...
Run Format

Source file src/runtime/mgc.go

Documentation: runtime

  // Copyright 2009 The Go Authors. All rights reserved.
  // Use of this source code is governed by a BSD-style
  // license that can be found in the LICENSE file.
  
  // Garbage collector (GC).
  //
  // The GC runs concurrently with mutator threads, is type accurate (aka precise), allows multiple
  // GC thread to run in parallel. It is a concurrent mark and sweep that uses a write barrier. It is
  // non-generational and non-compacting. Allocation is done using size segregated per P allocation
  // areas to minimize fragmentation while eliminating locks in the common case.
  //
  // The algorithm decomposes into several steps.
  // This is a high level description of the algorithm being used. For an overview of GC a good
  // place to start is Richard Jones' gchandbook.org.
  //
  // The algorithm's intellectual heritage includes Dijkstra's on-the-fly algorithm, see
  // Edsger W. Dijkstra, Leslie Lamport, A. J. Martin, C. S. Scholten, and E. F. M. Steffens. 1978.
  // On-the-fly garbage collection: an exercise in cooperation. Commun. ACM 21, 11 (November 1978),
  // 966-975.
  // For journal quality proofs that these steps are complete, correct, and terminate see
  // Hudson, R., and Moss, J.E.B. Copying Garbage Collection without stopping the world.
  // Concurrency and Computation: Practice and Experience 15(3-5), 2003.
  //
  // 1. GC performs sweep termination.
  //
  //    a. Stop the world. This causes all Ps to reach a GC safe-point.
  //
  //    b. Sweep any unswept spans. There will only be unswept spans if
  //    this GC cycle was forced before the expected time.
  //
  // 2. GC performs the "mark 1" sub-phase. In this sub-phase, Ps are
  // allowed to locally cache parts of the work queue.
  //
  //    a. Prepare for the mark phase by setting gcphase to _GCmark
  //    (from _GCoff), enabling the write barrier, enabling mutator
  //    assists, and enqueueing root mark jobs. No objects may be
  //    scanned until all Ps have enabled the write barrier, which is
  //    accomplished using STW.
  //
  //    b. Start the world. From this point, GC work is done by mark
  //    workers started by the scheduler and by assists performed as
  //    part of allocation. The write barrier shades both the
  //    overwritten pointer and the new pointer value for any pointer
  //    writes (see mbarrier.go for details). Newly allocated objects
  //    are immediately marked black.
  //
  //    c. GC performs root marking jobs. This includes scanning all
  //    stacks, shading all globals, and shading any heap pointers in
  //    off-heap runtime data structures. Scanning a stack stops a
  //    goroutine, shades any pointers found on its stack, and then
  //    resumes the goroutine.
  //
  //    d. GC drains the work queue of grey objects, scanning each grey
  //    object to black and shading all pointers found in the object
  //    (which in turn may add those pointers to the work queue).
  //
  // 3. Once the global work queue is empty (but local work queue caches
  // may still contain work), GC performs the "mark 2" sub-phase.
  //
  //    a. GC stops all workers, disables local work queue caches,
  //    flushes each P's local work queue cache to the global work queue
  //    cache, and reenables workers.
  //
  //    b. GC again drains the work queue, as in 2d above.
  //
  // 4. Once the work queue is empty, GC performs mark termination.
  //
  //    a. Stop the world.
  //
  //    b. Set gcphase to _GCmarktermination, and disable workers and
  //    assists.
  //
  //    c. Drain any remaining work from the work queue (typically there
  //    will be none).
  //
  //    d. Perform other housekeeping like flushing mcaches.
  //
  // 5. GC performs the sweep phase.
  //
  //    a. Prepare for the sweep phase by setting gcphase to _GCoff,
  //    setting up sweep state and disabling the write barrier.
  //
  //    b. Start the world. From this point on, newly allocated objects
  //    are white, and allocating sweeps spans before use if necessary.
  //
  //    c. GC does concurrent sweeping in the background and in response
  //    to allocation. See description below.
  //
  // 6. When sufficient allocation has taken place, replay the sequence
  // starting with 1 above. See discussion of GC rate below.
  
  // Concurrent sweep.
  //
  // The sweep phase proceeds concurrently with normal program execution.
  // The heap is swept span-by-span both lazily (when a goroutine needs another span)
  // and concurrently in a background goroutine (this helps programs that are not CPU bound).
  // At the end of STW mark termination all spans are marked as "needs sweeping".
  //
  // The background sweeper goroutine simply sweeps spans one-by-one.
  //
  // To avoid requesting more OS memory while there are unswept spans, when a
  // goroutine needs another span, it first attempts to reclaim that much memory
  // by sweeping. When a goroutine needs to allocate a new small-object span, it
  // sweeps small-object spans for the same object size until it frees at least
  // one object. When a goroutine needs to allocate large-object span from heap,
  // it sweeps spans until it frees at least that many pages into heap. There is
  // one case where this may not suffice: if a goroutine sweeps and frees two
  // nonadjacent one-page spans to the heap, it will allocate a new two-page
  // span, but there can still be other one-page unswept spans which could be
  // combined into a two-page span.
  //
  // It's critical to ensure that no operations proceed on unswept spans (that would corrupt
  // mark bits in GC bitmap). During GC all mcaches are flushed into the central cache,
  // so they are empty. When a goroutine grabs a new span into mcache, it sweeps it.
  // When a goroutine explicitly frees an object or sets a finalizer, it ensures that
  // the span is swept (either by sweeping it, or by waiting for the concurrent sweep to finish).
  // The finalizer goroutine is kicked off only when all spans are swept.
  // When the next GC starts, it sweeps all not-yet-swept spans (if any).
  
  // GC rate.
  // Next GC is after we've allocated an extra amount of memory proportional to
  // the amount already in use. The proportion is controlled by GOGC environment variable
  // (100 by default). If GOGC=100 and we're using 4M, we'll GC again when we get to 8M
  // (this mark is tracked in next_gc variable). This keeps the GC cost in linear
  // proportion to the allocation cost. Adjusting GOGC just changes the linear constant
  // (and also the amount of extra memory used).
  
  // Oblets
  //
  // In order to prevent long pauses while scanning large objects and to
  // improve parallelism, the garbage collector breaks up scan jobs for
  // objects larger than maxObletBytes into "oblets" of at most
  // maxObletBytes. When scanning encounters the beginning of a large
  // object, it scans only the first oblet and enqueues the remaining
  // oblets as new scan jobs.
  
  package runtime
  
  import (
  	"runtime/internal/atomic"
  	"runtime/internal/sys"
  	"unsafe"
  )
  
  const (
  	_DebugGC         = 0
  	_ConcurrentSweep = true
  	_FinBlockSize    = 4 * 1024
  
  	// sweepMinHeapDistance is a lower bound on the heap distance
  	// (in bytes) reserved for concurrent sweeping between GC
  	// cycles. This will be scaled by gcpercent/100.
  	sweepMinHeapDistance = 1024 * 1024
  )
  
  // heapminimum is the minimum heap size at which to trigger GC.
  // For small heaps, this overrides the usual GOGC*live set rule.
  //
  // When there is a very small live set but a lot of allocation, simply
  // collecting when the heap reaches GOGC*live results in many GC
  // cycles and high total per-GC overhead. This minimum amortizes this
  // per-GC overhead while keeping the heap reasonably small.
  //
  // During initialization this is set to 4MB*GOGC/100. In the case of
  // GOGC==0, this will set heapminimum to 0, resulting in constant
  // collection even when the heap size is small, which is useful for
  // debugging.
  var heapminimum uint64 = defaultHeapMinimum
  
  // defaultHeapMinimum is the value of heapminimum for GOGC==100.
  const defaultHeapMinimum = 4 << 20
  
  // Initialized from $GOGC.  GOGC=off means no GC.
  var gcpercent int32
  
  func gcinit() {
  	if unsafe.Sizeof(workbuf{}) != _WorkbufSize {
  		throw("size of Workbuf is suboptimal")
  	}
  
  	// No sweep on the first cycle.
  	mheap_.sweepdone = 1
  
  	// Set a reasonable initial GC trigger.
  	memstats.triggerRatio = 7 / 8.0
  
  	// Fake a heap_marked value so it looks like a trigger at
  	// heapminimum is the appropriate growth from heap_marked.
  	// This will go into computing the initial GC goal.
  	memstats.heap_marked = uint64(float64(heapminimum) / (1 + memstats.triggerRatio))
  
  	// Set gcpercent from the environment. This will also compute
  	// and set the GC trigger and goal.
  	_ = setGCPercent(readgogc())
  
  	work.startSema = 1
  	work.markDoneSema = 1
  }
  
  func readgogc() int32 {
  	p := gogetenv("GOGC")
  	if p == "off" {
  		return -1
  	}
  	if n, ok := atoi32(p); ok {
  		return n
  	}
  	return 100
  }
  
  // gcenable is called after the bulk of the runtime initialization,
  // just before we're about to start letting user code run.
  // It kicks off the background sweeper goroutine and enables GC.
  func gcenable() {
  	c := make(chan int, 1)
  	go bgsweep(c)
  	<-c
  	memstats.enablegc = true // now that runtime is initialized, GC is okay
  }
  
  //go:linkname setGCPercent runtime/debug.setGCPercent
  func setGCPercent(in int32) (out int32) {
  	lock(&mheap_.lock)
  	out = gcpercent
  	if in < 0 {
  		in = -1
  	}
  	gcpercent = in
  	heapminimum = defaultHeapMinimum * uint64(gcpercent) / 100
  	// Update pacing in response to gcpercent change.
  	gcSetTriggerRatio(memstats.triggerRatio)
  	unlock(&mheap_.lock)
  	return out
  }
  
  // Garbage collector phase.
  // Indicates to write barrier and synchronization task to perform.
  var gcphase uint32
  
  // The compiler knows about this variable.
  // If you change it, you must change builtin/runtime.go, too.
  // If you change the first four bytes, you must also change the write
  // barrier insertion code.
  var writeBarrier struct {
  	enabled bool    // compiler emits a check of this before calling write barrier
  	pad     [3]byte // compiler uses 32-bit load for "enabled" field
  	needed  bool    // whether we need a write barrier for current GC phase
  	cgo     bool    // whether we need a write barrier for a cgo check
  	alignme uint64  // guarantee alignment so that compiler can use a 32 or 64-bit load
  }
  
  // gcBlackenEnabled is 1 if mutator assists and background mark
  // workers are allowed to blacken objects. This must only be set when
  // gcphase == _GCmark.
  var gcBlackenEnabled uint32
  
  // gcBlackenPromptly indicates that optimizations that may
  // hide work from the global work queue should be disabled.
  //
  // If gcBlackenPromptly is true, per-P gcWork caches should
  // be flushed immediately and new objects should be allocated black.
  //
  // There is a tension between allocating objects white and
  // allocating them black. If white and the objects die before being
  // marked they can be collected during this GC cycle. On the other
  // hand allocating them black will reduce _GCmarktermination latency
  // since more work is done in the mark phase. This tension is resolved
  // by allocating white until the mark phase is approaching its end and
  // then allocating black for the remainder of the mark phase.
  var gcBlackenPromptly bool
  
  const (
  	_GCoff             = iota // GC not running; sweeping in background, write barrier disabled
  	_GCmark                   // GC marking roots and workbufs: allocate black, write barrier ENABLED
  	_GCmarktermination        // GC mark termination: allocate black, P's help GC, write barrier ENABLED
  )
  
  //go:nosplit
  func setGCPhase(x uint32) {
  	atomic.Store(&gcphase, x)
  	writeBarrier.needed = gcphase == _GCmark || gcphase == _GCmarktermination
  	writeBarrier.enabled = writeBarrier.needed || writeBarrier.cgo
  }
  
  // gcMarkWorkerMode represents the mode that a concurrent mark worker
  // should operate in.
  //
  // Concurrent marking happens through four different mechanisms. One
  // is mutator assists, which happen in response to allocations and are
  // not scheduled. The other three are variations in the per-P mark
  // workers and are distinguished by gcMarkWorkerMode.
  type gcMarkWorkerMode int
  
  const (
  	// gcMarkWorkerDedicatedMode indicates that the P of a mark
  	// worker is dedicated to running that mark worker. The mark
  	// worker should run without preemption.
  	gcMarkWorkerDedicatedMode gcMarkWorkerMode = iota
  
  	// gcMarkWorkerFractionalMode indicates that a P is currently
  	// running the "fractional" mark worker. The fractional worker
  	// is necessary when GOMAXPROCS*gcGoalUtilization is not an
  	// integer. The fractional worker should run until it is
  	// preempted and will be scheduled to pick up the fractional
  	// part of GOMAXPROCS*gcGoalUtilization.
  	gcMarkWorkerFractionalMode
  
  	// gcMarkWorkerIdleMode indicates that a P is running the mark
  	// worker because it has nothing else to do. The idle worker
  	// should run until it is preempted and account its time
  	// against gcController.idleMarkTime.
  	gcMarkWorkerIdleMode
  )
  
  // gcMarkWorkerModeStrings are the strings labels of gcMarkWorkerModes
  // to use in execution traces.
  var gcMarkWorkerModeStrings = [...]string{
  	"GC (dedicated)",
  	"GC (fractional)",
  	"GC (idle)",
  }
  
  // gcController implements the GC pacing controller that determines
  // when to trigger concurrent garbage collection and how much marking
  // work to do in mutator assists and background marking.
  //
  // It uses a feedback control algorithm to adjust the memstats.gc_trigger
  // trigger based on the heap growth and GC CPU utilization each cycle.
  // This algorithm optimizes for heap growth to match GOGC and for CPU
  // utilization between assist and background marking to be 25% of
  // GOMAXPROCS. The high-level design of this algorithm is documented
  // at https://golang.org/s/go15gcpacing.
  //
  // All fields of gcController are used only during a single mark
  // cycle.
  var gcController gcControllerState
  
  type gcControllerState struct {
  	// scanWork is the total scan work performed this cycle. This
  	// is updated atomically during the cycle. Updates occur in
  	// bounded batches, since it is both written and read
  	// throughout the cycle. At the end of the cycle, this is how
  	// much of the retained heap is scannable.
  	//
  	// Currently this is the bytes of heap scanned. For most uses,
  	// this is an opaque unit of work, but for estimation the
  	// definition is important.
  	scanWork int64
  
  	// bgScanCredit is the scan work credit accumulated by the
  	// concurrent background scan. This credit is accumulated by
  	// the background scan and stolen by mutator assists. This is
  	// updated atomically. Updates occur in bounded batches, since
  	// it is both written and read throughout the cycle.
  	bgScanCredit int64
  
  	// assistTime is the nanoseconds spent in mutator assists
  	// during this cycle. This is updated atomically. Updates
  	// occur in bounded batches, since it is both written and read
  	// throughout the cycle.
  	assistTime int64
  
  	// dedicatedMarkTime is the nanoseconds spent in dedicated
  	// mark workers during this cycle. This is updated atomically
  	// at the end of the concurrent mark phase.
  	dedicatedMarkTime int64
  
  	// fractionalMarkTime is the nanoseconds spent in the
  	// fractional mark worker during this cycle. This is updated
  	// atomically throughout the cycle and will be up-to-date if
  	// the fractional mark worker is not currently running.
  	fractionalMarkTime int64
  
  	// idleMarkTime is the nanoseconds spent in idle marking
  	// during this cycle. This is updated atomically throughout
  	// the cycle.
  	idleMarkTime int64
  
  	// markStartTime is the absolute start time in nanoseconds
  	// that assists and background mark workers started.
  	markStartTime int64
  
  	// dedicatedMarkWorkersNeeded is the number of dedicated mark
  	// workers that need to be started. This is computed at the
  	// beginning of each cycle and decremented atomically as
  	// dedicated mark workers get started.
  	dedicatedMarkWorkersNeeded int64
  
  	// assistWorkPerByte is the ratio of scan work to allocated
  	// bytes that should be performed by mutator assists. This is
  	// computed at the beginning of each cycle and updated every
  	// time heap_scan is updated.
  	assistWorkPerByte float64
  
  	// assistBytesPerWork is 1/assistWorkPerByte.
  	assistBytesPerWork float64
  
  	// fractionalUtilizationGoal is the fraction of wall clock
  	// time that should be spent in the fractional mark worker.
  	// For example, if the overall mark utilization goal is 25%
  	// and GOMAXPROCS is 6, one P will be a dedicated mark worker
  	// and this will be set to 0.5 so that 50% of the time some P
  	// is in a fractional mark worker. This is computed at the
  	// beginning of each cycle.
  	fractionalUtilizationGoal float64
  
  	_ [sys.CacheLineSize]byte
  
  	// fractionalMarkWorkersNeeded is the number of fractional
  	// mark workers that need to be started. This is either 0 or
  	// 1. This is potentially updated atomically at every
  	// scheduling point (hence it gets its own cache line).
  	fractionalMarkWorkersNeeded int64
  
  	_ [sys.CacheLineSize]byte
  }
  
  // startCycle resets the GC controller's state and computes estimates
  // for a new GC cycle. The caller must hold worldsema.
  func (c *gcControllerState) startCycle() {
  	c.scanWork = 0
  	c.bgScanCredit = 0
  	c.assistTime = 0
  	c.dedicatedMarkTime = 0
  	c.fractionalMarkTime = 0
  	c.idleMarkTime = 0
  
  	// If this is the first GC cycle or we're operating on a very
  	// small heap, fake heap_marked so it looks like gc_trigger is
  	// the appropriate growth from heap_marked, even though the
  	// real heap_marked may not have a meaningful value (on the
  	// first cycle) or may be much smaller (resulting in a large
  	// error response).
  	if memstats.gc_trigger <= heapminimum {
  		memstats.heap_marked = uint64(float64(memstats.gc_trigger) / (1 + memstats.triggerRatio))
  	}
  
  	// Re-compute the heap goal for this cycle in case something
  	// changed. This is the same calculation we use elsewhere.
  	memstats.next_gc = memstats.heap_marked + memstats.heap_marked*uint64(gcpercent)/100
  	if gcpercent < 0 {
  		memstats.next_gc = ^uint64(0)
  	}
  
  	// Ensure that the heap goal is at least a little larger than
  	// the current live heap size. This may not be the case if GC
  	// start is delayed or if the allocation that pushed heap_live
  	// over gc_trigger is large or if the trigger is really close to
  	// GOGC. Assist is proportional to this distance, so enforce a
  	// minimum distance, even if it means going over the GOGC goal
  	// by a tiny bit.
  	if memstats.next_gc < memstats.heap_live+1024*1024 {
  		memstats.next_gc = memstats.heap_live + 1024*1024
  	}
  
  	// Compute the total mark utilization goal and divide it among
  	// dedicated and fractional workers.
  	totalUtilizationGoal := float64(gomaxprocs) * gcGoalUtilization
  	c.dedicatedMarkWorkersNeeded = int64(totalUtilizationGoal)
  	c.fractionalUtilizationGoal = totalUtilizationGoal - float64(c.dedicatedMarkWorkersNeeded)
  	if c.fractionalUtilizationGoal > 0 {
  		c.fractionalMarkWorkersNeeded = 1
  	} else {
  		c.fractionalMarkWorkersNeeded = 0
  	}
  
  	// Clear per-P state
  	for _, p := range &allp {
  		if p == nil {
  			break
  		}
  		p.gcAssistTime = 0
  	}
  
  	// Compute initial values for controls that are updated
  	// throughout the cycle.
  	c.revise()
  
  	if debug.gcpacertrace > 0 {
  		print("pacer: assist ratio=", c.assistWorkPerByte,
  			" (scan ", memstats.heap_scan>>20, " MB in ",
  			work.initialHeapLive>>20, "->",
  			memstats.next_gc>>20, " MB)",
  			" workers=", c.dedicatedMarkWorkersNeeded,
  			"+", c.fractionalMarkWorkersNeeded, "\n")
  	}
  }
  
  // revise updates the assist ratio during the GC cycle to account for
  // improved estimates. This should be called either under STW or
  // whenever memstats.heap_scan, memstats.heap_live, or
  // memstats.next_gc is updated (with mheap_.lock held).
  //
  // It should only be called when gcBlackenEnabled != 0 (because this
  // is when assists are enabled and the necessary statistics are
  // available).
  func (c *gcControllerState) revise() {
  	// Compute the expected scan work remaining.
  	//
  	// Note that we currently count allocations during GC as both
  	// scannable heap (heap_scan) and scan work completed
  	// (scanWork), so this difference won't be changed by
  	// allocations during GC.
  	//
  	// This particular estimate is a strict upper bound on the
  	// possible remaining scan work for the current heap.
  	// You might consider dividing this by 2 (or by
  	// (100+GOGC)/100) to counter this over-estimation, but
  	// benchmarks show that this has almost no effect on mean
  	// mutator utilization, heap size, or assist time and it
  	// introduces the danger of under-estimating and letting the
  	// mutator outpace the garbage collector.
  	scanWorkExpected := int64(memstats.heap_scan) - c.scanWork
  	if scanWorkExpected < 1000 {
  		// We set a somewhat arbitrary lower bound on
  		// remaining scan work since if we aim a little high,
  		// we can miss by a little.
  		//
  		// We *do* need to enforce that this is at least 1,
  		// since marking is racy and double-scanning objects
  		// may legitimately make the expected scan work
  		// negative.
  		scanWorkExpected = 1000
  	}
  
  	// Compute the heap distance remaining.
  	heapDistance := int64(memstats.next_gc) - int64(atomic.Load64(&memstats.heap_live))
  	if heapDistance <= 0 {
  		// This shouldn't happen, but if it does, avoid
  		// dividing by zero or setting the assist negative.
  		heapDistance = 1
  	}
  
  	// Compute the mutator assist ratio so by the time the mutator
  	// allocates the remaining heap bytes up to next_gc, it will
  	// have done (or stolen) the remaining amount of scan work.
  	c.assistWorkPerByte = float64(scanWorkExpected) / float64(heapDistance)
  	c.assistBytesPerWork = float64(heapDistance) / float64(scanWorkExpected)
  }
  
  // endCycle computes the trigger ratio for the next cycle.
  func (c *gcControllerState) endCycle() float64 {
  	if work.userForced {
  		// Forced GC means this cycle didn't start at the
  		// trigger, so where it finished isn't good
  		// information about how to adjust the trigger.
  		// Just leave it where it is.
  		return memstats.triggerRatio
  	}
  
  	// Proportional response gain for the trigger controller. Must
  	// be in [0, 1]. Lower values smooth out transient effects but
  	// take longer to respond to phase changes. Higher values
  	// react to phase changes quickly, but are more affected by
  	// transient changes. Values near 1 may be unstable.
  	const triggerGain = 0.5
  
  	// Compute next cycle trigger ratio. First, this computes the
  	// "error" for this cycle; that is, how far off the trigger
  	// was from what it should have been, accounting for both heap
  	// growth and GC CPU utilization. We compute the actual heap
  	// growth during this cycle and scale that by how far off from
  	// the goal CPU utilization we were (to estimate the heap
  	// growth if we had the desired CPU utilization). The
  	// difference between this estimate and the GOGC-based goal
  	// heap growth is the error.
  	goalGrowthRatio := float64(gcpercent) / 100
  	actualGrowthRatio := float64(memstats.heap_live)/float64(memstats.heap_marked) - 1
  	assistDuration := nanotime() - c.markStartTime
  
  	// Assume background mark hit its utilization goal.
  	utilization := gcGoalUtilization
  	// Add assist utilization; avoid divide by zero.
  	if assistDuration > 0 {
  		utilization += float64(c.assistTime) / float64(assistDuration*int64(gomaxprocs))
  	}
  
  	triggerError := goalGrowthRatio - memstats.triggerRatio - utilization/gcGoalUtilization*(actualGrowthRatio-memstats.triggerRatio)
  
  	// Finally, we adjust the trigger for next time by this error,
  	// damped by the proportional gain.
  	triggerRatio := memstats.triggerRatio + triggerGain*triggerError
  
  	if debug.gcpacertrace > 0 {
  		// Print controller state in terms of the design
  		// document.
  		H_m_prev := memstats.heap_marked
  		h_t := memstats.triggerRatio
  		H_T := memstats.gc_trigger
  		h_a := actualGrowthRatio
  		H_a := memstats.heap_live
  		h_g := goalGrowthRatio
  		H_g := int64(float64(H_m_prev) * (1 + h_g))
  		u_a := utilization
  		u_g := gcGoalUtilization
  		W_a := c.scanWork
  		print("pacer: H_m_prev=", H_m_prev,
  			" h_t=", h_t, " H_T=", H_T,
  			" h_a=", h_a, " H_a=", H_a,
  			" h_g=", h_g, " H_g=", H_g,
  			" u_a=", u_a, " u_g=", u_g,
  			" W_a=", W_a,
  			" goalΔ=", goalGrowthRatio-h_t,
  			" actualΔ=", h_a-h_t,
  			" u_a/u_g=", u_a/u_g,
  			"\n")
  	}
  
  	return triggerRatio
  }
  
  // enlistWorker encourages another dedicated mark worker to start on
  // another P if there are spare worker slots. It is used by putfull
  // when more work is made available.
  //
  //go:nowritebarrier
  func (c *gcControllerState) enlistWorker() {
  	// If there are idle Ps, wake one so it will run an idle worker.
  	// NOTE: This is suspected of causing deadlocks. See golang.org/issue/19112.
  	//
  	//	if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
  	//		wakep()
  	//		return
  	//	}
  
  	// There are no idle Ps. If we need more dedicated workers,
  	// try to preempt a running P so it will switch to a worker.
  	if c.dedicatedMarkWorkersNeeded <= 0 {
  		return
  	}
  	// Pick a random other P to preempt.
  	if gomaxprocs <= 1 {
  		return
  	}
  	gp := getg()
  	if gp == nil || gp.m == nil || gp.m.p == 0 {
  		return
  	}
  	myID := gp.m.p.ptr().id
  	for tries := 0; tries < 5; tries++ {
  		id := int32(fastrandn(uint32(gomaxprocs - 1)))
  		if id >= myID {
  			id++
  		}
  		p := allp[id]
  		if p.status != _Prunning {
  			continue
  		}
  		if preemptone(p) {
  			return
  		}
  	}
  }
  
  // findRunnableGCWorker returns the background mark worker for _p_ if it
  // should be run. This must only be called when gcBlackenEnabled != 0.
  func (c *gcControllerState) findRunnableGCWorker(_p_ *p) *g {
  	if gcBlackenEnabled == 0 {
  		throw("gcControllerState.findRunnable: blackening not enabled")
  	}
  	if _p_.gcBgMarkWorker == 0 {
  		// The mark worker associated with this P is blocked
  		// performing a mark transition. We can't run it
  		// because it may be on some other run or wait queue.
  		return nil
  	}
  
  	if !gcMarkWorkAvailable(_p_) {
  		// No work to be done right now. This can happen at
  		// the end of the mark phase when there are still
  		// assists tapering off. Don't bother running a worker
  		// now because it'll just return immediately.
  		return nil
  	}
  
  	decIfPositive := func(ptr *int64) bool {
  		if *ptr > 0 {
  			if atomic.Xaddint64(ptr, -1) >= 0 {
  				return true
  			}
  			// We lost a race
  			atomic.Xaddint64(ptr, +1)
  		}
  		return false
  	}
  
  	if decIfPositive(&c.dedicatedMarkWorkersNeeded) {
  		// This P is now dedicated to marking until the end of
  		// the concurrent mark phase.
  		_p_.gcMarkWorkerMode = gcMarkWorkerDedicatedMode
  	} else {
  		if !decIfPositive(&c.fractionalMarkWorkersNeeded) {
  			// No more workers are need right now.
  			return nil
  		}
  
  		// This P has picked the token for the fractional worker.
  		// Is the GC currently under or at the utilization goal?
  		// If so, do more work.
  		//
  		// We used to check whether doing one time slice of work
  		// would remain under the utilization goal, but that has the
  		// effect of delaying work until the mutator has run for
  		// enough time slices to pay for the work. During those time
  		// slices, write barriers are enabled, so the mutator is running slower.
  		// Now instead we do the work whenever we're under or at the
  		// utilization work and pay for it by letting the mutator run later.
  		// This doesn't change the overall utilization averages, but it
  		// front loads the GC work so that the GC finishes earlier and
  		// write barriers can be turned off sooner, effectively giving
  		// the mutator a faster machine.
  		//
  		// The old, slower behavior can be restored by setting
  		//	gcForcePreemptNS = forcePreemptNS.
  		const gcForcePreemptNS = 0
  
  		// TODO(austin): We could fast path this and basically
  		// eliminate contention on c.fractionalMarkWorkersNeeded by
  		// precomputing the minimum time at which it's worth
  		// next scheduling the fractional worker. Then Ps
  		// don't have to fight in the window where we've
  		// passed that deadline and no one has started the
  		// worker yet.
  		//
  		// TODO(austin): Shorter preemption interval for mark
  		// worker to improve fairness and give this
  		// finer-grained control over schedule?
  		now := nanotime() - gcController.markStartTime
  		then := now + gcForcePreemptNS
  		timeUsed := c.fractionalMarkTime + gcForcePreemptNS
  		if then > 0 && float64(timeUsed)/float64(then) > c.fractionalUtilizationGoal {
  			// Nope, we'd overshoot the utilization goal
  			atomic.Xaddint64(&c.fractionalMarkWorkersNeeded, +1)
  			return nil
  		}
  		_p_.gcMarkWorkerMode = gcMarkWorkerFractionalMode
  	}
  
  	// Run the background mark worker
  	gp := _p_.gcBgMarkWorker.ptr()
  	casgstatus(gp, _Gwaiting, _Grunnable)
  	if trace.enabled {
  		traceGoUnpark(gp, 0)
  	}
  	return gp
  }
  
  // gcSetTriggerRatio sets the trigger ratio and updates everything
  // derived from it: the absolute trigger, the heap goal, mark pacing,
  // and sweep pacing.
  //
  // This can be called any time. If GC is the in the middle of a
  // concurrent phase, it will adjust the pacing of that phase.
  //
  // This depends on gcpercent, memstats.heap_marked, and
  // memstats.heap_live. These must be up to date.
  //
  // mheap_.lock must be held or the world must be stopped.
  func gcSetTriggerRatio(triggerRatio float64) {
  	// Set the trigger ratio, capped to reasonable bounds.
  	if triggerRatio < 0 {
  		// This can happen if the mutator is allocating very
  		// quickly or the GC is scanning very slowly.
  		triggerRatio = 0
  	} else if gcpercent >= 0 {
  		// Ensure there's always a little margin so that the
  		// mutator assist ratio isn't infinity.
  		maxTriggerRatio := 0.95 * float64(gcpercent) / 100
  		if triggerRatio > maxTriggerRatio {
  			triggerRatio = maxTriggerRatio
  		}
  	}
  	memstats.triggerRatio = triggerRatio
  
  	// Compute the absolute GC trigger from the trigger ratio.
  	//
  	// We trigger the next GC cycle when the allocated heap has
  	// grown by the trigger ratio over the marked heap size.
  	trigger := ^uint64(0)
  	if gcpercent >= 0 {
  		trigger = uint64(float64(memstats.heap_marked) * (1 + triggerRatio))
  		// Don't trigger below the minimum heap size.
  		minTrigger := heapminimum
  		if !gosweepdone() {
  			// Concurrent sweep happens in the heap growth
  			// from heap_live to gc_trigger, so ensure
  			// that concurrent sweep has some heap growth
  			// in which to perform sweeping before we
  			// start the next GC cycle.
  			sweepMin := atomic.Load64(&memstats.heap_live) + sweepMinHeapDistance*uint64(gcpercent)/100
  			if sweepMin > minTrigger {
  				minTrigger = sweepMin
  			}
  		}
  		if trigger < minTrigger {
  			trigger = minTrigger
  		}
  		if int64(trigger) < 0 {
  			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")
  			throw("gc_trigger underflow")
  		}
  	}
  	memstats.gc_trigger = trigger
  
  	// Compute the next GC goal, which is when the allocated heap
  	// has grown by GOGC/100 over the heap marked by the last
  	// cycle.
  	goal := ^uint64(0)
  	if gcpercent >= 0 {
  		goal = memstats.heap_marked + memstats.heap_marked*uint64(gcpercent)/100
  		if goal < trigger {
  			// The trigger ratio is always less than GOGC/100, but
  			// other bounds on the trigger may have raised it.
  			// Push up the goal, too.
  			goal = trigger
  		}
  	}
  	memstats.next_gc = goal
  	if trace.enabled {
  		traceNextGC()
  	}
  
  	// Update mark pacing.
  	if gcphase != _GCoff {
  		gcController.revise()
  	}
  
  	// Update sweep pacing.
  	if gosweepdone() {
  		mheap_.sweepPagesPerByte = 0
  	} else {
  		// Concurrent sweep needs to sweep all of the in-use
  		// pages by the time the allocated heap reaches the GC
  		// trigger. Compute the ratio of in-use pages to sweep
  		// per byte allocated, accounting for the fact that
  		// some might already be swept.
  		heapLiveBasis := atomic.Load64(&memstats.heap_live)
  		heapDistance := int64(trigger) - int64(heapLiveBasis)
  		// Add a little margin so rounding errors and
  		// concurrent sweep are less likely to leave pages
  		// unswept when GC starts.
  		heapDistance -= 1024 * 1024
  		if heapDistance < _PageSize {
  			// Avoid setting the sweep ratio extremely high
  			heapDistance = _PageSize
  		}
  		pagesSwept := atomic.Load64(&mheap_.pagesSwept)
  		sweepDistancePages := int64(mheap_.pagesInUse) - int64(pagesSwept)
  		if sweepDistancePages <= 0 {
  			mheap_.sweepPagesPerByte = 0
  		} else {
  			mheap_.sweepPagesPerByte = float64(sweepDistancePages) / float64(heapDistance)
  			mheap_.sweepHeapLiveBasis = heapLiveBasis
  			// Write pagesSweptBasis last, since this
  			// signals concurrent sweeps to recompute
  			// their debt.
  			atomic.Store64(&mheap_.pagesSweptBasis, pagesSwept)
  		}
  	}
  }
  
  // gcGoalUtilization is the goal CPU utilization for background
  // marking as a fraction of GOMAXPROCS.
  const gcGoalUtilization = 0.25
  
  // gcCreditSlack is the amount of scan work credit that can can
  // accumulate locally before updating gcController.scanWork and,
  // optionally, gcController.bgScanCredit. Lower values give a more
  // accurate assist ratio and make it more likely that assists will
  // successfully steal background credit. Higher values reduce memory
  // contention.
  const gcCreditSlack = 2000
  
  // gcAssistTimeSlack is the nanoseconds of mutator assist time that
  // can accumulate on a P before updating gcController.assistTime.
  const gcAssistTimeSlack = 5000
  
  // gcOverAssistWork determines how many extra units of scan work a GC
  // assist does when an assist happens. This amortizes the cost of an
  // assist by pre-paying for this many bytes of future allocations.
  const gcOverAssistWork = 64 << 10
  
  var work struct {
  	full  lfstack                  // lock-free list of full blocks workbuf
  	empty lfstack                  // lock-free list of empty blocks workbuf
  	pad0  [sys.CacheLineSize]uint8 // prevents false-sharing between full/empty and nproc/nwait
  
  	wbufSpans struct {
  		lock mutex
  		// free is a list of spans dedicated to workbufs, but
  		// that don't currently contain any workbufs.
  		free mSpanList
  		// busy is a list of all spans containing workbufs on
  		// one of the workbuf lists.
  		busy mSpanList
  	}
  
  	// Restore 64-bit alignment on 32-bit.
  	_ uint32
  
  	// bytesMarked is the number of bytes marked this cycle. This
  	// includes bytes blackened in scanned objects, noscan objects
  	// that go straight to black, and permagrey objects scanned by
  	// markroot during the concurrent scan phase. This is updated
  	// atomically during the cycle. Updates may be batched
  	// arbitrarily, since the value is only read at the end of the
  	// cycle.
  	//
  	// Because of benign races during marking, this number may not
  	// be the exact number of marked bytes, but it should be very
  	// close.
  	//
  	// Put this field here because it needs 64-bit atomic access
  	// (and thus 8-byte alignment even on 32-bit architectures).
  	bytesMarked uint64
  
  	markrootNext uint32 // next markroot job
  	markrootJobs uint32 // number of markroot jobs
  
  	nproc   uint32
  	tstart  int64
  	nwait   uint32
  	ndone   uint32
  	alldone note
  
  	// helperDrainBlock indicates that GC mark termination helpers
  	// should pass gcDrainBlock to gcDrain to block in the
  	// getfull() barrier. Otherwise, they should pass gcDrainNoBlock.
  	//
  	// TODO: This is a temporary fallback to work around races
  	// that cause early mark termination.
  	helperDrainBlock bool
  
  	// Number of roots of various root types. Set by gcMarkRootPrepare.
  	nFlushCacheRoots                               int
  	nDataRoots, nBSSRoots, nSpanRoots, nStackRoots int
  
  	// markrootDone indicates that roots have been marked at least
  	// once during the current GC cycle. This is checked by root
  	// marking operations that have to happen only during the
  	// first root marking pass, whether that's during the
  	// concurrent mark phase in current GC or mark termination in
  	// STW GC.
  	markrootDone bool
  
  	// Each type of GC state transition is protected by a lock.
  	// Since multiple threads can simultaneously detect the state
  	// transition condition, any thread that detects a transition
  	// condition must acquire the appropriate transition lock,
  	// re-check the transition condition and return if it no
  	// longer holds or perform the transition if it does.
  	// Likewise, any transition must invalidate the transition
  	// condition before releasing the lock. This ensures that each
  	// transition is performed by exactly one thread and threads
  	// that need the transition to happen block until it has
  	// happened.
  	//
  	// startSema protects the transition from "off" to mark or
  	// mark termination.
  	startSema uint32
  	// markDoneSema protects transitions from mark 1 to mark 2 and
  	// from mark 2 to mark termination.
  	markDoneSema uint32
  
  	bgMarkReady note   // signal background mark worker has started
  	bgMarkDone  uint32 // cas to 1 when at a background mark completion point
  	// Background mark completion signaling
  
  	// mode is the concurrency mode of the current GC cycle.
  	mode gcMode
  
  	// userForced indicates the current GC cycle was forced by an
  	// explicit user call.
  	userForced bool
  
  	// totaltime is the CPU nanoseconds spent in GC since the
  	// program started if debug.gctrace > 0.
  	totaltime int64
  
  	// initialHeapLive is the value of memstats.heap_live at the
  	// beginning of this GC cycle.
  	initialHeapLive uint64
  
  	// assistQueue is a queue of assists that are blocked because
  	// there was neither enough credit to steal or enough work to
  	// do.
  	assistQueue struct {
  		lock       mutex
  		head, tail guintptr
  	}
  
  	// sweepWaiters is a list of blocked goroutines to wake when
  	// we transition from mark termination to sweep.
  	sweepWaiters struct {
  		lock mutex
  		head guintptr
  	}
  
  	// cycles is the number of completed GC cycles, where a GC
  	// cycle is sweep termination, mark, mark termination, and
  	// sweep. This differs from memstats.numgc, which is
  	// incremented at mark termination.
  	cycles uint32
  
  	// Timing/utilization stats for this cycle.
  	stwprocs, maxprocs                 int32
  	tSweepTerm, tMark, tMarkTerm, tEnd int64 // nanotime() of phase start
  
  	pauseNS    int64 // total STW time this cycle
  	pauseStart int64 // nanotime() of last STW
  
  	// debug.gctrace heap sizes for this cycle.
  	heap0, heap1, heap2, heapGoal uint64
  }
  
  // GC runs a garbage collection and blocks the caller until the
  // garbage collection is complete. It may also block the entire
  // program.
  func GC() {
  	// We consider a cycle to be: sweep termination, mark, mark
  	// termination, and sweep. This function shouldn't return
  	// until a full cycle has been completed, from beginning to
  	// end. Hence, we always want to finish up the current cycle
  	// and start a new one. That means:
  	//
  	// 1. In sweep termination, mark, or mark termination of cycle
  	// N, wait until mark termination N completes and transitions
  	// to sweep N.
  	//
  	// 2. In sweep N, help with sweep N.
  	//
  	// At this point we can begin a full cycle N+1.
  	//
  	// 3. Trigger cycle N+1 by starting sweep termination N+1.
  	//
  	// 4. Wait for mark termination N+1 to complete.
  	//
  	// 5. Help with sweep N+1 until it's done.
  	//
  	// This all has to be written to deal with the fact that the
  	// GC may move ahead on its own. For example, when we block
  	// until mark termination N, we may wake up in cycle N+2.
  
  	gp := getg()
  
  	// Prevent the GC phase or cycle count from changing.
  	lock(&work.sweepWaiters.lock)
  	n := atomic.Load(&work.cycles)
  	if gcphase == _GCmark {
  		// Wait until sweep termination, mark, and mark
  		// termination of cycle N complete.
  		gp.schedlink = work.sweepWaiters.head
  		work.sweepWaiters.head.set(gp)
  		goparkunlock(&work.sweepWaiters.lock, "wait for GC cycle", traceEvGoBlock, 1)
  	} else {
  		// We're in sweep N already.
  		unlock(&work.sweepWaiters.lock)
  	}
  
  	// We're now in sweep N or later. Trigger GC cycle N+1, which
  	// will first finish sweep N if necessary and then enter sweep
  	// termination N+1.
  	gcStart(gcBackgroundMode, gcTrigger{kind: gcTriggerCycle, n: n + 1})
  
  	// Wait for mark termination N+1 to complete.
  	lock(&work.sweepWaiters.lock)
  	if gcphase == _GCmark && atomic.Load(&work.cycles) == n+1 {
  		gp.schedlink = work.sweepWaiters.head
  		work.sweepWaiters.head.set(gp)
  		goparkunlock(&work.sweepWaiters.lock, "wait for GC cycle", traceEvGoBlock, 1)
  	} else {
  		unlock(&work.sweepWaiters.lock)
  	}
  
  	// Finish sweep N+1 before returning. We do this both to
  	// complete the cycle and because runtime.GC() is often used
  	// as part of tests and benchmarks to get the system into a
  	// relatively stable and isolated state.
  	for atomic.Load(&work.cycles) == n+1 && gosweepone() != ^uintptr(0) {
  		sweep.nbgsweep++
  		Gosched()
  	}
  
  	// Callers may assume that the heap profile reflects the
  	// just-completed cycle when this returns (historically this
  	// happened because this was a STW GC), but right now the
  	// profile still reflects mark termination N, not N+1.
  	//
  	// As soon as all of the sweep frees from cycle N+1 are done,
  	// we can go ahead and publish the heap profile.
  	//
  	// First, wait for sweeping to finish. (We know there are no
  	// more spans on the sweep queue, but we may be concurrently
  	// sweeping spans, so we have to wait.)
  	for atomic.Load(&work.cycles) == n+1 && atomic.Load(&mheap_.sweepers) != 0 {
  		Gosched()
  	}
  
  	// Now we're really done with sweeping, so we can publish the
  	// stable heap profile. Only do this if we haven't already hit
  	// another mark termination.
  	mp := acquirem()
  	cycle := atomic.Load(&work.cycles)
  	if cycle == n+1 || (gcphase == _GCmark && cycle == n+2) {
  		mProf_PostSweep()
  	}
  	releasem(mp)
  }
  
  // gcMode indicates how concurrent a GC cycle should be.
  type gcMode int
  
  const (
  	gcBackgroundMode gcMode = iota // concurrent GC and sweep
  	gcForceMode                    // stop-the-world GC now, concurrent sweep
  	gcForceBlockMode               // stop-the-world GC now and STW sweep (forced by user)
  )
  
  // A gcTrigger is a predicate for starting a GC cycle. Specifically,
  // it is an exit condition for the _GCoff phase.
  type gcTrigger struct {
  	kind gcTriggerKind
  	now  int64  // gcTriggerTime: current time
  	n    uint32 // gcTriggerCycle: cycle number to start
  }
  
  type gcTriggerKind int
  
  const (
  	// gcTriggerAlways indicates that a cycle should be started
  	// unconditionally, even if GOGC is off or we're in a cycle
  	// right now. This cannot be consolidated with other cycles.
  	gcTriggerAlways gcTriggerKind = iota
  
  	// gcTriggerHeap indicates that a cycle should be started when
  	// the heap size reaches the trigger heap size computed by the
  	// controller.
  	gcTriggerHeap
  
  	// gcTriggerTime indicates that a cycle should be started when
  	// it's been more than forcegcperiod nanoseconds since the
  	// previous GC cycle.
  	gcTriggerTime
  
  	// gcTriggerCycle indicates that a cycle should be started if
  	// we have not yet started cycle number gcTrigger.n (relative
  	// to work.cycles).
  	gcTriggerCycle
  )
  
  // test returns true if the trigger condition is satisfied, meaning
  // that the exit condition for the _GCoff phase has been met. The exit
  // condition should be tested when allocating.
  func (t gcTrigger) test() bool {
  	if !memstats.enablegc || panicking != 0 {
  		return false
  	}
  	if t.kind == gcTriggerAlways {
  		return true
  	}
  	if gcphase != _GCoff || gcpercent < 0 {
  		return false
  	}
  	switch t.kind {
  	case gcTriggerHeap:
  		// Non-atomic access to heap_live for performance. If
  		// we are going to trigger on this, this thread just
  		// atomically wrote heap_live anyway and we'll see our
  		// own write.
  		return memstats.heap_live >= memstats.gc_trigger
  	case gcTriggerTime:
  		lastgc := int64(atomic.Load64(&memstats.last_gc_nanotime))
  		return lastgc != 0 && t.now-lastgc > forcegcperiod
  	case gcTriggerCycle:
  		// t.n > work.cycles, but accounting for wraparound.
  		return int32(t.n-work.cycles) > 0
  	}
  	return true
  }
  
  // gcStart transitions the GC from _GCoff to _GCmark (if
  // !mode.stwMark) or _GCmarktermination (if mode.stwMark) by
  // performing sweep termination and GC initialization.
  //
  // This may return without performing this transition in some cases,
  // such as when called on a system stack or with locks held.
  func gcStart(mode gcMode, trigger gcTrigger) {
  	// Since this is called from malloc and malloc is called in
  	// the guts of a number of libraries that might be holding
  	// locks, don't attempt to start GC in non-preemptible or
  	// potentially unstable situations.
  	mp := acquirem()
  	if gp := getg(); gp == mp.g0 || mp.locks > 1 || mp.preemptoff != "" {
  		releasem(mp)
  		return
  	}
  	releasem(mp)
  	mp = nil
  
  	// Pick up the remaining unswept/not being swept spans concurrently
  	//
  	// This shouldn't happen if we're being invoked in background
  	// mode since proportional sweep should have just finished
  	// sweeping everything, but rounding errors, etc, may leave a
  	// few spans unswept. In forced mode, this is necessary since
  	// GC can be forced at any point in the sweeping cycle.
  	//
  	// We check the transition condition continuously here in case
  	// this G gets delayed in to the next GC cycle.
  	for trigger.test() && gosweepone() != ^uintptr(0) {
  		sweep.nbgsweep++
  	}
  
  	// Perform GC initialization and the sweep termination
  	// transition.
  	semacquire(&work.startSema)
  	// Re-check transition condition under transition lock.
  	if !trigger.test() {
  		semrelease(&work.startSema)
  		return
  	}
  
  	// For stats, check if this GC was forced by the user.
  	work.userForced = trigger.kind == gcTriggerAlways || trigger.kind == gcTriggerCycle
  
  	// In gcstoptheworld debug mode, upgrade the mode accordingly.
  	// We do this after re-checking the transition condition so
  	// that multiple goroutines that detect the heap trigger don't
  	// start multiple STW GCs.
  	if mode == gcBackgroundMode {
  		if debug.gcstoptheworld == 1 {
  			mode = gcForceMode
  		} else if debug.gcstoptheworld == 2 {
  			mode = gcForceBlockMode
  		}
  	}
  
  	// Ok, we're doing it!  Stop everybody else
  	semacquire(&worldsema)
  
  	if trace.enabled {
  		traceGCStart()
  	}
  
  	if mode == gcBackgroundMode {
  		gcBgMarkStartWorkers()
  	}
  
  	gcResetMarkState()
  
  	work.stwprocs, work.maxprocs = gcprocs(), gomaxprocs
  	work.heap0 = atomic.Load64(&memstats.heap_live)
  	work.pauseNS = 0
  	work.mode = mode
  
  	now := nanotime()
  	work.tSweepTerm = now
  	work.pauseStart = now
  	systemstack(stopTheWorldWithSema)
  	// Finish sweep before we start concurrent scan.
  	systemstack(func() {
  		finishsweep_m()
  	})
  	// clearpools before we start the GC. If we wait they memory will not be
  	// reclaimed until the next GC cycle.
  	clearpools()
  
  	work.cycles++
  	if mode == gcBackgroundMode { // Do as much work concurrently as possible
  		gcController.startCycle()
  		work.heapGoal = memstats.next_gc
  
  		// Enter concurrent mark phase and enable
  		// write barriers.
  		//
  		// Because the world is stopped, all Ps will
  		// observe that write barriers are enabled by
  		// the time we start the world and begin
  		// scanning.
  		//
  		// Write barriers must be enabled before assists are
  		// enabled because they must be enabled before
  		// any non-leaf heap objects are marked. Since
  		// allocations are blocked until assists can
  		// happen, we want enable assists as early as
  		// possible.
  		setGCPhase(_GCmark)
  
  		gcBgMarkPrepare() // Must happen before assist enable.
  		gcMarkRootPrepare()
  
  		// Mark all active tinyalloc blocks. Since we're
  		// allocating from these, they need to be black like
  		// other allocations. The alternative is to blacken
  		// the tiny block on every allocation from it, which
  		// would slow down the tiny allocator.
  		gcMarkTinyAllocs()
  
  		// At this point all Ps have enabled the write
  		// barrier, thus maintaining the no white to
  		// black invariant. Enable mutator assists to
  		// put back-pressure on fast allocating
  		// mutators.
  		atomic.Store(&gcBlackenEnabled, 1)
  
  		// Assists and workers can start the moment we start
  		// the world.
  		gcController.markStartTime = now
  
  		// Concurrent mark.
  		systemstack(startTheWorldWithSema)
  		now = nanotime()
  		work.pauseNS += now - work.pauseStart
  		work.tMark = now
  	} else {
  		t := nanotime()
  		work.tMark, work.tMarkTerm = t, t
  		work.heapGoal = work.heap0
  
  		// Perform mark termination. This will restart the world.
  		gcMarkTermination(memstats.triggerRatio)
  	}
  
  	semrelease(&work.startSema)
  }
  
  // gcMarkDone transitions the GC from mark 1 to mark 2 and from mark 2
  // to mark termination.
  //
  // This should be called when all mark work has been drained. In mark
  // 1, this includes all root marking jobs, global work buffers, and
  // active work buffers in assists and background workers; however,
  // work may still be cached in per-P work buffers. In mark 2, per-P
  // caches are disabled.
  //
  // The calling context must be preemptible.
  //
  // Note that it is explicitly okay to have write barriers in this
  // function because completion of concurrent mark is best-effort
  // anyway. Any work created by write barriers here will be cleaned up
  // by mark termination.
  func gcMarkDone() {
  top:
  	semacquire(&work.markDoneSema)
  
  	// Re-check transition condition under transition lock.
  	if !(gcphase == _GCmark && work.nwait == work.nproc && !gcMarkWorkAvailable(nil)) {
  		semrelease(&work.markDoneSema)
  		return
  	}
  
  	// Disallow starting new workers so that any remaining workers
  	// in the current mark phase will drain out.
  	//
  	// TODO(austin): Should dedicated workers keep an eye on this
  	// and exit gcDrain promptly?
  	atomic.Xaddint64(&gcController.dedicatedMarkWorkersNeeded, -0xffffffff)
  	atomic.Xaddint64(&gcController.fractionalMarkWorkersNeeded, -0xffffffff)
  
  	if !gcBlackenPromptly {
  		// Transition from mark 1 to mark 2.
  		//
  		// The global work list is empty, but there can still be work
  		// sitting in the per-P work caches.
  		// Flush and disable work caches.
  
  		// Disallow caching workbufs and indicate that we're in mark 2.
  		gcBlackenPromptly = true
  
  		// Prevent completion of mark 2 until we've flushed
  		// cached workbufs.
  		atomic.Xadd(&work.nwait, -1)
  
  		// GC is set up for mark 2. Let Gs blocked on the
  		// transition lock go while we flush caches.
  		semrelease(&work.markDoneSema)
  
  		systemstack(func() {
  			// Flush all currently cached workbufs and
  			// ensure all Ps see gcBlackenPromptly. This
  			// also blocks until any remaining mark 1
  			// workers have exited their loop so we can
  			// start new mark 2 workers.
  			forEachP(func(_p_ *p) {
  				_p_.gcw.dispose()
  			})
  		})
  
  		// Check that roots are marked. We should be able to
  		// do this before the forEachP, but based on issue
  		// #16083 there may be a (harmless) race where we can
  		// enter mark 2 while some workers are still scanning
  		// stacks. The forEachP ensures these scans are done.
  		//
  		// TODO(austin): Figure out the race and fix this
  		// properly.
  		gcMarkRootCheck()
  
  		// Now we can start up mark 2 workers.
  		atomic.Xaddint64(&gcController.dedicatedMarkWorkersNeeded, 0xffffffff)
  		atomic.Xaddint64(&gcController.fractionalMarkWorkersNeeded, 0xffffffff)
  
  		incnwait := atomic.Xadd(&work.nwait, +1)
  		if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
  			// This loop will make progress because
  			// gcBlackenPromptly is now true, so it won't
  			// take this same "if" branch.
  			goto top
  		}
  	} else {
  		// Transition to mark termination.
  		now := nanotime()
  		work.tMarkTerm = now
  		work.pauseStart = now
  		getg().m.preemptoff = "gcing"
  		systemstack(stopTheWorldWithSema)
  		// The gcphase is _GCmark, it will transition to _GCmarktermination
  		// below. The important thing is that the wb remains active until
  		// all marking is complete. This includes writes made by the GC.
  
  		// Record that one root marking pass has completed.
  		work.markrootDone = true
  
  		// Disable assists and background workers. We must do
  		// this before waking blocked assists.
  		atomic.Store(&gcBlackenEnabled, 0)
  
  		// Wake all blocked assists. These will run when we
  		// start the world again.
  		gcWakeAllAssists()
  
  		// Likewise, release the transition lock. Blocked
  		// workers and assists will run when we start the
  		// world again.
  		semrelease(&work.markDoneSema)
  
  		// endCycle depends on all gcWork cache stats being
  		// flushed. This is ensured by mark 2.
  		nextTriggerRatio := gcController.endCycle()
  
  		// Perform mark termination. This will restart the world.
  		gcMarkTermination(nextTriggerRatio)
  	}
  }
  
  func gcMarkTermination(nextTriggerRatio float64) {
  	// World is stopped.
  	// Start marktermination which includes enabling the write barrier.
  	atomic.Store(&gcBlackenEnabled, 0)
  	gcBlackenPromptly = false
  	setGCPhase(_GCmarktermination)
  
  	work.heap1 = memstats.heap_live
  	startTime := nanotime()
  
  	mp := acquirem()
  	mp.preemptoff = "gcing"
  	_g_ := getg()
  	_g_.m.traceback = 2
  	gp := _g_.m.curg
  	casgstatus(gp, _Grunning, _Gwaiting)
  	gp.waitreason = "garbage collection"
  
  	// Run gc on the g0 stack. We do this so that the g stack
  	// we're currently running on will no longer change. Cuts
  	// the root set down a bit (g0 stacks are not scanned, and
  	// we don't need to scan gc's internal state).  We also
  	// need to switch to g0 so we can shrink the stack.
  	systemstack(func() {
  		gcMark(startTime)
  		// Must return immediately.
  		// The outer function's stack may have moved
  		// during gcMark (it shrinks stacks, including the
  		// outer function's stack), so we must not refer
  		// to any of its variables. Return back to the
  		// non-system stack to pick up the new addresses
  		// before continuing.
  	})
  
  	systemstack(func() {
  		work.heap2 = work.bytesMarked
  		if debug.gccheckmark > 0 {
  			// Run a full stop-the-world mark using checkmark bits,
  			// to check that we didn't forget to mark anything during
  			// the concurrent mark process.
  			gcResetMarkState()
  			initCheckmarks()
  			gcMark(startTime)
  			clearCheckmarks()
  		}
  
  		// marking is complete so we can turn the write barrier off
  		setGCPhase(_GCoff)
  		gcSweep(work.mode)
  
  		if debug.gctrace > 1 {
  			startTime = nanotime()
  			// The g stacks have been scanned so
  			// they have gcscanvalid==true and gcworkdone==true.
  			// Reset these so that all stacks will be rescanned.
  			gcResetMarkState()
  			finishsweep_m()
  
  			// Still in STW but gcphase is _GCoff, reset to _GCmarktermination
  			// At this point all objects will be found during the gcMark which
  			// does a complete STW mark and object scan.
  			setGCPhase(_GCmarktermination)
  			gcMark(startTime)
  			setGCPhase(_GCoff) // marking is done, turn off wb.
  			gcSweep(work.mode)
  		}
  	})
  
  	_g_.m.traceback = 0
  	casgstatus(gp, _Gwaiting, _Grunning)
  
  	if trace.enabled {
  		traceGCDone()
  	}
  
  	// all done
  	mp.preemptoff = ""
  
  	if gcphase != _GCoff {
  		throw("gc done but gcphase != _GCoff")
  	}
  
  	// Update GC trigger and pacing for the next cycle.
  	gcSetTriggerRatio(nextTriggerRatio)
  
  	// Update timing memstats
  	now := nanotime()
  	sec, nsec, _ := time_now()
  	unixNow := sec*1e9 + int64(nsec)
  	work.pauseNS += now - work.pauseStart
  	work.tEnd = now
  	atomic.Store64(&memstats.last_gc_unix, uint64(unixNow)) // must be Unix time to make sense to user
  	atomic.Store64(&memstats.last_gc_nanotime, uint64(now)) // monotonic time for us
  	memstats.pause_ns[memstats.numgc%uint32(len(memstats.pause_ns))] = uint64(work.pauseNS)
  	memstats.pause_end[memstats.numgc%uint32(len(memstats.pause_end))] = uint64(unixNow)
  	memstats.pause_total_ns += uint64(work.pauseNS)
  
  	// Update work.totaltime.
  	sweepTermCpu := int64(work.stwprocs) * (work.tMark - work.tSweepTerm)
  	// We report idle marking time below, but omit it from the
  	// overall utilization here since it's "free".
  	markCpu := gcController.assistTime + gcController.dedicatedMarkTime + gcController.fractionalMarkTime
  	markTermCpu := int64(work.stwprocs) * (work.tEnd - work.tMarkTerm)
  	cycleCpu := sweepTermCpu + markCpu + markTermCpu
  	work.totaltime += cycleCpu
  
  	// Compute overall GC CPU utilization.
  	totalCpu := sched.totaltime + (now-sched.procresizetime)*int64(gomaxprocs)
  	memstats.gc_cpu_fraction = float64(work.totaltime) / float64(totalCpu)
  
  	// Reset sweep state.
  	sweep.nbgsweep = 0
  	sweep.npausesweep = 0
  
  	if work.userForced {
  		memstats.numforcedgc++
  	}
  
  	// Bump GC cycle count and wake goroutines waiting on sweep.
  	lock(&work.sweepWaiters.lock)
  	memstats.numgc++
  	injectglist(work.sweepWaiters.head.ptr())
  	work.sweepWaiters.head = 0
  	unlock(&work.sweepWaiters.lock)
  
  	// Finish the current heap profiling cycle and start a new
  	// heap profiling cycle. We do this before starting the world
  	// so events don't leak into the wrong cycle.
  	mProf_NextCycle()
  
  	systemstack(startTheWorldWithSema)
  
  	// Flush the heap profile so we can start a new cycle next GC.
  	// This is relatively expensive, so we don't do it with the
  	// world stopped.
  	mProf_Flush()
  
  	// Prepare workbufs for freeing by the sweeper. We do this
  	// asynchronously because it can take non-trivial time.
  	prepareFreeWorkbufs()
  
  	// Free stack spans. This must be done between GC cycles.
  	systemstack(freeStackSpans)
  
  	// Print gctrace before dropping worldsema. As soon as we drop
  	// worldsema another cycle could start and smash the stats
  	// we're trying to print.
  	if debug.gctrace > 0 {
  		util := int(memstats.gc_cpu_fraction * 100)
  
  		var sbuf [24]byte
  		printlock()
  		print("gc ", memstats.numgc,
  			" @", string(itoaDiv(sbuf[:], uint64(work.tSweepTerm-runtimeInitTime)/1e6, 3)), "s ",
  			util, "%: ")
  		prev := work.tSweepTerm
  		for i, ns := range []int64{work.tMark, work.tMarkTerm, work.tEnd} {
  			if i != 0 {
  				print("+")
  			}
  			print(string(fmtNSAsMS(sbuf[:], uint64(ns-prev))))
  			prev = ns
  		}
  		print(" ms clock, ")
  		for i, ns := range []int64{sweepTermCpu, gcController.assistTime, gcController.dedicatedMarkTime + gcController.fractionalMarkTime, gcController.idleMarkTime, markTermCpu} {
  			if i == 2 || i == 3 {
  				// Separate mark time components with /.
  				print("/")
  			} else if i != 0 {
  				print("+")
  			}
  			print(string(fmtNSAsMS(sbuf[:], uint64(ns))))
  		}
  		print(" ms cpu, ",
  			work.heap0>>20, "->", work.heap1>>20, "->", work.heap2>>20, " MB, ",
  			work.heapGoal>>20, " MB goal, ",
  			work.maxprocs, " P")
  		if work.userForced {
  			print(" (forced)")
  		}
  		print("\n")
  		printunlock()
  	}
  
  	semrelease(&worldsema)
  	// Careful: another GC cycle may start now.
  
  	releasem(mp)
  	mp = nil
  
  	// now that gc is done, kick off finalizer thread if needed
  	if !concurrentSweep {
  		// give the queued finalizers, if any, a chance to run
  		Gosched()
  	}
  }
  
  // gcBgMarkStartWorkers prepares background mark worker goroutines.
  // These goroutines will not run until the mark phase, but they must
  // be started while the work is not stopped and from a regular G
  // stack. The caller must hold worldsema.
  func gcBgMarkStartWorkers() {
  	// Background marking is performed by per-P G's. Ensure that
  	// each P has a background GC G.
  	for _, p := range &allp {
  		if p == nil || p.status == _Pdead {
  			break
  		}
  		if p.gcBgMarkWorker == 0 {
  			go gcBgMarkWorker(p)
  			notetsleepg(&work.bgMarkReady, -1)
  			noteclear(&work.bgMarkReady)
  		}
  	}
  }
  
  // gcBgMarkPrepare sets up state for background marking.
  // Mutator assists must not yet be enabled.
  func gcBgMarkPrepare() {
  	// Background marking will stop when the work queues are empty
  	// and there are no more workers (note that, since this is
  	// concurrent, this may be a transient state, but mark
  	// termination will clean it up). Between background workers
  	// and assists, we don't really know how many workers there
  	// will be, so we pretend to have an arbitrarily large number
  	// of workers, almost all of which are "waiting". While a
  	// worker is working it decrements nwait. If nproc == nwait,
  	// there are no workers.
  	work.nproc = ^uint32(0)
  	work.nwait = ^uint32(0)
  }
  
  func gcBgMarkWorker(_p_ *p) {
  	gp := getg()
  
  	type parkInfo struct {
  		m      muintptr // Release this m on park.
  		attach puintptr // If non-nil, attach to this p on park.
  	}
  	// We pass park to a gopark unlock function, so it can't be on
  	// the stack (see gopark). Prevent deadlock from recursively
  	// starting GC by disabling preemption.
  	gp.m.preemptoff = "GC worker init"
  	park := new(parkInfo)
  	gp.m.preemptoff = ""
  
  	park.m.set(acquirem())
  	park.attach.set(_p_)
  	// Inform gcBgMarkStartWorkers that this worker is ready.
  	// After this point, the background mark worker is scheduled
  	// cooperatively by gcController.findRunnable. Hence, it must
  	// never be preempted, as this would put it into _Grunnable
  	// and put it on a run queue. Instead, when the preempt flag
  	// is set, this puts itself into _Gwaiting to be woken up by
  	// gcController.findRunnable at the appropriate time.
  	notewakeup(&work.bgMarkReady)
  
  	for {
  		// Go to sleep until woken by gcController.findRunnable.
  		// We can't releasem yet since even the call to gopark
  		// may be preempted.
  		gopark(func(g *g, parkp unsafe.Pointer) bool {
  			park := (*parkInfo)(parkp)
  
  			// The worker G is no longer running, so it's
  			// now safe to allow preemption.
  			releasem(park.m.ptr())
  
  			// If the worker isn't attached to its P,
  			// attach now. During initialization and after
  			// a phase change, the worker may have been
  			// running on a different P. As soon as we
  			// attach, the owner P may schedule the
  			// worker, so this must be done after the G is
  			// stopped.
  			if park.attach != 0 {
  				p := park.attach.ptr()
  				park.attach.set(nil)
  				// cas the worker because we may be
  				// racing with a new worker starting
  				// on this P.
  				if !p.gcBgMarkWorker.cas(0, guintptr(unsafe.Pointer(g))) {
  					// The P got a new worker.
  					// Exit this worker.
  					return false
  				}
  			}
  			return true
  		}, unsafe.Pointer(park), "GC worker (idle)", traceEvGoBlock, 0)
  
  		// Loop until the P dies and disassociates this
  		// worker (the P may later be reused, in which case
  		// it will get a new worker) or we failed to associate.
  		if _p_.gcBgMarkWorker.ptr() != gp {
  			break
  		}
  
  		// Disable preemption so we can use the gcw. If the
  		// scheduler wants to preempt us, we'll stop draining,
  		// dispose the gcw, and then preempt.
  		park.m.set(acquirem())
  
  		if gcBlackenEnabled == 0 {
  			throw("gcBgMarkWorker: blackening not enabled")
  		}
  
  		startTime := nanotime()
  
  		decnwait := atomic.Xadd(&work.nwait, -1)
  		if decnwait == work.nproc {
  			println("runtime: work.nwait=", decnwait, "work.nproc=", work.nproc)
  			throw("work.nwait was > work.nproc")
  		}
  
  		systemstack(func() {
  			// Mark our goroutine preemptible so its stack
  			// can be scanned. This lets two mark workers
  			// scan each other (otherwise, they would
  			// deadlock). We must not modify anything on
  			// the G stack. However, stack shrinking is
  			// disabled for mark workers, so it is safe to
  			// read from the G stack.
  			casgstatus(gp, _Grunning, _Gwaiting)
  			switch _p_.gcMarkWorkerMode {
  			default:
  				throw("gcBgMarkWorker: unexpected gcMarkWorkerMode")
  			case gcMarkWorkerDedicatedMode:
  				gcDrain(&_p_.gcw, gcDrainUntilPreempt|gcDrainFlushBgCredit)
  				if gp.preempt {
  					// We were preempted. This is
  					// a useful signal to kick
  					// everything out of the run
  					// queue so it can run
  					// somewhere else.
  					lock(&sched.lock)
  					for {
  						gp, _ := runqget(_p_)
  						if gp == nil {
  							break
  						}
  						globrunqput(gp)
  					}
  					unlock(&sched.lock)
  				}
  				// Go back to draining, this time
  				// without preemption.
  				gcDrain(&_p_.gcw, gcDrainNoBlock|gcDrainFlushBgCredit)
  			case gcMarkWorkerFractionalMode:
  				gcDrain(&_p_.gcw, gcDrainUntilPreempt|gcDrainFlushBgCredit)
  			case gcMarkWorkerIdleMode:
  				gcDrain(&_p_.gcw, gcDrainIdle|gcDrainUntilPreempt|gcDrainFlushBgCredit)
  			}
  			casgstatus(gp, _Gwaiting, _Grunning)
  		})
  
  		// If we are nearing the end of mark, dispose
  		// of the cache promptly. We must do this
  		// before signaling that we're no longer
  		// working so that other workers can't observe
  		// no workers and no work while we have this
  		// cached, and before we compute done.
  		if gcBlackenPromptly {
  			_p_.gcw.dispose()
  		}
  
  		// Account for time.
  		duration := nanotime() - startTime
  		switch _p_.gcMarkWorkerMode {
  		case gcMarkWorkerDedicatedMode:
  			atomic.Xaddint64(&gcController.dedicatedMarkTime, duration)
  			atomic.Xaddint64(&gcController.dedicatedMarkWorkersNeeded, 1)
  		case gcMarkWorkerFractionalMode:
  			atomic.Xaddint64(&gcController.fractionalMarkTime, duration)
  			atomic.Xaddint64(&gcController.fractionalMarkWorkersNeeded, 1)
  		case gcMarkWorkerIdleMode:
  			atomic.Xaddint64(&gcController.idleMarkTime, duration)
  		}
  
  		// Was this the last worker and did we run out
  		// of work?
  		incnwait := atomic.Xadd(&work.nwait, +1)
  		if incnwait > work.nproc {
  			println("runtime: p.gcMarkWorkerMode=", _p_.gcMarkWorkerMode,
  				"work.nwait=", incnwait, "work.nproc=", work.nproc)
  			throw("work.nwait > work.nproc")
  		}
  
  		// If this worker reached a background mark completion
  		// point, signal the main GC goroutine.
  		if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
  			// Make this G preemptible and disassociate it
  			// as the worker for this P so
  			// findRunnableGCWorker doesn't try to
  			// schedule it.
  			_p_.gcBgMarkWorker.set(nil)
  			releasem(park.m.ptr())
  
  			gcMarkDone()
  
  			// Disable preemption and prepare to reattach
  			// to the P.
  			//
  			// We may be running on a different P at this
  			// point, so we can't reattach until this G is
  			// parked.
  			park.m.set(acquirem())
  			park.attach.set(_p_)
  		}
  	}
  }
  
  // gcMarkWorkAvailable returns true if executing a mark worker
  // on p is potentially useful. p may be nil, in which case it only
  // checks the global sources of work.
  func gcMarkWorkAvailable(p *p) bool {
  	if p != nil && !p.gcw.empty() {
  		return true
  	}
  	if !work.full.empty() {
  		return true // global work available
  	}
  	if work.markrootNext < work.markrootJobs {
  		return true // root scan work available
  	}
  	return false
  }
  
  // gcMark runs the mark (or, for concurrent GC, mark termination)
  // All gcWork caches must be empty.
  // STW is in effect at this point.
  //TODO go:nowritebarrier
  func gcMark(start_time int64) {
  	if debug.allocfreetrace > 0 {
  		tracegc()
  	}
  
  	if gcphase != _GCmarktermination {
  		throw("in gcMark expecting to see gcphase as _GCmarktermination")
  	}
  	work.tstart = start_time
  
  	// Queue root marking jobs.
  	gcMarkRootPrepare()
  
  	work.nwait = 0
  	work.ndone = 0
  	work.nproc = uint32(gcprocs())
  
  	if work.full == 0 && work.nDataRoots+work.nBSSRoots+work.nSpanRoots+work.nStackRoots == 0 {
  		// There's no work on the work queue and no root jobs
  		// that can produce work, so don't bother entering the
  		// getfull() barrier.
  		//
  		// This will be the situation the vast majority of the
  		// time after concurrent mark. However, we still need
  		// a fallback for STW GC and because there are some
  		// known races that occasionally leave work around for
  		// mark termination.
  		//
  		// We're still hedging our bets here: if we do
  		// accidentally produce some work, we'll still process
  		// it, just not necessarily in parallel.
  		//
  		// TODO(austin): Fix the races and and remove
  		// work draining from mark termination so we don't
  		// need the fallback path.
  		work.helperDrainBlock = false
  	} else {
  		work.helperDrainBlock = true
  	}
  
  	if trace.enabled {
  		traceGCScanStart()
  	}
  
  	if work.nproc > 1 {
  		noteclear(&work.alldone)
  		helpgc(int32(work.nproc))
  	}
  
  	gchelperstart()
  
  	gcw := &getg().m.p.ptr().gcw
  	if work.helperDrainBlock {
  		gcDrain(gcw, gcDrainBlock)
  	} else {
  		gcDrain(gcw, gcDrainNoBlock)
  	}
  	gcw.dispose()
  
  	if debug.gccheckmark > 0 {
  		// This is expensive when there's a large number of
  		// Gs, so only do it if checkmark is also enabled.
  		gcMarkRootCheck()
  	}
  	if work.full != 0 {
  		throw("work.full != 0")
  	}
  
  	if work.nproc > 1 {
  		notesleep(&work.alldone)
  	}
  
  	// Record that at least one root marking pass has completed.
  	work.markrootDone = true
  
  	// Double-check that all gcWork caches are empty. This should
  	// be ensured by mark 2 before we enter mark termination.
  	for i := 0; i < int(gomaxprocs); i++ {
  		gcw := &allp[i].gcw
  		if !gcw.empty() {
  			throw("P has cached GC work at end of mark termination")
  		}
  		if gcw.scanWork != 0 || gcw.bytesMarked != 0 {
  			throw("P has unflushed stats at end of mark termination")
  		}
  	}
  
  	if trace.enabled {
  		traceGCScanDone()
  	}
  
  	cachestats()
  
  	// Update the marked heap stat.
  	memstats.heap_marked = work.bytesMarked
  
  	// Update other GC heap size stats. This must happen after
  	// cachestats (which flushes local statistics to these) and
  	// flushallmcaches (which modifies heap_live).
  	memstats.heap_live = work.bytesMarked
  	memstats.heap_scan = uint64(gcController.scanWork)
  
  	if trace.enabled {
  		traceHeapAlloc()
  	}
  }
  
  func gcSweep(mode gcMode) {
  	if gcphase != _GCoff {
  		throw("gcSweep being done but phase is not GCoff")
  	}
  
  	lock(&mheap_.lock)
  	mheap_.sweepgen += 2
  	mheap_.sweepdone = 0
  	if mheap_.sweepSpans[mheap_.sweepgen/2%2].index != 0 {
  		// We should have drained this list during the last
  		// sweep phase. We certainly need to start this phase
  		// with an empty swept list.
  		throw("non-empty swept list")
  	}
  	mheap_.pagesSwept = 0
  	unlock(&mheap_.lock)
  
  	if !_ConcurrentSweep || mode == gcForceBlockMode {
  		// Special case synchronous sweep.
  		// Record that no proportional sweeping has to happen.
  		lock(&mheap_.lock)
  		mheap_.sweepPagesPerByte = 0
  		unlock(&mheap_.lock)
  		// Sweep all spans eagerly.
  		for sweepone() != ^uintptr(0) {
  			sweep.npausesweep++
  		}
  		// Free workbufs eagerly.
  		prepareFreeWorkbufs()
  		for freeSomeWbufs(false) {
  		}
  		// All "free" events for this mark/sweep cycle have
  		// now happened, so we can make this profile cycle
  		// available immediately.
  		mProf_NextCycle()
  		mProf_Flush()
  		return
  	}
  
  	// Background sweep.
  	lock(&sweep.lock)
  	if sweep.parked {
  		sweep.parked = false
  		ready(sweep.g, 0, true)
  	}
  	unlock(&sweep.lock)
  }
  
  // gcResetMarkState resets global state prior to marking (concurrent
  // or STW) and resets the stack scan state of all Gs.
  //
  // This is safe to do without the world stopped because any Gs created
  // during or after this will start out in the reset state.
  func gcResetMarkState() {
  	// This may be called during a concurrent phase, so make sure
  	// allgs doesn't change.
  	lock(&allglock)
  	for _, gp := range allgs {
  		gp.gcscandone = false  // set to true in gcphasework
  		gp.gcscanvalid = false // stack has not been scanned
  		gp.gcAssistBytes = 0
  	}
  	unlock(&allglock)
  
  	work.bytesMarked = 0
  	work.initialHeapLive = atomic.Load64(&memstats.heap_live)
  	work.markrootDone = false
  }
  
  // Hooks for other packages
  
  var poolcleanup func()
  
  //go:linkname sync_runtime_registerPoolCleanup sync.runtime_registerPoolCleanup
  func sync_runtime_registerPoolCleanup(f func()) {
  	poolcleanup = f
  }
  
  func clearpools() {
  	// clear sync.Pools
  	if poolcleanup != nil {
  		poolcleanup()
  	}
  
  	// Clear central sudog cache.
  	// Leave per-P caches alone, they have strictly bounded size.
  	// Disconnect cached list before dropping it on the floor,
  	// so that a dangling ref to one entry does not pin all of them.
  	lock(&sched.sudoglock)
  	var sg, sgnext *sudog
  	for sg = sched.sudogcache; sg != nil; sg = sgnext {
  		sgnext = sg.next
  		sg.next = nil
  	}
  	sched.sudogcache = nil
  	unlock(&sched.sudoglock)
  
  	// Clear central defer pools.
  	// Leave per-P pools alone, they have strictly bounded size.
  	lock(&sched.deferlock)
  	for i := range sched.deferpool {
  		// disconnect cached list before dropping it on the floor,
  		// so that a dangling ref to one entry does not pin all of them.
  		var d, dlink *_defer
  		for d = sched.deferpool[i]; d != nil; d = dlink {
  			dlink = d.link
  			d.link = nil
  		}
  		sched.deferpool[i] = nil
  	}
  	unlock(&sched.deferlock)
  }
  
  // Timing
  
  //go:nowritebarrier
  func gchelper() {
  	_g_ := getg()
  	_g_.m.traceback = 2
  	gchelperstart()
  
  	if trace.enabled {
  		traceGCScanStart()
  	}
  
  	// Parallel mark over GC roots and heap
  	if gcphase == _GCmarktermination {
  		gcw := &_g_.m.p.ptr().gcw
  		if work.helperDrainBlock {
  			gcDrain(gcw, gcDrainBlock) // blocks in getfull
  		} else {
  			gcDrain(gcw, gcDrainNoBlock)
  		}
  		gcw.dispose()
  	}
  
  	if trace.enabled {
  		traceGCScanDone()
  	}
  
  	nproc := atomic.Load(&work.nproc) // work.nproc can change right after we increment work.ndone
  	if atomic.Xadd(&work.ndone, +1) == nproc-1 {
  		notewakeup(&work.alldone)
  	}
  	_g_.m.traceback = 0
  }
  
  func gchelperstart() {
  	_g_ := getg()
  
  	if _g_.m.helpgc < 0 || _g_.m.helpgc >= _MaxGcproc {
  		throw("gchelperstart: bad m->helpgc")
  	}
  	if _g_ != _g_.m.g0 {
  		throw("gchelper not running on g0 stack")
  	}
  }
  
  // itoaDiv formats val/(10**dec) into buf.
  func itoaDiv(buf []byte, val uint64, dec int) []byte {
  	i := len(buf) - 1
  	idec := i - dec
  	for val >= 10 || i >= idec {
  		buf[i] = byte(val%10 + '0')
  		i--
  		if i == idec {
  			buf[i] = '.'
  			i--
  		}
  		val /= 10
  	}
  	buf[i] = byte(val + '0')
  	return buf[i:]
  }
  
  // fmtNSAsMS nicely formats ns nanoseconds as milliseconds.
  func fmtNSAsMS(buf []byte, ns uint64) []byte {
  	if ns >= 10e6 {
  		// Format as whole milliseconds.
  		return itoaDiv(buf, ns/1e6, 0)
  	}
  	// Format two digits of precision, with at most three decimal places.
  	x := ns / 1e3
  	if x == 0 {
  		buf[0] = '0'
  		return buf[:1]
  	}
  	dec := 3
  	for x >= 100 {
  		x /= 10
  		dec--
  	}
  	return itoaDiv(buf, x, dec)
  }
  

View as plain text