mgcscavenge.go

// Copyright 2019 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.

// Scavenging free pages.
//
// This file implements scavenging (the release of physical pages backing mapped
// memory) of free and unused pages in the heap as a way to deal with page-level
// fragmentation and reduce the RSS of Go applications.
//
// Scavenging in Go happens on two fronts: there's the background
// (asynchronous) scavenger and the heap-growth (synchronous) scavenger.
//
// The former happens on a goroutine much like the background sweeper which is
// soft-capped at using scavengePercent of the mutator's time, based on
// order-of-magnitude estimates of the costs of scavenging. The background
// scavenger's primary goal is to bring the estimated heap RSS of the
// application down to a goal.
//
// That goal is defined as:
//   (retainExtraPercent+100) / 100 * (next_gc / last_next_gc) * last_heap_inuse
//
// Essentially, we wish to have the application's RSS track the heap goal, but
// the heap goal is defined in terms of bytes of objects, rather than pages like
// RSS. As a result, we need to take into account for fragmentation internal to
// spans. next_gc / last_next_gc defines the ratio between the current heap goal
// and the last heap goal, which tells us by how much the heap is growing and
// shrinking. We estimate what the heap will grow to in terms of pages by taking
// this ratio and multiplying it by heap_inuse at the end of the last GC, which
// allows us to account for this additional fragmentation. Note that this
// procedure makes the assumption that the degree of fragmentation won't change
// dramatically over the next GC cycle. Overestimating the amount of
// fragmentation simply results in higher memory use, which will be accounted
// for by the next pacing up date. Underestimating the fragmentation however
// could lead to performance degradation. Handling this case is not within the
// scope of the scavenger. Situations where the amount of fragmentation balloons
// over the course of a single GC cycle should be considered pathologies,
// flagged as bugs, and fixed appropriately.
//
// An additional factor of retainExtraPercent is added as a buffer to help ensure
// that there's more unscavenged memory to allocate out of, since each allocation
// out of scavenged memory incurs a potentially expensive page fault.
//
// The goal is updated after each GC and the scavenger's pacing parameters
// (which live in mheap_) are updated to match. The pacing parameters work much
// like the background sweeping parameters. The parameters define a line whose
// horizontal axis is time and vertical axis is estimated heap RSS, and the
// scavenger attempts to stay below that line at all times.
//
// The synchronous heap-growth scavenging happens whenever the heap grows in
// size, for some definition of heap-growth. The intuition behind this is that
// the application had to grow the heap because existing fragments were
// not sufficiently large to satisfy a page-level memory allocation, so we
// scavenge those fragments eagerly to offset the growth in RSS that results.

package runtime

import (
	"runtime/internal/atomic"
	"runtime/internal/sys"
	"unsafe"
)

const (
	// The background scavenger is paced according to these parameters.
	//
	// scavengePercent represents the portion of mutator time we're willing
	// to spend on scavenging in percent.
	scavengePercent = 1 // 1%

	// retainExtraPercent represents the amount of memory over the heap goal
	// that the scavenger should keep as a buffer space for the allocator.
	//
	// The purpose of maintaining this overhead is to have a greater pool of
	// unscavenged memory available for allocation (since using scavenged memory
	// incurs an additional cost), to account for heap fragmentation and
	// the ever-changing layout of the heap.
	retainExtraPercent = 10

	// maxPagesPerPhysPage is the maximum number of supported runtime pages per
	// physical page, based on maxPhysPageSize.
	maxPagesPerPhysPage = maxPhysPageSize / pageSize

	// scavengeCostRatio is the approximate ratio between the costs of using previously
	// scavenged memory and scavenging memory.
	//
	// For most systems the cost of scavenging greatly outweighs the costs
	// associated with using scavenged memory, making this constant 0. On other systems
	// (especially ones where "sysUsed" is not just a no-op) this cost is non-trivial.
	//
	// This ratio is used as part of multiplicative factor to help the scavenger account
	// for the additional costs of using scavenged memory in its pacing.
	scavengeCostRatio = 0.7 * (sys.GoosDarwin + sys.GoosIos)

	// scavengeReservationShards determines the amount of memory the scavenger
	// should reserve for scavenging at a time. Specifically, the amount of
	// memory reserved is (heap size in bytes) / scavengeReservationShards.
	scavengeReservationShards = 64
)

// heapRetained returns an estimate of the current heap RSS.
func heapRetained() uint64 {
	return memstats.heap_sys.load() - atomic.Load64(&memstats.heap_released)
}

// gcPaceScavenger updates the scavenger's pacing, particularly
// its rate and RSS goal.
//
// The RSS goal is based on the current heap goal with a small overhead
// to accommodate non-determinism in the allocator.
//
// The pacing is based on scavengePageRate, which applies to both regular and
// huge pages. See that constant for more information.
//
// mheap_.lock must be held or the world must be stopped.
func gcPaceScavenger() {
	// If we're called before the first GC completed, disable scavenging.
	// We never scavenge before the 2nd GC cycle anyway (we don't have enough
	// information about the heap yet) so this is fine, and avoids a fault
	// or garbage data later.
	if memstats.last_next_gc == 0 {
		mheap_.scavengeGoal = ^uint64(0)
		return
	}
	// Compute our scavenging goal.
	goalRatio := float64(atomic.Load64(&memstats.next_gc)) / float64(memstats.last_next_gc)
	retainedGoal := uint64(float64(memstats.last_heap_inuse) * goalRatio)
	// Add retainExtraPercent overhead to retainedGoal. This calculation
	// looks strange but the purpose is to arrive at an integer division
	// (e.g. if retainExtraPercent = 12.5, then we get a divisor of 8)
	// that also avoids the overflow from a multiplication.
	retainedGoal += retainedGoal / (1.0 / (retainExtraPercent / 100.0))
	// Align it to a physical page boundary to make the following calculations
	// a bit more exact.
	retainedGoal = (retainedGoal + uint64(physPageSize) - 1) &^ (uint64(physPageSize) - 1)

	// Represents where we are now in the heap's contribution to RSS in bytes.
	//
	// Guaranteed to always be a multiple of physPageSize on systems where
	// physPageSize <= pageSize since we map heap_sys at a rate larger than
	// any physPageSize and released memory in multiples of the physPageSize.
	//
	// However, certain functions recategorize heap_sys as other stats (e.g.
	// stack_sys) and this happens in multiples of pageSize, so on systems
	// where physPageSize > pageSize the calculations below will not be exact.
	// Generally this is OK since we'll be off by at most one regular
	// physical page.
	retainedNow := heapRetained()

	// If we're already below our goal, or within one page of our goal, then disable
	// the background scavenger. We disable the background scavenger if there's
	// less than one physical page of work to do because it's not worth it.
	if retainedNow <= retainedGoal || retainedNow-retainedGoal < uint64(physPageSize) {
		mheap_.scavengeGoal = ^uint64(0)
		return
	}
	mheap_.scavengeGoal = retainedGoal
}

// Sleep/wait state of the background scavenger.
var scavenge struct {
	lock       mutex
	g          *g
	parked     bool
	timer      *timer
	sysmonWake uint32 // Set atomically.
}

// readyForScavenger signals sysmon to wake the scavenger because
// there may be new work to do.
//
// There may be a significant delay between when this function runs
// and when the scavenger is kicked awake, but it may be safely invoked
// in contexts where wakeScavenger is unsafe to call directly.
func readyForScavenger() {
	atomic.Store(&scavenge.sysmonWake, 1)
}

// wakeScavenger immediately unparks the scavenger if necessary.
//
// May run without a P, but it may allocate, so it must not be called
// on any allocation path.
//
// mheap_.lock, scavenge.lock, and sched.lock must not be held.
func wakeScavenger() {
	lock(&scavenge.lock)
	if scavenge.parked {
		// Notify sysmon that it shouldn't bother waking up the scavenger.
		atomic.Store(&scavenge.sysmonWake, 0)

		// Try to stop the timer but we don't really care if we succeed.
		// It's possible that either a timer was never started, or that
		// we're racing with it.
		// In the case that we're racing with there's the low chance that
		// we experience a spurious wake-up of the scavenger, but that's
		// totally safe.
		stopTimer(scavenge.timer)

		// Unpark the goroutine and tell it that there may have been a pacing
		// change. Note that we skip the scheduler's runnext slot because we
		// want to avoid having the scavenger interfere with the fair
		// scheduling of user goroutines. In effect, this schedules the
		// scavenger at a "lower priority" but that's OK because it'll
		// catch up on the work it missed when it does get scheduled.
		scavenge.parked = false

		// Ready the goroutine by injecting it. We use injectglist instead
		// of ready or goready in order to allow us to run this function
		// without a P. injectglist also avoids placing the goroutine in
		// the current P's runnext slot, which is desireable to prevent
		// the scavenger from interfering with user goroutine scheduling
		// too much.
		var list gList
		list.push(scavenge.g)
		injectglist(&list)
	}
	unlock(&scavenge.lock)
}

// scavengeSleep attempts to put the scavenger to sleep for ns.
//
// Note that this function should only be called by the scavenger.
//
// The scavenger may be woken up earlier by a pacing change, and it may not go
// to sleep at all if there's a pending pacing change.
//
// Returns the amount of time actually slept.
func scavengeSleep(ns int64) int64 {
	lock(&scavenge.lock)

	// Set the timer.
	//
	// This must happen here instead of inside gopark
	// because we can't close over any variables without
	// failing escape analysis.
	start := nanotime()
	resetTimer(scavenge.timer, start+ns)

	// Mark ourself as asleep and go to sleep.
	scavenge.parked = true
	goparkunlock(&scavenge.lock, waitReasonSleep, traceEvGoSleep, 2)

	// Return how long we actually slept for.
	return nanotime() - start
}

// Background scavenger.
//
// The background scavenger maintains the RSS of the application below
// the line described by the proportional scavenging statistics in
// the mheap struct.
func bgscavenge(c chan int) {
	scavenge.g = getg()

	lockInit(&scavenge.lock, lockRankScavenge)
	lock(&scavenge.lock)
	scavenge.parked = true

	scavenge.timer = new(timer)
	scavenge.timer.f = func(_ interface{}, _ uintptr) {
		wakeScavenger()
	}

	c <- 1
	goparkunlock(&scavenge.lock, waitReasonGCScavengeWait, traceEvGoBlock, 1)

	// Exponentially-weighted moving average of the fraction of time this
	// goroutine spends scavenging (that is, percent of a single CPU).
	// It represents a measure of scheduling overheads which might extend
	// the sleep or the critical time beyond what's expected. Assume no
	// overhead to begin with.
	//
	// TODO(mknyszek): Consider making this based on total CPU time of the
	// application (i.e. scavengePercent * GOMAXPROCS). This isn't really
	// feasible now because the scavenger acquires the heap lock over the
	// scavenging operation, which means scavenging effectively blocks
	// allocators and isn't scalable. However, given a scalable allocator,
	// it makes sense to also make the scavenger scale with it; if you're
	// allocating more frequently, then presumably you're also generating
	// more work for the scavenger.
	const idealFraction = scavengePercent / 100.0
	scavengeEWMA := float64(idealFraction)

	for {
		released := uintptr(0)

		// Time in scavenging critical section.
		crit := float64(0)

		// Run on the system stack since we grab the heap lock,
		// and a stack growth with the heap lock means a deadlock.
		systemstack(func() {
			lock(&mheap_.lock)

			// If background scavenging is disabled or if there's no work to do just park.
			retained, goal := heapRetained(), mheap_.scavengeGoal
			if retained <= goal {
				unlock(&mheap_.lock)
				return
			}

			// Scavenge one page, and measure the amount of time spent scavenging.
			start := nanotime()
			released = mheap_.pages.scavenge(physPageSize, true)
			mheap_.pages.scav.released += released
			crit = float64(nanotime() - start)

			unlock(&mheap_.lock)
		})

		if released == 0 {
			lock(&scavenge.lock)
			scavenge.parked = true
			goparkunlock(&scavenge.lock, waitReasonGCScavengeWait, traceEvGoBlock, 1)
			continue
		}

		if released < physPageSize {
			// If this happens, it means that we may have attempted to release part
			// of a physical page, but the likely effect of that is that it released
			// the whole physical page, some of which may have still been in-use.
			// This could lead to memory corruption. Throw.
			throw("released less than one physical page of memory")
		}

		// On some platforms we may see crit as zero if the time it takes to scavenge
		// memory is less than the minimum granularity of its clock (e.g. Windows).
		// In this case, just assume scavenging takes 10 µs per regular physical page
		// (determined empirically), and conservatively ignore the impact of huge pages
		// on timing.
		//
		// We shouldn't ever see a crit value less than zero unless there's a bug of
		// some kind, either on our side or in the platform we're running on, but be
		// defensive in that case as well.
		const approxCritNSPerPhysicalPage = 10e3
		if crit <= 0 {
			crit = approxCritNSPerPhysicalPage * float64(released/physPageSize)
		}

		// Multiply the critical time by 1 + the ratio of the costs of using
		// scavenged memory vs. scavenging memory. This forces us to pay down
		// the cost of reusing this memory eagerly by sleeping for a longer period
		// of time and scavenging less frequently. More concretely, we avoid situations
		// where we end up scavenging so often that we hurt allocation performance
		// because of the additional overheads of using scavenged memory.
		crit *= 1 + scavengeCostRatio

		// If we spent more than 10 ms (for example, if the OS scheduled us away, or someone
		// put their machine to sleep) in the critical section, bound the time we use to
		// calculate at 10 ms to avoid letting the sleep time get arbitrarily high.
		const maxCrit = 10e6
		if crit > maxCrit {
			crit = maxCrit
		}

		// Compute the amount of time to sleep, assuming we want to use at most
		// scavengePercent of CPU time. Take into account scheduling overheads
		// that may extend the length of our sleep by multiplying by how far
		// off we are from the ideal ratio. For example, if we're sleeping too
		// much, then scavengeEMWA < idealFraction, so we'll adjust the sleep time
		// down.
		adjust := scavengeEWMA / idealFraction
		sleepTime := int64(adjust * crit / (scavengePercent / 100.0))

		// Go to sleep.
		slept := scavengeSleep(sleepTime)

		// Compute the new ratio.
		fraction := crit / (crit + float64(slept))

		// Set a lower bound on the fraction.
		// Due to OS-related anomalies we may "sleep" for an inordinate amount
		// of time. Let's avoid letting the ratio get out of hand by bounding
		// the sleep time we use in our EWMA.
		const minFraction = 1 / 1000
		if fraction < minFraction {
			fraction = minFraction
		}

		// Update scavengeEWMA by merging in the new crit/slept ratio.
		const alpha = 0.5
		scavengeEWMA = alpha*fraction + (1-alpha)*scavengeEWMA
	}
}

// scavenge scavenges nbytes worth of free pages, starting with the
// highest address first. Successive calls continue from where it left
// off until the heap is exhausted. Call scavengeStartGen to bring it
// back to the top of the heap.
//
// Returns the amount of memory scavenged in bytes.
//
// p.mheapLock must be held, but may be temporarily released if
// mayUnlock == true.
//
// Must run on the system stack because p.mheapLock must be held.
//
//go:systemstack
func (p *pageAlloc) scavenge(nbytes uintptr, mayUnlock bool) uintptr {
	assertLockHeld(p.mheapLock)

	var (
		addrs addrRange
		gen   uint32
	)
	released := uintptr(0)
	for released < nbytes {
		if addrs.size() == 0 {
			if addrs, gen = p.scavengeReserve(); addrs.size() == 0 {
				break
			}
		}
		r, a := p.scavengeOne(addrs, nbytes-released, mayUnlock)
		released += r
		addrs = a
	}
	// Only unreserve the space which hasn't been scavenged or searched
	// to ensure we always make progress.
	p.scavengeUnreserve(addrs, gen)
	return released
}

// printScavTrace prints a scavenge trace line to standard error.
//
// released should be the amount of memory released since the last time this
// was called, and forced indicates whether the scavenge was forced by the
// application.
func printScavTrace(gen uint32, released uintptr, forced bool) {
	printlock()
	print("scav ", gen, " ",
		released>>10, " KiB work, ",
		atomic.Load64(&memstats.heap_released)>>10, " KiB total, ",
		(atomic.Load64(&memstats.heap_inuse)*100)/heapRetained(), "% util",
	)
	if forced {
		print(" (forced)")
	}
	println()
	printunlock()
}

// scavengeStartGen starts a new scavenge generation, resetting
// the scavenger's search space to the full in-use address space.
//
// p.mheapLock must be held.
//
// Must run on the system stack because p.mheapLock must be held.
//
//go:systemstack
func (p *pageAlloc) scavengeStartGen() {
	assertLockHeld(p.mheapLock)

	if debug.scavtrace > 0 {
		printScavTrace(p.scav.gen, p.scav.released, false)
	}
	p.inUse.cloneInto(&p.scav.inUse)

	// Pick the new starting address for the scavenger cycle.
	var startAddr offAddr
	if p.scav.scavLWM.lessThan(p.scav.freeHWM) {
		// The "free" high watermark exceeds the "scavenged" low watermark,
		// so there are free scavengable pages in parts of the address space
		// that the scavenger already searched, the high watermark being the
		// highest one. Pick that as our new starting point to ensure we
		// see those pages.
		startAddr = p.scav.freeHWM
	} else {
		// The "free" high watermark does not exceed the "scavenged" low
		// watermark. This means the allocator didn't free any memory in
		// the range we scavenged last cycle, so we might as well continue
		// scavenging from where we were.
		startAddr = p.scav.scavLWM
	}
	p.scav.inUse.removeGreaterEqual(startAddr.addr())

	// reservationBytes may be zero if p.inUse.totalBytes is small, or if
	// scavengeReservationShards is large. This case is fine as the scavenger
	// will simply be turned off, but it does mean that scavengeReservationShards,
	// in concert with pallocChunkBytes, dictates the minimum heap size at which
	// the scavenger triggers. In practice this minimum is generally less than an
	// arena in size, so virtually every heap has the scavenger on.
	p.scav.reservationBytes = alignUp(p.inUse.totalBytes, pallocChunkBytes) / scavengeReservationShards
	p.scav.gen++
	p.scav.released = 0
	p.scav.freeHWM = minOffAddr
	p.scav.scavLWM = maxOffAddr
}

// scavengeReserve reserves a contiguous range of the address space
// for scavenging. The maximum amount of space it reserves is proportional
// to the size of the heap. The ranges are reserved from the high addresses
// first.
//
// Returns the reserved range and the scavenge generation number for it.
//
// p.mheapLock must be held.
//
// Must run on the system stack because p.mheapLock must be held.
//
//go:systemstack
func (p *pageAlloc) scavengeReserve() (addrRange, uint32) {
	assertLockHeld(p.mheapLock)

	// Start by reserving the minimum.
	r := p.scav.inUse.removeLast(p.scav.reservationBytes)

	// Return early if the size is zero; we don't want to use
	// the bogus address below.
	if r.size() == 0 {
		return r, p.scav.gen
	}

	// The scavenger requires that base be aligned to a
	// palloc chunk because that's the unit of operation for
	// the scavenger, so align down, potentially extending
	// the range.
	newBase := alignDown(r.base.addr(), pallocChunkBytes)

	// Remove from inUse however much extra we just pulled out.
	p.scav.inUse.removeGreaterEqual(newBase)
	r.base = offAddr{newBase}
	return r, p.scav.gen
}

// scavengeUnreserve returns an unscavenged portion of a range that was
// previously reserved with scavengeReserve.
//
// p.mheapLock must be held.
//
// Must run on the system stack because p.mheapLock must be held.
//
//go:systemstack
func (p *pageAlloc) scavengeUnreserve(r addrRange, gen uint32) {
	assertLockHeld(p.mheapLock)

	if r.size() == 0 || gen != p.scav.gen {
		return
	}
	if r.base.addr()%pallocChunkBytes != 0 {
		throw("unreserving unaligned region")
	}
	p.scav.inUse.add(r)
}

// scavengeOne walks over address range work until it finds
// a contiguous run of pages to scavenge. It will try to scavenge
// at most max bytes at once, but may scavenge more to avoid
// breaking huge pages. Once it scavenges some memory it returns
// how much it scavenged in bytes.
//
// Returns the number of bytes scavenged and the part of work
// which was not yet searched.
//
// work's base address must be aligned to pallocChunkBytes.
//
// p.mheapLock must be held, but may be temporarily released if
// mayUnlock == true.
//
// Must run on the system stack because p.mheapLock must be held.
//
//go:systemstack
func (p *pageAlloc) scavengeOne(work addrRange, max uintptr, mayUnlock bool) (uintptr, addrRange) {
	assertLockHeld(p.mheapLock)

	// Defensively check if we've recieved an empty address range.
	// If so, just return.
	if work.size() == 0 {
		// Nothing to do.
		return 0, work
	}
	// Check the prerequisites of work.
	if work.base.addr()%pallocChunkBytes != 0 {
		throw("scavengeOne called with unaligned work region")
	}
	// Calculate the maximum number of pages to scavenge.
	//
	// This should be alignUp(max, pageSize) / pageSize but max can and will
	// be ^uintptr(0), so we need to be very careful not to overflow here.
	// Rather than use alignUp, calculate the number of pages rounded down
	// first, then add back one if necessary.
	maxPages := max / pageSize
	if max%pageSize != 0 {
		maxPages++
	}

	// Calculate the minimum number of pages we can scavenge.
	//
	// Because we can only scavenge whole physical pages, we must
	// ensure that we scavenge at least minPages each time, aligned
	// to minPages*pageSize.
	minPages := physPageSize / pageSize
	if minPages < 1 {
		minPages = 1
	}

	// Helpers for locking and unlocking only if mayUnlock == true.
	lockHeap := func() {
		if mayUnlock {
			lock(p.mheapLock)
		}
	}
	unlockHeap := func() {
		if mayUnlock {
			unlock(p.mheapLock)
		}
	}

	// Fast path: check the chunk containing the top-most address in work,
	// starting at that address's page index in the chunk.
	//
	// Note that work.end() is exclusive, so get the chunk we care about
	// by subtracting 1.
	maxAddr := work.limit.addr() - 1
	maxChunk := chunkIndex(maxAddr)
	if p.summary[len(p.summary)-1][maxChunk].max() >= uint(minPages) {
		// We only bother looking for a candidate if there at least
		// minPages free pages at all.
		base, npages := p.chunkOf(maxChunk).findScavengeCandidate(chunkPageIndex(maxAddr), minPages, maxPages)

		// If we found something, scavenge it and return!
		if npages != 0 {
			work.limit = offAddr{p.scavengeRangeLocked(maxChunk, base, npages)}

			assertLockHeld(p.mheapLock) // Must be locked on return.
			return uintptr(npages) * pageSize, work
		}
	}
	// Update the limit to reflect the fact that we checked maxChunk already.
	work.limit = offAddr{chunkBase(maxChunk)}

	// findCandidate finds the next scavenge candidate in work optimistically.
	//
	// Returns the candidate chunk index and true on success, and false on failure.
	//
	// The heap need not be locked.
	findCandidate := func(work addrRange) (chunkIdx, bool) {
		// Iterate over this work's chunks.
		for i := chunkIndex(work.limit.addr() - 1); i >= chunkIndex(work.base.addr()); i-- {
			// If this chunk is totally in-use or has no unscavenged pages, don't bother
			// doing a more sophisticated check.
			//
			// Note we're accessing the summary and the chunks without a lock, but
			// that's fine. We're being optimistic anyway.

			// Check quickly if there are enough free pages at all.
			if p.summary[len(p.summary)-1][i].max() < uint(minPages) {
				continue
			}

			// Run over the chunk looking harder for a candidate. Again, we could
			// race with a lot of different pieces of code, but we're just being
			// optimistic. Make sure we load the l2 pointer atomically though, to
			// avoid races with heap growth. It may or may not be possible to also
			// see a nil pointer in this case if we do race with heap growth, but
			// just defensively ignore the nils. This operation is optimistic anyway.
			l2 := (*[1 << pallocChunksL2Bits]pallocData)(atomic.Loadp(unsafe.Pointer(&p.chunks[i.l1()])))
			if l2 != nil && l2[i.l2()].hasScavengeCandidate(minPages) {
				return i, true
			}
		}
		return 0, false
	}

	// Slow path: iterate optimistically over the in-use address space
	// looking for any free and unscavenged page. If we think we see something,
	// lock and verify it!
	for work.size() != 0 {
		unlockHeap()

		// Search for the candidate.
		candidateChunkIdx, ok := findCandidate(work)

		// Lock the heap. We need to do this now if we found a candidate or not.
		// If we did, we'll verify it. If not, we need to lock before returning
		// anyway.
		lockHeap()

		if !ok {
			// We didn't find a candidate, so we're done.
			work.limit = work.base
			break
		}

		// Find, verify, and scavenge if we can.
		chunk := p.chunkOf(candidateChunkIdx)
		base, npages := chunk.findScavengeCandidate(pallocChunkPages-1, minPages, maxPages)
		if npages > 0 {
			work.limit = offAddr{p.scavengeRangeLocked(candidateChunkIdx, base, npages)}

			assertLockHeld(p.mheapLock) // Must be locked on return.
			return uintptr(npages) * pageSize, work
		}

		// We were fooled, so let's continue from where we left off.
		work.limit = offAddr{chunkBase(candidateChunkIdx)}
	}

	assertLockHeld(p.mheapLock) // Must be locked on return.
	return 0, work
}

// scavengeRangeLocked scavenges the given region of memory.
// The region of memory is described by its chunk index (ci),
// the starting page index of the region relative to that
// chunk (base), and the length of the region in pages (npages).
//
// Returns the base address of the scavenged region.
//
// p.mheapLock must be held.
func (p *pageAlloc) scavengeRangeLocked(ci chunkIdx, base, npages uint) uintptr {
	assertLockHeld(p.mheapLock)

	p.chunkOf(ci).scavenged.setRange(base, npages)

	// Compute the full address for the start of the range.
	addr := chunkBase(ci) + uintptr(base)*pageSize

	// Update the scavenge low watermark.
	if oAddr := (offAddr{addr}); oAddr.lessThan(p.scav.scavLWM) {
		p.scav.scavLWM = oAddr
	}

	// Only perform the actual scavenging if we're not in a test.
	// It's dangerous to do so otherwise.
	if p.test {
		return addr
	}
	sysUnused(unsafe.Pointer(addr), uintptr(npages)*pageSize)

	// Update global accounting only when not in test, otherwise
	// the runtime's accounting will be wrong.
	nbytes := int64(npages) * pageSize
	atomic.Xadd64(&memstats.heap_released, nbytes)

	// Update consistent accounting too.
	stats := memstats.heapStats.acquire()
	atomic.Xaddint64(&stats.committed, -nbytes)
	atomic.Xaddint64(&stats.released, nbytes)
	memstats.heapStats.release()

	return addr
}

// fillAligned returns x but with all zeroes in m-aligned
// groups of m bits set to 1 if any bit in the group is non-zero.
//
// For example, fillAligned(0x0100a3, 8) == 0xff00ff.
//
// Note that if m == 1, this is a no-op.
//
// m must be a power of 2 <= maxPagesPerPhysPage.
func fillAligned(x uint64, m uint) uint64 {
	apply := func(x uint64, c uint64) uint64 {
		// The technique used it here is derived from
		// https://graphics.stanford.edu/~seander/bithacks.html#ZeroInWord
		// and extended for more than just bytes (like nibbles
		// and uint16s) by using an appropriate constant.
		//
		// To summarize the technique, quoting from that page:
		// "[It] works by first zeroing the high bits of the [8]
		// bytes in the word. Subsequently, it adds a number that
		// will result in an overflow to the high bit of a byte if
		// any of the low bits were initially set. Next the high
		// bits of the original word are ORed with these values;
		// thus, the high bit of a byte is set iff any bit in the
		// byte was set. Finally, we determine if any of these high
		// bits are zero by ORing with ones everywhere except the
		// high bits and inverting the result."
		return ^((((x & c) + c) | x) | c)
	}
	// Transform x to contain a 1 bit at the top of each m-aligned
	// group of m zero bits.
	switch m {
	case 1:
		return x
	case 2:
		x = apply(x, 0x5555555555555555)
	case 4:
		x = apply(x, 0x7777777777777777)
	case 8:
		x = apply(x, 0x7f7f7f7f7f7f7f7f)
	case 16:
		x = apply(x, 0x7fff7fff7fff7fff)
	case 32:
		x = apply(x, 0x7fffffff7fffffff)
	case 64: // == maxPagesPerPhysPage
		x = apply(x, 0x7fffffffffffffff)
	default:
		throw("bad m value")
	}
	// Now, the top bit of each m-aligned group in x is set
	// that group was all zero in the original x.

	// From each group of m bits subtract 1.
	// Because we know only the top bits of each
	// m-aligned group are set, we know this will
	// set each group to have all the bits set except
	// the top bit, so just OR with the original
	// result to set all the bits.
	return ^((x - (x >> (m - 1))) | x)
}

// hasScavengeCandidate returns true if there's any min-page-aligned groups of
// min pages of free-and-unscavenged memory in the region represented by this
// pallocData.
//
// min must be a non-zero power of 2 <= maxPagesPerPhysPage.
func (m *pallocData) hasScavengeCandidate(min uintptr) bool {
	if min&(min-1) != 0 || min == 0 {
		print("runtime: min = ", min, "\n")
		throw("min must be a non-zero power of 2")
	} else if min > maxPagesPerPhysPage {
		print("runtime: min = ", min, "\n")
		throw("min too large")
	}

	// The goal of this search is to see if the chunk contains any free and unscavenged memory.
	for i := len(m.scavenged) - 1; i >= 0; i-- {
		// 1s are scavenged OR non-free => 0s are unscavenged AND free
		//
		// TODO(mknyszek): Consider splitting up fillAligned into two
		// functions, since here we technically could get by with just
		// the first half of its computation. It'll save a few instructions
		// but adds some additional code complexity.
		x := fillAligned(m.scavenged[i]|m.pallocBits[i], uint(min))

		// Quickly skip over chunks of non-free or scavenged pages.
		if x != ^uint64(0) {
			return true
		}
	}
	return false
}

// findScavengeCandidate returns a start index and a size for this pallocData
// segment which represents a contiguous region of free and unscavenged memory.
//
// searchIdx indicates the page index within this chunk to start the search, but
// note that findScavengeCandidate searches backwards through the pallocData. As a
// a result, it will return the highest scavenge candidate in address order.
//
// min indicates a hard minimum size and alignment for runs of pages. That is,
// findScavengeCandidate will not return a region smaller than min pages in size,
// or that is min pages or greater in size but not aligned to min. min must be
// a non-zero power of 2 <= maxPagesPerPhysPage.
//
// max is a hint for how big of a region is desired. If max >= pallocChunkPages, then
// findScavengeCandidate effectively returns entire free and unscavenged regions.
// If max < pallocChunkPages, it may truncate the returned region such that size is
// max. However, findScavengeCandidate may still return a larger region if, for
// example, it chooses to preserve huge pages, or if max is not aligned to min (it
// will round up). That is, even if max is small, the returned size is not guaranteed
// to be equal to max. max is allowed to be less than min, in which case it is as if
// max == min.
func (m *pallocData) findScavengeCandidate(searchIdx uint, min, max uintptr) (uint, uint) {
	if min&(min-1) != 0 || min == 0 {
		print("runtime: min = ", min, "\n")
		throw("min must be a non-zero power of 2")
	} else if min > maxPagesPerPhysPage {
		print("runtime: min = ", min, "\n")
		throw("min too large")
	}
	// max may not be min-aligned, so we might accidentally truncate to
	// a max value which causes us to return a non-min-aligned value.
	// To prevent this, align max up to a multiple of min (which is always
	// a power of 2). This also prevents max from ever being less than
	// min, unless it's zero, so handle that explicitly.
	if max == 0 {
		max = min
	} else {
		max = alignUp(max, min)
	}

	i := int(searchIdx / 64)
	// Start by quickly skipping over blocks of non-free or scavenged pages.
	for ; i >= 0; i-- {
		// 1s are scavenged OR non-free => 0s are unscavenged AND free
		x := fillAligned(m.scavenged[i]|m.pallocBits[i], uint(min))
		if x != ^uint64(0) {
			break
		}
	}
	if i < 0 {
		// Failed to find any free/unscavenged pages.
		return 0, 0
	}
	// We have something in the 64-bit chunk at i, but it could
	// extend further. Loop until we find the extent of it.

	// 1s are scavenged OR non-free => 0s are unscavenged AND free
	x := fillAligned(m.scavenged[i]|m.pallocBits[i], uint(min))
	z1 := uint(sys.LeadingZeros64(^x))
	run, end := uint(0), uint(i)*64+(64-z1)
	if x<<z1 != 0 {
		// After shifting out z1 bits, we still have 1s,
		// so the run ends inside this word.
		run = uint(sys.LeadingZeros64(x << z1))
	} else {
		// After shifting out z1 bits, we have no more 1s.
		// This means the run extends to the bottom of the
		// word so it may extend into further words.
		run = 64 - z1
		for j := i - 1; j >= 0; j-- {
			x := fillAligned(m.scavenged[j]|m.pallocBits[j], uint(min))
			run += uint(sys.LeadingZeros64(x))
			if x != 0 {
				// The run stopped in this word.
				break
			}
		}
	}

	// Split the run we found if it's larger than max but hold on to
	// our original length, since we may need it later.
	size := run
	if size > uint(max) {
		size = uint(max)
	}
	start := end - size

	// Each huge page is guaranteed to fit in a single palloc chunk.
	//
	// TODO(mknyszek): Support larger huge page sizes.
	// TODO(mknyszek): Consider taking pages-per-huge-page as a parameter
	// so we can write tests for this.
	if physHugePageSize > pageSize && physHugePageSize > physPageSize {
		// We have huge pages, so let's ensure we don't break one by scavenging
		// over a huge page boundary. If the range [start, start+size) overlaps with
		// a free-and-unscavenged huge page, we want to grow the region we scavenge
		// to include that huge page.

		// Compute the huge page boundary above our candidate.
		pagesPerHugePage := uintptr(physHugePageSize / pageSize)
		hugePageAbove := uint(alignUp(uintptr(start), pagesPerHugePage))

		// If that boundary is within our current candidate, then we may be breaking
		// a huge page.
		if hugePageAbove <= end {
			// Compute the huge page boundary below our candidate.
			hugePageBelow := uint(alignDown(uintptr(start), pagesPerHugePage))

			if hugePageBelow >= end-run {
				// We're in danger of breaking apart a huge page since start+size crosses
				// a huge page boundary and rounding down start to the nearest huge
				// page boundary is included in the full run we found. Include the entire
				// huge page in the bound by rounding down to the huge page size.
				size = size + (start - hugePageBelow)
				start = hugePageBelow
			}
		}
	}
	return start, size
}