if gcBlackenEnabled == 0 {
  1950  			println("worker mode", pp.gcMarkWorkerMode)
  1951  			throw("gcBgMarkWorker: blackening not enabled")
  1952  		}
  1953  
  1954  		if pp.gcMarkWorkerMode == gcMarkWorkerNotWorker {
  1955  			throw("gcBgMarkWorker: mode not set")
  1956  		}
  1957  
  1958  		startTime := nanotime()
  1959  		pp.gcMarkWorkerStartTime = startTime
  1960  
  1961  		decnwait := atomic.Xadd(&work.nwait, -1)
  1962  		if decnwait == work.nproc {
  1963  			println("runtime: work.nwait=", decnwait, "work.nproc=", work.nproc)
  1964  			throw("work.nwait was > work.nproc")
  1965  		}
  1966  
  1967  		systemstack(func() {
  1968  			// Mark our goroutine preemptible so its stack
  1969  			// can be scanned. This lets two mark workers
  1970  			// scan each other (otherwise, they would
  1971  			// deadlock). We must not modify anything on
  1972  			// the G stack. However, stack shrinking is
  1973  			// disabled for mark workers, so it is safe to
  1974  			// read from the G stack.
  1975  			casgstatus(gp, _Grunning, _Gwaiting)
  1976  			switch pp.gcMarkWorkerMode {
  1977  			default:
  1978  				throw("gcBgMarkWorker: unexpected gcMarkWorkerMode")
  1979  			case gcMarkWorkerDedicatedMode:
  1980  				gcDrain(&pp.gcw, gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1981  				if gp.preempt {
  1982  					// We were preempted. This is
  1983  					// a useful signal to kick
  1984  					// everything out of the run
  1985  					// queue so it can run
  1986  					// somewhere else.
  1987  					lock(&sched.lock)
  1988  					for {
  1989  						gp, _ := runqget(pp)
  1990  						if gp == nil {
  1991  							break
  1992  						}
  1993  						globrunqput(gp)
  1994  					}
  1995  					unlock(&sched.lock)
  1996  				}
  1997  				// Go back to draining, this time
  1998  				// without preemption.
  1999  				gcDrain(&pp.gcw, gcDrainFlushBgCredit)
  2000  			case gcMarkWorkerFractionalMode:
  2001  				gcDrain(&pp.gcw, gcDrainFractional|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  2002  			case gcMarkWorkerIdleMode:
  2003  				gcDrain(&pp.gcw, gcDrainIdle|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  2004  			}
  2005  			casgstatus(gp, _Gwaiting, _Grunning)
  2006  		})
  2007  
  2008  		// Account for time.
  2009  		duration := nanotime() - startTime
  2010  		switch pp.gcMarkWorkerMode {
  2011  		case gcMarkWorkerDedicatedMode:
  2012  			atomic.Xaddint64(&gcController.dedicatedMarkTime, duration)
  2013  			atomic.Xaddint64(&gcController.dedicatedMarkWorkersNeeded, 1)
  2014  		case gcMarkWorkerFractionalMode:
  2015  			atomic.Xaddint64(&gcController.fractionalMarkTime, duration)
  2016  			atomic.Xaddint64(&pp.gcFractionalMarkTime, duration)
  2017  		case gcMarkWorkerIdleMode:
  2018  			atomic.Xaddint64(&gcController.idleMarkTime, duration)
  2019  		}
  2020  
  2021  		// Was this the last worker and did we run out
  2022  		// of work?
  2023  		incnwait := atomic.Xadd(&work.nwait, +1)
  2024  		if incnwait > work.nproc {
  2025  			println("runtime: p.gcMarkWorkerMode=", pp.gcMarkWorkerMode,
  2026  				"work.nwait=", incnwait, "work.nproc=", work.nproc)
  2027  			throw("work.nwait > work.nproc")
  2028  		}
  2029  
  2030  		// We'll releasem after this point and thus this P may run
  2031  		// something else. We must clear the worker mode to avoid
  2032  		// attributing the mode to a different (non-worker) G in
  2033  		// traceGoStart.
  2034  		pp.gcMarkWorkerMode = gcMarkWorkerNotWorker
  2035  
  2036  		// If this worker reached a background mark completion
  2037  		// point, signal the main GC goroutine.
  2038  		if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
  2039  			// We don't need the P-local buffers here, allow
  2040  			// preemption becuse we may schedule like a regular
  2041  			// goroutine in gcMarkDone (block on locks, etc).
  2042  			releasem(node.m.ptr())
  2043  			node.m.set(nil)
  2044  
  2045  			gcMarkDone()
  2046  		}
  2047  	}
  2048  }
  2049  
  2050  // gcMarkWorkAvailable reports whether executing a mark worker
  2051  // on p is potentially useful. p may be nil, in which case it only
  2052  // checks the global sources of work.
  2053  func gcMarkWorkAvailable(p *p) bool {
  2054  	if p != nil && !p.gcw.empty() {
  2055  		return true
  2056  	}
  2057  	if !work.full.empty() {
  2058  		return true // global work available
  2059  	}
  2060  	if work.markrootNext < work.markrootJobs {
  2061  		return true // root scan work available
  2062  	}
  2063  	return false
  2064  }
  2065  
  2066  // gcMark runs the mark (or, for concurrent GC, mark termination)
  2067  // All gcWork caches must be empty.
  2068  // STW is in effect at this point.
  2069  func gcMark(start_time int64) {
  2070  	if debug.allocfreetrace > 0 {
  2071  		tracegc()
  2072  	}
  2073  
  2074  	if gcphase != _GCmarktermination {
  2075  		throw("in gcMark expecting to see gcphase as _GCmarktermination")
  2076  	}
  2077  	work.tstart = start_time
  2078  
  2079  	// Check that there's no marking work remaining.
  2080  	if work.full != 0 || work.markrootNext < work.markrootJobs {
  2081  		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")
  2082  		panic("non-empty mark queue after concurrent mark")
  2083  	}
  2084  
  2085  	if debug.gccheckmark > 0 {
  2086  		// This is expensive when there's a large number of
  2087  		// Gs, so only do it if checkmark is also enabled.
  2088  		gcMarkRootCheck()
  2089  	}
  2090  	if work.full != 0 {
  2091  		throw("work.full != 0")
  2092  	}
  2093  
  2094  	// Clear out buffers and double-check that all gcWork caches
  2095  	// are empty. This should be ensured by gcMarkDone before we
  2096  	// enter mark termination.
  2097  	//
  2098  	// TODO: We could clear out buffers just before mark if this
  2099  	// has a non-negligible impact on STW time.
  2100  	for _, p := range allp {
  2101  		// The write barrier may have buffered pointers since
  2102  		// the gcMarkDone barrier. However, since the barrier
  2103  		// ensured all reachable objects were marked, all of
  2104  		// these must be pointers to black objects. Hence we
  2105  		// can just discard the write barrier buffer.
  2106  		if debug.gccheckmark > 0 {
  2107  			// For debugging, flush the buffer and make
  2108  			// sure it really was all marked.
  2109  			wbBufFlush1(p)
  2110  		} else {
  2111  			p.wbBuf.reset()
  2112  		}
  2113  
  2114  		gcw := &p.gcw
  2115  		if !gcw.empty() {
  2116  			printlock()
  2117  			print("runtime: P ", p.id, " flushedWork ", gcw.flushedWork)
  2118  			if gcw.wbuf1 == nil {
  2119  				print(" wbuf1=<nil>")
  2120  			} else {
  2121  				print(" wbuf1.n=", gcw.wbuf1.nobj)
  2122  			}
  2123  			if gcw.wbuf2 == nil {
  2124  				print(" wbuf2=<nil>")
  2125  			} else {
  2126  				print(" wbuf2.n=", gcw.wbuf2.nobj)
  2127  			}
  2128  			print("\n")
  2129  			throw("P has cached GC work at end of mark termination")
  2130  		}
  2131  		// There may still be cached empty buffers, which we
  2132  		// need to flush since we're going to free them. Also,
  2133  		// there may be non-zero stats because we allocated
  2134  		// black after the gcMarkDone barrier.
  2135  		gcw.dispose()
  2136  	}
  2137  
  2138  	// Update the marked heap stat.
  2139  	memstats.heap_marked = work.bytesMarked
  2140  
  2141  	// Flush scanAlloc from each mcache since we're about to modify
  2142  	// heap_scan directly. If we were to flush this later, then scanAlloc
  2143  	// might have incorrect information.
  2144  	for _, p := range allp {
  2145  		c := p.mcache
  2146  		if c == nil {
  2147  			continue
  2148  		}
  2149  		memstats.heap_scan += uint64(c.scanAlloc)
  2150  		c.scanAlloc = 0
  2151  	}
  2152  
  2153  	// Update other GC heap size stats. This must happen after
  2154  	// cachestats (which flushes local statistics to these) and
  2155  	// flushallmcaches (which modifies heap_live).
  2156  	memstats.heap_live = work.bytesMarked
  2157  	memstats.heap_scan = uint64(gcController.scanWork)
  2158  
  2159  	if trace.enabled {
  2160  		traceHeapAlloc()
  2161  	}
  2162  }
  2163  
  2164  // gcSweep must be called on the system stack because it acquires the heap
  2165  // lock. See mheap for details.
  2166  //
  2167  // The world must be stopped.
  2168  //
  2169  //go:systemstack
  2170  func gcSweep(mode gcMode) {
  2171  	assertWorldStopped()
  2172  
  2173  	if gcphase != _GCoff {
  2174  		throw("gcSweep being done but phase is not GCoff")
  2175  	}
  2176  
  2177  	lock(&mheap_.lock)
  2178  	mheap_.sweepgen += 2
  2179  	mheap_.sweepdone = 0
  2180  	mheap_.pagesSwept = 0
  2181  	mheap_.sweepArenas = mheap_.allArenas
  2182  	mheap_.reclaimIndex = 0
  2183  	mheap_.reclaimCredit = 0
  2184  	unlock(&mheap_.lock)
  2185  
  2186  	sweep.centralIndex.clear()
  2187  
  2188  	if !_ConcurrentSweep || mode == gcForceBlockMode {
  2189  		// Special case synchronous sweep.
  2190  		// Record that no proportional sweeping has to happen.
  2191  		lock(&mheap_.lock)
  2192  		mheap_.sweepPagesPerByte = 0
  2193  		unlock(&mheap_.lock)
  2194  		// Sweep all spans eagerly.
  2195  		for sweepone() != ^uintptr(0) {
  2196  			sweep.npausesweep++
  2197  		}
  2198  		// Free workbufs eagerly.
  2199  		prepareFreeWorkbufs()
  2200  		for freeSomeWbufs(false) {
  2201  		}
  2202  		// All "free" events for this mark/sweep cycle have
  2203  		// now happened, so we can make this profile cycle
  2204  		// available immediately.
  2205  		mProf_NextCycle()
  2206  		mProf_Flush()
  2207  		return
  2208  	}
  2209  
  2210  	// Background sweep.
  2211  	lock(&sweep.lock)
  2212  	if sweep.parked {
  2213  		sweep.parked = false
  2214  		ready(sweep.g, 0, true)
  2215  	}
  2216  	unlock(&sweep.lock)
  2217  }
  2218  
  2219  // gcResetMarkState resets global state prior to marking (concurrent
  2220  // or STW) and resets the stack scan state of all Gs.
  2221  //
  2222  // This is safe to do without the world stopped because any Gs created
  2223  // during or after this will start out in the reset state.
  2224  //
  2225  // gcResetMarkState must be called on the system stack because it acquires
  2226  // the heap lock. See mheap for details.
  2227  //
  2228  //go:systemstack
  2229  func gcResetMarkState() {
  2230  	// This may be called during a concurrent phase, so make sure
  2231  	// allgs doesn't change.
  2232  	lock(&allglock)
  2233  	for _, gp := range allgs {
  2234  		gp.gcscandone = false // set to true in gcphasework
  2235  		gp.gcAssistBytes = 0
  2236  	}
  2237  	unlock(&allglock)
  2238  
  2239  	// Clear page marks. This is just 1MB per 64GB of heap, so the
  2240  	// time here is pretty trivial.
  2241  	lock(&mheap_.lock)
  2242  	arenas := mheap_.allArenas
  2243  	unlock(&mheap_.lock)
  2244  	for _, ai := range arenas {
  2245  		ha := mheap_.arenas[ai.l1()][ai.l2()]
  2246  		for i := range ha.pageMarks {
  2247  			ha.pageMarks[i] = 0
  2248  		}
  2249  	}
  2250  
  2251  	work.bytesMarked = 0
  2252  	work.initialHeapLive = atomic.Load64(&memstats.heap_live)
  2253  }
  2254  
  2255  // Hooks for other packages
  2256  
  2257  var poolcleanup func()
  2258  
  2259  //go:linkname sync_runtime_registerPoolCleanup sync.runtime_registerPoolCleanup
  2260  func sync_runtime_registerPoolCleanup(f func()) {
  2261  	poolcleanup = f
  2262  }
  2263  
  2264  func clearpools() {
  2265  	// clear sync.Pools
  2266  	if poolcleanup != nil {
  2267  		poolcleanup()
  2268  	}
  2269  
  2270  	// Clear central sudog cache.
  2271  	// Leave per-P caches alone, they have strictly bounded size.
  2272  	// Disconnect cached list before dropping it on the floor,
  2273  	// so that a dangling ref to one entry does not pin all of them.
  2274  	lock(&sched.sudoglock)
  2275  	var sg, sgnext *sudog
  2276  	for sg = sched.sudogcache; sg != nil; sg = sgnext {
  2277  		sgnext = sg.next
  2278  		sg.next = nil
  2279  	}
  2280  	sched.sudogcache = nil
  2281  	unlock(&sched.sudoglock)
  2282  
  2283  	// Clear central defer pools.
  2284  	// Leave per-P pools alone, they have strictly bounded size.
  2285  	lock(&sched.deferlock)
  2286  	for i := range sched.deferpool {
  2287  		// disconnect cached list before dropping it on the floor,
  2288  		// so that a dangling ref to one entry does not pin all of them.
  2289  		var d, dlink *_defer
  2290  		for d = sched.deferpool[i]; d != nil; d = dlink {
  2291  			dlink = d.link
  2292  			d.link = nil
  2293  		}
  2294  		sched.deferpool[i] = nil
  2295  	}
  2296  	unlock(&sched.deferlock)
  2297  }
  2298  
  2299  // Timing
  2300  
  2301  // itoaDiv formats val/(10**dec) into buf.
  2302  func itoaDiv(buf []byte, val uint64, dec int) []byte {
  2303  	i := len(buf) - 1
  2304  	idec := i - dec
  2305  	for val >= 10 || i >= idec {
  2306  		buf[i] = byte(val%10 + '0')
  2307  		i--
  2308  		if i == idec {
  2309  			buf[i] = '.'
  2310  			i--
  2311  		}
  2312  		val /= 10
  2313  	}
  2314  	buf[i] = byte(val + '0')
  2315  	return buf[i:]
  2316  }
  2317  
  2318  // fmtNSAsMS nicely formats ns nanoseconds as milliseconds.
  2319  func fmtNSAsMS(buf []byte, ns uint64) []byte {
  2320  	if ns >= 10e6 {
  2321  		// Format as whole milliseconds.
  2322  		return itoaDiv(buf, ns/1e6, 0)
  2323  	}
  2324  	// Format two digits of precision, with at most three decimal places.
  2325  	x := ns / 1e3
  2326  	if x == 0 {
  2327  		buf[0] = '0'
  2328  		return buf[:1]
  2329  	}
  2330  	dec := 3
  2331  	for x >= 100 {
  2332  		x /= 10
  2333  		dec--
  2334  	}
  2335  	return itoaDiv(buf, x, dec)
  2336  }
  2337  

View as plain text