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  	// sweepMinHeapDistance is a lower bound on the heap distance
   143  	// (in bytes) reserved for concurrent sweeping between GC
   144  	// cycles.
   145  	sweepMinHeapDistance = 1024 * 1024
   146  )
   147  
   148  // heapminimum is the minimum heap size at which to trigger GC.
   149  // For small heaps, this overrides the usual GOGC*live set rule.
   150  //
   151  // When there is a very small live set but a lot of allocation, simply
   152  // collecting when the heap reaches GOGC*live results in many GC
   153  // cycles and high total per-GC overhead. This minimum amortizes this
   154  // per-GC overhead while keeping the heap reasonably small.
   155  //
   156  // During initialization this is set to 4MB*GOGC/100. In the case of
   157  // GOGC==0, this will set heapminimum to 0, resulting in constant
   158  // collection even when the heap size is small, which is useful for
   159  // debugging.
   160  var heapminimum uint64 = defaultHeapMinimum
   161  
   162  // defaultHeapMinimum is the value of heapminimum for GOGC==100.
   163  const defaultHeapMinimum = 4 << 20
   164  
   165  // Initialized from $GOGC.  GOGC=off means no GC.
   166  var gcpercent int32
   167  
   168  func gcinit() {
   169  	if unsafe.Sizeof(workbuf{}) != _WorkbufSize {
   170  		throw("size of Workbuf is suboptimal")
   171  	}
   172  
   173  	// No sweep on the first cycle.
   174  	mheap_.sweepdone = 1
   175  
   176  	// Set a reasonable initial GC trigger.
   177  	memstats.triggerRatio = 7 / 8.0
   178  
   179  	// Fake a heap_marked value so it looks like a trigger at
   180  	// heapminimum is the appropriate growth from heap_marked.
   181  	// This will go into computing the initial GC goal.
   182  	memstats.heap_marked = uint64(float64(heapminimum) / (1 + memstats.triggerRatio))
   183  
   184  	// Set gcpercent from the environment. This will also compute
   185  	// and set the GC trigger and goal.
   186  	_ = setGCPercent(readgogc())
   187  
   188  	work.startSema = 1
   189  	work.markDoneSema = 1
   190  }
   191  
   192  func readgogc() int32 {
   193  	p := gogetenv("GOGC")
   194  	if p == "off" {
   195  		return -1
   196  	}
   197  	if n, ok := atoi32(p); ok {
   198  		return n
   199  	}
   200  	return 100
   201  }
   202  
   203  // gcenable is called after the bulk of the runtime initialization,
   204  // just before we're about to start letting user code run.
   205  // It kicks off the background sweeper goroutine, the background
   206  // scavenger goroutine, and enables GC.
   207  func gcenable() {
   208  	// Kick off sweeping and scavenging.
   209  	c := make(chan int, 2)
   210  	go bgsweep(c)
   211  	go bgscavenge(c)
   212  	<-c
   213  	<-c
   214  	memstats.enablegc = true // now that runtime is initialized, GC is okay
   215  }
   216  
   217  //go:linkname setGCPercent runtime/debug.setGCPercent
   218  func setGCPercent(in int32) (out int32) {
   219  	// Run on the system stack since we grab the heap lock.
   220  	systemstack(func() {
   221  		lock(&mheap_.lock)
   222  		out = gcpercent
   223  		if in < 0 {
   224  			in = -1
   225  		}
   226  		gcpercent = in
   227  		heapminimum = defaultHeapMinimum * uint64(gcpercent) / 100
   228  		// Update pacing in response to gcpercent change.
   229  		gcSetTriggerRatio(memstats.triggerRatio)
   230  		unlock(&mheap_.lock)
   231  	})
   232  	// Pacing changed, so the scavenger should be awoken.
   233  	wakeScavenger()
   234  
   235  	// If we just disabled GC, wait for any concurrent GC mark to
   236  	// finish so we always return with no GC running.
   237  	if in < 0 {
   238  		gcWaitOnMark(atomic.Load(&work.cycles))
   239  	}
   240  
   241  	return out
   242  }
   243  
   244  // Garbage collector phase.
   245  // Indicates to write barrier and synchronization task to perform.
   246  var gcphase uint32
   247  
   248  // The compiler knows about this variable.
   249  // If you change it, you must change builtin/runtime.go, too.
   250  // If you change the first four bytes, you must also change the write
   251  // barrier insertion code.
   252  var writeBarrier struct {
   253  	enabled bool    // compiler emits a check of this before calling write barrier
   254  	pad     [3]byte // compiler uses 32-bit load for "enabled" field
   255  	needed  bool    // whether we need a write barrier for current GC phase
   256  	cgo     bool    // whether we need a write barrier for a cgo check
   257  	alignme uint64  // guarantee alignment so that compiler can use a 32 or 64-bit load
   258  }
   259  
   260  // gcBlackenEnabled is 1 if mutator assists and background mark
   261  // workers are allowed to blacken objects. This must only be set when
   262  // gcphase == _GCmark.
   263  var gcBlackenEnabled uint32
   264  
   265  const (
   266  	_GCoff             = iota // GC not running; sweeping in background, write barrier disabled
   267  	_GCmark                   // GC marking roots and workbufs: allocate black, write barrier ENABLED
   268  	_GCmarktermination        // GC mark termination: allocate black, P's help GC, write barrier ENABLED
   269  )
   270  
   271  //go:nosplit
   272  func setGCPhase(x uint32) {
   273  	atomic.Store(&gcphase, x)
   274  	writeBarrier.needed = gcphase == _GCmark || gcphase == _GCmarktermination
   275  	writeBarrier.enabled = writeBarrier.needed || writeBarrier.cgo
   276  }
   277  
   278  // gcMarkWorkerMode represents the mode that a concurrent mark worker
   279  // should operate in.
   280  //
   281  // Concurrent marking happens through four different mechanisms. One
   282  // is mutator assists, which happen in response to allocations and are
   283  // not scheduled. The other three are variations in the per-P mark
   284  // workers and are distinguished by gcMarkWorkerMode.
   285  type gcMarkWorkerMode int
   286  
   287  const (
   288  	// gcMarkWorkerDedicatedMode indicates that the P of a mark
   289  	// worker is dedicated to running that mark worker. The mark
   290  	// worker should run without preemption.
   291  	gcMarkWorkerDedicatedMode gcMarkWorkerMode = iota
   292  
   293  	// gcMarkWorkerFractionalMode indicates that a P is currently
   294  	// running the "fractional" mark worker. The fractional worker
   295  	// is necessary when GOMAXPROCS*gcBackgroundUtilization is not
   296  	// an integer. The fractional worker should run until it is
   297  	// preempted and will be scheduled to pick up the fractional
   298  	// part of GOMAXPROCS*gcBackgroundUtilization.
   299  	gcMarkWorkerFractionalMode
   300  
   301  	// gcMarkWorkerIdleMode indicates that a P is running the mark
   302  	// worker because it has nothing else to do. The idle worker
   303  	// should run until it is preempted and account its time
   304  	// against gcController.idleMarkTime.
   305  	gcMarkWorkerIdleMode
   306  )
   307  
   308  // gcMarkWorkerModeStrings are the strings labels of gcMarkWorkerModes
   309  // to use in execution traces.
   310  var gcMarkWorkerModeStrings = [...]string{
   311  	"GC (dedicated)",
   312  	"GC (fractional)",
   313  	"GC (idle)",
   314  }
   315  
   316  // gcController implements the GC pacing controller that determines
   317  // when to trigger concurrent garbage collection and how much marking
   318  // work to do in mutator assists and background marking.
   319  //
   320  // It uses a feedback control algorithm to adjust the memstats.gc_trigger
   321  // trigger based on the heap growth and GC CPU utilization each cycle.
   322  // This algorithm optimizes for heap growth to match GOGC and for CPU
   323  // utilization between assist and background marking to be 25% of
   324  // GOMAXPROCS. The high-level design of this algorithm is documented
   325  // at https://golang.org/s/go15gcpacing.
   326  //
   327  // All fields of gcController are used only during a single mark
   328  // cycle.
   329  var gcController gcControllerState
   330  
   331  type gcControllerState struct {
   332  	// scanWork is the total scan work performed this cycle. This
   333  	// is updated atomically during the cycle. Updates occur in
   334  	// bounded batches, since it is both written and read
   335  	// throughout the cycle. At the end of the cycle, this is how
   336  	// much of the retained heap is scannable.
   337  	//
   338  	// Currently this is the bytes of heap scanned. For most uses,
   339  	// this is an opaque unit of work, but for estimation the
   340  	// definition is important.
   341  	scanWork int64
   342  
   343  	// bgScanCredit is the scan work credit accumulated by the
   344  	// concurrent background scan. This credit is accumulated by
   345  	// the background scan and stolen by mutator assists. This is
   346  	// updated atomically. Updates occur in bounded batches, since
   347  	// it is both written and read throughout the cycle.
   348  	bgScanCredit int64
   349  
   350  	// assistTime is the nanoseconds spent in mutator assists
   351  	// during this cycle. This is updated atomically. Updates
   352  	// occur in bounded batches, since it is both written and read
   353  	// throughout the cycle.
   354  	assistTime int64
   355  
   356  	// dedicatedMarkTime is the nanoseconds spent in dedicated
   357  	// mark workers during this cycle. This is updated atomically
   358  	// at the end of the concurrent mark phase.
   359  	dedicatedMarkTime int64
   360  
   361  	// fractionalMarkTime is the nanoseconds spent in the
   362  	// fractional mark worker during this cycle. This is updated
   363  	// atomically throughout the cycle and will be up-to-date if
   364  	// the fractional mark worker is not currently running.
   365  	fractionalMarkTime int64
   366  
   367  	// idleMarkTime is the nanoseconds spent in idle marking
   368  	// during this cycle. This is updated atomically throughout
   369  	// the cycle.
   370  	idleMarkTime int64
   371  
   372  	// markStartTime is the absolute start time in nanoseconds
   373  	// that assists and background mark workers started.
   374  	markStartTime int64
   375  
   376  	// dedicatedMarkWorkersNeeded is the number of dedicated mark
   377  	// workers that need to be started. This is computed at the
   378  	// beginning of each cycle and decremented atomically as
   379  	// dedicated mark workers get started.
   380  	dedicatedMarkWorkersNeeded int64
   381  
   382  	// assistWorkPerByte is the ratio of scan work to allocated
   383  	// bytes that should be performed by mutator assists. This is
   384  	// computed at the beginning of each cycle and updated every
   385  	// time heap_scan is updated.
   386  	assistWorkPerByte float64
   387  
   388  	// assistBytesPerWork is 1/assistWorkPerByte.
   389  	assistBytesPerWork float64
   390  
   391  	// fractionalUtilizationGoal is the fraction of wall clock
   392  	// time that should be spent in the fractional mark worker on
   393  	// each P that isn't running a dedicated worker.
   394  	//
   395  	// For example, if the utilization goal is 25% and there are
   396  	// no dedicated workers, this will be 0.25. If the goal is
   397  	// 25%, there is one dedicated worker, and GOMAXPROCS is 5,
   398  	// this will be 0.05 to make up the missing 5%.
   399  	//
   400  	// If this is zero, no fractional workers are needed.
   401  	fractionalUtilizationGoal float64
   402  
   403  	_ cpu.CacheLinePad
   404  }
   405  
   406  // startCycle resets the GC controller's state and computes estimates
   407  // for a new GC cycle. The caller must hold worldsema.
   408  func (c *gcControllerState) startCycle() {
   409  	c.scanWork = 0
   410  	c.bgScanCredit = 0
   411  	c.assistTime = 0
   412  	c.dedicatedMarkTime = 0
   413  	c.fractionalMarkTime = 0
   414  	c.idleMarkTime = 0
   415  
   416  	// Ensure that the heap goal is at least a little larger than
   417  	// the current live heap size. This may not be the case if GC
   418  	// start is delayed or if the allocation that pushed heap_live
   419  	// over gc_trigger is large or if the trigger is really close to
   420  	// GOGC. Assist is proportional to this distance, so enforce a
   421  	// minimum distance, even if it means going over the GOGC goal
   422  	// by a tiny bit.
   423  	if memstats.next_gc < memstats.heap_live+1024*1024 {
   424  		memstats.next_gc = memstats.heap_live + 1024*1024
   425  	}
   426  
   427  	// Compute the background mark utilization goal. In general,
   428  	// this may not come out exactly. We round the number of
   429  	// dedicated workers so that the utilization is closest to
   430  	// 25%. For small GOMAXPROCS, this would introduce too much
   431  	// error, so we add fractional workers in that case.
   432  	totalUtilizationGoal := float64(gomaxprocs) * gcBackgroundUtilization
   433  	c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal + 0.5)
   434  	utilError := float64(c.dedicatedMarkWorkersNeeded)/totalUtilizationGoal - 1
   435  	const maxUtilError = 0.3
   436  	if utilError < -maxUtilError || utilError > maxUtilError {
   437  		// Rounding put us more than 30% off our goal. With
   438  		// gcBackgroundUtilization of 25%, this happens for
   439  		// GOMAXPROCS<=3 or GOMAXPROCS=6. Enable fractional
   440  		// workers to compensate.
   441  		if float64(c.dedicatedMarkWorkersNeeded) > totalUtilizationGoal {
   442  			// Too many dedicated workers.
   443  			c.dedicatedMarkWorkersNeeded--
   444  		}
   445  		c.fractionalUtilizationGoal = (totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)) / float64(gomaxprocs)
   446  	} else {
   447  		c.fractionalUtilizationGoal = 0
   448  	}
   449  
   450  	// In STW mode, we just want dedicated workers.
   451  	if debug.gcstoptheworld > 0 {
   452  		c.dedicatedMarkWorkersNeeded = int64(gomaxprocs)
   453  		c.fractionalUtilizationGoal = 0
   454  	}
   455  
   456  	// Clear per-P state
   457  	for _, p := range allp {
   458  		p.gcAssistTime = 0
   459  		p.gcFractionalMarkTime = 0
   460  	}
   461  
   462  	// Compute initial values for controls that are updated
   463  	// throughout the cycle.
   464  	c.revise()
   465  
   466  	if debug.gcpacertrace > 0 {
   467  		print("pacer: assist ratio=", c.assistWorkPerByte,
   468  			" (scan ", memstats.heap_scan>>20, " MB in ",
   469  			work.initialHeapLive>>20, "->",
   470  			memstats.next_gc>>20, " MB)",
   471  			" workers=", c.dedicatedMarkWorkersNeeded,
   472  			"+", c.fractionalUtilizationGoal, "\n")
   473  	}
   474  }
   475  
   476  // revise updates the assist ratio during the GC cycle to account for
   477  // improved estimates. This should be called either under STW or
   478  // whenever memstats.heap_scan, memstats.heap_live, or
   479  // memstats.next_gc is updated (with mheap_.lock held).
   480  //
   481  // It should only be called when gcBlackenEnabled != 0 (because this
   482  // is when assists are enabled and the necessary statistics are
   483  // available).
   484  func (c *gcControllerState) revise() {
   485  	gcpercent := gcpercent
   486  	if gcpercent < 0 {
   487  		// If GC is disabled but we're running a forced GC,
   488  		// act like GOGC is huge for the below calculations.
   489  		gcpercent = 100000
   490  	}
   491  	live := atomic.Load64(&memstats.heap_live)
   492  
   493  	var heapGoal, scanWorkExpected int64
   494  	if live <= memstats.next_gc {
   495  		// We're under the soft goal. Pace GC to complete at
   496  		// next_gc assuming the heap is in steady-state.
   497  		heapGoal = int64(memstats.next_gc)
   498  
   499  		// Compute the expected scan work remaining.
   500  		//
   501  		// This is estimated based on the expected
   502  		// steady-state scannable heap. For example, with
   503  		// GOGC=100, only half of the scannable heap is
   504  		// expected to be live, so that's what we target.
   505  		//
   506  		// (This is a float calculation to avoid overflowing on
   507  		// 100*heap_scan.)
   508  		scanWorkExpected = int64(float64(memstats.heap_scan) * 100 / float64(100+gcpercent))
   509  	} else {
   510  		// We're past the soft goal. Pace GC so that in the
   511  		// worst case it will complete by the hard goal.
   512  		const maxOvershoot = 1.1
   513  		heapGoal = int64(float64(memstats.next_gc) * maxOvershoot)
   514  
   515  		// Compute the upper bound on the scan work remaining.
   516  		scanWorkExpected = int64(memstats.heap_scan)
   517  	}
   518  
   519  	// Compute the remaining scan work estimate.
   520  	//
   521  	// Note that we currently count allocations during GC as both
   522  	// scannable heap (heap_scan) and scan work completed
   523  	// (scanWork), so allocation will change this difference will
   524  	// slowly in the soft regime and not at all in the hard
   525  	// regime.
   526  	scanWorkRemaining := scanWorkExpected - c.scanWork
   527  	if scanWorkRemaining < 1000 {
   528  		// We set a somewhat arbitrary lower bound on
   529  		// remaining scan work since if we aim a little high,
   530  		// we can miss by a little.
   531  		//
   532  		// We *do* need to enforce that this is at least 1,
   533  		// since marking is racy and double-scanning objects
   534  		// may legitimately make the remaining scan work
   535  		// negative, even in the hard goal regime.
   536  		scanWorkRemaining = 1000
   537  	}
   538  
   539  	// Compute the heap distance remaining.
   540  	heapRemaining := heapGoal - int64(live)
   541  	if heapRemaining <= 0 {
   542  		// This shouldn't happen, but if it does, avoid
   543  		// dividing by zero or setting the assist negative.
   544  		heapRemaining = 1
   545  	}
   546  
   547  	// Compute the mutator assist ratio so by the time the mutator
   548  	// allocates the remaining heap bytes up to next_gc, it will
   549  	// have done (or stolen) the remaining amount of scan work.
   550  	c.assistWorkPerByte = float64(scanWorkRemaining) / float64(heapRemaining)
   551  	c.assistBytesPerWork = float64(heapRemaining) / float64(scanWorkRemaining)
   552  }
   553  
   554  // endCycle computes the trigger ratio for the next cycle.
   555  func (c *gcControllerState) endCycle() float64 {
   556  	if work.userForced {
   557  		// Forced GC means this cycle didn't start at the
   558  		// trigger, so where it finished isn't good
   559  		// information about how to adjust the trigger.
   560  		// Just leave it where it is.
   561  		return memstats.triggerRatio
   562  	}
   563  
   564  	// Proportional response gain for the trigger controller. Must
   565  	// be in [0, 1]. Lower values smooth out transient effects but
   566  	// take longer to respond to phase changes. Higher values
   567  	// react to phase changes quickly, but are more affected by
   568  	// transient changes. Values near 1 may be unstable.
   569  	const triggerGain = 0.5
   570  
   571  	// Compute next cycle trigger ratio. First, this computes the
   572  	// "error" for this cycle; that is, how far off the trigger
   573  	// was from what it should have been, accounting for both heap
   574  	// growth and GC CPU utilization. We compute the actual heap
   575  	// growth during this cycle and scale that by how far off from
   576  	// the goal CPU utilization we were (to estimate the heap
   577  	// growth if we had the desired CPU utilization). The
   578  	// difference between this estimate and the GOGC-based goal
   579  	// heap growth is the error.
   580  	goalGrowthRatio := gcEffectiveGrowthRatio()
   581  	actualGrowthRatio := float64(memstats.heap_live)/float64(memstats.heap_marked) - 1
   582  	assistDuration := nanotime() - c.markStartTime
   583  
   584  	// Assume background mark hit its utilization goal.
   585  	utilization := gcBackgroundUtilization
   586  	// Add assist utilization; avoid divide by zero.
   587  	if assistDuration > 0 {
   588  		utilization += float64(c.assistTime) / float64(assistDuration*int64(gomaxprocs))
   589  	}
   590  
   591  	triggerError := goalGrowthRatio - memstats.triggerRatio - utilization/gcGoalUtilization*(actualGrowthRatio-memstats.triggerRatio)
   592  
   593  	// Finally, we adjust the trigger for next time by this error,
   594  	// damped by the proportional gain.
   595  	triggerRatio := memstats.triggerRatio + triggerGain*triggerError
   596  
   597  	if debug.gcpacertrace > 0 {
   598  		// Print controller state in terms of the design
   599  		// document.
   600  		H_m_prev := memstats.heap_marked
   601  		h_t := memstats.triggerRatio
   602  		H_T := memstats.gc_trigger
   603  		h_a := actualGrowthRatio
   604  		H_a := memstats.heap_live
   605  		h_g := goalGrowthRatio
   606  		H_g := int64(float64(H_m_prev) * (1 + h_g))
   607  		u_a := utilization
   608  		u_g := gcGoalUtilization
   609  		W_a := c.scanWork
   610  		print("pacer: H_m_prev=", H_m_prev,
   611  			" h_t=", h_t, " H_T=", H_T,
   612  			" h_a=", h_a, " H_a=", H_a,
   613  			" h_g=", h_g, " H_g=", H_g,
   614  			" u_a=", u_a, " u_g=", u_g,
   615  			" W_a=", W_a,
   616  			" goalΔ=", goalGrowthRatio-h_t,
   617  			" actualΔ=", h_a-h_t,
   618  			" u_a/u_g=", u_a/u_g,
   619  			"\n")
   620  	}
   621  
   622  	return triggerRatio
   623  }
   624  
   625  // enlistWorker encourages another dedicated mark worker to start on
   626  // another P if there are spare worker slots. It is used by putfull
   627  // when more work is made available.
   628  //
   629  //go:nowritebarrier
   630  func (c *gcControllerState) enlistWorker() {
   631  	// If there are idle Ps, wake one so it will run an idle worker.
   632  	// NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
   633  	//
   634  	//	if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
   635  	//		wakep()
   636  	//		return
   637  	//	}
   638  
   639  	// There are no idle Ps. If we need more dedicated workers,
   640  	// try to preempt a running P so it will switch to a worker.
   641  	if c.dedicatedMarkWorkersNeeded <= 0 {
   642  		return
   643  	}
   644  	// Pick a random other P to preempt.
   645  	if gomaxprocs <= 1 {
   646  		return
   647  	}
   648  	gp := getg()
   649  	if gp == nil || gp.m == nil || gp.m.p == 0 {
   650  		return
   651  	}
   652  	myID := gp.m.p.ptr().id
   653  	for tries := 0; tries < 5; tries++ {
   654  		id := int32(fastrandn(uint32(gomaxprocs - 1)))
   655  		if id >= myID {
   656  			id++
   657  		}
   658  		p := allp[id]
   659  		if p.status != _Prunning {
   660  			continue
   661  		}
   662  		if preemptone(p) {
   663  			return
   664  		}
   665  	}
   666  }
   667  
   668  // findRunnableGCWorker returns the background mark worker for _p_ if it
   669  // should be run. This must only be called when gcBlackenEnabled != 0.
   670  func (c *gcControllerState) findRunnableGCWorker(_p_ *p) *g {
   671  	if gcBlackenEnabled == 0 {
   672  		throw("gcControllerState.findRunnable: blackening not enabled")
   673  	}
   674  	if _p_.gcBgMarkWorker == 0 {
   675  		// The mark worker associated with this P is blocked
   676  		// performing a mark transition. We can't run it
   677  		// because it may be on some other run or wait queue.
   678  		return nil
   679  	}
   680  
   681  	if !gcMarkWorkAvailable(_p_) {
   682  		// No work to be done right now. This can happen at
   683  		// the end of the mark phase when there are still
   684  		// assists tapering off. Don't bother running a worker
   685  		// now because it'll just return immediately.
   686  		return nil
   687  	}
   688  
   689  	decIfPositive := func(ptr *int64) bool {
   690  		if *ptr > 0 {
   691  			if atomic.Xaddint64(ptr, -1) >= 0 {
   692  				return true
   693  			}
   694  			// We lost a race
   695  			atomic.Xaddint64(ptr, +1)
   696  		}
   697  		return false
   698  	}
   699  
   700  	if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
   701  		// This P is now dedicated to marking until the end of
   702  		// the concurrent mark phase.
   703  		_p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
   704  	} else if c.fractionalUtilizationGoal == 0 {
   705  		// No need for fractional workers.
   706  		return nil
   707  	} else {
   708  		// Is this P behind on the fractional utilization
   709  		// goal?
   710  		//
   711  		// This should be kept in sync with pollFractionalWorkerExit.
   712  		delta := nanotime() - gcController.markStartTime
   713  		if delta > 0 && float64(_p_.gcFractionalMarkTime)/float64(delta) > c.fractionalUtilizationGoal {
   714  			// Nope. No need to run a fractional worker.
   715  			return nil
   716  		}
   717  		// Run a fractional worker.
   718  		_p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode
   719  	}
   720  
   721  	// Run the background mark worker
   722  	gp := _p_.gcBgMarkWorker.ptr()
   723  	casgstatus(gp, _Gwaiting, _Grunnable)
   724  	if trace.enabled {
   725  		traceGoUnpark(gp, 0)
   726  	}
   727  	return gp
   728  }
   729  
   730  // pollFractionalWorkerExit reports whether a fractional mark worker
   731  // should self-preempt. It assumes it is called from the fractional
   732  // worker.
   733  func pollFractionalWorkerExit() bool {
   734  	// This should be kept in sync with the fractional worker
   735  	// scheduler logic in findRunnableGCWorker.
   736  	now := nanotime()
   737  	delta := now - gcController.markStartTime
   738  	if delta <= 0 {
   739  		return true
   740  	}
   741  	p := getg().m.p.ptr()
   742  	selfTime := p.gcFractionalMarkTime + (now - p.gcMarkWorkerStartTime)
   743  	// Add some slack to the utilization goal so that the
   744  	// fractional worker isn't behind again the instant it exits.
   745  	return float64(selfTime)/float64(delta) > 1.2*gcController.fractionalUtilizationGoal
   746  }
   747  
   748  // gcSetTriggerRatio sets the trigger ratio and updates everything
   749  // derived from it: the absolute trigger, the heap goal, mark pacing,
   750  // and sweep pacing.
   751  //
   752  // This can be called any time. If GC is the in the middle of a
   753  // concurrent phase, it will adjust the pacing of that phase.
   754  //
   755  // This depends on gcpercent, memstats.heap_marked, and
   756  // memstats.heap_live. These must be up to date.
   757  //
   758  // mheap_.lock must be held or the world must be stopped.
   759  func gcSetTriggerRatio(triggerRatio float64) {
   760  	// Compute the next GC goal, which is when the allocated heap
   761  	// has grown by GOGC/100 over the heap marked by the last
   762  	// cycle.
   763  	goal := ^uint64(0)
   764  	if gcpercent >= 0 {
   765  		goal = memstats.heap_marked + memstats.heap_marked*uint64(gcpercent)/100
   766  	}
   767  
   768  	// Set the trigger ratio, capped to reasonable bounds.
   769  	if triggerRatio < 0 {
   770  		// This can happen if the mutator is allocating very
   771  		// quickly or the GC is scanning very slowly.
   772  		triggerRatio = 0
   773  	} else if gcpercent >= 0 {
   774  		// Ensure there's always a little margin so that the
   775  		// mutator assist ratio isn't infinity.
   776  		maxTriggerRatio := 0.95 * float64(gcpercent) / 100
   777  		if triggerRatio > maxTriggerRatio {
   778  			triggerRatio = maxTriggerRatio
   779  		}
   780  	}
   781  	memstats.triggerRatio = triggerRatio
   782  
   783  	// Compute the absolute GC trigger from the trigger ratio.
   784  	//
   785  	// We trigger the next GC cycle when the allocated heap has
   786  	// grown by the trigger ratio over the marked heap size.
   787  	trigger := ^uint64(0)
   788  	if gcpercent >= 0 {
   789  		trigger = uint64(float64(memstats.heap_marked) * (1 + triggerRatio))
   790  		// Don't trigger below the minimum heap size.
   791  		minTrigger := heapminimum
   792  		if !isSweepDone() {
   793  			// Concurrent sweep happens in the heap growth
   794  			// from heap_live to gc_trigger, so ensure
   795  			// that concurrent sweep has some heap growth
   796  			// in which to perform sweeping before we
   797  			// start the next GC cycle.
   798  			sweepMin := atomic.Load64(&memstats.heap_live) + sweepMinHeapDistance
   799  			if sweepMin > minTrigger {
   800  				minTrigger = sweepMin
   801  			}
   802  		}
   803  		if trigger < minTrigger {
   804  			trigger = minTrigger
   805  		}
   806  		if int64(trigger) < 0 {
   807  			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")
   808  			throw("gc_trigger underflow")
   809  		}
   810  		if trigger > goal {
   811  			// The trigger ratio is always less than GOGC/100, but
   812  			// other bounds on the trigger may have raised it.
   813  			// Push up the goal, too.
   814  			goal = trigger
   815  		}
   816  	}
   817  
   818  	// Commit to the trigger and goal.
   819  	memstats.gc_trigger = trigger
   820  	memstats.next_gc = goal
   821  	if trace.enabled {
   822  		traceNextGC()
   823  	}
   824  
   825  	// Update mark pacing.
   826  	if gcphase != _GCoff {
   827  		gcController.revise()
   828  	}
   829  
   830  	// Update sweep pacing.
   831  	if isSweepDone() {
   832  		mheap_.sweepPagesPerByte = 0
   833  	} else {
   834  		// Concurrent sweep needs to sweep all of the in-use
   835  		// pages by the time the allocated heap reaches the GC
   836  		// trigger. Compute the ratio of in-use pages to sweep
   837  		// per byte allocated, accounting for the fact that
   838  		// some might already be swept.
   839  		heapLiveBasis := atomic.Load64(&memstats.heap_live)
   840  		heapDistance := int64(trigger) - int64(heapLiveBasis)
   841  		// Add a little margin so rounding errors and
   842  		// concurrent sweep are less likely to leave pages
   843  		// unswept when GC starts.
   844  		heapDistance -= 1024 * 1024
   845  		if heapDistance < _PageSize {
   846  			// Avoid setting the sweep ratio extremely high
   847  			heapDistance = _PageSize
   848  		}
   849  		pagesSwept := atomic.Load64(&mheap_.pagesSwept)
   850  		sweepDistancePages := int64(mheap_.pagesInUse) - int64(pagesSwept)
   851  		if sweepDistancePages <= 0 {
   852  			mheap_.sweepPagesPerByte = 0
   853  		} else {
   854  			mheap_.sweepPagesPerByte = float64(sweepDistancePages) / float64(heapDistance)
   855  			mheap_.sweepHeapLiveBasis = heapLiveBasis
   856  			// Write pagesSweptBasis last, since this
   857  			// signals concurrent sweeps to recompute
   858  			// their debt.
   859  			atomic.Store64(&mheap_.pagesSweptBasis, pagesSwept)
   860  		}
   861  	}
   862  
   863  	gcPaceScavenger()
   864  }
   865  
   866  // gcEffectiveGrowthRatio returns the current effective heap growth
   867  // ratio (GOGC/100) based on heap_marked from the previous GC and
   868  // next_gc for the current GC.
   869  //
   870  // This may differ from gcpercent/100 because of various upper and
   871  // lower bounds on gcpercent. For example, if the heap is smaller than
   872  // heapminimum, this can be higher than gcpercent/100.
   873  //
   874  // mheap_.lock must be held or the world must be stopped.
   875  func gcEffectiveGrowthRatio() float64 {
   876  	egogc := float64(memstats.next_gc-memstats.heap_marked) / float64(memstats.heap_marked)
   877  	if egogc < 0 {
   878  		// Shouldn't happen, but just in case.
   879  		egogc = 0
   880  	}
   881  	return egogc
   882  }
   883  
   884  // gcGoalUtilization is the goal CPU utilization for
   885  // marking as a fraction of GOMAXPROCS.
   886  const gcGoalUtilization = 0.30
   887  
   888  // gcBackgroundUtilization is the fixed CPU utilization for background
   889  // marking. It must be <= gcGoalUtilization. The difference between
   890  // gcGoalUtilization and gcBackgroundUtilization will be made up by
   891  // mark assists. The scheduler will aim to use within 50% of this
   892  // goal.
   893  //
   894  // Setting this to < gcGoalUtilization avoids saturating the trigger
   895  // feedback controller when there are no assists, which allows it to
   896  // better control CPU and heap growth. However, the larger the gap,
   897  // the more mutator assists are expected to happen, which impact
   898  // mutator latency.
   899  const gcBackgroundUtilization = 0.25
   900  
   901  // gcCreditSlack is the amount of scan work credit that can
   902  // accumulate locally before updating gcController.scanWork and,
   903  // optionally, gcController.bgScanCredit. Lower values give a more
   904  // accurate assist ratio and make it more likely that assists will
   905  // successfully steal background credit. Higher values reduce memory
   906  // contention.
   907  const gcCreditSlack = 2000
   908  
   909  // gcAssistTimeSlack is the nanoseconds of mutator assist time that
   910  // can accumulate on a P before updating gcController.assistTime.
   911  const gcAssistTimeSlack = 5000
   912  
   913  // gcOverAssistWork determines how many extra units of scan work a GC
   914  // assist does when an assist happens. This amortizes the cost of an
   915  // assist by pre-paying for this many bytes of future allocations.
   916  const gcOverAssistWork = 64 << 10
   917  
   918  var work struct {
   919  	full  lfstack          // lock-free list of full blocks workbuf
   920  	empty lfstack          // lock-free list of empty blocks workbuf
   921  	pad0  cpu.CacheLinePad // prevents false-sharing between full/empty and nproc/nwait
   922  
   923  	wbufSpans struct {
   924  		lock mutex
   925  		// free is a list of spans dedicated to workbufs, but
   926  		// that don't currently contain any workbufs.
   927  		free mSpanList
   928  		// busy is a list of all spans containing workbufs on
   929  		// one of the workbuf lists.
   930  		busy mSpanList
   931  	}
   932  
   933  	// Restore 64-bit alignment on 32-bit.
   934  	_ uint32
   935  
   936  	// bytesMarked is the number of bytes marked this cycle. This
   937  	// includes bytes blackened in scanned objects, noscan objects
   938  	// that go straight to black, and permagrey objects scanned by
   939  	// markroot during the concurrent scan phase. This is updated
   940  	// atomically during the cycle. Updates may be batched
   941  	// arbitrarily, since the value is only read at the end of the
   942  	// cycle.
   943  	//
   944  	// Because of benign races during marking, this number may not
   945  	// be the exact number of marked bytes, but it should be very
   946  	// close.
   947  	//
   948  	// Put this field here because it needs 64-bit atomic access
   949  	// (and thus 8-byte alignment even on 32-bit architectures).
   950  	bytesMarked uint64
   951  
   952  	markrootNext uint32 // next markroot job
   953  	markrootJobs uint32 // number of markroot jobs
   954  
   955  	nproc  uint32
   956  	tstart int64
   957  	nwait  uint32
   958  	ndone  uint32
   959  
   960  	// Number of roots of various root types. Set by gcMarkRootPrepare.
   961  	nFlushCacheRoots                               int
   962  	nDataRoots, nBSSRoots, nSpanRoots, nStackRoots int
   963  
   964  	// Each type of GC state transition is protected by a lock.
   965  	// Since multiple threads can simultaneously detect the state
   966  	// transition condition, any thread that detects a transition
   967  	// condition must acquire the appropriate transition lock,
   968  	// re-check the transition condition and return if it no
   969  	// longer holds or perform the transition if it does.
   970  	// Likewise, any transition must invalidate the transition
   971  	// condition before releasing the lock. This ensures that each
   972  	// transition is performed by exactly one thread and threads
   973  	// that need the transition to happen block until it has
   974  	// happened.
   975  	//
   976  	// startSema protects the transition from "off" to mark or
   977  	// mark termination.
   978  	startSema uint32
   979  	// markDoneSema protects transitions from mark to mark termination.
   980  	markDoneSema uint32
   981  
   982  	bgMarkReady note   // signal background mark worker has started
   983  	bgMarkDone  uint32 // cas to 1 when at a background mark completion point
   984  	// Background mark completion signaling
   985  
   986  	// mode is the concurrency mode of the current GC cycle.
   987  	mode gcMode
   988  
   989  	// userForced indicates the current GC cycle was forced by an
   990  	// explicit user call.
   991  	userForced bool
   992  
   993  	// totaltime is the CPU nanoseconds spent in GC since the
   994  	// program started if debug.gctrace > 0.
   995  	totaltime int64
   996  
   997  	// initialHeapLive is the value of memstats.heap_live at the
   998  	// beginning of this GC cycle.
   999  	initialHeapLive uint64
  1000  
  1001  	// assistQueue is a queue of assists that are blocked because
  1002  	// there was neither enough credit to steal or enough work to
  1003  	// do.
  1004  	assistQueue struct {
  1005  		lock mutex
  1006  		q    gQueue
  1007  	}
  1008  
  1009  	// sweepWaiters is a list of blocked goroutines to wake when
  1010  	// we transition from mark termination to sweep.
  1011  	sweepWaiters struct {
  1012  		lock mutex
  1013  		list gList
  1014  	}
  1015  
  1016  	// cycles is the number of completed GC cycles, where a GC
  1017  	// cycle is sweep termination, mark, mark termination, and
  1018  	// sweep. This differs from memstats.numgc, which is
  1019  	// incremented at mark termination.
  1020  	cycles uint32
  1021  
  1022  	// Timing/utilization stats for this cycle.
  1023  	stwprocs, maxprocs                 int32
  1024  	tSweepTerm, tMark, tMarkTerm, tEnd int64 // nanotime() of phase start
  1025  
  1026  	pauseNS    int64 // total STW time this cycle
  1027  	pauseStart int64 // nanotime() of last STW
  1028  
  1029  	// debug.gctrace heap sizes for this cycle.
  1030  	heap0, heap1, heap2, heapGoal uint64
  1031  }
  1032  
  1033  // GC runs a garbage collection and blocks the caller until the
  1034  // garbage collection is complete. It may also block the entire
  1035  // program.
  1036  func GC() {
  1037  	// We consider a cycle to be: sweep termination, mark, mark
  1038  	// termination, and sweep. This function shouldn't return
  1039  	// until a full cycle has been completed, from beginning to
  1040  	// end. Hence, we always want to finish up the current cycle
  1041  	// and start a new one. That means:
  1042  	//
  1043  	// 1. In sweep termination, mark, or mark termination of cycle
  1044  	// N, wait until mark termination N completes and transitions
  1045  	// to sweep N.
  1046  	//
  1047  	// 2. In sweep N, help with sweep N.
  1048  	//
  1049  	// At this point we can begin a full cycle N+1.
  1050  	//
  1051  	// 3. Trigger cycle N+1 by starting sweep termination N+1.
  1052  	//
  1053  	// 4. Wait for mark termination N+1 to complete.
  1054  	//
  1055  	// 5. Help with sweep N+1 until it's done.
  1056  	//
  1057  	// This all has to be written to deal with the fact that the
  1058  	// GC may move ahead on its own. For example, when we block
  1059  	// until mark termination N, we may wake up in cycle N+2.
  1060  
  1061  	// Wait until the current sweep termination, mark, and mark
  1062  	// termination complete.
  1063  	n := atomic.Load(&work.cycles)
  1064  	gcWaitOnMark(n)
  1065  
  1066  	// We're now in sweep N or later. Trigger GC cycle N+1, which
  1067  	// will first finish sweep N if necessary and then enter sweep
  1068  	// termination N+1.
  1069  	gcStart(gcTrigger{kind: gcTriggerCycle, n: n + 1})
  1070  
  1071  	// Wait for mark termination N+1 to complete.
  1072  	gcWaitOnMark(n + 1)
  1073  
  1074  	// Finish sweep N+1 before returning. We do this both to
  1075  	// complete the cycle and because runtime.GC() is often used
  1076  	// as part of tests and benchmarks to get the system into a
  1077  	// relatively stable and isolated state.
  1078  	for atomic.Load(&work.cycles) == n+1 && sweepone() != ^uintptr(0) {
  1079  		sweep.nbgsweep++
  1080  		Gosched()
  1081  	}
  1082  
  1083  	// Callers may assume that the heap profile reflects the
  1084  	// just-completed cycle when this returns (historically this
  1085  	// happened because this was a STW GC), but right now the
  1086  	// profile still reflects mark termination N, not N+1.
  1087  	//
  1088  	// As soon as all of the sweep frees from cycle N+1 are done,
  1089  	// we can go ahead and publish the heap profile.
  1090  	//
  1091  	// First, wait for sweeping to finish. (We know there are no
  1092  	// more spans on the sweep queue, but we may be concurrently
  1093  	// sweeping spans, so we have to wait.)
  1094  	for atomic.Load(&work.cycles) == n+1 && atomic.Load(&mheap_.sweepers) != 0 {
  1095  		Gosched()
  1096  	}
  1097  
  1098  	// Now we're really done with sweeping, so we can publish the
  1099  	// stable heap profile. Only do this if we haven't already hit
  1100  	// another mark termination.
  1101  	mp := acquirem()
  1102  	cycle := atomic.Load(&work.cycles)
  1103  	if cycle == n+1 || (gcphase == _GCmark && cycle == n+2) {
  1104  		mProf_PostSweep()
  1105  	}
  1106  	releasem(mp)
  1107  }
  1108  
  1109  // gcWaitOnMark blocks until GC finishes the Nth mark phase. If GC has
  1110  // already completed this mark phase, it returns immediately.
  1111  func gcWaitOnMark(n uint32) {
  1112  	for {
  1113  		// Disable phase transitions.
  1114  		lock(&work.sweepWaiters.lock)
  1115  		nMarks := atomic.Load(&work.cycles)
  1116  		if gcphase != _GCmark {
  1117  			// We've already completed this cycle's mark.
  1118  			nMarks++
  1119  		}
  1120  		if nMarks > n {
  1121  			// We're done.
  1122  			unlock(&work.sweepWaiters.lock)
  1123  			return
  1124  		}
  1125  
  1126  		// Wait until sweep termination, mark, and mark
  1127  		// termination of cycle N complete.
  1128  		work.sweepWaiters.list.push(getg())
  1129  		goparkunlock(&work.sweepWaiters.lock, waitReasonWaitForGCCycle, traceEvGoBlock, 1)
  1130  	}
  1131  }
  1132  
  1133  // gcMode indicates how concurrent a GC cycle should be.
  1134  type gcMode int
  1135  
  1136  const (
  1137  	gcBackgroundMode gcMode = iota // concurrent GC and sweep
  1138  	gcForceMode                    // stop-the-world GC now, concurrent sweep
  1139  	gcForceBlockMode               // stop-the-world GC now and STW sweep (forced by user)
  1140  )
  1141  
  1142  // A gcTrigger is a predicate for starting a GC cycle. Specifically,
  1143  // it is an exit condition for the _GCoff phase.
  1144  type gcTrigger struct {
  1145  	kind gcTriggerKind
  1146  	now  int64  // gcTriggerTime: current time
  1147  	n    uint32 // gcTriggerCycle: cycle number to start
  1148  }
  1149  
  1150  type gcTriggerKind int
  1151  
  1152  const (
  1153  	// gcTriggerHeap indicates that a cycle should be started when
  1154  	// the heap size reaches the trigger heap size computed by the
  1155  	// controller.
  1156  	gcTriggerHeap gcTriggerKind = iota
  1157  
  1158  	// gcTriggerTime indicates that a cycle should be started when
  1159  	// it's been more than forcegcperiod nanoseconds since the
  1160  	// previous GC cycle.
  1161  	gcTriggerTime
  1162  
  1163  	// gcTriggerCycle indicates that a cycle should be started if
  1164  	// we have not yet started cycle number gcTrigger.n (relative
  1165  	// to work.cycles).
  1166  	gcTriggerCycle
  1167  )
  1168  
  1169  // test reports whether the trigger condition is satisfied, meaning
  1170  // that the exit condition for the _GCoff phase has been met. The exit
  1171  // condition should be tested when allocating.
  1172  func (t gcTrigger) test() bool {
  1173  	if !memstats.enablegc || panicking != 0 || gcphase != _GCoff {
  1174  		return false
  1175  	}
  1176  	switch t.kind {
  1177  	case gcTriggerHeap:
  1178  		// Non-atomic access to heap_live for performance. If
  1179  		// we are going to trigger on this, this thread just
  1180  		// atomically wrote heap_live anyway and we'll see our
  1181  		// own write.
  1182  		return memstats.heap_live >= memstats.gc_trigger
  1183  	case gcTriggerTime:
  1184  		if gcpercent < 0 {
  1185  			return false
  1186  		}
  1187  		lastgc := int64(atomic.Load64(&memstats.last_gc_nanotime))
  1188  		return lastgc != 0 && t.now-lastgc > forcegcperiod
  1189  	case gcTriggerCycle:
  1190  		// t.n > work.cycles, but accounting for wraparound.
  1191  		return int32(t.n-work.cycles) > 0
  1192  	}
  1193  	return true
  1194  }
  1195  
  1196  // gcStart starts the GC. It transitions from _GCoff to _GCmark (if
  1197  // debug.gcstoptheworld == 0) or performs all of GC (if
  1198  // debug.gcstoptheworld != 0).
  1199  //
  1200  // This may return without performing this transition in some cases,
  1201  // such as when called on a system stack or with locks held.
  1202  func gcStart(trigger gcTrigger) {
  1203  	// Since this is called from malloc and malloc is called in
  1204  	// the guts of a number of libraries that might be holding
  1205  	// locks, don't attempt to start GC in non-preemptible or
  1206  	// potentially unstable situations.
  1207  	mp := acquirem()
  1208  	if gp := getg(); gp == mp.g0 || mp.locks > 1 || mp.preemptoff != "" {
  1209  		releasem(mp)
  1210  		return
  1211  	}
  1212  	releasem(mp)
  1213  	mp = nil
  1214  
  1215  	// Pick up the remaining unswept/not being swept spans concurrently
  1216  	//
  1217  	// This shouldn't happen if we're being invoked in background
  1218  	// mode since proportional sweep should have just finished
  1219  	// sweeping everything, but rounding errors, etc, may leave a
  1220  	// few spans unswept. In forced mode, this is necessary since
  1221  	// GC can be forced at any point in the sweeping cycle.
  1222  	//
  1223  	// We check the transition condition continuously here in case
  1224  	// this G gets delayed in to the next GC cycle.
  1225  	for trigger.test() && sweepone() != ^uintptr(0) {
  1226  		sweep.nbgsweep++
  1227  	}
  1228  
  1229  	// Perform GC initialization and the sweep termination
  1230  	// transition.
  1231  	semacquire(&work.startSema)
  1232  	// Re-check transition condition under transition lock.
  1233  	if !trigger.test() {
  1234  		semrelease(&work.startSema)
  1235  		return
  1236  	}
  1237  
  1238  	// For stats, check if this GC was forced by the user.
  1239  	work.userForced = trigger.kind == gcTriggerCycle
  1240  
  1241  	// In gcstoptheworld debug mode, upgrade the mode accordingly.
  1242  	// We do this after re-checking the transition condition so
  1243  	// that multiple goroutines that detect the heap trigger don't
  1244  	// start multiple STW GCs.
  1245  	mode := gcBackgroundMode
  1246  	if debug.gcstoptheworld == 1 {
  1247  		mode = gcForceMode
  1248  	} else if debug.gcstoptheworld == 2 {
  1249  		mode = gcForceBlockMode
  1250  	}
  1251  
  1252  	// Ok, we're doing it! Stop everybody else
  1253  	semacquire(&worldsema)
  1254  
  1255  	if trace.enabled {
  1256  		traceGCStart()
  1257  	}
  1258  
  1259  	// Check that all Ps have finished deferred mcache flushes.
  1260  	for _, p := range allp {
  1261  		if fg := atomic.Load(&p.mcache.flushGen); fg != mheap_.sweepgen {
  1262  			println("runtime: p", p.id, "flushGen", fg, "!= sweepgen", mheap_.sweepgen)
  1263  			throw("p mcache not flushed")
  1264  		}
  1265  	}
  1266  
  1267  	gcBgMarkStartWorkers()
  1268  
  1269  	systemstack(gcResetMarkState)
  1270  
  1271  	work.stwprocs, work.maxprocs = gomaxprocs, gomaxprocs
  1272  	if work.stwprocs > ncpu {
  1273  		// This is used to compute CPU time of the STW phases,
  1274  		// so it can't be more than ncpu, even if GOMAXPROCS is.
  1275  		work.stwprocs = ncpu
  1276  	}
  1277  	work.heap0 = atomic.Load64(&memstats.heap_live)
  1278  	work.pauseNS = 0
  1279  	work.mode = mode
  1280  
  1281  	now := nanotime()
  1282  	work.tSweepTerm = now
  1283  	work.pauseStart = now
  1284  	if trace.enabled {
  1285  		traceGCSTWStart(1)
  1286  	}
  1287  	systemstack(stopTheWorldWithSema)
  1288  	// Finish sweep before we start concurrent scan.
  1289  	systemstack(func() {
  1290  		finishsweep_m()
  1291  	})
  1292  	// clearpools before we start the GC. If we wait they memory will not be
  1293  	// reclaimed until the next GC cycle.
  1294  	clearpools()
  1295  
  1296  	work.cycles++
  1297  
  1298  	gcController.startCycle()
  1299  	work.heapGoal = memstats.next_gc
  1300  
  1301  	// In STW mode, disable scheduling of user Gs. This may also
  1302  	// disable scheduling of this goroutine, so it may block as
  1303  	// soon as we start the world again.
  1304  	if mode != gcBackgroundMode {
  1305  		schedEnableUser(false)
  1306  	}
  1307  
  1308  	// Enter concurrent mark phase and enable
  1309  	// write barriers.
  1310  	//
  1311  	// Because the world is stopped, all Ps will
  1312  	// observe that write barriers are enabled by
  1313  	// the time we start the world and begin
  1314  	// scanning.
  1315  	//
  1316  	// Write barriers must be enabled before assists are
  1317  	// enabled because they must be enabled before
  1318  	// any non-leaf heap objects are marked. Since
  1319  	// allocations are blocked until assists can
  1320  	// happen, we want enable assists as early as
  1321  	// possible.
  1322  	setGCPhase(_GCmark)
  1323  
  1324  	gcBgMarkPrepare() // Must happen before assist enable.
  1325  	gcMarkRootPrepare()
  1326  
  1327  	// Mark all active tinyalloc blocks. Since we're
  1328  	// allocating from these, they need to be black like
  1329  	// other allocations. The alternative is to blacken
  1330  	// the tiny block on every allocation from it, which
  1331  	// would slow down the tiny allocator.
  1332  	gcMarkTinyAllocs()
  1333  
  1334  	// At this point all Ps have enabled the write
  1335  	// barrier, thus maintaining the no white to
  1336  	// black invariant. Enable mutator assists to
  1337  	// put back-pressure on fast allocating
  1338  	// mutators.
  1339  	atomic.Store(&gcBlackenEnabled, 1)
  1340  
  1341  	// Assists and workers can start the moment we start
  1342  	// the world.
  1343  	gcController.markStartTime = now
  1344  
  1345  	// Concurrent mark.
  1346  	systemstack(func() {
  1347  		now = startTheWorldWithSema(trace.enabled)
  1348  		work.pauseNS += now - work.pauseStart
  1349  		work.tMark = now
  1350  	})
  1351  	// In STW mode, we could block the instant systemstack
  1352  	// returns, so don't do anything important here. Make sure we
  1353  	// block rather than returning to user code.
  1354  	if mode != gcBackgroundMode {
  1355  		Gosched()
  1356  	}
  1357  
  1358  	semrelease(&work.startSema)
  1359  }
  1360  
  1361  // gcMarkDoneFlushed counts the number of P's with flushed work.
  1362  //
  1363  // Ideally this would be a captured local in gcMarkDone, but forEachP
  1364  // escapes its callback closure, so it can't capture anything.
  1365  //
  1366  // This is protected by markDoneSema.
  1367  var gcMarkDoneFlushed uint32
  1368  
  1369  // debugCachedWork enables extra checks for debugging premature mark
  1370  // termination.
  1371  //
  1372  // For debugging issue #27993.
  1373  const debugCachedWork = false
  1374  
  1375  // gcWorkPauseGen is for debugging the mark completion algorithm.
  1376  // gcWork put operations spin while gcWork.pauseGen == gcWorkPauseGen.
  1377  // Only used if debugCachedWork is true.
  1378  //
  1379  // For debugging issue #27993.
  1380  var gcWorkPauseGen uint32 = 1
  1381  
  1382  // gcMarkDone transitions the GC from mark to mark termination if all
  1383  // reachable objects have been marked (that is, there are no grey
  1384  // objects and can be no more in the future). Otherwise, it flushes
  1385  // all local work to the global queues where it can be discovered by
  1386  // other workers.
  1387  //
  1388  // This should be called when all local mark work has been drained and
  1389  // there are no remaining workers. Specifically, when
  1390  //
  1391  //   work.nwait == work.nproc && !gcMarkWorkAvailable(p)
  1392  //
  1393  // The calling context must be preemptible.
  1394  //
  1395  // Flushing local work is important because idle Ps may have local
  1396  // work queued. This is the only way to make that work visible and
  1397  // drive GC to completion.
  1398  //
  1399  // It is explicitly okay to have write barriers in this function. If
  1400  // it does transition to mark termination, then all reachable objects
  1401  // have been marked, so the write barrier cannot shade any more
  1402  // objects.
  1403  func gcMarkDone() {
  1404  	// Ensure only one thread is running the ragged barrier at a
  1405  	// time.
  1406  	semacquire(&work.markDoneSema)
  1407  
  1408  top:
  1409  	// Re-check transition condition under transition lock.
  1410  	//
  1411  	// It's critical that this checks the global work queues are
  1412  	// empty before performing the ragged barrier. Otherwise,
  1413  	// there could be global work that a P could take after the P
  1414  	// has passed the ragged barrier.
  1415  	if !(gcphase == _GCmark && work.nwait == work.nproc && !gcMarkWorkAvailable(nil)) {
  1416  		semrelease(&work.markDoneSema)
  1417  		return
  1418  	}
  1419  
  1420  	// Flush all local buffers and collect flushedWork flags.
  1421  	gcMarkDoneFlushed = 0
  1422  	systemstack(func() {
  1423  		gp := getg().m.curg
  1424  		// Mark the user stack as preemptible so that it may be scanned.
  1425  		// Otherwise, our attempt to force all P's to a safepoint could
  1426  		// result in a deadlock as we attempt to preempt a worker that's
  1427  		// trying to preempt us (e.g. for a stack scan).
  1428  		casgstatus(gp, _Grunning, _Gwaiting)
  1429  		forEachP(func(_p_ *p) {
  1430  			// Flush the write barrier buffer, since this may add
  1431  			// work to the gcWork.
  1432  			wbBufFlush1(_p_)
  1433  			// For debugging, shrink the write barrier
  1434  			// buffer so it flushes immediately.
  1435  			// wbBuf.reset will keep it at this size as
  1436  			// long as throwOnGCWork is set.
  1437  			if debugCachedWork {
  1438  				b := &_p_.wbBuf
  1439  				b.end = uintptr(unsafe.Pointer(&b.buf[wbBufEntryPointers]))
  1440  				b.debugGen = gcWorkPauseGen
  1441  			}
  1442  			// Flush the gcWork, since this may create global work
  1443  			// and set the flushedWork flag.
  1444  			//
  1445  			// TODO(austin): Break up these workbufs to
  1446  			// better distribute work.
  1447  			_p_.gcw.dispose()
  1448  			// Collect the flushedWork flag.
  1449  			if _p_.gcw.flushedWork {
  1450  				atomic.Xadd(&gcMarkDoneFlushed, 1)
  1451  				_p_.gcw.flushedWork = false
  1452  			} else if debugCachedWork {
  1453  				// For debugging, freeze the gcWork
  1454  				// until we know whether we've reached
  1455  				// completion or not. If we think
  1456  				// we've reached completion, but
  1457  				// there's a paused gcWork, then
  1458  				// that's a bug.
  1459  				_p_.gcw.pauseGen = gcWorkPauseGen
  1460  				// Capture the G's stack.
  1461  				for i := range _p_.gcw.pauseStack {
  1462  					_p_.gcw.pauseStack[i] = 0
  1463  				}
  1464  				callers(1, _p_.gcw.pauseStack[:])
  1465  			}
  1466  		})
  1467  		casgstatus(gp, _Gwaiting, _Grunning)
  1468  	})
  1469  
  1470  	if gcMarkDoneFlushed != 0 {
  1471  		if debugCachedWork {
  1472  			// Release paused gcWorks.
  1473  			atomic.Xadd(&gcWorkPauseGen, 1)
  1474  		}
  1475  		// More grey objects were discovered since the
  1476  		// previous termination check, so there may be more
  1477  		// work to do. Keep going. It's possible the
  1478  		// transition condition became true again during the
  1479  		// ragged barrier, so re-check it.
  1480  		goto top
  1481  	}
  1482  
  1483  	if debugCachedWork {
  1484  		throwOnGCWork = true
  1485  		// Release paused gcWorks. If there are any, they
  1486  		// should now observe throwOnGCWork and panic.
  1487  		atomic.Xadd(&gcWorkPauseGen, 1)
  1488  	}
  1489  
  1490  	// There was no global work, no local work, and no Ps
  1491  	// communicated work since we took markDoneSema. Therefore
  1492  	// there are no grey objects and no more objects can be
  1493  	// shaded. Transition to mark termination.
  1494  	now := nanotime()
  1495  	work.tMarkTerm = now
  1496  	work.pauseStart = now
  1497  	getg().m.preemptoff = "gcing"
  1498  	if trace.enabled {
  1499  		traceGCSTWStart(0)
  1500  	}
  1501  	systemstack(stopTheWorldWithSema)
  1502  	// The gcphase is _GCmark, it will transition to _GCmarktermination
  1503  	// below. The important thing is that the wb remains active until
  1504  	// all marking is complete. This includes writes made by the GC.
  1505  
  1506  	if debugCachedWork {
  1507  		// For debugging, double check that no work was added after we
  1508  		// went around above and disable write barrier buffering.
  1509  		for _, p := range allp {
  1510  			gcw := &p.gcw
  1511  			if !gcw.empty() {
  1512  				printlock()
  1513  				print("runtime: P ", p.id, " flushedWork ", gcw.flushedWork)
  1514  				if gcw.wbuf1 == nil {
  1515  					print(" wbuf1=<nil>")
  1516  				} else {
  1517  					print(" wbuf1.n=", gcw.wbuf1.nobj)
  1518  				}
  1519  				if gcw.wbuf2 == nil {
  1520  					print(" wbuf2=<nil>")
  1521  				} else {
  1522  					print(" wbuf2.n=", gcw.wbuf2.nobj)
  1523  				}
  1524  				print("\n")
  1525  				if gcw.pauseGen == gcw.putGen {
  1526  					println("runtime: checkPut already failed at this generation")
  1527  				}
  1528  				throw("throwOnGCWork")
  1529  			}
  1530  		}
  1531  	} else {
  1532  		// For unknown reasons (see issue #27993), there is
  1533  		// sometimes work left over when we enter mark
  1534  		// termination. Detect this and resume concurrent
  1535  		// mark. This is obviously unfortunate.
  1536  		//
  1537  		// Switch to the system stack to call wbBufFlush1,
  1538  		// though in this case it doesn't matter because we're
  1539  		// non-preemptible anyway.
  1540  		restart := false
  1541  		systemstack(func() {
  1542  			for _, p := range allp {
  1543  				wbBufFlush1(p)
  1544  				if !p.gcw.empty() {
  1545  					restart = true
  1546  					break
  1547  				}
  1548  			}
  1549  		})
  1550  		if restart {
  1551  			getg().m.preemptoff = ""
  1552  			systemstack(func() {
  1553  				now := startTheWorldWithSema(true)
  1554  				work.pauseNS += now - work.pauseStart
  1555  			})
  1556  			goto top
  1557  		}
  1558  	}
  1559  
  1560  	// Disable assists and background workers. We must do
  1561  	// this before waking blocked assists.
  1562  	atomic.Store(&gcBlackenEnabled, 0)
  1563  
  1564  	// Wake all blocked assists. These will run when we
  1565  	// start the world again.
  1566  	gcWakeAllAssists()
  1567  
  1568  	// Likewise, release the transition lock. Blocked
  1569  	// workers and assists will run when we start the
  1570  	// world again.
  1571  	semrelease(&work.markDoneSema)
  1572  
  1573  	// In STW mode, re-enable user goroutines. These will be
  1574  	// queued to run after we start the world.
  1575  	schedEnableUser(true)
  1576  
  1577  	// endCycle depends on all gcWork cache stats being flushed.
  1578  	// The termination algorithm above ensured that up to
  1579  	// allocations since the ragged barrier.
  1580  	nextTriggerRatio := gcController.endCycle()
  1581  
  1582  	// Perform mark termination. This will restart the world.
  1583  	gcMarkTermination(nextTriggerRatio)
  1584  }
  1585  
  1586  func gcMarkTermination(nextTriggerRatio float64) {
  1587  	// World is stopped.
  1588  	// Start marktermination which includes enabling the write barrier.
  1589  	atomic.Store(&gcBlackenEnabled, 0)
  1590  	setGCPhase(_GCmarktermination)
  1591  
  1592  	work.heap1 = memstats.heap_live
  1593  	startTime := nanotime()
  1594  
  1595  	mp := acquirem()
  1596  	mp.preemptoff = "gcing"
  1597  	_g_ := getg()
  1598  	_g_.m.traceback = 2
  1599  	gp := _g_.m.curg
  1600  	casgstatus(gp, _Grunning, _Gwaiting)
  1601  	gp.waitreason = waitReasonGarbageCollection
  1602  
  1603  	// Run gc on the g0 stack. We do this so that the g stack
  1604  	// we're currently running on will no longer change. Cuts
  1605  	// the root set down a bit (g0 stacks are not scanned, and
  1606  	// we don't need to scan gc's internal state).  We also
  1607  	// need to switch to g0 so we can shrink the stack.
  1608  	systemstack(func() {
  1609  		gcMark(startTime)
  1610  		// Must return immediately.
  1611  		// The outer function's stack may have moved
  1612  		// during gcMark (it shrinks stacks, including the
  1613  		// outer function's stack), so we must not refer
  1614  		// to any of its variables. Return back to the
  1615  		// non-system stack to pick up the new addresses
  1616  		// before continuing.
  1617  	})
  1618  
  1619  	systemstack(func() {
  1620  		work.heap2 = work.bytesMarked
  1621  		if debug.gccheckmark > 0 {
  1622  			// Run a full non-parallel, stop-the-world
  1623  			// mark using checkmark bits, to check that we
  1624  			// didn't forget to mark anything during the
  1625  			// concurrent mark process.
  1626  			gcResetMarkState()
  1627  			initCheckmarks()
  1628  			gcw := &getg().m.p.ptr().gcw
  1629  			gcDrain(gcw, 0)
  1630  			wbBufFlush1(getg().m.p.ptr())
  1631  			gcw.dispose()
  1632  			clearCheckmarks()
  1633  		}
  1634  
  1635  		// marking is complete so we can turn the write barrier off
  1636  		setGCPhase(_GCoff)
  1637  		gcSweep(work.mode)
  1638  	})
  1639  
  1640  	_g_.m.traceback = 0
  1641  	casgstatus(gp, _Gwaiting, _Grunning)
  1642  
  1643  	if trace.enabled {
  1644  		traceGCDone()
  1645  	}
  1646  
  1647  	// all done
  1648  	mp.preemptoff = ""
  1649  
  1650  	if gcphase != _GCoff {
  1651  		throw("gc done but gcphase != _GCoff")
  1652  	}
  1653  
  1654  	// Record next_gc and heap_inuse for scavenger.
  1655  	memstats.last_next_gc = memstats.next_gc
  1656  	memstats.last_heap_inuse = memstats.heap_inuse
  1657  
  1658  	// Update GC trigger and pacing for the next cycle.
  1659  	gcSetTriggerRatio(nextTriggerRatio)
  1660  
  1661  	// Pacing changed, so the scavenger should be awoken.
  1662  	wakeScavenger()
  1663  
  1664  	// Update timing memstats
  1665  	now := nanotime()
  1666  	sec, nsec, _ := time_now()
  1667  	unixNow := sec*1e9 + int64(nsec)
  1668  	work.pauseNS += now - work.pauseStart
  1669  	work.tEnd = now
  1670  	atomic.Store64(&memstats.last_gc_unix, uint64(unixNow)) // must be Unix time to make sense to user
  1671  	atomic.Store64(&memstats.last_gc_nanotime, uint64(now)) // monotonic time for us
  1672  	memstats.pause_ns[memstats.numgc%uint32(len(memstats.pause_ns))] = uint64(work.pauseNS)
  1673  	memstats.pause_end[memstats.numgc%uint32(len(memstats.pause_end))] = uint64(unixNow)
  1674  	memstats.pause_total_ns += uint64(work.pauseNS)
  1675  
  1676  	// Update work.totaltime.
  1677  	sweepTermCpu := int64(work.stwprocs) * (work.tMark - work.tSweepTerm)
  1678  	// We report idle marking time below, but omit it from the
  1679  	// overall utilization here since it's "free".
  1680  	markCpu := gcController.assistTime + gcController.dedicatedMarkTime + gcController.fractionalMarkTime
  1681  	markTermCpu := int64(work.stwprocs) * (work.tEnd - work.tMarkTerm)
  1682  	cycleCpu := sweepTermCpu + markCpu + markTermCpu
  1683  	work.totaltime += cycleCpu
  1684  
  1685  	// Compute overall GC CPU utilization.
  1686  	totalCpu := sched.totaltime + (now-sched.procresizetime)*int64(gomaxprocs)
  1687  	memstats.gc_cpu_fraction = float64(work.totaltime) / float64(totalCpu)
  1688  
  1689  	// Reset sweep state.
  1690  	sweep.nbgsweep = 0
  1691  	sweep.npausesweep = 0
  1692  
  1693  	if work.userForced {
  1694  		memstats.numforcedgc++
  1695  	}
  1696  
  1697  	// Bump GC cycle count and wake goroutines waiting on sweep.
  1698  	lock(&work.sweepWaiters.lock)
  1699  	memstats.numgc++
  1700  	injectglist(&work.sweepWaiters.list)
  1701  	unlock(&work.sweepWaiters.lock)
  1702  
  1703  	// Finish the current heap profiling cycle and start a new
  1704  	// heap profiling cycle. We do this before starting the world
  1705  	// so events don't leak into the wrong cycle.
  1706  	mProf_NextCycle()
  1707  
  1708  	systemstack(func() { startTheWorldWithSema(true) })
  1709  
  1710  	// Flush the heap profile so we can start a new cycle next GC.
  1711  	// This is relatively expensive, so we don't do it with the
  1712  	// world stopped.
  1713  	mProf_Flush()
  1714  
  1715  	// Prepare workbufs for freeing by the sweeper. We do this
  1716  	// asynchronously because it can take non-trivial time.
  1717  	prepareFreeWorkbufs()
  1718  
  1719  	// Free stack spans. This must be done between GC cycles.
  1720  	systemstack(freeStackSpans)
  1721  
  1722  	// Ensure all mcaches are flushed. Each P will flush its own
  1723  	// mcache before allocating, but idle Ps may not. Since this
  1724  	// is necessary to sweep all spans, we need to ensure all
  1725  	// mcaches are flushed before we start the next GC cycle.
  1726  	systemstack(func() {
  1727  		forEachP(func(_p_ *p) {
  1728  			_p_.mcache.prepareForSweep()
  1729  		})
  1730  	})
  1731  
  1732  	// Print gctrace before dropping worldsema. As soon as we drop
  1733  	// worldsema another cycle could start and smash the stats
  1734  	// we're trying to print.
  1735  	if debug.gctrace > 0 {
  1736  		util := int(memstats.gc_cpu_fraction * 100)
  1737  
  1738  		var sbuf [24]byte
  1739  		printlock()
  1740  		print("gc ", memstats.numgc,
  1741  			" @", string(itoaDiv(sbuf[:], uint64(work.tSweepTerm-runtimeInitTime)/1e6, 3)), "s ",
  1742  			util, "%: ")
  1743  		prev := work.tSweepTerm
  1744  		for i, ns := range []int64{work.tMark, work.tMarkTerm, work.tEnd} {
  1745  			if i != 0 {
  1746  				print("+")
  1747  			}
  1748  			print(string(fmtNSAsMS(sbuf[:], uint64(ns-prev))))
  1749  			prev = ns
  1750  		}
  1751  		print(" ms clock, ")
  1752  		for i, ns := range []int64{sweepTermCpu, gcController.assistTime, gcController.dedicatedMarkTime + gcController.fractionalMarkTime, gcController.idleMarkTime, markTermCpu} {
  1753  			if i == 2 || i == 3 {
  1754  				// Separate mark time components with /.
  1755  				print("/")
  1756  			} else if i != 0 {
  1757  				print("+")
  1758  			}
  1759  			print(string(fmtNSAsMS(sbuf[:], uint64(ns))))
  1760  		}
  1761  		print(" ms cpu, ",
  1762  			work.heap0>>20, "->", work.heap1>>20, "->", work.heap2>>20, " MB, ",
  1763  			work.heapGoal>>20, " MB goal, ",
  1764  			work.maxprocs, " P")
  1765  		if work.userForced {
  1766  			print(" (forced)")
  1767  		}
  1768  		print("\n")
  1769  		printunlock()
  1770  	}
  1771  
  1772  	semrelease(&worldsema)
  1773  	// Careful: another GC cycle may start now.
  1774  
  1775  	releasem(mp)
  1776  	mp = nil
  1777  
  1778  	// now that gc is done, kick off finalizer thread if needed
  1779  	if !concurrentSweep {
  1780  		// give the queued finalizers, if any, a chance to run
  1781  		Gosched()
  1782  	}
  1783  }
  1784  
  1785  // gcBgMarkStartWorkers prepares background mark worker goroutines.
  1786  // These goroutines will not run until the mark phase, but they must
  1787  // be started while the work is not stopped and from a regular G
  1788  // stack. The caller must hold worldsema.
  1789  func gcBgMarkStartWorkers() {
  1790  	// Background marking is performed by per-P G's. Ensure that
  1791  	// each P has a background GC G.
  1792  	for _, p := range allp {
  1793  		if p.gcBgMarkWorker == 0 {
  1794  			go gcBgMarkWorker(p)
  1795  			notetsleepg(&work.bgMarkReady, -1)
  1796  			noteclear(&work.bgMarkReady)
  1797  		}
  1798  	}
  1799  }
  1800  
  1801  // gcBgMarkPrepare sets up state for background marking.
  1802  // Mutator assists must not yet be enabled.
  1803  func gcBgMarkPrepare() {
  1804  	// Background marking will stop when the work queues are empty
  1805  	// and there are no more workers (note that, since this is
  1806  	// concurrent, this may be a transient state, but mark
  1807  	// termination will clean it up). Between background workers
  1808  	// and assists, we don't really know how many workers there
  1809  	// will be, so we pretend to have an arbitrarily large number
  1810  	// of workers, almost all of which are "waiting". While a
  1811  	// worker is working it decrements nwait. If nproc == nwait,
  1812  	// there are no workers.
  1813  	work.nproc = ^uint32(0)
  1814  	work.nwait = ^uint32(0)
  1815  }
  1816  
  1817  func gcBgMarkWorker(_p_ *p) {
  1818  	gp := getg()
  1819  
  1820  	type parkInfo struct {
  1821  		m      muintptr // Release this m on park.
  1822  		attach puintptr // If non-nil, attach to this p on park.
  1823  	}
  1824  	// We pass park to a gopark unlock function, so it can't be on
  1825  	// the stack (see gopark). Prevent deadlock from recursively
  1826  	// starting GC by disabling preemption.
  1827  	gp.m.preemptoff = "GC worker init"
  1828  	park := new(parkInfo)
  1829  	gp.m.preemptoff = ""
  1830  
  1831  	park.m.set(acquirem())
  1832  	park.attach.set(_p_)
  1833  	// Inform gcBgMarkStartWorkers that this worker is ready.
  1834  	// After this point, the background mark worker is scheduled
  1835  	// cooperatively by gcController.findRunnable. Hence, it must
  1836  	// never be preempted, as this would put it into _Grunnable
  1837  	// and put it on a run queue. Instead, when the preempt flag
  1838  	// is set, this puts itself into _Gwaiting to be woken up by
  1839  	// gcController.findRunnable at the appropriate time.
  1840  	notewakeup(&work.bgMarkReady)
  1841  
  1842  	for {
  1843  		// Go to sleep until woken by gcController.findRunnable.
  1844  		// We can't releasem yet since even the call to gopark
  1845  		// may be preempted.
  1846  		gopark(func(g *g, parkp unsafe.Pointer) bool {
  1847  			park := (*parkInfo)(parkp)
  1848  
  1849  			// The worker G is no longer running, so it's
  1850  			// now safe to allow preemption.
  1851  			releasem(park.m.ptr())
  1852  
  1853  			// If the worker isn't attached to its P,
  1854  			// attach now. During initialization and after
  1855  			// a phase change, the worker may have been
  1856  			// running on a different P. As soon as we
  1857  			// attach, the owner P may schedule the
  1858  			// worker, so this must be done after the G is
  1859  			// stopped.
  1860  			if park.attach != 0 {
  1861  				p := park.attach.ptr()
  1862  				park.attach.set(nil)
  1863  				// cas the worker because we may be
  1864  				// racing with a new worker starting
  1865  				// on this P.
  1866  				if !p.gcBgMarkWorker.cas(0, guintptr(unsafe.Pointer(g))) {
  1867  					// The P got a new worker.
  1868  					// Exit this worker.
  1869  					return false
  1870  				}
  1871  			}
  1872  			return true
  1873  		}, unsafe.Pointer(park), waitReasonGCWorkerIdle, traceEvGoBlock, 0)
  1874  
  1875  		// Loop until the P dies and disassociates this
  1876  		// worker (the P may later be reused, in which case
  1877  		// it will get a new worker) or we failed to associate.
  1878  		if _p_.gcBgMarkWorker.ptr() != gp {
  1879  			break
  1880  		}
  1881  
  1882  		// Disable preemption so we can use the gcw. If the
  1883  		// scheduler wants to preempt us, we'll stop draining,
  1884  		// dispose the gcw, and then preempt.
  1885  		park.m.set(acquirem())
  1886  
  1887  		if gcBlackenEnabled == 0 {
  1888  			throw("gcBgMarkWorker: blackening not enabled")
  1889  		}
  1890  
  1891  		startTime := nanotime()
  1892  		_p_.gcMarkWorkerStartTime = startTime
  1893  
  1894  		decnwait := atomic.Xadd(&work.nwait, -1)
  1895  		if decnwait == work.nproc {
  1896  			println("runtime: work.nwait=", decnwait, "work.nproc=", work.nproc)
  1897  			throw("work.nwait was > work.nproc")
  1898  		}
  1899  
  1900  		systemstack(func() {
  1901  			// Mark our goroutine preemptible so its stack
  1902  			// can be scanned. This lets two mark workers
  1903  			// scan each other (otherwise, they would
  1904  			// deadlock). We must not modify anything on
  1905  			// the G stack. However, stack shrinking is
  1906  			// disabled for mark workers, so it is safe to
  1907  			// read from the G stack.
  1908  			casgstatus(gp, _Grunning, _Gwaiting)
  1909  			switch _p_.gcMarkWorkerMode {
  1910  			default:
  1911  				throw("gcBgMarkWorker: unexpected gcMarkWorkerMode")
  1912  			case gcMarkWorkerDedicatedMode:
  1913  				gcDrain(&_p_.gcw, gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1914  				if gp.preempt {
  1915  					// We were preempted. This is
  1916  					// a useful signal to kick
  1917  					// everything out of the run
  1918  					// queue so it can run
  1919  					// somewhere else.
  1920  					lock(&sched.lock)
  1921  					for {
  1922  						gp, _ := runqget(_p_)
  1923  						if gp == nil {
  1924  							break
  1925  						}
  1926  						globrunqput(gp)
  1927  					}
  1928  					unlock(&sched.lock)
  1929  				}
  1930  				// Go back to draining, this time
  1931  				// without preemption.
  1932  				gcDrain(&_p_.gcw, gcDrainFlushBgCredit)
  1933  			case gcMarkWorkerFractionalMode:
  1934  				gcDrain(&_p_.gcw, gcDrainFractional|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1935  			case gcMarkWorkerIdleMode:
  1936  				gcDrain(&_p_.gcw, gcDrainIdle|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  1937  			}
  1938  			casgstatus(gp, _Gwaiting, _Grunning)
  1939  		})
  1940  
  1941  		// Account for time.
  1942  		duration := nanotime() - startTime
  1943  		switch _p_.gcMarkWorkerMode {
  1944  		case gcMarkWorkerDedicatedMode:
  1945  			atomic.Xaddint64(&gcController.dedicatedMarkTime, duration)
  1946  			atomic.Xaddint64(&gcController.dedicatedMarkWorkersNeeded, 1)
  1947  		case gcMarkWorkerFractionalMode:
  1948  			atomic.Xaddint64(&gcController.fractionalMarkTime, duration)
  1949  			atomic.Xaddint64(&_p_.gcFractionalMarkTime, duration)
  1950  		case gcMarkWorkerIdleMode:
  1951  			atomic.Xaddint64(&gcController.idleMarkTime, duration)
  1952  		}
  1953  
  1954  		// Was this the last worker and did we run out
  1955  		// of work?
  1956  		incnwait := atomic.Xadd(&work.nwait, +1)
  1957  		if incnwait > work.nproc {
  1958  			println("runtime: p.gcMarkWorkerMode=", _p_.gcMarkWorkerMode,
  1959  				"work.nwait=", incnwait, "work.nproc=", work.nproc)
  1960  			throw("work.nwait > work.nproc")
  1961  		}
  1962  
  1963  		// If this worker reached a background mark completion
  1964  		// point, signal the main GC goroutine.
  1965  		if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
  1966  			// Make this G preemptible and disassociate it
  1967  			// as the worker for this P so
  1968  			// findRunnableGCWorker doesn't try to
  1969  			// schedule it.
  1970  			_p_.gcBgMarkWorker.set(nil)
  1971  			releasem(park.m.ptr())
  1972  
  1973  			gcMarkDone()
  1974  
  1975  			// Disable preemption and prepare to reattach
  1976  			// to the P.
  1977  			//
  1978  			// We may be running on a different P at this
  1979  			// point, so we can't reattach until this G is
  1980  			// parked.
  1981  			park.m.set(acquirem())
  1982  			park.attach.set(_p_)
  1983  		}
  1984  	}
  1985  }
  1986  
  1987  // gcMarkWorkAvailable reports whether executing a mark worker
  1988  // on p is potentially useful. p may be nil, in which case it only
  1989  // checks the global sources of work.
  1990  func gcMarkWorkAvailable(p *p) bool {
  1991  	if p != nil && !p.gcw.empty() {
  1992  		return true
  1993  	}
  1994  	if !work.full.empty() {
  1995  		return true // global work available
  1996  	}
  1997  	if work.markrootNext < work.markrootJobs {
  1998  		return true // root scan work available
  1999  	}
  2000  	return false
  2001  }
  2002  
  2003  // gcMark runs the mark (or, for concurrent GC, mark termination)
  2004  // All gcWork caches must be empty.
  2005  // STW is in effect at this point.
  2006  func gcMark(start_time int64) {
  2007  	if debug.allocfreetrace > 0 {
  2008  		tracegc()
  2009  	}
  2010  
  2011  	if gcphase != _GCmarktermination {
  2012  		throw("in gcMark expecting to see gcphase as _GCmarktermination")
  2013  	}
  2014  	work.tstart = start_time
  2015  
  2016  	// Check that there's no marking work remaining.
  2017  	if work.full != 0 || work.markrootNext < work.markrootJobs {
  2018  		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")
  2019  		panic("non-empty mark queue after concurrent mark")
  2020  	}
  2021  
  2022  	if debug.gccheckmark > 0 {
  2023  		// This is expensive when there's a large number of
  2024  		// Gs, so only do it if checkmark is also enabled.
  2025  		gcMarkRootCheck()
  2026  	}
  2027  	if work.full != 0 {
  2028  		throw("work.full != 0")
  2029  	}
  2030  
  2031  	// Clear out buffers and double-check that all gcWork caches
  2032  	// are empty. This should be ensured by gcMarkDone before we
  2033  	// enter mark termination.
  2034  	//
  2035  	// TODO: We could clear out buffers just before mark if this
  2036  	// has a non-negligible impact on STW time.
  2037  	for _, p := range allp {
  2038  		// The write barrier may have buffered pointers since
  2039  		// the gcMarkDone barrier. However, since the barrier
  2040  		// ensured all reachable objects were marked, all of
  2041  		// these must be pointers to black objects. Hence we
  2042  		// can just discard the write barrier buffer.
  2043  		if debug.gccheckmark > 0 || throwOnGCWork {
  2044  			// For debugging, flush the buffer and make
  2045  			// sure it really was all marked.
  2046  			wbBufFlush1(p)
  2047  		} else {
  2048  			p.wbBuf.reset()
  2049  		}
  2050  
  2051  		gcw := &p.gcw
  2052  		if !gcw.empty() {
  2053  			printlock()
  2054  			print("runtime: P ", p.id, " flushedWork ", gcw.flushedWork)
  2055  			if gcw.wbuf1 == nil {
  2056  				print(" wbuf1=<nil>")
  2057  			} else {
  2058  				print(" wbuf1.n=", gcw.wbuf1.nobj)
  2059  			}
  2060  			if gcw.wbuf2 == nil {
  2061  				print(" wbuf2=<nil>")
  2062  			} else {
  2063  				print(" wbuf2.n=", gcw.wbuf2.nobj)
  2064  			}
  2065  			print("\n")
  2066  			throw("P has cached GC work at end of mark termination")
  2067  		}
  2068  		// There may still be cached empty buffers, which we
  2069  		// need to flush since we're going to free them. Also,
  2070  		// there may be non-zero stats because we allocated
  2071  		// black after the gcMarkDone barrier.
  2072  		gcw.dispose()
  2073  	}
  2074  
  2075  	throwOnGCWork = false
  2076  
  2077  	cachestats()
  2078  
  2079  	// Update the marked heap stat.
  2080  	memstats.heap_marked = work.bytesMarked
  2081  
  2082  	// Update other GC heap size stats. This must happen after
  2083  	// cachestats (which flushes local statistics to these) and
  2084  	// flushallmcaches (which modifies heap_live).
  2085  	memstats.heap_live = work.bytesMarked
  2086  	memstats.heap_scan = uint64(gcController.scanWork)
  2087  
  2088  	if trace.enabled {
  2089  		traceHeapAlloc()
  2090  	}
  2091  }
  2092  
  2093  // gcSweep must be called on the system stack because it acquires the heap
  2094  // lock. See mheap for details.
  2095  //go:systemstack
  2096  func gcSweep(mode gcMode) {
  2097  	if gcphase != _GCoff {
  2098  		throw("gcSweep being done but phase is not GCoff")
  2099  	}
  2100  
  2101  	lock(&mheap_.lock)
  2102  	mheap_.sweepgen += 2
  2103  	mheap_.sweepdone = 0
  2104  	if mheap_.sweepSpans[mheap_.sweepgen/2%2].index != 0 {
  2105  		// We should have drained this list during the last
  2106  		// sweep phase. We certainly need to start this phase
  2107  		// with an empty swept list.
  2108  		throw("non-empty swept list")
  2109  	}
  2110  	mheap_.pagesSwept = 0
  2111  	mheap_.sweepArenas = mheap_.allArenas
  2112  	mheap_.reclaimIndex = 0
  2113  	mheap_.reclaimCredit = 0
  2114  	unlock(&mheap_.lock)
  2115  
  2116  	if !_ConcurrentSweep || mode == gcForceBlockMode {
  2117  		// Special case synchronous sweep.
  2118  		// Record that no proportional sweeping has to happen.
  2119  		lock(&mheap_.lock)
  2120  		mheap_.sweepPagesPerByte = 0
  2121  		unlock(&mheap_.lock)
  2122  		// Sweep all spans eagerly.
  2123  		for sweepone() != ^uintptr(0) {
  2124  			sweep.npausesweep++
  2125  		}
  2126  		// Free workbufs eagerly.
  2127  		prepareFreeWorkbufs()
  2128  		for freeSomeWbufs(false) {
  2129  		}
  2130  		// All "free" events for this mark/sweep cycle have
  2131  		// now happened, so we can make this profile cycle
  2132  		// available immediately.
  2133  		mProf_NextCycle()
  2134  		mProf_Flush()
  2135  		return
  2136  	}
  2137  
  2138  	// Background sweep.
  2139  	lock(&sweep.lock)
  2140  	if sweep.parked {
  2141  		sweep.parked = false
  2142  		ready(sweep.g, 0, true)
  2143  	}
  2144  	unlock(&sweep.lock)
  2145  }
  2146  
  2147  // gcResetMarkState resets global state prior to marking (concurrent
  2148  // or STW) and resets the stack scan state of all Gs.
  2149  //
  2150  // This is safe to do without the world stopped because any Gs created
  2151  // during or after this will start out in the reset state.
  2152  //
  2153  // gcResetMarkState must be called on the system stack because it acquires
  2154  // the heap lock. See mheap for details.
  2155  //
  2156  //go:systemstack
  2157  func gcResetMarkState() {
  2158  	// This may be called during a concurrent phase, so make sure
  2159  	// allgs doesn't change.
  2160  	lock(&allglock)
  2161  	for _, gp := range allgs {
  2162  		gp.gcscandone = false  // set to true in gcphasework
  2163  		gp.gcscanvalid = false // stack has not been scanned
  2164  		gp.gcAssistBytes = 0
  2165  	}
  2166  	unlock(&allglock)
  2167  
  2168  	// Clear page marks. This is just 1MB per 64GB of heap, so the
  2169  	// time here is pretty trivial.
  2170  	lock(&mheap_.lock)
  2171  	arenas := mheap_.allArenas
  2172  	unlock(&mheap_.lock)
  2173  	for _, ai := range arenas {
  2174  		ha := mheap_.arenas[ai.l1()][ai.l2()]
  2175  		for i := range ha.pageMarks {
  2176  			ha.pageMarks[i] = 0
  2177  		}
  2178  	}
  2179  
  2180  	work.bytesMarked = 0
  2181  	work.initialHeapLive = atomic.Load64(&memstats.heap_live)
  2182  }
  2183  
  2184  // Hooks for other packages
  2185  
  2186  var poolcleanup func()
  2187  
  2188  //go:linkname sync_runtime_registerPoolCleanup sync.runtime_registerPoolCleanup
  2189  func sync_runtime_registerPoolCleanup(f func()) {
  2190  	poolcleanup = f
  2191  }
  2192  
  2193  func clearpools() {
  2194  	// clear sync.Pools
  2195  	if poolcleanup != nil {
  2196  		poolcleanup()
  2197  	}
  2198  
  2199  	// Clear central sudog cache.
  2200  	// Leave per-P caches alone, they have strictly bounded size.
  2201  	// Disconnect cached list before dropping it on the floor,
  2202  	// so that a dangling ref to one entry does not pin all of them.
  2203  	lock(&sched.sudoglock)
  2204  	var sg, sgnext *sudog
  2205  	for sg = sched.sudogcache; sg != nil; sg = sgnext {
  2206  		sgnext = sg.next
  2207  		sg.next = nil
  2208  	}
  2209  	sched.sudogcache = nil
  2210  	unlock(&sched.sudoglock)
  2211  
  2212  	// Clear central defer pools.
  2213  	// Leave per-P pools alone, they have strictly bounded size.
  2214  	lock(&sched.deferlock)
  2215  	for i := range sched.deferpool {
  2216  		// disconnect cached list before dropping it on the floor,
  2217  		// so that a dangling ref to one entry does not pin all of them.
  2218  		var d, dlink *_defer
  2219  		for d = sched.deferpool[i]; d != nil; d = dlink {
  2220  			dlink = d.link
  2221  			d.link = nil
  2222  		}
  2223  		sched.deferpool[i] = nil
  2224  	}
  2225  	unlock(&sched.deferlock)
  2226  }
  2227  
  2228  // Timing
  2229  
  2230  // itoaDiv formats val/(10**dec) into buf.
  2231  func itoaDiv(buf []byte, val uint64, dec int) []byte {
  2232  	i := len(buf) - 1
  2233  	idec := i - dec
  2234  	for val >= 10 || i >= idec {
  2235  		buf[i] = byte(val%10 + '0')
  2236  		i--
  2237  		if i == idec {
  2238  			buf[i] = '.'
  2239  			i--
  2240  		}
  2241  		val /= 10
  2242  	}
  2243  	buf[i] = byte(val + '0')
  2244  	return buf[i:]
  2245  }
  2246  
  2247  // fmtNSAsMS nicely formats ns nanoseconds as milliseconds.
  2248  func fmtNSAsMS(buf []byte, ns uint64) []byte {
  2249  	if ns >= 10e6 {
  2250  		// Format as whole milliseconds.
  2251  		return itoaDiv(buf, ns/1e6, 0)
  2252  	}
  2253  	// Format two digits of precision, with at most three decimal places.
  2254  	x := ns / 1e3
  2255  	if x == 0 {
  2256  		buf[0] = '0'
  2257  		return buf[:1]
  2258  	}
  2259  	dec := 3
  2260  	for x >= 100 {
  2261  		x /= 10
  2262  		dec--
  2263  	}
  2264  	return itoaDiv(buf, x, dec)
  2265  }
  2266  

View as plain text