1 // Copyright 2014 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
11 "runtime/internal/atomic"
12 "runtime/internal/sys"
16 // set using cmd/go/internal/modload.ModInfoProg
19 // Goroutine scheduler
20 // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
22 // The main concepts are:
24 // M - worker thread, or machine.
25 // P - processor, a resource that is required to execute Go code.
26 // M must have an associated P to execute Go code, however it can be
27 // blocked or in a syscall w/o an associated P.
29 // Design doc at https://golang.org/s/go11sched.
31 // Worker thread parking/unparking.
32 // We need to balance between keeping enough running worker threads to utilize
33 // available hardware parallelism and parking excessive running worker threads
34 // to conserve CPU resources and power. This is not simple for two reasons:
35 // (1) scheduler state is intentionally distributed (in particular, per-P work
36 // queues), so it is not possible to compute global predicates on fast paths;
37 // (2) for optimal thread management we would need to know the future (don't park
38 // a worker thread when a new goroutine will be readied in near future).
40 // Three rejected approaches that would work badly:
41 // 1. Centralize all scheduler state (would inhibit scalability).
42 // 2. Direct goroutine handoff. That is, when we ready a new goroutine and there
43 // is a spare P, unpark a thread and handoff it the thread and the goroutine.
44 // This would lead to thread state thrashing, as the thread that readied the
45 // goroutine can be out of work the very next moment, we will need to park it.
46 // Also, it would destroy locality of computation as we want to preserve
47 // dependent goroutines on the same thread; and introduce additional latency.
48 // 3. Unpark an additional thread whenever we ready a goroutine and there is an
49 // idle P, but don't do handoff. This would lead to excessive thread parking/
50 // unparking as the additional threads will instantly park without discovering
53 // The current approach:
55 // This approach applies to three primary sources of potential work: readying a
56 // goroutine, new/modified-earlier timers, and idle-priority GC. See below for
57 // additional details.
59 // We unpark an additional thread when we submit work if (this is wakep()):
60 // 1. There is an idle P, and
61 // 2. There are no "spinning" worker threads.
63 // A worker thread is considered spinning if it is out of local work and did
64 // not find work in the global run queue or netpoller; the spinning state is
65 // denoted in m.spinning and in sched.nmspinning. Threads unparked this way are
66 // also considered spinning; we don't do goroutine handoff so such threads are
67 // out of work initially. Spinning threads spin on looking for work in per-P
68 // run queues and timer heaps or from the GC before parking. If a spinning
69 // thread finds work it takes itself out of the spinning state and proceeds to
70 // execution. If it does not find work it takes itself out of the spinning
71 // state and then parks.
73 // If there is at least one spinning thread (sched.nmspinning>1), we don't
74 // unpark new threads when submitting work. To compensate for that, if the last
75 // spinning thread finds work and stops spinning, it must unpark a new spinning
76 // thread. This approach smooths out unjustified spikes of thread unparking,
77 // but at the same time guarantees eventual maximal CPU parallelism
80 // The main implementation complication is that we need to be very careful
81 // during spinning->non-spinning thread transition. This transition can race
82 // with submission of new work, and either one part or another needs to unpark
83 // another worker thread. If they both fail to do that, we can end up with
84 // semi-persistent CPU underutilization.
86 // The general pattern for submission is:
87 // 1. Submit work to the local run queue, timer heap, or GC state.
88 // 2. #StoreLoad-style memory barrier.
89 // 3. Check sched.nmspinning.
91 // The general pattern for spinning->non-spinning transition is:
92 // 1. Decrement nmspinning.
93 // 2. #StoreLoad-style memory barrier.
94 // 3. Check all per-P work queues and GC for new work.
96 // Note that all this complexity does not apply to global run queue as we are
97 // not sloppy about thread unparking when submitting to global queue. Also see
98 // comments for nmspinning manipulation.
100 // How these different sources of work behave varies, though it doesn't affect
101 // the synchronization approach:
102 // * Ready goroutine: this is an obvious source of work; the goroutine is
103 // immediately ready and must run on some thread eventually.
104 // * New/modified-earlier timer: The current timer implementation (see time.go)
105 // uses netpoll in a thread with no work available to wait for the soonest
106 // timer. If there is no thread waiting, we want a new spinning thread to go
108 // * Idle-priority GC: The GC wakes a stopped idle thread to contribute to
109 // background GC work (note: currently disabled per golang.org/issue/19112).
110 // Also see golang.org/issue/44313, as this should be extended to all GC
120 //go:linkname runtime_inittask runtime..inittask
121 var runtime_inittask initTask
123 //go:linkname main_inittask main..inittask
124 var main_inittask initTask
126 // main_init_done is a signal used by cgocallbackg that initialization
127 // has been completed. It is made before _cgo_notify_runtime_init_done,
128 // so all cgo calls can rely on it existing. When main_init is complete,
129 // it is closed, meaning cgocallbackg can reliably receive from it.
130 var main_init_done chan bool
132 //go:linkname main_main main.main
135 // mainStarted indicates that the main M has started.
138 // runtimeInitTime is the nanotime() at which the runtime started.
139 var runtimeInitTime int64
141 // Value to use for signal mask for newly created M's.
142 var initSigmask sigset
144 // The main goroutine.
148 // Racectx of m0->g0 is used only as the parent of the main goroutine.
149 // It must not be used for anything else.
152 // Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
153 // Using decimal instead of binary GB and MB because
154 // they look nicer in the stack overflow failure message.
155 if goarch.PtrSize == 8 {
156 maxstacksize = 1000000000
158 maxstacksize = 250000000
161 // An upper limit for max stack size. Used to avoid random crashes
162 // after calling SetMaxStack and trying to allocate a stack that is too big,
163 // since stackalloc works with 32-bit sizes.
164 maxstackceiling = 2 * maxstacksize
166 // Allow newproc to start new Ms.
169 if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
171 newm(sysmon, nil, -1)
175 // Lock the main goroutine onto this, the main OS thread,
176 // during initialization. Most programs won't care, but a few
177 // do require certain calls to be made by the main thread.
178 // Those can arrange for main.main to run in the main thread
179 // by calling runtime.LockOSThread during initialization
180 // to preserve the lock.
184 throw("runtime.main not on m0")
187 // Record when the world started.
188 // Must be before doInit for tracing init.
189 runtimeInitTime = nanotime()
190 if runtimeInitTime == 0 {
191 throw("nanotime returning zero")
194 if debug.inittrace != 0 {
195 inittrace.id = getg().goid
196 inittrace.active = true
199 doInit(&runtime_inittask) // Must be before defer.
201 // Defer unlock so that runtime.Goexit during init does the unlock too.
211 main_init_done = make(chan bool)
213 if _cgo_thread_start == nil {
214 throw("_cgo_thread_start missing")
216 if GOOS != "windows" {
217 if _cgo_setenv == nil {
218 throw("_cgo_setenv missing")
220 if _cgo_unsetenv == nil {
221 throw("_cgo_unsetenv missing")
224 if _cgo_notify_runtime_init_done == nil {
225 throw("_cgo_notify_runtime_init_done missing")
227 // Start the template thread in case we enter Go from
228 // a C-created thread and need to create a new thread.
229 startTemplateThread()
230 cgocall(_cgo_notify_runtime_init_done, nil)
233 doInit(&main_inittask)
235 // Disable init tracing after main init done to avoid overhead
236 // of collecting statistics in malloc and newproc
237 inittrace.active = false
239 close(main_init_done)
244 if isarchive || islibrary {
245 // A program compiled with -buildmode=c-archive or c-shared
246 // has a main, but it is not executed.
249 fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
255 // Make racy client program work: if panicking on
256 // another goroutine at the same time as main returns,
257 // let the other goroutine finish printing the panic trace.
258 // Once it does, it will exit. See issues 3934 and 20018.
259 if runningPanicDefers.Load() != 0 {
260 // Running deferred functions should not take long.
261 for c := 0; c < 1000; c++ {
262 if runningPanicDefers.Load() == 0 {
268 if panicking.Load() != 0 {
269 gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
279 // os_beforeExit is called from os.Exit(0).
281 //go:linkname os_beforeExit os.runtime_beforeExit
282 func os_beforeExit() {
288 // start forcegc helper goroutine
293 func forcegchelper() {
295 lockInit(&forcegc.lock, lockRankForcegc)
298 if forcegc.idle != 0 {
299 throw("forcegc: phase error")
301 atomic.Store(&forcegc.idle, 1)
302 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
303 // this goroutine is explicitly resumed by sysmon
304 if debug.gctrace > 0 {
307 // Time-triggered, fully concurrent.
308 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
314 // Gosched yields the processor, allowing other goroutines to run. It does not
315 // suspend the current goroutine, so execution resumes automatically.
321 // goschedguarded yields the processor like gosched, but also checks
322 // for forbidden states and opts out of the yield in those cases.
325 func goschedguarded() {
326 mcall(goschedguarded_m)
329 // Puts the current goroutine into a waiting state and calls unlockf on the
332 // If unlockf returns false, the goroutine is resumed.
334 // unlockf must not access this G's stack, as it may be moved between
335 // the call to gopark and the call to unlockf.
337 // Note that because unlockf is called after putting the G into a waiting
338 // state, the G may have already been readied by the time unlockf is called
339 // unless there is external synchronization preventing the G from being
340 // readied. If unlockf returns false, it must guarantee that the G cannot be
341 // externally readied.
343 // Reason explains why the goroutine has been parked. It is displayed in stack
344 // traces and heap dumps. Reasons should be unique and descriptive. Do not
345 // re-use reasons, add new ones.
346 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
347 if reason != waitReasonSleep {
348 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
352 status := readgstatus(gp)
353 if status != _Grunning && status != _Gscanrunning {
354 throw("gopark: bad g status")
357 mp.waitunlockf = unlockf
358 gp.waitreason = reason
359 mp.waittraceev = traceEv
360 mp.waittraceskip = traceskip
362 // can't do anything that might move the G between Ms here.
366 // Puts the current goroutine into a waiting state and unlocks the lock.
367 // The goroutine can be made runnable again by calling goready(gp).
368 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
369 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
372 func goready(gp *g, traceskip int) {
374 ready(gp, traceskip, true)
379 func acquireSudog() *sudog {
380 // Delicate dance: the semaphore implementation calls
381 // acquireSudog, acquireSudog calls new(sudog),
382 // new calls malloc, malloc can call the garbage collector,
383 // and the garbage collector calls the semaphore implementation
385 // Break the cycle by doing acquirem/releasem around new(sudog).
386 // The acquirem/releasem increments m.locks during new(sudog),
387 // which keeps the garbage collector from being invoked.
390 if len(pp.sudogcache) == 0 {
391 lock(&sched.sudoglock)
392 // First, try to grab a batch from central cache.
393 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
394 s := sched.sudogcache
395 sched.sudogcache = s.next
397 pp.sudogcache = append(pp.sudogcache, s)
399 unlock(&sched.sudoglock)
400 // If the central cache is empty, allocate a new one.
401 if len(pp.sudogcache) == 0 {
402 pp.sudogcache = append(pp.sudogcache, new(sudog))
405 n := len(pp.sudogcache)
406 s := pp.sudogcache[n-1]
407 pp.sudogcache[n-1] = nil
408 pp.sudogcache = pp.sudogcache[:n-1]
410 throw("acquireSudog: found s.elem != nil in cache")
417 func releaseSudog(s *sudog) {
419 throw("runtime: sudog with non-nil elem")
422 throw("runtime: sudog with non-false isSelect")
425 throw("runtime: sudog with non-nil next")
428 throw("runtime: sudog with non-nil prev")
430 if s.waitlink != nil {
431 throw("runtime: sudog with non-nil waitlink")
434 throw("runtime: sudog with non-nil c")
438 throw("runtime: releaseSudog with non-nil gp.param")
440 mp := acquirem() // avoid rescheduling to another P
442 if len(pp.sudogcache) == cap(pp.sudogcache) {
443 // Transfer half of local cache to the central cache.
444 var first, last *sudog
445 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
446 n := len(pp.sudogcache)
447 p := pp.sudogcache[n-1]
448 pp.sudogcache[n-1] = nil
449 pp.sudogcache = pp.sudogcache[:n-1]
457 lock(&sched.sudoglock)
458 last.next = sched.sudogcache
459 sched.sudogcache = first
460 unlock(&sched.sudoglock)
462 pp.sudogcache = append(pp.sudogcache, s)
466 // called from assembly
467 func badmcall(fn func(*g)) {
468 throw("runtime: mcall called on m->g0 stack")
471 func badmcall2(fn func(*g)) {
472 throw("runtime: mcall function returned")
475 func badreflectcall() {
476 panic(plainError("arg size to reflect.call more than 1GB"))
479 var badmorestackg0Msg = "fatal: morestack on g0\n"
482 //go:nowritebarrierrec
483 func badmorestackg0() {
484 sp := stringStructOf(&badmorestackg0Msg)
485 write(2, sp.str, int32(sp.len))
488 var badmorestackgsignalMsg = "fatal: morestack on gsignal\n"
491 //go:nowritebarrierrec
492 func badmorestackgsignal() {
493 sp := stringStructOf(&badmorestackgsignalMsg)
494 write(2, sp.str, int32(sp.len))
502 func lockedOSThread() bool {
504 return gp.lockedm != 0 && gp.m.lockedg != 0
508 // allgs contains all Gs ever created (including dead Gs), and thus
511 // Access via the slice is protected by allglock or stop-the-world.
512 // Readers that cannot take the lock may (carefully!) use the atomic
517 // allglen and allgptr are atomic variables that contain len(allgs) and
518 // &allgs[0] respectively. Proper ordering depends on totally-ordered
519 // loads and stores. Writes are protected by allglock.
521 // allgptr is updated before allglen. Readers should read allglen
522 // before allgptr to ensure that allglen is always <= len(allgptr). New
523 // Gs appended during the race can be missed. For a consistent view of
524 // all Gs, allglock must be held.
526 // allgptr copies should always be stored as a concrete type or
527 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
528 // even if it points to a stale array.
533 func allgadd(gp *g) {
534 if readgstatus(gp) == _Gidle {
535 throw("allgadd: bad status Gidle")
539 allgs = append(allgs, gp)
540 if &allgs[0] != allgptr {
541 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
543 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
547 // allGsSnapshot returns a snapshot of the slice of all Gs.
549 // The world must be stopped or allglock must be held.
550 func allGsSnapshot() []*g {
551 assertWorldStoppedOrLockHeld(&allglock)
553 // Because the world is stopped or allglock is held, allgadd
554 // cannot happen concurrently with this. allgs grows
555 // monotonically and existing entries never change, so we can
556 // simply return a copy of the slice header. For added safety,
557 // we trim everything past len because that can still change.
558 return allgs[:len(allgs):len(allgs)]
561 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
562 func atomicAllG() (**g, uintptr) {
563 length := atomic.Loaduintptr(&allglen)
564 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
568 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
569 func atomicAllGIndex(ptr **g, i uintptr) *g {
570 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
573 // forEachG calls fn on every G from allgs.
575 // forEachG takes a lock to exclude concurrent addition of new Gs.
576 func forEachG(fn func(gp *g)) {
578 for _, gp := range allgs {
584 // forEachGRace calls fn on every G from allgs.
586 // forEachGRace avoids locking, but does not exclude addition of new Gs during
587 // execution, which may be missed.
588 func forEachGRace(fn func(gp *g)) {
589 ptr, length := atomicAllG()
590 for i := uintptr(0); i < length; i++ {
591 gp := atomicAllGIndex(ptr, i)
598 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
599 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
603 // cpuinit extracts the environment variable GODEBUG from the environment on
604 // Unix-like operating systems and calls internal/cpu.Initialize.
606 const prefix = "GODEBUG="
610 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
611 cpu.DebugOptions = true
613 // Similar to goenv_unix but extracts the environment value for
615 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
617 for argv_index(argv, argc+1+n) != nil {
621 for i := int32(0); i < n; i++ {
622 p := argv_index(argv, argc+1+i)
623 s := *(*string)(unsafe.Pointer(&stringStruct{unsafe.Pointer(p), findnull(p)}))
625 if hasPrefix(s, prefix) {
626 env = gostring(p)[len(prefix):]
634 // Support cpu feature variables are used in code generated by the compiler
635 // to guard execution of instructions that can not be assumed to be always supported.
638 x86HasPOPCNT = cpu.X86.HasPOPCNT
639 x86HasSSE41 = cpu.X86.HasSSE41
640 x86HasFMA = cpu.X86.HasFMA
643 armHasVFPv4 = cpu.ARM.HasVFPv4
646 arm64HasATOMICS = cpu.ARM64.HasATOMICS
650 // The bootstrap sequence is:
654 // make & queue new G
655 // call runtime·mstart
657 // The new G calls runtime·main.
659 lockInit(&sched.lock, lockRankSched)
660 lockInit(&sched.sysmonlock, lockRankSysmon)
661 lockInit(&sched.deferlock, lockRankDefer)
662 lockInit(&sched.sudoglock, lockRankSudog)
663 lockInit(&deadlock, lockRankDeadlock)
664 lockInit(&paniclk, lockRankPanic)
665 lockInit(&allglock, lockRankAllg)
666 lockInit(&allpLock, lockRankAllp)
667 lockInit(&reflectOffs.lock, lockRankReflectOffs)
668 lockInit(&finlock, lockRankFin)
669 lockInit(&trace.bufLock, lockRankTraceBuf)
670 lockInit(&trace.stringsLock, lockRankTraceStrings)
671 lockInit(&trace.lock, lockRankTrace)
672 lockInit(&cpuprof.lock, lockRankCpuprof)
673 lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
674 // Enforce that this lock is always a leaf lock.
675 // All of this lock's critical sections should be
677 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
679 // raceinit must be the first call to race detector.
680 // In particular, it must be done before mallocinit below calls racemapshadow.
683 gp.racectx, raceprocctx0 = raceinit()
686 sched.maxmcount = 10000
688 // The world starts stopped.
694 cpuinit() // must run before alginit
695 alginit() // maps, hash, fastrand must not be used before this call
696 fastrandinit() // must run before mcommoninit
697 mcommoninit(gp.m, -1)
698 modulesinit() // provides activeModules
699 typelinksinit() // uses maps, activeModules
700 itabsinit() // uses activeModules
701 stkobjinit() // must run before GC starts
703 sigsave(&gp.m.sigmask)
704 initSigmask = gp.m.sigmask
706 if offset := unsafe.Offsetof(sched.timeToRun); offset%8 != 0 {
708 throw("sched.timeToRun not aligned to 8 bytes")
717 sched.lastpoll.Store(nanotime())
719 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
722 if procresize(procs) != nil {
723 throw("unknown runnable goroutine during bootstrap")
727 // World is effectively started now, as P's can run.
730 // For cgocheck > 1, we turn on the write barrier at all times
731 // and check all pointer writes. We can't do this until after
732 // procresize because the write barrier needs a P.
733 if debug.cgocheck > 1 {
734 writeBarrier.cgo = true
735 writeBarrier.enabled = true
736 for _, pp := range allp {
741 if buildVersion == "" {
742 // Condition should never trigger. This code just serves
743 // to ensure runtime·buildVersion is kept in the resulting binary.
744 buildVersion = "unknown"
746 if len(modinfo) == 1 {
747 // Condition should never trigger. This code just serves
748 // to ensure runtime·modinfo is kept in the resulting binary.
753 func dumpgstatus(gp *g) {
755 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
756 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
759 // sched.lock must be held.
761 assertLockHeld(&sched.lock)
763 if mcount() > sched.maxmcount {
764 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
765 throw("thread exhaustion")
769 // mReserveID returns the next ID to use for a new m. This new m is immediately
770 // considered 'running' by checkdead.
772 // sched.lock must be held.
773 func mReserveID() int64 {
774 assertLockHeld(&sched.lock)
776 if sched.mnext+1 < sched.mnext {
777 throw("runtime: thread ID overflow")
785 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
786 func mcommoninit(mp *m, id int64) {
789 // g0 stack won't make sense for user (and is not necessary unwindable).
791 callers(1, mp.createstack[:])
802 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
803 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
807 // Same behavior as for 1.17.
808 // TODO: Simplify ths.
809 if goarch.BigEndian {
810 mp.fastrand = uint64(lo)<<32 | uint64(hi)
812 mp.fastrand = uint64(hi)<<32 | uint64(lo)
816 if mp.gsignal != nil {
817 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + _StackGuard
820 // Add to allm so garbage collector doesn't free g->m
821 // when it is just in a register or thread-local storage.
824 // NumCgoCall() iterates over allm w/o schedlock,
825 // so we need to publish it safely.
826 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
829 // Allocate memory to hold a cgo traceback if the cgo call crashes.
830 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
831 mp.cgoCallers = new(cgoCallers)
835 var fastrandseed uintptr
837 func fastrandinit() {
838 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
842 // Mark gp ready to run.
843 func ready(gp *g, traceskip int, next bool) {
845 traceGoUnpark(gp, traceskip)
848 status := readgstatus(gp)
851 mp := acquirem() // disable preemption because it can be holding p in a local var
852 if status&^_Gscan != _Gwaiting {
854 throw("bad g->status in ready")
857 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
858 casgstatus(gp, _Gwaiting, _Grunnable)
859 runqput(mp.p.ptr(), gp, next)
864 // freezeStopWait is a large value that freezetheworld sets
865 // sched.stopwait to in order to request that all Gs permanently stop.
866 const freezeStopWait = 0x7fffffff
868 // freezing is set to non-zero if the runtime is trying to freeze the
872 // Similar to stopTheWorld but best-effort and can be called several times.
873 // There is no reverse operation, used during crashing.
874 // This function must not lock any mutexes.
875 func freezetheworld() {
876 atomic.Store(&freezing, 1)
877 // stopwait and preemption requests can be lost
878 // due to races with concurrently executing threads,
879 // so try several times
880 for i := 0; i < 5; i++ {
881 // this should tell the scheduler to not start any new goroutines
882 sched.stopwait = freezeStopWait
883 sched.gcwaiting.Store(true)
884 // this should stop running goroutines
886 break // no running goroutines
896 // All reads and writes of g's status go through readgstatus, casgstatus
897 // castogscanstatus, casfrom_Gscanstatus.
900 func readgstatus(gp *g) uint32 {
901 return atomic.Load(&gp.atomicstatus)
904 // The Gscanstatuses are acting like locks and this releases them.
905 // If it proves to be a performance hit we should be able to make these
906 // simple atomic stores but for now we are going to throw if
907 // we see an inconsistent state.
908 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
911 // Check that transition is valid.
914 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
916 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
922 if newval == oldval&^_Gscan {
923 success = atomic.Cas(&gp.atomicstatus, oldval, newval)
927 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
929 throw("casfrom_Gscanstatus: gp->status is not in scan state")
931 releaseLockRank(lockRankGscan)
934 // This will return false if the gp is not in the expected status and the cas fails.
935 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
936 func castogscanstatus(gp *g, oldval, newval uint32) bool {
942 if newval == oldval|_Gscan {
943 r := atomic.Cas(&gp.atomicstatus, oldval, newval)
945 acquireLockRank(lockRankGscan)
951 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
952 throw("castogscanstatus")
956 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
957 // and casfrom_Gscanstatus instead.
958 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
959 // put it in the Gscan state is finished.
962 func casgstatus(gp *g, oldval, newval uint32) {
963 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
965 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
966 throw("casgstatus: bad incoming values")
970 acquireLockRank(lockRankGscan)
971 releaseLockRank(lockRankGscan)
973 // See https://golang.org/cl/21503 for justification of the yield delay.
974 const yieldDelay = 5 * 1000
977 // loop if gp->atomicstatus is in a scan state giving
978 // GC time to finish and change the state to oldval.
979 for i := 0; !atomic.Cas(&gp.atomicstatus, oldval, newval); i++ {
980 if oldval == _Gwaiting && gp.atomicstatus == _Grunnable {
981 throw("casgstatus: waiting for Gwaiting but is Grunnable")
984 nextYield = nanotime() + yieldDelay
986 if nanotime() < nextYield {
987 for x := 0; x < 10 && gp.atomicstatus != oldval; x++ {
992 nextYield = nanotime() + yieldDelay/2
996 // Handle tracking for scheduling latencies.
997 if oldval == _Grunning {
998 // Track every 8th time a goroutine transitions out of running.
999 if gp.trackingSeq%gTrackingPeriod == 0 {
1005 if oldval == _Grunnable {
1006 // We transitioned out of runnable, so measure how much
1007 // time we spent in this state and add it to
1010 gp.runnableTime += now - gp.runnableStamp
1011 gp.runnableStamp = 0
1013 if newval == _Grunnable {
1014 // We just transitioned into runnable, so record what
1015 // time that happened.
1017 gp.runnableStamp = now
1018 } else if newval == _Grunning {
1019 // We're transitioning into running, so turn off
1020 // tracking and record how much time we spent in
1023 sched.timeToRun.record(gp.runnableTime)
1029 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1030 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1031 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1032 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1033 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1036 func casgcopystack(gp *g) uint32 {
1038 oldstatus := readgstatus(gp) &^ _Gscan
1039 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1040 throw("copystack: bad status, not Gwaiting or Grunnable")
1042 if atomic.Cas(&gp.atomicstatus, oldstatus, _Gcopystack) {
1048 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1050 // TODO(austin): This is the only status operation that both changes
1051 // the status and locks the _Gscan bit. Rethink this.
1052 func casGToPreemptScan(gp *g, old, new uint32) {
1053 if old != _Grunning || new != _Gscan|_Gpreempted {
1054 throw("bad g transition")
1056 acquireLockRank(lockRankGscan)
1057 for !atomic.Cas(&gp.atomicstatus, _Grunning, _Gscan|_Gpreempted) {
1061 // casGFromPreempted attempts to transition gp from _Gpreempted to
1062 // _Gwaiting. If successful, the caller is responsible for
1063 // re-scheduling gp.
1064 func casGFromPreempted(gp *g, old, new uint32) bool {
1065 if old != _Gpreempted || new != _Gwaiting {
1066 throw("bad g transition")
1068 return atomic.Cas(&gp.atomicstatus, _Gpreempted, _Gwaiting)
1071 // stopTheWorld stops all P's from executing goroutines, interrupting
1072 // all goroutines at GC safe points and records reason as the reason
1073 // for the stop. On return, only the current goroutine's P is running.
1074 // stopTheWorld must not be called from a system stack and the caller
1075 // must not hold worldsema. The caller must call startTheWorld when
1076 // other P's should resume execution.
1078 // stopTheWorld is safe for multiple goroutines to call at the
1079 // same time. Each will execute its own stop, and the stops will
1082 // This is also used by routines that do stack dumps. If the system is
1083 // in panic or being exited, this may not reliably stop all
1085 func stopTheWorld(reason string) {
1086 semacquire(&worldsema)
1088 gp.m.preemptoff = reason
1089 systemstack(func() {
1090 // Mark the goroutine which called stopTheWorld preemptible so its
1091 // stack may be scanned.
1092 // This lets a mark worker scan us while we try to stop the world
1093 // since otherwise we could get in a mutual preemption deadlock.
1094 // We must not modify anything on the G stack because a stack shrink
1095 // may occur. A stack shrink is otherwise OK though because in order
1096 // to return from this function (and to leave the system stack) we
1097 // must have preempted all goroutines, including any attempting
1098 // to scan our stack, in which case, any stack shrinking will
1099 // have already completed by the time we exit.
1100 casgstatus(gp, _Grunning, _Gwaiting)
1101 stopTheWorldWithSema()
1102 casgstatus(gp, _Gwaiting, _Grunning)
1106 // startTheWorld undoes the effects of stopTheWorld.
1107 func startTheWorld() {
1108 systemstack(func() { startTheWorldWithSema(false) })
1110 // worldsema must be held over startTheWorldWithSema to ensure
1111 // gomaxprocs cannot change while worldsema is held.
1113 // Release worldsema with direct handoff to the next waiter, but
1114 // acquirem so that semrelease1 doesn't try to yield our time.
1116 // Otherwise if e.g. ReadMemStats is being called in a loop,
1117 // it might stomp on other attempts to stop the world, such as
1118 // for starting or ending GC. The operation this blocks is
1119 // so heavy-weight that we should just try to be as fair as
1122 // We don't want to just allow us to get preempted between now
1123 // and releasing the semaphore because then we keep everyone
1124 // (including, for example, GCs) waiting longer.
1127 semrelease1(&worldsema, true, 0)
1131 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1132 // until the GC is not running. It also blocks a GC from starting
1133 // until startTheWorldGC is called.
1134 func stopTheWorldGC(reason string) {
1136 stopTheWorld(reason)
1139 // startTheWorldGC undoes the effects of stopTheWorldGC.
1140 func startTheWorldGC() {
1145 // Holding worldsema grants an M the right to try to stop the world.
1146 var worldsema uint32 = 1
1148 // Holding gcsema grants the M the right to block a GC, and blocks
1149 // until the current GC is done. In particular, it prevents gomaxprocs
1150 // from changing concurrently.
1152 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1153 // being changed/enabled during a GC, remove this.
1154 var gcsema uint32 = 1
1156 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1157 // The caller is responsible for acquiring worldsema and disabling
1158 // preemption first and then should stopTheWorldWithSema on the system
1161 // semacquire(&worldsema, 0)
1162 // m.preemptoff = "reason"
1163 // systemstack(stopTheWorldWithSema)
1165 // When finished, the caller must either call startTheWorld or undo
1166 // these three operations separately:
1168 // m.preemptoff = ""
1169 // systemstack(startTheWorldWithSema)
1170 // semrelease(&worldsema)
1172 // It is allowed to acquire worldsema once and then execute multiple
1173 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1174 // Other P's are able to execute between successive calls to
1175 // startTheWorldWithSema and stopTheWorldWithSema.
1176 // Holding worldsema causes any other goroutines invoking
1177 // stopTheWorld to block.
1178 func stopTheWorldWithSema() {
1181 // If we hold a lock, then we won't be able to stop another M
1182 // that is blocked trying to acquire the lock.
1184 throw("stopTheWorld: holding locks")
1188 sched.stopwait = gomaxprocs
1189 sched.gcwaiting.Store(true)
1192 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1194 // try to retake all P's in Psyscall status
1195 for _, pp := range allp {
1197 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1209 pp, _ := pidleget(now)
1213 pp.status = _Pgcstop
1216 wait := sched.stopwait > 0
1219 // wait for remaining P's to stop voluntarily
1222 // wait for 100us, then try to re-preempt in case of any races
1223 if notetsleep(&sched.stopnote, 100*1000) {
1224 noteclear(&sched.stopnote)
1233 if sched.stopwait != 0 {
1234 bad = "stopTheWorld: not stopped (stopwait != 0)"
1236 for _, pp := range allp {
1237 if pp.status != _Pgcstop {
1238 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1242 if atomic.Load(&freezing) != 0 {
1243 // Some other thread is panicking. This can cause the
1244 // sanity checks above to fail if the panic happens in
1245 // the signal handler on a stopped thread. Either way,
1246 // we should halt this thread.
1257 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1258 assertWorldStopped()
1260 mp := acquirem() // disable preemption because it can be holding p in a local var
1261 if netpollinited() {
1262 list := netpoll(0) // non-blocking
1272 p1 := procresize(procs)
1273 sched.gcwaiting.Store(false)
1274 if sched.sysmonwait.Load() {
1275 sched.sysmonwait.Store(false)
1276 notewakeup(&sched.sysmonnote)
1289 throw("startTheWorld: inconsistent mp->nextp")
1292 notewakeup(&mp.park)
1294 // Start M to run P. Do not start another M below.
1299 // Capture start-the-world time before doing clean-up tasks.
1300 startTime := nanotime()
1305 // Wakeup an additional proc in case we have excessive runnable goroutines
1306 // in local queues or in the global queue. If we don't, the proc will park itself.
1307 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1315 // usesLibcall indicates whether this runtime performs system calls
1317 func usesLibcall() bool {
1319 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1322 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1327 // mStackIsSystemAllocated indicates whether this runtime starts on a
1328 // system-allocated stack.
1329 func mStackIsSystemAllocated() bool {
1331 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1335 case "386", "amd64", "arm", "arm64":
1342 // mstart is the entry-point for new Ms.
1343 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1346 // mstart0 is the Go entry-point for new Ms.
1347 // This must not split the stack because we may not even have stack
1348 // bounds set up yet.
1350 // May run during STW (because it doesn't have a P yet), so write
1351 // barriers are not allowed.
1354 //go:nowritebarrierrec
1358 osStack := gp.stack.lo == 0
1360 // Initialize stack bounds from system stack.
1361 // Cgo may have left stack size in stack.hi.
1362 // minit may update the stack bounds.
1364 // Note: these bounds may not be very accurate.
1365 // We set hi to &size, but there are things above
1366 // it. The 1024 is supposed to compensate this,
1367 // but is somewhat arbitrary.
1370 size = 8192 * sys.StackGuardMultiplier
1372 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1373 gp.stack.lo = gp.stack.hi - size + 1024
1375 // Initialize stack guard so that we can start calling regular
1377 gp.stackguard0 = gp.stack.lo + _StackGuard
1378 // This is the g0, so we can also call go:systemstack
1379 // functions, which check stackguard1.
1380 gp.stackguard1 = gp.stackguard0
1383 // Exit this thread.
1384 if mStackIsSystemAllocated() {
1385 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1386 // the stack, but put it in gp.stack before mstart,
1387 // so the logic above hasn't set osStack yet.
1393 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1394 // so that we can set up g0.sched to return to the call of mstart1 above.
1401 throw("bad runtime·mstart")
1404 // Set up m.g0.sched as a label returning to just
1405 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1406 // We're never coming back to mstart1 after we call schedule,
1407 // so other calls can reuse the current frame.
1408 // And goexit0 does a gogo that needs to return from mstart1
1409 // and let mstart0 exit the thread.
1410 gp.sched.g = guintptr(unsafe.Pointer(gp))
1411 gp.sched.pc = getcallerpc()
1412 gp.sched.sp = getcallersp()
1417 // Install signal handlers; after minit so that minit can
1418 // prepare the thread to be able to handle the signals.
1423 if fn := gp.m.mstartfn; fn != nil {
1428 acquirep(gp.m.nextp.ptr())
1434 // mstartm0 implements part of mstart1 that only runs on the m0.
1436 // Write barriers are allowed here because we know the GC can't be
1437 // running yet, so they'll be no-ops.
1439 //go:yeswritebarrierrec
1441 // Create an extra M for callbacks on threads not created by Go.
1442 // An extra M is also needed on Windows for callbacks created by
1443 // syscall.NewCallback. See issue #6751 for details.
1444 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1451 // mPark causes a thread to park itself, returning once woken.
1456 notesleep(&gp.m.park)
1457 noteclear(&gp.m.park)
1460 // mexit tears down and exits the current thread.
1462 // Don't call this directly to exit the thread, since it must run at
1463 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1464 // unwind the stack to the point that exits the thread.
1466 // It is entered with m.p != nil, so write barriers are allowed. It
1467 // will release the P before exiting.
1469 //go:yeswritebarrierrec
1470 func mexit(osStack bool) {
1474 // This is the main thread. Just wedge it.
1476 // On Linux, exiting the main thread puts the process
1477 // into a non-waitable zombie state. On Plan 9,
1478 // exiting the main thread unblocks wait even though
1479 // other threads are still running. On Solaris we can
1480 // neither exitThread nor return from mstart. Other
1481 // bad things probably happen on other platforms.
1483 // We could try to clean up this M more before wedging
1484 // it, but that complicates signal handling.
1485 handoffp(releasep())
1491 throw("locked m0 woke up")
1497 // Free the gsignal stack.
1498 if mp.gsignal != nil {
1499 stackfree(mp.gsignal.stack)
1500 // On some platforms, when calling into VDSO (e.g. nanotime)
1501 // we store our g on the gsignal stack, if there is one.
1502 // Now the stack is freed, unlink it from the m, so we
1503 // won't write to it when calling VDSO code.
1507 // Remove m from allm.
1509 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1515 throw("m not found in allm")
1518 // Delay reaping m until it's done with the stack.
1520 // If this is using an OS stack, the OS will free it
1521 // so there's no need for reaping.
1522 atomic.Store(&mp.freeWait, 1)
1523 // Put m on the free list, though it will not be reaped until
1524 // freeWait is 0. Note that the free list must not be linked
1525 // through alllink because some functions walk allm without
1526 // locking, so may be using alllink.
1527 mp.freelink = sched.freem
1532 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1535 handoffp(releasep())
1536 // After this point we must not have write barriers.
1538 // Invoke the deadlock detector. This must happen after
1539 // handoffp because it may have started a new M to take our
1546 if GOOS == "darwin" || GOOS == "ios" {
1547 // Make sure pendingPreemptSignals is correct when an M exits.
1549 if atomic.Load(&mp.signalPending) != 0 {
1550 pendingPreemptSignals.Add(-1)
1554 // Destroy all allocated resources. After this is called, we may no
1555 // longer take any locks.
1559 // Return from mstart and let the system thread
1560 // library free the g0 stack and terminate the thread.
1564 // mstart is the thread's entry point, so there's nothing to
1565 // return to. Exit the thread directly. exitThread will clear
1566 // m.freeWait when it's done with the stack and the m can be
1568 exitThread(&mp.freeWait)
1571 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1572 // If a P is currently executing code, this will bring the P to a GC
1573 // safe point and execute fn on that P. If the P is not executing code
1574 // (it is idle or in a syscall), this will call fn(p) directly while
1575 // preventing the P from exiting its state. This does not ensure that
1576 // fn will run on every CPU executing Go code, but it acts as a global
1577 // memory barrier. GC uses this as a "ragged barrier."
1579 // The caller must hold worldsema.
1582 func forEachP(fn func(*p)) {
1584 pp := getg().m.p.ptr()
1587 if sched.safePointWait != 0 {
1588 throw("forEachP: sched.safePointWait != 0")
1590 sched.safePointWait = gomaxprocs - 1
1591 sched.safePointFn = fn
1593 // Ask all Ps to run the safe point function.
1594 for _, p2 := range allp {
1596 atomic.Store(&p2.runSafePointFn, 1)
1601 // Any P entering _Pidle or _Psyscall from now on will observe
1602 // p.runSafePointFn == 1 and will call runSafePointFn when
1603 // changing its status to _Pidle/_Psyscall.
1605 // Run safe point function for all idle Ps. sched.pidle will
1606 // not change because we hold sched.lock.
1607 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1608 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1610 sched.safePointWait--
1614 wait := sched.safePointWait > 0
1617 // Run fn for the current P.
1620 // Force Ps currently in _Psyscall into _Pidle and hand them
1621 // off to induce safe point function execution.
1622 for _, p2 := range allp {
1624 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1634 // Wait for remaining Ps to run fn.
1637 // Wait for 100us, then try to re-preempt in
1638 // case of any races.
1640 // Requires system stack.
1641 if notetsleep(&sched.safePointNote, 100*1000) {
1642 noteclear(&sched.safePointNote)
1648 if sched.safePointWait != 0 {
1649 throw("forEachP: not done")
1651 for _, p2 := range allp {
1652 if p2.runSafePointFn != 0 {
1653 throw("forEachP: P did not run fn")
1658 sched.safePointFn = nil
1663 // runSafePointFn runs the safe point function, if any, for this P.
1664 // This should be called like
1666 // if getg().m.p.runSafePointFn != 0 {
1670 // runSafePointFn must be checked on any transition in to _Pidle or
1671 // _Psyscall to avoid a race where forEachP sees that the P is running
1672 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1673 // nor the P run the safe-point function.
1674 func runSafePointFn() {
1675 p := getg().m.p.ptr()
1676 // Resolve the race between forEachP running the safe-point
1677 // function on this P's behalf and this P running the
1678 // safe-point function directly.
1679 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1682 sched.safePointFn(p)
1684 sched.safePointWait--
1685 if sched.safePointWait == 0 {
1686 notewakeup(&sched.safePointNote)
1691 // When running with cgo, we call _cgo_thread_start
1692 // to start threads for us so that we can play nicely with
1694 var cgoThreadStart unsafe.Pointer
1696 type cgothreadstart struct {
1702 // Allocate a new m unassociated with any thread.
1703 // Can use p for allocation context if needed.
1704 // fn is recorded as the new m's m.mstartfn.
1705 // id is optional pre-allocated m ID. Omit by passing -1.
1707 // This function is allowed to have write barriers even if the caller
1708 // isn't because it borrows pp.
1710 //go:yeswritebarrierrec
1711 func allocm(pp *p, fn func(), id int64) *m {
1714 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1715 // disable preemption to ensure it is not stolen, which would make the
1716 // caller lose ownership.
1721 acquirep(pp) // temporarily borrow p for mallocs in this function
1724 // Release the free M list. We need to do this somewhere and
1725 // this may free up a stack we can use.
1726 if sched.freem != nil {
1729 for freem := sched.freem; freem != nil; {
1730 if freem.freeWait != 0 {
1731 next := freem.freelink
1732 freem.freelink = newList
1737 // stackfree must be on the system stack, but allocm is
1738 // reachable off the system stack transitively from
1740 systemstack(func() {
1741 stackfree(freem.g0.stack)
1743 freem = freem.freelink
1745 sched.freem = newList
1753 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1754 // Windows and Plan 9 will layout sched stack on OS stack.
1755 if iscgo || mStackIsSystemAllocated() {
1758 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1762 if pp == gp.m.p.ptr() {
1767 allocmLock.runlock()
1771 // needm is called when a cgo callback happens on a
1772 // thread without an m (a thread not created by Go).
1773 // In this case, needm is expected to find an m to use
1774 // and return with m, g initialized correctly.
1775 // Since m and g are not set now (likely nil, but see below)
1776 // needm is limited in what routines it can call. In particular
1777 // it can only call nosplit functions (textflag 7) and cannot
1778 // do any scheduling that requires an m.
1780 // In order to avoid needing heavy lifting here, we adopt
1781 // the following strategy: there is a stack of available m's
1782 // that can be stolen. Using compare-and-swap
1783 // to pop from the stack has ABA races, so we simulate
1784 // a lock by doing an exchange (via Casuintptr) to steal the stack
1785 // head and replace the top pointer with MLOCKED (1).
1786 // This serves as a simple spin lock that we can use even
1787 // without an m. The thread that locks the stack in this way
1788 // unlocks the stack by storing a valid stack head pointer.
1790 // In order to make sure that there is always an m structure
1791 // available to be stolen, we maintain the invariant that there
1792 // is always one more than needed. At the beginning of the
1793 // program (if cgo is in use) the list is seeded with a single m.
1794 // If needm finds that it has taken the last m off the list, its job
1795 // is - once it has installed its own m so that it can do things like
1796 // allocate memory - to create a spare m and put it on the list.
1798 // Each of these extra m's also has a g0 and a curg that are
1799 // pressed into service as the scheduling stack and current
1800 // goroutine for the duration of the cgo callback.
1802 // When the callback is done with the m, it calls dropm to
1803 // put the m back on the list.
1807 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1808 // Can happen if C/C++ code calls Go from a global ctor.
1809 // Can also happen on Windows if a global ctor uses a
1810 // callback created by syscall.NewCallback. See issue #6751
1813 // Can not throw, because scheduler is not initialized yet.
1814 write(2, unsafe.Pointer(&earlycgocallback[0]), int32(len(earlycgocallback)))
1818 // Save and block signals before getting an M.
1819 // The signal handler may call needm itself,
1820 // and we must avoid a deadlock. Also, once g is installed,
1821 // any incoming signals will try to execute,
1822 // but we won't have the sigaltstack settings and other data
1823 // set up appropriately until the end of minit, which will
1824 // unblock the signals. This is the same dance as when
1825 // starting a new m to run Go code via newosproc.
1830 // Lock extra list, take head, unlock popped list.
1831 // nilokay=false is safe here because of the invariant above,
1832 // that the extra list always contains or will soon contain
1834 mp := lockextra(false)
1836 // Set needextram when we've just emptied the list,
1837 // so that the eventual call into cgocallbackg will
1838 // allocate a new m for the extra list. We delay the
1839 // allocation until then so that it can be done
1840 // after exitsyscall makes sure it is okay to be
1841 // running at all (that is, there's no garbage collection
1842 // running right now).
1843 mp.needextram = mp.schedlink == 0
1845 unlockextra(mp.schedlink.ptr())
1847 // Store the original signal mask for use by minit.
1848 mp.sigmask = sigmask
1850 // Install TLS on some platforms (previously setg
1851 // would do this if necessary).
1854 // Install g (= m->g0) and set the stack bounds
1855 // to match the current stack. We don't actually know
1856 // how big the stack is, like we don't know how big any
1857 // scheduling stack is, but we assume there's at least 32 kB,
1858 // which is more than enough for us.
1861 gp.stack.hi = getcallersp() + 1024
1862 gp.stack.lo = getcallersp() - 32*1024
1863 gp.stackguard0 = gp.stack.lo + _StackGuard
1865 // Initialize this thread to use the m.
1869 // mp.curg is now a real goroutine.
1870 casgstatus(mp.curg, _Gdead, _Gsyscall)
1874 var earlycgocallback = []byte("fatal error: cgo callback before cgo call\n")
1876 // newextram allocates m's and puts them on the extra list.
1877 // It is called with a working local m, so that it can do things
1878 // like call schedlock and allocate.
1880 c := atomic.Xchg(&extraMWaiters, 0)
1882 for i := uint32(0); i < c; i++ {
1886 // Make sure there is at least one extra M.
1887 mp := lockextra(true)
1895 // oneNewExtraM allocates an m and puts it on the extra list.
1896 func oneNewExtraM() {
1897 // Create extra goroutine locked to extra m.
1898 // The goroutine is the context in which the cgo callback will run.
1899 // The sched.pc will never be returned to, but setting it to
1900 // goexit makes clear to the traceback routines where
1901 // the goroutine stack ends.
1902 mp := allocm(nil, nil, -1)
1904 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
1905 gp.sched.sp = gp.stack.hi
1906 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
1908 gp.sched.g = guintptr(unsafe.Pointer(gp))
1909 gp.syscallpc = gp.sched.pc
1910 gp.syscallsp = gp.sched.sp
1911 gp.stktopsp = gp.sched.sp
1912 // malg returns status as _Gidle. Change to _Gdead before
1913 // adding to allg where GC can see it. We use _Gdead to hide
1914 // this from tracebacks and stack scans since it isn't a
1915 // "real" goroutine until needm grabs it.
1916 casgstatus(gp, _Gidle, _Gdead)
1922 gp.goid = sched.goidgen.Add(1)
1924 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
1926 // put on allg for garbage collector
1929 // gp is now on the allg list, but we don't want it to be
1930 // counted by gcount. It would be more "proper" to increment
1931 // sched.ngfree, but that requires locking. Incrementing ngsys
1932 // has the same effect.
1935 // Add m to the extra list.
1936 mnext := lockextra(true)
1937 mp.schedlink.set(mnext)
1942 // dropm is called when a cgo callback has called needm but is now
1943 // done with the callback and returning back into the non-Go thread.
1944 // It puts the current m back onto the extra list.
1946 // The main expense here is the call to signalstack to release the
1947 // m's signal stack, and then the call to needm on the next callback
1948 // from this thread. It is tempting to try to save the m for next time,
1949 // which would eliminate both these costs, but there might not be
1950 // a next time: the current thread (which Go does not control) might exit.
1951 // If we saved the m for that thread, there would be an m leak each time
1952 // such a thread exited. Instead, we acquire and release an m on each
1953 // call. These should typically not be scheduling operations, just a few
1954 // atomics, so the cost should be small.
1956 // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
1957 // variable using pthread_key_create. Unlike the pthread keys we already use
1958 // on OS X, this dummy key would never be read by Go code. It would exist
1959 // only so that we could register at thread-exit-time destructor.
1960 // That destructor would put the m back onto the extra list.
1961 // This is purely a performance optimization. The current version,
1962 // in which dropm happens on each cgo call, is still correct too.
1963 // We may have to keep the current version on systems with cgo
1964 // but without pthreads, like Windows.
1966 // Clear m and g, and return m to the extra list.
1967 // After the call to setg we can only call nosplit functions
1968 // with no pointer manipulation.
1971 // Return mp.curg to dead state.
1972 casgstatus(mp.curg, _Gsyscall, _Gdead)
1973 mp.curg.preemptStop = false
1976 // Block signals before unminit.
1977 // Unminit unregisters the signal handling stack (but needs g on some systems).
1978 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
1979 // It's important not to try to handle a signal between those two steps.
1980 sigmask := mp.sigmask
1984 mnext := lockextra(true)
1986 mp.schedlink.set(mnext)
1990 // Commit the release of mp.
1993 msigrestore(sigmask)
1996 // A helper function for EnsureDropM.
1997 func getm() uintptr {
1998 return uintptr(unsafe.Pointer(getg().m))
2002 var extraMCount uint32 // Protected by lockextra
2003 var extraMWaiters uint32
2005 // lockextra locks the extra list and returns the list head.
2006 // The caller must unlock the list by storing a new list head
2007 // to extram. If nilokay is true, then lockextra will
2008 // return a nil list head if that's what it finds. If nilokay is false,
2009 // lockextra will keep waiting until the list head is no longer nil.
2012 func lockextra(nilokay bool) *m {
2017 old := atomic.Loaduintptr(&extram)
2022 if old == 0 && !nilokay {
2024 // Add 1 to the number of threads
2025 // waiting for an M.
2026 // This is cleared by newextram.
2027 atomic.Xadd(&extraMWaiters, 1)
2033 if atomic.Casuintptr(&extram, old, locked) {
2034 return (*m)(unsafe.Pointer(old))
2042 func unlockextra(mp *m) {
2043 atomic.Storeuintptr(&extram, uintptr(unsafe.Pointer(mp)))
2047 // allocmLock is locked for read when creating new Ms in allocm and their
2048 // addition to allm. Thus acquiring this lock for write blocks the
2049 // creation of new Ms.
2052 // execLock serializes exec and clone to avoid bugs or unspecified
2053 // behaviour around exec'ing while creating/destroying threads. See
2058 // newmHandoff contains a list of m structures that need new OS threads.
2059 // This is used by newm in situations where newm itself can't safely
2060 // start an OS thread.
2061 var newmHandoff struct {
2064 // newm points to a list of M structures that need new OS
2065 // threads. The list is linked through m.schedlink.
2068 // waiting indicates that wake needs to be notified when an m
2069 // is put on the list.
2073 // haveTemplateThread indicates that the templateThread has
2074 // been started. This is not protected by lock. Use cas to set
2076 haveTemplateThread uint32
2079 // Create a new m. It will start off with a call to fn, or else the scheduler.
2080 // fn needs to be static and not a heap allocated closure.
2081 // May run with m.p==nil, so write barriers are not allowed.
2083 // id is optional pre-allocated m ID. Omit by passing -1.
2085 //go:nowritebarrierrec
2086 func newm(fn func(), pp *p, id int64) {
2087 // allocm adds a new M to allm, but they do not start until created by
2088 // the OS in newm1 or the template thread.
2090 // doAllThreadsSyscall requires that every M in allm will eventually
2091 // start and be signal-able, even with a STW.
2093 // Disable preemption here until we start the thread to ensure that
2094 // newm is not preempted between allocm and starting the new thread,
2095 // ensuring that anything added to allm is guaranteed to eventually
2099 mp := allocm(pp, fn, id)
2101 mp.sigmask = initSigmask
2102 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2103 // We're on a locked M or a thread that may have been
2104 // started by C. The kernel state of this thread may
2105 // be strange (the user may have locked it for that
2106 // purpose). We don't want to clone that into another
2107 // thread. Instead, ask a known-good thread to create
2108 // the thread for us.
2110 // This is disabled on Plan 9. See golang.org/issue/22227.
2112 // TODO: This may be unnecessary on Windows, which
2113 // doesn't model thread creation off fork.
2114 lock(&newmHandoff.lock)
2115 if newmHandoff.haveTemplateThread == 0 {
2116 throw("on a locked thread with no template thread")
2118 mp.schedlink = newmHandoff.newm
2119 newmHandoff.newm.set(mp)
2120 if newmHandoff.waiting {
2121 newmHandoff.waiting = false
2122 notewakeup(&newmHandoff.wake)
2124 unlock(&newmHandoff.lock)
2125 // The M has not started yet, but the template thread does not
2126 // participate in STW, so it will always process queued Ms and
2127 // it is safe to releasem.
2137 var ts cgothreadstart
2138 if _cgo_thread_start == nil {
2139 throw("_cgo_thread_start missing")
2142 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2143 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2145 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2148 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2150 execLock.rlock() // Prevent process clone.
2151 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2155 execLock.rlock() // Prevent process clone.
2160 // startTemplateThread starts the template thread if it is not already
2163 // The calling thread must itself be in a known-good state.
2164 func startTemplateThread() {
2165 if GOARCH == "wasm" { // no threads on wasm yet
2169 // Disable preemption to guarantee that the template thread will be
2170 // created before a park once haveTemplateThread is set.
2172 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2176 newm(templateThread, nil, -1)
2180 // templateThread is a thread in a known-good state that exists solely
2181 // to start new threads in known-good states when the calling thread
2182 // may not be in a good state.
2184 // Many programs never need this, so templateThread is started lazily
2185 // when we first enter a state that might lead to running on a thread
2186 // in an unknown state.
2188 // templateThread runs on an M without a P, so it must not have write
2191 //go:nowritebarrierrec
2192 func templateThread() {
2199 lock(&newmHandoff.lock)
2200 for newmHandoff.newm != 0 {
2201 newm := newmHandoff.newm.ptr()
2202 newmHandoff.newm = 0
2203 unlock(&newmHandoff.lock)
2205 next := newm.schedlink.ptr()
2210 lock(&newmHandoff.lock)
2212 newmHandoff.waiting = true
2213 noteclear(&newmHandoff.wake)
2214 unlock(&newmHandoff.lock)
2215 notesleep(&newmHandoff.wake)
2219 // Stops execution of the current m until new work is available.
2220 // Returns with acquired P.
2224 if gp.m.locks != 0 {
2225 throw("stopm holding locks")
2228 throw("stopm holding p")
2231 throw("stopm spinning")
2238 acquirep(gp.m.nextp.ptr())
2243 // startm's caller incremented nmspinning. Set the new M's spinning.
2244 getg().m.spinning = true
2247 // Schedules some M to run the p (creates an M if necessary).
2248 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2249 // May run with m.p==nil, so write barriers are not allowed.
2250 // If spinning is set, the caller has incremented nmspinning and startm will
2251 // either decrement nmspinning or set m.spinning in the newly started M.
2253 // Callers passing a non-nil P must call from a non-preemptible context. See
2254 // comment on acquirem below.
2256 // Must not have write barriers because this may be called without a P.
2258 //go:nowritebarrierrec
2259 func startm(pp *p, spinning bool) {
2260 // Disable preemption.
2262 // Every owned P must have an owner that will eventually stop it in the
2263 // event of a GC stop request. startm takes transient ownership of a P
2264 // (either from argument or pidleget below) and transfers ownership to
2265 // a started M, which will be responsible for performing the stop.
2267 // Preemption must be disabled during this transient ownership,
2268 // otherwise the P this is running on may enter GC stop while still
2269 // holding the transient P, leaving that P in limbo and deadlocking the
2272 // Callers passing a non-nil P must already be in non-preemptible
2273 // context, otherwise such preemption could occur on function entry to
2274 // startm. Callers passing a nil P may be preemptible, so we must
2275 // disable preemption before acquiring a P from pidleget below.
2283 // The caller incremented nmspinning, but there are no idle Ps,
2284 // so it's okay to just undo the increment and give up.
2285 if sched.nmspinning.Add(-1) < 0 {
2286 throw("startm: negative nmspinning")
2295 // No M is available, we must drop sched.lock and call newm.
2296 // However, we already own a P to assign to the M.
2298 // Once sched.lock is released, another G (e.g., in a syscall),
2299 // could find no idle P while checkdead finds a runnable G but
2300 // no running M's because this new M hasn't started yet, thus
2301 // throwing in an apparent deadlock.
2303 // Avoid this situation by pre-allocating the ID for the new M,
2304 // thus marking it as 'running' before we drop sched.lock. This
2305 // new M will eventually run the scheduler to execute any
2312 // The caller incremented nmspinning, so set m.spinning in the new M.
2316 // Ownership transfer of pp committed by start in newm.
2317 // Preemption is now safe.
2323 throw("startm: m is spinning")
2326 throw("startm: m has p")
2328 if spinning && !runqempty(pp) {
2329 throw("startm: p has runnable gs")
2331 // The caller incremented nmspinning, so set m.spinning in the new M.
2332 nmp.spinning = spinning
2334 notewakeup(&nmp.park)
2335 // Ownership transfer of pp committed by wakeup. Preemption is now
2340 // Hands off P from syscall or locked M.
2341 // Always runs without a P, so write barriers are not allowed.
2343 //go:nowritebarrierrec
2344 func handoffp(pp *p) {
2345 // handoffp must start an M in any situation where
2346 // findrunnable would return a G to run on pp.
2348 // if it has local work, start it straight away
2349 if !runqempty(pp) || sched.runqsize != 0 {
2353 // if there's trace work to do, start it straight away
2354 if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
2358 // if it has GC work, start it straight away
2359 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2363 // no local work, check that there are no spinning/idle M's,
2364 // otherwise our help is not required
2365 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2370 if sched.gcwaiting.Load() {
2371 pp.status = _Pgcstop
2373 if sched.stopwait == 0 {
2374 notewakeup(&sched.stopnote)
2379 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2380 sched.safePointFn(pp)
2381 sched.safePointWait--
2382 if sched.safePointWait == 0 {
2383 notewakeup(&sched.safePointNote)
2386 if sched.runqsize != 0 {
2391 // If this is the last running P and nobody is polling network,
2392 // need to wakeup another M to poll network.
2393 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2399 // The scheduler lock cannot be held when calling wakeNetPoller below
2400 // because wakeNetPoller may call wakep which may call startm.
2401 when := nobarrierWakeTime(pp)
2410 // Tries to add one more P to execute G's.
2411 // Called when a G is made runnable (newproc, ready).
2413 if sched.npidle.Load() == 0 {
2416 // be conservative about spinning threads
2417 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2423 // Stops execution of the current m that is locked to a g until the g is runnable again.
2424 // Returns with acquired P.
2425 func stoplockedm() {
2428 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2429 throw("stoplockedm: inconsistent locking")
2432 // Schedule another M to run this p.
2437 // Wait until another thread schedules lockedg again.
2439 status := readgstatus(gp.m.lockedg.ptr())
2440 if status&^_Gscan != _Grunnable {
2441 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2442 dumpgstatus(gp.m.lockedg.ptr())
2443 throw("stoplockedm: not runnable")
2445 acquirep(gp.m.nextp.ptr())
2449 // Schedules the locked m to run the locked gp.
2450 // May run during STW, so write barriers are not allowed.
2452 //go:nowritebarrierrec
2453 func startlockedm(gp *g) {
2454 mp := gp.lockedm.ptr()
2456 throw("startlockedm: locked to me")
2459 throw("startlockedm: m has p")
2461 // directly handoff current P to the locked m
2465 notewakeup(&mp.park)
2469 // Stops the current m for stopTheWorld.
2470 // Returns when the world is restarted.
2474 if !sched.gcwaiting.Load() {
2475 throw("gcstopm: not waiting for gc")
2478 gp.m.spinning = false
2479 // OK to just drop nmspinning here,
2480 // startTheWorld will unpark threads as necessary.
2481 if sched.nmspinning.Add(-1) < 0 {
2482 throw("gcstopm: negative nmspinning")
2487 pp.status = _Pgcstop
2489 if sched.stopwait == 0 {
2490 notewakeup(&sched.stopnote)
2496 // Schedules gp to run on the current M.
2497 // If inheritTime is true, gp inherits the remaining time in the
2498 // current time slice. Otherwise, it starts a new time slice.
2501 // Write barriers are allowed because this is called immediately after
2502 // acquiring a P in several places.
2504 //go:yeswritebarrierrec
2505 func execute(gp *g, inheritTime bool) {
2508 if goroutineProfile.active {
2509 // Make sure that gp has had its stack written out to the goroutine
2510 // profile, exactly as it was when the goroutine profiler first stopped
2512 tryRecordGoroutineProfile(gp, osyield)
2515 // Assign gp.m before entering _Grunning so running Gs have an
2519 casgstatus(gp, _Grunnable, _Grunning)
2522 gp.stackguard0 = gp.stack.lo + _StackGuard
2524 mp.p.ptr().schedtick++
2527 // Check whether the profiler needs to be turned on or off.
2528 hz := sched.profilehz
2529 if mp.profilehz != hz {
2530 setThreadCPUProfiler(hz)
2534 // GoSysExit has to happen when we have a P, but before GoStart.
2535 // So we emit it here.
2536 if gp.syscallsp != 0 && gp.sysblocktraced {
2537 traceGoSysExit(gp.sysexitticks)
2545 // Finds a runnable goroutine to execute.
2546 // Tries to steal from other P's, get g from local or global queue, poll network.
2547 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2548 // reader) so the caller should try to wake a P.
2549 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2552 // The conditions here and in handoffp must agree: if
2553 // findrunnable would return a G to run, handoffp must start
2558 if sched.gcwaiting.Load() {
2562 if pp.runSafePointFn != 0 {
2566 // now and pollUntil are saved for work stealing later,
2567 // which may steal timers. It's important that between now
2568 // and then, nothing blocks, so these numbers remain mostly
2570 now, pollUntil, _ := checkTimers(pp, 0)
2572 // Try to schedule the trace reader.
2573 if trace.enabled || trace.shutdown {
2576 casgstatus(gp, _Gwaiting, _Grunnable)
2577 traceGoUnpark(gp, 0)
2578 return gp, false, true
2582 // Try to schedule a GC worker.
2583 if gcBlackenEnabled != 0 {
2584 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2586 return gp, false, true
2591 // Check the global runnable queue once in a while to ensure fairness.
2592 // Otherwise two goroutines can completely occupy the local runqueue
2593 // by constantly respawning each other.
2594 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2596 gp := globrunqget(pp, 1)
2599 return gp, false, false
2603 // Wake up the finalizer G.
2604 if fingwait && fingwake {
2605 if gp := wakefing(); gp != nil {
2609 if *cgo_yield != nil {
2610 asmcgocall(*cgo_yield, nil)
2614 if gp, inheritTime := runqget(pp); gp != nil {
2615 return gp, inheritTime, false
2619 if sched.runqsize != 0 {
2621 gp := globrunqget(pp, 0)
2624 return gp, false, false
2629 // This netpoll is only an optimization before we resort to stealing.
2630 // We can safely skip it if there are no waiters or a thread is blocked
2631 // in netpoll already. If there is any kind of logical race with that
2632 // blocked thread (e.g. it has already returned from netpoll, but does
2633 // not set lastpoll yet), this thread will do blocking netpoll below
2635 if netpollinited() && atomic.Load(&netpollWaiters) > 0 && sched.lastpoll.Load() != 0 {
2636 if list := netpoll(0); !list.empty() { // non-blocking
2639 casgstatus(gp, _Gwaiting, _Grunnable)
2641 traceGoUnpark(gp, 0)
2643 return gp, false, false
2647 // Spinning Ms: steal work from other Ps.
2649 // Limit the number of spinning Ms to half the number of busy Ps.
2650 // This is necessary to prevent excessive CPU consumption when
2651 // GOMAXPROCS>>1 but the program parallelism is low.
2652 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2655 sched.nmspinning.Add(1)
2658 gp, inheritTime, tnow, w, newWork := stealWork(now)
2660 // Successfully stole.
2661 return gp, inheritTime, false
2664 // There may be new timer or GC work; restart to
2670 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2671 // Earlier timer to wait for.
2676 // We have nothing to do.
2678 // If we're in the GC mark phase, can safely scan and blacken objects,
2679 // and have work to do, run idle-time marking rather than give up the P.
2680 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2681 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2683 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2685 casgstatus(gp, _Gwaiting, _Grunnable)
2687 traceGoUnpark(gp, 0)
2689 return gp, false, false
2691 gcController.removeIdleMarkWorker()
2695 // If a callback returned and no other goroutine is awake,
2696 // then wake event handler goroutine which pauses execution
2697 // until a callback was triggered.
2698 gp, otherReady := beforeIdle(now, pollUntil)
2700 casgstatus(gp, _Gwaiting, _Grunnable)
2702 traceGoUnpark(gp, 0)
2704 return gp, false, false
2710 // Before we drop our P, make a snapshot of the allp slice,
2711 // which can change underfoot once we no longer block
2712 // safe-points. We don't need to snapshot the contents because
2713 // everything up to cap(allp) is immutable.
2714 allpSnapshot := allp
2715 // Also snapshot masks. Value changes are OK, but we can't allow
2716 // len to change out from under us.
2717 idlepMaskSnapshot := idlepMask
2718 timerpMaskSnapshot := timerpMask
2720 // return P and block
2722 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2726 if sched.runqsize != 0 {
2727 gp := globrunqget(pp, 0)
2729 return gp, false, false
2731 if releasep() != pp {
2732 throw("findrunnable: wrong p")
2734 now = pidleput(pp, now)
2737 // Delicate dance: thread transitions from spinning to non-spinning
2738 // state, potentially concurrently with submission of new work. We must
2739 // drop nmspinning first and then check all sources again (with
2740 // #StoreLoad memory barrier in between). If we do it the other way
2741 // around, another thread can submit work after we've checked all
2742 // sources but before we drop nmspinning; as a result nobody will
2743 // unpark a thread to run the work.
2745 // This applies to the following sources of work:
2747 // * Goroutines added to a per-P run queue.
2748 // * New/modified-earlier timers on a per-P timer heap.
2749 // * Idle-priority GC work (barring golang.org/issue/19112).
2751 // If we discover new work below, we need to restore m.spinning as a signal
2752 // for resetspinning to unpark a new worker thread (because there can be more
2753 // than one starving goroutine). However, if after discovering new work
2754 // we also observe no idle Ps it is OK to skip unparking a new worker
2755 // thread: the system is fully loaded so no spinning threads are required.
2756 // Also see "Worker thread parking/unparking" comment at the top of the file.
2757 wasSpinning := mp.spinning
2760 if sched.nmspinning.Add(-1) < 0 {
2761 throw("findrunnable: negative nmspinning")
2764 // Note the for correctness, only the last M transitioning from
2765 // spinning to non-spinning must perform these rechecks to
2766 // ensure no missed work. We are performing it on every M that
2767 // transitions as a conservative change to monitor effects on
2768 // latency. See golang.org/issue/43997.
2770 // Check all runqueues once again.
2771 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
2775 sched.nmspinning.Add(1)
2779 // Check for idle-priority GC work again.
2780 pp, gp := checkIdleGCNoP()
2784 sched.nmspinning.Add(1)
2786 // Run the idle worker.
2787 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2788 casgstatus(gp, _Gwaiting, _Grunnable)
2790 traceGoUnpark(gp, 0)
2792 return gp, false, false
2795 // Finally, check for timer creation or expiry concurrently with
2796 // transitioning from spinning to non-spinning.
2798 // Note that we cannot use checkTimers here because it calls
2799 // adjusttimers which may need to allocate memory, and that isn't
2800 // allowed when we don't have an active P.
2801 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
2804 // Poll network until next timer.
2805 if netpollinited() && (atomic.Load(&netpollWaiters) > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
2806 sched.pollUntil.Store(pollUntil)
2808 throw("findrunnable: netpoll with p")
2811 throw("findrunnable: netpoll with spinning")
2817 delay = pollUntil - now
2823 // When using fake time, just poll.
2826 list := netpoll(delay) // block until new work is available
2827 sched.pollUntil.Store(0)
2828 sched.lastpoll.Store(now)
2829 if faketime != 0 && list.empty() {
2830 // Using fake time and nothing is ready; stop M.
2831 // When all M's stop, checkdead will call timejump.
2836 pp, _ := pidleget(now)
2845 casgstatus(gp, _Gwaiting, _Grunnable)
2847 traceGoUnpark(gp, 0)
2849 return gp, false, false
2853 sched.nmspinning.Add(1)
2857 } else if pollUntil != 0 && netpollinited() {
2858 pollerPollUntil := sched.pollUntil.Load()
2859 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
2867 // pollWork reports whether there is non-background work this P could
2868 // be doing. This is a fairly lightweight check to be used for
2869 // background work loops, like idle GC. It checks a subset of the
2870 // conditions checked by the actual scheduler.
2871 func pollWork() bool {
2872 if sched.runqsize != 0 {
2875 p := getg().m.p.ptr()
2879 if netpollinited() && atomic.Load(&netpollWaiters) > 0 && sched.lastpoll.Load() != 0 {
2880 if list := netpoll(0); !list.empty() {
2888 // stealWork attempts to steal a runnable goroutine or timer from any P.
2890 // If newWork is true, new work may have been readied.
2892 // If now is not 0 it is the current time. stealWork returns the passed time or
2893 // the current time if now was passed as 0.
2894 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
2895 pp := getg().m.p.ptr()
2899 const stealTries = 4
2900 for i := 0; i < stealTries; i++ {
2901 stealTimersOrRunNextG := i == stealTries-1
2903 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
2904 if sched.gcwaiting.Load() {
2905 // GC work may be available.
2906 return nil, false, now, pollUntil, true
2908 p2 := allp[enum.position()]
2913 // Steal timers from p2. This call to checkTimers is the only place
2914 // where we might hold a lock on a different P's timers. We do this
2915 // once on the last pass before checking runnext because stealing
2916 // from the other P's runnext should be the last resort, so if there
2917 // are timers to steal do that first.
2919 // We only check timers on one of the stealing iterations because
2920 // the time stored in now doesn't change in this loop and checking
2921 // the timers for each P more than once with the same value of now
2922 // is probably a waste of time.
2924 // timerpMask tells us whether the P may have timers at all. If it
2925 // can't, no need to check at all.
2926 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
2927 tnow, w, ran := checkTimers(p2, now)
2929 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2933 // Running the timers may have
2934 // made an arbitrary number of G's
2935 // ready and added them to this P's
2936 // local run queue. That invalidates
2937 // the assumption of runqsteal
2938 // that it always has room to add
2939 // stolen G's. So check now if there
2940 // is a local G to run.
2941 if gp, inheritTime := runqget(pp); gp != nil {
2942 return gp, inheritTime, now, pollUntil, ranTimer
2948 // Don't bother to attempt to steal if p2 is idle.
2949 if !idlepMask.read(enum.position()) {
2950 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
2951 return gp, false, now, pollUntil, ranTimer
2957 // No goroutines found to steal. Regardless, running a timer may have
2958 // made some goroutine ready that we missed. Indicate the next timer to
2960 return nil, false, now, pollUntil, ranTimer
2963 // Check all Ps for a runnable G to steal.
2965 // On entry we have no P. If a G is available to steal and a P is available,
2966 // the P is returned which the caller should acquire and attempt to steal the
2968 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
2969 for id, p2 := range allpSnapshot {
2970 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
2972 pp, _ := pidleget(0)
2978 // Can't get a P, don't bother checking remaining Ps.
2986 // Check all Ps for a timer expiring sooner than pollUntil.
2988 // Returns updated pollUntil value.
2989 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
2990 for id, p2 := range allpSnapshot {
2991 if timerpMaskSnapshot.read(uint32(id)) {
2992 w := nobarrierWakeTime(p2)
2993 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3002 // Check for idle-priority GC, without a P on entry.
3004 // If some GC work, a P, and a worker G are all available, the P and G will be
3005 // returned. The returned P has not been wired yet.
3006 func checkIdleGCNoP() (*p, *g) {
3007 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3008 // must check again after acquiring a P. As an optimization, we also check
3009 // if an idle mark worker is needed at all. This is OK here, because if we
3010 // observe that one isn't needed, at least one is currently running. Even if
3011 // it stops running, its own journey into the scheduler should schedule it
3012 // again, if need be (at which point, this check will pass, if relevant).
3013 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3016 if !gcMarkWorkAvailable(nil) {
3020 // Work is available; we can start an idle GC worker only if there is
3021 // an available P and available worker G.
3023 // We can attempt to acquire these in either order, though both have
3024 // synchronization concerns (see below). Workers are almost always
3025 // available (see comment in findRunnableGCWorker for the one case
3026 // there may be none). Since we're slightly less likely to find a P,
3027 // check for that first.
3029 // Synchronization: note that we must hold sched.lock until we are
3030 // committed to keeping it. Otherwise we cannot put the unnecessary P
3031 // back in sched.pidle without performing the full set of idle
3032 // transition checks.
3034 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3035 // the assumption in gcControllerState.findRunnableGCWorker that an
3036 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3038 pp, now := pidleget(0)
3044 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3045 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3051 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3055 gcController.removeIdleMarkWorker()
3061 return pp, node.gp.ptr()
3064 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3065 // going to wake up before the when argument; or it wakes an idle P to service
3066 // timers and the network poller if there isn't one already.
3067 func wakeNetPoller(when int64) {
3068 if sched.lastpoll.Load() == 0 {
3069 // In findrunnable we ensure that when polling the pollUntil
3070 // field is either zero or the time to which the current
3071 // poll is expected to run. This can have a spurious wakeup
3072 // but should never miss a wakeup.
3073 pollerPollUntil := sched.pollUntil.Load()
3074 if pollerPollUntil == 0 || pollerPollUntil > when {
3078 // There are no threads in the network poller, try to get
3079 // one there so it can handle new timers.
3080 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3086 func resetspinning() {
3089 throw("resetspinning: not a spinning m")
3091 gp.m.spinning = false
3092 nmspinning := sched.nmspinning.Add(-1)
3094 throw("findrunnable: negative nmspinning")
3096 // M wakeup policy is deliberately somewhat conservative, so check if we
3097 // need to wakeup another P here. See "Worker thread parking/unparking"
3098 // comment at the top of the file for details.
3102 // injectglist adds each runnable G on the list to some run queue,
3103 // and clears glist. If there is no current P, they are added to the
3104 // global queue, and up to npidle M's are started to run them.
3105 // Otherwise, for each idle P, this adds a G to the global queue
3106 // and starts an M. Any remaining G's are added to the current P's
3108 // This may temporarily acquire sched.lock.
3109 // Can run concurrently with GC.
3110 func injectglist(glist *gList) {
3115 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3116 traceGoUnpark(gp, 0)
3120 // Mark all the goroutines as runnable before we put them
3121 // on the run queues.
3122 head := glist.head.ptr()
3125 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3128 casgstatus(gp, _Gwaiting, _Grunnable)
3131 // Turn the gList into a gQueue.
3137 startIdle := func(n int) {
3138 for ; n != 0 && sched.npidle.Load() != 0; n-- {
3143 pp := getg().m.p.ptr()
3146 globrunqputbatch(&q, int32(qsize))
3152 npidle := int(sched.npidle.Load())
3155 for n = 0; n < npidle && !q.empty(); n++ {
3161 globrunqputbatch(&globq, int32(n))
3168 runqputbatch(pp, &q, qsize)
3172 // One round of scheduler: find a runnable goroutine and execute it.
3178 throw("schedule: holding locks")
3181 if mp.lockedg != 0 {
3183 execute(mp.lockedg.ptr(), false) // Never returns.
3186 // We should not schedule away from a g that is executing a cgo call,
3187 // since the cgo call is using the m's g0 stack.
3189 throw("schedule: in cgo")
3196 // Safety check: if we are spinning, the run queue should be empty.
3197 // Check this before calling checkTimers, as that might call
3198 // goready to put a ready goroutine on the local run queue.
3199 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3200 throw("schedule: spinning with local work")
3203 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3205 // This thread is going to run a goroutine and is not spinning anymore,
3206 // so if it was marked as spinning we need to reset it now and potentially
3207 // start a new spinning M.
3212 if sched.disable.user && !schedEnabled(gp) {
3213 // Scheduling of this goroutine is disabled. Put it on
3214 // the list of pending runnable goroutines for when we
3215 // re-enable user scheduling and look again.
3217 if schedEnabled(gp) {
3218 // Something re-enabled scheduling while we
3219 // were acquiring the lock.
3222 sched.disable.runnable.pushBack(gp)
3229 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3230 // wake a P if there is one.
3234 if gp.lockedm != 0 {
3235 // Hands off own p to the locked m,
3236 // then blocks waiting for a new p.
3241 execute(gp, inheritTime)
3244 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3245 // Typically a caller sets gp's status away from Grunning and then
3246 // immediately calls dropg to finish the job. The caller is also responsible
3247 // for arranging that gp will be restarted using ready at an
3248 // appropriate time. After calling dropg and arranging for gp to be
3249 // readied later, the caller can do other work but eventually should
3250 // call schedule to restart the scheduling of goroutines on this m.
3254 setMNoWB(&gp.m.curg.m, nil)
3255 setGNoWB(&gp.m.curg, nil)
3258 // checkTimers runs any timers for the P that are ready.
3259 // If now is not 0 it is the current time.
3260 // It returns the passed time or the current time if now was passed as 0.
3261 // and the time when the next timer should run or 0 if there is no next timer,
3262 // and reports whether it ran any timers.
3263 // If the time when the next timer should run is not 0,
3264 // it is always larger than the returned time.
3265 // We pass now in and out to avoid extra calls of nanotime.
3267 //go:yeswritebarrierrec
3268 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3269 // If it's not yet time for the first timer, or the first adjusted
3270 // timer, then there is nothing to do.
3271 next := int64(atomic.Load64(&pp.timer0When))
3272 nextAdj := int64(atomic.Load64(&pp.timerModifiedEarliest))
3273 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3278 // No timers to run or adjust.
3279 return now, 0, false
3286 // Next timer is not ready to run, but keep going
3287 // if we would clear deleted timers.
3288 // This corresponds to the condition below where
3289 // we decide whether to call clearDeletedTimers.
3290 if pp != getg().m.p.ptr() || int(atomic.Load(&pp.deletedTimers)) <= int(atomic.Load(&pp.numTimers)/4) {
3291 return now, next, false
3295 lock(&pp.timersLock)
3297 if len(pp.timers) > 0 {
3298 adjusttimers(pp, now)
3299 for len(pp.timers) > 0 {
3300 // Note that runtimer may temporarily unlock
3302 if tw := runtimer(pp, now); tw != 0 {
3312 // If this is the local P, and there are a lot of deleted timers,
3313 // clear them out. We only do this for the local P to reduce
3314 // lock contention on timersLock.
3315 if pp == getg().m.p.ptr() && int(atomic.Load(&pp.deletedTimers)) > len(pp.timers)/4 {
3316 clearDeletedTimers(pp)
3319 unlock(&pp.timersLock)
3321 return now, pollUntil, ran
3324 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3325 unlock((*mutex)(lock))
3329 // park continuation on g0.
3330 func park_m(gp *g) {
3334 traceGoPark(mp.waittraceev, mp.waittraceskip)
3337 casgstatus(gp, _Grunning, _Gwaiting)
3340 if fn := mp.waitunlockf; fn != nil {
3341 ok := fn(gp, mp.waitlock)
3342 mp.waitunlockf = nil
3346 traceGoUnpark(gp, 2)
3348 casgstatus(gp, _Gwaiting, _Grunnable)
3349 execute(gp, true) // Schedule it back, never returns.
3355 func goschedImpl(gp *g) {
3356 status := readgstatus(gp)
3357 if status&^_Gscan != _Grunning {
3359 throw("bad g status")
3361 casgstatus(gp, _Grunning, _Grunnable)
3370 // Gosched continuation on g0.
3371 func gosched_m(gp *g) {
3378 // goschedguarded is a forbidden-states-avoided version of gosched_m
3379 func goschedguarded_m(gp *g) {
3381 if !canPreemptM(gp.m) {
3382 gogo(&gp.sched) // never return
3391 func gopreempt_m(gp *g) {
3398 // preemptPark parks gp and puts it in _Gpreempted.
3401 func preemptPark(gp *g) {
3403 traceGoPark(traceEvGoBlock, 0)
3405 status := readgstatus(gp)
3406 if status&^_Gscan != _Grunning {
3408 throw("bad g status")
3410 gp.waitreason = waitReasonPreempted
3412 if gp.asyncSafePoint {
3413 // Double-check that async preemption does not
3414 // happen in SPWRITE assembly functions.
3415 // isAsyncSafePoint must exclude this case.
3416 f := findfunc(gp.sched.pc)
3418 throw("preempt at unknown pc")
3420 if f.flag&funcFlag_SPWRITE != 0 {
3421 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3422 throw("preempt SPWRITE")
3426 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3427 // be in _Grunning when we dropg because then we'd be running
3428 // without an M, but the moment we're in _Gpreempted,
3429 // something could claim this G before we've fully cleaned it
3430 // up. Hence, we set the scan bit to lock down further
3431 // transitions until we can dropg.
3432 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3434 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3438 // goyield is like Gosched, but it:
3439 // - emits a GoPreempt trace event instead of a GoSched trace event
3440 // - puts the current G on the runq of the current P instead of the globrunq
3446 func goyield_m(gp *g) {
3451 casgstatus(gp, _Grunning, _Grunnable)
3453 runqput(pp, gp, false)
3457 // Finishes execution of the current goroutine.
3468 // goexit continuation on g0.
3469 func goexit0(gp *g) {
3473 casgstatus(gp, _Grunning, _Gdead)
3474 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3475 if isSystemGoroutine(gp, false) {
3479 locked := gp.lockedm != 0
3482 gp.preemptStop = false
3483 gp.paniconfault = false
3484 gp._defer = nil // should be true already but just in case.
3485 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3492 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3493 // Flush assist credit to the global pool. This gives
3494 // better information to pacing if the application is
3495 // rapidly creating an exiting goroutines.
3496 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3497 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3498 gcController.bgScanCredit.Add(scanCredit)
3499 gp.gcAssistBytes = 0
3504 if GOARCH == "wasm" { // no threads yet on wasm
3506 schedule() // never returns
3509 if mp.lockedInt != 0 {
3510 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3511 throw("internal lockOSThread error")
3515 // The goroutine may have locked this thread because
3516 // it put it in an unusual kernel state. Kill it
3517 // rather than returning it to the thread pool.
3519 // Return to mstart, which will release the P and exit
3521 if GOOS != "plan9" { // See golang.org/issue/22227.
3524 // Clear lockedExt on plan9 since we may end up re-using
3532 // save updates getg().sched to refer to pc and sp so that a following
3533 // gogo will restore pc and sp.
3535 // save must not have write barriers because invoking a write barrier
3536 // can clobber getg().sched.
3539 //go:nowritebarrierrec
3540 func save(pc, sp uintptr) {
3543 if gp == gp.m.g0 || gp == gp.m.gsignal {
3544 // m.g0.sched is special and must describe the context
3545 // for exiting the thread. mstart1 writes to it directly.
3546 // m.gsignal.sched should not be used at all.
3547 // This check makes sure save calls do not accidentally
3548 // run in contexts where they'd write to system g's.
3549 throw("save on system g not allowed")
3556 // We need to ensure ctxt is zero, but can't have a write
3557 // barrier here. However, it should always already be zero.
3559 if gp.sched.ctxt != nil {
3564 // The goroutine g is about to enter a system call.
3565 // Record that it's not using the cpu anymore.
3566 // This is called only from the go syscall library and cgocall,
3567 // not from the low-level system calls used by the runtime.
3569 // Entersyscall cannot split the stack: the save must
3570 // make g->sched refer to the caller's stack segment, because
3571 // entersyscall is going to return immediately after.
3573 // Nothing entersyscall calls can split the stack either.
3574 // We cannot safely move the stack during an active call to syscall,
3575 // because we do not know which of the uintptr arguments are
3576 // really pointers (back into the stack).
3577 // In practice, this means that we make the fast path run through
3578 // entersyscall doing no-split things, and the slow path has to use systemstack
3579 // to run bigger things on the system stack.
3581 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3582 // saved SP and PC are restored. This is needed when exitsyscall will be called
3583 // from a function further up in the call stack than the parent, as g->syscallsp
3584 // must always point to a valid stack frame. entersyscall below is the normal
3585 // entry point for syscalls, which obtains the SP and PC from the caller.
3588 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3589 // If the syscall does not block, that is it, we do not emit any other events.
3590 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3591 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3592 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3593 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3594 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3595 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3596 // and we wait for the increment before emitting traceGoSysExit.
3597 // Note that the increment is done even if tracing is not enabled,
3598 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3601 func reentersyscall(pc, sp uintptr) {
3604 // Disable preemption because during this function g is in Gsyscall status,
3605 // but can have inconsistent g->sched, do not let GC observe it.
3608 // Entersyscall must not call any function that might split/grow the stack.
3609 // (See details in comment above.)
3610 // Catch calls that might, by replacing the stack guard with something that
3611 // will trip any stack check and leaving a flag to tell newstack to die.
3612 gp.stackguard0 = stackPreempt
3613 gp.throwsplit = true
3615 // Leave SP around for GC and traceback.
3619 casgstatus(gp, _Grunning, _Gsyscall)
3620 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3621 systemstack(func() {
3622 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3623 throw("entersyscall")
3628 systemstack(traceGoSysCall)
3629 // systemstack itself clobbers g.sched.{pc,sp} and we might
3630 // need them later when the G is genuinely blocked in a
3635 if sched.sysmonwait.Load() {
3636 systemstack(entersyscall_sysmon)
3640 if gp.m.p.ptr().runSafePointFn != 0 {
3641 // runSafePointFn may stack split if run on this stack
3642 systemstack(runSafePointFn)
3646 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3647 gp.sysblocktraced = true
3652 atomic.Store(&pp.status, _Psyscall)
3653 if sched.gcwaiting.Load() {
3654 systemstack(entersyscall_gcwait)
3661 // Standard syscall entry used by the go syscall library and normal cgo calls.
3663 // This is exported via linkname to assembly in the syscall package and x/sys.
3666 //go:linkname entersyscall
3667 func entersyscall() {
3668 reentersyscall(getcallerpc(), getcallersp())
3671 func entersyscall_sysmon() {
3673 if sched.sysmonwait.Load() {
3674 sched.sysmonwait.Store(false)
3675 notewakeup(&sched.sysmonnote)
3680 func entersyscall_gcwait() {
3682 pp := gp.m.oldp.ptr()
3685 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3691 if sched.stopwait--; sched.stopwait == 0 {
3692 notewakeup(&sched.stopnote)
3698 // The same as entersyscall(), but with a hint that the syscall is blocking.
3701 func entersyscallblock() {
3704 gp.m.locks++ // see comment in entersyscall
3705 gp.throwsplit = true
3706 gp.stackguard0 = stackPreempt // see comment in entersyscall
3707 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3708 gp.sysblocktraced = true
3709 gp.m.p.ptr().syscalltick++
3711 // Leave SP around for GC and traceback.
3715 gp.syscallsp = gp.sched.sp
3716 gp.syscallpc = gp.sched.pc
3717 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3721 systemstack(func() {
3722 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3723 throw("entersyscallblock")
3726 casgstatus(gp, _Grunning, _Gsyscall)
3727 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3728 systemstack(func() {
3729 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3730 throw("entersyscallblock")
3734 systemstack(entersyscallblock_handoff)
3736 // Resave for traceback during blocked call.
3737 save(getcallerpc(), getcallersp())
3742 func entersyscallblock_handoff() {
3745 traceGoSysBlock(getg().m.p.ptr())
3747 handoffp(releasep())
3750 // The goroutine g exited its system call.
3751 // Arrange for it to run on a cpu again.
3752 // This is called only from the go syscall library, not
3753 // from the low-level system calls used by the runtime.
3755 // Write barriers are not allowed because our P may have been stolen.
3757 // This is exported via linkname to assembly in the syscall package.
3760 //go:nowritebarrierrec
3761 //go:linkname exitsyscall
3762 func exitsyscall() {
3765 gp.m.locks++ // see comment in entersyscall
3766 if getcallersp() > gp.syscallsp {
3767 throw("exitsyscall: syscall frame is no longer valid")
3771 oldp := gp.m.oldp.ptr()
3773 if exitsyscallfast(oldp) {
3774 // When exitsyscallfast returns success, we have a P so can now use
3776 if goroutineProfile.active {
3777 // Make sure that gp has had its stack written out to the goroutine
3778 // profile, exactly as it was when the goroutine profiler first
3779 // stopped the world.
3780 systemstack(func() {
3781 tryRecordGoroutineProfileWB(gp)
3785 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
3786 systemstack(traceGoStart)
3789 // There's a cpu for us, so we can run.
3790 gp.m.p.ptr().syscalltick++
3791 // We need to cas the status and scan before resuming...
3792 casgstatus(gp, _Gsyscall, _Grunning)
3794 // Garbage collector isn't running (since we are),
3795 // so okay to clear syscallsp.
3799 // restore the preemption request in case we've cleared it in newstack
3800 gp.stackguard0 = stackPreempt
3802 // otherwise restore the real _StackGuard, we've spoiled it in entersyscall/entersyscallblock
3803 gp.stackguard0 = gp.stack.lo + _StackGuard
3805 gp.throwsplit = false
3807 if sched.disable.user && !schedEnabled(gp) {
3808 // Scheduling of this goroutine is disabled.
3817 // Wait till traceGoSysBlock event is emitted.
3818 // This ensures consistency of the trace (the goroutine is started after it is blocked).
3819 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
3822 // We can't trace syscall exit right now because we don't have a P.
3823 // Tracing code can invoke write barriers that cannot run without a P.
3824 // So instead we remember the syscall exit time and emit the event
3825 // in execute when we have a P.
3826 gp.sysexitticks = cputicks()
3831 // Call the scheduler.
3834 // Scheduler returned, so we're allowed to run now.
3835 // Delete the syscallsp information that we left for
3836 // the garbage collector during the system call.
3837 // Must wait until now because until gosched returns
3838 // we don't know for sure that the garbage collector
3841 gp.m.p.ptr().syscalltick++
3842 gp.throwsplit = false
3846 func exitsyscallfast(oldp *p) bool {
3849 // Freezetheworld sets stopwait but does not retake P's.
3850 if sched.stopwait == freezeStopWait {
3854 // Try to re-acquire the last P.
3855 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
3856 // There's a cpu for us, so we can run.
3858 exitsyscallfast_reacquired()
3862 // Try to get any other idle P.
3863 if sched.pidle != 0 {
3865 systemstack(func() {
3866 ok = exitsyscallfast_pidle()
3867 if ok && trace.enabled {
3869 // Wait till traceGoSysBlock event is emitted.
3870 // This ensures consistency of the trace (the goroutine is started after it is blocked).
3871 for oldp.syscalltick == gp.m.syscalltick {
3885 // exitsyscallfast_reacquired is the exitsyscall path on which this G
3886 // has successfully reacquired the P it was running on before the
3890 func exitsyscallfast_reacquired() {
3892 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
3894 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
3895 // traceGoSysBlock for this syscall was already emitted,
3896 // but here we effectively retake the p from the new syscall running on the same p.
3897 systemstack(func() {
3898 // Denote blocking of the new syscall.
3899 traceGoSysBlock(gp.m.p.ptr())
3900 // Denote completion of the current syscall.
3904 gp.m.p.ptr().syscalltick++
3908 func exitsyscallfast_pidle() bool {
3910 pp, _ := pidleget(0)
3911 if pp != nil && sched.sysmonwait.Load() {
3912 sched.sysmonwait.Store(false)
3913 notewakeup(&sched.sysmonnote)
3923 // exitsyscall slow path on g0.
3924 // Failed to acquire P, enqueue gp as runnable.
3926 // Called via mcall, so gp is the calling g from this M.
3928 //go:nowritebarrierrec
3929 func exitsyscall0(gp *g) {
3930 casgstatus(gp, _Gsyscall, _Grunnable)
3934 if schedEnabled(gp) {
3941 // Below, we stoplockedm if gp is locked. globrunqput releases
3942 // ownership of gp, so we must check if gp is locked prior to
3943 // committing the release by unlocking sched.lock, otherwise we
3944 // could race with another M transitioning gp from unlocked to
3946 locked = gp.lockedm != 0
3947 } else if sched.sysmonwait.Load() {
3948 sched.sysmonwait.Store(false)
3949 notewakeup(&sched.sysmonnote)
3954 execute(gp, false) // Never returns.
3957 // Wait until another thread schedules gp and so m again.
3959 // N.B. lockedm must be this M, as this g was running on this M
3960 // before entersyscall.
3962 execute(gp, false) // Never returns.
3965 schedule() // Never returns.
3968 // Called from syscall package before fork.
3970 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
3972 func syscall_runtime_BeforeFork() {
3975 // Block signals during a fork, so that the child does not run
3976 // a signal handler before exec if a signal is sent to the process
3977 // group. See issue #18600.
3979 sigsave(&gp.m.sigmask)
3982 // This function is called before fork in syscall package.
3983 // Code between fork and exec must not allocate memory nor even try to grow stack.
3984 // Here we spoil g->_StackGuard to reliably detect any attempts to grow stack.
3985 // runtime_AfterFork will undo this in parent process, but not in child.
3986 gp.stackguard0 = stackFork
3989 // Called from syscall package after fork in parent.
3991 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
3993 func syscall_runtime_AfterFork() {
3996 // See the comments in beforefork.
3997 gp.stackguard0 = gp.stack.lo + _StackGuard
3999 msigrestore(gp.m.sigmask)
4004 // inForkedChild is true while manipulating signals in the child process.
4005 // This is used to avoid calling libc functions in case we are using vfork.
4006 var inForkedChild bool
4008 // Called from syscall package after fork in child.
4009 // It resets non-sigignored signals to the default handler, and
4010 // restores the signal mask in preparation for the exec.
4012 // Because this might be called during a vfork, and therefore may be
4013 // temporarily sharing address space with the parent process, this must
4014 // not change any global variables or calling into C code that may do so.
4016 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4018 //go:nowritebarrierrec
4019 func syscall_runtime_AfterForkInChild() {
4020 // It's OK to change the global variable inForkedChild here
4021 // because we are going to change it back. There is no race here,
4022 // because if we are sharing address space with the parent process,
4023 // then the parent process can not be running concurrently.
4024 inForkedChild = true
4026 clearSignalHandlers()
4028 // When we are the child we are the only thread running,
4029 // so we know that nothing else has changed gp.m.sigmask.
4030 msigrestore(getg().m.sigmask)
4032 inForkedChild = false
4035 // pendingPreemptSignals is the number of preemption signals
4036 // that have been sent but not received. This is only used on Darwin.
4038 var pendingPreemptSignals atomic.Int32
4040 // Called from syscall package before Exec.
4042 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4043 func syscall_runtime_BeforeExec() {
4044 // Prevent thread creation during exec.
4047 // On Darwin, wait for all pending preemption signals to
4048 // be received. See issue #41702.
4049 if GOOS == "darwin" || GOOS == "ios" {
4050 for pendingPreemptSignals.Load() > 0 {
4056 // Called from syscall package after Exec.
4058 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4059 func syscall_runtime_AfterExec() {
4063 // Allocate a new g, with a stack big enough for stacksize bytes.
4064 func malg(stacksize int32) *g {
4067 stacksize = round2(_StackSystem + stacksize)
4068 systemstack(func() {
4069 newg.stack = stackalloc(uint32(stacksize))
4071 newg.stackguard0 = newg.stack.lo + _StackGuard
4072 newg.stackguard1 = ^uintptr(0)
4073 // Clear the bottom word of the stack. We record g
4074 // there on gsignal stack during VDSO on ARM and ARM64.
4075 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4080 // Create a new g running fn.
4081 // Put it on the queue of g's waiting to run.
4082 // The compiler turns a go statement into a call to this.
4083 func newproc(fn *funcval) {
4086 systemstack(func() {
4087 newg := newproc1(fn, gp, pc)
4089 pp := getg().m.p.ptr()
4090 runqput(pp, newg, true)
4098 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4099 // address of the go statement that created this. The caller is responsible
4100 // for adding the new g to the scheduler.
4101 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4103 fatal("go of nil func value")
4106 mp := acquirem() // disable preemption because we hold M and P in local vars.
4110 newg = malg(_StackMin)
4111 casgstatus(newg, _Gidle, _Gdead)
4112 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4114 if newg.stack.hi == 0 {
4115 throw("newproc1: newg missing stack")
4118 if readgstatus(newg) != _Gdead {
4119 throw("newproc1: new g is not Gdead")
4122 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4123 totalSize = alignUp(totalSize, sys.StackAlign)
4124 sp := newg.stack.hi - totalSize
4128 *(*uintptr)(unsafe.Pointer(sp)) = 0
4130 spArg += sys.MinFrameSize
4133 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4136 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4137 newg.sched.g = guintptr(unsafe.Pointer(newg))
4138 gostartcallfn(&newg.sched, fn)
4139 newg.gopc = callerpc
4140 newg.ancestors = saveAncestors(callergp)
4141 newg.startpc = fn.fn
4142 if isSystemGoroutine(newg, false) {
4145 // Only user goroutines inherit pprof labels.
4147 newg.labels = mp.curg.labels
4149 if goroutineProfile.active {
4150 // A concurrent goroutine profile is running. It should include
4151 // exactly the set of goroutines that were alive when the goroutine
4152 // profiler first stopped the world. That does not include newg, so
4153 // mark it as not needing a profile before transitioning it from
4155 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4158 // Track initial transition?
4159 newg.trackingSeq = uint8(fastrand())
4160 if newg.trackingSeq%gTrackingPeriod == 0 {
4161 newg.tracking = true
4163 casgstatus(newg, _Gdead, _Grunnable)
4164 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4166 if pp.goidcache == pp.goidcacheend {
4167 // Sched.goidgen is the last allocated id,
4168 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4169 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4170 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4171 pp.goidcache -= _GoidCacheBatch - 1
4172 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4174 newg.goid = pp.goidcache
4177 newg.racectx = racegostart(callerpc)
4178 if newg.labels != nil {
4179 // See note in proflabel.go on labelSync's role in synchronizing
4180 // with the reads in the signal handler.
4181 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4185 traceGoCreate(newg, newg.startpc)
4192 // saveAncestors copies previous ancestors of the given caller g and
4193 // includes infor for the current caller into a new set of tracebacks for
4194 // a g being created.
4195 func saveAncestors(callergp *g) *[]ancestorInfo {
4196 // Copy all prior info, except for the root goroutine (goid 0).
4197 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4200 var callerAncestors []ancestorInfo
4201 if callergp.ancestors != nil {
4202 callerAncestors = *callergp.ancestors
4204 n := int32(len(callerAncestors)) + 1
4205 if n > debug.tracebackancestors {
4206 n = debug.tracebackancestors
4208 ancestors := make([]ancestorInfo, n)
4209 copy(ancestors[1:], callerAncestors)
4211 var pcs [_TracebackMaxFrames]uintptr
4212 npcs := gcallers(callergp, 0, pcs[:])
4213 ipcs := make([]uintptr, npcs)
4215 ancestors[0] = ancestorInfo{
4217 goid: callergp.goid,
4218 gopc: callergp.gopc,
4221 ancestorsp := new([]ancestorInfo)
4222 *ancestorsp = ancestors
4226 // Put on gfree list.
4227 // If local list is too long, transfer a batch to the global list.
4228 func gfput(pp *p, gp *g) {
4229 if readgstatus(gp) != _Gdead {
4230 throw("gfput: bad status (not Gdead)")
4233 stksize := gp.stack.hi - gp.stack.lo
4235 if stksize != uintptr(startingStackSize) {
4236 // non-standard stack size - free it.
4245 if pp.gFree.n >= 64 {
4251 for pp.gFree.n >= 32 {
4252 gp := pp.gFree.pop()
4254 if gp.stack.lo == 0 {
4261 lock(&sched.gFree.lock)
4262 sched.gFree.noStack.pushAll(noStackQ)
4263 sched.gFree.stack.pushAll(stackQ)
4264 sched.gFree.n += inc
4265 unlock(&sched.gFree.lock)
4269 // Get from gfree list.
4270 // If local list is empty, grab a batch from global list.
4271 func gfget(pp *p) *g {
4273 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4274 lock(&sched.gFree.lock)
4275 // Move a batch of free Gs to the P.
4276 for pp.gFree.n < 32 {
4277 // Prefer Gs with stacks.
4278 gp := sched.gFree.stack.pop()
4280 gp = sched.gFree.noStack.pop()
4289 unlock(&sched.gFree.lock)
4292 gp := pp.gFree.pop()
4297 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4298 // Deallocate old stack. We kept it in gfput because it was the
4299 // right size when the goroutine was put on the free list, but
4300 // the right size has changed since then.
4301 systemstack(func() {
4308 if gp.stack.lo == 0 {
4309 // Stack was deallocated in gfput or just above. Allocate a new one.
4310 systemstack(func() {
4311 gp.stack = stackalloc(startingStackSize)
4313 gp.stackguard0 = gp.stack.lo + _StackGuard
4316 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4319 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4322 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4328 // Purge all cached G's from gfree list to the global list.
4329 func gfpurge(pp *p) {
4335 for !pp.gFree.empty() {
4336 gp := pp.gFree.pop()
4338 if gp.stack.lo == 0 {
4345 lock(&sched.gFree.lock)
4346 sched.gFree.noStack.pushAll(noStackQ)
4347 sched.gFree.stack.pushAll(stackQ)
4348 sched.gFree.n += inc
4349 unlock(&sched.gFree.lock)
4352 // Breakpoint executes a breakpoint trap.
4357 // dolockOSThread is called by LockOSThread and lockOSThread below
4358 // after they modify m.locked. Do not allow preemption during this call,
4359 // or else the m might be different in this function than in the caller.
4362 func dolockOSThread() {
4363 if GOARCH == "wasm" {
4364 return // no threads on wasm yet
4367 gp.m.lockedg.set(gp)
4368 gp.lockedm.set(gp.m)
4373 // LockOSThread wires the calling goroutine to its current operating system thread.
4374 // The calling goroutine will always execute in that thread,
4375 // and no other goroutine will execute in it,
4376 // until the calling goroutine has made as many calls to
4377 // UnlockOSThread as to LockOSThread.
4378 // If the calling goroutine exits without unlocking the thread,
4379 // the thread will be terminated.
4381 // All init functions are run on the startup thread. Calling LockOSThread
4382 // from an init function will cause the main function to be invoked on
4385 // A goroutine should call LockOSThread before calling OS services or
4386 // non-Go library functions that depend on per-thread state.
4387 func LockOSThread() {
4388 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4389 // If we need to start a new thread from the locked
4390 // thread, we need the template thread. Start it now
4391 // while we're in a known-good state.
4392 startTemplateThread()
4396 if gp.m.lockedExt == 0 {
4398 panic("LockOSThread nesting overflow")
4404 func lockOSThread() {
4405 getg().m.lockedInt++
4409 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4410 // after they update m->locked. Do not allow preemption during this call,
4411 // or else the m might be in different in this function than in the caller.
4414 func dounlockOSThread() {
4415 if GOARCH == "wasm" {
4416 return // no threads on wasm yet
4419 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4428 // UnlockOSThread undoes an earlier call to LockOSThread.
4429 // If this drops the number of active LockOSThread calls on the
4430 // calling goroutine to zero, it unwires the calling goroutine from
4431 // its fixed operating system thread.
4432 // If there are no active LockOSThread calls, this is a no-op.
4434 // Before calling UnlockOSThread, the caller must ensure that the OS
4435 // thread is suitable for running other goroutines. If the caller made
4436 // any permanent changes to the state of the thread that would affect
4437 // other goroutines, it should not call this function and thus leave
4438 // the goroutine locked to the OS thread until the goroutine (and
4439 // hence the thread) exits.
4440 func UnlockOSThread() {
4442 if gp.m.lockedExt == 0 {
4450 func unlockOSThread() {
4452 if gp.m.lockedInt == 0 {
4453 systemstack(badunlockosthread)
4459 func badunlockosthread() {
4460 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4463 func gcount() int32 {
4464 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4465 for _, pp := range allp {
4469 // All these variables can be changed concurrently, so the result can be inconsistent.
4470 // But at least the current goroutine is running.
4477 func mcount() int32 {
4478 return int32(sched.mnext - sched.nmfreed)
4486 func _System() { _System() }
4487 func _ExternalCode() { _ExternalCode() }
4488 func _LostExternalCode() { _LostExternalCode() }
4489 func _GC() { _GC() }
4490 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4491 func _VDSO() { _VDSO() }
4493 // Called if we receive a SIGPROF signal.
4494 // Called by the signal handler, may run during STW.
4496 //go:nowritebarrierrec
4497 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4502 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4503 // We must check this to avoid a deadlock between setcpuprofilerate
4504 // and the call to cpuprof.add, below.
4505 if mp != nil && mp.profilehz == 0 {
4509 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4510 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4511 // the critical section, it creates a deadlock (when writing the sample).
4512 // As a workaround, create a counter of SIGPROFs while in critical section
4513 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4514 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4515 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4516 if f := findfunc(pc); f.valid() {
4517 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4518 cpuprof.lostAtomic++
4522 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4523 // runtime/internal/atomic functions call into kernel
4524 // helpers on arm < 7. See
4525 // runtime/internal/atomic/sys_linux_arm.s.
4526 cpuprof.lostAtomic++
4531 // Profiling runs concurrently with GC, so it must not allocate.
4532 // Set a trap in case the code does allocate.
4533 // Note that on windows, one thread takes profiles of all the
4534 // other threads, so mp is usually not getg().m.
4535 // In fact mp may not even be stopped.
4536 // See golang.org/issue/17165.
4537 getg().m.mallocing++
4539 var stk [maxCPUProfStack]uintptr
4541 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4543 // Check cgoCallersUse to make sure that we are not
4544 // interrupting other code that is fiddling with
4545 // cgoCallers. We are running in a signal handler
4546 // with all signals blocked, so we don't have to worry
4547 // about any other code interrupting us.
4548 if atomic.Load(&mp.cgoCallersUse) == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4549 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4552 copy(stk[:], mp.cgoCallers[:cgoOff])
4553 mp.cgoCallers[0] = 0
4556 // Collect Go stack that leads to the cgo call.
4557 n = gentraceback(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, 0, &stk[cgoOff], len(stk)-cgoOff, nil, nil, 0)
4562 n = gentraceback(pc, sp, lr, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
4566 // Normal traceback is impossible or has failed.
4567 // See if it falls into several common cases.
4569 if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4570 // Libcall, i.e. runtime syscall on windows.
4571 // Collect Go stack that leads to the call.
4572 n = gentraceback(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), 0, &stk[0], len(stk), nil, nil, 0)
4574 if n == 0 && mp != nil && mp.vdsoSP != 0 {
4575 n = gentraceback(mp.vdsoPC, mp.vdsoSP, 0, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
4578 // If all of the above has failed, account it against abstract "System" or "GC".
4581 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4582 } else if pc > firstmoduledata.etext {
4583 // "ExternalCode" is better than "etext".
4584 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4587 if mp.preemptoff != "" {
4588 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4590 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4596 // Note: it can happen on Windows that we interrupted a system thread
4597 // with no g, so gp could nil. The other nil checks are done out of
4598 // caution, but not expected to be nil in practice.
4599 var tagPtr *unsafe.Pointer
4600 if gp != nil && gp.m != nil && gp.m.curg != nil {
4601 tagPtr = &gp.m.curg.labels
4603 cpuprof.add(tagPtr, stk[:n])
4607 if gp != nil && gp.m != nil {
4608 if gp.m.curg != nil {
4613 traceCPUSample(gprof, pp, stk[:n])
4615 getg().m.mallocing--
4618 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4619 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4620 func setcpuprofilerate(hz int32) {
4621 // Force sane arguments.
4626 // Disable preemption, otherwise we can be rescheduled to another thread
4627 // that has profiling enabled.
4631 // Stop profiler on this thread so that it is safe to lock prof.
4632 // if a profiling signal came in while we had prof locked,
4633 // it would deadlock.
4634 setThreadCPUProfiler(0)
4636 for !atomic.Cas(&prof.signalLock, 0, 1) {
4640 setProcessCPUProfiler(hz)
4643 atomic.Store(&prof.signalLock, 0)
4646 sched.profilehz = hz
4650 setThreadCPUProfiler(hz)
4656 // init initializes pp, which may be a freshly allocated p or a
4657 // previously destroyed p, and transitions it to status _Pgcstop.
4658 func (pp *p) init(id int32) {
4660 pp.status = _Pgcstop
4661 pp.sudogcache = pp.sudogbuf[:0]
4662 pp.deferpool = pp.deferpoolbuf[:0]
4664 if pp.mcache == nil {
4667 throw("missing mcache?")
4669 // Use the bootstrap mcache0. Only one P will get
4670 // mcache0: the one with ID 0.
4673 pp.mcache = allocmcache()
4676 if raceenabled && pp.raceprocctx == 0 {
4678 pp.raceprocctx = raceprocctx0
4679 raceprocctx0 = 0 // bootstrap
4681 pp.raceprocctx = raceproccreate()
4684 lockInit(&pp.timersLock, lockRankTimers)
4686 // This P may get timers when it starts running. Set the mask here
4687 // since the P may not go through pidleget (notably P 0 on startup).
4689 // Similarly, we may not go through pidleget before this P starts
4690 // running if it is P 0 on startup.
4694 // destroy releases all of the resources associated with pp and
4695 // transitions it to status _Pdead.
4697 // sched.lock must be held and the world must be stopped.
4698 func (pp *p) destroy() {
4699 assertLockHeld(&sched.lock)
4700 assertWorldStopped()
4702 // Move all runnable goroutines to the global queue
4703 for pp.runqhead != pp.runqtail {
4704 // Pop from tail of local queue
4706 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
4707 // Push onto head of global queue
4710 if pp.runnext != 0 {
4711 globrunqputhead(pp.runnext.ptr())
4714 if len(pp.timers) > 0 {
4715 plocal := getg().m.p.ptr()
4716 // The world is stopped, but we acquire timersLock to
4717 // protect against sysmon calling timeSleepUntil.
4718 // This is the only case where we hold the timersLock of
4719 // more than one P, so there are no deadlock concerns.
4720 lock(&plocal.timersLock)
4721 lock(&pp.timersLock)
4722 moveTimers(plocal, pp.timers)
4725 pp.deletedTimers = 0
4726 atomic.Store64(&pp.timer0When, 0)
4727 unlock(&pp.timersLock)
4728 unlock(&plocal.timersLock)
4730 // Flush p's write barrier buffer.
4731 if gcphase != _GCoff {
4735 for i := range pp.sudogbuf {
4736 pp.sudogbuf[i] = nil
4738 pp.sudogcache = pp.sudogbuf[:0]
4739 for j := range pp.deferpoolbuf {
4740 pp.deferpoolbuf[j] = nil
4742 pp.deferpool = pp.deferpoolbuf[:0]
4743 systemstack(func() {
4744 for i := 0; i < pp.mspancache.len; i++ {
4745 // Safe to call since the world is stopped.
4746 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
4748 pp.mspancache.len = 0
4750 pp.pcache.flush(&mheap_.pages)
4751 unlock(&mheap_.lock)
4753 freemcache(pp.mcache)
4758 if pp.timerRaceCtx != 0 {
4759 // The race detector code uses a callback to fetch
4760 // the proc context, so arrange for that callback
4761 // to see the right thing.
4762 // This hack only works because we are the only
4768 racectxend(pp.timerRaceCtx)
4773 raceprocdestroy(pp.raceprocctx)
4780 // Change number of processors.
4782 // sched.lock must be held, and the world must be stopped.
4784 // gcworkbufs must not be being modified by either the GC or the write barrier
4785 // code, so the GC must not be running if the number of Ps actually changes.
4787 // Returns list of Ps with local work, they need to be scheduled by the caller.
4788 func procresize(nprocs int32) *p {
4789 assertLockHeld(&sched.lock)
4790 assertWorldStopped()
4793 if old < 0 || nprocs <= 0 {
4794 throw("procresize: invalid arg")
4797 traceGomaxprocs(nprocs)
4800 // update statistics
4802 if sched.procresizetime != 0 {
4803 sched.totaltime += int64(old) * (now - sched.procresizetime)
4805 sched.procresizetime = now
4807 maskWords := (nprocs + 31) / 32
4809 // Grow allp if necessary.
4810 if nprocs > int32(len(allp)) {
4811 // Synchronize with retake, which could be running
4812 // concurrently since it doesn't run on a P.
4814 if nprocs <= int32(cap(allp)) {
4815 allp = allp[:nprocs]
4817 nallp := make([]*p, nprocs)
4818 // Copy everything up to allp's cap so we
4819 // never lose old allocated Ps.
4820 copy(nallp, allp[:cap(allp)])
4824 if maskWords <= int32(cap(idlepMask)) {
4825 idlepMask = idlepMask[:maskWords]
4826 timerpMask = timerpMask[:maskWords]
4828 nidlepMask := make([]uint32, maskWords)
4829 // No need to copy beyond len, old Ps are irrelevant.
4830 copy(nidlepMask, idlepMask)
4831 idlepMask = nidlepMask
4833 ntimerpMask := make([]uint32, maskWords)
4834 copy(ntimerpMask, timerpMask)
4835 timerpMask = ntimerpMask
4840 // initialize new P's
4841 for i := old; i < nprocs; i++ {
4847 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
4851 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
4852 // continue to use the current P
4853 gp.m.p.ptr().status = _Prunning
4854 gp.m.p.ptr().mcache.prepareForSweep()
4856 // release the current P and acquire allp[0].
4858 // We must do this before destroying our current P
4859 // because p.destroy itself has write barriers, so we
4860 // need to do that from a valid P.
4863 // Pretend that we were descheduled
4864 // and then scheduled again to keep
4867 traceProcStop(gp.m.p.ptr())
4881 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
4884 // release resources from unused P's
4885 for i := nprocs; i < old; i++ {
4888 // can't free P itself because it can be referenced by an M in syscall
4892 if int32(len(allp)) != nprocs {
4894 allp = allp[:nprocs]
4895 idlepMask = idlepMask[:maskWords]
4896 timerpMask = timerpMask[:maskWords]
4901 for i := nprocs - 1; i >= 0; i-- {
4903 if gp.m.p.ptr() == pp {
4911 pp.link.set(runnablePs)
4915 stealOrder.reset(uint32(nprocs))
4916 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
4917 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
4919 // Notify the limiter that the amount of procs has changed.
4920 gcCPULimiter.resetCapacity(now, nprocs)
4925 // Associate p and the current m.
4927 // This function is allowed to have write barriers even if the caller
4928 // isn't because it immediately acquires pp.
4930 //go:yeswritebarrierrec
4931 func acquirep(pp *p) {
4932 // Do the part that isn't allowed to have write barriers.
4935 // Have p; write barriers now allowed.
4937 // Perform deferred mcache flush before this P can allocate
4938 // from a potentially stale mcache.
4939 pp.mcache.prepareForSweep()
4946 // wirep is the first step of acquirep, which actually associates the
4947 // current M to pp. This is broken out so we can disallow write
4948 // barriers for this part, since we don't yet have a P.
4950 //go:nowritebarrierrec
4956 throw("wirep: already in go")
4958 if pp.m != 0 || pp.status != _Pidle {
4963 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
4964 throw("wirep: invalid p state")
4968 pp.status = _Prunning
4971 // Disassociate p and the current m.
4972 func releasep() *p {
4976 throw("releasep: invalid arg")
4979 if pp.m.ptr() != gp.m || pp.status != _Prunning {
4980 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
4981 throw("releasep: invalid p state")
4984 traceProcStop(gp.m.p.ptr())
4992 func incidlelocked(v int32) {
4994 sched.nmidlelocked += v
5001 // Check for deadlock situation.
5002 // The check is based on number of running M's, if 0 -> deadlock.
5003 // sched.lock must be held.
5005 assertLockHeld(&sched.lock)
5007 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5008 // there are no running goroutines. The calling program is
5009 // assumed to be running.
5010 if islibrary || isarchive {
5014 // If we are dying because of a signal caught on an already idle thread,
5015 // freezetheworld will cause all running threads to block.
5016 // And runtime will essentially enter into deadlock state,
5017 // except that there is a thread that will call exit soon.
5018 if panicking.Load() > 0 {
5022 // If we are not running under cgo, but we have an extra M then account
5023 // for it. (It is possible to have an extra M on Windows without cgo to
5024 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5027 if !iscgo && cgoHasExtraM {
5028 mp := lockextra(true)
5029 haveExtraM := extraMCount > 0
5036 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5041 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5042 throw("checkdead: inconsistent counts")
5046 forEachG(func(gp *g) {
5047 if isSystemGoroutine(gp, false) {
5050 s := readgstatus(gp)
5051 switch s &^ _Gscan {
5058 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5059 throw("checkdead: runnable g")
5062 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5063 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5064 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5067 // Maybe jump time forward for playground.
5069 if when := timeSleepUntil(); when < maxWhen {
5072 // Start an M to steal the timer.
5073 pp, _ := pidleget(faketime)
5075 // There should always be a free P since
5076 // nothing is running.
5077 throw("checkdead: no p for timer")
5081 // There should always be a free M since
5082 // nothing is running.
5083 throw("checkdead: no m for timer")
5085 // M must be spinning to steal. We set this to be
5086 // explicit, but since this is the only M it would
5087 // become spinning on its own anyways.
5088 sched.nmspinning.Add(1)
5091 notewakeup(&mp.park)
5096 // There are no goroutines running, so we can look at the P's.
5097 for _, pp := range allp {
5098 if len(pp.timers) > 0 {
5103 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5104 fatal("all goroutines are asleep - deadlock!")
5107 // forcegcperiod is the maximum time in nanoseconds between garbage
5108 // collections. If we go this long without a garbage collection, one
5109 // is forced to run.
5111 // This is a variable for testing purposes. It normally doesn't change.
5112 var forcegcperiod int64 = 2 * 60 * 1e9
5114 // needSysmonWorkaround is true if the workaround for
5115 // golang.org/issue/42515 is needed on NetBSD.
5116 var needSysmonWorkaround bool = false
5118 // Always runs without a P, so write barriers are not allowed.
5120 //go:nowritebarrierrec
5127 lasttrace := int64(0)
5128 idle := 0 // how many cycles in succession we had not wokeup somebody
5132 if idle == 0 { // start with 20us sleep...
5134 } else if idle > 50 { // start doubling the sleep after 1ms...
5137 if delay > 10*1000 { // up to 10ms
5142 // sysmon should not enter deep sleep if schedtrace is enabled so that
5143 // it can print that information at the right time.
5145 // It should also not enter deep sleep if there are any active P's so
5146 // that it can retake P's from syscalls, preempt long running G's, and
5147 // poll the network if all P's are busy for long stretches.
5149 // It should wakeup from deep sleep if any P's become active either due
5150 // to exiting a syscall or waking up due to a timer expiring so that it
5151 // can resume performing those duties. If it wakes from a syscall it
5152 // resets idle and delay as a bet that since it had retaken a P from a
5153 // syscall before, it may need to do it again shortly after the
5154 // application starts work again. It does not reset idle when waking
5155 // from a timer to avoid adding system load to applications that spend
5156 // most of their time sleeping.
5158 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5160 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5161 syscallWake := false
5162 next := timeSleepUntil()
5164 sched.sysmonwait.Store(true)
5166 // Make wake-up period small enough
5167 // for the sampling to be correct.
5168 sleep := forcegcperiod / 2
5169 if next-now < sleep {
5172 shouldRelax := sleep >= osRelaxMinNS
5176 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5181 sched.sysmonwait.Store(false)
5182 noteclear(&sched.sysmonnote)
5192 lock(&sched.sysmonlock)
5193 // Update now in case we blocked on sysmonnote or spent a long time
5194 // blocked on schedlock or sysmonlock above.
5197 // trigger libc interceptors if needed
5198 if *cgo_yield != nil {
5199 asmcgocall(*cgo_yield, nil)
5201 // poll network if not polled for more than 10ms
5202 lastpoll := sched.lastpoll.Load()
5203 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5204 sched.lastpoll.CompareAndSwap(lastpoll, now)
5205 list := netpoll(0) // non-blocking - returns list of goroutines
5207 // Need to decrement number of idle locked M's
5208 // (pretending that one more is running) before injectglist.
5209 // Otherwise it can lead to the following situation:
5210 // injectglist grabs all P's but before it starts M's to run the P's,
5211 // another M returns from syscall, finishes running its G,
5212 // observes that there is no work to do and no other running M's
5213 // and reports deadlock.
5219 if GOOS == "netbsd" && needSysmonWorkaround {
5220 // netpoll is responsible for waiting for timer
5221 // expiration, so we typically don't have to worry
5222 // about starting an M to service timers. (Note that
5223 // sleep for timeSleepUntil above simply ensures sysmon
5224 // starts running again when that timer expiration may
5225 // cause Go code to run again).
5227 // However, netbsd has a kernel bug that sometimes
5228 // misses netpollBreak wake-ups, which can lead to
5229 // unbounded delays servicing timers. If we detect this
5230 // overrun, then startm to get something to handle the
5233 // See issue 42515 and
5234 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5235 if next := timeSleepUntil(); next < now {
5239 if scavenger.sysmonWake.Load() != 0 {
5240 // Kick the scavenger awake if someone requested it.
5243 // retake P's blocked in syscalls
5244 // and preempt long running G's
5245 if retake(now) != 0 {
5250 // check if we need to force a GC
5251 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && atomic.Load(&forcegc.idle) != 0 {
5255 list.push(forcegc.g)
5257 unlock(&forcegc.lock)
5259 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5261 schedtrace(debug.scheddetail > 0)
5263 unlock(&sched.sysmonlock)
5267 type sysmontick struct {
5274 // forcePreemptNS is the time slice given to a G before it is
5276 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5278 func retake(now int64) uint32 {
5280 // Prevent allp slice changes. This lock will be completely
5281 // uncontended unless we're already stopping the world.
5283 // We can't use a range loop over allp because we may
5284 // temporarily drop the allpLock. Hence, we need to re-fetch
5285 // allp each time around the loop.
5286 for i := 0; i < len(allp); i++ {
5289 // This can happen if procresize has grown
5290 // allp but not yet created new Ps.
5293 pd := &pp.sysmontick
5296 if s == _Prunning || s == _Psyscall {
5297 // Preempt G if it's running for too long.
5298 t := int64(pp.schedtick)
5299 if int64(pd.schedtick) != t {
5300 pd.schedtick = uint32(t)
5302 } else if pd.schedwhen+forcePreemptNS <= now {
5304 // In case of syscall, preemptone() doesn't
5305 // work, because there is no M wired to P.
5310 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5311 t := int64(pp.syscalltick)
5312 if !sysretake && int64(pd.syscalltick) != t {
5313 pd.syscalltick = uint32(t)
5314 pd.syscallwhen = now
5317 // On the one hand we don't want to retake Ps if there is no other work to do,
5318 // but on the other hand we want to retake them eventually
5319 // because they can prevent the sysmon thread from deep sleep.
5320 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5323 // Drop allpLock so we can take sched.lock.
5325 // Need to decrement number of idle locked M's
5326 // (pretending that one more is running) before the CAS.
5327 // Otherwise the M from which we retake can exit the syscall,
5328 // increment nmidle and report deadlock.
5330 if atomic.Cas(&pp.status, s, _Pidle) {
5347 // Tell all goroutines that they have been preempted and they should stop.
5348 // This function is purely best-effort. It can fail to inform a goroutine if a
5349 // processor just started running it.
5350 // No locks need to be held.
5351 // Returns true if preemption request was issued to at least one goroutine.
5352 func preemptall() bool {
5354 for _, pp := range allp {
5355 if pp.status != _Prunning {
5365 // Tell the goroutine running on processor P to stop.
5366 // This function is purely best-effort. It can incorrectly fail to inform the
5367 // goroutine. It can inform the wrong goroutine. Even if it informs the
5368 // correct goroutine, that goroutine might ignore the request if it is
5369 // simultaneously executing newstack.
5370 // No lock needs to be held.
5371 // Returns true if preemption request was issued.
5372 // The actual preemption will happen at some point in the future
5373 // and will be indicated by the gp->status no longer being
5375 func preemptone(pp *p) bool {
5377 if mp == nil || mp == getg().m {
5381 if gp == nil || gp == mp.g0 {
5387 // Every call in a goroutine checks for stack overflow by
5388 // comparing the current stack pointer to gp->stackguard0.
5389 // Setting gp->stackguard0 to StackPreempt folds
5390 // preemption into the normal stack overflow check.
5391 gp.stackguard0 = stackPreempt
5393 // Request an async preemption of this P.
5394 if preemptMSupported && debug.asyncpreemptoff == 0 {
5404 func schedtrace(detailed bool) {
5411 print("SCHED ", (now-starttime)/1e6, "ms: gomaxprocs=", gomaxprocs, " idleprocs=", sched.npidle.Load(), " threads=", mcount(), " spinningthreads=", sched.nmspinning.Load(), " idlethreads=", sched.nmidle, " runqueue=", sched.runqsize)
5413 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5415 // We must be careful while reading data from P's, M's and G's.
5416 // Even if we hold schedlock, most data can be changed concurrently.
5417 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5418 for i, pp := range allp {
5420 h := atomic.Load(&pp.runqhead)
5421 t := atomic.Load(&pp.runqtail)
5423 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5429 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5431 // In non-detailed mode format lengths of per-P run queues as:
5432 // [len1 len2 len3 len4]
5438 if i == len(allp)-1 {
5449 for mp := allm; mp != nil; mp = mp.alllink {
5451 print(" M", mp.id, ": p=")
5463 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5464 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5472 forEachG(func(gp *g) {
5473 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5480 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5490 // schedEnableUser enables or disables the scheduling of user
5493 // This does not stop already running user goroutines, so the caller
5494 // should first stop the world when disabling user goroutines.
5495 func schedEnableUser(enable bool) {
5497 if sched.disable.user == !enable {
5501 sched.disable.user = !enable
5503 n := sched.disable.n
5505 globrunqputbatch(&sched.disable.runnable, n)
5507 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5515 // schedEnabled reports whether gp should be scheduled. It returns
5516 // false is scheduling of gp is disabled.
5518 // sched.lock must be held.
5519 func schedEnabled(gp *g) bool {
5520 assertLockHeld(&sched.lock)
5522 if sched.disable.user {
5523 return isSystemGoroutine(gp, true)
5528 // Put mp on midle list.
5529 // sched.lock must be held.
5530 // May run during STW, so write barriers are not allowed.
5532 //go:nowritebarrierrec
5534 assertLockHeld(&sched.lock)
5536 mp.schedlink = sched.midle
5542 // Try to get an m from midle list.
5543 // sched.lock must be held.
5544 // May run during STW, so write barriers are not allowed.
5546 //go:nowritebarrierrec
5548 assertLockHeld(&sched.lock)
5550 mp := sched.midle.ptr()
5552 sched.midle = mp.schedlink
5558 // Put gp on the global runnable queue.
5559 // sched.lock must be held.
5560 // May run during STW, so write barriers are not allowed.
5562 //go:nowritebarrierrec
5563 func globrunqput(gp *g) {
5564 assertLockHeld(&sched.lock)
5566 sched.runq.pushBack(gp)
5570 // Put gp at the head of the global runnable queue.
5571 // sched.lock must be held.
5572 // May run during STW, so write barriers are not allowed.
5574 //go:nowritebarrierrec
5575 func globrunqputhead(gp *g) {
5576 assertLockHeld(&sched.lock)
5582 // Put a batch of runnable goroutines on the global runnable queue.
5583 // This clears *batch.
5584 // sched.lock must be held.
5585 // May run during STW, so write barriers are not allowed.
5587 //go:nowritebarrierrec
5588 func globrunqputbatch(batch *gQueue, n int32) {
5589 assertLockHeld(&sched.lock)
5591 sched.runq.pushBackAll(*batch)
5596 // Try get a batch of G's from the global runnable queue.
5597 // sched.lock must be held.
5598 func globrunqget(pp *p, max int32) *g {
5599 assertLockHeld(&sched.lock)
5601 if sched.runqsize == 0 {
5605 n := sched.runqsize/gomaxprocs + 1
5606 if n > sched.runqsize {
5609 if max > 0 && n > max {
5612 if n > int32(len(pp.runq))/2 {
5613 n = int32(len(pp.runq)) / 2
5618 gp := sched.runq.pop()
5621 gp1 := sched.runq.pop()
5622 runqput(pp, gp1, false)
5627 // pMask is an atomic bitstring with one bit per P.
5630 // read returns true if P id's bit is set.
5631 func (p pMask) read(id uint32) bool {
5633 mask := uint32(1) << (id % 32)
5634 return (atomic.Load(&p[word]) & mask) != 0
5637 // set sets P id's bit.
5638 func (p pMask) set(id int32) {
5640 mask := uint32(1) << (id % 32)
5641 atomic.Or(&p[word], mask)
5644 // clear clears P id's bit.
5645 func (p pMask) clear(id int32) {
5647 mask := uint32(1) << (id % 32)
5648 atomic.And(&p[word], ^mask)
5651 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5653 // Ideally, the timer mask would be kept immediately consistent on any timer
5654 // operations. Unfortunately, updating a shared global data structure in the
5655 // timer hot path adds too much overhead in applications frequently switching
5656 // between no timers and some timers.
5658 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5659 // running P (returned by pidleget) may add a timer at any time, so its mask
5660 // must be set. An idle P (passed to pidleput) cannot add new timers while
5661 // idle, so if it has no timers at that time, its mask may be cleared.
5663 // Thus, we get the following effects on timer-stealing in findrunnable:
5665 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5666 // (for work- or timer-stealing; this is the ideal case).
5667 // - Running Ps must always be checked.
5668 // - Idle Ps whose timers are stolen must continue to be checked until they run
5669 // again, even after timer expiration.
5671 // When the P starts running again, the mask should be set, as a timer may be
5672 // added at any time.
5674 // TODO(prattmic): Additional targeted updates may improve the above cases.
5675 // e.g., updating the mask when stealing a timer.
5676 func updateTimerPMask(pp *p) {
5677 if atomic.Load(&pp.numTimers) > 0 {
5681 // Looks like there are no timers, however another P may transiently
5682 // decrement numTimers when handling a timerModified timer in
5683 // checkTimers. We must take timersLock to serialize with these changes.
5684 lock(&pp.timersLock)
5685 if atomic.Load(&pp.numTimers) == 0 {
5686 timerpMask.clear(pp.id)
5688 unlock(&pp.timersLock)
5691 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5692 // to nanotime or zero. Returns now or the current time if now was zero.
5694 // This releases ownership of p. Once sched.lock is released it is no longer
5697 // sched.lock must be held.
5699 // May run during STW, so write barriers are not allowed.
5701 //go:nowritebarrierrec
5702 func pidleput(pp *p, now int64) int64 {
5703 assertLockHeld(&sched.lock)
5706 throw("pidleput: P has non-empty run queue")
5711 updateTimerPMask(pp) // clear if there are no timers.
5712 idlepMask.set(pp.id)
5713 pp.link = sched.pidle
5716 if !pp.limiterEvent.start(limiterEventIdle, now) {
5717 throw("must be able to track idle limiter event")
5722 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
5724 // sched.lock must be held.
5726 // May run during STW, so write barriers are not allowed.
5728 //go:nowritebarrierrec
5729 func pidleget(now int64) (*p, int64) {
5730 assertLockHeld(&sched.lock)
5732 pp := sched.pidle.ptr()
5734 // Timer may get added at any time now.
5738 timerpMask.set(pp.id)
5739 idlepMask.clear(pp.id)
5740 sched.pidle = pp.link
5741 sched.npidle.Add(-1)
5742 pp.limiterEvent.stop(limiterEventIdle, now)
5747 // runqempty reports whether pp has no Gs on its local run queue.
5748 // It never returns true spuriously.
5749 func runqempty(pp *p) bool {
5750 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
5751 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
5752 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
5753 // does not mean the queue is empty.
5755 head := atomic.Load(&pp.runqhead)
5756 tail := atomic.Load(&pp.runqtail)
5757 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
5758 if tail == atomic.Load(&pp.runqtail) {
5759 return head == tail && runnext == 0
5764 // To shake out latent assumptions about scheduling order,
5765 // we introduce some randomness into scheduling decisions
5766 // when running with the race detector.
5767 // The need for this was made obvious by changing the
5768 // (deterministic) scheduling order in Go 1.5 and breaking
5769 // many poorly-written tests.
5770 // With the randomness here, as long as the tests pass
5771 // consistently with -race, they shouldn't have latent scheduling
5773 const randomizeScheduler = raceenabled
5775 // runqput tries to put g on the local runnable queue.
5776 // If next is false, runqput adds g to the tail of the runnable queue.
5777 // If next is true, runqput puts g in the pp.runnext slot.
5778 // If the run queue is full, runnext puts g on the global queue.
5779 // Executed only by the owner P.
5780 func runqput(pp *p, gp *g, next bool) {
5781 if randomizeScheduler && next && fastrandn(2) == 0 {
5787 oldnext := pp.runnext
5788 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
5794 // Kick the old runnext out to the regular run queue.
5799 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
5801 if t-h < uint32(len(pp.runq)) {
5802 pp.runq[t%uint32(len(pp.runq))].set(gp)
5803 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
5806 if runqputslow(pp, gp, h, t) {
5809 // the queue is not full, now the put above must succeed
5813 // Put g and a batch of work from local runnable queue on global queue.
5814 // Executed only by the owner P.
5815 func runqputslow(pp *p, gp *g, h, t uint32) bool {
5816 var batch [len(pp.runq)/2 + 1]*g
5818 // First, grab a batch from local queue.
5821 if n != uint32(len(pp.runq)/2) {
5822 throw("runqputslow: queue is not full")
5824 for i := uint32(0); i < n; i++ {
5825 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
5827 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
5832 if randomizeScheduler {
5833 for i := uint32(1); i <= n; i++ {
5834 j := fastrandn(i + 1)
5835 batch[i], batch[j] = batch[j], batch[i]
5839 // Link the goroutines.
5840 for i := uint32(0); i < n; i++ {
5841 batch[i].schedlink.set(batch[i+1])
5844 q.head.set(batch[0])
5845 q.tail.set(batch[n])
5847 // Now put the batch on global queue.
5849 globrunqputbatch(&q, int32(n+1))
5854 // runqputbatch tries to put all the G's on q on the local runnable queue.
5855 // If the queue is full, they are put on the global queue; in that case
5856 // this will temporarily acquire the scheduler lock.
5857 // Executed only by the owner P.
5858 func runqputbatch(pp *p, q *gQueue, qsize int) {
5859 h := atomic.LoadAcq(&pp.runqhead)
5862 for !q.empty() && t-h < uint32(len(pp.runq)) {
5864 pp.runq[t%uint32(len(pp.runq))].set(gp)
5870 if randomizeScheduler {
5871 off := func(o uint32) uint32 {
5872 return (pp.runqtail + o) % uint32(len(pp.runq))
5874 for i := uint32(1); i < n; i++ {
5875 j := fastrandn(i + 1)
5876 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
5880 atomic.StoreRel(&pp.runqtail, t)
5883 globrunqputbatch(q, int32(qsize))
5888 // Get g from local runnable queue.
5889 // If inheritTime is true, gp should inherit the remaining time in the
5890 // current time slice. Otherwise, it should start a new time slice.
5891 // Executed only by the owner P.
5892 func runqget(pp *p) (gp *g, inheritTime bool) {
5893 // If there's a runnext, it's the next G to run.
5895 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
5896 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
5897 // Hence, there's no need to retry this CAS if it falls.
5898 if next != 0 && pp.runnext.cas(next, 0) {
5899 return next.ptr(), true
5903 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
5908 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
5909 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
5915 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
5916 // Executed only by the owner P.
5917 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
5918 oldNext := pp.runnext
5919 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
5920 drainQ.pushBack(oldNext.ptr())
5925 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
5931 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
5935 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
5939 // We've inverted the order in which it gets G's from the local P's runnable queue
5940 // and then advances the head pointer because we don't want to mess up the statuses of G's
5941 // while runqdrain() and runqsteal() are running in parallel.
5942 // Thus we should advance the head pointer before draining the local P into a gQueue,
5943 // so that we can update any gp.schedlink only after we take the full ownership of G,
5944 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
5945 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
5946 for i := uint32(0); i < qn; i++ {
5947 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
5954 // Grabs a batch of goroutines from pp's runnable queue into batch.
5955 // Batch is a ring buffer starting at batchHead.
5956 // Returns number of grabbed goroutines.
5957 // Can be executed by any P.
5958 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
5960 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
5961 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
5966 // Try to steal from pp.runnext.
5967 if next := pp.runnext; next != 0 {
5968 if pp.status == _Prunning {
5969 // Sleep to ensure that pp isn't about to run the g
5970 // we are about to steal.
5971 // The important use case here is when the g running
5972 // on pp ready()s another g and then almost
5973 // immediately blocks. Instead of stealing runnext
5974 // in this window, back off to give pp a chance to
5975 // schedule runnext. This will avoid thrashing gs
5976 // between different Ps.
5977 // A sync chan send/recv takes ~50ns as of time of
5978 // writing, so 3us gives ~50x overshoot.
5979 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
5982 // On some platforms system timer granularity is
5983 // 1-15ms, which is way too much for this
5984 // optimization. So just yield.
5988 if !pp.runnext.cas(next, 0) {
5991 batch[batchHead%uint32(len(batch))] = next
5997 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6000 for i := uint32(0); i < n; i++ {
6001 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6002 batch[(batchHead+i)%uint32(len(batch))] = g
6004 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6010 // Steal half of elements from local runnable queue of p2
6011 // and put onto local runnable queue of p.
6012 // Returns one of the stolen elements (or nil if failed).
6013 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6015 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6020 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6024 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6025 if t-h+n >= uint32(len(pp.runq)) {
6026 throw("runqsteal: runq overflow")
6028 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6032 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6033 // be on one gQueue or gList at a time.
6034 type gQueue struct {
6039 // empty reports whether q is empty.
6040 func (q *gQueue) empty() bool {
6044 // push adds gp to the head of q.
6045 func (q *gQueue) push(gp *g) {
6046 gp.schedlink = q.head
6053 // pushBack adds gp to the tail of q.
6054 func (q *gQueue) pushBack(gp *g) {
6057 q.tail.ptr().schedlink.set(gp)
6064 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6066 func (q *gQueue) pushBackAll(q2 gQueue) {
6070 q2.tail.ptr().schedlink = 0
6072 q.tail.ptr().schedlink = q2.head
6079 // pop removes and returns the head of queue q. It returns nil if
6081 func (q *gQueue) pop() *g {
6084 q.head = gp.schedlink
6092 // popList takes all Gs in q and returns them as a gList.
6093 func (q *gQueue) popList() gList {
6094 stack := gList{q.head}
6099 // A gList is a list of Gs linked through g.schedlink. A G can only be
6100 // on one gQueue or gList at a time.
6105 // empty reports whether l is empty.
6106 func (l *gList) empty() bool {
6110 // push adds gp to the head of l.
6111 func (l *gList) push(gp *g) {
6112 gp.schedlink = l.head
6116 // pushAll prepends all Gs in q to l.
6117 func (l *gList) pushAll(q gQueue) {
6119 q.tail.ptr().schedlink = l.head
6124 // pop removes and returns the head of l. If l is empty, it returns nil.
6125 func (l *gList) pop() *g {
6128 l.head = gp.schedlink
6133 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6134 func setMaxThreads(in int) (out int) {
6136 out = int(sched.maxmcount)
6137 if in > 0x7fffffff { // MaxInt32
6138 sched.maxmcount = 0x7fffffff
6140 sched.maxmcount = int32(in)
6148 func procPin() int {
6153 return int(mp.p.ptr().id)
6162 //go:linkname sync_runtime_procPin sync.runtime_procPin
6164 func sync_runtime_procPin() int {
6168 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6170 func sync_runtime_procUnpin() {
6174 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6176 func sync_atomic_runtime_procPin() int {
6180 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6182 func sync_atomic_runtime_procUnpin() {
6186 // Active spinning for sync.Mutex.
6188 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6190 func sync_runtime_canSpin(i int) bool {
6191 // sync.Mutex is cooperative, so we are conservative with spinning.
6192 // Spin only few times and only if running on a multicore machine and
6193 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6194 // As opposed to runtime mutex we don't do passive spinning here,
6195 // because there can be work on global runq or on other Ps.
6196 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6199 if p := getg().m.p.ptr(); !runqempty(p) {
6205 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6207 func sync_runtime_doSpin() {
6208 procyield(active_spin_cnt)
6211 var stealOrder randomOrder
6213 // randomOrder/randomEnum are helper types for randomized work stealing.
6214 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6215 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6216 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6217 type randomOrder struct {
6222 type randomEnum struct {
6229 func (ord *randomOrder) reset(count uint32) {
6231 ord.coprimes = ord.coprimes[:0]
6232 for i := uint32(1); i <= count; i++ {
6233 if gcd(i, count) == 1 {
6234 ord.coprimes = append(ord.coprimes, i)
6239 func (ord *randomOrder) start(i uint32) randomEnum {
6243 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6247 func (enum *randomEnum) done() bool {
6248 return enum.i == enum.count
6251 func (enum *randomEnum) next() {
6253 enum.pos = (enum.pos + enum.inc) % enum.count
6256 func (enum *randomEnum) position() uint32 {
6260 func gcd(a, b uint32) uint32 {
6267 // An initTask represents the set of initializations that need to be done for a package.
6268 // Keep in sync with ../../test/initempty.go:initTask
6269 type initTask struct {
6270 // TODO: pack the first 3 fields more tightly?
6271 state uintptr // 0 = uninitialized, 1 = in progress, 2 = done
6274 // followed by ndeps instances of an *initTask, one per package depended on
6275 // followed by nfns pcs, one per init function to run
6278 // inittrace stores statistics for init functions which are
6279 // updated by malloc and newproc when active is true.
6280 var inittrace tracestat
6282 type tracestat struct {
6283 active bool // init tracing activation status
6284 id uint64 // init goroutine id
6285 allocs uint64 // heap allocations
6286 bytes uint64 // heap allocated bytes
6289 func doInit(t *initTask) {
6291 case 2: // fully initialized
6293 case 1: // initialization in progress
6294 throw("recursive call during initialization - linker skew")
6295 default: // not initialized yet
6296 t.state = 1 // initialization in progress
6298 for i := uintptr(0); i < t.ndeps; i++ {
6299 p := add(unsafe.Pointer(t), (3+i)*goarch.PtrSize)
6300 t2 := *(**initTask)(p)
6305 t.state = 2 // initialization done
6314 if inittrace.active {
6316 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6320 firstFunc := add(unsafe.Pointer(t), (3+t.ndeps)*goarch.PtrSize)
6321 for i := uintptr(0); i < t.nfns; i++ {
6322 p := add(firstFunc, i*goarch.PtrSize)
6323 f := *(*func())(unsafe.Pointer(&p))
6327 if inittrace.active {
6329 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6332 f := *(*func())(unsafe.Pointer(&firstFunc))
6333 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6336 print("init ", pkg, " @")
6337 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6338 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6339 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6340 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6344 t.state = 2 // initialization done