func RunSchedLocalQueueEmptyTest(iters int) {
// Test that runq is not spuriously reported as empty.
// Runq emptiness affects scheduling decisions and spurious emptiness
// can lead to underutilization (both runnable Gs and idle Ps coexist
// for arbitrary long time).
done := make(chan bool, 1)
p := new(p)
gs := make([]g, 2)
ready := new(uint32)
for i := 0; i < iters; i++ {
*ready = 0
next0 := (i & 1) == 0
next1 := (i & 2) == 0
runqput(p, &gs[0], next0)
go func() {
for atomic.Xadd(ready, 1); atomic.Load(ready) != 2; {
}
if runqempty(p) {
println("next:", next0, next1)
throw("queue is empty")
}
done <- true
}()
for atomic.Xadd(ready, 1); atomic.Load(ready) != 2; {
}
runqput(p, &gs[1], next1)
runqget(p)
<-done
runqget(p)
}
}
func semrelease(addr *uint32) {
root := semroot(addr)
atomic.Xadd(addr, 1)
// Easy case: no waiters?
// This check must happen after the xadd, to avoid a missed wakeup
// (see loop in semacquire).
if atomic.Load(&root.nwait) == 0 {
return
}
// Harder case: search for a waiter and wake it.
lock(&root.lock)
if atomic.Load(&root.nwait) == 0 {
// The count is already consumed by another goroutine,
// so no need to wake up another goroutine.
unlock(&root.lock)
return
}
s := root.head
for ; s != nil; s = s.next {
if s.elem == unsafe.Pointer(addr) {
atomic.Xadd(&root.nwait, -1)
root.dequeue(s)
break
}
}
unlock(&root.lock)
if s != nil {
if s.releasetime != 0 {
s.releasetime = cputicks()
}
goready(s.g, 4)
}
}
// Called from runtime.
func semacquire(addr *uint32, profile semaProfileFlags) {
gp := getg()
if gp != gp.m.curg {
throw("semacquire not on the G stack")
}
// Easy case.
if cansemacquire(addr) {
return
}
// Harder case:
// increment waiter count
// try cansemacquire one more time, return if succeeded
// enqueue itself as a waiter
// sleep
// (waiter descriptor is dequeued by signaler)
s := acquireSudog()
root := semroot(addr)
t0 := int64(0)
s.releasetime = 0
s.acquiretime = 0
if profile&semaBlockProfile != 0 && blockprofilerate > 0 {
t0 = cputicks()
s.releasetime = -1
}
if profile&semaMutexProfile != 0 && mutexprofilerate > 0 {
if t0 == 0 {
t0 = cputicks()
}
s.acquiretime = t0
}
for {
lock(&root.lock)
// Add ourselves to nwait to disable "easy case" in semrelease.
atomic.Xadd(&root.nwait, 1)
// Check cansemacquire to avoid missed wakeup.
if cansemacquire(addr) {
atomic.Xadd(&root.nwait, -1)
unlock(&root.lock)
break
}
// Any semrelease after the cansemacquire knows we're waiting
// (we set nwait above), so go to sleep.
root.queue(addr, s)
goparkunlock(&root.lock, "semacquire", traceEvGoBlockSync, 4)
if cansemacquire(addr) {
break
}
}
if s.releasetime > 0 {
blockevent(s.releasetime-t0, 3)
}
releaseSudog(s)
}
// Get a full work buffer off the work.full list.
// If nothing is available wait until all the other gc helpers have
// finished and then return nil.
// getfull acts as a barrier for work.nproc helpers. As long as one
// gchelper is actively marking objects it
// may create a workbuffer that the other helpers can work on.
// The for loop either exits when a work buffer is found
// or when _all_ of the work.nproc GC helpers are in the loop
// looking for work and thus not capable of creating new work.
// This is in fact the termination condition for the STW mark
// phase.
//go:nowritebarrier
func getfull(entry int) *workbuf {
b := (*workbuf)(lfstackpop(&work.full))
if b != nil {
b.logget(entry)
b.checknonempty()
return b
}
incnwait := atomic.Xadd(&work.nwait, +1)
if incnwait > work.nproc {
println("runtime: work.nwait=", incnwait, "work.nproc=", work.nproc)
throw("work.nwait > work.nproc")
}
for i := 0; ; i++ {
if work.full != 0 {
decnwait := atomic.Xadd(&work.nwait, -1)
if decnwait == work.nproc {
println("runtime: work.nwait=", decnwait, "work.nproc=", work.nproc)
throw("work.nwait > work.nproc")
}
b = (*workbuf)(lfstackpop(&work.full))
if b != nil {
b.logget(entry)
b.checknonempty()
return b
}
incnwait := atomic.Xadd(&work.nwait, +1)
if incnwait > work.nproc {
println("runtime: work.nwait=", incnwait, "work.nproc=", work.nproc)
throw("work.nwait > work.nproc")
}
}
if work.nwait == work.nproc && work.markrootNext >= work.markrootJobs {
return nil
}
_g_ := getg()
if i < 10 {
_g_.m.gcstats.nprocyield++
procyield(20)
} else if i < 20 {
_g_.m.gcstats.nosyield++
osyield()
} else {
_g_.m.gcstats.nsleep++
usleep(100)
}
}
}
// sweeps one span
// returns number of pages returned to heap, or ^uintptr(0) if there is nothing to sweep
//go:nowritebarrier
func sweepone() uintptr {
_g_ := getg()
// increment locks to ensure that the goroutine is not preempted
// in the middle of sweep thus leaving the span in an inconsistent state for next GC
_g_.m.locks++
sg := mheap_.sweepgen
for {
idx := atomic.Xadd(&sweep.spanidx, 1) - 1
if idx >= uint32(len(work.spans)) {
mheap_.sweepdone = 1
_g_.m.locks--
return ^uintptr(0)
}
s := work.spans[idx]
if s.state != mSpanInUse {
s.sweepgen = sg
continue
}
if s.sweepgen != sg-2 || !atomic.Cas(&s.sweepgen, sg-2, sg-1) {
continue
}
npages := s.npages
if !s.sweep(false) {
npages = 0
}
_g_.m.locks--
return npages
}
}
// sweeps one span
// returns number of pages returned to heap, or ^uintptr(0) if there is nothing to sweep
//go:nowritebarrier
func sweepone() uintptr {
_g_ := getg()
// increment locks to ensure that the goroutine is not preempted
// in the middle of sweep thus leaving the span in an inconsistent state for next GC
_g_.m.locks++
sg := mheap_.sweepgen
for {
idx := atomic.Xadd(&sweep.spanidx, 1) - 1
if idx >= uint32(len(work.spans)) {
mheap_.sweepdone = 1
_g_.m.locks--
if debug.gcpacertrace > 0 && idx == uint32(len(work.spans)) {
print("pacer: sweep done at heap size ", memstats.heap_live>>20, "MB; allocated ", mheap_.spanBytesAlloc>>20, "MB of spans; swept ", mheap_.pagesSwept, " pages at ", mheap_.sweepPagesPerByte, " pages/byte\n")
}
return ^uintptr(0)
}
s := work.spans[idx]
if s.state != mSpanInUse {
s.sweepgen = sg
continue
}
if s.sweepgen != sg-2 || !atomic.Cas(&s.sweepgen, sg-2, sg-1) {
continue
}
npages := s.npages
if !s.sweep(false) {
npages = 0
}
_g_.m.locks--
return npages
}
}
// lockextra locks the extra list and returns the list head.
// The caller must unlock the list by storing a new list head
// to extram. If nilokay is true, then lockextra will
// return a nil list head if that's what it finds. If nilokay is false,
// lockextra will keep waiting until the list head is no longer nil.
//go:nosplit
func lockextra(nilokay bool) *m {
const locked = 1
incr := false
for {
old := atomic.Loaduintptr(&extram)
if old == locked {
yield := osyield
yield()
continue
}
if old == 0 && !nilokay {
if !incr {
// Add 1 to the number of threads
// waiting for an M.
// This is cleared by newextram.
atomic.Xadd(&extraMWaiters, 1)
incr = true
}
usleep(1)
continue
}
if atomic.Casuintptr(&extram, old, locked) {
return (*m)(unsafe.Pointer(old))
}
yield := osyield
yield()
continue
}
}
// pop removes and returns a span from buffer b, or nil if b is empty.
// pop is safe to call concurrently with other pop operations, but NOT
// to call concurrently with push.
func (b *gcSweepBuf) pop() *mspan {
cursor := atomic.Xadd(&b.index, -1)
if int32(cursor) < 0 {
atomic.Xadd(&b.index, +1)
return nil
}
// There are no concurrent spine or block modifications during
// pop, so we can omit the atomics.
top, bottom := cursor/gcSweepBlockEntries, cursor%gcSweepBlockEntries
blockp := (**gcSweepBlock)(add(b.spine, sys.PtrSize*uintptr(top)))
block := *blockp
s := block.spans[bottom]
// Clear the pointer for block(i).
block.spans[bottom] = nil
return s
}
// gcDrainN blackens grey objects until it has performed roughly
// scanWork units of scan work or the G is preempted. This is
// best-effort, so it may perform less work if it fails to get a work
// buffer. Otherwise, it will perform at least n units of work, but
// may perform more because scanning is always done in whole object
// increments. It returns the amount of scan work performed.
//
// The caller goroutine must be in a preemptible state (e.g.,
// _Gwaiting) to prevent deadlocks during stack scanning. As a
// consequence, this must be called on the system stack.
//
//go:nowritebarrier
//go:systemstack
func gcDrainN(gcw *gcWork, scanWork int64) int64 {
if !writeBarrier.needed {
throw("gcDrainN phase incorrect")
}
// There may already be scan work on the gcw, which we don't
// want to claim was done by this call.
workFlushed := -gcw.scanWork
gp := getg().m.curg
for !gp.preempt && workFlushed+gcw.scanWork < scanWork {
// See gcDrain comment.
if work.full == 0 {
gcw.balance()
}
// This might be a good place to add prefetch code...
// if(wbuf.nobj > 4) {
// PREFETCH(wbuf->obj[wbuf.nobj - 3];
// }
//
b := gcw.tryGetFast()
if b == 0 {
b = gcw.tryGet()
}
if b == 0 {
// Try to do a root job.
//
// TODO: Assists should get credit for this
// work.
if work.markrootNext < work.markrootJobs {
job := atomic.Xadd(&work.markrootNext, +1) - 1
if job < work.markrootJobs {
markroot(gcw, job)
continue
}
}
// No heap or root jobs.
break
}
scanobject(b, gcw)
// Flush background scan work credit.
if gcw.scanWork >= gcCreditSlack {
atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
workFlushed += gcw.scanWork
gcw.scanWork = 0
}
}
// Unlike gcDrain, there's no need to flush remaining work
// here because this never flushes to bgScanCredit and
// gcw.dispose will flush any remaining work to scanWork.
return workFlushed + gcw.scanWork
}
func semrelease(addr *uint32) {
root := semroot(addr)
atomic.Xadd(addr, 1)
// Easy case: no waiters?
// This check must happen after the xadd, to avoid a missed wakeup
// (see loop in semacquire).
if atomic.Load(&root.nwait) == 0 {
return
}
// Harder case: search for a waiter and wake it.
lock(&root.lock)
if atomic.Load(&root.nwait) == 0 {
// The count is already consumed by another goroutine,
// so no need to wake up another goroutine.
unlock(&root.lock)
return
}
s := root.head
for ; s != nil; s = s.next {
if s.elem == unsafe.Pointer(addr) {
atomic.Xadd(&root.nwait, -1)
root.dequeue(s)
break
}
}
if s != nil {
if s.acquiretime != 0 {
t0 := cputicks()
for x := root.head; x != nil; x = x.next {
if x.elem == unsafe.Pointer(addr) {
x.acquiretime = t0
}
}
mutexevent(t0-s.acquiretime, 3)
}
}
unlock(&root.lock)
if s != nil { // May be slow, so unlock first
readyWithTime(s, 5)
}
}
//go:nosplit
func semawakeup(mp *m) {
atomic.Xadd(&mp.waitsemacount, 1)
ret := thrwakeup(uintptr(unsafe.Pointer(&mp.waitsemacount)), 1)
if ret != 0 && ret != _ESRCH {
// semawakeup can be called on signal stack.
systemstack(func() {
print("thrwakeup addr=", &mp.waitsemacount, " sem=", mp.waitsemacount, " ret=", ret, "\n")
})
}
}
//go:nosplit
func semawakeup(mp *m) {
atomic.Xadd(&mp.waitsemacount, 1)
// From NetBSD's _lwp_unpark(2) manual:
// "If the target LWP is not currently waiting, it will return
// immediately upon the next call to _lwp_park()."
ret := lwp_unpark(int32(mp.procid), unsafe.Pointer(&mp.waitsemacount))
if ret != 0 && ret != _ESRCH {
// semawakeup can be called on signal stack.
systemstack(func() {
print("thrwakeup addr=", &mp.waitsemacount, " sem=", mp.waitsemacount, " ret=", ret, "\n")
})
}
}
func dopanic(unused int) {
gp := getg()
if gp.sig != 0 {
signame := signame(gp.sig)
if signame != "" {
print("[signal ", signame)
} else {
print("[signal ", hex(gp.sig))
}
print(" code=", hex(gp.sigcode0), " addr=", hex(gp.sigcode1), " pc=", hex(gp.sigpc), "]\n")
}
level, all, docrash := gotraceback()
_g_ := getg()
if level > 0 {
if gp != gp.m.curg {
all = true
}
if gp != gp.m.g0 {
print("\n")
goroutineheader(gp)
traceback(0)
} else if level >= 2 || _g_.m.throwing > 0 {
print("\nruntime stack:\n")
traceback(0)
}
if !didothers && all {
didothers = true
tracebackothers(gp)
}
}
unlock(&paniclk)
if atomic.Xadd(&panicking, -1) != 0 {
// Some other m is panicking too.
// Let it print what it needs to print.
// Wait forever without chewing up cpu.
// It will exit when it's done.
lock(&deadlock)
lock(&deadlock)
}
if docrash {
crash()
}
exit(2)
}
func startpanic() {
_g_ := getg()
// Uncomment when mheap_ is in Go.
// if mheap_.cachealloc.size == 0 { // very early
// print("runtime: panic before malloc heap initialized\n")
// _g_.m.mallocing = 1 // tell rest of panic not to try to malloc
// } else
if _g_.m.mcache == nil { // can happen if called from signal handler or throw
_g_.m.mcache = allocmcache()
}
switch _g_.m.dying {
case 0:
_g_.m.dying = 1
_g_.writebuf = nil
atomic.Xadd(&panicking, 1)
lock(&paniclk)
if debug.schedtrace > 0 || debug.scheddetail > 0 {
schedtrace(true)
}
freezetheworld()
return
case 1:
// Something failed while panicing, probably the print of the
// argument to panic(). Just print a stack trace and exit.
_g_.m.dying = 2
print("panic during panic\n")
dopanic(0)
exit(3)
fallthrough
case 2:
// This is a genuine bug in the runtime, we couldn't even
// print the stack trace successfully.
_g_.m.dying = 3
print("stack trace unavailable\n")
exit(4)
fallthrough
default:
// Can't even print! Just exit.
exit(5)
}
}
func queuefinalizer(p unsafe.Pointer, fn *funcval, nret uintptr, fint *_type, ot *ptrtype) {
lock(&finlock)
if finq == nil || finq.cnt == uint32(len(finq.fin)) {
if finc == nil {
finc = (*finblock)(persistentalloc(_FinBlockSize, 0, &memstats.gc_sys))
finc.alllink = allfin
allfin = finc
if finptrmask[0] == 0 {
// Build pointer mask for Finalizer array in block.
// Check assumptions made in finalizer1 array above.
if (unsafe.Sizeof(finalizer{}) != 5*sys.PtrSize ||
unsafe.Offsetof(finalizer{}.fn) != 0 ||
unsafe.Offsetof(finalizer{}.arg) != sys.PtrSize ||
unsafe.Offsetof(finalizer{}.nret) != 2*sys.PtrSize ||
unsafe.Offsetof(finalizer{}.fint) != 3*sys.PtrSize ||
unsafe.Offsetof(finalizer{}.ot) != 4*sys.PtrSize) {
throw("finalizer out of sync")
}
for i := range finptrmask {
finptrmask[i] = finalizer1[i%len(finalizer1)]
}
}
}
block := finc
finc = block.next
block.next = finq
finq = block
}
f := &finq.fin[finq.cnt]
atomic.Xadd(&finq.cnt, +1) // Sync with markroots
f.fn = fn
f.nret = nret
f.fint = fint
f.ot = ot
f.arg = p
fingwake = true
unlock(&finlock)
}
// gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system
// stack. This is a separate function to make it easier to see that
// we're not capturing anything from the user stack, since the user
// stack may move while we're in this function.
//
// gcAssistAlloc1 indicates whether this assist completed the mark
// phase by setting gp.param to non-nil. This can't be communicated on
// the stack since it may move.
//
//go:systemstack
func gcAssistAlloc1(gp *g, scanWork int64) {
// Clear the flag indicating that this assist completed the
// mark phase.
gp.param = nil
if atomic.Load(&gcBlackenEnabled) == 0 {
// The gcBlackenEnabled check in malloc races with the
// store that clears it but an atomic check in every malloc
// would be a performance hit.
// Instead we recheck it here on the non-preemptable system
// stack to determine if we should preform an assist.
// GC is done, so ignore any remaining debt.
gp.gcAssistBytes = 0
return
}
// Track time spent in this assist. Since we're on the
// system stack, this is non-preemptible, so we can
// just measure start and end time.
startTime := nanotime()
decnwait := atomic.Xadd(&work.nwait, -1)
if decnwait == work.nproc {
println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
throw("nwait > work.nprocs")
}
// gcDrainN requires the caller to be preemptible.
casgstatus(gp, _Grunning, _Gwaiting)
gp.waitreason = "GC assist marking"
// drain own cached work first in the hopes that it
// will be more cache friendly.
gcw := &getg().m.p.ptr().gcw
workDone := gcDrainN(gcw, scanWork)
// If we are near the end of the mark phase
// dispose of the gcw.
if gcBlackenPromptly {
gcw.dispose()
}
casgstatus(gp, _Gwaiting, _Grunning)
// Record that we did this much scan work.
//
// Back out the number of bytes of assist credit that
// this scan work counts for. The "1+" is a poor man's
// round-up, to ensure this adds credit even if
// assistBytesPerWork is very low.
gp.gcAssistBytes += 1 + int64(gcController.assistBytesPerWork*float64(workDone))
// If this is the last worker and we ran out of work,
// signal a completion point.
incnwait := atomic.Xadd(&work.nwait, +1)
if incnwait > work.nproc {
println("runtime: work.nwait=", incnwait,
"work.nproc=", work.nproc,
"gcBlackenPromptly=", gcBlackenPromptly)
throw("work.nwait > work.nproc")
}
if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
// This has reached a background completion point. Set
// gp.param to a non-nil value to indicate this. It
// doesn't matter what we set it to (it just has to be
// a valid pointer).
gp.param = unsafe.Pointer(gp)
}
duration := nanotime() - startTime
_p_ := gp.m.p.ptr()
_p_.gcAssistTime += duration
if _p_.gcAssistTime > gcAssistTimeSlack {
atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
_p_.gcAssistTime = 0
}
}
// gcDrain scans roots and objects in work buffers, blackening grey
// objects until all roots and work buffers have been drained.
//
// If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
// is set. This implies gcDrainNoBlock.
//
// If flags&gcDrainIdle != 0, gcDrain returns when there is other work
// to do. This implies gcDrainNoBlock.
//
// If flags&gcDrainNoBlock != 0, gcDrain returns as soon as it is
// unable to get more work. Otherwise, it will block until all
// blocking calls are blocked in gcDrain.
//
// If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
// credit to gcController.bgScanCredit every gcCreditSlack units of
// scan work.
//
//go:nowritebarrier
func gcDrain(gcw *gcWork, flags gcDrainFlags) {
if !writeBarrier.needed {
throw("gcDrain phase incorrect")
}
gp := getg().m.curg
preemptible := flags&gcDrainUntilPreempt != 0
blocking := flags&(gcDrainUntilPreempt|gcDrainIdle|gcDrainNoBlock) == 0
flushBgCredit := flags&gcDrainFlushBgCredit != 0
idle := flags&gcDrainIdle != 0
initScanWork := gcw.scanWork
// idleCheck is the scan work at which to perform the next
// idle check with the scheduler.
idleCheck := initScanWork + idleCheckThreshold
// Drain root marking jobs.
if work.markrootNext < work.markrootJobs {
for !(preemptible && gp.preempt) {
job := atomic.Xadd(&work.markrootNext, +1) - 1
if job >= work.markrootJobs {
break
}
markroot(gcw, job)
if idle && pollWork() {
goto done
}
}
}
// Drain heap marking jobs.
for !(preemptible && gp.preempt) {
// Try to keep work available on the global queue. We used to
// check if there were waiting workers, but it's better to
// just keep work available than to make workers wait. In the
// worst case, we'll do O(log(_WorkbufSize)) unnecessary
// balances.
if work.full == 0 {
gcw.balance()
}
var b uintptr
if blocking {
b = gcw.get()
} else {
b = gcw.tryGetFast()
if b == 0 {
b = gcw.tryGet()
}
}
if b == 0 {
// work barrier reached or tryGet failed.
break
}
scanobject(b, gcw)
// Flush background scan work credit to the global
// account if we've accumulated enough locally so
// mutator assists can draw on it.
if gcw.scanWork >= gcCreditSlack {
atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
if flushBgCredit {
gcFlushBgCredit(gcw.scanWork - initScanWork)
initScanWork = 0
}
idleCheck -= gcw.scanWork
gcw.scanWork = 0
if idle && idleCheck <= 0 {
idleCheck += idleCheckThreshold
if pollWork() {
break
}
}
}
}
// In blocking mode, write barriers are not allowed after this
// point because we must preserve the condition that the work
// buffers are empty.
done:
// Flush remaining scan work credit.
if gcw.scanWork > 0 {
atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
if flushBgCredit {
gcFlushBgCredit(gcw.scanWork - initScanWork)
}
gcw.scanWork = 0
}
}
// push adds span s to buffer b. push is safe to call concurrently
// with other push operations, but NOT to call concurrently with pop.
func (b *gcSweepBuf) push(s *mspan) {
// Obtain our slot.
cursor := uintptr(atomic.Xadd(&b.index, +1) - 1)
top, bottom := cursor/gcSweepBlockEntries, cursor%gcSweepBlockEntries
// Do we need to add a block?
spineLen := atomic.Loaduintptr(&b.spineLen)
var block *gcSweepBlock
retry:
if top < spineLen {
spine := atomic.Loadp(unsafe.Pointer(&b.spine))
blockp := add(spine, sys.PtrSize*top)
block = (*gcSweepBlock)(atomic.Loadp(blockp))
} else {
// Add a new block to the spine, potentially growing
// the spine.
lock(&b.spineLock)
// spineLen cannot change until we release the lock,
// but may have changed while we were waiting.
spineLen = atomic.Loaduintptr(&b.spineLen)
if top < spineLen {
unlock(&b.spineLock)
goto retry
}
if spineLen == b.spineCap {
// Grow the spine.
newCap := b.spineCap * 2
if newCap == 0 {
newCap = gcSweepBufInitSpineCap
}
newSpine := persistentalloc(newCap*sys.PtrSize, sys.CacheLineSize, &memstats.gc_sys)
if b.spineCap != 0 {
// Blocks are allocated off-heap, so
// no write barriers.
memmove(newSpine, b.spine, b.spineCap*sys.PtrSize)
}
// Spine is allocated off-heap, so no write barrier.
atomic.StorepNoWB(unsafe.Pointer(&b.spine), newSpine)
b.spineCap = newCap
// We can't immediately free the old spine
// since a concurrent push with a lower index
// could still be reading from it. We let it
// leak because even a 1TB heap would waste
// less than 2MB of memory on old spines. If
// this is a problem, we could free old spines
// during STW.
}
// Allocate a new block and add it to the spine.
block = (*gcSweepBlock)(persistentalloc(unsafe.Sizeof(gcSweepBlock{}), sys.CacheLineSize, &memstats.gc_sys))
blockp := add(b.spine, sys.PtrSize*top)
// Blocks are allocated off-heap, so no write barrier.
atomic.StorepNoWB(blockp, unsafe.Pointer(block))
atomic.Storeuintptr(&b.spineLen, spineLen+1)
unlock(&b.spineLock)
}
// We have a block. Insert the span.
block.spans[bottom] = s
}
// notifyListAdd adds the caller to a notify list such that it can receive
// notifications. The caller must eventually call notifyListWait to wait for
// such a notification, passing the returned ticket number.
//go:linkname notifyListAdd sync.runtime_notifyListAdd
func notifyListAdd(l *notifyList) uint32 {
// This may be called concurrently, for example, when called from
// sync.Cond.Wait while holding a RWMutex in read mode.
return atomic.Xadd(&l.wait, 1) - 1
}
// gcDrain scans roots and objects in work buffers, blackening grey
// objects until all roots and work buffers have been drained.
//
// If flags&gcDrainUntilPreempt != 0, gcDrain returns when g.preempt
// is set. This implies gcDrainNoBlock.
//
// If flags&gcDrainNoBlock != 0, gcDrain returns as soon as it is
// unable to get more work. Otherwise, it will block until all
// blocking calls are blocked in gcDrain.
//
// If flags&gcDrainFlushBgCredit != 0, gcDrain flushes scan work
// credit to gcController.bgScanCredit every gcCreditSlack units of
// scan work.
//
//go:nowritebarrier
func gcDrain(gcw *gcWork, flags gcDrainFlags) {
if !writeBarrier.needed {
throw("gcDrain phase incorrect")
}
gp := getg()
preemtible := flags&gcDrainUntilPreempt != 0
blocking := flags&(gcDrainUntilPreempt|gcDrainNoBlock) == 0
flushBgCredit := flags&gcDrainFlushBgCredit != 0
// Drain root marking jobs.
if work.markrootNext < work.markrootJobs {
for blocking || !gp.preempt {
job := atomic.Xadd(&work.markrootNext, +1) - 1
if job >= work.markrootJobs {
break
}
// TODO: Pass in gcw.
markroot(job)
}
}
initScanWork := gcw.scanWork
// Drain heap marking jobs.
for !(preemtible && gp.preempt) {
// If another proc wants a pointer, give it some.
if work.nwait > 0 && work.full == 0 {
gcw.balance()
}
var b uintptr
if blocking {
b = gcw.get()
} else {
b = gcw.tryGet()
}
if b == 0 {
// work barrier reached or tryGet failed.
break
}
// If the current wbuf is filled by the scan a new wbuf might be
// returned that could possibly hold only a single object. This
// could result in each iteration draining only a single object
// out of the wbuf passed in + a single object placed
// into an empty wbuf in scanobject so there could be
// a performance hit as we keep fetching fresh wbufs.
scanobject(b, gcw)
// Flush background scan work credit to the global
// account if we've accumulated enough locally so
// mutator assists can draw on it.
if gcw.scanWork >= gcCreditSlack {
atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
if flushBgCredit {
gcFlushBgCredit(gcw.scanWork - initScanWork)
initScanWork = 0
}
gcw.scanWork = 0
}
}
// In blocking mode, write barriers are not allowed after this
// point because we must preserve the condition that the work
// buffers are empty.
// Flush remaining scan work credit.
if gcw.scanWork > 0 {
atomic.Xaddint64(&gcController.scanWork, gcw.scanWork)
if flushBgCredit {
gcFlushBgCredit(gcw.scanWork - initScanWork)
}
gcw.scanWork = 0
}
}
func parfordo(desc *parfor) {
// Obtain 0-based thread index.
tid := atomic.Xadd(&desc.thrseq, 1) - 1
if tid >= desc.nthr {
print("tid=", tid, " nthr=", desc.nthr, "\n")
throw("parfor: invalid tid")
}
// If single-threaded, just execute the for serially.
body := desc.body
if desc.nthr == 1 {
for i := uint32(0); i < desc.cnt; i++ {
body(desc, i)
}
return
}
me := &desc.thr[tid]
mypos := &me.pos
for {
for {
// While there is local work,
// bump low index and execute the iteration.
pos := atomic.Xadd64(mypos, 1)
begin := uint32(pos) - 1
end := uint32(pos >> 32)
if begin < end {
body(desc, begin)
continue
}
break
}
// Out of work, need to steal something.
idle := false
for try := uint32(0); ; try++ {
// If we don't see any work for long enough,
// increment the done counter...
if try > desc.nthr*4 && !idle {
idle = true
atomic.Xadd(&desc.done, 1)
}
// ...if all threads have incremented the counter,
// we are done.
extra := uint32(0)
if !idle {
extra = 1
}
if desc.done+extra == desc.nthr {
if !idle {
atomic.Xadd(&desc.done, 1)
}
goto exit
}
// Choose a random victim for stealing.
var begin, end uint32
victim := fastrand1() % (desc.nthr - 1)
if victim >= tid {
victim++
}
victimpos := &desc.thr[victim].pos
for {
// See if it has any work.
pos := atomic.Load64(victimpos)
begin = uint32(pos)
end = uint32(pos >> 32)
if begin+1 >= end {
end = 0
begin = end
break
}
if idle {
atomic.Xadd(&desc.done, -1)
idle = false
}
begin2 := begin + (end-begin)/2
newpos := uint64(begin) | uint64(begin2)<<32
if atomic.Cas64(victimpos, pos, newpos) {
begin = begin2
break
}
}
if begin < end {
// Has successfully stolen some work.
if idle {
throw("parfor: should not be idle")
}
atomic.Store64(mypos, uint64(begin)|uint64(end)<<32)
me.nsteal++
me.nstealcnt += uint64(end) - uint64(begin)
break
}
// Backoff.
if try < desc.nthr {
// nothing
} else if try < 4*desc.nthr {
me.nprocyield++
procyield(20)
} else if !desc.wait {
// If a caller asked not to wait for the others, exit now
// (assume that most work is already done at this point).
if !idle {
atomic.Xadd(&desc.done, 1)
}
goto exit
} else if try < 6*desc.nthr {
me.nosyield++
osyield()
} else {
me.nsleep++
usleep(1)
}
}
}
exit:
atomic.Xadd64(&desc.nsteal, int64(me.nsteal))
atomic.Xadd64(&desc.nstealcnt, int64(me.nstealcnt))
atomic.Xadd64(&desc.nprocyield, int64(me.nprocyield))
atomic.Xadd64(&desc.nosyield, int64(me.nosyield))
atomic.Xadd64(&desc.nsleep, int64(me.nsleep))
me.nsteal = 0
me.nstealcnt = 0
me.nprocyield = 0
me.nosyield = 0
me.nsleep = 0
}
// gcAssistAlloc performs GC work to make gp's assist debt positive.
// gp must be the calling user gorountine.
//
// This must be called with preemption enabled.
//go:nowritebarrier
func gcAssistAlloc(gp *g) {
// Don't assist in non-preemptible contexts. These are
// generally fragile and won't allow the assist to block.
if getg() == gp.m.g0 {
return
}
if mp := getg().m; mp.locks > 0 || mp.preemptoff != "" {
return
}
// Compute the amount of scan work we need to do to make the
// balance positive. We over-assist to build up credit for
// future allocations and amortize the cost of assisting.
debtBytes := -gp.gcAssistBytes + gcOverAssistBytes
scanWork := int64(gcController.assistWorkPerByte * float64(debtBytes))
retry:
// Steal as much credit as we can from the background GC's
// scan credit. This is racy and may drop the background
// credit below 0 if two mutators steal at the same time. This
// will just cause steals to fail until credit is accumulated
// again, so in the long run it doesn't really matter, but we
// do have to handle the negative credit case.
bgScanCredit := atomic.Loadint64(&gcController.bgScanCredit)
stolen := int64(0)
if bgScanCredit > 0 {
if bgScanCredit < scanWork {
stolen = bgScanCredit
gp.gcAssistBytes += 1 + int64(gcController.assistBytesPerWork*float64(stolen))
} else {
stolen = scanWork
gp.gcAssistBytes += debtBytes
}
atomic.Xaddint64(&gcController.bgScanCredit, -stolen)
scanWork -= stolen
if scanWork == 0 {
// We were able to steal all of the credit we
// needed.
return
}
}
// Perform assist work
completed := false
systemstack(func() {
if atomic.Load(&gcBlackenEnabled) == 0 {
// The gcBlackenEnabled check in malloc races with the
// store that clears it but an atomic check in every malloc
// would be a performance hit.
// Instead we recheck it here on the non-preemptable system
// stack to determine if we should preform an assist.
// GC is done, so ignore any remaining debt.
gp.gcAssistBytes = 0
return
}
// Track time spent in this assist. Since we're on the
// system stack, this is non-preemptible, so we can
// just measure start and end time.
startTime := nanotime()
decnwait := atomic.Xadd(&work.nwait, -1)
if decnwait == work.nproc {
println("runtime: work.nwait =", decnwait, "work.nproc=", work.nproc)
throw("nwait > work.nprocs")
}
// drain own cached work first in the hopes that it
// will be more cache friendly.
gcw := &getg().m.p.ptr().gcw
workDone := gcDrainN(gcw, scanWork)
// If we are near the end of the mark phase
// dispose of the gcw.
if gcBlackenPromptly {
gcw.dispose()
}
// Record that we did this much scan work.
//
// Back out the number of bytes of assist credit that
// this scan work counts for. The "1+" is a poor man's
// round-up, to ensure this adds credit even if
// assistBytesPerWork is very low.
gp.gcAssistBytes += 1 + int64(gcController.assistBytesPerWork*float64(workDone))
// If this is the last worker and we ran out of work,
// signal a completion point.
incnwait := atomic.Xadd(&work.nwait, +1)
if incnwait > work.nproc {
println("runtime: work.nwait=", incnwait,
"work.nproc=", work.nproc,
"gcBlackenPromptly=", gcBlackenPromptly)
throw("work.nwait > work.nproc")
}
if incnwait == work.nproc && !gcMarkWorkAvailable(nil) {
// This has reached a background completion
// point.
completed = true
}
duration := nanotime() - startTime
_p_ := gp.m.p.ptr()
_p_.gcAssistTime += duration
if _p_.gcAssistTime > gcAssistTimeSlack {
atomic.Xaddint64(&gcController.assistTime, _p_.gcAssistTime)
_p_.gcAssistTime = 0
}
})
if completed {
gcMarkDone()
}
if gp.gcAssistBytes < 0 {
// We were unable steal enough credit or perform
// enough work to pay off the assist debt. We need to
// do one of these before letting the mutator allocate
// more to prevent over-allocation.
//
// If this is because we were preempted, reschedule
// and try some more.
if gp.preempt {
Gosched()
goto retry
}
// Add this G to an assist queue and park. When the GC
// has more background credit, it will satisfy queued
// assists before flushing to the global credit pool.
//
// Note that this does *not* get woken up when more
// work is added to the work list. The theory is that
// there wasn't enough work to do anyway, so we might
// as well let background marking take care of the
// work that is available.
lock(&work.assistQueue.lock)
// If the GC cycle is over, just return. This is the
// likely path if we completed above. We do this
// under the lock to prevent a GC cycle from ending
// between this check and queuing the assist.
if atomic.Load(&gcBlackenEnabled) == 0 {
unlock(&work.assistQueue.lock)
return
}
oldHead, oldTail := work.assistQueue.head, work.assistQueue.tail
if oldHead == 0 {
work.assistQueue.head.set(gp)
} else {
oldTail.ptr().schedlink.set(gp)
}
work.assistQueue.tail.set(gp)
gp.schedlink.set(nil)
// Recheck for background credit now that this G is in
// the queue, but can still back out. This avoids a
// race in case background marking has flushed more
// credit since we checked above.
if atomic.Loadint64(&gcController.bgScanCredit) > 0 {
work.assistQueue.head = oldHead
work.assistQueue.tail = oldTail
if oldTail != 0 {
oldTail.ptr().schedlink.set(nil)
}
unlock(&work.assistQueue.lock)
goto retry
}
// Park for real.
goparkunlock(&work.assistQueue.lock, "GC assist wait", traceEvGoBlock, 2)
// At this point either background GC has satisfied
// this G's assist debt, or the GC cycle is over.
}
}
